Ninth

Edition

BUSINESS FORECASTING

John E. Hanke
Eastern Washington University, Emeritus

Dean W. Wichern
Texas A&M University, Emeritus

UPPER SADDLE RIVER, NEW JERSEY 07458

Library of Congress Cataloging-in-Publication Data
Hanke, John E., 1940–
Business forecasting / John E. Hanke, Dean W. Wichern. – 9th ed.
p. cm.
ISBN-13: 978-0-13-230120-6 (hardcover)
ISBN-10: 0-13-230120-2 (hardcover)
1. Business forecasting. I. Wichern, Dean W. II. Title.
HD30.27.H37 2009
658.4’0355—dc22
2007044011

AVP/Executive Editor: Mark Pfaltzgraff

AVP/Editor-in-Chief: Eric Svendsen
Manager, Product Development: Ashley Santora
Permissions Manager: Charles Morris
Editorial Project Manager: Susie Abraham
Editorial Assistant: Vanessa Bain
Marketing Manager: Anne Howard
Marketing Assistant: Susan Osterlitz
Senior Managing Editor: Judy Leale
Production Project Manager: Debbie Ryan
Sr. Operations Supervisor: Arnold Vila
Operations Specialist: Michelle Klein
Cover Design: Jayne Conte
Designer: Kiwi Design
Cover Illustration/Photo: Getty Images, Inc.
Composition Full-Service Project Management: GGS Book Services
Printer/Binder: Courier/Westford
Typeface: Times Ten-Roman, 10pt.

Credits and acknowledgments borrowed from other sources and reproduced, with permission, in this textbook
appear on appropriate page within text.

Microsoft® and Windows® are registered trademarks of the Microsoft Corporation in the U.S.A. and other
countries. Screen shots and icons reprinted with permission from the Microsoft Corporation. This book is not
sponsored or endorsed by or affiliated with the Microsoft Corporation.

Portions of the input and output contained in this publication/book are printed with permission of Minitab Inc.
All material remains the exclusive property and copyright of Minitab Inc. All rights reserved.

Copyright © 2009, 2005, 2001, 1998, 1995 by Pearson Education, Inc., Upper Saddle River, New Jersey 07458.
Pearson Prentice Hall. All rights reserved. Printed in the United States of America. This publication is protected
by Copyright and permission should be obtained from the publisher prior to any prohibited reproduction, storage
in a retrieval system, or transmission in any form or by any means, electronic, mechanical, photocopying,
recording, or likewise. For information regarding permission(s), write to: Rights and Permissions Department.

Pearson Prentice Hall™ is a trademark of Pearson Education, Inc.

Pearson® is a registered trademark of Pearson plc
Prentice Hall® is a registered trademark of Pearson Education, Inc.

Pearson Education LTD. Pearson Education Australia PTY, Limited

Pearson Education Singapore, Pte. Ltd Pearson Education North Asia Ltd
Pearson Education, Canada, Ltd Pearson Educación de Mexico, S.A. de C.V.
Pearson Education–Japan Pearson Education Malaysia, Pte. Ltd.

10 9 8 7 6 5 4 3 2 1
ISBN-13: 978-0-13-230120-6
ISBN-10: 0-13-230120-2
 Dedicated to the
memory of Harry
(who really didn’t want to read it);

Geri
(who doesn’t need to);

Donna and Dorothy

(who probably won’t);

Katrina, Michael, and Andrew

(who might);

Kevin
(who says he did);

and all of our students

(who better).
This page intentionally left blank
Brief Contents

Preface xv

CHAPTER 1 Introduction to Forecasting 1

CHAPTER 2 A Review of Basic Statistical Concepts 15
CHAPTER 3 Exploring Data Patterns and an Introduction to
Forecasting Techniques 61
CHAPTER 4 Moving Averages and Smoothing Methods 107
CHAPTER 5 Time Series and Their Components 165
CHAPTER 6 Simple Linear Regression 221
CHAPTER 7 Multiple Regression Analysis 281
CHAPTER 8 Regression with Time Series Data 339
CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 399
CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 481
CHAPTER 11 Managing the Forecasting Process 503
APPENDIX A Data for Case 7-1 521
APPENDIX B Tables 523
APPENDIX C Data Sets and Databases 533

Index 547

v
This page intentionally left blank
Contents

Preface xv

CHAPTER 1 Introduction to Forecasting 1

The History of Forecasting 1
Is Forecasting Necessary? 2
Types of Forecasts 2
Macroeconomic Forecasting Considerations 3
Choosing a Forecasting Method 4
Forecasting Steps 4
Managing the Forecasting Process 6
Forecasting Software 6
Online Information 7
Forecasting Examples 7
Summary 9
Case 1-1: Mr. Tux 10
Case 1-2: Consumer Credit Counseling 10
Minitab Applications 11
Excel Applications 12
References 12

CHAPTER 2 A Review of Basic Statistical Concepts 15

Describing Data with Numerical Summaries 15
Displays of Numerical Information 19
Probability Distributions 22
Sampling Distributions 26
Inference from a Sample 29
Estimation 29
Hypothesis Testing 30
p-Value 32

vii
viii Contents

Correlation Analysis 34
Scatter Diagrams 34
Correlation Coefficient 37
Fitting a Straight Line 39
Assessing Normality 42
Application to Management 44
Glossary 44
Key Formulas 45
Problems 46
Case 2-1: Alcam Electronics 53
Case 2-2: Mr. Tux 54
Case 2-3: Alomega Food Stores 56
Minitab Applications 56
Excel Applications 58
References 60

CHAPTER 3 Exploring Data Patterns and an Introduction

to Forecasting Techniques 61
Exploring Time Series Data Patterns 62
Exploring Data Patterns with Autocorrelation Analysis 64
Are the Data Random? 69
Do the Data Have a Trend? 72
Are the Data Seasonal? 76
Choosing a Forecasting Technique 76
Forecasting Techniques for Stationary Data 78
Forecasting Techniques for Data with a Trend 78
Forecasting Techniques for Seasonal Data 79
Forecasting Techniques for Cyclical Series 79
Other Factors to Consider When Choosing a Forecasting Technique 79
Empirical Evaluation of Forecasting Methods 81
Measuring Forecast Error 81
Determining the Adequacy of a Forecasting Technique 84
Application to Management 86
Glossary 87
Key Formulas 87
Problems 88
Case 3-1A: Murphy Brothers Furniture 94
Case 3-1B: Murphy Brothers Furniture 96
Case 3-2: Mr. Tux 97
Case 3-3: Consumer Credit Counseling 98
Case 3-4: Alomega Food Stores 99
 Contents ix
Case 3-5: Surtido Cookies 100
Minitab Applications 101
Excel Applications 103
References 105

CHAPTER 4 Moving Averages and Smoothing Methods 107

Naive Models 108
Forecasting Methods Based on Averaging 111
Simple Averages 111
Moving Averages 113
Double Moving Averages 116
Exponential Smoothing Methods 119
Exponential Smoothing Adjusted for Trend: Holt’s Method 126
Exponential Smoothing Adjusted for Trend and Seasonal
Variation: Winter’s Method 130
Application to Management 135
Glossary 136
Key Formulas 136
Problems 138
Case 4-1: The Solar Alternative Company 145
Case 4-2: Mr. Tux 147
Case 4-3: Consumer Credit Counseling 148
Case 4-4: Murphy Brothers Furniture 148
Case 4-5: Five-Year Revenue Projection for Downtown Radiology 149
Case 4-6: Web Retailer 154
Case 4-7: Southwest Medical Center 158
Case 4-8: Surtido Cookies 159
Minitab Applications 159
Excel Applications 161
References 163

CHAPTER 5 Time Series and Their Components 165

Decomposition 166
Trend 168
Additional Trend Curves 171
Forecasting Trend 174
Seasonality 175
Seasonally Adjusted Data 179
Cyclical and Irregular Variations 180
Summary Example 180
Business Indicators 184
x Contents

Forecasting a Seasonal Time Series 185

The Census II Decomposition Method 187
Application to Management 189
Appendix: Price Index 190
Glossary 192
Key Formulas 192
Problems 193
Case 5-1: The Small Engine Doctor 201
Case 5-2: Mr. Tux 202
Case 5-3: Consumer Credit Counseling 206
Case 5-4: Murphy Brothers Furniture 207
Case 5-5: AAA Washington 210
Case 5-6: Alomega Food Stores 212
Case 5-7: Surtido Cookies 213
Case 5-8: Southwest Medical Center 214
Minitab Applications 214
Excel Applications 217
References 219

CHAPTER 6 Simple Linear Regression 221

Regression Line 222
Standard Error of the Estimate 226
Forecasting Y 227
Decomposition of Variance 230
Coefficient of Determination 234
Hypothesis Testing 236
Analysis of Residuals 239
Computer Output 241
Variable Transformations 243
Growth Curves 246
Application to Management 250
Glossary 252
Key Formulas 253
Problems 254
Case 6-1: Tiger Transport 266
Case 6-2: Butcher Products, Inc. 268
Case 6-3: Ace Manufacturing 269
Case 6-4: Mr. Tux 270
Case 6-5: Consumer Credit Counseling 270
Case 6-6: AAA Washington 271
Minitab Applications 274
 Contents xi
Excel Applications 277
References 279

CHAPTER 7 Multiple Regression Analysis 281

Several Predictor Variables 281
Correlation Matrix 282
Multiple Regression Model 283
Statistical Model for Multiple Regression 283
Interpreting Regression Coefficients 285
Inference for Multiple Regression Models 286
Standard Error of the Estimate 287
Significance of the Regression 288
Individual Predictor Variables 290
Forecast of a Future Response 291
Computer Output 292
Dummy Variables 293
Multicollinearity 297
Selecting the “Best” Regression Equation 300
All Possible Regressions 302
Stepwise Regression 304
Final Notes on Stepwise Regression 306
Regression Diagnostics and Residual Analysis 307
Forecasting Caveats 309
Overfitting 309
Useful Regression, Large F Ratios 310
Application to Management 310
Glossary 312
Key Formulas 312
Problems 313
Case 7-1: The Bond Market 324
Case 7-2: AAA Washington 328
Case 7-3: Fantasy Baseball (A) 330
Case 7-4: Fantasy Baseball (B) 334
Minitab Applications 336
Excel Applications 337
References 338

CHAPTER 8 Regression with Time Series Data 339

Time Series Data and the Problem of Autocorrelation 339
Autocorrelation and the Durbin-Watson Test 343
Solutions to Autocorrelation Problems 347
xii Contents

Model Specification Error (Omitting a Variable) 348

Regression with Differences 350
Autocorrelated Errors and Generalized Differences 354
Autoregressive Models 357
Summary 358
Time Series Data and the Problem of Heteroscedasticity 358
Using Regression to Forecast Seasonal Data 361
Econometric Forecasting 364
Cointegrated Time Series 365
Application to Management 367
Glossary 367
Key Formulas 367
Problems 369
Case 8-1: Company of Your Choice 378
Case 8-2: Business Activity Index for Spokane County 379
Case 8-3: Restaurant Sales 383
Case 8-4: Mr. Tux 385
Case 8-5: Consumer Credit Counseling 388
Case 8-6: AAA Washington 389
Case 8-7: Alomega Food Stores 392
Case 8-8: Surtido Cookies 393
Case 8-9: Southwest Medical Center 394
Minitab Applications 395
Excel Applications 396
References 398

CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 399

Box-Jenkins Methodology 399
Autoregressive Models 404
Moving Average Models 405
Autoregressive Moving Average Models 407
Summary 407
Implementing the Model-Building Strategy 407
Step 1: Model Identification 407
Step 2: Model Estimation 409
Step 3: Model Checking 410
Step 4: Forecasting with the Model 411
Model-Building Caveats 430
Model Selection Criteria 431
ARIMA Models for Seasonal Data 432
 Contents xiii
Simple Exponential Smoothing and an ARIMA Model 442
Advantages and Disadvantages of ARIMA Models 443
Application to Management 444
Glossary 445
Key Formulas 445
Problems 446
Case 9-1: Restaurant Sales 457
Case 9-2: Mr. Tux 459
Case 9-3: Consumer Credit Counseling 460
Case 9-4: The Lydia E. Pinkham Medicine Company 461
Case 9-5: City of College Station 463
Case 9-6: UPS Air Finance Division 466
Case 9-7: AAA Washington 469
Case 9-8: Web Retailer 471
Case 9-9: Surtido Cookies 474
Case 9-10: Southwest Medical Center 476
Minitab Applications 478
References 480

CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 481

Judgmental Forecasting 483
The Delphi Method 483
Scenario Writing 485
Combining Forecasts 486
Forecasting and Neural Networks 488
Summary of Judgmental Forecasting 490
Other Tools Useful in Making Judgments About the Future 491
Key Formulas 496
Problems 496
Case 10-1: Golden Gardens Restaurant 497
Case 10-2: Alomega Food Stores 497
Case 10-3: The Lydia E. Pinkham Medicine Company 498
References 501

CHAPTER 11 Managing the Forecasting Process 503

The Forecasting Process 503
Monitoring Forecasts 504
Forecasting Steps Reviewed 509
Forecasting Responsibility 510
Forecasting Costs 511
xiv Contents

Forecasting and Management Information Systems 511

Selling Management on Forecasting 512
The Future of Forecasting 512
Problems 513
Case 11-1: Boundary Electronics 513
Case 11-2: Busby Associates 514
Case 11-3: Consumer Credit Counseling 517
Case 11-4: Mr. Tux 518
Case 11-5: Alomega Food Stores 519
Case 11-6: Southwest Medical Center 520
References 520
APPENDIX A Data for Case 7-1 521
APPENDIX B Tables 523
Table B-1 Individual Terms of the Binomial Distribution 523
Table B-2 Areas for Standard Normal Probability Distribution 525
Table B-3 Critical Values of t 526
Table B-4 Critical Values of Chi-Square 527
Table B-5 F Distribution 529
Table B-6 Durbin-Watson Test Bounds 530
APPENDIX C Data Sets and Databases 533
Index 547
 Preface

The goal of the ninth edition of Business Forecasting remains the same as that of the
previous editions: To present the basic statistical techniques that are useful for prepar-
ing individual business forecasts and long-range plans. The book is written in a simple
straightforward style and makes extensive use of practical business examples. Sixty-
five cases appear at the end of chapters to provide the student with the necessary link
between theoretical concepts and their real-world applications. The emphasis is on the
application of techniques by management for decision making. Students are assumed
to have taken an introductory course in statistics and to be comfortable with using the
computer to access software packages such as word processors and spreadsheets.

ORGANIZATION

All chapters have been revised to enhance the clarity of the writing and to increase
teaching and learning effectiveness. The content has been organized into six sections.
The first section (Chapters 1 and 2) presents background material. The nature of
forecasting and a quick review of basic statistical concepts set the stage for the cover-
age of techniques that begins in the second section.
The second section (Chapter 3) emphasizes the exploration of data patterns and
the choosing of a forecasting technique. The third section (Chapters 4 and 5) covers
averaging, smoothing techniques, and an introduction to time series decomposition in
terms of underlying components. The fourth section (Chapters 6 and 7) emphasizes
causal forecasting techniques, such as correlation, regression, and multiple regression
analysis.
The fifth section (Chapters 8 and 9) looks at techniques used to forecast time series
data. The book concludes with a final section (Chapters 10 and 11) on judgmental fore-
casting and forecast adjustments, along with a discussion of managing and monitoring
the forecasting process.

CHANGES IN THE NINTH EDITION

The ninth edition has been largely rewritten. Although the “flavor” of earlier editions
has been retained, added emphasis has been placed on the most recent theoretical
developments and empirical findings. Outdated material has been eliminated and
the book has been completely reorganized with the addition of new problems, examples,
data sets, and cases.
The following features are either new or improved in this edition:
• Twelve new cases have been added.
• Thirty-two new problems have been added.

xv
xvi Preface

• Data sets have been updated.

• A section on Cointegrated Time Series has been added to Chapter 8.
• An example of how to use Minitab 15 in a forecasting situation is demonstrated at
the end of most chapters.
• An example of how to use Excel 2003 to solve a forecasting problem is demon-
strated at the end of most chapters.
• A downloadable version of the Instructor’s Manual has been made available by
visiting www.prenhall.com/hanke
• Data sets for Minitab, Excel, or other programs have been made available on the
Internet.

THE ROLE OF THE COMPUTER

In the first eight editions, the computer was recognized as a powerful tool in fore-
casting. The computer is even more important now with the availability of powerful
forecasting software and easy access to data via networking capabilities and the
Internet.
A nationwide research study of all AACSB member institutions conducted by
the authors to determine what faculty do about using computers for teaching fore-
casting showed that (1) most forecasting faculty (94.2%) attempt to provide
students with hands-on experience in using the computer, and (2) several statistical
packages and specific personal computer forecasting packages were mentioned in
the survey. The packages mentioned most frequently were Minitab, SAS, and
spreadsheets.
The authors have tried several different approaches to help faculty and students
use the computer for forecasting. This edition features the following:
1. Minitab instructions presented at the end of most chapters.
2. Excel instructions presented at the end of most chapters.
3. Three data collections available on the Internet (Minitab, Excel, other programs).
Each collection contains data from the text examples and problems. Each collec-
tion also contains several new data series. To access the data sets on the Internet go
to the Prentice Hall Web site at www.prenhall.com/hanke
4. Examples of different computer outputs are placed throughout the text.

ACKNOWLEDGMENTS

The authors are indebted to the many instructors around the world who have used the
first eight editions and have provided invaluable suggestions for improving the book.
Special thanks go to: Professor Frank Forest (Marquette University); Professor
William Darrow (Townsend State University); Susan Winters (Northwestern State
University); Professor Shik Chun Young (Eastern Washington University); and Mark
Craze, Judy Johnson, Steve Brandon, and Dorothy Mercer for providing cases.
Portions of this text, particularly some data sets, are adapted from those that
appeared in the second edition of Understanding Business Statistics by Hanke and
Reitsch, published by Richard D. Irwin, Inc., whom we here credit for this reuse.
We also thank reviewers John Liechty, University of Michigan; John Tamura,
University of Washington; and Ted Taukahara, St. Mary’s University for their construc-
tive comments on previous editions of this book. Other reviewers deserving of our
 Preface xvii
thanks are Perry Sadorski, York University; Shady Kholdy, California State
Polytechnic University-Pomona; Michael Niemira, New York University; Fred
Zufryden, University of Southern California; and Haizheg Li, Georgia Institute of
Technology.
Finally, we owe a large debt to desktop computers and word processing software.
Modern technology has made textbook writing considerably easier. However, we, not
the computers, are responsible for any errors.
This page intentionally left blank
CHAPTER

1 INTRODUCTION TO
FORECASTING
This book is concerned with methods used to predict the uncertain nature of business
trends in an effort to help managers make better decisions and plans. Such efforts often
involve the study of historical data and the manipulation of these data in the search for
patterns that can be effectively extrapolated to produce forecasts.
In this text, we regularly remind readers that sound judgment must be used along
with numerical results if good forecasting is to result.The examples in this chapter and the
cases at the end of this and the remaining text chapters emphasize this point. There are
more discussions of the role of judgment in this chapter and in the concluding chapter.

THE HISTORY OF FORECASTING

In a book on the history of risk, author Peter Bernstein (1996) notes that the development
of business forecasting in the seventeenth century was a major innovation. He writes:

forecasting—long denigrated as a waste of time at best and a sin at worst—

became an absolute necessity in the course of the seventeenth century for
adventuresome entrepreneurs who were willing to take the risk of shaping the
future according to their own design. (p. 95)

Over the next 300 years, significant advances in data-based forecasting methods
occurred, with much of the development coming in the twentieth century. Regression
analysis, decomposition, smoothing, and autoregressive moving average methods are
examples of data-based forecasting procedures discussed in this book. These proce-
dures have proved to be highly effective and routinely appear in the menus of readily
available forecasting software.
Along with the development of data-based methods, the role of judgment and
judgmental approaches to forecasting has grown significantly over the last 25 years.
Without any data history, human judgment may be the only way to make predictions
about the future. In cases where data are available, judgment should be used to review
and perhaps modify the forecasts produced by quantitative procedures. Although the
majority of this book is devoted to data-based forecasting methods, there is a discus-
sion of judgmental forecasting methods in Chapter 10.
With the proliferation of powerful personal computers and the availability of
sophisticated software packages, forecasts of future values for variables of interest are
easily generated. However, this ease of computation does not replace clear thinking.
Lack of managerial oversight and improper use of forecasting techniques can lead to
costly decisions.
1
2 CHAPTER 1 Introduction to Forecasting

New forecasting procedures continue to be developed as the need for accurate

forecasts accelerates.1 Particular attention is being paid to the forecasting process
in organizations with the need to coordinate objectives, methods, assessment, and
interpretation.

IS FORECASTING NECESSARY?
In spite of the inherent inaccuracies in trying to predict the future, forecasts necessar-
ily drive policy setting and planning. How can the Federal Reserve Board realistically
adjust interest rates without some notion of future economic growth and inflationary
pressures? How can an operations manager realistically set production schedules with-
out some estimate of future sales? How can a company determine staffing for its call
centers without some guess of the future demand for service? How can a bank make
realistic plans without some forecast of future deposits and loan balances? Everyone
requires forecasts. The need for forecasts cuts across all functional lines as well as all
types of organizations. Forecasts are absolutely necessary to move forward in today’s
ever-changing and highly interactive business environment.
This book discusses various ways of generating forecasts that rely on logical meth-
ods of manipulating data that have been generated by historical events. But it is our
belief that the most effective forecaster is able to formulate a skillful mix of quantita-
tive forecasting and good judgment and to avoid the extremes of total reliance on
either. At one extreme, we find the executive who, through ignorance and fear of quan-
titative techniques and computers, relies solely on intuition and feel. At the other
extreme is the forecaster skilled in the latest sophisticated data manipulation tech-
niques but unable or unwilling to relate the forecasting process to the needs of the
organization and its decision makers. We view the quantitative forecasting techniques
discussed in most of this book to be only the starting point in the effective forecasting
of outcomes important to the organization: Analysis, judgment, common sense, and
business experience must be brought to bear on the process through which these
important techniques have generated their results.
Another passage from Bernstein (1996) effectively summarizes the role of fore-
casting in organizations.

You do not plan to ship goods across the ocean, or to assemble merchandise
for sale, or to borrow money without first trying to determine what the future
may hold in store. Ensuring that the materials you order are delivered on time,
seeing to it that the items you plan to sell are produced on schedule, and get-
ting your sales facilities in place all must be planned before that moment when
the customers show up and lay their money on the counter. The successful
business executive is a forecaster first; purchasing, producing, marketing, pric-
ing, and organizing all follow. (pp. 21–22)

TYPES OF FORECASTS
When managers are faced with the need to make decisions in an atmosphere of uncer-
tainty, what types of forecasts are available to them? Forecasting procedures might
first be classified as long term or short term. Long-term forecasts are necessary to set

1A recent review of the current state of forecasting is available in a special issue of the International Journal
of Forecasting, edited by R. J. Hyndman and J. K. Ord (2006).
 CHAPTER 1 Introduction to Forecasting 3
the general course of an organization for the long run; thus, they become the particular
focus of top management. Short-term forecasts are needed to design immediate strate-
gies and are used by middle and first-line management to meet the needs of the imme-
diate future.
Forecasts might also be classified in terms of their position on a micro–macro
continuum, that is, in terms of the extent to which they involve small details versus
large summary values. For example, a plant manager might be interested in forecasting
the number of workers needed for the next several months (a micro forecast), whereas
the federal government is forecasting the total number of people employed in the
entire country (a macro forecast). Again, different levels of management in an organi-
zation tend to focus on different levels of the micro–macro continuum. Top manage-
ment would be interested in forecasting the sales of the entire company, for example,
whereas individual salespersons would be much more interested in forecasting their
own sales volumes.
Forecasting procedures can also be classified according to whether they tend to be
more quantitative or qualitative. At one extreme, a purely qualitative technique is one
requiring no overt manipulation of data. Only the “judgment” of the forecaster is used.
Even here, of course, the forecaster’s “judgment” may actually be the result of the men-
tal manipulation of historical data. At the other extreme, purely quantitative techniques
need no input of judgment; they are mechanical procedures that produce quantitative
results. Some quantitative procedures require a much more sophisticated manipulation
of data than do others, of course. This book emphasizes the quantitative forecasting
techniques because a broader understanding of these very useful procedures is
needed in the effective management of modern organizations. However, we emphasize
again that judgment and common sense must be used along with mechanical and data-
manipulative procedures. Only in this way can intelligent forecasting take place.
Finally, forecasts might be classified according to the nature of the output. One
must decide if the forecast will be a single number best guess (a point forecast), a range
of numbers within which the future value is expected to fall (an interval forecast), or an
entire probability distribution for the future value (a density forecast). Since unpre-
dictable “shocks” will affect future values (the future is never exactly like the past),
nonzero forecast errors will occur even from very good forecasts. Thus, there is some
uncertainty associated with a particular point forecast. The uncertainty surrounding
point forecasts suggests the usefulness of an interval forecast. However, if forecasts
are solely the result of judgment, point forecasts are typically the only recourse. In
judgmental situations, it is extremely difficult to accurately describe the uncertainty
associated with the forecast.

MACROECONOMIC FORECASTING CONSIDERATIONS

We usually think of forecasting in terms of predicting important variables for an indi-

vidual company or perhaps for one component of a company. Monthly company sales,
unit sales for one of a company’s stores, and absent hours per employee per month in a
factory are examples.
By contrast, there is growing interest in forecasting important variables for the
entire economy of a country. Much work has been done in evaluating methods for
doing this kind of overall economic forecasting, called macroeconomic forecasting.
Examples of interest to the federal government of the United States are the unemploy-
ment rate, gross domestic product, and prime interest rate. Economic policy is based, in
part, on projections of important economic indicators such as these. For this reason,
4 CHAPTER 1 Introduction to Forecasting

there is great interest in improving forecasting methods that focus on overall measures
of a country’s economic performance.
One of the chief difficulties in developing accurate forecasts of overall economic
activity is the unexpected and significant shift in a key economic factor. Significant
changes in oil prices, inflation surges, and broad policy changes by a country’s govern-
ment are examples of shifts in a key factor that can affect the global economy.
The possibility of such significant shifts in the economic scene has raised a key
question in macroeconomic forecasting: Should the forecasts generated by the fore-
casting model be modified using the forecaster’s judgment? Current work on forecast-
ing methodology often involves this question.
Theoretical and practical work on macroeconomic forecasting continues.
Considering the importance of accurate economic forecasting to economic policy for-
mulation in this country and others, increased attention to this kind of forecasting can
be expected in the future. A good introductory reference for macroeconomic forecast-
ing is Pindyck and Rubinfeld (1998).

CHOOSING A FORECASTING METHOD

The preceding discussion suggests several factors to be considered in choosing a fore-
casting method. The level of detail must be considered. Are forecasts of specific details
needed (a micro forecast)? Or is the future status of some overall or summary factor
needed (a macro forecast)? Is the forecast needed for some point in the near future
(a short-term forecast) or for a point in the distant future (a long-term forecast)? To what
extent are qualitative (judgment) and quantitative (data-manipulative) methods appropri-
ate? And, finally, what form should the forecast take (point, interval, or density forecast)?
The overriding consideration in choosing a forecasting method is that the results
must facilitate the decision-making process of the organization’s managers. Rarely
does one method work for all cases. Different products (for example, new versus estab-
lished), goals (for example, simple prediction versus the need to control an important
business driver of future values), and constraints (for example, cost, required expertise,
and immediacy) must be considered when selecting a forecasting method. With the
availability of current forecasting software, it is best to think of forecasting methods as
generic tools that can be applied simultaneously. Several methods can be tried in a
given situation. The methodology producing the most accurate forecasts in one case
may not be the best methodology in another situation. However, the method(s) chosen
should produce a forecast that is accurate, timely, and understood by management so
that the forecast can help produce better decisions.
Choosing a forecasting technique is discussed in Chapter 3 and summarized
in Table 3-6. The forecasting methodologies considered in this book are summarized in
Table 11-1. These tables, along with additional discussion available in Chase (1997), can
help the forecaster select an initial set of forecasting procedures to be considered.

FORECASTING STEPS
All formal forecasting procedures involve extending the experiences of the past into
the future. Thus, they involve the assumption that the conditions that generated past
relationships and data are indistinguishable from the conditions of the future.
A human resource department is hiring employees, in part, on the basis of a com-
pany entrance examination score because, in the past, that score seemed to be an impor-
tant predictor of job performance rating. To the extent that this relation continues to
 CHAPTER 1 Introduction to Forecasting 5
hold, forecasts of future job performance—hence hiring decisions—can be improved by
using examination scores. If, for some reason, the association between examination
score and job performance changes, then forecasting job performance ratings from
examination scores using the historical model will yield inaccurate forecasts and poten-
tially poor hiring decisions. This is what makes forecasting difficult. The future is not
always like the past. To the extent it is, quantitative forecasting methods work well. To
the extent it isn’t, inaccurate forecasts can result. However, it is generally better to have
some reasonably constructed forecast than no forecast.
The recognition that forecasting techniques operate on the data generated by histor-
ical events leads to the identification of the following five steps in the forecasting process:

1. Problem formulation and data collection

2. Data manipulation and cleaning
3. Model building and evaluation
4. Model implementation (the actual forecast)
5. Forecast evaluation

In step 1, problem formulation and data collection are treated as a single step
because they are intimately related. The problem determines the appropriate data. If a
quantitative forecasting methodology is being considered, the relevant data must be
available and correct. Often accessing and assembling appropriate data is a challenging
and time-consuming task. If appropriate data are not available, the problem may have
to be redefined or a nonquantitative forecasting methodology employed. Collection
and quality control problems frequently arise whenever it becomes necessary to obtain
pertinent data for a business forecasting effort.
Step 2, data manipulation and cleaning, is often necessary. It is possible to have too
much data as well as too little in the forecasting process. Some data may not be rele-
vant to the problem. Some data may have missing values that must be estimated. Some
data may have to be reexpressed in units other than the original units. Some data may
have to be preprocessed (for example, accumulated from several sources and
summed). Other data may be appropriate but only in certain historical periods (for
example, in forecasting the sales of small cars, one may wish to use only car sales data
since the oil embargo of the 1970s rather than sales data over the past 60 years).
Ordinarily, some effort is required to get data into the form that is required for using
certain forecasting procedures.
Step 3, model building and evaluation, involves fitting the collected data into a
forecasting model that is appropriate in terms of minimizing forecasting error. The sim-
pler the model is, the better it is in terms of gaining acceptance of the forecasting
process by managers who must make the firm’s decisions. Often a balance must be
struck between a sophisticated forecasting approach that offers slightly more accuracy
and a simple approach that is easily understood and gains the support of—and is
actively used by—the company’s decision makers. Obviously, judgment is involved in
this selection process. Since this book discusses numerous forecasting models and their
applicability, the reader’s ability to exercise good judgment in the choice and use of
appropriate forecasting models will increase after studying this material.
Step 4, model implementation, is the generation of the actual model forecasts once
the appropriate data have been collected and cleaned and an appropriate forecasting
model has been chosen. Data for recent historical periods are often held back and later
used to check the accuracy of the process.
Step 5, forecast evaluation, involves comparing forecast values with actual histori-
cal values. After implementation of the forecasting model is complete, forecasts are
6 CHAPTER 1 Introduction to Forecasting

made for the most recent historical periods where data values were known but held
back from the data set being analyzed. These forecasts are then compared with the
known historical values, and any forecasting errors are analyzed. Some forecasting pro-
cedures sum the absolute values of the errors and may report this sum, or they divide
this sum by the number of forecast attempts to produce the average forecast error.
Other procedures produce the sum of squared errors, which is then compared with
similar figures from alternative forecasting methods. Some procedures also track and
report the magnitude of the error terms over the forecasting period. Examination of
error patterns often leads the analyst to modify the forecasting model. Specific meth-
ods of measuring forecasting errors are discussed near the end of Chapter 3.

MANAGING THE FORECASTING PROCESS

The discussion in this chapter serves to underline our belief that management ability
and common sense must be involved in the forecasting process. The forecaster should
be thought of as an advisor to the manager rather than as a monitor of an automatic
decision-making device. Unfortunately, the latter is sometimes the case in practice,
especially with the aura of the computer. Again, quantitative forecasting techniques
must be seen as what they really are, namely, tools to be used by the manager in arriv-
ing at better decisions. According to Makridakis (1986):

The usefulness and utility of forecasting can be improved if management

adopts a more realistic attitude. Forecasting should not be viewed as a substi-
tute for prophecy but rather as the best way of identifying and extrapolating
established patterns or relationships in order to forecast. If such an attitude is
accepted, forecasting errors must be considered inevitable and the circum-
stances that cause them investigated. (p. 33)

Given that, several key questions should always be raised if the forecasting process is
to be properly managed:
• Why is a forecast needed?
• Who will use the forecast, and what are their specific requirements?
• What level of detail or aggregation is required, and what is the proper time horizon?
• What data are available, and will the data be sufficient to generate the needed
forecast?
• What will the forecast cost?
• How accurate can we expect the forecast to be?
• Will the forecast be made in time to help the decision-making process?
• Does the forecaster clearly understand how the forecast will be used in the
organization?
• Is a feedback process available to evaluate the forecast after it is made and to
adjust the forecasting process accordingly?

FORECASTING SOFTWARE
Today, there are a large number of computer software packages specifically designed
to provide the user with various forecasting methods. Two types of computer packages
are of primary interest to forecasters: (1) general statistical packages that include
regression analysis, time series analysis, and other techniques used frequently by
 CHAPTER 1 Introduction to Forecasting 7
forecasters and (2) forecasting packages that are specifically designed for forecasting
applications. In addition, some forecasting tools are available in Enterprise Resource
Planning (ERP) systems.
Graphical capabilities, interfaces to spreadsheets and external data sources, numeri-
cally and statistically reliable methods, and simple automatic algorithms for the selection
and specification of forecasting models are now common features of business forecasting
software. However, although development and awareness of forecasting software have
increased dramatically in recent years, the majority of companies still use spreadsheets
(perhaps with add-ins) to generate forecasts and develop business plans.
Examples of stand-alone software packages with forecasting tools include
Minitab, SAS, and SPSS. There are many add-ins or supplemental programs that pro-
vide forecasting tools in a spreadsheet environment. For example, the Analysis
ToolPak add-in for Microsoft Excel provides some regression analysis and smoothing
capabilities. There are currently several more comprehensive add-ins that provide a
(almost) full range of forecasting capabilities.2
It is sometimes the case, particularly in a spreadsheet setting, that “automatic”
forecasting is available. That is, the software selects the best model or procedure for
forecasting and immediately generates forecasts. We caution, however, that this con-
venience comes at a price. Automatic procedures produce numbers but rarely provide
the forecaster with real insight into the nature and quality of the forecasts. The genera-
tion of meaningful forecasts requires human intervention, a give and take between
problem knowledge and forecasting procedures (software).
Many of the techniques in this book will be illustrated with Minitab 15 and
Microsoft Excel 2003 (with the Analysis ToolPak add-in). Minitab 15 was chosen for its
ease of use and widespread availability. Excel, although limited in its forecasting func-
tionality, is frequently the tool of choice for calculating projections.

ONLINE INFORMATION
Information of interest to forecasters is available on the World Wide Web. Perhaps the
best way to learn about what’s available in cyberspace is to spend some time searching for
whatever interests you, using a browser such as Netscape or Microsoft Internet Explorer.
Any list of websites for forecasters is likely to be outdated by the time this edition
appears; however, there are two websites that are likely to remain available for some
time. B&E DataLinks, available at www.econ-datalinks.org, is a website maintained
by the Business and Economic Statistics Section of the American Statistical Association.
This website contains extensive links to economic and financial data sources of interest
to forecasters. The second site, Resources for Economists on the Internet, sponsored
by the American Economic Association and available at rfe.org, contains an extensive set
of links to data sources, journals, professional organizations, and so forth.

FORECASTING EXAMPLES
Discussions in this chapter emphasize that forecasting requires a great deal of judg-
ment along with the mathematical manipulation of collected data. The following exam-
ples demonstrate the kind of thinking that often precedes a forecasting effort in a real

2At the time this edition was written, the Institute for Forecasting Education provided reviews of
forecasting software on its website. These reviews can be accessed at www.forecastingeducation.com/
forecastingsoftwarereviews.asp.
8 CHAPTER 1 Introduction to Forecasting

firm. Notice that the data values that will produce useful forecasts, even if they exist,
may not be apparent at the beginning of the process and may or may not be identified
as the process evolves. In other words, the initial efforts may turn out to be useless and
another approach required.
The results of the forecasting efforts for the two examples discussed here are not
shown, as they require topics that are described throughout the text. Look for the tech-
niques described in later chapters to be applied to these data. For the moment, we hope
these examples illustrate the forecasting effort that real managers face.

Example 1.1 Alomega Food Stores

Alomega Food Stores is a retail food provider with 27 stores in a midwestern state. The
company engages in various kinds of advertising and, until recently, had never studied the
effect its advertising dollars have on sales, although some data had been collected and
stored for 3 years.
The executives at Alomega decided to begin tracking their advertising efforts along
with the sales volumes for each month. Their hope was that after several months the col-
lected data could be examined to possibly reveal relationships that would help in determin-
ing future advertising expenditures.
The accounting department began extending its historical records by recording the
sales volume for each month along with the advertising dollars for both newspaper ads and
TV spots. They also recorded both sales and advertising values that had been lagged for one
and two months. This was done because some people on the executive committee thought
that sales might depend on advertising expenditures in previous months rather than in the
month the sales occurred.
The executives also believed that sales experienced a seasonal effect. For this reason, a
dummy or categorical variable was used to indicate each month. In addition, they wondered
about any trend in sales volume.
Finally, the executives believed that Alomega’s advertising dollars might have an effect
on its major competitors’ advertising budgets the following month. For each following
month, it was decided that competitors’ advertising could be classified as a (1) small amount,
(2) a moderate amount, or (3) a large amount.
After a few months of collecting data and analyzing past records, the accounting
department completed a data array for 48 months using the following variables:

• Sales dollars
• Newspaper advertising dollars
• TV advertising dollars
• Month code where January  1, February  2, through December  12
• A series of 11 dummy variables to indicate month
• Newspaper advertising lagged one month
• Newspaper advertising lagged two months
• TV advertising lagged one month
• TV advertising lagged two months
• Month number from 1 to 48
• Code 1, 2, or 3 to indicate competitors’ advertising efforts the following month

Alomega managers, especially Julie Ruth, the company president, now want to learn
anything they can from the data they have collected. In addition to learning about
the effects of advertising on sales volumes and competitors’ advertising, Julie wonders
about any trend and the effect of season on sales. However, the company’s production
manager, Jackson Tilson, does not share her enthusiasm. At the end of the forecasting
planning meeting, he makes the following statement: “I’ve been trying to keep my
mouth shut during this meeting, but this is really too much. I think we’re wasting a lot of
people’s time with all this data collection and fooling around with computers. All you have
to do is talk with our people on the floor and with the grocery store managers to understand
what’s going on. I’ve seen this happen around here before, and here we go again. Some
of you people need to turn off your computers, get out of your fancy offices, and talk with a
few real people.”
 CHAPTER 1 Introduction to Forecasting 9
Example 1.2 Large Web-based Retailer
One of the goals of a large Internet-based retailer is to be the world’s most consumer-
centric company. The company recognizes that the ability to establish and maintain long-
term relationships with customers and to encourage repeat visits and purchases depends, in
part, on the strength of its customer service operations. For service matters that cannot be
handled using website features, customer service representatives located in contact centers
are available 24 hours a day to field voice calls and emails.
Because of its growing sales and its seasonality (service volume is relatively low in the
summer and high near the end of the year), a challenge for the company is to appropriately
staff its contact centers. The planning problem involves making decisions about hiring and
training at internally managed centers and about allocating work to outsourcers based on
the volume of voice calls and emails. The handling of each contact type must meet a targeted
service level every week.
To make the problem even more difficult, the handling time for each voice call and
email is affected by a number of contact attributes, including type of product, customer, and
purchase type. These attributes are used to classify the contacts into categories: in this case,
one “primary” category and seven “specialty” categories. Specific skill sets are needed to
resolve the different kinds of issues that arise in the various categories. Since hiring and
training require a 6-week lead time, forecasts of service contacts are necessary in order to
have the required number of service representatives available 24 hours a day, 7 days a week
throughout the year.
Pat Niebuhr and his team are responsible for developing a global staffing plan for the
contact centers. His initial challenge is to forecast contacts for the primary and specialty cat-
egories. Pat must work with monthly forecasts of total orders (which, in turn, are derived
from monthly revenue forecasts) and contacts per order (CPO) numbers supplied by the
finance department. Pat recognizes that contacts are given by

Contacts  Orders  CPO

For staff planning purposes, Pat must have forecasts of contacts on a weekly basis.
Fortunately, there is a history of actual orders, actual contacts, actual contacts per order, and
other relevant information, in some cases, recorded by day of the week. This history is
organized in a spreadsheet. Pat is considering using this historical information to develop
the forecasts he needs.

Summary
The purpose of a forecast is to reduce the range of uncertainty within which manage-
ment judgments must be made. This purpose suggests two primary rules to which the
forecasting process must adhere:

1. The forecast must be technically correct and produce forecasts accurate enough to
meet the firm’s needs.
2. The forecasting procedure and its results must be effectively presented to manage-
ment so that the forecasts are used in the decision-making process to the firm’s
advantage; results must also be justified on a cost-benefit basis.

Forecasters often pay particular attention to the first rule and expend less effort on the
second. Yet if well-prepared and cost-effective forecasts are to benefit the firm, those
who have the decision-making authority must use them. This raises the question of
what might be called the “politics” of forecasting. Substantial and sometimes major
expenditures and resource allocations within a firm often rest on management’s view
of the course of future events. Because the movement of resources and power within
an organization is often based on the perceived direction of the future (forecasts), it is
not surprising to find a certain amount of political intrigue surrounding the forecasting
process. The need to be able to effectively sell forecasts to management is at least as
important as the need to be able to develop the forecasts.
10 CHAPTER 1 Introduction to Forecasting

The remainder of this book discusses various forecasting models and procedures.
First, we review basic statistical concepts and provide an introduction to correlation
and regression analysis. Then, one chapter is devoted to methods of collecting data and
exploring data sets for underlying patterns. Many specific forecasting methods are
detailed in the chapters that follow, and the final two chapters are devoted to aspects of
judgmental forecasting and management of the forecasting process.

CASES

CASE 1-1 MR. TUX

John Mosby owns several Mr. Tux rental stores, most long-term debt. He has sources for both types of
of them in the Spokane, Washington, area.3 His growth financing, but investors and bankers are inter-
Spokane store also makes tuxedo shirts, which he dis- ested in a concrete way of forecasting future sales.
tributes to rental shops across the country. Because Although they trust John, his word that the future of
rental activity varies from season to season due to his business “looks great” leaves them uneasy.
proms, reunions, and other activities, John knows that As a first step in building a forecasting model,
his business is seasonal. He would like to measure this John directs one of his employees, McKennah Lane,
seasonal effect, both to assist him in managing his to collect monthly sales data for the past several
business and to use in negotiating a loan repayment years. In the chapters that follow, various techniques
with his banker. are used to forecast these sales data for Mr. Tux. In
Of even greater interest to John is finding a way Chapter 11, these efforts are summarized, and John
of forecasting his monthly sales. His business contin- Mosby attempts to choose the forecasting technique
ues to grow, which, in turn, requires more capital and that will best meet his needs.

CASE 1-2 CONSUMER CREDIT

COUNSELING

Consumer Credit Counseling (CCC), a private, non- financial difficulties or who want to improve their
profit corporation, was founded in 1982.4 The pur- money management skills. Money management
pose of CCC is to provide consumers with assistance educational programs are provided for schools, com-
in planning and following budgets, with assistance in munity groups, and businesses. A debt management
making arrangements with creditors to repay delin- program is offered as an alternative to bankruptcy.
quent obligations, and with money management Through this program, CCC negotiates with credi-
education. tors on behalf of the client for special payment
Private financial counseling is provided at no arrangements. The client makes a lump-sum pay-
cost to families and individuals who are experiencing ment to CCC that is then disbursed to creditors.

3We are indebted to John Mosby, the owner of Mr. Tux rental stores, for his help in preparing this case.
4We are indebted to Marv Harnishfeger, executive director of Consumer Credit Counseling of Spokane, and
Dorothy Mercer, president of its board of directors, for their help in preparing the case. Dorothy is a former
M.B.A. student of JH who has consistently kept us in touch with the use of quantitative methods in the real
world of business.
 CHAPTER 1 Introduction to Forecasting 11
CCC has a blend of paid and volunteer staff; in A major portion of corporate support comes
fact, volunteers outnumber paid staff three to one. from a local utility that provides funding to support
Seven paid staff provide management, clerical sup- a full-time counselor position as well as office space
port, and about half of the counseling needs for for counseling at all offices.
CCC. Twenty-one volunteer counselors fulfill the In addition, client fees are a source of funding.
other half of the counseling needs of the service. Clients who participate in debt management pay a
CCC depends primarily on corporate funding to monthly fee of $15 to help cover the administrative
support operations and services. The Fair Share cost of this program. (Fees are reduced or waived
Funding Program allows creditors who receive pay- for clients who are unable to afford them.)
ments from client debt management programs to This background will be used in the chapters
donate back to the service a portion of the funds that follow as CCC faces difficult problems related
returned to them through these programs. to forecasting important variables.

Minitab Applications

Minitab is a sophisticated statistical program that improves with each release.

Described here is Release 15.
Figure 1-1 shows you four important aspects of Minitab. The menu bar is where
you choose commands. For instance, click on Stat and a pull-down menu appears that
contains all of the statistical techniques available. The toolbar displays buttons for
commonly used functions. Note that these buttons change depending on which
Minitab window is open. There are two separate windows on the Minitab screen: the
data window, where you enter, edit, and view the column data for each worksheet; and
the session window, which displays text output, such as tables of statistics.
In the following chapters, specific instructions will be given to enable you to enter
data into the Minitab spreadsheet and to activate forecasting procedures to produce
needed forecasts.

FIGURE 1-1 Basic Minitab Screen

Menu bar
Toolbar

Session
window

Data
window
12 CHAPTER 1 Introduction to Forecasting

Menu bar
Toolbar

Formula
bar

Worksheet

FIGURE 1-2 Basic Excel Screen

Excel Applications

Excel is a popular spreadsheet program that is frequently used for forecasting. Figure 1-2
shows the opening screen of Version 2003. Data are entered in the rows and columns of
the spreadsheet, and then commands are issued to perform various operations on the
entered data.
For example, annual salaries for a number of employees could be entered into col-
umn A and the average of these values calculated by Excel. As another example,
employee ages could be placed in column B and the relationship between age and
salary examined. The chapters that follow will show you how to use Excel to solve
these and other forecasting problems.
There are several statistical functions available on Excel that may not be on the
drop-down menus on your screen. To activate these functions, click on the following:

Tools>Add-Ins

The Add-Ins dialog box appears. Select Analysis ToolPak and click on OK. The func-
tions available under ToolPak will be called on in later chapters.
It is strongly recommended that an Excel add-in be used to help with the multi-
tude of statistical computations required by the forecasting techniques discussed in this
textbook.

References
Bernstein, P. L. Against the Gods: The Remarkable Chase, C. W., Jr. “Selecting the Appropriate
Story of Risk. New York: Wiley, 1996. Forecasting Method.” Journal of Business
Carlberg, C. “Use Excel’s Forecasting to Get Forecasting 15 (Fall 1997): 2.
Terrific Projections.” Denver Business Journal 47 Diebold, F. X. Elements of Forecasting, 3rd ed.
(18) (1996): 2B. Cincinnati, Ohio: South-Western, 2004.
 CHAPTER 1 Introduction to Forecasting 13
Georgoff, D. M., and R. G. Mardick. “Manager’s Newbold, P., and T. Bos. Introductory Business and
Guide to Forecasting.” Harvard Business Review Economic Forecasting, 2nd ed. Cincinnati, Ohio:
1 (1986): 110–120. South-Western, 1994.
Hogarth, R. M., and S. Makridakis. “Forecasting and Ord, J. K., and S. Lowe. “Automatic Forecasting.”
Planning: An Evaluation.” Management Science American Statistician 50 (1996): 88–94.
27 (2) (1981): 115–138. Perry, S. “Applied Business Forecasting.”
Hyndman, R. J., and J. K. Ord, eds. “Special Issue: Management Accounting 72 (3) (1994): 40.
Twenty Five Years of Forecasting.” International Pindyck, R. S., and D. L. Rubinfeld. Econometric
Journal of Forecasting 22 (2006): 413–636. Models and Economic Forecasts, 4th ed. New
Levenbach, H., and J. P. Cleary. Forecasting Practice York: McGraw-Hill, 1998.
and Process for Demand Management. Belmont, Wright, G., and P. Ayton, eds. Judgemental
Calif.: Thomson Brooks/Cole, 2006. Forecasting. New York: Wiley, 1987.
Makridakis, S. “The Art and Science of Forecasting.”
International Journal of Forecasting 2 (1986): 15–39.
This page intentionally left blank
C H A P T E R

2 A REVIEW OF BASIC
STATISTICAL CONCEPTS
Most forecasting techniques are based on fundamental statistical concepts that are the
subject of business statistics textbooks and introductory statistics courses. This chapter
reviews some of the basic concepts that will serve as a foundation for much of the
material in the remainder of the text.
Most statistical procedures make inferences about the items of interest, called the
population, after selecting and measuring a subgroup of these items, called the sample.
Careful selection of a representative sample and the use of a sufficiently large sample
size are important components of a statistical inference process that has an acceptably
low degree of risk.
In forecasting, we usually work with historical data in an attempt to predict, or
forecast, the uncertain future. For this reason, we will concentrate on examining
sample data, manipulating these data in some way, and using the results to make
forecasts.

DESCRIBING DATA WITH NUMERICAL SUMMARIES

The purpose of descriptive statistical procedures is to briefly describe a large collection
of measurements with a few key summary values. The most common summary is
obtained by averaging the values. In statistics, the process of averaging is usually
accomplished by computing the mean, which involves adding all values and dividing by
the number of values. Although the mean indicates the “middle” in some sense, its cal-
culation is sensitive to a few measurements that are either quite a bit larger or quite a
bit smaller than the rest. These unusual measurements are called outliers. Example 2.1
provides an illustration of this phenomenon.
The sample mean (X or X-bar) is computed using Equation 2.1.

©X
X = (2.1)
n

where
X = the sample mean
©X = the sum of all the values of the sample
n = the sample size
To simplify the computations in this text, some shorthand notation is used. In the
simplified notation for summing all X values (see Equation 2.1), the summations

15
16 CHAPTER 2 A Review of Basic Statistical Concepts

are understood to extend from 1 to n. A more formal notation system for this
procedure is
n

a Xi
i=1

where the subscript i varies from its initial value of 1 to n in increments of 1. Since
almost all sums run from 1 to n, the starting and ending (n) indices will be suppressed,
and the simpler notation will be used except where the more complete notation is
needed for clarity.
In addition to the central tendency of a group of values, measured by computing
the mean, the extent to which the values are dispersed around the mean is fre-
quently of interest. The standard deviation can be thought of as a unit of measure-
ment for measuring distance from the mean. The formula for the standard deviation
is given by

1©X22
©1X - X 22
©X 2 -
n
S = = (2.2)
D n - 1 T n - 1

where the numerator is the sum of the squared differences between the measured val-
ues and their mean.
Many statistical procedures make use of the sample variance. The variance of a
collection of measurements is the standard deviation squared. Thus, the sample vari-
ance (S 2) is computed as

1©X22
©1X - X 22
©X 2 -
n
S2 = = (2.3)
n - 1 n - 1

Example 2.1 Calculation of the Mean, Variance, and Standard Deviation

Consider the following collection of people’s ages:

23, 38, 42, 25, 60, 55, 50, 42, 32, 35

For this sample, n = 10 and

n
Á + X10 = 23 + 38 + Á + 35 = 402
a Xi = X1 + X2 +
i=1

©X 402
X = = = 40.2
n 10

©1X - X 22 123 - 40.222 + Á + 135 - 40.222 1,339.6

S2 = = = = 148.84
n - 1 10 - 1 9

S = 2S2 = 1148.84 = 12.2

The computations are shown in Table 2-1. The sample mean is 40.2 years, the sample vari-
ance is 148.84, and the sample standard deviation is 12.2 years.
 CHAPTER 2 A Review of Basic Statistical Concepts 17

TABLE 2-1 Calculation of

S2 (X = 40.2 )

X X-X 1X - X 22
23 -17.2 295.84
38 -2.2 4.84
42 1.8 3.24
25 -15.2 231.04
60 19.8 392.04
55 14.8 219.04
50 9.8 96.04
42 1.8 3.24
32 -8.2 67.24
35 -5.2 27.04

©1X - X 2 2 = 1,339.60
1,339.60
S2 = = 148.84
10 - 1

To indicate the sensitivity of the mean to an outlying measurement, suppose we replace age
60 with a much larger age—say, 90. With this replacement age, the mean is

©X 23 + Á + 25 + 90 + 55 + Á 35 432
X = = = = 43.2
n 10 10

Increasing the size of only 1 of the 10 original ages results in a 3-year increase in the mean age.

The term degrees of freedom is used to indicate the number of data items that are
free of each other in the sense that they cannot be calculated from one another and
therefore carry unique pieces of information. For example, suppose the following three
statements are made:

I am thinking of the number 5. I am thinking of the number 7. The sum of

the two numbers I am thinking of is 12.

At first glance, three pieces of information are presented. However, if any two of these
statements are known, the other one can be determined. It could be said that there are
only two unique pieces of information in the three statements, or to use the statistical
term, there are only two degrees of freedom because only two of the values are free to
vary; the third is not.

Degrees of freedom refers to the number of data items that are independent of
one another and that carry unique pieces of information.

In the example presented in Table 2-1, the ages of 10 people constitute a sample with
10 degrees of freedom. Anyone’s age could have been included in the sample, and
therefore, each of the ages is free to vary. When the mean is calculated, all 10 ages are
used to account for a mean age equal to 40.2 years.
18 CHAPTER 2 A Review of Basic Statistical Concepts

The computation of the sample standard deviation differs. When the sample stan-
dard deviation is calculated, an estimate of the population mean (the sample mean X )
is used. By using the sample mean as an estimate of the population mean in the compu-
tation, a standard deviation that is smaller than the population standard deviation will
usually be obtained. However, this problem can be corrected by dividing the value
©1X - X 22 by the appropriate degrees of freedom. Once the sample mean is com-
puted, only nine of the deviations X - X are required, in this example, to calculate the
sample standard deviation. Given nine of the deviations, the last deviation is fixed,
since ©1X - X 2 must equal zero. Consequently, we say that the sample standard devi-
ation (or sample variance) has nine degrees of freedom. In general, whenever a sample
statistic is used as an estimate of a population parameter in a computation, one degree
of freedom is lost.
Summary statistics can be defined for populations. To distinguish population statis-
tics from sample statistics, we use different notation. Table 2-2 shows the symbols used
for both population and sample statistics.
The mean and standard deviation are the measures most commonly used to
describe sample data in a brief and meaningful way. However, there are other descrip-
tive summary measures available. For example, the median is often used to indicate the
middle or central value in a data collection. The median is that value that divides a set
of ordered measurements in half. That is, half the measurements are less than the
median and half are greater. Since the median simply divides a set of measurements in
half, it is unchanged if, for example, the largest measurement is increased. We say the
median is resistant to the effects of outlying observations.
Sometimes the range of the data is presented as a rough measure of dispersion.
The range is the difference between the largest and smallest values. The range of ages
in Table 2-1, for example, is 37 160 - 232.
The quartiles divide the data collection into four equal parts, after the numerical
values have been ordered or arrayed from smallest to largest. As we discussed, the
median divides the array into two equal parts and is sometimes called the second quar-
tile, denoted by Q2. The first quartile (Q1) divides the lower half of the array into two
equal parts, and the third quartile (Q3) divides the upper half into two equal parts. The
collection of ages in Table 2-1, for example, has a first quartile of Q1 = 30.25, a median
(or second quartile) of Q2 = 40, and a third quartile of Q3 = 51.25.
Finally, the interquartile range is another indicator of the variability of a data set. It
is simply the difference between the third quartile and the first quartile (Q3 - Q1) or
the range for the middle 50% of the data values. For the age data, the interquartile
range is 21 (51.25 - 30.25).
Minitab and Excel can be used to compute most of the descriptive statistics
presented so far. Figure 2-1 shows the Minitab output for the age data presented in
Table 2-1. The instructions for computing descriptive statistics using Minitab and
Excel are presented in the Minitab and Excel Applications sections at the end of
this chapter.

TABLE 2-2 Notation for Population

and Sample Statistics

Statistic Population Symbol Sample Symbol

Mean μ X
Variance σ2 S2
Standard deviation σ S
 CHAPTER 2 A Review of Basic Statistical Concepts 19

FIGURE 2-1 Minitab Output for Descriptive Statistics

DISPLAYS OF NUMERICAL INFORMATION

In forecasting, we are concerned with two types of data: cross-sectional data, which
consist of observations from the same time frame, and time series data, which consist of
a sequence of observations over time. Table 2-3 shows an example of cross-sectional
data: the net income as a percentage of equity for a sample of 209 companies from a
particular Fortune 500 survey. Other examples are corporation executives’ yearly
incomes for 2006 and selling prices for homes in a certain city on a given day. Some
examples of time series data are the number of visitors at the Grand Coulee Dam
Visitor Center per month for several years, monthly sales for the Sears Corporation for
the last 10 years, daily IBM stock price for 2006, and annual U.S. wheat production for
the period 2000–2005.
One of the most important things to do when first exploring a variable is to visual-
ize the data through charts and graphs. Basic features of the data, including unusual
observations and unique patterns, are most easily seen visually. At times, graphs even
suggest possible explanations for some of the variation in the data.
A dot plot is one of the simplest ways to visualize data graphically. In Figure 2-2, a
horizontal axis displays the range of values for net income as a percentage of equity from
Table 2-3. Each observation is represented by a dot placed above the axis. Dot plots show
the details of the data and also allow a forecaster to compare two or more sets of data.
A box plot, also called a box-and-whisker plot, is useful for displaying the distribu-
tional characteristics of data. Figure 2-3 shows a box plot for the net income as a per-
centage of equity data from Table 2-3. A line is drawn across the box at the median.
20 CHAPTER 2 A Review of Basic Statistical Concepts

TABLE 2-3 Net Income as a Percentage of Equity for a Sample

of 209 Companies from a Fortune 500 Survey

17 23 22 18 8 7 12 2 49 14
14 36 16 7 3 8 10 11 20 17
15 25 18 12 20 7 5 11 0 22
14 10 14 19 8 12 13 21 3 22
11 18 2 18 14 11 36 16 7 14
12 14 10 8 20 13 8 23 6 21
9 23 7 14 25 12 12 8 11 5
18 13 14 9 16 2 19 21 18 9
14 2 20 17 11 16 13 12 22 16
7 6 14 10 1 21 35 20 18 28
17 15 9 12 5 10 14 1 17 14
14 14 6 22 16 13 14 8 12 6
15 10 22 19 16 4 20 18 2 3
20 7 15 39 4 3 10 7 15 16
12 13 12 11 18 10 13 7 13 12
14 8 11 17 11 22 16 11 12 11
9 11 13 0 12 3 9 9 13 27
1 16 18 12 11 0 10 9 12 22
18 44 4 3 17 12 8 16 7 16
27 11 19 12 22 3 14 14 7 8
11 1 3 17 8 7 5 19 22

Dot Plot of % of Equity

FIGURE 2-2 Dot Plot for Net Income as a Percentage of Equity for a
Sample of 209 Companies from a Fortune 500 Survey

This line divides the data into two equal sections. The lower edge of the box is the first
quartile (Q1), and the upper edge is the third quartile (Q3). Additional limits using the
interquartile range (Q3 - Q1) can be constructed. The lower limit is located at
Q1 - 1.51Q3 - Q12, and the upper limit is located at Q3 + 1.51Q3 - Q12. Outliers are
identified as those points outside of the lower and upper limits and are plotted with
asterisks. In Figure 2-3, the first quartile is 8, the median is 13, the third quartile is 17,
 CHAPTER 2 A Review of Basic Statistical Concepts 21

Box Plot for % of Equity

Outliers

Highest Value within

Upper Limit (28)

Third Quartile
Median

First Quartile Lowest Value within

Lower Limit (0)

FIGURE 2-3 Box Plot for Net Income as a Percentage of Equity for
a Sample of 209 Companies from a Fortune 500 Survey

and the interquartile range is 9 (17 - 8). The lower limit is -5.5 (8 - 1.5 * 9), and the
upper limit is 30.5 (17 + 1.5 * 9). Note that the lowest value within the lowest limit,
-5.5, is 0 and the highest value within the highest limit, 30.5, is 28. Six values (35, 36, 36,
39, 44, and 49) exceed the upper limit of 30.5 and are the outliers.
The histogram condenses the data by grouping similar values into classes. A his-
togram can be constructed by placing the variable of interest on the horizontal axis and
the frequency, relative frequency, or percent frequency on the vertical axis. By looking
at a histogram, such as Figure 2-4, you can tell the proportion of the total area above an
interval of the horizontal axis. A total of 23 out of the 209 companies or 11.0% had net

FIGURE 2-4 Histogram for Net Income as a Percentage of Equity

for a Sample of 209 Companies from a Fortune 500 Survey
22 CHAPTER 2 A Review of Basic Statistical Concepts

FIGURE 2-5 Monthly Sales for Alomega Food Stores (from Table 3-11)

income as a percentage of equity in the interval between 6.25% and 8.75%.

Consequently, the fourth rectangle in Figure 2-4 contains .110 of the total area repre-
sented by the shaded vertical bars.
Although histograms provide good visual descriptions of data sets, especially very
large data sets, they do not let us identify individual observations. In contrast, in a dot
plot each of the original values is visible.
Scatter diagrams are used to visualize the relationship between two variables. They
will be discussed later in the chapter in the section on correlation analysis.
For chronological data, the most frequently used graphical form is a time series
plot, in which the data are plotted over time. Figure 2-5 shows a time series plot of
monthly sales for Alomega Food Stores, discussed in Example 1.1. A time series plot
reveals the variability of the data and the time at which peaks and valleys occur. It also
shows the relative size of the peaks and valleys compared with the rest of the series.
One important step in selecting an appropriate forecasting technique is to identify
the data patterns that exist within a time series. Once the data patterns have been iden-
tified, then the forecasting methods most appropriate to those patterns can be utilized.
Four types of time series data patterns can be identified: horizontal, trend, cyclical, and
seasonal. Each type will be discussed in depth in Chapter 3.
A tool that is often used to help identify time series data patterns is the correlogram
or autocorrelation function. It is a graphical tool for displaying the correlations between
various time lags of a time series. Correlograms will also be discussed in Chapter 3.

PROBABILITY DISTRIBUTIONS
Random variable is the name given to a quantity that is capable of taking on different
values from trial to trial in an experiment, the exact outcome being a chance or random
event. If only certain specified values are possible, the random variable is called a
discrete variable. Examples include the number of rooms in a house, the number
of people arriving at a supermarket checkout stand in an hour, and the number of
defective units in a batch of electronic parts. If any value of the random variable is
 CHAPTER 2 A Review of Basic Statistical Concepts 23
possible within some range, it is called a continuous variable. Examples of this type of
variable are the weight of a person, the length of a manufactured part, and the time
between car arrivals at a tollbooth.

A discrete random variable can assume only values from a predetermined set.
These outcomes are often represented numerically by integers. A continuous
random variable can assume any value within a specified range. These outcomes
are represented numerically by intervals of values.

The probability distribution of a discrete random variable lists all possible values
that the variable can take on, along with the probability of each. The expected value of
a random variable is the mean value that the variable assumes over many trials. The
expected value, E(X), for a discrete probability distribution can be found by multiply-
ing each possible X value by its probability, P(X), and then summing these products.
Equation 2.4 shows this calculation:

E1X2 = ©3X * P1X24 (2.4)

The expected value of a random variable is the mean value of the variable over
many trials or observations.

Example 2.2 A Discrete Probability Distribution

The number of no-sales days for a salesperson during a month is described by the probabil-
ity distribution shown in Table 2-4. These values are based on the salesperson’s past experi-
ence and will be used to forecast future sales activity. The X column lists all values (no-sales
days) that are possible, whereas the P(X) column lists the corresponding probabilities. Note
that, since all possible values of X are listed, the probabilities sum to 1.00, or 100%. This is
true for all probability distributions, disregarding rounding errors.

For the probability distribution given in Table 2-4, the expected value is
E1X2 = 11.12 + 21.22 + 31.252 + 41.152 + 51.302 = 3.35
Thus, if this salesperson was observed for a very large number of months and the num-
bers of no-sales days were recorded, the mean would be 3.35 no-sales days if future
activity is correctly predicted by the historical data on which the probability distribu-
tion is based.

TABLE 2-4 Discrete Probability Distribution

for Example 2.2

X P(X)

1 .10
2 .20
3 .25
4 .15
5 .30
24 CHAPTER 2 A Review of Basic Statistical Concepts

Notice that the mean falls near the middle of the X values. It is pulled toward the
upper end of the probability distribution because of the relatively large probability
associated with X = 5.
For a continuous distribution, the probability of obtaining a specific value
approaches zero. For instance, the probability of anyone weighing 150 pounds may be
considered zero, since this would mean that this weight is exactly 150.0000—no matter
how accurate a scale is used. In the continuous distribution case, probabilities are
assigned to intervals or ranges of values. The probability that a person’s weight falls in
the interval of 149 to 151 pounds might be computed, for example.
Some theoretical distributions occur over and over again in practical statistical
applications, and for this reason, it is important to examine their properties. One of
these is the binomial distribution, often used to represent a discrete random variable.
The requirements for a binomial experiment are as follows:

1. There are n identical trials, each of which results in one of two possible outcomes—
say, success and failure.
2. The probability of success for each outcome remains fixed from trial to trial.
3. The trials are independent (not related).

The interest is in finding the probability of X successful occurrences in n trials,

where a successful occurrence is arbitrarily defined to be one of the two possi-
ble outcomes. The various values of X, along with their probabilities, form the
binomial distribution. These probabilities can be found from the following binomial
formula:

P1X2 = a bpX11 - p2n - X

n
for X = 0, 1, Á , n (2.5)
X

where

a b
n
= the number of combinations of n things taken X at a time
X
p = the probability of success on each trial
X = the particular number of successes of interest
n = the number of trials

The binomial distribution is a discrete probability distribution describing the like-

lihood of X successes in n independent trials of a binomial experiment.

An easier way to find binomial probabilities than using Equation 2.5 is to refer to
a binomial distribution table such as Table B-1, found in Appendix B. Blocks that
represent n group the probabilities. Each block has a column headed by p and rows
indicated by X.1

1Minitab and Excel will also calculate binomial probabilities.

 CHAPTER 2 A Review of Basic Statistical Concepts 25

FIGURE 2-6 A Normal Distribution

Example 2.3 A Binomial Probability

Suppose eight items are randomly drawn from a production line that is known to produce
defective parts 5% of the time. What is the probability of getting exactly zero defectives?
The answer, from Table B-1, is .6634. (Here, n = 8, p = .05, and X = 0.)

An important continuous distribution of interest, because many useful populations

of numbers can be approximated by it, is the normal distribution. Knowledge of the
mean and the standard deviation is necessary to identify a specific normal distribution.
A normal curve is symmetrical and bell-shaped, as shown in Figure 2-6. This distribu-
tion represents many real-life variables that are measured on a continuous scale.
Probabilities of values drawn from a normal distribution falling into various inter-
vals are found by first converting all intervals to units of standard deviation called
Z-scores.2 The Z-score of any X value is the number of standard deviations from the
central value of the curve (m) to that value. Thus, the formula is
X - m
Z = (2.6)
s
where
X = the value of interest
m = the mean
s = the standard deviation
After the Z-score has been computed, the normal curve table can be consulted to
find the area under the curve between the center of the original curve (m) and the
value of interest (X).3

The normal distribution has a bell shape and is completely determined by its
mean and standard deviation.

2It can be shown that, if the random variable X has a normal distribution with mean m and standard devia-
tion σ, then the random variable Z = 1X - m2>s has a normal distribution with mean 0 and standard devi-
ation 1. This particular normal distribution is called the standard normal distribution.
3Normal probabilities can be computed using Minitab or Excel.
26 CHAPTER 2 A Review of Basic Statistical Concepts

Population of
Part Weights

9 12
 = 10
= 2

FIGURE 2-7 Normal Curve Areas for Example 2.4

Example 2.4 Calculating a Normal Probability

The weights of a population of parts made by a certain machine are normally distributed
with a mean of 10 pounds and a standard deviation of 2 pounds. What is the probability that
a part drawn at random from the machine weighs between 9 and 12 pounds? The normal
curve with the appropriate area shaded is shown in Figure 2-7.
Since normal curve tables are often designed to give areas from the center of the curve
to some point, two separate areas must be found: one on each side of the mean. Then these
areas are added together. This process will produce the probability of a value falling in this
interval. The two Z-scores are

X1 - m 9 - 10
Z1 = = = - .50
s 2
X2 - m 12 - 10
Z2 = = = 1.00
s 2
Consequently, the area under the normal curve between 9 and 12 in Figure 2-7 is the same
as the area under the normal curve of Z (see footnote 2) between -.5 and 1. For the normal
table in this book, the negative sign on the first Z-score is disregarded, since the area under
the normal curve for Z between -.5 and 0 is the same as the area under the curve between 0
and .5. When these two Z-scores are located in Table B-2, a normal curve table, they yield
the following two areas, which are then added together:

Z1 = - .50 : .1915
Z2 = 1.00 : .3413
.5328

It is concluded that there is about a 53% chance that a part randomly drawn from this pop-
ulation of parts will weigh between 9 and 12 pounds.

SAMPLING DISTRIBUTIONS

In most statistical applications, a random sample is taken from the population under
investigation, a statistic is computed from the sample data, and conclusions are
drawn about the population on the basis of this sample. A sampling distribution is
the distribution of all possible values of the sample statistic that can be obtained
from the population for a given sample size. For instance, a random sample of 100
 CHAPTER 2 A Review of Basic Statistical Concepts 27

Mean = 

FIGURE 2-8 Sampling Distribution of X

persons might be taken from a population and weighed and then their mean weight
computed. This sample mean (X ) can be thought of as having been drawn from the
distribution of all possible sample means of sample size 100 that could be taken
from the population. More generally, each sample statistic that can be com-
puted from sample data can be considered as having been drawn from a sampling
distribution.

A sampling distribution is the array of all possible values of a sample statistic that
can be drawn from a population for a given sample size.

The central limit theorem states that, as the sample size becomes larger, the sam-
pling distribution of sample means tends toward the normal distribution and that the
mean of this normal distribution is m, the population mean, and the standard deviation
is s> 1n, the population standard deviation divided by the square root of the sample
size. The quantity s> 1n is known as the standard error of the sample mean. Thus, the
sampling distribution of the sample mean will tend toward normality regardless of the
shape of the population distribution from which the samples were drawn. Figure 2-8
demonstrates how such a sampling distribution might appear.
The central limit theorem is of particular importance in statistics, since it allows the
analyst to compute the probability of various sample results through knowledge of
normal curve probabilities.

Example 2.5 Calculating a Probability Using the Sampling Distribution of X

What is the probability that the mean of a random sample of 100 weights drawn from a
population will be within 2 pounds of the true population mean weight if the standard
deviation of the population is estimated to be 15 pounds? Figure 2-9 illustrates the appro-
priate sampling distribution.
The standard error is s> 1n = 15> 1100 = 1.5, yielding a Z-score of Z = 1X - m2>
1s> 1n2 = 2>1.5 = 1.33 and an area from the normal curve table of .4082. Doubling this
area to account for both sides of the mean results in a total area of .8164. Chances are
about 82% that the sample mean will be within 2 pounds of the true mean, given the sam-
ple size of 100 and the estimated variability of the population, s = 15. As will be seen, this
ability to calculate probabilities of sample results will enable an analyst to make infer-
ences about the population from the sample, with direct applicability to forecasting.
28 CHAPTER 2 A Review of Basic Statistical Concepts

X
−2 +2


FIGURE 2-9 Sampling Distribution Areas for Example 2.5

For small sample sizes, the sample means may not be normally distributed.
However, if the population from which the sample is selected is essentially normal and
the population standard deviation is estimated by the sample standard deviation, S, the
sampling distribution of the ratio

X - m
t = (2.7)
S> 1n

is known. It is called Student’s t distribution (or simply the t distribution) with n - 1

degrees of freedom (abbreviated df). The t distribution is centered at 0 and has a bell
shape but with fatter tails than a normal distribution. However, as the number of degrees
of freedom (equivalently, the sample size, n) increases, the t distribution looks more and
more like a normal distribution with a mean of 0 and a standard deviation of 1.
Table B-3 (in Appendix B) shows values taken from the t distribution. Note that
only one value need be specified before referring to the table, namely, the degrees of
freedom. Once the degrees of freedom is known, the t values that exclude desired
percentages of the area under the curve can be found. For example, if the t distribu-
tion of interest has 12 degrees of freedom, then a t value of 2.179 on each side of
0 will include 95% of the area under the curve and exclude 5% of it, that is, 2.5% in
each tail.
Example 2.6 Using the t Distribution
A manager of a credit card company claims the mean time to resolve disputed claims is 30
days. To test this claim, data were collected on a sample of 15 disputed claims. The 15 claims
had a mean time to resolve of 35.9 days and a standard deviation of 10.2 days.
If the data are regarded as a random sample from a normal population, do they cast
suspicion on the manager’s claim?
If the manager’s claim is correct, the population mean is m = 30, and the random
variable

X - m
t =
S> 1n

has Student’s t distribution with n - 1 = 15 - 1 = 14 degrees of freedom. The manager’s

claim will be suspect if the value of this variable is too large or too small because this occurs
when the value of X is far from m. Consulting Table B-3 for 14 degrees of freedom, we
see that t.025 = 2.145. Before the sample is obtained, the probability that t will exceed
2.145 is .025. Similarly, the probability is .025 that t will be less than - 2.145. Adding
these two equal probabilities, the probability is .025 + .025 = .05 that t will be larger
 CHAPTER 2 A Review of Basic Statistical Concepts 29
than 2.145 or smaller than -2.145. Now we are ready to determine whether the value of
X , 35.9, is far from m, 30. The observed value of t is
35.9 - 30
t = = 2.240
10.2> 115
which exceeds t.025 = 2.145. Rather than accepting the explanation that an event with small
probability has occurred, we question the claim that the mean is 30 days. A larger value for
the mean would be more consistent with the data.

INFERENCE FROM A SAMPLE

Estimation
When it is not feasible, or even possible, to measure the entire population of interest, a
sample from the population is selected and used to learn about the population. This
learning (called inference) typically takes two forms. The first of these is called
estimation, where the sample results are used to estimate an unknown population char-
acteristic. Although estimation is the common statistical term for this task, it could as
well be called forecasting in many business situations because the sample consists of
collected historical observations and a value (estimate, or forecast) for a future obser-
vation is required. The second form of learning about the population from a sample is
called hypothesis testing and will be discussed in the next section.
A point estimate (a forecast) of a population quantity, such as the population
mean, is a single value calculated from the sample data that estimates the value of
the unknown population quantity. In statistics, a population quantity is called a pop-
ulation parameter. A point estimate is a “best guess” of a population parameter
computed from the sample. Often the best guess of a population parameter is pro-
vided by the corresponding sample quantity. For example, the best guess of the
value of the population mean is given by the value of the sample mean. Table 2-2
contained three population parameters and the sample statistics that provide point
estimates of them.
An interval estimate is an interval within which the population parameter of
interest is likely to lie. It is found by constructing an interval around the point estimate
of the form

Point estimate  Multiple  (Estimated standard deviation of point estimate) (2.8)

where the “multiple” is often a normal distribution percentage point or a t distribution

percentage point. Interval estimates are preferred over point estimates because they
take account of the uncertainty associated with an estimate.
Ideally, we would like to obtain a sample and then determine an interval that
would, for instance, definitely contain the population mean, m. However, because of
sample-to-sample variability, this goal is unattainable. Instead, we specify a high prob-
ability—say, .95 or .99—that a proposed interval will cover the true value of the popu-
lation parameter of interest. Because this probability pertains to the interval before
the sample is observed, it is called the level of confidence. An interval obtained by a
procedure satisfying the probability requirement is called a confidence interval. The
quantity to the right of the ± sign in Equation 2.8 is sometimes called the error
allowance, and its size will depend directly on the level of confidence (through the
multiple) and the information in the sample (through the estimated standard devia-
tion of point estimate).
30 CHAPTER 2 A Review of Basic Statistical Concepts

A point estimate is a single-valued estimate of a population parameter. An interval

estimate is an interval within which the population parameter is likely to lie.

Example 2.7 Calculating Point and Interval Estimates

A random sample of 500 shoppers is selected from Northgate Shopping Center to deter-
mine the average distance they traveled to the mall. An analysis of the sample results
reveals that X = 23.5 miles and S = 10.4 miles.
The point estimate for the unknown population mean distance for all shoppers at the
mall is 23.5 miles, the sample mean.
The estimated standard deviation of X is S/1n , and for large random samples (n > 30,
say), the multiple in the construction of a confidence interval (see Equation 2.8) for the
population mean, m, is Z, a standard normal percentage point. Consequently, an interval
estimate for m is

S
X ; Z (2.9)
1n
If we set the confidence level at 95%, then Z = 1.96, the upper .025 (2.5%) point of a stan-
dard normal distribution (see footnote 2). Using Equation 2.9 with our sample results pro-
duces the interval

10.4
23.5 ; 1.96
1500
23.5 ; 1.961.4652

23.5 ; .91 : 122.6, 24.42

Thus, it can be stated, with 95% confidence, that the mean number of miles traveled to
the shopping mall by the population of shoppers is somewhere between 22.6 miles and 24.4
miles. The 95% confidence refers to the fact that if, say, 100 samples of size 500 were
selected, sample means and standard deviations computed, and interval estimates con-
structed, 95 out of 100 of the intervals would, in fact, contain the actual population mean.
Consequently, it is highly likely that the particular interval constructed above contains the
population mean miles traveled to the mall.
If the sample size had been small, perhaps n = 20, then a 95% confidence interval for
m can be constructed from Equation 2.9, using a t percentage point with 19 df in place of Z.
Since the upper .025 (2.5%) point of a t distribution with 19 df is t.025 = 2.093 and
S> 1n = 10.4> 120 = 2.326, the small-sample 95% confidence interval will be considerably
wider than the large-sample interval calculated above. This is reasonable because there is
more uncertainty associated with a small sample. Note that for small samples we also make
the additional assumption that miles traveled is approximately normally distributed.

Hypothesis Testing
In many inferential situations, including forecasting, the interest is in testing some
claim about the population rather than estimating or forecasting one of its parameters.
This procedure is called hypothesis testing and is the second major form of examining
sample evidence. Hypothesis testing involves the following steps:
Step 1. Formulate the hypothesis being tested (called the null hypothesis, symbol
H0), and state the alternative hypothesis (the one accepted if H0 is
rejected, symbol H1).
Step 2. Collect a random sample of items from the population, measure them, and
compute the appropriate sample test statistic.
 CHAPTER 2 A Review of Basic Statistical Concepts 31

TABLE 2-5 Results of a Hypothesis Test

Action
Do Not Reject H0 Reject H0

H0 True Correct decision Type I error: probability α

State of Nature
H0 False Type II error: probability β Correct decision

Step 3. Assume the null hypothesis is true, and determine the sampling distribu-
tion of the test statistic.
Step 4. Compute the probability that a value of the sample statistic at least as
large as the one observed could have been drawn from this sampling
distribution.
Step 5. If this probability is high, do not reject the null hypothesis; if this probabil-
ity is low, the null hypothesis is discredited and can be rejected with small
chance of error.
When these steps are followed, two types of error can occur, as shown in Table 2-5.
It is hoped that the correct decision concerning the null hypothesis will be reached
after examining sample evidence, but there is always a possibility of rejecting a true H0
and failing to reject a false H0. The probabilities of these events are known as alpha (α)
and beta (β), respectively. Alpha is also known as the significance level of the test.

Example 2.8 Constructing a Hypothesis Test

A job-shop foreman thinks the mean weight of parts produced by a certain machine has
decreased. It is desired to test the hypothesis that the mean weight of parts produced by the
machine is still 50 pounds, the mean weight of the parts in past years. A random sample of
100 parts is taken. It is assumed that the standard deviation of part weights is 5 pounds
regardless of the mean weight, since this value has remained constant in past studies of
parts. We will use the sample mean as our test statistic and take as the null hypothesis the
status quo; that is, the mean part weight, m, is 50 pounds. Consequently, if the null hypothe-
sis, H0, is true, the central limit theorem says the sampling distribution of X is, in this case,
normal with mean m = 50 and standard deviation s> 1n = 5> 1100 = .5. A value of the
sample mean, X, close to or larger than 50 is support for H0; a value of X far below 50 is
evidence against H0, since such an event is unlikely if H0 is true. A sensible decision rule then
is one that rejects H0 for a value of X far below 50. However, with this rule, it’s possible we
could make a type I error; that is, we could reject a true H0. We can control the chance of
rejecting a true H0 when selecting α, the significance level. Suppose we make α small—say,
.05 (5%). Then we will reject H0 if the observed value of our test statistic, X, falls beyond the
lower .05 (5%) point of its sampling distribution under H0. The test is shown in Figure 2-10.
The decision rule for this test is as follows:

If X 6 49.18, reject the null hypothesis H0: m = 50

Suppose that the sample mean turned out to be 49.6 pounds. Then we could not reject
H0 at the 5% level. On the other hand, if the sample mean weight turned out to be
48.6 pounds, we would reject H0. Note that for X = 48.6 , the probability of getting a
value this extreme (far below 50) if H0 is true is P1X 6 48.62 = P 1Z 6 48.6 .5- 502 =
P1Z 6 - 2.82 = P1Z 7 2.82 = 1.5000 - .49742 = .0026 which is an unlikely event
indeed.

In hypothesis testing, there is always some ambiguity associated with selecting the
null and alternative hypotheses. In general, the null hypothesis is the status quo or “no
32 CHAPTER 2 A Review of Basic Statistical Concepts

H0:   50 pounds (the null hypothesis)

H1:   50 pounds (the alternative hypothesis)

Reject H0 Fail to Reject H0

(the sampling
distribution under the
assumption H0 is true)
.05

X
49.18 50
a

Z
−1.645 0

  5  .5
n 100

a  50 − 1.645(.5)   .05 (the chosen

 49.18 significance level)

FIGURE 2-10 Hypothesis Test for Example 2.8

change” hypothesis. The alternative hypothesis is the research or change hypothesis. To

discover if what we have done has made a difference or resulted in a change from the
status quo or current procedures, we formulate the difference or change as the alterna-
tive hypothesis. The goal is to see if we can reject the null hypothesis, with a small
chance of a type I error, in favor of the alternative.
The hypothesis test described in Example 2.8 is referred to as one-sided or one-
tailed test. This terminology follows because in this example the alternative hypothesis
specifies values of the population parameter to one side of the value specified in the
null hypothesis. Tests for which the alternative hypothesis specifies values both larger
and smaller than the value under the null hypothesis are called two-sided or two-tailed
tests. In this case, values of the test statistic much larger or much smaller than the value
of the population parameter under H0 lead to rejection.

p-Value
In Example 2.8, we calculated the probability of getting a value of the sample mean
as extreme as X = 48.6 if the null hypothesis H0: m = 50 is true. The probability
turned out to be .0026. This probability is called the prob-value or simply the p-value
of the test. Rather than selecting a significance level, it is now common practice to
compute and report the p-value for the test; in fact, statistical software packages
routinely report p-values associated with test statistics. p-Values and their use in
hypothesis testing are, in fact, discussed in steps 4 and 5 in our hypothesis-testing
procedure at the beginning of this section. A small p-value means a strong rejection
of the null hypothesis. The p-value can be regarded as evidence for (large p-value) or
against (small p-value) the null hypothesis. Typical cutoff points for small p-values
are .05 and .01.
 CHAPTER 2 A Review of Basic Statistical Concepts 33
Using the information in Example 2.8, if a two-tailed test is performed, H0: m = 50 versus
H1: m Z 50, and X = 48.6 , the p-value is P1X 6 48.6 or X 7 51.42 = 2P1Z 7 51.4 .5- 502,
or twice the area under the standard normal curve to the right of Z = 2.8. In this exam-
ple, the p-value equals 21.00262 = .0052.

The p-value or significance probability is the probability of getting at least as

extreme a sample result as the one actually observed if H0 is true. Equivalently,
the p-value may be regarded as the smallest α for which the observed test statistic
leads to the rejection of H0.

Example 2.9 Calculating a p-Value

Suppose it is desired to test the hypothesis that the mean score of students on a national
examination is 500 against the alternative hypothesis that it is less than 500. A random sample
of 15 students is taken from the population and produces a sample mean score of
X = 475 . The standard deviation of the population is estimated by the sample standard
deviation, S = 35. Assume the population of test scores is normally distributed.
We want to know if mean test scores are less than their historical value of 500. We will
make a decision on the basis of the sample evidence and, in particular, the value of the sam-
ple mean. The competing hypotheses are

H0 : m = 500
H1 : m 6 500
If H0 is true, we would expect to see a value of the sample average close to or larger than
500. If H0 is false (H1 is true), we would expect to see a value of the sample average quite a
bit less than 500. Since the sample size is small, the population is normal, and the population
standard deviation is estimated by the sample standard deviation, an appropriate test statis-
tic is the t statistic given in Equation 2.7, with n - 1 = 15 - 1 = 14 df. A large negative
value of the t statistic is evidence against H0 (indicating the sample average is much less
than the hypothesized population average of 500). First, let’s construct the test for a signifi-
cance level of a = .05. The test (decision rule) is shown in Figure 2-11.

H0 :  = 500 t Distribution Sample Results: n = 15

H1 :  < 500 df = 14 x = 475
s = 35
 = .05

Sampling distribution of (X − )/(S/ n)

assuming H0 true, and assuming that
the population is normally distributed

 = .05
t = −1.761
0
Reject H0

t from t table with

 = .05 and df = (n − 1) = 14 X −  475 − 500 = −2.77
=
= 1.761 S/ n 35/ 15

FIGURE 2-11 Hypothesis Test for Example 2.9

34 CHAPTER 2 A Review of Basic Statistical Concepts

Because the decision rule is as follows:

If observed t 6 - 1.76, reject H0 : m = 500

and the observed t statistic is

X - m 475 - 500
t = = = - 2.77
S> 1n 35> 115
we would reject the null hypothesis and conclude that the mean test score of students on the
national exam is less than 500. The p-value in this case is P1t 6 - 2.772. Using Table B-3
with 14 df and the symmetry of the t distribution, this probability is between .010 and .005.
Our observed test statistic is very unlikely if H0 is true. The p-value tells us we should reject
the null hypothesis.

CORRELATION ANALYSIS
In building statistical models for forecasting, it is often useful to examine the relation-
ship between two variables. Two techniques, correlation and regression analysis, are
reviewed here. In addition, special cases of correlation and regression are considered in
later chapters. This emphasis on correlation and regression is justified in view of the
widespread use of these techniques in all sorts of forecasting applications.

Scatter Diagrams
A study of the relationship between variables begins with the simplest case, that of the
relationship existing between two variables. Suppose two measurements have been
taken on each of several objects. An analyst wishes to determine whether one of these
measurable variables, called Y, tends to increase or decrease as the other variable,
called X, changes. For instance, suppose both age and income have been measured for
several individuals, as shown in Table 2-6. What can be said about the relationship
between X and Y?
From Table 2-6, it appears that Y and X have a definite relationship. As
X increases, Y tends to increase. From this sample of five persons, it appears as if
the older a person becomes, the more money that person makes. Of course, it is
dangerous to reach conclusions on the basis of a small sample size, a subject to be
pursued later. Yet given these observations, a definite relationship appears to exist
between Y and X.
These five data points can be plotted on a two-dimensional scale, with values of X
along the horizontal axis and values of Y along the vertical axis. Such a plot is called a
scatter diagram or scatter plot and appears in Figure 2-12.

TABLE 2-6 Income and Age Measurements

Person Income ($1,000s) Y Age (years) X

.
1 27.8 22
2 28.5 23
3 30.0 26
4 35.0 27
5 36.4 35
 CHAPTER 2 A Review of Basic Statistical Concepts 35

Scatter Plot of Income vs Age

FIGURE 2-12 Scatter Diagram for Age and Income

A scatter diagram is a plot of X–Y data points on a two-dimensional graph.

The scatter diagram helps to illustrate what intuition suggested when the raw data
were first observed—namely, the appearance of a linear relationship between Y and X.
This relationship is called a positive linear relationship because as X increases, so does Y.
In other situations involving two variables, different scatter diagram patterns
might emerge. Consider the plots in Figure 2-13.
Figure 2-13(a) suggests what is called a perfect, positive linear relationship. As X
increases, Y increases also, and in a perfectly predictable way. That is, the X and Y data
points appear to lie on a straight line. Figure 2-13(b) suggests a perfect, negative linear
relationship. As X increases, Y decreases in a perfectly predictable way.
Figures 2-13(c) and (d) illustrate imperfect, positive and negative linear relation-
ships, respectively. As X increases in these scatter diagrams, Y increases (part c) or
decreases (part d) but not in a perfectly predictable way. Thus, Y might be slightly
higher or lower than “expected.” That is, the X–Y points do not lie on a straight line.
Scatter diagrams in Figures 2-13(a) through (d) illustrate linear relationships. The
X–Y relationship, be it perfect or imperfect, can be summarized by a straight line. In
comparison, a curved relationship appears in Figure 2-13(e).
Finally, Figure 2-13(f) suggests that no relationship of any kind exists between vari-
ables X and Y. As X increases, Y does not appear to either increase or decrease in any
predictable way. On the basis of the sample evidence that appears in Figure 2-13(f), it
might be concluded that in the world containing all the X–Y data points, there exists no
relationship, linear or otherwise, between variables X and Y.
Now consider the two scatter diagrams in Figure 2-14. Both scatter diagrams sug-
gest imperfect, positive linear relationships between Y and X. Figure 2-14(a) shows a
strong relationship because the data points are all quite close to the straight line that
36 CHAPTER 2 A Review of Basic Statistical Concepts

Y Y

X X
(a) Perfect, Positive, Linear (b) Perfect, Negative, Linear

Y Y

X X
(c) Imperfect, Positive, Linear (d) Imperfect, Negative, Linear

Y Y

X X
(e) Curved (f) No Relationship

FIGURE 2-13 Patterns for X–Y Data Plots

Y Y

X X
(a) Strong (b) Weak

FIGURE 2-14 Strong and Weak Linear Association for X–Y Data Plots
 CHAPTER 2 A Review of Basic Statistical Concepts 37
passes through them. In Figure 2-14(b), the data points are farther away from the
straight line that passes through them, suggesting a weaker linear relationship. Later in
this chapter, we will demonstrate how to measure the strength of the relationship that
exists between two variables.
As the two scatter diagrams in Figure 2-14 suggest, it is frequently desirable to
summarize the relationship between two variables by fitting a straight line through the
data points. You will learn how this is done presently, but at the moment, it can be said
that a straight line can be fitted to the points in a scatter diagram so that a “good fit”
results. A question of interest for forecasting is this: How rapidly does this straight line
rise or fall?
Answering this question requires the calculation of the slope of the line. The
slope of any straight line is defined as the change in Y associated with a one-unit
increase in X.
To summarize, when investigating a relationship between two variables, one must
first know whether the relationship is linear (illustrated by a straight line) or nonlin-
ear. If it is linear, one needs to know whether the relationship is positive or negative
and how sharply the line that fits the data points rises or falls. Finally, one needs to
know the strength of the relationship—that is, how close the data points are to the line
that best fits them.

Correlation Coefficient
The strength of the linear relationship that exists between two variables is measured by
the correlation that exists between them. The correlation coefficient measures strength
as follows. Two variables with a perfect negative relationship have a correlation coeffi-
cient equal to -1. At the other extreme, two variables with a perfect positive relation-
ship have a correlation coefficient equal to 1. Thus, the correlation coefficient varies
between -1 and 1 inclusive, depending on the amount of association between the two
variables being measured.

The correlation coefficient measures the extent to which two variables are
linearly related to each other.

The scatter diagram in Figure 2-13(a) illustrates a situation that produces a corre-
lation coefficient of 1. The scatter diagram in Figure 2-13(b) has a correlation coeffi-
cient of -1. Figures 2-13(e) and (f) demonstrate two variables that are not linearly
related. The correlation coefficients for these relationships are equal to 0; that is, no lin-
ear relationship is present.
Forecasters are concerned with both population and sampled data. In the
population that contains all of the X–Y data points of interest, there is a correlation
coefficient whose symbol is ρ, the Greek letter rho. If a random sample of these X–Y
data points is drawn, the correlation coefficient for these sample data is denoted by r.
Frequently, X and Y are measured in different units, such as pounds and dollars,
sales units and sales dollars, or unemployment rate and GNP dollars. In spite of these
differing units of measure for X and Y, it is still important to measure the extent to
which X and Y are related. This measurement is done by first converting variables X
and Y to standardized units, or Z-scores.
38 CHAPTER 2 A Review of Basic Statistical Concepts

After the X–Y measurements are converted to Z-scores, the Z-scores for each
X–Y measurement are multiplied, providing cross products for each case. These cross
products are of interest because the mean of these values is the correlation coefficient.
The calculation of the correlation coefficient as (essentially) the mean cross product of
Z-scores will produce the correct value, but in most cases, the correlation coefficient is
computed directly from the X–Y values. Equation 2.10 shows how to compute the sam-
ple correlation coefficient, r, from the Z-scores and from the X–Y measurements. Here,
Zx = 1X - X 2>SX and ZY = 1Y - Y 2>SY.

1 ©1X - X 21Y - Y 2
r = ©ZXZY =
n - 1 2©1X - X22 2©1Y - Y22
n©XY - 1©X21©Y2
= (2.10)
2n©X 2 - 1©X22 2n©Y2 - 1©Y22

A table such as Table 2-7 facilitates the calculation of r.

Example 2.10 Calculating a Sample Correlation Coefficient

If the relationship between age and income is being studied, it might be of interest to know
the value of r for these data. The required computations appear in Table 2-7. The totals from
Table 2-7 are substituted into Equation 2.10, providing

n©XY - ©X©Y
r =
2n©X - 1©X22 2n©Y2 - 1©Y22
2

514,266.12 - 113321157.72
=
2513,6432 - 113322 2515,035.052 - 1157.722

21,330.5 - 20,974.1 356.4

= = = .89
2526 2305.96 122.932117.492

It can be seen that the sample correlation coefficient confirms what was observed in
Figure 2-12. The value for r is positive, suggesting a positive linear relationship between age
and income. Also, on a scale of 0 to 1, the value of r is fairly high (.89). This result suggests a
strong linear relationship rather than a weak one. The remaining question is whether the
combination of sample size and correlation coefficient is strong enough to make meaningful
statements about the population from which the data values were drawn.

Two important points about correlation should now be made. First, it must always
be kept in mind that correlation, not causation, is being measured. It may be perfectly
valid to say that two variables are related on the basis of a high correlation coefficient.

TABLE 2-7 Computations for the Correlation of Age

and Income in Example 2.10

Person Y X Y2 X2 XY

1 27.8 22 772.84 484 611.6

2 28.5 23 812.25 529 655.5
3 30.0 26 900.00 676 780.0
4 35.0 27 1,225.00 729 945.0
5 36.4 35 1,324.96 1,225 1,274.0
Totals 157.7 133 5,035.05 3,643 4,266.1
 CHAPTER 2 A Review of Basic Statistical Concepts 39
It may or may not be valid to say that one variable is causing the movement of the
other; that is a question for the analyst’s judgment. For instance, it may be true that the
sales volume of a country store in a sparsely populated area is highly correlated with
the average stock market price in New York City. It might be concluded after examin-
ing a large sample of these two variables that such a high correlation exists. It is proba-
bly not true that one of these variables is causing the movement of the other. In fact,
the movements of both these variables are probably caused by a third factor, the gen-
eral state of the economy. Politicians, advertisers, and others frequently make the error
of assuming causation on the basis of correlation.
Second, note that the correlation coefficient is measuring a linear relationship
between two variables. In the case in which the correlation coefficient is low, it can be
concluded that the two variables are not closely related in a linear way. It may be that
they are closely related in a nonlinear or curved fashion. Thus, a low correlation coeffi-
cient does not mean that the two variables are not related, only that a linear or
straight-line relationship does not appear to exist.

FITTING A STRAIGHT LINE

As mentioned earlier, it is often of interest to fit a straight line through a collection of X–Y
data points in such a way that a “good fit” results. As will be shown in later chapters, such
a good-fitting line can be used to forecast unknown values of Y, the variable of interest.
In practice, a well-defined mathematical procedure is used to calculate both the
Y-intercept and the slope of the best-fitting straight line. The most common approach
to determining such a best-fitting line is called the method of least squares. With this
procedure, a line is constructed that minimizes the sum of the squared distances
between the data points and the line, as measured in the vertical (Y) direction.
The method of least squares chooses the values for the slope and the Y-intercept that
minimize the sum of squared errors (distances), SSE, between the Y values and the line.
N 22 = ©1Y - b0 - b1X22
SSE = ©1Y - Y (2.11)

where YN = b0 + b1 X is the estimated Y value on the fitted line.

Using calculus, specific algebraic expressions can be derived for the least squares
values. In particular,
©1X - X 21Y - Y2 n©XY - ©X©Y
b1 = =
n©X2 - 1©X22
(2.12)
©1X - X2 2

©Y b1 ©X
b0 = Y - b1X = - (2.13)
n n

where
b1 = the slope of the line
b0 = the Y-intercept

The method of least squares is used to calculate the equation of a straight line that
minimizes the sum of squared distances between the X–Y data points and the
line, as measured in the vertical (Y) direction.
40 CHAPTER 2 A Review of Basic Statistical Concepts

FIGURE 2-15 Fitted Straight Line for Example 2.11

Example 2.11 Calculating a Fitted Line Using the Method of Least Squares
Example 2.10 suggested a strong positive linear relationship between age and income.
Substituting the totals from Table 2-7 into Equations 2.12 and 2.13, the equation of a
straight line that best fits the points is computed as follows:

n©XY - ©X©Y 514,266.12 - 113321157.72 356.4

b1 = = = = .678
n©X 2 - 1©X22 513,6432 - 113322 526

©Y b1 ©X 157.7 1.678211332
b0 = - = - = 31.540 - 18.035 = 13.505
n n 5 5

The line that best fits the data, YN = 13.505 + .678 X, is shown in Figure 2-15.

The equation computed in Example 2.11, along with other values that can be
calculated from the sample data, might be profitably used by managers to forecast
future values of an important variable and to assess in advance just how accurate
such forecasts might be. In a later chapter, you will learn how to extract a great
deal of information from sample data and use it to make forecasts with regression
analysis.
The least squares slope coefficient is related to the sample correlation coefficient.
Specifically,

2©1Y - Y22
b1 = r (2.14)
2©1X - X22

As a result, b1 and r are proportional to one another and have the same sign.
 CHAPTER 2 A Review of Basic Statistical Concepts 41
Example 2.12 Demonstrating the Relationship between the Correlation Coefficient
and the Slope Coefficient
Using the results in Table 2-7 and Examples 2.10 and 2.11, we can verify Equation 2.14
numerically. Using Equation 2.14, we have

2©1Y - Y22 261.192

b1 = r = .89 = 1.76321.892 = .679
2©1X - X2 2
2105.2

This value, within rounding error, agrees with the value of the slope coefficient computed
directly in Example 2.11, using Equation 2.12.
Example 2.13
Suppose the CEO of a large construction firm suspects that the estimated expenses of his
company’s construction projects are not very close to the actual expenses at project comple-
tion. The data shown in Table 2-8 are collected for the past few projects in order to analyze
the relationship between estimated and actual costs. Since Minitab is available, the data are
analyzed using this program.
Figure 2-16 shows that the correlation between estimated and actual construction costs,
based on the sample data, is r = .912. The company CEO is surprised to learn that it is this
high.
Figure 2-17 shows the data plotted as a scatter diagram and the line that best fits these data:
YN = .683 + .922X. The CEO can now forecast an actual construction cost (Y) after the esti-
mate for the project (X) is prepared.

TABLE 2-8 Estimated and Actual Construction

Project Costs for Example 2.13

Row Actual Estimate

1 0.918 0.575
2 7.214 6.127
3 14.577 11.215
4 30.028 28.195
5 38.173 30.100
6 15.320 21.091
7 14.837 8.659
8 51.284 40.630
9 34.100 37.800
10 2.003 1.803
11 20.099 18.048
12 4.324 8.102
13 10.523 10.730
14 13.371 8.947
15 1.553 3.157
16 4.069 3.540
17 27.973 37.400
18 7.642 7.650
19 3.692 13.700
20 29.522 29.003
21 15.317 14.639
22 5.292 5.292
23 0.707 0.960
24 1.246 1.240
25 1.143 1.419
26 21.571 38.936
42 CHAPTER 2 A Review of Basic Statistical Concepts

FIGURE 2-16 Minitab Output for Example 2.13

FIGURE 2-17 Fitted Line Plot for Example 2.13

ASSESSING NORMALITY
Many statistical techniques, including some of those used in forecasting, require the
assumption that a data set follows a normal distribution. For this reason, statistical
techniques have been developed that test the hypothesis that a collection of sample
data was drawn from a normally distributed population.
Consider the monthly return rates on the Standard & Poor’s (S&P) 500 stock mar-
ket index shown in Table 2-9. Can it be assumed that these data follow a normal distri-
bution? The answer, produced by Minitab, is contained in Figure 2-18.
The straight line displayed in Figure 2-18 shows what points from a perfect normal
curve would look like if plotted using this special scale.4 As shown, the data from Table 2-9
lie very close to this line, suggesting a good fit between the S&P data and a normal
distribution.

4Other normal probability plots are available. One such plot, called a normal scores plot, is often used. In all
of these plots, normality is indicated if the data plotted lie close to a straight line.
 CHAPTER 2 A Review of Basic Statistical Concepts 43

TABLE 2-9 Monthly Rates of Return (%)

on Standard & Poor’s 500 Index

Year
1 2 3 4

Jan. — 6.87 -7.13 4.07

Feb. —
4.10 -2.94 0.85 6.51
Mar. -3.39 2.06 2.40 2.20
Apr. 0.94 4.89 -2.73 0.03
May 0.32 3.45 8.80 3.79
Jun. 4.23 -0.80 -0.89 -4.91
Jul. -0.54 8.47 -0.52 4.39
Aug. -3.94 1.54 -9.91 1.95
Sep. 3.90 -0.66 -5.25 -1.93
Oct. 2.56 -2.55 -0.67 1.18
Nov. -1.91 1.64 5.82 -4.49
Dec. 1.46 2.12 2.45 10.58

FIGURE 2-18 Normal Probability Plot for Monthly Rates of Return on

Standard & Poor’s 500 Index

The Minitab default normality test is the Anderson-Darling test, the results of
which are shown in Figure 2-18. The details of this test need not concern us at present,
but note the p-value (labeled P-Value in the box to the upper right of the figure) of
.927. In this case, the large p-value suggests the sample S&P data are consistent with
the null hypothesis (certainly not unusual if the null hypothesis is true). To reject the
null hypothesis of normality would result in almost certain error. Therefore, the null
hypothesis should not be rejected, and the assumption that the S&P data follow a
normal distribution can safely be made.
44 CHAPTER 2 A Review of Basic Statistical Concepts

APPLICATION TO MANAGEMENT
Many of the concepts in this review chapter may be considered background material
necessary for understanding the more advanced forecasting techniques found through-
out the remainder of this book. However, the concepts in this chapter also have value
themselves in many statistical applications. Although some of these applications might
not logically fall under the heading “forecasting,” they nonetheless involve using
collected data to answer questions about the uncertainties of the business operation,
especially the uncertain outcomes of the future.
The descriptive statistical procedures mentioned early in the chapter are widely
used whenever large amounts of data must be cogently described so that they can be
used in the decision-making process. It would be nearly impossible to think of a sin-
gle area involving numerical measurements in which data collections are not rou-
tinely summarized using descriptive statistics. This fact applies particularly to the
mean, commonly referred to as the “average,” and—to a somewhat lesser extent—to
the standard deviation. Averages are commonplace and have been used for many
years to provide central measurements of data arrays. The recent emphasis on qual-
ity requires an understanding of variation, and consequently, measures of dispersion,
such as the standard deviation, are appearing more and more frequently in business
practice.
The binomial and normal distributions are good examples of theoretical distribu-
tions that serve as models of many real-life situations. As such, they are widely used in
applications, including forecasting.
Estimation and hypothesis testing are the two mainstays of basic statistical infer-
ence. Forecasting, or estimating, population values of interest from a random sample is
routinely employed whenever time and cost constraints preclude an examination of all
items under consideration. Sampling is especially widespread in auditing. Hypothesis
testing is concerned with two competing statements made about population parame-
ters, so it is often used to answer such questions as these:

• How does the mean sales for this period compare to that of the previous period?
• How does the mean asset/liability ratio for healthy firms compare to those of
failed firms?
• Is the mean output per hour from this production process less than what is has been?

Correlation is widely used to examine the relationships between pairs of numerical

variables. As will be seen in later chapters, these relationships are of great importance
in forecasting because forecasting a variable often requires the values of relevant
related variables. Both simple and multiple regression analyses are forecasting tech-
niques that fall into this category.

Glossary
Binomial distribution. The binomial distribution is Correlation coefficient. The correlation coeffi-
a discrete probability distribution describing the cient measures the extent to which two variables
likelihood of X successes in n independent trials are linearly related to each other.
of a binomial experiment. Degrees of freedom. Degrees of freedom refers to
Continuous random variable. A continuous ran- the number of data items that are independent of one
dom variable can assume any value within a speci- another and that carry unique pieces of information.
fied range. These outcomes are represented Discrete random variable. A discrete random vari-
numerically by intervals of values. able can assume only values from a predetermined
 CHAPTER 2 A Review of Basic Statistical Concepts 45
set. These outcomes are often represented numeri- Point estimate. A point estimate is a single-valued
cally by integers. estimate of a population parameter.
Expected value. The expected value of a random p-Value. The p-value or significance probability is
variable is the mean value of the variable over the probability of getting at least as extreme a
many trials or observations. sample result as the one actually observed if H0 is
Interval estimate. An interval estimate is an interval true. Equivalently, the p-value may be regarded as
within which the population parameter is likely to lie. the smallest α for which the observed test statistic
Method of least squares. The method of least leads to the rejection of H0.
squares is used to calculate the equation of a Sampling distribution. A sampling distribution is
straight line that minimizes the sum of the squared the array of all possible values of a sample statistic
distances between the X–Y data points and the that can be drawn from a population for a given
line, as measured in the vertical (Y ) direction. sample size.
Normal distribution. The normal distribution has a Scatter diagram. A scatter diagram is a plot of
bell shape and is completely determined by its X–Y data points on a two-dimensional graph.
mean and standard deviation.

Key Formulas

Sample mean

©X
X = (2.1)
n

Sample standard deviation

1©X22
©X 2 -
©1X - X2 2
n
S = = (2.2)
A n - 1 Q n - 1
Sample variance

1©X22
1X - X2 2 ©X 2 -
n
S2 = = (2.3)
n - 1 n - 1

Expected value

E1X2 = © C X * P1X2 D (2.4)

Binomial probability distribution

P1X2 = a b pX11 - p2n - X

n
for X = 0, 1, Á , n (2.5)
X

Z-score

X - m
Z = (2.6)
s
46 CHAPTER 2 A Review of Basic Statistical Concepts

t-test statistic

X - m
t = (2.7)
S> 1n

Interval estimate for population mean (large sample)

S
X ; Z (2.9)
1n

Interval estimate for population mean (small sample, normal population)

S
X ; t
1n

Correlation coefficient

1 ©1X - X21Y - Y2
r = ©ZXZY =
n - 1 2©1X - X22 2©1Y - Y22
n©XY - 1©X21©Y2
= (2.10)
2n©X 2 - 1©X22 2n©Y2 - 1©Y22

Method of least squares criterion

SSE = ©1Y - YN 22 = ©1Y - b0 - b1X22 (2.11)

Fitted regression line slope equation

©1X - X2(Y - Y ) n©XY - ©X©Y

b1 = =
n©X 2 - 1©X22
(2.12)
©1X - X22

Fitted regression line Y-intercept equation

©Y b1 ©X
b0 = Y - b1X = - (2.13)
n n

Alternative regression equation slope equation

2©1Y - Y22
b1 = § ¥r (2.14)
2©1X - X22

Problems

1. Dick Hoover, owner of Modern Office Equipment, is concerned about freight costs
and clerical costs incurred on small orders. In an effort to reduce expenditures in
 CHAPTER 2 A Review of Basic Statistical Concepts 47
this area, he has decided to introduce a discount policy rewarding orders over $40 in
the hope that this will cause customers to consolidate a number of small orders into
large orders. The following data show the amounts per transaction for a sample of
28 customers:
10, 15, 20, 25, 15, 17, 41, 50, 5, 9,
12, 14, 35, 18, 19, 17, 28, 29, 11, 11,
43, 54, 7, 8, 16, 13, 37, 18
a. Compute the sample mean.
b. Compute the sample standard deviation.
c. Compute the sample variance.
d. If the policy is successful, will the mean of the distribution increase, decrease, or
remain unaffected?
e. If the policy is successful, will the standard deviation of the distribution
increase, decrease, or remain unaffected?
f. Given the data above, forecast the amount of the next customer order.
2. Sandy James thinks that housing prices have stabilized in the past few months. To
convince her boss, she intends to compare current prices with last year’s prices. She
collects 12 housing prices from the want ads:
125,900 253,000 207,500 146,950 121,450 135,450
175,000 200,000 210,950 166,700 185,000 191,950
She then calculates the mean and standard deviation of the prices she has found.
What are these two summary values?
3. A large construction company is trying to establish a useful way to view typical
profits from jobs obtained from competitive bidding. Because the jobs vary sub-
stantially in size and the final amount of the successful bid, the company has
decided to express profits as percent earnings:
Earnings
Percent earnings = 100 *
Actual construction costs
When money is lost on a project, the earnings are negative and so is the resulting
net profit. A sample of 30 jobs yields these percent earnings:

15.9 21.3 -1.8 6.6 .4 53.6 19.7 -.5 6.7 -2.3

11.9 -.3 19.0 12.8 -9.6 26.8 21.0 32.0 -.4 10.9
6.9 -8.5 3.5 3.5 -1.9 4.0 13.0 15.1 9.7 33.9

a. Calculate an estimate of the mean percent earnings for the population of jobs
or for all potential jobs.
b. Construct a 95% confidence interval for the mean percent earnings for the pop-
ulation of jobs using a large-sample argument.
c. Construct a 95% confidence interval for the mean percent earnings for the pop-
ulation of jobs assuming 30 is a small sample size. What additional assumption
do you need to make in this case?
d. Compare the two intervals in parts b and c. Explain why a sample size of 30 is
often taken as the cutoff between large and small samples.
48 CHAPTER 2 A Review of Basic Statistical Concepts

4. From data on a large sample of sales transactions, a small business owner reports
that a 95% confidence interval for the mean profit per transaction, m, is (23.41,
102.59). Use these data to determine
a. A point estimate (best guess) of the mean, m, and its 95% error margin.
b. A 90% confidence interval for the mean, m.
5. We want to forecast whether the mean number of absent days per year has
increased for our large workforce. A year ago the mean was known to be 12.1. A
recent sample of 100 employees reveals a sample mean of 13.5 with a sample
standard deviation of 1.7 days. Test at the .05 significance level to determine if the
population mean has increased or if the difference between 13.5 and 12.1 simply
represents sampling error.
6. New Horizons Airlines wants to forecast the mean number of unoccupied seats per
flight to Germany next year.To develop this forecast, the records of 49 flights are ran-
domly selected from the files for the past year, and the number of unoccupied seats is
noted for each flight. The sample mean and standard deviation are 8.1 seats and 5.7
seats, respectively. Develop a point and 95% interval estimate of the mean number of
unoccupied seats per flight during the past year. Forecast the mean number of unoccu-
pied seats per flight to Germany next year. Discuss the accuracy of this forecast.
7. Over a period of years, a toothpaste has received a mean rating of 5.9, on a 7-point
scale, for overall customer satisfaction with the product. Because of a minor unadver-
tised change in the product, there is concern that the customer satisfaction may have
changed. Suppose the satisfaction ratings from a sample of 60 customers have a mean
of 5.60 and a standard deviation of .87. Do these data indicate that the mean satisfac-
tion rating is different from 5.9? Test with  = .05. What is the p-value for the test?
8. The manager of a frozen yogurt store claims that a medium-size serving contains
an average of more than 4 ounces of yogurt. From a random sample of 14 servings,
she obtains a mean of 4.31 ounces and a standard deviation of .52 ounce. Test, with
 = .05, the manager’s claim. Find the p-value for the test. Assume that the distri-
bution of weight per serving is normal.
9. Based on past experience, the California Power Company forecasts that the mean
residential electricity usage per household will be 700 kwh next January. In
January, the company selects a simple random sample of 50 households and com-
putes a mean and standard deviation of 715 and 50, respectively. Test at the .05 sig-
nificance level to determine whether California Power’s forecast is reasonable.
Calculate and interpret the p-value for the test.
10. Population experts indicate that family size has decreased in the last few years. Ten
years ago the average family size was 2.9. Consider the population of 200 family
sizes given in Table P-10. Randomly select a sample of 30 family sizes and test the
hypothesis that the average family size has not changed in the last 10 years.
11. James Dobbins, maintenance supervisor for the Atlanta Transit Authority, would
like to determine whether there is a positive relationship between the annual
maintenance cost of a bus and its age. If a relationship exists, James feels that he
can do a better job of predicting the annual bus maintenance budget. He collects
the data shown in Table P-11.
a. Plot a scatter diagram.
b. What kind of relationship exists between these two variables?
c. Compute the correlation coefficient.
 CHAPTER 2 A Review of Basic Statistical Concepts 49

TABLE P-10
(1) 3 (35) 1 (69) 2 (102) 1 (135) 5 (168) 6
(2) 2 (36) 2 (70) 4 (103) 2 (136) 2 (169) 3
(3) 7 (37) 4 (71) 3 (104) 5 (137) 1 (170) 2
(4) 3 (38) 1 (72) 7 (105) 3 (138) 4 (171) 3
(5) 4 (39) 4 (73) 2 (106) 2 (139) 2 (172) 4
(6) 2 (40) 2 (74) 6 (107) 1 (140) 4 (173) 2
(7) 3 (41) 1 (75) 2 (108) 2 (141) 1 (174) 2
(8) 1 (42) 3 (76) 7 (109) 2 (142) 2 (175) 1
(9) 5 (43) 5 (77) 3 (110) 1 (143) 4 (176) 5
(10) 3 (44) 2 (78) 6 (111) 4 (144) 1 (177) 3
(11) 2 (45) 1 (79) 4 (112) 1 (145) 2 (178) 2
(12) 3 (46) 4 (80) 2 (113) 1 (146) 2 (179) 4
(13) 4 (47) 3 (81) 3 (114) 2 (147) 5 (180) 3
(14) 1 (48) 5 (82) 5 (115) 2 (148) 3 (181) 5
(15) 2 (49) 2 (83) 2 (116) 1 (149) 1 (182) 3
(16) 2 (50) 4 (84) 1 (117) 4 (150) 2 (183) 1
(17) 4 (51) 1 (85) 3 (118) 2 (151) 6 (184) 2
(18) 4 (52) 6 (86) 3 (119) 1 (152) 2 (185) 4
(19) 3 (53) 2 (87) 2 (120) 3 (153) 5 (186) 3
(20) 2 (54) 5 (88) 4 (121) 5 (154) 1 (187) 2
(21) 1 (55) 4 (89) 1 (122) 1 (155) 2 (188) 5
(22) 5 (56) 1 (90) 2 (123) 2 (156) 1 (189) 3
(23) 2 (57) 2 (91) 3 (124) 3 (157) 4 (190) 4
(24) 1 (58) 1 (92) 3 (125) 4 (158) 2 (191) 3
(25) 4 (59) 5 (93) 2 (126) 3 (159) 2 (192) 2
(26) 3 (60) 2 (94) 4 (127) 2 (160) 7 (193) 3
(27) 2 (61) 7 (95) 1 (128) 1 (161) 4 (194) 2
(28) 3 (62) 1 (96) 2 (129) 6 (162) 2 (195) 5
(29) 6 (63) 2 (97) 4 (130) 1 (163) 1 (196) 3
(30) 1 (64) 6 (98) 3 (131) 2 (164) 7 (197) 3
(31) 2 (65) 4 (99) 2 (132) 5 (165) 2 (198) 2
(32) 4 (66) 1 (100) 6 (133) 2 (166) 7 (199) 5
(33) 3 (67) 2 (101) 4 (134) 1 (167) 4 (200) 1
(34) 2 (68) 1

TABLE P-11

Maintenance Age
Cost ($) (years)
Bus Y X
1 859 8
2 682 5
3 471 3
4 708 9
5 1,094 11
6 224 2
7 320 1
8 651 8
9 1,049 12
50 CHAPTER 2 A Review of Basic Statistical Concepts

TABLE P-12
Shelf Space
Week Books Sold Y X
1 275 6.8
2 142 3.3
3 168 4.1
4 197 4.2
5 215 4.8
6 188 3.9
7 241 4.9
8 295 7.7
9 125 3.1
10 266 5.9
11 200 5.0

12. Anna Sheehan is the manager of the Spendwise supermarket chain. She would like
to be able to predict paperback book sales (books per week) based on the amount
of shelf display space (feet) provided. Anna gathers data for a sample of 11 weeks,
as shown in Table P-12.
a. Plot a scatter diagram.
b. What kind of relationship exists between these two variables?
c. Compute the correlation coefficient. Determine the equation of the least
squares line by calculating the slope and Y-intercept. Use this equation to fore-
cast the number of books sold if 5.2 feet of shelf space is used (i.e., X = 5.2).
13. Consider the population of 200 weekly observations presented in Table P-13. The
independent variable X is the average weekly temperature of Spokane,
Washington. The dependent variable Y is the number of shares of Sunshine Mining
Stock traded on the Spokane exchange in a week. Randomly select data for 16
weeks and compute the coefficient of correlation. (Hint: Make sure your sample is
randomly drawn from the population.) Then determine the least squares line and
forecast Y for an average weekly temperature of 63.
14. A real estate investor collects the following data on a random sample of apart-
ments on the west side of College Station, Texas.
a. Plot the data as a scatter diagram with Y = rent and X = size.
b. Determine the equation of the fitted line relating rent to size.
c. What is the estimated increase in rent for an additional square foot of space?
d. Forecast the monthly rent for an apartment with 750 square feet.

Rent Size Rent Size

($ per month) (sq. ft.) ($ per month) (sq. ft.)

720 1,000 650 800

595 900 748 960
915 1,200 685 650
760 810 755 970
1,000 1,210 815 1,000
790 860 745 1,000
880 1,135 715 1,000
845 960 885 1,180
 CHAPTER 2 A Review of Basic Statistical Concepts 51

TABLE P-13
OBS. Y X OBS. Y X OBS. Y X OBS. Y X

(1) 50 37 (51) 54 86 (101) 22 43 (151) 79 85

(2) 90 77 (52) 76 48 (102) 32 5 (152) 79 27
(3) 46 55 (53) 55 48 (103) 24 13 (153) 48 61
(4) 47 27 (54) 12 15 (104) 63 3 (154) 5 7
(5) 12 49 (55) 5 70 (105) 16 58 (155) 24 79
(6) 23 23 (56) 2 9 (106) 4 13 (156) 47 49
(7) 65 18 (57) 77 52 (107) 79 18 (157) 65 71
(8) 37 1 (58) 6 71 (108) 5 5 (158) 56 27
(9) 87 41 (59) 67 38 (109) 59 26 (159) 52 15
(10) 83 73 (60) 30 69 (110) 99 9 (160) 17 88
(11) 87 61 (61) 3 13 (111) 76 96 (161) 45 38
(12) 39 85 (62) 6 63 (112) 15 94 (162) 45 31
(13) 28 16 (63) 70 65 (113) 10 30 (163) 90 35
(14) 97 46 (64) 33 87 (114) 20 41 (164) 69 78
(15) 69 88 (65) 13 18 (115) 37 1 (165) 62 93
(16) 87 87 (66) 10 4 (116) 56 27 (166) 0 51
(17) 52 82 (67) 21 29 (117) 6 73 (167) 8 68
(18) 52 56 (68) 56 21 (118) 86 19 (168) 47 30
(19) 15 22 (69) 74 9 (119) 27 94 (169) 7 81
(20) 85 49 (70) 47 8 (120) 67 5 (170) 48 30
(21) 41 44 (71) 34 18 (121) 22 31 (171) 59 46
(22) 82 33 (72) 38 84 (122) 32 13 (172) 76 99
(23) 98 77 (73) 75 64 (123) 90 11 (173) 54 98
(24) 99 87 (74) 0 81 (124) 88 50 (174) 95 11
(25) 23 54 (75) 51 98 (125) 35 40 (175) 7 6
(26) 77 8 (76) 47 55 (126) 57 80 (176) 24 83
(27) 42 64 (77) 63 40 (127) 73 44 (177) 55 49
(28) 60 24 (78) 7 14 (128) 13 63 (178) 41 39
(29) 22 29 (79) 6 11 (129) 18 74 (179) 14 16
(30) 91 40 (80) 68 42 (130) 70 40 (180) 24 13
(31) 68 35 (81) 72 43 (131) 9 53 (181) 36 31
(32) 36 37 (82) 95 73 (132) 93 79 (182) 62 44
(33) 22 28 (83) 82 45 (133) 41 9 (183) 77 11
(34) 92 56 (84) 91 16 (134) 17 52 (184) 32 60
(35) 34 33 (85) 83 21 (135) 10 82 (185) 12 82
(36) 34 82 (86) 27 85 (136) 69 37 (186) 85 7
(37) 63 89 (87) 13 37 (137) 5 57 (187) 90 68
(38) 30 78 (88) 6 89 (138) 18 62 (188) 78 10
(39) 31 24 (89) 76 76 (139) 88 21 (189) 60 27
(40) 84 53 (90) 55 71 (140) 99 94 (190) 96 90
(41) 56 61 (91) 13 53 (141) 86 99 (191) 51 6
(42) 48 18 (92) 50 13 (142) 95 45 (192) 9 62
(43) 0 45 (93) 60 12 (143) 78 19 (193) 93 78
(44) 58 4 (94) 61 30 (144) 3 76 (194) 61 22
(45) 27 23 (95) 73 57 (145) 38 81 (195) 5 99
(46) 78 68 (96) 20 66 (146) 57 95 (196) 88 51
(47) 78 79 (97) 36 27 (147) 77 30 (197) 45 44
(48) 72 66 (98) 85 41 (148) 25 59 (198) 34 86
(49) 21 80 (99) 49 20 (149) 99 93 (199) 28 47
(50) 73 99 (100) 83 66 (150) 9 28 (200) 44 49
52 CHAPTER 2 A Review of Basic Statistical Concepts

15. Abbott & Sons needs to forecast the mean age, m, of its hourly workforce. A ran-
dom sample of personnel files is pulled with the results below. Prepare both a point
estimate and a 98% confidence interval for the mean age of the entire workforce.
Test the hypothesis H0: m  44 versus H1: m  44 at the 2% level. Are the results of
the hypothesis test consistent with the confidence interval for m? Would you
expect them to be?
X = 45.2 S = 10.3 n = 175
16. In each of the following situations, state an appropriate null hypothesis, H0, and alter-
native hypothesis, H1. Identify the parameter that you use to state the hypotheses.
a. Census Bureau data show that the mean household income in the area served
by a shopping mall is $63,700 per year. A market research firm surveys shoppers
at the mall to find out whether the mean household income of mall shoppers is
higher than that of the general population.
b. Last year the local fire department took an average of 4.3 minutes to respond to
calls. Do this year’s data show a different average response time?
c. The mean area of several thousand apartments in a new development is adver-
tised to be 1,300 square feet. A tenant group thinks that the apartments are
smaller than advertised. They hire an engineer to measure a sample of apart-
ments to test their suspicion.
17. An investor with a substantial stock portfolio sued her broker and brokerage firm
because lack of diversification in her portfolio led to poor performance. The rates of
return for the 39 months that the account was managed by the broker produced these
summary statistics: X = - 1.10%, S = 5.99% . Consider the 39 monthly returns as a
random sample from the population of returns the brokerage would generate if it
managed the account forever. Using the sample results, construct a 95% confidence
interval for the mean monthly market return. Let the S&P 500 represent the market,
and suppose the mean S&P 500 return for the same period is 0.94%. Is this a realistic
value for the population mean of the client’s account? Explain.
18. Table P-18 gives weekly wages (WAGES) in dollars and length of service (LOS) in
months, at a specific point in time, for 16 women who hold customer service jobs in
Texas banks.
TABLE P-18

Employee WAGES LOS

1 389 94
2 377 60
3 315 45
4 316 39
5 324 20
6 403 76
7 390 106
8 441 62
9 418 68
10 413 55
11 396 27
12 353 58
13 483 139
14 437 98
15 341 70
16 476 60
 CHAPTER 2 A Review of Basic Statistical Concepts 53
a. Plot the data in Table P-18 as a scatter diagram with WAGES along the vertical
(Y) axis and LOS along the horizontal (X) axis.
b. Calculate the sample correlation coefficient, r. Using the sign and magnitude of r,
describe the nature of the linear association between WAGES and LOS. Can
you think of other variables that might affect weekly wages in addition to length
of service?
c. Calculate the fitted line that can be used to predict WAGES from LOS. If LOS
is 80 months, what is the predicted WAGES?

CASES

CASE 2-1 ALCAM ELECTRONICS

Jarrick Tilby recently received a degree in business users of this component to determine if any lifetime
administration from a small university and went to records were available. Fortunately, he found three
work for Alcam Electronics, a manufacturer of vari- companies that had used the transistor in the past
ous electronic components for industry. After a few and that had limited records on component life-
weeks on the job, he was called into the office of times. In total, he received data on 38 transistors
Alcam’s owner and manager, McKennah Labrum, whose failure times were known. Since these transis-
who asked him to investigate a question regarding a tors were manufactured using the current process,
certain transistor manufactured by Alcam because he reasoned that the results of this sample could be
a large TV company was interested in a major used to make inferences about the units in inventory
purchase. and those yet to be produced.
McKennah wanted to forecast the average life- The results of the computations Jarrick per-
time of this type of transistor, a matter of great con- formed on his sample of lifetime data follow.
cern to the TV company. Units currently in stock
could represent those that would be produced over n = 38
the lifetime of the new contract, should it be
Average lifetime X = 4,805 hours
accepted.
Jarrick decided to take a random sample of the Standard deviation of lifetimes S = 675 hours
transistors in question and formulated a plan to
accomplish this task. He numbered the storage bins After finding that the sample average life-
holding the transistors, drew random bin numbers, time was only 4,805 hours, Jarrick was concerned
and sampled all transistors in each selected bin for because he knew the other supplier of components
the sample. Since each bin contained about 20 tran- was guaranteeing an average lifetime of 5,000
sistors, he selected 10 random numbers, which gave hours. Although his sample average was a bit below
him a final sample size of 205 transistors. Because he 5,000 hours, he realized that the sample size was
had selected 10 of 55 bins, he thought he had a good only 38 and that this did not constitute positive
representative sample and could use the results of proof that Alcam’s quality was inferior to that of
this sample to generalize to the entire population of the other supplier. He decided to test the hypothe-
transistors in inventory as well as to units yet to be sis that the average lifetime of all transistors was
manufactured by the same process. 5,000 hours against the alternative that it was less.
Jarrick then considered the question of the aver- Following are the calculations he performed using
age lifetime of the units. Because the unit’s lifetime a = .01:
can extend to several years, he realized that none of H0: m = 5,000
the sampled units could be tested if a timely answer
was desired. Therefore, he decided to contact several H1: m 6 5,000
54 CHAPTER 2 A Review of Basic Statistical Concepts

If S = 675, then the decision point is thought this would be good news to McKennah
Labrum and included a summary of his findings in
675 his final report. A few days after he gave his written
5,000 - 2.33 = 4,744.9 and verbal report to her, McKennah called him into
138
her office to compliment him on a good job and to
and the decision rule is as follows: share a concern she had regarding his findings. She
said, “I am concerned about the very low signifi-
If X 6 4,744.9, reject H0 cance level of your hypothesis test. You took only a
1% chance of rejecting the null hypothesis if it is
Since the sample mean (4,805) was not below true. This strikes me as very conservative. I am con-
the decision rule point for rejection (4,744.9), Jarrick cerned that we will enter into a contract and then
failed to reject the hypothesis that the mean lifetime find that our quality level does not meet the desired
of all components was equal to 5,000 hours. He 5,000-hour specification.”

QUESTION
1. How would you respond to McKennah Labrum’s
comment?

CASE 2-2 MR. TUX

John Mosby, owner of several Mr. Tux rental stores, 1998 and divides by 12). John also computes the
is interested in forecasting his monthly sales volume standard deviation for the 12 monthly values for
(see Case 1-1). As a first step, John collects monthly each year. The results are shown in Table 2-11. John
sales data for the years 1998 through 2005, as shown also decides to construct a time series plot, given in
in Table 2-10. Figure 2-19. He plots the mean monthly sales values
Next, John computes the average monthly sales on the Y-axis and time on the X-axis.
value for each year (i.e., he adds up the 12 values for

TABLE 2-10 Mr. Tux Monthly Sales Data

1998 1999 2000 2001 2002 2003 2004 2005

Jan. 6,028 16,850 15,395 27,773 31,416 51,604 58,843 71,043

Feb. 5,927 12,753 30,826 36,653 48,341 80,366 82,386 152,930
Mar. 10,515 26,901 25,589 51,157 85,651 208,938 224,803 250,559
Apr. 32,276 61,494 103,184 217,509 242,673 263,830 354,301 409,567
May 51,920 147,862 197,608 206,229 289,554 252,216 328,263 394,747
Jun. 31,294 57,990 68,600 110,081 164,373 219,566 313,647 272,874
Jul. 23,573 51,318 39,909 102,893 160,608 149,082 214,561 230,303
Aug. 36,465 53,599 91,368 128,857 176,096 213,888 337,192 375,402
Sep. 18,959 23,038 58,781 104,776 142,363 178,947 183,482 195,409
Oct. 13,918 41,396 59,679 111,036 114,907 133,650 144,618 173,518
Nov. 17,987 19,330 33,443 63,701 113,552 116,946 139,750 181,702
Dec. 15,294 22,707 53,719 82,657 127,042 164,154 184,546 258,713
 CHAPTER 2 A Review of Basic Statistical Concepts 55

TABLE 2-11 Mr. Tux Average

Monthly Sales Values
Average Standard
(Mean) Deviation
Year ($) ($)
1998 22,013 13,165
1999 44,603 35,290
2000 64,841 47,217
2001 103,610 57,197
2002 141,381 70,625
2003 169,432 63,376
2004 213,866 96,387
2005 247,231 99,153

FIGURE 2-19 Mr. Tux Mean Monthly Sales

QUESTIONS
1. What forecasting ideas come to mind when you calculated in answering this question. Based on
study John’s mean monthly sales values for the your analysis, would you encourage John to
years of his data? continue searching for a more accurate fore-
2. Suppose John draws a straight line freehand casting method? John has the latest version of
through his scatter diagram so that it “fits well” Minitab on his computer. Do you think he
and then extends this line into the future, using should use the regression analysis feature of
points along the line as his monthly forecasts. Minitab to calculate a least squares line? If he
How accurate do you think these forecasts will did, what X variable should he use to forecast
be? Use the standard deviation values John monthly sales (Y )?
56 CHAPTER 2 A Review of Basic Statistical Concepts

CASE 2-3 ALOMEGA FOOD STORES

In Example 1.1, the president of Alomega, Julie Ruth, least squares line using sales as the dependent vari-
had collected data from her company’s operations. able and monthly TV ad dollars as the predictor
She found several months of sales data along with sev- variable. The results of this run were
eral possible predictor variables (review this situation
Sales = 341,663 + .336 1monthly TV ads2
in Example 1.1). While her analysis team was working
with the data in an attempt to forecast monthly sales,
she became impatient and wondered which of the pre- r-squared = .36136%2 p-value = .000
dictor variables was best for this purpose.
Because she had a statistical program on her
Julie had to dig out her college statistics textbook
desktop computer, she decided to have a look at the
to interpret the r-squared and p-value results from her
data herself. First, she found the correlation coeffi-
printout. After reading, she recalled that r-squared
cients between the monthly sales variable and several
(which is the square of the correlation coefficient, r)
of the potential predictor variables. Specifically, she
measures the percentage of the variability in sales that
was interested in the correlations between monthly
can be explained by the variability in monthly TV ad
sales and monthly newspaper ad dollars, monthly TV
dollars (this will be explained in Chapter 6). Also, the
ad dollars, newspaper ad dollars lagged one and two
p-value indicates that the slope coefficient (.336) is
months, TV ad dollars lagged one and two months,
significant; that is, the hypothesis that it is zero in the
and her competitors’ advertising ratings. The r values
population from which the sample was drawn can be
(correlation coefficients) were as follows:
rejected with almost no chance of error.
Correlation coefficient (r) between Julie concluded that the regression equation she
variable sales and found was significant and could be used to forecast
Monthly newspaper ad dollars .45 monthly sales if the TV ad budget is known. Since
Monthly TV ad dollars .60 TV ad expenditures are under the company’s con-
Newspaper ad dollars lagged one month -.32 trol, she felt she had a good way to forecast future
Newspaper ad dollars lagged two months .21 sales. In a brief conversation with the head of her
TV ad dollars lagged one month -.06 data management department, Roger Jackson, she
TV ad dollars lagged two months .03 mentioned her findings. He replied, “Yeah, we found
Competitors’ advertising ratings -.18 that, too. But realize that TV ads explain only about
a third of sales variability. OK, 36%. We really don’t
Julie was not surprised to find that the highest think that is high enough, and we’re trying to use
correlation was between monthly sales and monthly several variables together to try and get that
TV advertising dollars (r = .60), but she was hoping r-squared value higher. Plus, we think we’re onto a
for a stronger correlation. She decided to use a method that will do a better job than regression
regression feature to calculate the equation of the analysis anyway.”

QUESTIONS
1. What do you think of Julie Ruth’s analysis? values predicted by the straight line. How might
2. Define the residuals (errors) to be the differ- you examine the residuals to decide if Julie’s
ences between the actual sales values and the straight-line representation is adequate?

Minitab Applications

The problem. In Example 2.1, a collection of ages was analyzed using descriptive
statistics.
 CHAPTER 2 A Review of Basic Statistical Concepts 57
Minitab Solution
1. Enter the variable name Ages below C1.
2. Enter the data in column C1.
3. Click on the menus shown below:

Stat>Basic Statistics>Graphical Summary

4. The Graphical Summary dialog box appears.

a. Click on Variables and select Ages (C1), as shown in Figure 2-20.
b. Click on OK on the Graphical Summary dialog box, and the summary shown in
Figure 2-21 appears.
c. To print this graph, click on the following menus:

File>Print Graph

5. Click on the menus shown below:

Stat>Basic Statistics>Display Descriptive Statistics

6. The Display Descriptive Statistics dialog box appears.

a. Click on Variables and select Ages (C1).
b. Click on OK and the summary shown in Figure 2-1 appears in the Session win-
dow. (To get additional descriptive statistics, click on Statistics and select addi-
tional descriptive measures before clicking on OK.)

FIGURE 2-20 Minitab Display of Graphical Summary

Dialog Box
58 CHAPTER 2 A Review of Basic Statistical Concepts

FIGURE 2-21 Minitab Display of Graphical Summary

c. To print the session window that contains a summary of the descriptive statis-
tics, click on the following menus:

File>Print Session Window

The results are shown in Figure 2-1.

7. The following commands will allow you to develop a histogram, dot plot, and box
plot:

Graph>Histogram
Graph>Dotplot
Graph>Boxplot

Instructions on how to use Minitab to run correlation and regression analyses are
presented at the end of Chapter 6.

Excel Applications

The problem. In Problem 1, Dick Hoover, owner of Modern Office Equipment, is con-
cerned about freight costs and clerical costs incurred on small orders.

Excel Solution
1. The Excel program is entered, and the spreadsheet screen shown in Figure 1-2
appears. Move the cursor to highlight cell A1 in the upper left corner of the
spreadsheet.
 CHAPTER 2 A Review of Basic Statistical Concepts 59

FIGURE 2-22 Excel Worksheet for Problem 1

2. Enter the first value, 10, followed by the return key, then the next data value, and
so on.
3. After all 28 data values are keyed into column A, the cursor is placed in the cell
where the results of the first calculation are desired, A30.
4. The average of the data in cells A1 through A28 is computed by placing the
formula in A30. In order to enter a formula, the = sign must proceed it. The
formula is = Average1A1:A282. Note: A30 is shown to the left of the formula bar
and = Average1A1:A282 to the right above the spreadsheet.
5. The same approach is used to compute the standard deviation. The formula 
Stdev1A1:A282 is entered into cell A31. The results are shown in Figure 2-22.
The average and standard deviation can also be computed using either the
Insert function or the Data Analysis tool. These approaches will be discussed in
later chapters.
Instructions on how to use Excel to run correlation and regression analyses are
presented at the end of Chapter 6.
If you are using Minitab or Excel, you are encouraged to try different data sets
and statistical routines to familiarize yourself with these powerful programs. The
skill you gain will be very useful as you learn about the forecasting procedures in
this text.
60 CHAPTER 2 A Review of Basic Statistical Concepts

References
Anderson, D. R., D. J. Sweeney, and T. A. Williams. Moore, D. S., G. P. McCabe, W. M. Duckworth, and
Essentials of Modern Business Statistics, 3rd ed. S. L. Sclove. The Practice of Business Statistics.
Belmont, Calif.: Thomson/South-Western, 2007. New York: Freeman, 2003.
Keller, G. Statistics for Management and Economics,
7th ed. Belmont, Calif.: Thomson/South-Western,
2005.
C H A P T E R

3 EXPLORING DATA PATTERNS

AND AN INTRODUCTION TO
FORECASTING TECHNIQUES
One of the most time-consuming and difficult parts of forecasting is the collection of
valid and reliable data. Data processing personnel are fond of using the expression
“garbage in, garbage out” (GIGO). This expression also applies to forecasting. A fore-
cast can be no more accurate than the data on which it is based. The most sophisticated
forecasting model will fail if it is applied to unreliable data.
Modern computer power and capacity have led to an accumulation of an incred-
ible amount of data on almost all subjects. The difficult task facing most forecasters
is how to find relevant data that will help solve their specific decision-making
problems.
Four criteria can be applied when determining whether data will be useful:
1. Data should be reliable and accurate. Proper care must be taken that data are col-
lected from a reliable source with proper attention given to accuracy.
2. Data should be relevant. The data must be representative of the circumstances for
which they are being used.
3. Data should be consistent. When definitions concerning data collection change,
adjustments need to be made to retain consistency in historical patterns. This
can be a problem, for example, when government agencies change the mix or
“market basket” used in determining a cost-of-living index. Years ago personal
computers were not part of the mix of products being purchased by consumers;
now they are.
4. Data should be timely. Data collected, summarized, and published on a timely
basis will be of greatest value to the forecaster. There can be too little data (not
enough history on which to base future outcomes) or too much data (data from
irrelevant historical periods far in the past).
Generally, two types of data are of interest to the forecaster. The first is data
collected at a single point in time, be it an hour, a day, a week, a month, or a quarter.
The second is observations of data made over time. When all observations are
from the same time period, we call them cross-sectional data. The objective is to
examine such data and then to extrapolate or extend the revealed relationships to
the larger population. Drawing a random sample of personnel files to study the cir-
cumstances of the employees of a company is one example. Gathering data on
the age and current maintenance cost of nine Spokane Transit Authority buses
is another. A scatter diagram such as Figure 3-1 helps us visualize the relationship
and suggests age might be used to help in forecasting the annual maintenance
budget.

61
62 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

FIGURE 3-1 Scatter Diagram of Age and Maintenance Cost

for Nine Spokane Transit Authority Buses

Cross-sectional data are observations collected at a single point in time.

Any variable that consists of data that are collected, recorded, or observed over
successive increments of time is called a time series. Monthly U.S. beer production is an
example of a time series.

A time series consists of data that are collected, recorded, or observed over suc-
cessive increments of time.

EXPLORING TIME SERIES DATA PATTERNS

One of the most important steps in selecting an appropriate forecasting method for
time series data is to consider the different types of data patterns. There are typically
four general types of patterns: horizontal, trend, seasonal, and cyclical.
When data collected over time fluctuate around a constant level or mean, a
horizontal pattern exists. This type of series is said to be stationary in its mean. Monthly
sales for a food product that do not increase or decrease consistently over an extended
period would be considered to have a horizontal pattern.
When data grow or decline over several time periods, a trend pattern exists. Figure 3-2
shows the long-term growth (trend) of a time series variable (such as housing costs)
with data points one year apart. A linear trend line has been drawn to illustrate this
growth. Although the variable housing costs have not increased every year, the move-
ment of the variable has been generally upward between periods 1 and 20. Examples of
the basic forces that affect and help explain the trend of a series are population growth,
price inflation, technological change, consumer preferences, and productivity increases.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 63

25

Cyclical Peak

20

Cost
15
Cyclical Valley
Trend Line

10

0 10 20
Year

FIGURE 3-2 Trend and Cyclical Components of an Annual

Time Series Such as Housing Costs

Many macroeconomic variables, like the U.S. gross national product (GNP),
employment, and industrial production exhibit trendlike behavior. Figure 3-10 (see
p. 74) contains another example of a time series with a prevailing trend. This figure
shows the growth of operating revenue for Sears from 1955 to 2004.

The trend is the long-term component that represents the growth or decline in
the time series over an extended period of time.

When observations exhibit rises and falls that are not of a fixed period, a cyclical
pattern exists. The cyclical component is the wavelike fluctuation around the trend that
is usually affected by general economic conditions. A cyclical component, if it exists,
typically completes one cycle over several years. Cyclical fluctuations are often influ-
enced by changes in economic expansions and contractions, commonly referred to as
the business cycle. Figure 3-2 also shows a time series with a cyclical component. The
cyclical peak at year 9 illustrates an economic expansion and the cyclical valley at year
12 an economic contraction.

The cyclical component is the wavelike fluctuation around the trend.

When observations are influenced by seasonal factors, a seasonal pattern exists.

The seasonal component refers to a pattern of change that repeats itself year after year.
For a monthly series, the seasonal component measures the variability of the series
each January, each February, and so on. For a quarterly series, there are four seasonal
elements, one for each quarter. Figure 3-3 shows that electrical usage for Washington
64 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

Electrical Usage for Washington Water Power: 1980 –1991

1,100

1,000

900

Kilowatts
800

700

600

500

1980 1982 1984 1986 1988 1990

Year

FIGURE 3-3 Electrical Usage for Washington Water Power

Company, 1980–1991

Water Power residential customers is highest in the first quarter (winter months) of
each year. Figure 3-14 (see p. 77) shows that the quarterly sales for Coastal Marine are
typically low in the first quarter of each year. Seasonal variation may reflect weather
conditions, school schedules, holidays, or length of calendar months.

The seasonal component is a pattern of change that repeats itself year after year.

EXPLORING DATA PATTERNS WITH AUTOCORRELATION ANALYSIS

When a variable is measured over time, observations in different time periods are fre-
quently related or correlated. This correlation is measured using the autocorrelation
coefficient.

Autocorrelation is the correlation between a variable lagged one or more time

periods and itself.

Data patterns, including components such as trend and seasonality, can be studied
using autocorrelations. The patterns can be identified by examining the autocorrela-
tion coefficients at different time lags of a variable.
The concept of autocorrelation is illustrated by the data presented in Table 3-1.
Note that variables Yt - 1 and Yt - 2 are actually the Y values that have been lagged by
one and two time periods, respectively. The values for March, which are shown on the
row for time period 3, are March sales, Yt = 125; February sales, Yt - 1 = 130; and
January sales, Yt - 2 = 123.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 65

TABLE 3-1 VCR Data for Example 3.1

Time Original Data Y Lagged One Y Lagged Two

t Month Yt Period Yt-1 Periods Yt-2

1
1]
January 123
2 February 130 123
3 March 125 130 123
4 April 138 125 130
5 May 145 138 125
6 June 142 145 138
7 July 141 142 145
8 August 146 141 142
9 September 147 146 141
10 October 157 147 146
11 November 150 157 147
12 December 160 150 157

Equation 3.1 is the formula for computing the lag k autocorrelation coefficient (rk)
between observations, Yt and Yt - k that are k periods apart.

a 1Yt - Y 21Yt - k - Y 2
t=k+1
rk = n k = 0, 1, 2, Á (3.1)
a 1Yt - Y 2
2
t=1

where
rk = the autocorrelation coefficient for a lag of k periods
Y = the mean of the values of the series
Yt = the observation in time period t
Yt - k = the observation k time periods earlier or at time period t - k

Example 3.1
Harry Vernon has collected data on the number of VCRs sold last year by Vernon’s Music
Store. The data are presented in Table 3-1. Table 3-2 shows the computations that lead to the
calculation of the lag 1 autocorrelation coefficient. Figure 3-4 contains a scatter diagram of
the pairs of observations (Yt, Yt - 1). It is clear from the scatter diagram that the lag 1 correla-
tion will be positive.
The lag 1 autocorrelation coefficient (r1), or the autocorrelation between Yt and Yt-1 ,
is computed using the totals from Table 3-2 and Equation 3.1. Thus,

a 1Yt-1 - Y 21Yt - Y2 843

t=1+1
r1 = n = = .572
a 1Yt - Y2
2 1,474
t=1

As suggested by the plot in Figure 3-4, positive lag 1 autocorrelation exists in this time
series. The correlation between Yt and Yt-1 or the autocorrelation for time lag 1, is .572. This
means that the successive monthly sales of VCRs are somewhat correlated with each other.
This information may give Harry valuable insights about his time series, may help him pre-
pare to use an advanced forecasting method, and may warn him about using regression
analysis with his data. All of these ideas will be discussed in subsequent chapters.
66 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

TABLE 3-2 Computation of the Lag 1 Autocorrelation Coefficient

for the Data in Table 3-1

1Yt - Y 2 1Yt - 1 - Y 2 1Y - Y 22 1Yt - Y2 (Yt–1 — Y )

—
Time t Yt Yt - 1
1
le02]
123 — –19 — 361 —
2 130 123 -12 -19 144 228
3 125 130 -17 -12 289 204
4 138 125 -4 -17 16 68
5 145 138 3 -4 9 -12
6 142 145 0 3 0 0
7 141 142 -1 0 1 0
8 146 141 4 -1 16 -4
9 147 146 5 4 25 20
10 157 147 15 5 225 75
11 150 157 8 15 64 120
12 160 150 18 8 324 144
Total 1,704 0 1,474 843
1,704
Y = = 142
12
843
r1 = = .572
1,474

The second-order autocorrelation coefficient (r2), or the correlation between Yt and

Yt-2 for Harry’s data, is also computed using Equation 3.1.

a 1Yt - Y21Yt - 2 - Y2 682

t=2+1
r2 = n = = .463
a 1Yt - Y2
2 1,474
t=1
Yt

Yt1

FIGURE 3-4 Scatter Diagram for Vernon’s Music Store Data for
Example 3.1
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 67
It appears that moderate autocorrelation exists in this time series lagged two time peri-
ods. The correlation between Yt and Yt - 2, or the autocorrelation for time lag 2, is .463. Notice
that the autocorrelation coefficient at time lag 2 (.463) is less than the autocorrelation coef-
ficient at time lag 1 (.572). Generally, as the number of time lags (k) increases, the magni-
tudes of the autocorrelation coefficients decrease.

Figure 3-5 shows a plot of the autocorrelations versus time lags for the Harry Vernon
data used in Example 3.1.The horizontal scale on the bottom of the graph shows each time
lag of interest: 1, 2, 3, and so on.The vertical scale on the left shows the possible range of an
autocorrelation coefficient, -1 to 1. The horizontal line in the middle of the graph repre-
sents autocorrelations of zero.The vertical line that extends upward above time lag 1 shows
an autocorrelation coefficient of .57, or r1 = .57. The vertical line that extends upward
above time lag 2 shows an autocorrelation coefficient of .46, or r2 = .46. The dotted lines
and the T (test) and LBQ (Ljung-Box Q) statistics displayed in the Session window will be
discussed in Examples 3.2 and 3.3. Patterns in a correlogram are used to analyze key fea-
tures of the data, a concept demonstrated in the next section. The Minitab computer pack-
age (see the Minitab Applications section at the end of the chapter for specific instructions)
can be used to compute autocorrelations and develop correlograms.

The correlogram or autocorrelation function is a graph of the autocorrelations for

various lags of a time series.

FIGURE 3-5 Correlogram or Autocorrelation Function for the Data

Used in Example 3.1
68 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

With a display such as that in Figure 3-5, the data patterns, including trend and sea-
sonality, can be studied. Autocorrelation coefficients for different time lags for a vari-
able can be used to answer the following questions about a time series:
1. Are the data random?
2. Do the data have a trend (are they nonstationary)?
3. Are the data stationary?
4. Are the data seasonal?
If a series is random, the autocorrelations between Yt and Yt - k for any time lag k
are close to zero. The successive values of a time series are not related to each other.
If a series has a trend, successive observations are highly correlated, and the auto-
correlation coefficients typically are significantly different from zero for the first sev-
eral time lags and then gradually drop toward zero as the number of lags increases.
The autocorrelation coefficient for time lag 1 will often be very large (close to 1). The
autocorrelation coefficient for time lag 2 will also be large. However, it will not be as
large as for time lag 1.
If a series has a seasonal pattern, a significant autocorrelation coefficient will occur
at the seasonal time lag or multiples of the seasonal lag. The seasonal lag is 4 for quar-
terly data and 12 for monthly data.
How does an analyst determine whether an autocorrelation coefficient is signifi-
cantly different from zero for the data of Table 3-1? Quenouille (1949) and others have
demonstrated that the autocorrelation coefficients of random data have a sampling dis-
tribution that can be approximated by a normal curve with a mean of zero and an
approximate standard deviation of 1> 1n. Knowing this, the analyst can compare the
sample autocorrelation coefficients with this theoretical sampling distribution and deter-
mine whether, for given time lags, they come from a population whose mean is zero.
Actually, some software packages use a slightly different formula, as shown in
Equation 3.2, to compute the standard deviations (or standard errors) of the autocor-
relation coefficients. This formula assumes that any autocorrelation before time lag k is
different from zero and any autocorrelation at time lags greater than or equal to k is
zero. For an autocorrelation at time lag 1, the standard error 1> 1n is used.

k-1
1 + 2 a r 2i
SE1rk2 =
i=1
(3.2)
T n

where
SE1rk2 = the standard error 1estimated standard deviation2
of the autocorrelation at time lag k
ri = the autocorrelation at time lag i
k = the time lag
n = the number of observations in the time series
This computation will be demonstrated in Example 3.2. If the series is truly random,
almost all of the sample autocorrelation coefficients should lie within a range specified
by zero, plus or minus a certain number of standard errors. At a specified confidence
level, a series can be considered random if each of the calculated autocorrelation coef-
ficients is within the interval about 0 given by 0 ± t × SE(rk), where the multiplier t is an
appropriate percentage point of a t distribution.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 69
Although testing each rk to see if it is individually significantly different from 0 is
useful, it is also good practice to examine a set of consecutive rk’s as a group. We can
use a portmanteau test to see whether the set, say, of the first 10 rk values, is signifi-
cantly different from a set in which all 10 values are zero.
One common portmanteau test is based on the Ljung-Box Q statistic (Equation
3.3). This test is usually applied to the residuals of a forecast model. If the autocor-
relations are computed from a random (white noise) process, the statistic Q has a
chi-square distribution with m (the number of time lags to be tested) degrees of
freedom. For the residuals of a forecast model, however, the statistic Q has a chi-
square distribution with the degrees of freedom equal to m minus the number of
parameters estimated in the model. The value of the Q statistic can be compared
with the chi-square table (Table B-4) to determine if it is larger than we would
expect it to be under the null hypothesis that all the autocorrelations in the set are
zero. Alternatively, the p-value generated by the test statistic Q can be computed
and interpreted. The Q statistic is given in Equation 3.3. It will be demonstrated in
Example 3.3.

m
r 2k
Q = n1n + 22 a (3.3)
k=1 n - k

where
n = the number of observations in the time series
k = the time lag
m = the number of time lags to be tested
rk = the sample autocorrelation function of the residuals lagged k time periods

Are the Data Random?

A simple random model, often called a white noise model, is displayed in Equation 3.4.
Observation Yt is composed of two parts: c, the overall level, and t, which is the ran-
dom error component. It is important to note that the t component is assumed to be
uncorrelated from period to period.

Yt = c + t (3.4)

Are the data in Table 3-1 consistent with this model? This issue will be explored in
Examples 3.2 and 3.3.

Example 3.2
A hypothesis test is developed to determine whether a particular autocorrelation coefficient
is significantly different from zero for the correlogram shown in Figure 3-5. The null and
alternative hypotheses for testing the significance of the lag 1 population autocorrelation
coefficient are

H0 : r1 = 0
H1 : r1 Z 0

If the null hypothesis is true, the test statistic

r1 - r1 r1 - 0 r1
t = = =
SE1r12 SE1r12 SE1r12
(3.5)
70 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

has a t distribution with df = n - 1. Here, n - 1 = 12 - 1 = 11, so for a 5% significance

level, the decision rule is as follows:

If t 6 2.2 or t 7 2.2, reject H0 and conclude the lag 1 autocorrelation is significantly

different from 0.

The critical values 2.2 are the upper and lower .025 points of a t distribution with 11
degrees of freedom. The standard error of r1 is SE1r12 = 11>12 = 1.083 = .289, and the
value of the test statistic becomes

r1 .572
t = = = 1.98
SE1r12 .289

Using the decision rule above, H0 : r1 = 0 cannot be rejected, since -2.2 < 1.98 < 2.2.
Notice the value of our test statistic, t = 1.98, is the same as the quantity in the Lag 1 row
under the heading T in the Minitab output in Figure 3-5. The T values in the Minitab out-
put are simply the values of the test statistic for testing for zero autocorrelation at the
various lags.
To test for zero autocorrelation at time lag 2, we consider

H0 : r2 = 0
H1 : r2 Z 0

and the test statistic

r2 - r2 r2 - 0 r2
t = = =
SE1r22 SE122 SE1r22

Using Equation 3.2,

k-1 2-1
1 + 2 a r 2i 1 + 2 a r 2i
1 + 21.57222
SE1r22 = T
i=1 i=1 1.6544
= T = = = 1.138 = .371
n n Q 12 A 12

and

.463
t = = 1.25
.371

This result agrees with the T value for Lag 2 in the Minitab output in Figure 3-5.
Using the decision rule above, H0 : r1 = 0 cannot be rejected at the .05 level, since
-2.2  1.25  2.2. An alternative way to check for significant autocorrelation is to construct,
say, 95% confidence limits centered at 0. These limits for time lags 1 and 2 are as follows:

lag 1: 0 ; t.025 * SE1r12 or 0 ; 2.21.2892 : 1-.636, .6362

lag 2: 0 ; t.025 * SE1r22 or 0 ; 2.21.3712 : 1-.816, .8162

Autocorrelation significantly different from 0 is indicated whenever a value for rk falls out-
side the corresponding confidence limits. The 95% confidence limits are shown in Figure 3-5
by the dashed lines in the graphical display of the autocorrelation function.

Example 3.3
Minitab was used to generate the time series of 40 pseudo-random three-digit numbers
shown in Table 3-3. Figure 3-6 shows a time series graph of these data. Because these data are
random (independent of one another and all from the same population), autocorrelations for
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 71

TABLE 3-3 Time Series of 40 Random Numbers for Example 3.3

t Yt t Yt t Yt t Yt

1
[Tchble03]
343 11 946 21 704 31 555
2 574 12 142 22 291 32 476
3 879 13 477 23 43 33 612
4 728 14 452 24 118 34 574
5 37 15 727 25 682 35 518
6 227 16 147 26 577 36 296
7 613 17 199 27 834 37 970
8 157 18 744 28 981 38 204
9 571 19 627 29 263 39 616
10 72 20 122 30 424 40 97

all time lags should theoretically be equal to zero. Of course, the 40 values in Table 3-3 are
only one set of a large number of possible samples of size 40. Each sample will produce
different autocorrelations. Most of these samples will produce sample autocorrelation coeffi-
cients that are close to zero. However, it is possible that a sample will produce an autocorre-
lation coefficient that is significantly different from zero just by chance.
Next, the autocorrelation function shown in Figure 3-7 is constructed using Minitab. Note
that the two dashed lines show the 95% confidence limits. Ten time lags are examined, and all
the individual autocorrelation coefficients lie within these limits. There is no reason to doubt
that each of the autocorrelations for the first 10 lags is zero. However, even though the individ-
ual sample autocorrelations are not significantly different from zero, are the magnitudes of
the first 10 rk’s as a group larger than one would expect under the hypothesis of no autocorre-
lation at any lag? This question is answered by the Ljung-Box Q (LBQ in Minitab) statistic.
If there is no autocorrelation at any lag, the Q statistic has a chi-square distribution
with, in this case, df = 10. Consequently, a large value for Q (in the tail of the chi-square dis-
tribution) is evidence against the null hypothesis. From Figure 3-7, the value of Q (LBQ) for
10 time lags is 7.75. From Table B-4, the upper .05 point of a chi-square distribution with 10
degrees of freedom is 18.31. Since 7.75 < 18.31, the null hypothesis cannot be rejected at the
5% significance level. These data are uncorrelated at any time lag, a result consistent with
the model in Equation 3.4.
Yt

FIGURE 3-6 Time Series Plot of 40 Random Numbers Used

in Example 3.3
72 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

FIGURE 3-7 Autocorrelation Function for the Data Used in Example 3.3

Do the Data Have a Trend?

If a series has a trend, a significant relationship exists between successive time series
values. The autocorrelation coefficients are typically large for the first several time lags
and then gradually drop toward zero as the number of lags increases.
A stationary time series is one whose basic statistical properties, such as the mean
and variance, remain constant over time. Consequently, a series that varies about a fixed
level (no growth or decline) over time is said to be stationary. A series that contains a
trend is said to be nonstationary. The autocorrelation coefficients for a stationary series
decline to zero fairly rapidly, generally after the second or third time lag. On the other
hand, sample autocorrelations for a nonstationary series remain fairly large for several
time periods. Often, to analyze nonstationary series, the trend is removed before addi-
tional modeling occurs. The procedures discussed in Chapter 9 use this approach.
A method called differencing can often be used to remove the trend from a nonsta-
tionary series. The VCR data originally presented in Table 3-1 are shown again in
Figure 3-8, column A. The Yt values lagged one period, Yt - 1, are shown in column B.
The differences, Yt - Yt - 1 (column A - column B), are shown in column C. For
example, the first value for the differences is Y2 - Y1 = 130 - 123 = 7 . Note the
upward growth or trend of the VCR data shown in Figure 3-9, plot A. Now observe
the stationary pattern of the differenced data in Figure 3-9, plot B. Differencing the
data has removed the trend.
Example 3.4
Maggie Trymane, an analyst for Sears, is assigned the task of forecasting operating revenue
for 2005. She gathers the data for the years 1955 to 2004, shown in Table 3-4. The data
are plotted as a time series in Figure 3-10. Notice that, although Sears operating revenues
were declining over the 2000–2004 period, the general trend over the entire 1955–2004 time
frame is up. First, Maggie computes a 95% confidence interval for the autocorrelation coef-
ficients at time lag 1 using 0 ; Z.02511> 1n2 where, for large samples, the standard normal
.025 point has replaced the corresponding t distribution percentage point:
0 ; 1.96111>502
0 ; .277
Next, Maggie runs the data on Minitab and produces the autocorrelation function
shown in Figure 3-11. Upon examination, she notices that the autocorrelations for the first
FIGURE 3-8 Excel Results for Differencing the VCR Data of Example 3.1
Yt

FIGURE 3-9 Time Series Plots of the VCR Data and the
Differenced VCR Data of Example 3.1

73
74 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

TABLE 3-4 Yearly Operating Revenue for Sears, 1955–2004, for Example 3.4

Year Yt Year Yt Year Yt Year Yt Year Yt

1955
[T]
3,307 1967 7,296 1979 17,514 1991 57,242 2003 23,253
1956 3,556 1968 8,178 1980 25,195 1992 52,345 2004 19,701
1957 3,601 1969 8,844 1981 27,357 1993 50,838
1958 3,721 1970 9,251 1982 30,020 1994 54,559
1959 4,036 1971 10,006 1983 35,883 1995 34,925
1960 4,134 1972 10,991 1984 38,828 1996 38,236
1961 4,268 1973 12,306 1985 40,715 1997 41,296
1962 4,578 1974 13,101 1986 44,282 1998 41,322
1963 5,093 1975 13,639 1987 48,440 1999 41,071
1964 5,716 1976 14,950 1988 50,251 2000 40,937
1965 6,357 1977 17,224 1989 53,794 2001 36,151
1966 6,769 1978 17,946 1990 55,972 2002 30,762

Source: Industry Surveys, various years.

FIGURE 3-10 Time Series Plot of Sears Operating Revenue

for Example 3.4

four time lags are significantly different from zero (.96, .92, .87, and .81) and that the values
then gradually drop to zero. As a final check, Maggie looks at the Q statistic for 10 time lags.
The LBQ is 300.56, which is greater than the chi-square value 18.3 (the upper .05 point of a
chi-square distribution with 10 degrees of freedom). This result indicates the autocorrela-
tions for the first 10 lags as a group are significantly different from zero. She decides that the
data are highly autocorrelated and exhibit trendlike behavior.
Maggie suspects that the series can be differenced to remove the trend and to create a
stationary series. She differences the data (see the Minitab Applications section at the end
of the chapter), and the results are shown in Figure 3-12. The differenced series shows no
evidence of a trend, and the autocorrelation function, shown in Figure 3-13, appears to sup-
port this conclusion. Examining Figure 3-13, Maggie notes that the autocorrelation coeffi-
cient at time lag 3, .32, is significantly different from zero (tested at the .05 significance
level). The autocorrelations at lags other than lag 3 are small, and the LBQ statistic for 10
lags is also relatively small, so there is little evidence to suggest the differenced data are
autocorrelated. Yet Maggie wonders whether there is some pattern in these data that can be
modeled by one of the more advanced forecasting techniques discussed in Chapter 9.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 75

FIGURE 3-11 Autocorrelation Function for Sears Operating

Revenue for Example 3.4

FIGURE 3-12 Time Series Plot of the First Differences of the

Sears Operating Revenue for Example 3.4

FIGURE 3-13 Autocorrelation Function for the First Differences

of the Sears Operating Revenue for Example 3.4
76 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

TABLE 3-5 Quarterly Sales for Coastal Marine,

1994–2006, for Example 3.5

Fiscal Year December 31 March 31 June 30 September 30

1994
[Tc
147.6 251.8 273.1 249.1
1995 139.3 221.2 260.2 259.5
1996 140.5 245.5 298.8 287.0
1997 168.8 322.6 393.5 404.3
1998 259.7 401.1 464.6 479.7
1999 264.4 402.6 411.3 385.9
2000 232.7 309.2 310.7 293.0
2001 205.1 234.4 285.4 258.7
2002 193.2 263.7 292.5 315.2
2003 178.3 274.5 295.4 286.4
2004 190.8 263.5 318.8 305.5
2005 242.6 318.8 329.6 338.2
2006 232.1 285.6 291.0 281.4

Are the Data Seasonal?

If a series is seasonal, a pattern related to the calendar repeats itself over a particular
interval of time (usually a year). Observations in the same position for different sea-
sonal periods tend to be related. If quarterly data with a seasonal pattern are analyzed,
first quarters tend to look alike, second quarters tend to look alike, and so forth, and a
significant autocorrelation coefficient will appear at time lag 4. If monthly data are
analyzed, a significant autocorrelation coefficient will appear at time lag 12. That is,
January will correlate with other Januarys, February will correlate with other
Februarys, and so on. Example 3.5 discusses a series that is seasonal.
Example 3.5
Perkin Kendell is an analyst for the Coastal Marine Corporation. He has always felt that
sales were seasonal. Perkin gathers the data shown in Table 3-5 for the quarterly sales of the
Coastal Marine Corporation from 1994 to 2006 and plots them as the time series graph
shown in Figure 3-14. Next, he computes a large-sample 95% confidence interval for the
autocorrelation coefficient at time lag 1:

0 ; 1.96 11>52
0 ; .272

Then Perkin computes the autocorrelation coefficients shown in Figure 3-15. He notes
that the autocorrelation coefficients at time lags 1 and 4 are significantly different from zero
(r1 = .39 7 .272 and r4 = .74 7 .333). He concludes that Coastal Marine sales are seasonal
on a quarterly basis.

CHOOSING A FORECASTING TECHNIQUE

This text is devoted mostly to explaining various forecasting techniques and demon-
strating their usefulness. First, the important job of choosing among several forecasting
techniques is addressed.
Some of the questions that must be considered before deciding on the most appro-
priate forecasting technique for a particular problem follow:
• Why is a forecast needed?
• Who will use the forecast?
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 77

FIGURE 3-14 Time Series Plot of Quarterly Sales for Coastal

Marine for Example 3.5

FIGURE 3-15 Autocorrelation Function for Quarterly Sales for

Coastal Marine for Example 3.5

• What are the characteristics of the available data?

• What time period is to be forecast?
• What are the minimum data requirements?
• How much accuracy is desired?
• What will the forecast cost?
To select the appropriate forecasting technique properly, the forecaster must be
able to accomplish the following:
• Define the nature of the forecasting problem.
• Explain the nature of the data under investigation.
• Describe the capabilities and limitations of potentially useful forecasting techniques.
• Develop some predetermined criteria on which the selection decision can be made.
A major factor influencing the selection of a forecasting technique is the identifica-
tion and understanding of historical patterns in the data. If trend, cyclical, or seasonal
78 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

patterns can be recognized, then techniques that are capable of effectively extrapolating
these patterns can be selected.
Forecasting Techniques for Stationary Data
A stationary series was defined earlier as one whose mean value is not changing over
time. Such situations arise when the demand patterns influencing the series are rela-
tively stable. It is important to recognize that stationary data do not necessarily vary
randomly about a mean level. Stationary series can be autocorrelated, but the nature
of the association is such that the data do not wander away from the mean for any
extended period of time. In its simplest form, forecasting a stationary series involves
using the available history of the series to estimate its mean value, which then becomes
the forecast for future periods. More-sophisticated techniques allow the first few fore-
casts to be somewhat different from the estimated mean but then revert to the mean
for additional future periods. Forecasts can be updated as new information becomes
available. Updating is useful when initial estimates are unreliable or when the stability
of the average is in question. In the latter case, updating provides some degree of
responsiveness to a potential change in the underlying level of the series.
Stationary forecasting techniques are used in the following circumstances:
• The forces generating a series have stabilized, and the environment in which the
series exists is relatively unchanging. Examples are the number of breakdowns per
week on an assembly line having a uniform production rate, the unit sales of a
product or service in the maturation stage of its life cycle, and the number of sales
resulting from a constant level of effort.
• A very simple model is needed because of a lack of data or for ease of explanation or
implementation. An example is when a business or organization is new and very
few historical data are available.
• Stability may be obtained by making simple corrections for factors such as popula-
tion growth or inflation. Examples are changing income to per capita income and
changing dollar sales to constant dollar amounts.
• The series may be transformed into a stable one. Examples are transforming a
series by taking logarithms, square roots, or differences.
• The series is a set of forecast errors from a forecasting technique that is considered
adequate. (See Example 3.7 on p. 85.)
Techniques that should be considered when forecasting stationary series include
naive methods, simple averaging methods, moving averages, and autoregressive mov-
ing average (ARMA) models (Box-Jenkins methods).
Forecasting Techniques for Data with a Trend
Simply stated, a trend in a time series is a persistent, long-term growth or decline. For a
trending time series, the level of the series is not constant. It is common for economic
time series to contain a trend.
Forecasting techniques for trending data are used in the following circumstances:
• Increased productivity and new technology lead to changes in lifestyle. Examples
are the demands for electronic components, which increased with the advent of the
computer, and railroad usage, which decreased with the advent of the airplane.
• Increasing population causes increases in demand for goods and services. Examples
are the sales revenues of consumer goods, demand for energy consumption, and use
of raw materials.
• The purchasing power of the dollar affects economic variables due to inflation.
Examples are salaries, production costs, and prices.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 79
• Market acceptance increases. An example is the growth period in the life cycle of a
new product.
Techniques that should be considered when forecasting trending series include
moving averages, Holt’s linear exponential smoothing, simple regression, growth
curves, exponential models, and autoregressive integrated moving average (ARIMA)
models (Box-Jenkins methods).

Forecasting Techniques for Seasonal Data

A seasonal series was defined earlier as a time series with a pattern of change that
repeats itself year after year. One way to develop seasonal forecasts is to estimate sea-
sonal indexes from the history of the series. For example, with monthly data, there is an
index for January, an index for February, and so forth. These indexes are then used to
include seasonality in forecasts or to remove such effects from the observed values. The
latter process is referred to as seasonally adjusting the data and is discussed with the
time series decomposition methods in Chapter 5.
Forecasting techniques for seasonal data are used in the following circumstances:
• Weather influences the variable of interest. Examples are electrical consumption,
summer and winter activities (e.g., sports such as skiing), clothing, and agricultural
growing seasons.
• The annual calendar influences the variable of interest. Examples are retail sales
influenced by holidays, three-day weekends, and school calendars.
Techniques that should be considered when forecasting seasonal series include
classical decomposition, Census X-12, Winter’s exponential smoothing, multiple
regression, and ARIMA models (Box-Jenkins methods).

Forecasting Techniques for Cyclical Series

The cyclical effect was defined earlier as the wavelike fluctuation around the trend.
Cyclical patterns are difficult to model because their patterns are typically not stable.
The up-down wavelike fluctuations around the trend rarely repeat at fixed intervals of
time, and the magnitude of the fluctuations also tends to vary. Decomposition methods
(Chapter 5) can be extended to analyze cyclical data. However, because of the irregu-
lar behavior of cycles, analyzing the cyclical component of a series, if it exists, often
requires finding coincident or leading economic indicators.
Forecasting techniques for cyclical data are used in the following circumstances:
• The business cycle influences the variable of interest. Examples are economic, mar-
ket, and competition factors.
• Shifts in popular tastes occur. Examples are fashions, music, and food.
• Shifts in population occur. Examples are wars, famines, epidemics, and natural
disasters.
• Shifts in product life cycle occur. Examples are introduction, growth, maturation
and market saturation, and decline.
Techniques that should be considered when forecasting cyclical series include clas-
sical decomposition, economic indicators, econometric models, multiple regression,
and ARIMA models (Box-Jenkins methods).
Other Factors to Consider When Choosing a Forecasting Technique
The time horizon for a forecast has a direct bearing on the selection of a forecasting
technique. For short- and intermediate-term forecasts, a variety of quantitative tech-
niques can be applied. As the forecasting horizon increases, however, a number of
80 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

these techniques become less applicable. For instance, moving averages, exponential
smoothing, and ARIMA models are poor predictors of economic turning points,
whereas econometric models are more useful. Regression models are appropriate for
the short, intermediate, and long terms. Means, moving averages, classical decomposi-
tion, and trend projections are quantitative techniques that are appropriate for the
short and intermediate time horizons. The more complex Box-Jenkins and economet-
ric techniques are also appropriate for short- and intermediate-term forecasts.
Qualitative methods are frequently used for longer time horizons (see Chapter 10).
The applicability of forecasting techniques is generally something a forecaster
bases on experience. Managers frequently need forecasts in a relatively short time.
Exponential smoothing, trend projection, regression models, and classical decomposi-
tion methods have an advantage in this situation. (See Table 3-6.)
Computer costs are no longer a significant part of technique selection. Desktop
computers (microprocessors) and forecasting software packages are becoming com-
monplace for many organizations. Due to these developments, other criteria will likely
overshadow computer cost considerations.
Ultimately, a forecast will be presented to management for approval and use in the
planning process. Therefore, ease of understanding and interpreting the results is an
important consideration. Regression models, trend projections, classical decomposi-
tion, and exponential smoothing techniques all rate highly on this criterion.
It is important to point out that the information displayed in Table 3-6 should be
used as a guide for the selection of a forecasting technique. It is good practice to try

TABLE 3-6 Choosing a Forecasting Technique

Minimal Data Requirements

Pattern of Time Type of
Method Data Horizon Model Nonseasonal Seasonal

Naive
6]
ST, T, S S TS 1
Simple averages ST S TS 30
Moving averages ST S TS 4–20
Exponential smoothing ST S TS 2
Linear exponential smoothing T S TS 3
Quadratic exponential smoothing T S TS 4
Seasonal exponential smoothing S S TS 2×s
Adaptive filtering S S TS 5×s
Simple regression T I C 10
Multiple regression C, S I C 10 × V
Classical decomposition S S TS 5×s
Exponential trend models T I, L TS 10
S-curve fitting T I, L TS 10
Gompertz models T I, L TS 10
Growth curves T I, L TS 10
Census X-12 S S TS 6×s
Box-Jenkins ST, T, C, S S TS 24 3×s
Leading indicators C S C 24
Econometric models C S C 30
Time series multiple regression T, S I, L C 6×s
Pattern of the data: ST, stationary; T, trending; S, seasonal; C, cyclical
Time horizon: S, short term (less than three months); I, intermediate term; L, long term
Type of model: TS, time series; C, causal
Seasonal: s, length of seasonality
Variable: V, number of variables
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 81
more than one forecasting method for a particular problem, holding out some recent
data, and then to compute forecasts of these holdout observations using the different
methods. The performance of the methods for these holdout test cases can be
determined using one or more of the accuracy measures defined in Equations 3.7
through 3.11, discussed below. Assuming an adequate fit to the data, the most accurate
method (the one with the smallest forecast error) is a reasonable choice for the “best”
method. It may not be the best method in the next situation.
Empirical Evaluation of Forecasting Methods
Empirical research has found that the forecast accuracy of simple methods is often as
good as that of complex or statistically sophisticated techniques (see Fildes et al., 1997;
Makridakis et al., 1993; and Makridakis and Hibon, 2000). Results of the M3–IJF
Competition, where different experts using their favorite forecasting methodology
each generated forecasts for 3,003 different time series, tended to support this finding
(Makridakis and Hibon, 2000). It would seem that the more statistically complex a
technique is, the better it should predict time series patterns. Unfortunately, estab-
lished time series patterns can and do change in the future. Thus, having a model that
best represents the historical data (the thing complex methods do well) does not neces-
sarily guarantee more accuracy in future predictions. Of course, the ability of the fore-
caster also plays an important role in the development of a good forecast.
The M3–IJF Competition was held in 1997. The forecasts produced by the various
forecasting techniques were compared across the sample of 3,003 time series, with the
accuracy assessed using a range of measures on a holdout set. The aim of the 1997
study was to check the four major conclusions of the original M-Competition on a
larger data set (see Makridakis et al., 1982). Makridakis and Hibon (2000) summarized
the latest competition as follows:
1. As discussed previously, statistically sophisticated or complex methods do not nec-
essarily produce more accurate forecasts than simpler methods.
2. Various accuracy measures produce consistent results when used to evaluate dif-
ferent forecasting methods.
3. The combination of three exponential smoothing methods outperforms, on aver-
age, the individual methods being combined and does well in comparison with
other methods.
4. The performance of the various forecasting methods depends on the length of the
forecasting horizon and the kind of data (yearly, quarterly, monthly) analyzed.
Some methods perform more accurately for short horizons, whereas others are
more appropriate for longer ones. Some methods work better with yearly data, and
others are more appropriate for quarterly and monthly data.
As part of the final selection, each technique must be evaluated in terms of its reli-
ability and applicability to the problem at hand, its cost effectiveness and accuracy
compared with competing techniques, and its acceptance by management. Table 3-6
summarizes forecasting techniques appropriate for particular data patterns. As we
have pointed out, this table represents a place to start—that is, methods to consider for
data with certain characteristics. Ultimately, any chosen method should be continu-
ously monitored to be sure it is adequately doing the job for which it was intended.

MEASURING FORECAST ERROR

Because quantitative forecasting techniques frequently involve time series data, a
mathematical notation is developed to refer to each specific time period. The letter Y
will be used to denote a time series variable unless there is more than one variable
82 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

involved. The time period associated with an observation is shown as a subscript. Thus,
Yt refers to the value of the time series at time period t. The quarterly data for the
Coastal Marine Corporation presented in Example 3.5 (see p. 76) would be denoted
Y1 = 147.6, Y2 = 251.8, Y3 = 273.1, . . . , Y52 = 281.4.
Mathematical notation must also be developed to distinguish between an actual
value of the time series and the forecast value. A ^ (hat) will be placed above a value to
indicate that it is being forecast. The forecast value for Yt is YN t. The accuracy of a fore-
casting technique is frequently judged by comparing the original series, Y1, Y2, . . . . , to
the series of forecast values, YN1, YN2, . . . .
Basic forecasting notation is summarized as follows:
Yt = the value of a time series at period t
YNt = the forecast value of Yt
et = Yt - YNt = the residual or forecast error
Several methods have been devised to summarize the errors generated by a partic-
ular forecasting technique. Most of these measures involve averaging some function of
the difference between an actual value and its forecast value. These differences
between observed values and forecast values are often referred to as residuals.

A residual is the difference between an actual observed value and its forecast
value.

Equation 3.6 is used to compute the error or residual for each forecast period.
et = Yt - YNt (3.6)
where
et = the forecast error in time period t
Yt = the actual value in time period t
YNt = the forecast value for time period t
One method for evaluating a forecasting technique uses the sum of the absolute
errors. The mean absolute deviation (MAD) measures forecast accuracy by averaging
the magnitudes of the forecast errors (the absolute values of the errors). The MAD is
in the same units as the original series and provides an average size of the “miss”
regardless of direction. Equation 3.7 shows how the MAD is computed.
1 n
MAD = ƒ Yt - YNt ƒ (3.7)
n ta
=1

The mean squared error (MSE) is another method for evaluating a forecasting
technique. Each error or residual is squared; these are then summed and divided by the
number of observations. This approach penalizes large forecasting errors, since the
errors are squared. This is important because a technique that produces moderate
errors may well be preferable to one that usually has small errors but occasionally
yields extremely large ones. The MSE is given by Equation 3.8.
1 n
MSE = 1Yt - YNt 22 (3.8)
n ta
=1

The square root of the MSE, or the root mean squared error (RMSE), is also used
to evaluate forecasting methods. The RMSE, like the MSE, penalizes large errors but
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 83
has the same units as the series being forecast so its magnitude is more easily inter-
preted. The RMSE is displayed below.
1 n
RMSE = 1Yt - YNt22 (3.9)
A n ta
=1

Sometimes it is more useful to compute the forecast errors in terms of percentages

rather than amounts. The mean absolute percentage error (MAPE) is computed by
finding the absolute error in each period, dividing this by the actual observed value for
that period, and averaging these absolute percentage errors. The final result is then
multiplied by 100 and expressed as a percentage. This approach is useful when the
error relative to the respective size of the time series value is important in evaluating
the accuracy of the forecast. The MAPE is especially useful when the Yt values are
large. The MAPE has no units of measurement (it is a percentage) and can be used to
compare the accuracy of the same or different techniques on two entirely different
series. Equation 3.10 shows how MAPE is computed.

1 n ƒ Yt - YNt ƒ
MAPE = (3.10)
n ta
=1 Yt
Notice that MAPE cannot be calculated if any of the Yt are zero.
Sometimes it is necessary to determine whether a forecasting method is biased (con-
sistently forecasting low or high). The mean percentage error (MPE) is used in these
cases. It is computed by finding the error in each period, dividing this by the actual value
for that period, and then averaging these percentage errors. The result is typically multi-
plied by 100 and expressed as a percentage. If the forecasting approach is unbiased, the
MPE will produce a number that is close to zero. If the result is a large negative percent-
age, the forecasting method is consistently overestimating. If the result is a large positive
percentage, the forecasting method is consistently underestimating. The MPE is given by

1 n 1Yt - YNt 2
MPE = (3.11)
n ta
=1 Yt

The decision to use a particular forecasting technique is based, in part, on the

determination of whether the technique will produce forecast errors that are judged to
be sufficiently small. It is certainly realistic to expect a good forecasting technique to
produce relatively small forecast errors on a consistent basis.
The five measures of forecast accuracy just described are used
• To compare the accuracy of two (or more) different techniques.
• To measure a particular technique’s usefulness or reliability.
• To help search for an optimal technique.

Example 3.6 will illustrate how each of these error measurements is computed.

Example 3.6
Table 3-7 shows the data for the daily number of customers requiring repair work, Yt, and
a forecast of these data, YNt, for Gary’s Chevron station. The forecasting technique used the
number of customers serviced in the previous period as the forecast for the current period.
This simple technique will be discussed in Chapter 4. The following computations were
employed to evaluate this model using the MAD, MSE, RMSE, MAPE, and MPE.

1 n 34
MAD = ƒ Yt - YNt ƒ = = 4.3
n ta
=1 8
84 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

1 n
1Yt - YNt22 =
188
MSE = = 23.5
n ta
=1 8

RMSE = 1MSE = 123.5 = 4.8

1 n ƒ Yt - YNt ƒ
= .0695 16.95%2
.556
MAPE = a =
n t=1 Yt 8

1 n 1Yt - YNt 2
= .0203 12.03%2
.162
MPE = =
n ta
=1 Yt 8

TABLE 3-7 Data Used in Computations to Measure Forecast

Error for Example 3.6
Forecast
Time t Customers Yt YN t
Error et |et| et2 |et|/Yt et /Yt

1
7]
58 — — — — — —
2 54 58 -4 4 16 .074 -.074
3 60 54 6 6 36 .100 .100
4 55 60 -5 5 25 .091 -.091
5 62 55 7 7 49 .113 .113
6 62 62 0 0 0 .000 .000
7 65 62 3 3 9 .046 .046
8 63 65 -2 2 4 .032 -.032
9 70 63 7 7 49 .100 .100
Totals 12 34 188 .556 .162

The MAD indicates that each forecast deviated by an average of 4.3 customers. The MSE
of 23.5 (or the RMSE of 4.8) and the MAPE of 6.95% would be compared to the MSE
(RMSE) and the MAPE for any other method used to forecast these data. Finally, the small
MPE of 2.03% indicates that the technique is not biased: Since the value is close to zero, the
technique does not consistently over- or underestimate the number of customers serviced daily.

DETERMINING THE ADEQUACY OF A FORECASTING TECHNIQUE

Before forecasting with a selected technique, the adequacy of the choice should be
evaluated. The forecaster should answer the following questions:
• Are the autocorrelation coefficients of the residuals indicative of a random series?
This question can be answered by examining the autocorrelation function for the
residuals, such as the one to be demonstrated in Example 3.7.
• Are the residuals approximately normally distributed? This question can be
answered by analyzing a histogram of the residuals or a normal probability plot.
• Do all parameter estimates have significant t ratios? The t distribution was
reviewed in Chapter 2, and applications of t ratios are demonstrated in Example
3.2 and Chapters 6 to 9.
• Is the technique simple to use, and can planners and policy makers understand it?
The basic requirement that the residual pattern be random is verified by examining
the autocorrelation coefficients of the residuals. There should be no significant autocor-
relation coefficients. Example 3.2 illustrated how the autocorrelation coefficients can be
used to determine whether a series is random. The Ljung-Box Q statistic is also used to
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 85
test that the autocorrelations for all lags up to a given lag K equal zero. Example 3.7 illus-
trates this procedure with the residuals from two fitted models.
Example 3.7
Maggie Trymane, the analyst for Sears, has been asked to forecast sales for 2005. The data
are shown in Table 3-4 for 1955 to 2004. First, Maggie tries to forecast the data using a five-
month moving average (see Chapter 4 for a description of this technique). The residuals,
the differences between the actual values and the predicted values, are computed and
stored. The autocorrelation coefficients for these residuals are shown in Figure 3-16. An
examination of these autocorrelation coefficients indicates that two are significantly differ-
ent from zero, r1 = .77 and r2 = .58. Significant autocorrelation coefficients indicate some
association or pattern in the residuals. Furthermore, the autocorrelation function itself has
a pattern of smoothly declining coefficients. Examining the first 10 autocorrelations as a
group, the Q statistic for 10 lags is 73.90, much greater than the upper .05 value of a chi-
square variable with 10 degrees of freedom, 18.3. The hypothesis that the first 10 autocor-
relations are consistent with those for a random series is clearly rejected at the 5% level.
Since one of the basic requirements for a good forecasting technique is that it give a resid-
ual or error series that is essentially random, Maggie judges the five-month moving aver-
age technique to be inadequate.
Maggie now tries Holt’s linear exponential smoothing (see Chapter 4 for a descrip-
tion of this technique). The autocorrelation function for the residual series generated
by this technique is shown in Figure 3-17. An examination of these autocorrelation

FIGURE 3-16 Autocorrelation Function for Residuals with a Pattern

for Example 3.7

FIGURE 3-17 Autocorrelation Function for Residuals

That Are Essentially Random for Example 3.7
86 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

coefficients indicates that none is significantly different from zero (at the 5% level).
The Q statistic for 10 time lags is also examined. The LBQ value of 7.40 in the Minitab
output is less than the upper .05 value of a chi-square variable with eight degrees of
freedom, 15.5. (In this case, the degrees of freedom are equal to the number of lags to be
tested minus the number of parameters in the linear exponential smoothing model that
have been fitted to the data.) As a group, the first 10 residual autocorrelations are not
unlike those for a completely random series. Maggie decides to consider Holt’s linear
exponential smoothing technique as a possible model to forecast 2005 operating revenue
for Sears.

APPLICATION TO MANAGEMENT
The concepts in this chapter provide a basis for selecting a proper forecasting tech-
nique in a given situation. Many of the most important forecasting techniques are
discussed and applied to forecasting situations in the chapters that follow. It is impor-
tant to note that in many practical situations, more than one forecasting method or
model may produce acceptable and nearly indistinguishable forecasts. In fact, it is
good practice to try several reasonable forecasting techniques. Often judgment,
based on ease of use, cost, external environmental conditions, and so forth, must
be used to select a particular set of forecasts from, say, two sets of nearly indistin-
guishable values.
The following are a few examples of situations constantly arising in the busi-
ness world where a sound forecasting technique would help the decision-making
process:

• A soft-drink company wants to project the demand for its major product over the
next two years, by month.
• A major telecommunications company wants to forecast the quarterly dividend
payments of its chief rival for the next three years.
• A university needs to forecast student credit hours by quarter for the next four
years in order to develop budget projections for the state legislature.
• A public accounting firm needs monthly forecasts of dollar billings so it can plan
for additional accounting positions and begin recruiting.
• The quality control manager of a factory that makes aluminum ingots needs a
weekly forecast of production defects for top management of the company.
• A banker wants to see the projected monthly revenue of a small bicycle manufac-
turer that is seeking a large loan to triple its output capacity.
• A federal government agency needs annual projections of average miles per gallon
of American-made cars over the next 10 years in order to make regulatory
recommendations.
• A personnel manager needs a monthly forecast of absent days for the company
workforce in order to plan overtime expenditures.
• A savings and loan company needs a forecast of delinquent loans over the next
two years in an attempt to avoid bankruptcy.
• A company that makes computer chips needs an industry forecast for the number
of personal computers sold over the next 5 years in order to plan its research and
development budget.
• An Internet company needs forecasts of requests for service over the next six
months in order to develop staffing plans for its call centers.
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 87

Glossary
Autocorrelation. Autocorrelation is the correla- Seasonal component. The seasonal component is a
tion between a variable lagged one or more peri- pattern of change that repeats itself year after
ods and itself. year.
Correlogram or autocorrelation function. The cor- Stationary series. A stationary series is one whose
relogram (autocorrelation function) is a graph of basic statistical properties, such as the mean and
the autocorrelations for various lags of a time series. variance, remain constant over time.
Cross-sectional. Cross-sectional data are observa- Time series. A time series consists of data that are
tions collected at a single point in time. collected, recorded, or observed over successive
Cyclical component. The cyclical component is the increments of time.
wavelike fluctuation around the trend. Trend. The trend is the long-term component that
Residual. A residual is the difference between an represents the growth or decline in the time series
actual observed value and its forecast value. over an extended period of time.

Key Formulas

kth-order autocorrelation coefficient

n

a 1Yt - Y21Yt - k - Y 2
t=k+1
rk = n (3.1)
a 1Yt - Y 2
2
t=1

Standard error of autocorrelation coefficient

k-1
1 + 2 a r 2i
SE1rk2 =
i=1
(3.2)
T n

Ljung-Box Q statistic
m
r k2
Q = n1n + 22 a (3.3)
k = 1n - k

Random model

Yt = c + t (3.4)

t statistic for testing the significance of lag 1 autocorrelation

r1
t =
SE1r12
(3.5)

Forecast error or residual

et = Yt - YNt (3.6)
88 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

Mean absolute deviation

1 n
MAD = ƒ Yt - YNt ƒ (3.7)
n ta
=1

Mean squared error

1 n
MSE = 1Yt - YNt 22 (3.8)
n ta
=1

Root mean squared error

1 n
RMSE = 1Yt - YN t22 (3.9)
A n ta
=1

Mean absolute percentage error

1 n ƒ Yt - YNt ƒ
MAPE = (3.10)
n ta
=1 Yt

Mean percentage error

1 n 1Yt - YNt 2
MPE = (3.11)
n ta
=1 Yt

Problems

1. Explain the differences between qualitative and quantitative forecasting techniques.

2. What is a time series?
3. Describe each of the components in a time series.
4. What is autocorrelation?
5. What does an autocorrelation coefficient measure?
6. Describe how correlograms are used to analyze autocorrelations for various lags
of a time series.
7. Indicate whether each of the following statements describes a stationary or a non-
stationary series.
a. A series that has a trend
b. A series whose mean and variance remain constant over time
c. A series whose mean value is changing over time
d. A series that contains no growth or decline
8. Descriptions are provided for several types of series: random, stationary, trending,
and seasonal. Identify the type of series that each describes.
a. The series has basic statistical properties, such as the mean and variance, that
remain constant over time.
b. The successive values of a time series are not related to each other.
c. A high relationship exists between each successive value of a series.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 89
d. A significant autocorrelation coefficient appears at time lag 4 for quarterly
data.
e. The series contains no growth or decline.
f. The autocorrelation coefficients are typically significantly different from zero
for the first several time lags and then gradually decrease toward zero as the
number of lags increases.
9. List some of the forecasting techniques that should be considered when forecasting
a stationary series. Give examples of situations in which these techniques would be
applicable.
10. List some of the forecasting techniques that should be considered when forecasting
a trending series. Give examples of situations in which these techniques would be
applicable.
11. List some of the forecasting techniques that should be considered when forecasting
a seasonal series. Give examples of situations in which these techniques would be
applicable.
12. List some of the forecasting techniques that should be considered when forecasting
a cyclical series. Give examples of situations in which these techniques would be
applicable.
13. The number of marriages in the United States is given in Table P-13. Compute the
first differences for these data. Plot the original data and the difference data as a
time series. Is there a trend in either of these series? Discuss.
14. Compute the 95% confidence interval for the autocorrelation coefficient for time
lag 1 for a series that contains 80 terms.
15. Which measure of forecast accuracy should be used in each of the following
situations?
a. The analyst needs to determine whether a forecasting method is biased.
b. The analyst feels that the size or magnitude of the forecast variable is important
in evaluating the accuracy of the forecast.
c. The analyst needs to penalize large forecasting errors.

TABLE P-13

Year Marriages (1,000s) Year Marriages (1,000s)

1985
]
2,413 1995 2,336
1986 2,407 1996 2,344
1987 2,403 1997 2,384
1988 2,396 1998 2,244
1989 2,403 1999 2,358
1990 2,443 2000 2,329
1991 2,371 2001 2,345
1992 2,362 2002 2,254
1993 2,334 2003 2,245
1994 2,362 2004 2,279

Source: Based on Statistical Abstract of the United States, various years.

90 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

16. Which of the following statements is true concerning the accuracy measures used
to evaluate forecasts?
a. The MAPE takes into consideration the magnitude of the values being forecast.
b. The MSE and RMSE penalize large errors.
c. The MPE is used to determine whether a model is systematically predicting too
high or too low.
d. The advantage of the MAD method is that it relates the size of the error to the
actual observation.
17. Allie White, the chief loan officer for the Dominion Bank, would like to ana-
lyze the bank’s loan portfolio for the years 2001 to 2006. The data are shown in
Table P-17.
a. Compute the autocorrelations for time lags 1 and 2. Test to determine whether
these autocorrelation coefficients are significantly different from zero at the
.05 significance level.
b. Use a computer program to plot the data and compute the autocorrelations for
the first six time lags. Is this time series stationary?

TABLE P-17 Quarterly Loans ($ millions) for

Dominion Bank, 2001–2006

Calendar Mar. 31 Jun. 30 Sep. 30 Dec. 31

2001
[]
2,313 2,495 2,609 2,792
2002 2,860 3,099 3,202 3,161
2003 3,399 3,471 3,545 3,851
2004 4,458 4,850 5,093 5,318
2005 5,756 6,013 6,158 6,289
2006 6,369 6,568 6,646 6,861
Source: Based on Dominion Bank records.

18. This question refers to Problem 17. Compute the first differences of the quarterly
loan data for Dominion Bank.
a. Compute the autocorrelation coefficient for time lag 1 using the differenced
data.
b. Use a computer program to plot the differenced data and compute the autocor-
relations for the differenced data for the first six time lags. Is this time series
stationary?
19. Analyze the autocorrelation coefficients for the series shown in Figures 3-18
through 3-21. Briefly describe each series.
20. An analyst would like to determine whether there is a pattern to earnings per
share for the Price Company, which operated a wholesale/retail cash-and-carry
business in several states under the name Price Club. The data are shown in Table
P-20. Describe any patterns that exist in these data.
a. Find the forecast value of the quarterly earnings per share for Price Club for
each quarter by using the naive approach (i.e., the forecast for first-quarter 1987
is the value for fourth-quarter 1986, .32).
b. Evaluate the naive forecast using MAD.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 91
c. Evaluate the naive forecast using MSE and RMSE.
d. Evaluate the naive forecast using MAPE.
e. Evaluate the naive forecast using MPE.
f. Write a memo summarizing your findings.

Autocorrelation Function for Trade

1.0
0.8
0.6
Autocorrelation

0.4
0.2
0.0
−0.2
−0.4
−0.6
−0.8
−1.0

5 10 15

FIGURE 3-18 Autocorrelation Function for Problem 19

Autocorrelation Function for Acid

1.0
0.8
0.6
Autocorrelation

0.4
0.2
0.0
−0.2
−0.4
−0.6
−0.8
−1.0

2 7 12 17

FIGURE 3-19 Autocorrelation Function for Problem 19

Autocorrelation Function for Trade First Differenced

1.0
0.8
0.6
Autocorrelation

0.4
0.2
0.0
−0.2
−0.4
−0.6
−0.8
−1.0

5 15 25

FIGURE 3-20 Autocorrelation Function for Problem 19

92 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

Autocorrelation Function for Fuel

1.0
0.8
0.6

Autocorrelation
0.4
0.2
0.0
−0.2
−0.4
−0.6
−0.8
−1.0

5 10 15

FIGURE 3-21 Autocorrelation Function for Problem 19

TABLE P-20 Quarterly Earnings per Share for

Price Club: 1986–1993 Quarter
Year 1st 2nd 3rd 4th
1986
]
.40 .29 .24 .32
1987 .47 .34 .30 .39
1988 .63 .43 .38 .49
1989 .76 .51 .42 .61
1990 .86 .51 .47 .63
1991 .94 .56 .50 .65
1992 .95 .42 .57 .60
1993 .93 .38 .37 .57

Source: The Value Line Investment Survey (New York: Value Line,
1994), p. 1646.

TABLE P-21 Weekly Sales of a Food Item

(Read across)
[1]

2649.9 2898.7 2897.8 3054.3 3888.1 3963.6 3258.9 3199.6

3504.3 2445.9 1833.9 2185.4 3367.4 1374.1 497.5 1699.0
1425.4 1946.2 1809.9 2339.9 1717.9 2420.3 1396.5 1612.1
1367.9 2176.8 2725.0 3723.7 2016.0 862.2 1234.6 1166.5
1759.5 1039.4 2404.8 2047.8 4072.6 4600.5 2370.1 3542.3
2273.0 3596.6 2615.8 2253.3 1779.4 3917.9 3329.3 1864.4
3318.9 3342.6 2131.9 3003.2

21. Table P-21 contains the weekly sales for a food item for 52 consecutive weeks.
a. Plot the sales data as a time series.
b. Do you think this series is stationary or nonstationary?
c. Using Minitab or a similar program, compute the autocorrelations of the sales
series for the first 10 time lags. Is the behavior of the autocorrelations consistent
with your choice in part b? Explain.
CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 93
22. This question refers to Problem 21.
a. Fit the random model given by Equation 3.4 to the data in Table P-21 by esti-
mating c with the sample mean Y so YNt = Y. Compute the residuals using
et = Yt - YNt - Y .
b. Using Minitab or a similar program, compute the autocorrelations of the resid-
uals from part c for the first 10 time lags. Is the random model adequate for the
sales data? Explain.

23. Table P-23 contains Southwest Airlines’ quarterly income before extraordinary
items ($MM) for the years 1988–1999.
a. Plot the income data as a time series and describe any patterns that exist.
b. Is this series stationary or nonstationary? Explain.
c. Using Minitab or a similar program, compute the autocorrelations of the
income series for the first 10 time lags. Is the behavior of the autocorrelations
consistent with your choice in part b? Explain.

24. This question refers to Problem 23.

a. Use Minitab or Excel to compute the fourth differences of the income data in
Table P-23. Fourth differences are computed by differencing observations four time
periods apart.With quarterly data, this procedure is sometimes useful for creating a
stationary series from a nonstationary series (see Chapter 9). Consequently,
the fourth differenced data will be Y5 - Y1 = 19.64 - .17 = 19.47, Y6 - Y2 =
19.24 - 15.13 = 4.11, . . . , and so forth.
b. Plot the time series of fourth differences. Does this time series appear to be sta-
tionary or nonstationary? Explain.
25. This question refers to Problem 23.
a. Consider a naive forecasting method where the first-quarter income is used to
forecast first-quarter income for the following year, second-quarter income is
used to forecast second-quarter income, and so forth. For example, a forecast of
first-quarter income for 1998 is provided by the first-quarter income for 1997,
50.87 (see Table P-23). Use this naive method to calculate forecasts of quarterly
income for the years 1998–1999.

TABLE P-23 Quarterly Income for Southwest Airlines ($MM)

Year 1st 2nd 3rd 4th
1988
[]
0.17 15.13 26.59 16.07
1989 19.64 19.24 24.57 8.11
1990 5.09 23.53 23.04 -4.58
1991 -8.21 10.57 15.72 8.84
1992 13.48 23.48 26.89 27.17
1993 24.93 42.15 48.83 38.37
1994 41.85 58.52 58.62 20.34
1995 11.83 59.72 67.72 43.36
1996 33.00 85.32 60.86 28.16
1997 50.87 93.83 92.51 80.55
1998 70.01 133.39 129.64 100.38
1999 95.85 157.76 126.98 93.80

Source: Based on Compustat Industrial Quarterly Data Base.

94 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

b. Using the forecasts in part a, calculate the MAD, RMSE, and MAPE.
c. Given the results in part b and the nature of the patterns in the income series,
do you think this naive forecasting method is viable? Can you think of another
naive method that might be better?

CASES

CASE 3-1A MURPHY BROTHERS

FURNITURE

In 1958, the Murphy brothers established a furniture many federal publications. After looking through a
store in downtown Dallas. Over the years, they were recent copy of the Survey of Current Business, she
quite successful and extended their retail coverage found the history on monthly sales for all retail
throughout the West and Midwest. By 1996, their stores in the United States and decided to use this
chain of furniture stores had become well estab- variable as a substitute for her variable of interest,
lished in 36 states. Murphy Brothers sales dollars. She reasoned that,
Julie Murphy, the daughter of one of the if she could establish accurate forecasts for national
founders, had recently joined the firm. Her sales, she could relate these forecasts to Murphy’s
father and uncle were sophisticated in many ways own sales and come up with the forecasts she
but not in the area of quantitative skills. In particu- wanted.
lar, they both felt that they could not accu- Table 3-8 shows the data that Julie collected,
rately forecast the future sales of Murphy Brothers and Figure 3-22 shows a data plot provided by Julie’s
using modern computer techniques. For this rea- computer program. Julie began her analysis by using
son, they appealed to Julie for help as part of her the computer to develop a plot of the autocorrela-
new job. tion coefficients.
Julie first considered using Murphy sales dollars After examining the autocorrelation function
as her variable but found that several years of his- produced in Figure 3-23, it was obvious to Julie that
tory were missing. She asked her father, Glen, about her data contain a trend. The early autocorrelation
this, and he told her that at the time he “didn’t think coefficients are very large, and they drop toward
it was that important.” Julie explained the impor- zero very slowly with time. To make the series
tance of past data to Glen, and he indicated that he stationary so that various forecasting methods could
would save future data. be considered, Julie decided to first difference her
Julie decided that Murphy sales were probably data to see if the trend could be removed. The auto-
closely related to national sales figures and decided correlation function for the first differenced data is
to search for an appropriate variable in one of the shown in Figure 3-24.

QUESTIONS
1. What should Julie conclude about the retail 3. What forecasting techniques should Julie try?
sales series? 4. How will Julie know which technique works
2. Has Julie made good progress toward finding a best?
forecasting technique?
 TABLE 3-8 Monthly Sales ($ billions) for All Retail Stores, 1983–1995
1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995
Jan.
]
81.3 93.1 98.8 105.6 106.4 113.6 122.5 132.6 130.9 142.1 148.4 154.6 167.0
Feb. 78.9 93.7 95.6 99.7 105.8 115.0 118.9 127.3 128.6 143.1 145.0 155.8 164.0
Mar. 93.8 103.3 110.2 114.2 120.4 131.6 141.3 148.3 149.3 154.7 164.6 184.2 192.1
Apr. 93.8 103.9 113.1 115.7 125.4 130.9 139.8 145.0 148.5 159.1 170.3 181.8 187.5
May 97.8 111.8 120.3 125.4 129.1 136.0 150.3 154.1 159.8 165.8 176.1 187.2 201.4
Jun. 100.6 112.3 115.0 120.4 129.0 137.5 149.0 153.5 153.9 164.6 175.7 190.1 202.6
Jul. 99.4 106.9 115.5 120.7 129.3 134.1 144.6 148.9 154.6 166.0 177.7 185.8 194.9
Aug. 100.1 111.2 121.1 124.1 131.5 138.7 153.0 157.4 159.9 166.3 177.1 193.8 204.2
Sep. 97.9 104.0 113.8 124.4 124.5 131.9 144.1 145.6 146.7 160.6 171.1 185.9 192.8
Oct. 100.7 109.6 115.8 123.8 128.3 133.8 142.3 151.5 152.1 168.7 176.4 189.7 194.0
Nov. 103.9 113.5 118.1 121.4 126.9 140.2 148.8 156.1 155.6 167.2 180.9 194.7 202.4
Dec. 125.8 132.3 138.6 152.1 157.2 171.0 176.5 179.7 181.0 204.1 218.3 233.3 238.0
Source: Based on Survey of Current Business, various years.

FIGURE 3-22 Time Series Graph of Monthly Sales

for All U.S. Retail Stores, 1983–1995

FIGURE 3-23 Autocorrelation Function for Monthly Sales

for All U.S. Retail Stores

95
96 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

FIGURE 3-24 Autocorrelation Function for Monthly Sales

for All U.S. Retail Stores First Differenced

TABLE 3-9 Monthly Sales for Murphy Brothers Furniture, 1992–1995

Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1992
]
4,906 5,068 4,710 4,792 4,638 4,670 4,574 4,477 4,571 4,370 4,293 3,911
1993 5,389 5,507 4,727 5,030 4,926 4,847 4,814 4,744 4,844 4,769 4,483 4,120
1994 5,270 5,835 5,147 5,354 5,290 5,271 5,328 5,164 5,372 5,313 4,924 4,552
1995 6,283 6,051 5,298 5,659 5,343 5,461 5,568 5,287 5,555 5,501 5,201 4,826

Source: Murphy Brothers sales records.

CASE 3-1B MURPHY BROTHERS FURNITURE

Glen Murphy was not happy to be chastised by his Table 3-9. He was surprised to find out that Julie did
daughter. He decided to conduct an intensive search of not share his enthusiasm. She knew that acquiring
Murphy Brothers’ records. Upon implementing an actual sales data for the past 4 years was a positive
investigation, he was excited to discover sales data for occurrence. Julie’s problem was that she was not quite
the past four years, 1992 through 1995, as shown in sure what to do with the newly acquired data.

QUESTIONS
1. What should Julie conclude about Murphy 3. Which data should Julie use to develop a fore-
Brothers’ sales data? casting model?
2. How does the pattern of the actual sales data
compare to the pattern of the retail sales data
presented in Case 3-1A?
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 97

CASE 3-2 MR. TUX

John Mosby, owner of several Mr. Tux rental stores, is 24, r12 = .68 and r24 = .42, respectively, are both sig-
beginning to forecast his most important business nificantly different from zero.
variable, monthly dollar sales (see the Mr. Tux cases Finally, John uses another computer program to
at the ends of Chapters 1 and 2). One of his employ- calculate the percentage of the variance in the origi-
ees, Virginia Perot, has gathered the sales data shown nal data explained by the trend, seasonal, and ran-
in Case 2-2. John decides to use all 96 months of data dom components.
he has collected. He runs the data on Minitab and The program calculates the percentage of the
obtains the autocorrelation function shown in Figure variance in the original data explained by the factors
3-25. Since all the autocorrelation coefficients are in the analysis:
positive and they are trailing off very slowly, John
concludes that his data have a trend. FACTOR % EXPLAINED
Next, John asks the program to compute the Data
]
100
first differences of the data. Figure 3-26 shows the Trend 6
autocorrelation function for the differenced data. Seasonal 45
Random 49
The autocorrelation coefficients for time lags 12 and

QUESTIONS
1. Summarize the results of John’s analysis in one 3. How would you explain the line “Random 49%.”?
paragraph that a manager, not a forecaster, can 4. Consider the significant autocorrelations, r12 and
understand. r24, for the differenced data. Would you conclude
2. Describe the trend and seasonal effects that that the sales first differenced have a seasonal
appear to be present in the sales data for component? If so, what are the implications for
Mr. Tux. forecasting, say, the monthly changes in sales?

FIGURE 3-25 Autocorrelation Function for Mr. Tux Data

98 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

FIGURE 3-26 Autocorrelation Function for Mr. Tux Data First Differenced

CASE 3-3 CONSUMER CREDIT COUNSELING

The Consumer Credit Counseling (CCC) operation Dorothy, who worked for a local utility and was
was described in Case 1-2. familiar with various data exploration techniques,
Marv Harnishfeger, executive director, was con- agreed to analyze the problem. She asked Marv to
cerned about the size and scheduling of staff for the provide monthly data for the number of new clients
remainder of 1993. He explained the problem to seen. Marv provided the monthly data shown in
Dorothy Mercer, recently elected president of the Table 3-10 for the number of new clients seen by
executive committee. Dorothy thought about the CCC for the period January 1985 through March
problem and concluded that CCC needed to analyze 1993. Dorothy then analyzed these data using a time
the number of new clients it acquired each month. series plot and autocorrelation analysis.

TABLE 3-10 Number of New Clients Seen by CCC from January 1985
through March 1993
Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1985
[]
182 136 99 77 75 63 87 73 83 82 74 75
1986 102 121 128 128 112 122 104 108 97 141 97 87
1987 145 103 113 150 100 131 96 92 88 118 102 98
1988 101 153 138 107 100 114 78 106 94 93 103 104
1989 150 102 151 100 100 98 97 120 98 135 141 67
1990 127 146 175 110 153 117 121 121 131 147 121 110
1991 171 185 172 168 142 152 151 141 128 151 121 126
1992 166 138 175 108 112 147 168 149 145 149 169 138
1993 152 151 199
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 99

QUESTIONS
1. Explain how Dorothy used autocorrelation 3. What type of forecasting technique did Dorothy
analysis to explore the data pattern for the num- recommend for this data set?
ber of new clients seen by CCC.
2. What did she conclude after she completed this
analysis?

CASE 3-4 ALOMEGA FOOD STORES

In Example 1.1, the president of Alomega, Julie which of the predictor variables was best for this
Ruth, had collected data from her company’s opera- purpose.
tions. She found several months of sales data along In Case 2-3, Julie investigated the relationships
with several possible predictor variables (review this between sales and the possible predictor variables.
situation in Example 1.1). While her analysis team She now realizes that this step was premature
was working with the data in an attempt to forecast because she doesn’t even know the data pattern of
monthly sales, she became impatient and wondered her sales. (See Table 3-11.)

QUESTION
1. What did Julie conclude about the data pattern
of Alomega Sales?

TABLE 3-11 Monthly Sales for 27 Alomega

Food Stores, 2003–2006

Month 2003 2004 2005 2006

Jan.
6]
425,075 629,404 655,748 455,136
Feb. 315,305 263,467 270,483 247,570
Mar. 432,101 468,612 429,480 732,005
Apr. 357,191 313,221 260,458 357,107
May 347,874 444,404 528,210 453,156
Jun. 435,529 386,986 379,856 320,103
Jul. 299,403 414,314 472,058 451,779
Aug. 296,505 253,493 254,516 249,482
Sep. 426,701 484,365 551,354 744,583
Oct. 329,722 305,989 335,826 421,186
Nov. 281,783 315,407 320,408 397,367
Dec. 166,391 182,784 276,901 269,096
100 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

CASE 3-5 SURTIDO COOKIES

GAMESA Company is the largest producer of Surtido is a mixture of different cookies in one
cookies, snacks, and pastry in Mexico and Latin presentation. Jame knows that this product is com-
America. monly served at meetings, parties, and reunions. It is
Jame Luna, an analyst for SINTEC, which is an also popular during the Christmas holidays. So Jame
important supply-chain consulting firm based in is pretty sure there will be a seasonal component in
Mexico, is working with GAMESA on an optimal the time series of monthly sales, but he is not sure if
vehicle routing and docking system for all the distri- there is a trend in sales. He decides to plot the sales
bution centers in Mexico. Forecasts of the demand series and use autocorrelation analysis to help him
for GAMESA products will help to produce a plan determine if these data have a trend and a seasonal
for the truck float and warehouse requirements component.
associated with the optimal vehicle routing and Karin, one of the members of Jame’s team, sug-
docking system. gests that forecasts of future monthly sales might be
As a starting point, Jame decides to focus on generated by simply using the average sales for each
one of the main products of the Cookies Division of month. Jame decides to test this suggestion by “hold-
GAMESA. He collects data on monthly aggregated ing out” the monthly sales for 2003 as test cases. Then
demand (in kilograms) for Surtido cookies from the average of the January sales for 2000–2002 will
January 2000 through May 2003. The results are be used to forecast January sales for 2003 and so
shown in Table 3-12. forth.

QUESTIONS
1. What should Jame conclude about the data 2. What did Jame learn about the forecast accu-
pattern of Surtido cookie sales? racy of Karin’s suggestion?

TABLE 3-12 Monthly Sales (in kgs) for Surtido Cookies,

January 2000–May 2003
Month 2000 2001 2002 2003
Jan.
[]
666,922 802,365 574,064 1,072,617
Feb. 559,962 567,637 521,469 510,005
Mar. 441,071 527,987 522,118 579,541
Apr. 556,265 684,457 716,624 771,350
May 529,673 747,335 632,066 590,556
Jun. 564,722 658,036 633,984
Jul. 638,531 874,679 543,636
Aug. 478,899 793,355 596,131
Sep. 917,045 1,819,748 1,606,798
Oct. 1,515,695 1,600,287 1,774,832
Nov. 1,834,695 2,177,214 2,102,863
Dec. 1,874,515 1,759,703 1,819,749
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 101

Minitab Applications

The problem. In Example 3.4, Maggie Trymane, an analyst for Sears, wants to forecast
sales for 2005. She needs to determine the pattern for the sales data for the years from
1955 to 2004.
Minitab Solution

1. Enter the Sears data shown in Table 3-4 into column C1. Since the data are already
stored in a file called Tab3-4.MTW, click on

File>Open Worksheet

and double-click the Minitab Ch3 file. Click on Tab3-4.MTW and Open the file.
The Sears data will appear in C1.
2. To construct an autocorrelation function, click on the following menus, as shown in
Figure 3-27:

Stat>Time Series>Autocorrelation

3. The Autocorrelation Function dialog box shown in Figure 3-28 appears.

a. Double-click on the variable Revenue and it will appear to the right of Series.
b. Enter a Title in the appropriate space and click on OK. The resulting autocorre-
lation function is shown in Figure 3-11.
4. In order to difference the data, click on the following menus:
Stat>Time Series>Differences

FIGURE 3-27 Minitab Autocorrelation Menu

102 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

FIGURE 3-28 Minitab Autocorrelation

Function Dialog Box

FIGURE 3-29 Minitab Differences Dialog Box

The Differences option is above the Autocorrelation option shown in Figure 3-27.
5. The Differences dialog box shown in Figure 3-29 appears.
a. Double-click on the variable Revenue and it will appear to the right of Series.
b. Tab to Store differences in: and enter C2. The differenced data will now appear
in the worksheet in column C2.
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 103

Excel Applications

The problem. Harry Vernon wants to use Excel to compute the autocorrelation coeffi-
cients and a correlogram for the data presented in Table 3-1 (see p. 65).

Excel Solution
1. Create a new file by clicking on the following menus:

File>New

2. Position the mouse at A1. Notice that, whenever you position the mouse on a cell,
it is highlighted. Type the heading VERNON’S MUSIC STORE. Position the
mouse at A2 and type NUMBER OF VCRS SOLD.
3. Position the mouse at A4 and type Month. Press <enter> and the A5 cell is high-
lighted. Now enter each month, starting with January in A5 and ending with
December in A16.
4. Position the mouse at B4 and type Y. Enter the data from Table 3-1, beginning in
cell B5. Position the mouse at C4 and type Z.
5. Highlight cells B4:C16 and click on the following menus:

Insert>Name>Create

In the Create Names dialog box, click on the Top row check box, and click on OK.
This step creates the name Y for the range B5:B16 and the name Z for the range
C5:C16.
6. Highlight C5 and enter the formula

=(B5-AVERAGE(Y))/STDEV(Y)

Copy C5 to the rest of the column by highlighting it and then clicking on the Fill
handle in the lower right corner and dragging it down to cell C16. With cells
C5:C16 still highlighted, click the Decrease Decimal button (shown in Figure 3-30
toward the upper middle) until three decimal places are displayed. The Decrease
Decimal button is on the Formatting taskbar. This taskbar can be displayed by
right-clicking File and then left-clicking Formatting.
7. Enter the labels LAG and ACF in cells E4 and F4. In order to examine the first six
time lags, enter the digits 1 through 6 in cells E5:E10.
8. Highlight F5 and enter the formula

=SUMPRODUCT(OFFSET(Z,E5,0,12–E5),OFFSET(Z,0,0,12-E5))/11

Highlight F5, click the Fill handle in the lower right corner, and drag it down to cell
F10. With cells F5:F10 still highlighted, click the Decrease Decimal button until
three decimal places are displayed. The results are shown in Figure 3-30.
9. To develop the autocorrelation function, highlight cells F5:F10. Click on the
ChartWizard tool (shown in Figure 3-31 toward the upper middle).
10. The ChartWizard—Step 1 to 4 dialog box appears. In step 1, select a chart type by
clicking on Column and then on Next. In the dialog box for step 2, click on the Series
dialog box. In the blank next to Name, type Corr.. Click Next and the step 3 dialog
box appears. Under Chart title, delete Corr.. Under Category (X) axis, type Time
Lags. Now click on the Data Table dialog box and on the box next to Show data
table. Click on Next to obtain the step 4 dialog box and then on Finish to produce the
104 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques

FIGURE 3-30 Excel Spreadsheet for VCR Data

FIGURE 3-31 Excel Output for Autocorrelation Function

autocorrelation function shown in Figure 3-31. Click on one of the corners of the
chart, and move it outward in order to enlarge the autocorrelation function.
11. In order to save the data for use in Chapter 9, click on

File>Save As

In the Save As dialog box, type Tab3-1 in the space to the right of File name. Click
on Save and the file will be saved as Tab3-1.xls.
 CHAPTER 3 Exploring Data Patterns and an Introduction to Forecasting Techniques 105

References
Armstrong, J. S., ed. Principles of Forecasting: A Makridakis, S., C. Chatfield, M. Hibon, M. Lawrence,
Handbook for Researchers and Practitioners. T. Mills, J. K. Ord, and L. F. Simmons. “The M2-
Norwell, Mass.: Kluwer, 2001. Competition: A Real Time Judgmentally Based
Diebold, F. X. Elements of Forecasting, 3rd ed. Forecasting Study.” International Journal of
Cincinnati, Ohio: South-Western, 2004. Forecasting 9 (1993): 5–30.
Ermer, C. M. “Cost of Error Affects the Forecasting Makridakis, S., and M. Hibon. “The M3-
Model Selection.” Journal of Business Forecasting Competition: Results, Conclusions and
9 (Spring 1991): 10–12. Implications.” International Journal of Forecasting
Fildes, R., M. Hibon, S. Makridakis, and N. Meade. 16 (2000): 451–476.
“The Accuracy of Extrapolative Forecasting Newbold, P., and T. Bos. Introductory Business and
Methods: Additional Empirical Evidence.” Economic Forecasting, 2nd ed. Cincinnati, Ohio:
International Journal of Forecasting 13 (1997): 13. South-Western, 1994.
Makridakis, S., A. Andersen, R. Carbone, R. Fildes, Quenouille, M. H. “The Joint Distribution of Serial
M. Hibon, R. Lewandowski, J. Newton, E. Parzen, Correlation Coefficients.” Annals of
and R. Winkler. “The Accuracy of Extrapolation Mathematical Statistics 20 (1949): 561–571.
(Time Series) Methods: Results of a Forecasting Wilkinson, G. F. “How a Forecasting Model Is
Competition.” Journal of Forecasting 1 (1982): Chosen.” Journal of Business Forecasting 7
111–153. (Summer 1989): 7–8.
This page intentionally left blank
C H A P T E R

4 MOVING AVERAGES AND

SMOOTHING METHODS
This chapter will describe three simple approaches to forecasting a time series: naive,
averaging, and smoothing methods. Naive methods are used to develop simple models
that assume that very recent data provide the best predictors of the future. Averaging
methods generate forecasts based on an average of past observations. Smoothing
methods produce forecasts by averaging past values of a series with a decreasing
(exponential) series of weights.
Figure 4-1 outlines the forecasting methods discussed in this chapter. Visualize
yourself on a time scale. You are at point t in Figure 4-1 and can look backward over
past observations of the variable of interest 1Yt 2 or forward into the future. After you
select a forecasting technique, you adjust it to the known data and obtain forecast val-
ues 1YNt2. Once these forecast values are available, you compare them to the known
observations and calculate the forecast error 1e t2.
A good strategy for evaluating forecasting methods involves the following steps:

1. A forecasting method is selected based on the forecaster’s analysis of and intuition

about the nature of the data.
2. The data set is divided into two sections—an initialization or fitting section and a
test or forecasting section.
3. The selected forecasting technique is used to develop fitted values for the initial-
ization part of the data.
4. The technique is used to forecast the test part of the data, and the forecasting error
is determined and evaluated (refer to Chapter 3 for a review of measures of fore-
casting accuracy).
5. A decision is made. The decision might be to use the technique in its present form,
to modify the technique, or to develop a forecast using another technique and
compare the results.

FIGURE 4-1 Forecasting Outline

You are here

Past data t Periods to be forecast
^ ^ ^
. . . Yt −3, Yt−2, Yt −1, Yt , Yt +1, Yt +2, Yt +3, . . .
where Yt is the most recent observation of a variable
^
Yt +1 is the forecast for one period in the future

107
108 CHAPTER 4 Moving Averages and Smoothing Methods

NAIVE MODELS
Often young businesses face the dilemma of trying to forecast with very small data sets.
This situation creates a real problem, since many forecasting techniques require large
amounts of data. Naive forecasts are one possible solution, since they are based solely
on the most recent information available.
Naive forecasts assume that recent periods are the best predictors of the future.
The simplest model is

YNt + 1 = Yt (4.1)

where YNt + 1 is the forecast made at time t (the forecast origin) for time t + 1.
The naive forecast for each period is the immediately preceding observation. One
hundred percent of the weight is given to the current value of the series. The naive
forecast is sometimes called the “no change” forecast. In short-term weather forecast-
ing, the “no change” forecast occurs often. Tomorrow’s weather will be much like
today’s weather.
Since the naive forecast (Equation 4.1) discards all other observations, this scheme
tracks changes very rapidly. The problem with this approach is that random fluctua-
tions are tracked as faithfully as other fundamental changes.

Example 4.1
Figure 4-2 shows the quarterly sales of saws from 2000 to 2006 for the Acme Tool Company.
The naive technique is used to forecast sales for the next quarter to be the same as for the
previous quarter. Table 4-1 shows the data from 2000 to 2006. If the data from 2000 to 2005
are used as the initialization part and the data from 2006 as the test part, the forecast for the
first quarter of 2006 is

YN24 + 1 = Y24
YN25 = 650

FIGURE 4-2 Time Series Plot for Sales of Saws for Acme
Tool Company, 2000–2006, for Example 4.1
 CHAPTER 4 Moving Averages and Smoothing Methods 109

TABLE 4-1 Sales of Saws for Acme Tool

Company, 2000–2006, for
Example 4.1

Year Quarter t Sales

2000 1 1 500
2 2 350
3 3 250
4 4 400
2001 1 5 450
2 6 350
3 7 200
4 8 300
2002 1 9 350
2 10 200
3 11 150
4 12 400
2003 1 13 550
2 14 350
3 15 250
4 16 550
2004 1 17 550
2 18 400
3 19 350
4 20 600
2005 1 21 750
2 22 500
3 23 400
4 24 650
2006 1 25 850
2 26 600
3 27 450
4 28 700

The forecasting error is determined using Equation 3.6. The error for period 25 is

e25 = Y25 - YN25 = 850 - 650 = 200

In a similar fashion, the forecast for period 26 is 850, with an error of -250. Figure 4-2 shows
that these data have an upward trend and that there appears to be a seasonal pattern (the
first and fourth quarters are relatively high), so a decision is made to modify the naive model.

Examination of the data in Example 4.1 leads us to conclude that the values are
increasing over time. When data values increase over time, they are said to be
nonstationary in level or to have a trend. If Equation 4.1 is used, the projections will be
consistently low. However, the technique can be adjusted to take trend into considera-
tion by adding the difference between this period and the last period. The forecast
equation is

YNt + 1 = Yt + 1Yt - Yt - 12 (4.2)

Equation 4.2 takes into account the amount of change that occurred between quarters.
110 CHAPTER 4 Moving Averages and Smoothing Methods

Example 4.1 (cont.)

Using Equation 4.2, the forecast for the first quarter of 2006 is

YN24 + 1 = Y24 + (Y24 - Y24 - 1)

YN25 = Y24 + (Y24 - Y23)
N
Y = 650 + (650 - 400)
25

YN25 = 650 + 250 = 900

The forecast error with this model is

e25 = Y25 - YN = 850 - 900 = - 50

For some purposes, the rate of change might be more appropriate than the
absolute amount of change. If this is the case, it is reasonable to generate forecasts
according to
Yt
YNt + 1 = Yt (4.3)
Yt - 1
Visual inspection of the data in Table 4-1 indicates that seasonal variation seems to
exist. Sales in the first and fourth quarters are typically larger than those in the second
and third quarters. If the seasonal pattern is strong, then an appropriate forecast equa-
tion for quarterly data might be
YNt + 1 = Yt - 3 (4.4)
Equation 4.4 says that in next quarter the variable will take on the same value that
it did in the corresponding quarter one year ago.
The major weakness of this approach is that it ignores everything that has occurred
since last year and also any trend. There are several ways of introducing more recent
information. For example, the analyst can combine seasonal and trend estimates and
forecast the next quarter using
(Y1 - Yt - 1) + Á + 1Yt - 3 - Yt - 42 Yt - Yt - 4
YNt - 1 = Yt - 3 + = Yt - 3 + (4.5)
4 4
where the Yt - 3 term forecasts the seasonal pattern and the remaining term averages
the amount of change for the past four quarters and provides an estimate of the trend.
The naive forecasting models in Equations 4.4 and 4.5 are given for quarterly data.
Adjustments can be made for data collected over different time periods. For monthly
data, for example, the seasonal period is 12, not 4, and the forecast for the next period
(month) given by Equation 4.4 is Y N t + 1 = Yt - 11.
It is apparent that the number and complexity of possible naive models are limited
only by the ingenuity of the analyst, but use of these techniques should be guided by
sound judgment.
Naive methods are also used as the basis for making comparisons against which
the performance of more sophisticated methods is judged.

Example 4.1 (cont.)

The forecasts for the first quarter of 2006 using Equations 4.3, 4.4, and 4.5 are

Y24 Y24
YN24 + 1 = Y24 = Y24
Y24 - 1 Y23
650 (Equation 4.3)
YN25 = 650 = 1,056
400
 CHAPTER 4 Moving Averages and Smoothing Methods 111

YN24 + 1 = Y24 - 3 = Y21

(Equation 4.4)
YN25 = Y21 = 750
1Y24 - Y24 - 12 + # # # + 1Y24 - 3 - Y24 - 42 Y24 - Y2 - 4
YN24 + 1 = Y24 - 3 + = Y24 - 3 +
4 4
1Y24 - Y202 650 - 600
YN25 = Y21 + = 750 + (Equation 4.5)
4 4
YN 25 = 750 + 12.5 = 762.5

FORECASTING METHODS BASED ON AVERAGING

Frequently, management faces the situation in which forecasts need to be updated

daily, weekly, or monthly for inventories containing hundreds or thousands of items.
Often it is not possible to develop sophisticated forecasting techniques for each item.
Instead, some quick, inexpensive, very simple short-term forecasting tools are needed
to accomplish this task.
A manager facing such a situation is likely to use an averaging or smoothing tech-
nique. These types of techniques use a form of weighted average of past observations
to smooth short-term fluctuations. The assumption underlying these techniques is that
the fluctuations in past values represent random departures from some underlying
structure. Once this structure is identified, it can be projected into the future to pro-
duce a forecast.

Simple Averages
Historical data can be smoothed in many ways. The objective is to use past data to
develop a forecasting model for future periods. In this section, the method of simple
averages is considered. As with the naive methods, a decision is made to use the first t
data points as the initialization part and the remaining data points as the test part.
Next, Equation 4.6 is used to average (compute the mean of) the initialization part of
the data and to forecast the next period.
1 t
YNt + 1 = a Yi (4.6)
t i=1
When a new observation becomes available, the forecast for the next period, YNt + 2, is the
average or the mean computed using Equation 4.6 and this new observation.
When forecasting a large number of series simultaneously (e.g., for inventory man-
agement), data storage may be an issue. Equation 4.7 solves this potential problem.
Only the most recent forecast and the most recent observation need be stored as time
moves forward.
tYNt + 1 + Yt + 1
YNt + 2 = (4.7)
t + 1
The method of simple averages is an appropriate technique when the forces gener-
ating the series to be forecast have stabilized and the environment in which the series
exists is generally unchanging. Examples of this type of series are the quantity of sales
resulting from a consistent level of salesperson effort; the quantity of sales of a product
in the mature stage of its life cycle; and the number of appointments per week
requested of a dentist, doctor, or lawyer whose patient or client base is fairly stable.
112 CHAPTER 4 Moving Averages and Smoothing Methods

A simple average uses the mean of all relevant historical observations as the
forecast for the next period.

Example 4.2
The Spokane Transit Authority operates a fleet of vans used to transport both persons with
disabilities and the elderly. A record of the gasoline purchased for this fleet of vans is shown
in Table 4-2. The actual amount of gasoline consumed by a van on a given day is determined
by the random nature of the calls and the destinations. Examination of the gasoline pur-
chases plotted in Figure 4-3 shows the data are very stable. Since the data seem stationary,
the method of simple averages is used for weeks 1 to 28 to forecast gasoline purchases for
weeks 29 and 30. The forecast for week 29 is

TABLE 4-2 Gasoline Purchases for the Spokane Transit

Authority for Example 4.2

Week Gallons Week Gallons Week Gallons

t Yt t Yt t Yt

1 275 11 302 21 310

2 291 12 287 22 299
3 307 13 290 23 285
4 281 14 311 24 250
5 295 15 277 25 260
6 268 16 245 26 245
7 252 17 282 27 271
8 279 18 277 28 282
9 264 19 298 29 302
10 288 20 303 30 285

FIGURE 4-3 Time Series Plot of Weekly Gasoline Purchases

for the Spokane Transit Authority for Example 4.2
 CHAPTER 4 Moving Averages and Smoothing Methods 113

1 28
YN28 + 1 = Yi
28 ia
=1
7,874
YN29 = = 281.2
28
The forecast error is

e29 = Y29 - YN 29 = 302 - 281.2 = 20.8

The forecast for week 30 includes one more data point (302) added to the initialization
period. The forecast using Equation 4.7 is

28YN28 + 1 + Y28 + 1 28YN29 + Y29

YN28 + 2 = =
28 + 1 29
28(281.2) + 302
YN20 = = 281.9
29
The forecast error is

e30 = Y30 - YN30 = 285 - 281.9 = 3.1

Using the method of simple averages, the forecast of gallons of gasoline purchased for
week 31 is
1 30 8,461
YN30 + 1 = a Yi = = 282
30 i = 1 30

Moving Averages
The method of simple averages uses the mean of all the data to forecast. What if the
analyst is more concerned with recent observations? A constant number of data points
can be specified at the outset and a mean computed for the most recent observations.
The term moving average is used to describe this approach. As each new observation
becomes available, a new mean is computed by adding the newest value and dropping
the oldest. This moving average is then used to forecast the next period. Equation 4.8
gives the simple moving average forecast. A moving average of order k, MA(k), is
computed by

Yt + Yt - 1 + # # # + Yt - k + 1
YNt + 1 = (4.8)
k
where
YNt + 1 = the forecast value for the next period
Yt = the actual value at period t
k = the number of terms in the moving average
The moving average for time period t is the arithmetic mean of the k most recent
observations. In a moving average, equal weights are assigned to each observation.
Each new data point is included in the average as it becomes available, and the earliest
data point is discarded. The rate of response to changes in the underlying data pattern
depends on the number of periods, k, included in the moving average.
Note that the moving average technique deals only with the latest k periods of
known data; the number of data points in each average does not change as time
advances. The moving average model does not handle trend or seasonality very well,
although it does better than the simple average method.
114 CHAPTER 4 Moving Averages and Smoothing Methods

The analyst must choose the number of periods, k, in a moving average. A moving
average of order 1, MA(1), would take the current observation, Yt, and use it to forecast
Y for the next period. This is simply the naive forecasting approach of Equation 4.1.

A moving average of order k is the mean value of k consecutive observations. The

most recent moving average value provides a forecast for the next period.

Example 4.3
Table 4-3 demonstrates the moving average forecasting technique with the Spokane Transit
Authority data, using a five-week moving average.The moving average forecast for week 29 is

Y28 + Y28 - 1 + # # # + Y28 - 5 + 1

YN28 + 1 =
5

Y28 + Y27 + Y26 + Y25 + Y24

YN29 =
5

TABLE 4-3 Gasoline Purchases for the Spokane

Transit Authority for Example 4.3

t Gallons Ynt et

1 275 — —
2 291 — —
3 307 — —
4 281 — —
5 295 — —
6 268 289.8 -21.8
7 252 288.4 -36.4
8 279 280.6 -1.6
9 264 275.0 -11.0
10 288 271.6 16.4
11 302 270.2 31.8
12 287 277.0 10.0
13 290 284.0 6.0
14 311 286.2 24.8
15 277 295.6 -18.6
16 245 293.4 -48.4
17 282 282.0 0.0
18 277 281.0 -4.0
19 298 278.4 19.6
20 303 275.8 27.2
21 310 281.0 29.0
22 299 294.0 5.0
23 285 297.4 -12.4
24 250 299.0 -49.0
25 260 289.4 -29.4
26 245 280.8 -35.8
27 271 267.8 3.2
28 282 262.2 19.8
29 302 261.6 40.4
30 285 272.0 13.0
 CHAPTER 4 Moving Averages and Smoothing Methods 115
282 + 271 + 245 + 260 + 250 1,308
YN29 = = = 261.6
5 5

When the actual value for week 29 is known, the forecast error is calculated:

e29 = Y29 - YN29 = 302 - 261.6 = 40.4

The forecast for week 31 is

Y30 + Y30 - 1 + Á + Y30 - 5 + 1
YN30 + 1 =
5
Y30 + Y29 + Y28 + Y27 + Y26
YN31 =
5
285 + 302 + 282 + 271 + 245 1,385
YN31 = = = 277
5 5

Minitab can be used to compute a five-week moving average (see the Minitab
Applications section at the end of the chapter for instructions). Figure 4-4 shows the five-
week moving average plotted against the actual data; the MAPE, MAD, and MSD; and the
basic Minitab instructions. Note that Minitab calls the mean squared error MSD (mean
squared deviation).
Figure 4-5 shows the autocorrelation function for the residuals from the five-week
moving average method. Error limits for the individual autocorrelations centered at zero
and the Ljung-Box Q statistic (with six degrees of freedom, since no model parameters are
estimated) indicate that significant residual autocorrelation exists. That is, the residuals are
not random. The association contained in the residuals at certain time lags can be used to
improve the forecasting model.

The analyst must use judgment when determining how many days, weeks, months,
or quarters on which to base the moving average. The smaller the number, the larger
the weight given to recent periods. Conversely, the larger the number, the smaller the
weight given to more recent periods. A small number is most desirable when there are

FIGURE 4-4 Five-Week Moving Average Applied to

Weekly Gasoline Purchases for the Spokane
Transit Authority for Example 4.3
116 CHAPTER 4 Moving Averages and Smoothing Methods

FIGURE 4-5 Autocorrelation Function for the Residuals When a Five-Week

Moving Average Method Was Used with the Spokane Transit
Authority Data for Example 4.3

sudden shifts in the level of the series. A small number places heavy weight on recent
history, which enables the forecasts to catch up more rapidly to the current level.
A large number is desirable when there are wide, infrequent fluctuations in the series.
Moving averages are frequently used with quarterly or monthly data to help smooth
the components within a time series, as shown in Chapter 5. For quarterly data, a four-
quarter moving average, MA(4), yields an average of the four quarters, and for monthly
data, a 12-month moving average, MA(12), eliminates or averages out the seasonal
effects. The larger the order of the moving average, the greater the smoothing effect.
In Example 4.3, the moving average technique was used with stationary data. In
Example 4.4, we show what happens when the moving average method is used with
trending data. The double moving average technique, which is designed to handle
trending data, is introduced next.

Double Moving Averages

One way of forecasting time series data that have a linear trend is to use double mov-
ing averages. This method does what the name implies: One set of moving averages is
computed, and then a second set is computed as a moving average of the first set.
Table 4-4 shows the weekly rentals data for the Movie Video Store along with the
results of using a three-week moving average to forecast future sales. Examination
of the error column in Table 4-4 shows that every entry is positive, signifying that the
forecasts do not catch up to the trend. The three-week moving average and the double
moving average for these data are shown in Figure 4-6. Note how the three-week mov-
ing averages lag behind the actual values for comparable periods. This illustrates what
happens when the moving average technique is used with trending data. Note also that
the double moving averages lag behind the first set about as much as the first set lags
behind the actual values. The difference between the two sets of moving averages is
added to the three-week moving average to forecast the actual values.
Equations 4.9 through 4.12 summarize double moving average construction. First,
Equation 4.8 is used to compute the moving average of order k.
Yt + Yt - 1 + Yt - 2 + Á + Yt - k + 1
Mt = YNt + 1 =
k
 CHAPTER 4 Moving Averages and Smoothing Methods 117

TABLE 4-4 Weekly Rentals for the Movie Video Store for
Example 4.4

Weekly Units Three-Week Moving Average

t Rented Yt Moving Total Forecast YN t + 1 e

1 654 — — —
2 658 — — —
3 665 1,977 — —
4 672 1,995 659 13
5 673 2,010 665 8
6 671 2,016 670 1
7 693 2,037 672 21
8 694 2,058 679 15
9 701 2,088 686 15
10 703 2,098 696 7
11 702 2,106 699 3
12 710 2,115 702 8
13 712 2,124 705 7
14 711 2,133 708 3
15 728 2,151 711 17
16 — — 717 —

MSE  133

730 Rentals

720
Moving
Average
710

700
Rentals

690 Double
Moving
680 Average

670

660

650
5 10 15
Week

FIGURE 4-6 Three-Week Single and Double Moving Averages for the
Movie Video Store Data for Example 4.4

Then Equation 4.9 is used to compute the second moving average.

Mt + Mt - 1 + Mt - 2 + Á + Mt - k + 1
M t¿ = (4.9)
k
Equation 4.10 is used to develop a forecast by adding to the single moving average the
difference between the single and the second moving averages.
at = Mt + 1Mt - M t¿2 = 2Mt - M t¿ (4.10)
118 CHAPTER 4 Moving Averages and Smoothing Methods

Equation 4.11 is an additional adjustment factor, which is similar to a slope measure

that can change over the series.

1M - M t¿2
2
bt = (4.11)
k - 1 t
Finally, Equation 4.12 is used to make the forecast p periods into the future.

YNt+p = at + bt p (4.12)

where
k = the number of periods in the moving average
p = the number of periods ahead to be forecast

Example 4.4
The Movie Video Store operates several videotape rental outlets in Denver, Colorado. The
company is growing and needs to expand its inventory to accommodate the increasing
demand for its services. The president of the company assigns Jill Ottenbreit to forecast
rentals for the next month. Rental data for the last 15 weeks are available and are presented
in Table 4-5. At first, Jill attempts to develop a forecast using a three-week moving average.
The MSE for this model is 133 (see Table 4-4). Because the data are obviously trending, she
finds that her forecasts are consistently underestimating actual rentals. For this reason, she
decides to try a double moving average. The results are presented in Table 4-5. To under-
stand the forecast for week 16, the computations are presented next. Equation 4.8 is used to
compute the three-week moving average (column 3).
Y15 + Y15-1 + Y15-3+1
M15 = YN15+1 =
3
728 + 711 + 712
M15 = YN16 = = 717
3

TABLE 4-5 Double Moving Average Forecast for the Movie Video Store for
Example 4.4

(1) (2) (3) Three-Week (4) Double (5) (6) (7)

Time Weekly Moving Moving Average Value Value Forecast (8)
t Sales Yt Average Mt Mt ¿ of a of b a + bp (p = 1) et

1 654 — — — — — —
2 658 — — — — — —
3 665 659 — — — — —
4 672 665 — — — — —
5 673 670 665 675 5 — —
6 671 672 669 675 3 680 -9
7 693 679 674 684 5 678 15
8 694 686 679 693 7 689 5
9 701 696 687 705 9 700 1
10 703 699 694 704 5 714 -11
11 702 702 699 705 3 709 -7
12 710 705 702 708 3 708 2
13 712 708 705 711 3 711 1
14 711 711 708 714 3 714 -3
15 728 717 712 722 5 717 11
16 — — — — — 727 —

MSE  63.7
 CHAPTER 4 Moving Averages and Smoothing Methods 119
Then Equation 4.9 is used to compute the double moving average (column 4).

¿
M15 + M15-1 + M15-3+1
M 15 =
3
¿ 717 + 711 + 708
M 15 = = 712
3

Equation 4.10 is used to compute the difference between the two moving averages
(column 5).
¿
a15 = 2M15 - M 15 = 217172 - 712 = 722

Equation 4.11 is used to adjust the slope (column 6).

1M15 - M 152 = 1717 - 7122 = 5

2 ¿ 2
b15 =
3 - 1 2

Equation 4.12 is used to make the forecast one period into the future (column 7).

YN15+1 = a15 + b15 p = 722 + 5112 = 727

The forecast four weeks into the future is

YN15+4 = a15 + b15 p = 722 + 5142 = 742

Note that the MSE has been reduced from 133 to 63.7.

It seems reasonable that more-recent observations are likely to contain more-

important information. A procedure is introduced in the next section that gives more
emphasis to the most recent observations.

EXPONENTIAL SMOOTHING METHODS

Whereas the method of moving averages takes into account only the most recent
observations, simple exponential smoothing provides an exponentially weighted mov-
ing average of all previously observed values. The model is often appropriate for data
with no predictable upward or downward trend. The aim is to estimate the current
level. This level estimate is then used as the forecast of future values.
Exponential smoothing continually revises an estimate in the light of more-recent
experiences. This method is based on averaging (smoothing) past values of a series in
an exponentially decreasing manner. The most recent observation receives the largest
weight, a (where 0 6 a 6 1); the next most recent observation receives less weight,
a11 - a2 ; the observation two time periods in the past receives even less weight,
a11 - a22; and so forth.
In one representation of exponential smoothing, the new forecast (for time t + 1)
may be thought of as a weighted sum of the new observation (at time t) and the old
forecast (for time t). The weight a (0 6 a 6 1) is given to the newly observed value,
and the weight 11 - a2 is given to the old forecast. Thus,

New forecast = 3a * 1New observation24 + 311 - a2 * 1Old forecast24

More formally, the exponential smoothing equation is

YNt + 1 = aYt + 11 - a2YNt (4.13)

120 CHAPTER 4 Moving Averages and Smoothing Methods

where
YNt+1 = the new smoothed value or the forecast value for the next period
a = the smoothing constant 10 6 a 6 12
Yt = the new observation or the actual value of the series in period t
YNt = the old smoothed value or the forecast for period t
Equation 4.13 can be written as

YNt+1 = aYt + (1- a)YNt = aYt + YNt - aYNt

YNt+1 = YNt + a(Yt - YNt)

In this form, the new forecast (YN t + 1) is the old forecast (YN t) adjusted by a times the
error Yt - YN t in the old forecast.
In Equation 4.13, the smoothing constant, a, serves as the weighting factor. The
value of a determines the extent to which the current observation influences the fore-
cast of the next observation. When a is close to 1, the new forecast will be essentially
the current observation. (Equivalently, the new forecast will be the old forecast plus a
substantial adjustment for any error that occurred in the preceding forecast.)
Conversely, when a is close to zero, the new forecast will be very similar to the old fore-
cast, and the current observation will have very little impact.

Exponential smoothing is a procedure for continually revising a forecast in the

light of more-recent experience.

Finally, Equation 4.13 implies, for time t, that YNt = aYNt - 1 + (1 - a)YNt - 1, and substi-
tution for YNt in Equation 4.13 gives

YNt+1 = aYt + 11 - a2YNt = aYt + 11 - a23aYt-1 + 11 - a2YNt-14

YNt+1 = aYt + a11 - a2Yt-1 + 11 - a22YNt-1

Continued substitution (for YN t - 1 and so forth) shows YN t + 1 can be written as a sum of

current and previous Y’s with exponentially declining weights or

YNt+1 = aYt + a11 - a2Yt-1 + a11 - a22Yt-2 + a11 - a23Yt-3 + Á (4.14)

That is, YN t + 1 is an exponentially smoothed value. The speed at which past observations
lose their impact depends on the value of a, as demonstrated in Table 4-6.
Equations 4.13 and 4.14 are equivalent, but Equation 4.13 is typically used to cal-
culate the forecast YN t + 1 because it requires less data storage and is easily implemented.
The value assigned to a is the key to the analysis. If it is desired that predictions be
stable and random variations be smoothed, a small value of a is required. If a rapid
response to a real change in the pattern of observations is desired, a larger value of a is
appropriate. One method of estimating α is an iterative procedure that minimizes the
mean squared error (MSE ) given by Equation 3.8. Forecasts are computed for, say, α
equal to .1, .2, . . . , .9, and the sum of the squared forecast errors is computed for each.
 CHAPTER 4 Moving Averages and Smoothing Methods 121

TABLE 4-6 Comparison of Smoothing Constants

a = .1 a = .6
Period Calculations Weight Calculations Weight

t .100 .600
t -1 .9 * .1 .090 .4 * .6 .240
t -2 .9 * .9 * .1 .081 .4 * .4 * .6 .096
t -3 .9 * .9 * .9 * .1 .073 .4 * .4 * .4 * .6 .038
t -4 .9 * .9 * .9 * .9 * .1 .066 .4 * .4 * .4 * .4 * .6 .015
All others .590 .011

Totals 1.000 1.000

The value of α producing the smallest error is chosen for use in generating future
forecasts.
To start the algorithm for Equation 4.13, an initial value for the old smoothed
series must be set. One approach is to set the first smoothed value equal to the first
observation. Example 4.5 will illustrate this approach. Another method is to use the
average of the first five or six observations for the initial smoothed value.

Example 4.5
The exponential smoothing technique is demonstrated in Table 4-7 and Figure 4-7 for Acme
Tool Company for the years 2000 to 2006, using smoothing constants of .1 and .6. The data
for the first quarter of 2006 will be used as test data to help determine the best value of α
(among the two considered). The exponentially smoothed series is computed by initially set-
ting YN1 equal to 500. If earlier data are available, it might be possible to use them to develop
a smoothed series up to 2000 and then use this experience as the initial value for the
smoothed series. The computations leading to the forecast for periods 3 and 4 are demon-
strated next.
1. Using Equation 4.13, at time period 2, the forecast for period 3 with  = .1 is

YN2+1 = aY2 + 11 - a2YN2

YN3 = .113502 + .915002 = 485

2. The error in this forecast is

e3 = Y3 - YN3 = 250 - 485 = - 235

3. The forecast for period 4 is

YN3+1 = aY3 + 11 - a2YN3

YN4 = .112502 + .914852 = 461.5

From Table 4-7, when the smoothing constant is .1, the forecast for the first quarter of 2006
is 469, with a squared error of 145,161. When the smoothing constant is .6, the forecast for
the first quarter of 2006 is 576, with a squared error of 75,076. On the basis of this limited
evidence, exponential smoothing with a = .6 performs better than exponential smoothing
with a = .1.

In Figure 4-7, note how stable the smoothed values are for the .1 smoothing con-
stant. On the basis of minimizing the mean squared error, MSE (MSE is called MSD
on the Minitab output), over the first 24 quarters, the .6 smoothing constant is better.
122 CHAPTER 4 Moving Averages and Smoothing Methods

TABLE 4-7 Exponentially Smoothed Values for Acme Tool

Company Sales for Example 4.5
Actual Smoothed Forecast Smoothed Forecast
Time Value Value Error Value Error
Year Quarters Yt YN t(a = .1) et YN t(a = .6) et

2000 1 500 500.0 0.0 500.0 0.0

2 350 500.0 -150.0 500.0 -150.0
3 250 485.0 (1) -235.0 122 410.0 -160.0
4 400 461.5 (3) -61.5 314.0 86.0
2001 5 450 455.4 -5.4 365.6 84.4
6 350 454.8 -104.8 416.2 -66.2
7 200 444.3 -244.3 376.5 -176.5
8 300 419.9 -119.9 270.6 29.4
2002 9 350 407.9 -57.9 288.2 61.8
10 200 402.1 -202.1 325.3 -125.3
11 150 381.9 -231.9 250.1 -100.1
12 400 358.7 41.3 190.0 210.0
2003 13 550 362.8 187.2 316.0 234.0
14 350 381.6 -31.5 456.4 -106.4
15 250 378.4 -128.4 392.6 -142.6
16 550 365.6 184.4 307.0 243.0
2004 17 550 384.0 166.0 452.8 97.2
18 400 400.6 -0.6 511.1 -111.1
19 350 400.5 -50.5 444.5 -94.5
20 600 395.5 204.5 387.8 212.2
2005 21 750 415.9 334.1 515.1 234.9
22 500 449.3 -50.7 656.0 -156.0
23 400 454.4 -54.4 562.4 -162.4
24 650 449.0 201.0 465.0 185.0
2006 25 850 469.0 381.0 576.0 274.0

Note: The numbers in parentheses refer to the explanations given in the text in Example 4.5.

If the mean absolute percentage errors (MAPEs) are compared, the .6 smoothing
constant is also better. To summarize:

a = .1 MSE = 24,262 MAPE = 38.9%

a = .6 MSE = 22,248 MAPE = 36.5%

However, the MSE and MAPE are both large, and on the basis of these summary
statistics, it is apparent that exponential smoothing does not represent these data well.
As we shall see, a smoothing method that allows for seasonality does a better job of
predicting the Acme Tool Company saw sales.
A factor, other than the choice of α, that affects the values of subsequent forecasts
is the choice of the initial value, YN1 for the smoothed series. In Table 4-7 (see Example
4.5), YN1 = Y1 was used as the initial smoothed value. This choice tends to give Y1 too
much weight in later forecasts. Fortunately, the influence of the initial forecast dimin-
ishes greatly as t increases.
 CHAPTER 4 Moving Averages and Smoothing Methods 123

FIGURE 4-7 Exponential Smoothing for Acme Tool Company Data from
Example 4.5: (Top) a = .1 and (Bottom) a = .6

Another approach to initializing the smoothing procedure is to average the first k

observations. The smoothing then begins with

1 k
YN1 = a Yt
k t=1
Often k is chosen to be a relatively small number. For example, the default approach in
Minitab is to set k = 6.
124 CHAPTER 4 Moving Averages and Smoothing Methods

Example 4.6
The computation of the initial value as an average for the Acme Tool Company data
presented in Example 4.5 is shown next. If k is chosen to equal 6, then the initial value is

1 6
YN1 = a Yt = 1500 + 350 + 250 + 400 + 450 + 3502 = 383.3
1
6 t=1 6

The MSE and MAPE for each α when an initial smoothed value of 383.3 is used are
shown next.
a = .1 MSE = 21,091 MAPE = 32.1%
a = .6 MSE = 22,152 MAPE = 36.7%

The initial value, YN1 = 383.3, led to a decrease in the MSE and MAPE for a = .1 but did
not have much effect when a = .6. Now the best model, based on the MSE and MAPE sum-
mary measures, appears to be one that uses a = .1 instead of .6.
Figure 4-8 shows results for Example 4.5 when the data are run on Minitab (see the
Minitab Applications section at the end of the chapter for instructions). The smoothing con-
stant of a = .266 was automatically selected by minimizing the MSE. The MSE is reduced
to 19,447, the MAPE equals 32.2%, and although not shown, the MPE equals –6.4%. The
forecast for the first quarter of 2006 is 534.
Figure 4-9 shows the autocorrelation function for the residuals of the exponential
smoothing method using an alpha of .266. When the Ljung-Box test is conducted for six
time lags, the large value of LBQ (33.86) shows that the first six residual autocorrelations as
a group are larger than would be expected if the residuals were random. In particular, the
significantly large residual autocorrelations at lags 2 and 4 indicate that the seasonal varia-
tion in the data is not accounted for by simple exponential smoothing.

Exponential smoothing is often a good forecasting procedure when a nonrandom

time series exhibits trending behavior. It is useful to develop a measure that can be
used to determine when the basic pattern of a time series has changed. A tracking sig-
nal is one way to monitor change. A tracking signal involves computing a measure of
the forecast errors over time and setting limits so that, when the errors go outside those
limits, the forecaster is alerted.

FIGURE 4-8 Exponential Smoothing with a = .266 for Acme

Tool Company Data for Example 4.6
 CHAPTER 4 Moving Averages and Smoothing Methods 125

FIGURE 4-9 Autocorrelation Function for the Residuals When Exponential

Smoothing with a = .266 Is Used with Acme Tool Company Data
for Example 4.6

A tracking signal involves computing a measure of forecast errors over time and
setting limits so that, when the cumulative error goes outside those limits, the
forecaster is alerted.

For example, a tracking signal might be used to determine when the size of the
smoothing constant α should be changed. Since a large number of items are usually
being forecast, common practice is to continue with the same value of α for many peri-
ods before attempting to determine if a revision is necessary. Unfortunately, the sim-
plicity of using an established exponential smoothing model is a strong motivator for
not making a change. But at some point, it may be necessary to update α or abandon
exponential smoothing altogether. When the model produces forecasts containing a
great deal of error, a change is appropriate.
A tracking system is a method for monitoring the need for change. Such a system
contains a range of permissible deviations of the forecast from actual values. As long as
forecasts generated by exponential smoothing fall within this range, no change in α is
necessary. However, if a forecast falls outside the range, the system signals a need to
update α.
For instance, if things are going well, the forecasting technique should over-
and underestimate equally often. A tracking signal based on this rationale can be
developed.
Let U equal the number of underestimates out of the last n forecasts. In other
words, U is the number of errors out of the last k that are positive. If the process is in
control, the expected value of U is k/2, but sampling variability is involved, so values
close to k/2 would not be unusual. On the other hand, values that are not close to k/2
would indicate that the technique is producing biased forecasts.
Example 4.7
Suppose that Acme Tool Company has decided to use the exponential smoothing technique
with α equal to .1, as shown in Example 4.5 (see p. 121). If the process is in control and the
analyst decides to monitor the last 10 error values, U has an expected value of 5. Actually,
126 CHAPTER 4 Moving Averages and Smoothing Methods

a U value of 2, 3, 4, 6, 7, or 8 would not be unduly alarming. However, a value of 0, 1, 9, or 10

would be of concern, since the probability of obtaining such a value by chance alone would
be .022 (based on the binomial distribution). With this information, a tracking system can be
developed based on the following rules:
If 2 … U … 8, then the process is in control.
If U 6 2 or U 7 8 then the process is out of control.
Assume that, out of the next 10 forecasts using this technique, only one has a positive error.
Since the probability of obtaining only one positive error out of 10 is quite low (.011), the
process is considered to be out of control (overestimating), and the value of α should be
changed.

Another way of tracking a forecasting technique is to determine a range that

should contain the forecasting errors. This can be accomplished by using the MSE that
was established when the optimally sized α was determined. If the exponential
smoothing technique is reasonably accurate, the forecast error should be approxi-
mately normally distributed about a mean of zero. Under this condition, there is about
a 95% chance that the actual observation will fall within approximately two standard
deviations of the forecast. Using the RMSE as an estimate of the standard deviation of
the forecast error, approximate 95% error limits can be determined. Forecast errors
falling within these limits indicate no cause for alarm. Errors (particularly a sequence
of errors) outside the limits suggest a change. Example 4.8 illustrates this approach.
Example 4.8
In Example 4.6, on the Acme Tool Company, the optimal α was determined to be a = .266,
with MSE = 19,447 . An estimate of the standard deviation of the forecast errors is
RMSE = 219,447 = 139.5. If the forecast errors are approximately normally distributed
about a mean of zero, there is about a 95% chance that the actual observation will fall
within two standard deviations of the forecast or within

;2RMSE = ; 2219,477 = ; 21139.52 = ; 279

For this example, the permissible absolute error is 279. If for any future forecast the magni-
tude of the error is greater than 279, there is reason to believe that the optimal smoothing
constant α should be updated or a different forecasting method considered.

The preceding discussion on tracking signals also applies to the smoothing meth-
ods yet to be discussed in the rest of the chapter.
Simple exponential smoothing works well when the data vary about an infre-
quently changing level. Whenever a sustained trend exists, exponential smoothing will
lag behind the actual values over time. Holt’s linear exponential smoothing technique,
which is designed to handle data with a well-defined trend, addresses this problem and
is introduced next.

Exponential Smoothing Adjusted for Trend: Holt’s Method

In simple exponential smoothing, the level of the time series is assumed to be changing
occasionally, and an estimate of the current level is required. In some situations, the
observed data will be clearly trending and contain information that allows the anticipa-
tion of future upward movements.When this is the case, a linear trend forecast function is
needed. Since business and economic series rarely exhibit a fixed linear trend, we consider
the possibility of modeling evolving local linear trends over time. Holt (2004) developed
an exponential smoothing method, Holt’s linear exponential smoothing,1 that allows for
evolving local linear trends in a time series and can be used to generate forecasts.
1Holt’s linear exponential smoothing is sometimes called double exponential smoothing.
 CHAPTER 4 Moving Averages and Smoothing Methods 127
When a trend in the time series is anticipated, an estimate of the current slope, as
well as the current level, is required. Holt’s technique smoothes the level and slope
directly by using different smoothing constants for each. These smoothing constants
provide estimates of level and slope that adapt over time as new observations become
available. One of the advantages of Holt’s technique is that it provides a great deal of
flexibility in selecting the rates at which the level and trend are tracked.
The three equations used in Holt’s method are
1. The exponentially smoothed series, or current level estimate
Lt = Yt + 11 - 21Lt-1 + Tt-12 (4.15)
2. The trend estimate
Tt =
1Lt - Lt-12 + 11 -
)Tt-1 (4.16)
3. The forecast for p periods into the future
YNt+p = Lt + pTt (4.17)
where
Lt = the new smoothed value (estimate of current level)
a = 1 the smoothing constant for the level 10 6 a 6 12
Yt = the new observation or actual value of the series in period t
b = the smoothing constant for the trend estimate 10 6 b 6 12
Tt = the trend estimate
p = the periods to be forecast into the future
YNt+p  the forecast for p periods into the future

Equation 4.15 is very similar to the equation for simple exponential smoothing,
Equation 4.13, except that a term 1Tt-12 has been incorporated to properly update the level
when a trend exists. That is, the current level 1Lt2 is calculated by taking a weighted aver-
age of two estimates of level—one estimate is given by the current observation 1Yt2, and
the other estimate is given by adding the previous trend 1Tt-12 to the previously smoothed
level 1Lt-12. If there is no trend in the data, there is no need for the term Tt-1 in Equation
4.15, effectively reducing it to Equation 4.13. There is also no need for Equation 4.16.
A second smoothing constant, b, is used to create the trend estimate. Equation 4.16
shows that the current trend 1Tt2 is a weighted average (with weights b and 1 - b) of
two trend estimates—one estimate is given by the change in level from time t - 1 to t
1Lt - Lt-12, and the other estimate is the previously smoothed trend 1Tt-12. Equation
4.16 is similar to Equation 4.15, except that the smoothing is done for the trend rather
than the actual data.
Equation 4.17 shows the forecast for p periods into the future. For a forecast made
at time t, the current trend estimate 1Tt2 is multiplied by the number of periods to be
forecast 1p2, and then the product is added to the current level 1Lt2. Note that the fore-
casts for future periods lie along a straight line with slope Tt and intercept Lt.
As with simple exponential smoothing, the smoothing constants α and b can be
selected subjectively or generated by minimizing a measure of forecast error such as
the MSE. Large weights result in more rapid changes in the component; small weights
result in less rapid changes. Therefore, the larger the weights are, the more the
smoothed values follow the data; the smaller the weights are, the smoother the pattern
in the smoothed values is.
128 CHAPTER 4 Moving Averages and Smoothing Methods

We could develop a grid of values of α and b (e.g., each combination of

a = 0.1, 0.2, Á , 0.9 and b = 0.1, 0.2, Á , 0.9) and then select the combination that
provides the lowest MSE. Most forecasting software packages use an optimization
algorithm to minimize the MSE. We might insist that a = b , thus providing equal
amounts of smoothing for the level and the trend. In the special case where a = b ,
Holt’s approach is the same as Brown’s double exponential smoothing.
To get started, initial values for L and T in Equations 4.15 and 4.16 must be deter-
mined. One approach is to set the first estimate of the smoothed level equal to the first
observation.The trend is then estimated to be zero.A second approach is to use the aver-
age of the first five or six observations as the initial smoothed value L. The trend is then
estimated using the slope of a line fit to these five or six observations. Minitab develops a
regression equation using the variable of interest as Y and time as the independent vari-
able X. The constant from this equation is the initial estimate of the level component, and
the slope or regression coefficient is the initial estimate of the trend component.

Example 4.9
In Example 4.6, simple exponential smoothing did not produce successful forecasts of Acme
Tool Company saw sales. Because Figure 4-8 suggests that there might be a trend in these
data, Holt’s linear exponential smoothing is used to develop forecasts. To begin the computa-
tions shown in Table 4-8, two estimated initial values are needed, namely, the initial level and
the initial trend value. The estimate of the level is set equal to the first observation. The trend
is estimated to equal zero. The technique is demonstrated in Table 4-8 for a = .3 and b = .1.
The value for α used here is close to the optimal α (a = .266) for simple exponential
smoothing in Example 4.6. α is used to smooth the data to eliminate randomness and esti-
mate level. The smoothing constant b is like α, except that it is used to smooth the trend in
the data. Both smoothing constants are used to average past values and thus to remove ran-
domness. The computations leading to the forecast for period 3 are shown next.
1. Update the exponentially smoothed series or level:

Lt = aYt + 11 - a21Lt-1 + Tt-12

L2 = .3Y2 + 11 - .321L2-1 + T2-12
L2 = .313502 + .71500 + 02 = 455

2. Update the trend estimate:

Tt =
1Lt - Lt-12 + 11 -
2Tt-1
T2 = .11L2 - L2-12 + 11 - .12T2-1
T2 = .11455 - 5002 + .9102 = - 4.5

3. Forecast one period into the future:

YN t+p = Lt + pTt

YN2+1 = L2 + 1T2 = 455 + 11 -4.52 = 450.5

4. Determine the forecast error:

e3 = Y3 - YN3 = 250 -450.5 = - 200.5

The forecast for period 25 is computed as follows:

1. Update the exponentially smoothed series or level:

L24 = .3Y24 + 11 - .321L24-1 + T24-12

L24 = .316502 + .71517.6 + 9.82 = 564.2
 CHAPTER 4 Moving Averages and Smoothing Methods 129

TABLE 4-8 Exponentially Smoothed Values for Acme Tool

Company Sales, Holt’s Method, for Example 4.9

Year t Yt Lt Tt YN t+p et

2000 1 500 500.0 0.0 500.0 0.0

2 350 455.0 -4.5 500.0 -150.0
3 250 390.4 -10.5 450.5 -200.5
4 400 385.9 -9.9 379.8 20.2
2001 5 450 398.2 -7.7 376.0 74.0
6 350 378.3 -8.9 390.5 -40.5
7 200 318.6 -14.0 369.4 -169.4
8 300 303.2 -14.1 304.6 -4.6
2002 9 350 307.4 -12.3 289.1 60.9
10 200 266.6 -15.2 295.1 -95.1
11 150 221.0 -18.2 251.4 -101.4
12 400 262.0 -12.3 202.8 197.2
2003 13 550 339.8 -3.3 249.7 300.3
14 350 340.6 -2.9 336.5 13.5
15 250 311.4 -5.5 337.7 -87.7
16 550 379.1 1.8 305.9 244.1
2004 17 550 431.7 6.9 381.0 169.0
18 400 427.0 5.7 438.6 -38.6
19 350 407.9 3.3 432.7 -82.7
20 600 467.8 8.9 411.2 188.8
2005 21 750 558.7 17.1 476.8 273.2
22 500 553.1 14.8 575.9 -75.9
23 400 517.6 9.8 567.9 -167.9
24 650 564.2 13.5 527.4 122.6
2006 25 850 — — 577.7 272.3

MSE = 20,515.5

2. Update the trend estimate:

T24 = .11L24 - L24-12 + 11 - .12T24-1

T24 = .11564.2 - 517.62 + .919.82 = 13.5

3. Forecast one period into the future:

YN24+1 = L24 + 1T24

YN25 = 564.2 + 1113.52 = 577.7

On the basis of minimizing the MSE over the period 2000 to 2006, Holt’s linear
smoothing (with a = .3 and b = .1) does not reproduce the data any better than sim-
ple exponential smoothing that used a smoothing constant of .266. A comparison of the
MAPEs shows them to be about the same. When the forecasts for the actual sales for
130 CHAPTER 4 Moving Averages and Smoothing Methods

FIGURE 4-10 Holt’s Linear Exponential Smoothing for Acme Tool

Company Data for Example 4.9

the first quarter of 2006 are compared, again Holt smoothing and simple exponential
smoothing are comparable. To summarize:
a = .266 MSE = 19,447 MAPE = 32.2%
a = .3, b = .1 MSE = 20,516 MAPE = 35.4%
Figure 4-10 shows the results when Holt’s method using a = .3 and b = .1 is run
on Minitab.2 The autocorrelation function for the residuals from Holt’s linear expo-
nential smoothing is given in Figure 4-11. The autocorrelation coefficients at time lags
2 and 4 appear to be significant. Also, when the Ljung-Box Q statistic is computed for
six time lags, the large value of LBQ (36.33) shows that the residuals contain extensive
autocorrelation; they are not random. The large residual autocorrelations at lags 2 and
4 suggest a seasonal component may be present in Acme Tool Company data.
The results in Examples 4.6 and 4.9 (see Figures 4-8 and 4-10) are not much differ-
ent because the smoothing constant α is about the same in both cases and the smooth-
ing constant b in Example 4.9 is small. (For b = 0, Holt’s linear smoothing reduces to
simple exponential smoothing.)

Exponential Smoothing Adjusted for Trend

and Seasonal Variation: Winters’ Method
Examination of the data for Acme Tool Company in Table 4-8 indicates that sales are
consistently higher during the first and fourth quarters and lower during the third
quarter. A seasonal pattern appears to exist. Winters’ three-parameter linear and sea-
sonal exponential smoothing method, an extension of Holt’s method, might represent
the data better and reduce forecast error. In Winters’ method, one additional equation
is used to estimate seasonality. In the multiplicative version of Winters’ method, the
seasonality estimate is given as a seasonal index, and it is calculated with Equation
4.20. Equation 4.20 shows that to compute the current seasonal component, St, the

2In the Minitab program, the trend parameter gamma (g) is identical to our beta (b).
 CHAPTER 4 Moving Averages and Smoothing Methods 131

FIGURE 4-11 Autocorrelation Function for the Residuals from Holt’s Linear
Exponential Smoothing for Acme Tool Company Data for
Example 4.9

product of g and an estimate of the seasonal index given by Yt >Lt is added to 11 - g2

times the previous seasonal component, St-s. This procedure is equivalent to smoothing
current and previous values of Yt >Lt. Yt is divided by the current level estimate, Lt, to
create an index (ratio) that can be used in a multiplicative fashion to adjust a forecast
to account for seasonal peaks and valleys.
The four equations used in Winters’ (multiplicative) smoothing are
1. The exponentially smoothed series, or level estimate:

+ 11 - a21Lt-1 + Tt-12
Yt
Lt = a (4.18)
St-s
2. The trend estimate:
Tt = b1Lt - Lt-12 + 11 - b2Tt-1 (4.19)
3. The seasonality estimate:

+ 11 - 2St-s
Yt
St =  (4.20)
Lt
4. The forecast for p periods into the future:
YNt+p = 1Lt + pTt2St-s+p (4.21)

where
Lt = the new smoothed value or current level estimate
a = the smoothing constant for the level
Yt = the new observation or actual value in period t
b = the smoothing constant for the trend estimate
Tt = the trend estimate
g = the smoothing constant for the seasonality estimate
St = the seasonal estimate
132 CHAPTER 4 Moving Averages and Smoothing Methods

p = the periods to be forecast into the future

s = the length of seasonality
YNt+p = the forecast for p periods into the future
Equation 4.18 updates the smoothed series. A slight difference in this equation distin-
guishes it from the corresponding one in Holt’s procedure, Equation 4.15. In Equation
4.18, Yt is divided by St-s, which adjusts Yt for seasonality, thus removing the seasonal
effects that might exist in the original data, Yt.
After the trend estimate and seasonal estimate have been smoothed in Equations
4.19 and 4.20, a forecast is obtained with Equation 4.21. It is almost the same as the cor-
responding formula, Equation 4.17, used to obtain a forecast with Holt’s smoothing.
The difference is that this estimate for future periods, t + p, is multiplied by St-s+p. This
seasonal index is the last one available and is therefore used to adjust the forecast for
seasonality.
As with Holt’s linear exponential smoothing, the weights α, b, and g can be
selected subjectively or generated by minimizing a measure of forecast error, such as
the MSE. The most common approach for determining these values is to use an opti-
mization algorithm to find the optimal smoothing constants.
To begin the algorithm for Equation 4.18, the initial values for the smoothed series,
Lt; the trend, Tt; and the seasonal indices, St, must be set. One approach is to set the first
estimate of the smoothed series (level) equal to the first observation. The trend is then
estimated to equal zero, and each seasonal index is set to 1.0. Other approaches for ini-
tializing the level, trend, and seasonal estimates are available. Minitab, for example,
develops a regression equation using the variable of interest as Y and time as the inde-
pendent variable X. The constant from this equation is the initial estimate of the
smoothed series or level component, and the slope or regression coefficient is the ini-
tial estimate of the trend component. Initial values for the seasonal components are
obtained from a dummy variable regression using detrended data (see Chapter 8).
Example 4.10
Winters’ technique is demonstrated in Table 4-9 for a = .4, b = .1, and g = .3 for the Acme
Tool Company data. The value for α is similar to the one used for simple exponential
smoothing in Example 4.6, and it is used to smooth the data to create a level estimate. The
smoothing constant b is used to create a smoothed estimate of trend. The smoothing con-
stant g is used to create a smoothed estimate of the seasonal component in the data.
Minitab can be used to solve this example (see the Minitab Applications section at the
end of the chapter for the instructions).3 The results are shown in Table 4-9 and Figure 4-12.
The forecast for the first quarter of 2006 is 778.2. The computations leading to the forecast
value for the first quarter of 2006, or period 25, are shown next.
1. The exponentially smoothed series, or level estimate:

+ 11 - a21Lt - 1 + Tt - 12
Yt
Lt = a
St - s

+ 11 - .421L24 - 1 + T24 - 12
Y24
L24 = .4
S24 - 4

+ 11 - .421501.286 + 9.1482
650
L24 = .4
1.39628
L24 = .41465.522 + .61510.4342 = 492.469

3In the Minitab program, the trend parameter gamma (g) is identical to our beta (b) and the seasonal param-
eter delta (δ) is identical to our gamma (g) in Equations 4.19 and 4.20, respectively.
 CHAPTER 4 Moving Averages and Smoothing Methods 133

TABLE 4-9 Exponentially Smoothed Values for Acme Tool

Company Sales, Winters’ Method, for Example 4.10

Year t Yt Lt Tt St YN t+p et

2000 1 500 415.459 -41.9541 1.26744 563.257 -63.257

2 350 383.109 -40.9937 0.89040 328.859 21.141
3 250 358.984 -39.3068 0.66431 222.565 27.435
4 400 328.077 -38.4668 1.18766 375.344 24.656
2001 5 450 315.785 -35.8494 1.31471 367.063 82.937
6 350 325.194 -31.3235 0.94617 249.255 100.745
7 200 296.748 -31.0358 0.66721 195.221 4.779
8 300 260.466 -31.5604 1.17690 315.576 -15.576
2002 9 350 243.831 -30.0679 1.35093 300.945 49.055
10 200 212.809 -30.1632 0.94426 202.255 -2.255
11 150 199.515 -28.4764 0.69259 121.863 28.137
12 400 238.574 -21.7228 1.32682 201.294 198.706
2003 13 550 292.962 -14.1117 1.50886 292.949 257.051
14 350 315.575 -10.4393 0.99371 263.306 86.694
15 250 327.466 -8.2062 0.71385 211.335 38.665
16 550 357.366 -4.3956 1.39048 423.599 126.401
2004 17 550 357.588 -3.9339 1.51763 532.584 17.416
18 400 373.206 -1.9787 1.01713 351.428 48.572
19 350 418.856 2.7843 0.75038 264.999 85.001
20 600 425.586 3.1788 1.39628 586.284 13.716
2005 21 750 454.936 5.7959 1.55691 650.706 99.294
22 500 473.070 7.0297 1.02907 468.626 31.374
23 400 501.286 9.1484 0.76465 360.255 39.745
24 650 492.469 7.3518 1.37336 712.712 -62.712
2006 25 850 — — — 778.179 —
26 600 — — — 521.917 —
27 450 — — — 393.430 —
28 700 — — — 716.726 —

MSE  7,636.86

2. The trend estimate:

Tt = b1Lt - Lt-12 + 11 - b2Tt-1
T24 = .11L24 - L24 - 12 + 11 - .12T24 - 1
T24 = .11492.469 - 501.2862 + .919.1482
T24 = .11-8.8172 + .919.1482 = 7.352
3. The seasonality estimate:

+ 11 - 2St-s
Yt
St = 
Lt

+ 11 - .32S24 - 4
Y24
S24 = .3
L24
650
S24 = .3 + .711.396282
492.469
S24 = .311.31992 + .9774 = 1.3734
134 CHAPTER 4 Moving Averages and Smoothing Methods

FIGURE 4-12 Winters’ Exponential Smoothing for Acme Tool Company

Data for Example 4.10

4. The forecast for p = 1 period into the future:

YN24 + 1 = (L24 + 1T24)S24 - 4 + 1

YN 25 = 1492.469 + 117.352221.5569 = 778.17

For the parameter values considered, Winters’ technique is better than both of the pre-
vious smoothing procedures in terms of minimizing the MSE. When the forecasts for the
actual sales for the first quarter of 2006 are compared, Winters’ technique also appears to do
a better job. Figure 4-13 shows the autocorrelation function for the Winters’ exponential
smoothing residuals. None of the residual autocorrelation coefficients appears to be signifi-
cantly larger than zero. When the Ljung-Box Q statistic is calculated for six time lags, the
small value of LBQ (5.01) shows that the residual series is random. Winters’ exponential
smoothing method seems to provide adequate forecasts for the Acme Tool Company data.

Winters’ method provides an easy way to account for seasonality when data have a
seasonal pattern. An alternative method consists of first deseasonalizing or seasonally
adjusting the data. Deseasonalizing is a process that removes the effects of seasonality
from the raw data and will be demonstrated in Chapter 5. The forecasting model is
applied to the deseasonalized data, and if required, the seasonal component is rein-
serted to provide accurate forecasts.
Exponential smoothing is a popular technique for short-run forecasting. Its major
advantages are its low cost and simplicity. When forecasts are needed for inventory sys-
tems containing thousands of items, smoothing methods are often the only acceptable
approach.
Simple moving averages and exponential smoothing base forecasts on weighted
averages of past measurements. The rationale is that past values contain information
about what will occur in the future. Since past values include random fluctuations as
well as information concerning the underlying pattern of a variable, an attempt is made
 CHAPTER 4 Moving Averages and Smoothing Methods 135

FIGURE 4-13 Autocorrelation Function for the Residuals from Winters’

Multiplicative Exponential Smoothing Method for Acme Tool
Company Data for Example 4.10

to smooth these values. Smoothing assumes that extreme fluctuations represent ran-
domness in a series of historical observations.
Moving averages are the means of a certain number, k, of values of a variable. The
most recent average is then the forecast for the next period. This approach assigns an
equal weight to each past value involved in the average. However, a convincing argu-
ment can be made for using all the data but emphasizing the most recent values.
Exponential smoothing methods are attractive because they generate forecasts by
using all the observations and assigning weights that decline exponentially as the
observations get older.

APPLICATION TO MANAGEMENT
Forecasts are one of the most important inputs managers have to aid them in the
decision-making process. Virtually every important operating decision depends to
some extent on a forecast. The production department has to schedule employment
needs and raw material orders for the next month or two; the finance department must
determine the best investment opportunities; marketing must forecast the demand for
a new product. The list of forecasting applications is quite lengthy.
Executives are keenly aware of the importance of forecasting. Indeed, a great deal
of time is spent studying trends in economic and political affairs and how events might
affect demand for products and/or services. One issue of interest is the importance
executives place on quantitative forecasting methods compared to their own opinions.
This issue is especially sensitive when events that have a significant impact on demand
are involved. One problem with quantitative forecasting methods is that they depend
on historical data. For this reason, they are probably least effective in determining the
dramatic change that often results in sharply higher or lower demand.
The averaging and smoothing forecasting methods discussed in this chapter are
useful because of their relative simplicity. These simple methods tend to be less costly,
easier to implement, and easier to understand than complex methods. Frequently, the
cost and potential difficulties associated with constructing more sophisticated models
outweigh any gains in their accuracy. For these reasons, small businesses find simple
136 CHAPTER 4 Moving Averages and Smoothing Methods

methods practical. Businesses without personnel capable of handling statistical models

also turn to simple methods. Business managers frequently face the need to prepare
short-term forecasts for a number of different items. A typical example is the manager
who must schedule production on the basis of some forecast of demand for several
hundred different products in a product line. Also, new businesses without lengthy his-
torical databases find these simple approaches helpful.
With a judicious choice of order k, the moving average method can do a good job
of adjusting to shifts in levels. It is economical to update, and it does not require much
data storage. The moving average method is most frequently used when repeated fore-
casts are necessary.
Exponential smoothing is a popular technique whose strength lies in good short-
term accuracy combined with quick, low-cost updating. The technique is widely used
when regular monthly or weekly forecasts are needed for a large number, perhaps
thousands, of items. Inventory control is one example where exponential smoothing
methods are routinely used.

Glossary
Exponential smoothing. Exponential smoothing is Simple average. A simple average uses the mean
a procedure for continually revising a forecast in of all relevant historical observations as the fore-
the light of more-recent experience. cast for the next period.
Moving average. A moving average of order k is Tracking signal. A tracking signal involves com-
the mean value of k consecutive observations. The puting a measure of forecast errors over time and
most recent moving average value provides a fore- setting limits so that, when the cumulative error
cast for the next period. goes outside those limits, the forecaster is alerted.

Key Formulas

Naive model

YNt + 1 = Yt (4.1)
Naive trend model

YNt + 1 = Yt + 1Yt - Yt - 12 (4.2)

Naive rate of change model

Yt
YNt + 1 = Yt (4.3)
Yt - 1
Naive seasonal model for quarterly data

YNt + 1 = Yt - 3 (4.4)
Naive trend and seasonal model for quarterly data

Yt - Yt - 4
YNt + 1 = Yt - 3 + (4.5)
4
Simple average model
1 t
YNt + 1 = a Yi (4.6)
t i=1
 CHAPTER 4 Moving Averages and Smoothing Methods 137
Updated simple average, new period

tYNt + 1 + Yt + 1
YNt + 2 = (4.7)
t + 1
Moving average for k time periods
Yt + Yt - 1 + Á + Yt - k + 1
YNt + 1 = (4.8)
k
Double moving average
Mt + Mt-1 + Mt-2 + Á + Mt-k + 1
M ¿t = (4.9)
k
at = 2Mt - M t¿ (4.10)

1M - M ¿t2
2
bt = (4.11)
k - 1 t
YNt + p = at + bt p (4.12)

Simple exponential smoothing

YNt + 1 = aYt + 11 - a)YNt (4.13)

Equivalent alternative expression:

YNt + 1 = aYt + a11 - a2Yt - 1 + a11 - a22Yt - 2 + a11 - a23Yt - 3 + Á (4.14)

Holt’s linear smoothing

The exponentially smoothed series, or current level estimate:
Lt = aYt + 11 - a21Lt-1 + Tt-12 (4.15)
The trend estimate:
Tt = b1Lt - Lt-12 + 11 - b2Tt-1 (4.16)

The forecast for p periods into the future:

YNt + p = Lt + pTt (4.17)

Winters’ multiplicative smoothing

The exponentially smoothed series, or level estimate:

+ 11 - a21Lt-1 + Tt-12
Yt
Lt = a (4.18)
St-s
The trend estimate:
Tt = b1Lt - Lt-12 + 11 - b2Tt-1 (4.19)

The seasonality estimate:

+ 11 - 2St-s
Yt
St =  (4.20)
Lt

The forecast for p periods into the future:

YNt + p = 1Lt + pTt2St - s + p (4.21)

138 CHAPTER 4 Moving Averages and Smoothing Methods

Problems
1. Which forecasting technique continually revises an estimate in the light of more-
recent experiences?
2. Which forecasting technique uses the value for the current period as the forecast
for the next period?
3. Which forecasting technique assigns equal weight to each observation?
4. Which forecasting technique(s) should be tried if the data are trending?
5. Which forecasting technique(s) should be tried if the data are seasonal?
6. Apex Mutual Fund invests primarily in technology stocks. The price of the fund at
the end of each month for the 12 months of 2006 is shown in Table P-6.
a. Find the forecast value of the mutual fund for each month by using a naive
model (see Equation 4.1). The value for December 2005 was 19.00.
b. Evaluate this forecasting method using the MAD.
c. Evaluate this forecasting method using the MSE.
d. Evaluate this forecasting method using the MAPE.
e. Evaluate this forecasting method using the MPE.
f. Using a naive model, forecast the mutual fund price for January 2007.
g. Write a memo summarizing your findings.
7. Refer to Problem 6. Use a three-month moving average to forecast the mutual
fund price for January 2007. Is this forecast better than the forecast made using the
naive model? Explain.
8. Given the series Yt in Table P-8:
a. What is the forecast for period 9, using a five-period moving average?
b. If simple exponential smoothing with a smoothing constant of .4 is used, what is
the forecast for time period 4?
c. In part b, what is the forecast error for time period 3?
9. The yield on a general obligation bond for the city of Davenport fluctuates with
the market. The monthly quotations for 2006 are given in Table P-9.

TABLE P-6

Month Mutual Fund Price

January 19.39
February 18.96
March 18.20
April 17.89
May 18.43
June 19.98
July 19.51
August 20.63
September 19.78
October 21.25
November 21.18
December 22.14
 CHAPTER 4 Moving Averages and Smoothing Methods 139

TABLE P-8

Time Period Yt YNt et

1 200 200 —
2 210 — —
3 215 — —
4 216 — —
5 219 — —
6 220 — —
7 225 — —
8 226 — —

TABLE P-9

Month Yield

January 9.29
February 9.99
March 10.16
April 10.25
May 10.61
June 11.07
July 11.52
August 11.09
September 10.80
October 10.50
November 10.86
December 9.97

a. Find the forecast value of the yield for the obligation bonds for each month,
starting with April, using a three-month moving average.
b. Find the forecast value of the yield for the obligation bonds for each month,
starting with June, using a five-month moving average.
c. Evaluate these forecasting methods using the MAD.
d. Evaluate these forecasting methods using the MSE.
e. Evaluate these forecasting methods using the MAPE.
f. Evaluate these forecasting methods using the MPE.
g. Forecast the yield for January 2007 using the better technique.
h. Write a memo summarizing your findings.
10. This question refers to Problem 9. Use exponential smoothing with a smoothing
constant of .2 and an initial value of 9.29 to forecast the yield for January 2007. Is
this forecast better than the forecast made using the better moving average
model? Explain.
11. The Hughes Supply Company uses an inventory management method to deter-
mine the monthly demands for various products. The demand values for the last 12
months of each product have been recorded and are available for future forecast-
ing. The demand values for the 12 months of 2006 for one electrical fixture are pre-
sented in Table P-11.
140 CHAPTER 4 Moving Averages and Smoothing Methods

TABLE P-11

Month Demand

January 205
February 251
March 304
April 284
May 352
June 300
July 241
August 284
September 312
October 289
November 385
December 256

Source: Based on Hughes Supply Company records.

Use exponential smoothing with a smoothing constant of .5 and an initial value of

205 to forecast the demand for January 2007.
12. General American Investors Company, a closed-end regulated investment man-
agement company, invests primarily in medium- and high-quality stocks. Jim
Campbell is studying the asset value per share for this company and would like to
forecast this variable for the remaining quarters of 1996. The data are presented in
Table P-12.
Evaluate the ability to forecast the asset value per share variable of the follow-
ing forecasting methods: naive, moving average, and exponential smoothing. When
you compare techniques, take into consideration that the actual asset value per
share for the second quarter of 1996 was 26.47. Write a report for Jim indicating
which method he should use and why.

TABLE P-12 General American Investors Company

Assets per Share, 1985–1996

Quarter

Year 1 2 3 4

1985 16.98 18.47 17.63 20.65

1986 21.95 23.85 20.44 19.29
1987 22.75 23.94 24.84 16.70
1988 18.04 19.19 18.97 17.03
1989 18.23 19.80 22.89 21.41
1990 21.50 25.05 20.33 20.60
1991 25.33 26.06 28.89 30.60
1992 27.44 26.69 28.71 28.56
1993 25.87 24.96 27.61 24.75
1994 23.32 22.61 24.08 22.31
1995 22.67 23.52 25.41 23.94
1996 25.68 — — —

Source: The Value Line Investment Survey (New York: Value Line,
1990, 1993, 1996).
 CHAPTER 4 Moving Averages and Smoothing Methods 141

TABLE P-13 Southdown Revenues, 1986–1999

Quarter

Year 1 2 3 4

1986 77.4 88.8 92.1 79.8

1987 77.5 89.1 92.4 80.1
1988 74.7 185.2 162.4 178.1
1989 129.1 158.4 160.6 138.7
1990 127.2 149.8 151.7 132.9
1991 103.0 136.8 141.3 123.5
1992 107.3 136.1 138.6 123.7
1993 106.1 144.4 156.1 138.2
1994 111.8 149.8 158.5 141.8
1995 119.1 158.0 170.4 151.8
1996 127.4 178.2 189.3 169.5
1997 151.4 187.2 199.2 181.4
1998 224.9 317.7 341.4 300.7
1999 244.9 333.4 370.0 326.7

Source: The Value Line Investment Survey (New York: Value Line,
1990, 1993, 1996, 1999).

13. Southdown, Inc., the nation’s third largest cement producer, is pushing ahead
with a waste fuel burning program. The cost for Southdown will total about
$37 million. For this reason, it is extremely important for the company to have an
accurate forecast of revenues for the first quarter of 2000. The data are presented
in Table P-13.
a. Use exponential smoothing with a smoothing constant of .4 and an initial value
of 77.4 to forecast the quarterly revenues for the first quarter of 2000.
b. Now use a smoothing constant of .6 and an initial value of 77.4 to forecast the
quarterly revenues for the first quarter of 2000.
c. Which smoothing constant provides the better forecast?
d. Refer to part c. Examine the residual autocorrelations. Are you happy with sim-
ple exponential smoothing for this example? Explain.
14. The Triton Energy Corporation explores for and produces oil and gas. Company
president Gail Freeman wants to have her company’s analyst forecast the com-
pany’s sales per share for 2000. This will be an important forecast, since Triton’s
restructuring plans have hit a snag. The data are presented in Table P-14.
Determine the best forecasting method and forecast sales per share for 2000.
15. The Consolidated Edison Company sells electricity (82% of revenues), gas (13%),
and steam (5%) in New York City and Westchester County. Bart Thomas, company
forecaster, is assigned the task of forecasting the company’s quarterly revenues for
the rest of 2002 and all of 2003. He collects the data shown in Table P-15.
Determine the best forecasting technique and forecast quarterly revenue for
the rest of 2002 and all of 2003.
16. A job-shop manufacturer that specializes in replacement parts has no forecasting
system in place and manufactures products based on last month’s sales. Twenty-
four months of sales data are available and are given in Table P-16.
142 CHAPTER 4 Moving Averages and Smoothing Methods

TABLE P-14 Triton Sales per Share, 1974–1999

Year Sales per Share Year Sales per Share

1974 .93 1987 5.33

1975 1.35 1988 8.12
1976 1.48 1989 10.65
1977 2.36 1990 12.06
1978 2.45 1991 11.63
1979 2.52 1992 6.58
1980 2.81 1993 2.96
1981 3.82 1994 1.58
1982 5.54 1995 2.99
1983 7.16 1996 3.69
1984 1.93 1997 3.98
1985 5.17 1998 4.39
1986 7.72 1999 6.85

Source: The Value Line Investment Survey (New York: Value Line,
1990, 1993, 1996, 1999)

TABLE P-15 Quarterly Revenues for Consolidated Edison

($ millions), 1985–June 2002

Year Mar. 31 Jun. 30 Sept. 30 Dec. 31

1985 1,441 1,209 1,526 1,321

1986 1,414 1,187 1,411 1,185
1987 1,284 1,125 1,493 1,192
1988 1,327 1,102 1,469 1,213
1989 1,387 1,218 1,575 1,371
1990 1,494 1,263 1,613 1,369
1991 1,479 1,330 1,720 1,344
1992 1,456 1,280 1,717 1,480
1993 1,586 1,396 1,800 1,483
1994 1,697 1,392 1,822 1,461
1995 1,669 1,460 1,880 1,528
1996 1,867 1,540 1,920 1,632
1997 1,886 1,504 2,011 1,720
1998 1,853 1,561 2,062 1,617
1999 1,777 1,479 2,346 1,889
2000 2,318 2,042 2,821 2,250
2001 2,886 2,112 2,693 1,943
2002 2,099 1,900 — —

Source: The Value Line Investment Survey (New York: Value Line, 1990, 1993,
1996, 1999, 2001).

a. Plot the sales data as a time series. Are the data seasonal?
Hint: For monthly data, the seasonal period is s = 12. Is there a pattern (e.g.,
summer sales relatively low, fall sales relatively high) that tends to repeat
itself every 12 months?
 CHAPTER 4 Moving Averages and Smoothing Methods 143

TABLE P-16

Month Sales Month Sales

January 2005 430 January 2006 442

February 420 February 449
March 436 March 458
April 452 April 472
May 477 May 463
June 420 June 431
July 398 July 400
August 501 August 487
September 514 September 503
October 532 October 503
November 512 November 548
December 410 December 432

b. Use a naive model to generate monthly sales forecasts (e.g., the February
2005 forecast is given by the January 2005 value, and so forth). Compute the
MAPE.
c. Use simple exponential smoothing with a smoothing constant of .5 and an initial
smoothed value of 430 to generate sales forecasts for each month. Compute the
MAPE.
d. Do you think either of the models in parts b and c is likely to generate accurate
forecasts for future monthly sales? Explain.
e. Use Minitab and Winters’ multiplicative smoothing method with smoothing con-
stants a = b = g = .5 to generate a forecast for January 2007. Save the residuals.
f. Refer to part e. Compare the MAPE for Winters’ method from the computer
printout with the MAPEs in parts b and c. Which of the three forecasting proce-
dures do you prefer?
g. Refer to part e. Compute the autocorrelations (for six lags) for the residuals
from Winters’ multiplicative procedure. Do the residual autocorrelations sug-
gest that Winters’ procedure works well for these data? Explain.
17. Consider the gasoline purchases for the Spokane Transit Authority given in Table
4-2. In Example 4.3, a five-week moving average is used to smooth the data and
generate forecasts.
a. Use Minitab to smooth the Spokane Transit Authority data using a four-week
moving average. Which moving average length, four weeks or five weeks,
appears to represent the data better? Explain.
b. Use Minitab to smooth the Spokane Transit Authority data using simple expo-
nential smoothing. Compare your results with those in part a. Which procedure,
four-week moving average or simple exponential smoothing, do you prefer for
these data? Explain.
18. Table P-18 contains the number of severe earthquakes (those with a Richter scale
magnitude of 7 and above) per year for the years 1900–1999.
a. Use Minitab to smooth the earthquake data with moving averages of orders
k = 5, 10, and 15. Describe the nature of the smoothing as the order of the mov-
ing average increases. Do you think there might be a cycle in these data? If so,
provide an estimate of the length (in years) of the cycle.
144 CHAPTER 4 Moving Averages and Smoothing Methods

TABLE P-18 Number of Severe Earthquakes, 1900–1999

Year Number Year Number Year Number Year Number

1900 13 1927 20 1954 17 1981 14

1901 14 1928 22 1955 19 1982 10
1902 8 1929 19 1956 15 1983 15
1903 10 1930 13 1957 34 1984 8
1904 16 1931 26 1958 10 1985 15
1905 26 1932 13 1959 15 1986 6
1906 32 1933 14 1960 22 1987 11
1907 27 1934 22 1961 18 1988 8
1908 18 1935 24 1962 15 1989 7
1909 32 1936 21 1963 20 1990 12
1910 36 1937 22 1964 15 1991 11
1911 24 1938 26 1965 22 1992 23
1912 22 1939 21 1966 19 1993 16
1913 23 1940 23 1967 16 1994 15
1914 22 1941 24 1968 30 1995 25
1915 18 1942 27 1969 27 1996 22
1916 25 1943 41 1970 29 1997 20
1917 21 1944 31 1971 23 1998 16
1918 21 1945 27 1972 20 1999 23
1919 14 1946 35 1973 16
1920 8 1947 26 1974 21
1921 11 1948 28 1975 21
1922 14 1949 36 1976 25
1923 23 1950 39 1977 16
1924 18 1951 21 1978 18
1925 17 1952 17 1979 15
1926 19 1953 22 1980 18

Source: U.S. Geological Survey Earthquake Hazard Program.

b. Use Minitab to smooth the earthquake data using simple exponential smooth-
ing. Store the residuals and generate a forecast for the number of severe earth-
quakes in the year 2000. Compute the residual autocorrelations. Does simple
exponential smoothing provide a reasonable fit to these data? Explain.
c. Is there a seasonal component in the earthquake data? Why or why not?
19. Table P-23 in Chapter 3 contains the quarterly income before extraordinary items
for Southwest Airlines for the years 1988–1999.
a. Using Minitab, smooth the Southwest Airlines income data using Holt’s linear
smoothing and store the residuals. Compute the residual autocorrelations. Does
it appear as if Holt’s smoothing procedure represents these data well? If not,
what time series component (trend, cycle, seasonal) is not accounted for by
Holt’s method?
b. Use Minitab to smooth the Southwest Airlines income data using Winters’ mul-
tiplicative exponential smoothing. Store the residuals and generate forecasts of
income for the four quarters of 2000. Compute the residual autocorrelations.
Does Winters’ smoothing technique fit the income data well? Do the forecasts
seem reasonable? Discuss.
 CHAPTER 4 Moving Averages and Smoothing Methods 145

TABLE P-20 Quarterly Sales for The Gap, Fiscal Years 1980–2004

Year Q1 Q2 Q3 Q4

1980 65.3 72.1 107.1 125.3

1981 76.0 76.9 125.4 139.1
1982 84.4 86.0 125.4 139.1
1983 87.3 90.9 133.3 169.0
1984 98.0 105.3 141.4 189.4
1985 105.7 120.1 181.7 239.8
1986 160.0 164.8 224.8 298.4
1987 211.1 217.7 273.6 359.6
1988 241.3 264.3 322.8 423.7
1989 309.9 325.9 405.6 545.2
1990 402.4 405.0 501.7 624.7
1991 490.3 523.1 702.0 803.5
1992 588.9 614.1 827.2 930.2
1993 643.6 693.2 898.7 1,060.2
1994 751.7 773.1 988.3 1,209.8
1995 848.7 868.5 1,155.9 1,522.1
1996 1,113.2 1,120.3 1,383.0 1,667.9
1997 1,231.2 1,345.2 1,765.9 2,165.5
1998 1,719.7 1,905.0 2,399.9 3,029.8
1999 2,277.7 2,453.3 3,045.4 3,858.9
2000 2,732.0 2,947.7 3,414.7 4,579.1
2001 3,179.7 3,245.2 3,333.4 4,089.6
2002 2,890.8 3,268.3 3,645.0 4,650.6
2003 3,353.0 3,685.0 3,929.0 4,887.0
2004 3,668.0 3,721.0 3,980.0 4,898.0

Source: Based on The Value Line Investment Survey (New York: Value Line, various years),
and 10K filings with the Securities and Exchange Commission.

20. Table P-20 contains quarterly sales ($MM) of The Gap for fiscal years 1980–2004.
Plot The Gap sales data as a time series and examine its properties. The objec-
tive is to generate forecasts of sales for the four quarters of 2005. Select an appro-
priate smoothing method for forecasting and justify your choice.

CASES

CASE 4-1 THE SOLAR ALTERNATIVE

COMPANY4
The Solar Alternative Company is about to enter founded the company. The Johnsons started the
its third year of operation. Bob and Mary Johnson, Solar Alternative Company to supplement their
who both teach science in the local high school, teaching income. Based on their research into solar

4This case was contributed by William P. Darrow of Towson State University, Towson, Maryland.
146 CHAPTER 4 Moving Averages and Smoothing Methods

energy systems, they were able to put together a than the first. Many of the second-year customers
solar system for heating domestic hot water. The sys- are neighbors of people who had purchased the sys-
tem consists of a 100-gallon fiberglass storage tank, tem in the first year. Apparently, after seeing the
two 36-foot solar panels, electronic controls, PVC system operate successfully for a year, others were
pipe, and miscellaneous fittings. willing to try the solar concept. Sales occur through-
The payback period on the system is 10 years. out the year. Demand for the system is greatest in
Although this situation does not present an attrac- late summer and early fall, when homeowners typi-
tive investment opportunity from a financial point cally make plans to winterize their homes for the
of view, there is sufficient interest in the novelty of upcoming heating season.
the concept to provide a moderate level of sales. With the anticipated growth in the business,
The Johnsons clear about $75 on the $2,000 price of the Johnsons felt that they needed a sales forecast
an installed system, after costs and expenses. to manage effectively in the coming year. It usually
Material and equipment costs account for 75% of takes 60 to 90 days to receive storage tanks after
the installed system cost. An advantage that helps placing the order. The solar panels are available
to offset the low profit margin is the fact that the off the shelf most of the year. However, in the late
product is not profitable enough to generate any summer and throughout the fall, the lead time can
significant competition from heating contractors. be as much as 90 to 100 days. Although there is
The Johnsons operate the business out of their limited competition, lost sales are nevertheless a
home. They have an office in the basement, and real possibility if the potential customer is asked to
their one-car garage is used exclusively to store the wait several months for installation. Perhaps more
system components and materials. As a result, over- important is the need to make accurate sales pro-
head is at a minimum. The Johnsons enjoy a modest jections to take advantage of quantity discount
supplemental income from the company’s opera- buying. These factors, when combined with the
tion. The business also provides a number of tax high cost of system components and the limited
advantages. storage space available in the garage, make it nec-
Bob and Mary are pleased with the growth of essary to develop a reliable forecast. The sales his-
the business. Although sales vary from month to tory for the company’s first two years is given in
month, overall the second year was much better Table 4-10.

TABLE 4-10

Month 2005 2006 Month 2005 2006

January 5 17 July 23 44
February 6 14 August 26 41
March 10 20 September 21 33
April 13 23 October 15 23
May 18 30 November 12 26
June 15 38 December 14 17

ASSIGNMENT

1. Identify the model Bob and Mary should use as 2. Forecast sales for 2007.
the basis for their business planning in 2007, and
discuss why you selected this model.
 CHAPTER 4 Moving Averages and Smoothing Methods 147

CASE 4-2 MR. TUX

John Mosby, owner of several Mr. Tux rental stores, the program finds the optimum α value to be
is beginning to forecast his most important business .867. Again he records the appropriate error
variable, monthly dollar sales (see the Mr. Tux measurements:
cases in previous chapters). One of his employees,
MAD = 46,562
Virginia Perot, gathered the sales data shown in
Case 2-2. John now wants to create a forecast based MAPE = 44.0%
on these sales data using moving average and John asks the program to use Holt’s linear exponential
exponential smoothing techniques. smoothing on his data. This program uses the expo-
John used Minitab in Case 3-2 to determine nential smoothing method but can account for a trend
that these data are both trending and seasonal. He in the data as well. John has the program use a
has been told that simple moving averages and smoothing constant of .4 for both α and b. The two
exponential smoothing techniques will not work summary error measurements for Holt’s method are
with these data but decides to find out for himself.
MAD = 63,579
He begins by trying a three-month moving aver-
age. The program calculates several summary fore- MAPE = 59.0%
cast error measurements. These values summarize John is surprised to find larger measurement
the errors found in predicting actual historical data errors for this technique. He decides that the seasonal
values using a three-month moving average. John aspect of the data is the problem. Winters’ multiplica-
decides to record two of these error measurements: tive exponential smoothing is the next method John
tries. This method can account for seasonal factors as
MAD = 54,373
well as trend. John uses smoothing constants of
MAPE = 47.0% a = .2, b = .2, and g = .2. Error measurements are
The MAD (mean absolute deviation) is the average MAD = 25,825
absolute error made in forecasting past values. Each
MAPE = 22.0%
forecast using the three-month moving average
method is off by an average of 54,373. The MAPE When John sits down and begins studying the results
(mean absolute percentage error) shows the error as of his analysis, he is disappointed. The Winters’
a percentage of the actual value to be forecast. The method is a big improvement; however, the MAPE
average error using the three-month moving aver- is still 22%. He had hoped that one of the methods
age technique is 47%, or almost half as large as the he used would result in accurate forecasts of past
value to be forecast. periods; he could then use this method to forecast
Next, John tries simple exponential smoothing. the sales levels for coming months over the next
The program asks him to input the smoothing year. But the average absolute errors (MADs) and
constant (α) to be used or to ask that the optimum percentage errors (MAPEs) for these methods lead
α value be calculated. John does the latter, and him to look for another way of forecasting.

QUESTIONS
1. Summarize the forecast error level for the best methods. What should John do, for example, to
method John has found using Minitab. determine the adequacy of the Winters’ fore-
2. John used the Minitab default values for α, b, casting technique?
and g. John thinks there are other choices for 4. Although not calculated directly in Minitab, the
these parameters that would lead to smaller MPE (mean percentage error) measures fore-
error measurements. Do you agree? cast bias. What is the ideal value for the MPE?
3. Although disappointed with his initial results, What is the implication of a negative sign on the
this may be the best he can do with smoothing MPE?
148 CHAPTER 4 Moving Averages and Smoothing Methods

CASE 4-3 CONSUMER CREDIT

COUNSELING

The Consumer Credit Counseling (CCC) operation data for the number of new clients seen by CCC for
was described in Case 1-2. The executive director, the period from January 1985 through March 1993
Marv Harnishfeger, concluded that the most impor- (see Case 3-3). Dorothy then used autocorrelation
tant variable that CCC needed to forecast was the analysis to explore the data pattern. Use the results
number of new clients that would be seen in the rest of this investigation to complete the following
of 1993. Marv provided Dorothy Mercer monthly tasks.

ASSIGNMENT
1. Develop a naive model to forecast the number 4. Evaluate these forecasting methods using the
of new clients seen by CCC for the rest of 1993. forecast error summary measures presented in
2. Develop a moving average model to forecast Chapter 3.
the number of new clients seen by CCC for the 5. Choose the best model and forecast new clients
rest of 1993. for the rest of 1993.
3. Develop an exponential smoothing procedure 6. Determine the adequacy of the forecasting
to forecast the number of new clients seen by model you have chosen.
CCC for the rest of 1993.

CASE 4-4 MURPHY BROTHERS FURNITURE

Julie Murphy knows that most important operat- investment opportunities and must forecast the
ing decisions depend, to some extent, on a fore- demand for a new furniture line.
cast. For Murphy Brothers Furniture, sales fore- In Case 3-1A, Julie Murphy used monthly sales for
casts impact adding new furniture lines or all retail stores from 1983 through 1995 (see Table 3-8)
dropping old ones; planning purchases; setting to develop a pattern for Murphy Brothers Furniture
sales quotas; and making personnel, advertising, sales. In Case 3-1B, Glen Murphy discovered actual
and financial decisions. Specifically, Julie is aware sales data for the past four years, 1992 through 1995
of several current forecasting needs. She knows (see Table 3-9). Julie was not excited about her father’s
that the production department has to schedule discovery because she was not sure which set of data to
employees and determine raw material orders for use to develop a forecast for 1996. She determined that
the next month or two. She also knows that her sales for all retail stores had somewhat the same pat-
dad, Glen Murphy, needs to determine the best tern as actual Murphy Brothers’ sales data.

QUESTIONS
1. Do any of the forecasting models studied in this 3. Which data set and forecasting model should
chapter work with the national sales data? Julie use to forecast sales for 1996?
2. Do any of the forecasting models studied in this
chapter work with the actual Murphy Brothers’
sales data?
 CHAPTER 4 Moving Averages and Smoothing Methods 149

CASE 4-5 FIVE-YEAR REVENUE PROJECTION FOR

DOWNTOWN RADIOLOGY

Some years ago Downtown Radiology developed a METHODOLOGY

medical imaging center that was more complete and Medical Procedures
technologically advanced than any located in an area The following steps were implemented in order to
of eastern Washington and northern Idaho called the complete the five-year projection of revenue. An
Inland Empire.The equipment planned for the center analysis of the past number of procedures was per-
equaled or surpassed the imaging facilities of all med- formed. The appropriate forecasting model was
ical centers in the region. The center initially con- developed and used to determine a starting point for
tained a 9800 series CT scanner and nuclear magnetic the projection of each procedure.
resonance imaging (MRI) equipment.The center also
included ultrasound, nuclear medicine, digital sub- 1. The market area was determined for each type
traction angiography (DSA), mammography, and of procedure, and population forecasts were
conventional radiology and fluoroscopy equipment. obtained for 1986 and 1990.
Ownership interest was made available in a type of 2. Doctor referral patterns were studied to deter-
public offering, and Downtown Radiology used an mine the percentage of doctors who refer to
independent evaluation of the market. Professional Downtown Radiology and the average number
Marketing Associates, Inc., evaluated the market and of referrals per doctor.
completed a five-year projection of revenue. 3. National rates were acquired from the National
Center for Health Statistics. These rates were
STATEMENT OF THE PROBLEM compared with actual numbers obtained from
The purpose of this study is to forecast revenue for the Hospital Commission.
the next five years for the proposed medical imaging 4. Downtown Radiology’s market share was calcu-
center, assuming you are employed by Professional lated based on actual CT scans in the market
Marketing Associates, Inc., in the year 1984. area. (Market share for other procedures was
determined based on Downtown Radiology’s
OBJECTIVES share compared to rates provided by the
The objectives of this study are to National Center for Health Statistics.)
• Identify market areas for each type of medical Assumptions
procedure to be offered by the new facility. The following assumptions, which were necessary to
• Gather and analyze existing data on market develop a quantitative forecast, were made:
area revenue for each type of procedure to be
• The new imaging center will be operational,
offered by the new facility.
with all equipment functional except the MRI,
• Identify trends in the health care industry that will on January 1, 1985.
positively or negatively affect revenue of proce- • The nuclear magnetic resonance imaging equip-
dures to be provided by the proposed facility. ment will be functional in April 1985.
• Identify factors in the business, marketing, or • The offering of the limited partnership will be
facility planning of the new venture that will pos- successfully marketed to at least 50 physicians
itively or negatively impact revenue projections. in the service area.
• Physicians who have a financial interest in the
• Analyze past procedures of Downtown
new imaging center will increase their referrals
Radiology when compiling a database for the
to the center.
forecasting model to be developed.
• There will be no other MRIs in the market area
• Utilize the appropriate quantitative forecasting before 1987.
model to arrive at five-year revenue projections • The new imaging center will offer services at
for the proposed center. lower prices than the competition.
150 CHAPTER 4 Moving Averages and Smoothing Methods

• An effective marketing effort will take place, ANALYSIS OF PAST DATA

especially concentrating on large employers, Office X-Rays
insurance groups, and unions. The number of X-ray procedures performed was
• The MRI will replace approximately 60% of the analyzed from July 1981 to May 1984. The data
head scans that are presently done with the CT included diagnostic X-rays, gastrointestinal X-rays,
scanner during the first six months of operation breast imaging, injections, and special procedures.
and 70% during the next 12 months. Examination of these data indicated that no trend or
• The general public will continue to pressure the seasonal or cyclical pattern is present. For this rea-
health care industry to hold down costs. son, simple exponential smoothing was chosen as the
• Costs of outlays in the health care industry rose appropriate forecasting method. Various smoothing
13.2% annually from 1971 to 1981. The Health constants were examined, and a smoothing constant
Care Financing Administration estimates that of .3 was found to provide the best model.The results
the average annual rate of increase will be are presented in Figure 4-14. The forecast for period
reduced to approximately 11% to 12% between 36, June 1984, is 855 X-ray procedures.
1981 and 1990 (Industry Surveys, April 1984).
• Insurance firms will reimburse patients for (at Office Ultrasound
worst) 0% up to (at best) 100% of the cost of The number of ultrasound procedures performed
magnetic resonance imaging (Imaging News, was analyzed from July 1981 to May 1984. Figure 4-15
February 1984). shows the data pattern. Again, no trend or seasonal
or cyclical pattern was present. Exponential smooth-
ing with a smoothing constant of a = .5 was deter-
Models mined to provide the best model. The forecast for
A forecast was developed for each procedure, June 1984 plotted in Figure 4-15 is 127 ultrasound
based on past experience, industry rates, and rea- procedures.
sonable assumptions. The models were developed The number of ultrasound procedures performed
based on the preceding assumptions; however, if the by the two mobile units owned by Downtown
assumptions are not valid, the models will not be Radiology was analyzed from July 1981 to May 1984.
accurate. Figure 4-16 shows the data pattern. An increasing

FIGURE 4-14 Simple Exponential Smoothing: Downtown

Radiology X-Rays
 CHAPTER 4 Moving Averages and Smoothing Methods 151

FIGURE 4-15 Simple Exponential Smoothing: Downtown Radiology

Ultrasound

FIGURE 4-16 Holt’s Linear Exponential Smoothing:

Downtown Radiology Nonoffice Ultrasound

trend is apparent and can be modeled using Holt’s Downtown Radiology was analyzed from August
two-parameter linear exponential smoothing. Smooth- 1982 to May 1984. Figure 4-17 shows the data pat-
ing constants of a = .5 and b = .1 are used, and the tern. The data were not seasonal and had no trend or
forecast for period 36, June 1984, is 227. cyclical pattern. For this reason, simple exponential
smoothing was chosen as the appropriate forecast-
Nuclear Medicine Procedures ing method. A smoothing factor of a = .5 was found
The number of nuclear medicine procedures to provide the best model. The forecast for period
performed by the two mobile units owned by 23, June 1984, is 48 nuclear medicine procedures.
152 CHAPTER 4 Moving Averages and Smoothing Methods

FIGURE 4-17 Simple Exponential Smoothing: Downtown

Radiology Nonoffice Nuclear Medicine

Office CT Scans (see Chapters 8 and 9) was examined and compared

The number of CT scans performed was also analyzed to an exponential smoothing model with a smoothing
from July 1981 to May 1984. Seasonality was not constant of a = .461. The larger smoothing constant
found, and the number of CT scans did not seem to gives the most recent observation more weight in the
have a trend. However, a cyclical pattern seemed to be forecast.The exponential smoothing model was deter-
present. Knowing how many scans were performed mined to be better than the autoregressive model,
last month would be important in forecasting what is and Figure 4-18 shows the projection of the number of
going to happen this month. An autoregressive model CT scans for period 36, June 1984, to be 221.

FIGURE 4-18 Simple Exponential Smoothing: Downtown Radiology

CT Scans
 CHAPTER 4 Moving Averages and Smoothing Methods 153

MARKET AREA ANALYSIS referrals per procedure is extremely large. A few

Market areas were determined for procedures cur- doctors referred an extremely large percentage of
rently done by Downtown Radiology by examining the procedures done by Downtown Radiology. If a
patient records and doctor referral patterns. Market few new many-referral doctors are recruited, they
areas were determined for procedures not currently can have a major effect on the total number of pro-
done by Downtown Radiology by investigating the cedures done by Downtown Radiology.
competition and analyzing the geographical areas Finally, the effect that a new imaging center will
they served. have on Downtown Radiology’s market share must
be estimated. The new imaging center will have the
CT Scanner Market Area best equipment and will be prepared to do the total
The market area for CT scanning for the proposed spectrum of procedures at a lower cost. The number
medical imaging center includes Spokane, Whitman, of new doctors referring should increase on the basis
Adams, Lincoln, Stevens, and Pend Oreille Counties of word of mouth from the new investing doctors.
in Washington and Bonner, Boundary, Kootenai, If insurance companies, large employers, and/or
Benewah, and Shoshone Counties in Idaho. Based unions enter into agreements with the new imaging
on the appropriate percentage projections, the CT center, Downtown Radiology should be able to
scanning market area will have a population of increase its share of the market by at least 4% in
630,655 in 1985 and 696,018 in 1990. 1985, 2% in 1986, and 1% in 1987 and retain this
market share in 1988 and 1989. This market share
Quantitative Estimates increase will be referred to as the total imaging effect
To project revenue, it is necessary to determine in the rest of this report.
certain quantitative estimates. The most important
estimate involves the number of doctors who will par- Revenue Projections
ticipate in the limited partnership. The estimate used Revenue projections were completed for every pro-
in future computations is that at least 8% of the doc- cedure. Only the projections for the CT scanner are
tor population of Spokane County will participate. shown in this case.
The next uncertainty that must be quantified
involves the determination of how the referral pat- CT Scan Projections
tern will be affected by the participation of 50 doc- Based on the exponential smoothing model and
tors in the limited partnership. It is assumed that 30 what has already taken place in the first five months
of the doctors who presently refer to Downtown of 1984, the forecast of CT scans for 1984 (January
Radiology will join the limited partnership. Of the 1984 to January 1985) is 2,600.
30 who join, it is assumed that 10 will not increase The National Center for Health Statistics reports
their referrals and 20 will double their referrals. It is a rate of 261 CT scans per 100,000 population per
also assumed that 20 doctors who had never month. Using the population of 630,655 projected
referred to Downtown Radiology will join the lim- for the CT scan market area, the market should be
ited partnership and will begin to refer at least half 19,752 procedures for all of 1985. The actual number
of their work to Downtown Radiology. of CT scans performed in the market area during
The quantification of additional doctor referrals 1983 was estimated to be 21,600. This estimate was
should be clarified with some qualitative observa- based on actual known procedures for Downtown
tions. The estimate of 50 doctors joining the pro- Radiology (2,260), Sacred Heart (4,970), Deaconess
posed limited partnership is conservative. There is a (3,850), Valley (2,300), and Kootenai (1,820) and on
strong possibility that doctors from areas outside of estimates for Radiation Therapy (2,400) and
Spokane County may join. Traditionally, the doctor Northwest Imaging (4,000). If the estimates are
referral pattern changes very slowly. However, the accurate, Downtown Radiology had a market share
sudden competitive nature of the marketplace will of approximately 10.5% in 1983. The actual values
probably have an impact on doctor referrals. If the were also analyzed for 1982, and Downtown
limited partnership is marketed to doctors in spe- Radiology was projected to have approximately
cialties with high radiology referral potential, the 15.5% of the CT scan market during that year.
number of referrals should increase more than Therefore, Downtown Radiology is forecast to aver-
projected. The variability in the number of doctor age about 13% of the market.
154 CHAPTER 4 Moving Averages and Smoothing Methods

Based on the increased referrals from doctors

TABLE 4-11 Five-Year Projected Revenue for
belonging to the limited partnership and an analysis CT Scans
of the average number of referrals of CT scans, an
increase of 320 CT scans is projected for 1985 from Year Procedures Revenue ($)
this source. If actual values for 1983 are used, the
1985 3,138 1,129,680
rate for the Inland Empire CT scan market area is
1986 2,531 1,012,400
3,568 (21,600/6.054) per 100,000 population. If this 1987 2,716
.
1,205,904
pattern continues, the number of CT scans in the 1988 2,482 1,223,626
market area will increase to 22,514 (3,568 * 6.31) in 1989 2,529 1,383,363
1985. Therefore, Downtown Radiology’s market
share is projected to be 13% (2,920/22,514). When
the 4% increase in market share based on total
imaging is added, Downtown Radiology’s market for the next five years.The cost of these procedures is
share increases to 17.0%, and its projected number estimated to increase approximately 11% per year.
of CT scans is 3,827 (22,514 * .17). Without the effect of the MRI, the projection
However, research seems to indicate that the MRI for CT scans in 1986 is estimated to be 4,363
will eventually replace a large number of CT head ( 6.31 * 1.02 * 3,568 * .19). However, if 60% are
scans (Applied Radiology, May/June 1983, and CT head scans and the MRI replaces 70% of the
Diagnostic Imaging, February 1984). The National head scans, the projected number of CT scans
Center for Health Statistics indicated that 60% of all should drop to 2,531 [4,363 - 14,363 * .60 * .702].
CT scans were of the head. Downtown Radiology The projection of CT scans without the MRI
records showed that 59% of its CT scans in 1982 were effect for 1987 is 4,683 (6.31 * 1.04 * 3,568 * .20).
head scans and 54% in 1983. If 60% of Downtown The forecast with the MRI effect is 2,716
Radiology’s CT scans are of the head and the MRI [4,683 - 14,683 * .60 * .702].
approach replaces approximately 60% of them, new The projection of CT scans without the MRI
projections for CT scans in 1985 are necessary. Since effect for 1988 is 4,773 (6.31 * 1.06 * 3,568 * .20).
the MRI will operate for only half the year, a reduc- The forecast with the MRI effect is 2,482 [4,773 -
tion of 689 (3,827>2 * .60 * .60) CT scans is forecast. 14,773 * .60 * .802].
The projected number of CT scans for 1985 is The projection of CT scans without the MRI
3,138. The average cost of a CT scan is $360, and the effect for 1989 is 4,863 (6.31 * 1.08 * 3,568 * .20).
projected revenue from CT scans is $1,129,680. Table The forecast with the MRI effect is 2,529 [4,863 -
4-11 shows the projected revenue from CT scans 14,863 * .60 * .802].

QUESTION
1. Downtown Radiology’s accountant projected Downtown Radiology’s management must
that revenue would be considerably higher make a decision concerning the accuracy of
than that provided by Professional Marketing Professional Marketing Associates’ projections.
Associates. Since ownership interest will be You are asked to analyze the report. What rec-
made available in some type of public offering, ommendations would you make?

CASE 4-6 WEB RETAILER

Example 1.2 introduced Pat Niebuhr and his team, needs to take a monthly forecast of the total orders
who are responsible for developing a global staffing and contacts per order (CPO) supplied by the finance
plan for the contact centers of a large web retailer. Pat department and ultimately forecast the number of
 CHAPTER 4 Moving Averages and Smoothing Methods 155
customer contacts (phone, email, and so forth) arriv- Pat thinks it might be a good idea to use the
ing at the retailer’s contact centers weekly.The contact historical data to generate forecasts of orders and
centers are open 24 hours 7 days a week and must be contacts per order directly. He is interested in
appropriately staffed to maintain a high service level. determining whether these forecasts are more accu-
The retailer recognizes that excellent customer service rate than the forecasts for these quantities that the
will likely result in repeat visits and purchases. finance department derives from revenue projec-
The key equation for Pat and his team is tions. As a start, Pat and his team are interested in
the data patterns of the monthly historical orders
Contacts = Orders * CPO
and contacts per order and decide to plot these time
Historical data provide the percentage of contacts series and analyze the autocorrelations. The data
for each day of the week. For example, historically are given in Table 4-12 and plotted in Figures 4-19
9.10% of the weekly contacts occur on Sundays, and 4-20. The autocorrelations are shown in Figures
17.25% of the weekly contacts occur on Mondays, 4-21 and 4-22.
and so forth. Keeping in mind the number of Pat is intrigued with the time series plots and
Sundays, Mondays, and so on in a given month, the autocorrelation functions and feels a smoothing
monthly forecasts of contacts can be converted to procedure might be the right tool for fitting the time
weekly forecasts of contacts. It is the weekly fore- series for orders and contacts per order and for gen-
casts that are used for staff planning purposes. erating forecasts.

TABLE 4-12 Orders and Contacts per

Order (CPO) for Web Retailer,
June 2001–June 2003

Month Orders CPO

Jun-01 3,155,413 0.178

Jul-01 3,074,723 0.184
Aug-01 3,283,838 0.146
Sep-01 2,772,971 0.144
Oct-01 3,354,889 0.144
Nov-01 4,475,792 0.152
Dec-01 5,944,348 0.152
Jan-02 3,742,334 0.174
Feb-02 3,681,370 0.123
Mar-02 3,546,647 0.121
Apr-02 3,324,321 0.117
May-02 3,318,181 0.116
Jun-02 3,181,115 0.129
Jul-02 3,022,091 0.131
Aug-02 3,408,870 0.137
Sep-02 3,501,779 0.140
Oct-02 3,712,424 0.144
Nov-02 4,852,090 0.129
Dec-02 7,584,065 0.124
Jan-03 4,622,233 0.136
Feb-03 3,965,540 0.116
Mar-03 3,899,108 0.111
Apr-03 3,670,589 0.108
May-03 3,809,110 0.101
Jun-03 4,159,358 0.105
156 CHAPTER 4 Moving Averages and Smoothing Methods

FIGURE 4-19 Time Series Plot of Orders, June 2001–June 2003

FIGURE 4-20 Time Series Plot of Contacts per Order (CPO), June
2001–June 2003

QUESTIONS
1. What did Pat and his team learn about the data 2. Fit an appropriate smoothing procedure to the
patterns for orders and contacts per order from the orders time series and generate forecasts for
time series plots and autocorrelation functions? the next four months. Justify your choice.
 CHAPTER 4 Moving Averages and Smoothing Methods 157

FIGURE 4-21 Autocorrelation Function for Orders

FIGURE 4-22 Autocorrelation Function for Contacts per Order (CPO)

3. Fit an appropriate smoothing procedure to the contacts directly instead of multiplying forecasts
contacts per order time series and generate fore- of orders and contacts per order to get a fore-
casts for the next four months. Justify your choice. cast. Does this seem reasonable? Why?
4. Use the results for Questions 2 and 3 to generate 6. Many orders consist of more than one item
forecasts of contacts for the next four months. (unit). Would it be better to focus on number of
5. Pat has access to a spreadsheet with historical units and contacts per unit to get a forecast of
actual contacts. He is considering forecasting contacts? Discuss.
158 CHAPTER 4 Moving Averages and Smoothing Methods

CASE 4-7 SOUTHWEST MEDICAL CENTER

Mary Beasley is responsible for keeping track of the If so, how many new doctors should be hired and/or
number of billable visits to the Medical Oncology be reassigned from other areas?
group at Southwest Medical Center. Her anecdotal To provide some insight into the nature of the
evidence suggests that the number of visits has been demand for service, Mary opens her Excel spreadsheet
increasing and some parts of the year seem to be and examines the total number of billable visits on a
busier than others. Some doctors are beginning to monthly basis for the last several fiscal years. The data
complain about the work load and suggest they are listed in Table 4-13.
don’t always have enough time to interact with indi- A time series plot of Mary’s data is shown in
vidual patients. Will additional medical staff be Figure 4-23. As expected, the time series shows
required to handle the apparently increasing demand? an upward trend, but Mary is not sure if there is a

TABLE 4-13 Total Billable Visits to Medical Oncology, FY1995–FY2004

Year Sept. Oct. Nov. Dec. Jan. Feb. Mar. Apr. May Jun. Jul. Aug.

FY1994–95 725 789 893 823 917 811 1,048 970 1,082 1,028 1,098 1,062
FY1995–96 899 1,022 895 828 1,011 868 991 970 934 784 1,028 956
FY1996–97 916 988 921 865 998 963 992 1,118 1,041 1,057 1,200 1,062
FY1997–98 1,061 1,049 829 1,029 1,120 1,084 1,307 1,458 1,295 1,412 1,553 1,480
FY1998–99 1,554 1,472 1,326 1,231 1,251 1,092 1,429 1,399 1,341 1,409 1,367 1,483
FY1999–00 1,492 1,650 1,454 1,373 1,466 1,477 1,466 1,182 1,208 1,132 1,094 1,061
FY2000–01 1,018 1,233 1,112 1,107 1,305 1,181 1,391 1,324 1,259 1,236 1,227 1,294
FY2001–02 1,083 1,404 1,329 1,107 1,313 1,156 1,184 1,404 1,310 1,200 1,396 1,373
FY2002–03 1,259 1,295 1,100 1,097 1,357 1,256 1,350 1,318 1,271 1,439 1,441 1,352
FY2003–04 1,339 1,351 1,197 1,333 1,339 1,307 — — — — — —

FIGURE 4-23 Time Series Plot of Total Visits to Medical

Oncology, FY1995–FY2004
 CHAPTER 4 Moving Averages and Smoothing Methods 159
seasonal component in the total visits series. She and so forth. Mary knows from a course in her
decides to investigate this issue by constructing the Executive MBA program that Winters’ smoothing
autocorrelation function. If a seasonal component procedure might be a good way to generate fore-
exists with monthly data, Mary expects to see fairly casts of future visits if trend and seasonal compo-
large autocorrelations at the seasonal lags: 12, 24, nents are present.

QUESTIONS
1. What did Mary’s autocorrelation analysis 3. Given the results in Question 2, do you think it
show? is likely another forecasting method would gen-
2. Fit an appropriate smoothing procedure to erate “better” forecasts? Discuss.
Mary’s data, examine the residual autocor- 4. Do you think additional medical staff might be
relations, and generate forecasts for the remain- needed to handle future demand? Write a brief
der of FY2003–04. Do these forecasts seem report summarizing Mary’s data analysis and
reasonable? the implications for additional staff.

CASE 4-8 SURTIDO COOKIES

In Case 3-5, Jame Luna investigated the data pattern nizes that important first steps in selecting a fore-
of monthly Surtido cookie sales (see Table 3-12). In casting method are plotting the sales time series and
that case, Karin, one of the members of Jame’s team, conducting an autocorrelation analysis. He knows
suggested forecasts of future sales for a given month you can often learn a lot by simply examining a plot
might be generated by simply using the historical of your time series. Moreover, the autocorrelations
average sales for that month. After learning some- tend to reinforce the pattern observed from the plot.
thing about smoothing methods however, Jame Jame is ready to begin with the goal of generating
thinks a smoothing procedure might be a better way forecasts of monthly cookie sales for the remaining
to construct forecasts of future sales. Jame recog- months of 2003.

QUESTIONS
1. What pattern(s) did Jame observe from a time and produce forecasts for the remaining months
series plot of Surtido cookie sales? of 2003.
2. Are the autocorrelations consistent with the pat- 4. Use Karin’s historical monthly average sugges-
tern(s) Jame observed in the time series plot? tion to construct forecasts for the remaining
3. Select and justify an appropriate smoothing months of 2003. Which forecasts, yours or
procedure for forecasting future cookie sales Karin’s, do you prefer? Why?

Minitab Applications

The problem. In Example 4.3, the Spokane Transit Authority data need to be forecast
using a five-week moving average.
Minitab Solution
1. Enter the Spokane Transit Authority data shown in Table 4-2 (see p. 112) into
column C1. Click on the following menus:
Stat>Time Series>Moving Average
160 CHAPTER 4 Moving Averages and Smoothing Methods

2. The Moving Average dialog box appears.

a. Double-click on the variable Gallons and it will appear to the right of Variable.
b. Since we want a five-month moving average, indicate 5 for MA length.
c. Do not click on the Center moving average box. We will use a centered moving
average to smooth data in Chapter 5.
d. Click on Generate forecasts and indicate 1 to the right of Number of forecasts.
e. Click on OK and Figure 4-4 will appear.
The problem. In Example 4.6, the Acme Tool Company data need to be forecast
using single exponential smoothing.
Minitab Solution
1. Enter the Acme Tool Company data shown in Table 4-1 (see p. 109) for the years
2000 through 2006 into column C1. Click on the following menus:
Stat>Time Series>Single Exponential Smoothing
2. The Single Exponential Smoothing dialog box Appears.
a. Double-click on the variable Saws and it will appear to the right of Variable.
b. Under Weight to Use in Smoothing, choose Optimal ARIMA and then click on
OK. The result is shown in Figure 4-8 (see p. 129).
The problem. In Example 4.10, the Acme Tool Company data need to be forecast
using exponential smoothing adjusted for trend and seasonality.
Minitab Solution
1. Enter the Acme Tool Company data shown in Table 4-1 (see p. 109) for the years
2000 through 2006 in column C1. Click on the following menus:
Stat>Time Series>Winters’ Method

FIGURE 4-24 Minitab Winters’ Method Dialog Box

 CHAPTER 4 Moving Averages and Smoothing Methods 161
2. The Winters’ Method dialog box appears, as shown in Figure 4-24.
a. The variable of interest is Saws.
b. Since the data are quarterly, indicate 4 for Seasonal length.
c. The Weights to Use in Smoothing are Level: 0.4; Trend: 0.1; and Seasonal: 0.3.
d. Click on Generate forecasts, and for Number of forecasts, indicate 4.
e. Click on Storage.
3. The Winters’ Method Storage dialog box appears.
a. Click on Level estimates, Trend estimates, Seasonal estimates, Fits (one-period-
ahead forecasts), and Residuals.
b. Click on OK on both the Winters’ Method Storage dialog box and the Winters’
Method dialog box. The results are shown in Table 4-9 and Figure 4-12. The
forecast for the first quarter of 2007 is 778.2.
4. To store the data for further use, click on the following menus:
File>Save Worksheet As
5. The Save Worksheet As dialog box appears.
a. Type a name such as Saws in the File Name space.
b. The Save as Type space allows you to choose how you want to save your file.
Most of the time you will select Minitab. However, you can save your file so sev-
eral software programs can use it. As an example, you could choose to save the
file as an Excel file. The file is saved as Saws.xls and will be used in the Excel
Applications section.

Excel Applications

The problem. In Example 4.5, the Acme Tool Company data were forecast using sin-
gle exponential smoothing with a smoothing constant equal to .6.
Excel Solution
1. Open the file containing the data presented in Table 4-1 (see p. 109) by clicking on
the following menus:
File>Open

Look for the file called Saws.xls.

2. Click on the following menus:
Tools>Data Analysis

The Data Analysis dialog box appears. Under Analysis Tools, choose Exponential
Smoothing and click on OK. The Exponential Smoothing dialog box, shown in
Figure 4-25, appears.
a. Enter A2:A25 in the Input Range edit box.
b. Check the Labels box.
c. Enter .4 in the Damping factor edit box, since the damping factor (1 - a) is
defined as the complement of the smoothing constant.
d. Enter B3 in the Output Range edit box. (This will put the forecast YNt opposite
the corresponding value in column A.)
e. Check the Chart Output box.
f. Now click on OK.
4. The results (column B) and the graph are shown in Figure 4-26. Note that the
Exponential Smoothing analysis tool puts formulas in the worksheet. Cell B5 is high-
lighted and the formula = 0.6 * A4 + 0.4 * B4 is shown on the formula toolbar.
162 CHAPTER 4 Moving Averages and Smoothing Methods

FIGURE 4-25 Excel Exponential Smoothing Dialog Box

FIGURE 4-26 Excel Exponential Smoothing: Example 4.5 Results

5. Notice that, although Excel starts exponential smoothing differently from

the way the smoothing was started in Table 4-7 (see the first few values in the
column corresponding to a = .6), after six or seven iterations the numbers in
 CHAPTER 4 Moving Averages and Smoothing Methods 163
column B in Figure 4-26 and the numbers in the column corresponding to
a = .6 in Table 4-7 are exactly the same. Fluctuations due to different choices of
starting values die out fairly rapidly .

References
Aaker, D. A., and R. Jacobson. “The Sophistication Work Force. Englewood Cliffs, N.J.: Prentice-Hall,
of ‘Naive’ Modeling.” International Journal of 1960.
Forecasting 3 (314) (1987): 449–452. Koehler, A. B., R. D. Snyder, and D. K. Ord.
Bowerman, B. L., R. T. O’Connell, and A. B. Koehler. “Forecasting Models and Prediction Intervals for
Forecasting, Time Series and Regression, 4th ed. the Multiplicative Holt-Winters Method.”
Belmont, CA: Thomson Brooks/Cole, 2005. International Journal of Forecasting 17 (2001):
Dalrymple, D. J., and B. E. King. “Selecting 269–286.
Parameters for Short-Term Forecasting Ledolter, J., and B. Abraham. “Some Comments on
Techniques.” Decision Sciences 12 (1981): the Initialization of Exponential Smoothing.”
661–669. Journal of Forecasting 3 (1) (1984): 79–84.
Gardner, E. S., Jr. “Exponential Smoothing: The State Makridakis, S., S. C. Wheelwright, and R. Hyndman.
of the Art.” Journal of Forecasting 4 (1985): 1–28. Forecasting Methods and Applications. New York:
Gardner, E. S., Jr., and D. G. Dannenbring. Wiley, 1998.
“Forecasting with Exponential Smoothing: Some McKenzie, E. “An Analysis of General Exponential
Guidelines for Model Selection.” Decision Smoothing.” Operations Research 24 (1976):
Sciences 11 (1980): 370–383. 131–140.
Holt, C. C. “Forecasting Seasonals and Trends by Newbold, P., and T. Bos. Introductory Business and
Exponentially Weighted Moving Averages.” Economic Forecasting, 2nd ed. Cincinnati, Ohio:
International Journal of Forecasting 20 (2004): South-Western, 1994.
5–10. Winters, P. R. “Forecasting Sales by Exponentially
Holt, C. C., F. Modigliani, J. F. Muth, and H. A. Weighted Moving Averages.” Management
Simon. Planning Production Inventories and Science 6 (1960): 324–342.
This page intentionally left blank
C H A P T E R

5 TIME SERIES AND THEIR

COMPONENTS
As we have pointed out in earlier chapters, observations of a variable Y that become
available over time are called time series data, or simply a time series. These observa-
tions are often recorded at fixed time intervals. For example, Y might represent sales,
and the associated time series could be a sequence of annual sales figures. Other exam-
ples of time series include quarterly earnings, monthly inventory levels, and weekly
exchange rates. In general, time series do not behave like random samples and require
special methods for their analysis. Observations of a time series are typically related to
one another (autocorrelated). This dependence produces patterns of variability that
can be used to forecast future values and assist in the management of business opera-
tions. Consider these situations.
American Airlines (AA) compares current reservations with forecasts based on
projections of historical patterns. Whether current reservations are lagging behind or
exceeding the projections, AA adjusts the proportion of discounted seats accordingly.
The adjustments are made for each flight segment in the AA system.
A Canadian importer of cut flowers buys from growers in the United States,
Mexico, Central America, and South America. However, because these sources pur-
chase their growing stock and chemicals from the United States, all the selling prices
are quoted in U.S. dollars at the time of the sale. An invoice is not paid immediately,
and since the Canadian–U.S. exchange rate fluctuates, the cost to the importer in
Canadian dollars is not known at the time of purchase. If the exchange rate does not
change before the invoice is paid, there is no monetary risk to the importer. If the index
rises, the importer loses money for each U.S. dollar of purchase. If the index drops, the
importer gains. The importer uses forecasts of the weekly Canadian-dollar-to-U.S.-dollar
exchange rate to manage the inventory of cut flowers.
Although time series are often generated internally and are unique to the organi-
zation, many time series of interest in business can be obtained from external sources.
Publications such as Statistical Abstract of the United States, Survey of Current
Business, Monthly Labor Review, and Federal Reserve Bulletin contain time series of
all types. These and other publications provide time series data on prices, production,
sales, employment, unemployment, hours worked, fuel used, energy produced, earn-
ings, and so on, reported on a monthly, quarterly, or annual basis. Today, extensive col-
lections of time series are available on the World Wide Web at sites maintained by U.S.
government agencies, statistical organizations, universities, and individuals.
It is important that managers understand the past and use historical data and
sound judgment to make intelligent plans to meet the demands of the future. Properly
constructed time series forecasts help eliminate some of the uncertainty associated
with the future and can assist management in determining alternative strategies.

165
166 CHAPTER 5 Time Series and Their Components

The alternative, of course, is not to plan ahead. In a dynamic business environment,

however, this lack of planning might be disastrous. A mainframe computer manufac-
turer that some years ago ignored the trend to personal computers and workstations
would have lost a large part of its market share rather quickly.
Although we will focus our attention on a model-based approach to time series
analysis that relies primarily on the data, a subjective review of the forecasting effort is
very important. The past is relevant in the search for clues about the future only to the
extent that causal conditions previously in effect continue to hold in the period ahead.
In economic and business activity, causal conditions seldom remain constant. The
multitude of causal factors at work tends to be constantly shifting, so the connections
among the past, the present, and the future must be continually reevaluated.
The techniques of time series provide a conceptual approach to forecasting that
has proved to be very useful. Forecasts are made with the aid of a set of specific formal
procedures, and judgments that follow are indicated explicitly.

DECOMPOSITION
One approach to the analysis of time series data involves an attempt to identify the
component factors that influence each of the values in a series. This identification pro-
cedure is called decomposition. Each component is identified separately. Projections of
each of the components can then be combined to produce forecasts of future values of
the time series. Decomposition methods are used for both short-run and long-run fore-
casting. They are also used to simply display the underlying growth or decline of a
series or to adjust the series by eliminating one or more of the components.
Analyzing a time series by decomposing it into its component parts has a long
history. Recently, however, decomposition methods of forecasting have lost some of
their luster. Projecting the individual components into the future and recombining
these projections to form a forecast of the underlying series often does not work very
well in practice. The difficulty lies in getting accurate forecasts of the components. The
development of more-flexible model-based forecasting procedures (some of which we
discuss in later chapters) has made decomposition primarily a tool for understanding a
time series rather than a forecasting method in its own right.
To understand decomposition, we start with the four components of a time series
that were introduced in Chapter 3. These are the trend component, the cyclical compo-
nent, the seasonal component, and the irregular or random component.

1. Trend. The trend is the component that represents the underlying growth (or
decline) in a time series. The trend may be produced, for example, by consistent
population change, inflation, technological change, and productivity increases. The
trend component is denoted by T.
2. Cyclical. The cyclical component is a series of wavelike fluctuations or cycles of
more than one year’s duration. Changing economic conditions generally produce
cycles. C denotes the cyclical component.
In practice, cycles are often difficult to identify and are frequently regarded as
part of the trend. In this case, the underlying general growth (or decline) compo-
nent is called the trend-cycle and denoted by T. We use the notation for the trend
because the cyclical component often cannot be separated from the trend.
3. Seasonal. Seasonal fluctuations are typically found in quarterly, monthly, or weekly
data. Seasonal variation refers to a more or less stable pattern of change that
appears annually and repeats itself year after year. Seasonal patterns occur
 CHAPTER 5 Time Series and Their Components 167
because of the influence of the weather or because of calendar-related events such
as school vacations and national holidays. S denotes the seasonal component.
4. Irregular. The irregular component consists of unpredictable or random fluctua-
tions. These fluctuations are the result of a myriad of events that individually may
not be particularly important but whose combined effect could be large. I denotes
the irregular component.

To study the components of a time series, the analyst must consider how the
components relate to the original series. This task is accomplished by specifying a model
(mathematical relationship) that expresses the time series variable Y in terms of the
components T, C, S, and I. A model that treats the time series values as a sum of the
components is called an additive components model. A model that treats the time series
values as the product of the components is called a multiplicative components model.
Both models are sometimes referred to as unobserved components models, since, in
practice, although we observe the values of the time series, the values of the components
are not observed. The approach to time series analysis described in this chapter involves
an attempt, given the observed series, to estimate the values of the components. These
estimates can then be used to forecast or to display the series unencumbered by the sea-
sonal fluctuations. The latter process is called seasonal adjustment.
It is difficult to deal with the cyclical component of a time series. To the extent that
cycles can be determined from historical data, both their lengths (measured in years)
and their magnitudes (differences between highs and lows) are far from constant. This
lack of a consistent wavelike pattern makes distinguishing cycles from smoothly evolv-
ing trends difficult. Consequently, to keep things relatively simple, we will assume any
cycle in the data is part of the trend. Initially, then, we consider only three components,
T, S, and I. A brief discussion of one way to handle cyclical fluctuations in the decom-
position approach to time series analysis is available in the Cyclical and Irregular
Variations section of this chapter (see p. 180).
The two simplest models relating the observed value (Yt) of a time series to the
trend (Yt ), seasonal (St), and irregular (It) components are the additive components
model

Yt = Tt + St + It (5.1)

and the multiplicative components model

Yt = Tt * St * It (5.2)

The additive components model works best when the time series being analyzed
has roughly the same variability throughout the length of the series. That is, all the val-
ues of the series fall essentially within a band of constant width centered on the trend.
The multiplicative components model works best when the variability of the time
series increases with the level.1 That is, the values of the series spread out as the trend
increases, and the set of observations has the appearance of a megaphone or funnel. A
time series with constant variability and a time series with variability increasing with

1It is possible to convert a multiplicative decomposition to an additive decomposition by working

with the logarithms of the data. Using Equation 5.2 and the properties of logarithms, we have
log Y = log (T * S * I ) = logT + logS + logI. Decomposition of logged data is explored in Problem 5.15.
168 CHAPTER 5 Time Series and Their Components

1,000

900

Milk Production
800

700

600

50 100 150
Month

900

800

700

600
Monthly Sales

500

400

300

200

100

0
10 20 30 40 50 60 70
Month

FIGURE 5-1 Time Series with Constant Variability (Top) and a Time
Series with Variability Increasing with Level (Bottom)

level are shown in Figure 5-1. Both of these monthly series have an increasing trend
and a clearly defined seasonal pattern.2

Trend
Trends are long-term movements in a time series that can sometimes be described by a
straight line or a smooth curve. Examples of the basic forces producing or affecting the
trend of a series are population change, price change, technological change, productiv-
ity increases, and product life cycles.

2Variants of the decomposition models (see Equations 5.1 and 5.2) exist that contain both multiplicative and
additive terms. For example, some software packages do “multiplicative” decomposition using the model
Y = T * S + I.
 CHAPTER 5 Time Series and Their Components 169
A population increase may cause retail sales of a community to rise each year for
several years. Moreover, the sales in current dollars may have been pushed upward
during the same period because of general increases in the prices of retail goods—even
though the physical volume of goods sold did not change.
Technological change may cause a time series to move upward or downward. The
development of high-speed computer chips, enhanced memory devices, and improved
display panels, accompanied by improvements in telecommunications technology, has
resulted in dramatic increases in the use of personal computers and cellular
telephones. Of course, the same technological developments have led to a downward
trend in the production of mechanical calculators and rotary telephones.
Productivity increases—which, in turn, may be due to technological change—give
an upward slope to many time series. Any measure of total output, such as manufactur-
ers’ sales, is affected by changes in productivity.
For business and economic time series, it is best to view the trend (or trend-cycle)
as smoothly changing over time. Rarely can we realistically assume that the trend can
be represented by some simple function such as a straight line over the whole period
for which the time series is observed. However, it is often convenient to fit a trend
curve to a time series for two reasons: (1) It provides some indication of the general
direction of the observed series, and (2) it can be removed from the original series to
get a clearer picture of the seasonality.
If the trend appears to be roughly linear—that is, if it increases or decreases like a
straight line—then it is represented by the equation
TNt = b0 + b1t (5.3)

Here, TNt is the predicted value for the trend at time t. The symbol t represents time, the
independent variable, and ordinarily assumes integer values 1, 2, 3, . . . corresponding
to consecutive time periods. The slope coefficient, b1, is the average increase or
decrease in T for each one-period increase in time.
Time trend equations, including the straight-line trend, can be fit to the data using
the method of least squares. Recall from Chapter 2 that the method of least squares
selects the values of the coefficients in the trend equation (b0 and b1 in the straight-line
case) so that the estimated trend values (TNt) are close to the actual values (Yt ) as mea-
sured by the sum of squared errors criterion

SSE = ©1Yt - TNt22 (5.4)

Example 5.1
Data on annual registrations of new passenger cars in the United States from 1960 to 1992
are shown in Table 5-1 and plotted in Figure 5-2. The values from 1960 to 1992 are used to
develop the trend equation. Registrations is the dependent variable, and the independent
variable is time t, coded as 1960 = 1, 1961 = 2, and so on.
The fitted trend line has the equation
TNt = 7.988 + .0687t
The slope of the trend equation indicates that registrations are estimated to increase an
average of 68,700 each year. Figure 5-3 shows the straight-line trend fitted to the actual data.
Figure 5-3 also shows forecasts of new car registrations for the years 1993 and 1994 (t = 34
and t = 35), which were obtained by extrapolating the trend line. We will say more about
forecasting trend shortly.
The estimated trend values for passenger car registrations from 1960 to 1992 are
shown in Table 5-1 under TN . For example, the trend equation estimates registrations in 1992
(t = 33) to be

TN33 = 7.988 + .06871332 = 10.255

170 CHAPTER 5 Time Series and Their Components

TABLE 5-1 Registration of New Passenger

Cars in the United States,
1960–1992, for Example 5.1

Registrations Trend Estimates Error

(millions) (millions) (millions)
Year Y Time t TN Y - TN

1960 6.577 1 8.0568 -1.4798

1961 5.855 2 8.1255 -2.2705
1962 6.939 3 8.1942 -1.2552
1963 7.557 4 8.2629 -0.7059
1964 8.065 5 8.3316 -0.2666
1965 9.314 6 8.4003 0.9138
1966 9.009 7 8.4690 0.5401
1967 8.357 8 8.5376 -0.1807
1968 9.404 9 8.6063 0.7977
1969 9.447 10 8.6750 0.7720
1970 8.388 11 8.7437 -0.3557
1971 9.831 12 8.8124 1.0186
1972 10.409 13 8.8811 1.5279
1973 11.351 14 8.9498 2.4012
1974 8.701 15 9.0185 -0.3175
1975 8.168 16 9.0872 -0.9192
1976 9.752 17 9.1559 0.5961
1977 10.826 18 9.2246 1.6014
1978 10.946 19 9.2933 1.6527
1979 10.357 20 9.3620 0.9950
1980 8.761 21 9.4307 -0.6697
1981 8.444 22 9.4994 -1.0554
1982 7.754 23 9.5681 -1.8141
1983 8.924 24 9.6368 -0.7128
1984 10.118 25 9.7055 0.4125
1985 10.889 26 9.7742 1.1148
1986 11.140 27 9.8429 1.2971
1987 10.183 28 9.9116 0.2714
1988 10.398 29 9.9803 0.4177
1989 9.833 30 10.0490 -0.2160
1990 9.160 31 10.1177 -0.9577
1991 9.234 32 10.1863 -0.9524
1992 8.054 33 10.2550 -2.2010

Source: Data from U.S. Department of Commerce, Survey of Current

Business (various years).

or 10,255,000 registrations. However, registrations of new passenger cars were actually

8,054,000 in 1992. Thus, for 1992, the trend equation overestimates registrations by approxi-
mately 2.2 million. This error and the remaining estimation errors are listed in Table 5-1
under Y - TN . These estimation errors were used to compute the measures of fit, the MAD,
MSD, and MAPE shown in Figure 5-3. These are the same measures of forecast accuracy that
are given by Equations 3.7, 3.8, and 3.10, respectively, in Chapter 3. (Minitab commands used
to produce the results in Example 5.1 are given in the Minitab Applications section at the end
of this chapter.)
 CHAPTER 5 Time Series and Their Components 171

FIGURE 5-2 Car Registrations Time Series for Example 5.1

FIGURE 5-3 Trend Line for the Car Registrations Time Series for Example 5.1

Additional Trend Curves

The life cycle of a new product has three stages: introduction, growth, and maturity and
saturation. A curve representing sales (in dollars or units) over a new product’s life
cycle is shown in Figure 5-4. Time, shown on the horizontal axis, can vary from days to
years, depending on the nature of the market. A straight-line trend would not work for
these data. Linear models assume that a variable is increasing (or decreasing) by a con-
stant amount each time period. The increases per time period in the product’s life cycle
172 CHAPTER 5 Time Series and Their Components

Trend

Sales, Units

Introduction Growth Maturity and Saturation

Time

FIGURE 5-4 Life Cycle of a Typical New Product

curve are quite different depending on the stage of the cycle. A curve, other than a
straight line, is needed to model the trend over a new product’s life cycle.
A simple function that allows for curvature is the quadratic trend
TNt = b0 + b1t + b2t2 (5.5)

As an illustration, Figure 5-5 shows a quadratic trend curve fit to the passenger car reg-
istrations data of Example 5.1 using the SSE criterion. The quadratic trend can be pro-
jected beyond the data for, say, two additional years, 1993 and 1994. We will consider
the implications of this projection in the next section, Forecasting Trend.

FIGURE 5-5 Quadratic Trend Curve for the Car Registrations

Time Series for Example 5.1
 CHAPTER 5 Time Series and Their Components 173
Based on the MAPE, MAD, and MSD accuracy measures, a quadratic trend appears
to be a better representation of the general direction of the car registrations series than
the linear trend in Figure 5-3. Which trend model is appropriate? Before considering this
issue, we will introduce a few additional trend curves that have proved useful.
When a time series starts slowly and then appears to be increasing at an increasing
rate (see Figure 5-4) such that the percentage difference from observation to observa-
tion is constant, an exponential trend can be fitted. The exponential trend is given by

TNt = b0bt1 (5.6)

The coefficient b1 is related to the growth rate. If the exponential trend is fit to annual
data, the annual growth rate is estimated to be 1001b1 - 12%.
Figure 5-6 indicates the number of mutual fund salespeople employed by a partic-
ular company for several consecutive years. The increase in the number of salespeople
is not constant. It appears as if increasingly larger numbers of people are being added
in the later years.
An exponential trend curve fit to the salespeople data has the equation

TNt = 10.01611.3132t

implying an annual growth rate of about 31%. Consequently, if the model estimates
51 salespeople for 2005, the increase for 2006 would be 16 151 * .312 for an estimated
total of 67. This can be compared to the actual value of 68 salespeople.
A linear trend fit to the salespeople data would indicate a constant average
increase of about nine salespeople per year. This trend overestimates the actual
increase in the earlier years and underestimates the increase in the last year. It does not
model the apparent trend in the data as well as the exponential curve does.
It is clear that extrapolating an exponential trend with a 31% growth rate will quickly
result in some very big numbers. This is a potential problem with an exponential trend
model. What happens when the economy cools off and stock prices begin to retreat? The

FIGURE 5-6 Graph of Mutual Fund Salespeople

174 CHAPTER 5 Time Series and Their Components

demand for mutual fund salespeople will decrease, and the number of salespeople could
even decline. The trend forecast by the exponential curve will be much too high.
Growth curves of the Gompertz or logistic type represent the tendency of many
industries and product lines to grow at a declining rate as they mature. If the plotted
data reflect a situation in which sales begin low, then increase as the product catches on,
and finally ease off as saturation is reached, the Gompertz curve or Pearl–Reed logistic
model might be appropriate. Figure 5-7 shows a comparison of the general shapes of
(a) the Gompertz curve and (b) the Pearl–Reed logistic model. Note that the logistic
curve is very similar to the Gompertz curve but has a slightly gentler slope. Figure 5-7
shows how the Y intercepts and maximum values for these curves are related to some of
the coefficients in their functional forms. The formulas for these trend curves are com-
plex and not within the scope of this text. Many statistical software packages, including
Minitab, allow one to fit several of the trend models discussed in this section.
Although there are some objective criteria for selecting an appropriate trend, in
general, the correct choice is a matter of judgment and therefore requires experience
and common sense on the part of the analyst. As we will discuss in the next section, the
line or curve that best fits a set of data points might not make sense when projected as
the trend of the future.

Forecasting Trend
Suppose we are presently at time t = n (end of series) and we want to use a trend model
to forecast the value of Y, p steps ahead. The time period at which we make the forecast,
n in this case, is called the forecast origin. The value p is called the lead time. For the lin-
ear trend model, we can produce a forecast by evaluating TNn + p = b0 + b11n + p2.
Using the trend line fitted to the car registration data in Example 5.1, a forecast of the
trend for 1993 (t = 34) made in 1992 (t = n = 33) would be the p = 1 step ahead forecast

TN33 + 1 = 7.988 + .0687133 + 12 = 7.988 + .06871342 = 10.324

Similarly, the p = 2 step ahead (1994) forecast is given by

TN33 + 2 = 7.988 + .0687133 + 22 = 7.988 + .06871352 = 10.393

These two forecasts are shown in Figure 5-3 as extrapolations of the fitted trend line.

^ ^
Tt Tt

b0 1
b0

1
b 0b 1 b0 + b1
t t
0 0
(a) Gompertz Trend Curve (b) Logistic (Pearl-Reed) Trend Curve

FIGURE 5-7 S-Shaped Growth Curves

 CHAPTER 5 Time Series and Their Components 175
Figure 5-5 shows the fitted quadratic trend curve for the car registration data.
Using the equation shown in the figure, we can calculate forecasts of the trend for 1993
and 1994 by setting t = 33 + 1 = 34 and t = 33 + 2 = 35. The reader can verify that
TN33 + 1 = 8.688 and TN33 + 2 = 8.468. These numbers were plotted in Figure 5-5 as extrapo-
lations of the quadratic trend curve.
Recalling that car registrations are measured in millions, the two forecasts of trend
produced from the quadratic curve are quite different from the forecasts produced by
the linear trend equation. Moreover, they are headed in the opposite direction. If we
were to extrapolate the linear and quadratic trends for additional time periods, their
differences would be magnified.
The car registration example illustrates why great care must be exercised in using
fitted trend curves for the purpose of forecasting future trends. Two equations, both of
which may reasonably represent the observed time series, can give very different
results when projected over future time periods. These differences can be substantial
for large lead times (long-run forecasting).
Trend curve models are based on the following assumptions:
1. The correct trend curve has been selected.
2. The curve that fits the past is indicative of the future.
These assumptions suggest that judgment and expertise play a substantial role in the
selection and use of a trend curve. To use a trend curve for forecasting, we must be able
to argue that the correct trend has been selected and that, in all likelihood, the future
will be like the past.
There are objective criteria for selecting a trend curve. We will discuss two of these
criteria, the Akaike information criterion (AIC) and the Bayesian information crite-
rion (BIC), in Chapter 9. However, although these and other criteria help to determine
an appropriate model, they do not replace good judgment.

Seasonality
A seasonal pattern is one that repeats itself year after year. For annual data, seasonal-
ity is not an issue because there is no chance to model a within-year pattern with data
recorded once per year. Time series consisting of weekly, monthly, or quarterly obser-
vations, however, often exhibit seasonality.
The analysis of the seasonal component of a time series has immediate short-term
implications and is of greatest importance to mid- and lower-level management.
Marketing plans, for example, have to take into consideration expected seasonal pat-
terns in consumer purchases.
Several methods for measuring seasonal variation have been developed. The basic
idea in all of these methods is to first estimate and remove the trend from the original
series and then smooth out the irregular component. Keeping in mind our decomposi-
tion models, this leaves data containing only seasonal variation. The seasonal values
are then collected and summarized to produce a number (generally an index number)
for each observed interval of the year (week, month, quarter, and so on).
Thus, the identification of the seasonal component in a time series differs from
trend analysis in at least two ways:
1. The trend is determined directly from the original data, but the seasonal compo-
nent is determined indirectly after eliminating the other components from the
data, so that only the seasonality remains.
2. The trend is represented by one best-fitting curve, or equation, but a separate sea-
sonal value has to be computed for each observed interval (week, month, quarter)
of the year and is often in the form of an index number.
176 CHAPTER 5 Time Series and Their Components

If an additive decomposition is employed, estimates of the trend, seasonal, and

irregular components are added together to produce the original series. If a multiplica-
tive decomposition is used, the individual components must be multiplied together to
reconstruct the original series, and in this formulation, the seasonal component is rep-
resented by a collection of index numbers. These numbers show which periods within
the year are relatively low and which periods are relatively high. The seasonal indexes
trace out the seasonal pattern.

Index numbers are percentages that show changes over time.

With monthly data, for example, a seasonal index of 1.0 for a particular month means
the expected value for that month is 1/12 the total for the year. An index of 1.25 for a dif-
ferent month implies the observation for that month is expected to be 25% more than
1/12 of the annual total.A monthly index of 0.80 indicates that the expected level of activ-
ity that month is 20% less than 1/12 of the total for the year, and so on.The index numbers
indicate the expected ups and downs in levels of activity over the course of a year after the
effects due to the trend (or trend-cycle) and irregular components have been removed.
To highlight seasonality, we must first estimate and remove the trend. The trend
can be estimated with one of the trend curves we discussed previously, or it can be esti-
mated using a moving average, as discussed in Chapter 4.
Assuming a multiplicative decomposition model, the ratio to moving average is a
popular method for measuring seasonal variation. In this method, the trend is esti-
mated using a centered moving average. We illustrate the ratio-to-moving-average
method using the monthly sales of the Cavanaugh Company, shown in Figure 5.1
(Bottom) in the next example.

Example 5.2
To illustrate the ratio-to-moving-average method, we use two years of the monthly sales of
the Cavanaugh Company.3 Table 5-2 gives the monthly sales from January 2004 to
December 2005 to illustrate the beginning of the computations. The first step for monthly
data is to compute a 12-month moving average (for quarterly data, a four-month moving
average would be computed). Because all of the months of the year are included in the cal-
culation of this moving average, effects due to the seasonal component are removed, and
the moving average itself contains only the trend and the irregular components.
The steps (identified in Table 5-2) for computing seasonal indexes by the ratio-to-moving-
average method follow:
Step 1. Starting at the beginning of the series, compute the 12-month moving total, and
place the total for January 2004 through December 2004 between June and July
2004.
Step 2. Compute a two-year moving total so that the subsequent averages are centered on
July rather than between months.
Step 3. Since the two-year total contains the data for 24 months (January 2004 once,
February 2004 to December 2004 twice, and January 2005 once), this total is cen-
tered on July 2004.
Step 4. Divide the two-year moving total by 24 in order to obtain the 12-month centered
moving average.
Step 5. The seasonal index for July is calculated by dividing the actual value for July by the
12-month centered moving average.4

3The units have been omitted and the dates and name have been changed to protect the identity of the company.
4This is the ratio-to-moving-average operation that gives the procedure its name.
 CHAPTER 5 Time Series and Their Components 177

TABLE 5-2 Sales of the Cavanaugh Company,

2004–2005, for Example 5.2

12-Month
12-Month Two-Year Centered
Moving Moving Moving Seasonal
Period Sales Total Total Average Index

2004

4
January 518
February 404
March 300
April 210
May 196
June

4
186 1 4,869 2
July 247 4,964 9,833} 3 409.7} 4 0.60} 5
August 343 4,952 9,916 413.2 0.83
September 464 4,925 9,877 411.5 1.13
October 680 5,037 9,962 415.1 1.64
November 711 5,030 10,067 419.5 1.69
December 610 10,131 422.1 1.45
2005
January 613 5,101 10,279 428.3 1.43
February 392 5,178 10,417 434.0 0.90
March 273 5,239 10,691 445.5 0.61
April 322 5,452 11,082 461.8 0.70
May 189 5,630 11,444 476.8 0.40
June 257 5,814 11,682 486.8 0.53
July 324 5,868
August 404
September 677
October 858
November 895
December 664

Repeat steps 1 to 5 beginning with the second month of the series, August 2004, and so on.
The process ends when a full 12-month moving total can no longer be calculated.
Because there are several estimates (corresponding to different years) of the seasonal
index for each month, they must be summarized to produce a single number. The median,
rather than the mean, is used as the summary measure. Using the median eliminates the
influence of data for a month in a particular year that are unusually large or small. A sum-
mary of the seasonal ratios, along with the median value for each month, is contained in
Table 5-3.
The monthly seasonal indexes for each year must sum to 12, so the median for each
month must be adjusted to get the final set of seasonal indexes.5 Since this multiplier should
be greater than 1 if the total of the median ratios before adjustment is less than 12 and
smaller than 1 if the total is greater than 12, the multiplier is defined as

12
Multiplier =
Actual total

5The monthly indexes must sum to 12 so that the expected annual total equals the actual annual total.
178 CHAPTER 5 Time Series and Their Components

TABLE 5-3 A Summary of the Monthly Seasonal Indexes for the Cavanaugh
Company for Example 5.2

Adjusted
Seasonal
Index (Median
Month 2000 2001 2002 2003 2004 2005 2006 Median  1.0044)
January — 1.208 1.202 1.272 1.411 1.431 — 1.272 1.278
February — 0.700 0.559 0.938 1.089 0.903 — 0.903 0.907
March — 0.524 0.564 0.785 0.800 0.613 — 0.613 0.616
April — 0.444 0.433 0.480 0.552 0.697 — 0.480 0.482
May — 0.424 0.365 0.488 0.503 0.396 — 0.424 0.426
June — 0.490 0.459 0.461 0.465 0.528 0.465 0.467
July 0.639 0.904 0.598 0.681 0.603 0.662 0.651 0.654
August 1.115 0.913 0.889 0.799 0.830 0.830 0.860 0.864
September 1.371 1.560 1.346 1.272 1.128 1.395 1.359 1.365
October 1.792 1.863 1.796 1.574 1.638 1.771 1.782 1.790
November 1.884 2.012 1.867 1.697 1.695 1.846 1.857 1.865
December 1.519 1.088 1.224 1.282 1.445 — 1.282 1.288
11.948 12.002

Using the information in Table 5-3,

12
Multiplier = = 1.0044
11.948
The final column in Table 5-3 contains the final seasonal index for each month, determined
by making the adjustment (multiplying by 1.0044) to each of the median ratios.6 The final
seasonal indexes, shown in Figure 5-8, represent the seasonal component in a multiplicative
decomposition of the sales of the Cavanaugh Company time series. The seasonality in sales

Seasonal Indices
1.9

1.4
Index

0.9

0.4

1 2 3 4 5 6 7 8 9 10 11 12
Month

FIGURE 5-8 Seasonal Indexes for the Cavanaugh Company

for Example 5.2

6The seasonal indexes are sometimes multiplied by 100 and expressed as percentages.
 CHAPTER 5 Time Series and Their Components 179
is evident from Figure 5-8. Sales for this company are periodic, with relatively low sales in
the late spring and relatively high sales in the late fall.

Our analysis of the sales series in Example 5.2 assumed that the seasonal pattern
remained constant from year to year. If the seasonal pattern appears to change over
time, then estimating the seasonal component with the entire data set can produce mis-
leading results. It is better, in this case, either (1) to use only recent data (from the last
few years) to estimate the seasonal component or (2) to use a time series model that
allows for evolving seasonality. We will discuss models that allow for evolving seasonal-
ity in a later chapter.
The seasonal analysis illustrated in Example 5.2 is appropriate for a multiplicative
decomposition model. However, the general approach outlined in steps 1 to 5 works
for an additive decomposition if, in step 5, the seasonality is estimated by subtracting
the trend from the original series rather than dividing by the trend (moving average) to
get an index. In an additive decomposition, the seasonal component is expressed in the
same units as the original series.
In addition, it is apparent from our sales example that using a centered moving
average to determine trend results in some missing values at the ends of the series. This
is particularly problematic if forecasting is the objective. To forecast future values using
a decomposition approach, alternative methods for estimating the trend must be used.
The results of a seasonal analysis can be used to (1) eliminate the seasonality in
data; (2) forecast future values; (3) evaluate current positions in, for example, sales,
inventory, and shipments; and (4) schedule production.

Seasonally Adjusted Data

Once the seasonal component has been isolated, it can be used to calculate seasonally
adjusted data. For an additive decomposition, the seasonally adjusted data are com-
puted by subtracting the seasonal component:

Yt - St = Tt + It (5.7a)

For a multiplicative decomposition, the seasonally adjusted data are computed by

dividing the original observations by the seasonal component:

Yt
= Tt * It (5.7b)
St
Most economic series published by government agencies are seasonally adjusted
because seasonal variation is not of primary interest. Rather, it is the general pattern of
economic activity, independent of the normal seasonal fluctuations, that is of interest.
For example, new car registrations might increase by 10% from May to June, but is this
increase an indication that new car sales are completing a banner quarter? The answer
is “no” if the 10% increase is typical at this time of year largely due to seasonal factors.
In a survey concerned with the acquisition of seasonally adjusted data, Bell and
Hillmer (1984) found that a wide variety of users value seasonal adjustment. They
identified three motives for seasonal adjustment:

1. Seasonal adjustment allows reliable comparison of values at different points in

time.
2. It is easier to understand the relationships among economic or business variables
once the complicating factor of seasonality has been removed from the data.
3. Seasonal adjustment may be a useful element in the production of short-term fore-
casts of future values of a time series.
180 CHAPTER 5 Time Series and Their Components

Bell and Hillmer concluded that “seasonal adjustment is done to simplify data so that
they may be more easily interpreted by statistically unsophisticated users without a sig-
nificant loss of information” (p. 301).

Cyclical and Irregular Variations

Cycles are long-run, wavelike fluctuations that occur most frequently in macro indica-
tors of economic activity. As we have discussed, to the extent that they can be mea-
sured, cycles do not have a consistent pattern. However, some insight into the cyclical
behavior of a time series can be obtained by eliminating the trend and seasonal compo-
nents to give, using a multiplicative decomposition,7

Yt Tt * Ct * St * It
= = Ct * It (5.8)
Tt * St Tt * St

A moving average can be used to smooth out the irregularities, It, leaving the cyclical
component, Ct. To eliminate the centering problem encountered when a moving aver-
age with an even number of time periods is used, the irregularities are smoothed using a
moving average with an odd number of time periods. For monthly data, a 5-, a 7-, a
9-, or even an 11-period moving average will work. For quarterly data, an estimate of C
can be computed using a three-period moving average of the values.8
Finally, the irregular component is estimated by

Ct * It
It = (5.9)
Ct

The irregular component represents the variability in the time series after the other
components have been removed. It is sometimes called the residual or error. With a
multiplicative decomposition, both the cyclical and the irregular components are
expressed as indexes.

Summary Example
One reason for decomposing a time series is to isolate and examine the components of
the series. After the analyst is able to look at the trend, seasonal, cyclical, and irregular
components of a series one at a time, insights into the patterns in the original data val-
ues may be gained. Also, once the components have been isolated, they can be recom-
bined or synthesized to produce forecasts of future values of the time series.
Example 5.3
In Example 3.5, Perkin Kendell, the analyst for the Coastal Marine Corporation, used auto-
correlation analysis to determine that sales were seasonal on a quarterly basis. Now he uses
decomposition to understand the quarterly sales variable. Perkin uses Minitab (see the
Minitab Applications section at the end of the chapter) to produce Table 5-4 and Figure 5-9.
In order to keep the seasonal pattern current, only the last seven years (2000 to 2006) of
sales data, Y, were analyzed.
The original data are shown in the chart in the upper left of Figure 5-10. The trend is
computed using the linear model: TNt = 261.24 + .759t. Since 1 represented the first quarter
of 2000, Table 5-4 shows the trend value equal to 262.000 for this time period, and estimated
sales (the T column) increased by .759 each quarter.

7Notice that we have added the cyclical component, C, to the multiplicative decomposition shown in
Equation 5.2.
8For annual data, there is no seasonal component, and the cyclical * irregular component is obtained by
simply removing the trend from the original series.
 CHAPTER 5 Time Series and Their Components 181

TABLE 5-4 The Multiplicative Decomposition for Coastal Marine

Sales for Example 5.3

t Year Quarter Sales T SCI S TCI CI C I

1 2000 1 232.7 262.000 0.888 0.780 298.458 1.139 — —

2 2 309.2 262.759 1.177 1.016 304.434 1.159 1.118 1.036
3 3 310.7 263.518 1.179 1.117 278.239 1.056 1.078 0.980
4 4 293.0 264.276 1.109 1.088 269.300 1.019 1.022 0.997
5 2001 1 205.1 265.035 0.774 0.780 263.059 0.993 0.960 1.034
6 2 234.4 265.794 0.882 1.016 230.787 0.868 0.940 0.924
7 3 285.4 266.553 1.071 1.117 255.582 0.959 0.906 1.059
8 4 258.7 267.312 0.968 1.088 237.775 0.890 0.924 0.962
9 2002 1 193.2 268.071 0.721 0.780 247.796 0.924 0.927 0.998
10 2 263.7 268.830 0.981 1.016 259.635 0.966 0.954 1.012
11 3 292.5 269.589 1.085 1.117 261.940 0.972 1.003 0.969
12 4 315.2 270.348 1.166 1.088 289.705 1.072 0.962 1.114
13 2003 1 178.3 271.107 0.658 0.780 228.686 0.844 0.970 0.870
14 2 274.5 271.866 1.010 1.016 270.269 0.994 0.936 1.062
15 3 295.4 272.625 1.084 1.117 264.537 0.970 0.976 0.994
16 4 286.4 273.383 1.048 1.088 263.234 0.963 0.942 1.022
17 2004 1 190.8 274.142 0.696 0.780 244.718 0.893 0.933 0.957
18 2 263.5 274.901 0.959 1.016 259.438 0.944 0.957 0.986
19 3 318.8 275.660 1.157 1.117 285.492 1.036 0.998 1.038
20 4 305.3 276.419 1.104 1.088 280.605 1.015 1.058 0.960
21 2005 1 242.6 277.178 0.875 0.780 311.156 1.123 1.089 1.031
22 2 318.8 277.937 1.147 1.016 313.886 1.129 1.104 1.023
23 3 329.6 278.696 1.183 1.117 295.164 1.059 1.100 0.963
24 4 338.2 279.455 1.210 1.088 310.844 1.112 1.078 1.032
25 2006 1 232.1 280.214 0.828 0.780 297.689 1.062 1.059 1.004
26 2 285.6 280.973 1.016 1.016 281.198 1.001 0.996 1.005
27 3 291.0 281.732 1.033 1.117 260.597 0.925 0.947 0.977
28 4 281.4 282.491 0.996 1.088 258.639 0.916 — —

The chart in the upper right of Figure 5-10 shows the detrended data.These data are also
shown in the SCI column of Table 5-4. The detrended value for the first quarter of 2000 was9

Y 232.7
SCI = = = .888
T 262.000

The seasonally adjusted data are shown in the TCI column in Table 5-4 and in Figure
5-10. The seasonally adjusted value for the first quarter of 2000 was

232.7
TCI = = 298.458
.77967

Sales in the first quarter of 2005 were 242.6. However, examination of the seasonally
adjusted column shows that the sales for this quarter were actually high when the data were
adjusted for the fact that the first quarter is typically a very weak quarter.

9To simplify the notation in this example, we omit the subscript t on the original data, Y, and each of its com-
ponents, T, S, C, and I. We also omit the multiplication sign, , between components, since it is clear we are
considering a multiplicative decomposition.
 FIGURE 5-9 Minitab Output for Decomposition of Coastal Marine
Quarterly Sales for Example 5.3

FIGURE 5-10 Component Analysis for Coastal Marine Sales for Example 5.3

182
 CHAPTER 5 Time Series and Their Components 183
The seasonal indexes in Figure 5-9 were

First quarter = .77967 : 78.0%

Second quarter = 1.01566 : 101.6%
Third quarter = 1.11667 : 111.7%
Fourth quarter = 1.08800 : 108.8%

The chart on the left of Figure 5-11 shows the seasonal components relative to 1.0. We see
that first-quarter sales are 22% below average, second-quarter sales are about as expected,
third-quarter sales are almost 12% above average, and fourth-quarter sales are almost 9%
above normal.
The cyclical-irregular value for first quarter of 2000 was10

Y 232.7
CI = = = 1.139
TS 1262.00021.779672

In order to calculate the cyclical column, a three-period moving average was computed. The
value for the second quarter of 2000 was

1.139
1.159
1.056
3.354 3.354/3 = 1.118

Notice how smooth the C column is compared to the CI column. The reason is that using the
moving average has smoothed out the irregularities.
Finally, the I column was computed. For example, for the second quarter of 2000,

CI 1.159
I = = = 1.036
C 1.118

Examination of the I column shows that there were some large changes in the irregular
component. The irregular index dropped from 111.4% in the fourth quarter of 2002 to 87%

FIGURE 5-11 Seasonal Analysis for Coastal Marine Sales for Example 5.3

10Minitab calculates the cyclical * irregular component (or simply the irregular component if no cyclical
component is contemplated) by subtracting the trend * seasonal component from the original data. In sym-
bols, Minitab sets CI = Y - TS. The Minitab CI component is shown in the lower right-hand chart of Figure
5-10. Moreover, Minitab fits the trend line to the seasonally adjusted data. That is, the seasonal adjustment is
done before the trend is determined.
184 CHAPTER 5 Time Series and Their Components

in the first quarter of 2003 and then increased to 106.2% in the second quarter of 2003. This
behavior results from the unusually low sales in the first quarter of 2003.

Business Indicators
The cyclical indexes can be used to answer the following questions:
1. Does the series cycle?
2. If so, how extreme is the cycle?
3. Does the series follow the general state of the economy (business cycle)?
One way to investigate cyclical patterns is through the study of business indicators.
A business indicator is a business-related time series that is used to help assess the gen-
eral state of the economy, particularly with reference to the business cycle. Many busi-
nesspeople and economists systematically follow the movements of such statistical
series to obtain economic and business information in the form of an unfolding picture
that is up to date, comprehensive, relatively objective, and capable of being read and
understood with a minimum expenditure of time.

Business indicators are business-related time series that are used to help assess
the general state of the economy.

The most important list of statistical indicators originated during the sharp busi-
ness setback from 1937 to 1938. Secretary of the Treasury Henry Morgenthau
requested the National Bureau of Economic Research (NBER) to devise a system that
would signal when the setback was nearing an end. Under the leadership of Wesley
Mitchell and Arthur F. Burns, NBER economists selected 21 series that from past per-
formance promised to be fairly reliable indicators of business revival. Since then, the
business indicator list has been revised several times. The current list consists of 21
indicators—10 classified as leading, 4 as coincident, and 7 as lagging.
1. Leading indicators. In practice, components of the leading series are studied to help
anticipate turning points in the economy. The Survey of Current Business publishes
this list each month, along with the actual values of each series for the past several
months and the most recent year. Also, a composite index of leading indicators is
computed for each month and year; the most recent monthly value is frequently
reported in the popular press to indicate the general direction of the future economy.
Examples of leading indicators are building permits and an index of stock prices.
2. Coincident indicators. The four coincident indicators provide a measure of how the
U.S. economy is currently performing. An index of these four series is computed
each month. Examples of coincident indicators are industrial production and man-
ufacturing and trade sales.
3. Lagging indicators. The lagging indicators tend to lag behind the general state of
the economy, both on the upswing and on the downswing. A composite index is
also computed for this list. Examples of lagging indicators are the prime interest
rate and the unemployment rate.
Cycles imply turning points. That is, turning points come into existence only as a
consequence of a subsequent decline or gain in the business cycle. Leading indicators
change direction ahead of turns in general business activity, coincident indicators turn at
about the same time as the general economy, and turns in the lagging indicators follow
 CHAPTER 5 Time Series and Their Components 185
those of the general economy. However, it is difficult to identify cyclical turning points
at the time they occur, since all areas of the economy do not expand at the same time
during periods of expansion or contract at the same time during periods of contraction.
Hence, several months may go by before a genuine cyclical upturn or downturn is finally
identified with any assurance.
Leading indicators are the most useful predictive tool, since they attempt to signal
economic changes in advance. Coincident and lagging indicators are of minimal inter-
est from a forecasting perspective, but they are used to assess the effectiveness of cur-
rent and past economic policy and so to help formulate future policy.
In their article entitled “Early Warning Signals for the Economy,” Geoffrey H.
Moore and Julius Shiskin (1976) have the following to say on the usefulness of business
cycle indicators:

It seems clear from the record that business cycle indicators are helpful in
judging the tone of current business and short-term prospects. But because of
their limitations, the indicators must be used together with other data and with
full awareness of the background of business and consumer confidence and
expectations, governmental policies, and international events. We also must
anticipate that the indicators will often be difficult to interpret, that interpreta-
tions will sometimes vary among analysts, and that the signals they give will
not be correctly interpreted. Indicators provide a sensitive and revealing pic-
ture of the ebb and flow of economic tides that a skillful analyst of the eco-
nomic, political, and international scene can use to improve his chances of
making a valid forecast of short-run economic trends. If the analyst is aware of
their limitations and alert to the world around him, he will find the indicators
useful guideposts for taking stock of the economy and its needs. (p. 81)

Cyclical components of individual time series generally conform only loosely—

and sometimes not at all—to the business cycle as identified by the NBER indicators.
However, if a cyclical component for a given time series is estimated, it should always
be plotted over time to get some indication of the magnitudes and lengths of any cycles
that appear to exist. In addition, the plot can be examined for any relation to the ups
and downs of general economic activity.
The discussion so far shows how the factors that create variation in a time series
can be separated and studied individually. Analysis is the process for taking the time
series apart; synthesis is the process for putting it back together. Next, we shall put the
components of the time series back together to do forecasting.

FORECASTING A SEASONAL TIME SERIES

In forecasting a seasonal time series, the decomposition process is reversed. Instead of
separating the series into individual components for examination, we recombine the com-
ponents to develop the forecasts for future periods. We will use the multiplicative model
and the results of Example 5.3 to develop forecasts for Coastal Marine Corporation sales.
Example 5.4
Forecasts of Coastal Marine Corporation sales for the four quarters of 2007 can be devel-
oped using Table 5-4.

1. Trend. The quarterly trend equation is TNt = 261.24 + .759t. The forecast origin is the
fourth quarter of 2006, or time period t = n = 28. Sales for the first quarter of 2007
occurred in time period t = 28 + 1 = 29. This notation shows we are forecasting p = 1
186 CHAPTER 5 Time Series and Their Components

period ahead from the end of the time series. Setting t = 29, the trend projection is
then

TN29 = 261.24 + .7591292 = 283.251

2. Seasonal. The seasonal index for the first quarter, .77967, is given in Figure 5-9.
3. Cyclical. The cyclical projection must be determined from the estimated cyclical pat-
tern (if any) and from any other information generated by indicators of the general
economy for 2007. Projecting the cyclical pattern for future time periods is fraught with
uncertainty, and as we indicated earlier, it is generally assumed, for forecasting pur-
poses, to be included in the trend. To demonstrate the completion of this example, we
set the cyclical index to 1.0.
4. Irregular. Irregular fluctuations represent random variation that can’t be explained by
the other components. For forecasting, the irregular component is set to the average
value 1.0.11
The forecast for the first quarter of 2007 is

YN29 = T29 * S29 * C29 * I29 = 1283.25121.77967211.0211.02 = 220.842

The forecasts for the rest of 2007 are

Second quarter = 288.455

Third quarter = 317.990
Fourth quarter = 310.654

The multiplicative decomposition fit for Coastal Marine Corporation sales, along with
the forecasts for 2007, is shown in Figure 5-12. We can see from the figure that the fit, con-
structed from the trend and seasonal components, represents the actual data reasonably
well. However, the fit is not good for the last two quarters of 2006, time periods 27 and 28.
The forecasts for 2007 mimic the pattern of the fit.

FIGURE 5-12 Decomposition Fit and Forecasts for Coastal

Marine Sales for Example 5.4

11For forecasts generated from an additive model, the irregular index is set to the average value 0.
 CHAPTER 5 Time Series and Their Components 187
Forecasts produced by an additive or a multiplicative decomposition model reflect
the importance of the individual components. If a variable is highly seasonal, then the
forecasts will have a strong seasonal pattern. If, in addition, there is a trend, the fore-
casts will follow the seasonal pattern superimposed on the extrapolated trend. If one
component dominates the analysis, it alone might provide a practical, accurate short-
term forecast.

THE CENSUS II DECOMPOSITION METHOD

Time series decomposition methods have a long history. In the 1920s and early 1930s,
the Federal Reserve Board and the National Bureau of Economic Research were
heavily involved in the seasonal adjustment and smoothing of economic time series.
However, before the development of computers, the decomposition calculations were
laborious, and practical application of the methods was limited. In the early 1950s,
Julius Shiskin, chief economic statistician at the U.S. Bureau of the Census, developed
a large-scale computer program to decompose time series. The first computer program
essentially approximated the hand methods that were used up to that time and was
replaced a year later by an improved program known as Method II. Over the years,
improved variants of Method II followed. The current variant of the Census Bureau
time series decomposition program is known as X-12-ARIMA. This program is avail-
able from the Census Bureau at no charge and is widely used by government agencies
and private companies.12
Census II decomposition is usually multiplicative, since most economic time series
have seasonal variation that increases with the level of the series. Also, the decomposi-
tion assumes three components: trend-cycle, seasonal, and irregular.
The Census II method iterates through a series of steps until the components are
successfully isolated. Many of the steps involve the application of weighted moving
averages to the data. This results in inevitable loss of data at the beginning and end of
the series because of the averaging. The ARIMA part of X-12-ARIMA provides the
facility to extend the original series in both directions with forecasts so that more of
the observations are adjusted using the full weighted moving averages. These forecasts
are generated from an ARIMA time series model (see Chapter 9).
The steps for each iteration of the Census II method as implemented in X-12-
ARIMA are outlined next. It may seem that the method is complicated because of the
many steps involved. However, the basic idea is quite simple—to isolate the trend-
cycle, seasonal, and irregular components one by one. The various iterations are
designed to improve the estimate of each component. Good references for additional
study are Forecasting: Methods and Applications (Makridakis, Wheelwright, and
Hyndman 1998), Forecasting: Practice and Process for Demand Management
(Levenbach and Cleary 2006), and “New Capabilities and Methods of the X-12-
ARIMA Seasonal-Adjustment Program” (Findley et al. 1998).
Step 1. An s-period moving average is applied to the original data to get a rough
estimate of the trend-cycle. (For monthly data, s = 12; for quarterly data,
s = 4; and so forth.)
Step 2. The ratios of the original data to these moving average values are calculated
as in classical multiplicative decomposition, illustrated in Example 5.2.

12The PC version of the X-12-ARIMA program can be downloaded from the U.S. Census Bureau website.
At the time this book was written, the web address for the download page was www.census.gov/srd/www/
x12a/x12down_pc.html.
188 CHAPTER 5 Time Series and Their Components

Step 3. The ratios from step 2 contain both the seasonal and the irregular compo-
nents. They also include extreme values resulting from unusual events
such as strikes or wars. The ratios are divided by a rough estimate of the
seasonal component to give an estimate of the irregular component.
A large value for an irregular term indicates an extreme value in the orig-
inal data. These extreme values are identified and the ratios in step 2
adjusted accordingly. This effectively eliminates values that do not fit the
pattern of the remaining data. Missing values at the beginning and end of
the series are also replaced by estimates at this stage.
Step 4. The ratios created from the modified data (with extreme values replaced
and estimates for missing values) are smoothed using a moving average to
eliminate irregular variation. This creates a preliminary estimate of the
seasonal component.
Step 5. The original data are then divided by the preliminary seasonal component
from step 4 to get the preliminary seasonally adjusted series. This season-
ally adjusted series contains the trend-cycle and irregular components. In
symbols,

Yt Tt * St * It
= = Tt * It
St St

Step 6. The trend-cycle is estimated by applying a weighted moving average to

the preliminary seasonally adjusted series. This moving average eliminates
irregular variation and gives a smooth curve that indicates the preliminary
trend-cycle in the data.
Step 7. Repeat step 2 with this new estimate of the trend-cycle. That is, new ratios,
containing only the seasonal and irregular components, are obtained by
dividing the original observations by the trend-cycle from step 6. These
are the final seasonal-irregular ratios. Mathematically,
Yt Tt * St * It
= = St * It
Tt Tt

Step 8. Repeat step 3 using the new ratios computed in step 7.

Step 9. Repeat step 4 to get a new estimate of the seasonal component.
Step 10. Repeat step 5 with the seasonal component from step 9.
Step 11. Divide the seasonally adjusted data from step 10 by the trend-cycle
obtained in step 6 to get the estimated irregular component.
Step 12. Extreme values of the irregular component are replaced as in step 3.
A series of modified data is obtained by multiplying the trend-cycle, sea-
sonal component, and adjusted irregular component together. These data
reproduce the original data except for the extreme values.
The preceding 12 steps are repeated, beginning with the modified data from step
12 rather than the original data. Some of the lengths of the moving averages used in the
various steps are changed, depending on the variability in the data.
The final seasonally adjusted series is determined by dividing the final seasonal
component into the original data. The result contains only the product of the trend-
cycle and irregular components.
The value of each of the final components is printed out and plotted. A series of
diagnostic tests is available to determine if the decomposition was successful.
 CHAPTER 5 Time Series and Their Components 189
The X-12-ARIMA program contains many additional features that we have not
described. For example, adjustments can be made for different numbers of trading days
and for holiday effects, missing values within the series can be estimated and replaced,
the effects of outliers can be removed before decomposition, and other changes in
trend such as shifts in level or temporary ramp effects can be modeled.

APPLICATION TO MANAGEMENT

Time series analysis is a widely used statistical tool for forecasting future events that
are intertwined with the economy in some fashion. Manufacturers are extremely inter-
ested in the boom–bust cycles of our economy as well as of foreign economies so that
they can better predict demand for their products, which, in turn, impacts their inven-
tory levels, employment needs, cash flows, and almost all other business activities
within the firm.
The complexity of these problems is enormous. Take, for example, the problem of
predicting demand for oil and its by-products. In the late 1960s, the price of oil per bar-
rel was very low, and there seemed to be an insatiable worldwide demand for gas and
oil. Then came the oil price shocks of the early and mid-1970s. What would the future
demand for oil be? What about prices? Firms such as Exxon and General Motors were
obviously very interested in the answers to these questions. If oil prices continued to
escalate, would the demand for large cars diminish? What would be the demand for
electricity? By and large, analysts predicted that the demand for energy, and therefore
oil, would be very inelastic; thus, prices would continue to outstrip inflation. However,
these predictions did not take into account a major downswing in the business cycle in
the early 1980s and the greater elasticity of consumer demand for energy than pre-
dicted. By 1980, the world began to see a glut of oil on the market and radically falling
prices. At the time, it seemed hard to believe that consumers were actually benefiting
once again from gasoline price wars. At the time this edition was written, substantial
unrest in the Middle East created a shortage of oil once again. The price of a barrel of
oil and the cost of a gallon of gas in the United States were at record highs.
Oil demand is affected not only by long-term cyclical events but also by seasonal and
random events, as are most other forecasts of demand for any type of product or service.
For instance, consider the service and retail industries. We have witnessed a continued
movement of employment away from manufacturing to the retail and service fields.
However, retailing (whether in-store, catalog, or web-based) is an extremely seasonal
and cyclical business and demand and inventory projections are critical to retailers, so
time series analysis will be used more widely by increasingly sophisticated retailers.
Manufacturers will have a continued need for statistical projections of future
events. Witness the explosive growth in the technology and telecommunications fields
during the 1990s and the substantial contraction of these industries in the early 2000s.
This growth and contraction resulted, to a large extent, from projections of demand that
never completely materialized. Questions that all manufacturers must address include
these: What will the future inflation rate be? How will it affect the cost-of-living adjust-
ments that may be built into a company’s labor contract? How will these adjustments
affect prices and demand? What is the projected pool of managerial skills for 2025?
What will be the effect of the government’s spending and taxing strategies?
What will the future population of young people look like? What will the ethnic mix
be? These issues affect almost all segments of our economy. Demographers are closely
watching the current fertility rate and using almost every available time series forecast-
ing technique to try to project population variables. Very minor miscalculations will
190 CHAPTER 5 Time Series and Their Components

have major impacts on everything from the production of babies’ toys to the financial
soundness of the Social Security system. Interestingly, demographers are looking at very
long-term business cycles (20 years or more per cycle) in trying to predict what this gen-
eration’s population of women of childbearing age will do with regard to having chil-
dren. Will they have one or two children, as families in the 1960s and 1970s did, or will
they return to having two or three, as preceding generations did? These decisions will
determine the age composition of our population for the next 50 to 75 years.
Political scientists are interested in using time series analysis to study the changing
patterns of government spending on defense and social welfare programs. Obviously,
these trends have great impact on the future of whole industries.
Finally, one interesting microcosm of applications of time series analysis has shown
up in the legal field. Lawyers are making increasing use of expert witnesses to testify
about the present value of a person’s or a firm’s future income, the cost incurred from
the loss of a job due to discrimination, and the effect on a market of an illegal strike.
These questions can often be best answered through the judicious use of time series
analysis.
Satellite technology and the World Wide Web have made the accumulation and
transmission of information almost instantaneous. The proliferation of personal comput-
ers, the availability of easy-to-use statistical software programs, and increased access to
databases have brought information processing to the desktop. Business survival during
periods of major competitive change requires quick, data-driven decision making. Time
series analysis and forecasting play a major role in these decision-making processes.

APPENDIX: PRICE INDEX

Several of the series on production, sales, and other economic situations contain data
available only in dollar values. These data are affected by both the physical quantity of
goods sold and their prices. Inflation and widely varying prices over time can cause
analysis problems. For instance, an increased dollar volume may hide decreased sales
in units when prices are inflated. Thus, it is frequently necessary to know how much of
the change in dollar value represents a real change in physical quantity and how much
is due to change in price because of inflation. It is desirable in these instances to
express dollar values in terms of constant dollars.
The concept of purchasing power is important. The current purchasing power of $1
is defined as follows:

100
Current purchasing power of $1 = (5.10)
Consumer Price Index
Thus, if in November 2006 the consumer price index (with 2002 as 100) reaches 150, the
current purchasing power of the November 2006 consumer dollar is

100
Current purchasing power of $1 = = .67
150
The 2006 dollar purchased only two-thirds of the goods and services that could have
been purchased with a base period (2002) dollar.
To express dollar values in terms of constant dollars, Equation 5.11 is used.

Deflated dollar value = 1Dollar value2 * 1Purchasing power of $12 (5.11)

 CHAPTER 5 Time Series and Their Components 191
Suppose that car sales rose from $300,000 in 2005 to $350,000 in 2006, while the
new-car price index (with 2002 as the base) rose from 135 to 155. Deflated sales for
2005 and 2006 would be

Deflated 2005 sales = 1$300,0002 ¢

100
≤ = $222,222
135

Deflated 2006 sales = 1$350,0002 ¢

100
≤ = $225,806
155

Note that actual dollar sales had a sizable increase of $350,000 - $300,000 = $50,000.
However, deflated sales increased by only $225,806 - $222,222 = $3,584.
The purpose of deflating dollar values is to remove the effect of price changes.
This adjustment is called price deflation or is referred to as expressing a series in con-
stant dollars.

Price deflation is the process of expressing values in a series in constant dollars.

The deflating process is relatively simple. To adjust prices to constant dollars, an

index number computed from the prices of commodities whose values are to be
deflated is used. For example, shoe store sales should be deflated by an index of shoe
prices, not by a general price index. For deflated dollar values that represent more than
one type of commodity, the analyst should develop a price index by combining the
appropriate price indexes together in the proper mix.
Example 5.5
Mr. Burnham wishes to study the long-term growth of the Burnham Furniture Store. The
long-term trend of his business should be evaluated using the physical volume of sales. If
this evaluation cannot be done, price changes reflected in dollar sales will follow no consis-
tent pattern and will merely obscure the real growth pattern. If sales dollars are to be used,
actual dollar sales need to be divided by an appropriate price index to obtain sales that are
measured in constant dollars.
The consumer price index (CPI) is not suitable for Burnham because it contains ele-
ments such as rent, food, and personal services not sold by the store, but some components
of this index may be appropriate. Burnham is aware that 70% of sales are from furniture
and 30% from appliances. He can therefore multiply the CPI retail furniture component by
.70, multiply the appliance component by .30, and then add the results to obtain a combined
price index. Table 5-5 illustrates this approach, in which the computations for 1999 are

90.11.702 + 94.61.302 = 91.45

The sales are deflated for 1999 in terms of 2002 purchasing power, so that

Deflated 1999 sales = 142.12 ¢

100
≤ = 46.0
91.45

Table 5-5 shows that, although actual sales gained steadily from 1999 to 2006, physical vol-
ume remained rather stable from 2004 to 2006. Evidently, the sales increases were due to
price markups that were generated, in turn, by the inflationary tendency of the economy.
192 CHAPTER 5 Time Series and Their Components

TABLE 5-5 Burnham Furniture Sales Data, 1999–2006,

for Example 5.5

Retail
Burnham Retail Furniture Appliance Price Price
Sales Price Index Index Indexa Deflated Salesb
Year ($1,000s) (2002=100) (2002=100) (2002=100) ($1,000s of 2002)

1999 42.1 90.1 94.6 91.45 46.0

2000 47.2 95.4 97.2 95.94 49.2
2001 48.4 97.2 98.4 97.56 49.6
2002 50.6 100.0 100.0 100.00 50.6
2003 55.2 104.5 101.1 103.48 53.3
2004 57.9 108.6 103.2 106.98 54.1
2005 59.8 112.4 104.3 109.97 54.4
2006 60.7 114.0 105.6 111.48 54.4

aConstructed for furniture (weight 70%) and appliances (weight 30%).

bSales divided by price index times 100.

Glossary
Business indicators. Business indicators are business- Price deflation. Price deflation is the process of
related time series that are used to help assess the expressing values in a series in constant dollars.
general state of the economy.
Index numbers. Index numbers are percentages that
show changes over time.

Key Formulas

Time series additive decomposition

Yt = Tt + St + It (5.1)

Time series multiplicative decomposition

Yt = Tt * St * It (5.2)

Linear trend
TNt = b0 + b1t (5.3)

Sum of squared errors criterion (with trend TN )

SSE = a 1Yt - TNt22 (5.4)

Quadratic trend
TNt = b0 + b1t + b2t2 (5.5)

Exponential trend
TNt = b0bt1 (5.6)
 CHAPTER 5 Time Series and Their Components 193
Seasonally adjusted data (additive decomposition)

Yt - St = Tt + It (5.7a)

Seasonally adjusted data (multiplicative decomposition)

Yt
= Tt * It (5.7b)
St

Cyclical-irregular component (multiplicative decomposition)

Yt
Ct * It = (5.8)
Tt * St

Irregular component (multiplicative decomposition)

Ct * It
It = (5.9)
Ct

Current purchasing power of $1

100
(5.10)
Consumer Price Index

Deflated dollar value

1Dollar value2 * 1Purchasing power of $12 (5.11)

Problems

1. Explain the concept of decomposing a time series.

2. Explain when a multiplicative decomposition may be more appropriate than an
additive decomposition.
3. What are some basic forces that affect the trend-cycle of most variables?
4. What kind of trend model should be used in each of the following cases?
a. The variable is increasing by a constant rate.
b. The variable is increasing by a constant rate until it reaches saturation and
levels out.
c. The variable is increasing by a constant amount.
5. What are some basic forces that affect the seasonal component of most variables?
6. Value Line estimates of sales and earnings growth for individual companies are
derived by correlating sales, earnings, and dividends to appropriate components of
the National Income Accounts such as capital spending. Jason Black, an analyst for
Value Line, is examining the trend of the capital spending variable from 1977 to
1993. The data are given in Table P-6.
a. Plot the data and determine the appropriate trend model for the years 1977 to
1993.
b. If the appropriate model is linear, compute the linear trend model for the years
1977 to 1993.
194 CHAPTER 5 Time Series and Their Components

TABLE P-6 Capital Spending ($ billions), 1977–1993

Year $Billions Year $Billions Year $Billions

1977 214 1983 357 1989 571
1978 259 1984 416 1990 578
1979 303 1985 443 1991 556
1980 323 1986 437 1992 566
1981 369 1987 443 1993 623
1982 367 1988 545 1994 680a
a
Value Line estimate.
Source:The Value Line Investment Survey (New York:Value Line, 1988, 1990, 1994).

c. What has the average increase in capital spending per year been since 1977?
d. Estimate the trend value for capital spending in 1994.
e. Compare your trend estimate with Value Line’s.
f. What factor(s) influence the trend of capital spending?
7. A large company is considering cutting back on its TV advertising in favor of busi-
ness videos to be given to its customers. This action is being considered after the
company president read a recent article in the popular press touting business
videos as today’s “hot sales weapon.” One thing the president would like to inves-
tigate prior to taking this action is the history of TV advertising in this country,
especially the trend-cycle.
Table P-7 contains the total dollars spent on U.S. TV advertising (in millions of dollars).
a. Plot the time series of U.S. TV advertising expenditures.
b. Fit a linear trend to the advertising data and plot the fitted line on the time
series graph.
c. Forecast TV advertising dollars for 1998.
d. Given the results in part b, do you think there may be a cyclical component in
TV advertising dollars? Explain.
8. Assume the following specific percentage seasonal indexes for March based on the
ratio-to-moving-average method:
102.2 105.9 114.3 122.4 109.8 98.9

What is the seasonal index for March using the median?

TABLE P-7

Year Y Year Y

1980 11,424 1989 26,891

1981 12,811 1990 29,073
1982 14,566 1991 28,189
1983 16,542 1992 30,450
1984 19,670 1993 31,698
1985 20,770 1994 35,435
1986 22,585 1995 37,828
1987 23,904 1996 42,484
1988 25,686 1997 44,580

Source: Statistical Abstract of the United States, various years.

 CHAPTER 5 Time Series and Their Components 195
9. The expected trend value for October is $850. Assuming an October seasonal
index of 1.12 (112%) and the multiplicative model given by Equation 5.2, what
would be the forecast for October?
10. The following specific percentage seasonal indexes are given for the month of
December:
75.4 86.8 96.9 72.6 80.0 85.4

Assume a multiplicative decomposition model. If the expected trend for

December is $900 and the median seasonal adjustment is used, what is the forecast
for December?
11. A large resort near Portland, Maine, has been tracking its monthly sales for several
years but has never analyzed these data. The resort computes the seasonal indexes
for its monthly sales. Which of the following statements about the index are correct?
a. The sum of the 12 monthly index numbers, expressed as percentages, should be
1,200.
b. An index of 85 for May indicates that sales are 15% lower than the average
monthly sales.
c. An index of 130 for January indicates that sales are 30% above the average
monthly sales.
d. The index for any month must be between zero and 200.
e. The average percent index for each of the 12 months should be 100.
12. In preparing a report for June Bancock, manager of the Kula Department Store,
you include the statistics from last year’s sales (in thousands of dollars) shown in
Table P-12. Upon seeing them, Ms. Bancock says, “This report confirms what I’ve
been telling you: Business is getting better and better.” Is this statement accurate?
Why or why not?
13. The quarterly sales levels (measured in millions of dollars) for Goodyear Tire are
shown in Table P-13. Does there appear to be a significant seasonal effect in these
sales levels? Analyze this time series to get the four seasonal indexes and deter-
mine the extent of the seasonal component in Goodyear’s sales.

TABLE P-12
Adjusted
Seasonal
Month Sales ($1,000s) Index (%)

January 125 51
February 113 50
March 189 87
April 201 93
May 206 95
June 241 99
July 230 96
August 245 89
September 271 103
October 291 120
November 320 131
December 419 189

Source: Based on Kula Department Store records.

196 CHAPTER 5 Time Series and Their Components

TABLE P-13
Quarter

Year 1 2 3 4

1985 2,292 2,450 2,363 2,477

1986 2,063 2,358 2,316 2,366
1987 2,268 2,533 2,479 2,625
1988 2,616 2,793 2,656 2,746
1989 2,643 2,811 2,679 2,736
1990 2,692 2,871 2,900 2,811
1991 2,497 2,792 2,838 2,780
1992 2,778 3,066 3,213 2,928
1993 2,874 3,000 2,913 2,916
1994 2,910 3,052 3,116 3,210
1995 3,243 3,351 3,305 3,267
1996 3,246 3,330 3,340a 3,300a

aValue Line estimates.

Source: The Value Line Investment Survey (New York: Value Line,
1988, 1989, 1993, 1994, 1996).

a. Would you use the trend component, the seasonal component, or both to forecast?
b. Forecast for third and fourth quarters of 1996.
c. Compare your forecasts to Value Line’s.
14. The monthly sales of the Cavanaugh Company, pictured in Figure 5.1 (bottom),
are given in Table P-14.
a. Perform a multiplicative decomposition of the Cavanaugh Company sales time
series, assuming trend, seasonal, and irregular components.
b. Would you use the trend component, the seasonal component, or both to forecast?
c. Provide forecasts for the rest of 2006.
15. Construct a table similar to Table P-14 with the natural logarithms of monthly
sales. For example, the value for January 2000 is ln11542 = 5.037.
a. Perform an additive decomposition of ln(sales), assuming the model
Y = T + S + I.

TABLE P-14
Month 2000 2001 2002 2003 2004 2005 2006

January 154 200 223 346 518 613 628

February 96 118 104 261 404 392 308
March 73 90 107 224 300 273 324
April 49 79 85 141 210 322 248
May 36 78 75 148 196 189 272
June 59 91 99 145 186 257
July 95 167 135 223 247 324
August 169 169 211 272 343 404
September 210 289 335 445 464 677
October 278 347 460 560 680 858
November 298 375 488 612 711 895
December 245 203 326 467 610 664
 CHAPTER 5 Time Series and Their Components 197
b. Would you use the trend component, the seasonal component, or both to forecast?
c. Provide forecasts of ln(sales) for the remaining months of 2006.
d. Take the antilogs of the forecasts calculated in part c to get forecasts of the
actual sales for the remainder of 2006.
e. Compare the forecasts in part d with those in Problem 14, part c. Which set of
forecasts do you prefer? Why?
16. Table P-16 contains the quarterly sales (in millions of dollars) for the Disney
Company from the first quarter of 1980 to the third quarter of 1995.
a. Perform a multiplicative decomposition of the time series consisting of Disney’s
quarterly sales.
b. Does there appear to be a significant trend? Discuss the nature of the seasonal
component.
c. Would you use both trend and seasonal components to forecast?
d. Forecast sales for the fourth quarter of 1995 and the four quarters of 1996.
17. The monthly gasoline demand (in thousands of barrels/day) for Yukong Oil
Company of South Korea for the period from January 1986 to September 1996 is
contained in Table P-17.
a. Plot the gasoline demand time series. Do you think an additive or a multiplica-
tive decomposition would be appropriate for this time series? Explain.
b. Perform a decomposition analysis of gasoline demand.
c. Interpret the seasonal indexes.
d. Forecast gasoline demand for the last three months of 1996.
18. Table P-18 contains data values that represent the monthly sales (in billions of dol-
lars) of all retail stores in the United States. Using the data through 1994, perform
a decomposition analysis of this series. Comment on all three components of the
series. Forecast retail sales for 1995 and compare your results with the actual val-
ues provided in the table.

TABLE P-16

Quarter

Year 1 2 3 4

1980 218.1 245.4 265.5 203.5

1981 235.1 258.0 308.4 211.8
1982 247.7 275.8 295.0 270.1
1983 315.7 358.5 363.0 302.2
1984 407.3 483.3 463.2 426.5
1985 451.5 546.9 590.4 504.2
1986 592.4 647.9 726.4 755.5
1987 766.4 819.4 630.1 734.6
1988 774.5 915.7 1,013.4 1,043.6
1989 1,037.9 1,167.6 1,345.1 1,288.2
1990 1,303.8 1,539.5 1,712.2 1,492.4
1991 1,439.0 1,511.6 1,739.4 1,936.6
1992 1,655.1 1,853.5 2,079.1 2,391.4
1993 2,026.5 1,936.8 2,174.5 2,727.3
1994 2,275.8 2,353.6 2,698.4 3,301.7
1995 2,922.8 2,764.0 3,123.6
198 CHAPTER 5 Time Series and Their Components

TABLE P-17

Month 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996

January 15.5 20.4 26.9 36.0 52.1 64.4 82.3 102.7 122.2 145.8 170.0
February 17.8 20.8 29.4 39.0 53.1 68.1 83.6 102.2 121.4 144.4 176.3
March 18.1 22.2 29.9 42.2 56.5 68.5 85.5 104.7 125.6 145.2 174.2
April 20.5 24.1 32.4 44.3 58.4 72.3 91.0 108.9 129.7 148.6 176.1
May 21.3 25.5 33.3 46.6 61.7 74.1 92.1 112.2 133.6 153.7 185.3
June 19.8 25.9 34.5 46.1 61.0 77.6 95.8 109.7 137.5 157.9 182.7
July 20.5 26.1 34.8 48.5 65.5 79.9 98.3 113.5 143.0 169.7 197.0
August 22.3 27.5 39.1 52.6 71.0 86.7 102.2 120.4 149.0 184.2 216.1
September 22.9 25.8 39.0 52.2 68.1 84.4 101.5 124.6 149.9 163.2 192.2
October 21.1 29.8 36.5 50.8 67.5 81.4 98.5 116.7 139.5 155.4
November 22.0 27.4 37.5 51.9 68.8 85.1 101.1 120.6 147.7 168.9
December 22.8 29.7 39.7 55.1 68.1 81.7 102.5 124.9 154.7 178.3

TABLE P-18

Month 1988 1989 1990 1991 1992 1993 1994 1995

January 113.6 122.5 132.6 130.9 142.1 148.4 154.6 167.0

February 115.0 118.9 127.3 128.6 143.1 145.0 155.8 164.0
March 131.6 141.3 148.3 149.3 154.7 164.6 184.2 192.1
April 130.9 139.8 145.0 148.5 159.1 170.3 181.8 187.5
May 136.0 150.3 154.1 159.8 165.8 176.1 187.2 201.4
June 137.5 149.0 153.5 153.9 164.6 175.7 190.1 202.6
July 134.1 144.6 148.9 154.6 166.0 177.7 185.8 194.9
August 138.7 153.0 157.4 159.9 166.3 177.1 193.8 204.2
September 131.9 144.1 145.6 146.7 160.6 171.1 185.9 192.8
October 133.8 142.3 151.5 152.1 168.7 176.4 189.7 194.0
November 140.2 148.8 156.1 155.6 167.2 180.9 194.7 202.4
December 171.0 176.5 179.7 181.0 204.1 218.3 233.3 238.0

Source: Based on Survey of Current Business, 1989, 1993, 1996.

TABLE P-19

Adjusted Adjusted
Month Seasonal Index Month Seasonal Index

January 120 July 153

February 137 August 151
March 100 September 95
April 33 October 60
May 47 November 82
June 125 December 97

Source: Based on Mt. Spokane Resort Hotel records.

19. The adjusted seasonal indexes presented in Table P-19 reflect the changing volume
of business of the Mt. Spokane Resort Hotel, which caters to family tourists in the
summer and skiing enthusiasts during the winter months. No sharp cyclical varia-
tions are expected during 2007.
 CHAPTER 5 Time Series and Their Components 199

TABLE P-24

Commodity
Sales Price Index
Volume ($) (2001 = 100)

2005 January 358,235 118.0

February 297,485 118.4
March 360,321 118.7
April 378,904 119.2
May 394,472 119.7
June 312,589 119.6
July 401,345 119.3

a. If 600 tourists were at the resort in January 2007, what is a reasonable estimate
for February?
b. The monthly trend equation is TN = 140 + 5t where t = 0 represents January
15, 2001. What is the forecast for each month of 2007?
c. What is the average number of new tourists per month?
20. Discuss the performance of the composite index of leading indicators as a barom-
eter of business activity in recent years.
21. What is the present position of the business cycle? Is it expanding or contracting?
When will the next turning point occur?
22. What is the purpose of deflating a time series that is measured in dollars?
23. In the base period of June, the price of a selected quantity of goods was $1,289.73.
In the most recent month, the price index for these goods was 284.7. How much
would the selected goods cost if purchased in the most recent month?
24. Deflate the dollar sales volumes in Table P-24 using the commodity price index.
These indexes are for all commodities, with 2001 = 100.
25. Table P-25 contains the number (in thousands) of men 16 years of age and
older who were employed in the United States for the months from January

TABLE P-25
Year Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.
1993 63,344 63,621 64,023 64,482 65,350 66,412 67,001 66,861 65,808 65,961 65,779 65,545
1994 64,434 64,564 64,936 65,492 66,340 67,230 67,649 67,717 66,997 67,424 67,313 67,292
1995 65,966 66,333 66,758 67,018 67,227 68,384 68,750 68,326 67,646 67,850 67,219 67,049
1996 66,006 66,481 66,961 67,415 68,258 69,298 69,819 69,533 68,614 69,099 68,565 68,434
1997 67,640 67,981 68,573 69,105 69,968 70,619 71,157 70,890 69,890 70,215 70,328 69,849
1998 68,932 69,197 69,506 70,348 70,856 71,618 72,049 71,537 70,866 71,219 71,256 70,930
1999 69,992 70,084 70,544 70,877 71,470 72,312 72,803 72,348 71,603 71,825 71,797 71,699
2000 71,862 72,177 72,501 73,006 73,236 74,267 74,420 74,352 73,391 73,616 73,497 73,338
2001 72,408 72,505 72,725 73,155 73,313 74,007 74,579 73,714 73,483 73,228 72,690 72,547
2002 71,285 71,792 71,956 72,483 73,230 73,747 74,210 73,870 73,596 73,513 72,718 72,437
2003 71,716 72,237 72,304 72,905 73,131 73,894 74,269 74,032 73,715 73,979

Source: Based on Labor force statistics from the Current Population Survey.
200 CHAPTER 5 Time Series and Their Components

1993 to October 2003. Use Minitab to do a multiplicative decomposition of

these data and generate forecasts for the next 12 months. Does a multiplicative
decomposition appear to be appropriate in this case? Explain. Is there a strong
seasonal component in these data? Are you surprised? Do the forecasts seem
reasonable?
26. Refer to Problem 25. The default multiplicative decomposition in Minitab
assumes a linear trend. Plot the employed men data in Table P-25 and examine
the years 1993–2000 and 2001–2003. Is a linear trend appropriate? If not, can
you suggest a trend curve that might be appropriate? Fit your suggested trend
curve and store the residuals. Compute the autocorrelation function of the
residuals. Do the residual autocorrelations suggest a seasonal component?
Explain.
27. Table P-27 gives quarterly sales (in millions of dollars) for Wal-Mart Stores from
1990–2004. Use Minitab to do a multiplicative decomposition of the Wal-Mart
sales time series for the years 1990–2003 and generate forecasts for the four quar-
ter of 2004. Does a multiplicative decomposition appear to be appropriate for the
Wal-Mart data? Explain. Is there a strong seasonal component? Are you sur-
prised? Compare the quarterly forecasts for 2004 with the actual sales. Do the
results reinforce the choice of a multiplicative decomposition?
28. Refer to Problem 27. The default multiplicative decomposition in Minitab assumes
a linear trend. Fit and plot a linear trend line to the Wal-Mart sales. Is a linear trend
appropriate for these data? If not, can you suggest a trend curve that might be
appropriate? Fit your suggested trend curve and store the residuals. Compute the
residual autocorrelations. Do the residual autocorrelations suggest a seasonal
component? Explain.

TABLE P-27

Quarter

Year 1 2 3 4
1990 6,768 7,544 7,931 10,359
1991 9,281 10,340 10,628 13,639
1992 11,649 13,028 13,684 17,122
1993 13,920 16,237 16,827 20,361
1994 17,686 19,942 20,418 24,448
1995 20,440 22,723 22,914 27,550
1996 22,772 25,587 25,644 30,856
1997 25,409 28,386 28,777 35,386
1998 29,819 33,521 33,509 40,785
1999 35,129 28,913 40,899 51,868
2000 43,447 46,588 46,181 57,079
2001 48,565 53,187 53,185 64,735
2002 52,126 56,781 55,765 66,905
2003 57,224 63,231 63,036 75,190
2004 65,443 70,466 69,261 82,819
Source: Based on S&P Compustat North American Industrial Quarterly database.
 CHAPTER 5 Time Series and Their Components 201

CASES

CASE 5-1 THE SMALL ENGINE DOCTOR13

The Small Engine Doctor is the name of a business the dealer had to back-order one or more parts for
developed by Thomas Brown, who is a mail carrier any given repair job. Parts ordered from the manu-
for the U.S. Postal Service. He had been a tinkerer facturer had lead times of anywhere from 30 to 120
since childhood, always taking discarded household days. As a result, Tom changed his policy and began
gadgets apart in order to understand “what made to order parts directly from the factory. He found
them tick.” As Tom grew up and became a typical that shipping and handling charges ate into his prof-
suburbanite, he acquired numerous items of lawn its, even though the part price was only 60% of retail.
and garden equipment. When Tom found out about a However, the two most important problems created
course in small engine repair offered at a local com- by the replacement parts were lost sales and storage
munity college, he jumped at the opportunity. Tom space. Tom attracted customers because of his qual-
started small engine repair by dismantling his own ity service and reasonable repair charges, which were
equipment, overhauling it, and then reassembling it. possible because of his low overhead. Unfortunately,
Soon after completing the course in engine repair, he many potential customers would go to equipment
began to repair lawn mowers, rototillers, snowblow- dealers rather than wait several months for repair.
ers, and other lawn and garden equipment for friends The most pressing problem was storage space. While
and neighbors. In the process, he acquired various a piece of equipment was waiting for spare parts, it
equipment manuals and special tools. had to be stored on the premises. It did not take long
It was not long before Tom decided to turn his for both his workshop and his one-car garage to
hobby into a part-time business. He placed an adver- overflow with equipment while he was waiting for
tisement in a suburban shopping circular under the spare parts. In the second year of operation, Tom
name of the Small Engine Doctor. Over the last two actually had to suspend advertising as a tactic to limit
years, the business has grown enough to provide a customers due to lack of storage space.
nice supplement to his regular salary. Although the Tom has considered stocking inventory for his
growth was welcomed, as the business is about to third year of operation. This practice will reduce pur-
enter its third year of operation there are a number of chasing costs by making it possible to obtain quantity
concerns.The business is operated out of Tom’s home. discounts and more favorable shipping terms. He also
The basement is partitioned into a family room, a hopes that it will provide much better turnaround
workshop, and an office. Originally, the office area time for the customers, improving both cash flow and
was used to handle the advertising, order processing, sales.The risks in this strategy are uncontrolled inven-
and bookkeeping. All engine repair was done in the tory carrying costs and part obsolescence.
workshop. Tom’s policy has been to stock only a lim- Before committing himself to stocking spare
ited number of parts, ordering replacement parts as parts, Tom wants to have a reliable forecast for busi-
they are needed. This seemed to be the only practical ness activity in the forthcoming year. He is confident
way of dealing with the large variety of parts involved enough in his knowledge of product mix to use an
in repairing engines made by the dozen or so manu- aggregate forecast of customer repair orders as a
facturers of lawn and garden equipment. basis for selectively ordering spare parts. The fore-
Spare parts have proved to be the most aggra- cast is complicated by seasonal demand patterns and
vating problem in running the business. Tom started a trend toward increasing sales.
his business by buying parts from equipment dealers. Tom plans to develop a sales forecast for the
This practice had several disadvantages. First, he had third year of operation. A sales history for the first
to pay retail for the parts. Second, most of the time two years is given in Table 5-6.

13This case was contributed by William P. Darrow of the Towson State University, Towson, Maryland.
202 CHAPTER 5 Time Series and Their Components

TABLE 5-6 Small Engine Doctor Sales History

2005 2006 2005 2006

Month (units) (units) Month (units) (units)

January 5 21 July 28 46
February 8 20 August 20 32
March 10 29 September 14 27
April 18 32 October 8 13
May 26 44 November 6 11
June 35 58 December 26 52

ASSIGNMENT
1. Plot the data on a two-year time horizon from constants: 1 = .1,
= .12, 1 = .25,
= .252,
2005 through 2006. Connect the data points to and 1 = .5,
= .52. Plot the three sets of
make a time series plot. smoothed values on the time series graph.
2. Develop a trend-line equation using linear Generate forecasts through the end of the third
regression and plot the results. year for each of the trend-adjusted exponential
3. Estimate the seasonal adjustment factors for smoothing possibilities considered.
each month by dividing the average demand 5. Calculate the MAD values for the two models
for corresponding months by the average of that visually appear to give the best fits (the
the corresponding trend-line forecasts. Plot the most accurate one-step-ahead forecasts).
fitted values and forecasts for 2007 given by 6. If you had to limit your choice to one of the models
Trend * Seasonal. in Questions 2 and 4, identify the model you would
4. Smooth the time series using Holt’s linear expo- use for your business planning in 2007, and dis-
nential smoothing with three sets of smoothing cuss why you selected that model over the others.

CASE 5-2 MR. TUX

John Mosby has been looking forward to the decom- solid forecasts on which to base their investment
position of his time series, monthly sales dollars. He decisions. John knows that his business is improving
knows that the series has a strong seasonal effect and that future prospects look bright, but investors
and would like it measured for two reasons. First, his want documentation.
banker is reluctant to allow him to make variable The monthly sales volumes for Mr. Tux for the
monthly payments on his loan. John has explained years 1999 through 2005 are entered into Minitab.
that because of the seasonality of his sales and Since 1998 was the first year of business, the sales
monthly cash flow, he would like to make extra pay- volumes were extremely low compared to the rest of
ments in some months and reduced payments, and the years. For this reason, John decides to eliminate
even no payments, in others. His banker wants to see these values from the analysis. The seasonal indexes
some verification of John’s assertion that his sales are shown in Table 5-7. The rest of the computer
have a strong seasonal effect. printout is shown in Table 5-8.
Second, John wants to be able to forecast his John is not surprised to see the seasonal indexes
monthly sales. He needs such forecasts for planning shown in Table 5-7, and he is pleased to have some
purposes, especially since his business is growing. hard numbers to show his banker. After reviewing
Both bankers and venture capitalists want some these figures, the banker agrees that John will make
 CHAPTER 5 Time Series and Their Components 203

TABLE 5-7 Summary of the Monthly

Seasonal Indexes for
Mr. Tux
Time Series Decomposition
Trend-Line Equation
TNt = 12,133 + 3,033t

Seasonal Index

Period Index

1 0.3144
2 0.4724
3 0.8877
4 1.7787
5 1.9180
6 1.1858
7 1.0292
8 1.2870
9 0.9377
10 0.8147
11 0.6038
12 0.7706

Accuracy of Model

MAPE: 20
MAD: 21,548
MSD: 9.12E  08

TABLE 5-8 Calculations for the Short-Term Components for Mr. Tux

t Year Month Sales T SCI TCI CI C I

1 1999 January 16,850 15,166 1.1111 53,589 3.5336 — —

2 February 12,753 18,198 0.7708 26,997 1.4835 — —
3 March 26,901 21,231 1.2671 30,306 1.4274 2.1387 0.6674
4 April 61,494 24,264 2.5344 34,572 1.4249 1.7545 0.8121
5 May 147,862 27,297 5.4169 77,092 2.8242 1.7567 1.6077
6 June 57,990 30,329 1.9120 48,902 1.6124 1.7001 0.9484
7 July 51,318 33,362 1.5382 49,862 1.4946 1.5397 0.9707
8 August 53,599 36,395 1.4727 41,647 1.1443 1.2142 0.9424
9 September 23,038 39,428 0.5843 24,568 0.6231 1.0325 0.6035
10 October 41,396 42,460 0.9749 50,815 1.1968 0.8550 1.3997
11 November 19,330 45,493 0.4249 32,014 0.7037 0.8161 0.8623
12 December 22,707 48,526 0.4679 29,466 0.6072 0.9305 0.6526
13 2000 January 15,395 51,559 0.2986 48,961 0.9496 0.7912 1.2002
14 February 30,826 54,592 0.5647 65,257 1.1954 0.8418 1.4201
15 March 25,589 57,624 0.4441 28,828 0.5003 1.0439 0.4793
16 April 103,184 60,657 1.7011 58,011 0.9564 1.0273 0.9309
17 May 197,608 63,690 3.1027 103,029 1.6177 0.8994 1.7985
18 June 68,600 66,723 1.0281 57,850 0.8670 0.9945 0.8719
19 July 39,909 69,755 0.5721 38,777 0.5559 0.9685 0.5740
20 August 91,368 72,788 1.2553 70,994 0.9754 0.8308 1.1740
204 CHAPTER 5 Time Series and Their Components

TABLE 5-8 (Continued)

t Year Month Sales T SCI TCI CI C I

21 September 58,781 75,821 0.7753 62,684 0.8267 0.7927 1.0430

22 October 59,679 78,854 0.7568 73,257 0.9290 0.8457 1.0986
23 November 33,443 81,887 0.4084 55,387 0.6764 0.8515 0.7944
24 December 53,719 84,919 0.6326 69,709 0.8209 0.8567 0.9582
25 2001 January 27,773 87,952 0.3158 88,327 1.0043 0.7935 1.2657
26 February 36,653 90,985 0.4029 77,592 0.8528 0.9102 0.9370
27 March 51,157 94,018 0.5441 57,632 0.6130 0.9609 0.6380
28 April 217,509 97,050 2.2412 122,285 1.2600 0.9401 1.3403
29 May 206,229 100,083 2.0606 107,523 1.0743 0.9579 1.1216
30 June 110,081 103,116 1.0676 92,830 0.9003 1.0187 0.8837
31 July 102,893 106,149 0.9693 99,973 0.9418 0.9658 0.9751
32 August 128,857 109,182 1.1802 100,124 0.9170 0.9875 0.9287
33 September 104,776 112,214 0.9337 111,732 0.9957 0.9858 1.0100
34 October 111,036 115,247 0.9635 136,299 1.1827 0.9743 1.2139
35 November 63,701 118,280 0.5386 105,499 0.8920 0.9516 0.9373
36 December 82,657 121,313 0.6814 107,260 0.8842 0.9131 0.9683
37 2002 January 31,416 124,345 0.2527 99,913 0.8035 0.8246 0.9745
38 February 48,341 127,378 0.3795 102,335 0.8034 0.8507 0.9444
39 March 85,651 130,411 0.6568 96,492 0.7399 0.8951 0.8266
40 April 242,673 133,444 1.8185 136,432 1.0224 0.9331 1.0957
41 May 289,554 136,477 2.1216 150,967 1.1062 0.9914 1.1158
42 June 164,373 139,509 1.1782 138,614 0.9936 1.0314 0.9634
43 July 160,608 142,542 1.1267 156,051 1.0948 1.0312 1.0616
44 August 176,096 145,575 1.2097 136,829 0.9399 0.9960 0.9437
45 September 142,363 148,608 0.9580 151,815 1.0216 1.0405 0.9819
46 October 114,907 151,640 0.7578 141,051 0.9302 1.0306 0.9026
47 November 113,552 154,673 0.7341 188,061 1.2159 1.0468 1.1615
48 December 127,042 157,706 0.8056 164,856 1.0453 1.0502 0.9953
49 2003 January 51,604 160,739 0.3210 164,118 1.0210 1.1464 0.8906
50 February 80,366 163,771 0.4907 170,129 1.0388 1.0779 0.9637
51 March 208,938 166,804 1.2526 235,383 1.4111 1.0210 1.3821
52 April 263,830 169,837 1.5534 148,327 0.8734 1.0273 0.8501
53 May 252,216 172,870 1.4590 131,500 0.7607 0.9815 0.7751
54 June 219,566 175,903 1.2482 185,157 1.0526 0.8819 1.1936
55 July 149,082 178,935 0.8332 144,852 0.8095 0.9135 0.8862
56 August 213,888 181,968 1.1754 166,194 0.9133 0.9359 0.9759
57 September 178,947 185,001 0.9673 190,828 1.0315 0.9281 1.1114
58 October 133,650 188,034 0.7108 164,059 0.8725 0.9857 0.8852
59 November 116,946 191,066 0.6121 193,682 1.0137 0.9929 1.0210
60 December 164,154 194,099 0.8457 213,015 1.0975 0.9609 1.1422
61 2004 January 58,843 197,132 0.2985 187,140 0.9493 1.0356 0.9167
62 February 82,386 200,165 0.4116 174,405 0.8713 1.0261 0.8492
63 March 224,803 203,198 1.1063 253,256 1.2464 0.9701 1.2847
64 April 354,301 206,230 1.7180 199,190 0.9659 1.0295 0.9382
65 May 328,263 209,263 1.5687 171,149 0.8179 1.0488 0.7798
66 June 313,647 212,296 1.4774 264,495 1.2459 1.0395 1.1985
 CHAPTER 5 Time Series and Their Components 205

TABLE 5-8 (Continued)

t Year Month Sales T SCI TCI CI C I

67 July 214,561 215,329 0.9964 208,473 0.9682 1.0231 0.9463

68 August 337,192 218,361 1.5442 262,003 1.1999 1.0177 1.1790
69 September 183,482 221,394 0.8288 195,664 0.8838 0.9721 0.9092
70 October 144,618 224,427 0.6644 177,522 0.7910 0.9862 0.8020
71 November 139,750 227,460 0.6144 231,449 1.0175 0.9398 1.0828
72 December 184,546 230,493 0.8007 239,476 1.0390 1.0367 1.0022
73 2005 January 71,043 233,525 0.3042 225,940 0.9675 1.1142 0.8684
74 February 152,930 236,558 0.6465 323,742 1.3686 1.1005 1.2436
75 March 250,559 239,591 1.0458 282,272 1.1781 1.0602 1.1112
76 April 409,567 242,624 1.6881 230,261 0.9490 1.0518 0.9023
77 May 394,747 245,656 1.6069 205,813 0.8378 0.9559 0.8765
78 June 272,874 248,689 1.0973 230,111 0.9253 0.9492 0.9748
79 July 230,303 251,722 0.9149 223,768 0.8890 0.9211 0.9651
80 August 375,402 254,755 1.4736 291,693 1.1450 0.9169 1.2488
81 September 195,409 257,787 0.7580 208,383 0.8084 0.9599 0.8421
82 October 173,518 260,820 0.6653 212,998 0.8166 1.0337 0.7900
83 November 181,702 263,853 0.6887 300,928 1.1405 — —
84 December 258,713 266,886 0.9694 335,719 1.2579 — —

double payments on his loan in April, May, June, and 2005 November 104
August and make no payments at all in January, December 105
February, November, and December. His banker 2006 January 105
asks for a copy of the seasonal indexes to show his February 106
boss and include in John’s loan file. March 107
Turning to a forecast for the first six months of April 109
2006, John begins by projecting trend values using May 110
the trend equation TNt = 12,133 + 3,033t. The trend June 111
estimate for January 2006 is
TN85 = 12,133 + 3,0331852 = 269,938
Turning to the irregular (I ) value for these
Next, John obtains the seasonal index from months, John does not foresee any unusual events
Table 5-7. The index for January is 31.44%. John has except for March 2006. In that month, he plans to hold
been reading The Wall Street Journal and watching an open house and reduce the rates in one of his stores
the business news talk shows on a regular basis, so that he is finishing remodeling. Because of this promo-
he already has an idea of the overall nature of the tion, to be accompanied with radio and TV advertis-
economy and its future course. He also belongs to a ing, he expects sales in that store to be 50% higher
business service club that features talks by local eco- than normal. For his overall monthly sales, he thinks
nomic experts on a regular basis. As he studies the C this effect will result in about 15% higher sales overall.
column of his computer output, showing the cyclical Using all the figures he has estimated, along
history of his series, he thinks about how he will with his computer output, John makes the forecasts
forecast this value for the first six months of 2006. for Mr. Tux sales for the first six months of 2006,
Since the forecasts of national and local experts call shown in Table 5-9.
for an improvement in business in 2006 and since After studying the 2006 forecasts, John is
the last C value for October 2005 has turned up disturbed to see the wide range of monthly sales
(103.4%), he decides to use the following C values projections—from $89,112 to $595,111. Although he
for his forecasts: knew his monthly volume had considerable variability,
206 CHAPTER 5 Time Series and Their Components

TABLE 5-9 Forecasts for Mr. Tux

Sales Forecast = T  S  C  I

January 89,112 = 269,938 .3144 1.05 1.00

February 136,689 = 272,971 .4724 1.06 1.00
March 301,483 = 276,004 .8877 1.07 1.15
April 540,992 = 279,037 1.7787 1.09 1.00
May 595,111 = 282,070 1.9180 1.10 1.00
June 375,263 = 285,103 1.1858 1.11 1.00

he is concerned about such wide fluctuations. John His real concern, however, is focused on his
has been thinking about expanding from his current worst months, January and February. He has recently
Spokane location into the Seattle area. He has been considering buying a tuxedo-shirt-making
recently discovered that there are several “dress up” machine that he saw at a trade show, thinking that he
events in Seattle that make it different from his pres- might be able to concentrate on that activity during
ent Spokane market. Formal homecoming dances, in the winter months. If he receives a positive reaction
particular, are big in Seattle but not in Spokane. from potential buyers of shirts for this period of time,
Since these take place in the fall, when his Spokane he might be willing to give it a try. As it is, the
business is slower (see the seasonal indexes for seasonal indexes on his computer output have
October and November), he sees the advantage of focused his attention on the extreme swings in his
leveling his business by entering the Seattle market. monthly sales levels.

QUESTIONS
1. Suppose John’s banker asked for two sentences Determine the Seattle seasonal indexes that
to show his boss that would justify John’s would be ideal to balance out the monthly rev-
request to make extra loan payments in some enues for Mr. Tux.
months and no payments in others. Write these 3. Disregarding Seattle, how much volume would
two sentences. John have to realize from his shirt-making
2. Assume that John will do exactly twice as much machine to make both January and February
business in Seattle as Spokane next year. “average”?

CASE 5-3 CONSUMER CREDIT

COUNSELING

The Consumer Credit Counseling (CCC) operation Dorothy gave you these data and asked you to
was described in Case 1-2. complete a time series decomposition analysis. She
The executive director, Marv Harnishfeger, con- emphasized that she wanted to understand com-
cluded that the most important variable that CCC pletely the trend and seasonality components.
needed to forecast was the number of new clients Dorothy wanted to know the importance of each
that would be seen for the rest of 1993. Marv pro- component. She also wanted to know if any unusual
vided Dorothy Mercer monthly data for the number irregularities appeared in the data. Her final
of new clients seen by CCC for the period from instruction required you to forecast for the rest of
January 1985 through March 1993 (see Case 3-3). 1993.
 CHAPTER 5 Time Series and Their Components 207

ASSIGNMENT
Write a report that provides Dorothy with the infor-
mation she has requested.

CASE 5-4 MURPHY BROTHERS FURNITURE

In Case 4-4, Julie Murphy developed a naive model production requirements, Julie is very anxious to
that combined seasonal and trend estimates (similar prepare short-term forecasts for the company that
to Equation 4.5). One of the major reasons why she are based on the best available information concern-
chose this naive model was its simplicity. Julie knew ing demand.
that her father, Glen, would need to understand the For forecasting purposes, Julie has decided to
forecasting model used by the company. use only the data gathered since 1996, the first full
It is now October of 2002, and a lot has changed. year Murphy Brothers manufactured its own line of
Glen Murphy has retired. Julie has completed sev- furniture (Table 5-10). Julie can see (Figure 5-13)
eral business courses, including business forecasting, that her data have both trend and seasonality. For
at the local university. Murphy Brothers Furniture this reason, she decides to use a time series decom-
built a factory in Dallas and began to manufacture position approach to analyze her sales variable.
its own line of furniture in October 1995. Since Figure 5-13 shows that the time series she
Monthly sales data for Murphy Brothers Furni- is analyzing has roughly the same variability
ture from 1996 to October 2002 are shown in Table throughout the length of the series, Julie decides to
5-10. As indicated by the pattern of these data use an additive components model to forecast. She
demonstrated in Figure 5-13, sales have grown dra- runs the model Yt = Tt + St + It. A summary of the
matically since 1996. Unfortunately, Figure 5-13 also results is shown in Table 5-11. Julie checks the auto-
demonstrates that one of the problems with demand correlation pattern of the residuals (see Figure 5-14)
is that it is somewhat seasonal. The company’s gen- for randomness. The residuals are not random, and
eral policy is to employ two shifts during the sum- the model does not appear to be adequate.
mer and early fall months and then work a single Julie is stuck. She has tried a naive model that
shift through the remainder of the year. Thus, sub- combined seasonal and trend estimates, Winters’
stantial inventories are developed in the late sum- exponential smoothing, and classical decomposition.
mer and fall months until demand begins to pick up Julie finally decides to adjust seasonality out of the
in November and December. Because of these data so that forecasting techniques that cannot

TABLE 5-10 Monthly Sales for Murphy Brothers Furniture, 1996–2002

Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1996 4,964 4,968 5,601 5,454 5,721 5,690 5,804 6,040 5,843 6,087 6,469 7,002
1997 5,416 5,393 5,907 5,768 6,107 6,016 6,131 6,499 6,249 6,472 6,946 7,615
1998 5,876 5,818 6,342 6,143 6,442 6,407 6,545 6,758 6,485 6,805 7,361 8,079
1999 6,061 6,187 6,792 6,587 6,918 6,920 7,030 7,491 7,305 7,571 8,013 8,727
2000 6,776 6,847 7,531 7,333 7,685 7,518 7,672 7,992 7,645 7,923 8,297 8,537
2001 7,005 6,855 7,420 7,183 7,554 7,475 7,687 7,922 7,426 7,736 8,483 9,329
2002 7,120 7,124 7,817 7,538 7,921 7,757 7,816 8,208 7,828

Source: Murphy Brothers sales records.

208 CHAPTER 5 Time Series and Their Components

FIGURE 5-13 Murphy Brothers Furniture Monthly Sales, 1996–2002

TABLE 5-11 Summary of the

Decomposition Model for
Murphy Brothers
Furniture

Time Series Decomposition

Trend-Line Equation
TNt = 5,672 + 31.4t

Seasonal Index

Period Index
1 - 674.60
2 - 702.56
3 - 143.72
4 - 366.64
5 - 53.52
6 - 173.27
7 - 42.74
8 222.32
9 - 57.95
10 145.76
11 612.30
12 1234.63
Accuracy Measures

MAPE: 1.9
MAD: 135.1
MSD: 30,965.3
 CHAPTER 5 Time Series and Their Components 209

FIGURE 5-14 Autocorrelation Function for Residuals Using an

Additive Time Series Decomposition Model for
Murphy Brothers Furniture

TABLE 5-12 Seasonally Adjusted Monthly Sales for Murphy Brothers Furniture,
1996–2002

Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1996 5,621 5,671 5,745 5,821 5,775 5,863 5,847 5,818 5,901 5,941 5,857 5,767
1997 6,091 6,096 6,051 6,135 6,161 6,189 6,174 6,277 6,307 6,326 6,334 6,380
1998 6,551 6,521 6,486 6,510 6,496 6,580 6,588 6,536 6,543 6,659 6,749 6,844
1999 6,736 6,890 6,936 6,954 6,972 7,093 7,073 7,269 7,363 7,425 7,401 7,492
2000 7,451 7,550 7,675 7,700 7,739 7,691 7,715 7,770 7,703 7,777 7,685 7,302
2001 7,680 7,558 7,564 7,550 7,608 7,648 7,730 7,700 7,484 7,590 7,871 8,094
2002 7,795 7,827 7,961 7,905 7,975 7,930 7,859 7,986 7,886

handle seasonal data can be applied. Julie deseason- adds 674.60 to the data for each January and sub-
alizes the data by adding or subtracting the seasonal tracts 1,234.63 from the data for each December.
index for the appropriate month. For example, she Table 5-12 shows the seasonally adjusted data.

ASSIGNMENT
1. Using the data through 2001 in Table 5-12, 3. Forecast sales for October 2002 using the same
develop a model to forecast the seasonally procedure as in part 2.
adjusted sales data and generate forecasts for 4. Compare the pattern of the retail sales data pre-
the first nine months of 2002. sented in Case 3-1A with the pattern of the
2. Using the forecasts from part 1, forecast sales actual sales data from 1992 through 1995 pre-
for the first nine months of 2002 by adding or sented in Case 4-4 with the pattern of the actual
subtracting the appropriate seasonal index in sales data from 1996 through 2001 presented in
Table 5-11. Are these forecasts accurate when this case.
compared with the actual values?
210 CHAPTER 5 Time Series and Their Components

CASE 5-5 AAA WASHINGTON14

In 1993, AAA Washington was one of the two agency; and an insurance agency. The club provided
regional automobile clubs affiliated with the Ameri- these services through a network of offices located
can Automobile Association (AAA or Triple A) in Bellevue, Bellingham, Bremerton, Everett,
operating in Washington State. At that time, 69% of Lynnwood, Olympia, Renton, Seattle, Tacoma, the
all people belonging to automobile clubs were mem- Tri-Cities (Pasco, Richland, and Kennewick),
bers of the American Automobile Association, mak- Vancouver, Wenatchee, and Yakima, Washington.
ing it the largest automobile club in North America. Club research had consistently shown that the
AAA was a national association that serviced its emergency road service benefit was the primary rea-
individual members through a federation of approx- son that people join AAA. The importance of emer-
imately 150 regional clubs that chose to be affiliated gency road service in securing members was
with the national association. The national associa- reflected in the three types of memberships offered
tion set a certain number of minimum standards by AAA Washington: Basic, AAA Plus, and AAA
with which the affiliated clubs had to comply in Plus RV. Basic membership provided members five
order to retain their affiliation with the association. miles of towing from the point at which their vehicle
Each regional club was administered locally by its was disabled. AAA Plus provided members with 100
own board of trustees and management staff. The miles of towing from the point at which their vehicle
local trustees and managers were responsible for was disabled. AAA Plus RV provided the 100-mile
recruiting and retaining members within their towing service to members who own recreational
assigned territories and for ensuring the financial vehicles in addition to passenger cars and light
health of the regional club. Beyond compliance with trucks. Providing emergency road service was also
the minimum standards set by the AAA, each the club’s single largest operating expense. It was
regional club was free to determine what additional projected that delivering emergency road service
products and services it would offer and how it would cost $9.5 million, 37% of the club’s annual
would price these products and services. operating budget, in the next fiscal year.
AAA Washington was founded in 1904. Its ser- Michael DeCoria, a CPA and MBA graduate of
vice territory consisted of the 26 Washington coun- Eastern Washington University, had recently joined
ties west of the Columbia River. The club offered its the club’s management team as vice president of
members a variety of automobile and automobile- operations. One of the responsibilities Michael
travel-related services. Member benefits provided in assumed was the management of emergency road
cooperation with the national association included service. Early in his assessment of the emergency
emergency road services; a rating service for lodging, road service operation, Mr. DeCoria discovered that
restaurants, and automotive repair shops; tour guides emergency road service costs had increased at a rate
to AAA-approved lodging, restaurants, camping, and faster than could be justified by the rate of inflation
points of interest; and advocacy for legislation and and the growth in club membership. Michael began
public spending in the best interests of the motor- by analyzing the way the club delivered emergency
ing public. In addition to these services, AAA road service to determine if costs could be con-
Washington offered its members expanded protec- trolled more tightly in this area.
tion plans for emergency road service; financial ser- Emergency road service was delivered in one of
vices, including affinity credit cards, personal lines of four ways: the AAA Washington service fleet, con-
credit, checking and savings accounts, time deposits, tracting companies, reciprocal reimbursement, and
and no-fee American Express Travelers Cheques; direct reimbursement. AAA Washington’s fleet of
access to a fleet of mobile diagnostic vans for deter- service vehicles responded to emergency road ser-
mining the “health” of a member’s vehicle; a travel vice calls from members who became disabled in the

14Thiscase was provided by Steve Branton, former student and MBA graduate, Eastern Washington
University.
 CHAPTER 5 Time Series and Their Components 211
downtown Seattle area. Within AAA Washington’s association. Finally, members could contact a towing
service area, but outside of downtown Seattle, com- company of their choice directly, paying for the
mercial towing companies that had contracted with towing service and then submitting a request for
AAA Washington to provide this service responded reimbursement to the club. AAA Washington
to emergency road service calls. Members arranged reimbursed the actual cost of the towing or $50,
for both of these types of emergency road service by whichever was less, directly to the member. After a
calling the club’s dispatch center. Should a member careful examination of the club’s four service deliv-
become disabled outside of AAA Washington’s ser- ery methods, Michael concluded that the club was
vice area, the member could call the local AAA- controlling the cost of service delivery as tightly as
affiliated club to receive emergency road service. was practical.
The affiliate club paid for this service and then billed Another possible source of the increasing costs
AAA Washington for reciprocal reimbursement was a rise in the use of emergency road service.
through a clearing service provided by the national Membership had been growing steadily for several

TABLE 5-13 Emergency Road Call Volume by

Month for AAA Washington

Year Month Calls Year Month Calls

1988 May 20,002 1991 January 23,441

June 21,591 February 19,205
July 22,696 March 20,386
August 21,509 April 19,988
September 22,123 May 19,077
October 21,449 June 19,141
November 23,475 July 20,883
December 23,529 August 20,709
1989 January 23,327 September 19,647
February 24,050 October 22,013
March 24,010 November 22,375
April 19,735 December 22,727
May 20,153 1992 January 22,367
June 19,512 February 21,155
July 19,892 March 21,209
August 20,326 April 19,286
September 19,378 May 19,725
October 21,263 June 20,276
November 21,443 July 20,795
December 23,366 August 21,126
1990 January 23,836 September 20,251
February 23,336 October 22,069
March 22,003 November 23,268
April 20,155 December 26,039
May 20,070 1993 January 26,127
June 19,588 February 20,067
July 20,804 March 19,673
August 19,644 April 19,142
September 17,424
October 20,833
November 22,490
December 24,861
212 CHAPTER 5 Time Series and Their Components

years, but the increased cost was more than what road service calls per member grew by 3.28%, from
could be attributed to simple membership growth. an average of 0.61 calls per member to 0.63 calls.
Michael then checked to see if there was a growth in Concerned that a continuation of this trend would
emergency road service use on a per-member basis. have a negative impact on the club financially,
He discovered that between fiscal year 1990 and fis- Mr. DeCoria gathered the data on emergency road
cal year 1991, the average number of emergency service call volume presented in Table 5-13.

ASSIGNMENT
1. Perform a time series decomposition on the road service call volume that you discovered
AAA emergency road service calls data. from your time series decomposition analysis.
2. Write a memo to Mr. DeCoria summarizing the
important insights into changes in emergency

CASE 5-6 ALOMEGA FOOD STORES

In Example 1.1, Julie Ruth, Alomega Food Stores After reviewing the results of this regression
president, had collected data on her company’s analysis, including the low r2 value (36%), she
monthly sales along with several other variables she decided to try time series decomposition on the sin-
thought might be related to sales (review Example gle variable, monthly sales. Figure 5-15 shows the
1.1). Case 2-3 explained how Julie used her Minitab plot of sales data (see Case 3-4) that she obtained. It
program to calculate a simple regression equation looked like sales were too widely scattered around
using the best predictor for monthly sales. the trend line for accurate forecasts. This impression

FIGURE 5-15 Trend Analysis for Alomega Food Store Sales

 CHAPTER 5 Time Series and Their Components 213

FIGURE 5-16 Decomposition Fit for Alomega Food Store Sales

was confirmed when she looked at the MAPE value She also noted that the MAPE had dropped to
of 28. She interpreted this to mean that the average 12%, a definite improvement over the value
percentage error between the actual values and the obtained using the trend equation alone.
trend line was 28%, a value she considered too high. Finally, Julie had the program provide forecasts
She next tried a multiplicative decomposition of the for the next 12 months, using the trend equation pro-
data. The results are shown in Figure 5-16. jections modified by the seasonal indexes. She
In addition to the trend equation shown on the thought she might use these as forecasts for her
printout, Julie was interested in the seasonal planning purposes but wondered if another forecast-
(monthly) indexes that the program calculated. She ing method might produce better forecasts. She was
noted that the lowest sales month was December also concerned about what her production manager,
(month 12, index = 0.49) and the highest was Jackson Tilson, might say about her forecasts, espe-
January (month 1, index = 1.74). She was aware of cially since he had expressed concern about using
the wide swing between December and January but the computer to make predictions (see his remarks
didn’t realize how extreme it was. at the end of Example 1.1).

QUESTION
1. What might Jackson Tilson say about her fore-
casts?

CASE 5-7 SURTIDO COOKIES

The monthly sales (in kilograms) for Surtido seasonal component. He also considered a smooth-
Cookies were introduced in Case 3-5. In Case 4-8, ing method to generate forecasts of future sales.
Jame Luna examined the autocorrelations for A member of Jame’s team has had some experience
cookie sales to determine if there might be a with decomposition analysis and suggests that they
214 CHAPTER 5 Time Series and Their Components

try a multiplicative decomposition of the cookie year. Jame doesn’t have a lot of faith in a procedure
sales data. Not only will they get an indication of the that tries to estimate components that cannot be
trend in sales, but also they will be able to look at the observed directly. However, since Minitab is avail-
seasonal indexes. The latter can be important pieces able, he agrees to give decomposition a try and to
of information for determining truck float and ware- use it to generate forecasts of sales for the remaining
house (inventory) requirements throughout the months of 2003.

QUESTIONS
1. Perform a multiplicative decomposition of the 3. Compute the residual autocorrelations. Examin-
Surtido Cookie sales data, store the residuals, ing the residual autocorrelations and the fore-
and generate forecasts of sales for the remain- casts of sales for the remainder of 2003, should
ing months of 2003. Jame change his thinking about the value of
2. What did Jame learn about the trend in sales? decomposition analysis? Explain.
What did the seasonal indexes tell him?

CASE 5-8 SOUTHWEST MEDICAL CENTER

Mary Beasley and her goal of predicting the number she wanted to have hard evidence that some periods
of future total billable visits to Medical Oncology at of the year were busier than others. (Scheduling
Southwest Medical Center were introduced in Case doctors is never an easy task.) A colleague of Mary’s
4-7. In that case, Mary learned there was a seasonal suggested she consider decomposition analysis,
component in her data and considered using a technique with which Mary was not familiar.
Winters’ smoothing method to generate forecasts of However, since she had a statistical software pack-
future visits. Mary was not completely satisfied with age available that included decomposition, Mary
this analysis, since there appeared to be some signif- was willing to give it a try. She also realized she had
icant residual autocorrelations. Moreover, Mary was to understand decomposition well enough to sell the
interested in isolating the seasonal indexes because results to central administration if necessary.

QUESTIONS
1. Write a brief memo to Mary explaining decom- 3. Interpret the trend component for Mary. What
position of a time series. did she learn from the seasonal indexes?
2. Perform a multiplicative decomposition of total 4. Compute the residual autocorrelations. Given
billable visits, save the residuals, and generate the residual autocorrelations and the forecasts,
forecasts of visits for the next 12 months, using should Mary be pleased with the decomposition
February FY2003–04 as the forecast origin. analysis of total billable visits? Explain.

Minitab Applications

The problem. In Example 5.1, a trend equation was developed for annual registration
of new passenger cars in the United States from 1960 to 1992.
 CHAPTER 5 Time Series and Their Components 215
Minitab Solution

1. After the new passenger car registration data are entered into column C1 of the
worksheet, click on the following menus to run the trend analysis:
Stat>Time Series>Trend Analysis

2. The Trend Analysis dialog box appears.

a. The Variable is Cars.
b. Click on Linear for Model Type.
c. Click on Generate forecasts, and place a 1 in the Number of forecasts box in
order to forecast 1993.
d. Click on Options. In the space provided for a TITLE, place Linear Trend for
Annual Car Registrations.
e. Click on OK for the Options dialog box. Click on OK again, and the graph that
was shown in Figure 5-3 appears.

The problem. Table 5-1 was constructed to show the trend estimates and errors com-
puted for the new passenger car registration data (see p. 170).

Minitab Solution
1. Column C1 is labeled Year, C2 is labeled Y, C3 is labeled t, C4 is labeled Estimates,
and C5 is labeled Error. Clicking on the following menus creates the years:
Calc>Make Patterned Data>Simple Set of Numbers

2. The Simple Set of Numbers dialog box appears.

a. The following responses are given:
Store patterned data in: C1
From first value: 1960
To last value: 1992
In steps of: 1
b. Click on OK, and the years appear in C1.
c. The new passenger car registration data are entered into C2.
3. The time-coded data t are entered into C3, using the Simple Set of Numbers dialog
box.
4. The trend estimates are entered into C4 by clicking on the same menus for the
trend analysis used to solve Example 5.1, with one additional step.
5. Click on the Storage menu, and obtain the Trend Analysis-Storage dialog box.
a. Under Storage, click on Fits (trend line) and Residuals (detrended data).
b. Click on OK for this dialog box and then the Trend Analysis dialog box. The
trend estimates will appear in C4, and the errors (residuals) will appear in C5.

The problem. In Examples 5.3 and 5.4, Perkin Kendell, the analyst for the Coastal
Marine Corporation, wanted to forecast quarterly sales for 2007.

Minitab Solution

1. Enter the appropriate years into column C1, quarters into C2, and the data into
C3. To run a decomposition model, click on the following menus:
Stat>Time Series>Decomposition
216 CHAPTER 5 Time Series and Their Components

FIGURE 5-17 Minitab Decomposition Dialog Box

2. The Decomposition dialog box shown in Figure 5-17 appears.

a. The Variable is C3 or Sales.
b. Since the data are quarterly, the Seasonal length is 4.
c. The Model type is Multiplicative, and the Model Components are Trend plus
seasonal.
d. Click on Options. The initial seasonal period is 1. Click on OK.
e. Click on Generate forecasts, and enter 4 for the Number of forecasts.
3. Click on the Storage menu, and obtain the Decomposition-Storage dialog box
shown in Figure 5-18.
a. Under Storage, click on Trend line, Detrended data, Seasonals, and Seasonally
adjusted data.
b. Click on OK for both this dialog box and the Decomposition dialog box. Table
5-4 shows the trend estimates in C4 (labeled T), the detrended data in C5
(labeled SCI), the seasonals in C6 (labeled S), and the seasonally adjusted data
in C7 (labeled TCI).
4. Figures 5-10, 5-11, and 5-12 appear on the screen and can be printed one at a time
using the following menus:

File>Print Graph

5. After the graphs have been printed, click on

File>Print Session Window

and the forecasts shown in Figure 5-9 will print.

The CI, C, and I columns can be computed using Minitab, as the next steps demon-
strate.
 CHAPTER 5 Time Series and Their Components 217

FIGURE 5-18 Minitab Decomposition-Storage Dialog Box

6. Label column C8 CI. Using the Calc>Calculator menu in Minitab, generate col-
umn C8 by dividing column C7 by column C4, so CI = TCI>T.
7. Label column C9 C. Generate column C9 by taking a centered moving average of
order 3 of the values in column C8. Use the menu

Stat>Time Series>Moving Average

Be sure to check the Center the moving averages box.

8. Label column C10 I. Using the Calc>Calculator menu in Minitab, generate column
C10 by dividing column C8 by column C9, so I = CI>C. The columns required for
Table 5-4 are now complete.

Excel Applications

The problem. Figure 5-6 shows data and a graph for mutual fund salespeople (see
p. 173). An exponential trend model is needed to fit these data.
Excel Solution

1. Enter Y in A1 and the salespeople data in A2:A8.

2. Enter X in B1 and the X-coded variable (1, 2, 3, 4, 5, 6, 7) in B2:B8.
3. Enter Log Y in C1, and create the logarithms (to the base 10) of the Y variable by
entering the formula = LOG10(A2) in cell C2. Copy this formula to the rest of
the column.
4. Click on the following menus to compute the exponential trend:

Tools>Data Analysis

5. The Data Analysis dialog box appears. Under Analysis Tools, choose Regression
and click on OK. The Regression dialog box shown in Figure 5-19 appears.
218 CHAPTER 5 Time Series and Their Components

FIGURE 5-19 Excel Regression Dialog Box

FIGURE 5-20 Excel Regression Output for the Mutual Fund Salespeople
Example
 CHAPTER 5 Time Series and Their Components 219
a. Enter C1:C8 as the Input Y Range.
b. Enter B1:B8 as the Input X Range.
c. Select the Labels check box.
d. Enter Figure 5-20 as the name of the New Worksheet Ply. Click on OK.
Figure 5-20 represents Excel output for an exponential model of mutual fund
salespeople. The equation is

log TN = 1.00069 + 0.11834t

The antilogs of the regression coefficients in this equation are

b0 = antilog 1.00069 = 10.016

b1 = antilog 0.11834 = 1.313

Thus, the fitted exponential trend equation is

YN = 110.016211.3132t

References
Bell, W. R., and S. C. Hillmer. “Issues Involved with Program,” Journal of Business and Economic
the Seasonal Adjustment of Economic Time Statistics 16 (1998): 127–152.
Series,” Journal of Business and Economic Levenbach, H., and J. P. Cleary. Forecasting: Practice
Statistics 2 (1984): 291–320. and Process for Demand Management. Belmont,
Bowerman, B. L., R. T. O’connell, and A. B. Koehler. Calif.: Thomson Brooks/Cole, 2006.
Forecasting, Time Series and Regression, 4th ed. Makridakis, S., S. C. Wheelwright, and R. J.
Belmont, Calif.: Thomson Brooks/Cole, 2005. Hyndman. Forecasting Methods and Applications,
Diebold, F. X. Elements of Forecasting, 3rd ed. 3rd ed. New York: Wiley, 1998.
Cincinnati, Ohio: South-Western, 2004. Moore, G. H., and J. Shiskin. “Early Warning Signals
Findley, D. F., B. C. Monsell, W. R. Bell, M. C. Otto, for the Economy.” In Statistics: A Guide to
and B. Chen. “New Capabilities and Methods of Business and Economics, J. M. Tanur et al., eds.
the X-12-ARIMA Seasonal-Adjustment San Francisco: Holden-Day, 1976.
This page intentionally left blank
C H A P T E R

6 SIMPLE LINEAR
REGRESSION
In Chapter 2, the linear association between two numerical variables (correlation) was
discussed. Linear association implies a straight-line relationship, and we showed how
to fit a straight line to pairs of observations on the two variables using the method of
least squares. In this chapter, simple linear regression (straight line) models are
discussed in some detail. Once a linear relationship is established, knowledge of the
independent variable can be used to forecast the dependent variable.
To review the analysis of the relationship between two variables discussed in
Chapter 2, consider the following example.
Example 6.1
Suppose Mr. Bump observes the selling price and sales volume of milk gallons for 10 ran-
domly selected weeks. The data he has collected are presented in Table 6-1.
First, he constructs a scatter diagram of the data, shown in Figure 6-1. It appears from this
scatter diagram that a negative linear relationship exists between Y, the number of milk gal-
lons sold, and X, the price of each gallon. It seems that, as price goes up, volume goes down.
Bump now wishes to measure the degree of this apparent relationship by calculating the
sample correlation coefficient, r. Using Equation 2.10 and the calculations in Table 6-2, he finds

n©X Y - 1©X21©Y2
r =
2n©X 2 - 1©X22 2n©Y2 - 1©Y22
101149.32 - 114.4211122
=
210121.562 - 114.422 21011,4882 - 111222
- 119.8
= = - .86
138.7

TABLE 6-1 Data for Milk Gallons for Example 6.1

Weekly Sales Level, Y Selling Price
Week (1,000s of gallons) X ($)

1 10 1.30
2 6 2.00
3 5 1.70
4 12 1.50
5 10 1.60
6 15 1.20
7 5 1.60
8 12 1.40
9 17 1.00
10 20 1.10

221
222 CHAPTER 6 Simple Linear Regression

30

20

Gallons
10

0 X
1.00 1.50 2.00
Price (in dollars)

FIGURE 6-1 Scatter Diagram for Example 6.1

TABLE 6-2 Computations Needed for Example 6.1

Y X XY X2 Y2
n = 10 10 1.30 13.0 1.69 100
6 2.00 12.0 4.00 36
5 1.70 8.5 2.89 25
12 1.50 18.0 2.25 144
10 1.60 16.0 2.56 100
15 1.20 18.0 1.44 225
5 1.60 8.0 2.56 25
12 1.40 16.8 1.96 144
17 1.00 17.0 1.00 289
20 1.10 22.0 1.21 400
Totals 112 14.40 149.3 21.56 1,488

The sample correlation coefficient of -.86 indicates a fairly strong negative relation-
ship between Y and X—as the price of a gallon of milk goes up, the number of gallons sold
goes down.
The question that may occur next is, how much does the volume drop as price is
raised? This question suggests drawing a straight line through the data points displayed on
the scatter diagram. After this line has been drawn, the slope of the line will show the aver-
age decrease in volume Y for each dollar increase in price X.

REGRESSION LINE
Mr. Bump might visually draw a straight line through the data points, attempting to fit
the line to the points as closely as possible. However, someone else might draw a differ-
ent line. A better procedure is to find the best straight line using a criterion that, for a
given set of data, produces the same line regardless of the person doing the fitting. As
 CHAPTER 6 Simple Linear Regression 223
we pointed out in Chapter 2, one popular criterion for determining the best-fitting
straight line is the least squares criterion.

The line that best fits a collection of X–Y data points is the line that minimizes the
sum of the squared distances from the points to the line as measured in the verti-
cal, or Y, direction. This line is known as the least squares line or fitted regression
line, and its equation is called the fitted regression equation.

The fitted straight line is of the form YN = b0 + b1X . The first term, b0 , is the
Y-intercept, and the second term, b1, is the slope. Recall that the slope represents the
amount of change in Y when X increases by one unit. The immediate objective is to
determine the values of b0 and b1.
Recall from Chapter 2 (see Equation 2.11) that the method of least squares chooses
the values for b0 and b1 that minimize the sum of the squared errors (distances).

SSE = ©1Y - YN 22 = ©1Y - b0 - b1X22 (6.1)

Using a little calculus, specific algebraic expressions can be derived for the least
squares values. In particular,

n©XY - ©X©Y ©1X - X21Y - Y2

b1 = =
n©X - 1©X2
(6.2)
2 2
©1X - X22

©Y b1 ©X
b0 = - = Y - b1 X (6.3)
n n
As you might guess, the least squares slope coefficient is related to the sample
correlation coefficient. Specifically,

2©1Y - Y22
b1 = r (6.4)
2©1X - X22

So b1 and r are proportional to one another and have the same sign.
The differences between the Y values actually observed and the corresponding fit-
ted Y values, the YN ’s, are called the residuals. The residuals are the vertical distances
(positive or negative) of the data points from the least squares line.
We have the identity

Observation = Fit + Residual

or, in symbols,

Y = YN + 1Y - YN 2 (6.5)

In this context, the fit represents the overall pattern in the data, and the residuals
represent the deviations from this pattern. The split into fit plus residual applies
to patterns other than straight lines, and we will make use of it repeatedly in later
chapters.
224 CHAPTER 6 Simple Linear Regression

Example 6.2
The least squares coefficients for a straight line fit to Mr. Bump’s data (see Figure 6-1) are
easily computed using Equations 6.2 and 6.3 and the information in Table 6-2. We have

101149.32 - 114.4211122 -119.8

b1 = = = - 14.54
10121.562 - 114.422 8.24

- 1 -14.542
112 14.4
b0 = = 11.2 + 14.5411.442 = 32.14
10 10

The fitted least squares line has the equation

YN = b0 + b1X
(6.6)
YN = 32.14 - 14.54X

This equation is called the fitted regression equation.

Mr. Bump may now wish to interpret the values in this equation. The Y-intercept, b0, is
the value Y when X is equal to zero. A strict interpretation would suggest that the average
number of gallons sold when X = 0 (that is, when the price of a gallon of milk is zero) is
32,140 gallons. This interpretation does not agree with common sense, since one would
expect more milk to be “sold” if it were free. The problem illustrated here involves predict-
ing a value for Y based on an X value about which no sample evidence has been collected.
That is, none of the sample points has an X value at or near zero. In this case, as in many
regression analysis cases, a useful interpretation of the Y-intercept is not possible.
In more general terms, it is often not wise to predict Y values for any X beyond the
range of the X’s collected in the sample data. A regression function should be interpreted as
a useful approximation to real-world behavior over the region for which there are data.
Extrapolation beyond the data requires the strong assumption that the nature of the rela-
tion between Y and X does not change.1
The slope coefficient, b1, may be interpreted as the average change in Y that occurs
when X increases by one unit. In this example, Y decreases by an average of 14.54 (that is,
14,540 fewer gallons are sold) when X increases by 1 (that is, the cost of a gallon of milk is
raised by $1). Each dollar increase in the price of a gallon of milk reduces the quantity pur-
chased by an average of 14,540 gallons. To express this statement in more meaningful units,
the sample evidence indicates that each increase of 1 cent in the price of a gallon of milk
reduces the quantity purchased by an average of 145.4 gallons.
The X–Y relationship can be illustrated by drawing the straight line that best fits the
data points on the scatter diagram. The result is shown in Figure 6-2.
Notice that the vertical distances from the points to the line have been shown as dotted
lines. If these distances were squared and added, the resulting sum would be smaller than
that computed for any other line that could be drawn through the points.2 In accordance
with the least squares procedure then, this line represents the best possible fit to the
10 sample data points.

Statistical ideas must be introduced in the study of the relation between two vari-
ables when the points in the scatter plot do not lie exactly on a line, as in Figure 6-2.
We think of the data in the scatter plot as a sample of observations on an underlying
relation that holds in the population of X–Y values. The statistical model for simple
linear regression assumes that for all the values of X, the observed values of the
dependent variable, Y, are normally distributed about a mean, my = b 0 + b 1X, that
depends linearly on X. That is, as X changes, the means of the distributions of

1We encountered a similar situation when we discussed the usefulness of extrapolating trend curves to fore-
cast future values of a time series in Chapter 5.
2For Bump’s data, this smallest sum of squared distances is SSE = 59.14.
 CHAPTER 6 Simple Linear Regression 225

b0 = 32.14

30 ^
Y = 32.14 − 14.54X

20

Gallons
10

0 X
1.00 2.00
Price (in dollars)

FIGURE 6-2 Fitted Regression Line for Mr. Bump’s Data for Example 6.2

y = 0 + 1X

FIGURE 6-3 The Statistical Model for Simple Linear Regression

possible Y values lie along a straight line. This line is known as the population regres-
sion line. Observed Y’s will vary about these means because of the influence of
unmeasured factors. The model assumes that this variation, measured by the stan-
dard deviation, s, is the same for all values of X. Finally, the deviation (distance)
between a Y value and its mean is known as error and is represented by , the Greek
letter epsilon.
In the simple linear regression model, the response, or dependent, variable Y is the
sum of its mean and a random deviation ( ) from the mean. The deviations ( ) repre-
sent variation in Y due to other unobserved factors that prevent the X–Y values from
lying exactly on a straight line in the scatter plot.
The statistical model for simple linear regression is illustrated schematically in
Figure 6-3.
226 CHAPTER 6 Simple Linear Regression

Statistical Model for Straight-Line Regression

The response, or dependent, variable Y is related to the controlled, or indepen-
dent, variable X by
Y = b 0 + b 1X +
Here, b 0 + b 1X is the mean response for a given X. The deviations ( ) are
assumed to be independent and normally distributed, with mean 0 and standard
deviation s. The unknown constants are b 0, b 1, and s.

The fitted (sample) regression line, YN = b0 + b1X, may be regarded as an estimate

of the population regression line, my = b 0 + b 1X, and the residuals, e = Y - YN ,
may be regarded as estimates of the error components, . This implies the following
correspondence:
Population: Y = b 0 + b 1X +
Sample: Y = b0 + b1X + e
Most computer software packages perform correlation and regression analysis.
The instructions for using Excel to run Examples 6.1 and 6.2 are presented in the Excel
Applications section at the end of this chapter.

STANDARD ERROR OF THE ESTIMATE

Having computed the fitted straight line, Mr. Bump might next be interested in mea-
suring the extent to which the sample data points are spread around the fitted regres-
sion function. A measure of dispersion analogous to the sample standard deviation can
be developed. This measure, called the standard error of the estimate, measures the
spread of the data points about the fitted line in the Y direction. The standard error of
the estimate is denoted by sy # x and is given by

©1Y - YN 22
sy # x = (6.7)
D n - 2

The standard error of the estimate measures the amount by which the actual Y
values differ from the estimated, or YN , values. For relatively large samples, we
would expect about 67% of the differences Y - YN to be within sy # x of 0 and about
95% of these differences to be within 2 sy # x of 0.

The standard error of the estimate is similar to the sample standard deviation
introduced in Chapter 2. It can be used to estimate a population standard deviation. In
fact, sy # x estimates the standard deviation, s, of the error term, , in the statistical model
for simple linear regression. Equivalently, sy # x estimates the common standard devia-
tion, s, of the normal distribution of Y values about the population regression line,
my = b 0 + b 1X, at each X (see Figure 6-3).
 CHAPTER 6 Simple Linear Regression 227
A regression analysis with a small standard error of the estimate means that all the
data points lie very close to the fitted regression line.3 If the standard error of the esti-
mate is large, the data points are widely dispersed about the fitted line.
For computational purposes, Equation 6.7 can be converted to

©Y2 - b0 ©Y - b1 ©XY
sy # x = (6.8)
D n - 2
For the Bump example, the standard error of the estimate is

1,488 - 132.14211122 - 1- 14.5421149.32

sy # x =
D 10 - 2
59.14
= = 17.39 = 2.72
A 8
With YN values in the 3–18 range (see Figure 6-2), sy # x is moderately large and indicates
that an appreciable amount of the variation in Y (gallons sold) is not explained by X
(price). We investigate this assertion further in the section on Coefficient of
Determination.

FORECASTING Y
Next, the fitted regression line can be used to estimate the value of Y for a given value
of X. To obtain a point forecast or forecast for a given value of X, simply evaluate the
estimated regression function at X.
Example 6.3
Suppose Mr. Bump wished to forecast the quantity of milk sold if the price were set at $1.63.
From Equation 6.6, the forecast is

YN = 32.14 - 14.54X
YN = 32.14 - 14.5411.632 = 8.440

or 8,440 gallons. Note that this forecast is a value of YN ; that is, the forecast is the Y coordi-
nate of the point on the fitted regression line where X = 1.63.

Of course, Mr. Bump realizes that the actual Y values corresponding to settings of
the X’s are unlikely to lie exactly on the regression line. In fact, they will be spread
about the line as measured by sy # x. Moreover, the sample (fitted) regression line is an
estimate of the population regression line based on a sample of 10 data points. Other
random samples of 10 would produce different fitted regression lines, similar to the case
in which many samples drawn from the same population have different sample means.
There are two sources of uncertainty then associated with a point forecast gener-
ated by the fitted regression equation:

1. Uncertainty due to the dispersion of the data points about the sample regression
line
2. Uncertainty due to the dispersion of the sample regression line about the popula-
tion regression line

3If all the data points lie exactly on the fitted line, Y = YN for all X and sy # x = 0.
228 CHAPTER 6 Simple Linear Regression

An interval forecast of Y can be constructed that takes these two sources of uncertainty
into account.

The standard error of the forecast, sf, measures the variability of the predicted Y
about the actual Y for a given value of X.

The standard error of the forecast is

1 1X - X22
sf = s 2y # x + s 2y # x ¢ + ≤
D n ©1X - X22

1 1X - X22
sf = sy # x 1 + + (6.9)
D n ©1X - X22
The first term under the first radical sign in Equation 6.9, s2y # x, measures the dispersion of
the data points about the sample regression line (first source of uncertainty). The second
term under the radical sign measures the dispersion of the sample regression line about
the population regression line (second source of uncertainty). Notice that the standard
error of the forecast depends on X, the value of X for which a forecast of Y is desired.Also,
notice that sf is smallest when X = X , since then the numerator in the third term under the
radical in Equation 6.9 (bottom) will be 1X - X22 = 0.4 All things equal, the farther X is
from X , the larger the standard error of the forecast is.
If the statistical model for simple linear regression is appropriate, a prediction
interval for Y is given by

YN ; t sf (6.10)

where t is a percentage point of Student’s t distribution with df = n - 2. If the sample

size is large (n Ú 30), the t percentage point can be replaced with the corresponding
percentage point Z of the standard normal distribution. For example, a large-sample
95% prediction interval for Y is essentially

YN ; 2 sf (6.11)

Example 6.4
Pictorially, Mr. Bump’s 95% prediction interval for Y at various values of X would look as
pictured in Figure 6-4.
Using the results in Table 6-3 and Equation 6.9 with X = 1.44 , the standard error of the
forecast at X = 1.63 is

1 11.63 - 1.4422
sf = 2.72 1 + + = 2.7211.0692 = 2.91
D 10 .824

From Example 6.3, YN = 8.440 when X = 1.63, and using Equation 6.10, a 95% predic-
tion interval for Y is

YN ; t sf = 8.44 ; 2.30612.912 = 8.44 ; 6.71

4For the choice X = X, the predicted Y is YN = Y .

 CHAPTER 6 Simple Linear Regression 229

Y
Sample
Regression
Line

Y
95% Prediction
Interval for Y

X
X

FIGURE 6-4 Prediction Interval for Mr. Bump’s Data for Example 6.4

TABLE 6-3 Calculation of ©1X - X22

for Example 6.4
X 1X - X22
1.30 .0196
2.00 .3136
1.70 .0676
1.50 .0036
1.60 .0256
1.20 .0576
1.60 .0256
1.40 .0016
1.00 .1936
1.10 .1156
©1X - X 22 = .8240

or (1.73, 15.15); that is, 1,730 gallons to 15,150 gallons. Here, 2.306 = t.025 is the upper 2.5%
point of a t distribution with df = 8.
The prediction interval is very wide—so wide as to be virtually worthless in forecasting
Y—because of the small sample size and the relatively large value for sf. The amount of
uncertainty reflected by the wide prediction interval is not apparent in the point forecast
computed from the fitted regression function.The major advantage of the interval estimate is
that it explicitly accounts for the uncertainty associated with the forecast.

In general, it is dangerous to use the fitted regression function to predict Y values

beyond the range of the available data. Mr. Bump is justified in trying to forecast Y
when X = 1.63 because some of the original X values are near 1.63. On the other
hand, it may not be wise to forecast Y when X = 3.00. No data have been collected for
X values this large, and for this reason, any prediction involving such an X value would
be highly suspect.5 To estimate the quantity of milk sold when the price per gallon is
$3.00, Mr. Bump has to assume the straight-line model is still valid. He may have good
reason to make this assumption, but he has no direct evidence to support it.

5Moreover, the standard error of the forecast would be large, since the quantity 1X - X 22 would be rela-
tively large.
230 CHAPTER 6 Simple Linear Regression

It is useful to end this section by reviewing the assumptions underlying the statisti-
cal model for linear regression.

1. For a given value of X, the population of Y values is normally distributed about the
population regression line. This condition is shown in Figure 6-3. In practice, rea-
sonably accurate results are obtained as long as the Y values are approximately
normally distributed.
2. The dispersion of population data points around the population regression line
remains constant everywhere along the line. That is, the population variance does
not become larger or smaller as the X values of the data points increase. A viola-
tion of this assumption is called heteroscedasticity; an example of this condition
and its cure appear in Chapter 8.
3. The error terms ( ) are independent of each other. This assumption implies a ran-
dom sample of X–Y data points. When the X–Y data points are recorded over
time, this assumption is often violated. Rather than being independent, consecu-
tive observations are serially correlated. Methods for handling the serial correla-
tion problem are considered in Chapter 8.
4. A linear relationship exists between X and Y in the population. There are extensions
of simple linear regression for dealing with X–Y relationships that are nonlinear;
some of these are discussed later in this chapter.

DECOMPOSITION OF VARIANCE
From Equation 6.5
Y = YN + 1Y - YN 2
or
Y = b0 + b1X + 1Y - b0 - b1X2

Observed Y value Explained by Residual or

linear relation deviation from
linear relation
In an ideal situation in which all the points lie exactly on the fitted line, the residuals
are all zero, and the Y values are completely accounted for or explained by the linear
relation with X.
Subtracting Y from both sides of the previous expression,
1Y - Y2 = 1YN - Y 2 + 1Y - YN 2
It can be shown with a little algebra that the sums of squares add:
©1Y - Y22 = ©1YN - Y 22 + ©1Y - YN 22 (6.12)
or
SST = SSR + SSE
where
SST = ©1Y - Y22
SSR = ©1YN - Y 22
SSE = ©1Y - YN 22
 CHAPTER 6 Simple Linear Regression 231
Here, SS stands for sum of squares, and T, R, and E represent total, regression, and
error, respectively. These sums of squares have associated degrees of freedom:

df1SST2 = n - 1
df1SSR2 = 1
df1SSE2 = n - 2

Corresponding to the sums of squares, the degrees of freedom are related by

n - 1 = 1 + 1n - 22 (6.13)

If there is no linear relation, Y does not depend on X, and the variation in Y is

described by the sample variance:

1
s 2y = ©1Y - Y22
n - 1

If, on the other hand, Y is related to X, some of the differences in the Y values are due
to this relationship.
The regression sum of squares, SSR, measures that part of the variation in Y
explained by the linear relation. The error sum of squares, SSE, is the remaining varia-
tion in Y, or the variation in Y not explained by the linear relation.

Decomposition of Variability

SST  SSR  SSE

Total Variability Residual or
variability of Y explained by unexplained
linear relation variability

The sums of squares associated with the decomposition of the variability of Y and their
corresponding degrees of freedom can be set out as shown in Table 6-4, which is known
as an analysis of variance or ANOVA table.
The final column in the ANOVA table is the mean square column. The mean
square regression, MSR, is the regression sum of squares divided by its degrees of free-
dom. Similarly, the mean square error, MSE, is the error sum of squares divided by its
degrees of freedom.

TABLE 6-4 ANOVA Table for

Straight-Line Regression
Sum of
Source Squares df Mean Square
Regression SSR 1 MSR = SSR>1
Error SSE n -2 MSE = SSE>1n -22
Total SST n -1
232 CHAPTER 6 Simple Linear Regression

TABLE 6-5 Residuals for Mr. Bump’s Data with

Predictor Y for Example 6.5

Actual Y Predicted Y ( Y ) Residual ( Y - Y) 1Y - Y 22

10 11.2 -1.2 1.44
6 11.2 -5.2 27.04
5 11.2 -6.2 38.44
12 11.2 .8 .64
10 11.2 -1.2 1.44
15 11.2 3.8 14.44
5 11.2 -6.2 38.44
12 11.2 .8 .64
17 11.2 5.8 33.64
20 11.2 8.8 77.44
Totals 0.0 233.60

Notice from Equation 6.7 that the mean square error is

SSE ©1Y - YN 22
MSE = = = s 2y #x
n - 2 n - 2
the square of the standard error of the estimate. The mean square ratios will be used
for another purpose later in this chapter.
Example 6.5
Mr. Bump’s analysis began with knowledge of only 10 weekly sales volume quantities (the
Y variable). If no further information were available, Mr. Bump might employ the sample
average, Y = 11.2, as a predictor of gallons of milk sold each week. The errors, or residu-
als, associated with this forecast are Y - Y , and the sum of the squared errors is
©1Y - Y22. Notice that this latter quantity, ©1Y - Y22, is exactly SST, the total sum of
squares introduced in Equation 6.12. Thus, SST measures the variability of Y about a pre-
dictor that uses only the Y values in its calculation.6 The predictions (Y ), the residuals
(Y - Y ), and the total sum of squares (SST = ©1Y - Y22) are shown in Table 6-5.7
Mr. Bump also has information about a variable X, the price per gallon of milk, that is
related to Y, the weekly volume of milk sold. (Recall from Example 6.1 that r = - .86). With
this additional variable, he expects to be able to explain some of the variation (differences)
in the Y values beyond that explained by the predictor, Y .
From Example 6.2, the line fit to the scatter plot of X–Y observations has the equation
YN = 32.14 - 14.54X. A table similar to Table 6-5 can be constructed if YN is used to predict Y.
The result is Table 6-6.8
A comparison of Tables 6-5 and 6-6 shows that the use of YN as a predictor of Y has
resulted in generally smaller residuals (in absolute value) and a considerably smaller residual
(error) sum of squares than when Y is used as a predictor. Use of the related variable X
reduces the prediction, or forecast, errors. That is, knowledge of X helps to further explain
the differences in the Y’s. How much does X help? The decomposition of variability pro-
vides an answer to this question.

6If the analysis were to stop at this point, the variability in Y would be measured by the sample variance,
s2y = ©1Y - Y 22>1n - 12 , rather than SST = ©1Y - Y 22 . The sample variance is the usual measure of
variability for measurements on a single variable.
7The residuals, Y - Y , always sum to zero because the average Y is the mathematical center of the Y values.

N 2 is always zero.
8If an intercept term is included in the regression equation, the sum of the residuals ©1Y - Y
 CHAPTER 6 Simple Linear Regression 233

TABLE 6-6 Residuals for Mr. Bump’s Data with

Predictor YN for Example 6.5

Predicted Y(YN ) Using Residual

X Y YN = 32.14 - 14.54X (Y - YN ) (Y - YN 22

1.30 10 13.238 -3.238 10.48

2.00 6 3.060 2.940 8.64
1.70 5 7.422 -2.422 5.87
1.50 12 10.330 1.670 2.79
1.60 10 8.876 1.124 1.26
1.20 15 14.692 .308 .09
1.60 5 8.876 -3.876 15.02
1.40 12 11.784 .216 .05
1.00 17 17.600 -.600 .36
1.10 20 16.146 3.854 14.85
Totals 0.000 59.41

Using Tables 6-5 and 6-6 and Equation 6.12,

SST = ©1Y - Y 22 = 233.60

SSE = ©1Y - YN 22 = 59.41
and, consequently,

SSR = ©1YN - Y 22 = 233.60 - 59.41 = 174.19

The decomposition of variability is

SST  SSR  SSE

233.60  174.19  59.41
Total Explained Unexplained
variation variation variation

Of the variability remaining after predicting Y with Y , Mr. Bump sees that a proportion

SSR 174.19
= = .75
SST 233.60
of it has been explained by the relationship of Y with X. A proportion, 1 - .75 = .25, of the
variation in Y about Y remains unexplained. From this perspective, knowledge of the
related variable X results in better predictions of Y than can be obtained from Y , a quantity
that doesn’t depend on X.
The decomposition of variability for Mr. Bump’s data can be set out in an analysis of
variance table as the next example demonstrates.

Example 6.6
Mr. Bump constructs an ANOVA table for his data using the format in Table 6-4 and the
sums of squares calculations from Example 6.5, as shown in Table 6-7.
The decomposition of variability is clearly shown in the sum of squares column. Notice
that, within rounding error, MSE = 7.43 = 12.7222 = s2y # x.
234 CHAPTER 6 Simple Linear Regression

TABLE 6-7 ANOVA Table for Mr. Bump’s Data

for Example 6.6
Source Sum of Squares df Mean Square
Regression 174.19 1 174.19
Error 59.41 8 7.43
Total 233.60 9

COEFFICIENT OF DETERMINATION
The identity

1Y - Y2 = 1YN - Y 2 + 1Y - YN 2

leading to the decomposition of variance given by Equation 6.12 is shown graphically

for Mr. Bump’s fitted regression line and a hypothetical data point in Figure 6-5.
If Y did not depend on X, Mr. Bump would expect Y to be close to Y and the devi-
ation Y - Y to simply reflect random variation. However, Y does depend on X in a
manner suggested by the fitted regression function. In the figure, the observed X value
is greater than X , and it is known that X and Y have a fairly strong negative correlation
(r = - .86). Of the total vertical distance Y - Y , the amount YN - Y is therefore
“explained” by the movement in X, whereas the remaining vertical distance, Y - YN , is
“unexplained” by the movement in X.
As indicated in the previous section, SST measures the total variation about Y, and
that part of the total that is explained by movement in X is SSR. The remaining or
unexplained variation is SSE. The ratio of the explained to total variation is called the
sample coefficient of determination and is denoted by r 2.

Y
32.14 = b0
30 Bump's
Regression Line:
^
Y = 32.14 − 14.54X
Gallons

20

11.2 = Y
10
Y − Y (Total) Y^ − Y (Explained by X )
^
Y − Y (Unexplained by X )
0 X
1.00 X = 1.44 2.00
Price (in dollars)

FIGURE 6-5 Explained and Unexplained Variation in Mr. Bump’s Data

 CHAPTER 6 Simple Linear Regression 235

Explained variation SSR ©1YN - Y22

r2 = = =
Total variation SST ©1Y - Y22

Unexplained variation SSE ©1Y - YN 22

= 1 - = 1 - = 1 - (6.14)
Total variation SST ©1Y - Y22

The coefficient of determination measures the percentage of variability in Y that

can be explained through knowledge of the variability (differences) in the
independent variable X.

Example 6.7
The coefficient of determination, r 2, for Mr. Bump’s data was calculated in Example 6.5,
although, at that point, it was not labeled as such. The coefficient of determination is also
readily available for the ANOVA table, Table 6-7, in Example 6.6. Recall that
SST = ©1Y - Y22 = 233.60
SSR = ©1YN - Y22 = 174.19
SSE = ©1Y - YN 22 = 59.41

and
174.19
r2 = = .746
233.60
Alternatively, r 2 may be calculated as
59.41
r2 = 1 - = 1 - .254 = .746
233.60
About 75% of the variability in gallons of milk sold (Y) can be explained by differences in
price per gallon (X ). About 25% of the variability in quantity of milk sold cannot be
explained by price. This portion of the variability must be explained by factors that have not
been considered in this regression analysis (for example, amount of advertising, availability
of substitute products, quality of milk).

For straight-line regression, r 2 is the square of r, the coefficient of correlation.

Coefficient of determination = 1Coefficient of correlation22
r 2 = 1r22
Thus, for Mr. Bump’s data and ignoring rounding error,
.746 = 1 - .86322
Why is it necessary to specifically identify both r and r 2 in a regression analysis? The
answer is that they convey different information.
The correlation coefficient reveals the strength and the direction of the linear rela-
tion. In the case of the data that Mr. Bump has collected, a negative relationship exists
(r = - .86). In other cases, a positive relationship might be indicated by the value for r.
As will be seen in the next chapter, it is useful to identify the nature of the relationships
that exist between certain pairs of variables when dealing with a large collection of
variables. Note that, when the coefficient of correlation is squared, the value is always
positive, so the nature of the relationship is lost.
The coefficient of determination, r 2, measures the strength of the relationship
between Y and X in a different way than r does. The value for r 2 measures the
236 CHAPTER 6 Simple Linear Regression

Y Y ^
Y

^
Y = Y Y

X X
^ ^
r2 = 1 − Σ(Y − Y )2 / Σ (Y − Y )2 r2 = 1 − Σ(Y − Y )2 / Σ (Y − Y )2
=1−1=0 =1−0=1
(a) No Linear Correlation (b) Perfect Linear Correlation

FIGURE 6-6 Extreme Values for r 2

percentage of the variability in Y that is explained by differences in X. This useful inter-

pretation can be generalized to relationships between Y and more than one X.
Figure 6-6 illustrates the two extreme cases for r2, r2 = 0 and r2 = 1. In the former
case, none of variability in Y is explained by X: The scatter diagram suggests no linear
relationship between X and Y. When r 2 = 1, all of the variability in Y is explained
when X is known: The sample data points all lie on the fitted regression line.

HYPOTHESIS TESTING
The fitted regression line is produced by a sample of X–Y values. The statistical model
for simple linear regression suggests that the straight-line relationship between Y and
X holds for all choices of X–Y pairs. That is, there is a true relation between X and Y of
the form my = b 0 + b 1X. Given the sample evidence, can we conclude that the true
relation holds for all X and Y?
Consider the hypothesis
H0 : b 1 = 0
where b 1 is the slope of the population regression line. Notice that, if this hypothesis is
true, there is no relation between Y and X in the population. Failing to reject H0 means
that, in spite of the fact that the sample has produced a fitted line with a nonzero value
for b1, we must conclude that there is not sufficient evidence to indicate Y is related to
X. That is, we cannot rule out the possibility that the population regression line is flat
(horizontal).9
How could b 1 be zero while b1 is nonzero? Consider Figure 6-7 where a population
of data points is shown from which a sample of five has been selected (sampled data
points are indicated by ). As this scatter diagram suggests, if enough sample data points
are selected, it will become obvious that the population of data points has a regression
line with a zero slope. However, the five randomly selected data points lie fairly close to
an upward-trending regression line. It might be erroneously concluded from this evi-
dence that X and Y are related in a positive linear way. However, if the hypothesis
b 1 = 0 is tested with the sample data, the forecaster will probably not be able to reject it.

9A flat population regression line (i.e., b = 0) is also equivalent to the statement H : r = 0 where ρ is the
1 0
population correlation coefficient.
 CHAPTER 6 Simple Linear Regression 237

Y
b1 ≠ 0
Sample Regression Line

1 = 0
True Regression Line

FIGURE 6-7 Population and Sample Data Points

b
If H0 : b 1 = 0 is true, the test statistic t with value t = sb1 has a t distribution with
1

df = n - 2. Here, sb1 is the estimated standard deviation (or standard error) of b1,
given by sb1 = sy # x> 2©1X - X22.

This result provides a way to test H0 : b 1 = 0, as the next example illustrates.

Example 6.8
Mr. Bump would like to test

H0 : b 1 = 0 1no linear relation2

versus

H1 : b 1 Z 0 1a linear relation with a nonzero slope2

He computes10

sb1 = sy # x> 2©1X - X22 = 2.72> 2.824 = 3.00

and forms the ratio (test statistic)

b1 - 14.54
t = = = - 4.8
sb1 3.00
Is the value t = - 4.8 an unusual result if H0 is true? Mr. Bump checks the t table with
n - 2 = 8 degrees of freedom and finds

t.005 = 3.355
t.025 = 2.306

Since ƒ t ƒ = 4.8 7 3.355, Mr. Bump rejects H0 at the 1% level of significance. He concludes
that his regression is significant since t = - 4.8 is highly unlikely if H0 is true. A t value of
this magnitude would occur less than one time in a 100 if there were no linear relation
between Y and X.

10The values used in the following calculation were computed earlier in this chapter.
238 CHAPTER 6 Simple Linear Regression

For very large sample sizes, it is possible to reject H0 and conclude there is a linear
relation between Y and X even though r 2 may be small—say, 10%. Similarly, for small
samples and a very large r2—say, 95%—it is possible to conclude that the regression is
significant.A small r 2 means the fitted regression equation is unlikely to have much pre-
dictive power. On the other hand, a large r 2 with a very small sample size may leave the
analyst uncomfortable and require more sample evidence before the fitted function is
used for forecasting. There can be a difference between statistical significance and prac-
tical significance. At times, judgment coupled with subject matter knowledge is neces-
sary to determine if a fitted regression function is likely to be a useful forecasting tool.
An alternative test of H0 : b 1 = 0 is available from the ANOVA table.
If the assumptions of the statistical model for linear regression are appropriate and
if the null hypothesis H0 : b 1 = 0 is true, the ratio

Regression mean square MSR

F = = (6.15)
Error mean square MSE

has an F distribution with df = 1, n - 2. When H0 is true, MSR and MSE are each
estimators of s 2, the variance of the error term ( ) in the statistical model for straight-
line regression. On the other hand, if H0 : b 1 Z 0 is true, the numerator in the F ratio
tends to be larger than the denominator. Large F ratios then are consistent with the
alternative hypothesis.

In the straight-line regression model, the test of the hypotheses H0 : b 1 = 0

versus H0 : b 1 Z 0 can be based on the F ratio F = MSR
MSE with df = 1, n - 2. At
level a, the rejection region is F 7 Fa.

As discussed in the next chapter, the F test can be extended to check the signifi-
cance of regression models with more than one independent variable.
Example 6.9
Table 6-7 is the ANOVA table for Mr. Bump’s data. From this table,
MSR 174.19
F = = = 23.4
MSE 7.43
and with 1 = 1 and 2 = 8 degrees of freedom,

F.05 = 5.32
F.01 = 11.26

Since F = 23.4 7 11.26, H0 : b 1 = 0 is rejected at the 1% level. The regression is significant.

It is no accident that the results of the F test are consistent with the results of the t test
in Example 6.8. In fact, within rounding error,

F = 23.4 = 1-4.822 = t2

Moreover,

F.01 = 11.26 = 13.35522 = 1t.00522

Thus, for a given significance level, the t test rejects H0 : b 1 = 0 whenever the F test
rejects and vice versa. This relationship between the t and F tests holds only for the straight-
line regression model.
 CHAPTER 6 Simple Linear Regression 239

The F statistic can also be expressed in terms of the coefficient of determination, r 2.

r 21n - 22
F = (6.16)
1 - r2
Everything else being equal, the value of F increases as r 2 increases. As more of the
variation in Y is explained by the fitted regression function, it becomes more likely that
the regression will be judged significant by the F test.

ANALYSIS OF RESIDUALS
Fitting a model by least squares, constructing prediction intervals, and testing
hypotheses do not complete a regression study. These steps are only half the story: the
inferences that can be made when the assumed model is adequate. In most studies, it is
not obvious that a particular model is correct.
Inferences can be seriously misleading if the assumptions made in the model
formulation are grossly incompatible with the data. It is essential to check the data
carefully for indications of any violation of the assumptions. Recall that the assump-
tions for the straight-line regression model are as follows:
1. The underlying relation is linear.
2. The errors are independent.
3. The errors have constant variance.
4. The errors are normally distributed.
The information on variation that cannot be explained by the fitted regression
function is contained in the residuals, e = Y - YN . To check the merits of a tentative
model, we can examine the residuals by plotting them in various ways:
1. Plot a histogram of the residuals.
2. Plot the residuals against the fitted values.
3. Plot the residuals against the explanatory variable.
4. Plot the residuals over time if the data are chronological.
A histogram of the residuals provides a check on the normality assumption.
Typically, moderate departures from a bell-shaped curve do not impair the conclusions
from tests or prediction intervals based on the t distribution, particularly if the data set
is large. A violation of the normality assumption alone is ordinarily not as serious as a
violation of any of the other assumptions.
If a plot of the residuals against the fitted values indicates that the general nature
of the relationship between Y and X forms a curve rather than a straight line, a suitable
transformation of the data may reduce a nonlinear relation to one that is approxi-
mately linear. Variable transformations are considered in a later section of this chapter.
A transformation may also help to stabilize the variance. Figure 6-8 shows a resid-
ual plot that indicates the spread of the residuals increases as the magnitude of the
fitted values increases. That is, the variability of the data points about the least squares
line is larger for large responses than it is for small responses. This implies that the
constant variance assumption may not hold. In this situation, relating the logarithm of
Y to X may produce residual variation that is more consistent with a constant variance.
The assumption of independence is the most critical. Lack of independence can
drastically distort the conclusions drawn from t tests. The independence assumption is
particularly risky for time series data—the data frequently arising in business and
economic forecasting problems.
240 CHAPTER 6 Simple Linear Regression

^
e=Y−Y

^
Y

FIGURE 6-8 A Residual Plot Showing Increasing Spread

For time series residuals—that is, for residuals produced using regression meth-
ods on time-ordered data—independence can be checked with a plot of the residuals
over time. There should be no systematic pattern, such as a string of high values
followed by a string of low values. In addition, the sample autocorrelations of the
residuals

a et et - k
rk1e2 =
t=k+1
n k = 1, 2, Á , K (6.17)
2
a et
t=1

where n is the number of residuals and K is typically n/4, should all be small.
Specifically, independence is indicated if the residual autocorrelation coefficients are
each in the interval 0 ; 2> 1n for all time lags k.

Example 6.10
Using the fitted values and residuals shown in Table 6-6, Mr. Bump constructed a histogram
of the residuals and a plot of the residuals against the fitted values. These and other results
are displayed in Figure 6-9.
The histogram is centered at zero, and although it is symmetric, it doesn’t appear to be
bell-shaped. However, with only 10 observations, a histogram such as the one pictured in
Figure 6-9 is not unusual for normally distributed data. The normal assumption appears to
be reasonable.
The points in the normal probability plot in Figure 6-9 lie nearly along a straight line.
As we pointed out in Chapter 2, this straight line–like behavior suggests a good fit between
the data (in this case, the residuals) and a normal distribution. The normal probability plot
suggests there is no reason to doubt the normal assumption.
The second plot in the first row of Figure 6-9 also looks good. When the residuals are
plotted against the fitted values, the spread about zero in the vertical direction should be
about the same for all values along the horizontal axis. That is, the magnitudes of the resid-
uals for small fitted values should be about the same as the magnitudes of the residuals for
intermediate fitted values and about the same as the magnitudes of the residuals for large
fitted values. This ideal behavior suggests two things: (1) The underlying relation between Y
and X is linear, and (2) the error variability is constant (Y’s at different values of X have the
same spread about the regression line).
Mr. Bump is pleased that the plot of the residuals versus the fitted values is not
“bowed”—for example, a string of positive residuals followed by a string of negative resid-
uals followed by a string of positive residuals. This behavior would suggest a nonlinear
 CHAPTER 6 Simple Linear Regression 241

FIGURE 6-9 Residual Plots for Mr. Bump’s Data for Example 6.10

relation between Y and X. He is also pleased his plot did not have the cone-shaped appear-
ance of the plot in Figure 6-8, which indicates nonconstant (increasing) variability.
Although the Y’s are weekly sales of milk, the weeks were selected at random and were
not ordered in time. Consequently, plotting the residuals over time or computing the resid-
ual autocorrelations was not appropriate.
Mr. Bump is satisfied with his residual analysis. He feels his straight-line regression
model adequately describes the relation between weekly sales volume and price.

COMPUTER OUTPUT
Mr. Bump’s regression analysis problem (using the data from Table 6-1) is run on
Minitab (see the Minitab Applications section at the end of this chapter for instruc-
tions), and the output is presented in Table 6-8.
To explain the terminology used in the computer output, definitions and computa-
tions are presented in the following list. These definitions and calculations are keyed to
Table 6-8.

1. Correlation = - .86. The sample correlation coefficient (r) indicates the relation-
ship between X and Y, or price and sales, respectively.
2. Regression coefficient 1Coef2 = - 14.54. This value (b1) is the change in Y (sales)
when X (price) increases by one unit. When price increases $1, estimated sales
decrease by 14,539 units.
242 CHAPTER 6 Simple Linear Regression

TABLE 6-8 Minitab Output for Mr. Bump’s Data

Correlations: Sales, Price

Pearson correlation of Sales (Y) and Price 1X2 = - 0.863 (1)
Regression Analysis: Sales versus Price
The regression equation is
Sales 1Y2 = 32.1 - 14.5 Price (X) (5) and (2)
Predictor Coef SE Coef T P
Constant 32.136 (5) 4.409 7.29 0.000
Price (X) -14.539 (2) 3.002 (3) - 4.84 (4) 0.001 (7)
S = 2.72545 (6) R -Sq = 74.6% (8) R -Sq1adj2 = 71.4% (9)
Analysis of Variance
Source DF SS MS F P
Regression 1 174.18 174.18 23.45 (12) 0.001
Residual Error 8 59.42 (10) 7.43 (6)
Total 9 233.60 (11)

3. Standard error of regression coefficient 1SE Coef2 = 3.0. This value is the esti-
mated standard deviation of the regression coefficient value (b1).

sb1 = sy # x> 2©1X - X22 = 2.725> 2.824 = 2.725>.908 = 3.002

4. Computed t value = - 4.84. The computed t value is used to test whether the pop-
ulation regression coefficient ( b 1) is significantly different from zero.

b1 - 14.54
t = = = - 4.84
sb1 3.002

5. Constant = 32.14. This value is the Y-intercept (b0). Therefore, the entire regres-
sion equation is

YN = 32.14 - 14.54X

6. Standard error of estimate = 2.725. The standard error of the estimate indicates
that the Y values fall typically about 2.725 units from the regression line.

©1Y - YN 22 SSE
sy # x = = = 2MSE = 27.43 = 2.725
D n - 2 An - 2

7. The p value, .001, is the probability of obtaining a t value as large as

ƒ t ƒ = ƒ -4.84 ƒ = 4.84 by chance if H0 : b 1 = 0 is true. Since the p value is extremely
small, it is concluded that the regression slope coefficient is significantly different
from 0.
 CHAPTER 6 Simple Linear Regression 243

8. r2 1R - Sq2 = .746. The fitted regression line explains 74.6% of the sales volume
variance.

SSR SSE 59.42

r2 = = 1 - = 1 - = 1 - .254 = .746
SST SST 233.60

9. Adjusted r 2 1R - Sq1adj22 = .714. r 2 is adjusted for the appropriate degrees of

freedom.

©1Y - YN 22>1n - 22 SSE>1n - 22 59.42>8

r2 = 1 - = 1 - = 1 -
©1Y - Y2 >1n - 12
2 SST>1n - 12 233.60>9
7.428
= 1 - = 1 - .286 = .714
25.956

10. Residual sum of squares 1Residual Error2 = 59.42 . The sum of squares
residual is the sum of squared differences between the actual Y’s and the predicted
Y’s (YN ’s).

SSE = ©1Y - YN 22 = 59.42

11. Total sum of squares = 233.60. This value is the sum of the squared deviations of
the Y’s from their mean.

SST = ©1Y - Y22 = 233.60

12. Analysis of variance and F ratio: The F ratio (23.45 = 174.18>7.43) in this ANOVA
table tests the null hypothesis that the regression is not significant; that is,
H0 : b 1 = 0. A large F value will allow rejection of this hypothesis, suggesting a
significant regression. The F value (23.45) becomes larger as a larger portion of the
total sum of squared deviations (SST) is explained by the regression. In this case,
the tabulated F value (df = 1, 8; a = .01) is 11.26. The hypothesis of no significant
regression is rejected at the 1% significance level, since F = 23.45 7 11.26. (See
Example 6.9.)

VARIABLE TRANSFORMATIONS
Although the simple linear regression model assumes a straight-line relationship
between Y and X, in general, a linear regression model refers to a model that is linear
in the unknown b ’s. As long as the regression function is linear in the b ’s (for exam-
ple, b 21 is not present), the predictor variables (the X’s) can take various forms, and the
standard regression methodology is still appropriate. Regression models then can be
used to model complex relationships between Y and X (or several X’s) or to model a
straight-line relationship between Y and some function (transformation) of X.
When a scatter diagram indicates there is a nonlinear relationship between Y and
X, there are two basic approaches for dealing with this case. The first is to fit the data
with a regression function that plots as a curve and use the fitted relationship for fore-
casting purposes. The second approach involves the transformation of the X variable to
another form so that the resulting relationship with Y is linear.
244 CHAPTER 6 Simple Linear Regression

Four of the most common transformations (functions) that are used to generate
new predictor variables are the reciprocal, the log, the square root, and the square:

1
, logX, 1X, X2
X
When these variables are each plotted against Y, the hope is that the nonlinear rela-
tionship between Y and X will become a linear relationship between Y and one of the
transformed X’s. If so, Y and this new variable can be treated using the straight-line
model discussed in this chapter, including calculation of the correlation coefficient and
the fitted regression equation.
In the following example, Minitab (see the Minitab Applications section at the end of
the chapter) is used to plot a simple X and Y relationship that appears to be nonlinear.The
program is then instructed to calculate the four transformations described previously.
These variables are then each plotted against Y to produce the data plots shown.
Example 6.11
Gilbert Garcia owns a chain of hardware stores in Chicago, Illinois. He is interested in
predicting his monthly sales using knowledge of the corresponding monthly advertising
expenditures. Gil suspects that sales will increase as the amount spent on advertising
increases. However, he also believes that after a certain point sales will begin to increase
at a slower rate. Gil feels that, after spending a certain amount on advertising, he will
reach a point where there will be little to gain in sales by further spending on advertising.
Gil selected a random sample of 14 months of data from company records. The data
appear in Figure 6-10, and a scatter diagram of the data is shown in Figure 6-11. Gil notes
that sales do appear to level off after a certain amount is spent on advertising. He fits the
linear regression equation shown in Figure 6-12 and notes that the equation explains 77.7%
of the variability in sales.
A plot of the residuals versus the fitted values from the straight-line fit is shown in
Figure 6-13. This plot indicates that a straight line does not adequately describe the relation
between sales and advertising expenditures. The residuals are all negative for small
predicted values, they are all positive for midlevel predicted values, and they are all negative

FIGURE 6-10 Data for Example 6.11

 CHAPTER 6 Simple Linear Regression 245

FIGURE 6-11 Scatter Diagram for Example 6.11

FIGURE 6-12 Fitted Regression Line for Example 6.11

again for large predicted values. The residuals are not uniformly distributed about the esti-
mated regression line. Clearly, a straight line does not capture the curvature in the data.
Next, Gil considers several transformations of X (see Figure 6-10). He constructs a
multiple scatter diagram with plots of sales (Y) versus the logarithms of advertising expendi-
tures (log X ), sales (Y ) versus the square roots of advertising expenditures ( 1X ), sales
246 CHAPTER 6 Simple Linear Regression

FIGURE 6-13 Plot of Residuals versus Fitted Values for

Straight-Line Fit for Example 6.11

(Y ) versus the squares of advertising expenditures (X2), and finally sales (Y ) versus the
reciprocals of advertising expenditures (1/X). The result is shown in Figure 6-14.
After examining the multiple scatter diagram, Gil concludes that the relationships
between sales and the logarithms of advertising expenditures, sales and the square roots of
advertising expenditures, and sales and the squares of advertising expenditures all show
some degree of curvature. These relationships are not linear. However, the relationship
between sales and the reciprocals of advertising expenditures does appear to be linear with
a negative slope. Using Minitab, Gil fits a simple linear regression model to the sales and
reciprocals of advertising expenditures data. He determines the fitted equation
YN = 4.29 - 12.711>X2, as shown in Table 6-9, and sees that r 2 is equal to 98.1%. A residual
analysis indicates this linear model is appropriate.

GROWTH CURVES
Growth curves were introduced in Chapter 5 (see Figure 5-7) in our discussion of mod-
eling the trend in the decomposition of a time series. Often the trend (long-term
change) is of interest in itself. A forecaster, for example, might be interested in project-
ing the cost per computer transaction well into the future without regard to the tech-
nology that might bring about changes in this cost. As another example, a life insurance
analyst might be interested in long-term projections of the life expectancy at birth for
the U.S. population without regard to the economic and environmental conditions that
might be responsible for changes in life expectancy.
Growth curves are curvilinear relations between the variable of interest and time.
Growth curves are typically fit to annual data, since long-term forecasts are required.
Even though inaccurate forecasts are likely when growth curves fit to historical
data are extrapolated to predict the future, this forecasting method can be of great ben-
efit to managers, since it concentrates attention on the long-term aspects of the business.
Moreover, growth curves indicate the annual rate of growth that must be maintained in
 CHAPTER 6 Simple Linear Regression 247

(a) (b)

(c) (d)

FIGURE 6-14 Scatter Diagram of Sales (Y) versus (a) Log X,

(b) Square Root X, (c) X Squared, and (d) Reciprocal of X

TABLE 6-9 Minitab Output for Example 6.11

Regression Analysis: Sales versus 1/X
The regression equation is
Sales = 4.29 - 12.7 1/X
Predictor Coef SE Coef T P
Constant 4.28587 0.07695 55.69 0.000
1/X -12.7132 0.5092 -24.97 0.000

S = 0.134159 R-Sq = 98.1% R- Sq1adj2 = 98.0%

Analysis of Variance
Source DF SS MS F P
Regression 1 11.221 11.221 623.44 0.000
Residual Error 12 0.216 0.018
Total 13 11.437

order to reach projected future levels. This annual growth rate may or may not be rea-
sonable and can be debated in executive retreat or “think tank” sessions.
If a variable measured over time increases by the same percentage each time
period, it is said to exhibit exponential growth. If a variable increases by the same
amount each time period, it is said to exhibit linear growth. Sometimes a simple trans-
formation will convert a variable having exponential growth to a variable having linear
growth. If this is the case, the regression methods discussed in this chapter can be used
to model exponential growth.
248 CHAPTER 6 Simple Linear Regression

For example, suppose a variable Y, measured annually, follows the exponential

trend curve (see Equation 5.6).
Y = b0bt1
Here, 100(b1 - 1)% is the year-to-year percentage increase in Y. Taking the logarithms
of both sides of the exponential trend equation gives
~ ~
logY = log(b0bt1) = log b0 + t log b1 = b0 + b1 t
~
Thus, log Y versus time (t) plots as a straight line with slope b1 = log b1 and intercept
~ ~
b0 = log b0. Log Y increases by a constant amount, b1, from one year to the next. As
the next example illustrates, forecasts of Y and an estimate of the growth rate, b1, can
be obtained from the straight-line equation relating log Y to t.
Example 6.12
Jill Johnson, an administrative assistant at a major university, is pleased with the acceptance
of an Internet-based course registration system. She has collected the data in Table 6-10 and
sees that the growth in system usage has been phenomenal. If the present rate of growth
continues, the entire campus should be registering for courses online within a year, and the
old telephone course registration system can be eliminated completely.
The data in Table 6-10 show the numbers of students using the Internet-based course
registration system for the 2002–2006 period. The last column in the table contains the loga-
rithms to the base 10 of the numbers of users.
Jill has had some experience with simple linear regression and wonders if this method-
ology can be used to forecast the number of users for 2007 and provide her with an estimate
of the annual growth rate. A plot of the number of users versus year (see Figure 6-15(a))
quickly suggests to Jill that a simple linear regression model is not appropriate for her data.
The growth in usage appears to be exponential, with an annual percentage increase of about
70% per year. The increase in the number of users from year to year, however, is quite dif-
ferent. The increases range in magnitude from a low of 2,500 over the 2002–2003 period to a
high of 10,000 over the 2005–2006 period.
Jill has always been good at manipulating logarithms and remembers that a constant
percentage change in a variable over some time period is equivalent to a constant increase
in magnitude in the log of the variable over the same time period. With this in mind, Jill
transforms the numbers of online course registration users by taking their logarithms and
plots log (users) versus year (see Figure 6-15(b)).
As she suspected, when the numbers of users have been transformed by taking their
logarithms, exponential growth is converted to linear growth. The curvature representing
exponential growth in Figure 6-15(a) has been straightened out in Figure 6-15(b) by the log-
arithmic transformation. In Figure 6-15(b), the increase in the magnitude of log (users) from
year to year is roughly the same, about .23 per year. Jill fits a straight line to the data in
Figure 6-15(b) with log (users) as the dependent variable Y and time (t) as the independent
variable. The results are shown in Figure 6-16.

TABLE 6-10 Data for Example 6.12

Year Time Users Log (Users)

2002 1 3,000 3.477

2003 2 5,500 3.740
2004 3 9,000 3.954
2005 4 16,500 4.217
2006 5 26,500 4.423
 CHAPTER 6 Simple Linear Regression 249

(a)

Scatterplot of Log (Users) vs Year

Log (Users)

(b)

FIGURE 6-15 Scatter Diagrams of Jill

Johnson’s Data for Example 6.12

Using the fitted equation logYN = 3.252 + .2369 t with t = 6 for the year 2007, Jill com-
putes the forecast

logYN 2007 = 3.252 + .2369162 = 4.673

YN 2007 = 104.673 = 47,100

Jill predicts 47,100 students will use the Internet course registration system in 2007. Since
the university’s total enrollment is expected to be about 45,000 students, Jill concludes that
all students are likely to use the new course registration system and that the old telephone
system can be discontinued.
Finally, Jill estimates the annual growth rate in online course registration users to be

1001antilog1.23692 - 12% = 100110.2369 - 12% = 10011.725 - 12% = 72.5%

Jill recognizes that with the cap on university enrollment, it doesn’t make sense to extrapo-
late the growth in online course registration users beyond 2007.

In some situations, a straight-line growth curve can be fit directly to the original data.
250 CHAPTER 6 Simple Linear Regression

Log (Users)
Log (Users)

FIGURE 6-16 Fitted Regression Line for Example 6.12

Example 6.13
Suppose in 1999 a company concerned with servicing the health needs of elderly persons is
interested in a long-term projection of U.S. longevity at birth so it can formulate long-term
corporate strategy. It finds projections for this variable from 1970 to 1997 in the Statistical
Abstract of the United States. These data are shown in Table 6-11. After coding the years as
1 = 1970, 2 = 1971, Á , 28 = 1997 and keying the data into Minitab, the Regression-Fitted
Line Plot routine produces the graph shown in Figure 6-17.
The company is interested not in examining the reasons for the obvious increase in
longevity over the past several years but rather in projecting the fitted regression line far
into the future. Neither is it interested in speculating about whether the upward trend will
continue or by what means such an increase might be achieved. It merely wants to complete
this statement: “If present trends continue, the life expectancy at birth will reach Y at X
years in the future.” It chooses the year 2050 (X = 53 years beyond 1997) as its target future
year. Using 28 + 53 = 81 as the value for time in the regression equation shown in Figure
6-17 and solving for Y, the company arrives at a life expectancy of 86.6 years. This age obvi-
ously represents a substantial increase in longevity over the next decades, and the company
begins working on strategic plans to take advantage of this increased expected lifetime.

APPLICATION TO MANAGEMENT
Regression analysis is the most widely used statistical tool employed by management
when there is a need to evaluate the impact of a single independent variable on a
dependent variable. Regression analysis, along with correlation analysis, helps the fore-
caster characterize the relationships among variables. The forecaster can determine
both the importance and the direction of the relationship between variables.
Most problems utilizing regression analysis involve the more sophisticated version
called multiple regression analysis (to be described in the next chapter) because most
 CHAPTER 6 Simple Linear Regression 251

TABLE 6-11 Expectation of Life at Birth, Total

Population, United States, for
Example 6.13

1970 70.8 1984 74.7

1971 71.2* 1985 74.7
1972 71.5* 1986 74.7
1973 71.9* 1987 74.9
1974 72.4* 1988 74.9
1975 72.6 1989 75.1
1976 72.8* 1990 75.4
1977 73.0* 1991 75.5
1978 73.3* 1992 75.8
1979 73.5* 1993 75.5
1980 73.7 1994 75.7
1981 74.1* 1995 75.8
1982 74.5 1996 76.1
1983 74.6 1997 76.5

Source: Based on Statistical Abstract of the United States, 119th ed.,

1999.
*Interpolated.

FIGURE 6-17 Fitted Regression Line for Example 6.13

252 CHAPTER 6 Simple Linear Regression

relationships involve the study of the relation between a dependent variable and more
than just one independent variable. Nevertheless, simple regression and correlation
analysis are often useful. A few examples of multiple regression situations follow:
• Product consumption. A manufacturer might try to forecast how much beer a
person drinks per week from looking at variables such as income, age, education,
and demographic class.
• Sales. A retailer might try to forecast the sales of a product for one store versus
another on the basis of price differentials, the relative income of the surrounding
community, the relative friendliness of store personnel, and the number and
strength of competitors in each market.
• Stock prices. A stock analyst for a regional brokerage firm might try to forecast
the price of a new issue for a local firm on the basis of the regional economy,
income, population, and visibility of the firm.
• Bad debts. An accountant might try to forecast the bad debt a firm might
encounter in the next fiscal quarter on the basis of number of people unemployed,
outstanding credit, interest rates, and expected sales.
• Employment needs. A personnel director of a large manufacturing firm might try
to forecast the coming year’s staffing requirements from the average age of its
employees, its wage scale compared with that of the surrounding community,
expected new sales contracts, and the availability of competitive jobs.
• Shopping center demand. The manager of a new shopping center might try to
anticipate demand by analyzing the income of the surrounding community, the size
of the population, and the proximity and size of competitive shopping centers.
Once the relationship between the independent and the dependent variables is
determined, management can, in some cases, try to control the dependent variable with
this knowledge. For instance, suppose that a marketing manager determines that there
is a significant positive relationship between advertising expenditures and sales. The
regression equation might be
Sales = $43,000 + .31Advertising expenditures2
From this equation, the marketing manager can try to control sales by increasing or
decreasing advertising expenditures by the amount that would maximize profits.
Whenever the manager has control over the independent variable and the cause-and-
effect relationship implied by the regression equation exists, there is the opportunity
for partial control of the dependent variable. The regression equation and the coeffi-
cient of determination help management determine if such control is worthwhile.

Glossary
Coefficient of determination. The coefficient of Standard error of the estimate. The standard error
determination measures the percentage of variabil- of the estimate measures the amount by which the
ity in Y that can be explained through knowledge actual Y values differ from the estimated, or YN ,
of the variability (differences) in the independent values. It is an estimate of the standard deviation
variable X. of the error term, , in the simple linear regression
Fitted regression line. A fitted regression line is model.
the line that best fits a collection of X–Y data Standard error of the forecast. The standard error
points. It minimizes the sum of the squared dis- of the forecast, sf, measures the variability of the
tances from the points to the line as measured in predicted Y about the actual Y for a given value
the vertical, or Y, direction. of X.
 CHAPTER 6 Simple Linear Regression 253

Key Formulas

Sum of squared errors

SSE = ©1Y - YN 22 = ©1Y - b0 - b1X22 (6.1)

Method of least squares: slope formula

n©XY - ©X©Y ©1X - X21Y - Y2
b1 = =
n©X - 1©X2
(6.2)
2 2
©1X - X22

Method of least squares: Y-intercept formula

©Y b1 ©X
b0 = - = Y - b1X (6.3)
n n

Relation between slope coefficient and correlation coefficient

2©1Y - Y22
b1 = r (6.4)
2©1X - X22
Observation  Fit  Residual

Y = YN + 1Y - YN 2 (6.5)

Fitted regression equation

YN = b0 + b1X (6.6)

Standard error of estimate: definitional formula

©1Y - YN 22
sy # x = (6.7)
A n - 2

Standard error of estimate: computational formula

©Y2 - b0 ©Y - b1 ©XY
sy # x = (6.8)
A n - 2
Standard error of the forecast

sf = sy # x 1 + 1 + 1X - X2
2
(6.9)
A n ©1X - X22
Prediction interval
YN ; t sf (6.10)

Large-sample 95% prediction interval

YN ; 2sf (6.11)

Sum of squares and degrees of freedom decomposition

SST = SSR + SSE
©1Y - Y22 = ©1YN - Y22 + ©1Y - YN 22 (6.12)
df : n - 1 = 1 + (n - 2) (6.13)
254 CHAPTER 6 Simple Linear Regression

Coefficient of determination
©1YN - Y22 ©1Y - YN 22
r2 = = 1 - (6.14)
©1Y - Y22 ©1Y - Y22

t statistic for testing H0 : b 1 = 0

1see shaded box on p. 2372

b1
t =
sb1

Standard error of the regression coefficient

sb1 = sy # x> 2©1X - X22 1see shaded box on p. 2372

F statistic
Regression mean square MSR
F = = (6.15)
Error mean square MSE

Relation of F statistic to coefficient of determination

r 21n - 22
F = (6.16)
1 - r2

Residual autocorrelation coefficient

n

a etet - k
rk1e2 =
t=k+1
n k = 1, 2, Á , K (6.17)
2
a et
t=1

Problems

Note: Most of the following problems contain data that are to be manipulated using
regression analysis procedures. Although it is possible, even useful, to work one or
two of these by hand, it is important for you to learn how to use computer software to
solve such problems. In the next chapter, you will learn about multiple regression
analysis, in which it is not feasible to solve the problems by hand. You should become
familiar with regression analysis software while solving the following problems. If you
have access to Minitab or Excel, see the section at the end of this chapter for instruc-
tions on their use.
1. Which of the following situations is inconsistent?
a. YN = 499 + .21X and r = .75
b. YN = 100 + .9X and r = - .70
c. YN = - 20 + 1X and r = .40
d. YN = - 7 - 4X and r = - .90
2. AT&T (American Telegraph and Telephone) earnings in billions of dollars are
estimated using GNP (gross national product). The regression equation is
YN = .078 + .06X, where GNP is measured in billions of dollars.
a. Interpret the slope.
b. Interpret the Y-intercept.
 CHAPTER 6 Simple Linear Regression 255

TABLE P-3
Y ($) X ($) Y ($) X ($)

1,250 41 1,300 46
1,380 54 1,400 62
1,425 63 1,510 61
1,425 54 1,575 64
1,450 48 1,650 71

3. Consider the data in Table P-3 where X = weekly advertising expenditures and
Y = weekly sales.
a. Does a significant relationship exist between advertising expenditures and
sales?
b. State the prediction equation.
c. Forecast sales for an advertising expenditure of $50.
d. What percentage of the variation in sales can be explained with the prediction
equation?
e. State the amount of unexplained variation.
f. State the amount of total variation.
4. The times required to check out customers in a supermarket and the correspond-
ing values of purchases are shown in Table P-4. Answer parts a, b, e, and f of
Problem 3 using these data. Give a point forecast and a 99% interval forecast for Y
if X = 3.0.
5. Lori Franz, maintenance supervisor for the Baltimore Transit Authority, would like
to determine whether there is a positive relationship between the annual mainte-
nance cost of a bus and its age. If a relationship exists, Lori feels that she can do a
better job of forecasting the annual bus maintenance budget. She collects the data
shown in Table P-5.
a. Plot a scatter diagram.
b. What kind of relationship exists between these two variables?
c. Compute the correlation coefficient.
d. Determine the least squares line.
e. Test for the significance of the slope coefficient at the .05 significance level. Is
the correlation significant? Explain.
f. Forecast the annual maintenance cost for a five-year-old bus.
6. Andrew Vazsonyi is the manager of the Spendwise supermarket chain. He would
like to be able to forecast paperback book sales (books per week) based on the

TABLE P-4

Time Required Value of Time Required Value of

for Checkout Purchase for Checkout Purchase
(minutes) ($) (minutes) ($)
3.6 30.6 1.8 6.2
4.1 30.5 4.3 40.1
.8 2.4 .2 2.0
5.7 42.2 2.6 15.5
3.4 21.8 1.3 6.5
256 CHAPTER 6 Simple Linear Regression

TABLE P-5

Maintenance Cost ($) Age (years)

Bus Y X

1 859 8
2 682 5
3 471 3
4 708 9
5 1,094 11
6 224 2
7 320 1
8 651 8
9 1,049 12

TABLE P-6

Number of Books Sold Feet of Shelf Space

Week Y X

1 275 6.8
2 142 3.3
3 168 4.1
4 197 4.2
5 215 4.8
6 188 3.9
7 241 4.9
8 295 7.7
9 125 3.1
10 266 5.9
11 200 5.0

amount of shelf display space (feet) provided. Andrew gathers data for a sample of
11 weeks, as shown in Table P-6.
a. Plot a scatter diagram.
b. What kind of relationship exists between these two variables?
c. Compute the correlation coefficient.
d. Determine the least squares line.
e. Test for the significance of the slope coefficient at the .10 significance level. Is
the correlation significant? Explain.
f. Plot the residuals versus the fitted values. Based on this plot, is the simple linear
regression model appropriate for these data?
g. Forecast the paperback book sales for a week during which four feet of shelf
space is provided.
7. Information supplied by a mail-order business for 12 cities is shown in Table P-7.
a. Determine the fitted regression line.
b. Calculate the standard error of the estimate.
c. Determine the ANOVA table.
d. What percentage of the variation in mail orders is explained by the number of
catalogs distributed?
 CHAPTER 6 Simple Linear Regression 257

TABLE P-7

Number of Number of Number of Number of

Mail Orders Catalogs Mail Orders Catalogs
Received Distributed Received Distributed
(1,000s) (1,000s) (1,000s) (1,000s)
City Y X City Y X

A 24 6 G 18 15
B 16 2 H 18 3
C 23 5 I 35 11
D 15 1 J 34 13
E 32 10 K 15 2
F 25 7 L 32 12

e. Test to determine whether the regression slope coefficient is significantly differ-

ent from zero. (Use the .01 significance level.)
f. Test for the significance of the regression using the F statistic from the ANOVA
table. (Use the .01 significance level.) Is the result consistent with the result in
part e? Should it be?
g. Forecast the number of mail orders received when 10,000 catalogs are distrib-
uted with a 90% prediction interval.
8. In a regression of investment on the interest rate, the results in Table P-8 were
observed over a 10-year period.
a. Is the relationship between these variables significant?
b. Can an effective prediction equation be developed?
c. If the average interest rate is 4% five years from now, can yearly investment be
forecast?
d. Calculate and interpret r 2.
e. Discuss correlation and causation in this example.
9. The ABC Investment Company is in the business of making bids on investments
offered by various firms that want additional financing. ABC has tabulated its bids
on the last 25 issues in terms of their percentage of par value. The bids of ABC’s
major competitor, as a percentage of par value, are also tabulated on these issues.
ABC now wonders if it is using the same rationale as its competitor in preparing
bids. In other words, could ABC’s bid be used to forecast the competitor’s bid? If
not, then the competitor must be evaluating issues differently. The data are given
in Table P-9.

TABLE P-8

Yearly Investment Average Interest Yearly Investment Average Interest

($1,000s) Rate (%) ($1,000s) Rate (%)

1,060 4.8 2,050 3.8

940 5.1 2,070 3.7
920 5.9 2,030 4.5
1,110 5.1 1,780 4.9
1,590 4.8 1,420 6.2
258 CHAPTER 6 Simple Linear Regression

TABLE P-9

Issue ABC Bid Competitor Bid Issue ABC Bid Competitor Bid

1 99.035 100.104 14 100.542 99.936

2 104.358 105.032 15 96.842 95.834
3 99.435 99.517 16 99.200 99.863
4 96.932 95.808 17 101.614 102.010
5 98.904 98.835 18 99.501 99.432
6 101.635 101.563 19 100.898 99.965
7 100.001 101.237 20 97.001 96.838
8 98.234 99.123 21 100.025 100.804
9 93.849 94.803 22 103.014 104.300
10 99.412 100.063 23 98.702 99.010
11 99.949 99.564 24 101.834 100.936
12 104.012 103.889 25 102.903 103.834
13 99.473 99.348

a. To what extent are the two firms using the same rationale in preparing their bids?
b. Forecast the competitor’s bid if ABC bids 101% of par value. Give both a point
forecast and an interval prediction.
c. Under part b, what is the probability of ABC winning this particular bid (the
lowest bid wins)?
10. Evaluate the following statements:
a. A high r 2 means a significant regression.
b. A very large sample size in a regression problem will always produce useful results.
11. Ed Bogdanski, owner of the American Precast Company, has hired you as a part-
time analyst. Ed was extremely pleased when you determined that there is a posi-
tive relationship between the number of building permits issued and the amount of
work available to his company. Now he wonders if it is possible to use knowledge
of interest rates on first mortgages to predict the number of building permits that
will be issued each month. You collect a random sample of nine months of data, as
shown in Table P-11.

TABLE P-11

Building Permits Interest Rate (%)

Month Y X

1 786 10.2
2 494 12.6
3 289 13.5
4 892 9.7
5 343 10.8
6 888 9.5
7 509 10.9
8 987 9.2
9 187 14.2
 CHAPTER 6 Simple Linear Regression 259
a. Plot the data as a scatter diagram.
b. Determine the fitted regression function.
c. Test for the significance of the slope coefficient at the .05 significance level.
d. When the interest rate increases by 1%, what is the average decrease in the
number of building permits issued?
e. Compute the coefficient of determination.
f. Write a sentence that Ed can understand, interpreting the number computed in
part e.
g. Write Ed a memo explaining the results of your analysis.
12. Consider the population of 140 observations that are presented in Table P-12. The
Marshall Printing Company wishes to estimate the relationship between the num-
ber of copies produced by an offset printing technique (X) and the associated
direct labor cost (Y). Select a random sample of 20 observations.
a. Construct a scatter diagram.
b. Compute the sample correlation coefficient.
c. Determine the fitted regression line.
d. Plot the fitted line on the scatter diagram.
e. Compute the standard error of estimate.
f. Compute the coefficient of determination and interpret its value.
g. Test the hypothesis that the slope, b 1, of the population regression line is zero.
h. Calculate a point forecast and a 90% prediction interval for the direct labor cost
if the project involves 250 copies.
i. Examine the residuals. Does it appear as if a simple linear regression model is
appropriate for these data? Explain.
13. Harry Daniels is a quality control engineer for the Specific Electric Corporation.
Specific manufactures electric motors. One of the steps in the manufacturing
process involves the use of an automatic milling machine to produce slots in the
shafts of the motors. Each batch of shafts is tested, and all shafts that do not meet
required dimensional tolerances are discarded. The milling machine must be read-
justed at the beginning of each new batch because its cutter head wears slightly
during production. Harry is assigned the job of forecasting how the size of a batch
affects the number of defective shafts in the batch so that he can select the best
batch size. He collects the data for the average batch size of 13 batches shown in
Table P-13 and assigns you to analyze it.
a. Plot the data as a scatter diagram.
b. Fit a simple linear regression model.
c. Test for the significance of the slope coefficient.
d. Examine the residuals.
e. Develop a curvilinear model by fitting a simple linear regression model to some
transformation of the independent variable.
f. Test for the significance of the slope coefficient of the transformed variable.
g. Examine the residuals.
h. Forecast the number of defectives for a batch size of 300.
i. Which of the models in parts b and e do you prefer?
j. Write Harry a memo summarizing your results.
14. The data in Table P-14 were collected as part of a study of real estate property
evaluation. The numbers are observations on X = assessed value (in thousands of
dollars) on the city assessor’s books and Y = market value (selling price in thou-
sands of dollars) for n = 30 parcels of land that sold in a particular calendar year
in a certain geographical area.
260 CHAPTER 6 Simple Linear Regression

TABLE P-12

Obs. Y X Obs. Y X Obs. Y X

(1) 1.0 10 (48) 2.2 180 (95) 2.0 330

(2) 0.9 10 (49) 2.4 180 (96) 2.4 340
(3) 0.8 10 (50) 1.6 180 (97) 2.2 340
(4) 1.3 20 (51) 1.8 190 (98) 2.0 340
(5) 0.9 20 (52) 4.1 190 (99) 2.5 350
(6) 0.6 30 (53) 2.0 190 (100) 2.8 350
(7) 1.1 30 (54) 1.5 200 (101) 2.3 350
(8) 1.0 30 (55) 2.1 200 (102) 2.7 350
(9) 1.4 40 (56) 2.5 200 (103) 2.8 360
(10) 1.4 40 (57) 1.7 220 (104) 3.1 360
(11) 1.2 40 (58) 2.0 220 (105) 2.5 370
(12) 1.7 50 (59) 2.3 220 (106) 2.9 370
(13) 0.9 50 (60) 1.8 220 (107) 2.6 370
(14) 1.2 50 (61) 1.3 230 (108) 3.0 380
(15) 1.3 50 (62) 1.6 230 (109) 3.2 380
(16) 0.7 60 (63) 2.8 230 (110) 2.9 390
(17) 1.0 60 (64) 2.2 230 (111) 2.6 390
(18) 1.3 70 (65) 2.6 230 (112) 2.5 390
(19) 1.5 70 (66) 1.4 240 (113) 2.7 400
(20) 2.0 70 (67) 1.6 240 (114) 3.1 400
(21) 0.8 80 (68) 1.7 240 (115) 2.4 400
(22) 0.6 80 (69) 1.5 250 (116) 3.0 400
(23) 1.8 80 (70) 2.2 250 (117) 3.4 420
(24) 1.0 90 (71) 2.5 250 (118) 3.5 420
(25) 2.0 100 (72) 2.4 260 (119) 3.1 420
(26) 0.5 100 (73) 2.0 260 (120) 2.9 420
(27) 1.5 100 (74) 2.7 260 (121) 2.8 430
(28) 1.3 110 (75) 2.0 270 (122) 3.3 430
(29) 1.7 110 (76) 2.2 270 (123) 2.5 440
(30) 1.2 110 (77) 2.4 270 (124) 2.8 440
(31) 0.8 110 (78) 1.8 280 (125) 2.4 450
(32) 1.0 120 (79) 2.8 290 (126) 2.6 450
(33) 1.8 120 (80) 2.2 290 (127) 3.0 450
(34) 2.1 120 (81) 2.4 290 (128) 3.4 460
(35) 1.5 130 (82) 2.1 290 (129) 3.0 460
(36) 1.9 130 (83) 1.9 290 (130) 3.3 470
(37) 1.7 140 (84) 2.4 300 (131) 3.4 470
(38) 1.2 150 (85) 2.5 300 (132) 3.1 470
(39) 1.4 150 (86) 2.9 300 (133) 3.6 480
(40) 2.1 150 (87) 2.0 300 (134) 3.0 480
(41) 0.9 160 (88) 1.9 310 (135) 2.9 480
(42) 1.1 160 (89) 2.5 310 (136) 3.2 480
(43) 1.7 160 (90) 2.6 310 (137) 2.6 490
(44) 2.0 160 (91) 3.2 320 (138) 3.8 490
(45) 1.6 170 (92) 2.8 320 (139) 3.3 490
(46) 1.9 170 (93) 2.4 320 (140) 2.9 500
(47) 1.7 170 (94) 2.5 320
 CHAPTER 6 Simple Linear Regression 261

TABLE P-13

Number of Defectives Batch Size

Batch Y X

1 4 25
2 8 50
3 6 75
4 16 100
5 22 125
6 27 150
7 36 175
8 49 200
9 53 225
10 70 250
11 82 275
12 95 300
13 109 325

TABLE P-14

Parcel Assessed Market Parcel Assessed Market

1 68.2 87.4 16 74.0 88.4

2 74.6 88.0 17 72.8 93.6
3 64.6 87.2 18 80.4 92.8
4 80.2 94.0 19 74.2 90.6
5 76.0 94.2 20 80.0 91.6
6 78.0 93.6 21 81.6 92.8
7 76.0 88.4 22 75.6 89.0
8 77.0 92.2 23 79.4 91.8
9 75.2 90.4 24 82.2 98.4
10 72.4 90.4 25 67.0 89.8
11 80.0 93.6 26 72.0 97.2
12 76.4 91.4 27 73.6 95.2
13 70.2 89.6 28 71.4 88.8
14 75.8 91.8 29 81.0 97.4
15 79.2 94.8 30 80.6 95.4

a. Plot the market value against the assessed value as a scatter diagram.
b. Assuming a simple linear regression model, determine the least squares line
relating market value to assessed value.
c. Determine r 2 and interpret its value.
d. Is the regression significant? Explain.
e. Predict the market value of a property with an assessed value of 90.5. Is there
any danger in making this prediction?
f. Examine the residuals. Can you identify any observations that have a large
influence on the location of the least squares line?
15. Player costs (X ) and operating expenses (Y ) for n = 26 major league baseball
teams for the 1990–1991 season are given in Table P-15.
262 CHAPTER 6 Simple Linear Regression

TABLE P-15

Player Costs Operating Expenses

Team X ($ millions) Y ($ millions)

1 29.8 59.6
2 36.0 72.0
3 35.2 70.4
4 29.7 62.4
5 35.4 70.8
6 15.8 39.5
7 18.0 60.0
8 23.2 46.4
9 29.0 58.0
10 20.7 47.6
11 30.4 60.8
12 21.7 43.4
13 39.2 66.6
14 34.3 61.7
15 33.3 53.3
16 27.1 48.8
17 24.4 48.8
18 12.1 31.5
19 24.9 49.8
20 31.1 54.4
21 20.4 40.8
22 24.1 48.2
23 17.4 41.8
24 26.4 50.2
25 19.5 46.8
26 21.8 43.6

a. Assuming a simple linear regression model, determine the equation for the fit-
ted straight line.
b. Determine r 2 and comment on the strength of the linear relation.
c. Test for the significance of the regression with a level of significance of .10.
d. Can we conclude that, as a general rule, operating expenses are about twice
player costs? Discuss.
e. Forecast operating expenses, with a large-sample 95% prediction interval, if
player costs are $30.5 million.
f. Using the residuals as a guide, identify any unusual observations. That is, do
some teams have unusually low or unusually high player costs as a component
of operating expenses?
16. Table P-16 contains data for 23 cities on newsprint consumption (Y ) and the num-
ber of families in the city (X) during a particular year.
a. Plot newsprint consumption against number of families as a scatter diagram.
b. Is a simple linear regression model appropriate for the data in Table P-16? Be
sure your answer includes an analysis of the residuals.
c. Consider a log transformation of newsprint consumption and a simple linear
regression model relating Y = log (newsprint consumption) to X = number of
families. Fit this model.
 CHAPTER 6 Simple Linear Regression 263

TABLE P-16

Newsprint Number of Newsprint Number of

Consumption Families Consumption Families
City Y (metric tons) X City Y (metric tons) X

1 961 8,600 13 878 8,330

2 469 6,870 14 637 9,010
3 556 9,880 15 3,291 11,790
4 1,252 12,370 16 2,470 18,910
5 902 6,920 17 916 8,550
6 1,399 13,760 18 525 8,850
7 1,877 7,450 19 1,159 8,540
8 921 6,700 20 1,138 6,910
9 494 7,420 21 979 7,060
10 530 6,930 22 1,899 10,920
11 488 7,400 23 5,022 14,800
12 1,253 7,420

d. Examine the residuals from the regression in part c. Which model, the one in
part b or the one in part c, is better? Justify your answer.
e. Using the fitted function in part c, forecast the amount of newsprint consumed
in a year if a city has 10,000 families.
f. Can you think of other variables that are likely to influence the amount of
newsprint consumed in a year?
17. Outback Steakhouse grew explosively during its first years of operation. The num-
bers of Outback Steakhouse locations for the period 1988–1993 are given below.

Year 1988 1989 1990 1991 1992 1993

Number of locations 2 9 23 49 87 137

Source: Outback Steakhouse.

a. Does there appear to be linear or exponential growth in the number of

steakhouses?
b. Estimate the annual growth rate for Outback Steakhouse over the 1988–1993
period.
c. Forecast the number of Outback locations for 2007. Does this number seem rea-
sonable? Explain.
18. On The Double, a chain of campus copy centers, began operations with a single
store in 1993. The number of copy centers in operation, Y, is recorded for 14 con-
secutive years in Table P-18.
a. Plot number of copy centers versus year. Has On The Double experienced lin-
ear or exponential growth?
b. Determine the annual growth rate for On The Double.
c. Predict the number of copy centers in operation in 2012. Does this number
seem reasonable? Why or why not?
19. The number of employees (X) and profits per employee (Y) for n = 16 publishing
firms are given in Table P-19. Employees are recorded in thousands and profits per
264 CHAPTER 6 Simple Linear Regression

TABLE P-18

Year Time Copy Centers

1993 1 1
1994 2 2
1995 3 2
1996 4 6
1997 5 10
1998 6 16
1999 7 25
2000 8 41
2001 9 60
2002 10 97
2003 11 150
2004 12 211
2005 13 382
2006 14 537

TABLE P-19

Firm Profits ($1,000s) Employees (1,000s)

1 33.5 9.4
2 31.4 6.3
3 25.0 10.7
4 23.1 7.4
5 14.2 17.1
6 11.7 21.2
7 10.8 36.8
8 10.5 28.5
9 9.8 10.7
10 9.1 9.9
11 8.5 26.1
12* 8.3 70.5
13 4.8 14.8
14 3.2 21.3
15 2.7 14.6
16 –9.5 26.8

*Dun and Bradstreet.

employee in thousands of dollars. A portion of the Minitab output from a straight-

line regression analysis is given below.

The regression equation is

Profits = 18.0 - 0.271 Employees
Predictor Coef SE Coef T P
Constant 17.954 4.457 4.03 0.001
Employees -0.2715 0.1726 - 1.57 0.138
S = 10.6112 R - Sq = 15.0% R - Sq1adj2 = 9.0%
 CHAPTER 6 Simple Linear Regression 265
a. Identify the least squares estimates of the slope and intercept coefficients.
b. Test the hypothesis H0 : b 1 = 0 with a = .10. Does there appear to be a rela-
tionship between profits per employee and number of employees? Explain.
c. Identify and interpret r 2.
d. Is this fitted regression function likely to be a good tool for forecasting profits
per employee for a given number of publishing firm employees? Discuss.
20. Refer to Problem 19. Observation 12 corresponds to Dun and Bradstreet. Redo
the straight-line regression analysis omitting this observation. Do your conclusions
in parts b and d of Problem 19 change? What, if anything, does this imply about the
influence of a single observation on a regression analysis when the number of
observations is fairly small? Do you think it is reasonable to throw out Dun and
Bradstreet? Discuss.
21. Data on the actual cost (Y) and the estimated cost (X), in millions of dollars, for
n = 26 construction projects are given in Table 2-8 in the context of Example 2.13.
Figure 2-17 shows the fitted line plot for these data.
a. Reanalyze the cost data using a simple linear regression model. Construct a plot
of the residuals versus the fitted values.
b. Is the regression of actual cost on estimated cost significant? Justify your
answer.
c. Identify and interpret r 2.
d. What are the appropriate values for the slope and intercept coefficients in the
population regression function if estimated cost is a perfect predictor of
expected actual cost? Are the estimated coefficients in the equation for your fit-
ted straight line consistent with these values? Discuss.
e. Consider the plot of the residuals versus the fitted values. Does it appear as if
estimated costs are, in general, more accurate predictors of actual costs for rela-
tively inexpensive projects than for expensive projects? Why or why not?
22. Refer to Problem 21. The cost data can be transformed by taking the natural loga-
rithms of both estimated costs and actual costs. Consider a simple linear regression
model for the transformed costs. A portion of the Minitab output for this analysis
is shown below.

The regression equation is

LnActual = 0.003 + 0.968 LnEstimate
Predictor Coef SE Coef T P
Constant 0.0026 0.1625 0.02 0.987
LnEstimate 0.96811 0.06583 14.71 0.000
S = 0.412124 R - Sq = 90.0% R - Sq1adj2 = 89.6%

a. Is this regression significant? Explain.

b. Identify and interpret r 2.
c. Are the sample data consistent with the population values b 0 = 0 and b 1 = 1?
Explain. What is the significance of these particular choices for b 0 and b 1?
d. Construct a forecast of actual cost when estimated cost is $24 million.
266 CHAPTER 6 Simple Linear Regression

CASES

CASE 6-1 TIGER TRANSPORT

Tiger Transport Company is a trucking firm that assuming that all trucks are the same; in fact, they
moves household goods locally and across the coun- are nearly identical in terms of size, gross-weight
try. Its current concern involves the price charged for capacity, and engine size. You also assume that every
moving small loads over long distances. It has rates it driver will get the same truck mileage over a long
is happy with for full truckloads; these rates are based trip. Tiger’s chief accountant feels that these
on the variable costs of driver, fuel, and maintenance, assumptions are reasonable.
plus overhead and profit. When a truck is less than You are then left with only one variable that
fully loaded, however, there is some question about might affect the miles per gallon of long-haul trucks:
the proper rate to be charged on goods needed to fill cargo weight. You find that the accounting depart-
the truck. To forecast future fuel needs and prepare ment has records for every trip made by a Tiger truck
long-range budgets,Tiger would like to determine the over the past several years. These records include the
cost of adding cargo to a partially filled truck. total cargo weight, the distance covered, and the
Tiger feels that the only additional cost incurred number of gallons of diesel fuel used.A ratio of these
if extra cargo is added to the truck is the cost of last two figures is the miles per gallon for the trip.
additional fuel, since the miles per gallon of the You select trips made over the past four years as
truck would then be lowered. As one of the factors your population; there are a total of 5,428 trips. You
used to determine rates for small loads, the company then select 40 random numbers from a random
would like to know its out-of-pocket fuel costs asso- number table, and since the trips are recorded one
ciated with additional cargo. after another, you assign the number 1 to the first
You are a recent business school graduate work- recorded trip, 2 to the second, and so on. Your 40
ing in the cost accounting department, and you are random numbers thus produce a random selection
assigned the job of investigating this matter and of 40 trips to be examined. The cargo weight and
advising top management on the considerations nec- miles per gallon for these trips are recorded and
essary for a sound rate decision. You begin by appear in Table 6-12.

TABLE 6-12 Data for Trip Cargo Weight and Miles per Gallon for Tiger Transport

Weight Miles per Weight Miles per Weight Miles per Weight Miles per
(1,000s of lbs.) Gallon (1,000s of lbs.) Gallon (1,000s of lbs.) Gallon (1,000s of lbs.) Gallon

60 5.3 58 4.9 63 5.0 63 5.0

55 5.0 60 5.1 65 4.9 62 4.9
80 4.0 74 4.5 72 4.6 77 4.6
72 4.2 80 4.3 81 4.0 76 4.5
75 4.5 53 5.9 64 5.3 51 5.7
63 5.1 61 5.5 78 4.4 74 4.2
48 7.2 80 3.5 62 4.9 78 4.3
79 3.9 68 4.1 83 3.8 50 6.1
82 3.8 76 4.5 79 4.1 79 4.3
72 4.4 75 4.4 61 4.8 55 4.7
 CHAPTER 6 Simple Linear Regression 267

TABLE 6-13 Regression Analysis Output for Tiger Transport

Regression Analysis: MPG versus Weight

The regression equation is
MPG = 8.85 - 0.0604 Weight

Predictor Coef SE Coef T P

Constant 8.8484 0.3840 23.04 0.000
Weight -0.060399 0.005538 -10.91 0.000

S = 0.3534 R-Sq = 75.8% R-Sq1adj2 = 75.1%

Analysis of Variance

Source DF SS MS F P
Regression 1 14.853 14.853 118.93 0.000
Residual Error 38 4.746 0.125
Total 39 19.599

Since your desktop computer has software with Mean miles per gallon = 4.71from Table 6-122
a regression analysis package, you fit a simple linear
regression model to the data in Table 6-12. The 10012.552
Cost of 100 miles = = $54.26
resulting printout appears in Table 6-13. 4.7
After studying the printout in Table 6-13, you
decide that the sample data have produced a useful Cost of same trip with an additional
regression equation. This conclusion is based on a
relatively high r 2 (76%), a large negative t value 10012.552
= $54.96
14.7 - .06042
1,000 pounds is
( -10.9), and a high F value (119). From the printout,
you write down the equation of the fitted line
Thus,
YN = 8.8484 - .0604X
where Y is measured in miles per gallon and X is mea- Incremental cost of 1,000 pounds carried 100 miles
sured in thousands of pounds. The slope of the regres-  $.70
sion equation ( -.0604) is interpreted as follows: Each
additional 1,000 pounds of cargo reduces the mileage of You now believe you have completed part of your
a truck by an average of .0604 mile per gallon. assignment, namely, determination of the out-of-
Tiger is currently paying approximately $2.55 pocket fuel costs associated with adding cargo
per gallon for diesel fuel. You can therefore calcu- weight to a less-than-full truck. You realize, of
late the cost of hauling an additional 1,000 pounds of course, that other factors bear on a rate decision for
cargo 100 miles, as follows: small loads.

ASSIGNMENT
1. Prepare a memo for Tiger’s top management on the extent to which your work will improve
that summarizes the analysis. Include comments forecasts for fuel needs and truck revenue.
268 CHAPTER 6 Simple Linear Regression

CASE 6-2 BUTCHER PRODUCTS, INC.

Gene Butcher is owner and president of Butcher appear as follows; Y represents the number of units
Products, Inc., a small company that makes fiber- produced, while X represents the absolute difference
glass ducting for electrical cable installations. Gene (negative signs eliminated) between the day’s high
has been studying the number of duct units manu- temperature and 65 degrees:
factured per day over the past two and a half years
and is concerned about the wide variability in this Y X Y X
figure. To forecast production output, costs, and rev- 485 12 327 15
enues properly, Gene needs to establish a relation- 512 10 308 25
ship between output and some other variable. 625 3 603 8
585 4 321 35
Based on his experience with the company, Gene 318 27 426 5
is unable to come up with any reason for the variabil- 405 10 410 12
379 18 515 2
ity in output until he begins thinking about weather 497 12 498 7
conditions. His reasoning is that the outside tempera- 316 27 357 17
ture may have something to do with the productivity 351 20 429 8
525 4 401 12
of his workforce and the daily output achieved. 395 11
He randomly selects several days from his records
and records the number of ducting units produced for Gene performs a simple linear regression analysis
each of these days. He then goes to the local weather using his company’s computer and the Minitab soft-
bureau and, for each of the selected days, records the ware program.
high temperature for the day. He is then ready to run a Gene is pleased to see the results of his regression
correlation study between these two figures when he analysis, as presented in Table 6-14. The t values corre-
realizes that output would probably be related to sponding to the estimated intercept and slope coeffi-
deviation from an ideal temperature rather than to cients are large (in absolute value), and their p-values
the temperature itself.That is, he thinks that a day that are very small. Both coefficients (552 and 8.9) in the
is either too hot or too cold would have a negative sample regression equation are clearly significant.
effect on production when compared with a day that Turning to r 2, Gene is somewhat disappointed
has an ideal temperature. He decides to convert his to find that this value, although satisfactory, is not as
temperature readings to deviations from 65 degrees high as he had hoped (64.2%). However, he decides
Fahrenheit, a temperature he understands is ideal in that it is high enough to begin thinking about ways
terms of generating high worker output. His data to increase daily production levels.

TABLE 6-14 Regression Analysis Output for Butcher Products, Inc.

Regression Analysis: Y versus X

The regression equation is
Y = 552 - 8.91 X
Predictor Coef SE Coef T P
Constant 552.04 22.85 24.16 0.000
X 8.911 1.453 6.13 0.000
S = 59.41 R -Sq = 64.2% R -Sq1adj2 = 62.5%
Analysis of Variance
Source DF SS MS F P
Regression 1 132,758 132,758 37.62 0.000
Residual Error 21 74,109 3,529
Total 22 206,866
 CHAPTER 6 Simple Linear Regression 269

QUESTIONS
1. How many units would you forecast for a day in Gene to take in order to increase daily
which the high temperature is 89 degrees? output?
2. How many units would you forecast for a day in 4. Do you think Gene has developed an effective
which the high temperature is 41 degrees? forecasting tool?
3. Based on the results of the regression analysis
as shown earlier, what action would you advise

CASE 6-3 ACE MANUFACTURING

Ace Manufacturing Company employs several thou- sample, he intends to take a sample of 200 or 300
sand people in the manufacture of keyboards, equip- folders and formulate a good prediction equation.
ment cases, and cables for the small-computer industry. The following table contains the data values col-
The president of Ace has recently become concerned lected in the initial sample. The number of absent
with the absentee rate among the company’s employ- days during the past fiscal year is represented by Y,
ees and has asked the personnel department to look while X represents employee age.
into this matter. Personnel realizes that an effective
method of forecasting absenteeism would greatly
strengthen its ability to plan properly. Y X Y X
Bill McGone, the personnel director, decides to
3 25 9 56
take a look at a few personnel folders in an attempt
4 36 12 60
to size up the problem. He decides to randomly
7 41 8 51
select 15 folders and record the number of absent
4 27 5 33
days during the past fiscal year, along with employee
3 35 6 37
age. After reading an article in a recent personnel
3 31 2 31
journal, he believes that age may have a significant
5 35 2 29
effect on absenteeism. If he finds that age and
7 41
absent days show a good correlation in his small

QUESTIONS
1. How well are absent days and age correlated? proper statistical procedures to support your
Can this correlation be generalized to the entire answer.
workforce? 5. Suppose a newly hired person is 24 years old.
2. What is the forecasting equation for absent days How many absent days would you forecast for
using age as a predictor variable? this person during the fiscal year?
3. What percentage of the variability in absent 6. Should Bill McGone proceed to take a large
days can be explained through knowledge of sample of company employees based on the
age? preliminary results of his sample?
4. Is there a significant relation between absent 7. Has an effective forecasting method been
days and age? In answering this question, use developed?
270 CHAPTER 6 Simple Linear Regression

CASE 6-4 MR. TUX

John Mosby has heard that regression analysis is is consistent with this result (John recalls that F = t 2
often used to forecast time series variables, and for straight-line regression), and the null hypothesis
since he has a personal computer with a regression that the regression is not significant must be
software package, he decides to give it a try. The rejected.
monthly sales volumes for Mr. Tux for the years 1998 John is disappointed with the relatively low r 2
through 2005 are the dependent variable. As a first (56.3%). He had hoped for a higher value so that his
try, he decides to use the period number as the pre- simple regression equation could be used to accu-
dictor, or X, variable. His first Y sales value, $6,028, rately forecast his sales. He realizes that this low
will have an X value 1, the second will have an X value must be due to the seasonality of his monthly
value of 2, and so on. His reasoning is that the sales, a fact he knew about before he began his fore-
upward trend he knows exists in his data will be casting efforts. Considerable seasonal variation
accounted for by using an ever-rising X value to would result in monthly data points that do not clus-
explain his sales data. ter about the linear regression line, resulting in an
After he performs the regression analysis on his unsatisfactory r 2 value.
computer, John records the following values: Another thing troubles John about his regres-
sion results. On the printout is this statement:
t = 11.01 r 2 = .563
Durbin-Watson = .99. He does not understand this
F = 121.14 YN = - 6, 495 + 2, 729.2X statement and calls the statistics professor he had in
college. After he describes the values on his regres-
The high t value indicates to John that his fitted sion printout, the professor says, “I have a class to
regression line slope (2,729.2) is significant; that is, teach right now, but your low Durbin-Watson statis-
he rejects the null hypothesis that the slope of the tic means that one of the assumptions of regression
population regression line is zero. The high F value analysis does not hold.”

QUESTIONS
1. Comment on John’s belief that his monthly sales 3. How do John’s data violate one of the assump-
are highly seasonal and therefore lead to a tions of regression analysis?
“low” r 2 value.
2. What is your opinion regarding the adequacy of
John’s forecasting method?

CASE 6-5 CONSUMER CREDIT COUNSELING

The Consumer Credit Counseling (CCC) operation the number of new clients seen by CCC for the
was described in Case 1-2. period from January 1985 through March 1993 (see
The executive director, Marv Harnishfeger, Case 3-3). In Case 3-3, Dorothy used autocorrela-
concluded that the most important variable that tion analysis to explore the data pattern. In Case
CCC needed to forecast was the number of new 4-3, she used moving average and exponential
clients that would be seen for the rest of 1993. Marv smoothing methods to forecast the remaining
provided Dorothy Mercer with monthly data for months of 1993.
 CHAPTER 6 Simple Linear Regression 271

Year Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1985 110 108 105 104 104 102 103 101 102 102 99 99
1986 102 105 106 107 105 106 105 105 103 105 103 101
1987 108 105 105 108 104 108 105 105 103 105 104 104
1988 104 109 109 103 103 104 99 102 101 101 102 102
1989 107 106 109 105 104 103 103 106 105 106 107 99
1990 103 106 110 108 110 105 105 106 107 107 111 112
1991 114 116 118 119 118 120 120 121 119 121 119 120
1992 122 118 123 118 118 120 122 120 122 123 124 122
1993 125 125 130

Dorothy wondered if regression analysis could 1989 24,450 24,761 25,397 25,617 25,283 25,242
be used to develop a good forecasting model. She 25,163 25,184 25,417 25,411 25,565 26,543
asked Marv if he could think of any potential predic- 1990 26,784 27,044 27,567 28,080 28,142 28,412
28,161 27,936 28,423 28,366 29,029 29,035
tor variables. Marv felt that the number of people on
food stamps might be related to the number of new 1991 29,254 29,962 30,499 30,879 30,995 31,356
30,863 31,288 31,492 31,577 31,912 32,050
clients seen.
1992 32,383 32,625 33,499 34,076 34,191 33,788
Dorothy could find data for the number of peo- 33,556 33,751 33,777 33,769 34,077 34,232
ple on food stamps only from January 1989 to
December 1992. These data follow.
Marv was also acquainted with a business activ- activity index was an indicator of the relative
ity index computed for the county by the local changes in overall business conditions for the region.
Economic Development Council. The business The data for this index are at the top of the page.

QUESTIONS
1. Determine whether there is a significant rela- 4. Would the business activity index be a good pre-
tionship between the number of new clients dictor of the number of new clients?
seen and the number of people on food stamps 5. The data consist of a time series. Does this mean
and/or the business activity index. Don’t forget the independence assumption has been
the possibility of data transformations. violated?
2. Develop a regression equation and use it to 6. Assume that you developed a good regression
forecast the number of new clients for the first equation. Would you be able to use this equa-
three months of 1993. tion to forecast for the rest of 1993? Explain
3. Compare the results of your forecast with the your answer.
actual observations for the first three months of
1993.

CASE 6-6 AAA WASHINGTON

An overview of AAA Washington was provided in showed that the pattern Michael DeCoria had
Case 5-5 when students were asked to prepare a observed in emergency road service call volume was
time series decomposition of the emergency road probably somewhat cyclical in nature. Michael
service calls received by the club over five years. The would like to be able to predict emergency road
time series decomposition performed in Case 5-5 service call volume for future years.
272 CHAPTER 6 Simple Linear Regression

Other research done by the club identified sev- the number of members in the club. Finally, Michael
eral factors that have an impact on emergency road feels that the number of calls received is related to
service call volume. Among these factors are aver- the general economic cycle. The unemployment rate
age daily temperature and the amount of rainfall for Washington State is used as a good surrogate
received in a day. This research has shown that measurement for the general state of Washington’s
emergency road service calls increase as rainfall economy. Data on the unemployment rate, average
increases and as average daily temperature falls. The monthly temperature, monthly rainfall, and number
club also believes that the total number of emer- of members in the club have been gathered and are
gency road service calls it receives is dependent on presented in Table 6-15.

TABLE 6-15 Data for AAA Washington

Unemployment Temp.
Year Month Calls Rate (%) (deg, F.) Rain (in.) Members

1988 May 20,002 5.7867 55.1 3.75

June 21,591 5.7592 59 1.95
July 22,696 5.5718 63.8 0.89
August 21,509 5.2939 63.8 0.51 384,746
September 22,123 5.4709 59.1 2.31 388,652
October 21,449 5.5049 54.6 3.12 392,241
November 23,475 5.863 45.4 8.42 393,115
December 23,529 6.1349 41 4.44 392,631
1989 January 23,327 7.5474 40.3 4.3 396,975
February 24,050 7.8157 34.3 3.18 395,186
March 24,010 7.139 43.2 6.57 397,791
April 19,735 6.2637 52.5 2.39 397,764
May 20,153 5.8332 55.3 2.83 399,348
June 19,512 5.8077 62.4 1.3 401,949
July 19,892 5.6713 62.9 0.83 404,866
August 20,326 5.4977 63.5 1.53 405,341
September 19,378 5.2989 60.9 0.32 407,479
October 21,263 5.6028 51.9 3.44 405,430
November 21,443 5.9143 46.2 7.24 412,134
December 23,366 5 41.8 4.72 415,342
1990 January 23,836 6.1917 41.8 9.55 416,255
February 23,336 6.3775 38.9 5.73 423,001
March 22,003 5.7234 46.3 3.4 428,559
April 20,155 4.7792 51.7 2.91 431,429
May 20,070 4.5715 54.9 2.15 434,675
June 19,588 4.3899 59.8 3.55 435,864
July 20,804 4.2559 66.7 0.59 437,969
August 19,644 3.9359 66.4 1.33 440,565
September 17,424 3.9048 61.9 0.24 441,936
October 20,833 4.4294 50.4 1.17 448,595
November 22,490 5.1523 45.8 10.66 446,291
 CHAPTER 6 Simple Linear Regression 273

TABLE 6-15 (continued)

Unemployment Temp.
Year Month Calls Rate (%) (deg, F.) Rain (in.) Members
December 24,861 5.5102 33.9 7.93 446,455
1991 January 23,441 6.8901 37.9 4.4 445,392
February 19,205 7.0308 46.9 5.42 445,787
March 20,386 6.7186 43.4 4.35 445,746
April 19,988 6.128 49.1 5.69 446,430
May 19,077 5.8146 54.3 2.12 450,001
June 19,141 5.948 58.2 1.61 452,303
July 20,883 5.9026 65.4 0.51 456,551
August 20,709 5.7227 66 2.8 455,747
September 19,647 5.6877 60.9 0.2 456,764
October 22,013 6.2922 51 1.7 462,340
November 22,375 7.0615 46.2 6.5 460,492
December 22,727 7.437 42.4 3.45 465,361
1992 January 22,367 8.4513 43 7.26 465,492
February 21,155 8.7699 46 3.59 466,775
March 21,209 8.0728 48.9 1.47 467,168
April 19,286 7.2392 52.7 4.35 464,575
May 19,725 7.0461 58.3 0.6 459,019
June 20,276 7.0478 63.6 1.84 463,665
July 20,795 7.108 64.9 1.41 463,775
August 21,126 6.7824 65 1.01 466,230
September 20,251 6.7691 58.4 2.16
October 22,069 7.5896 53.2 2.55
November 23,268 7.9908 44.8 6.23
December 26,039 8.246 37.8 4.38
1993 January 26,127 9.5301 34.9 4.08
February 20,067 9.279
March 19,673 8.6802
April 19,142 7.7815

A conversation with the manager of the emer- temperature drops a few degrees from an average
gency road service call center has led to two important temperature in the 30s than it does when a similar
observations: (1) Automakers seem to design cars to drop occurs with an average temperature in the 60s.
operate best at 65 degrees Fahrenheit and (2) call vol- This information suggests that the effect of tempera-
ume seems to increase more sharply when the average ture on emergency road service is nonlinear.
274 CHAPTER 6 Simple Linear Regression

QUESTIONS
1. Run four simple linear regression models using 65 degrees should make vehicles operate less
total number of emergency road service calls as reliably. To accomplish a transformation of the
the dependent variable and unemployment rate, temperature data that simulates this effect,
temperature, rainfall, and number of members begin by subtracting 65 from the average
as the four independent variables. Would any of monthly temperature values. This repositions
these independent variables be useful for pre- “zero” to 65 degrees Fahrenheit. Should
dicting the total number of emergency road absolute values of this new temperature vari-
service calls? able be used?
2. Create a new temperature variable and relate it 3. Develop a scatter diagram. Is there a linear rela-
to emergency road service. Remember that tem- tionship between calls and the new temperature
perature is a relative scale and that the selection variable?
of the zero point is arbitrary. If vehicles are 4. If a nonlinear relationship exists between calls
designed to operate best at 65 degrees and the new temperature variable, develop the
Fahrenheit, then every degree above or below best model.

Minitab Applications

The problem. In Example 6.2, Mr. Bump wanted to run a regression analysis with the
data shown in Table 6-1.

Minitab Solution
1. Enter the data from Table 6-1 onto the worksheet: Gallons of milk sales go in col-
umn C1 and selling price in column C2.
2. In order to run a regression model, click on the following menus:
Stat>Regression>Regression
3. The Regression dialog box shown in Figure 6-18 appears.
a. Sales is selected as the Response, or dependent variable.
b. Price is selected as the Predictor, or independent variable.
4. In order to store residuals and Y estimates, click on Storage.
The Regression-Storage dialog box shown in Figure 6-19 appears.
a. Click on Residuals under Diagnostic Measures to store the residuals in C3.
b. Click on Fits under Characteristics of Estimated Equation to store the pre-
dicted values of Y in C4.
c. Click on OK to close the Regression-Storage dialog box.
5. In order to run residual plots, click on Graphs in the Regression dialog box.
The Regression-Graphs dialog box shown in Figure 6-20 appears.
a. Click on Four in one to include all four graphs.
b. Click on OK to close the Regression-Graphs dialog box.
c. Click on OK on the Regression dialog box, and the regression analysis dis-
played in Table 6-8 is presented in the session window and the graph shown in
Figure 6-9 appears on the screen.
The problem. In Example 6.11, Gilbert Garcia wanted to forecast sales using adver-
tising expenditures.

Minitab Solution
1. Enter the data from Figure 6-10 onto the worksheet: Sales go in column C1 and
advertising expenditures in column C2.
 CHAPTER 6 Simple Linear Regression 275

FIGURE 6-18 Minitab Regression Dialog Box

FIGURE 6-19 Minitab Regression-Storage Dialog Box

2. In order to develop the scatter diagram shown in Figure 6-11, click on

Graph>Scatterplot

3. A choice of Scatterplots appears. Select Simple, and click on OK.

4. The Scatterplot-Simple dialog box appears.
276 CHAPTER 6 Simple Linear Regression

FIGURE 6-20 Minitab Regression-Graphs Dialog Box

a. Select C1, Sales, as the Y variable and C2, Advertising Expenditures, as the X
variable.
b. Click on OK, and the scatter diagram shown in Figure 6-11 will appear.
5. In order to run a fitted model such as the one shown in Figure 6-12, click on
Stat>Regression>Fitted Line Plot

6. The Fitted Line Plot dialog box appears.

a. The dependent or Response (Y ) variable is Sales.
b. The independent or Predictor (X ) variable is Advertising Expenditures.
c. The type of regression model is Linear.
d. Click on OK. The results are shown in Figure 6-12.
7. Next, convert the X variable to the natural log of X by clicking on the following
menus:

Calc>Calculator

8. The Calculator dialog box shown in Figure 6-21 appears.

a. Enter C3 in the space next to Store result in variable.
b. In order to perform the transformation, highlight Natural log (log base e) on the
Functions: screen.
c. Click on Select, and the LN(number) under Select appears in the Expression: space.
d. Since Advertising Expenditures is the variable to be transformed, C2 replaces
“number” in this expression.
e. Click on OK, and the natural log of X appears in C3 of the data worksheet.
 CHAPTER 6 Simple Linear Regression 277

FIGURE 6-21 Minitab Calculator Dialog Box

9. Transformations for the square root of X, the square of X, and the reciprocal of X
are also accomplished using the Calculator dialog box.
10. The complete Minitab worksheet is presented in Figure 6-10.

Excel Applications

The problem. In Mr. Bump’s situation, Example 6.1, regression analysis is used to deter-
mine whether selling price could be used to forecast weekly sales for gallons of milk.

Excel Solution
1. Enter weekly sales (see Table 6-1) into A1 through A10 and selling price into B1
through B10 of the worksheet.
2. Click on the following menus to perform regression analysis:
Tools>Data Analysis
3. The Data Analysis dialog box appears. Under Analysis Tools, choose Regression,
and click on OK. The Regression dialog box shown in Figure 6-22 appears.
a. Enter A1:A10 in the Input Y Range.
b. Enter B1:B10 in the Input X Range.
c. Click on Output Range, and enter C1 in the next space.
d. Click on OK, and the output presented in Figure 6-23 appears.
278 CHAPTER 6 Simple Linear Regression

FIGURE 6-22 Excel Regression Dialog Box

FIGURE 6-23 Excel Regression Output for Example 6.1

 CHAPTER 6 Simple Linear Regression 279

References
Abraham, B., and J. Ledolter. Introduction to Kutner, M. H., C. J. Nachtsheim, and J. Neter.
Regression Modeling. Belmont, Calif.: Thomson Applied Linear Regression Models, 4th ed. New
Brooks/Cole, 2006. York: McGraw-Hill, 2004.
Draper, N., and H. Smith. Applied Regression Moore, D. S., G. P. McCabe, W. M. Duckworth, and
Analysis, 3rd ed. New York: Wiley, 1998. S. L. Sclove. The Practice of Business Statistics.
Flaherty, W. P. “Using Regression Analysis to Pick the New York: Freeman, 2003.
Best Targets,” M&A (March–April 1991): 47–49.
This page intentionally left blank
C H A P T E R

7 MULTIPLE REGRESSION
ANALYSIS
In simple linear regression, the relationship between a single independent variable and
a dependent variable is investigated. The relationship between two variables frequently
allows one to accurately predict the dependent variable from knowledge of the inde-
pendent variable. Unfortunately, many real-life forecasting situations are not so simple.
More than one independent variable is usually necessary in order to predict a depen-
dent variable accurately. Regression models with more than one independent variable
are called multiple regression models. Most of the concepts introduced in simple linear
regression carry over to multiple regression. However, some new concepts arise because
more than one independent variable is used to predict the dependent variable.

Multiple regression involves the use of more than one independent variable to
predict a dependent variable.

SEVERAL PREDICTOR VARIABLES

As an example, return to the problem in which sales volume of gallons of milk is fore-
cast from knowledge of price per gallon. Mr. Bump is faced with the problem of making
a prediction that is not entirely accurate. He can explain almost 75% of the differences
in gallons of milk sold by using one independent variable. Thus, 25% 11 - r22 of the
total variation is unexplained. In other words, from the sample evidence Mr. Bump
knows 75% of what he must know to forecast sales volume perfectly. To do a more accu-
rate job of forecasting, he needs to find another predictor variable that will enable him
to explain more of the total variation. If Mr. Bump can reduce the unexplained varia-
tion, his forecast will involve less uncertainty and be more accurate.
A search must be conducted for another independent variable that is related to sales
volume of gallons of milk. However, this new independent, or predictor, variable cannot
relate too highly to the independent variable (price per gallon) already in use. If the two
independent variables are highly related to each other, they will explain the same varia-
tion, and the addition of the second variable will not improve the forecast.1 In fields such
as econometrics and applied statistics, there is a great deal of concern with this problem
of intercorrelation among independent variables, often referred to as multicollinearity.

1Interrelated predictor variables essentially contain much of the same information and therefore do not
contribute “new” information about the behavior of the dependent variable. Ideally, the effects of separate
predictor variables on the dependent variable should be unrelated to one another.

281
282 CHAPTER 7 Multiple Regression Analysis

The simple solution to the problem of two highly related independent variables is merely
not to use both of them together. The multicollinearity problem will be discussed later in
this chapter.

CORRELATION MATRIX
Mr. Bump decides that advertising expense might help improve his forecast of weekly
sales volume. He investigates the relationships among advertising expense, sales vol-
ume, and price per gallon by examining a correlation matrix. The correlation matrix is
constructed by computing the simple correlation coefficients for each combination of
pairs of variables.
An example of a correlation matrix is illustrated in Table 7-1. The correlation coef-
ficient that indicates the relationship between variables 1 and 2 is represented as r12.
Note that the first subscript, 1, also refers to the row and the second subscript, 2, also
refers to the column in the table. This approach allows one to determine, at a glance, the
relationship between any two variables. Of course, the correlation between, say, vari-
ables 1 and 2 is exactly the same as the correlation between variables 2 and 1; that is,
r12 = r21. Therefore, only half of the correlation matrix is necessary. In addition, the cor-
relation of a variable with itself is always 1, so that, for example, r11 = r22 = r33 = 1.
Mr. Bump runs his data on the computer, and the correlation matrix shown in
Table 7-2 results. An investigation of the relationships among advertising expense,
sales volume, and price per gallon indicates that the new independent variable should
contribute to improved prediction. The correlation matrix shows that advertising
expense has a high positive relationship 1r13 = .892 with the dependent variable, sales
volume, and a moderate negative relationship 1r23 = - .652 with the independent vari-
able, price per gallon. This combination of relationships should permit advertising
expenses to explain some of the total variation of sales volume that is not already being
explained by price per gallon. As will be seen, when both price per gallon and advertis-
ing expense are used to estimate sales volume, R 2 increases to 93.2%.
The analysis of the correlation matrix is an important initial step in the solution of
any problem involving multiple independent variables.

TABLE 7-1 Correlation Matrix

Variables

Variables 1 2 3

1 r11 r12 r13

2 r21 r22 r23
3 r31 r32 r33

TABLE 7-2 Correlation Matrix for Mr. Bump’s data

Variables
Sales Price Advertising
Variables
1 2 3
Sales, 1 1.00 -.86 .89
Price, 2 1.00 -.65
Advertising, 3 1.00
 CHAPTER 7 Multiple Regression Analysis 283

MULTIPLE REGRESSION MODEL

In simple regression, the dependent variable can be represented by Y and the inde-
pendent variable by X. In multiple regression analysis, X’s with subscripts are used to
represent the independent variables. The dependent variable is still represented by Y,
and the independent variables are represented by X1, X2, . . ., Xk. Once the initial set of
independent variables has been determined, the relationship between Y and these X’s
can be expressed as a multiple regression model.
In the multiple regression model, the mean response is taken to be a linear func-
tion of the explanatory variables:
mY = b 0 + b 1X1 + b 2X2 + Á + b kXk (7.1)
This expression is the population multiple regression function. As was the case with
simple linear regression, we cannot directly observe the population regression function
because the observed values of Y vary about their means. Each combination of values
for all of the X’s defines the mean for a subpopulation of responses Y. We assume that
the Y’s in each of these subpopulations are normally distributed about their means
with the same standard deviation, σ.
The data for simple linear regression consist of observations 1Xi, Yi2 on the two
variables. In multiple regression, the data for each case consist of an observation on the
response and an observation on each of the independent variables. The ith observation
on the jth predictor variable is denoted by Xi j. With this notation, data for multiple
regression have the form given in Table 7-3. It is convenient to refer to the data for the
ith case as simply the ith observation. With this convention, n is the number of observa-
tions and k is the number of predictor variables.

Statistical Model for Multiple Regression

The response, Y, is a random variable that is related to the independent (predictor)
variables, X1,X2, Á ,Xk, by
Y = b 0 + b 1X1 + b 2X2 + Á + b kXk +

where
1. For the ith observation, Y = Yi and X1, X2, Á , Xk are set at values Xi 1, Xi 2, Á , Xik.

TABLE 7-3 Data Structure for Multiple

Regression

Predictor Variables Response

Case X1 X2 ............ Xk Y

1 X11 X12 ............ X1k Y1

2 X21 X22 ............ X2k Y2
. . . ............ . .
. . . ............ . .
i Xi1 Xi2 ............ Xik Yi
. . . ............ . .
. . . ............ . .
n Xn1 Xn2 ............ Xnk Yn
284 CHAPTER 7 Multiple Regression Analysis

2. The ’s are error components that represent the deviations of the response from
the true relation. They are unobservable random variables accounting for the
effects of other factors on the response. The errors are assumed to be independent,
and each is normally distributed with mean 0 and unknown standard deviation σ.
3. The regression coefficients, b 0,b 1, Á ,b k, that together locate the regression function
are unknown.
Given the data, the regression coefficients can be estimated using the principle of
least squares. The least squares estimates are denoted by b0, b1, Á , bk and the estimated
regression function by

N = b + bX + Á + bX
Y (7.2)
0 1 1 k k

The residuals, e = Y - YN , are estimates of the error component and similar to the sim-
ple linear regression situation; the correspondence between population and sample is

Population: Y = b 0 + b 1X1 + b 2X2 + Á + b kXk +

Sample: Y = b0 + b1X1 + b2X2 + Á + bkXk + e

The calculations in multiple regression analysis are ordinarily performed using

computer software packages such as Minitab and Excel (see the Minitab and Excel
Applications sections at the end of the chapter).

Example 7.1
For the data shown in Table 7-4, Mr. Bump considers a multiple regression model relating
sales volume (Y) to price 1X12 and advertising 1X22:

Y = b 0 + b 1X1 + b 2X2 +

Mr. Bump determines the fitted regression function:

YN = 16.41 - 8.25X1 + .59X2

The least squares values— b0 = 16.41, b1 = - 8.25, and b2 = .59—minimize the sum of
squared errors:

SSE = a 1Yi - b0 - b1Xi1 - b2Xi222 = a 1Yi - YN i22

i i

for all possible choices of b0 , b1 , and b2. Here, the best-fitting function is a plane (see
Figure 7-1). The data points are plotted in three dimensions along the Y, X1, and X2 axes.
The points fall above and below the plane in such a way that ©1Y - YN 22 is a minimum.
The fitted regression function can be used to forecast next week’s sales. If plans call for
a price per gallon of $1.50 and advertising expenditures of $1,000, the forecast is 9.935 thou-
sands of gallons; that is,

YN = 16.41 - 8.25X1 + .59X2

YN = 16.41 - 8.2511.52 + .591102 = 9.935
 CHAPTER 7 Multiple Regression Analysis 285

TABLE 7-4 Mr. Bump’s Data for Example 7.1

Week Sales (1,000s) Price per Gallon ($) Advertising ($100s)

Y X1 X2

1 10 1.30 9
2 6 2.00 7
3 5 1.70 5
4 12 1.50 14
5 10 1.60 15
6 15 1.20 12
7 5 1.60 6
8 12 1.40 10
9 17 1.00 15
10 20 1.10 21
Totals 112 14.40 114
Means 11.2 1.44 11.4

Y
Sales

17
16.41
^
Y = 16.41 − 8.25 X1 + .59X2
B

X2
A 6 15 20 Advertising
$1.00
$1.60
$2.00
X1
Price
Point Week Sales Price Advertising
A 7 5 $1.60 6
B 9 17 $1.00 15

FIGURE 7-1 Fitted Regression Plane for Mr. Bump’s Data for Example 7.1

INTERPRETING REGRESSION COEFFICIENTS

Consider the interpretation of b0, b1, and b2 in Mr. Bump’s fitted regression function.
The value b0 is again the Y-intercept. However, now it is interpreted as the value of YN
when both X1 and X2 are equal to zero. The coefficients b1 and b2 are referred to as the
partial, or net, regression coefficients. Each measures the average change in Y per unit
change in the relevant independent variables. However, because the simultaneous
286 CHAPTER 7 Multiple Regression Analysis

influence of all independent variables on Y is being measured by the regression func-

tion, the partial or net effect of X1 (or any other X) must be measured apart from any
influence of other variables. Therefore, it is said that b1 measures the average change in
Y per unit change in X1, holding the other independent variables constant.

The partial, or net, regression coefficient measures the average change in the
dependent variable per unit change in the relevant independent variable, holding
the other independent variables constant.

In the present example, the b1 value of - 8.25 indicates that each increase of 1 cent
in the price of a gallon of milk when advertising expenditures are held constant reduces
the quantity purchased by an average of 82.5 gallons. Similarly, the b2 value of .59
means that, if advertising expenditures are increased by $100 when the price per gallon
is held constant, then sales volume will increase an average of 590 gallons.
Example 7.2
To illustrate the net effects of individual X’s on the response, consider the situation in which
price is to be $1.00 per gallon and $1,000 is to be spent on advertising. Then

YN = 16.41 - 8.25X1 + .59X2

= 16.41 - 8.2511.002 + .591102
= 16.41 - 8.25 + 5.9 = 14.06

Sales are forecast to be 14,060 gallons of milk.

What is the effect on sales of a 1-cent price increase if $1,000 is still spent on advertising?

YN = 16.41 - 8.2511.012 + .591102

= 16.41 - 8.3325 + 5.9 = 13.9775

Note that sales decrease by 82.5 gallons 114.06 - 13.9775 = .08252.

What is the effect on sales of a $100 increase in advertising if price remains constant
at $1.00?

YN = 16.41 - 8.2511.002 + .591112

= 16.41 - 8.25 + 6.49 = 14.65

Note that sales increase by 590 gallons 114.65 - 14.06 = .592.

INFERENCE FOR MULTIPLE REGRESSION MODELS

Inference for multiple regression models is analogous to that for simple linear regres-
sion. The least squares estimates of the model parameters, their estimated standard
errors, the t statistics used to examine the significance of individual terms in the regres-
sion model, and the F statistic used to check the significance of the regression are all
provided in output from standard statistical software packages. Determining these
quantities by hand for a multiple regression analysis of any size is not practical, and the
computer must be used for calculations.
As in Chapter 6, any observation Y can be written

Observation = Fit + Residual

 CHAPTER 7 Multiple Regression Analysis 287
or
Y = YN + 1Y - YN 2
where
YN = b0 + b1X1 + b2X2 + Á + bkXk
is the fitted regression function. Recall that YN is an estimate of the population regres-
sion function. It represents that part of Y explained by the relation of Y with the X’s.
The residual, Y - YN , is an estimate of the error component of the model. It represents
that part of Y not explained by the predictor variables.
The sum of squares decomposition and the associated degrees of freedom are

©1Y - Y22 = ©1YN - Y22 + ©1Y - YN 22

SST = SSR + SSE
df : n - 1 = k + n - k - 1 (7.3)

The total variation in the response, SST, consists of two components: SSR, the variation
explained by the predictor variables through the estimated regression function, and
SSE, the unexplained or error variation. The information in Equation 7.3 can be set out
in an analysis of variance (ANOVA) table, which is discussed in a later section.

Standard Error of the Estimate

The standard error of the estimate is the standard deviation of the residuals. It mea-
sures the typical scatter of the Y values about the fitted regression function.2 The stan-
dard error of the estimate is

©1Y - YN 22 SSE
sy #x¿s = = = 1MSE (7.4)
Dn - k - 1 Dn - k - 1

where
n = the number of observations
k = the number of independent variables in the regression function
SSE = ©1Y - YN 22 = the residual sum of squares
MSE = SSE>1n - k - 12 = the residual mean square

The standard error of the estimate is the standard deviation of the residuals. It mea-
sures the amount the actual values (Y) differ from the estimated values (YN ). For rel-
atively large samples, we would expect about 67% of the differences Y - YN to be
within sy #x¿s of zero and about 95% of these differences to be within 2 sy #x¿s of zero.

Example 7.3
The quantities required to calculate the standard error of the estimate for Mr. Bump’s data
are given in Table 7-5.

2The standard error of the estimate is an estimate of s, the standard deviation of the error term, , in the
multiple regression model.
288 CHAPTER 7 Multiple Regression Analysis

TABLE 7-5 Residuals from the Model for Mr. Bump’s Data for
Example 7.3

N Using
Predicted Y (Y) Residual
Y X1 X2 YN  16.406  8.248X1  .585X2 (Y - YN ) (Y - YN ) 2

10 1.30 9 10.95 -.95 .90

6 2.00 7 4.01 1.99 3.96
5 1.70 5 5.31 -.31 .10
12 1.50 14 12.23 -.23 .05
10 1.60 15 11.99 -1.99 3.96
15 1.20 12 13.53 1.47 2.16
5 1.60 6 6.72 -1.72 2.96
12 1.40 10 10.71 1.29 1.66
17 1.00 15 16.94 .06 .00
20 1.10 21 19.62 .38 .14
Totals .00 15.90

The standard error of the estimate is

15.90
sy #x¿s = = 22.27 = 1.51
A 10 - 2 - 1
With a single predictor, X1 = price, the standard error of the estimate was sy #x = 2.72
(see Chapter 6). With the additional predictor, X2 = advertising, Mr. Bump has reduced the
standard error of the estimate by almost 50%. The differences between the actual volumes
of milk sold and their forecasts obtained from the fitted regression equation are consider-
ably smaller with two predictor variables than they were with a single predictor. That is, the
two-predictor equation comes a lot closer to reproducing the actual Y ’s than the single-
predictor equation.

Significance of the Regression

The ANOVA table based on the decomposition of the total variation in Y (SST) into its
explained (SSR) and unexplained (SSE) parts (see Equation 7.3) is given in Table 7-6.
Consider the hypothesis H0: b 1 = b 2 = Á = b k = 0. This hypothesis means that
Y is not related to any of the X’s (the coefficient attached to every X is zero). A test of
H0 is referred to as a test of the significance of the regression. If the regression model
assumptions are appropriate and H0 is true, the ratio
MSR
F =
MSE
has an F distribution with df = k, n - k - 1. Thus, the F ratio can be used to test the
significance of the regression.

TABLE 7-6 ANOVA Table for Multiple Regression

Source Sum of Squares df Mean Square F Ratio

MSR
Regression SSR k MSR = SSR>k F =
MSE
Error SSE n - k - 1 MSE = SSE>(n - k - 12
Total SST n - 1
 CHAPTER 7 Multiple Regression Analysis 289
In simple linear regression, there is only one predictor variable. Consequently, test-
ing for the significance of the regression using the F ratio from the ANOVA table is
equivalent to the two-sided t test of the hypothesis that the slope of the regression line
is zero. For multiple regression, the t tests (to be introduced shortly) examine the sig-
nificance of individual X’s in the regression function, and the F test examines the
significance of all the X’s collectively.

F Test for the Significance of the Regression

In the multiple regression model, the hypotheses
H0: b 1 = b 2 = Á = b k = 0
H1: at least one b j Z 0
are tested by the F ratio:
MSR
F =
MSE
with df = k, n - k - 1. At significance level α, the rejection region is
F 7 Fa
where Fa is the upper α percentage point of an F distribution with 1 = k,
2 = n - k - 1 degrees of freedom.

The coefficient of determination, R2, is given by

SSR ©1YN - Y22

R =
2
=
SST ©1Y - Y22

SSE ©1Y - YN 22
= 1 - = 1 - (7.5)
SST ©1Y - Y22

and has the same form and interpretation as r2 does for simple linear regression. It rep-
resents the proportion of variation in the response, Y, explained by the relationship of
Y with the X’s.
A value of R 2 = 1 says that all the observed Y’s fall exactly on the fitted regression
function. All of the variation in the response is explained by the regression. A value of
R2 = 0 says that YN = Y—that is, SSR = 0—and none of the variation in Y is explained
by the regression. In practice, 0 … R2 … 1, and the value of R 2 must be interpreted relative
to the extremes, 0 and 1.
The quantity

R = 2R 2 (7.6)

is called the multiple correlation coefficient and is the correlation between the
responses, Y, and the fitted values, YN . Since the fitted values predict the responses, R is
always positive, so that 0 … R … 1.
290 CHAPTER 7 Multiple Regression Analysis

For multiple regression,

R2 n - k - 1
F = ¢ ≤ (7.7)
1 - R2 k

so, everything else equal, significant regressions (large F ratios) are associated with rel-
atively large values for R2.
The coefficient of determination 1R22 can always be increased by adding an addi-
tional independent variable, X, to the regression function, even if this additional vari-
able is not important.3 For this reason, some analysts prefer to interpret R2 adjusted for
the number of terms in the regression function. The adjusted coefficient of determina-
tion, R 2, is given by

n - 1
R 2 = 1 - 11 - R22 ¢ ≤ (7.8)
n - k - 1

Like R2, R2 is a measure of the proportion of variability in the response, Y, explained

by the regression. It can be shown that 0 … R2 … R2. When the number of observa-
tions (n) is large relative to the number of independent variables (k), R2 L R2. If k = 0,
YN = Y and R2 = R2. In many practical situations, there is not much difference between
the magnitudes of R2 and R2.

Example 7.4
Using the total sum of squares in Table 7-6 and the residual sum of squares from Example 7.3,
the sum of squares decomposition for Mr. Bump’s problem is

SST = SST + SSE

a 1Y - Y2 = a 1Y - Y2 + a 1Y - Y2
2 N N 2
2

233.6 = 217.7 + 15.9

Hence, using both forms of Equation 7.5 to illustrate the calculations,

217.7 15.9
R2 = = 1 - = .932
233.6 233.6

and the multiple correlation coefficient is R = 2R2 = 2.932 = .965.

Here, about 93% of the variation in sales volume is explained by the regression, that is,
the relation of sales to price and advertising expenditures. In addition, the correlation
between sales and fitted sales is about .965, indicating close agreement between the actual
and predicted values. A summary of the analysis of Mr. Bump’s data to this point is given in
Table 7-7.

Individual Predictor Variables

The coefficient of an individual X in the regression function measures the partial or net
effect of that X on the response, Y, holding the other X’s in the equation constant. If the
regression is judged significant, then it is of interest to examine the significance of
the individual predictor variables. The issue is this: Given the other X’s, is the effect of
this particular X important, or can this X term be dropped from the regression func-
tion? This question can be answered by examining an appropriate t value.

3Here, “not important” means “not significant.” That is, the coefficient of X is not significantly different from
zero (see the Individual Predictor Variables section that follows).
 CHAPTER 7 Multiple Regression Analysis 291

TABLE 7-7 Summary of the Analysis of

Mr. Bump’s Data for Example 7.4
Variables Used to Explain
Variability in Y R2 π(Y - YN )2

None .00 233.6

Price .75 59.4
Price and Advertising expense .93 15.9

If H0: b j = 0 is true, the test statistic, t, with the value t = bj>sbj has a t distribution
with df = n - k - 1.4

To judge the significance of the jth term, j = 0, 1, Á , k, in the regression function,

the test statistic, t, is compared with a percentage point of a t distribution with
n - k - 1 degrees of freedom. For an α level test of
H0: b j = 0
H1: b j Z 0

reject H0 if ƒ t ƒ 7 ta>2 . Here, ta>2 is the upper α/2 percentage point of a t distribution
with df = n - k - 1.
Some care must be exercised in dropping from the regression function those pre-
dictor variables that are judged to be insignificant by the t test 1H0: b j = 0 cannot be
rejected). If the X’s are related (multicollinear), the least squares coefficients and the
corresponding t values can change, sometimes appreciably, if a single X is deleted
from the regression function. For example, an X that was previously insignificant may
become significant. Consequently, if there are several small (insignificant) t values,
predictor variables should be deleted one at a time (starting with the variable having
the smallest t value) rather than in bunches. The process stops when the regression is
significant and all the predictor variables have large (significant) t statistics.

Forecast of a Future Response

A forecast, YN *, of a future response, Y, for new values of the X’s—say, X1 = X*1,
X2 = X*2, Á , Xk = X*k—is given by evaluating the fitted regression function at the X*’s:

YN * = b0 + b1X*1 + b2X*2 + Á + bkX*k (7.9)

With confidence level 1 - α, a prediction interval for Y takes the form

YN * ; ta>2 * 1Standard error of the forecast2
The standard error of the forecast is a complicated expression, but the standard error of
the estimate, sy # x’s, is an important component. In fact, if n is large and all the X’s are quite
variable, an approximate 10011 - α)% prediction interval for a new response Y is

1YN * - ta>2 sy #x’s , YN * + ta>2 sy #x’s2 (7.10)

4Here, b is the least squares coefficient for the jth predictor variable, X , and s is its estimated standard
j j bj
deviation (standard error). These two statistics are ordinarily obtained with computer software such as
Minitab.
292 CHAPTER 7 Multiple Regression Analysis

COMPUTER OUTPUT
The computer output for Mr. Bump’s problem is presented in Table 7-8. Examination of
this output leads to the following observations (explanations are keyed to Table 7-8).
1. The regression coefficients are - 8.25 for price and .585 for advertising expense.The
fitted regression equation is YN = 16.4 - 8.25X1 + .585X2.
2. The regression equation explains 93.2% of the variation in sales volume.
3. The standard error of the estimate is 1.5072 gallons. This value is a measure of the
amount the actual values differ from the fitted values.
4. In Chapter 6, the regression slope coefficient was tested to determine whether it
was different from zero. In the current situation, the large t statistic of - 3.76 for
the price variable, X1, and its small p-value (.007) indicate that the coefficient of
price is significantly different from zero (reject H0: b 1 = 02. Given the advertising
variable, X2, in the regression function, price cannot be dropped from the regres-
sion function. Similarly, the large t statistic of 4.38 for the advertising variable, X2,
and its small p-value (.003) indicate that the coefficient of advertising is signifi-
cantly different from zero (reject H0: b 2 = 02. Given the price variable, X1, in the
regression function, the advertising variable cannot be dropped from the regres-
sion function. (As a reference point for the magnitude of the t values, with seven
degrees of freedom, Table B-3 gives t.01 = 2.998.) In summary, the coefficients of
both predictor variables are significantly different from zero.
5. The p-value .007 is the probability of obtaining a t value at least as large as - 3.76 if
the hypothesis H0: b 1 = 0 is true. Since this probability is extremely small, H0 is
unlikely to be true, and it is rejected.The coefficient of price is significantly different
from zero. The p-value .003 is the probability of obtaining a t value at least as large
as 4.38 if H0: b 2 = 0 is true. Since a t value of this magnitude is extremely unlikely,
H0 is rejected. The coefficient of advertising is significantly different from zero.

TABLE 7-8 Minitab Output for Mr. Bump’s Data

Correlations: Y, X1, X2
Y X1
X1 -0.863
X2 0.891 -0.654 162
Regression Analysis: Y versus X1, X2
The regression equation is
Y = 16.4 - 8.25 X1 + 0.585 X2 112
Presictor Coef SE Coef T P
Constant 16.406 (1) 4.343 3.78 0.007
X1 -8.248 112 2.196 -3.76 142 0.007 (5)
X2 0.5851 (1) 0.1337 4.38 (4) 0.003 (5)
S = 1.50720 132 R-Sq = 93.2% 122 R -Sq1adj2 = 91.2% 192
Analysis of Variance
Source DF SS MS F P
Regression 2 217.70 (7) 108.85 47.92 (8) 0.000
Residual Error 7 15.90 (7) 2.27
Total 9 233.60 (7)
 CHAPTER 7 Multiple Regression Analysis 293
6. The correlation matrix was demonstrated in Table 7-2.
7. The sum of squares decomposition, SST = SSR + SSE 1sum of squares total =
sum of squares regression + sum of squares error2, was given in Example 7.4.
8. The computed F value (47.92) is used to test for the significance of the regression.
The large F ratio and its small p-value (.000) show the regression is significant
(reject H0: b 1 = b 2 = 02. The F ratio is computed from
MSR 108.85
F = = = 47.92
MSE 2.27
As a reference for the magnitude of the F ratio, Table B-5 gives the upper 1%
point of an F distribution with two and seven degrees of freedom as F.01 = 9.55.
The regression function explains a significant amount of the variability in sales, Y.
9. The computation for the corrected or adjusted R2, R 2, is
n - 1
R2 = 1 - 11 - R22 ¢ ≤
n - k - 1
10 - 1
= 1 - 11 - .9322 ¢ ≤ = 1 - 1.068211.2862 = .912
10 - 2 - 1

DUMMY VARIABLES
Consider the following example.
Example 7.5
Suppose an analyst wishes to investigate how well a particular aptitude test predicts job per-
formance. Eight women and seven men have taken the test, which measures manual dexter-
ity in using the hands with tiny objects. Each subject then went through a month of intensive
training as an electronics assembler, followed by a month of actual assembly, during which
his or her productivity was evaluated by an index having values ranging from 0 to 10 (zero
means unproductive).
The data are shown in Table 7-9. A scatter diagram is presented in Figure 7-2. Each
female worker is represented by a 0 and each male by a 1.
It is immediately evident from observing Figure 7-2 that the relationship of this apti-
tude test to job performance follows two distinct patterns, one applying to women and the
other to men.

It is sometimes necessary to determine how a dependent variable is related to an

independent variable when a qualitative factor is influencing the situation. This relation-
ship is accomplished by creating a dummy variable. There are many ways to identify
quantitatively the classes of a qualitative variable.The values 0 and 1 are used in this text.

Dummy, or indicator, variables are used to determine the relationships between

qualitative independent variables and a dependent variable.

The dummy variable technique is illustrated in Figure 7-3. The data points for
females are shown as 0’s; the 1’s represent males. Two parallel lines are constructed for
the scatter diagram. The top one fits the data for females; the bottom one fits the male
data points.
Each of these lines was obtained from a fitted regression function of the form

YN = b0 + b1X1 + b2X2
294 CHAPTER 7 Multiple Regression Analysis

TABLE 7-9 Electronics Assemblers Dummy Variable

Data for Example 7.5

Job Performance Rating Aptitude Test Score Gender

Subject Y X1 X2

1 5 60 0(F)
2 4 55 0(F)
3 3 35 0(F)
4 10 96 0(F)
5 2 35 0(F)
6 7 81 0(F)
7 6 65 0(F)
8 9 85 0(F)
9 9 99 1(M)
10 2 43 1(M)
11 8 98 1(M)
12 6 91 1(M)
13 7 95 1(M)
14 3 70 1(M)
15 6 85 1(M)
Totals 87 1,093
YF = the mean female job performance rating = 5.75
YM = the mean male job performance rating = 5.86
XF = the mean female aptitude test score = 64
XM = the mean male aptitude test score = 83

Y
10 0
0 = Females
1 = Males
9 0 1

8 1
Job Performance Rating

7 0 1

6 0 1 1

5 0

4 0

3 0 1

2 0 1

X
0 10 20 30 40 50 60 70 80 90 100
Aptitude Test Score

FIGURE 7-2 Scatter Diagram for Data in Example 7.5

 CHAPTER 7 Multiple Regression Analysis 295

Y
10 0
^
9 0 = Females 0 Y = −1.96 + .12X1
1 = Males

8 1

0 1

Job Performance Rating

7 ^
Y = −4.14 + .12X1
6 0 1 1

5 0

4 0

3 0 1

2 0 1

X1
0 10 20 30 40 50 60 70 80 90 100
Aptitude Test Score

FIGURE 7-3 Regression Lines Corresponding to Dummy Variables

in Example 7.5

where
X1 = the test score
0 for females
X2 = b dummy variable
1 for males
The single equation is equivalent to the following two equations:

YN = b0 + b1X1 for females

YN = b0 + b1X1 + b2 = 1b0 + b22 + b1X1 for males

Note that b2 represents the effect of a male on job performance and that b1 represents
the effect of differences in aptitude test scores (the b1 value is assumed to be the same
for both males and females). The important point is that one multiple regression equa-
tion will yield the two estimated lines shown in Figure 7-3. The top line is the estimated
relation for females, and the lower line is the estimated relation for males. One might
envisage X2 as a “switching” variable that is “on” when an observation is made for a
male and “off” when it is made for a female.

Example 7.6
The estimated multiple regression equation for the data of Example 7.5 is shown in the
Minitab computer output in Table 7-10. It is

YN = - 1.96 + .12X1 - 2.18X2

296 CHAPTER 7 Multiple Regression Analysis

TABLE 7-10 Minitab Output for Example 7.6

Correlations: Ratings, Test, Gender

Rating Test
Test 0.876
Gender 0.021 0.428
Regression Analysis: Rating versus Test, Gender
The regression equation is
Rating = - 1.96 + 0.120 Test -2.18 Gender
Predictor Coef SE Coef T P
Constant -1.9565 0.7068 -2.77 0.017
Test 0.12041 0.01015 11.86 0.000
Gender -2.1807 0.4503 -4.84 0.000

S = 0.7863 R-sq = 92.1% R -sq1adj2 = 90.8%

Analysis of Variance
Source DF SS MS F P
Regression 2 86.981 43.491 70.35 0.000
Residual Error 12 7.419 0.618
Total 14 94.400

For the two values (0 and 1) of X2, the fitted equation becomes

YN = - 1.96 + .12X1 - 2.18102 = - 1.96 + .12X1 for females

and

YN = - 1.96 + .12X1 - 2.18112 = - 4.14 + .12X1 for males

These equations may be interpreted in the following way: The regression coefficient
value b1 = .12, which is the slope of each of the lines, is the estimated average increase in
performance rating for each one-unit increase in aptitude test score. This coefficient applies
to both males and females.
The other regression coefficient, b2 = - 2.18, applies only to males. For a male test
taker, the estimated job performance rating is reduced, relative to female test takers, by 2.18
units when the aptitude score is held constant.
An examination of the means of the Y and X1 variables, classified by gender, helps one
understand this result. Table 7-9 shows that the mean job performance ratings were approx-
imately equal for males, 5.86, and females, 5.75. However, the males scored significantly
higher (83) on the aptitude test than did the females (64). Therefore, if two applicants, one
male and one female, took the aptitude test and both scored 70, the female’s estimated job
performance rating would be 2.18 points higher than the male’s, since

Female: YN = - 1.96 + .12X1 = - 1.96 + .121702 = 6.44

Male: YN = - 4.14 + .12X1 = - 4.14 + .121702 = 4.26

A look at the correlation matrix in Table 7-10 provides some interesting insights.
A strong linear relationship exists between job performance and the aptitude test because
r12 = .876 . If the aptitude test score alone were used to predict performance, it would
explain about 77% 1.8762 = .7672 of the variation in job performance scores.
The correlation coefficient r13 = .02 indicates virtually no relationship between gender
and job performance. This conclusion is also evident from the fact that the mean perform-
ance ratings for males and females are nearly equal (5.86 versus 5.75). At first glance, one
might conclude that knowledge of whether an applicant is male or female is not useful
 CHAPTER 7 Multiple Regression Analysis 297
information. However, the moderate relationship, r23 = .428, between gender and aptitude
test score indicates that the test might discriminate between sexes. Males seem to do better
on the test than do females (83 versus 64). Perhaps some element of strength is required on
the test that is not required on the job.
When both test results and gender are used to forecast job performance, 92% of the vari-
ance is explained. This result suggests that both variables make a valuable contribution to pre-
dicting performance. The aptitude test scores explain 77% of the variance, and gender used in
conjunction with the aptitude test scores adds another 15%. The computed t statistics, 11.86
1p-value = .0002 and -4.84 1p-value = .0002, for aptitude test score and gender, respec-
tively, indicate that both predictor variables should be included in the final regression function.

MULTICOLLINEARITY
In many regression problems, data are routinely recorded rather than generated from
preselected settings of the independent variables. In these cases, the independent vari-
ables are frequently linearly dependent or multicollinear. For example, in appraisal
work, the selling price of a home may be related to predictor variables such as age, liv-
ing space in square feet, number of bathrooms, number of rooms other than bath-
rooms, lot size, and an index of construction quality. Living space, number of rooms,
and number of bathrooms should certainly “move together.” If one of these variables
increases, the others will generally increase.
If this linear dependence is less than perfect, the least squares estimates of the
regression model coefficients can still be obtained. However, these estimates tend to be
unstable—their values can change dramatically with slight changes in the data—and
inflated—their values are larger than expected. In particular, individual coefficients
may have the wrong sign, and the t statistics for judging the significance of individual
terms may all be insignificant, yet the F test will indicate the regression is significant.
Finally, the calculation of the least squares estimates is sensitive to rounding errors.

Multicollinearity is the situation in which independent variables in a multiple

regression equation are highly intercorrelated. That is, a linear relation exists
between two or more independent variables.

The strength of the multicollinearity is measured by the variance inflation factor

(VIF):5
1
VIFj = j = 1, 2, Á , k (7.11)
1 - R2j
Here, R 2j is the coefficient of determination from the regression of the jth independent
variable on the remaining k - 1 independent variables. For k = 2 independent vari-
2
ables, R j is the square of their sample correlation, r.
If the jth predictor variable, Xj, is not related to the remaining X ’s, R 2j = 0 and VIF j = 1.
If there is a relationship, then VIFj>1. For example, when R 2j is equal to .90,
VIFj = 1>11 - .902 = 10.
A VIF near 1 suggests that multicollinearity is not a problem for that independent
variable. Its estimated coefficient and associated t value will not change much as the other
independent variables are added or deleted from the regression equation. A VIF much

5The variance inflation factor (VIF) gets its name from the fact that sbj r VIFj. The estimated standard
deviation (standard error) of the least squares coefficient, bj, increases as VIFj increases.
298 CHAPTER 7 Multiple Regression Analysis

greater than 1 indicates that the estimated coefficient attached to that independent vari-
able is unstable. Its value and associated t statistic may change considerably as the other
independent variables are added or deleted from the regression equation. A large VIF
means essentially that there is redundant information among the predictor variables. The
information being conveyed by a variable with a large VIF is already being explained by
the remaining predictor variables. Thus, multicollinearity makes interpreting the effect of
an individual predictor variable on the response (dependent variable) difficult.
Example 7.7
A large component of the cost of owning a newspaper is the cost of newsprint. Newspaper
publishers are interested in factors that determine annual newsprint consumption. In one
study (see Johnson and Wichern, 1997), data on annual newsprint consumption (Y), the num-
ber of newspapers in a city 1X12, the logarithm6 of the number of families in a city 1X22, and
the logarithm of total retail sales in a city 1X32 were collected for n = 15 cities. The
correlation array for the three predictor variables and the Minitab output from a regression
analysis relating newsprint consumption to the predictor variables are in Table 7-11.
The F statistic (18.54) and its p-value (.000) clearly indicate that the regression is signifi-
cant. The t statistic for each of the independent variables is small with a relatively large
p-value. It must be concluded, for example, that the variable LnFamily is not significant, pro-
vided the other predictor variables remain in the regression function. This suggests that the
term b 2X2 can be dropped from the regression function if the remaining terms, b 1X1 and
b 3X3, are retained. Similarly, it appears as if b 3X3 can be dropped if b 1X1 and b 2X2 remain in
the regression function. The t value (1.69) associated with papers is marginally significant, but
the term b 1X1 might also be dropped if the other predictor variables remain in the equation.
Here, the regression is significant, but each of the predictor variables is not significant. Why?
The VIF column in Table 7-11 provides the answer. Since VIF = 1.7 for Papers, this pre-
dictor variable is very weakly related (VIF near 1) to the remaining predictor variables,
LnFamily and LnRetSales. The VIF = 7.4 for LnFamily is relatively large, indicating this

TABLE 7-11 Minitab Output for Example 7.7—Three Predictor Variables

Correlations: Papers, LnFamily, LnRetSales
Papers LnFamily
LnFamily 0.600
LnRetSales 0.643 0.930
Regression Analysis: Newsprint versus Papers, LnFamily, LnRetSales
The regression equation is
Newsprint = - 56388 + 2385 Papers + 1859 LnFamily + 3455 LnRetSales

Predictor Coef SE Coef T P VIF

Constant -56388 13206 -4.27 0.001
Papers 2385 1410 1.69 0.119 1.7
LnFamily 1859 2346 0.79 0.445 7.4
LnRetSales 3455 2590 1.33 0.209 8.1
S = 1849 R-Sq = 83.8% R -Sq1adj2 = 79.0%
Analysis of Variance
Source DF SS MS F P
Regression 3 190239371 63413124 18.54 0.000
Residual Error 11 37621478 3420134
Total 14 227860849

6Logarithmsof the number of families and the total retail sales are used to make the numbers less positively
skewed and more manageable.
 CHAPTER 7 Multiple Regression Analysis 299

TABLE 7-12 Minitab Output for Example 7.7—Two Predictor

Variables
Regression Analysis: Newsprint versus Papers, LnRetSales
The regression equation is
Newsprint = - 59766 + 2393 Papers + 5279 LnRetSales
Predictor Coef SE Coef T P VIF
Constant - 59766 12304 -4.86 0.000
Papers 2393 1388 1.72 0.110 1.7
LnRetSales 5279 1171 4.51 0.001 1.7
S = 1820 R-sq = 82.5% R -sq1adj2 = 79.6%
Analysis of Variance
Source DF SS MS F P
Regression 2 188090489 94045244 28.38 0.000
Residual Error 12 39770360 3314197
Total 14 227860849

variable is linearly related to the remaining predictor variables. Also, the VIF = 8.1 for
LnRetSales indicates that LnRetSales is related to the remaining predictor variables. Since
Papers is weakly related to LnFamily and LnRetSales, the relationship among the predictor
variables is essentially the relationship between LnFamily and LnRetSales. In fact, the sample
correlation between LnFamily and LnRetSales is r = .93, showing strong linear association.
The variables LnFamily and LnRetSales are very similar in their ability to explain
newsprint consumption. We need only one, but not both, in the regression function. The
Minitab output from a regression analysis with LnFamily (smallest t statistic) deleted from
the regression function is shown in Table 7-12.
Notice that the coefficient of Papers is about the same for the two regressions. The coef-
ficients of LnRetSales, however, are considerably different (3,455 for k = 3 predictors and
5,279 for k = 2 predictors). Also, for the second regression, the variable LnRetSales is
clearly significant 1t = 4.51 with p-valve = .0012. With Papers in the model, LnRetSales is
an additional important predictor of newsprint consumption. The R2’s for the two regres-
sions are nearly the same, approximately .83, as are the standard errors of the estimates,
sy # x’s = 1,849 and sy #x’s = 1,820, respectively. Finally, the common VIF = 1.7 for the two pre-
dictors in the second model indicates that multicollinearity is no longer a problem. As a
residual analysis confirms, for the variables considered, the regression of Newsprint on
Papers and LnRetSales is entirely adequate.

If estimating the separate effects of the predictor variables is important and multi-
collinearity appears to be a problem, what should be done? There are several ways to
deal with severe multicollinearity, as follows. None of them may be completely satisfac-
tory or feasible.
'
• Create new X variables (call them X ) by scaling all the independent variables
according to the formula

' Xi j - X j
X = j = 1, 2, Á , k; i = 1, 2, Á , n (7.12)
1Xij - Xj2 2
Aa
i

These new variables will each have a sample mean of 0 and the same sample
standard deviation. The regression calculations with the new X’s are less sensitive
to round-off error in the presence of severe multicollinearity.
• Identify and eliminate one or more of the redundant independent variables from
the regression function. (This approach was used in Example 7.7.)
300 CHAPTER 7 Multiple Regression Analysis

• Consider estimation procedures other than least squares.7

• Regress the response, Y, on new X’s that are uncorrelated with each other. It is
possible to construct linear combinations of the original X’s that are uncorrelated.8
• Carefully select potential independent variables at the beginning of the study. Try
to avoid variables that “say the same thing.”

SELECTING THE “BEST” REGRESSION EQUATION

How does one develop the best multiple regression equation to forecast a variable of
interest? The first step involves the selection of a complete set of potential predictor
variables. Any variable that might add to the accuracy of the forecast should be
included. In the selection of a final equation, one is usually faced with the dilemma of
providing the most accurate forecast for the smallest cost. In other words, when choos-
ing predictor variables to include in the final equation, the analyst must evaluate them
by using the following two opposed criteria:
1. The analyst wants the equation to include as many useful predictor variables as
possible.9
2. Given that it costs money to obtain and monitor information on a large number of
X’s, the equation should include as few predictors as possible. The simplest equa-
tion is usually the best equation.
The selection of the best regression equation usually involves a compromise between
these extremes, and judgment will be a necessary part of any solution.
After a seemingly complete list of potential predictors has been compiled, the
second step is to screen out the independent variables that do not seem appropriate. An
independent variable (1) may not be fundamental to the problem (there should be
some plausible relation between the dependent variable and an independent variable),
(2) may be subject to large measurement errors, (3) may duplicate other independent
variables (multicollinearity), or (4) may be difficult to measure accurately (accurate
data are unavailable or costly).
The third step is to shorten the list of predictors so as to obtain a “best” selection of
independent variables. Techniques currently in use are discussed in the material that
follows. None of the search procedures can be said to yield the “best” set of independent
variables. Indeed, there is often no unique “best” set. To add to the confusion, the various
techniques do not all necessarily lead to the same final prediction equation.The entire vari-
able selection process is very subjective.The primary advantage of automatic-search proce-
dures is that analysts can then focus their judgments on the pivotal areas of the problem.
To demonstrate various search procedures, a simple example is presented that has
five potential independent variables.
Example 7.8
Pam Weigand, the personnel manager of the Zurenko Pharmaceutical Company, is inter-
ested in forecasting whether a particular applicant will become a good salesperson. She
decides to use the first month’s sales as the dependent variable (Y), and she chooses to
analyze the following independent variables:

7Alternative procedures for estimating the regression parameters are beyond the scope of this book. The
interested reader should consult the work of Draper and Smith (1998).
8Again, the procedures for creating linear combinations of the X’s that are uncorrelated are beyond the
scope of this book. Draper and Smith (1998) discuss these techniques.
9Recall that, whenever a new predictor variable is added to a multiple regression equation, R2 increases.
Therefore, it is important that a new predictor variable make a significant contribution to the regression equation.
 CHAPTER 7 Multiple Regression Analysis 301
X1 = the selling aptitude test score
X2 = the age, in years
X3 = the anxiety test score
X4 = the experience, in years
X5 = the high school GPA 1grade point average2
The personnel manager collects the data shown in Table 7-13, and she assigns the task of
obtaining the “best” set of independent variables for forecasting sales ability to her analyst.
The first step is to obtain a correlation matrix for all the variables from a computer pro-
gram. This matrix will provide essential knowledge about the basic relationships among the
variables.
Examination of the correlation matrix in Table 7-14 reveals that the selling aptitude
test score, age, experience, and GPA are positively related to sales ability and have potential
as good predictor variables. The anxiety test score shows a low negative correlation
with sales, and it is probably not an important predictor. Further analysis indicates that
age is moderately correlated with both GPA and experience. It is the presence of these

TABLE 7-13 Zurenko Pharmaceutical Data for Example 7.8

One Month’s Aptitude Age Anxiety Experience High

Sales (units) Test Score (years) Test Score (years) School GPA
44 10 22.1 4.9 0 2.4
47 19 22.5 3.0 1 2.6
60 27 23.1 1.5 0 2.8
71 31 24.0 .6 3 2.7
61 64 22.6 1.8 2 2.0
60 81 21.7 3.3 1 2.5
58 42 23.8 3.2 0 2.5
56 67 22.0 2.1 0 2.3
66 48 22.4 6.0 1 2.8
61 64 22.6 1.8 1 3.4
51 57 21.1 3.8 0 3.0
47 10 22.5 4.5 1 2.7
53 48 22.2 4.5 0 2.8
74 96 24.8 .1 3 3.8
65 75 22.6 .9 0 3.7
33 12 20.5 4.8 0 2.1
54 47 21.9 2.3 1 1.8
39 20 20.5 3.0 2 1.5
52 73 20.8 .3 2 1.9
30 4 20.0 2.7 0 2.2
58 9 23.3 4.4 1 2.8
59 98 21.3 3.9 1 2.9
52 27 22.9 1.4 2 3.2
56 59 22.3 2.7 1 2.7
49 23 22.6 2.7 1 2.4
63 90 22.4 2.2 2 2.6
61 34 23.8 .7 1 3.4
39 16 20.6 3.1 1 2.3
62 32 24.4 .6 3 4.0
78 94 25.0 4.6 5 3.6
302 CHAPTER 7 Multiple Regression Analysis

TABLE 7-14 Correlations: Sales, Aptitude, Age, Anxiety, Experience, GPA

Correlations: Sales, Aptitude, Age, Anxiety, Experience, GPA
Sales Aptitude Age Anxiety Experience
Aptitude 0.676
Age 0.798 0.228
Anxiety -0.296 -0.222 -0.287
Experience 0.550 0.350 0.540 -0.279
GPA 0.622 0.318 0.695 -0.244 0.312

interrelationships that must be dealt with in attempting to find the best possible set of
explanatory variables.
Two procedures are demonstrated: all possible regressions and stepwise regression.

All Possible Regressions

The procedure calls for the investigation of all possible regression equations that
involve the potential independent variables. The analyst starts with an equation
containing no independent variables and then proceeds to analyze every possible
combination in order to select the best set of predictors.
Different criteria for comparing the various regression equations may be used with
the all possible regressions approach. Only the R2 technique, which involves four steps,
is discussed here.
This procedure first requires the fitting of every possible regression model that
involves the dependent variable and any number of independent variables. Each
independent variable can either be or not be in the equation (two possible out-
comes), and this fact is true for every independent variable. Thus, altogether there
are 2k equations (where k equals the number of independent variables). So, if there
are eight independent variables to consider 1k = 82, then 28 = 256 equations must
be examined.
The second step in the procedure is to divide the equations into sets according to
the number of parameters to be estimated.

Example 7.9
The results from the all possible regressions runs for the Zurenko Pharmaceutical Company
are presented in Table 7-15. Notice that Table 7-15 is divided into six sets of regression equa-
tion outcomes. This breakdown coincides with the number of parameters contained in each
equation.

The third step involves the selection of the best independent variable (or vari-
ables) for each parameter grouping. The equation with the highest R2 is considered
best. Using the results from Example 7.9, the best equation from each set listed in
Table 7-15 is presented in Table 7-16.
The fourth step involves making the subjective decision: “Which equation is the
best?” On the one hand, the analyst desires the highest R2 possible; on the other
hand, he or she wants the simplest equation possible. The all possible regressions
approach assumes that the number of data points, n, exceeds the number of parame-
ters, k + 1.

Example 7.10
The analyst is attempting to find the point at which adding additional independent variables
for the Zurenko Pharmaceutical problem is not worthwhile because it leads to a very small
increase in R2. The results in Table 7-16 clearly indicate that adding variables after selling
 CHAPTER 7 Multiple Regression Analysis 303

TABLE 7-15 R2 Values for All Possible Regressions for

Zurenko Pharmaceutical for Example 7.9

Independent Number of Error Degrees

Variables Used Parameters of Freedom R2
None 1 29 .0000
X1 2 28 .4570
X2 2 28 .6370
X3 2 28 .0880
X4 2 28 .3020
X5 2 28 .3870
X1, X2 3 27 .8948
X1, X3 3 27 .4790
X1, X4 3 27 .5690
X1, X5 3 27 .6410
X2, X3 3 27 .6420
X2, X4 3 27 .6570
X2, X5 3 27 .6460
X3, X4 3 27 .3240
X3, X5 3 27 .4090
X4, X5 3 27 .5270
X1, X2, X3 4 26 .8951
X1, X2, X4 4 26 .8948
X1, X2, X5 4 26 .8953
X1, X3, X4 4 26 .5750
X1, X3, X5 4 26 .6460
X1, X4, X5 4 26 .7010
X2, X3, X4 4 26 .6590
X2, X3, X5 4 26 .6500
X2, X4, X5 4 26 .6690
X3, X4, X5 4 26 .5310
X1, X2, X3, X4 5 25 .8951
X1, X2, X3, X5 5 25 .8955
X1, X2, X4, X5 5 25 .8953
X1, X3, X4, X5 5 25 .7010
X2, X3, X4, X5 5 25 .6710
X1, X2, X3, X4, X5 6 24 .8955

TABLE 7-16 Best Regression Equations for

Zurenko Pharmaceutical
from Example 7.9
Number of Independent Error Degrees
Parameters Variables of Freedom R2

1 None 29 .0000
2 X2 28 .6370
3 X1, X2 27 .8948
4 X1, X2, X5 26 .8953
5 X1, X2, X3, X5 25 .8955
6 X1, X2, X3, X4, X5 24 .8955
304 CHAPTER 7 Multiple Regression Analysis

aptitude test (X1) and age (X2) is not necessary. Therefore, the final fitted regression equa-
tion is of the form

YN = b0 + b1X1 + b2X2

and it explains 89.48% of the variation in Y.

The all possible regressions procedure is best summed up by Draper and Smith (1998):

In general the analysis of all regressions is quite unwarranted. While it means

that the investigator has “looked at all possibilities” it also means he has exam-
ined a large number of regression equations that intelligent thought would
often reject out of hand. The amount of computer time used is wasteful and
the sheer physical effort of examining all the computer printouts is enormous
when more than a few variables are being examined. Some sort of selection
procedure that shortens this task is preferable. (p. 333)

Stepwise Regression
The stepwise regression procedure adds one independent variable at a time to the
model, one step at a time. A large number of independent variables can be handled on
the computer in one run when using this procedure.
Stepwise regression can best be described by listing the basic steps (algorithm)
involved in the computations.

1. All possible simple regressions are considered. The predictor variable that explains
the largest significant proportion of the variation in Y (has the largest correlation
with the response) is the first variable to enter the regression equation.
2. The next variable to enter the equation is the one (out of those not included) that
makes the largest significant contribution to the regression sum of squares. The sig-
nificance of the contribution is determined by an F test. The value of the F statistic
that must be exceeded before the contribution of a variable is deemed significant
is often called the F to enter.
3. Once an additional variable has been included in the equation, the individual con-
tributions to the regression sum of squares of the other variables already in the
equation are checked for significance using F tests. If the F statistic is less than a
value called the F to remove, the variable is deleted from the regression equation.
4. Steps 2 and 3 are repeated until all possible additions are nonsignificant and all
possible deletions are significant. At this point, the selection stops.

Stepwise regression permits predictor variables to enter or leave the regression

function at different stages of its development. An independent variable is
removed from the model if it doesn’t continue to make a significant contribution
when a new variable is added.

The user of a stepwise regression program supplies the values that decide when a
variable is allowed to enter and when a variable is removed. Since the F statistics used in
stepwise regression are such that F = t2 where t is the t statistic for checking the signifi-
cance of a predictor variable, F = 4 1corresponding to ƒ t ƒ = 22 is a common choice for
both the F to enter and the F to remove. An F to enter of 4 is essentially equivalent to
testing for the significance of a predictor variable at the 5% level. The Minitab stepwise
 CHAPTER 7 Multiple Regression Analysis 305
program allows the user to choose an α level to enter and to remove variables or the F
value to enter and to remove variables. Using an α value of .05 is approximately equiv-
alent to using F = 4. The current default values in Minitab are a = .15 and F = 4.
The result of the stepwise procedure is a model that contains only independent
variables with t values that are significant at the specified level. However, because of
the step-by-step development, there is no guarantee that stepwise regression will
select, for example, the best three variables for prediction. In addition, an automatic
selection method is not capable of indicating when transformations of variables are
useful, nor does it necessarily avoid a multicollinearity problem. Finally, stepwise
regression cannot create important variables that are not supplied by the user. It is nec-
essary to think carefully about the collection of independent variables that is supplied
to a stepwise regression program.
The stepwise procedure is illustrated in Example 7.11.

Example 7.11
Let’s “solve” the Zurenko Pharmaceutical problem using stepwise regression.
Pam examines the correlation matrix shown in Table 7-14 and decides that, when she
runs the stepwise analysis, the age variable will enter the model first because it has the
largest correlation with sales 1r1,3 = .7982 and will explain 63.7% 1.79822 of the variation
in sales.
She notes that the aptitude test score will probably enter the model second because it is
strongly related to sales 1r1,2 = .6762 but not highly related to the age variable 1r2,3 = .2282
already in the model.
Pam also notices that the other variables will probably not qualify as good predictor
variables. The anxiety test score will not be a good predictor because it is not well related to
sales 1r1,4 = - .2962 . The experience and GPA variables might have potential as good
predictor variables 1r1,5 = .550 and r1,6 = .622, respectively2 . However, both of these
predictor variables have a potential multicollinearity problem with the age variable
1r3,5 = .540 and r3,6 = .695, respectively2.
The Minitab commands to run a stepwise regression analysis for this example are
demonstrated in the Minitab Applications section at the end of the chapter. The output
from this stepwise regression run is shown in Table 7-17. The stepwise analysis proceeds
according to the steps that follow.

TABLE 7-17 Stepwise Regression for Example 7.11: Sales

versus Aptitude, Age, Anxiety, Experience, GPA
Alpha-to-Enter: 0.05 Alpha-to-Remove: 0.05
Response is Sales on 5 predictors, with N = 30
Step 1 2
Constant -100.85 -86.79
Age 6.97 5.93
T-Value 7.01 10.60
P-Value 0.000 0.000
Aptitude 0.200
T-Value 8.13
P-Value 0.000
S 6.85 3.75
R-Sq 63.70 89.48
R-Sq(adj) 62.41 88.70
Mallows Cp 57.4 0.2
306 CHAPTER 7 Multiple Regression Analysis

Step 1. The model after step 1 is

Sales = - 100.85 + 6.97 1Age2

As Pam thought, the age variable entered the model first and explains 63.7%
of the sales variance. Since the p-value of .000 is less than the a value of .05, age is
added to the model. Remember that the p-value is the probability of obtaining a
t statistic as large as 7.01 by chance alone. The Minitab decision rule that Pam
selected is to enter a variable if the p-value is less than a = .05.
Note that t = 7.01 6 2.048, the upper .025 point of a t distribution with
28 1n - k - 1 = 30 - 1 - 12 degrees of freedom. Thus, at the .05 significance
level, the hypothesis H0: b 1 = 0 is rejected in favor of H1: b 1 Z 0 . Since
t2 = F or 2.0482 = 4.19, an F to enter of 4 is also essentially equivalent to testing
for the significance of a predictor variable at the 5% level. In this case, since the
coefficient of the age variable is clearly significantly different from zero, age
enters the regression equation, and the procedure now moves to step 2.
Step 2. The model after step 2 is
Sales = - 86.79 + 5.93 1Age2 + 0.200 1Aptitude2

This model explains 89.48% of the variation in sales.

The null and alternative hypotheses to determine whether the aptitude test
score’s regression coefficient is significantly different from zero are

H 0: b 2 = 0
H1: b 2 Z 0

Again, the p-value of .000 is less than the α value of .05, and aptitude test score is
added to the model. The aptitude test score’s regression coefficient is significantly
different from zero, and the probability that this occurred by chance sampling
error is approximately zero. This result means that the aptitude test score is an
important variable when used in conjunction with age.
The critical t statistic based on 27 1n - k - 1 = 30 - 2 - 12 degrees of
freedom is 2.052.10 The computed t ratio found on the Minitab output is 8.13,
which is greater than 2.052. Using a t test, the null hypothesis is also rejected.
Note that the p-value for the age variable’s t statistic, .000, remains very small.
Age is still a significant predictor of sales. The procedure now moves on to step 3.
Step 3. The computer now considers adding a third predictor variable, given that X1 (age)
and X2 (aptitude test score) are in the regression equation. None of the remaining
independent variables is significant (has a p-value less than .05) when run in
combination with X1 and X2, so the stepwise procedure is completed.
Pam’s final model selected by the stepwise procedure is the two-predictor
variable model given in step 2.

Final Notes on Stepwise Regression

The stepwise regression technique is extremely easy to use. Unfortunately, it is also
extremely easy to misuse. Analysts developing a regression model often produce a
large set of potential independent variables and then let the stepwise procedure deter-
mine which ones are significant. The problem is that, when a large set of independent
variables is analyzed, many t tests are performed, and it is likely that a type I error
(adding a nonsignificant variable) will result. That is, the final model might contain a
variable that is not linearly related to the dependent variable and entered the model
just by chance.

10Again, since 2.0522 = 4.21, using an F to enter of 4 is roughly equivalent to testing for the significance of a
predictor variable at the .05 level.
 CHAPTER 7 Multiple Regression Analysis 307
As mentioned previously, another problem involves the initial selection of poten-
tial independent variables. When these variables are selected, higher-order terms
(curvilinear, nonlinear, and interaction) are often omitted to keep the number of vari-
ables manageable. Consequently, several important variables may be initially omitted
from the model. It becomes obvious that an analyst’s intuitive choice of the initial
independent variables is critical to the development of a successful regression model.

REGRESSION DIAGNOSTICS AND RESIDUAL ANALYSIS

A regression analysis is not complete until one is convinced the model is an adequate
representation of the data. It is imperative to examine the adequacy of the model
before it becomes part of the decision-making apparatus.
An examination of the residuals is a crucial component of the determination of
model adequacy. Several residual plots designed to check the regression model assump-
tions were introduced in Chapter 6. Also, if regression models are used with time series
data (to be discussed in Chapter 8), it is important to compute the residual autocorrela-
tions to check the independence assumption. Inferences (and decisions) made with
models that do not approximately conform to the regression assumptions can be grossly
misleading. For example, it may be concluded that the manipulation of a predictor vari-
able will produce a specified change in the response when, in fact, it will not. It may be
concluded that a forecast is very likely (95% confidence) to be within 2% of the future
response when, in fact, the actual confidence is much less, and so forth.
In this section, some additional tools that can be used to evaluate a regression
model will be discussed. These tools are designed to identify observations that are out-
lying or extreme (observations that are well separated from the remainder of the data).
Outlying observations are often hidden by the fitting process and may not be easily
detected from an examination of residual plots. Yet they can have a major role in deter-
mining the fitted regression function. It is important to study outlying observations to
decide whether they should be retained or eliminated and, if retained, whether their
influence should be reduced in the fitting process or the regression function revised.
A measure of the influence of the ith data point on the location of the fitted
regression function is provided by the leverage hii. The leverage depends only on the
predictors; it does not depend on the response, Y. For simple linear regression with one
predictor variable, X,

1 1Xi - X22
hii = +
a 1Xi - X2
2
(7.13)
n

With k predictors, the expression for the ith leverage is more complicated; however,
one can show that 0 6 hii 6 1 and that the mean leverage is h = 1k + 12>n.
If the ith data point has high leverage 1hii is close to 1), the fitted response, YNi, at
these X’s is almost completely determined by Yi, with the remaining data having very
little influence. The high leverage data point is also an outlier among the X’s (far from
other combinations of X values).11 A rule of thumb suggests that hii is large enough to
merit checking if hii Ú 31k + 12>n.
The detection of outlying or extreme Y values is based on the size of the residuals,
e = Y - YN . Large residuals indicate a Y value that is “far” from its fitted or predicted

11The converse is not necessarily true. That is, an outlier among the X’s may not be a high leverage point.
308 CHAPTER 7 Multiple Regression Analysis

value, YN . A large residual will show up in a histogram of the residuals as a value far (in
either direction) from zero. A large residual will show up in a plot of the residuals ver-
sus the fitted values as a point far above or below the horizontal axis.
Software packages such as Minitab flag data points with extreme Y values by com-
puting “standardized” residuals and identifying points with large standardized residuals.
One standardization is based on the fact that the residuals have estimated stan-
dard deviations:

sei = sy #x¿s 21 - hii

where sy # x¿s = 2MSE is the standard error of the estimate and hii is the leverage
associated with the ith data point. The standardized residual12 is then

ei ei
= (7.14)
sei sy # x¿s 21 - hii

The standardized residuals all have a variance of 1. A standardized residual is consid-

ered large (the response extreme) if

` ` 7 2
ei
sei

The Y values corresponding to data points with large standardized residuals can heav-
ily influence the location of the fitted regression function.

Example 7.12
Chief executive officer (CEO) salaries in the United States are of interest because of their
relationship to salaries of CEOs in international firms and to salaries of top professionals
outside corporate America. Also, for an individual firm, the CEO compensation directly, or
indirectly, influences the salaries of managers in positions below that of CEO. CEO salary
varies greatly from firm to firm, but data suggest that salary can be explained in terms of a
firm’s sales and the CEO’s amount of experience, educational level, and ownership stake in
the firm. In one study, 50 firms were used to develop a multiple regression model linking
CEO compensation to several predictor variables such as sales, profits, age, experience,
professional background, educational level, and ownership stake.
After eliminating unimportant predictor variables, the final fitted regression func-
tion was

YN = 5.52 - .467X1 + .263X2

where

Y = the logarithm of CEO compensation

X1 = the indicator variable for educational level
X2 = the logarithm of company sales

Minitab identified three observations from this regression analysis that have either
large standardized residuals or large leverage.

12Some software packages may call the standardized residual given by Equation 7.14 the Studentized
residual.
 CHAPTER 7 Multiple Regression Analysis 309
Unusual Observations
Obs Educate LnComp Fit StDev Fit Residual St Resid
14 1.00 6.0568 7.0995 0.0949 1.0427 2.09R
25 0.00 8.1342 7.9937 0.2224 0.1405 0.31X
33 0.00 6.3969 7.3912 0.2032 0.9943 2.13R
R denotes an observation with a large standardized residual.
X denotes an observation whose X value gives it large
influence.

Observations 14 and 33 have large standardized residuals. The fitted regression function is
predicting (log) compensation that is too large for these two CEOs. An examination of the
full data set shows that these CEOs each own relatively large percentages of their compa-
nies’ stock. Case 14 owns more than 10% of the company’s stock, and case 33 owns more
than 17% of the company’s stock. These individuals are receiving much of their remunera-
tion through long-term compensation, such as stock incentives, rather than through annual
salary and bonuses. Since amount of stock owned (or stock value) is not included as a vari-
able in the regression function, it cannot be used to adjust the prediction of compensation
determined by CEO education and company sales. Although education and (log) sales do
not predict the compensation of these two CEOs as well as the others, there appears to be
no reason to eliminate them from consideration.
Observation 25 is singled out because the leverage for this data point is greater than
31k + 12>n = 3132>50 = .18. This CEO has no college degree 1Educate = 02 but is with a
company with relatively large sales 1LnSales = 9.3942. The combination (0, 9.394) is far
from the point 1X1,X22 ; therefore, it is an outlier among the pairs of X’s. The response asso-
ciated with these X’s will have a large influence on the determination of the fitted regres-
sion function. (Notice that the standardized residual for this data point is small, indicating
that the predicted or fitted (log) compensation is close to the actual value.) This particular
CEO has 30 years of experience as a CEO, more experience than all but one of the CEOs in
the data set. This observation is influential, but there is no reason to delete it.

Leverage tells us if an observation has unusual predictors, and a standardized

residual tells us if an observation has an unusual response. These quantities can be
combined into one overall measure of influence known as Cook’s distance. Cook’s
distances can be printed out in most statistical software packages, but additional
discussion is beyond the scope of this book.13

FORECASTING CAVEATS
We finish this discussion of multiple regression with some general comments. These
comments are oriented toward the practical application of regression analysis.

Overfitting

Overfitting refers to adding independent variables to the regression function that,

to a large extent, account for all the eccentricities of the sample data under analysis.

When such a model is applied to new sets of data selected from the same population, it
does not forecast as well as the initial fit might suggest.
Overfitting is more likely to occur when the sample size is small, especially if a large
number of independent variables are included in the model. Some practitioners have

13A good discussion of Cook’s distance is provided by Draper and Smith (1998).
310 CHAPTER 7 Multiple Regression Analysis

suggested that there should be at least 10 observations for each independent variable. (If
there are four independent variables, a sample size n of at least 40 is suggested.)
One way to guard against overfitting is to develop the regression function from
one part of the data and then apply it to a “holdout” sample. Use the fitted regression
function to forecast the holdout responses and calculate the forecast errors. If the fore-
cast errors are substantially larger than the fitting errors as measured by, say, compara-
ble mean squared errors, then overfitting has occurred.

Useful Regressions, Large F Ratios

A regression that is statistically significant is not necessarily useful. With a relatively large
sample size (i.e., when n is large relative to k, the number of predictors), it is not unusual to
get a significant F ratio and a small R2. That is, the regression is significant, yet it explains
only a small proportion of the variation in the response. One rule of thumb suggests that
with a significance level of .05, the F ratio should be at least four times the corresponding
critical value before the regression is likely to be of much use for prediction purposes.14
The “four times” criterion comes from the argument that the range of the predic-
tions (over all the X’s) should be about four times the (average) prediction error
before the regression is likely to yield a worthwhile interpretation.15
As an example, with k = 3 predictors, n = 25 observations, and a significance level
of .05, the computed F from the ANOVA table would have to exceed the critical value
F = 3.07 (see Table B-5 with 1 = k = 3, 2 = n - k - 1 = 21 degrees of freedom)
for the regression to be significant. (Using Equation 7.7, the critical F = 3.07 corre-
sponds to an R2 of about 30%, not a particularly large number.) However, the “four
times” rule suggests that the computed F should exceed 413.072 = 12.28 in order for the
regression to be worthwhile from a practical point of view.

APPLICATION TO MANAGEMENT

Multiple regression analysis has been used extensively to help forecast the economic
activity of the various segments of the economy. Many of the reports and forecasts
about the future of our economy that appear in the Wall Street Journal, Fortune,
Business Week, and other similar sources are based on econometric (regression) mod-
els. The U.S. government makes wide use of regression analysis in predicting future
revenues, expenditures, income levels, interest rates, birthrates, unemployment, and
Social Security benefits requirements as well as a multitude of other events. In fact,
almost every major department in the U.S. government makes use of the tools
described in this chapter.
Similarly, business entities have adopted and, when necessary, modified regression
analysis to help in the forecasting of future events. Few firms can survive in today’s
environment without a fairly accurate forecast of tomorrow’s sales, expenditures, capi-
tal requirements, and cash flows. Although small or less sophisticated firms may be
able to get by with intuitive forecasts, larger and/or more sophisticated firms have
turned to regression analysis to study the relationships among several variables and to
determine how these variables are likely to affect their future.
Unfortunately, the very notoriety that regression analysis receives for its usefulness
as a tool in predicting the future tends to overshadow an equally important asset: its

14Some authors argue that the “four times” rule is not enough and should be replaced by a “ten times”
criterion.
15This assumes that no other defect is detected in the fit.
 CHAPTER 7 Multiple Regression Analysis 311
ability to help evaluate and control the present. Because a fitted regression equation
provides the researcher with both strength and direction information, management can
evaluate and change current strategies.
Suppose, for example, a manufacturer of jams wants to know where to direct its
marketing efforts when introducing a new flavor. Regression analysis can be used to
help determine the profile of heavy users of jams. For instance, a company might try to
predict the number of flavors of jam a household might have at any one time on the
basis of a number of independent variables, such as the following:

Number of children living at home

Age of children
Gender of children
Home ownership versus rental
Time spent shopping
Income

Even a superficial reflection on the jam example quickly leads the researcher to
realize that regression analysis has numerous possibilities for use in market segmenta-
tion studies. In fact, many companies use regression to study market segments to deter-
mine which variables seem to have an impact on market share, purchase frequency,
product ownership, and product and brand loyalty as well as on many other areas.
Agricultural scientists use regression analysis to explore the relationship of product
yield (e.g., number of bushels of corn per acre) to fertilizer type and amount, rainfall,
temperature, days of sun, and insect infestation. Modern farms are equipped with mini-
and microcomputers complete with software packages to help them in this process.
Medical researchers use regression analysis to seek links between blood pressure
and independent variables such as age, social class, weight, smoking habits, and race.
Doctors explore the impact of communications, number of contacts, and age of patient
on patient satisfaction with service.
Personnel directors explore the relationship of employee salary levels to geo-
graphic location, unemployment rates, industry growth, union membership, industry
type, and competitive salaries. Financial analysts look for causes of high stock prices by
analyzing dividend yields, earnings per share, stock splits, consumer expectations of
interest rates, savings levels, and inflation rates.
Advertising managers frequently try to study the impact of advertising budgets,
media selection, message copy, advertising frequency, and spokesperson choice on
consumer attitude change. Similarly, marketers attempt to determine sales from adver-
tising expenditures, price levels, competitive marketing expenditures, and consumer
disposable income as well as a wide variety of other variables.
A final example further illustrates the versatility of regression analysis. Real estate
site location analysts have found that regression analysis can be very helpful in pin-
pointing geographic areas of over- and underpenetration of specific types of retail
stores. For instance, a hardware store chain might look for a potential city in which to
locate a new store by developing a regression model designed to predict hardware sales
in any given city. Researchers could concentrate their efforts on those cities where the
model predicted higher sales than actually achieved (as can be determined from many
sources). The hypothesis is that sales of hardware are not up to potential in these cities.
In summary, regression analysis has provided management with a powerful and
versatile tool for studying the relationships between a dependent variable and multiple
independent variables. The goal is to better understand and perhaps control present
events as well as to better predict future events.
312 CHAPTER 7 Multiple Regression Analysis

Glossary

Dummy variables. Dummy, or indicator, variables are Partial, or net, regression coefficient. The partial,
used to determine the relationships between qualita- or net, regression coefficient measures the aver-
tive independent variables and a dependent variable. age change in the dependent variable per unit
Multicollinearity. Multicollinearity is the situation change in the relevant independent variable, hold-
in which independent variables in a multiple ing the other independent variables constant.
regression equation are highly intercorrelated. Standard error of the estimate. The standard error
That is, a linear relation exists between two or of the estimate is the standard deviation
more independent variables. of the residuals. It measures the amount the actual
Multiple regression. Multiple regression involves values (Y) differ from the estimated values 1YN 2.
the use of more than one independent variable to Stepwise regression. Stepwise regression permits
predict a dependent variable. predictor variables to enter or leave the regression
Overfitting. Overfitting refers to adding indepen- function at different stages of its development.
dent variables to the regression function that, to a An independent variable is removed from the
large extent, account for all the eccentricities of model if it doesn’t continue to make a significant
the sample data under analysis. contribution when a new variable is added.

Key Formulas

Population multiple regression function

mY = b 0 + b 1X1 + b 2X2 + Á + b kXk (7.1)

Estimated (fitted) regression function

YN = b0 + b1X1 + Á + bkXk (7.2)

Sum of squares decomposition and associated degrees of freedom

©1Y - Y22 = ©1YN - Y22 + ©1Y - YN 22
SST = SSR + SSE (7.3)
df : n - 1 = k + n - k - 1

Standard error of the estimate

a 1Y - Y 2
N 2 SSE
sy #x’s = = = 2MSE (7.4)
D n - k - 1 Dn - k - 1

F statistic for testing the significance of the regression

MSR
F = (seebox on p. 289)
MSE

Coefficient of determination
SSR ©1YN - Y22
R2 = =
SST ©1Y - Y22

SSE ©1Y - YN 22
= 1 - = 1 - (7.5)
SST ©1Y - Y22
 CHAPTER 7 Multiple Regression Analysis 313
Multiple correlation coefficient
R = 2R2 (7.6)

Relation between F statistic and R2

R2 n - k - 1
F = ¢ ≤ (7.7)
1 - R2 k
Adjusted coefficient of determination
n - 1
R 2 = 1 - 11 - R 22 ¢ ≤ (7.8)
n - k - 1
t statistic for testing H0: Bj = 0
bj
t = 1see box on p. 2912
sbj
Forecast of a future value
YN * = b0 + b1X*1 + b2X*2 + Á + bkX*k (7.9)

Large-sample prediction interval for a future response

1YN * - ta>2 sy #x¿s , YN * + ta>2 sy #x¿s2 (7.10)

Variance inflation factors

1
VIFj = j = 1, 2, Á , k (7.11)
1 - R2j
Standardized independent variable values
' Xi j - Xj
X = j = 1, 2, Á , k; i = 1, 2, Á , n (7.12)
a 1Xi j - Xj2
2
A i
Leverage (one-predictor variable)

1 1Xi - X 22
hii = +
a 1Xi - X2
2
(7.13)
n
Standardized residual
ei ei
= (7.14)
sei sy #x¿s 21 - hii
Problems

1. What are the characteristics of a good predictor variable?

2. What are the assumptions associated with the multiple regression model?
3. What does the partial, or net, regression coefficient measure in multiple regression?
4. What does the standard error of the estimate measure in multiple regression?
5. Your estimated multiple regression equation is YN = 7.52 + 3X1 - 12.2X2. Predict
the value of Y if X1 = 20 and X2 = 7.
314 CHAPTER 7 Multiple Regression Analysis

TABLE P–7

Variable Number
Variable Number 1 2 3 4 5 6
1 1.00 .55 .20 -.51 .79 .70
2 1.00 .27 .09 .39 .45
3 1.00 .04 .17 .21
4 1.00 -.44 -.14
5 1.00 .69
6 1.00

6. Explain each of the following concepts:

a. Correlation matrix
b. R2
c. Multicollinearity
d. Residual
e. Dummy variable
f. Stepwise regression
7. Most computer solutions for multiple regression begin with a correlation matrix.
Examining this matrix is often the first step when analyzing a regression problem
that involves more than one independent variable. Answer the following questions
concerning the correlation matrix given in Table P-7.
a. Why are all the entries on the main diagonal equal to 1.00?
b. Why is the bottom half of the matrix below the main diagonal blank?
c. If variable 1 is the dependent variable, which independent variables have the
highest degree of linear association with variable 1?
d. What kind of association exists between variables 1 and 4?
e. Does this correlation matrix show any evidence of multicollinearity?
f. In your opinion, which variable or variables will be included in the best fore-
casting model? Explain.
g. If the data given in this correlation matrix are run on a stepwise program, which
independent variable (2, 3, 4, 5, or 6) will be the first to enter the regression function?
8. Jennifer Dahl, supervisor of the Circle O discount chain, would like to forecast the
time it takes to check out a customer. She decides to use the following independent
variables: number of purchased items and total amount of the purchase. She
collects data for a sample of 18 customers, shown in Table P-8.
a. Determine the best regression equation.
b. When an additional item is purchased, what is the average increase in the
checkout time?
c. Compute the residual for customer 18.
d. Compute the standard error of the estimate.
e. Interpret part d in terms of the variables used in this problem.
f. Compute a forecast of the checkout time if a customer purchases 14 items that
amount to $70.
g. Compute a 95% interval forecast for your prediction in part f.
h. What should Jennifer conclude?
9. Table P-9 contains data on food expenditures, annual income, and family size for a
sample of 10 families.
 CHAPTER 7 Multiple Regression Analysis 315

TABLE P-8

Checkout Time Amount ($) Number

Customer (minutes) Y X1 of items X2
1 3.0 36 9
2 1.3 13 5
3 .5 3 2
4 7.4 81 14
5 5.9 78 13
6 8.4 103 16
7 5.0 64 12
8 8.1 67 11
9 1.9 25 7
10 6.2 55 11
11 .7 13 3
12 1.4 21 8
13 9.1 121 21
14 .9 10 6
15 5.4 60 13
16 3.3 32 11
17 4.5 51 15
18 2.4 28 10

TABLE P-9

Annual Food
Expenditures Annual Income Family Size
Family ($100s) Y ($1,000s) X1 X2

A 24 11 6
B 8 3 2
C 16 4 1
D 18 7 3
E 24 9 5
F 23 8 4
G 11 5 2
H 15 7 2
I 21 8 3
J 20 7 2

a. Construct the correlation matrix for the three variables in Table P-9. Interpret
the correlations in the matrix.
b. Fit a multiple regression model relating food expenditures to income and family
size. Interpret the partial regression coefficients of income and family size. Do
they make sense?
c. Compute the variance inflation factors (VIFs) for the independent variables. Is
multicollinearity a problem for these data? If so, how might you modify the
regression model?
10. Beer sales at the Shapiro One-Stop Store are analyzed using temperature
and number of people (age 21 or over) on the street as independent variables.
316 CHAPTER 7 Multiple Regression Analysis

TABLE P-10 Minitab Output

Correlations
Y X1
X1 0.827
X2 0.822 0.680
Regression Analysis
The regression equation is
Y = - 26.7 + .782 X1 + .068 X2
Predictor Coef SE Coef T P
Constant -26.706
X1 .78207 .22694
X2 .06795 .02026
S = R-Sq = R-Sq1adj2 =
Analysis of Variance

Source DF SS MS F
Regression 2 11589.035 5794.516 36.11
Residual Error 17 2727.914 160.466
Total 19 14316.949

A random sample of 20 days is selected, and the following variables are

measured:
Y = the number of six-packs of beer sold each day
X1 = the daily high temperature
X2 = the daily traffic count
The data are analyzed using multiple regression. The partial computer output is in
Table P-10.
a. Analyze the correlation matrix.
b. Test the hypothesis H0: b j = 0 , j = 1, 2, at the .01 significance level.
c. Forecast the volume of beer sold if the high temperature is 60 degrees and the
traffic count is 500 people.
d. Calculate R2, and interpret its meaning in terms of this problem.
e. Calculate the standard error of the estimate.
f. Explain how beer sales are affected by an increase of one degree in the high
temperature.
g. State your conclusions for this analysis concerning the accuracy of the forecast-
ing equation and also the contributions of the independent variables.
11. A taxi company is interested in the relationship between mileage, measured in
miles per gallon, and the age of cars in its fleet. The 12 fleet cars are the same make
and size and are in good operating condition as a result of regular maintenance.
The company employs both male and female drivers, and it is believed that some
of the variability in mileage may be due to differences in driving techniques
between the groups of drivers of opposite gender. In fact, other things being equal,
women tend to get better mileage than men. Data are generated by randomly
assigning the 12 cars to five female and seven male drivers and computing miles
per gallon after 300 miles. The data appear in Table P-11.
a. Construct a scatter diagram with Y as the vertical axis and X1 as the horizontal
axis. Identify the points corresponding to male and female drivers, respectively.
 CHAPTER 7 Multiple Regression Analysis 317

TABLE P-11
Miles per Age of Car Gender (0  male,
Gallon Y (years) X1 1  female) X2

22.3 3 0
22.0 4 1
23.7 3 1
24.2 2 0
25.5 1 1
21.1 5 0
20.6 4 0
24.0 1 0
26.0 1 1
23.1 2 0
24.8 2 1
20.2 5 0

b. Fit the regression model

Y = b 0 + b 1X1 + b 2X2 +

and interpret the least squares coefficient, b2.

c. Compute the fitted values for each of the 1X1, X22 pairs, and plot the fitted val-
ues on the scatter diagram. Draw straight lines through the fitted values for
male drivers and female drivers, respectively. Specify the equations for these
two straight lines.
d. Suppose gender is ignored. Fit the simple linear regression model,
Y = b 0 + b 1X1 + , and plot the fitted straight line on the scatter diagram. Is
it important to include the effects of gender in this case? Explain.
12. The sales manager of a large automotive parts distributor, Hartman Auto Supplies,
wants to develop a model to forecast as early as May the total annual sales of a
region. If regional sales can be forecast, then the total sales for the company can be
forecast. The number of retail outlets in the region stocking the company’s parts
and the number of automobiles registered for each region as of May 1 are the two
independent variables investigated. The data appear in Table P-12.
a. Analyze the correlation matrix.
b. How much error is involved in the prediction for region 1?
c. Forecast the annual sales for region 12, given 2,500 retail outlets and 20.2 million
automobiles registered.
d. Discuss the accuracy of the forecast made in part c.
e. Show how the standard error of the estimate was computed.
f. Give an interpretation of the partial regression coefficients. Are these regression
coefficients sensible?
g. How can this regression equation be improved?
13. The sales manager of Hartman Auto Supplies decides to investigate a new inde-
pendent variable, personal income by region (see Problem 12). The data for this
new variable are presented in Table P-13.
a. Does personal income by region make a contribution to the forecasting of sales?
318 CHAPTER 7 Multiple Regression Analysis

TABLE P-12

Annual Sales Number of

($ millions) Retail Outlets Number of Automobiles
Region X2 Y X1 Registered ($ millions)
1 52.3 2,011 24.6
2 26.0 2,850 22.1
3 20.2 650 7.9
4 16.0 480 12.5
5 30.0 1,694 9.0
6 46.2 2,302 11.5
7 35.0 2,214 20.5
8 3.5 125 4.1
9 33.1 1,840 8.9
10 25.2 1,233 6.1
11 38.2 1,699 9.5

TABLE P-13
Region Personal Income Region Personal Income
($ billions) ($ billions)

1 98.5 7 67.6
2 31.1 8 19.7
3 34.8 9 67.9
4 32.7 10 61.4
5 68.8 11 85.6
6 94.7

b. Forecast annual sales for region 12 for personal income of $40 billion and the val-
ues for retail outlets and automobiles registered given in part c of Problem 12.
c. Discuss the accuracy of the forecast made in part b.
d. Which independent variables would you include in your final forecast model?
Why?
14. The Nelson Corporation decides to develop a multiple regression equation to
forecast sales performance. A random sample of 14 salespeople is interviewed and
given an aptitude test. Also, an index of effort expended is calculated for each
salesperson on the basis of a ratio of the mileage on his or her company car to the
total mileage projected for adequate coverage of territory. Regression analysis
yields the following results:
YN = 16.57 + .65 X1 + 20.6 X2
1.052 11.692

The quantities in parentheses are the standard errors of the partial regression
coefficients. The standard error of the estimate is 3.56. The standard deviation of
the sales variable is sy = 16.57. The variables are
Y = the sales performance, in thousands
X1 = the aptitude test score
X2 = the effort index
 CHAPTER 7 Multiple Regression Analysis 319
a. Are the partial regression coefficients significantly different from zero at the .01
significance level?
b. Interpret the partial regression coefficient for the effort index.
c. Forecast the sales performance for a salesperson who has an aptitude test score
of 75 and an effort index of .5.
d. Calculate the sum of squared residuals, ©1Y - YN 22.
e. Calculate the total sum of squares, ©1Y - Y22.
f. Calculate R2, and interpret this number in terms of this problem.
g. Calculate the adjusted coefficient of determination, R2.
15. We might expect credit card purchases to differ from cash purchases at the same
store. Table P-15 contains daily gross sales and items sold for cash purchases and
daily gross sales and items sold for credit card purchases at the same consignment
store for 25 consecutive days.
a. Make a scatter diagram of daily gross sales, Y, versus items sold for cash pur-
chases, X1. Using a separate plot symbol or color, add daily gross sales and
items sold for credit card purchases, X2 . Visually compare the relationship
between sales and number of items sold for cash with that for credit card
purchases.

TABLE P-15

Gross Number of Gross Credit Number of

Day Cash ($) Items Card ($) Items

1 348 55 148 4
2 42 8 111 6
3 61 9 62 7
4 94 16 0 0
5 60 11 39 5
6 165 26 7 1
7 126 27 143 26
8 111 19 27 5
9 26 5 14 2
10 109 18 71 12
11 180 27 116 21
12 212 36 50 9
13 58 10 13 2
14 115 20 105 16
15 15 8 19 3
16 97 15 44 14
17 61 10 0 0
18 85 15 24 3
19 157 24 144 10
20 88 15 63 11
21 96 19 0 0
22 202 33 14 3
23 108 23 0 0
24 158 21 24 4
25 176 43 253 28
320 CHAPTER 7 Multiple Regression Analysis

b. Define the dummy variable

1 if cash purchase
X2 = b
0 if credit card purchase
and fit the regression model
Y = b 0 + b 1X1 + b 2X2 +
c. Analyze the fit in part b. Be sure to include an analysis of the residuals. Are you
happy with your model?
d. Using the fitted model from part b, generate a forecast of daily sales for an
individual that purchases 25 items and pays cash. Construct a large-sample 95%
prediction interval for daily sales.
e. Describe the nature of the fitted function in part b. Do you think it is better to
fit two separate straight lines, one for the cash sales and another for the credit
card sales, to the data in Table P-15? Discuss.
16. Cindy Lawson just bought a major league baseball team. She has been receiving a lot
of advice about what she should do to create a winning ball club. Cindy asks you to
study this problem and write a report. You decide to use multiple regression analysis
to determine which statistics are important in developing a winning team (measured
by the number of games won during the 1991 season). You gather data for six statis-
tics from the Sporting News 1992 Baseball Yearbook, as shown in Table P-16, and run

TABLE P-16

Team Wins ERA SO BA Runs HR SB

Giants 75 4.03 905 .246 649 141 95
Mets 77 3.56 1,028 .244 640 117 153
Cubs 77 4.03 927 .253 695 159 123
Reds 74 3.83 997 .258 689 164 124
Pirates 98 3.44 919 .263 768 126 124
Cardinals 84 3.69 822 .255 651 68 202
Phillies 78 3.86 988 .241 629 111 92
Astros 65 4.00 1,033 .244 605 79 125
Dodgers 93 3.06 1,028 .253 665 108 126
Expos 71 3.64 909 .246 579 95 221
Braves 94 3.49 969 .258 749 141 165
Padres 84 3.57 921 .244 636 121 101
Red Sox 84 4.01 999 .269 731 126 59
White Sox 87 3.79 923 .262 758 139 134
Yankees 71 4.42 936 .256 674 147 109
Tigers 84 4.51 739 .247 817 209 109
Orioles 67 4.59 868 .254 686 170 50
Brewers 83 4.14 859 .271 799 116 106
Indians 57 4.23 862 .254 576 79 84
Blue Jays 91 3.50 971 .257 684 133 148
Mariners 83 3.79 1,003 .255 702 126 97
Rangers 85 4.47 1,022 .270 829 177 102
Athletics 84 4.57 892 .248 760 159 151
Royals 82 3.92 1,004 .264 727 117 119
Angels 81 3.69 990 .255 653 115 94
Twins 95 3.69 876 .280 776 140 107
 CHAPTER 7 Multiple Regression Analysis 321
a stepwise regression program, assuming a multiple regression model with “Wins” as
the dependent variable.
a. Discuss the importance of each independent variable.
b. What equation should Cindy use to forecast wins?
c. Write a report for Cindy.
d. Collect data from the most recent Sporting News Baseball Yearbook or other
source of baseball statistics. Run a stepwise regression, and compare your results.
17. Ms. Haight, a real estate broker, wishes to forecast the importance of four factors
in determining the prices of lots. She accumulates data on price, area, elevation,
and slope and rates the view for 50 lots. She runs the data on a correlation program
and obtains the correlation matrix given in Table P-17. Ms. Haight then runs the
data on a stepwise multiple regression program.
a. Determine which variable will enter the model first, second, third, and last.
b. Which variable or variables will be included in the best prediction equation?
18. The scores for two within-term examinations, X1 and X2; the current grade point
average (GPA), X3; and the final exam score, Y, for 20 students in a business statis-
tics class are listed in Table P-18.
a. Fit a multiple regression model to predict the final exam score from the scores
on the within-term exams and GPA. Is the regression significant? Explain.
b. Predict the final exam score for a student with within-term exam scores of 86
and 77 and a GPA of 3.4.
c. Compute the VIFs and examine the t statistics for checking the significance of
the individual predictor variables. Is multicollinearity a problem? Explain.

TABLE P-17
Variable
Variable Price Area Elevation Slope View
Price 1.00 .59 .66 .68 .88
Area 1.00 .04 .64 .41
Elevation 1.00 .13 .76
Slope 1.00 .63
View 1.00

TABLE P-18
X1 X2 X3 Y X1 X2 X3 Y
87 85 2.7 91 93 60 3.2 54
100 84 3.3 90 92 69 3.1 63
91 82 3.5 83 100 86 3.6 96
85 60 3.7 93 80 87 3.5 89
56 64 2.8 43 100 96 3.8 97
81 48 3.1 75 69 51 2.8 50
77 67 3.1 63 80 75 3.6 74
86 73 3.0 78 74 70 3.1 58
79 90 3.8 98 79 66 2.9 87
96 69 3.7 99 95 83 3.3 57
322 CHAPTER 7 Multiple Regression Analysis

d. Compute the mean leverage. Are any of the observations high leverage points?
e. Compute the standardized residuals. Identify any observation with a large stan-
dardized residual. Does the fitted model under- or overpredict the response for
these observations?
19. Refer to the data in Table P-18. Find the “best” regression model using the step-
wise regression procedure and the all possible regressions procedure. Compare the
results. Are you confident using a regression model to predict the final exam score
with fewer than the original three independent variables?
20. Recall Example 7.12. The full data set related to CEO compensation is contained
in Appendix C. (See pages 533–545.) Use stepwise regression to select the “best”
model with k = 3 predictor variables. Fit the stepwise model, and interpret the
estimated coefficients. Examine the residuals. Identify and explain any influential
observations. If you had to choose between this model and the k = 2 predictor
model discussed in Example 7.12, which one would you choose? Why?
21. Table P-21 contains the number of accounts (in thousands) and the assets (in bil-
lions of dollars) for 10 online stock brokerages. Plot the assets versus the number
of accounts. Investigate the possibility the relationship is curved by running a mul-
tiple regression to forecast assets using the number of accounts and the square of
the number of accounts as independent variables.
a. Give the fitted regression function. Is the regression significant? Explain.
b. Test for the significance of the coefficient of the squared term. Summarize your
conclusion.
c. Rerun the analysis without the quadratic (squared) term. Explain why the coeffi-
cient of the number of accounts is not the same as the one you found for part a.
22. The quality of cheese is determined by tasters whose scores are summarized in a
dependent variable called Taste. The independent (predictor) variables are three
chemicals that are present in the cheese: acetic acid, hydrogen sulfide (H2S), and
lactic acid. The 15 cases in the data set are given in Table P-22. Analyze these data
using multiple regression methods. Be sure to include only significant independent
variables in your final model and interpret R2. Include an analysis of the residuals.
23. Refer to Problem 22. Using your final fitted regression function, forecast Taste (qual-
ity) for Acetic = 5.750, H 2S = 7.300, and Lactic = 1.85. (All three independent
variable values may not be required.) Although n in this case is small, construct the

TABLE P-21
Assets Number of accounts
($ billions) X (1,000s) Y
219.0 2,500
21.1 909
38.8 615
5.5 205
160.0 2,300
19.5 428
11.2 590
5.9 134
1.3 130
6.8 125
 TABLE P-22
Case Taste Y Acetic X1 H2S X2 Lactic X3
1 40.9 6.365 9.588 1.74
2 15.9 4.787 3.912 1.16
3 6.4 5.412 4.700 1.49
4 18.0 5.247 6.174 1.63
5 38.9 5.438 9.064 1.99
6 14.0 4.564 4.949 1.15
7 15.2 5.298 5.220 1.33
8 32.0 5.455 9.242 1.44
9 56.7 5.855 10.199 2.01
10 16.8 5.366 3.664 1.31
11 11.6 6.043 3.219 1.46
12 26.5 6.458 6.962 1.72
13 0.7 5.328 3.912 1.25
14 13.4 5.802 6.685 1.08
15 5.5 6.176 4.787 1.25

TABLE P-24 Accounting Categories for Major League Baseball ($ millions)

GtReceit MediaRev StadRev TotRev PlayerCt OpExpens OpIncome FranValu Franchise

19.4 61.0 7.4 90.0 29.8 59.6 30.4 200 NyYank

26.6 32.5 18.0 79.3 36.0 72.0 7.3 180 LADodge
22.9 50.0 16.0 91.1 35.2 70.4 20.7 170 NYMets
44.5 30.0 12.0 88.7 29.7 62.4 26.3 160 TorBlJay
24.5 40.5 14.3 81.5 35.4 70.8 10.7 160 BosRdSox
19.0 24.4 5.0 50.6 15.8 39.5 11.1 140 BaltOrio
27.5 25.7 19.3 78.0 18.0 60.0 18.0 140 ChWhSox
19.9 25.0 12.0 59.1 23.2 46.4 12.7 132 StLCard
22.8 27.5 12.0 64.5 29.0 58.0 6.5 132 ChCubs
19.0 25.5 14.0 61.5 20.7 47.6 13.9 123 TxRanger
16.9 20.5 14.0 53.6 30.4 60.8 7.2 117 KCRoyals
15.2 23.2 7.6 48.2 21.7 43.4 4.8 115 PhilPhil
25.7 27.0 10.0 64.9 39.2 66.6 1.7 115 OakAthlt
19.0 27.9 5.0 54.1 34.3 61.7 -7.6 103 CalAngls
15.5 26.2 5.0 48.9 33.3 53.3 4.4 99 SFGiants
17.1 24.4 5.3 49.0 27.1 48.8 0.2 98 CinnReds
15.6 23.1 7.5 48.4 24.4 48.8 0.4 96 SDPadres
10.6 25.2 8.0 46.0 12.1 31.5 14.5 95 HousAstr
16.2 21.9 5.5 45.8 24.9 49.8 4.0 87 PittPirt
15.6 28.8 5.0 51.6 31.1 54.4 2.8 85 DetTiger
15.4 18.9 3.8 40.3 20.4 40.8 0.5 83 AtBraves
18.2 20.5 3.2 44.1 24.1 48.2 4.1 83 MinnTwin
15.5 22.0 5.0 44.7 17.4 41.8 2.9 79 SeatMar
14.2 19.4 3.0 38.8 26.4 50.2 11.4 77 MilBrews
9.5 23.7 6.6 42.0 19.5 46.8 4.8 77 ClevIndn
10.7 23.5 3.0 39.4 21.8 43.6 4.2 75 MonExpos
Source: M. Ozanian and S. Taub, “Big Leagues, Bad Business,” Financial World, July 7, 1992, pp. 34–51.

323
324 CHAPTER 7 Multiple Regression Analysis

large-sample approximate 95% prediction interval for your forecast. Do you feel your
regression analysis has yielded a useful tool for forecasting cheese quality? Explain.
24. The 1991 accounting numbers for major league baseball are given in Table P-24.
All figures are in millions of dollars. The numerical variables are GtReceit (Gate
Receipts), MediaRev (Media Revenue), StadRev (Stadium Revenue), TotRev
(Total Revenue), PlayerCt (Player Costs), OpExpens (Operating Expenses),
OpIncome 1Operating Income = Total Revenue - Operating Expenses2, and
FranValu (Franchise Value).
a. Construct the correlation matrix for the variables GtReceit, MediaRev, . . . ,
FranValu. From the correlation matrix, can you determine a variable that is
likely to be a good predictor of FranValu? Discuss.
b. Use stepwise regression to build a model for predicting franchise value using
the remaining variables. Are you surprised at the result? Explain.
c. Can we conclude that, as a general rule, franchise value is about twice total rev-
enue? Discuss.
d. Player costs are likely to be a big component of operating expenses. Develop an
equation for forecasting operating expenses from player costs. Comment on the
strength of the relation. Using the residuals as a guide, identify teams that have
unusually low or unusually high player costs as a component of operating
expenses.
e. Consider the variables other than FranValu. Given their definitions, are there
groups of variables that are multicollinear? If so, identify these sets.

CASES

CASE 7-1 THE BOND MARKET16

Judy Johnson, vice president of finance of a large, Judy called a meeting of her financial staff to
private, investor-owned utility in the Northwest, was discuss the bond market problem. One member of
faced with a financing problem. The company her staff, Ron Peterson, a recent M.B.A. graduate,
needed money both to pay off short-term debts said he thought a multiple regression model could
coming due and to continue construction of a coal- be developed to forecast the bond rates. Since the
fired plant. vice president was not familiar with multiple regres-
Judy’s main concern was estimating the 10- or sion, she steered the discussion in another direction.
30-year bond market; the company needed to decide After an hour of unproductive interaction, Judy
whether to use equity financing or long-term debt. then asked Ron to have a report on her desk the fol-
To make this decision, the utility needed a reliable lowing Monday.
forecast of the interest rate it would pay at the time Ron knew that the key to the development of a
of bond issuance. good forecasting model is the identification of

16Thedata for this case study were provided by Dorothy Mercer, an Eastern Washington University M.B.A.
student. The analysis was done by M.B.A. students Tak Fu, Ron Hand, Dorothy Mercer, Mary Lou
Redmond, and Harold Wilson.
 CHAPTER 7 Multiple Regression Analysis 325
independent variables that relate to the interest high positive relationship between the interest rate
rate paid by utilities at the time of bond issuance. paid by the utility at the time of bond issuance and
After discussing the problem with various people at the U.S. Treasury bond rate, r = .883. He also
the utility, Ron decided to investigate the following expected a fairly high positive relationship between
variables: a utility’s bond rating (Moody’s), a util- the dependent variable and the prime lending rate
ity’s ratio of earnings to fixed charges, the U.S. 1r = .5962. He was not too surprised to discover that
Treasury bond rates, the bond maturity (10 or 30 these two predictor variables were also related to
years), and the prime lending rate at the time of each other (potential multicollinearity, r = .7132.
issuance. The negative relationship between the dependent
Ron gathered data he believed might correlate variable and length of bond maturity (10 or 30
with bond interest rates for utility bond issues dur- years), r = - .221, was also a result that made sense
ing two previous years. At first, he was uncertain to Ron.
how to handle the utility bond rating. He finally Next, Ron ran a full model containing all the
decided to consider only utilities whose ratings predictor variables. Examination of the computed
were the same or slightly higher than that of his t values and/or the p-values, which are presented in
company. This decision provided him with a sample Table 7-19, indicated that perhaps the variable of
of 93 issues to analyze. But he was still worried ratio of earnings to fixed charges and certainly
about the validity of using the bond ratings as inter- the variable of prime interest rate were not
val data. He called his former statistics professor making a contribution to the forecast of the
and learned that dummy variables would solve the interest rate paid by the utility at the time of bond
problem. Thus, he coded the bond ratings in the fol- issuance.
lowing way: To check his interpretation of the full regression
results, Ron decided to run a stepwise regression.
X1 = 1 if the utility bond rating is A; The output is shown in Table 7-20.
0 otherwise Although the p-value, .035, associated with the
X2 = 1 if the utility bond rating is AA; t value for the ratio of earnings to fixed charges is
0 otherwise less than .05, Ron decided to drop the variable
Ratio from his regression equation. Further support
If the utility bond rating is BAA, both X1 and X2 are for this decision was provided by the small correla-
0. The next step was for Ron to select a multiple tion 1r = .0972 of Ratio to the dependent variable
regression program from the computer library and Interest Rate. The results of the Minitab run for
input the data. The following variables were Ron’s final model are shown in Table 7-21.
included in the full-model equation: Ron’s report to Judy included the following
comments:
Variable 1: Y  the interest rate paid by the utility
at the time of bond issuance 1. The best model, Interest = –1.28 –.929 AA – 1.18
Variable 2: X1  1 if the utility’s bond rating is A AA + 1.23 Bond rates + 0.0615 Maturity,
Variable 3: X2  1 if the utility’s bond rating is explains 90.6% of the interest rate variation.
AA 2. The standard error of the estimate is .53 based
Variable 4: X3  the utility’s ratio of earnings to on n = 93 observations. Thus, for a given fore-
fixed charges cast, YN , a 95% prediction interval for the actual
Variable 5: X4  the U.S. Treasury bond rates (for interest rate is YN ; 21.532 or YN ; 1.06.
10 and 30 years) at the time of bond issuance 3. The coefficients of the independent variables
Variable 6: X5  the bond maturity (10 or 30 are all significant (very small p-values) and
years) appear to be reliable. Multicollinearity is not a
Variable 7: X6  the prime lending rate at the problem.
time of issuance
Ron was very pleased with his effort and felt that
The actual data are presented in Appendix A. Judy would also be pleased.
Ron decided to analyze the correlation matrix
shown in Table 7-18. He was not surprised to find a
326 CHAPTER 7 Multiple Regression Analysis

TABLE 7-18 Correlation Matrix for Bond Market Study

Correlations
Interest Rate A AA Ratio Bond Rates Maturity
A –0.347
AA 0.173 0.399
Ratio 0.097 0.037 0.577
Bond rates 0.883 0.256 0.291 0.253
Maturity 0.221 0.278 0.010 0.094 0.477
Prime 0.596 0.152 0.342 0.255 0.713 0.314

TABLE 7-19 Full-Model Run for Bond Market Study

Regression Analysis
The regression equation is
Interest Rates = - 1.03 - 0.829 A - 0.889 AA - 0.242 Ratio + Bond Rates
+ 0.0628 Maturity - 0.0031 Prime
Predictor Coef SE Coef T P
Constant -1.0263 0.6572 -1.56 0.122
A -0.8285 0.1342 -6.17 0.000
AA -0.8894 0.2248 -3.96 0.000
Ratio -0.2417 0.1135 -2.13 0.036
Rates 1.25753 0.05964 21.08 0.000
Maturity 0.062839 0.006589 9.54 0.000
Prime -0.00313 0.02720 -0.12 0.909
S  0.5235 RSq  91.1% Rsq(adj)  90.5%
Analysis of Variance
Source DF SS MS F P
Regression 6 240.907 40.151 146.52 0.000
Residual Error 86 23.567 0.274
Total 92 264.474
Unusual Observations

Obs A Rate Fit St Dev Fit Residual St Resid

64 1.00 15.5000 13.7645 0.1627 1.7355 3.49R
66 1.00 11.3000 13.4780 0.1357 -2.1780 -4.31R
67 1.00 12.3800 13.7897 0.1297 -1.4097 -2.78R
68 1.00 12.1500 13.7097 0.0947 -1.5597 -3.03R
R denotes an observation with a large standardized residual.

QUESTION
1. What questions do you think Judy will have for
Ron?
 TABLE 7-20 Stepwise Regression for Bond Market Study
Stepwise Regression: Int rate versus A, AA, . . .
Alpha-to-Enter: 0.05 Alpha-to-Remove: 0.05
Response is Int rate on 6 predictors, with N = 93
Step 1 2 3 4 5
Constant 1.9818 -0.7641 -0.7372 -0.2046 -1.0244
Bondrate 1.029 1.173 1.267 1.231 1.253
T-Value 17.95 20.44 23.43 24.78 27.21
P-Value 0.000 0.000 0.000 0.000 0.000
Maturity 0.0439 0.0537 0.0588 0.0629
T-Value 5.25 7.02 8.37 9.60
P-Value 0.000 0.000 0.000 0.000
Ratio -0.547 -0.518 -0.241
T-Value -5.08 -5.29 -2.14
P-Value 0.000 0.000 0.035
A -0.56 -0.83
T-Value -4.49 -6.28
P-Value 0.000 0.000
AA -0.89
T-Value -4.10
P-Value 0.000
S 0.800 0.704 0.623 0.565 0.521
R–Sq 77.97 83.14 86.93 89.37 91.09
R–Sq(adj) 77.73 82.76 86.49 88.88 90.58
Mallows Cp 123.6 75.7 41.2 19.6 5.0

TABLE 7-21 Final Model for Bond Market Study

Regression Analysis
The regression equation is
Interest Rates = - 1.28 - 0.929 A - 1.18 AA + 1.23 Bond Rates + 0.0615 Maturity
Predictor Coef SE Coef T P
Constant -1.2765 0.6554 - 1.95 0.055
A -0.9293 0.1264 - 7.35 0.000
AA -1.1751 0.1781 - 6.60 0.000
Rates 1.23308 0.05964 26.81 0.000
Maturity 0.061474 0.006649 9.25 0.000
S = 0.5310 R -sq = 90.6% R-sq1adj2 = 90.2%
Analysis of Variance
Source DF SS MS F P
Regression 4 239.665 59.916 212.53 0.000
Residual Error 88 24.809 0.282
Total 92 264.474
Unusual Observations
Obs A Rate Fit St Dev Fit Residual St Resid
64 1.00 15.5000 13.6956 0.0894 1.8044 3.45R
66 1.00 11.3000 13.6462 0.0898 -2.3462 -4.48R
67 1.00 12.3800 13.6462 0.0898 -1.2662 -4.42R
68 1.00 12.1500 13.6462 0.0898 -1.4962 -2.86R
R denotes an observation with a large standardized residual.

327
328 CHAPTER 7 Multiple Regression Analysis

CASE 7-2 AAA WASHINGTON

An overview of AAA Washington was provided in operate best at 65 degrees Fahrenheit and (2) call vol-
Case 5-5 when students were asked to prepare a ume seems to increase more sharply when the average
time series decomposition of the emergency road temperature drops a few degrees from an average
service calls received by the club over five years. The temperature in the 30s than it does when a similar
analysis performed in Case 5-5 showed that the pat- drop occurs with an average temperature in the 60s.
tern in emergency road service call volume was This information suggested that the effect of tempera-
probably somewhat cyclical in nature. In Case 6-6, ture on emergency road service was nonlinear.
four variables were investigated: unemployment Michael DeCoria had stated in Case 6-6 that he
rate, average daily temperature, amount of rainfall, believes the number of road service calls received
and number of members in the club. is related to the general economic cycle and that the
Average daily temperature, rainfall, and per- Washington State unemployment rate is a good
haps, unemployment rate were determined to be sig- surrogate measurement for the general state of
nificant variables. The number of members in the Washington’s economy. Now he has observed that the
club was not a significant variable. cyclical trend of the time series seems to be lagging
A conversation with the manager of the emer- behind the general economic cycle. Data on the aver-
gency road service call center led to two important age monthly temperature and the Washington State
observations: (1) Automakers seem to design cars to unemployment rate are presented in Table 7-22.

TABLE 7-22 Data for AAA Washington

Year Month Calls Unemployment Rate Temp.
1987 June 6.9940
July 6.8137
August 6.3648
September 6.5435
October 6.7826
November 6.9225
December 7.1560
1988 January 7.9143
February 7.7576
March 7.0641
April 6.0977
May 20,002 5.7867 55.1
June 21,591 5.7592 59
July 22,696 5.5718 63.8
August 21,509 5.2939 63.8
September 22,123 5.4709 59.1
October 21,449 5.5049 54.6
November 23,475 5.8630 45.4
December 23,529 6.1349 41
1989 January 23,327 7.5474 40.3
February 24,050 7.8157 34.3
March 24,010 7.1390 43.2
April 19,735 6.2637 52.5
May 20,153 5.8332 55.3
June 19,512 5.8077 62.4
July 19,892 5.6713 62.9
 CHAPTER 7 Multiple Regression Analysis 329

TABLE 7-22 (Continued)

Year Month Calls Unemployment Rate Temp.
August 20,326 5.4977 63.5
September 19,378 5.2989 60.9
October 21,263 5.6028 51.9
November 21,443 5.9143 46.2
December 23,366 5 41.8
1990 January 23,836 6.1917 41.8
February 23,336 6.3775 38.9
March 22,003 5.7234 46.3
April 20,155 4.7792 51.7
May 20,070 4.5715 54.9
June 19,588 4.3899 59.8
July 20,804 4.2559 66.7
August 19,644 3.9359 66.4
September 17,424 3.9048 61.9
October 20,833 4.4294 50.4
November 22,490 5.1523 45.8
December 24,861 5.5102 33.9
1991 January 23,441 6.8901 37.9
February 19,205 7.0308 46.9
March 20,386 6.7186 43.4
April 19,988 6.1280 49.1
May 19,077 5.8146 54.3
June 19,141 5.9480 58.2
July 20,883 5.9026 65.4
August 20,709 5.7227 66
September 19,647 5.6877 60.9
October 22,013 6.2922 51
November 22,375 7.0615 46.2
December 22,727 7.4370 42.4
1992 January 22,367 8.4513 43
February 21,155 8.7699 46
March 21,209 8.0728 48.9
April 19,286 7.2392 52.7
May 19,725 7.0461 58.3
June 20,276 7.0478 63.6
July 20,795 7.1080 64.9
August 21,126 6.7824 65
September 20,251 6.7691 58.4
October 22,069 7.5896 53.2
November 23,268 7.9908 44.8
December 26,039 8.2460 37.8
1993 January 26,127 9.5301 34.9

QUESTIONS
1. Develop a multiple regression equation using 2. Create a new temperature variable and relate it to
unemployment rate and average monthly tem- emergency road service. Remember that tempera-
perature to predict emergency road service calls. ture is a relative scale and that the selection of the
330 CHAPTER 7 Multiple Regression Analysis

zero point is arbitrary. If vehicles are designed to the month (1) three months prior to the current
operate best at 65 degrees Fahrenheit, then every month and (2) 11 months prior to the current
degree above or below 65 degrees should make month as the data for the unemployment inde-
vehicles operate less reliably.To accomplish a trans- pendent variable. Which model is better for pre-
formation of the temperature data that simulates diction? Are the signs on the coefficients for the
this effect, begin by subtracting 65 from the average independent variables what you would expect
monthly temperature values. This repositions them to be? Are the coefficients of the indepen-
“zero” to 65 degrees Fahrenheit. Should absolute dent variables significantly different from zero?
values of this new temperature variable be used? 4. Develop a multiple regression equation using the
3. Create a new unemployment rate variable and transformed average temperature variable cre-
relate it to emergency road service. Give unem- ated in step 2 and the lagged unemployment vari-
ployment a lagged effect on emergency road able created in step 3. Is this a good model? Have
service by using the unemployment rate for any of the underlying assumptions been violated?

CASE 7-3 FANTASY BASEBALL (A)

It was February, and John Hanke, a retired professor accurately forecast pitching performances for start-
of statistics, was preparing for another fantasy base- ing pitchers.
ball season. In past years, his fellow players had The three categories that Hanke wished to
always teased him about using his knowledge of sta- research were wins (WINS), earned run average
tistics to gain an advantage. Unfortunately, it had (ERA), and walks and hits given up per innings
never been true. Teaching, researching, publishing, pitched (WHIP). He had spent a considerable amount
and committee work had kept him far too busy. of time downloading baseball statistics for starting
Now, having recently retired, he finally had the time pitchers from the 2006 season.17 He intended to
to apply his knowledge of statistics to the annual develop a multiple regression model to forecast each
rotisserie draft. In this type of fantasy league, each of the three categories of interest. He had often
manager has $260 with which to bid on and purchase preached to his students that the initial variable selec-
23 players (14 hitters and 9 pitchers). Each team is tion was the most important aspect of developing a
then ranked (based on actual player statistics from regression model. He knew that, if he didn’t have good
the previous season) in eight statistical categories. predictor variables, he wouldn’t end up with useful
Dr. Hanke was very concerned with choosing play- prediction equations. After a considerable amount of
ers who would perform well on three out of the four work, Dr. Hanke chose the five potential predictor
pitching categories. In past years, his pitching staff, variables that follow. He also decided to include only
especially his starting pitchers, had been the laugh- starting pitchers who had pitched at least 100 innings
ing stock of the league. The 2007 season was going to during the season. A portion of the data for the 138
be different. He intended to develop models to starting pitchers selected is presented in Table 7-23.

TABLE 7-23 Pitching Statistics for 138 Starting Pitchers

Pitcher TEAM ERA WHIP CMD K/9 HR/9 OBA THROWS
1 Armas, Jr., T WAS 5.03 1.50 5.7 1.5 1.1 339 1
2 Arroyo, B CIN 3.29 1.19 6.9 2.9 1.2 289 1
3 Batista, M ARI 4.58 1.53 4.8 1.3 0.8 343 1
4 Beckett, J BOS 5.01 1.29 6.9 2.1 1.6 307 1
5 Bedard, E BAL 3.76 1.35 7.8 2.5 0.7 316 0

17Baseballpitching statistics were downloaded from Ron Shandler’s BaseballHQ website at

www.BaseballHQ.com.
 TABLE 7-23 (Continued)

Pitcher TEAM ERA WHIP CMD K/9 HR/9 OBA THROWS

6 Benson, K BAL 4.82 1.40 4.3 1.5 1.6 325 1
7 Blanton, J OAK 4.82 1.54 5.0 1.8 0.8 345 1
8 Bonderman, J DET 4.08 1.30 8.5 3.2 0.8 308 1
9 Bonser, B MIN 4.22 1.28 7.5 3.5 1.6 304 1
10 Buchholz, T HOU 5.89 1.25 6.1 2.3 1.7 299 1
11 Buehrle, M CHW 4.99 1.45 4.3 2.0 1.6 331 0
12 Burnett, A TOR 3.98 1.30 7.8 3.0 0.9 309 1
13 Bush, D MIL 4.41 1.14 7.1 4.4 1.1 280 1
14 Byrd, P CLE 4.88 1.51 4.4 2.3 1.3 341 1
15 Cabrera, D BAL 4.74 1.58 9.5 1.5 0.7 351 1
16 Cain, M SF 4.15 1.28 8.4 2.1 0.8 305 1
17 Capuano, C MIL 4.03 1.25 7.1 3.7 1.2 299 0
18 Carpenter, C STL 3.09 1.07 7.5 4.3 0.9 268 1
19 Carrasco, H LAA 3.41 1.20 6.5 2.7 0.9 291 1
20 Chacon, S PIT 6.36 1.72 5.1 1.0 1.9 370 1
: : : : : : : : :
: : : : : : : : :
119 Tomko, B LAD 4.73 1.35 6.1 2.6 1.4 317 1
120 Trachsel, S NYM 4.97 1.60 4.3 1.0 1.3 354 1
121 Vargas, C ARI 4.83 1.41 6.6 2.4 1.4 326 1
122 Vazquez, J CHW 4.84 1.29 8.2 3.3 1.0 307 1
123 Verlander, J DET 3.63 1.33 6.0 2.1 1.0 313 1
124 Wakefield, T BOS 4.63 1.33 5.8 1.8 1.2 313 1
125 Wang, C NYY 3.63 1.31 3.1 1.5 0.5 309 1
126 Washburn, J SEA 4.67 1.35 5.0 1.9 1.2 317 0
127 Weaver, J STL 5.76 1.51 5.6 2.3 1.8 341 1
128 Weaver, J LAA 2.56 1.03 7.7 3.2 1.1 261 1
129 Webb, B ARI 3.10 1.13 6.8 3.6 0.6 279 1
130 Westbrook, J CLE 4.17 1.43 4.6 2.0 0.6 329 1
131 Williams, W SD 3.65 1.29 4.5 2.1 1.3 306 1
132 Willis, D FLA 3.87 1.42 6.4 1.9 0.8 327 0
133 Woods, J SEA 4.20 1.60 5.7 1.2 1.0 354 0
134 Wright, J SF 5.19 1.48 4.6 1.2 0.9 336 1
135 Wright, J NYY 4.49 1.52 5.4 1.5 0.6 343 1
136 Young, C SD 3.46 1.13 8.2 2.4 1.4 279 1
137 Zambrano, C CHC 3.41 1.29 8.8 1.8 0.8 307 1
138 Zito, B OAK 3.83 1.40 6.1 1.5 1.1 324 0
*The full data set is available on website: www.prenhall.com/hanke under Chapter 7 Case 7-3.

The variables are defined as K/9: How many batters a pitcher strikes out per
game (nine innings pitched)
HR/9: Opposition homeruns per game (nine
ERA: Earned run average or the number of
innings pitched)
earned runs allowed per game (nine innings
OBA: Opposition batting average
pitched)
THROWS: Right-handed pitcher (1) or left-
WHIP: Number of walks plus hits given up per
handed pitcher (0)
inning pitched
CMD: Command of pitches, the ratio strikeouts/ The next step in the analysis was the creation
walks of the correlation matrix shown in Table 7-24.
331
332 CHAPTER 7 Multiple Regression Analysis

TABLE 7-24 Correlations: ERA, THROWS, WHIP, K/9, CMD, HR/9, OBA

ERA THROWS WHIP K/9 CMD HR/9

THROWS 0.036
WHIP 0.825 -0.039
K/9 -0.445 -0.061 -0.503
CMD -0.462 0.121 -0.664 0.570
HR/9 0.562 0.023 0.341 -0.238 -0.184
OBA 0.825 -0.039 0.998 -0.510 -0.673 0.341

Dr. Hanke found the correlations between ERA variables, with the exception of WHIP, as indepen-
and WHIP and between ERA and OBA to be the dent variables. The results are shown in Table 7-25.
same, .825. Moreover, the correlation between From Table 7-25, the independent variables
WHIP and OBA, .998, indicated these variables are THROWS and K/9 are not significant, given the
strongly linearly related. Consequently, pitchers who other variables in the regression function. Moreover,
performed well on one of these variables should the small VIFs suggest that THROWS and K/9 can be
perform well on the other two. Dr. Hanke decided dropped together and the coefficients for the remain-
that ERA is the best indicator of performance and ing variables will not change much. Table 7-26 shows
decided to run a regression to see how well the vari- the result when both THROWS and K/9 are left out
ables in his collection predicted ERA. He knew, of the model.The R2 is 78.1%, and the equation looks
however, that the very high correlation between good.The t statistic for each of the predictor variables
WHIP and OBA would create a multicollinearity is large with a very small p-value. The VIFs are small
problem, so only one of these variables would be for the three predictors, indicating that multi-
required in the regression function. Tossing a coin, collinearity is no longer a problem.
Dr. Hanke selected OBA and ran a regression with Dr. Hanke decided that he has a good model and
ERA as the dependent variable and the remaining developed the residual plots shown in Figure 7-4.

TABLE 7-25 Minitab Regression Output Using All Predictor Variables Except WHIP

Regression Analysis: ERA versus THROWS, K/9, CMD, HR/9, OBA

The regression equation is
ERA = - 5.54 + 0.0779 THROWS - 0.0303 K>9 + 0.125 CMD + 0.840 HR>9 + 0.0280 OBA
Predictor Coef SE Coef T P VIF
Constant -5.5439 0.7395 -7.50 0.000
THROWS 0.07792 0.08507 0.92 0.361 1.042
K/9 -0.03028 0.02977 -1.02 0.311 1.602
CMD 0.12517 0.04671 2.68 0.008 2.181
HR/9 0.8404 0.1198 7.02 0.000 1.151
OBA 0.027991 0.002025 13.82 0.000 2.068
S = 0.437305 R -Sq = 78.4% R- Sq1adj2 = 77.6%
Analysis of Variance
Source DF SS MS F P
Regression 5 91.783 18.357 95.99 0.000
Residual Error 132 25.243 0.191
Total 137 117.026
 CHAPTER 7 Multiple Regression Analysis 333

TABLE 7-26 Minitab Final Regression Output for Forecasting ERA

Regression Analysis: ERA versus CMD, HR/9, OBA
The regression equation is
ERA = - 5.81 + 0.113 CMD + 0.857 HR>9 + 0.0284 OBA
Predictor Coef SE Coef T P VIF
Constant - 5.8088 0.6862 - 8.46 0.000
CMD 0.11298 0.04290 2.63 0.009 1.837
HR/9 0.8573 0.1191 7.20 0.000 1.137
OBA 0.028425 0.001997 14.23 0.000 2.009
S = 0.437628 R-Sq = 78.1% R- Sq1adj2 = 77.6%
Analysis of Variance
Source DF SS MS F P
Regression 3 91.363 30.454 159.01 0.000
Residual Error 134 25.663 0.192
Total 137 117.026

FIGURE 7-4 Residual Plots for Forecasting ERA

334 CHAPTER 7 Multiple Regression Analysis

QUESTIONS
1. Comment on the model that Dr. Hanke devel- 3. Develop a model to forecast ERA using the
oped to forecast ERA. Examine the residual predictor variable WHIP instead of OBA.
plots shown in Figure 7-4 and determine whether Which model do you prefer, the one with OBA
this model is valid. as a predictor variable or the one with WHIP as
2. Are there any nonlinear relationships between a predictor variable? Why?
the predictor variables and earned run average?
If so, develop a new model including the appro-
priate variable transformation.

CASE 7-4 FANTASY BASEBALL (B)18

Now that Dr. Hanke felt he had successfully devel- 3. Bullpen support
oped a model for forecasting ERA, he was ready to 4. Team defense
tackle WINS.19 However, the expert consensus indi-
cated the project was doomed to failure. Comments Dr. Hanke was ready to develop a predictor database.
found on Ron Shandler’s BaseballHQ website indi- In order to project a team’s offense, he chose
cated: “There’s no way to project pitchers accurately runs scored by a team (RUNS)20 as the most impor-
from year to year” (Bill James); “Your most valuable tant variable. In order to indicate how good a team
commodity is a starting pitcher you can count on. was overall, Dr. Hanke chose team wins (TMWINS).
The only problem is, you can’t count on any of Six of the seven variables used in Case 7-3 were used
them” (Peter Golenbock); “Where else in the realm to indicate a pitcher’s effectiveness: ERA, CMD,
of fantasy sports can you have worse odds on suc- K/9, HR/9, OBA, and THROWS. For team defense,
cess than from the wonderful world of the pitcher?” he used the total number of team errors
(Rod Beaton); “Starting pitchers are the most unre- (ERRORS). For bullpen support, he tried saves
liable, unpredictable, unpleasant group of people in (SAVES).
the world, statistically speaking that is” (John A partial data set for the pitching effective-
Benson). ness variables is given in Table 7-23. Some of the
Dr. Hanke investigated and found a statistical data for the rest of the variables are presented in
model that could be used. According to Ron Table 7-27. The remaining data for both pitcher
Shandler’s BaseballHQ, four variables have an effectiveness and team statistics are available
important impact on pitching win totals: on the website: www.prenhall.com/hanke under
Chapter 7 Case 7-4.
1. Team offense
2. Pitching effectiveness

18Case
7-4 is based on a model developed by Ron Shandler’s BaseballHQ.
19Baseball pitching statistics were downloaded from Ron Shandler’s BaseballHQ website at
www.BaseballHQ.com.
20Statistics
for RUNS, TMWINS, and SAVES are from SportsTicker® and appear at www.sports.espn.go.
com/mlb/stats.
 CHAPTER 7 Multiple Regression Analysis 335

TABLE 7-27 Team Statistics for Starting Pitchers

Pitcher WINS TM WINS RUNS ERRORS SAVES

1 Armas Jr., T 9 71 746 131 32
2 Arroyo, B 14 80 749 128 36
3 Batista, M 11 76 773 104 34
4 Beckett, J 16 86 820 66 46
5 Bedard, E 15 70 768 102 35
6 Benson, K 11 70 768 102 35
7 Blanton, J 16 93 771 84 54
8 Bonderman, J 14 95 822 106 46
9 Bonser, B 7 96 801 84 40
10 Buchholz, T 6 82 735 80 42
11 Buehrle, M 12 90 868 80 46
12 Burnett, A. 10 87 809 99 42
13 Bush, D 12 75 730 117 43
14 Byrd, P 10 78 870 118 24
15 Cabrera, D 9 70 768 102 35
16 Cain, M 13 76 746 91 37
17 Capuano, C 11 75 730 117 43
18 Carpenter, C 15 83 781 98 38
19 Carrasco, H 7 89 766 124 50
20 Chacon, S 7 67 691 104 39
: : : : : : :
: : : : : : :
119 Tomko, B 8 88 820 115 40
120 Trachsel, S 15 97 834 104 43
121 Vargas, C 12 76 773 104 34
122 Vazquez,J 11 90 868 80 46
123 Verlander, J 17 95 822 106 46
124 Wakefield, T 7 86 820 66 46
125 Wang, C 19 97 930 104 43
126 Washburn, J 8 78 756 88 47
127 Weaver, J 8 83 781 98 38
128 Weaver, J 11 89 766 124 50
129 Webb, B 16 76 773 104 34
130 Westbrook, J 15 78 870 118 24
131 Williams, W 12 88 731 92 50
132 Willis, D 12 78 758 126 41
133 Woods, J 7 78 756 88 47
134 Wright, J 6 76 746 91 37
135 Wright, J 11 97 930 104 43
136 Young, C 11 88 731 92 50
137 Zambrano, C 16 66 716 106 29
138 Zito, B 16 93 771 84 54
*The full data set is available on website: www.prenhall.com/hanke under Chapter 7
Case 7-4.
336 CHAPTER 7 Multiple Regression Analysis

QUESTION
1. The expert consensus indicated the project was
doomed to failure. Were the experts right?

Minitab Applications

The problem. In Example 7.11, Pam Weigand wanted to run stepwise regression on
the data for the Zurenko Pharmaceutical Company to forecast which applicant would
become a good salesperson.

Minitab Solution

1. If the data are on file, open the file (Tab 7-13) using the following menus:

File>Open Worksheet

If the data are not on file, enter them on the worksheet. To run stepwise regression,
click on the following menus:

Stat>Regression>Stepwise

2. The Stepwise Regression dialog box shown in Figure 7-5 appears.

a. The dependent or Response variable is C1 or Sales.
b. The predictor variables are C2-C6 or Aptitude-GPA.
c. Click on Methods to provide the α level.

FIGURE 7-5 Minitab Stepwise Regression Dialog Box

 CHAPTER 7 Multiple Regression Analysis 337

FIGURE 7-6 Minitab Stepwise-Methods Dialog Box

3. The Stepwise-Methods dialog box shown in Figure 7-6 appears.

a. Click on Use alpha values.
b. Click on Stepwise (forward and backward).
c. Change the Alpha to enter and Alpha to remove from .15 to .05.
d. Click on OK and then on OK on the Stepwise Regression dialog box. The
results are shown in Table 7-17.

Excel Applications

The problem. In Example 7.1, multiple regression analysis was used to determine
whether selling price and advertising expense could be used to forecast weekly sales
for gallons of milk (see Table 7-4).

Excel Solution

1. Enter the label Sales in A1, Price in B1, and Adv. in C1. Enter weekly sales into A2
through A11, selling prices into B2 through B11, and advertising expenses into C2
through C11 of the worksheet. The two predictor variables must be in adjacent
columns.
2. Click on the following menus:

Tools>Data Analysis
338 CHAPTER 7 Multiple Regression Analysis

3. The Data Analysis dialog box appears. Under Analysis Tools, choose Regression.
The Regression dialog box shown in Figure 6-22 appears.
a. Enter A1 to A11 in the Input Y Range.
b. Enter B1 to C11 in the Input X Range for the two predictor variables, selling
price and advertising expense.
c. Click on Labels.
d. Click on Output Range and enter D1.
e. Click on OK, and the output appears in the worksheet.

References
Abraham, B., and J. Ledolter. Introduction to Draper, N. R., and H. Smith. Applied Regression
Regression Modeling. Belmont, Calif.: Thomson Analysis, 3rd ed. New York: Wiley, 1998.
Brooks/Cole, 2006. Johnson, R. A., and D. W. Wichern. Business
Belsley, D. A. Conditioning, Diagnostics, Collinearity Statistics: Decision Making with Data. New York:
and Weak Data in Regression. New York: Wiley, Wiley, 1997.
1991. Kutner, M. H., C. J. Nachtsheim, and J. Neter.
Bowerman, B. L., R. T. O’Connell, and A. B. Koehler. Applied Linear Regression Models, 4th ed. New
Forecasting, TimeSeries and Regression, 4th ed. York: McGraw-Hill, 2004.
Belmont, Calif: Thomson Brooks/Cole, 2005.
C H A P T E R

8 REGRESSION WITH TIME

SERIES DATA
Many business and economic applications of forecasting involve time series data.
Regression models can be fit to monthly, quarterly, or yearly data, using the techniques
described in previous chapters. However, because data collected over time tend to
exhibit trends, seasonal patterns, and so forth, observations in different time periods are
related or autocorrelated.That is, for time series data, the sample of observations cannot
be regarded as a random sample. Problems of interpretation can arise when standard
regression methods are applied to observations that are related to one another over
time. Fitting regression models to time series data must be done with considerable care.

TIME SERIES DATA AND THE PROBLEM OF AUTOCORRELATION

The regression models discussed in Chapters 6 and 7 assume that the errors, ε, are
independent (or uncorrelated) random variables. This means that the different values
of the response variable, Y, can be related to the values of the predictor variables, the
X ’s, but not to one another. The usual interpretations of the results of a regression
analysis depend heavily on the assumption of independence.
With time series data, the assumption of independence rarely holds. Consider the
annual base price for a particular model of a new car. Can you imagine the chaos that
would exist if the new car prices from one year to the next were indeed unrelated (inde-
pendent) of one another? In such a world, prices would be determined like numbers drawn
from a random number table. Knowledge of the price in one year would not tell you any-
thing about the price in the next year. In the real world, price in the current year is related
to (correlated with) the price in the previous year, and maybe the price two years ago, and
so forth.That is, the prices in different years are autocorrelated; they are not independent.
From the discussion of autocorrelation in Chapter 3:

Autocorrelation exists when successive observations over time are related to one
another.

Autocorrelation can occur because the effect of a predictor variable on the

response is distributed over time. For example, an increase in salary may affect your
consumption (or saving) not only in the current period but also in several future peri-
ods. A change in price may affect sales in the current period and in future periods.
A current labor contract may affect the cost of production for some time to come. Over
time, relationships tend to be dynamic (evolving), not static.

339
340 CHAPTER 8 Regression with Time Series Data

From a forecasting perspective, autocorrelation is not all bad. If values of a

response, Y, in one time period are related to Y values in previous time periods, then
previous Y’s can be used to predict future Y’s.1 In a regression framework, autocor-
relation is handled by “fixing up” the standard regression model. To accommodate
autocorrelation, sometimes it is necessary to change the mix of predictor variables
and/or the form of the regression function. More typically, however, autocorrelation
is handled by changing the nature of the error term.
A common kind of autocorrelation, sometimes called first-order serial correlation,
is one in which the error term in the current time period is directly related to the error
term in the previous time period. In this case, with the subscript t representing time, the
simple linear regression model takes the form

Yt = b 0 + b 1Xt + t (8.1)

with

t = r t - 1 + t (8.2)

where
t = the error at time t
r = the parameter1lag 1 autocorrelation coefficient2 that measures correlation
between adjacent error terms
t = the normally distributed independent error term with mean zero
and variance s2
Equation 8.2 says that the level of one error term 1 t - 12 directly affects the level of the
next error term 1 t2. The magnitude of the autocorrelation coefficient, ρ, where
- 1 … r 6 1, indicates the strength of the serial correlation. If ρ is zero, then there is no
serial correlation, and the error terms are independent 1 t = t2.
Figure 8-1 illustrates the effect of positive serial correlation in a simple linear
regression model. Suppose the true relation between Y and X, indicated by the solid
line in the figure, is increasing over time. If the first Y value is above the true regression
line, then the next several Y values are likely to be above the line because of the positive
autocorrelation (if the first error is positive, the second error is likely to be positive, and
so forth). Eventually, there may be a sequence of Y’s below the true regression line
(a negative error is likely to be followed by negative error).The data are “tilted” relative
to the true X, Y relationship. However, the least squares line, by its very nature, will pass
through the observations, as indicated by the dotted line in the figure. Using the dotted
line to make inferences about the solid line or using the dotted line to generate forecasts
of future Y’s could be very misleading.
It should also be clear from Figure 8-1 that, in this case, the scatter about the least
squares line is tighter than it is about the true regression line. Consequently, the standard
error of the estimate, sy #x, will underestimate the variability of the Y’s about the true regres-
sion line or, equivalently, underestimate the standard deviation, σ, of the error term, ε.
Strong autocorrelation can make two unrelated variables appear to be related.
Standard regression procedures applied to observations on these variables can pro-
duce a significant regression. In this case, the estimated relationship is spurious, and an

1Thisidea is considered again in a later section of this chapter when autoregressive models are discussed and
then developed more fully in Chapter 9.
 CHAPTER 8 Regression with Time Series Data 341

Y
True Regression Line

Fitted Regression Line

^
Y
Y

FIGURE 8-1 Positive Serial Correlation and the Least

Squares Procedure

Time Series Plot of Yt , Xt

40 Variable
Yt
Xt
30

20
Data

10

−10

1 10 20 30 40 50 60 70 80 90 100
Time

FIGURE 8-2 Time Series Plots of Two Unrelated Series, Yt (top) and Xt (bottom)

examination of the residuals will ordinarily reveal the problem.2 However, with an
uncritical application of standard procedures, the spurious regression may go unde-
tected, resulting in a serious misinterpretation of the results.
Example 8.1
Figure 8-2 contains plots of two computer-generated time series, Yt and Xt . These two
series were formed in such a way that the first series 1Yt2 is not related to the second series

2Two (or more) autocorrelated time series can be related, but special care is required to uncover the rela-
tionship. One special case is briefly considered in the Cointegrated Time Series section later in this chapter.
342 CHAPTER 8 Regression with Time Series Data

FIGURE 8-3 Autocorrelations for the Yt Series

1Xt2.3 At the same time, each sequence of observations is highly autocorrelated. The auto-
correlations for the first series are shown in Figure 8-3. The autocorrelations for the sec-
ond series (not shown) are very similar.
Figure 8-2 indicates that the two time series appear to move together. In fact, it might
be possible to relate the top series to the bottom series, using a simple linear regression
model. A scatter diagram of the data is shown in Figure 8-4 along with the least squares line.
The statistic R2 is also shown in the figure. The estimated regression is significant (a full
analysis gives F = 83.08 and p-value = .000), with Xt explaining about 46% of the variabil-
ity in Yt. Yet the Yt series was generated independently of the Xt series. That is, the X’s were
not used to generate the Y’s. The estimated regression in this case is spurious.
An examination of the residuals would reveal problems with this regression analysis.
For example, the residual autocorrelations are large for several lags, indicating that the
assumption of independent errors is wrong and that the initial regression model should be
modified. In this case, if the modification is done correctly, the spurious relation between Yt
and Xt would disappear.

If regression models are used with autocorrelated (time series) data, it is especially
important to examine the residuals. If this is not done, it is possible to reach conclusions

3The first series was constructed by selecting a random sample of 100 values from a normal distribution with
mean 0 and a standard deviation 2 and then by forming partial sums. For example, the first observation in the
series was the first value selected in the sample, the second observation in the series was the sum of the first two
values, the third observation was the sum of the first three values, and so forth.The second series was constructed
in the same way, beginning with a different random sample of 100 values from the same normal distribution.
 CHAPTER 8 Regression with Time Series Data 343

Fitted Line Plot

Yt = 12.92 + 1.111 Xt

35 S 6.62467
R-Sq 45.9%
R-Sq(adj) 45.3%
30

25

20
Yt
15

10

−10 −5 0 5 10 15
Xt

FIGURE 8-4 Simple Linear Regression Results for Two Highly Autocorrelated Time
Series, Yt and Xt

that are not justified. The fault is not with the least squares procedure. The fault lies in
applying the standard regression model in a situation that does not correspond to the
usual regression assumptions. The technical problems that arise include the following:

1. The standard error of the estimate can seriously underestimate the variability of the
error terms.
2. The usual inferences based on the t and F statistics are no longer strictly applicable.
3. The standard errors of the regression coefficients underestimate the variability of the
estimated regression coefficients. Spurious regressions can result.

AUTOCORRELATION AND THE DURBIN-WATSON TEST

As we discussed in Chapter 3, autocorrelation can be examined by constructing the
autocorrelation coefficients and comparing them with their standard errors (see
Equations 3.1 and 3.2). In addition, an overall test for significant autocorrelation is
provided by the Ljung-Box Q statistic (Equation 3.3). In the present context, we
would apply these procedures to the residuals from a regression analysis with time
series data to check for randomness. Equation 6.17 gives the residual autocorrelation
coefficients.
Examining the residual autocorrelations directly is good practice and should be
employed. However, Minitab and other computer programs provide the option of
directly computing a statistic that is useful for detecting first order-serial correlation,
or lag 1 autocorrelation, known as the Durbin-Watson (DW) statistic. A test for
significant first-order serial correlation based on this statistic is known as the
344 CHAPTER 8 Regression with Time Series Data

Durbin-Watson test.4 The test involves the determination of whether the autocorrela-
tion parameter, ρ, shown in Equation 8.2, is zero.
Consider
t = r t - 1 + t
The hypotheses to be tested are
H0: r = 0
H1: r 7 0
The alternative hypothesis is r 7 0, since business and economic time series typically
tend to show positive autocorrelation.
If a regression model does not properly account for autocorrelation, the residuals
will be autocorrelated. As a result, the Durbin-Watson test is carried out using the
residuals from the regression analysis.
The Durbin-Watson statistic is defined as
n

a 1et - et - 12
2
t=2
DW = n (8.3)
2
a et
t=1
where
et = Yt - YNt = the residual for time period t
et - 1 = Yt - 1 - YNt - 1 = the residual for time period t - 1
For positive serial correlation, successive residuals tend to be alike. and the sum of
squared differences in the numerator of the Durbin-Watson statistic will be relatively
small. Small values of the Durbin-Watson statistic are consistent with positive serial
correlation.
The autocorrelation coefficient, ρ, can be estimated by the lag 1 residual autocor-
relation, r11e2 (see Equation 6.17), and with a little bit of mathematical maneuvering,
the Durbin-Watson statistic can be related to r11e2. For moderate to large samples,

DW = 211 - r11e22 (8.4)

Since - 1 6 r11e2 6 1, Equation 8.4 shows that 0 6 DW 6 4. For r11e2 close to zero,
the DW statistic will be close to 2. Positive lag 1 residual autocorrelation is associated
with DW values less than 2, and negative lag 1 residual autocorrelation is associated with
DW values above 2.
A useful, but sometimes not definitive, test for serial correlation can be performed
by comparing the calculated value of the Durbin-Watson statistic with lower 1dL2 and
upper 1dU2 bounds. The decision rules are as follows:
1. When the Durbin-Watson statistic is larger than the upper 1dU2 bound, the auto-
correlation coefficient, ρ, is equal to zero (there is no positive autocorrelation).
2. When the Durbin-Watson statistic is smaller than the lower 1dL2 bound, the auto-
correlation coefficient, ρ, is greater than zero (there is positive autocorrelation).
3. When the Durbin-Watson statistic lies within the lower and upper bounds, the test
is inconclusive (we don’t know whether there is positive autocorrelation).

4See Durbin and Watson (1951). This test is not directly applicable if the regression equation does not con-
tain a constant term.
 CHAPTER 8 Regression with Time Series Data 345

The Durbin-Watson test can be used to determine whether positive lag 1

autocorrelation is present.
If DW 7 dU, conclude H0: r = 0.
If DW 6 dL, conclude H1: r 7 0.
If DW lies within the lower and upper bounds ( dL  DW  dU), the test is
inconclusive.

The critical bounds for dL and dU are given in Table B-6. To find the appropriate dL
and dU, the analyst needs to know the sample size, level of significance, and number of
independent variables. In Table B-6, the sample size is given in the left-hand column, and
the number of independent variables is determined from the top of each column. If three
independent variables were used, for instance, one would look in the k = 3 column.5
As indicated in Equation 8.4, one can infer the sign and magnitude of the lag 1
residual autocorrelation coefficient from the DW statistic and vice versa. Thus, for situ-
ations in which the Durbin-Watson test is inconclusive, the significance of the serial
correlation can be investigated by comparing r11e2 with ; 2> 1n.6 If r11e2 falls in the
interval 0 ;2> 1n, conclude the autocorrelation is small and can be ignored.
Example 8.2
Suppose an analyst is engaged in forward planning for Reynolds Metals Company, an alu-
minum producer, and wishes to establish a quantitative basis for projecting future sales.
Since the company sells regionally, a measure of disposable personal income for the region
should be closely related to sales. Table 8-1 shows sales and income for the period from 1986
to 2006. Also shown in the table are the columns necessary to calculate the DW statistic (see
the Minitab Applications section at end of chapter). The residuals come from a least squares
line fit to the data, as shown in Figure 8-5.
Before using the least squares line for forecasting, the analyst performs the Durbin-
Watson test for positive serial correlation. The computations for the last three columns for
1987 are as follows:

et - et - 1 = - 47.53 - 1 -76.362 = 28.83

1et - et - 122 = 28.832 = 831.17
e 2t = 1-47.5322 = 2,259.1
The Durbin-Watson statistic is computed as
21

a 1et - et - 12
2
t=2 1,926,035.14
DW = 21
= = .87
2
2,210,641.78
a et
t=1

Using a .01 level of significance for a sample of n = 21 and k = 1 independent variable, one
obtains

dL = .97
dU = 1.16

5Itis also possible to test for negative autocorrelation. In this case, H1 : r 6 0, and the test statistic, DW, is
compared to 4 - dL and 4 - dU . The null hypothesis, H0 : r = 0 , is rejected if DW 7 4 - dL and not
rejected if DW 6 4-dU. The test is inconclusive for DW between 4 - dU and 4 - dL.
r11e2 is approximately 1> 2n (see the discussion of auto-
6If there is no autocorrelation, the standard error of
correlation in Chapter 3 and the discussion of residual autocorrelation in Chapter 6).
346 CHAPTER 8 Regression with Time Series Data

TABLE 8-1 Reynolds Metals Sales Data and the Durbin-Watson

Calculations for Example 8.2
Sales Income Residuals
Year Yt Xt et et – et - 1 1et - et - 122 e2t

1986 295 273.4 - 76.36 –– –– 5,830.85

1987 400 291.3 - 47.53 28.83 831.17 2,259.10
1988 390 306.9 -123.91 -76.38 5,833.90 15,353.69
1989 425 317.1 - 132.32 - 8.41 70.73 17,508.58
1990 547 336.1 - 91.16 41.16 1,694.15 8,310.15
1991 555 349.4 - 139.76 -48.60 2,361.96 19,532.86
1992 620 362.9 -132.20 7.56 57.15 17,476.84
1993 720 383.9 - 121.56 10.64 113.21 14,776.83
1994 880 402.8 - 41.98 79.58 6,332.98 1,762.32
1995 1,050 437.0 - 17.51 24.47 598.78 306.60
1996 1,290 472.2 72.71 90.22 8,139.65 5,286.74
1997 1,528 510.4 148.16 75.45 5,692.70 21,951.39
1998 1,586 544.5 61.06 -87.10 7,586.41 3,728.32
1999 1,960 588.1 249.53 188.47 35,520.94 62,265.22
2000 2,118 630.4 227.54 -21.99 483.56 51,774.45
2001 2,116 685.9 -10.62 -238.16 56,720.19 112.78
2002 2,477 742.8 108.26 118.88 14,132.45 11,720.23
2003 3,199 801.3 581.33 473.07 223,795.22 337,944.57
2004 3,702 903.1 651.16 69.83 4,876.23 424,009.35
2005 3,316 983.6 -77.38 - 728.54 530,770.53 5,987.66
2006 2,702 1,076.7 - 1,087.54 - 1,010.16 1,020,423.23 1,182,743.25
Totals 1,926,035.14 2,210,641.78

FIGURE 8-5 Regression Plot for Reynolds Metals Data for Example 8.2
 CHAPTER 8 Regression with Time Series Data 347
Since DW = .87 falls below dL = .97, the null hypothesis H0 : r = 0 is rejected, and it is
concluded that the errors are positively correlated 1r 7 02. The regression model should
be modified before it is used for forecasting.

SOLUTIONS TO AUTOCORRELATION PROBLEMS

After autocorrelation has been discovered in a regression of time series data, it is nec-
essary to remove it, or model it, before the regression function can be evaluated for its
effectiveness. The appropriate method for treating autocorrelation depends on what
caused it in the first place. Autocorrelation can arise because of a specification error
such as an omitted variable, or it can arise because the independent error terms are
correlated in an otherwise correctly specified model.
The solution to the problem of autocorrelation begins with an evaluation of the
model specification. Is the functional form correct? Were any important variables
omitted? Are there effects that might have some pattern over time that could have
introduced autocorrelation into the errors?
Because a major cause of autocorrelated errors in the regression model is the
omission of one or more key variables, the best approach to solving the problem is to
find them. This effort is sometimes referred to as improving the model specification.
Model specification involves not only finding the important predictor variables but
also entering these variables in the regression function in the right way. Unfortunately,
it is not always possible to improve the model specification because an important miss-
ing variable may not be quantifiable or, if it is quantifiable, the data may not be avail-
able. For example, one may suspect that business investment in future periods is
related to the attitude of potential investors. However, it is difficult to quantify the
variable “attitude.” Nevertheless, whenever possible, the model should be specified in
accordance with theoretically sound insight.
Only after the specification of the equation has been carefully reviewed should the
possibility of an adjustment be considered. Several techniques for eliminating autocor-
relation will be discussed.
One approach to eliminating autocorrelation is to add an omitted variable to the
regression function that explains the association in the response from one period to
the next.
Another approach to eliminating autocorrelation involves the general notion of
differencing. In this approach, the regression model is specified in terms of changes
rather than levels. Using the Reynolds Metals data, for example, the change in sales
from one year to the next (this change is 105 for the period 1986 to 1987) is related to
the corresponding change in income (this change is 17.9 for the period 1986 to 1987).
On occasion, the original variables may be expressed in terms of logarithms, and
changes in the logarithms are used in the regression. This procedure is equivalent to
regressing the percentage change in the response on the percentage changes in the
predictor variables. Finally, instead of using simple or first differences in the regres-
sion model, it may be the case that generalized differences are required to eliminate
autocorrelation.
The autoregressive model approach to eliminating autocorrelation generates pre-
dictor variables by using the response variable, Y, lagged one or more periods. In the
first-order autoregressive model case, the only predictor variable is the Y variable
lagged one time period. Using the Reynolds Metals data again, the value used to pre-
dict 1987 sales is the sales for 1986 (295).
The next examples illustrate these methods for eliminating autocorrelation.
348 CHAPTER 8 Regression with Time Series Data

Model Specification Error (Omitting a Variable)

Example 8.3 shows how incorporating a missing variable can eliminate serial correlation.
Example 8.3
The Novak Corporation wishes to develop a forecasting model for the projection of future
sales. Since the corporation has outlets throughout the region, disposable personal income
on a regionwide basis is chosen as a possible predictor variable. Table 8-2 shows Novak sales
for 1990 to 2006. The table also shows disposable personal income and the unemployment
for the region.
From Table 8-3, the Durbin-Watson statistic is .72, and using a significance level of .01
with n = 17 and k = 1, Table B-6 gives
dL = .87
dU = 1.10
Since DW = .72 6 dL = .87, positive serial correlation is indicated. A key variable that
accounts for the remaining association in sales from one year to the next may be missing
from the model. This result may be true even though the Minitab output says that dispos-
able income explains 99.5% of the variability in sales.
The unemployment rate may be an important missing predictor of sales. Table 8-4
shows the results of the regression analysis when the unemployment rate is added to the
model. The fitted model now explains 99.9% of the variability in sales. Although the inter-
cept is not significant, refitting the model without the intercept leaves the Durbin-Watson
statistic of 1.98 essentially unchanged.
With a .01 level of significance, n = 17, and k = 2, Table B-6 gives
dL = .77
dU = 1.25
Since DW = 1.98 7 dU = 1.25 , there is no evidence of first-order serial correlation.
Moreover, the lag 1 residual autocorrelation is r11e2 = .005, which is well within
; 2> 1n = ; 2> 117 = ; .49 of zero. Notice that DW = 1.98 ⬵ 211 - r11e22 =
211 - .0052 = 1.99, so the results of the Durbin-Watson test are consistent with no lag 1
residual autocorrelation, as expected.

TABLE 8-2 Novak Sales Data for Examples 8.3 and 8.6
Sales ($ millions) Income Unemployment
Row Year (Y) ($ millions) Rate Y-Lagged

1 1990 8.0 336.1 5.5 —

2 1991 8.2 349.4 5.5 8.0
3 1992 8.5 362.9 6.7 8.2
4 1993 9.2 383.9 5.5 8.5
5 1994 10.2 402.8 5.7 9.2
6 1995 11.4 437.0 5.2 10.2
7 1996 12.8 472.2 4.5 11.4
8 1997 13.6 510.4 3.8 12.8
9 1998 14.6 544.5 3.8 13.6
10 1999 16.4 588.1 3.6 14.6
11 2000 17.8 630.4 3.5 16.4
12 2001 18.6 685.9 4.9 17.8
13 2002 20.0 742.8 5.9 18.6
14 2003 21.9 801.3 5.6 20.0
15 2004 24.9 903.1 4.9 21.9
16 2005 27.3 983.6 5.6 24.9
17 2006 29.1 1076.7 8.5 27.3
 CHAPTER 8 Regression with Time Series Data 349

TABLE 8-3 Minitab Output for Novak Sales and Disposable Personal Income
for Example 8.3
Regression Analysis: Sales (Y) versus Income
The regression equation is
Sales 1Y2 = - 1.50 + 0.0292 Income

Predictor Coef SE Coef T P

Constant -1.5046 0.3290 -4.57 0.000
Income 0.0291916 0.0005129 56.92 0.000

S = 0.476669 R-Sq = 99.5% R-Sq1adj2 = 99.5%

Analysis of Variance
Source DF SS MS F P
Regression 1 736.15 736.15 3239.89 0.000
Residual Error 15 3.41 0.23
Total 16 739.56
Durbin-Watson statistic = 0.72

TABLE 8-4 Minitab Output for Novak Sales, Disposable Income,

and Unemployment Rate for Example 8.3

Regression Analysis: Sales (Y) versus Income, Rate

The regression equation is
Sales 1Y2 = - 0.014 + 0.0297 Income - 0.350 Rate
Predictor Coef SE Coef T P
Constant -0.0140 0.2498 -0.06 0.956
Income 0.0297492 0.0002480 119.96 0.000
Rate -0.34987 0.04656 -7.51 0.000
S = 0.219930 R-Sq = 99.9% R-Sq1adj2 = 99.9%
Analysis of Variance
Source DF SS MS F P
Regression 2 738.88 369.44 7637.91 0.000
Residual Error 14 0.68 0.05
Total 16 739.56
Durbin-Watson statistic = 1.98

The function YN = - .014 + .03X1 - .35X2 can be used to predict Novak sales with the
knowledge that the errors are independent.7 Expert estimates of disposable personal
income ($1,185 million) and the unemployment rate (7.8%) for the region are used to gen-
erate a forecast of Novak sales for 2007. The forecast is

YN = - .014 + .0311,1852 - .3517.82 = 32.7

or $32.7 million.

7The usual residual plots indicate that there is no reason to doubt any of the regression model assumptions.
350 CHAPTER 8 Regression with Time Series Data

Regression with Differences

For highly autocorrelated data, modeling changes rather than levels can often elimi-
nate the serial correlation. That is, instead of formulating the regression equation in
terms of Y and X1, X2, Á , Xk, the regression equation is written in terms of the
differences, Yt¿ = Yt - Yt - 1 and X t1¿ = Xt1 - Xt - 1,1, X t2¿ = Xt2 - Xt - 1,2, and so forth.
Differences should be considered when the Durbin-Watson statistic associated with
the regression involving the original variables is close to zero.8
One rationale for differencing comes from the following argument. Suppose
Equations 8.1 and 8.2 hold, so that

Yt = b 0 + b 1Xt + t

with

t = r t - 1 + t

where
r = the correlation between consecutive errors
t = the random error
t = t when r = 0
The model holds for any time period, so

Yt - 1 = b 0 + b 1Xt - 1 + t - 1

Multiplying this equation on both sides by ρ and then subtracting equals from equals in
Equation 8.1 gives

Yt = b 0 + b 1Xt + t 1Equation 8.12

rYt - 1 = rb 0 + rb 1Xt - 1 + r t - 1 1multiply Yt - 1 by r2
Yt - rYt - 1 = b 0 - rb 0 + 1b 1Xt - rb 1Xt - 12 + 1 t - r t - 12 1subtract2

or

Yt¿ = b 011 - r2 + b 1X t¿ + t (8.5)

where the “prime” indicates the generalized differences

Yt¿ = Yt - rYt - 1
X t¿ = Xt - rXt - 1 (8.6)

The model in Equation 8.5 has errors, t, that are independently distributed with the
mean equal to zero and a constant variance. Thus, the usual regression methods can be
applied to this model.

8An autocorrelation pattern for the Y variable or the X variables such as the one pictured in Figure 8-3 also
indicates that a regression function with differences may eliminate (or greatly reduce) problems caused by
serial correlation.
 CHAPTER 8 Regression with Time Series Data 351
If the correlation between consecutive errors is strong (ρ is near 1), then the gen-
eralized differences are essentially simple or first differences:

Yt¿ = Yt - Yt - 1
X t¿ = Xt - Xt - 1 (8.7)

and the intercept term in the model (Equation 8.5) is near zero (it disappears).
Using regression models constructed with generalized differences can frequently
eliminate serial correlation. If the serial correlation is especially strong, then simple dif-
ferences can be used. Example 8.4 shows what can happen if strong autocorrelation is
ignored.

Example 8.4
Some years ago, Fred Gardner was engaged in forecasting Sears Roebuck sales in thousands
of dollars for the western region. He had chosen disposable personal income for the region
as his independent variable. Fred wanted to relate sales to disposable income using a log
linear regression model, since that would allow him to also estimate the income elasticity of
sales. The elasticity measures the percentage change in sales for a 1% change in income.
The log linear regression model assumes that income is related to sales by the equation

Sales = g1Income2b1

Taking the natural logarithms (Ln) of both sides of the foregoing equation gives

Ln1Sales2 = Ln g + b 1 Ln1Income2

Adding an error term to account for the influence of variables other than income on sales,
the previous expression becomes a log linear regression model of the form

Ln Yt = b 0 + b 1 Ln Xt + t (8.8)

where
Ln Yt = Ln1Sales2 = the natural logarithm of sales
Ln Xt = Ln1Income2 = the natural logarithm of income
t = the error term
b 0 = Lng = the intercept coefficient
b 1 = the slope coefficient = the income elasticity of sales
Table 8-5 shows Sears sales, disposable income, their logarithms, and the differences in the
logarithms of sales and disposable income for the 1976–1996 period.
A portion of the Minitab output showing the regression of Ln(Sales) on Ln(Income) is
shown in Table 8-6. Fred noticed that 99.2% of the variability in the logarithm of Sears sales for
the western region can be explained by its relationship with the logarithm of disposable income
for the same region. The regression is highly significant. Also, the income elasticity is estimated
to be b1 = 1.117, with a standard error of sb1 = .023. However, the Durbin-Watson statistic of
.50 is small and less than dL = .97, the lower .01 level critical value for n = 21 and k = 1. Fred
concluded that the correlation between successive errors is positive and large (close to 1).
Because of the large serial correlation, Fred decided to model the changes or differ-
ences in the logarithms of sales and income, respectively. He knew that the slope coefficient
in the model for the differences is the same slope coefficient as the one in the original model
involving the logarithms. Therefore, he could still estimate the income elasticity directly.
The intercept coefficient in the regression model for the differences is likely to be small and
was omitted. The Minitab results for the changes are shown in Table 8-7.
Table 8-7 shows that the regression is significant. The income elasticity is estimated to be
b1 = 1.010, with a standard error of sb1 = .093. The elasticity estimate, b1, did not change too
much from the first regression (a 1% increase in disposable income leads to an approximate
352 CHAPTER 8 Regression with Time Series Data

TABLE 8-5 Sears Sales and U.S. Disposable Income,

1976–1996, Along with Transformed
Data, for Example 8.4

Sales Income
($1,000s) ($ millions) Differences
Year Yt Xt Ln Yt Ln Xt Yt Xt
1976 3,307 273.4 8.1038 5.6109 — —
1977 3,556 291.3 8.1764 5.6744 .0726 .0634
1978 3,601 306.9 8.1890 5.7265 .0126 .0522
1979 3,721 317.1 8.2218 5.7592 .0328 .0327
1980 4,036 336.1 8.3030 5.8174 .0813 .0582
1981 4,134 349.4 8.3270 5.8562 .0240 .0388
1982 4,268 362.9 8.3589 5.8941 .0319 .0379
1983 4,578 383.9 8.4290 5.9504 .0701 .0563
1984 5,093 402.8 8.5356 5.9984 .1066 .0481
1985 5,716 437.0 8.6510 6.0799 .1154 .0815
1986 6,357 472.2 8.7573 6.1574 .1063 .0775
1987 6,769 510.4 8.8201 6.2352 .0628 .0778
1988 7,296 544.5 8.8951 6.2999 .0750 .0647
1989 8,178 588.1 9.0092 6.3769 .1141 .0770
1990 8,844 630.4 9.0875 6.4464 .0783 .0695
1991 9,251 685.9 9.1325 6.5307 .0450 .0844
1992 10,006 742.8 9.2109 6.6104 .0785 .0797
1993 11,200 801.3 9.3237 6.6862 .1127 .0758
1994 12,500 903.1 9.4335 6.8058 .1098 .1196
1995 13,101 983.6 9.4804 6.8912 .0470 .0854
1996 13,640 1,076.7 9.5208 6.9817 .0403 .0904

TABLE 8-6 Minitab Output for the Regression of the

Logarithms of Sears Sales on the Logarithms
of Disposable Income for Example 8.4
Regression Analysis: Ln(Sales) versus Ln(Income)
The regression equation is
Ln1Sales2 = 1.82 + 1.12 Ln1Income2
Predictor Coef SE Coef T P
Constant 1.8232 0.1434 12.71 0.000
Ln(Income) 1.11727 0.02305 48.47 0.000
S = 0.0436806 R-Sq = 99.2% R-Sq1adj2 = 99.2%
Analysis of Variance
Source DF SS MS F P
Regression 1 4.4821 4.4821 2349.13 0.000
Residual Error 19 0.0363 0.0019
Total 20 4.5184
Durbin-Watson statistic = 0.50
 CHAPTER 8 Regression with Time Series Data 353

TABLE 8-7 Minitab Output for the Regression of the Changes in Logarithms of
Sears Sales on the Changes in Logarithms of Disposable Income for
Example 8.4
Regression Analysis: Change in Ln(Sales) versus Change in Ln(Income)
The regression equation is
Change in Ln1Sales2 = 1.01 Change in Ln1Income2
Predictor Coef SE Coef T P
Noconstant
Change in Ln(Income) 1.00989 0.09304 10.85 0.000
S = 0.0297487
Analysis of Variance
Source DF SS MS F P
Regression 1 0.10428 0.10428 117.83 0.000
Residual Error 19 0.01681 0.00088
Total 20 0.12109
Durbin-Watson statistic = 1.28

FIGURE 8-6 The Residual Autocorrelations for the Regression Analysis

in Table 8-7 for Example 8.4

1% increase in annual sales in both cases), but its current standard error 1sb1 = .0932 is about
four times as large as the previous standard error 1sb1 = .0232. The previous standard error is
likely to understate the true standard error due to the serial correlation.
Checking the Durbin-Watson statistic for n = 20, k = 1, and a significance level of
.05, Fred found that dL = 1.20 6 DW = 1.28 6 dU = 1.41, so the test for positive serial
correlation is inconclusive. However, a check of the residual autocorrelations, shown in
Figure 8-6, indicates that they are all well within their two standard error limits (the dashed
lines in the figure) for the first few lags. Fred concluded that serial correlation had been
eliminated, and he used the fitted equation for forecasting.
To use the final model for forecasting, Fred wrote

YN t¿ = Ln YNt - Ln YNt - 1
YN t¿ = b1X t¿ where
X t¿ = Ln Xt - Ln Xt - 1
354 CHAPTER 8 Regression with Time Series Data

Substituting for YN t¿ and X t¿ and rearranging terms,

Ln YN t = Ln YN t - 1 + b11Ln Xt - Ln Xt - 12 (8.9)

The forecast for Sears sales in 1997 was obtained by setting b1 = 1.01 and t = 22:

Ln YN 22 = Ln YN 21 + 1.011Ln X22 - Ln X212

Sales in 1996 were known, so YN21 = Y21 = 13,640. Disposable income for 1996 was known,
so X21 = 1,076.7 . To continue, Fred needed disposable income for 1997. An economist
familiar with the western region sent Fred an estimate of $1,185 million for 1997 disposable
income. Fred used this expert’s estimate and set X22 = 1,185. The forecasting equation
became

LnYN22 = Ln113,6402 + 1.011Ln11,1852 - Ln11,076.722

= 9.5208 + 1.0117.0775 - 6.98172 = 9.6176

or, taking antilogs,

YN22 = e 9.6176 = 15,027

Fred’s forecast of Sears 1997 sales for the western region was $15,027 thousands. Fred could
use Equation 8.9 and the procedure described earlier to generate forecasts for the years
1998, 1999, and so forth, but to do so, he would need estimates of disposable personal
incomes for these years.

Autocorrelated Errors and Generalized Differences

The objective is to adequately describe the nature of the relationship between the
variables Y and X when serial correlation is present.
Consider again the regression model with serially correlated errors (see Equations 8.1
and 8.2):

Yt = b 0 + b 1Xt + t
t = r t - 1 + t

The errors, t, are said to follow a first-order autoregressive or AR(1) model.9

Recall that, after some algebraic manipulation, the system of equations above can
be written as a simple linear regression model involving the generalized differences
Yt¿ = Yt - rYt - 1 and X t¿ = Xt - rXt - 1 (see Equation 8.5):

Yt¿ = b 011 - r2 + b 1X t¿ + t

The regression model involving the generalized differences is specifically constructed to

eliminate serial correlation in the errors. The errors, t, are assumed to be independent
and normally distributed with mean zero and constant variance, s2. Consequently, it
seems reasonable to fit this model directly, using the principle of least squares. However,
the autocorrelation coefficient, ρ, is unknown, so Yt¿ and X t¿ cannot be determined.
Therefore, the model cannot be fit using ordinary least squares.
There are two general approaches to estimating b 0 and b 1, the parameters of
primary interest. One approach is to use the model corrected for serial correlation

9First-orderautoregressive models are formally introduced in the next section of this chapter and discussed
again in Chapter 9.
 CHAPTER 8 Regression with Time Series Data 355

FIGURE 8-7 E-Views Output for the Regression of Sears Sales

on Disposable Income for Example 8.5

(Equation 8.5) and estimate the parameters b 0, b 1, and ρ directly, using a numerical
technique called nonlinear least squares. This approach uses a search routine to find the
parameter values that minimize the error sum of squares, ©2t . The other approach is to
estimate ρ; use the estimate, rN, to construct the generalized differences; and then fit the
model involving the generalized differences using ordinary least squares.10
The next example illustrates the nonlinear least squares approach using output
from E-Views, a popular software package for econometric modeling.
Example 8.5
The Sears data are given in Table 8-5. Ordinary least squares is used to fit a simple linear
regression model relating sales to disposable personal income. The E-Views output is shown
in Figure 8-7. From Figure 8-7, the fitted regression equation is

YNt = - 524.33 + 14.05Xt

where

b1 = 14.05
sb1 = .319
t = b1>sb1 = 44.11
r 2 = .99
DW = .63

10 Techniques for estimating ρ and accounting for serial correlation are discussed by Pindyck and Rubinfeld
(1998).
356 CHAPTER 8 Regression with Time Series Data

FIGURE 8-8 E-Views Output for the Regression of Generalized

Differences of Sears Sales on Generalized Differences of
Disposable Income for Example 8.5

Since the DW statistic is close to zero, positive autocorrelation is indicated.11 At this

point, we allow for serial correlated (AR(1)) errors and fit the model involving the general-
ized differences:

Yt¿ = b 011 - r2 + b 1X t¿ + t
with Yt¿ = Yt - rYt - 1 and X t¿ = Xt - rXt - 1. E-Views is used to estimate the parameters in
this model directly. The E-Views output is shown in Figure 8-8.
The fitted regression function is

YN t¿ = 54,48311 - .9972 + 9.26X t¿

where
rN = .997
N
Yt¿ = Yt - .997Yt - 1, X t¿ = Xt - .997Xt - 1
b0 = 54,483, b1 = 9.26
sb1 = 7.241
t = b1>sb1 = 1.28
DW = 1.12
The value of the Durbin-Watson statistic, 1.12, is in the “inconclusive” region at the .01
level. Note that the estimates of the slope coefficient in the two regressions are similar

11For n = 21, k = 1, and a = .01, dL = .97. Since DW = .63 6 dL = .97, we reject H0: r = 0 in favor of
H1: r 7 0.
 CHAPTER 8 Regression with Time Series Data 357
(14.05 and 9.26). However, the standard error associated with b1 in the second regression is
considerably larger than the corresponding standard error in the first regression (7.241
versus .319). Thus, the t statistic for testing the significance of the slope coefficient in the
second regression is much smaller than it is in the first regression (1.28 versus 44.11). In fact,
the p-value associated with the t statistic in the second regression is .218, indicating the slope
coefficient is not significantly different from zero. The strong serial correlation has little
effect on the estimate of the slope coefficient in the relation between Y and X. However, the
strong (positive) serial correlation does result in severe underestimation of the standard
error of the estimated slope coefficient. Indeed, one reason for adjusting for serial correla-
tion is to avoid making mistakes of inference because of t values that are too large.
Finally, rN = .997 is very close to 1. This suggests that the relationship between Y and X
might be represented by a simple linear regression model with the differences
Yt¿ = Yt - Yt - 1 and X t¿ = Xt - Xt - 1. This issue is explored in Problem 17.

Autoregressive Models
Autocorrelation implies that values of the dependent variable in one time period are
linearly related to values of the dependent variable in another time period. Thus, one
way to solve the problem of serial correlation is to model the association in different
time periods directly. This can be done in a regression framework, using the dependent
variable lagged one or more time periods as the predictor or independent variable.
Regression models formulated in this way are called autoregressive models. The first-
order autoregressive model is written

Yt = b 0 + b 1Yt - 1 + t (8.10)

where the errors, t, are assumed to have the usual regression model properties. Once
this model has been fit to the data by least squares, the forecasting equation becomes

YNt = b0 + b1Yt - 1 (8.11)

An autoregressive model expresses a forecast as a function of previous values of

the time series.

Autoregressive models are a subset of the autoregressive integrated moving aver-

age (ARIMA) models, which are discussed in more detail in Chapter 9.
Example 8.6
The Novak Corporation sales data introduced in Example 8.3 and shown in Table 8-2 will be
used to demonstrate the development of an autoregressive model. Note that, after sales
were lagged as shown in Table 8-2, one year of data is lost, since Novak sales for 1989 are not
known. The sample size is n = 16 instead of n = 17. A first-order autoregressive model is
developed with Novak sales lagged one year (Y-lagged) as the predictor variable.
The Minitab output that results from fitting the autoregressive model to the Novak
sales data is shown in Table 8-8. A forecast for 2007 sales and the 95% prediction interval
are also shown in the table.
The forecast of Novak sales for 2007 is $31.722 million. This figure is computed, within
rounding error, as follows:

YN = b0 + b1Yt - 1 = - .109 + 1.094Yt - 1

N
Y18 = - .109 + 1.094Y18 - 1 = - .109 + 1.094Y17
YN18 = - .109 + 1.094129.12 = 31.726
358 CHAPTER 8 Regression with Time Series Data

TABLE 8-8 Minitab Output for Novak Sales: Autoregressive

Model Fit for Example 8.6
Regression Analysis: Sales (Y) versus Y-Lagged
The regression equation is
Sales1Y2 = - 0.109 + 1.09 Y - Lagged
Predictor Coef SE Coef T P
Constant -0.1093 0.3367 -0.32 0.750
Y-Lagged 1.09388 0.02063 53.01 0.000
S = 0.487455 R-Sq = 99.5% R-Sq1adj2 = 99.5%
Analysis of Variance
Source DF SS MS F P
Regression 1 667.73 667.73 2810.16 0.000
Residual Error 14 3.33 0.24
Total 15 671.05
Predicted Values for New Observations
New Obs Fit SE Fit 95% CI 95% PI
1 31.722 0.311 (31.055, 32.390) (30.482, 32.963)

The Durbin-Watson test cannot be used in this example. When a lagged dependent vari-
able is included in the regression as a predictor variable, the Durbin-Watson statistic is biased
toward 2. Instead, a test for serial correlation can be based on the Durbin-Watson h statistic.12
The intercept coefficient in this regression is small and not significantly different from
zero. Rerunning the regression without an intercept term leaves the estimate of the slope
coefficient, and subsequent forecasts, essentially unchanged. The fact that the estimated slope
coefficient is about 1 suggests that a forecast of next year’s sales is very nearly this year’s sales.

Summary
When regression analysis is applied to time series data, the residuals are frequently
autocorrelated. Since regression analysis assumes that the errors are independent,
problems can arise. The R2 for a regression with data containing autocorrelation can be
artificially high. Furthermore, the standard errors of the regression coefficients can
be seriously underestimated and the corresponding t statistics inflated.
One cause of autocorrelated residuals is the omission of one or more key predictor
variables. This omission usually means that an important part of the dependent vari-
able variation has not been adequately explained. One solution to this problem is to
search for the missing variable(s) to include in the model. Other solutions to the
problems caused by autocorrelation are to consider either regression models with
differenced data or autoregressive models.

TIME SERIES DATA AND THE PROBLEM OF HETEROSCEDASTICITY

The variability in some time series tends to increase with the level of the series.
The Cavanaugh Company sales time series shown in Figure 5-1 (bottom) and discussed
in Example 5.2 is one example. Variability can increase if a variable is growing at a

12The h test for serial correlation is described by Pindyck and Rubinfeld (1998).
 CHAPTER 8 Regression with Time Series Data 359
constant rate rather than a constant amount over time. Nonconstant variability is
called heteroscedasticity.
In a regression framework, heteroscedasticity occurs if the variance of the error
term, ε, is not constant. If the variability for recent time periods is larger than it was for
past time periods, then the standard error of the estimate, sy #x’s, underestimates the cur-
rent standard deviation of the error term. If the standard deviation of the estimate is
then used to set forecast limits for future observations, these limits can be too narrow
for the stated confidence level.
Sometimes the problem of heteroscedasticity can be solved by simple transforma-
tions of the data. For example, in the case of two variables, the log linear model shown
in Equation 8.8 might be used to reduce the heteroscedasticity. Also, if the variables
are expressed as dollar amounts, converting current dollars to constant dollars (see the
discussion of price deflation in Chapter 5) may overcome the problem of increasing
error variability.
Example 8.7
Consider again the Reynolds Metals sales data introduced in Example 8.2 and shown in
Table 8-1. The result of a simple linear regression of sales on disposable personal income is
given in Figure 8-5. A time sequence plot of the residuals from this regression is pictured in
Figure 8-9.
In addition to the positive autocorrelation in the residuals (string of negative residuals
followed by string of positive residuals; see the discussion in Example 8.2), it is clear from
Figure 8-9 that the size of the residuals is increasing over time. One approach to this prob-
lem is to try a log linear model (Equation 8.8) for the Reynolds Metals data.
The results of a log linear model fit are given in Figure 8-10. Comparing Figure 8-10
with Figure 8-5, it can be seen that the residuals (deviations from the fitted line) for the log
linear regression are more uniform in size throughout the time period under study, but the

FIGURE 8-9 Time Sequence Plot of Residuals from Regression of Reynolds

Metals Sales Data on Disposable Income for Example 8.7
360 CHAPTER 8 Regression with Time Series Data

FIGURE 8-10 Regression Plot of Log Linear Model Fit to Reynolds Metals
Data for Example 8.7

FIGURE 8-11 Time Sequence Plot of Residuals from Fitting Log Linear Model
with Quadratic Term to Reynolds Metals Data for Example 8.7
 CHAPTER 8 Regression with Time Series Data 361
fitted straight line does not capture the curvature in the data. An additional predictor vari-
able, X2 = X 21 = 1Ln Income22, was added, and the model13

Ln Sales = b 0 + b 1 Ln Income + b 2 1Ln Income22 +

was fit to the data. A time sequence plot of the residuals from this regression is displayed in
Figure 8-11.
The residuals in Figure 8-11 appear to be randomly distributed about zero with con-
stant variability. It appears as if the final regression adequately represents the Reynolds
Metals data. For this model, there is no reason to doubt the error term assumptions.

USING REGRESSION TO FORECAST SEASONAL DATA

Decomposition models for time series with seasonal patterns were discussed in
Chapter 5. A regression model for representing seasonal data that is closely aligned
with an additive decomposition is given next. In this model, the seasonality is handled
by using dummy variables in the regression function.
A seasonal model for quarterly data with a time trend is
Yt = b 0 + b 1t + b 2S2 + b 3S3 + b 4S4 + t (8.12)
where
Yt = the variable to be forecast
t = the time index
S2 = a dummy variable that is 1 for the second quarter of the year;
0 otherwise
S3 = a dummy variable that is 1 for the third quarter of the year;
0 otherwise
S4 = a dummy variable that is 1 for the fourth quarter of the year;
0 otherwise
t = the errors, assumed to be independent and normally distributed
with mean zero and a constant variance
b 0, b 1, b 2, b 3, b 4 = the coefficients to be estimated
Note that the four levels of the qualitative variable (quarter or season) are
described with only three dummy variables. The final quarter (the first quarter in this
case) is handled by the intercept term, b 0. To see this, for first-quarter data,
S2 = S3 = S4 = 0, and the expected level is
E1Yt2 = b 0 + b 1t
For second-quarter data, S2 = 1, S3 = S4 = 0, and the expected level is
E1Yt2 = b 0 + b 1t + b 2 = 1b 0 + b 22 + b 1t
Similar patterns emerge for the third and fourth quarters.
In Equation 8.12, different quarters give different intercept terms. The trend and
the seasonal pattern are modeled as a series of straight lines, one for each quarter. The
four lines are postulated to have the same slope but different intercepts. The regression
coefficients for the dummy variables represent changes in the intercept relative to the

13A regression model with predictor variables X, X 2, X 3, . . . is called a polynomial regression model.
362 CHAPTER 8 Regression with Time Series Data

intercept for the first quarter, b 0 . If there is a seasonal pattern but no trend, then
Equation 8.12 applies, with b 1 = 0.
Example 8.8
At one time, James Brown, forecaster for the Washington Water Power Company, wanted to
forecast electrical usage for residential customers for the third and fourth quarters of 1996.
He knew the data were seasonal and decided to use Equation 8.12 to develop a forecasting
equation. He used quarterly data from 1980 through the second quarter of 1996. The data
for electrical usage measured in millions of kilowatt-hours are given in Table 8-9.
James created dummy variables S2, S3 , and S4 , representing the second, third, and
fourth quarters, respectively. The data for the four quarters of 1980 are given in Table 8-10.

TABLE 8-9 Electrical Usage Data for Washington Water

Power, 1980–1996, for Example 8.8
Kilowatt-Hours Kilowatt-Hours
Year Quarter (millions) Year Quarter (millions)

1980 1 1,071 1989 1 1,036

2 648 2 612
3 480 3 503
4 746 4 710
1981 1 965 1990 1 952
2 661 2 628
3 501 3 534
4 768 4 733
1982 1 1,065 1991 1 1,085
2 667 2 692
3 486 3 568
4 780 4 783
1983 1 926 1992 1 928
2 618 2 655
3 483 3 590
4 757 4 814
1984 1 1,047 1993 1 1,018
2 667 2 670
3 495 3 566
4 794 4 811
1985 1 1,068 1994 1 962
2 625 2 647
3 499 3 630
4 850 4 803
1986 1 975 1995 1 1,002
2 623 2 887
3 496 3 615
4 728 4 828
1987 1 933 1996 1 1,003
2 582 2 706
3 490
4 708
1988 1 953
2 604
3 508
4 708

Source: Based on Washington Water Power Annual Report, various years.

 CHAPTER 8 Regression with Time Series Data 363

TABLE 8-10 Electrical Usage Data for

Washington Water Power
for Four Quarters of 1980
for Example 8.8
Yt S2 S3 S4

1,071 0 0 0
648 1 0 0
480 0 1 0
746 0 0 1

TABLE 8-11 Computer Output for Washington Water Power for Example 8.8
Regression Analysis: Hours versus Time, 2nd Qt., 3rd Qt., 4th Qt.
The regression equation is
Hours = 968 + 0.938 Time - 342 2nd Qt. - 472 3rd Qt. - 230 4th Qt.
Predictor Coef SE Coef T P
Constant 968.39 16.88 57.38 0.000
Time 0.9383 0.3377 2.78 0.007
2nd Qt. -341.94 17.92 -19.08 0.000
3rd Qt. -471.60 18.20 -25.91 0.000
4th Qt. -230.23 18.20 -12.65 0.000
S = 52.2488 R-Sq = 92.4% R-Sq1adj2 = 91.9%
Analysis of Variance
Source DF SS MS F P
Regression 4 2012975 503244 184.34 0.000
Residual Error 61 166526 2730
Total 65 2179502
Durbin-Watson statistic = 1.48
Predicted Values for New Observations
New Obs Fit SE Fit 95% CI 95% PI
1 559.65 17.39 (524.87, 594.43) (449.54, 669.76)

Values of Predictors for New Observations

New Obs Time 2nd Qt. 3rd Qt. 4th Qt.
1 67.0 0 1 0

The Minitab commands to run the seasonal analysis are shown in the Minitab
Applications section at the end of the chapter. The results are shown in Table 8-11.
The fitted seasonal regression model is

YNt = 968 + .938t - 342S2 - 472S3 - 230S4

where S2 = 2nd Qt . S3 = 3rd Qt . and S4 = 4th Qt. James noted that the model explains
92.4% of the variability in electrical usage for residential customers measured in kilowatt-
hours. The forecast for the third quarter of 1996 was computed as
Hours = 968 + .9381672 - 342102 - 472112 - 230102 = 559
The forecast for the fourth quarter of 1996 was
Hours = 968 + .9381682 - 342102 - 472102 - 230112 = 802
364 CHAPTER 8 Regression with Time Series Data

James saw that the forecasts for the different quarters would lie along four straight
lines. The lines all have the same slope (.938), but the intercepts change depending on
the quarter. The first-quarter forecasts lie along a line with intercept 968. The second-
quarter forecasts lie along a line with intercept 968 - 342 = 626. The intercept for third-
quarter forecasts is 968 - 472 = 496 , and the intercept for fourth-quarter forecasts is
968 - 230 = 738 . James was pleased that the forecasting model captured the seasonal
pattern and slight upward trend observed in the series. Within a given year, forecasts of
electrical usage are highest for the first quarter, lower for the second quarter, lowest for the
third quarter, and second highest for the fourth quarter.

ECONOMETRIC FORECASTING
When regression analysis is applied to economic data, the predictions developed from
such models are referred to as economic forecasts. However, since economic theory fre-
quently suggests that the values taken by the quantities of interest are determined
through the simultaneous interaction of different economic forces, it may be necessary to
model this interaction with a set of simultaneous equations. This idea leads to the con-
struction of simultaneous equation econometric models. These models involve individual
equations that look like regression equations. However, in a simultaneous system, the
individual equations are related, and the econometric model allows the joint determina-
tion of a set of dependent variables in terms of several independent variables. This con-
trasts with the usual regression situation, in which a single equation determines the
expected value of one dependent variable in terms of the independent variables.
A simultaneous equation econometric model determines jointly the values of a set
of dependent variables, called endogenous variables by econometricians, in terms of the
values of the independent variables, called exogenous variables.The values of the exoge-
nous variables are assumed to influence the endogenous variables but not the other way
around. A complete simultaneous equation model will involve the same number of
equations as endogenous variables.
Simultaneity in the econometric system creates some problems that require special
statistical treatment. A full treatment of econometric models is beyond the scope of
this book.14 However, a two-equation model will illustrate some of the concepts.
Economic theory holds that, in equilibrium, the quantity supplied is equal to the
quantity demanded at a particular price. That is, the quantity demanded, the quantity
supplied, and the price are determined simultaneously. In one study of the price elastic-
ity of demand, the model was specified as
Qt = a0 + a1Pt + a2It + a3Tt + t
Pt = b 0 + b 1Qt + b 2Lt + t
where
Qt = a measure of the demand 1quantity sold2
Pt = a measure of price 1deflated dollars2
It = a measure of income per capita
Tt = a measure of temperature
Lt = a measure of labor cost
t, t = the independent error terms, which are uncorrelated with each other

14Pindyck and Rubinfeld (1998) provide an introductory account of simultaneous equation econometric
models.
 CHAPTER 8 Regression with Time Series Data 365
Notice in this model that the price and quantity variables, Pt and Qt, appear in both
equations. In the first equation, quantity sold is partially determined by price, and in the
second equation, price is partially determined by quantity sold. Price and quantity are
endogenous variables whose values are determined within the system. The remaining
variables, income and temperature in the first equation and labor cost in the second
equation, are exogenous variables, whose values are determined outside the system.
Given adequate estimates for the coefficients in the model (the identification problem),
forecasts of, say, future demand (sales) can be generated. Of course, to estimate future
demand, future values of the exogenous variables must be specified or estimated from
outside the system. In addition, future values of the price variable must be determined.
Large-scale econometric models are being used today to model the behavior of
specific firms within an industry, selected industries within the economy, and the total
economy. Econometric models can include any number of simultaneous multiple
regression–like equations. Econometric models are used to understand how the econ-
omy works and to generate forecasts of key economic variables. Econometric models
are important aids in policy formulation.

COINTEGRATED TIME SERIES

In Example 8.1, we pointed out that a regression analysis involving two highly autocor-
related nonstationary (no natural mean level) time series can lead to spurious regres-
sions. To guard against spurious regressions, regressions among nonstationary variables
should be done as regressions among changes in the variables (see Granger and
Newbold 1974).15 If the original regression is spurious, there should be no relationship
among the changes. However, there are special situations in which two nonstationary
variables are related in the sense they tend to cling to one another, and a regression
that involves only changes in the variables misses the mechanism that connects the
clinging variables.
To illustrate clinging variables, both inventory level and unit sales of a product can
be nonstationary (due to growth, say), yet the difference between sales and inventory
level can be stationary (vary about a fixed level), a natural by-product of schemes that
adjust inventories to uncertain sales. That is, because of the continual adjustments, the
difference between inventory and sales can bounce around from one period to the next
but tends to be relatively constant. Sales and inventory cling to one another.
To proceed more formally, let Yt and Xt be two nonstationary time series, each of
which can be converted to a stationary time series by simple differencing. Nonstationary
series that become stationary when differenced once are said to be integrated of order 1.16
To make things simple, suppose the stationary differences for each series are random
errors. Then we can write

Yt - Yt - 1 = t

and

Xt - Xt - 1 = t

15Also,a spurious regression can typically be detected with a careful examination of the residuals and the
residual autocorrelations from the regression involving the original variables.
16Nonstationary series that become stationary when differenced n times are said to be integrated of order n.
366 CHAPTER 8 Regression with Time Series Data

where t and t are independent, normally distributed error terms, each with mean zero
and variances s2 and s2, respectively.17 As they stand, Yt and Xt are not connected. But
as in the sales inventory example earlier, suppose the difference Yt - Xt is stationary of
the following form:

Yt - Xt = ht
where ht is independently and normally distributed error with mean zero and variance
s2h. The linear combination Yt - Xt that links Yt and Xt is called a cointegrating relation-
ship of order 0.18 In general, there is no way to form a linear combination (weighted
average) of two nonstationary series, Yt and Xt, to produce a stationary series. However,
in the special case where such a relationship does exist, we say Yt and Xt are cointe-
grated, with the cointegrating relationship given by the linear combination.19

A set of nonstationary time series for which simple differencing produces a sta-
tionary series in each case is said to be cointegrated if and only if some linear
combination of the series is stationary. The stationary linear combination of
cointegrated time series is called the cointegrating relationship.

Cointegration is a probabilistic concept. In our example, the difference between

sales and inventory is not fixed but fluctuates as a stationary random variable about
some level.
Now the deviation h = Y - X at any time may influence the future evolution of
the variables Y and X. For example, too little inventory relative to sales may lead to
future shortages and reduced sales and may also result in an increase in future inven-
tory. We can accommodate the notion of recent linkage behavior influencing future
evolution by modeling Yt - Yt - 1 and Xt - Xt - 1 as a function of lagged values of
ht = Yt - Xt, the error correction term.

Yt - Yt - 1 = b 0 + b 11Yt - 1 - Xt - 12 + t (8.13a)

Xt - Xt - 1 = b 0¿ + b 1¿ 1Yt - 1 - Xt - 12 + t (8.13b)

where b 0, b 0¿ , b 1, and b 1¿ are parameters to be estimated.

The models in Equations 8.13a and 8.13b are the simplest of the models for coin-
tegrated time series and are meant to illustrate the concepts. In general, identifying,
estimating, and checking models for cointegrated variables pose special challenges.
Moreover, forecasts of cointegrated series and measures of forecast accuracy should
preserve the cointegrating relationships. If, in attempting to model the relationship
between two or more cointegrated time series, one misses the error correction mech-
anism that connects cointegrated variables, spurious effects may be observed.
Cointegration and the corresponding error correction models are of interest to
forecasters in business, finance, economics, and government because linkages among

17The Yt and Xt series in this case are said to be random walks—a change in position in each case is equally
likely to be positive or negative. The two time series pictured in Figure 8-2 are random walks.
18The cointegrating relationship is of order 0 because it involves the original variables, not the differenced
variables.
19Cointegration does not require the difference in the variables to be stationary. Two nonstationary variables
are cointegrated if there exists any linear combination of the variables that is stationary.
 CHAPTER 8 Regression with Time Series Data 367
inherently nonstationary variables are often created by management action, by gov-
ernment or regulating body policy, and as the natural by-products of long-run theoret-
ical relationships.
The reader interested in learning more about cointegrated time series is referred
to Murray (1994), Diebold (2004), and other references at the end of this chapter.

APPLICATION TO MANAGEMENT
The applications described in Chapter 5 are also appropriate for this chapter. The tech-
niques described in this chapter permit the analyst to detect and correct for the prob-
lem of serial correlation and thus to develop better forecasting models. The net result is
that management and/or economists can deal with a far greater variety of time-
dependent data and feel confident that the predictions are sound. Areas in which these
techniques are particularly helpful include the following:

Sales forecasting
Stock and bond price projections
Raw materials cost projections
New-product penetration projections
Personnel needs estimates
Advertising–sales relationship studies
Inventory control

Because these applications involve variables that evolve over time, the variables are
likely to be autocorrelated. Forecasting models based on the techniques presented in
this chapter should provide more reliable forecasts than some of the techniques con-
sidered earlier that ignore autocorrelation.

Glossary
Autocorrelation (serial correlation). Autocorrelation cointegrated if and only if some linear combina-
exists when successive observations over time are tion of the series is stationary.
related to one another. Cointegrating relationship. The stationary linear
Autoregressive model. An autoregressive model combination of cointegrated time series is called
expresses a forecast as a function of previous the cointegrating relationship.
values of the time series. Durbin-Watson test. The Durbin-Watson test can
Cointegrated time series. A set of nonstationary be used to determine whether positive lag 1 auto-
time series for which simple differencing pro- correlation is present.
duces a stationary series in each case is said to be

Key Formulas

Simple linear regression model

Yt = b 0 + b 1Xt + t (8.1)

First-order serial correlation

t = r t - 1 + t (8.2)
368 CHAPTER 8 Regression with Time Series Data

Durbin-Watson statistic
n

a 1et - et - 12
2
t=2
DW = n (8.3)
2
a et
t=1

Relation of Durbin-Watson statistic to lag 1 residual autocorrelation (n large)

DW = 211 - r11e22 (8.4)

Transformed simple linear regression model

Yt¿ = b 011 - r2 + b 1X t¿ + t (8.5)

Generalized differences

Yt¿ = Yt - rYt - 1 (8.6)

X t¿ = Xt - rXt - 1
Simple or first differences

Yt¿ = Yt - Yt - 1 (8.7)
X t¿ = Xt - Xt - 1

Log linear regression model

Ln Yt = b 0 + b 1 Ln Xt + t (8.8)

Forecasting equation for the differenced form of log linear regression model
Ln YN t = Ln YN t - 1 + b11Ln Xt - Ln Xt - 12 (8.9)

First-order autoregressive model

Yt = b 0 + b 1Yt - 1 + t (8.10)

Forecasting equation for first-order autoregressive model

YNt = b0 + b1Yt - 1 (8.11)

Seasonal model with dummy variables for quarterly data

Yt = b 0 + b 1t + b 2S2 + b 3S3 + b 4S4 + t (8.12)

Error correction models

Yt - Yt - 1 = b 0 + b 11Yt - 1 - X t - 12 + t (8.13a)

Xt - Xt - 1 = b 0¿ + b 1¿ 1Yt - 1 - X t - 12 + t (8.13b)

Standardized coefficients (see Case 8-2)

sx
B = b¢ ≤ (8.14)
sy
 CHAPTER 8 Regression with Time Series Data 369

Problems

1. What is serial correlation, and why can it be a problem when time series data are
analyzed?
2. What is a major cause of serial correlation?
3. Which underlying regression assumption is often violated when time series vari-
ables are analyzed?
4. Name a statistic that is commonly used to detect serial correlation.
5. You test for serial correlation, at the .01 level, with 32 residuals from a regression
with two independent variables. If the calculated Durbin-Watson statistic is equal
to 1.0, what is your conclusion?
6. You test for serial correlation, at the .05 level, with 61 residuals from a regression
with one independent variable. If the calculated Durbin-Watson statistic is equal
to 1.6, what is your conclusion?
7. Suggest ways to solve the problem of serial correlation.
8. What are the predictor variables in an autoregressive model?
9. Tamson Russell, an economist working for the government, was trying to deter-
mine the demand function for passenger car motor fuel in the United States.
Tamson developed a model that used the actual price of a gallon of regular gaso-
line to predict motor fuel consumed per year. After adding a variable representing
the population of the United States to the model, she was able to explain 76.6% of
the variation in fuel consumption. Did Tamson have a problem with serial correla-
tion? The data are shown in Table P-9.
10. Decision Science Associates was asked to do a feasibility study for a proposed des-
tination resort to be located within half a mile of the Grand Coulee Dam. Mark
Craze was not happy with the regression model that used the price of a regular gal-
lon of gasoline to predict the number of visitors to the Grand Coulee Dam Visitors
Center. After plotting the data on a scatter diagram, Mark decided to use a dummy
variable to represent significant celebrations in the general area. Mark used a 1 to
represent a celebration and a 0 to represent no celebration. Note that the 1 in 1974
represents the Expo 74 World’s Fair celebrated in Spokane, Washington, the 1 in
1983 represents the celebration of the 50th anniversary of the construction of
Grand Coulee Dam, and the 1 in 1986 represents the World’s Fair held in
Vancouver, Canada. Mark also decided to use time as a predictor variable. The
data are shown in Table P-10. Suppose you were asked to write a report for Mark
to present to his boss. Indicate whether serial correlation is a problem. Also indi-
cate what additional information would be important in deciding whether to rec-
ommend that the destination resort be built.
11. Jim Jackson, a rate analyst for the Washington Water Power Company, while
preparing for a rate case needed to forecast electric residential revenue for 1996.
Jim decided to investigate three potential predictor variables: residential use in
kilowatt-hours (kWh), residential charge per kWh (cents/kWh), and number of
residential electric customers. He collected data from 1968 to 1995. The data are
shown in Table P-11. Jim testified before the Idaho Rate Commission and was
asked if serial correlation was a problem. He didn’t know the answer and has
asked you to write a response to the commission’s question.
370 CHAPTER 8 Regression with Time Series Data

TABLE P-9

Motor Fuel U.S.

Consumed by Cars Price of Population
(billions of gallons) Gasoline ($/gallon) (millions)
Year Y X1 X2

1973 78.8 .39 211.9

1974 75.1 .53 213.9
1975 76.4 .57 216.0
1976 79.7 .59 218.0
1977 80.4 .62 220.2
1978 81.7 .63 222.6
1979 77.1 .86 225.1
1980 71.9 1.19 227.7
1981 71.0 1.33 230.1
1982 70.1 1.26 232.5
1983 69.9 1.22 234.8
1984 68.7 1.21 236.3
1985 69.3 1.16 238.5
1986 71.4 .92 240.7
1987 70.6 .95 242.8
1988 71.7 .95 245.1

Source: Statistical Abstract of the United States, various years.

TABLE P-10

Number of Price of
Visitors Time Gasoline ($/gallon) Celebration
Year Y X1 X2 X3

1973 268,528 1 .39 0

1974 468,136 2 .53 1
1975 390,129 3 .57 0
1976 300,140 4 .59 0
1977 271,140 5 .62 0
1978 282,752 6 .63 0
1979 244,006 7 .86 0
1980 161,524 8 1.19 0
1981 277,134 9 1.31 0
1982 382,343 10 1.22 0
1983 617,737 11 1.16 1
1984 453,881 12 1.13 0
1985 471,417 13 .86 0
1986 654,147 14 .90 1

Source: Based on Grand Coulee Dam Visitors Center and Statistical Abstract of the
United States, 1988.
 CHAPTER 8 Regression with Time Series Data 371

TABLE P-11

Revenue Use per Charge Number of

($ millions) kWh (cents/kWh) Customers
Year Y X1 X2 X3

1968 19.3 10,413 1.33 139,881

1969 20.4 11,129 1.29 142,806
1970 20.9 11,361 1.25 146,616
1971 21.9 11,960 1.21 151,640
1972 23.4 12,498 1.19 157,205
1973 24.5 12,667 1.19 162,328
1974 25.8 12,857 1.21 166,558
1975 30.5 13,843 1.29 170,317
1976 33.3 14,223 1.33 175,536
1977 37.2 14,427 1.42 181,553
1978 42.5 14,878 1.52 188,325
1979 48.8 15,763 1.59 194,237
1980 55.4 15,130 1.84 198,847
1981 64.3 14,697 2.17 201,465
1982 78.9 15,221 2.55 203,444
1983 86.5 14,166 2.97 205,533
1984 114.6 14,854 3.70 208,574
1985 129.7 14,997 4.10 210,811
1986 126.1 13,674 4.34 212,865
1987 132.0 13,062 4.71 214,479
1988 138.1 13,284 4.82 215,610
1989 141.2 13,531 4.81 217,164
1990 143.7 13,589 4.81 219,968
1991 149.2 13,800 4.84 223,364
1992 146.1 13,287 4.83 227,575
1993 153.9 13,406 4.91 233,795
1994 146.9 12,661 4.84 239,733
1995 156.8 12,434 4.98 253,364

Source: Based on “Financial and Operating Supplement,” Washington Water

Power Annual Report, various years.

12. Paul Raymond, president of Washington Water Power, was worried about the pos-
sibility of a takeover attempt and the fact that the number of common sharehold-
ers has been decreasing since 1983. Suppose he instructed you to study the number
of common shareholders since 1968 and be prepared to compute a forecast for
1996. You decide to investigate three potential predictor variables: earnings per
share (common), dividends per share (common), and payout ratio. You collect the
data from 1968 to 1995, as shown in Table P-12.
a. Run these data on the computer, and determine the best model using your cho-
sen predictor variables.
b. Is serial correlation a problem in this model?
c. If serial correlation is a problem, write a memo to Paul that discusses various solu-
tions to the autocorrelation problem and includes your final recommendation.
13. Thompson Airlines has determined that 5% of the total number of U.S. domestic
airline passengers fly on Thompson planes. You are given the task of forecasting
372 CHAPTER 8 Regression with Time Series Data

TABLE P-12

Common Earnings Dividends Payout

Shareholders per Share ($) per Share ($) Ratio
Year Y X1 X2 X3

1968 26,472 1.68 1.21 72

1969 28,770 1.70 1.28 73
1970 29,681 1.80 1.32 73
1971 30,481 1.86 1.36 72
1972 30,111 1.96 1.39 71
1973 31,052 2.02 1.44 71
1974 30,845 2.11 1.49 71
1975 32,012 2.42 1.53 63
1976 32,846 2.79 1.65 55
1977 32,909 2.38 1.76 74
1978 34,593 2.95 1.94 61
1979 34,359 2.78 2.08 75
1980 36,161 2.33 2.16 93
1981 38,892 3.29 2.28 69
1982 46,278 3.17 2.40 76
1983 47,672 3.02 2.48 82
1984 45,462 2.46 2.48 101
1985 45,599 3.03 2.48 82
1986 41,368 2.06 2.48 120
1987 38,686 2.31 2.48 107
1988 37,072 2.54 2.48 98
1989 36,968 2.70 2.48 92
1990 34,348 3.46 2.48 72
1991 34,058 2.68 2.48 93
1992 34,375 2.74 2.48 91
1993 33,968 2.88 2.48 86
1994 34,120 2.56 2.48 97
1995 33,138 2.82 2.48 88

Source: Based on “Financial and Operating Supplement,” Washington Water Power

Annual Report, various years.

the number of passengers who will fly on Thompson Airlines in 2007. The data are
presented in Table P-13.
a. Develop a time series regression model, using time as the independent variable
and the number of passengers as the dependent variable. Fit this model.
b. Is the assumption of independent errors for this model viable?
c. Fit the model in part a with the logarithms of the number of passengers as the
dependent variable.
d. Repeat part a with the time represented by an exponential trend (see
Equation 5.6).
e. Which of the models in parts c and d do you prefer? Why?
f. Do the errors for either of the models in parts c and d appear to be indepen-
dent? If not, what problem(s) might arise when using one (or both) of these fitted
models for forecasting?
g. Using your preferred model, forecast the number of Thompson Airlines passen-
gers for 2007.
 CHAPTER 8 Regression with Time Series Data 373

TABLE P-13

Number of Number of Number of Number of

Passengers Passengers Passengers Passengers
Year (1,000s) Year (1,000s) Year (1,000s) Year (1,000s)

1982 22.8 1989 45.4 1996 79.9 2003 144.8

1983 26.1 1990 46.3 1997 96.3 2004 147.9
1984 29.4 1991 45.8 1998 109.0 2005 150.1
1985 34.5 1992 48.0 1999 116.0 2006 151.9
1986 37.6 1993 54.6 2000 117.2
1987 40.3 1994 61.9 2001 124.9
1988 39.5 1995 69.9 2002 136.6

14. Thomas Furniture Company concludes that production scheduling can be

improved by developing an accurate method of predicting quarterly sales. The
company analyst, Mr. Estes, decides to investigate the relationship between hous-
ing construction permits and furniture sales in the Springfield area. Estes feels that
permits will lead sales by one or two quarters. In addition, he wonders if furniture
sales are seasonal. Estes decides to consider another independent variable:

0 if first- or second-quarter sales

X2 = b
1 if third- or fourth-quarter sales

The data are given in Table P-14.

a. Develop a regression model that uses housing construction permits as the pre-
dictor variable.
b. Test for autocorrelation in this model.
c. Develop a regression model that uses both permits and the seasonal dummy as
predictor variables.
d. Is there a significant seasonal pattern in these data? (Test at the .05 level.)
e. Is there an autocorrelation problem with the multiple regression model devel-
oped in part c? If so, how might it be corrected?

TABLE P-14

Sales Sales
(1,000s) Permits (1,000s) Permits
Year Quarter Y X1 Year Quarter Y X1

2002 3 19 2005 1 120 72

4 3 2 150 31
2003 1 120 35 3 660 19
2 80 11 4 270 14
3 400 11 2006 1 200 75
4 200 16 2 280 41
2004 1 75 32 3 800 17
2 120 10 4 320 10
3 270 12 2007
4 155 21
374 CHAPTER 8 Regression with Time Series Data

TABLE P-15

Quarter

Year 1 2 3 4
1985 16.3 17.7 28.1 34.3
1986 17.3 16.7 32.2 42.3
1987 17.4 16.9 30.9 36.5
1988 17.5 16.5 28.6 45.5
1989 24.3 24.2 33.8 45.2
1990 20.6 18.7 28.1 59.6
1991 19.5 22.5 38.3 81.2
1992 24.9 17.5 26.8 59.1
1993 22.4 14.3 24.7 57.2
1994 16.2 16.5 35.5 59.8
1995 18.0 15.9 28.0 57.3
1996 17.1 17.0

Source: The Value Line Investment Survey (New York: Value

Line, 1996), p. 132.

f. Using the model from part a, forecast Thomas Furniture Company sales for the
four quarters of 2007. (Notice you will need some additional values for con-
struction permits to develop forecasts for the four quarters of 2007.)
15. National Presto is a manufacturer of small electrical appliances and housewares,
including pressure cookers, heaters, canners, fry pans, griddles, roaster ovens, deep
fryers, corn poppers, can openers, coffee makers, slicers, hand mixers, and portable
ranges. Its quarterly sales in millions of dollars for several years are shown in Table
P-15. Presto does most of its business at Christmas, so there is a strong seasonal
effect. Develop a multiple regression model using dummy variables to forecast sales
for the third and fourth quarters of 1996. Write a report summarizing your results.
16. The data in Table P-16 show seasonally adjusted quarterly sales for Dickson
Corporation and for the entire industry for 20 quarters.
a. Fit a linear regression model, and store the residuals. Plot the residuals against
time, and obtain the residual autocorrelations. What do you find?
b. Calculate the Durbin-Watson statistic, and determine whether autocorrelation
exists.
c. Estimate the regression coefficient, b 1, using generalized differences. (Estimate
r with the lag 1 residual autocorrelation coefficient.)
d. Compare the standard errors of the two estimates of b 1 obtained using the original
data and the generalized differences. Which estimate is more accurate? Explain.
17. Refer to Example 8.5. Using the Sears data in Table 8-5, convert the sales and dis-
posable income values to simple differences. That is, create the numbers
Yt¿ = Yt - Yt - 1 and X t¿ = Xt - Xt - 1. Fit a simple linear regression model to the dif-
ferenced data. Compare your results to the results obtained by the method of gen-
eralized differences in Example 8.5. Did you expect them to be much different?
Explain.
18. Although the time series data in Table P-18 are old, they provide the basis for
some interesting regression modeling. Using the data in Table P-18, attempt to
 CHAPTER 8 Regression with Time Series Data 375

TABLE P-16
Dickson Industry Dickson Industry
Sales Sales Sales Sales
(1,000s) (millions) (1,000s) (millions)
Year Quarter Y X1 Year Quarter Y X1

2002 1 83.8 31.8 2004 3 98.2 37.1

2 85.6 32.5 4 97.2 36.6
3 87.8 33.2 2005 1 100.1 37.6
4 86.1 32.4 2 102.6 38.3
2003 1 89.6 33.8 3 105.4 39.3
2 91.0 34.3 4 107.9 40.2
3 93.9 35.3 2006 1 110.1 41.1
4 94.6 35.7 2 111.1 41.4
2004 1 96.4 36.4 3 110.1 41.1
2 96.0 36.3 4 111.1 41.4

TABLE P-18

Personal Personal Personal Personal Personal Personal

Year Savings Y Income X Year Savings Y Income X Year Savings Y Income X

1935 2 60 1942 28 123 1949 9 207

1936 4 69 1943 33 151 1950 13 279
1937 4 74 1944 37 165 1951 18 257
1938 1 68 1945 30 171 1952 19 273
1939 3 73 1946 15 179 1953 20 288
1940 4 78 1947 7 191 1954 19 290
1941 11 96 1948 13 210

relate personal savings to personal income (in billions of dollars) for the time
period from 1935 to 1954.
a. Fit a simple linear regression model to the data in Table P-18, using personal
income to predict personal savings. (1) Test for the significance of the regression
at the a = .01 level; (2) calculate r 2, and interpret this quantity; and (3) test for
autocorrelation using the Durbin-Watson test with a = .05. Should you modify
your conclusion in part 1? How can the model be improved?
b. Construct a dummy variable, X2, for the WWII years. Let X2 = 0 for peacetime
and X2 = 1 for wartime.The war years are from 1941 to 1945. Fit a multiple linear
regression model using personal income and the war years dummy variable as
predictor variables for personal savings. Evaluate the results. Specifically
(1) test to determine whether knowledge of the war years makes a significant con-
tribution to the prediction of personal savings beyond that provided by personal
income (set a = .01) and (2) test for autocorrelation. Is the multiple regression
model better than the simple linear regression model of part a? Discuss.
19. Circuit City Inc. is a retailer of video and audio equipment and other consumer
electronics and office products. Recently, sales have been weak, declining by a total
376 CHAPTER 8 Regression with Time Series Data

TABLE P-19

Year May 31 Aug 31 Nov 30 Feb 28

1996 1,742 1,806 1,999 2,897

1997 1,851 1,948 2,039 3,156
1998 2,078 2,369 2,222 3,289
1999 2,205 2,423 2,495 3,476
2000 2,449 2,566 2,336 3,177
2001 1,882 2,037 2,280 3,391
2002 2,118 2,221 2,422 3,239
2003 2,150 2,350 2,600 3,400

Source: The Value Line Investment Survey (New York: Value Line, 2002), p. 1725.

of 5% in December. Among the reasons were a fiercely competitive retailing envi-

ronment among consumer-electronics sellers, price deflation in many products, a
slump in store traffic, and sluggish demand for most computer hardware items.20
Jim Lowe has the task of forecasting sales for 2003. Jim has access to estimates
provided by The Value Line Investment Survey (see Table P-19); however, he is
afraid that they are optimistic.
a. Plot the data from 1996 through 2002 as a time series. Do these data appear to
be seasonal? Calculate the autocorrelation function.
b. From the results in part a, an autoregressive model with the predictor variable
sales lagged four time periods might be appropriate. Why?
c. Let Yt be Circuit City sales. Using the data through 2002, fit an autoregressive
model of the form Yt = b 0 + b 1Yt - 4 + t, and store the residuals. Compute the
residual autocorrelations.
d. Use the fitted model from part c to generate forecasts of sales for the four quar-
ters of 2003. Compare your forecasts with the Value Line estimates.
e. Was Jim correct to think the Value Line estimates for 2003 were optimistic?
f. Do you think the autoregressive model in part c can be improved? Discuss.
20. The demand for a commodity typically depends on the income of the consumer,
the real price of the commodity, and the real price of complementary or competing
products. Table P-20 gives the per capita consumption of chicken in the United
States (in pounds); the per capita disposable income (in dollars); and the retail
prices for chicken, pork, and beef (in cents per pound) for several years.
a. Compute the correlation matrix for all the variables, using both the original
units and the log-transformed units. Comment on the implied strength of the
linear association between chicken consumed and each of the remaining vari-
ables. Can you think of a reason why one must be careful interpreting the mag-
nitudes of correlation coefficients constructed from time series data?
b. Using the original data, run a stepwise regression program with chicken con-
sumption as the dependent variable and the remaining variables as predictor
variables. Set alpha to enter = alpha to remove = .05.

20In late 2006, Wal-Mart decided to slash the price of its 42-inch flat-panel TV to less than $1,000. This move
triggered a financial meltdown among some consumer-electronics retailers during the following several
months. Circuit City closed 70 U.S. stores, laid off 3,400 employees, and put its 800 Canadian stores on the
block. Source: MSN Money, April 23, 2007.
 CHAPTER 8 Regression with Time Series Data 377

TABLE P-20

Per Capita
Chicken Disposable Chicken
Consumption Income Price Pork Price Beef Price
Year (lbs.) ($) (c/lb.) (c/lb.) (c/lb.)

1 28.0 397.5 42.2 50.7 78.3

2 29.9 413.3 38.1 52.0 79.2
3 30.1 439.2 40.3 54.0 79.2
4 30.9 459.7 39.5 55.3 79.2
5 31.4 492.9 37.3 54.7 77.4
6 33.7 528.6 38.1 63.7 80.2
7 35.6 560.3 39.3 69.8 80.4
8 35.6 624.6 37.8 65.9 83.9
9 37.1 666.4 38.4 64.5 85.5
10 38.5 717.8 40.1 70.0 93.7
11 40.3 768.2 38.6 73.2 106.1
12 40.2 843.3 39.8 67.8 104.8
13 41.7 911.6 39.7 79.1 114.0
14 39.9 931.1 52.1 95.4 124.1
15 39.7 1,021.5 48.9 94.2 127.6
16 39.0 1,165.9 58.3 123.5 142.9
17 39.1 1,349.6 57.9 129.9 143.6
18 42.8 1,449.4 56.5 117.6 139.2
19 44.9 1,575.5 63.7 130.9 165.5
20 48.3 1,759.1 61.6 129.8 203.3
21 49.0 1,994.2 58.9 128.0 219.6
22 49.4 2,258.1 66.4 141.0 221.6
23 49.6 2,478.7 70.4 168.2 232.6

c. Run a full regression of chicken consumption on the remaining variables. Be

sure to delete variables one at a time that you deem to be not significant until
you are satisfied with your final model. Is your result consistent with the result
of the stepwise procedure in part b? Is serial correlation likely to be a problem
in this regression analysis?
21. Repeat parts b and c of Problem 20 with the log-transformed data. Give an inter-
pretation of the coefficients of income and chicken price in terms of elasticities.
Using your final fitted regression function, indicate how a forecast of the following
year’s chicken consumption would be obtained.
22. Refer to Problem 20. Consider only the variables chicken consumption, income,
and chicken price in the original units. Compute simple differences for each of the
variables. Using the differenced data, regress the change in chicken consumption
on the change in income and the change in chicken price. Examine the residuals.
Interpret the results of your regression analysis. Do either of the predictor vari-
ables have much explanatory power? Explain.
23. Refer to Problem 20. Run a simple linear regression of chicken consumption on
chicken consumption lagged one time period. Examine the residuals. Interpret
378 CHAPTER 8 Regression with Time Series Data

the results of your regression analysis. Is this year’s chicken consumption likely
to be a good predictor of next year’s chicken consumption? Explain. Can we
infer the effect of a change in chicken price on chicken consumption with this
model?
24. Consider the bivariate system
Xt = Xt - 1 + t
Yt = Xt + t

where t and t are each independently distributed with mean zero and variance s2.
Develop an expression for Yt - Yt - 1, and show that X and Y are cointegrated.
What is the cointegrating linear combination in this case?

CASES

CASE 8-1 COMPANY OF YOUR CHOICE

A company’s health can be examined every month, for a company of your choice and then to analyze
quarter, and/or year with measurements on an array the patterns in the data using autocorrelation
of variables. For any one of these variables, there analysis. In addition, you can use an appropriate
may be several other variables that can provide computer program to develop an equation that
insights into its behavior and that can be used as can be used to forecast future values of your time
predictor variables in a forecasting equation. series variable.
The purpose of this case is to simulate the
identification of an important time series variable

ASSIGNMENT
1. Identify a company or organization that inter- differenced data. Describe the resulting pat-
ests you. The company can be a local or national terns in the time series of first differences.
company that has published records, including 5. Identify several potential predictor variables for
the measurement of time series variables. your dependent variable. You may use company
Identify a key variable for your chosen com- records and other data sources in this process.
pany, and record its values for several years, 6. Develop a forecasting equation for your
quarters, or months. dependent variable using one or more of the
2. Calculate several autocorrelation coefficients, identified predictor variables.
and plot the autocorrelation function. 7. Examine the residuals from your fitted model.
3. Based on the pattern of the autocorrelation In particular, check for autocorrelation. Once
function, describe the patterns in your time you are satisfied with your forecasting equation,
series. generate forecasts for the next six time periods.
4. Compute the first differences for your data, and If possible, compare your forecasts with the
construct the autocorrelation function for the actual values.
 CHAPTER 8 Regression with Time Series Data 379

CASE 8-2 BUSINESS ACTIVITY INDEX

FOR SPOKANE COUNTY

Prior to 1973, Spokane County, Washington, had no local business conditions. Personal income measures
up-to-date measurement of general business activ- the total income received by households before per-
ity. What happens in this area as a whole, however, sonal taxes are paid. Since productive activities are
affects every local business, government agency, and typically remunerated by monetary means, personal
individual. Plans and policies made by an economic income may indeed be viewed as a reasonable proxy
unit would be incomplete without some reliable for the general economic performance. Why then is
knowledge about the recent performance of the it necessary to construct another index if personal
economy of which the unit is a component part. A income can serve as a good business activity indica-
Spokane business activity index should serve as a tor? Unfortunately, personal income data at the
vital input in the formulation of strategies and deci- county level are estimated by the U.S. Department
sions in private as well as in public organizations. of Commerce on an annual basis and are released 16
A business activity index is an indicator of the months too late. Consequently, these data are of lit-
relative changes in overall business conditions tle use for short-term planning. Young’s task is to
within a specified region. At the national level, the establish an up-to-date business activity index.
gross domestic product (GDP, computed by the The independent variables are drawn from those
Department of Commerce) and the industrial pro- local data that are readily available on a monthly
duction index (compiled by the Federal Reserve basis. Currently, about 50 series of such monthly data
Board) are generally considered excellent indica- are available, ranging from employment, bank
tors. Each of these series is based on thousands of activities, and real estate transactions to electrical
pieces of information—the collecting, editing, and consumption. If each series were to be included in
computing of which are costly and time-consuming the regression analysis, the effort would not be very
undertakings. For a local area such as Spokane productive because only a handful of these series
County, Washington, a simplified version, capable of would likely be statistically significant. Therefore,
providing reasonably accurate and current informa- some knowledge of the relationship between per-
tion at moderate cost, is very desirable. sonal income and the available data is necessary in
Multiple regression can be used to construct a order to determine which independent variables are
business activity index. There are three essential to be included in the regression equation. From
questions that must be answered in order to con- Young’s knowledge of the Spokane economy, the
struct such an index: following 10 series are selected:

• What are the components of the index? X1, total employment

• Do these components adequately represent the X2, manufacturing employment
changes in overall business conditions? X3, construction employment
• What weight should be assigned to each of the X4, wholesale and retail trade employment
chosen components? X5, service employment
X6, bank debits
Answers to these questions can be obtained through X7, bank demand deposits
regression analysis. X8, building permits issued
Dr. Shik Chun Young, professor of economics at X9, real estate mortgages
Eastern Washington University, is attempting to X10, total electrical consumption
develop a business activity index for Spokane
County. Young selects personal income as the The first step in the analysis is to fit the model
dependent variable. At the county level, personal
income is judged as the best available indicator of E1Y2 = b 0 + b 1X1 + b 2X2 + Á + b 10X10
380 CHAPTER 8 Regression with Time Series Data

where ¢X7, change in demand deposits

¢X8, change in building permits issued
Y = personal income
¢X9, change in real estate mortgages
b 0 = the Y-intercept ¢X10, change in total electrical consumption
b 1, b 2, Á , b 10 = the coefficients of the respec-
tive independent variables The regression model becomes
When the preceding model is fit to the data, the
E1¢Y2 = b 0 + b 1 ¢X1 + b 2 ¢X2 + # # # + b 10 ¢X10
adjusted R2 is .96, which means that the 10 variables
used together explain 96% of the variability in the
dependent variable, personal income. However, other where
regression statistics indicate problems. First, of these ¢Y = the change in personal income
10 independent variables, only three—namely, total b 0 = the ¢Y-intercept
employment, service employment, and bank debits—
b 1, b 2, Á , b 10 = the coefficients of the respec-
have a computed t value significant at the .05 level.
tive independent variables
Second, the correlation matrix shows a high degree
of interdependence among several of the indepen- A regression run using this model, based on the first
dent variables—multicollinearity.21 For example, total difference data, produces a Durbin-Watson statistic
employment and bank debits have a correlation coef- of 1.71. It indicates that no serious autocorrelation
ficient of .88; total electrical consumption and bank remains.
demand deposits, .76; and building permits issued and The next step is to determine which of the 10
real estate mortgages, .68. Third, a test for autocorre- variables are significant predictors of the dependent
lation using the DurbinWatson statistic of .91 indi- variable. The dependent variable, ¢Y, is regressed
cates that successive values of the dependent variable against several possible combinations of the 10
are positively correlated. Of course, autocorrelation is potential predictors in order to select the best equa-
rather common in time series data; in general, obser- tion. The criteria used in the selection are
vations in the same series tend to be related to one
another. • A satisfactorily high R2
Since one of the basic assumptions in regression • Low correlation coefficients among the inde-
analysis is that the observations of the dependent pendent variables
variable are random, Young chooses to deal with the • Significant (at the .05 level) coefficients for each
autocorrelation problem first. He decides to calculate of the independent variables
first differences, or changes, in an attempt to minimize
the interdependence among the observations in each After careful scrutiny of the regression results,
of the time series. The 10 independent variables are Young finds that the equation that contains
now measured by the difference between the periods ¢X4, ¢X5, and ¢X10 as independent variables best
rather than by the absolute value for each period. So meets the foregoing criteria.
that the sets of data can be distinguished, a new desig- However, Young reasons that (in addition to
nation for the independent variables is used: commercial and industrial uses) total electrical con-
sumption includes residential consumption, which
¢X1, change in total employment should not have a significant relation to business
¢X2, change in manufacturing employment activity in the near term. To test this hypothesis,
¢X3, change in construction employment Young subdivides the total electrical consumption
¢X4, change in wholesale and retail trade into four variables:
employment
¢X5, change in service employment ¢X11, change in residential electricity use
¢X6, change in bank debits ¢X12, change in commercial electricity use

21Some care must be exercised in interpreting a correlation between two time series variables, as autocorre-
lation in the individual series can produce spurious correlation—see Example 8.1.
 CHAPTER 8 Regression with Time Series Data 381

TABLE 8-12 Young’s Regression Variables TABLE 8-13 Correlation Coefficient Matrix

Independent Dependent X4 X5 X14

Equation Variables Variable
X4 1.00
A ¢X4, ¢X5, ¢X11 ¢Y
B ¢X4, ¢X5, ¢X12 ¢Y X5 .45 1.00
C ¢X4, ¢X5, ¢X13 ¢Y X14 .11 .12 1.00
D ¢X4, ¢X5, ¢X14 ¢Y

(Recall that the regression coefficient represents the

¢X13, change in industrial electricity use average change in the dependent variable for a one-
¢X14, change in commercial and industrial elec- unit increase in the independent variable.) However,
tricity use because the variables in the regression equation are
not measured in the same units (for example, ¢Y is
All four variables, combined with ¢X4 and ¢X5, measured in thousands of dollars and ¢X14 in thou-
are used to produce four new regression equations sands of kilowatt-hours), the regression coefficients
(see Table 8-12). must be transformed into relative values. This trans-
Statistical analysis indicates that Equation D in formation is accomplished by computing their
Table 8-12 is the best. Compared with the previous standardized or B coefficients.
equation, which contained ¢X4, ¢X5, and ¢X10 as
independent variables, Equation A is the only one sx
that shows a deterioration in statistical significance. B = b¢ ≤ (8.14)
sy
This result confirms Young’s notion that commercial
and industrial electricity uses are better predictors
of personal income than total electrical consump- where
tion, which includes residential electricity use. b = the independent variable’s regression
Therefore, Equation D is selected as the final coefficient
regression equation, and the results are sx = the independent variable’s standard deviation
¢YN = - 1.86 + 17.10¢X4 + 23.01¢X5 + .007¢X14 sy = the dependent variable’s standard deviation
(4.07) (5.61) (.002) The values of all these statistics are typically
n = 15 F = 26.26 available from the regression computer output.
Hence, the standardized coefficients of the three
DW = 1.77 R2 = .835 independent variables are

The figures in parentheses below the regression B4 = .4959

coefficients are the standard errors of the estimated
B5 = .4833
coefficients. The t values of the coefficients are 4.20,
4.10, and 3.50 for ¢X4, ¢X5, and ¢X14, respectively. B14 = .3019
The R 2 indicates that nearly 84% of the variance in Total = 1.2811
change in personal income is explained by the three
independent variables. The Durbin-Watson (DW) Because the sum of the weights in an index must be
statistic shows that autocorrelation is not a problem. 100%, the standardized coefficients are normalized
In addition, the correlation coefficient matrix of as shown in Table 8-14.
Table 8-13 demonstrates a low level of interdepen- After the components and their respective
dence among the three independent variables. weights have been determined, the following steps
For index construction purposes, the independent give the index:
variables in the final regression equation become the
index components. The weights of the components 1. Compute the percentage change of each com-
can be determined from the regression coefficients. ponent since the base period.
382 CHAPTER 8 Regression with Time Series Data

140
Spokane Index

130

120
Percentage

U.S. GNP

110

100

90

1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976

FIGURE 8-12 Spokane County Business Activity Index and the

U.S. GNP, in Constant Dollars 11967 = 1002

2. Multiply the percentage change by the appro-

TABLE 8-14 Standardized Coefficients
priate weight.
Component Weight 3. Sum the weighted percentage changes obtained
in step 2.
¢X4 .4959
= .3871
1.2811
The completed Spokane County activity index, for
¢X5 .4833
the time period considered, is compared with the
= .3772
1.2811 U.S. GNP, in constant dollars 11967 = 1002, in
¢X14 .3019 Figure 8-12.
= .2357
1.2811
Total = 1.000

QUESTIONS
1. Why did Young choose to solve the autocorrela- 4. Should the regression done on the first differ-
tion problem first? ences have been through the origin?
2. Would it have been better to eliminate multi- 5. Is there any potential for the use of lagged data?
collinearity first and then tackle autocorrelation? 6. What conclusions can be drawn from a compar-
3. How does the small sample size affect the ison of the Spokane County business activity
analysis? index and the GNP?
 CHAPTER 8 Regression with Time Series Data 383

7,500

6,000 2 32
3
3
3 2
Sales

4,500

2
2 2

3,000

2
1,500
0 25 50 75 100 125
Time

FIGURE 8-13 Restaurant Sales, January 1981–December 1982

CASE 8-3 RESTAURANT SALES22

Jim Price, at the time he was completing his M.B.A. The mean weekly sales for the 104 weeks turned
degree, worked at a small restaurant near Marquette out to be $4,862. Figure 8-13 is a graph of the weekly
University in Milwaukee, Wisconsin. One day the sales over time. The graph indicates that weekly
restaurant manager asked Jim to report to her sales were quite volatile, ranging from $1,870 to
office. She indicated that she was very interested in $7,548, with very little trend. Since Jim had recently
forecasting weekly sales and wanted to know completed a course on regression analysis, he
whether Jim would help. Since Jim had just taken an decided to use weekly sales as the dependent vari-
advanced statistics course, he said that he would able and see if he could find some useful indepen-
enjoy the challenge. dent or predictor variables.
Jim asked the restaurant manager to provide Jim tested three predictors. The first predictor
him with whatever historical records she had avail- was time. The second predictor was a dummy vari-
able. She indicated that the restaurant compiled the able indicating whether or not Marquette University
previous week’s sales every Monday morning. Jim was in full session that week 10 = not in full session,
began his analysis by obtaining weekly sales data 1 = in full session2. Examination of the sales data
from the week ending Sunday, January 1, 1981, in Figure 8-13 revealed that weekly sales always
through the week ending Sunday, December 29, dropped when Marquette was not in full session,
1982—a total of 104 observations. namely, during the Christmas break, the spring

22Frank G. Forst, Marquette University, Milwaukee, Wisconsin, contributed this case study.
384 CHAPTER 8 Regression with Time Series Data

TABLE 8-15 Restaurant Sales Correlation Matrix

Current Dummy Sales Lagged

Sales Time Variable One Week

Current sales 1.000 .049 .772 .580

Time 1.000 .048 .120
2
Dummy variable 1.000 .490
Sales lagged one week 1.000

TABLE 8-16 Restaurant Regression Models

Autocorrelation
Durbin-Watson Significant at the Amount of
Model Predictor(s) R2 Statistic .05 Level? Collinearity

(1) Time .0024 .81 Yes None

(2) Dummy .5960 1.30 Yes None
(3) Sales lagged one week .3360 N.A.* No None
(4) Time and dummy .6030 1.32 Yes Very little
(5) Time and sales lagged one .3360 N.A.* No Little
week
(6) Dummy and sales lagged .6490 N.A.* No Little
one week
(7) Time, dummy, and sales .6510 N.A.* No Moderate
lagged one week
* N.A. indicates that the Durbin-Watson statistic is not applicable because it cannot be used with a
lagged dependent variable.

break, and the summer break. Jim was not sur- current week’s sales. The current week’s sales were
prised, since the restaurant is located near Mar- moderately related, r = .580, to sales lagged one
quette’s campus and most of its customers are week. However, Jim also noticed that the dummy
members of the Marquette community. The third variable was moderately related, r = .490, to sales
predictor Jim tried was sales lagged one week, lagged one week.
since examination of Figure 8-13 indicated that Jim experimented with several regression
sales for two adjacent weeks were frequently models. The results of the various regression
similar. models are presented in Table 8-16. Since the sales
Jim computed the simple correlations among data have almost no trend, the predictor “time”
the three potential predictors and the dependent adds very little predictive power to the regression
variable, weekly sales. The results are presented in model. Note that model 4 has only a slightly higher
the correlation matrix shown in Table 8-15. As Jim R2 than model 2 and that the errors from both
expected, there was almost no trend in the weekly models appear to have a significant amount of
sales, as indicated by the correlation coefficient of autocorrelation. Models 3 and 5 have the same R2,
.049. However, the dummy variable was strongly whereas model 7 has only a slightly higher R2 than
correlated with current sales, r = .772 ; in other model 6. On the other hand, the predictor “sales
words, whether or not Marquette University is in lagged one week” adds a fair amount of predictive
full session had good potential as a predictor of the power to a regression model. Finally, model 6 has a
 CHAPTER 8 Regression with Time Series Data 385

higher R2 than model 2, without a significant R2 means 64.9% of the variation in weekly sales can
amount of residual autocorrelation. be explained by whether or not Marquette was in full
Jim decided to select regression model 6 to fore- session and by what the previous week’s sales were.
cast weekly sales for the following reasons: The regression equation implies that the weekly
sales average is about $1,611 higher when Marquette
1. Model 6 had the second largest R 2, only .002 is in full session, holding the previous week’s sales
below that of model 7. constant.
2. The parameters of model 6 were each signifi- Jim was pleased with his effort but wondered if
cantly different from zero at the .01 level. another type of forecasting model might not be
3. Autocorrelation is not a problem for model 6. more effective. For this reason, he decided to take a
4. Model 6 is simpler than model 7 and does not forecasting course.
have as much multicollinearity.
The fitted regression function that Jim used was
YN = 2,614.3 + 1,610.71dummy variable2
+ .26051sales lagged one week2

QUESTIONS
1. Was Jim’s use of a dummy variable correct? 3. Do you agree with Jim’s conclusions?
2. Was it correct for Jim to use sales lagged one 4. Would another type of forecasting model be
week as a predictor variable? more effective for forecasting weekly sales?

CASE 8-4 MR. TUX

John Mosby ran a simple regression analysis using 87.7% of the sales variable variance. However, a
time as the predictor variable and observed a disap- MAPE of over 20% seems high. Next, John generated
pointing r 2 of .563 (see Case 6-4). Since he knew the autocorrelations for the residuals of the model,
that his sales data have a strong seasonal compo- which are shown in Figure 8-14.
nent, he decided to fit a seasonal multiple regression The residual autocorrelations have a spike at
model. lag 12, the seasonal lag. Somewhat dissatisfied
John created 11 dummy variables for the sea- with the results of his seasonal regression model,
sonal effect. Since the Mr. Tux data are monthly, he John decided to try an autoregressive model to
coded S1 as 1 if the data were from January; 0 other- forecast his monthly sales (the data appeared in
wise. John did this for each month, ending with S11, Case 2-2).
which was 1 for November and 0 otherwise. Because of his data’s strong seasonal compo-
John ran a model using all 12 predictor vari- nent, John decided to try to model this component
ables: 1 for time and the other 11 representing the using an autoregressive model with the Y values
monthly seasonal effect. The results are shown in lagged 12 months. Lagging the data 12 periods
Table 8-17. leaves John with a sample size of 84. The Minitab
John hand-calculated the MAPE for the last 12 output is shown in Table 8-18. The autocorrelation
months of data and found it to be 21.25%. He is not function for the residuals from the autoregressive
sure whether he has a good model.The model explains model is shown in Figure 8-15.
 TABLE 8-17 Minitab Output for Mr. Tux Seasonal Regression Model
Regression Analysis: Sales versus Time, S1, . . .
The regression equation is
Sales = - 35023 + 2752 Time - 48459 S1 - 29808 S2 + 21681 S3 + 119019 S4
+ 139212 S5 + 57713 S6 + 21689 S7 + 74014 S8 + 7872 S9
- 9009 S10 - 25050 S11
Predictor Coef SE Coef T P
Constant - 35023 15441 - 2.27 0.026
Time 2752.4 141.0 19.52 0.000
S1 - 48459 19059 - 2.54 0.013
S2 - 29808 19048 - 1.56 0.121
S3 21681 19039 1.14 0.258
S4 119019 19030 6.25 0.000
S5 139212 19022 7.32 0.000
S6 57713 19015 3.04 0.003
S7 21689 19009 1.14 0.257
S8 74014 19005 3.89 0.000
S9 7872 19001 0.41 0.680
S10 - 9009 18998 - 0.47 0.637
S11 - 25050 18997 - 1.32 0.191
S = 37992.3 R - Sq = 87.7% R - Sq1adj2 = 85.9%
Analysis of Variance
Source DF SS MS F P
Regression 12 8.55392E+11 71282630871 49.38 0.000
Residual Error 83 1.19804E+11 1443416727
Total 95 9.75195E+11
Durbin-Watson statistic = 1.41

FIGURE 8-14 Autocorrelation Function for Mr. Tux Seasonal Regression

Model Residuals

386
 CHAPTER 8 Regression with Time Series Data 387

TABLE 8-18 Minitab Output for Mr. Tux Seasonal Autoregressive Model
Regression Analysis: Sales versus Lag12Sales
The regression equation is
Sales = 24786 + 1.07 Lag12Sales
84 cases used, 12 cases contain missing values
Predictor Coef SE Coef T P
Constant 24786 5322 4.66 0.000
Lag12Sales 1.06807 0.03803 28.08 0.000
S = 30784.9 R - Sq = 90.6% R - Sq1adj2 = 90.5%
Analysis of Variance
Source DF SS MS F P
Regression 1 7.47470E + 11 7.47470E + 11 788.71 0.000
Residual Error 82 77712386803 947712034
Total 83 8.25183E + 11

FIGURE 8-15 Autocorrelation Function for Mr. Tux Autoregressive

Model Residuals

ASSIGNMENT
1. Write a memo to John with a careful analysis of evaluation of model fit, potential forecast
the results of his two attempts to develop a accuracy, and any remaining problems—for
seasonal forecasting model. Which model is example, autocorrelation.
better? Be sure your discussion includes an
388 CHAPTER 8 Regression with Time Series Data

CASE 8-5 CONSUMER CREDIT COUNSELING

The Consumer Credit Counseling (CCC) operation data on the building permits issued from January
was described in Cases 1-2 and 3-3. 1986 to December 1992 are shown in Table 8-20.
The executive director, Marv Harnishfeger, con- Dorothy developed a multiple regression model
cluded that the most important variable that CCC that used the number of people on food stamps, the
needed to forecast was the number of new clients to be business activity index, the number of bankruptcies
seen in the rest of 1993. Marv provided Dorothy filed, and the number of building permits issued. She
Mercer monthly data for the number of new clients also created a model based solely on the assumption
seen by CCC for the period from January 1985 through that the data were seasonal.
March 1993 (see Case 3-3). In Case 3-3, Dorothy used Dorothy was recently informed that serial correla-
autocorrelation analysis to explore the data pattern. In tion is often a problem when regression is performed
Case 6-5, she tried both the number of people on food with time series data. She became worried that some
stamps and a business activity index to develop a of the regression models developed to predict the
regression model to forecast the rest of 1993. number of clients seen were affected by this condition.
Dorothy was not happy with the results of her Since she liked the report you submitted that used the
regression model. She decided to try multiple regres- time series decomposition analysis model, she
sion and asked Marv to think of other variables that assigned you the task of checking on this situation.
might be related to the number of clients seen. Marv You vaguely remembered developing an autore-
indicated that she might try the number of bankrupt- gressive model at one point in your educational expe-
cies filed and the number of building permits issued. rience and asked Dorothy if she would like to have
Data on the bankruptcies filed from January you look into that possibility. She decided that it
1986 to December 1992 are shown in Table 8-19, and would be a good way for you to spend your time.

TABLE 8-19 Bankruptcies Filed, January 1986 to December 1992

Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.
1986 160 170 140 173 162 160 150 145 134 155 145 152
1987 171 206 173 195 165 177 168 165 131 169 166 157
1988 174 162 196 178 169 170 143 177 192 195 164 146
1989 180 149 200 165 168 177 143 180 169 170 160 161
1990 172 146 185 143 156 173 140 160 131 169 152 135
1991 136 167 179 181 166 151 165 129 132 162 140 140
1992 130 165 172 145 129 166 146 143 127 186 157 94

TABLE 8-20 Building Permits Issued, January 1986 to December 1992

Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.
1986 49 60 149 214 191 193 161 174 168 203 111 64
1987 44 76 137 145 150 168 168 170 158 146 104 48
1988 32 32 127 149 128 178 132 164 152 92 121 77
1989 61 48 108 161 200 204 164 175 200 171 153 79
1990 111 92 166 189 245 261 207 238 198 247 164 108
1991 67 118 205 253 296 300 284 282 282 261 237 130
1992 133 210 298 334 312 333 311 327 352 387 243 140
 CHAPTER 8 Regression with Time Series Data 389

ASSIGNMENT
1. Analyze the significance of the variables in do you feel is the better candidate for generating
Dorothy’s regression model. Develop a regres- forecasts for the rest of 1993? Write Dorothy a
sion model (be sure to include additive dummy memo that provides her with the information she
variables for the seasonal component, if nec- has requested concerning the problem of serial
essary), and use it to forecast the number correlation. Include an analysis of the results
of new clients for the first three months of of your efforts to develop an appropriate model
1993. Compare your forecasts with the actual to forecast the number of new clients for the
observations. remainder of 1993.
2. Develop an autoregressive model, and generate
forecasts for the first three months of 1993.Which
model (multiple regression or autoregression)

CASE 8-6 AAA WASHINGTON23

An overview of AAA Washington was provided in average monthly temperature, monthly rainfall, and
Case 5-5 when students were asked to prepare a Washington State unemployment rate were gath-
time series decomposition of the emergency road ered and are presented in Table 8-21. A conversation
service calls received by the club over five years. The with the manager of the emergency road service call
time series decomposition performed in Case 5-5 center led to two important observations: (1) Auto-
showed that the pattern Michael DeCoria had makers seem to design cars to operate best at
observed in emergency road service call volume was 65 degrees Fahrenheit and (2) call volume seems to
probably cyclical in nature. Michael wanted to be increase more sharply when the average tempera-
able to predict the cyclical effect on emergency road ture drops a few degrees from an average tempera-
service call volume for future years. ture in the 30s than it does when a similar drop
Other research done by the club identified occurs with an average temperature in the 60s. This
several factors that have an impact on emergency information suggested that the effect of temperature
road service call volume. Among these factors are on emergency road service was nonlinear.
the average daily temperature and the amount of In Case 6-6, four linear regression models using
rainfall received in a day. This research has shown the total number of emergency road service calls as
that emergency road service calls increase as rain- the dependent variable and the unemployment rate,
fall increases and as the average daily temper- temperature, rainfall, and number of members as
ature falls. The club also believes that the total the four independent variables were investigated.
number of emergency road service calls it receives The temperature variable was transformed by
is dependent on the number of members in the subtracting 65 degrees from the average monthly
club. temperature values. A nonlinear relationship was
Michael had observed that the cyclical trend of then researched.
the time series seemed to be lagging behind the In Case 7-2, a multiple regression model was
general economic cycle. He suggested that the developed. Variables such as the rainfall, number
unemployment rate for Washington State would of members, exponentially transformed average
be a good surrogate measurement for the gen- monthly temperature, and unemployment rate
eral state of Washington’s economy. Data on the lagged 11 months were tested.

23Thiscase was provided by Steve Branton, former student and M.B.A. graduate, Eastern Washington
University.
390 CHAPTER 8 Regression with Time Series Data

TABLE 8-21 Data for AAA Washington

Year Month Calls Unemployment Rate Temperature Rainfall Members
1987 June — 6.9940 — — —
July — 6.8137 — — —
August — 6.3648 — — —
September — 6.5435 — — —
October — 6.7826 — — —
November — 6.9225 — — —
December — 7.1560 — — —
1988 January — 7.9143 — — —
February — 7.7576 — — —
March — 7.0641 — — —
April — 6.0977 — — —
May 20,002 5.7867 55.1 3.75 —
June 21,591 5.7592 59.0 1.95 —
July 22,696 5.5718 63.8 0.89 —
August 21,509 5.2939 63.8 0.51 384,746
September 22,123 5.4709 59.1 2.31 388,652
October 21,449 5.5049 54.6 3.12 392,241
November 23,475 5.8630 45.4 8.42 393,115
December 23,529 6.1349 41.0 4.44 392,631
1989 January 23,327 7.5474 40.3 4.30 396,975
February 24,050 7.8157 34.3 3.18 395,186
March 24,010 7.1390 43.2 6.57 397,791
April 19,735 6.2637 52.5 2.39 397,764
May 20,153 5.8332 55.3 2.83 399,348
June 19,512 5.8077 62.4 1.30 401,949
July 19,892 5.6713 62.9 0.83 404,866
August 20,326 5.4977 63.5 1.53 405,341
September 19,378 5.2989 60.9 0.32 407,479
October 21,263 5.6028 51.9 3.44 405,430
November 21,443 5.9143 46.2 7.24 412,134
December 23,366 5.0000 41.8 4.72 415,342
1990 January 23,836 6.1917 41.8 9.55 416,255
February 23,336 6.3775 38.9 5.73 423,001
March 22,003 5.7234 46.3 3.40 428,559
April 20,155 4.7792 51.7 2.91 431,429
May 20,070 4.5715 54.9 2.15 434,675
June 19,588 4.3899 59.8 3.55 435,864
July 20,804 4.2559 66.7 0.59 437,969
August 19,644 3.9359 66.4 1.33 440,565
September 17,424 3.9048 61.9 0.24 441,936
October 20,833 4.4294 50.4 1.17 448,595
November 22,490 5.1523 45.8 10.66 446,291
December 24,861 5.5102 33.9 7.93 446,455
1991 January 23,441 6.8901 37.9 4.40 445,392
February 19,205 7.0308 46.9 5.42 445,787
March 20,386 6.7186 43.4 4.35 445,746
April 19,988 6.1280 49.1 5.69 446,430
 CHAPTER 8 Regression with Time Series Data 391

TABLE 8-21 Data for AAA Washington

(Continued)
Year Month Calls Unemployment Rate Temperature Rainfall Members
May 19,077 5.8146 54.3 2.12 450,001
June 19,141 5.9480 58.2 1.61 452,303
July 20,883 5.9026 65.4 0.51 456,551
August 20,709 5.7227 66.0 2.80 455,747
September 19,647 5.6877 60.9 0.20 456,764
October 22,013 6.2922 51.0 1.70 462,340
November 22,375 7.0615 46.2 6.50 460,492
December 22,727 7.4370 42.4 3.45 465,361
1992 January 22,367 8.4513 43.0 7.26 465,492
February 21,155 8.7699 46.0 3.59 466,775
March 21,209 8.0728 48.9 1.47 467,168
April 19,286 7.2392 52.7 4.35 464,575
May 19,725 7.0461 58.3 0.60 459,019
June 20,276 7.0478 63.6 1.84 463,665
July 20,795 7.1080 64.9 1.41 463,775
August 21,126 6.7824 65.0 1.01 466,230
September 20,251 6.7691 58.4 2.16 —
October 22,069 7.5896 53.2 2.55 —
November 23,268 7.9908 44.8 6.23 —
December 26,039 8.2460 37.8 4.38 —
1993 January 26,127 9.5301 34.9 4.08 —
February 20,067 9.2790 — — —
March 19,673 8.6802 — — —
April 19,142 7.7815 — — —
May — 7.4338 — — —
June — 7.3701 — — —
July — 7.2442 — — —
-- = Data not available

QUESTIONS
1. Using the data in Table 8-21 and the indepen- try running a regression without these indepen-
dent variables described in Case 7-2, which dent variables. Try experimenting with different
regression model is the best for prediction? Are powers for the exponential transformation.
the signs on the coefficients for the independent 3. Prepare a memo to Michael recommending the
variables what you would expect them to be? regression model you believe is more appropri-
Are the coefficients of the independent vari- ate for predicting the cyclical nature of emer-
ables significantly different from zero? gency road service call volume.
2. Is serial correlation a problem? If any coeffi-
cients are not significantly different from zero,
392 CHAPTER 8 Regression with Time Series Data

CASE 8-7 ALOMEGA FOOD STORES

In Example 1.1, Julie Ruth collected monthly sales paper advertising (Papers), TV advertising (TV),
data on her company along with several other vari- and eleven 0–1 dummy variables, with December
ables she thought might be related to sales. Her fore- serving as the base month, to account for the
casting efforts using these data along with Minitab monthly seasonal effects. Some of the Minitab
are detailed in Cases 2-3, 3-4, and 5-6. Her desire to output from Julie’s regression run is shown in
improve the accuracy of her forecasts suggested the Table 8-22.
use of multiple regression. Julie was quite pleased to see an R2 value of
Julie recognized from her decomposition analy- about 91%, a significant improvement over the r 2
sis of sales that her data had a strong seasonal value for the regression model containing the sin-
component (see Case 5-6). Since she also felt that gle predictor TV (review the regression results in
sales were influenced by the amounts spent on Case 2-3). Moreover, the variance inflation factor
newspaper and TV advertising (see Case 2-3), she (VIF) for each of the predictor variables was small,
decided to run a regression relating sales to news- and the Durbin-Watson statistic was close to 2.

TABLE 8-22 Minitab Output for Julie Ruth’s Regression Model for
Alomega Sales
Regression Analysis: Sales versus Paper, TV, . . .
The regression equation is
Sales = 184393 + 0.363 Paper + 0.315 TV + 200847 Duml + 55491 Dum2
+ 199556 Dum3 + 100151 Dum4 + 190293 Dum5 + 135441 Dum6
+ 156609 Dum7 + 51586 Dum8 + 183619 Dum9 + 109096 Dum10
+ 96206 Dum11
Predictor Coef SE Coef T P VIF
Constant 184393 23402 7.88 0.000
Paper 0.36319 0.06854 5.30 0.000 2.1
TV 0.31530 0.03638 8.67 0.000 1.6
Dum1 200847 39151 5.13 0.000 2.7
Dum2 55491 32399 1.71 0.096 1.8
Dum3 199556 34147 5.84 0.000 2.0
Dum4 100151 32388 3.09 0.004 1.8
Dum5 190293 32822 5.80 0.000 1.9
Dum6 135441 32581 4.16 0.000 1.9
Dum7 156609 32699 4.79 0.000 1.9
Dum8 51586 32420 1.59 0.121 1.8
Dum9 183619 36522 5.03 0.000 2.3
Dum10 109096 32439 3.36 0.002 1.8
Dum11 96206 32417 2.97 0.005 1.8
S = 45801.4. R - Sq = 90.8% R - Sq1adj2 = 87.3%
Analysis of Variance
Source DF SS MS F P
Regression 13 7.06412E + 11 54339402344 25.90 0.000
Residual Error 34 71324031968 2097765646
Total 47 7.77736E + 11
Durbin-Watson statistic = 2.27
 CHAPTER 8 Regression with Time Series Data 393
Julie realized an examination of the residuals these variables could then be chosen in advance to
was required before she could be completely pleased generate forecasts of future sales.
with her results, and she intended to do a complete Julie was almost ready to confront her tormen-
residual analysis before proposing her regression tor, Jackson Tilson, but she had a few more loose
model as a useful tool for forecasting sales. ends to wrap up.
Julie liked the fact that the predictors Papers
and TV were under the company control. Values of

QUESTIONS
1. Julie has collected data on other variables that forecasting model? How would you “sell” the
were not included in her multiple regression model to management (and Jackson Tilson)?
model. Should one or more of these other vari- 3. How might Julie’s model be used to determine
ables be included in her model? More generally, future amounts spent on newspaper and TV
how can Julie be sure she has the “right” set of advertising?
predictor variables? 4. What conditions might prompt Julie to reexam-
2. Assuming there are no additional important ine her regression model or, perhaps, to look for
predictor variables, are you satisfied with Julie’s another method of forecasting sales?

CASE 8-8 SURTIDO COOKIES

Jame Luna’s efforts to understand the trend and sea- February sales for predicting current February sales,
sonality in Surtido Cookies monthly sales have been and so forth.
examined in Cases 3-5, 4-8, and 5-7. The fact that Coding the first month as 1, the second month
cookie sales are seasonal is now well established. as 2, and so forth, Jame defines a time trend variable.
Jame and his team have tried a smoothing proce- He then constructs a series of dummy variables with
dure, decomposition, and now are interested in try- December as the base month. Finally, Jame fits a
ing regression methods in their continuing attempt regression model with a time trend and monthly
to come up with a procedure that produces the dummy variables to his sales data (see Table 3-12).
“best” forecasts of future sales. Jame is not sure how The Minitab output is shown in Table 8-23. Jan is the
to formulate the regression model(s). He has not dummy variable representing January sales, with
identified potential predictor variables for sales but similar definitions for the remaining months. After
thinks all is not lost because the seasonal pattern is examining the residuals, Jame is quite pleased with
well established and a model with dummy variables the results of his regression analysis. He is ready to
representing different months may work. In addi- consider another regression model, using sales
tion, the previous January sales may be useful for lagged an appropriate number of time periods as a
predicting current January sales, the previous predictor variable, but he needs your help.

QUESTIONS
1. All the coefficients of the dummy variables in 3. Using Jame’s fitted model in Table 8-23, gener-
Jame’s regression are negative except that for ate forecasts for the remaining seven months of
Nov. Does this make sense? Explain. 2003.
2. Are you happy with Jame’s regression model? 4. Fit an autoregressive model to Jame’s data with
What changes would you make, if any? sales lagged 12 months as the predictor variable.
394 CHAPTER 8 Regression with Time Series Data

Is this model reasonable? Generate forecasts 5. Which model, the dummy variable regression or
for the remaining seven months of 2003 using the autoregression, do you prefer? Why?
your autoregressive model.

TABLE 8-23 Minitab Output for Jame Luna’s Regression Model

for Surtido Cookie Sales
Regression Analysis: SurtidoSales versus Time, Jan, ...
The regression equation is
SurtidoSales = 1703762 + 4759 Time - 1015200 Jan - 1259183 Feb -
1286031 Mar - 1126296 Apr - 1188322 May - 1170518 Jun -
1108576 Jul - 1176156 Aug - 355847 Sep - 178199 Oct +
225028 Nov
Predictor Coef SE Coef T P VIF
Constant 1703762 107455 15.86 0.000
Time 4759 2172 2.19 0.037 1.0
Jan -1015200 124780 -8.14 0.000 2.1
Feb -1259183 124610 -10.11 0.000 2.1
Mar -1286031 124477 -10.33 0.000 2.1
Apr -1126296 124383 -9.06 0.000 2.1
May -1188322 124326 -9.56 0.000 2.1
Jun -1170518 133527 -8.77 0.000 1.9
Jul -1108576 133332 -8.31 0.000 1.9
Aug -1176156 133173 -8.83 0.000 1.9
Sep -355847 133049 -2.67 0.012 1.9
Oct -178199 132960 -1.34 0.191 1.9
Nov 225028 132907 1.69 0.102 1.9
S = 162756 R - Sq = 93.8% R - Sq1adj2 = 91.1%
Analysis of Variance
Source DF SS MS F P
Regression 12 1.11783E + 13 9.31528E + 11 35.17 0.000
Residual Error 28 7.41703E + 11 26489395582
Total 40 1.19200E + 13
Durbin-Watson statistic = 1.50

CASE 8-9 SOUTHWEST MEDICAL CENTER

Mary Beasley had a fair amount of regression analysis in begin September FY 1994–1995 (see Table 4-13), she
her Executive M.B.A.program and decided a regression created 11 dummy variables, using August as the base
model might represent her data well. She knew total month. She then ran a regression with billable visits as
billable visits to Medical Oncology were trending the dependent variable and the 11 dummy variables
upward and perhaps had a seasonal component. Mary plus time (coded 1, 2, 3, . . .) as the independent vari-
was familiar with using dummy variables to represent ables. After looking at her Minitab output, including a
seasonality in a regression context. Since Mary’s thorough analysis of the residuals, Mary was disap-
monthly data are organized on a fiscal year basis and pointed. Maybe regression was not the way to go.
 CHAPTER 8 Regression with Time Series Data 395

QUESTIONS
1. Repeat Mary’s regression analysis. Did Mary and/or suggesting another regression model she
have a right to be disappointed? Explain. might try.
2. Write Mary a brief report describing how she
might improve her original regression model

Minitab Applications

The problem. In Example 8.8, James Brown is trying to forecast electrical usage for resi-
dential customers for the third and fourth quarters of 1996 for Washington Water Power.

Minitab Solution
1. Enter the variable Hours from Table 8-9 in column C1. Enter Time in column C2
(1, 2, 3, . . . , 66). Enter the dummy variables shown in Table 8-10 to columns C3, C4,
and C5.
2. In order to run the seasonal regression model, click on the following menus:
Stat>Regression>Regression

3. The Regression dialog box similar to the one shown in Figure 6-18 appears.
a. Hours is selected as the Response or dependent variable.
b. Time-4th Qt. are entered as the Predictor or independent variables.
c. Click on Options to obtain the Regression-Options dialog box.
4. The Regression-Options dialog box shown in Figure 8-16 appears.

FIGURE 8-16 Minitab Regression-Options Dialog Box

396 CHAPTER 8 Regression with Time Series Data

a. Click on the Durbin-Watson statistic box.

b. In the space under Prediction intervals for new observations:, type the new
observations for third quarter: 67 0 1 0.
c. Click on OK and then on OK on the Regression dialog box. The results dis-
played in Table 8-11 are presented in the session window.
Note that in Example 8.4 Fred Gardner developed a regression model through the
origin (there was no Y-intercept). In order to run this type of model, click off the
check mark in the Fit intercept box shown in Figure 8-16.

Excel Applications

The problem. In Example 8.6, a first-order autoregressive model was developed for the
Novak Corporation sales data.

Excel Solution

1. Enter the Minitab file that contains the data shown in Table 8-2, and highlight the
Sales column. Click on the following menus:

Edit>Copy Cells

Now enter your Excel spreadsheet, highlight A3, and click on the following menus:

Edit>Paste

The data for Sales appear in column A. After the heading for this column is
entered in the first row, the spreadsheet looks like Figure 8-17 without the lagged
variable.
2. In order to create the Sales variable lagged one period, position the mouse at A3
and highlight through A18. Click on the following menus:

Edit>Copy

Now highlight B4 and click on

Edit>Paste

The result appears in Figure 8-17.

3. Click on the following menus to run the autoregressive model:

Tools>Data Analysis

The Data Analysis dialog box appears. Under Analysis Tools, choose Regression.
The Regression dialog box was shown in Figure 6-22.
a. Enter A4:A19 in the Input Y Range.
b. Enter B4:B19 in the Input X Range.
c. Click on the button next to New Worksheet Ply.
d. Click on OK, and the output appears as displayed in Figure 8-18.
FIGURE 8-17 Excel Spreadsheet After Lagging the Variable Sales One Period

FIGURE 8-18 Excel Output for an Autoregressive Model

397
398 CHAPTER 8 Regression with Time Series Data

References
Diebold, F. X. Elements of Forecasting, 3rd ed. Murray, M. P. “A Drunk and Her Dog: An
Cincinnati, Ohio: South-Western, 2004. Illustration of Cointegration and Error
Durbin, J., and G. S. Watson. “Testing for Serial Correction.” American Statistician 48 (1) (1994):
Correlation in Least Squares Regression II.” 37–39.
Biometrika 38 (1951): 159–178. Newbold, P., and T. Bos. Introductory Business and
Engle, R. F., and C. W. J. Granger. “Co-Integration Economic Forecasting, 2nd ed. Cincinnati, Ohio:
and Error Correction: Representation, Estimation South-Western, 1994.
and Testing.” Econometrica 55 (1987): 251–276. Pindyck, R. S., and D. L. Rubinfeld. Econometric
Granger, C. W. J., and P. Newbold. “Spurious Models and Economic Forecasts, 4th ed. New
Regressions in Econometrics.” Journal of York: McGraw-Hill, 1998.
Econometrics 2 (1974): 111–120. Young, R. M. “Forecasting with an Econometric
Levenbach, H., and J. P. Cleary. Forecasting: Practice Model: The Issue of Judgemental Adjustment.”
and Process for Demand Management. Belmont, Journal of Forecasting 1 (2) (1982): 189–204.
Calif.: Thomson Brooks/Cole, 2006.
C H A P T E R

9 THE BOX-JENKINS (ARIMA)

METHODOLOGY
Several approaches for analyzing and forecasting time series have been discussed.
In Chapter 4, smoothing was introduced, and time series forecasts were generated by a
particular smoothing (averaging) mechanism. In Chapter 5, time series were decom-
posed into trend, seasonal, and irregular components, and forecasts were produced by
extrapolating the estimated trend and seasonality. In Chapter 8, regression models
appropriate for time series data were considered. Forecasts of the dependent variable, Y,
generated from these models ordinarily require forecasts of future values of the inde-
pendent variables, the X’s.
This chapter introduces a class of models that can produce accurate forecasts based
on a description of historical patterns in the data. Autoregressive integrated moving
average (ARIMA) models are a class of linear models that are capable of representing
stationary as well as nonstationary time series. Recall that stationary processes vary
about a fixed level and that nonstationary processes have no natural constant mean
level. The autoregressive models discussed in Chapter 8 are actually a special subset of
ARIMA models that are useful for modeling stationary time series.
ARIMA models do not involve independent variables in their construction. Rather,
they make use of the information in the series itself to generate forecasts. For example,
an ARIMA model for monthly sales would project the historical sales pattern to produce
a forecast of next month’s sales.
ARIMA models rely heavily on autocorrelation patterns in the data. The method-
ology for identifying, fitting, and checking appropriate ARIMA models was greatly
advanced by the work of two statisticians, G. E. P. Box and G. M. Jenkins. For this reason,
ARIMA modeling and forecasting is often referred to as the Box-Jenkins methodology.

BOX-JENKINS METHODOLOGY
The Box-Jenkins methodology of forecasting is different from most methods because it
does not assume any particular pattern in the historical data of the series to be forecast.
It uses an iterative approach of identifying a possible model from a general class of
models. The chosen model is then checked against the historical data to see whether it
accurately describes the series. The model fits well if the residuals are generally small
and randomly distributed and contain no useful information. If the specified model is
not satisfactory, the process is repeated using a new model designed to improve on the
original one. This iterative procedure continues until a satisfactory model is found.
At that point, the model can be used for forecasting. Figure 9-1 illustrates the Box-
Jenkins model-building strategy.

399
400 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Postulate General Class

of Models

Identify Model to Be
Tentatively Entertained

Estimate Parameters in
Tentatively Entertained
Model

Diagnostic Checking
(Is the model adequate?)

No Source: Box, G. E. P., Jenkins, G. M.,

Yes
and G. C. Reinsel, Time Series
Analysis: Forecasting and Control
Use Model for (3rd ed.), Upper Saddle River,
Forecasting New Jersey: Prentice Hall, 1994,
p.17. Reprinted with permission.

FIGURE 9-1 Flow Diagram for the Box-Jenkins

Model-Building Strategy

The initial selection of an ARIMA model is based on an examination of a plot of

the time series (to observe its general character) and an examination of its autocorrela-
tions for several time lags. Specifically, the pattern of sample autocorrelations calculated
from the time series is matched with the known autocorrelation pattern associated with
a particular ARIMA model. This matching is done for both autocorrelations and partial
autocorrelations.1 Theoretical autocorrelation coefficients for some of the more com-
mon ARIMA models are shown in Figures 9-2, 9-3, and 9-4. The significance of the pat-
terns shown in these figures will be evident later in the chapter.

The Box-Jenkins methodology refers to a set of procedures for identifying, fitting,

and checking ARIMA models with time series data. Forecasts follow directly
from the form of the fitted model.

In selecting a model, remember that the autocorrelations calculated from the data
will not exactly match any set of theoretical autocorrelations associated with an
ARIMA model. Autocorrelations calculated from the data are subject to sampling
variation. However, you should be able to adequately match most time series data with
an ARIMA model. If the initial selection is not quite right, inadequacies will show up

1A partial autocorrelation at time lag k is the correlation between Yt and Yt - k—the responses for periods t
and t - k, respectively—after adjusting for the effects of the intervening values, Yt - 1, Yt - 2, Á , Yt - k + 1. Partial
autocorrelations are derived and discussed by Newbold and Bos (1994) and by Box, Jenkins, and Reinsel
(1994).
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 401

Autocorrelation Partial Autocorrelation

1 1

0 k 0 k

AR(1): Yt = φ0 + φ1Yt −1 + εt
−1 −1
(a)

1 1

0 k 0 k

−1 −1
(b)

1 1

0 k 0 k
AR(2): Yt = φ0 + φ1Yt − 1 + φ2Yt − 2 + εt

−1 −1
(c)

1 1

0 k 0 k

−1 −1
(d)

FIGURE 9-2 Autocorrelation and Partial Autocorrelation Coefficients

of AR(1) and AR(2) Models
402 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Autocorrelation Partial Autocorrelation

1 1

0 k 0 k

MA(1): Yt = μ + εt − ω1εt −1
−1 −1
(a)

1 1

0 k 0 k

−1 −1
(b)

1 1

0 k 0 k
MA(2): Yt = μ + εt − ω1εt − 1 − ω2 εt − 2

−1 −1
(c)

1 1

0 k 0 k

−1 −1
(d) (d)

FIGURE 9-3 Autocorrelation and Partial Autocorrelation Coefficients

of MA(1) and MA(2) Models
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 403

Autocorrelation Partial Autocorrelation

1 1

0 k 0 k

−1 −1
(a)

1 1
ARMA(1,1): Yt = φ0 + φ1Yt −1 + εt − ω1εt −1

0 k 0 k

−1 −1
(b)

1 1

0 k 0 k

−1 −1
(c)

1 1

0 k 0 k

−1 −1
(d)

FIGURE 9-4 Autocorrelation and Partial Autocorrelation Coefficients

of a Mixed ARMA(1, 1) Model
404 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

in an analysis of the residuals (model checking), and the original model can be modi-
fied. As you gain experience, this iterative model-building task becomes much easier.

Autoregressive Models
A first-order autoregressive model was introduced in Chapter 8. A pth-order autore-
gressive model takes the form

Yt = f0 + f1Yt - 1 + f2Yt - 2 + Á + fpYt - p + t (9.1)

where
Yt = the response (dependent) variable at time t
Yt - 1, Yt - 2, Á , Yt - p = the response variable at time lags t - 1, t - 2, Á , t - p,
respectively; these Y’s play the role of independent variables
f0, f1, f2, Á , fp = the coefficients to be estimated2
t = the error term at time t, which represents the effects of variables
not explained by the model; the assumptions about the error
term are the same as those for the standard regression model
The model in Equation 9.1 has the appearance of a regression model with lagged val-
ues of the dependent variable in the independent variable positions, hence the name
autoregressive model. Autoregressive models are appropriate for stationary time series,
and the coefficient f0 is related to the constant level of the series. If the data vary about
zero or are expressed as deviations from the mean, Yt - Y, the coefficient f0 is not
required.
The equations of an AR model of order 1, or AR(1) model, and an AR model of
order 2, or AR(2) model, are shown in Figure 9-2. Figures 9-2(a) and (b) illustrate the
behavior of the theoretical autocorrelation and partial autocorrelation functions,
respectively, for an AR(1) model. Notice how differently the autocorrelation and par-
tial autocorrelation functions behave. The autocorrelation coefficients trail off to zero
gradually, whereas the partial autocorrelation coefficients drop to zero after the first
time lag. Figures 9-2(c) and (d) show the autocorrelations and partial autocorrelations,
respectively, for an AR(2) model. Again, the autocorrelation coefficients trail off to
zero, whereas the partial autocorrelation coefficients drop to zero after the second
time lag. This type of pattern will generally hold for any AR( p) model. It must be
remembered that sample autocorrelation functions are going to differ from these the-
oretical functions because of sampling variation.

Example 9.1
Forecasting with an AR(2) model is demonstrated using the data set consisting of 75 process
readings shown in Table 9-5. Only the last five observations in Table 9-5 will be used to con-
struct the forecasts shown in Table 9-1.
An AR(2) model is chosen for the process data, and the Minitab program computes the
least squares estimates fN 0 = 115.2, fN 1 = - .535, and fN 2 = .0055 . Suppose that at time
t -1 = 75 a forecast of the observation for the next period, t = 76, is required. Since the best
guess of the error term is its mean value zero, the forecast for period t = 76 is

YNt = fN0 + fN1Yt - 1 + fN 2Yt - 2

N
Y76 = 115.2 - .535Y75 + .0055Y74
YN76 = 115.2 - .5351722 + .00551992 = 77.2

2The coefficient f0 is related to the mean of the process, μ, by f0 = m11 - f1 - f2 - Á - fp2.

 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 405

TABLE 9-1 Forecasting with an AR(2) Autoregressive

Model for Example 9.1

Time Values Forecasts Residuals

Period t Yt YN t et
t -5 71 90 — —
t -4 72 78 — —
t -3 73 87 74.0 13.0
t -2 74 99 69.1 29.9
t -1 75 72 62.7 9.3
t 76 77.2

For autoregressive models, forecasts depend on observed values in previous time peri-
ods. For AR(2) models, forecasts of the next value depend on observed values in the two
previous time periods. For AR(3) models, forecasts of the next value depend on observed
values in the three previous time periods, and so forth.

Moving Average Models

A qth-order moving average model takes the form
Yt = m + t - v1 t - 1 - v2 t - 2 - Á - vq t - q (9.2)

where
Yt = the response (dependent) variable at time t
m = the constant mean of the process
v1, v2, Á , vq = the coefficients to be estimated
t = the error term, which represents the effects of variables not
explained by the model; the assumptions about the error term
are the same as those for the standard regression model
t - 1, t - 2, Á , t - q = the errors in previous time periods that, at time t, are
incorporated in the response, Yt
Equation 9.2 is similar to Equation 9.1 except that the dependent variable, Yt, depends
on previous values of the errors rather than on the variable itself. Moving average
(MA) models provide forecasts of Yt based on a linear combination of a finite number
of past errors, whereas autoregressive (AR) models forecast Yt as a linear function of a
finite number of past values of Yt.
The term moving average for the model in Equation 9.2 is historical and should not
be confused with the moving average procedures discussed in Chapter 4. Here, moving
average refers to the fact that the deviation of the response from its mean, Yt - m, is a
linear combination of current and past errors and that, as time moves forward, the
errors involved in this linear combination move forward as well.

Yt - m = t - v1 t - 1 - v2 t - 2 - Á - vq t - q
Yt + 1 - m = t + 1 - v1 t - v2 t - 1 - Á - vq t - q + 1

The weights v1, v2, Á , vq do not necessarily sum to 1 and may be positive or negative,
although they are each preceded by a minus sign in the specification of the model.
Figure 9-3 shows the equations of a moving average model of order 1, or MA(1)
model, and a moving average model of order 2, or MA(2) model. Error terms can be
added sequentially to get an MA(q) model where q is the number of past error terms
406 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

to be included in the forecast of the next observation. Figures 9-3(a) and (b) also
illustrate the behavior of the theoretical autocorrelation and partial autocorrelation
coefficients for the MA(1) model. Note that the autocorrelation and partial autocorre-
lation functions of AR and MA models behave very differently. This is fortunate
because AR and MA models can then be distinguished by their different autocorrela-
tion and partial autocorrelation patterns.
The autocorrelation coefficients for the MA(1) model drop to zero after the first
time lag, whereas the partial autocorrelation coefficients trail off to zero gradually.
Furthermore, the autocorrelation coefficients for the MA(2) model are zero after the
second time lag, whereas the partial autocorrelations trail off gradually (see Figures 9-3(c)
and (d)). Again, it must be mentioned that sample autocorrelation functions are going
to differ from these theoretical functions because of sampling variation.

Example 9.2
The 75 process readings given in Table 9-5 can be used to demonstrate forecasting with an
MA(2) model. Forecasting is demonstrated using only the last five observations and residu-
als shown in Table 9-2.
An MA(2) model is selected, and the Minitab program computes the least squares esti-
mates3 of the model coefficients, getting mN = 75.4, vN 1 = .5667, and vN 2 = - .3560. Again,
suppose that at time t - 1 = 75 a forecast of the observation in period t = 76 is required.
Since at time t - 1 the best guess of the error term in the next period is its mean value, zero,
and the best guesses of errors in the current and previous time periods are the correspon-
ding residuals, the forecast for period t = 76 is

YNt = mN - vN 1et - 1 - vN 2et - 2

N
Y76 = 75.4 - .5667e75 + .3560e74
YN76 = 75.4 - .566717.72 + .35601272 = 80.6

Notice that the two residuals, e75 and e74, are substituted for the errors, 75 and 74, in the
computation of the forecast for period 76. To compute forecasts from moving average mod-
els, errors corresponding to time periods that have already occurred are replaced by residu-
als for those time periods. The number of residuals involved in the forecast of the next
observation is equal to the order of the moving average model.

TABLE 9-2 Forecasting with an MA(2) Moving

Average Model for Example 9.2

Time Values Forecasts Residuals

Period t Yt YNt et
t -5 71 90 — 13.9
t -4 72 78 — 8.9
t -3 73 87 75.3 11.7
t -2 74 99 72.0 27.0
t -1 75 72 64.3 7.7
t 76 80.6

3Least squares estimates for coefficients in moving average models, or models involving moving average
terms, must be obtained iteratively, using a nonlinear least squares algorithm. Using an initial starting point,
a nonlinear least squares algorithm generates improved coefficient estimates that have a smaller sum of
squared errors. Estimates are continually improved until the sum of squared errors cannot be made appre-
ciably smaller. Although autoregressive models can be fit with standard regression packages, least squares
estimates of the coefficients in autoregressive models are often obtained using a nonlinear least squares
procedure.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 407

Autoregressive Moving Average Models

A model with autoregressive terms can be combined with a model having moving
average terms to get a “mixed” autoregressive–moving average model. It is convenient
to use the notation ARMA(p, q), where p is the order of the autoregressive part and q
is the order of the moving average part, to represent these models. An ARMA(p, q)
model has the general form

Yt = f0 + f1Yt - 1 + f2Yt - 2 + Á + fpYt - p + t - v1 t - 1 - v2 t - 2 - Á - vq t - q

(9.3)
ARMA( p, q) models can describe a wide variety of behaviors for stationary time
series.4 Forecasts generated by an ARMA(p, q) model will depend on current and past
values of the response, Y, as well as current and past values of the errors (residuals), e.
Figure 9-4 shows the equation of an ARMA(1, 1) model and the possible behav-
iors of the theoretical autocorrelations and partial autocorrelations. In this case, both
the autocorrelations and the partial autocorrelations die out; neither cuts off.

Summary
The autocorrelation and partial autocorrelation patterns for autoregressive–moving
average processes can be summarized as follows:
Autocorrelations Partial Autocorrelations
MA(q) Cut off after the order Die out
q of the process

AR(p) Die out Cut off after the order

p of the process

ARMA(p, q) Die out Die out

The numbers of autoregressive and moving average terms (orders p and q) in an

ARMA model are determined from the patterns of the sample autocorrelations and
partial autocorrelations and the values of the model selection criteria that are discussed
in a later section of this chapter. In practice, the values of p and q each rarely exceed 2.

IMPLEMENTING THE MODEL-BUILDING STRATEGY

As shown in Figure 9-1, the Box-Jenkins approach uses an iterative model-building
strategy that consists of selecting an initial model (model identification), estimating the
model coefficients (parameter estimation), and analyzing the residuals (model check-
ing). If necessary, the initial model is modified and the process is repeated until the
residuals indicate no further modification is necessary. At this point, the fitted model
can be used for forecasting.
The steps in the model-building strategy are examined in some detail next.

Step 1: Model Identification

1. The first step in model identification is to determine whether the series is stationary—
that is, whether the time series appears to vary about a fixed level. It is useful to look
at a plot of the series along with the sample autocorrelation function.A nonstationary

4Notice that when q = 0 , the ARMA(p, 0) model reduces to a pure autoregressive model of order p.
Similarly, when p = 0, the ARMA(0, q) model is a pure moving average model of order q.
408 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

time series is indicated if the series appears to grow or decline over time and the sam-
ple autocorrelations fail to die out rapidly. The time series pictured in Figure 8-2 are
nonstationary, and the pattern of sample autocorrelations shown in Figure 8-3 is typi-
cal for a nonstationary series.
If the series is not stationary, it can often be converted to a stationary series
by differencing. That is, the original series is replaced by a series of differences. An
ARMA model is then specified for the differenced series. In effect, the analyst is
modeling changes rather than levels.
As an example, suppose that the original series, Yt, is generally increasing over
time but that the first differences, ¢Yt = Yt - Yt - 1, vary about a fixed level. It may
be appropriate to model the stationary differences using an ARMA model of, say,
order p = 1 and q = 1. In this case, the model is5
¢Yt = f1 ¢Yt - 1 + t - v1 t - 1
or
1Yt - Yt - 12 = f11Yt - 1 - Yt - 22 + t - v1 t - 1
In some cases, it may be necessary to difference the differences before stationary
data are obtained. Simple differencing is done twice, and the stationary data are
¢ 2Yt = ¢1¢Yt2 = ¢1Yt - Yt - 12 = Yt - 2Yt - 1 + Yt - 2
Differencing is done until a plot of the data indicates that the series varies about a
fixed level and the sample autocorrelations die out fairly rapidly. The number of dif-
ferences required to achieve stationarity is denoted by d.
Models for nonstationary series are called autoregressive integrated moving
average models and denoted by ARIMA(p, d, q).6 Here, p indicates the order of
the autoregressive part, d indicates the amount of differencing, and q indicates the
order of the moving average part. If the original series is stationary, then d = 0,
and the ARIMA models reduce to the ARMA models. Consequently, from this
point on, the ARIMA(p, d, q) notation is used to indicate models for both station-
ary (d = 0) and nonstationary (d 7 0) time series.
Although ARIMA models involve differences, forecasts for the original series
can always be computed directly from the fitted model.
2. Once a stationary series has been obtained, the analyst must identify the form of
the model to be used.
Identifying the model form is accomplished by comparing the autocorrelations
and partial autocorrelations computed from the data to the theoretical autocorrela-
tions and partial autocorrelations for the various ARIMA models. The theoretical
correlations for some of the more common ARIMA models are shown in Figures 9-2,
9-3, and 9-4 to help in selecting an appropriate model.
Each ARIMA model has a unique set of autocorrelations and partial autocor-
relations, and the analyst should be able to match the corresponding sample values
to one of the theoretical patterns.
There may be some ambiguity in determining an appropriate ARIMA model
from the patterns of the sample autocorrelations and partial autocorrelations. Thus,
the initial model selection should be regarded as tentative. Analyses can be done
during steps 2 and 3 to determine if the model is adequate. If not, an alternative

5When an ARMA model is used for a differenced series, the constant term, f0, may not be required.
6The term integrated means that the differences must be summed (or integrated) to get the original series.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 409
model can be tried. With a little practice, the analyst should become more adept at
identifying an adequate model.
Remember that, if the sample autocorrelations die out exponentially to zero
and the sample partial autocorrelations cut off, the model will require autoregres-
sive terms. If the sample autocorrelations cut off and the sample partial autocorre-
lations die out, the model will require moving average terms. If both the sample
autocorrelations and the sample partial autocorrelations die out, both the autore-
gressive and the moving average terms are indicated. By counting the number of
significant sample autocorrelations and partial autocorrelations, the orders of the
MA and AR parts can be determined. To judge their significance, both the sample
autocorrelations and the sample partial autocorrelations are usually compared
with ; 2> 2n where n is the number of observations in the time series. These limits
work well when n is large.
All things being equal, simple models are preferred to complex models. This is
known as the principle of parsimony. With a limited amount of data, it is relatively
easy to find a model with a large number of parameters that fits the data well.
However, forecasts from such a model are likely to be poor because much of the
variation in the data is due to the random error that is being modeled. The goal is
to develop the simplest model that provides an adequate description of the major
features of the data.

The principle of parsimony refers to the preference for simple models over
complex ones.

Step 2: Model Estimation

1. Once a tentative model has been selected, the parameters for that model must be
estimated.
The parameters in ARIMA models are estimated by minimizing the sum of
squares of the fitting errors. These least squares estimates must, in general, be
obtained using a nonlinear least squares procedure. A nonlinear least squares pro-
cedure is simply an algorithm that finds the minimum of the sum of squared errors
function. Once the least squares estimates and their standard errors are deter-
mined, t values can be constructed and interpreted in the usual way. Parameters
that are judged significantly different from zero are retained in the fitted model;
parameters that are not significant are dropped from the model.
For example, suppose that an ARIMA(1, 0, 1) has been fit to a time series of
100 observations and the fitted equation is

YNt = 33.4 + .25Yt - 1 - .5 N t-1

(7.02) (.17) (.21)

where the numbers in parentheses under the estimated coefficients are their stan-
dard errors. Since the t ratio for the coefficient of the autoregressive term is
t = .25>.17 = 1.47 with a p-value of .14, the hypothesis H0 : f1 = 0 is not rejected,
and this term could be deleted from the model. An ARIMA(0, 0, 1) model—that is,
an MA(1) model—could then be fit to the data.
2. The residual mean square error, an estimate of the variance of the error, t, is
computed.
410 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

The residual mean square error is defined as7

n n

a 1Yt - Yt2
2 N 2
a et
t=1 t=1
s2 = = (9.4)
n - r n - r
where
et = Yt - YNt = the residual at time t
n = the number of residuals
r = the total number of parameters estimated
3. The residual mean square error is useful for assessing fit and comparing different
models. It is also used to calculate forecast error limits.

Step 3: Model Checking

Before using the model for forecasting, it must be checked for adequacy. Basically, a
model is adequate if the residuals cannot be used to improve the forecasts. That is, the
residuals should be random.
1. Many of the same residual plots that are useful in regression analysis can be devel-
oped for the residuals from an ARIMA model. A histogram and normal probabil-
ity plot (to check for normality) and a time sequence plot (to check for outliers)
are particularly helpful.
2. The individual residual autocorrelations should be small and generally within
; 2> 2n of zero. Significant residual autocorrelations at low lags or seasonal lags
suggest that the model is inadequate and that a new or modified model should be
selected.
3. The residual autocorrelations as a group should be consistent with those produced
by random errors.
An overall check of model adequacy is provided by a chi-square (2) test based on
the Ljung-Box Q statistic. This test looks at the sizes of the residual autocorrelations as
a group. The test statistic Q is
m r 2 1e2
k
Q = n(n + 2) a (9.5)
k=1 n - k

which is approximately distributed as a chi-square random variable with m - r

degrees of freedom where r is the total number of parameters estimated in the
ARIMA model. In Equation 9.5,

rk1e2 = the residual autocorrelation at lag k

n = the number of residuals
k = the time lag
m = the number of time lags to be tested

If the p-value associated with the Q statistic is small (say, p-value 6 .05), the model is
considered inadequate. The analyst should consider a new or modified model and con-
tinue the analysis until a satisfactory model has been determined.
Judgment plays a large role in the model-building effort. Two simple competing
models may adequately describe the data, and a choice may be made on the basis of the

7The square root of s 2, s = 2s 2, is analogous to the standard error of the estimate (see Equation 7.4).
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 411
nature of the forecasts. Also, a few large residuals may be ignored if they can be
explained by unusual circumstances and the model is adequate for the remainder of the
observations.

Step 4: Forecasting with the Model

1. Once an adequate model has been found, forecasts for one period or several periods
into the future can be made.
Prediction intervals based on the forecasts can also be constructed. In general,
for a given confidence level, the longer the forecast lead time, the larger the predic-
tion interval. This is sensible, since the uncertainty is expected to be greater for a
forecast of a distant value than it is for a forecast of, say, the next observation.
Calculating forecasts and prediction intervals is tedious and best left to the com-
puter. Computer programs that fit ARIMA models generate forecasts and predic-
tion intervals at the analyst’s request.
2. As more data become available, the same ARIMA model can be used to generate
revised forecasts from another time origin.
3. If the pattern of the series appears to be changing over time, the new data may be
used to reestimate the model parameters or, if necessary, to develop an entirely
new model.
It is a good idea to monitor the forecast errors. If the most recent errors tend to be
consistently larger than the previous errors, it may be time to reevaluate the model. At
this point, another iteration of the model-building strategy may be required. The same
holds true if the recent forecast errors tend to be consistently positive (underpredict-
ing) or negative (overpredicting).
Example 9.3
The Cameron Consulting Corporation specializes in portfolio investment services. Lynn
Stephens, the company analyst, was assigned the task of developing more-sophisticated
techniques for forecasting Dow Jones averages. Lynn recently attended a workshop on Box-
Jenkins methodology and decided to try the technique on the Dow Jones Transportation
Index. Table 9-3 represents 65 daily closing averages of the Transportation Index during the
summer months.

TABLE 9-3 Daily Closing Averages of the Transportation Index

for Example 9.3

Time Closing Time Closing Time Closing Time Closing Time Closing
Period Average Period Average Period Average Period Average Period Average

1 222.34 15 223.07 29 246.74 43 249.76 57 268.21

2 222.24 16 225.36 30 248.73 44 251.66 58 272.16
3 221.17 17 227.60 31 248.83 45 253.41 59 272.79
4 218.88 18 226.82 32 248.78 46 252.04 60 275.03
5 220.05 19 229.69 33 249.61 47 248.78 61 278.49
6 219.61 20 229.30 34 249.90 48 247.76 62 281.75
7 216.40 21 228.96 35 246.45 49 249.27 63 285.70
8 217.33 22 229.99 36 247.57 50 247.95 64 286.33
9 219.69 23 233.05 37 247.76 51 251.41 65 288.57
10 219.32 24 235.00 38 247.81 52 254.67
11 218.25 25 236.17 39 250.68 53 258.62
12 220.30 26 238.31 40 251.80 54 259.25
13 222.54 27 241.14 41 251.07 55 261.49
14 223.56 28 241.48 42 248.05 56 264.95
412 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Lynn began the analysis by looking at a plot of the data shown in Figure 9-5. There
appeared to be an upward trend in the series. Her next step in identifying a tentative model
was to look at the sample autocorrelations of the data shown in Figure 9-6. When Lynn
observed that the first several autocorrelations were persistently large and trailed off to
zero rather slowly, she recognized that her original observation was correct: This time series
was nonstationary and did not vary about a fixed level.
Lynn decided to difference the data to see if she could eliminate the trend and create a
stationary series. A plot of the differenced data (not shown) appears to vary about a fixed
level. In fact, the sample mean of the differences was 1.035. The sample autocorrelations for

FIGURE 9-5 Daily Closing Averages of the Dow Jones

Transportation Index for Example 9.3

FIGURE 9-6 Sample Autocorrelation Function for the Transportation

Index for Example 9.3
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 413
the differences are shown in Figure 9-7. The sample partial autocorrelations for the differ-
ences are displayed in Figure 9-8.
Lynn was perplexed. Comparing the autocorrelations with their error limits, the only
significant autocorrelation was at lag 1. Similarly, only the lag 1 partial autocorrelation was
significant. The autocorrelations appear to cut off after lag 1, indicating MA(1) behavior. At
the same time, the partial autocorrelations appear to cut off after lag 1, indicating AR(1)

FIGURE 9-7 Sample Autocorrelation Function for the First Differences

of the Transportation Index for Example 9.3

FIGURE 9-8 Sample Partial Autocorrelations for the First Differences

of the Transportation Index for Example 9.3
414 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

behavior. Neither pattern appears to die out in a declining manner at low lags. Lynn decided
to fit both ARIMA(1, 1, 0) and ARIMA(0, 1, 1) models to the Transportation Index. She
decided to include a constant term in each model to allow for the fact that the series of dif-
ferences appears to vary about a level greater than zero. If Yt denotes the Transportation
Index, then the differenced series is ¢Yt = Yt - Yt - 1, and Lynn’s models are

ARIMA11, 1, 02: ¢Yt = f0 + f1 ¢Yt - 1 + t

ARIMA10, 1, 12: ¢Yt = m + t - v1 t - 1

The Minitab outputs for Lynn’s models are shown in Table 9-4. The residual autocorrela-
tions for the ARIMA(1, 1, 0) model fit are shown in Figure 9-9.
Both models fit the data equally well. The residual mean squares (MS) are

ARIMA11, 1, 02: s 2 = 3.536

ARIMA10, 1, 12: s 2 = 3.538

Lynn also noticed that the constant in the ARIMA(0, 1, 1) model is estimated to be
mN = 1.0381, (essentially) the sample mean of the differences.
Figure 9-9 shows there is no significant residual autocorrelation for the ARIMA(1, 1, 0)
model. Although the residual autocorrelation function for the ARIMA(0, 1, 1) model is not
shown, the results are similar. The Ljung-Box Q statistics computed for groups of lags
m = 12, 24, 36, and 48 are not significant, as indicated by the large p-values for each model.
Lynn decided that either model is adequate. Moreover, the one-step-ahead forecasts pro-
vided by the two models are nearly the same.
To resolve her dilemma, Lynn opted for the ARIMA(1, 1, 0) model on the basis of its
slightly better fit. She checks the forecast for period 66 for this model as follows:

Yt - Yt - 1 = f0 + f11Yt - 1 - Yt - 22 + t

or

Yt = Yt - 1 + f0 + f11Yt - 1 - Yt - 22 + t

FIGURE 9-9 Residual Autocorrelations: ARIMA(1, 1, 0) Model Fit

to Transportation Index for Example 9.3
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 415

With fN 0 = .741 and fN 1 = .284, the forecasting equation becomes

YN66 = Y65 + .741 + .2841Y65 - Y642

= 288.57 + .741 + .2841288.57 - 286.232 = 289.947

The forecast agrees with the result in Table 9-4. The prediction interval for the actual value
in period 66 calculated by Minitab is (286.3, 293.6).

TABLE 9-4 Minitab Output for ARIMA(1, 1, 0) and ARIMA(0, 1, 1) Models

Fit to Transportation Index for Example 9.3

ARIMA(1, 1, 0): Model for Transportation Index

Final Estimates of Parameters

Type Coef SE Coef T P
AR 1 0.2844 0.1221 2.33 0.023
Constant 0.7408 0.2351 3.15 0.003
Differencing: 1 regular difference
Number of observations: Original series 65, after differencing 64
Residuals: SS = 219.223 (backforecasts excluded)
MS = 3.536 DF = 62
Modified Box-Pierce (Ljung-Box) Chi-Square statistic
Lag 12 24 36 48
Chi-Square 11.8 29.1 37.1 48.1
DF 10 22 34 46
P-Value 0.297 0.141 0.328 0.389

Forecasts from period 65

95% Limits
Period Forecast Lower Upper Actual
66 289.948 286.262 293.634
ARIMA(0, 1, 1): Model for Transportation Index
Final Estimates of Parameters
Type Coef SE Coef T P
MA 1 -0.2913 0.1226 -2.38 0.021
Constant 1.0381 0.3035 3.42 0.001
Differencing: 1 regular difference
Number of observations: Original series 65, after differencing 64
Residuals: SS = 219.347(backforecasts excluded)
MS = 3.538 DF = 62
Modified Box-Pierce (Ljung-Box) Chi-Square statistic
Lag 12 24 36 48
Chi-Square 11.6 32.0 41.0 51.4
DF 10 22 34 46
P-Value 0.310 0.077 0.189 0.270
Forecasts from period 65
95% Limits
Period Forecast Lower Upper Actual
66 290.053 286.366 293.740
416 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-5 Readings for the Atron Process

for Example 9.4 (read by
columns)

60.0 99.0 75.0 79.5 61.5 88.5 72.0 90.0

81.0 25.5 78.0 64.5 81.0 51.0 66.0 78.0
72.0 93.0 66.0 99.0 76.5 85.5 73.5 87.0
78.0 75.0 97.5 72.0 84.0 58.5 66.0 99.0
61.5 57.0 60.0 78.0 57.0 90.0 73.5 72.0
78.0 88.5 97.5 63.0 84.0 60.0 103.5
57.0 76.5 61.5 66.0 73.5 78.0 60.0
84.0 82.5 96.0 84.0 78.0 66.0 81.0
72.0 72.0 79.5 66.0 49.5 97.5 87.0
67.8 76.5 72.0 87.0 78.0 64.5 73.5

FIGURE 9-10 Readings for the Atron Process

Example 9.4
The analyst for Atron Corporation, Jim White, had a time series of readings for an industrial
process that needed to be forecast. The data are shown in Table 9-5. The readings are plot-
ted in Figure 9-10. Jim believed that the Box-Jenkins methodology might work best for
these data.
Jim began to identify a tentative model by looking at the plot of the data and the sam-
ple autocorrelations shown in Figure 9-11. The time series of readings appears to vary about
a fixed level of around 80, and the autocorrelations die out rapidly toward zero. Jim con-
cluded that the time series is stationary.
The first sample autocorrelation coefficient ( -.53) is significantly different from zero
at the 5% level, since it lies outside the range

1 1
0 ; 2 = 0 ; 2 = 0 ; 21.1152 = 0 ; .23
2n 275

The autocorrelation at lag 2 is close to significant at the 5% level and opposite in sign from
the lag 1 autocorrelation, r1. The remaining autocorrelations are small and well within their
individual error limits. The pattern of the autocorrelations suggests either an AR(1) model
(see Figure 9-2(b)) or perhaps an MA(2) model if the autocorrelations cut off (are zero)
after lag 2. Jim decided to examine the sample partial autocorrelations shown in Figure 9-12.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 417

FIGURE 9-11 Sample Autocorrelation Function for the Atron Readings

for Example 9.4

FIGURE 9-12 Sample Partial Autocorrelations for the Atron Readings

for Example 9.4

Jim noticed that the first partial autocorrelation ( - .53) was significantly different from
zero but none of the other partial autocorrelations approached significance.8 Jim felt the auto-
correlations and partial autocorrelations suggest an AR(1) (or, equivalently, an ARIMA(1, 0, 0))
model, but to play it safe, he decided to fit an MA(2) (ARIMA(0, 0, 2)) model as well. If both
models are adequate, he will decide the issue using the principle of parsimony.
Table 9-6 shows the results of using Minitab to fit the AR(1) and MA(2) models to the
readings for the Atron process. A constant term is included in both models to allow for the
fact that the readings vary about a level other than zero.9

8The value of the first partial autocorrelation is always equal to the value of the first autocorrelation.
9If the data had been expressed as deviations from the sample mean, a constant term would not be required
in either model.
418 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-6 Minitab Output for MA(2) and AR(1) Models for the Atron
Readings for Example 9.4

ARIMA(0, 0, 2) Model: Atron readings

Final Estimates of Parameters

Type Coef SE Coef T P
MA 1 0.5667 0.1107 5.12 0.000
MA 2 -0.3560 0.1146 -3.11 0.003
Constant 75.410 1.061 71.08 0.000
Mean 75.410 1.061
Number of observations: 75
Residuals: SS = 9724.97 (backforecasts excluded)
MS = 135.07 DF = 72
Modified Box-Pierce (Ljung-Box) Chi-Square statistic
Lag 12 24 36 48
Chi-Square 7.0 23.8 31.8 46.9
DF 9 21 33 45
P-Value 0.638 0.305 0.526 0.394

Forecasts from period 75

95% Limits
Period Forecast Lower Upper Actual
76 80.648 57.864 103.431
77 78.169 51.982 104.356
78 75.410 47.996 102.825
ARIMA (1, 0, 0) Model: Atron readings
Final Estimates of Parameters
Type Coef SE Coef T P
AR 1 -0.5379 0.0986 -5.46 0.000
Constant 115.842 1.356 85.44 0.000
Mean 75.3269 0.8817
Number of observations: 75
Residuals: SS = 10065.2 (backforecasts excluded)
MS = 137.9 DF = 73
Modified Box-Pierce (Ljung-Box) Chi-Square statistic
Lag 12 24 36 48
Chi-Square 9.3 29.9 37.3 58.3
DF 10 22 34 46
P-Value 0.503 0.121 0.319 0.105

Forecasts from period 75

95% Limits
Period Forecast Lower Upper Actual
76 77.116 54.097 100.136
77 74.364 48.227 100.502
78 75.845 48.872 102.817
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 419

FIGURE 9-13 Autocorrelation Function for the Residuals

of the AR(1) Model for Example 9.4

Both models appear to fit the data well. The estimated coefficients are significantly dif-
ferent from zero. The mean square errors are similar:

MA122: s 2 = 135.1

AR112: s 2 = 137.9

The one- and two-step-ahead forecasts from the two models are somewhat different, but the
three-step-ahead (period 78) forecasts are nearly the same. For a fixed forecast origin, fore-
casts for stationary processes will eventually equal the estimated mean level. In this case, the
mean level is estimated to be approximately mN = 75 for both models.
The Ljung-Box Q statistic is not significant for collections of lags m = 12, 24, 36, and 48
for either model.The residual autocorrelations for the AR(1) model are shown in Figure 9-13.
The individual residual autocorrelations are small and well within their error limits. The
residual autocorrelation function for the MA(2) model is similar. For each model, there is no
reason to believe the errors are not random.
Since the AR(1) model has two parameters (including the constant term) and the
MA(2) model has three parameters (including the constant term), Jim appealed to the prin-
ciple of parsimony and decided to use the simpler AR(1) model to forecast future readings.
The AR(1) forecasting equation is10

YNt = 115.842 + 1- .5382Yt - 1 = 115.842 - .538Yt - 1

so, for period 76,

YN76 = 115.842 - .538Y75 = 115.842 - .5381722 = 77.11

Moreover, the two-step-ahead forecast is

YN77 = 115.842 - .538YN76 = 115.842 - .538177.112 = 74.36

Jim was pleased to see that these results agreed with those on the Minitab output.

10The error term, t, is dropped because, for the forecast YNt, the best guess of t is zero.
420 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Example 9.5
Jim White was happy with the results of his forecasts for the time series of readings shown
in Table 9-5. He decided to use the Box-Jenkins methodology to attempt to forecast the
errors (deviations from target) resulting from the quality control of a manufacturing
process under his administration. The data are given in Table 9-7, and the time series of
errors is plotted in Figure 9-14.
Jim began the identification process by examining the plot of the error time series and
by checking the autocorrelations and partial autocorrelations shown in Figures 9-15 and
9-16. The time series plot and the autocorrelation functions indicate the series is stationary.

TABLE 9-7 Errors for Atron Quality Control

for Example 9.5 (read by columns)

.23 .20 1.93 .97 .10

.63 .21 1.87 .83 .62
.48 .91 .97 .33 2.27
.83 .36 .46 .91 .62
.03 .48 2.12 1.13 .74
1.31 .61 2.11 2.22 .16
.86 1.38 .70 .80 1.34
1.28 .04 .69 1.95 1.83
0 .90 .24 2.61 .31
.63 1.79 .34 .59 1.13
.08 .37 .60 .71 .87
1.30 .40 .15 .84 1.45
1.48 1.19 .02 .11 1.95
.28 .98 .46 1.27 .51
.79 1.51 .54 .80 .41
1.86 .90 .89 .76 .49
.07 1.56 1.07 1.58 1.54
.09 2.18 .20 .38 .96

FIGURE 9-14 Errors (Deviations from Target) for Atron

Quality Control for Example 9.5
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 421

FIGURE 9-15 Sample Autocorrelations for Atron Quality Control Errors

for Example 9.5

FIGURE 9-16 Sample Partial Autocorrelations for Atron Quality Control

Errors for Example 9.5

There is one significant autocorrelation of –.50 at lag 1 and the autocorrelations at the
remaining lags are small and well within their error limits; the sample autocorrelations
appear to cut off after lag 1. Beginning with a significant partial autocorrelation at lag 1, the
first three sample partial autocorrelations are all negative and decay toward zero. The behav-
iors of the sample autocorrelations and sample partial autocorrelations look much like the
theoretical quantities for an MA(1) (ARIMA(0, 0, 1)) process shown in Figure 9-3(a). Jim
was convinced his series could be represented by an MA(1) model.
Jim used Minitab to fit an MA(1) model to his data. The results are displayed in Table 9-8.
The parameters in the MA(1) model are estimated to be m N = .1513 and vN 1 = .5875. Each is
significantly different from zero. The residual autocorrelations shown in Figure 9-17, and the
Ljung-Box Q statistics indicate the errors are random.
The MA(1) forecasting equation is

YNt = mN - vN 1 t - 1 = .1513 - .5875 t - 1

422 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-8 Minitab Output for MA(1) Model for Atron

Quality Control for Example 9.5
ARIMA (0, 0, 1) Model: Atron Quality Control

Final Estimates of Parameters

Type Coef SE Coef T P
MA 1 0.5875 0.0864 6.80 0.000
Constant 0.15129 0.04022 3.76 0.000
Mean 0.15129 0.04022
Number of observations: 90
Residuals: SS = 74.4933 (backforecasts excluded)
MS = 0.8465 DF = 88

Modified Box-Pierce (Ljung-Box) Chi-Square statistic

Lag 12 24 36 48
Chi-Square 9.1 10.8 17.3 31.5
DF 10 22 34 46
P-Value 0.524 0.977 0.992 0.950

Forecasts from period 90

95% Limits
Period Forecast Lower Upper Actual
91 0.43350 1.37018 2.23719
92 0.15129 1.94064 2.24322

FIGURE 9-17 Autocorrelation Function for the Residuals of the MA(1)

Model for Example 9.5
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 423
where t - 1 is estimated by the corresponding residual, et - 1. To forecast the error (deviation
from target) for period 91, the residual for period 90, e90 = - .4804, is required. Jim computed

YN91 = .1513 - .58751 -.48042 = .4335

Jim’s forecast of the quality control error in period 92 is simply the estimated mean of the
series because, at forecast origin, t = 90, the best guess of the error term in period 91, 91, is
zero. Thus,

YN92 = .1513 - .5875102 = .1513

Jim was satisfied that his calculated forecasts agreed with those from Minitab.

Example 9.6
The analyst for Atron Corporation, Jim White, was delighted with the forecasts of the
quality control errors that he developed using the Box-Jenkins methodology discussed in
Example 9.5. Jim met his old friend Ed Jones at a conference and told him about his success.
Ed had been struggling with a similar application and decided to give Box-Jenkins a try.
Ed’s data are given in Table 9-9, and a time series plot of the data is shown in Figure 9-18.
The plot of the series and the sample autocorrelations, shown in Figure 9-19, suggest
that the series of quality control errors is stationary. The errors appear to vary about a fixed
level of zero, and the autocorrelations die out rather quickly.
Ed noticed the first two autocorrelations were significantly different from zero and,
perhaps more important, the autocorrelations for the first few time lags fell off toward zero
much like the theoretical pattern for an AR(1) process (see Figure 9-2(a)). Ed also exam-
ined the sample partial autocorrelations shown in Figure 9-20. As he suspected, the partial
autocorrelations seemed to cut off after the significant partial autocorrelation at lag 1.11
Taken together, the sample autocorrelation and sample partial autocorrelation patterns

TABLE 9-9 Errors for Ed Jones’ Quality Control

for Example 9.6 (read by columns)

.77 1.04 2.46 .73 .23

.33 1.02 .37 .10 1.05
2.15 2.03 .80 1.47 .66
2.50 2.54 .49 .89 .25
1.36 .23 .50 .53 .63
.48 .49 .07 .20 .91
2.05 .87 1.92 .70 .21
1.46 .61 1.00 .27 .24
1.13 .20 2.16 .39 .05
2.85 .98 .04 .07 .85
2.67 .78 1.91 .89 1.55
2.71 .80 .43 .37 .40
1.30 .86 .32 .75 1.82
.88 1.72 .48 1.24 .81
.07 .15 .13 .62 .28
1.47 1.15 2.26 .54 1.06

11Ed attributed the significant partial autocorrelation at lag 5 to sampling error, since he could find no phys-
ical reason why the quality control errors five time periods apart should be correlated.
424 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-18 Errors (Deviations from Target) for Ed Jones’ Quality

Control for Example 9.6

FIGURE 9-19 Sample Autocorrelations for Ed Jones’ Quality Control Errors

for Example 9.6

were consistent with an AR(1) process, and Ed felt comfortable fitting an AR(1)
(ARIMA(1, 0, 0)) model to his time series of errors (deviations from target).
The Minitab output for Ed’s initial attempt with the Box-Jenkins methodology is given
in Table 9-10. Since the sample mean of the error series is extremely small (close to zero)
compared with its standard deviation, Ed did not include a constant term in his AR(1)
model.
The parameter in the AR(1) model is estimated to be fN 1 = .501, and it is significantly
different from zero (t = 5.11). The residual mean square error is s 2 = 1.0998. Plots of the
residuals (not shown), the Ljung-Box chi-square statistics, and the residual autocorrelations
displayed in Figure 9-21 suggest the AR(1) model is adequate. There is no reason to doubt
the usual error term assumptions.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 425

FIGURE 9-20 Sample Partial Autocorrelations for Ed Jones’ Quality Control

Errors for Example 9.6

TABLE 9-10 Minitab Output for AR(1) Model for Ed Jones’

Quality Control Errors for Example 9.6

ARIMA (1, 0, 0) Model: Ed Jones’ Quality Control Errors

Final Estimates of Parameters

Type Coef SE Coef T P
AR 1 0.5008 0.0980 5.11 0.000
Number of observations: 80
Residuals: SS = 86.8808(backforecasts excluded)
MS = 1.0998 DF = 79

Modified Box-Pierce (Ljung-Box) Chi-Square statistic

Lag 12 24 36 48
Chi-Square 10.7 19.5 36.2 44.2
DF 11 23 35 47
P-Value 0.468 0.669 0.410 0.591

Forecasts from period 80

95% Limits
Period Forecast Lower Upper Actual

81 0.53088 1.52498 2.58673

82 0.26588 2.03340 2.56515
426 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-21 Autocorrelation Function for the Residuals of the AR(1)

Model for Example 9.6

The AR(1) forecasting equation is

YNt = .501Yt - 1

so the forecasts for time periods 81 and 82 are

YN81 = .501Y80 = .50111.062 = .531

YN82 = .501YN81 = .5011.5312 = .266

Ed’s computed forecasts agree with the one- and two-step-ahead forecasts generated by
Minitab.
Because it was his first experience with the Box-Jenkins methodology, Ed decided to fit a
slightly more complex model to see if the results would confirm his choice of an AR(1) model for
his data. He decided to add an additional parameter and fit an ARIMA(1, 1) (ARIMA(1, 0, 1))
model to the quality control errors. He reasoned that, if his AR(1) model was correct, the
moving average parameter in the ARIMA(1, 1) model should not be significant.
Table 9-11 shows the output for the ARIMA(1, 1) model fit to Ed’s series. Ed was
pleased to see that the MA(1) parameter was not significantly different from zero (t = 1.04),
implying that this term is not needed. Of course, since it is a slightly more general model
than the AR(1) model, it still adequately represents the data as indicated by s 2 = 1.0958
and the random behavior of the residuals.

Example 9.7
Jill Blake, the analyst for the ISC Corporation, was asked to develop forecasts for the clos-
ing prices of ISC stock. The stock had been languishing for some time with little growth, and
senior management wanted some projections to discuss with the board of directors. The ISC
stock prices are plotted in Figure 9-22 and listed in Table 9-12.
Somewhat surprisingly, since stock prices tend to be nonstationary, the plot of the ISC
stock prices tends to vary about a fixed level of approximately 250 (see Figure 9-22). Jill
concludes her price series is stationary.
Jill believes the Box-Jenkins methodology might suit her purpose and immediately
generates the sample autocorrelations and sample partial autocorrelations for the stock
price series. The results are displayed in Figures 9-23 and 9-24.
Jill notices that the sample autocorrelations alternate in sign and decline to zero for
low lags. The sample partial autocorrelations are similar but clearly cut off (are zero) after
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 427

TABLE 9-11 Minitab Output for ARIMA(1, 0, 1) Model for Ed Jones’

Quality Control Errors for Example 9.6

ARIMA (1, 0, 1) Model: Ed Jones’ Errors

Final Estimates of Parameters

Type Coef SE Coef T P
AR 1 0.6671 0.1628 4.10 0.000
MA 1 0.2204 0.2121 1.04 0.302

Number of observations: 80
Residuals: SS = 85.4710 (backforecasts excluded)
MS = 1.0958 DF = 78

Modified Box-Pierce (Ljung-Box) Chi-Square statistic

Lag 12 24 36 48
Chi-Square 8.8 17.5 32.2 38.7
DF 10 22 34 46
P-Value 0.547 0.737 0.555 0.770

Forecasts from period 80

95% Limits
Period Forecast Lower Upper Actual
81 0.52814 1.52399 2.58027
82 0.35230 1.89526 2.59985

FIGURE 9-22 Closing Prices for ISC Corporation Stock for Example 9.7

lag 2. Jill remembers the possible patterns for the theoretical autocorrelations and partial
autocorrelations for an AR(2) process (see Figure 9-2(d)) and decides the corresponding
sample quantities are consistent with one of these patterns. She identifies an AR(2)
(ARIMA(2, 0, 0)) model for her data.
428 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-12 Closing Prices for ISC Corporation Stock

for Example 9.7 (read by columns)

235 200 250 270 275

320 290 225 240 205
115 220 125 275 265
355 400 295 225 245
190 275 250 285 170
320 185 355 250 175
275 370 280 310 270
205 255 370 220 225
295 285 250 320 340
240 250 290 215 190
355 300 225 260 250
175 225 270 190 300
285 285 180 295 195

FIGURE 9-23 Sample Autocorrelations for the ISC Stock Prices for Example 9.7

Jill uses Minitab to estimate the parameters in her model. She includes a constant term
to allow for a nonzero level. The output is shown in Table 9-13.
The parameter estimates are fN 0 = 284.9, fN1 = - .324, and fN2 = .219. The estimated
coefficient fN 2 is not significant (t = 1.75) at the 5% level, but it is significant at the 10%
level. Jill decides to keep this parameter in her model. The residual mean square is
s 2 = 2,808, and the residual autocorrelations given in Figure 9-25 are each well within their
two standard error limits. In addition, the values of the Ljung-Box Q statistics for m = 12,
24, 36, and 48 lags are all small. Jill decides her model is adequate.
Jill uses her model to generate forecasts for periods 66 and 67 as follows.
From forecast origin t = 65, the forecast for period 66 is generated by the equation

YNt = fN0 + fN1Yt - 1 + fN 2Yt - 2

 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 429

FIGURE 9-24 Sample Partial Autocorrelations for the ISC Stock Prices for
Example 9.7

TABLE 9-13 Minitab Output for AR(2) Model for ISC

Closing Stock Prices for Example 9.7

ARIMA (2, 0, 0) Model: ISC Closing Stock Prices

Final Estimates of Parameters

Type Coef SE Coef T P
AR 1 0.3243 0.1246 2.60 0.012
AR 2 0.2192 0.1251 1.75 0.085
Constant 284.903 6.573 43.34 0.000
Mean 257.828 5.949

Number of observations: 65
Residuals: SS = 174093 (backforecasts excluded)
MS = 2808 DF = 62

Modified Box-Pierce (Ljung-Box) Chi-Square statistic

Lag 12 24 36 48
Chi-Square 6.3 13.3 18.2 29.1
DF 9 21 33 45
P-Value 0.707 0.899 0.983 0.969

Forecasts from period 65

95% Limits
Period Forecast Lower Upper Actual
66 287.446 183.565 391.328
67 234.450 125.244 343.656
430 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-25 Autocorrelation Function for the Residuals of the AR(2)

Model for Example 9.7

so
YN66 = 284.9 + 1- .3242Y65 + .219Y64
= 284.9 - .32411952 + .21913002 = 287.4

Similarly,
YN67 = 284.9 + 1- .3242YN66 + .219Y65
= 284.9 - .3241287.42 + .21911952 = 234.5

The results agree with those shown in the Minitab output in Table 9-13. For stationary series,
the 95% prediction limits are approximately
YN ; 2 s (9.6)
where YN is the forecast and s is the square root of the mean square error. For example, the
approximate 95% prediction limits for period 66 are

287.4 ; 222808 = 287.4 ; 106 or 1181.4, 393.42

This interval is close to the 95% interval given by Minitab for period 66 in Table 9-13.

Model-Building Caveats
In ARIMA modeling, it is not good practice to include AR and MA parameters to
“cover all the possibilities” suggested by the sample autocorrelation and sample partial
autocorrelation functions. That is, when in doubt, start with a model containing few
rather than many parameters. The need for additional parameters will be evident from
an examination of the residual autocorrelations and partial autocorrelations. If MA
behavior is apparent in the residual autocorrelations and partial autocorrelations, add
an MA parameter and fit the revised model. If the residual autocorrelations look like
those of an AR process, add an AR term and refit the model.
Least squares estimates of autoregressive and moving average parameters in
ARIMA models tend to be highly correlated. When there are more parameters than
necessary, this leads to “trade-offs” among the parameters and unstable models that
can produce poor forecasts.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 431
To summarize, it is good practice to start with a small number of clearly justifiable
parameters and add one parameter at a time as needed. On the other side, if parame-
ters in a fitted ARIMA model are not significant (as judged by their t ratios), delete
one parameter at a time and refit the model. Because of the high correlation among
estimated parameters, it may be the case that a previously nonsignificant parameter
becomes significant.

Model Selection Criteria

ARIMA models are identified (selected) by looking at a plot of the series and by
matching sample autocorrelation and sample partial autocorrelation patterns with the
known theoretical patterns of ARIMA processes. However, there is some subjectivity
involved in this procedure, and it is possible that two (or more) initial models may be
consistent with the patterns of the sample autocorrelations and partial autocorrela-
tions. Moreover, after estimation and checking, both models may adequately represent
the data. If the models contain the same number of parameters, the model with the
smallest mean square error, s 2, is ordinarily preferred. (This was the case for Lynn
Stephens in Example 9.3.) If the models contain different numbers of parameters, the
parsimony principle leads to the selection of the simpler model. However, the model
with more parameters may have an appreciably smaller mean square error.
An approach to model selection that considers both the model fit and the number
of parameters has been developed. The information criterion of Akaike (1974), or AIC,
selects the best model from a group of candidate models as the one that minimizes

2
AIC = ln s
N2 + r (9.7)
n
where
ln = the natural log
sN 2 = the residual sum of squares divided by the number of observations
n = the number of observations (residuals)
r = the total number of parameters (including the constant term) in the ARIMA
model
The Bayesian information criterion developed by Schwarz (1978), or BIC, selects the
model that minimizes

ln n
BIC = ln s
N2 + r (9.8)
n
The second term in both the AIC and the BIC is a factor that “penalizes” the inclusion
of additional parameters in the model. Since the BIC criterion imposes a greater
penalty for the number of parameters than does the AIC criterion, use of the minimum
BIC for model selection will result in a model whose number of parameters is no
greater than that chosen by the AIC. Often the two criteria produce the same result.
The AIC and the BIC should be viewed as additional procedures to assist in model
selection. They should not be used as substitutes for a careful examination of the sample
autocorrelations and partial autocorrelations.

Example 9.8
In Example 9.4, Jim White found that two ARIMA models appeared to provide an adequate
description of the readings for the Atron Corporation process. One was an AR(1) model
with r = 2 estimated parameters (including a constant term) and sN 2 = 10,065>75 = 134.2.
432 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

The second model was an MA(2) model with r = 3 parameters (including a constant term)
and sN 2 = 9,725>75 = 129.7. Jim computed the AIC and the BIC criteria as follows.

122 = 4.8993 + .0533 = 4.9526

2
AIC = ln1134.22 +
75
AR112:
ln1752
BIC = ln1134.22 + 122 = 4.8993 + .1151 = 5.0144
75

132 = 4.8652 + .0800 = 4.9452

2
AIC = ln1129.72 +
75
MA122:
ln1752
BIC = ln1129.72 + 132 = 4.8652 + .1727 = 5.0379
75

The AIC and the BIC give conflicting results. The AIC is smaller for the MA(2) model,
whereas the BIC is smaller for the AR(1) model. Jim remembered that the BIC will not
select a model with more parameters than the AIC because of the larger penalty for addi-
tional parameters. Jim was happy to see that his initial choice of the AR(1) model, based on
the parsimony principle, was supported by the BIC criterion.

ARIMA Models for Seasonal Data

Seasonal data have a distinct pattern that repeats itself every year. The monthly sales of
the Cavanaugh Company shown in Figure 5.1 always peak in the late fall (October and
November). For monthly data with an annual seasonal pattern, observations for the
same months in different years should be correlated. That is, January in one year should
be similar to January in the next year, February in one year should be similar to February
the next year, and so forth. Thus, not only are observations within the year related to one
another (correlated), but also observations between years are related to one another
(correlated). If the length of the seasonal period is S, so that S = 12 for monthly data and
S = 4 for quarterly data, the autocorrelations and partial autocorrelations for seasonal
processes are nonzero at low lags (within-year association) and at lags that are multiples
of the seasonal period S (between-year association). The interpretations of the autocor-
relations and partial autocorrelations at the seasonal lags are the same as the interpreta-
tions of the autocorrelations and partial autocorrelations at low lags.
Seasonal ARIMA models contain regular autoregressive and moving average
terms that account for the correlation at low lags and seasonal autoregressive and
moving average terms that account for the correlation at the seasonal lags. In addition,
for nonstationary seasonal series, an additional seasonal difference is often required to
completely specify the model.
The next examples illustrate the development of seasonal ARIMA models. The
same model-building strategy used for nonseasonal models is used for seasonal models.
Once an adequate model has been established, it can be used for forecasting.
Example 9.9
Kathy Fisher was given the responsibility for forecasting sales for the Keytron Corporation.
Kathy was able to collect 115 consecutive months of sales data. The data begin in January
1997 and are shown in Table 9-14.
Kathy examined the time series plot in Figure 9-26 and noticed the pronounced sea-
sonal pattern along with a general upward trend. She felt the series was nonstationary and
decided to develop a seasonal ARIMA model for her data.
Kathy began the model identification step by looking at the sample autocorrelations
shown in Figure 9-27. Kathy saw that the autocorrelations at the small lags appeared to cut off
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 433

TABLE 9-14 Keytron Sales for Example 9.9 (read by columns)

1,736.8 1,627.6 1,895.4 1,768.0 2,202.2 1,708.2 2,288.0

1,297.4 1,575.6 1,822.6 1,840.8 2,449.2 806.0 2,275.0
559.0 1,682.2 2,054.0 1,804.4 2,090.4 2,028.0 2,581.8
1,455.6 1,710.8 1,544.4 2,007.2 2,184.0 2,236.0 2,540.2
1,526.2 1,853.8 600.6 2,067.0 2,267.2 2,028.0 2,519.4
1,419.6 1,788.8 1,604.2 2,048.8 1,705.6 2,100.8 2,267.2
1,484.6 1,822.4 1,796.6 2,314.0 962.0 2,327.0 2,615.6
1,651.0 1,838.2 1,822.6 2,072.2 1,929.2 2,225.6 2,163.2
1,661.4 1,635.4 1,835.6 1,952.6 2,202.2 2,321.8 899.6
1,851.2 618.8 1,944.8 2,134.6 1,903.2 2,275.0 2,210.0
1,617.2 1,593.8 2,009.8 1,799.2 2,337.4 2,171.0 2,376.4
1,614.6 1,898.0 2,116.4 756.6 2,022.8 2,431.0 2,259.4
1,757.6 1,911.0 1,994.2 1,890.2 2,225.6 2,165.8 2,584.4
1,302.6 1,695.0 1,895.4 2,256.8 2,441.4 780.0
572.0 1,757.6 1,947.4 2,111.2 2,113.8 2,056.6
1,458.6 1,944.8 1,770.6 2,080.0 2,035.8 2,340.0
1,567.8 2,108.6 626.6 2,191.8 2,152.8 2,033.2

FIGURE 9-26 Keytron Sales for Example 9.9

after lag 1, although there was a small blip at lag 3. Kathy also saw that the autocorrelations at
the seasonal lags—12, 24, and 36 (not shown)—were large and failed to die out quickly. This
suggested the series was nonstationary and confirmed the impression Kathy had from the
time series plot. Before continuing, Kathy decided to difference the series with respect to the
seasonal lag to see if she could convert the nonstationary series to a stationary one.
The seasonal difference for period S = 12 is defined as
¢ 12 = Yt - Yt - 12
For Keytron sales, the first seasonal difference that can be computed is
Y13 - Y1 = 1757.6 - 1736.8 = 20.8
The remaining seasonal differences are computed, and the series consisting of the seasonal
differences of sales is shown in Figure 9-28.
434 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-27 Sample Autocorrelations for the Keytron Sales for Example 9.9

FIGURE 9-28 Seasonally Differenced Keytron Sales for

Example 9.9

The sample autocorrelations and sample partial autocorrelations for the differenced
series are given in Figures 9-29 and 9-30, respectively. Figure 9-28 indicates that the seasonally
differenced data are stationary and seem to vary about a level of roughly 100. The autocorre-
lations (Figure 9-29) have one significant spike at lag 12 (cuts off), and the sample partial auto-
correlations (Figure 9-30) have significant spikes at lags 12 and 24 that get progressively
smaller (die out). This behavior suggests an MA(1) term at the seasonal lag 12.
Kathy identified an ARIMA(0, 0, 0)10, 1, 1212 model for her sales data. This notation
means the model has
p = 0 regular autoregressive terms
d = 0 regular differences
q = 0 regular moving average terms
P = 0 seasonal autoregressive terms
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 435

FIGURE 9-29 Sample Autocorrelations for the Seasonally Differenced

Keytron Sales for Example 9.9

FIGURE 9-30 Sample Partial Autocorrelations for the Seasonally

Differenced Keytron Sales for Example 9.9

D = 1 seasonal difference at lag 12

Q = 1 seasonal moving average term at lag 12
Since the seasonally differenced series varies about a nonzero level, Kathy included a con-
stant term in her model. The identified model has the form

Yt - Yt - 12 = m + t - Æ 1 t - 12 (9.9)

where μ is the mean level of the seasonally differenced process and Æ 1 is the seasonal mov-
ing average parameter.
436 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-15 Minitab Output for ARIMA(0, 0, 0)(0, 1, 1)12 Model

for Keytron Sales for Example 9.9
ARIMA (0,0,0)(0,1,1)12 Model: Keytron Sales
Final Estimates of Parameters
Type Coef SE Coef T P
SMA 12 0.8180 0.0881 9.28 0.000
Constant 85.457 2.910 29.36 0.000
Differencing: 0 regular, 1 seasonal of order 12
Number of observations: Original series 115, after
differencing 103
Residuals: SS = 1172652 (backforecasts excluded)
MS = 11610 DF = 101
Modified Box-Pierce (Ljung-Box) Chi-Square statistic
Lag 12 24 36 48
Chi-Square 12.5 27.1 34.6 46.4
DF 10 22 34 46
P-Value 0.250 0.209 0.439 0.456
Forecasts from period 115
95% Limits
Period Forecast Lower Upper Actual
116 2419.69 2208.46 2630.93
117 2504.31 2293.07 2715.54

Kathy used Minitab to estimate the parameters in her model, compute the residual
autocorrelations, and generate forecasts. The output is shown in Table 9-15. The residual
autocorrelations are given in Figure 9-31, and forecasts for the next 12 months are
appended to the end of the sales series in Figure 9-32.
Kathy was pleased to see her initial model fit the data very well.The Ljung-Box chi-square
statistics, for groups of lags m = 12, 24, 36, and 48, were clearly not significant, as evidenced by
their large p-values. The residual autocorrelations were uniformly small with no apparent pat-
tern. Kathy was ready to use her fitted model to verify the forecasts produced by Minitab.

FIGURE 9-31 Autocorrelation Function for the Residuals for the ARIMA
(0, 0, 0)(0, 1, 1)12 Model for Keytron Sales for Example 9.9
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 437

FIGURE 9-32 Sales and Forecasts for the Next 12 Months

for Keytron for Example 9.9

N 1 = .818. Using these values,

The estimated parameters were mN = 85.457 and Æ
Equation 9.9 can be solved for Yt to give

Yt = Yt - 12 + 85.457 + t - .818 t - 12

To forecast sales for period 116, Kathy set t = 116 and noted that, at the time she was mak-
ing the forecast, her best guess of 116 (the error for the next period) was zero. Thus, the
forecasting equation is

YN116 = Y104 + 85.457 - .818e104

where e104 = - 72.418 is the residual (estimated error) for period 104. Thus,

YN116 = 2,275 + 85.457 - .8181-72.4182 = 2,419.7

Similarly,

YN117 = Y105 + 85.457 - .818e105

YN117 = 2,581.8 + 85.457 - .8181199.2142 = 2,504.3

Kathy’s forecasts agreed with those from the Minitab output. She was convinced she under-
stood how to use the fitted model to produce forecasts. Kathy was delighted with the forecasts of
sales for the next 12 months shown in Figure 9-32.The forecasts seemed to be entirely consistent
with the behavior of the series. She felt they captured the seasonal pattern and the growth in
sales quite well.

Example 9.10
In Example 3.5, Perkin Kendell, analyst for Coastal Marine Corporation, used autocorrela-
tion analysis to conclude that quarterly sales were seasonal. He has decided to forecast
sales for 2007 using the Box-Jenkins methodology. The data were plotted in Figure 3.14.
The time series plot shows a pretty clear seasonal pattern with perhaps a slight upward
trend.
Perkin began his analysis by looking at the sample autocorrelations for the original
series and the series’ various differences. Perkin’s friend, an expert in the Box-Jenkins
438 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

methodology, told him this was a good idea whenever there was a possibility the series was
nonstationary. Once the simplest model suggested by an autocorrelation pattern has been
tentatively identified, the identification can be confirmed by examining the sample partial
autocorrelations. Perkin had Minitab compute the autocorrelation functions for the follow-
ing series:

Yt = the original sales series

¢Yt = Yt - Yt - 1 = the series of first 1regular2 differences
¢ 4Yt = Yt - Yt - 4 = the series of seasonal differences of order S = 4
¢ ¢ 4Yt = ¢1¢ 4Yt2 = ¢1Yt - Yt -42
c d  the series consisting of a seasonal difference
= Yt - Yt - 1 - Yt - 4 + Yt - 5 followed by a regular difference

On occasion, it is necessary to take one seasonal difference and one regular difference
before the resulting series is stationary.
The sample autocorrelations for the original series, for the seasonally differenced series,
and for the series with one regular and one seasonal difference are shown in Figures 9-33
(also Figure 3.15), 9-34, and 9-35.
The autocorrelation function of the original series has (significantly) large spikes at
lags 1 and 4. However, the autocorrelations at the seasonal lags 4, 8, and 12 decay toward
zero. This may indicate a nonstationary series and the need for a seasonal difference.
Perkin’s attempt to fit an ARIMA model to the original series with one constant term, one
regular moving average term, and one seasonal autoregressive term was not successful. The
estimated seasonal autoregressive parameter turned out to be close to one (implying that a
seasonal difference was needed), and the residual autocorrelations were generally large and
not consistent with random errors. Perkin next turned to the autocorrelation function of the
seasonally differenced series in Figure 9-34.
The autocorrelations for the seasonally differenced data are large at low lags and
decline rather slowly in the form of a wavelike pattern. Perkin decided the series may
still be nonstationary and felt that a regular difference in addition to the seasonal differ-
ence might be required to achieve stationarity. Consequently, he had Minitab produce
the sample autocorrelations for the first differenced and seasonally differenced sales in
Figure 9-35.

FIGURE 9-33 Sample Autocorrelations for Coastal Marine Sales for

Example 9.10
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 439

FIGURE 9-34 Sample Autocorrelations for Coastal Marine Sales

Seasonally Differenced for Example 9.10

FIGURE 9-35 Sample Autocorrelations for Coastal Marine Sales First

Differenced and Seasonally Differenced for Example 9.10

Examining the autocorrelations in Figure 9-35, Perkin noticed only two significant val-
ues, those at lags 1 and 8. In addition, he saw that the autocorrelations for the first two lags
alternated in sign. Perkin felt that an ARIMA model with a regular autoregressive term and
perhaps a seasonal moving average term at lag 8 might be appropriate. Before continuing,
Perkin wanted to confirm his initial choice by examining the partial autocorrelations for the
first differenced and seasonally differenced series.
The sample partial autocorrelations for the first differenced and seasonally differenced
data are shown in Figure 9-36. The partial autocorrelations seem to cut off after the first lag,
consistent with the AR(1) behavior Perkin observed in the autocorrelations, and there is a
significant partial autocorrelation at lag 8.
Perkin was not sure how to handle the seasonal parameters, so he decided to keep it
simple and fit an ARIMA(1, 1, 0)10, 1, 024 model. This model allows for one regular and
one seasonal difference and a regular autoregressive term but no seasonal autoregressive
or moving average coefficients. A plot of the series after differencing shows it varies about
zero, so no constant term is included in the model. Perkin reasoned that, if seasonal
440 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-36 Sample Partial Autocorrelations for Coastal Marine Sales

First Differenced and Seasonally Differenced for Example

parameters were needed, the evidence would show up in the residual autocorrelations
from his initial model.
The output for the ARIMA(1, 1, 0)10, 1, 024 model is shown at the top of Table 9-16, and
the residual autocorrelations are shown at the top of Figure 9-37. The model seems to fit fairly
well with an estimated AR(1) coefficient of fN 1 = .352 and a residual mean square error of
s 2 = 1,040.7. However, a significant spike occurs at lag 8 in the residual autocorrelations, and
the Ljung-Box statistic for the first m = 12 lags is Q = 19.7 with a p-value of .05. The latter
suggests that, as a group, the first 12 residual autocorrelations are larger than would be
expected for random errors.
Perkin decided to modify his initial model and include moving average terms corre-
sponding to the seasonal lags 4 and 8. The evidence suggests that only the term at lag 8 is
necessary; however, the Minitab program requires inclusion of the seasonal AR or MA
parameters in the model up to and including the highest multiple of the seasonal lag that is
needed. In this case, the seasonal lag is S = 4, and a moving average parameter at lag
2 * 4 = 8 is needed.
Perkin used Minitab to fit an ARIMA(1, 1, 0)10, 1, 224 model to his data.12 The out-
put is shown in Table 9-16 (bottom), and the residual autocorrelations are given in Figure
9-37 (bottom). Perkin was pleased with the results. The residual mean square error has
been reduced to s 2 = 716.5, and the residual autocorrelations, as judged by the individual
values and the Ljung-Box chi-square statistics, are consistent with random errors.
Residual plots (not shown) suggest that the other error term assumptions are appropriate.
To gain additional experience with the Box-Jenkins methodology, Perkin wanted to
check the forecasts of sales for the next two quarters shown on the Minitab output. Using
parameter estimates in place of the true values, the final model is

Wt = - .351Wt - 1 + t - .239 t - 4 - .672 t - 8

where

Wt = ¢¢ 4Yt = Yt - Yt - 1 - Yt - 4 + Yt - 5

12The Minitab commands to run the ARIMA11, 1, 0210, 1, 224 model are demonstrated at the end of this
chapter.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 441

TABLE 9-16 Minitab Output for ARIMA(1, 1, 0)(0, 1, 0)4 Model (top) and
ARIMA(1, 1,0)(0, 1, 2)4 Model (bottom) for Coastal Marine
Sales for Example 9.10

ARIMA (1,1,0)(0,1,0)4 Model: Coastal Marine Sales

Final Estimates of Parameters
Type Coef SE Coef T P
AR 1 -0.3517 0.1384 -2.54 0.015

Differencing: 1 regular, 1 seasonal of order 4

Number of observations: Original series 52, after differencing 47
Residuals: SS = 47874.1 (backforecasts excluded)
MS = 1040.7 DF = 46

Modified Box-Pierce (Ljung-Box) Chi-Square statistic

Lag 12 24 36 48
Chi-Square 19.7 23.7 32.1 *
DF 11 23 35 *
P-Value 0.050 0.421 0.610 *
ARIMA (1, 1, 0)(0,1,2)4(Model: Coastal Marine Sales
Final Estimates of Parameters
Type Coef SE Coef T P
AR 1 -0.3505 0.1423 -2.46 0.018
SMA 4 0.2392 0.1340 1.78 0.081
SMA 8 0.6716 0.1404 4.78 0.000
Differencing: 1 regular, 1 seasonal of order 4
Number of observations: Original series 52, after differencing 47
Residuals: SS = 31525.4 (backforecasts excluded)
MS = 716.5 DF = 44
Modified Box-Pierce (Ljung-Box) Chi-Square statistic
Lag 12 24 36 48
Chi-Square 8.4 10.7 22.0 *
DF 9 21 33 *
P-Value 0.493 0.969 0.928 *
Forecasts from period 52
95% Limits
Period Forecast Lower Upper Actual
53 169.401 116.926 221.875
54 230.617 168.047 293.188

is the appropriately differenced series. At forecast origin t - 1, the forecasting equation,

written in terms of the Y’s, takes the form
YNt = Yt - 1 + Yt - 4 - Yt - 5 - .3511Yt - 1 - Yt - 2 - Yt - 5 + Yt - 62 - .239et - 4 - .672et - 8
where et - 4 and et - 8 are the residuals (estimated errors) for periods t - 4 and t - 8, respectively.
The forecast of sales for period 53 is
YN53 = Y52 + Y49 - Y48 - .3511Y52 - Y51 - Y48 + Y472 - .239 e49 - .672 e45
YN 53 = 281.4 + 232.1 - 338.2 - .3511281.4 - 291.0 - 338.2 + 329.62
- .2391 -1.7262 - .672118.8982
= 169.4
442 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-37 Autocorrelations of Residuals from the ARIMA(1, 1, 0)(0,1,0)4

Model (top) and Autocorrelations of Residuals from the
ARIMA(1, 1, 0)(0,1,2)4 Model (bottom) for Coastal Marine
Sales for Example 9.10

and the forecast for period 54 is

YN54 = YN53 + Y50 - Y49 - .3511YN53 - Y52 - Y49 + Y482 - .239 e50 - .672 e46
YN54 = 169.4 + 285.6 - 232.1 - .3511169.4 - 281.4 - 232.1 + 338.22
- .2391-37.0402 - .67214.7812
= 230.6
Perkin was happy to see that his calculated forecasts agreed with those produced by Minitab.
Forecasts for the next four quarters are plotted in Figure 9-38. The forecasts seem reasonable.
The pattern of the forecasts is much like the historical pattern. Perkin felt his forecasting effort
was a success and planned on recommending the Box-Jenkins methodology to his colleagues.

Simple Exponential Smoothing and an ARIMA Model

Several ARIMA models produce forecasts that are the same or nearly the same as
those resulting from the smoothing methods discussed in Chapter 4. To illustrate one
case, consider the ARIMA(0, 1, 1) model
Yt - Yt - 1 = t - v1 t - 1 (9.10)
Suppose the forecast origin is t and a forecast of Yt + 1 is required. Substituting t + 1 for
t in Equation 9.10 and solving for Yt + 1 gives
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 443

FIGURE 9-38 Forecasts for the Next Four Quarters from

the ARIMA(1, 1, 0)(0,1,2)4 Model for
Coastal Marine Sales for Example 9.10

Yt + 1 = Yt + t + 1 - v1 t
Because, at time t, the best guess of t + 1 is zero and t is estimated by the residual
et = Yt - YNt, the forecasting equation is
YNt + 1 = Yt - v11Yt - YNt2 = 11 - v12Yt + v1YNt (9.11)
Let a = 1 - v1, and Equation 9.11 is identical to the simple exponential smoothing
Equation 4.13:
YNt + 1 = aYt + 11 - a2YNt

Forecasting using simple exponential smoothing is equivalent to generating fore-

casts using an ARIMA(0, 1, 1) model with parameter v1 = 1 - a. Note that the
ARIMA(0, 1, 1) model represents a nonstationary process. Simple exponential smooth-
ing should work well for those time series that can be adequately described by an
ARIMA(0, 1, 1) model. For those time series that cannot be adequately described by
this model, forecasts generated by exponential smoothing may not be particularly good.
For simple exponential smoothing, the parameter a is usually restricted to the
range 0 6 a 6 1. The moving average parameter, v1, in the ARIMA(0, 1, 1) model is
restricted to the range - 1 6 v1 6 1, so, strictly speaking, the two forecasting methods
are equivalent for positive values of the parameters a and v1.

Advantages and Disadvantages of ARIMA Models

The Box-Jenkins approach to time series analysis is a very powerful tool for providing
accurate short-range forecasts. ARIMA models are quite flexible and can represent a
wide range of characteristics of time series that occur in practice. Formal procedures
for testing the adequacy of the model are available. Moreover, forecasts and prediction
intervals follow directly from the fitted model.
However, ARIMA modeling has some drawbacks.
1. A relatively large amount of data is required. It should be recognized that, if the
data are seasonal with, say, a seasonal period of S = 12, then monthly observations
444 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

for one year essentially constitute one data point (one look at the seasonal pat-
tern), not 12. Generally speaking, for nonseasonal data, about 40 observations or
more are required to develop an ARIMA model. For seasonal data, about 6 to
10 years of data—depending on the length of the seasonal period—are required to
construct an ARIMA model.
2. There are no easy ways to update the parameters of an ARIMA model as new
data become available, as there are in some smoothing methods. The model has
to be periodically completely refitted, and sometimes a new model must be
developed.
3. The construction of a satisfactory ARIMA model often requires a large investment
of time and other resources. The costs of data manipulation and model develop-
ment can be substantially higher for ARIMA models than for the more traditional
forecasting techniques such as smoothing.

APPLICATION TO MANAGEMENT
According to Bernstein (1996), forecasts are one of the most important inputs man-
agers develop to aid them in the decision-making process. Virtually every important
operating decision depends to some extent on a forecast. Inventory accumulation is
related to the forecast of expected demand; the production department has to schedule
employment needs and raw materials orders for the next month or two; the finance
department must arrange short-term financing for the next quarter; the personnel
department must determine hiring and layoff requirements. The list of forecasting
applications is quite lengthy.
Executives are keenly aware of the importance of forecasting. Indeed, a great deal
of time is spent studying trends in economic and political affairs and how events might
affect demand for products and/or services. One issue of interest is the importance
executives place on quantitative forecasting methods versus forecasts based strictly on
their judgments. This issue is especially sensitive when events that have a significant
impact on demand are involved. One problem with quantitative forecasting methods is
that they depend on historical data. For this reason, they are probably least effective in
calling the turn that often results in sharply higher or lower demand.
One type of problem that business managers frequently face is the need to prepare
short-term forecasts for a number of different items. A typical example is the manager
who must schedule production on the basis of some forecast of demand for several
hundred different products in a product line. The techniques that are used most fre-
quently in this situation are the smoothing methods.
The major advantages of exponential smoothing are its low cost and its simplic-
ity. It may not be as accurate as more sophisticated methods, such as general
ARIMA modeling. However, when forecasts are needed for inventory systems con-
taining thousands of items, smoothing methods are often the only reasonable
approach.
Accurate forecasting from a time series perspective depends on the assumption
that the future will be like the past and that past patterns can be described adequately.
Time series techniques are most useful for forecasting variables that are likely to con-
tinue to fluctuate in some stable pattern in the future.
The Box-Jenkins methodology is a very powerful tool for providing accurate
short-range forecasts. Managers must be aware that building a satisfactory ARIMA
model with the Box-Jenkins technique requires a fair amount of historical data and a
potentially high investment in the analyst’s time.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 445
Appropriate applications for Box-Jenkins methodology are many. ARIMA mod-
els have been used to
• Estimate a change in price structure in the U.S. telephone industry
• Investigate the relationship among ammonia concentration, flow rate, and temper-
ature in rivers
• Forecast annual stock volume
• Forecast the number of active oil wells
• Analyze the number of private housing units started
• Analyze the daily observations on percent advances of number of stocks traded
• Analyze the competition between rail and airline routes
• Forecast employment
• Analyze a large number of energy time series for a utility company
• Analyze the effect of promotions on the sales of consumer products
• Forecast different categories of product quality assurance.

Glossary
Box-Jenkins methodology. The Box-Jenkins method- Principle of parsimony. The principle of parsi-
ology refers to a set of procedures for identifying, fit- mony refers to the preference for simple models
ting, and checking ARIMA models with time series over complex ones.
data. Forecasts follow directly from the form of the
fitted model.

Key Formulas

ARIMA model: pth-order autoregressive model

Yt = f0 + f1Yt - 1 + f2Yt - 2 + Á + fpYt - p + t (9.1)

ARIMA model: qth-order moving average model

Yt = m + t - v1 t - 1 - v2 t - 2 - Á - vq t - q (9.2)

ARIMA model: ARIMA(p, q) model

Yt = f0 + f1Yt - 1 + f2Yt - 2 + Á
(9.3)
+ fpYt - p + t - v1 t - 1 - v2 t - 2 - Á - vq t - q

Residual mean square error

n n

a 1Yt - Yt2
2 N 2
a et
t=1 t=1
s = 2
= (9.4)
n - r n - r
Ljung-Box Q statistic
m r 2 1e2
k
Q = n1n + 22 a (9.5)
k=1 n - k

Approximate 95% prediction interval for stationary series

YN ; 2 s (9.6)
446 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Akaike information criterion (AIC)

2
N2 +
AIC = ln s r (9.7)
n
Bayesian (Schwarz) information criterion (BIC)
ln n
BIC = ln s
N2 + r (9.8)
n
Seasonal moving average model (monthly data)
Yt - Yt - 12 = m + t - Æ 1 t - 12 (9.9)

ARIMA(0, 1, 1) model
Yt - Yt - 1 = t - v1 t - 1 (9.10)

ARIMA(0, 1, 1) model forecasting equation in exponential smoothing form

YNt + 1 = Yt - v11Yt - YNt2 = 11 - v12Yt + v1YNt (9.11)

Problems

1. a. For a sample of 100 observations of random data, calculate a 95% confidence

interval for the autocorrelation coefficient at any lag.
b. If all the autocorrelation coefficients are within their individual 95% confidence
intervals and they show no particular pattern, what conclusion about the
process can be drawn?
c. If the first three autocorrelations are all positive and significantly different from
zero and, in general, the autocorrelation pattern declines to zero, what conclu-
sion can be drawn about the process?
d. A process is observed quarterly. If the autocorrelations r4, r8, and r12 are signifi-
cantly different from zero, what conclusion can be drawn?
2. Suppose the following time series model has been fit to historical data and found
to be an adequate model.
Yt = 35 + t + .25 t - 1 - .30 t - 2

The first four observations are Y1 = 32.5, Y2 = 36.6, Y3 = 33.3, and Y4 = 31.9.
Assuming YN1 = 35 and 0 = 0, calculate forecasts for periods 5, 6, and 7 if period 4
is the forecast origin.
3. A time series model has been fit and checked with historical data yielding
Yt = 50 + .45Yt - 1 + t

Suppose at time t = 60 the observation is Y60 = 57.

a. Determine forecasts for periods 61, 62, and 63 from origin 60.
b. Suppose the observed value of Y61 is 59. Update the forecasts for periods 62 and 63.
c. Suppose the estimate for the variance of the error term is s 2 = 3.2. Compute a
95% prediction interval about the forecast for period 61.
4. Fill in the missing information in Table P-4, indicating whether the theoretical
autocorrelations and partial autocorrelations die out or cut off for these models.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 447

TABLE P-4

Model Autocorrelations Partial Autocorrelations

MA
AR
ARIMA

5. Given the graphs in Figure P-5 of the sample autocorrelations and the sample par-
tial autocorrelations, tentatively identify an ARIMA model from each pair of
graphs.
6. An ARIMA(1, 1, 0) model (AR(1) model for first differences) is fit to observations
of a time series. The first 12 residual autocorrelations are shown in Figure P-6. The
model was fit with a constant.
a. Based on an inspection of the residual autocorrelations, does the model appear
to be adequate? Why or why not?
b. If you decide the model is not adequate, indicate how it might be modified to
eliminate any inadequacy.
7. Chips Bakery has been having trouble forecasting the demand for its special high-
fiber bread and would like your assistance. The data for the weekly demand and
the autocorrelations of the original data and various differences of the data are
shown in Tables P-7A–D.
a. Inspect the autocorrelation plots and suggest a tentative model for these data.
How did you decide on this model? (Note: Do not difference more than neces-
sary. Too much differencing is indicated if low lag sample autocorrelations tend
to increase in magnitude.)
b. Using a computer program for ARIMA modeling, fit and check your identified
model with the bread demand data.
c. Write down the equation for forecasting the demand for high-fiber bread for
period 53.
d. Using a computer program for ARIMA modeling, forecast the demand for
high-fiber bread for the next four periods from forecast origin 52. Also, con-
struct 95% prediction intervals.
8. Table P-8 contains a time series with 126 observations. Using a computer pro-
gram for ARIMA modeling, obtain a plot of the data, the sample autocorrela-
tions, and the sample partial autocorrelations. Develop an appropriate ARIMA
model, and generate forecasts for the next three time periods from forecast ori-
gin t = 126.
9. Table P-9 contains a time series of 80 observations. Using a computer program for
ARIMA modeling, obtain a plot of the data, the sample autocorrelations, and the
sample partial autocorrelations. Develop an appropriate ARIMA model, and gen-
erate forecasts for the next three time periods from forecast origin t = 80.
10. Table P-10 contains a time series of 80 observations. Using a computer program for
ARIMA modeling, obtain a plot of the data, the sample autocorrelations, and the
sample partial autocorrelations. Develop an appropriate ARIMA model, and gen-
erate forecasts for the next three time periods from forecast origin t = 80.
11. Table P-11 contains a time series of 96 monthly observations. Using a computer pro-
gram for ARIMA modeling, obtain a plot of the data, the sample autocorrelations,
448 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Autocorrelation Partial Autocorrelation

1.0 1.0

0.0 0.0

Time Lags Time Lags

−1.0 −1.0

(a)

Autocorrelation Partial Autocorrelation

1.0 1.0

0.0 0.0

Time Lags Time Lags

−1.0 −1.0

(b)

Autocorrelation Partial Autocorrelation

1.0 1.0

0.0 0.0

Time Lags Time Lags

−1.0 −1.0

(c)

FIGURE P-5 Sample Autocorrelations and Sample Partial

Autocorrelations
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 449
Residual Autocorrelations: ARIMA(1, 1, 0)

1.0
0.8
0.6

Autocorrelation
0.4
0.2
0.0
−0.2
−0.4
−0.6
−0.8
−1.0

2 7 12
Lag Corr T LBQ Lag Corr T LBQ
1 −0.21 −1.51 2.40 8 0.04 0.22 31.72
2 −0.53 −3.65 18.07 9 −0.18 −0.86 33.72
3 0.20 1.12 20.37 10 −0.23 −1.10 37.11
4 0.28 1.50 24.77 11 0.25 1.19 41.40
5 −0.23 −1.20 27.91 12 0.20 0.94 44.30
6 −0.08 −0.41 28.31
7 0.23 1.16 31.59

FIGURE P-6 Autocorrelations for the Residuals of an ARIMA(1, 1, 0)

Model

and the sample partial autocorrelations. Develop an appropriate ARIMA model,

and generate forecasts for the next 12 time periods from forecast origin t = 96.
12. The data in Table P-12 are weekly prices for IBM stock.
a. Using a program for ARIMA modeling, obtain a plot of the data, the sample
autocorrelations, and the sample partial autocorrelations. Use this information
to tentatively identify an appropriate ARIMA model for this series.

TABLE P-7A Weekly Sales Demand for High-Fiber Bread

(in 1,000s)

Week Demand Week Demand Week Demand Week Demand

1 22.46 14 30.21 27 39.29 40 47.31

2 20.27 15 30.09 28 39.61 41 50.08
3 20.97 16 33.04 29 41.02 42 50.25
4 23.68 17 31.21 30 42.52 43 49.00
5 23.25 18 32.44 31 40.83 44 49.97
6 23.48 19 34.73 32 42.15 45 52.52
7 24.81 20 34.92 33 43.91 46 53.39
8 25.44 21 33.37 34 45.67 47 52.37
9 24.88 22 36.91 35 44.53 48 54.06
10 27.38 23 37.75 36 45.23 49 54.88
11 27.74 24 35.46 37 46.35 50 54.82
12 28.96 25 38.48 38 46.28 51 56.23
13 28.48 26 37.72 39 46.70 52 57.54
450 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE P-7B Sample Autocorrelations

for Original Data

Lag Autocorrelation Lag Autocorrelation

1 .94 7 .59
2 .88 8 .53
3 .82 9 .48
4 .77 10 .43
5 .71 11 .38
6 .65 12 .32

TABLE P-7C Autocorrelations for the

First-Differenced Series

Lag Autocorrelation Lag Autocorrelation

1 -.40 7 .20
2 -.29 8 - .03
3 .17 9 - .03
4 .21 10 - .23
5 -.22 11 .21
6 -.05 12 .14

TABLE P-7D Autocorrelations for the

Second-Differenced Series

Lag Autocorrelation Lag Autocorrelation

1 -.53 7 .16
2 -.10 8 - .05
3 .11 9 .06
4 .18 10 - .23
5 -.20 11 .16
6 -.04 12 .13

TABLE P-8

t Yt t Yt t Yt t Yt t Yt

1 60.9 7 54.2 13 55.6 19 59.6 25 50.9

2 64.2 8 56.9 14 52.2 20 55.6 26 59.6
3 64.2 9 58.2 15 58.2 21 60.2 27 54.2
4 59.6 10 56.2 16 62.9 22 62.2 28 50.9
5 62.2 11 60.9 17 57.6 23 58.2 29 52.2
6 60.9 12 56.9 18 58.9 24 51.6 30 57.6
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 451

TABLE P-8 (Continued)

t Yt t Yt t Yt t Yt
31 51.6 55 58.2 79 51.6 103 64.2
32 46.9 56 60.9 80 48.2 104 58.2
33 51.6 57 51.6 81 47.6 105 52.9
34 57.6 58 54.9 82 50.2 106 56.9
35 60.2 59 66.2 83 58.2 107 51.6
36 64.2 60 57.6 84 65.6 108 48.2
37 62.2 61 48.9 85 53.6 109 47.6
38 53.6 62 46.2 86 55.6 110 50.2
39 50.9 63 50.9 87 61.6 111 58.2
40 54.2 64 57.6 88 57.6 112 65.6
41 56.2 65 54.9 89 56.2 113 53.6
42 59.6 66 51.6 90 60.9 114 55.6
43 66.2 67 50.2 91 57.6 115 61.6
44 57.6 68 50.9 92 51.6 116 57.6
45 48.9 69 56.9 93 56.2 117 56.2
46 50.9 70 50.2 94 52.2 118 60.9
47 60.2 71 54.2 95 50.2 119 57.6
48 64.2 72 58.2 96 56.9 120 51.6
49 56.9 73 56.9 97 56.9 121 56.2
50 56.9 74 55.6 98 50.2 122 52.2
51 63.6 75 64.2 99 54.2 123 50.2
52 58.2 76 58.2 100 58.2 124 56.9
53 56.9 77 52.9 101 56.9 125 56.9
54 58.9 78 56.9 102 55.6 126 50.2

TABLE P-9

t Yt t Yt t Yt t Yt

1 61 21 50 41 59 61 57
2 50 22 69 42 49 62 56
3 62 23 53 43 64 63 53
4 47 24 57 44 55 64 55
5 64 25 52 45 48 65 55
6 40 26 66 46 61 66 66
7 76 27 47 47 47 67 49
8 38 28 67 48 58 68 57
9 75 29 51 49 46 69 50
10 41 30 57 50 58 70 68
11 74 31 55 51 57 71 42
12 47 32 64 52 52 72 77
13 72 33 48 53 62 73 30
14 47 34 65 54 46 74 88
15 62 35 52 55 72 75 37
16 57 36 65 56 37 76 88
17 56 37 47 57 71 77 32
18 53 38 68 58 33 78 90
19 58 39 48 59 71 79 31
20 61 40 61 60 47 80 85
 TABLE P-10

t Yt t Yt t Yt t Yt

1 91.5 21 181.5 41 125.1 61 155.6

2 93.0 22 183.0 42 119.0 62 169.3
3 106.8 23 167.8 43 117.4 63 173.8
4 109.8 24 143.3 44 115.9 64 170.8
5 114.4 25 132.7 45 94.5 65 158.6
6 106.8 26 134.2 46 90.0 66 151.0
7 105.2 27 134.2 47 135.7 67 151.0
8 115.9 28 126.6 48 114.4 68 158.6
9 123.5 29 122.0 49 106.8 69 169.3
10 131.2 30 131.2 50 91.5 70 199.8
11 138.8 31 151.0 51 96.1 71 222.7
12 140.3 32 161.7 52 106.8 72 233.3
13 120.5 33 163.2 53 108.3 73 250.1
14 125.1 34 163.2 54 93.0 74 271.4
15 134.2 35 152.5 55 93.0 75 273.0
16 138.8 36 131.2 56 120.5 76 273.0
17 152.5 37 122.0 57 131.2 77 269.9
18 164.7 38 126.6 58 129.6 78 260.8
19 161.7 39 129.6 59 143.3 79 260.8
20 163.2 40 125.1 60 151.0 80 266.9

TABLE P-11

t Yt t Yt t Yt t Yt

1 97,575 26 106,650 51 149,850 76 166,950

2 97,755 27 122,550 52 149,850 77 181,500
3 105,825 28 124,650 53 160,350 78 173,925
4 104,700 29 128,925 54 161,325 79 174,450
5 110,700 30 126,038 55 153,225 80 179,475
6 108,060 31 129,300 56 159,375 81 180,300
7 111,825 32 127,050 57 149,325 82 177,975
8 114,525 33 125,025 58 157,200 83 174,525
9 108,375 34 129,675 59 163,350 84 220,650
10 117,900 35 134,250 60 190,650 85 152,625
11 117,600 36 161,475 61 141,600 86 143,100
12 142,800 37 115,350 62 136,725 87 164,925
13 109,500 38 111,750 63 158,625 88 167,175
14 103,995 39 134,025 64 163,050 89 181,725
15 119,250 40 134,325 65 173,925 90 174,150
16 121,200 41 140,175 66 165,900 91 174,675
17 123,900 42 144,300 67 226,650 92 179,700
18 124,350 43 143,475 68 175,050 93 180,525
19 125,775 44 143,700 69 164,700 94 178,200
20 126,900 45 140,325 70 167,625 95 174,750
21 121,650 46 144,375 71 171,225 96 220,875
22 126,600 47 149,175 72 203,550
23 124,350 48 182,400 73 152,400
24 152,400 49 132,900 74 142,875
25 108,150 50 133,950 75 164,700

452
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 453

TABLE P-12

Period IBM Period IBM Period IBM Period IBM

Jan. 6 267 Apr. 7 241 Jul. 7 258 Oct. 6 279

13 267 14 244 14 259 13 287
20 268 21 254 21 268 20 276
27 264 28 262 28 276 27 273
Feb. 3 263 May 5 261 Aug. 4 285 Nov. 3 270
10 260 12 265 11 288 10 264
17 256 19 261 18 295 17 261
24 256 26 261 25 297 24 268
Mar. 3 252 Jun. 2 257 Sep. 1 292 Dec. 1 270
10 245 9 268 8 299 8 276
17 243 16 270 15 294 15 274
24 240 23 266 22 284 22 284
31 238 30 259 29 277 29 304

b. Is the IBM series stationary? What correction would you recommend if the
series is nonstationary?
c. Fit an ARIMA model to the IBM series. Interpret the result. Are successive
changes random?
d. Perform diagnostic checks to determine the adequacy of your fitted model.
e. After a satisfactory model has been found, forecast the IBM stock price for
the first week of January of the next year. How does your forecast differ from the
naive forecast, which says that the forecast for the first week of January is the
price for the last week in December (current price)?
13. The data in Table P-13 are closing stock quotations for the DEF Corporation for
150 days. Determine the appropriate ARIMA model, and forecast the stock price

TABLE P-13

Period DEF Period DEF Period DEF

1 136.0 16 136.0 31 140.0

2 132.8 17 138.4 32 144.8
3 130.4 18 137.6 33 140.0
4 128.8 19 139.2 34 139.2
5 136.8 20 134.4 35 139.2
6 135.2 21 136.8 36 136.8
7 134.4 22 139.2 37 140.8
8 139.2 23 139.2 38 141.6
9 136.8 24 140.0 39 139.2
10 136.0 25 139.2 40 142.4
11 133.6 26 140.8 41 140.8
12 139.2 27 139.2 42 140.0
13 137.6 28 138.4 43 132.0
14 139.2 29 136.0 44 142.4
15 139.2 30 142.4 45 138.4
454 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE P-13 (Continued)

Period DEF Period DEF Period DEF

46 138.4 81 134.4 116 133.6
47 136.8 82 133.6 117 134.4
48 139.2 83 131.2 118 130.4
49 135.2 84 132.0 119 132.8
50 138.4 85 131.2 120 134.4
51 140.8 86 132.8 121 135.2
52 135.2 87 132.0 122 136.8
53 133.6 88 133.6 123 134.4
54 134.4 89 131.2 124 136.0
55 134.4 90 131.2 125 137.6
56 137.6 91 129.6 126 138.4
57 134.4 92 131.2 127 137.6
58 140.8 93 130.4 128 138.4
59 137.6 94 131.2 129 137.6
60 132.8 95 136.0 130 137.6
61 136.8 96 135.2 131 140.0
62 135.2 97 136.8 132 135.2
63 132.8 98 136.8 133 135.2
64 144.0 99 133.6 134 135.2
65 137.6 100 135.2 135 136.0
66 138.4 101 136.0 136 132.0
67 136.0 102 137.6 137 133.6
68 135.2 103 131.2 138 134.4
69 138.4 104 136.0 139 133.6
70 134.4 105 136.0 140 133.6
71 138.4 106 133.6 141 132.8
72 139.2 107 129.6 142 132.0
73 141.6 108 132.8 143 136.0
74 134.4 109 135.2 144 133.6
75 135.2 110 132.0 145 133.6
76 136.0 111 132.8 146 135.2
77 135.2 112 132.8 147 139.2
78 136.0 113 136.0 148 136.8
79 132.8 114 136.8 149 136.0
80 133.6 115 136.8 150 134.4

five days ahead from forecast origin t = 145. Compare your forecasts with the
actual prices using the MAPE. How accurate are your forecasts?
14. The data in Table P-14 are weekly automobile accident counts for the years 2006
and 2007 in Havana County. Determine the appropriate ARIMA model, and
forecast accidents for the 91st week. Comment on the accuracy of your forecast.
15. Table P-15 gives the 120 monthly observations on the price in cents per bushel of
corn in Omaha, Nebraska. Determine the best ARIMA model for these data.
Generate forecasts of the price of corn for the next 12 months. Comment on the
pattern of the forecasts.
16. Use the Box-Jenkins methodology to model and forecast the monthly sales of the
Cavanaugh Company given in Table P-14 in Chapter 5. (Hint: Consider a log trans-
formation before modeling these data.)
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 455

TABLE P-14

Period Observation Period Observation Period Observation

1 101 31 16 61 79
2 84 32 17 62 53
3 54 33 21 63 40
4 39 34 28 64 27
5 26 35 30 65 31
6 40 36 51 66 56
7 99 37 62 67 78
8 148 38 57 68 114
9 147 39 46 69 140
10 134 40 40 70 112
11 106 41 32 71 82
12 83 42 23 72 80
13 76 43 20 73 70
14 63 44 18 74 55
15 57 45 24 75 37
16 37 46 33 76 23
17 32 47 52 77 20
18 22 48 66 78 39
19 20 49 78 79 71
20 23 50 83 80 110
21 30 51 87 81 112
22 50 52 64 82 93
23 61 53 44 83 75
24 59 54 24 84 60
25 64 55 29 85 63
26 58 56 73 86 46
27 44 57 138 87 32
28 26 58 154 88 23
29 24 59 119 89 53
30 18 60 102 90 90

17. Use the Box-Jenkins methodology to model and forecast the quarterly sales of
Disney Company given in Table P-16 in Chapter 5. (Hint: Consider a log transfor-
mation before modeling these data.)
18. Use the Box-Jenkins methodology to model and forecast the monthly gasoline
demand of Yukong Oil Company shown in Table P-17 in Chapter 5.
19. Use the Box-Jenkins methodology to model and forecast the monthly numbers of
employed men 16 years of age and older in the United States shown in Table P-25
in Chapter 5.
20. Use the Box-Jenkins methodology to model and forecast the quarterly sales of
Wal-Mart stores shown in Table P-27 in Chapter 5.
21. Use the Box-Jenkins methodology to model and forecast the annual numbers of
severe earthquakes shown in Table P-18 in Chapter 4.
22. Use the Box-Jenkins methodology to model and forecast the quarterly sales for
The Gap shown in Table P-20 in Chapter 4.
456 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE P-15

t Yt t Yt t Yt

1 125 41 206 81 282

2 126 42 252 82 250
3 126 43 215 83 231
4 129 44 238 84 247
5 132 45 249 85 252
6 136 46 250 86 250
7 139 47 250 87 253
8 137 48 275 88 255
9 155 49 292 89 236
10 144 50 304 90 221
11 142 51 301 91 204
12 153 52 265 92 187
13 159 53 271 93 187
14 160 54 289 94 182
15 156 55 319 95 210
16 153 56 377 96 221
17 151 57 350 97 222
18 158 58 378 98 236
19 150 59 358 99 256
20 129 60 359 100 253
21 117 61 333 101 250
22 112 62 310 102 232
23 109 63 293 103 221
24 126 64 285 104 214
25 123 65 284 105 221
26 121 66 283 106 225
27 123 67 291 107 224
28 125 68 312 108 228
29 128 69 288 109 233
30 128 70 280 110 239
31 130 71 255 111 251
32 129 72 258 112 261
33 139 73 266 113 268
34 131 74 271 114 304
35 131 75 272 115 285
36 166 76 265 116 274
37 159 77 281 117 281
38 163 78 294 118 259
39 158 79 300 119 270
40 165 80 284 120 251

23. Figure P-23 is a time series plot of the weekly number of Influenza A Positive cases
in a region of Texas from the week of September 2, 2003, to the week of April 10,
2007. The number of cases peaks during the November–April period and is zero
during the remaining weeks from late spring to early fall. Although medical data,
this series shares many of the characteristics of the demand for certain spare parts.
The number of flu cases and the demand for spare parts are “lumpy,” with periods
of activity followed by relatively long and variable periods of no activity. Can you
think of any problems a series like the flu series might pose for ARIMA modeling?
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 457

FIGURE P-23 Time Series Plot of Weekly Influenza

A Positive Cases

Can you think of a simple nonseasonal ARIMA model that might produce a rea-
sonable forecast of the number of Influenza A Positive cases one week ahead?

CASES

CASE 9-1 RESTAURANT SALES

This case refers to the sales data and situation for the begin with the week ending Sunday, January 4, 1981,
restaurant discussed in Case 8-3. Jim Price has now and continue through the week ending Sunday,
completed a course in forecasting and is anxious to December 26, 1982. Table 9-17B contains new data
apply the Box-Jenkins methodology to the restau- for the week ending January 2, 1983, through the
rant sales data. These data, shown in Table 9-17A, week ending October 30, 1983.

TABLE 9-17A Restaurant Sales: Old Data

Week Ending Sales ($) Week Ending Sales ($) Week Ending Sales ($)

1/4/81 1,688 9/6/81 5,739 5/9/82 5,837

1/11/81 2,514 9/13/81 5,560 5/16/82 7,162
1/18/81 5,843 9/20/81 5,335 5/23/82 4,997
1/25/81 4,912 9/27/81 5,305 5/30/82 4,063
2/1/81 5,133 10/4/81 5,364 6/6/82 3,942
2/8/81 4,563 10/11/81 5,511 6/13/82 4,011
2/15/81 5,416 10/18/81 5,698 6/20/82 3,999
2/22/81 6,416 10/25/81 5,382 6/27/82 4,794
3/1/81 5,879 11/1/81 5,629 7/4/82 4,956
458 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-17A (Continued)

Week Ending Sales ($) Week Ending Sales ($) Week Ending Sales ($)
3/8/81 3,460 11/8/81 5,617 7/11/82 3,885
3/15/81 4,517 11/15/81 5,742 7/18/82 4,209
3/22/81 5,188 11/22/81 3,747 7/25/82 3,614
3/29/81 5,944 11/29/81 4,159 8/1/82 3,722
4/5/81 5,842 12/6/81 4,853 8/8/82 4,307
4/12/81 6,589 12/13/81 5,607 8/15/82 3,322
4/19/81 5,447 12/20/81 3,946 8/22/82 5,962
4/26/81 7,548 12/27/81 1,919 8/29/82 6,784
5/3/81 6,403 1/3/82 1,898 9/5/82 6,069
5/10/81 4,103 1/10/82 1,870 9/12/82 5,897
5/17/81 6,594 1/17/82 3,962 9/19/82 5,916
5/24/81 5,742 1/24/82 5,973 9/26/82 4,998
5/31/81 3,714 1/31/82 5,009 10/3/82 5,111
6/7/81 3,399 2/7/82 5,328 10/10/82 5,612
6/14/81 3,376 2/14/82 5,014 10/17/82 5,906
6/21/81 3,627 2/21/82 4,986 10/24/82 6,010
6/28/81 4,201 2/28/82 5,213 10/31/82 5,937
7/5/81 3,515 3/7/82 4,807 11/7/82 6,004
7/12/81 3,645 3/14/82 3,964 11/14/82 5,959
7/19/81 3,416 3/21/82 5,201 11/21/82 4,223
7/26/81 3,565 3/28/82 4,863 11/28/82 4,679
8/2/81 2,428 4/4/82 5,019 12/5/82 5,307
8/9/81 3,292 4/11/82 4,868 12/12/82 6,101
8/16/81 3,460 4/18/82 5,777 12/19/82 6,896
8/23/81 6,212 4/25/82 6,543 12/26/82 2,214
8/30/81 6,057 5/2/82 6,352

TABLE 9-17B Restaurant Sales: New Data

Week Ending Sales ($) Week Ending Sales ($) Week Ending Sales ($)

1/2/83 2,431 4/17/83 6,357 7/31/83 3,558

1/9/83 2,796 4/24/83 7,273 8/7/83 3,791
1/16/83 4,432 5/1/83 8,678 8/14/83 3,946
1/23/83 5,714 5/8/83 7,418 8/21/83 3,054
1/30/83 5,324 5/15/83 10,388 8/28/83 6,893
2/6/83 5,011 5/22/83 4,940 9/4/83 8,037
2/13/83 5,336 5/29/83 4,195 9/11/83 6,884
2/20/83 4,999 6/5/83 3,895 9/18/83 7,143
2/27/83 5,340 6/12/83 3,762 9/25/83 8,318
3/6/83 5,009 6/19/83 3,739 10/2/83 6,871
3/13/83 5,590 6/26/83 3,975 10/9/83 6,514
3/20/83 3,308 7/3/83 4,634 10/16/83 6,656
3/27/83 6,558 7/10/83 4,891 10/23/83 6,484
4/3/83 4,991 7/17/83 3,463 10/30/83 6,125
4/10/83 6,952 7/24/83 3,536
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 459

QUESTIONS
1. What is the appropriate Box-Jenkins model to 4. How does your Box-Jenkins model compare to
use on the original data? the regression models used in Chapter 8?
2. What are your forecasts for the first four weeks 5. Would you use the same Box-Jenkins model if
of January 1983? the new data were combined with the old data?
3. How do these forecasts compare with actual sales?

CASE 9-2 MR. TUX

John Mosby has decided to try the Box-Jenkins John computed the sample autocorrelation
method for forecasting his monthly sales data. He function for his data in Case 3-2 (see Figure 3-25)
understands that this procedure is more complex and determined that his data were trending (non-
than the simpler methods he has been trying. He stationary). He then computed the autocorrelation
also knows that more accurate forecasts are possi- function for the first differences of his sales data.
ble using this advanced method. Also, he has access The result is shown in Figure 3-26. From this plot,
to the Minitab computer program that can fit John immediately noticed the significant spikes at
ARIMA models. And since he already has the data lags 12 and 24, indicating seasonality and perhaps
collected and stored in his computer, he decides to the need for a seasonal difference. The sample
give it a try. autocorrelation function for the data after a regu-
John decides to use his entire data set, consisting lar (first) difference and a seasonal difference is
of 96 months of sales data. John knows that an displayed in Figure 9-39. The sample partial auto-
ARIMA model can handle the seasonal structure in correlations for the differenced data are shown in
his time series as well as model the month-to-month Figure 9-40.
correlation.

FIGURE 9-39 Autocorrelations for Mr. Tux: Regular and Seasonally

Differenced Series
460 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-40 Partial Autocorrelations for Mr. Tux: Regular and

Seasonally Differenced Series

QUESTIONS
1. Discuss the problems, if any, of explaining the include in an ARIMA model for Mr. Tux sales?
Box-Jenkins method to John’s banker and oth- What seasonal terms might he include in the
ers on his management team. model?
2. Given the autocorrelations in Figure 9-39 and 3. Using a program for ARIMA modeling, fit and
the partial autocorrelations in Figure 9-40, check an ARIMA model for Mr. Tux sales.
what regular (nonseasonal) terms might John Generate forecasts for the next 12 months.

CASE 9-3 CONSUMER CREDIT COUNSELING

The Consumer Credit Counseling (CCC) operation Dorothy was very happy with your forecasting
was described in Cases 1-2 and 3-3. work so far. However, you are not completely satis-
The executive director, Marv Harnishfeger, con- fied and decide that now is the time to really impress
cluded that the most important variable that CCC her. You tell her that she should try one of the most
needed to forecast was the number of new clients that powerful techniques used in forecasting, the Box-
would be seen for the rest of 1993. Marv provided Jenkins methodology and ARIMA models. Dorothy
Dorothy Mercer with monthly data for the number of has never heard of Box-Jenkins but is willing to have
new clients seen by CCC for the period from January you give it a try.
1985 through March 1993 (see Case 3-3).

ASSIGNMENT
1. Write Dorothy a memo that explains the Box- 3. Write Dorothy a second memo that summarizes
Jenkins methodology. the results of this analysis.
2. Develop an ARIMA model using the Box-
Jenkins methodology, and forecast the monthly
number of new clients for the rest of 1993.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 461

CASE 9-4 THE LYDIA E. PINKHAM MEDICINE

COMPANY13

The Lydia E. Pinkham Medicine Company was a with the causal approaches (see, e.g., Kyle 1978). The
family-owned concern whose income was derived sales data are appealing to study for two reasons:
largely from the sale of Lydia Pinkham’s Vegetable
Compound. Perhaps students today could use some 1. The product itself was unchanged for the span
of the compound to relieve stress; unfortunately, it is of the data; that is, there are no shifts in the
no longer sold. Lydia Pinkham’s picture was on the series caused by changes in the product.
label, and the compound was marketed to women. 2. There was no change in the sales force over the
Ads for the compound included this invitation: span of the data, and the proportion of revenues
“write freely and fully to Mrs. Pinkham, at Lynn, spent on advertising was fairly constant. Thus,
Mass., and secure the advice which she offers free of there are no shifts in the data caused by special
charge to all women. This is the advice that has promotions or other marketing phenomena.
brought sunshine into many homes which nervous-
Typically, actual data are not this “clean” in terms of
ness and irritability had nearly wrecked.” In fact, the
product and marketing continuity.
company ensured that a female employee answered
The task at hand, then, is to determine which
every letter. Women did write to Mrs. Pinkham. Their
Box-Jenkins (ARIMA) model is the “best” for these
claims included this one: “Without [Lydia Pinkham’s
data. The model will be developed using the
Vegetable Compound] I would by this time have
1907–1948 data and tested using the 1949–1960 data
been dead or, worse, insane . . . . I had given up on
shown in Table 9-18.
myself; as I had tried so many things, I believed noth-
ing would ever do me any good. But, thanks to your
medicine, I am now well and strong; in fact, another MODEL IDENTIFICATION
person entirely.” This testimonial and others were A computer program capable of ARIMA modeling
reproduced in print ads for the compound. was used to examine the data for 1907 through 1948;
The unique nature of the company—one domi- the data for 1949 through 1960 are used to examine
nant product that accounted for most of the com- the forecasting ability of the selected model.
pany’s sales, no sales staff, and a large proportion of Preliminary tests suggest that the data are stationary
sales revenues invested in advertising—and the (that is, there is no apparent trend), so differencing is
availability of data on both sales and advertising led not employed. After examining the autocorrelations
the Committee on Price Determination of the and partial autocorrelations, it was determined that
National Bureau of Economic Research (NBER) in an AR model was most appropriate for the data.
1943 to recommend that the data be subjected to (The autocorrelations (ACF) and partial autocorre-
thorough analysis. The research was not undertaken lations (PACF) for 10 periods are given in Table 9-19.)
for several years (Palda 1964). Analysts have studied The autocorrelations and partial autocorrelations
the data using causal models that include the adver- seemed to be consistent with those of an AR(2)
tising data and other economic variables (similar to process. To verify the order p of the AR component,
those presented in Chapter 8). However, several Akaike’s information criterion (AIC) (see Equation
researchers have suggested that Box-Jenkins 9.7) was used with autoregressive models of orders
approaches using only the sales data provide compa- p = 1, 2, and 3. The AIC leads to the choice of an
rable, or even superior, predictions when compared AR(2) model for the Lydia Pinkham data.

13This case was contributed by Dr. Susan C. White, George Washington University, Washington, D.C. For
more information, see Susan C. White, “Predicting Time Series with Neural Networks Versus Statistical
Models: The Lydia Pinkham Data,” Proceedings of the 24th Annual Conference of the Decision Sciences
Institute, Southwest Region, 1993, 108–110.
462 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-18 Lydia E. Pinkham Medicine Data

Year Sales (thousands Year Sales (thousands Year Sales (thousands

of current dollars) of current dollars) of current dollars)

1907 1,016 1925 3,438 1943 2,602

1908 921 1926 2,917 1944 2,518
1909 934 1927 2,359 1945 2,637
1910 976 1928 2,240 1946 2,177
1911 930 1929 2,196 1947 1,920
1912 1,052 1930 2,111 1948 1,910
1913 1,184 1931 1,806 1949 1,984
1914 1,089 1932 1,644 1950 1,787
1915 1,087 1933 1,814 1951 1,689
1916 1,154 1934 1,770 1952 1,866
1917 1,330 1935 1,518 1953 1,896
1918 1,980 1936 1,103 1954 1,684
1919 2,223 1937 1,266 1955 1,633
1920 2,203 1938 1,473 1956 1,657
1921 2,514 1939 1,423 1957 1,569
1922 2,726 1940 1,767 1958 1,390
1923 3,185 1941 2,161 1959 1,397
1924 3,352 1942 2,336 1960 1,289

significance level. The residual autocorrelations were

TABLE 9-19
all small, and each was within its 95% error limits.
Lag k ACF PACF The Ljung-Box chi-square statistics had p-values of
.63, .21, and .64 for groups of lags m = 12, 24, and 36,
1 0.915 0.915 respectively. The AR(2) model appears to be ade-
2 0.765 -0.440 quate for the Lydia Pinkham sales data.
3 0.597 -0.061
4 0.424 -0.124 FORECASTING WITH THE MODEL
5 0.243 -0.168
The final step in the analysis of the AR(2) model for
6 0.059 -0.141
these data is to forecast the values for 1949 through
7 - 0.099 0.024
8 -0.240 -0.170 1960 one period ahead. (For example, data through
9 -0.356 -0.036
10 -0.440 -0.040 TABLE 9-20

Year Actual Forecast Error

MODEL ESTIMATION AND TESTING
OF MODEL ADEQUACY 1949 1,984 1,896.4 87.6
A computer program was used to estimate the 1950 1,787 2,006.7 -219.7
parameters (including a constant term) of the AR(2) 1951 1,689 1,687.8 1.2
model using the data for 1907 through 1948. Using 1952 1,866 1,651.0 215.0
the estimated parameters, the resulting model is 1953 1,896 1,953.9 -57.9
1954 1,684 1,904.4 -220.4
Yt = 178.6 + 1.423Yt - 1 - .521Yt - 2 + t 1955 1,633 1,587.1 45.9
(.137) (.136) 1956 1,657 1,625.0 32.0
1957 1,569 1,685.7 -116.7
where the numbers in parentheses beneath the 1958 1,390 1,548.0 -158.0
autoregressive coefficients are their estimated stan- 1959 1,387 1,339.1 47.9
dard deviations. Each autoregressive coefficient is 1960 1,289 1,428.1 -139.1
significantly different from zero for any reasonable
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 463

1958 are used in computing the forecast for 1959.) current dollars). These figures compare favorably to
The forecasting equation is the accuracy measures from the causal models
YNt = 178.6 + 1.423Yt - 1 - .521Yt - 2 developed by other researchers.
and the one-step-ahead forecasts and the forecast
errors are shown in Table 9-20. SUMMARY AND CONCLUSIONS
In addition to the one-step-ahead forecasts, A parsimonious (smallest number of parameters)
some accuracy measures were computed. The fore- AR(2) model has been fit to the Lydia Pinkham
casts from the AR(2) model have a mean absolute data for the years 1907 through 1948. This model has
percentage error (MAPE) of 6.9% and a mean produced fairly accurate one-step-ahead forecasts
absolute deviation (MAD) of $112 (thousands of for the years 1949 through 1960.

QUESTIONS
1. After this analysis was completed, the figure for 1960.Which model,AR(2) or ARIMA(1, 1, 0), do
sales in 1961 became available: $1,426. What is you think is better for the Lydia Pinkham data?
the model’s forecast for 1961? If this year were 3. The Lydia Pinkham data are interesting due to the
added to the testing data set, what would the unique (unchanging) nature of the product and
revised MAPE and MAD be? marketing for the 54-year period represented.
2. There is some evidence that the Lydia Pinkham What factors might affect annual sales data for
data may be nonstationary. For example, the sam- automobiles and copper over this same period?
ple autocorrelations tend to be large (persist) for Why?
several lags. Difference the data. Construct a For further reading on the Lydia E. Pinkham
time series plot of the differences. Using Minitab Medicine Company and Lydia Pinkham, see
or a similar program, fit an ARIMA(1, 1, 0) Sarah Stage, Female Complaints: Lydia Pinkham
model to annual sales for 1907 to 1948. Generate and the Business of Women’s Medicine (New
one-step-ahead forecasts for the years 1949 to York: Norton, 1979).

CASE 9-5 CITY OF COLLEGE STATION

The City of College Station relies on monthly sales A strong seasonal component is evident in a plot
tax revenue to fund operations. The sales tax rev- of the series and in the sample autocorrelations and
enue generated accounts for approximately 44% of sample partial autocorrelations shown in Figures 9-42
the total General Fund budget. It is important to be and 9-43, respectively. The autocorrelations at multi-
able to forecast the amount of sales tax revenue the ples of lag 3 suggest that there is a within-year quar-
city will receive each month from the state of Texas terly seasonal pattern (the second monthly payment
so that anticipated expenditures can be matched to the city within any quarter is relatively large) as
with projected revenue. well as an annual pattern. In addition, the series con-
Charles Lemore, an analyst for the city, was tains a clear upward trend.
asked to develop a forecasting model for monthly Charles decides to difference the revenue series
sales tax revenue. Charles has recently completed a with respect to the seasonal period S = 12.The sample
forecasting course at the local university and decides autocorrelations for the seasonally differenced series
to try the Box-Jenkins methodology. He begins by are shown in Figure 9-44. A plot of the seasonally dif-
looking at the data and the sample autocorrelation ferenced series (not shown) appears to vary about a
and sample partial autocorrelation functions. constant level of about 50.
The monthly sales tax revenues (in thousands of Charles is ready to fit and check an ARIMA
dollars) for the period from January 1991 to model for the sales tax revenue series, but he needs
November 1999 are listed in Table 9-21. A time your help. He would also like to generate forecasts for
series plot of revenues is given in Figure 9-41. the next 12 months, once he has an adequate model.
 Sales Tax Revenues Jan. 1991–Nov. 1999
1,500

Revenues
1,000

500

20 40 60 80 100
Months

FIGURE 9-41 City of College Station Sales Tax Revenues

FIGURE 9-42 Sample Autocorrelations for Sales Tax Revenues

464
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 465

ASSIGNMENT
1. Develop an ARIMA model for sales tax rev- forecasts to the end of the series, and plot the
enue using the Box-Jenkins methodology. results. Are you happy with the pattern of the
2. Using your model, generate forecasts of rev- forecasts?
enue for the next 12 months. Append these 3. Write a brief memo summarizing your findings.

TABLE 9-21

Year Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1991 382 872 408 443 679 406 409 662 388 570 879 417
1992 468 895 457 568 685 492 465 713 518 572 891 519
1993 544 1,005 543 568 785 596 488 799 518 644 968 544
1994 691 1,125 531 576 868 541 571 876 579 801 960 546
1995 675 1,139 684 564 851 595 641 902 596 735 1,014 652
1996 663 1,174 714 698 842 675 665 948 670 753 1,073 676
1997 646 1,219 680 628 943 756 649 951 742 814 1,100 759
1998 715 1,487 763 793 955 780 777 1,052 775 957 1,117 757
1999 839 1,459 828 862 1,147 840 814 1,064 766 997 1,187

FIGURE 9-43 Sample Partial Autocorrelations for Sales Tax Revenues

466 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-44 Sample Autocorrelations for the Seasonally

Differenced Sales Tax Revenues

CASE 9-6 UPS AIR FINANCE DIVISION

The Air Finance Division is one of several divisions greater fee income than the UPS segment; however,
within United Parcel Service Financial Corporation, it is more costly to pursue.
a wholly owned subsidiary of United Parcel Service Funding forecasts for the non-UPS market seg-
(UPS). The Air Finance Division has the responsi- ment of the Air Finance Division have been highly
bility of providing financial services to the company subjective and not as reliable as the forecasts for
as a whole with respect to all jet aircraft acquisitions. other segments. Table 9-22 lists 10 years of monthly
In addition, this division provides financing to inde- funding requirements for the non-UPS Air Finance
pendent outside entities for aviation purchases as a Division segment. These data were collected from
separate and distinct operation. Historically, the management reports during the period from
non-UPS financing segment of the Air Finance January 1989 through December 1998 and are the
Division has provided a higher rate of return and month-end figures in millions of dollars.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 467

TABLE 9-22

Year Jan. Feb. Mar. Apr. May Jun. Jul. Aug. Sep. Oct. Nov. Dec.

1989 16.2 16.7 18.7 18.8 20.6 22.5 23.3 23.8 22.3 22.3 22.1 23.6
1990 20.1 21.6 21.6 21.9 23.4 25.9 26.0 26.2 24.7 23.5 23.4 23.9
1991 20.0 20.4 20.9 21.6 23.2 25.6 26.6 26.3 23.7 22.2 22.7 23.6
1992 20.2 21.1 21.5 22.2 23.4 25.7 26.3 26.2 23.6 22.8 22.8 23.3
1993 21.0 21.7 22.2 23.1 24.8 26.6 27.4 27.1 25.3 23.6 23.5 24.7
1994 21.2 22.5 22.7 23.6 25.1 27.6 28.2 27.7 25.7 24.3 23.7 24.9
1995 21.8 21.9 23.1 23.2 24.2 27.2 28.0 27.6 25.2 24.1 23.6 24.1
1996 20.7 22.0 22.5 23.6 25.2 27.6 28.2 28.0 26.3 25.9 25.9 27.1
1997 22.9 23.8 24.8 25.4 27.0 29.9 31.2 30.7 28.9 28.3 28.0 29.1
1998 25.6 26.5 27.2 27.9 29.4 31.8 32.7 32.4 30.4 29.5 29.3 30.3

The data in Table 9-22 are plotted in Figure 9-45. autocorrelations for various differences of the series
The funding requirement time series has a strong are examined. The autocorrelation function for the
seasonal component along with a general upward series with one regular difference and one seasonal
trend. The sample autocorrelations and sample par- difference of order S = 12 is shown in Figure 9-48.
tial autocorrelations are shown in Figures 9-46 and That is, these autocorrelations are computed for the
9-47, respectively. differences
The Box-Jenkins methodology will be used to
develop an ARIMA model for forecasting future Wt = ¢¢ 12Yt = ¢1Yt - Yt - 122
funding requirements for the non-UPS segment of = Yt - Yt - 1 - Yt - 12 + Yt - 13
the Air Finance Division.
Because of the nonstationary character of There is clearly a significant spike at lag 12 in the
the time series, the autocorrelations and partial sample autocorrelations shown in Figure 9-48. This

Air Finance Non-UPS Funding Requirements

35
Funding

25

15
20 40 60 80 100 120
Months

FIGURE 9-45 Funding Requirements for Non-UPS Segment

of Air Finance Division
468 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-46 Sample Autocorrelations for Funding

Requirements

suggests an ARIMA model for the Wt’s (the process Consequently, an ARIMA10, 1, 0210, 1, 1212 model
with a regular difference and a seasonal difference) might be a good initial choice for the funding
that has a moving average term at the seasonal lag 12. requirements data.

QUESTIONS
1. Using Minitab or equivalent software, fit an and refit the initial model until you feel you
ARIMA10, 1, 0210, 1, 1212 model to the data in have achieved a satisfactory model.
Table 9-22. Do you think a constant term is 3. Using the model you have developed in Question
required in the model? Explain. 2, generate forecasts for the funding require-
2. Is the model suggested in Question 1 adequate? ments for the next 12 months.
Discuss with reference to residual plots, residual 4. Write a report summarizing your findings. Include
autocorrelations, and the Ljung-Box chi-square in your report a plot of the original series and
statistics. If the model is not adequate, modify the forecasts.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 469

FIGURE 9-47 Sample Partial Autocorrelations for Funding

Requirements

CASE 9-7 AAA WASHINGTON

An overview of AAA Washington was provided in the dependent variable and the unemployment rate,
Case 5-5 when students were asked to prepare a temperature, rainfall, and number of members as
time series decomposition of the emergency road the four independent variables were investigated.
service calls received by the club over five years. The The temperature variable was transformed by sub-
time series decomposition performed in Case 5-5 tracting 65 degrees from the average monthly tem-
showed that the pattern Michael DeCoria had perature values. A nonlinear relationship was then
observed in emergency road service call volume was researched.
probably cyclical in nature. Michael would like to be In Case 7-2, a multiple regression model was
able to predict the cyclical effect on emergency road developed. Variables such as the rainfall, number
service call volume for future years. of members, exponentially transformed average
In Case 6-6, four linear regression models using monthly temperature, and unemployment rate
the total number of emergency road service calls as lagged 11 months were tested.
470 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-48 Sample Autocorrelations for the Regularly

Differenced and Seasonally Differenced
Funding Requirements

In Case 8-6, the multiple regression model of the variance of the emergency road service call
developed in Case 7-2 was checked for serial corre- volume variable. He has asked you to try the Box-
lation, and a model was recommended to Michael Jenkins methodology to forecast the road call vol-
that was believed to be most appropriate for pre- ume data presented in Table 9-23. You decide to
dicting the cyclical nature of emergency road service divide the data set into two sections. The values from
call volume. May 1988 through December 1992 will be the initial-
Michael is still not satisfied. The model pre- ization or fitting section, and the first four months of
sented in Case 8-6 explained only about 71 percent 1993 will be the test or forecasting section.

ASSIGNMENT
1. Develop an ARIMA model for the emergency
road service call volume data. Write Michael a
memo summarizing your findings.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 471

TABLE 9-23 Emergency Road Service Call Volume by Month

Year Month Calls Year Month Calls

1988 May 20,002 1991 January 23,441

June 21,591 February 19,205
July 22,696 March 20,386
August 21,509 April 19,988
September 22,123 May 19,077
October 21,449 June 19,141
November 23,475 July 20,883
December 23,529 August 20,709
1989 January 23,327 September 19,647
February 24,050 October 22,013
March 24,010 November 22,375
April 19,735 December 22,727
May 20,153 1992 January 22,367
June 19,512 February 21,155
July 19,892 March 21,209
August 20,326 April 19,286
September 19,378 May 19,725
October 21,263 June 20,276
November 21,443 July 20,795
December 23,366 August 21,126
1990 January 23,836 September 20,251
February 23,336 October 22,069
March 22,003 November 23,268
April 20,155 December 26,039
May 20,070 1993 January 26,127
June 19,588 February 20,067
July 20,804 March 19,673
August 19,644 April 19,142
September 17,424
October 20,833
November 22,490
December 24,861

CASE 9-8 WEB RETAILER

In Case 4-6, Pat Niebuhr and his team were inter- those supplied by the finance department.
ested in developing alternative forecasts for total However, the appropriate variable for planning
orders and contacts per order (CPO) as part of a purposes is
staffing plan project for the contact centers of his
employer, a large web retailer. Monthly forecasts of
total orders and CPO were supplied by the finance Contacts = Orders * CPO
department for Pat’s use, but since he had some his-
torical data available, he thought he would try to and Pat was convinced, since the data were available,
generate alternative forecasts using smoothing that forecasting contacts directly would be better
methods. He could then compare his forecasts with than generating a contacts forecast by multiplying an
472 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

TABLE 9-24 Contacts for Web Retailer,

June 2001–June 2003

Month Contacts

Jun-01 561,516
Jul-01 566,056
Aug-01 478,739
Sep-01 398,672
Oct-01 482,174
Nov-01 681,482
Dec-01 906,251
Jan-02 649,727
Feb-02 454,580
Mar-02 430,809
Apr-02 387,877
May-02 384,972
Jun-02 409,512
Jul-02 394,388
Aug-02 465,686
Sep-02 489,443
Oct-02 534,183
Nov-02 625,216
Dec-02 940,097
Jan-03 627,356
Feb-03 459,306
Mar-03 434,151
Apr-03 397,658
May-03 383,449
Jun-03 437,019

orders forecast by a CPO forecast. Pat had some The monthly contacts data are given in Table 9-24.
training in Box-Jenkins methods and thought he A plot of the contacts time series is shown in
should give ARIMA modeling a try. But Pat recog- Figure 9-49. Even though he had only about two
nized that he didn’t have very many observations and years’ worth of data, Pat concluded from the time
wondered if ARIMA modeling was even feasible. series plot that the number of contacts is seasonal.

FIGURE 9-49 Monthly Number of Contacts for a Web Retailer

 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 473
Pat was unsure about stationarity. The sample The contact differences appear to vary about
autocorrelations for the contacts series are shown in zero, and the autocorrelation function shows no
Figure 9-50.There is a significant autocorrelation at lag 1, significant autocorrelation, although the error
but the autocorrelation coefficients are not slow to die bands are fairly wide because of the small sample
out like those for a nonstationary series. Because the size. There is a reasonably large (almost signifi-
time series plot seems to show a slight downward cant) autocorrelation at lag 12, and since the
trend, Pat decides to difference the contacts series and time series plot of contacts shows a seasonal
look at a plot and the autocorrelations of the differ- pattern, Pat decided to try a seasonal model for
ences.The results are shown in Figures 9-51 and 9-52. contacts.

FIGURE 9-50 Sample Autocorrelation Function for Contacts

FIGURE 9-51 Time Series Plot of the Differences of Contacts

474 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-52 Sample Autocorrelation Function for the Differenced Contacts

QUESTIONS
1. Using Minitab or equivalent software, fit an and refit the initial model until you feel you
ARIMA10, 1, 0210, 0, 1212 model to the data in have achieved a satisfactory model.
Table 9-24. Why does this model seem to be a 3. Using the model you have developed in Question
reasonable initial choice? Is there another can- 2, generate forecasts of contacts for the next 12
didate model that you might consider? months. Include the forecasts on a plot of the orig-
2. Is the model suggested in Question 1 adequate? inal contacts series.Write a brief report comment-
Discuss with reference to residual plots, residual ing on the reasonableness of the forecasts.
autocorrelations, and the Ljung-Box chi-square 4. Do you have any reservations about using
statistics. If the model is not adequate, modify ARIMA modeling in this situation? Discuss.

CASE 9-9 SURTIDO COOKIES

Jame Luna’s efforts to model and forecast Surtido He learned that to identify an appropriate ARIMA
Cookies monthly sales have been documented in model, it is usually a good idea to examine plots of the
Cases 3-5, 4-8, 5-7, and 8-8. A time series plot of sales time series and its various differences. Moreover, the
and the corresponding autocorrelations were consid- sample autocorrelations and sample partial autocor-
ered in Case 4-8. In Case 5-7, a decomposition analysis relations for the series and its differences provide use-
of sales confirmed the existence of a trend component ful model identification information. Jame decides to
and a seasonal component. A multiple regression try to fit an ARIMA model to his data and generate
model with a time trend and seasonal dummy vari- forecasts. Since Surtido Cookies sales are seasonal,
ables was considered in Case 8-8. At a recent forecast- S = 12, Jame constructed plots and autocorrelation
ing seminar, Jame encountered ARIMA modeling. functions for the original series, the differences of the
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 475
original series (DiffSales), the seasonal differences of autocorrelation function for Diff12Sales are shown in
the series (Diff12Sales), and the series with one regu- Figures 9-53 and 9-54, respectively. Looking at this
lar difference followed by one seasonal difference information, Jame has selected an initial ARIMA
(DiffDiff12 Sales). A time series plot and the sample model but needs your help.

FIGURE 9-53 Time Series Plot of Seasonally Differenced Surtido

Cookies Sales

FIGURE 9-54 Sample Autocorrelations for the Seasonally Differenced

Cookie Sales
476 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

QUESTIONS
1. Using Minitab or equivalent software, fit an statistics. If the model is not adequate, modify
ARIMA10, 0, 0210, 1, 1212 model to the data in and refit the initial model until you feel you
Table 3-12. Why does this model seem to be a rea- have achieved a satisfactory model.
sonable initial choice for Jame? Looking at the 3. Using the model you have developed in
plots and autocorrelations of the original series Question 2, generate forecasts of cookie sales
and differences other than Diff12Sales, is there for the next 12 months. Include the forecasts on
another candidate model that you might consider? a plot of the original Surtido Cookies sales
2. Is the model suggested in Question 1 adequate? series. Write a brief report describing the
Discuss with reference to residual plots, residual nature of the forecasts and whether they seem
autocorrelations, and the Ljung-Box chi-square reasonable.

CASE 9-10 SOUTHWEST MEDICAL CENTER

Mary Beasley has been disappointed in her forecast- decides to give ARIMA modeling a try. She has
ing efforts so far. She has tried Winters’ smoothing nothing to lose at this point.
(see Case 4-7), decomposition (see Case 5-8), and Following the prescribed procedure of looking
regression (see Case 8-9) with less than satisfactory at plots, autocorrelations, and partial autocorrela-
results. Mary knows her time series of monthly total tions of the original series and various differences,
billable visits to Medical Oncology is nonstationary Mary generates these quantities for the original
and seasonal. She knows from one of her Executive series (TotVisits), the differences of the series
M.B.A. courses that ARIMA models can describe (DiffTotVisits), the seasonal differences of the series
time series that are nonstationary and seasonal and (Diff12TotVisits), and the series with one regular

FIGURE 9-55 Time Series Plot of DiffDiff12TotVisits

 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 477

FIGURE 9-56 Sample Autocorrelations for DiffDiff12TotVisits

difference followed by one seasonal difference Figure 9-56 appears to have two significant negative
(DiffDiff12TotVisits). A plot of DiffDiff12TotVisits autocorrelations at lags 1 and 12.This suggests a mov-
is shown in Figure 9-55. The corresponding autocor- ing average parameter at lag 1 and a seasonal moving
relation function is given in Figure 9-56. average parameter at lag 12 in an ARIMA model
Examining the time series plot in Figure 9-55, with one regular and one seasonal difference. Mary is
Mary sees that the differences appear to be stationary ready to fit an ARIMA10, 1, 1210, 1, 1212 model to
and vary about zero. The autocorrelation function in her data.

QUESTIONS
1. Using Mary Beasley’s data in Table 4-13, con- Beasley’s data, and generate forecasts for the
struct a time series plot, and compute the sample next 12 months. Examine the residuals to check
autocorrelations and sample partial autocorrela- for adequacy of fit. Should Mary be happy with
tions for the total billable visits, Yt; the differ- the forecasts of total billable visits now?
enced series, ¢Yt; and the seasonally differenced 3. Looking at a time series plot of total billable vis-
series, ¢ 12Yt. Given this information and the its to Medical Oncology, can you see any fea-
information in Figures 9-55 and 9-56, do you ture(s) of these data that might make modeling
think Mary’s initial model is reasonable? Discuss. difficult?
2. Using Minitab or equivalent software, fit
an ARIMA10, 1, 1210, 1, 1212 model to Mary
478 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

Minitab Applications

The problem. In Example 9.10, Perkin Kendell, an analyst for Coastal Marine
Corporation, wants to forecast sales for 2007.
Minitab Solution

1. If the data are stored in a file, open it using the following menus:

File>Open Worksheet

If the data are not on file, enter them on the worksheet.

2. To compute the autocorrelations for the variable sales, click on

Stat>Time Series>Autocorrelation

3. The Autocorrelation Function dialog box that was shown in Figure 3-28 appears.
a. Double-click on the variable Sales, and it will appear to the right of Series:.
b. Click on OK, and Figure 9-33 will appear.
4. In order to seasonally difference the data, click on the following menus:

Stat>Time Series>Differences

5. The Differences dialog box that was shown in Figure 3-29 appears.
a. Double-click on the variable Sales, and it will appear to the right of Series:.
b. Tab to Store differences in: and enter C2.
c. Tab to Lag: and enter a 4. Click on OK, and the fourth differences will appear in
column 2 beginning in row 5.
6. Label the C2 variable Diff4Sales. To compute the autocorrelations for this vari-
able, repeat step 2 using Diff4Sales as the variable to the right of Series:.
7. To compute the first differences for the Diff4Sales variable, repeat step 5, storing
the differences in C3 and using a Lag: of 1.
8. Label the C3 variable Diff1Diff4Sales. To compute the autocorrelations for this
variable, repeat step 2 using Diff1Diff4Sales as the variable to the right of
Series:.
9. To compute the partial autocorrelations for the variable Diff1Diff4Sales,
click on

Stat>Time Series>Partial Autocorrelation

10. The Partial Autocorrelation Function dialog box that is similar to Figure 3-28
appears.
a. Double-click on the variable Diff1Diff4Sales, and it will appear to the right of
Series:.
b. Click on OK, and Figure 9-36 will appear.
11. In order to run an ARIMA(1, 1, 0)(0, 1, 2) model, click on the following menus:

Stat>Time Series>ARIMA

12. The ARIMA dialog box shown in Figure 9-57 appears.

a. Double-click on the variable Sales, and it will appear to the right of Series:.
 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology 479

FIGURE 9-57 Minitab ARIMA Dialog Box

b. Click the check box to the left of Fit seasonal model, and indicate 4 to the right
of Period:.
c. Under Nonseasonal, place a 1 to the right of Autoregressive:, a 1 to the right of
Difference:, and a 0 to the right of Moving average:.
d. Under Seasonal, place a 0 to the right of Autoregressive:, a 1 to the right of
Difference:, and a 2 to the right of Moving average:.
e. Because the data have been differenced, click off the check box for Include
constant term in model.
f. Click on Forecasts, and the ARIMA-Forecasts dialog box appears. In order to
forecast two periods into the future, place a 2 to the right of Lead:. Click on OK.
g. Click on Storage, and the ARIMA-Storage dialog box appears. Click on the box
next to Residuals, and then click on OK. Click on OK on the ARIMA dialog
box, and the bottom portion of Table 9-16 appears.
h. To compute the autocorrelations for the residuals, repeat step 2 using RESI1 as
the variable to the right of Series:.
13. In order to develop a time series plot that includes a forecast, click on Graphs on
the ARIMA dialog box, and the ARIMA-Graphs dialog box shown in Figure 9-58
appears.
a. Click on the check box to the left of Time series plot.
b. An alternative method for obtaining the autocorrelations for the residuals is to
click on the check box to the left of ACF of residuals.
c. Click on OK. Click on OK on the ARIMA dialog box, and Figure 9-38 appears.
480 CHAPTER 9 The Box-Jenkins (ARIMA) Methodology

FIGURE 9-58 Minitab ARIMA-Graphs Dialog Box

References
Akaike, H. “A New Look at the Statistical Model Newbold, P., and T. Bos. Introductory Business and
Identification.” IEEE Transactions Automatic Economic Forecasting, 2nd ed. Cincinnati, Ohio:
Control 19 (1974): 716–723. South-Western, 1994.
Bernstein, P. Against the Gods: The Remarkable Ord, K., and S. Lowe. “Automatic Forecasting.”
Story of Risk. New York: Wiley, 1996. American Statistician 50 (1996): 88–94.
Bowerman, B. L., R. T. O’Connell, and A. B. Koehler. Pack, D. J. “In Defense of ARIMA Modeling.”
Forecasting, Time Series and Regression, 4th ed. International Journal of Forecasting 6 (2) (1990):
Belmont, Calif.: Thomson Brooks/Cole, 2005. 211–218.
Box, G. E. P., G. M. Jenkins, and G. C. Reinsel. Time Palda, K. S. The Measurement of Cumulative
Series Analysis: Forecasting and Control, 3rd ed. Advertising Effects. Englewood Cliffs, N.J.:
Upper Saddle River, N.J.: Prentice Hall, 1994. Prentice-Hall, 1964.
Diebold, F. X. Elements of Forecasting, 3rd ed. Pindyck, R. S., and D. L. Rubinfeld. Econometric
Cincinnati, Ohio: South-Western, 2004. Models and Economic Forecasts, 4th ed. New
Kyle, P. W. “Lydia Pinkham Revisited: A Box- York: McGraw-Hill, 1998.
Jenkins Approach.” Journal of Advertising Quenouille, M. H. “The Joint Distribution of Serial
Research 18 (2) (1978): 31–39. Correlation Coefficients.” Annuals of
Levenbach, H., and J. P. Cleary. Forecasting Practice Mathematical Statistics 20 (1949): 561–571.
and Process for Demand Management. Belmont, Schwarz, G. “Estimating the Dimension of a
Calif.: Thomson Brooks/Cole, 2006. Model.” Annals of Statistics 6 (1978): 461–464.
C H A P T E R

10 JUDGMENTAL FORECASTING
AND FORECAST ADJUSTMENTS
The forecasting techniques covered in this book all involve the manipulation of histor-
ical data to produce predictions or forecasts of important variables of interest. The dis-
cussions in previous chapters were concerned with complex data analyses and perhaps
implied that the forecaster’s judgment was not involved. To discount judgment in the
interpretation of the quantitative approach to forecasting is definitely misguided. In
fact, as emphasized in Chapter 1, the use of good judgment is an essential component
of all good forecasting techniques. Good judgment is required in deciding on the data
that are relevant to the problem, selecting an appropriate model, and interpreting the
results of the data analysis process. Judgment is sometimes a major portion of the
analysis itself.
This chapter discusses some of the important forecasting elements that are
adjuncts to or supplements for the methodical manipulation of such historical data. In
Chapter 11, several considerations regarding the management of the forecasting
process are described.
In many forecasting situations, only the analysis of historical data is used to gener-
ate the final forecast; the judgment or opinion of the analyst1 is not injected into the
process. This book is concerned primarily with such forecasting techniques and, as a
result, with short- and intermediate-term forecasts. Such forecasts are the essential
concern of most levels of management in an organization and are associated with most
of the critical decisions that must be made.
These forecasting procedures rely on the manipulation of historical data and
assume a past and a future that are indistinguishable except for the specific variables
identified as affecting the likelihood of future outcomes. This assumption precludes a
substantive shift in the technological base of the society—an assumption that many
recent developments suggest is erroneous. Consider, for example, the introduction and
proliferation of high-speed, inexpensive personal computers, cellular phones, hand-
held personal assistants, and so forth.
In some forecasting situations, the analyst supplements the data analysis process
after considering the unusual circumstances of the situation or after recognizing that
past history is not an accurate predictor of the future. The amount of judgment injected
into the forecasting process tends to increase when the historical data are few in num-
ber or are judged to be partially irrelevant. In the extreme case, it may be the analyst’s
opinion that no historical data are directly relevant to the forecasting process. Under

1Here, we use the term analyst to indicate anyone that may produce judgmental forecasts or judgmental
adjustments to model-based forecasts. In this context, the analyst may or may not be the person who gener-
ated the model-based forecasts.

481
482 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

these conditions, forecasts based purely on the opinions of “experts” are used to for-
mulate the forecast or scenario for the future.
It is convenient to think of forecasting in terms of the underlying information sets.
For smoothing methods, decomposition, and ARIMA procedures, the underlying
information set consists of the history of the time series being forecast. For regression
methods, the information set is again historical data—but historical data for more than
one variable. Judgmental forecasts or judgmental adjustments to forecasts are typically
based on domain knowledge. Domain knowledge is any information relevant to the
forecasting task other than time series data. That is, domain knowledge is non–time
series information. Now, in principle, more information should lead to better forecasts.
So forecasts based on historical data and domain knowledge should be better than
forecasts based solely on one information set or the other. However, the available evi-
dence is mixed.
Judgmental forecasts are sometimes the only alternative, but they can be better or
worse than quantitative forecasts when both are available. If judgmental forecasts are
the outcomes of organizational budgeting or incentive objectives (targets to be
achieved), they can be worse than quantitative forecasts. If judgmental forecasts are
produced with up-to-date knowledge, while the statistical methods may lag in data
availability, and/or if the domain knowledge represents a component that can’t be
modeled (e.g., a competitor’s actions), judgmental forecasts can be more accurate.
When judgment is used to modify a quantitative forecast, research suggests that,
in general, the judgmental modifications tend to reduce the accuracy of the forecast.
However, judgmental modifications can result in greater accuracy if domain knowl-
edge is particularly relevant and if the adjustments tend to be “large” rather than
“small.”
As we have mentioned, when few or no relevant historical data are available to
assist in the forecasting process, judgment must be relied on if forecasts or predictions
about the future are desired. Since such situations often arise, especially for top man-
agement, techniques have been developed to improve the accuracy of such forecasts by
utilizing the available executive judgment to the best advantage. The use of these tech-
niques is worthy of consideration, since executives frequently consider their own judg-
ment to be superior to other methods for predicting the future. As Makridakis (1986)
states, “People prefer making forecasts judgmentally. They believe that their knowl-
edge of the product, market, and customers as well as their insight and inside informa-
tion gives them a unique ability to forecast judgmentally” (p. 63).
Following are several questions, each of which suggests the use of imagination and
brainstorming rather than completely relying on the collection and manipulation of
historical data. For each of these questions, one of the forecasting techniques discussed
in this chapter may provide a company’s management team with valuable insights into
its firm’s future operating environment.

• What will be the age distribution of the United States in the year 2030?
• To what extent will U.S. citizens work in the home 25 years from now?
• What cities will be the major population and business centers in 20 years?
• To what extent will the United States outsource the manufacture of key consumer
items in 10 years?
• To what extent will shopping from the home, using television and computers, be
popular in 20 years?
• What kinds of recreation will occupy U.S. citizens in the year 2025?
• Will the United States begin pulling back from its commitments around the globe
over the next 25 years? If so, in what ways will this affect U.S. business?
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 483
We now describe some techniques for producing judgmental forecasts, so named
because judgment is the primary or sole component of the process.2 Some techniques
are also referred to as technological forecasting methods, since they frequently deal
with projecting the effects of technological changes into the uncertain future.

JUDGMENTAL FORECASTING
The Delphi Method
When experts are gathered in a single meeting location and asked about the future,
group dynamics can sometimes distort the process and result in a consensus that may
not be carefully thought out by all participants. The Delphi method, first used by an Air
Force–funded RAND Corporation project in the 1950s, attempts to remove the group
dynamic aspect from the deliberations of the forecasters. In the first round of the
method, the experts reply in writing to the questions posed by an organization’s inves-
tigating team. The team then summarizes the comments of the participants and mails
them back. Participants are then able to read the reactions of the others and to either
defend their original views or modify them based on the views of others.
This process continues through two or three rounds until the investigators are
satisfied that many viewpoints have been developed and carefully considered.
Participants may then be invited to meet together to share and debate their viewpoints.
At the conclusion of this process, the investigating team should have good insight into
the future and can begin to plan the organization’s posture accordingly.
As Rowe and Wright (1999) point out:

Delphi is not a procedure designed to challenge statistical or model-based

procedures, against which human judgment is generally shown to be inferior: it
is intended for use in judgment and forecasting situations in which pure
model-based statistical methods are not practical or possible because of the
lack of appropriate historical/economic/technical data . . . . (p. 354)

Any Delphi procedure has four key features: anonymity, iteration, controlled feed-
back, and aggregation to provide a group response. Anonymity is achieved through the
use of questionnaires administered by hand, by mail, over the phone, or via a computer
network. When experts express their opinions privately, they are free from group social
pressures and can concentrate only on the merits of each idea, proposal, or situation.
Moreover, with the iteration of the questionnaire over a number of rounds, the partic-
ipating experts are given the opportunity to change their opinions without losing face
in the eyes of the remaining (anonymous) group members.
After each iteration of the questionnaire, the opinions of each expert are briefly
summarized, and each group member is informed of the positions of their anonymous
colleagues. Consequently, feedback contains the opinions and judgments of all group
members and not just the most vocal. After the final round, the group judgment is sum-
marized (often using descriptive statistics such as an average or quartiles describing
the range) so that each judgment receives equal weight.
The first round of the Delphi method is often unstructured, allowing the experts
free rein to identify, and comment on, those issues they see as important. The monitor

2For
an organized outline of many forecasting methods including judgmental forecasting, see Georgoff and
Murdick (1986).
484 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

team then consolidates these issues into a single set of structured statements (the ques-
tionnaire) from which the opinions and judgments of the Delphi panelists may be gath-
ered in a quantitative manner in subsequent rounds. The results from each round are
summarized and provided to each expert for further consideration. After the second
round, members are given the opportunity to change prior opinions based on the pro-
vided feedback. In most applications of the Delphi technique, the number of rounds
rarely exceeds three. Empirical studies of the Delphi method suggest that accuracy
tends to increase over Delphi rounds and that Delphi panelists tend to be more accu-
rate than unstructured interacting groups.

Example 10.1
Applied Biosystems supplies life science firms and research institutions with a host of prod-
ucts and services for researching genes and proteins, studying how drugs interact with the
body’s systems and genetic makeup, testing food and the environment for contaminants,
and performing DNA-based identification. In the early years of the company, instrument
systems and other equipment generated most of the sales revenue. In recent years, the side
of the business selling consumables (reagents, services, assays) has been growing rapidly,
and consumables now account for about half of the company’s revenue. Applied Biosystems
does most of its business in the United States but is considering expanding its presence in
Europe, Japan, and Australia, where it currently has relatively small percentages of the
biosystems equipment market. Before committing to the capital expenditures associated
with global expansion, the company is interested in developing forecasts of sales growth in
Europe, Japan, and Australia for the next 10 years. With little historical sales data in these
regions to guide it, Applied Biosystems has decided to hire three experts familiar with life
science research and general international economic conditions.
Expert A is a geneticist who has extensive experience with the pharmaceutical industry
and has spent some time studying the growth of genetically linked research in Europe.
Expert B is a noted economist and has extensive experience studying, in particular, the
economy of Japan. Expert C, a native of Australia, has spent some time working at the
European headquarters of the World Health Organization in the health technologies and
pharmaceuticals area. She currently works as a consultant to Australia’s government and
industry.
Applied Biosystems would like a forecast of the growth in sales for both its equipment
and its consumables groups over the next 10 years in each area—Europe, Japan, and
Australia. To begin, each expert is asked to provide his or her estimates of sales growth for
the instrumentation/equipment and consumables groups by region over the next 10 years.
Each expert is provided with a current level of annual sales and estimated market share by
group for Europe, Japan, and Australia. For the money budgeted, Applied Biosystems is
able to obtain the commitment of the three experts for two rounds of reports. The anony-
mous (to each other) experts, located in different parts of the world, are connected to the
project manager at Applied Biosystems by computer.
The results after the first round are summarized in Table 10-1.
In the first round, Expert A does not see much sales growth for the products and ser-
vices of Applied Biosystems in Europe over the next 10 years. Although genetically linked
research in Europe is likely to grow substantially, the competition from European firms and
from non-European firms with a strong presence in the biologicals market is likely to limit
the opportunities for Applied Biosystems in spite of its excellent reputation. Similarly,
Expert A sees relatively limited growth for Applied Biosystems’ equipment/instrumentation
sales in Japan because of the tendency of Japanese researchers to use the excellent Japanese
equipment. He does, however, feel there is a chance for appreciable growth in consumables
sales in Japan. Expert A does not have an opinion about the growth of Applied Biosystems’
sales in Australia.
Expert B sees limited growth for Applied Biosystems’ equipment/instrumentation sales
in Europe and Japan but a substantial growth opportunity in Australia, although the
Australian market as a whole is relatively small. Expert B is a little more optimistic about
European sales growth than Expert A, but both agree the sales growth, particularly in con-
sumables, is likely to be small for a 10-year period. Expert B feels there is likely to be tremen-
dous growth in sales for both equipment/instrumentation and consumables in Australia, with
the growth in consumables sales being particularly attractive. Expert B feels that Applied
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 485

TABLE 10-1 Results from the First Round of the Delphi Method for Example 10.1

Expert A Expert B Expert C

Equipment/Instrumentation—Europe +10% to +40% +30% to +70% +0% to +150%

Consumables—Europe +5% to +15% +10% to +25% +20% to +60%
Equipment/Instrumentation—Japan +0% to +50% - 10% to + 40% —
Consumables—Japan +40% to +100% +50% to +200% —
Equipment/Instrumentation—Australia — +100% to +200% +150% to +300%
Consumables—Australia — +200% to +400% +200% to +400%

Biosystems has no major competitors in Australia and that the Australian commitment to nur-
turing a healthy lifestyle and maintaining a quality environment will remain strong.
Expert C is more uncertain about equipment/instrumentation sales growth in Europe
than are Experts A and B and suggests a reasonably wide range of 0% to 150%. Her fore-
cast of consumables sales growth for Europe, although more optimistic than those of
Experts A and B, is still relatively small. Expert C has no opinion about Applied
Biosystems’ sales growth in Japan. She is very “bullish” on potential sales growth of Applied
Biosystems’ products and services in her native country of Australia. Her sales growth esti-
mates are large for both equipment/instrumentation and consumables and are of the same
order of magnitude as those of Expert B.
The information in Table 10-1 is supplied to each of the experts, and a second round of
sales growth forecasts is solicited. Each expert has the opportunity to adjust his or her initial
forecast after viewing the results of the first round. Expert A’s opinions remain essentially
unchanged. Expert B adjusts his sales growth range for equipment/instrumentation in the
European market to +20% to +60%, a slight downward adjustment. He also adjusts the
upper limit of his range for consumables sales growth in the Japanese market down to
150%. He leaves his sales growth forecasts for the Australian market unchanged. Expert
C adjusts the ranges of her sales growth forecasts for the European market to

Equipment/Instrumentation—Europe +10% to +90%

Consumables—Europe +15% to +40%

She leaves her forecasts for Australia unchanged.

A final phone call to the Delphi panelists with the results of the second round does not
yield any additional adjustments. With the results of the Delphi exercise in hand, Applied
Biosystems decides to maintain its current presence in Europe, with more emphasis on
equipment/instrumentation sales. The company decides to launch a major marketing effort
in Japan to increase the sales of products in its consumable groups, with the possibility of
partnering with a Japanese firm to help with the sales and distribution efforts. Australia
appears to represent a major opportunity, but the overall market is small, and Applied
Biosystems decides to do a little more cost-benefit analysis before committing to the
Australian market.

The advantage of the Delphi method is that noted experts can be asked to carefully
consider the subject of interest and to reply thoughtfully to the viewpoints of others
without the interference of group dynamics. The result, if the process is handled care-
fully, may be a good consensus of the future, along with several alternative scenarios.3

Scenario Writing
Scenario writing involves defining the particulars of an uncertain future by writing a
“script” for the environment of an organization over many years in the future. New
technology, population shifts, and changing consumer demands are among the factors

3For a detailed description of the Delphi method, see Parente and Anderson-Parente (1987).
486 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

that are considered and woven into this speculation to provoke the thinking of top
management.
A most likely scenario is usually written, along with one or more less likely, but
possible, scenarios. By considering the posture of the company for each of these possi-
ble future environments, top management is in a better position to react to actual busi-
ness environment changes as they occur and to recognize the long-range implications
of subtle changes that might otherwise go unnoticed. In this way, the organization is in
a better position to maintain its long-term profitability rather than concentrating on
short-term profits and ignoring the changing technological environment in which it
operates.
The scenario writing process is often followed by a discussion phase, sometimes by
a group other than the one that developed the scenarios. Discussion among the groups
can then be used to defend and modify viewpoints so that a solid consensus and alter-
native scenarios are developed. For example, scenarios might be developed by a com-
pany’s planning staff and then discussed by the top management team. Even if none of
the scenarios is subsequently proven to be totally true, this process encourages the
long-range thinking of the top management team and better prepares it to recognize
and react to important environmental changes.
Example 10.2
A company that manufactures industrial telephone and television cables decides to conduct
a scenario-writing exercise prior to its annual weekend retreat. Each member of the retreat
group is asked to write three scenarios that might face the company five years from now:
a worst-case, a most-likely, and a best-case scenario. After these writing assignments are
completed, and just before the weekend retreat, the president and his senior vice president
summarize the contributions into the following three scenarios on which they intend to
focus the group’s discussion during the two-day retreat:
1. Internet usage continues to grow rapidly but slowly moves away from cable in favor of
satellite access. Even telephone service relies increasingly on noncable means, as does
home television reception. The company sees its sales and profits on a nonending
decline and will soon be out of business.
2. Internet and home television service continues to grow rapidly but is provided by sev-
eral sources. Satellite service is widely used, but buried cables continue to be an integral
part of high-tech service, especially in large cities, both in private housing and in indus-
trial applications. The company’s leadership in cable development and deployment
results in increased sales and profits.
3. Due to technical problems and security issues, satellite usage for the Internet and tele-
vision service declines until it is used primarily in rural areas. Cable service grows rap-
idly in both residential and industrial applications, and the company prospers as its
well-positioned products and industry leadership result in industry dominance.
The company president and senior vice president intend to have extensive discussions
of each of these three scenarios. They want to have long-range strategies developed that will
accommodate all future possibilities and believe that focusing on these three cases will
energize both themselves and their management team.

Combining Forecasts
A developing branch of forecasting study involves the combination of two or more
forecasting methods to produce the final forecasts. An issue of the International
Journal of Forecasting contained a special section on this new technique. Portions of
the abstracts of three articles in this issue illustrate the developing nature of combining
forecasts:
1. According to Armstrong (1989), research from over 200 studies demonstrates that
combining forecasts produces consistent but modest gains in accuracy. However,
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 487
this research does not define well the conditions under which combining is most
effective or how methods should be combined in each situation.
2. Mahoud (1989) indicates that the amount of research on combining forecasts is
substantial. Yet relatively little is known about when and how managers combine
forecasts. Important managerial issues that require further study include manage-
rial adjustment of quantitative forecasts, the use of expert systems in combining
forecasts, and analyses of the costs of combining forecasts.
3. Considerable literature has accumulated over the years regarding the combination
of forecasts. The primary conclusion of this line of research, according to Clemen
(1989), is that forecast accuracy can be substantially improved through the combi-
nation of multiple individual forecasts. This paper provides a review and annotated
bibliography of that literature.
In a world in which information sets can be quickly and costlessly combined, it is
always optimal to combine information sets rather than forecasts. However, when the
pooling of information sets is impossible (qualitative versus quantitative information)
or prohibitively costly (deadlines must be met and timely forecasts produced), combin-
ing forecasts is the only alternative. Forecast combination can be viewed as a link
between the short-run, real-time forecast production process and the longer run, con-
tinuing process of model development based on all available information.
Example 10.3
A company that makes parts for large farm tractors wants to forecast the number of these
units that will be sold over the next 10 years. From this forecast, it will develop strategies to
remain competitive in its business. Specifically, the company is concerned about its plant
capacity. A forecast of future business would help it greatly both in developing expansion
plans and in dealing with its sources of financial support.
After an extensive data collection and forecasting effort, the company is faced with
deciding which of two forecasts to accept. Although they are not too far apart for most
years, there are some differences. A professional forecasting firm with a solid track record
generated the first forecast. The methods this firm used are unknown to company manage-
ment, but they have been told the process was “mathematically sophisticated.” The second
forecast resulted from an executive retreat attended by top management along with mar-
keting staff familiar with the expectations of its customers.
After discussions with company executives, the president decides to combine the two
forecasts. These discussions tended to favor the professional forecast over the in-house fore-
cast, so the president decides to weight the former 75% and the latter 25%. Table 10-2
shows the two forecasts for each of the next 10 years in units sold, in thousands, followed by

TABLE 10-2 Combining Forecasts of

Units Sold (1,000’s)

Year in Professional In-House Final

Future Forecast Forecast Forecast

1 328 335 329.8

2 342 340 341.5
3 340 345 341.3
4 348 350 348.5
5 350 352 350.5
6 360 355 358.8
7 366 365 365.8
8 371 370 370.8
9 385 375 382.5
10 390 385 388.8
488 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

the combined forecast. Each final forecast was computed by multiplying the first forecast by
.75 and the second by .25 and adding them together. For example, for year 1, the combined
forecast is given by

1.752328 + 1.252335 = 329.75

Notice that the final forecasts fall between the professional and in-house forecasts, but they
are closer to the professional forecasts, since the professional forecasts received the greater
weight.

One of the benefits of combining forecasts is that it minimizes the effects of bias,
which result from assigning undue weight to a particular forecasting method. One
strategy is to combine different forecasts by simply averaging the individual forecasts pro-
duced by different methods. If YN11, YN12, Á , YN1m are one-step-ahead forecasts produced
by m different methods, then the one-step-ahead combined forecast, YN1C, obtained by
simple averaging is

YN11 + YN12 + Á + YN1m

YN1C = (10.1)
m
Similar results hold for other forecast lead times.
As in Example 10.2, forecasts can be combined with unequal weights. In this case,
each one-step-ahead forecast, YN1i, receives weight 0 6 wi 6 1 where ©wi = 1. So, with m
one-step-ahead forecasts produced by m different methods, the combined forecast, YN1C, is

YN1C = w1YN11 + w2YN12 + Á + wmYN1m (10.2)

where
m

a wi = 1
i=1

There is no change in the weights as the forecast lead time changes. For example, the com-
bined forecast two steps ahead, YN2C, would be computed like the combined forecast in
Equation 10.2 with the two-step-ahead forecasts for each method replacing the one-step-
ahead forecasts.The fact that the weights sum to 1 ensures that the combined forecast will
be somewhere between the smallest and largest values of the individual forecasts.
There are several procedures for determining the weights, wi. If a record of past
performance for each forecasting method is available, the weights can be taken to be
inversely proportional to the sum of the squared forecasting errors. Alternatively, the
weights can be obtained by regression methods. Newbold and Bos (1994) provide addi-
tional discussion of the procedures for combining forecasts.
In the coming years, further research will likely be conducted on the advantages of
combining forecasts, along with the techniques for doing so. The objective of such com-
binations will be to develop accurate forecasts that are cost effective.

Forecasting and Neural Networks

Conventional forecasting methods, such as those discussed in this book, rely on histor-
ical data to develop a model and use it to project the variables of interest into the
future. These projected values become the forecasts that are then used to develop the
plans for the business. In these models, it is assumed that the future will be much like
the past, except for those variables specifically recognized by the model.
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 489
Conventional models sometimes make assumptions about the form of population
distributions, assumptions that may or may not be subject to verification. Interval esti-
mates using a regression model, for example, assume that the underlying population
follows a normal distribution.
The developing field of artificial intelligence attempts to duplicate the processes of
the human brain and nervous system using the computer. Although this field origi-
nated in biology and psychology, it is rapidly advancing into other areas, including
business and economics. The three main thrusts of artificial intelligence are language
processing, robotics, and artificial neural networks.This last field has the most commercial
applications, including forecasting.
In neural networks, many examples that capture the full range of past relationships
among all variables that might affect the outcome of the dependent variables are pro-
grammed into the computer. The neural network program then assimilates these exam-
ples and attempts to develop the underlying relationships by “learning” as it goes
along. This learning process, also called training, is analogous to a human trainee learn-
ing on the job.
A few forecasting researchers have noticed the similarity between neural network
techniques and conventional forecasting methods that attempt to find variables to suc-
cessfully predict the dependent variable. The theoretical advantage of neural network-
ing as a forecasting tool is that relationships need not be specified in advance, since the
method involves learning about relationships using the examples provided. Also,
neural networks do not require any assumptions about underlying population distribu-
tions, and unlike many conventional forecasting methods, they can operate with incom-
plete data.4
A successful neural network application is sometimes called plug compatible. This
means that the neural network program can quickly and easily replace the existing
model, such as regression analysis, without disruption. An improvement in perform-
ance, such as more accurate forecasts, can sometimes be achieved with a minimal
impact on ongoing operations. Neural networks are especially valuable when inputs
are highly correlated or missing and when the systems are highly nonlinear.
Following are some examples of the successful use of neural networks in practical
settings.5

• A Kodak plant in Texas reduced costs by $3 million per year while maintaining
product yield and quality. It began by collecting historical operating data, which
were used to train a neural network to forecast the quality of the product as a func-
tion of the various process parameters.
• One of the classic problems in interpreting seismic signals for oil exploration is
finding the first evidence of the shock wave from sensor recordings. Traditional sig-
nal processing methods have not been very successful in correctly identifying this
first evidence: only 30% for one algorithm. Several oil companies have used a net-
work program to correctly identify the first wave, with Amoco claiming a 95% suc-
cess rate. This has had a major impact on their ability to improve the processing of
seismic data for reservoir modeling.

4Although neural networks do not produce judgmental forecasts in the usual sense, neural net forecasts are
distinct from the data-based procedures discussed in Chapters 4–9 and so are included in this chapter. A neu-
ral network is an attempt to put the human thought process on a computer. Moreover, considerable judg-
ment is required to use a neural network effectively.
5These examples were provided by NeuralWare, a commercial supplier of neural network software located
at 230 East Main Street, Suite 200, Carnegie, PA 15106.
490 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

• A German manufacturer of automotive electric fan motors used master inspectors

to determine whether a completed motor was too noisy or exhibited “funny”
sounds. A neural network took one week to learn from a master inspector and is
currently doing the noise test work. The financial impact is a uniformly higher-
quality product with fewer complaints from customers. The master inspectors have
been released from this high-stress job to concentrate on other quality improve-
ment tasks in the factory.
• Since 1989, a neural network has been used to determine the suitability of cus-
tomers for instant credit from a loan company.
• Several companies have discovered that neural networks can be used to develop
indicators useful in trading. Results of the systems indicate that neural network
trading indicators, coupled with appropriate trading strategies, can consistently
make money in the marketplace.
• The Air Force is using a neural network to forecast component failures in Air
Force jets. A variety of data was collected on each aircraft, and the network was
trained to predict the likelihood of specific failures in specific planes. These results
are used for part stocking and preventive maintenance.
• A commercial product has been developed to forecast the outcome of horse races.
The system requires the user to input specific information about the horses in a
race. It wins 80% of the time it decides that the odds favor betting.
As the examples above indicate, neural networks have enjoyed much practical suc-
cess, but the forecasting performance of networks has been mixed. Although free from
the confining restrictions of specifying relationships in advance and assuming popula-
tion distributions, constructing an effective neural network is not a simple task. There is
often a tendency to make networks more complicated than they need to be—in effect,
incorporating noise as well as signal. As a result, the forecasting efficacy of the neural
network is reduced. For an example of a neural network forecasting application, see
Case 10-3.

Summary of Judgmental Forecasting

A potential danger in using most of the forecasting methods discussed in this text is
that they involve the manipulation of historical data to generate the forecast. This prac-
tice is a valid procedure if the forecast is for the near-term future but becomes increas-
ingly suspect as the length of the forecasting horizon increases. The methods discussed
in this chapter are valuable additions to the forecaster’s arsenal when the particular
concern of top management, the long-term forecast of the company’s environment, is
being considered.
Several authors have fascinated the general public with their speculations about
the long-term trends currently developing in our society and the future that they por-
tend, along with the managerial skills that will be needed to cope with them. Among
the more provocative, in no particular order, are these:
• Bill Gates, The Road Ahead (Viking, 1995)
• Rowan Gibson (editor), Alvin Toffler, and Heidi Toffler, Rethinking the Future:
Rethinking Business, Principles, Competition, Control and Complexity, Leadership,
Markets and the World (Nicholas Breasley, 1999)
• Di Kamp, The 21st Century Manager: Future-Focused Skills for the Next
Millennium (Kogan Page, 1999)
• Chuck Martin, Net Future: The 7 Cybertrends That Will Drive Your Business, Create
New Wealth, and Define Your Future (McGraw-Hill, 1998)
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 491
• Alvin Toffler, Future Shock (Random House, 1970)
• Alvin Toffler, The Third Wave (Bantam Books, 1980)
• John Naisbitt, Megatrends (Warner Books, 1982)
• John Naisbitt and Patricia Aburdene, Megatrends 2000 (William Morrow, 1990)
• John Naisbitt, Eight Asian Megatrends That Are Reshaping the World (audiotape)
(Simon & Schuster, 1996)
• John Naisbitt, Global Paradox (Avon, 1995)
• Jim Taylor, Watts Wacker, and Howard Means, The 500-Year Delta: What Happens
After What Comes Next (HarperBusiness, 1998)

The references above make for interesting reading, even though some of them are
25 to 35 years old, because these authors have a unique ability to speculate about the
future in provocative ways. A more formal approach to the kinds of changes they write
about is called technological forecasting and is becoming an increasingly important
field for many firms. This is not surprising, since an estimated 25% of existing technol-
ogy is replaced every year. Obviously, even a very sophisticated manipulation of historical
data to produce forecasts can miss the mark by a wide margin under these circum-
stances. Consider, for example, the following developing fields. As these technologies
unfold, the impact on many firms will be considerable.

• Artificial intelligence
• Genetic engineering
• Bioelectricity
• Multisensory robotics
• Lasers
• Fiber optics
• Microwaves
• Advanced satellites
• Solar energy
• Superconductors

The PricewaterhouseCoopers Company has recently made an internal publica-

tion, Technology Forecast, available to the general public. This annual document ana-
lyzes key information technologies and forecasts their directions in the next one to
three years. This is an example of one company’s intense concern about technological
advances and their effects on society as well as on that company.
The rate of technological development means that managers must be increasingly
alert to change. Still, historical patterns can provide at least some clues as to the future
in almost every business. As stressed in Chapter 1, there needs to be a prudent mix of
appropriate historical data manipulation with the judicious use of judgment and expe-
rience if useful forecasts are to be generated.

OTHER TOOLS USEFUL IN MAKING JUDGMENTS ABOUT THE FUTURE

A number of tools have been developed to assist decision makers in the process of
weighing information about the future and making the best possible decisions. One set
of tools is frequently discussed in textbooks and journal articles under the general
heading of “decision making” or sometimes “decision making under uncertainty.”
When managers are faced with the task of making decisions in the face of uncertainty,
their ability to forecast becomes a critical element in the decision-making process.
492 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

When the analysis of historical data has been completed, the decision maker must
reach a judgment regarding alterations in the firm’s course of action. In other words,
the analyst must weave the results of the forecasting process into the firm’s existing
decision-making procedures. A few of the elements of decision theory that are often
relevant at this stage of the process are briefly discussed next.6
The concept of expected value was described in Chapter 2. Decision makers fre-
quently use it, either explicitly or implicitly. Recall that this concept involves calculating
the mean value that a random numerical variable will assume over many trials. In Table
10-3, a discrete random variable, X, is displayed in a probability distribution; every pos-
sible future value that X can assume is shown along with the probability of each.
Notice that the sum of the probabilities in Table 10-3 is 1.00, or 100%, which means
that every possible value that X can assume has been identified. In Table 10-3, let X
represent the number of new major contracts that a firm will sign during the next fiscal
year. The question that is answered by the expected value is, How many new contracts
can be expected, on the average, if the probability distribution of Table 10-3 is valid?
Equation 10.3 is used to calculate the expected value of a probability distribution.

E1X2 = ©X 3P1X24 (10.3)

where
E1X2 = the expected value
X = the values that the random variable can assume
P1X2 = the probability of each X occurring
The expected value of Table 10-3 can be calculated using Equation 10.3 as follows:

E1X2 = 11.102 + 21.202 + 31.252 + 41.302 + 51.152 = 3.2

The expected value of the probability distribution shown in Table 10-3 is 3.2. If X in
this example represents the number of new major contracts for the coming fiscal year,
then, on average, if the chances for new contracts remain the same year after year, 3.2
new contracts would be signed. Notice that the value 3.2 is not possible in any one year;
only integer values—1, 2, 3, 4, 5—are possible. Nevertheless, the value 3.2 represents
the mean outcome over many trials. Decision makers are frequently interested in
expected values and use them as their best forecasts for critical numerical variables in
planning for the uncertain future.

TABLE 10-3 A Probability

Distribution

X P(X)

1 .10
2 .20
3 .25
4 .30
5 .15
1.00

6Hammond, Keeney, and Raiffa (1999) have written an excellent, and very readable, book on decision making.
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 493
Decision theory formally addresses the elements that comprise the decision-making
function of business leaders. Expected values are often woven into this more general
consideration of decision making. The decision tree diagram is used to help the decision
maker visualize a complex situation and to make rational decisions. Such a decision
tree diagram is shown in Figure 10-1.
Figure 10-1 reflects the uncertainty that exists about the nature of future sales and
incorporates the decision about whether to build a new plant or repair the old one. The
problem is that, if the company knew high demand would result, it would be better off
building the new plant; on the other hand, if it knew low demand would result, profits
would be higher if it repaired the old plant. Even in this simple example, the benefit of the
tree diagram can be seen: It enables the decision maker to see the various choices avail-
able, to identify those uncertainties beyond the firm’s control, and to explicitly determine
costs, profits, and the probabilities of future events. In more complicated situations, the
benefits of the tree diagram and formal decision theory are even more evident.
A statistical concept designed to revise preliminary probabilities on the basis of
sample evidence is Bayes’ theorem. This concept is often applicable in situations in
which estimates of the probabilities of unknown future events are determined and
then modified after collecting sample evidence. An example is the test market concept
used by many manufacturers of consumer products. Such a company might estimate
the probability of public acceptance of a new product as being quite high. However,
before risking the millions of dollars that a national campaign requires, a test market
might be undertaken in areas the company regards as good representative markets.

Payoffs

$150 million
d
an
m
De
gh
Hi .4
0

Low
De
ma
nd
.60
ant
Pl $40 million
w on
Ne illi $150 million
a m
i ld 5 0 a nt n
Bu :$ Pl llio
st nd m
i
Co a 0
E xp $4
:
st
Co
d Do
n’
an tE
xpa
Re m nd
pa De Pla
Co ir O gh nt
st: ld Hi .40 $40 million
$2 Pla
5m nt
illio
n
Lo
w
De
ma
.60 nd

$40 million

FIGURE 10-1 Decision Tree Diagram

494 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

The results of the test market are then used to modify the original estimates of product
success, and a decision about introducing the product nationally is made. A simplified
version of Bayes’ theorem is shown in Equation 10.4.7
P1A2P1B ƒ A2
P1A ƒ B2 = (10.4)
P1B2
where
P1A ƒ B2 = the probability of event A occurring, given that event B has occurred
P1B2 = the probability that event B occurs
P1A2 = the probability that event A occurs
P1B ƒ A2 = the probability of event B occurring, given that event A has occurred

Example 10.4
Figure 10-2 reflects a specific application of Bayes’ theorem in a test market situation. The
managers of a large consumer products company need to decide whether to introduce a
new product nationally. They estimate that their new product has a 50% chance of high sales
in the national market; that is, P1H2 = .50. They are considering the use of a test market to
determine whether they can do a better job of forecasting high or low sales of the new prod-
uct. Figure 10-2 shows a decision tree of the test market outcomes.
Past experience shows that, when a new product was introduced and high sales were
achieved, the test market was successful 80% of the time, or P1S ƒ H2 = .80. Past experience

= .40
Test Market
Prediction
s
es
cc
Su

0
.8

P (S ) = .40 + .075 = .475

National Fa .80(.50)
Market Sales ilur P (H |S ) = = .84
e .475
.20
h
ig
H

= .10
0
.5

Lo
w ss = .075
.5 ce
0 c
Su
.15 P (F ) = .10 + .425 = .525

Fa .85(.50)
P (L |F ) = = .81
ilu .525
re
.8
5

= .425

FIGURE 10-2 Bayes’ Decision for New-Product Introduction

7The general statement of Bayes’ theorem for n events, A

1, A2, Á , An, each of which can occur in conjunction with
another event, B, is
P1A i2P1B | A i2
P1A i | B2 = i = 1, 2, Á , n
P1A 12P1B | A 12 + P1A 22P1B | A22 + Á + P1A n2P1B | A n2
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 495
also shows that, when a new product was introduced and low sales were achieved, the test
market failed (low sales) 85% of the time, or P1F ƒ L2 = .85.
Bayes’ theorem can be used to compute the probability of the new product being suc-
cessful in the national market under different test conditions. Figure 10-2 shows the compu-
tation of the two probabilities of particular interest:
• If the test market predicts success, there is an 84% chance the product will have high
sales nationally. Equivalently,
P1H ƒ S2 = .84 = .801.502>.475 = P1S ƒ H2P1H2>P1S2
• If the test market predicts failure, there is an 81% chance the product will have low
sales nationally, or
P1L ƒ F2 = .81 = .851.502>.525 = P1F ƒ L2P1L2>P1F2

In Example 10.4, a test market would help management decide whether or not to
introduce the new product nationally. The test market accuracy is sufficient to change
the probability of high sales (or low sales) from the pretest value of 50%. If the test
market is successful, the chance of high sales increases to 84%. If the test market is a
failure, the chance of low sales increases to 81%. The probabilities were calculated
using Bayes’ theorem. The decision about product introduction will be much easier to
make than it would have been without the test market.
Another useful tool for forecasting is computer simulation. Simulation is a set of
numerical and programming techniques for modeling situations subject to uncertainty and
for conducting sampling experiments on those models using a computer. Each simulation
run (replication) produces one possible outcome (“forecast”) for the problem being stud-
ied. Many runs (replications) allow the decision maker the opportunity to observe the
complete set of possible outcomes as well as their likelihood of occurrence.These computer-
generated scenarios can then be summarized and used to make the best decision.
There are many reasons for using simulation, instead of mathematical analysis, to
gather information in an uncertain environment.

• Many realistic representations of actual business systems are much too complex to
be analyzed mathematically.
• The primary interest might be to experiment with the system or to find the best
levels for the variables that influence the system or to simply study the system.
Experimenting with the actual system may be impossible (combat situations) or
extremely expensive (design of nuclear generating facilities) or so time consuming
that only one replication is possible.
• The simulation effort is frequently useful in itself because it leads to a better
understanding of the system.
• As a tool, simulation carries a certain amount of credibility with management. It is
relatively easy to explain to management the efforts involved in a simulation study.

From a business perspective, the chief advantage of a computer simulation exer-

cise is that forecasts of the consequences of various management decisions can be gen-
erated before such decisions must be made. Simulation has been used in a wide variety
of settings, including modeling manufacturing operations, scheduling the workforce,
modeling investment alternatives, analyzing and pricing sophisticated investment
instruments, and modeling passenger flows at an airport terminal.8 Decision trees such
as the one in Figure 10-1 are often part of computer simulation experiments.

8Lawand Kelton (2000); Ross (2006); and Seila, Ceric, and Tadikamalla (2003) provide good discussions of
computer simulation.
496 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

Key Formulas

Combined forecast given by simple average

YN11 + YN12 + Á + YN1m

YN1C = (10.1)
m
Combined forecast given by weighted average
YN1C = w1YN11 + w2YN12 + Á + wmYN1m (10.2)

Expected value formula

E1X2 = ©X 3P1X24 (10.3)
Bayes’ theorem
P1A2P1B ƒ A2
P1A ƒ B2 = (10.4)
P1B2

Problems
1. Identify two business situations where the Delphi method might be used to gener-
ate forecasts. Can you think of any difficulties and pitfalls associated with using the
Delphi method?
2. Consider the actual sales shown in Table P-2 along with one-step-ahead forecasts
produced by Winters’ method and by a regression model.
a. Construct the combined forecasts of sales produced by taking a simple average
of the forecasts produced by Winters’ method and the regression model.
b. Construct the combined forecasts of sales produced by taking a weighted aver-
age of the Winters’ forecasts and the regression forecasts with weights w1 = .8
and w2 = 1 - w1 = .2.
c. Using the actual sales, determine the MAPEs for the Winters’ forecasts and the
regression forecasts.
d. Repeat part c using the combined forecasts from parts a and b. Based on the
MAPE measure, which set of forecasts do you prefer?

TABLE P-2

Actual Winters’ Regression

Month Sales Forecasts Forecasts

1 5,502 4,586 5,263

2 6,535 5,944 6,008
3 7,390 6,706 6,832
4 4,388 4,530 4,886
5 4,521 5,010 4,918
6 5,679 6,574 5,630
7 8,578 8,462 7,963
8 7,088 6,415 5,942
9 4,839 4,457 5,156
10 4,050 4,638 3,819
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 497

CASES

CASE 10-1 GOLDEN GARDENS RESTAURANT

Sue and Bill Golden have decided to open a restau- already conducted a series of three focus groups
rant in a city in the Midwest. They have spent over a with area residents who eat out regularly, and
year researching the area and visiting medium- to no consensus on this matter emerged. They have
high-price restaurants. They definitely believe that talked about the matter considerably between
there is room for another restaurant and have found themselves but now believe some other opinions
a good site that is available at a good price. would be valuable.
In addition, they have contacts with a number of After reading about some of the techniques
first-class chefs and believe that they can attract one of used in judgmental forecasting, they believe some of
them to their new restaurant.Their inquiries with local them might help them decide on the atmosphere for
bankers have convinced them that financing will be their new restaurant. They have identified a number
readily available, given their own financial resources of their friends and associates in other cities who
and their expertise in the restaurant business. would be willing to help them but are not certain
The only thing still troubling the Goldens is the how to utilize their talents.
atmosphere or motif for their restaurant. They have

QUESTIONS
1. What method would you suggest to the Goldens 2. Are there any other methods they have over-
in utilizing the expertise of their friends to looked in trying to research this matter?
decide on the atmosphere and motif for their
new restaurant?

CASE 10-2 ALOMEGA FOOD STORES

Example 1.1 described how Julie Ruth, the president manager, during a recent meeting (see Example 1.1).
of Alomega Food Stores, collected monthly sales Tilson said, “I’ve been trying to keep my mouth shut
data for her company along with several other vari- during this meeting, but this is really too much. I think
ables she thought might be related to sales. The we’re wasting a lot of people’s time with all this data
Alomega cases—Cases 2-3. 3-4, 5-6, and 8-7— collection and fooling around with computers. All
described her attempts to use various forecasting you have to do is talk with our people on the floor
procedures available in Minitab in an effort to pro- and with the grocery store managers to understand
duce meaningful forecasts of monthly sales. what’s going on. I’ve seen this happen around here
In Case 8-7, Julie developed a multiple regression before, and here we go again. Some of you people
model that explained almost 91% of the monthly need to turn off your computers, get out of your fancy
sales variable variance. She felt good about this offices, and talk with a few real people.”
model but was especially sensitive to the negative Julie decided that office politics dictated that she
comments made by Jackson Tilson, her production heed Jackson’s advice. She consulted with several
498 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

TABLE 10-4 Monthly Sales for 27

Alomega Food Stores,
2003–2006

Month 2003 2004 2005 2006

Jan. 425,075 629,404 655,748 455,136

Feb. 315,305 263,467 270,483 247,570
Mar. 432,101 468,612 429,480 732,005
Apr. 357,191 313,221 260,458 357,107
May 347,874 444,404 528,210 453,156
Jun. 435,529 386,986 379,856 320,103
Jul. 299,403 414,314 472,058 451,779
Aug. 296,505 253,493 254,516 249,482
Sep. 426,701 484,365 551,354 744,583
Oct. 329,722 305,989 335,826 421,186
Nov. 281,783 315,407 320,408 397,367
Dec. 166,391 182,784 276,901 269,096

people, including Tilson, to determine their opinions Based on this input, Julie developed a naive
of how to forecast sales for January 2007. A large forecasting model:
majority felt that using the sales figure for the previ-
YNt + 1 = Yt - 11
ous January would provide the best prediction.
Likewise, the forecast for February 2007 should be that used last year’s monthly value to predict this
based on the sales figure for February 2006. year’s monthly value.

QUESTIONS
1. How accurate is Julie’s naive forecasting forecasts. She feels that this approach would
model? counter office politics and still allow her to use a
2. How does the naive model compare to the mul- more scientific approach. Would this be a good
tiple regression model developed in Case 8-7? idea?
3. Until Julie can experience each of these two 4. Should Julie use a simple average or weighted
methods in action, she is considering combining average approach to combining the forecasts?

CASE 10-3 THE LYDIA E. PINKHAM MEDICINE

COMPANY

This case demonstrates an actual application of the The Lydia E. Pinkham Medicine Company and
usage of neural networks to forecast time series Lydia Pinkham’s Vegetable Compound were intro-
data. The authors understand that students have not duced in Case 9-4.
been provided with the background to completely There have been many attempts to use networks
understand this case. However, it is felt that benefits to forecast time series data. Most of the work has
will be derived from experiencing this actual case. been in the field of power utilization, since power
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 499
companies need accurate forecasts of hourly demand both the AR(2) model and the neural network
for their product. However, some research has model use the same “information” in computing the
focused on more traditional business time series, such one-step-ahead predictions.) The 4 indicates the
as micro- and macroeconomic series, demographic number of nodes, or processing units, in the hidden
data, and company-specific data. Virtually all this layer. (It is termed hidden because it is not directly
work has used a feed-forward network trained using connected to the “outside world,” as are the input
backpropagation.This case study will employ this type and output layers.) The number of nodes in the hid-
of network to forecast the Lydia Pinkham sales data. den layer is chosen arbitrarily in a sense: Too few
The resulting forecasts will be compared to the fore- hidden nodes restrict the network’s ability to “fit”
casts from the AR(2) model presented in Case 9-4. the data, and too many hidden nodes cause the net-
Figure 10-3 depicts a 2–4–1 feed-forward neural work to memorize the training (or estimation) data.
network—the network used for this study. The 2 in The memorization leads to very poor performance
the 2–4–1 indicates the number of inputs to the net- over the testing sample. In this case, the number of
work. In this case, the two inputs are Yt - 1 and Yt - 2. nodes in the hidden layer is simply twice the number
(Using the two previous periods to predict the cur- of inputs. Finally, the 1 indicates that one output
rent period is consistent with the AR(2) model; thus, node gives the one-step-ahead forecast, or Yt.

Y in Time t − 1 Y in Time t − 2

Input Layer Pass to Hidden Pass to Hidden

Layer Layer

Connecting Weights

Compute Compute Compute Compute

Weighted Sum Weighted Sum Hidden Layer Weighted Sum Weighted Sum
and Transform and Transform and Transform and Transform

Connecting Weights

Compute
Output Layer
Weighted Sum
for Output

Predicted
Y in Time t

FIGURE 10-3 A 2–4–1 Feed-Forward Network

500 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments

TABLE 10-5 Lydia E. Pinkham Medicine Neural Network Results

Estimation Data Training Data

Iterations MAD MAPE MSE MAD MAPE MSE

10,000 189.7 11.06% 51,923.0 182.2 11.26% 43,441.1

20,000 189.8 10.80% 53,926.3 169.6 10.18% 37,737.4
50,000 195.0 11.22% 54,311.3 176.3 10.51% 41,529.7
75,000 196.0 11.40% 54,424.2 178.1 10.66% 42,758.1
100,000 196.5 11.53% 54,725.5 181.9 11.02% 43,985.4

The neural network computes its output in the fol- how long to train the network; an overtrained net-
lowing manner: Each of the connecting arcs between work tends to memorize the training data and
nodes in two adjacent layers has an associated weight. perform poorly on the testing data. Thus, some
Each node in the hidden layer computes a weighted researchers have suggested simply stopping the train-
sum of its inputs. (The input layer nodes simply pass ing “early”—before the network has memorized the
the inputs on to the hidden layer.) This weighted sum is data. To determine the training effect, the network
then “transformed” in some fashion, such as was trained for 10,000, 20,000, 50,000, 75,000, and
Y = 1>11 + e -x2 where Y is the “transformed” data 100,000 iterations. (One iteration is the presentation
and X is the weighted sum. The Y is then passed on to of one observation; the iterations listed represent 250,
the output layer, where each node again computes a 500, 1,250, 1,875, and 2,500 passes through the com-
weighted sum. This final weighted sum is the output of plete training set, respectively.) This allows the analyst
the network.The network is trained by adjusting all the to assess the possible impact of overtraining. (The
connecting weights in an iterative fashion. work was originally done on an Intel 386SX-20–based
PC, and the “time” to train for 100,000 iterations was
approximately 20 minutes.)
THE NEURAL NETWORK MODEL
The neural network was trained using BrainMaker.9
For this study, the step size was set to 0.500 and the RESULTS
training tolerance to 0.001. (Other packages might The MAD, MAPE, and MSE for the various neural
require the user to specify a learning rate and a network models are presented in Table 10-5. They
momentum term.) As in Case 9-4, the first 42 observa- do not compare favorably with the AR(2) model.
tions are used to train the network; the last 12 are used The author of this case is currently experimenting
to assess the performance of the network predicting with a different type of neural network—a radial
one step ahead. One problem with using neural net- basis function neural network—which produces
works to forecast time series data lies in determining results comparable to those of the AR(2) model.10

QUESTIONS
1. Find an article that describes an application of forecasts than the 2–4–1 network presented
neural networks to time series forecasting. What here.
method did the authors use, and how successful 3. Why are neural networks viewed as a viable
was it? alternative to the other forecasting methods dis-
2. If you have access to a neural network pro- cussed in this text?
gram, try to find a network that produces better

9BrainMaker, a commercially available PC-based neural network simulation package, was developed by
California Scientific Software in 1990.
10For further reading on neural networks, see Haykin (1998) and Khanna (1990).
 CHAPTER 10 Judgmental Forecasting and Forecast Adjustments 501

References
Armstrong, J. S. “Combining Forecasts: The End of Kudlow, L. “The Case for Market-Based
the Beginning or the Beginning of the End?” Forecasting.” CATO Journal 12 (1) (Spring 1992):
International Journal of Forecasting 5 (4) (1989): 119.
585–592. Law, A. M., and W. D. Kelton. Simulation Modeling
Balkin, S. D., and J. K. Ord. “Automatic Neural and Analysis, 3rd ed. Boston: McGraw-Hill, 2000.
Network Modeling for Univariate Time Series.” Lawerence, M., P. Goodwin, M. O’Conner, and
International Journal of Forecasting 16 (4) (2000): D. Onkal. “Judgmental Forecasting: A Review of
509–515. Progress over the Last 25 Years.” International
Bopp, A. E. “On Combining Forecasts: Some Journal of Forecasting 22 (3) (2006): 493–518.
Extensions and Results.” Management Science 31 Levenbach, H., and J. P. Cleary. Forecasting Practice
(12) (December 1986): 1492–1497. and Process for Demand Management. Belmont,
Bunn, D., and G. Wright. “Interaction of Judgmental Calif.: Thomson Brooks/Cole, 2006.
and Statistical Forecasting Methods: Issues and Mahoud, E. “Combining Forecasts: Some
Analysis.” Management Science 37 (5) (May Managerial Issues.” International Journal of
1991): 501–516. Forecasting 5 (4) (1989): 599–600.
Clemen, R. T. “Combining Forecasts: A Review and Makridakis, S. “The Art and Science of
Annotated Bibliography.” International Journal Forecasting.” International Journal of Forecasting
of Forecasting 5 (4) (1989): 559–583. 2 (1986): 15–39.
Denton, J. W. “How Good Are Neural Networks for Newbold, P., and T. Bos. Introductory Business and
Causal Forecasting?” Journal of Business Economic Forecasting, 2nd ed. Cincinnati, Ohio:
Forecasting 14 (2) (Summer 1995): 17–20. South-Western, 1994.
Dull, R. “Delphi Forecasting: Market Research for Parente, F. J., and J. K. Anderson-Parente. “Delphi
the 1990’s.” Marketing News, August 1988, 17. Inquiry Systems.” In Judgmental Forecasting,
Edmundson, R. H. “Decomposition: A Strategy for G. Wright and P. Ayton, eds. New York: Wiley,
Judgmental Forecasting.” Journal of Business 1987, 129–156.
Forecasting 9 (Summer 1990): 305–315. Price Waterhouse World Technology Centre.
Georgoff, D. M., and R. G. Murdick. “Manager’s Technology Forecast. Menlo Park, Calif.: Price
Guide to Forecasting.” Harvard Business Review 1 Waterhouse World Technology Centre, 2006.
(January–February 1986): 110–120. Ross, S. Simulation, 4th ed. New York: Academic
Hammond, J. S., R. L. Keeney, and H. Raiffa. Smart Press, 2006.
Choices: A Practical Guide to Making Better Rowe, G., and G. Wright. “The Delphi Technique as
Decisions. Boston: Harvard Business School a Forecasting Tool: Issues and Analysis.”
Press, 1999. International Journal of Forecasting 15 (4) (1999):
Haykin, S. Neural Networks: A Comprehensive 353–375.
Foundation, 2nd ed. Upper Saddle River, N.J.: Seila, A. F., V. Ceric, and P. Tadikamalla. Applied
Prentice Hall, 1998. Simulation Modeling. Belmont, Calif.: Thomson
Jain, C. L. A Managerial Guide to Judgmental Brooks/Cole, 2003.
Forecasting. New York: Graceway, 1987. Walden, M. “How to Evaluate and Improve a
Jain, C. L. “Myths and Realities of Forecasting.” Forecasting Process.” Journal of Business
Journal of Business Forecasting 9 (Fall 1990): 18–22. Forecasting 15 (2) (Summer 1996): 22.
Khanna, T. Foundations of Neural Networks. Willard, T. “Forecasting: A Key to Business
Reading, Mass.: Addison-Wesley, 1990. Success.” Futurist (July–August 1991): 33–34.
This page intentionally left blank
C H A P T E R

11 MANAGING THE FORECASTING

PROCESS
Chapter 1 introduced the subject of forecasting with suggestions about its usefulness
within the modern organization. Organizations today must develop forecasts to make
timely decisions in the face of uncertainty. This process of making educated guesses
about the uncertain future (forecasting) usually involves a rational process of extend-
ing historical data and experiences into the future. It is now possible to review the basic
notions and important points made in Chapter 1 with the benefit of the numerous fore-
casting techniques covered in the intervening chapters.

THE FORECASTING PROCESS

The forecasting process consists of two distinct phases. One phase occurs on a strategic
level, and the other phase occurs on an operational level. On the strategic level, deci-
sions include what to forecast, how to use the forecasts, and who will be responsible for
generating the forecasts. The operational phase of forecasting consists of data gather-
ing, forecast generation, and the ongoing evaluation of the forecasts. The forecasting
process is like any other process: Left unmanaged and unchecked, it is likely to spiral
out of control.
One key point discussed in Chapter 1 is the importance of using good management
judgment along with quantitative techniques in developing good forecasts. Good judg-
ment is indeed important and can now be illustrated using several of the forecasting
techniques discussed in previous chapters. A judicious mixture of quantitative tech-
niques with common sense is always necessary if forecasts are to be accurate, under-
stood, and used by the firm’s decision makers.
Time series decomposition is a good example of a technique involving the neces-
sity of using sound judgment along with an analysis of past history. A company using a
time series analysis program with, say, monthly sales data that yields a historical analy-
sis of trend-cycle variation, seasonal variation, and irregular movements might recom-
bine the first two factors to produce a forecast of unit sales. Appropriate choice of the
trend function and the nature of the recombination require considerable judgment as
to their future behavior. In using the decomposition process, it is also important to
decide how many historical periods to include in the analysis.
Regression analysis requires judgment along with statistical analysis whenever
forecasting takes place. If a multiple regression were conducted using employee job
performance rating as the dependent variable and two variables—entry test score and
age—as the predictor variables, an R2 value of 60% might be obtained. In addition, the
t values for the predictor variables might both be significant along with the regression
F value. The forecaster is then tempted to measure the two-predictor variables on each
503
504 CHAPTER 11 Managing the Forecasting Process

job applicant and use them to predict job performance. However, three additional
questions need to be considered. First, is the 60% explained variation sufficient for
forecasting purposes? Perhaps “intuitive judgment” on the desirability of hiring a per-
son is a superior method, or perhaps more precision in the forecasting process is
needed and other predictor variables should be considered. Second, can it be assumed
that future job applicants are essentially identical to those sampled in the regression
study? If they differ in any substantive way, the forecasting model may not be valid.
Finally, is the cost of the forecasting process justified in terms of benefit received? The
company test may be expensive, especially if purchased from an outside testing agency,
and must be justified by the benefits of the forecast.
Regression of time series data is a common occurrence in organizations that track
important measures of performance on a weekly, monthly, or quarterly basis. As auto-
correlation is a common occurrence in such studies, an understanding of this condition
and its treatment becomes vital if the results of such analyses are to be valid in the
decision-making process. Unfortunately, such understanding is often lacking; this
shortcoming has become an increasing problem with the advent of inexpensive regres-
sion analysis software.
The Box-Jenkins techniques discussed in Chapter 9 illustrate a common problem
in forecasting discussed in Chapter 1. These procedures are often superior forecasting
methods, producing smaller forecasting errors in many complex situations. Their disad-
vantage is that some sophistication is required on the part of the user. If the managerial
decision maker cannot understand the process used to generate the forecasts, the fore-
casts may be disregarded regardless of their precision.
The short-, medium-, and long-term aspects of forecasting techniques as they relate
to different levels of management in a firm can be illustrated with time series analysis
and technological forecasting. First- and second-line management in a firm might be
interested in a time series analysis of monthly unit sales, using smoothing and/or
ARIMA models, with data collected over the past four years. Forecasts produced by
these procedures, perhaps modified by judgment, might be used to plan monthly unit
production for the next fiscal year. Midlevel managers might use the same time series
procedures, supplemented by a separate trend analysis, to analyze annual unit sales data
over the past eight years and to forecast annual sales five years into the future. In this
case, the objective might be to plan capital expenditure needs for the business during
the approaching five-year period. At the same time, top management might be engaged
in technological forecasting using the Delphi method, along with scenario writing. Their
purpose would be to evaluate the company’s current position in the market and to
search for technology or societal changes that would threaten its market niche over the
next 20 years or for opportunities not evident in day-to-day operations.
The data analysis techniques discussed in this book are summarized in Table 11-1.
This table provides descriptions, points out applications, estimates cost levels, and indi-
cates whether computer capabilities are necessary for the implementation of each
technique. Each technique is also referenced to the chapters in which it is discussed.
Summaries such as those in Table 11-1 should be viewed as guidelines and not as defi-
nite statements that cannot be challenged.

MONITORING FORECASTS
After the forecasting objectives are established, collecting data and selecting an
acceptable forecasting technique are among the next steps in implementing an effec-
tive, ongoing forecasting effort. Several additional steps in the operational phase of the
 CHAPTER 11 Managing the Forecasting Process 505

TABLE 11-1 Forecasting Models

Computer

Method Description Applications Cost Necessary? Chapter

Causal Forecasting Models

Regression Explanatory Short- and Low to Usually 6
analysis forecasting; medium-range medium
assumes a forecasting of existing
cause-and-effect products and services;
relationship marketing strategies, pro-
between the input duction, personnel hiring,
to a system and its and facility planning
output
Multiple Explanatory Same as above Low to Yes 7
regression forecasting; medium
assumes a
cause-and-effect
relationship
between more than
one input to a
system and its
output
Time Series Forecasting Models
Decomposition Explanatory Medium-range Low to Yes 5
method forecasting; forecasting for medium
assumes a new plant and
cause-and-effect equipment
relationship planning, financing,
between time and new-product
the output of a development, and
system; the system new methods of
is decomposed into assembly;
its components short-range
forecasting for
personnel,
advertising,
inventory,
financing, and
production
planning
Moving Used to eliminate Short-range forecasting Low No 4 and 5
averages randomness in a for operations such as
time series; forecast inventory, scheduling,
is based on projection control, pricing,
from time series and timing of
data smoothed by a special promotions;
moving average used to compute
both the seasonal
and the cyclical
components for
the short-term
decomposition method
506 CHAPTER 11 Managing the Forecasting Process

TABLE 11-1 (Continued)

Computer

Method Description Applications Cost Necessary? Chapter

Exponential Similar to moving Short-range fore- Low Yes 4

smoothing averages, but casting for operations
values are weighted such as inventory,
exponentially, scheduling, control,
giving more weight pricing, and timing
to the most recent data of special promotions

Autoregressive Employed with Short- and Medium Yes 8 and 9

models economic variables medium-range
to account for forecasting for
relationships economic data—
between adjacent including price,
observations in a inventory,
time series production, stocks,
and sales—ordered
in a time series

Box-Jenkins Does not assume any Same as above High Yes 9

techniques particular pattern in
the historical data
of the series to be
forecast; uses an
iterative approach
of identifying and
fitting a possibly
useful model from a
general class of
models
Neural Uses a sophisticated Increasing use High Yes 10
networks computer program in a variety of
to assimilate forecasting
relevant data and applications; is in
recognize patterns the development phase
by “learning” as
humans do

forecasting process have been described in this book, with an emphasis on learning the
techniques commonly used to generate the actual forecasts. The key operational steps
in the forecasting process are summarized in Figure 11-1.
The collection and examination of appropriate historical data were described ear-
lier in this book (Chapter 3), as were considerations in selecting a forecasting tech-
nique or model. As suggested by Figure 11-1, the next step is usually to forecast several
historical periods where the actual values are known. The resulting errors can be sum-
marized in several ways, as discussed in Chapter 3, and this process is continued until a
technique with a sufficient cost-benefit ratio is found. The model is then used to fore-
cast future periods, and the results are incorporated into the firm’s decision-making
process.
 CHAPTER 11 Managing the Forecasting Process 507

Collect Data

Examine Data Patterns

Choose Forecasting Method

Forecast Past Periods

Acceptable
Accuracy?
No Yes

Reexamine Forecast Future Periods

Data Patterns and Use Results in
Decision-Making Process

From Time to Time,

Check Accuracy of
Forecasts

Yes

Acceptable
Accuracy?

No

Examine Data Patterns

Using Updated
Historical Values

FIGURE 11-1 The Operational Phase of

the Forecasting Process

From time to time, it is necessary to pause in the forecasting process and recon-
sider the procedures being used. The usual steps are as follows:
1. The oldest historical values in the data being used by the forecasting technique are
discarded, and the most recent actual values are added to the data bank.
2. Following this data update, the parameters used in the forecasting model are rees-
timated. For example, the best value(s) of the weighting constant(s) used in expo-
nential smoothing may shift, possibly considerably, when more recent data values
are added. Or the coefficients in a regression analysis can change when different
data are used to fit the regression function.
508 CHAPTER 11 Managing the Forecasting Process

3. The forecasting model with new parameters is examined for adequate accuracy. If
this accuracy is judged to be sufficient, the model is then used as before until the next
update period. If forecasting accuracy is deemed inadequate or marginal, the pat-
terns in the new data can be examined with the possibility of choosing a new fore-
casting procedure. This process continues until the accuracy of the chosen model, as
judged by a comparison of forecasts with actual values, is deemed to be adequate.
The preceding process is summarized in the flow diagram of Figure 11-1 and consti-
tutes the kind of feedback loop commonly found in system designs of all types.
Forecasts are sometimes monitored constantly using a tracking signal, a concept
discussed in Chapter 4 following the material on simple exponential smoothing. The
idea is to establish limits within which the errors generated by the forecasts are
expected to fall if the forecasting process is adequate. As long as the errors fall within
these acceptable limits, the forecasting process continues. As soon as an error falls out-
side the acceptable range, management attention is focused on the forecasting process,
and the updating and revision steps outlined earlier are undertaken. This concept is
illustrated in Example 11.1.
Example 11.1
Sue Bradley is responsible for forecasting the monthly dollar sales of her company. Sue has
chosen a forecasting model that has an error rate acceptable to her managers. Specifically,
the standard error of this forecasting process is $935; that is, the forecast and actual values of
monthly dollar sales are typically $935 apart.
Sue assumes that forecast errors are normally distributed with a mean of zero and a
standard deviation of $935. She makes this assumption after examining a plot of past fore-
casting errors and finds that they follow a bell-shaped curve about zero. Using a 95% confi-
dence level, she then establishes the following limits within which she expects the forecast
error of each month to fall:
0 ; 11.96219352
0 ; 1, 833
Thus, Sue expects each monthly forecast to be within $1,833 of the actual value for the
month, with 95% confidence. If it is, the forecasting procedure will continue without her
attention. But if the error should be greater than $1,833, she will examine both the parame-
ters in her chosen forecasting technique and even consider using another technique.
To monitor the forecasting errors more easily, Sue designs a chart to track them. Over
the course of several months, Sue finds two plots that cause her to closely examine her fore-
casting procedures. The first (Figure 11-2) shows forecasting errors that appear to be

2,400
*

1,200
* *
Error

0 * *
*
*
*
*
−1,200
*
0.0 2.0 4.0 6.0 8.0 10.0
Time

FIGURE 11-2 First Data Error Plot for Example 11.1

 CHAPTER 11 Managing the Forecasting Process 509

2,400

Error
* * *
0 *
* *
* *
−1,200
* *
*
2.0 4.0 6.0 8.0 10.0 12.0
Time

FIGURE 11-3 Second Data Error Plot for Example 11.1

randomly distributed until the most recent period. This out-of-tolerance error leads Sue to
reestimate the parameters in her forecasting model after updating her database by adding
recent values and discarding the same number of older values. Some time later a second
error plot (Figure 11-3) again causes Sue to look at her forecasting process. Although none
of the errors has exceeded her tolerance limits, Sue notes that recent errors do not appear to
be randomly distributed. In fact, the errors are increasing in the positive direction, and it
seems obvious that they will soon be out of control. Sue updates her database and, after
carefully examining the data patterns, chooses a new forecasting technique.

FORECASTING STEPS REVIEWED

In Chapter 1, the point was made several times that sound judgment must be con-
stantly exercised along with quantitative analysis if useful and accurate forecasting is
to take place. Several questions were listed that should be raised if the forecasting
process is to be properly managed. They are repeated here:
• Why is a forecast needed?
• Who will use the forecast, and what are their specific requirements?
• What level of detail or aggregation is required, and what is the proper time horizon?
• What data are available, and will the data be sufficient to generate the needed
forecast?
• What will the forecast cost?
• How accurate can we expect the forecast to be?
• Will the forecast be made in time to help the decision-making process?
• Does the forecaster clearly understand how the forecast will be used in the
organization?
• Is a feedback process available to evaluate the forecast after it is made and adjust
the forecasting process accordingly?
When the preceding questions have been properly addressed and answered, the
actual forecasting process can begin. Chapter 1 listed the steps followed in such a
process:
Problem formulation and data collection
Data manipulation and cleaning
Model building and evaluation
510 CHAPTER 11 Managing the Forecasting Process

Model implementation (the actual forecast)

Forecast evaluation
To this list of steps might be added another: Feedback collection after the forecasting
process is under way to determine if sufficient accuracy has been achieved and if man-
agement is finding the forecast useful and cost effective in the decision-making
process.
Although the primary concern of this book has been with model building or selec-
tion of the appropriate forecasting method, the importance of managing the forecast-
ing process has been emphasized as well. The questions listed previously are important
in all forecasting situations and must be addressed if useful results are to be obtained.
We especially recommend study of the cases at the end of this chapter. They are
designed to emphasize the managerial aspects of forecasting.

FORECASTING RESPONSIBILITY

The location of the forecasting process within a firm varies depending on the size of
the firm, the importance attached to formal forecasting, and the nature of the firm’s
management style. The forecasting responsibility falls somewhere on the continuum
between a separate forecasting department and forecasting within small management
units without reference to other efforts within the firm.
Forecasting staffs are more common in large organizations than in small ones.
Large firms can afford to hire the experts needed for sophisticated forecasting and can
equip their staffs with modern computing and software capabilities. The advantage of
such a centralized effort is that such expertise is available to all units of the organiza-
tion. The disadvantage is that coordination between the forecasting staff and line man-
agers is often quite difficult to achieve. The forecasting staff may find itself spending
more time negotiating with users and explaining its role than in actual forecasting.
At the other extreme is the location of the forecasting process within each unit of
the firm without coordination or cooperation across units. The advantage of this
process is that there is no misunderstanding between those forecasting and those using
the forecasts: They are the same people. Forecasts generated under these conditions
tend to be accepted and used in the decision-making process. The disadvantage is that
sophistication and therefore forecasting accuracy may be difficult to achieve because
computer/software availability and technical expertise may not be uniformly spread
across many users. It is usually difficult to persuade top management to acquire appro-
priate hardware and software, for example, when they know they exist in other loca-
tions in the company.
Moreover, lack of coordination and cooperation can lead to problems when one
unit uses forecasts generated by another unit in the company. If units are not held
properly accountable for the quality of the forecasts, biases can result that lead to oper-
ational difficulties.
In one forecasting situation, the marketing organization owned the forecasts.
Marketing was measured by product availability, which was defined by the percentage
of time that a particular product was available for purchase by a customer. However,
marketing was not penalized for any excess inventory that might develop as a result of
production’s reliance on marketing’s forecasts. Although those forecasts were based on
market conditions, they were at the same time heavily influenced by quota and rev-
enue expectations. Consequently, the marketing group consistently generated opti-
mistic demand forecasts to ensure the product availability that would enable it to meet
its revenue targets. On the other hand, the production planning unit was measured on
 CHAPTER 11 Managing the Forecasting Process 511
both product availability and excess inventory. This group used statistical methods to
buffer against uncertainties and was highly motivated to keep the forecasts accurate
and unbiased.
Many organizations attempt to locate the responsibility for forecasting midway
between the extremes just mentioned. A small staff of forecasters may be assigned to
subunits within the firm to service the needs of several functional areas. The task of
such a forecasting staff involves proper coordination with clients as well as the genera-
tion of accurate forecasts. Sometimes this forecasting responsibility is combined with
other staff functions such as statistical support or computing support.
The availability of inexpensive small computers and forecasting software has
tended to move the forecasting function downward in the organization. It is now possi-
ble for managers to have access to sophisticated forecasting tools at a fraction of the
cost of such capability just a few years ago. However, the knowledge required to prop-
erly use this capability does not come with the hardware or software package; the need
to understand the proper use of forecasting techniques has increased as the computing
capability has moved out of the hands of the “experts” into those of the users in an
organization.

FORECASTING COSTS
Computing hardware and software, plus staff, are the obvious costs involved with cre-
ating forecasts. But additional costs are involved that may not be as obvious due to the
expenditure of company personnel time as well as money. The time of salaried persons
in various departments spent in gathering data for the forecasting process, monitoring
the process, and interpreting the results must be considered a cost of forecasting. Such
cost must be balanced against the benefits received if rational decisions regarding the
usefulness of the resulting forecasts are to be reached.
An alternative to producing forecasts internally is to use consultants for this pur-
pose.This practice is especially appealing if the need for a forecast is a one-time require-
ment rather than an ongoing one. Also, a forecasting requirement beyond the technical
capability of company personnel suggests the use of professional consultants. Such out-
side hiring of forecasting assistance makes the identification of cost an easy matter.

FORECASTING AND MANAGEMENT INFORMATION SYSTEMS

The management information systems (MISs) of modern firms have increased in
sophistication and usefulness in recent years. Their primary benefit to the forecasting
process involves their enormous capability to collect and record data throughout the
organization. The forecaster must resist the temptation to collect data for the forecast-
ing models being used if such data have already been collected and recorded in the
company’s MIS system.
The rapid development of computer networks within many firms has also
increased the network users’ access to such stored data. In addition, large integrative
software programs such as SAP1 have forecasting modules that make use of databases
that can be accessed with the program.

1SAP, headquartered in Walldorf, Germany, is, at the time of this writing, the world’s largest interenterprise
software company.
512 CHAPTER 11 Managing the Forecasting Process

Reliance on existing databases is important even if the available data are not in
precisely the format or time sequence desired by the forecaster. Modifications in the
forecasting model or in the available data should be considered before abandoning the
MIS data in favor of collecting new data. This advice presumes, of course, that collect-
ing new data would involve a considerable expenditure of time and money. If the data
needed for a forecasting model were easy to obtain in the correct format, this would be
preferable to using precollected data in a database that are not in the proper form or
that are out of date.
An additional advantage to using data available on the company’s MIS or network
system is that the forecasting process then becomes a component of this system. As
such, it enters the distribution and decision-making network already established by the
system and may become more readily incorporated into the company’s decision-making
process. This situation is in contrast to a forecasting procedure that attempts to infiltrate
the decision-making procedures already in use by company managers.

SELLING MANAGEMENT ON FORECASTING

Several factors are important considerations if the forecasts generated within a firm
are to become important aspects of the decision-making process. First, it must be rec-
ognized that effective managers are interested in practical and helpful results. In gen-
eral, forecasts must meet the needs of such managers.
Second, forecasts must be accurate enough to be useful. Good managers will not
stay with a forecasting process long, regardless of its sophistication, if accurate results
are not generated.
Finally, the cost-benefit instincts of all good managers must be recognized in the
forecasting process. The ability to analyze situations in terms of cost incurred versus
benefit received is the keystone of an effective management process and must be rec-
ognized by the forecasting staff. This situation often creates difficulties between the
forecaster and the user. The forecaster must always keep in mind that the purpose of
the forecasting process is to improve the efficiency of the business and that includes
maximizing the bottom-line profit.

THE FUTURE OF FORECASTING

As mentioned in Chapter 1, forecasting continues to gain in importance in modern
organizations due to the increasing complexity of the business world, along with the
availability of less expensive and more powerful computing equipment and software.
The continuing competition in the small computer and software areas is obvious to
everyone; less obvious may be the long-term trends that slowly change the makeup of
the business scene and that exert subtle but powerful pressures on the ways business-
people operate.
Consider some of the “megatrends” identified by Naisbitt (1982) in his book by
that title (see several additional references of this kind in Chapter 10). These important
trends are now mostly in place and reinforce the need for effective forecasting:
Industrial to information society
National to world economy
Short- to long-term thinking
Centralization to decentralization
Either/or to multiple options
 CHAPTER 11 Managing the Forecasting Process 513
Today’s economy is an economy of fierce competition on a global basis. Due to the
World Wide Web, an abundance of information is available at any time of the day.
Changes in the marketplace are swift and sudden, and likely to remain so. Product life
cycles continue to shorten. Conducting business electronically (e-business) is a way of
life, and transactions over the Web are going to continue to increase.
These trends emphasize the increasing need for management to deal with complex
issues and, in particular, the need to develop sophisticated methods for dealing with
future uncertainties. They emphasize, in other words, the growing importance of com-
bining good judgment and sophisticated data manipulation methods into sound business
forecasting. As Naisbitt’s trends and the increasingly dynamic business environment
continue to unfold, the ability of business leaders to react quickly and profitably to
changing events is brought into sharper focus. The basic business question—“What will
happen next?”—will assume even greater importance; it is with this question that busi-
ness forecasting is concerned. New challenges and opportunities for forecasters await.

Problems
1. Write a short response to each of the following assertions:
a. Forecasts are always wrong, so why place any emphasis on demand planning?
b. A good forecasting process is too expensive.
c. All we need to do is hire a “quantitative type” to do our forecasting.
2. Can you think of any future business trends, not discussed in this chapter, that may
influence the way forecasting is done in the future? Explain briefly.

CASES

CASE 11-1 BOUNDARY ELECTRONICS

Boundary Electronics is a large supplier of elec- After thinking about this matter, Guy plans a
tronic products for home use. Among its largest sell- Saturday retreat for the six members of his top man-
ers are home video recorders and satellite television agement team. He rents a meeting room in a local
systems. Because the company’s business has grown hotel and arranges for lunch and coffee breaks for
so rapidly, Guy Preston, Boundary’s president, is the day. When the team meets Saturday morning, he
concerned that a change in market conditions could introduces the topic of the day and then instructs
alter its sales pattern. each person to prepare a one- or two-page descrip-
In asking his managers about the future of the tion of the company’s operating environment over
company, Guy has discovered two things. First, most the next 20 years for each of the following situations:
of his managers are too busy thinking about the day-
to-day problems of meeting growing demand to give 1. The company’s environment will continue
much thought to the long-range future. Second, their essentially as it is now. Products demanded by
opinions vary considerably from quite optimistic to the market will be modifications of current
quite pessimistic. As president of the company, Guy products, and no new technology will intervene.
feels he has an obligation to seriously consider the 2. Major technological changes will render the
future environment of his company. company’s current line of products obsolete.
514 CHAPTER 11 Managing the Forecasting Process

New products will have to be developed to During this hour, Guy thinks about the rest of the day
meet the leisure demands of the American and what will happen. He hopes that there will be
population. some provocative ideas developed by his managers
3. Between these two extremes, what is the most and that subsequent discussions will prove lively and
likely scenario for the company’s operating interesting. In addition to gaining ideas for his own use,
environment? Guy hopes that the day’s exercise will help his man-
agers look beyond the company’s immediate problems
Guy allows one hour for each team member to and opportunities and give them a more long-range
develop the scenarios for each of these three situations. view of the company.

QUESTIONS
1. What process do you think Guy should use after 3. Do you think Guy will accomplish his objectives
the hour’s writing activities have been completed? with the Saturday meeting?
2. Is there some other approach that Guy might
have tried, given his objectives?

CASE 11-2 BUSBY ASSOCIATES

Jill Tilly was a recent graduate of a university busi- well with this dependent variable. She ended up with
ness school when she took a job with Busby values for four variables for 14 quarters.
Associates, a large exporter of farm equipment. She then computed three additional variables
Busby’s president noticed a forecasting course on from her dependent variable values: change in Y,
Jill’s résumé during the hiring process and decided percent change in Y, and Y lagged one period. By
to start Jill’s employment with a forecasting project the time she began to think about various ways to
that had been discussed many times by Busby’s top forecast her variable, she had collected the data
managers. shown in Table 11-2.
Busby’s president believed there was a strong Jill keyed her data into a computer program
relationship between the company’s export sales that performed regression analysis and computed
and the national figures for exports. The national fig- the correlation matrix for her seven variables. After
ures were readily available from government examining this matrix, she chose three regressions
sources, so Jill’s project was to forecast a good, rep- with a single predictor variable and six regressions
resentative export variable. If this effort was suc- with two predictor variables. She then ran these
cessful, Busby’s president believed the company regressions and chose the one she considered best: It
would have a powerful tool for forecasting its own used one predictor (Y lagged one period) with the
export sales. following results:
At the time, Jill located the most recent copy of
r 2 = .98
the Survey of Current Business in a local library and
recorded the quarterly figures for consumer goods t = 25.9
exports in billions of dollars. She believed this vari-
F = 671.6
able was a good representative of total national
exports. Anticipating the possibility of forecasting Jill examined the residual autocorrelations and
using regression analysis, she also recorded values for found them to be small. She concluded she did not
other variables she thought might possibly correlate need to correct for any autocorrelation in her
 CHAPTER 11 Managing the Forecasting Process 515

TABLE 11-2 Quarterly Time Series Data

Variable

Time Period 1 2 3 4 5 6 7

1987
1 18.2 128.3 306.2 110.0 — — —
2 19.8 45.8 311.6 109.7 1.6 8.79 18.2
3 20.9 66.1 320.7 109.9 1.1 5.56 19.8
4 22.1 129.7 324.2 109.7 1.2 5.74 20.9
1988
1 24.0 136.4 331.0 109.4 1.9 8.60 22.1
2 26.0 140.7 337.3 110.5 2.0 8.33 24.0
3 27.7 156.9 342.6 110.6 1.7 6.54 26.0
4 29.7 148.5 352.6 110.9 2.0 7.22 27.7
1989
1 33.6 189.8 351.5 113.4 3.9 13.13 29.7
2 35.0 168.9 357.6 112.4 1.4 4.17 33.6
3 35.0 154.5 365.2 111.9 0.0 0.00 35.0
4 38.0 174.1 366.3 111.0 3.0 8.57 35.0
1990
1 40.7 191.3 369.1 111.9 2.7 7.11 38.0
2 42.0 201.2 370.0a 112.1 1.3 3.19 40.7
a
Estimated.
Variable key:
1: Consumer goods, exports, billions of dollars
2: Gross personal savings, billions of dollars
3: National income retail trade, billions of dollars
4: Fixed weight price indexes for national defense purchases, military equipment, 1982 = 100
5: Change in dependent variable from previous period
6: Percent change in dependent variable from previous period
7: Dependent variable lagged one period
Source for variables 1 through 4: Survey of Current Business (U.S. Department of
Commerce) 70 (7) (July 1990).

regression. Jill believed that she had found a good Table 11-3, lagged the data one period, and again ran
predictor variable (Y lagged one period). a regression analysis using Y lagged one period as
Jill realized that her sample size was rather the predictor variable.
small: 13 quarters. She returned to the Survey of This time she again found good statistics in her
Current Business to collect more data points and was regression printout, except for the lag 1 residual auto-
disappointed to find that during the years in which correlation. That value, .47, was significantly different
she was interested the definition of her dependent from zero at the 5% level. Jill was unsure what to do.
variable changed, resulting in an inconsistent time She tried additional runs, including the period number
series. That is, the series took a jump upward halfway and the change in Y as additional predictor variables.
through the period that she was studying. But after examining the residuals, she was unable to
Jill pointed out this problem to her boss, and it conclude that the autocorrelation had been removed.
was agreed that total merchandise exports could be Jill decided to look at other forecasting techniques to
used as the dependent variable instead of consumer forecast her new dependent variable: total merchan-
goods exports. Jill found that this variable remained dise exports. She used the time series data shown in
consistent through several issues of the Survey of the Y column of Table 11-3. The time series plot of
Current Business and that several years of data total merchandise exports for the years 1984 through
could be collected. She collected the data shown in the second quarter of 1990 is shown in Figure 11-4.
516 CHAPTER 11 Managing the Forecasting Process

After studying Figure 11-4, Jill decided to use only

TABLE 11-3 Quarterly Time Series Data: Total
Merchandise Exports ($ billions) the last 16 data points in her forecasting effort. She rea-
soned that beginning with period 9 the series had
Year Quarter Y Y Lagged One Period shown a relatively steady increase, whereas before that
period it had exhibited an increase and a decline.
1984 1 219.3 —
2 223.1 219.3
An examination of her data revealed no season-
3 225.9 223.1 ality, so Jill decided to try two smoothing methods:
4 228.0 225.9
1985 1 225.0 228.0 Simple exponential smoothing
2 221.6 225.0 Holt’s linear exponential smoothing, which can
3 218.0 221.6 accommodate a trend in the data
4 218.6 218.0
1986 1 220.7 218.6 For simple exponential smoothing, Jill restricted
2 221.4 220.7
the smoothing constant, a, to lie between 0 and 1
3 225.7 221.4
4 230.4 225.7 and found the best a to be very close to 1. The
1987 1 234.5 230.4 Minitab program selected the optimal values for the
2 246.6 234.5 Holt smoothing constants, a and b. Error measures
3 261.6 246.6 and the choices for the smoothing constants are
4 281.3 261.6 shown in Table 11-4. Given the results shown in
1988 1 306.7 281.3 Table 11-4, Jill decided Holt’s linear smoothing
2 319.2 306.7
offered the most promise and used this procedure to
3 327.9 319.2
4 342.8 327.9 generate forecasts for the next four quarters.
1989 1 360.6 342.8 Jill noted that the optimum smoothing constant
2 373.2 360.6 using simple exponential smoothing was almost 1.00
3 367.3 373.2 (.99). Apparently, in order to track through the data
4 378.7 367.3 in this fashion, the program was basically using each
1990 1 394.2 378.7 data value to predict the next. This is equivalent to
2 394.4 394.2
using a simple naive method to forecast—that is, a
Source: Survey of Current Business, various issues. model that says consecutive differences are random.

FIGURE 11-4 Plot of Quarterly Data Values: Total Merchandise

Exports, First Quarter of 1984 to Second Quarter
of 1990 ($ billions)
 CHAPTER 11 Managing the Forecasting Process 517
Jill realized that with each passing quarter a new
TABLE 11-4
actual value of total merchandise exports would be
Optimum available and that the forecasts for future periods
Smoothing could be updated.
Technique MSE MAPE Constant(s) Jill then met with her boss to discuss her results.
Simple exponential 203.7 4.2 a = .99 She indicated that she thought she had a good way
Holt’s linear 35.8 1.5 a = 1.77, b = .14 to forecast the national variable, total merchandise
exports, using exponential smoothing with trend
Using Holt’s smoothing method, the forecasts adjustments. Her boss asked her to explain this
for the next four quarters beyond the end of her method, which she did. Her next assignment was to
data are use actual data to verify the hunch of Busby’s presi-
dent: that Busby’s exports were well correlated with
Period Forecast national exports. If she could establish this linkage,
Busby would have a good forecasting method for its
17 400.2 exports and could use the forecasts to plan future
18 408.8 operations.
19 417.4
20 425.9

QUESTIONS
1. Jill did not consider combining the forecasts 3. It’s possible that the choice of forecasting
generated by the two methods she analyzed. method could shift to another technique as new
How would she go about doing so? What would quarterly data are added to the database.
be the advantages and disadvantages of such Should Jill rerun her entire analysis once in a
action? while to check this? If so, how often should she
2. The optimum smoothing constants used by do this?
Holt’s linear exponential smoothing were 4. A colleague of Jill’s suggested she try the Box-
a = 1.77 and b = .14. As new data come in over Jenkins ARIMA methodology. What advice do
the next few quarters, Jill should probably rerun you have for Jill if she decides to try Box-
her data to see if these values change. How Jenkins? Can you identify a tentative ARIMA
often do you think she should do this? model?

CASE 11-3 CONSUMER CREDIT COUNSELING

The Consumer Credit Counseling (CCC) operation Dorothy, with your help, has tried several ways
was described in Cases 1-2 and 3-3. to forecast the most important variable. These
The executive director, Marv Harnishfeger, con- efforts are outlined in Cases 4-3, 5-3, 6-5, 8-5, and
cluded that the most important variable that CCC 9-3. Having completed these forecasting attempts,
needed to forecast was the number of new clients that Dorothy decides it is time to summarize these
would be seen for the rest of 1993. Marv provided efforts and to attempt to arrive at a method for fore-
Dorothy Mercer with monthly data for the number of casting the rest of the year.
new clients seen by CCC for the period from January
1985 through March 1993 (see Case 3-3).
518 CHAPTER 11 Managing the Forecasting Process

ASSIGNMENT
1. Assume that Dorothy assigns you to help her possible and that he can use in the everyday
with this forecasting problem. Write a report decision making for the organization. Be spe-
that recommends a course of action. Keep in cific about what you are recommending to
mind that Marv must develop forecasts for the Dorothy and Marv. Remember to consider the
number of clients seen that are as accurate as issues discussed in this chapter, such as cost.

CASE 11-4 MR. TUX

The owner of several Mr. Tux rental outlets, John measurements were unsatisfactory. He realized
Mosby, has tried several ways to forecast his most that these measurements, such as the average
important variable, monthly dollar sales. His efforts error and average percentage error, resulted
are outlined in Cases 1-1, 2-2, 3-2, 4-2, 5-2, 6-4, 8-4, from predicting past values of his variable. But
and 9-2. Having completed these forecasting attempts, since they were so high, he didn’t want to use
John decides it is time to summarize his efforts and these techniques to predict the unknown future.
attempt to arrive at a method for forecasting the • Case 5-2: John finally got some encouraging
future. He realizes that he should update both his results using the decomposition method to con-
data and his method of forecasting at some time in struct a trend line, seasonal indexes, and a cycli-
the future, but he needs to choose a way of forecast- cal component for his data. He was able to show
ing the next few months right away. his banker the seasonal indexes and make
To begin, John summarizes the results of the desired arrangements on his loan payments. He
methods he has tried so far: also generated forecasts for the next few
months by reassembling his estimated compo-
• Case 2-2: Using the annual average to forecast nents. However, John was somewhat disturbed
future annual sales might work, but since John by the wide range of his projections.
noticed an upward trend, he needs a sound way • Case 6-4: Simple regression analysis was the next
of extending these averages into the future. technique John tried, using the time period as
Also, John is quite concerned with the seasonal the independent or predictor variable. He rea-
effect, since he knows his sales vary consider- soned that this variable would account for the
ably by month. His efforts using annual aver- trend factor he knew was in his data. This
ages are not fruitful. method did not account for the seasonality in
• Case 3-2: John’s use of the Minitab program sales, and the r 2 value of 56% was unsatisfactory.
established that both a trend and a seasonal • Case 8-4: John next tried a multiple regression
effect existed in his data. Although he knew using both a time period number to reflect trend
these elements were there before he started, he and a series of dummy variables to account for
was pleased to see that the results from his com- the seasonal effect (months). His R2 value of
puter program established them statistically. 88% was a considerable improvement over his
The program also indicated that several auto- simple regression, but the forecast error for the
correlation coefficients were outside sampling last 12 months of his data, as measured by the
error limits, indicating to John that both trend mean absolute percentage error (MAPE), of
and seasonal effects needed to be reflected in 21% was unacceptable. With this kind of error,
his final forecasting model. John decided not to use multiple regression. He
• Case 4-2: When John used exponential smooth- also tried a seasonal autoregressive model,
ing, including methods that took account of which resulted in an R 2 value of 91%. John was
trend and seasonal factors, the resulting error quite pleased with this result.
 CHAPTER 11 Managing the Forecasting Process 519
• Case 9-2: The Box-Jenkins ARIMA methodol- In thinking over these efforts, John realized that
ogy bothered John from the outset because he time was running out and that he must generate
did not totally understand it. He recognized, forecasts of his monthly revenues soon. He had lim-
however, that his seasonal autoregressive model ited time to try modifications to the methods he
was a particular ARIMA model. He wanted to used and could think about combining two or more
know if he could improve on this model. He methods. But he did not have time to acquire new
knew he would have to explain whatever fore- software and try completely different methods. As
casts he came up with to investors and bankers he wondered what he should do, he looked at one of
in his attempt to gain capital for expansion, so his favorite sayings posted on his office wall: “Let’s
he wanted a forecasting method that was both do something, even if it’s not quite right.”
accurate and understandable.

ASSIGNMENT
1. Assume you have been hired to help John forecasts for monthly sales that are as accurate
Mosby with his forecasting problem. Write him as possible and that he can use in discussions
a memo that summarizes his efforts to date and with investors. Be very specific about what you
that recommends a course of action to him. are recommending to the owner of Mr. Tux.
Keep in mind that John must quickly develop

CASE 11-5 ALOMEGA FOOD STORES

Example 1.1 described how Julie Ruth, the president several potential candidates for the job and that
of Alomega Food Stores, collected monthly sales there might be some resentment among her man-
data for her company along with several other vari- agement team. She was especially sensitive to the
ables she thought might be related to sales. Cases negative comments made by Jackson Tilson, her
2-3, 3-4, 5-6, 8-7, and 10-2 described Julie’s efforts to production manager, during a recent meeting (see
use various Minitab procedures to make meaningful the end of Example 1.1).
forecasts of monthly sales. In reviewing her efforts, Julie decided to dis-
Julie knew that her technical staff were using card the simple regression analysis she performed
the same data to generate good forecasts but didn’t as summarized in Case 2-3. She was left with a
know how they were coming along. Besides, she choice among the decomposition analysis described
really wanted to come up with a good forecasting in Case 5-6, the multiple regression described in
method on her own. She knew that as the first Case 8-7, and a combination of methods described
woman president of Alomega, she had jumped over in Case 10-2.

QUESTIONS
1. Suppose you were recently hired by Alomega 2. Based on your choice in Question 1, write a
Food Stores and assigned to assist Julie in devel- detailed memo to Julie outlining your reasons
oping an effective forecasting method for for your choice, and indicate the extent to which
monthly sales. After reviewing Julie’s efforts to you think this forecasting method would be
date, which of the methods she tried would you effective.
recommend to her?
520 CHAPTER 11 Managing the Forecasting Process

3. In addition to the methods Julie was considering, software package to try other methods and com-
what other forecasting procedures would you pare them with your choice in Question 1.
suggest? Use Minitab or another forecasting 4. Should Julie combine forecasts?

CASE 11-6 SOUTHWEST MEDICAL CENTER

Mary Beasley is not very happy. She has tried sev- 1999–2000—three years in the middle of her data—
eral time series methods to develop forecasts for the is different from that for the remaining years.
monthly totals of billable visits to Medical Oncology Monthly total visits are generally larger and more
(see Cases 4-7, 5-8, 8-9, and 9-10). She has had the volatile for these middle fiscal years. It’s as if this
greatest success with an ARIMA model, although piece of the time series has been “lifted up” from the
the residual autocorrelations as judged by the rest of the series. Mary is not sure why that should
Ljung-Box Q statistic (see Equation 9.5) seem to be be the case and wonders if maybe there was a
collectively large. Examining a plot of her series (see change in procedure, or some other intervention,
Figure 4-23), Mary notices that the behavior of total that was responsible for changing the character of
visits for fiscal years 1997–1998, 1998–1999, and the time series during these middle years.

ASSIGNMENT
1. You have been hired to help Mary with her current procedures? If so, how should she “sell”
forecasting problem. Write her a memo that the forecasts to management? Given the some-
summarizes her efforts to date and that recom- what inconsistent character of her time series,
mends a course of action to her. Be sure your should Mary search for another forecasting
memo addresses the following issues: Should technique? Should she use one of her current
Mary accept the forecasts from one of her methods but with a modified data set?

References
Armstrong, J. S., ed. Principles of Forecasting: Naisbitt, J. Megatrends. New York: Warner Books,
A Handbook for Researchers and Practitioners. 1982.
Norwell, Mass.: Kluwer, 2001.
Appendix A

Data for Case 7-1

Y X1 X2 X3 X4 X5 X6 Y X1 X2 X3 X4 X5 X6
14.75 0.00 0.00 2.01 12.57 10.00 20.50 16.13 0.00 0.00 2.80 13.39 10.00 16.50
14.50 0.00 0.00 2.53 12.57 10.00 20.00 15.75 1.00 0.00 4.00 13.35 30.00 16.00
14.13 1.00 0.00 2.10 12.57 10.00 20.00 16.13 0.00 0.00 2.81 13.50 10.00 15.80
14.63 0.00 1.00 4.13 12.14 30.00 20.00 16.25 1.00 0.00 3.38 13.50 30.00 15.80
14.00 1.00 0.00 2.10 12.57 10.00 20.00 16.00 0.00 0.00 2.57 13.50 10.00 15.80
13.38 0.00 1.00 3.97 12.57 10.00 20.00 15.88 0.00 1.00 3.96 13.50 30.00 15.80
14.57 0.00 1.00 3.27 12.14 30.00 20.00 16.50 1.00 0.00 2.67 13.50 30.00 15.80
13.88 1.00 0.00 3.50 13.19 10.00 19.50 16.38 1.00 0.00 3.05 13.50 30.00 15.80
15.38 0.00 0.00 2.85 13.19 10.00 19.50 12.50 1.00 0.00 2.36 10.60 30.00 15.30
15.63 0.00 0.00 1.81 13.12 10.00 18.50 12.25 1.00 0.00 2.54 10.60 30.00 15.30
15.88 1.00 0.00 2.72 12.69 30.00 18.50 14.25 1.00 0.00 2.20 12.13 30.00 15.30
15.00 1.00 0.00 2.43 13.12 10.00 18.00 15.00 1.00 0.00 3.03 12.13 30.00 15.80
16.13 0.00 0.00 3.27 12.69 30.00 18.00 15.25 1.00 0.00 3.24 12.13 30.00 16.50
15.25 0.00 1.00 3.13 12.69 30.00 17.50 16.00 0.00 0.00 1.95 12.34 30.00 17.80
16.00 0.00 0.00 2.55 13.68 10.00 17.00 14.88 1.00 0.00 2.86 12.34 30.00 17.80
16.25 0.00 0.00 2.08 13.68 10.00 17.50 14.75 1.00 0.00 2.64 12.34 30.00 19.00
17.38 0.00 0.00 2.12 13.20 30.00 17.50 15.50 1.00 0.00 2.23 11.40 30.00 20.00
16.35 1.00 0.00 3.40 14.10 10.00 19.00 13.75 1.00 0.00 2.24 11.40 30.00 19.50
17.00 1.00 0.00 2.63 13.60 30.00 19.00 11.30 1.00 0.00 3.24 11.36 30.00 17.50
16.00 0.00 1.00 2.61 14.10 10.00 19.50 12.38 1.00 0.00 1.95 11.36 30.00 17.50
16.63 1.00 0.00 2.06 14.10 10.00 19.50 12.15 1.00 0.00 2.32 11.36 30.00 14.50
16.38 0.00 0.00 2.08 14.10 10.00 20.00 11.75 1.00 0.00 2.45 9.81 30.00 13.00
16.75 1.00 0.00 2.09 13.60 30.00 20.00 12.38 1.00 0.00 1.88 9.81 30.00 13.00
15.13 0.00 1.00 4.29 12.69 30.00 20.00 12.63 0.00 0.00 1.76 9.81 30.00 13.00
16.00 1.00 0.00 2.50 12.96 30.00 20.00 11.13 1.00 0.00 1.99 9.81 30.00 12.50
14.50 0.00 1.00 3.32 13.47 10.00 20.00 11.38 0.00 0.00 2.20 9.78 10.00 12.50
16.25 0.00 0.00 2.95 12.96 30.00 20.00 11.88 1.00 0.00 2.14 9.81 30.00 12.00
16.88 0.00 0.00 1.85 14.28 10.00 20.50 11.75 1.00 0.00 2.61 9.81 30.00 12.00
17.38 0.00 0.00 1.55 13.59 30.00 20.50 13.63 0.00 0.00 1.84 10.24 30.00 11.00
16.00 0.00 1.00 3.33 14.28 10.00 20.50 13.88 0.00 0.00 1.62 11.00 30.00 11.00
16.75 1.00 0.00 2.77 14.94 10.00 20.50 13.00 1.00 0.00 3.56 11.00 30.00 11.00
17.13 0.00 0.00 2.18 14.94 10.00 20.50 12.00 1.00 0.00 2.65 11.10 10.00 11.00
17.50 0.00 1.00 4.21 14.67 30.00 20.50 13.13 1.00 0.00 2.65 11.00 30.00 11.00
17.00 1.00 0.00 2.66 15.32 10.00 19.50 14.27 0.00 0.00 1.80 11.34 30.00 12.30
16.75 0.00 1.00 3.58 15.32 10.00 19.50 14.63 0.00 0.00 1.69 11.34 30.00 12.30
17.20 0.00 1.00 2.96 15.32 10.00 19.50 15.25 0.00 0.00 1.88 11.34 30.00 12.20
18.75 0.00 0.00 1.93 15.32 10.00 19.50 14.25 1.00 0.00 2.77 11.34 30.00 12.30
17.50 0.00 1.00 2.57 14.68 30.00 19.00 13.52 1.00 0.00 2.22 11.75 10.00 13.50
17.50 0.00 0.00 3.18 15.15 10.00 18.00 14.63 1.00 0.00 2.42 11.59 30.00 13.50
18.00 0.00 0.00 1.93 15.15 10.00 18.00 14.75 0.00 0.00 1.77 11.39 30.00 13.50
15.63 0.00 0.00 2.20 13.39 10.00 17.00 14.00 0.00 0.00 2.22 11.75 10.00 13.50
14.75 1.00 0.00 2.21 13.39 10.00 17.00 14.50 0.00 0.00 2.99 11.59 30.00 13.50
15.25 0.00 1.00 3.24 13.35 30.00 16.50 14.25 0.00 0.00 2.22 11.75 10.00 13.50
15.75 1.00 0.00 2.35 13.35 30.00 16.50 14.63 0.00 0.00 1.93 11.75 10.00 14.50
15.25 1.00 0.00 2.11 13.39 10.00 16.50 13.30 1.00 0.00 3.35 12.68 10.00 15.50
15.75 1.00 0.00 2.80 13.35 30.00 16.50 14.50 0.00 0.00 2.21 12.68 10.00 17.00
15.63 0.00 0.00 1.95 13.39 10.00 16.50

521
This page intentionally left blank
Appendix B

Tables

Table B-1 Individual Terms of the Binomial Distribution

Table B-2 Areas for Standard Normal Probability Distribution
Table B-3 Critical Values of t
Table B-4 Critical Values of Chi-Square
Table B-5 F Distribution
Table B-6 Durbin-Watson Test Bounds

TABLE B-1 Individual Terms of the Binomial Distribution

p

n x .05 .10 .15 .20 .25 .30 .35 .40 .45 .50 .55 .60 .65 .70 .75 .80 .85 .90 .95

1 0 .9500 .9000 .8500 .8000 .7500 .7000 .6500 .6000 .5500 .5000 .4500 .4000 .3500 .3000 .2500 .2000 .1500 .1000 .0500
1 .0500 .1000 .1500 .2000 .2500 .3000 .3500 .4000 .4500 .5000 .5500 .6000 .6500 .7000 .7500 .8000 .8500 .9000 .9500
2 0 .9025 .8100 .7225 .6400 .5625 .4900 .4225 .3600 .3025 .2500 .2025 .1600 .1225 .0900 .0625 .0400 .0225 .0100 .0025
1 .0950 .1800 .2550 .3200 .3750 .4200 .4550 .4800 .4950 .5000 .4950 .4800 .4550 .4200 .3750 .3200 .2550 .1800 .0950
2 .0025 .0100 .0225 .0400 .0625 .0900 .1225 .1600 .2025 .2500 .3025 .3600 .4225 .4900 .5625 .6400 .7225 .8100 .9025
3 0 .8574 .7290 .6141 .5120 .4219 .3430 .2746 .2160 .1664 .1250 .0911 .0640 .0429 .0270 .0156 .0900 .0034 .0010 .0001
1 .1354 .2430 .3251 .3840 .4219 .4410 .4436 .4320 .4084 .3750 .3341 .2880 .2389 .1890 .1406 .0960 .0574 .0270 .0071
2 .0071 .0270 .0574 .0960 .1406 .1890 .2389 .2880 .3341 .3750 .4084 .4320 .4436 .4410 .4219 .3840 .3251 .2430 .1354
3 .0001 .0010 .0034 .0080 .0156 .0270 .0429 .0640 .0911 .1250 .1664 .2160 .2746 .3430 .4219 .5120 .6141 .7290 .8574
4 0 .8145 .6561 .5220 .4096 .3164 .2401 .1785 .1296 .0915 .0625 .0410 .0256 .0150 .0081 .0039 .0016 .0005 .0001 .0000
1 .1715 .2916 .3685 .4096 .4219 .4116 .3845 .3456 .2995 .2500 .2005 .1536 .1115 .0756 .0469 .0256 .0115 .0036 .0005
2 .0135 .0486 .0975 .1536 .2109 .2646 .3105 .3456 .3675 .3750 .3675 .3456 .3105 .2646 .2109 .1536 .0975 .0486 .0135
3 .0005 .0036 .0115 .0256 .0469 .0756 .1115 .1536 .2005 .2500 .2995 .3456 .3845 .4116 .4219 .4096 .3685 .2916 .1715
4 .0000 .0001 .0005 .0016 .0039 .0081 .0150 .0256 .0410 .0625 .0915 .1296 .1785 .2401 .3164 .4096 .5220 .6561 .8145
5 0 .7738 .5905 .4437 .3277 .2373 .1681 .1160 .0778 .0503 .0313 .0185 .0102 .0053 .0024 .0010 .0003 .0001 .0000 .0000
1 .2036 .3281 .3915 .4096 .3955 .3602 .3124 .2592 .2059 .1563 .1128 .0768 .0488 .0284 .0146 .0064 .0022 .0004 .0000
2 .0214 .0729 .1382 .2048 .2637 .3087 .3364 .3456 .3369 .3125 .2757 .2304 .1811 .1323 .0879 .0512 .0244 .0081 .0011
3 .0011 .0081 .0244 .0512 .0879 .1323 .1811 .2304 .2757 .3125 .3369 .3456 .3364 .3087 .2637 .2048 .1382 .0729 .0214
4 .0000 .0004 .0022 .0064 .0146 .0283 .0488 .0768 .1128 .1562 .2059 .2592 .3124 .3601 .3955 .4096 .3915 .3281 .2036
5 .0000 .0000 .0001 .0003 .0010 .0024 .0053 .0102 .0185 .0312 .0503 .0778 .1160 .1681 .2373 .3277 .4437 .5905 .7738
6 0 .7351 .5314 .3771 .2621 .1780 .1176 .0754 .0467 .0277 .0156 .0083 .0041 .0018 .0007 .0002 .0001 .0000 .0000 .0000
1 .2321 .3543 .3993 .3932 .3560 .3025 .2437 .1866 .1359 .0938 .0609 .0369 .0205 .0102 .0044 .0015 .0004 .0001 .0000
2 .0305 .0984 .1762 .2458 .2966 .3241 .3280 .3110 .2780 .2344 .1861 .1382 .0951 .0595 .0330 .0154 .0055 .0012 .0001
3 .0021 .0146 .0415 .0819 .1318 .1852 .2355 .2765 .3032 .3125 .3032 .2765 .2355 .1852 .1318 .0819 .0415 .0146 .0021
4 .0001 .0012 .0055 .0154 .0330 .0595 .0951 .1382 .1861 .2344 .2780 .3110 .3280 .3241 .2966 .2458 .1762 .0984 .0305
5 .0000 .0001 .0004 .0015 .0044 .0102 .0205 .0369 .0609 .0937 .1359 .1866 .2437 .3025 .3560 .3932 .3993 .3543 .2321
6 .0000 .0000 .0000 .0001 .0002 .0007 .0018 .0041 .0083 .0156 .0277 .0467 .0754 .1176 .1780 .2621 .3771 .5314 .7351
7 0 .6983 .4783 .3206 .2097 .1335 .0824 .0490 .0280 .0152 .0078 .0037 .0016 .0006 .0002 .0001 .0000 .0000 .0000 .0000
1 .2573 .3720 .3960 .3670 .3115 .2471 .1848 .1306 .0872 .0547 .0320 .0172 .0084 .0036 .0013 .0004 .0001 .0000 .0000
2 .0406 .1240 .2097 .2753 .3115 .3177 .2985 .2613 .2140 .1641 .1172 .0774 .0466 .0250 .0115 .0043 .0012 .0002 .0000
3 .0036 .0230 .0617 .1147 .1730 .2269 .2679 .2903 .2918 .2734 .2388 .1935 .1442 .0972 .0577 .0287 .0109 .0026 .0002
4 .0002 .0026 .0109 .0287 .0577 .0972 .1442 .1935 .2388 .2734 .2918 .2903 .2679 .2269 .1730 .1147 .0617 .0230 .0036
5 .0000 .0002 .0012 .0043 .0115 .0250 .0466 .0774 .1172 .1641 .2140 .2613 .2985 .3177 .3115 .2753 .2097 .1240 .0406
6 .0000 .0000 .0001 .0004 .0013 .0036 .0084 .0172 .0320 .0547 .0872 .1306 .1848 .2471 .3115 .3670 .3960 .3720 .2573
7 .0000 .0000 .0000 .0000 .0001 .0002 .0006 .0016 .0037 .0078 .0152 .0280 .0490 .0824 .1335 .2097 .3206 .4783 .6983
8 0 .6634 .4305 .2725 .1678 .1001 .0576 .0319 .0168 .0084 .0039 .0017 .0007 .0002 .0001 .0000 .0000 .0000 .0000 .0000
1 .2793 .3826 .3847 .3355 .2670 .1977 .1373 .0896 .0548 .0313 .0164 .0079 .0033 .0012 .0004 .0001 .0000 .0000 .0000
2 .0515 .1488 .2376 .2936 .3115 .2965 .2587 .2090 .1569 .1094 .0703 .0413 .0217 .0100 .0038 .0011 .0002 .0000 .0000

523
524 APPENDIX B Tables

TABLE B-1 Individual

(Continued)
Terms of the Binomial Distribution
p

n x .05 .10 .15 .20 .25 .30 .35 .40 .45 .50 .55 .60 .65 .70 .75 .80 .85 .90 .95

3 .0054 .0331 .0839 .1468 .2076 .2541 .2786 .2787 .2568 .2188 .1719 .1239 .0808 .0467 .0231 .0092 .0026 .0004 .0000
4 .0004 .0046 .0185 .0459 .0865 .1361 .1875 .2322 .2627 .2734 .2627 .2322 .1875 .1361 .0865 .0459 .0185 .0046 .0004
5 .0000 .0004 .0026 .0092 .0231 .0467 .0808 .1239 .1719 .2188 .2568 .2787 .2786 .2541 .2076 .1468 .0839 .0331 .0054
6 .0000 .0000 .0002 .0011 .0038 .0100 .0217 .0413 .0703 .1094 .1569 .2090 .2587 .2965 .3115 .2936 .2376 .1488 .0515
7 .0000 .0000 .0000 .0001 .0004 .0012 .0033 .0079 .0164 .0312 .0548 .8996 .1373 .1977 .2670 .3355 .3847 .3826 .2793
8 .0000 .0000 .0000 .0000 .0000 .0001 .0002 .0007 .0017 .0039 .0084 .0168 .0319 .0576 .1001 .1678 .2725 .4305 .6634
9 0 .6302 .3874 .2316 .1342 .0751 .0404 .0207 .0101 .0046 .0020 .0008 .0003 .0001 .0000 .0000 .0000 .0000 .0000 .0000
1 .2986 .3874 .3679 .3020 .2253 .1556 .1004 .0605 .0339 .0176 .0083 .0035 .0013 .0004 .0001 .0000 .0000 .0000 .0000
2 .0629 .1722 .2597 .3020 .3003 .2668 .2162 .1612 .1110 .0703 .0407 .0212 .0098 .0039 .0012 .0003 .0000 .0000 .0000
3 .0077 .0446 .1069 .1762 .2336 .2668 .2716 .2508 .2119 .1641 .1160 .0743 .0424 .0210 .0087 .0028 .0006 .0001 .0000
4 .0006 .0074 .0283 .0661 .1168 .1715 .2194 .2508 .2600 .2461 .2128 .1672 .1181 .0735 .0389 .0165 .0050 .0008 .0000
5 .0000 .0008 .0050 .0165 .0389 .0735 .1181 .1672 .2128 .2461 .2600 .2508 .2194 .1715 .1168 .0661 .0283 .0074 .0006
6 .0000 .0001 .0006 .0028 .0087 .0210 .0424 .0743 .1160 .1641 .2119 .2508 .2716 .2668 .2336 .1762 .1069 .0446 .0077
7 .0000 .0000 .0000 .0003 .0012 .0039 .0098 .0212 .0407 .0703 .1110 .1612 .2162 .2668 .3003 .3020 .2597 .1722 .0629
8 .0000 .0000 .0000 .0000 .0001 .0004 .0013 .0035 .0083 .0176 .0339 .0605 .1004 .1556 .2253 .3020 .3679 .3874 .2986
9 .0000 .0000 .0000 .0000 .0000 .0000 .0001 .0003 .0008 .0020 .0046 .0101 .0207 .0404 .0751 .1342 .2316 .3874 .6302
10 0 .5987 .3487 .1969 .1074 .0563 .0282 .0135 .0060 .0025 .0010 .0003 .0001 .0000 .0000 .0000 .0000 .0000 .0000 .0000
1 .3151 .3874 .3474 .2684 .1877 .1211 .0725 .0403 .0207 .0098 .0042 .0016 .0005 .0001 .0000 .0000 .0000 .0000 .0000
2 .0746 .1937 .2759 .3020 .2816 .2335 .1757 .1209 .0763 .0439 .0229 .0106 .0043 .0014 .0004 .0001 .0000 .0000 .0000
3 .0105 .0574 .1298 .2013 .2503 .2668 .2522 .2150 .1665 .1172 .0746 .0425 .0212 .0090 .0031 .0008 .0001 .0000 .0000
4 .0010 .0112 .0401 .0881 .1460 .2001 .2377 .2508 .2384 .2051 .1596 .1115 .0689 .0368 .0162 .0055 .0012 .0001 .0000
5 .0001 .0015 .0085 .0264 .0584 .1029 .1536 .2007 .2340 .2461 .2340 .2007 .1536 .1029 .0584 .0264 .0085 .0015 .0001
6 .0000 .0001 .0012 .0055 .0162 .0368 .0689 .1115 .1596 .2051 .2384 .2508 .2377 .2001 .1460 .0881 .0401 .0112 .0010
7 .0000 .0000 .0001 .0008 .0031 .0090 .0212 .0425 .0746 .1172 .1665 .2150 .2522 .2668 .2503 .2013 .1298 .0574 .0105
8 .0000 .0000 .0000 .0001 .0004 .0014 .0043 .0106 .0229 .0439 .0763 .1209 .1757 .2335 .2816 .3020 .2759 .1937 .0746
9 .0000 .0000 .0000 .0000 .0000 .0001 .0005 .0016 .0042 .0098 .0207 .0403 .0725 .1211 .1877 .2684 .3474 .3874 .3151
10 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0001 .0003 .0010 .0025 .0060 .0135 .0282 .0563 .1074 .1969 .3487 .5987
11 0 .5688 .3138 .1673 .0859 .0422 .0198 .0088 .0036 .0014 .0005 .0002 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000
1 .3293 .3835 .3248 .2362 .1549 .0932 .0518 .0266 .0125 .0054 .0021 .0007 .0002 .0000 .0000 .0000 .0000 .0000 .0000
2 .0867 .2131 .2866 .2953 .2581 .1998 .1395 .0887 .0513 .0269 .0126 .0052 .0018 .0005 .0001 .0000 .0000 .0000 .0000
3 .0137 .0710 .1517 .2215 .2581 .2568 .2254 .1774 .1259 .0806 .0462 .0234 .0102 .0037 .0011 .0002 .0000 .0000 .0000
4 .0014 .0158 .0536 .1107 .1721 .2201 .2428 .2365 .2060 .1611 .1128 .0701 .0379 .0173 .0064 .0017 .0003 .0000 .0000
5 .0001 .0025 .0132 .0388 .0803 .1321 .1830 .2207 .2360 .2256 .1931 .1471 .0985 .0566 .0268 .0097 .0023 .0003 .0000
6 .0000 .0003 .0023 .0097 .0268 .0566 .0985 .1471 .1931 .2256 .2360 .2270 .1830 .1321 .0803 .0388 .0132 .0025 .0001
7 .0000 .0000 .0003 .0017 .0064 .0173 .0379 .0701 .1128 .1611 .2060 .2365 .2428 .2201 .1721 .1107 .0536 .0158 .0014
8 .0000 .0000 .0000 .0002 .0011 .0037 .0102 .0234 .0462 .0806 .1259 .1774 .2254 .2568 .2581 .2215 .1517 .0710 .0137
9 .0000 .0000 .0000 .0000 .0001 .0005 .0018 .0052 .0126 .0269 .0513 .0887 .1395 .1998 .2581 .2953 .2866 .2131 .0867
10 .0000 .0000 .0000 .0000 .0000 .0000 .0002 .0007 .0021 .0054 .0125 .0266 .0518 .0932 .1549 .2362 .3248 .3835 .3293
11 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0002 .0005 .0014 .0036 .0088 .0198 .0422 .0859 .1673 .3138 .5688
12 0 .5404 .2824 .1422 .0687 .0317 .0138 .0057 .0022 .0008 .0002 .0001 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000
1 .3413 .3766 .3012 .2062 .1267 .0712 .0368 .0174 .0075 .0029 .0010 .0003 .0001 .0000 .0000 .0000 .0000 .0000 .0000
2 .0988 .2301 .2924 .2835 .2323 .1678 .1088 .0639 .0339 .0161 .0068 .0025 .0008 .0002 .0000 .0000 .0000 .0000 .0000
3 .0173 .0852 .1720 .2362 .2581 .2397 .1954 .1419 .0923 .0537 .0277 .0125 .0048 .0015 .0004 .0001 .0000 .0000 .0000
4 .0021 .0213 .0683 .1329 .1936 .2311 .2367 .2128 .1700 .1208 .0762 .0420 .0199 .0078 .0024 .0005 .0001 .0000 .0000
5 .0002 .0038 .0193 .0532 .1032 .1585 .2039 .2270 .2225 .1934 .1489 .1009 .0591 .0291 .0115 .0033 .0006 .0000 .0000
6 .0000 .0005 .0040 .0155 .0401 .0792 .1281 .1766 .2124 .2256 .2124 .1766 .1281 .0792 .0401 .0155 .0040 .0005 .0000
7 .0000 .0000 .0006 .0033 .0115 .0291 .0591 .1009 .1489 .1934 .2225 .2270 .2039 .1585 .1032 .0532 .0193 .0038 .0002
8 .0000 .0000 .0001 .0005 .0024 .0078 .0199 .0420 .0762 .1208 .1700 .2128 .2367 .2311 .1936 .1329 .0683 .0213 .0021
9 .0000 .0000 .0000 .0001 .0004 .0015 .0048 .0125 .0277 .0537 .0923 .1419 .1954 .2397 .2581 .2362 .1720 .0852 .0173
10 .0000 .0000 .0000 .0000 .0000 .0002 .0008 .0025 .0068 .0161 .0339 .0639 .1088 .1678 .2323 .2835 .2924 .2301 .0988
11 .0000 .0000 .0000 .0000 .0000 .0000 .0001 .0003 .0010 .0029 .0075 .0174 .0368 .0712 .1267 .2062 .3012 .3766 .3413
12 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0000 .0001 .0002 .0008 .0022 .0057 .0138 .0317 .0687 .1422 .2824 .5404

Source: Table A, pages 464–466 in Business Statistics: Concepts and Applications by William J. Stevenson. Copyright © 1978 by
William J. Stevenson. Reprinted by permission of Harper &Row, Publishers, Inc.
 APPENDIX B Tables 525

TABLE B-2 Areas for Standard Normal Probability Distribution

0 z

For z  1.93, shaded area is .4732 out of total area of 1.

z .00 .01 .02 .03 .04 .05 .06 .07 .08 .09

0.0 .00000 .0040 .0080 .0120 .0160 .0199 .0239 .0279 .0319 .0359
0.1 .03980 .0438 .0478 .0517 .0557 .0596 .0636 .0675 .0714 .0753
0.2 .07930 .0832 .0871 .0910 .0948 .0987 .1026 .1064 .1103 .1141
0.3 .11790 .1217 .1255 .1293 .1331 .1368 .1406 .1443 .1480 .1517
0.4 .15540 .1591 .1628 .1664 .1700 .1736 .1772 .1808 .1844 .1879
0.5 .19150 .1950 .1985 .2019 .2054 .2088 .2123 .2157 .2190 .2224
0.6 .22570 .2291 .2324 .2357 .2389 .2422 .2454 .2486 .2518 .2549
0.7 .25800 .2612 .2642 .2673 .2704 .2734 .2764 .2794 .2823 .2852
0.8 .28810 .2910 .2939 .2967 .2995 .3023 .3051 .3078 .3106 .3133
0.9 .31590 .3186 .3212 .3238 .3264 .3289 .3315 .3340 .3365 .3389
1.0 .34130 .3438 .3461 .3485 .3508 .3531 .3554 .3577 .3599 .3621
1.1 .36430 .3665 .3686 .3708 .3729 .3749 .3770 .3790 .3810 .3830
1.2 .38490 .3869 .3888 .3907 .3925 .3944 .3962 .3980 .3997 .4015
1.3 .40320 .4049 .4066 .4082 .4099 .4115 .4131 .4147 .4162 .4177
1.4 .41920 .4207 .4222 .4236 .4251 .4265 .4279 .4292 .4306 .4319
1.5 .43320 .4345 .4357 .4370 .4382 .4394 .4406 .4418 .4429 .4441
1.6 .44520 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545
1.7 .45540 .4564 .4573 .4582 .4591 .4599 .4608 .4616 .4625 .4633
1.8 .46410 .4649 .4656 .4664 .4671 .4678 .4686 .4693 .4699 .4706
1.9 .47130 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .5761 .4767
2.0 .47720 .4778 .4783 .4788 .4793 .4798 .4803 .4808 .4812 .4817
2.1 .48210 .4826 .4830 .4834 .4838 .4842 .4846 .4850 .4854 .4857
2.2 .48610 .4864 .4868 .4871 .4875 .4878 .4881 .4884 .4887 .4890
2.3 .48930 .4896 .4898 .4901 .4904 .4906 .4909 .4911 .4913 .4916
2.4 .49180 .4920 .4922 .4925 .4927 .4929 .4931 .4932 .4934 .4936
2.5 .49380 .4940 .4941 .4943 .4945 .4946 .4948 .4949 .4951 .4952
2.6 .49530 .4955 .4956 .4957 .4959 .4960 .4961 .4962 .4963 .4964
2.7 .49650 .4966 .4967 .4968 .4969 .4970 .4971 .4972 .4973 .4974
2.8 .49740 .4975 .4976 .4977 .4977 .4978 .4979 .4979 .4898 .4981
2.9 .49810 .4982 .4982 .4983 .4984 .4984 .4985 .4985 .4986 .4986
3.0 .49865 .4987 .4987 .4988 .4989 .4989 .4989 .4989 .4990 .4990
4.0 .4999683
526 APPENDIX B Tables

TABLE B-3 Critical Values of t

0 tα

df t.100 t.050 t.025 t.010 t.005

1 3.078 6.314 12.706 31.821 63.657

2 1.886 2.920 4.303 6.965 9.925
3 1.638 2.353 3.182 4.541 5.841
4 1.533 2.132 2.776 3.747 4.604
5 1.476 2.015 2.571 3.365 4.032
6 1.440 1.943 2.447 3.143 3.707
7 1.415 1.895 2.365 2.998 3.499
8 1.397 1.860 2.306 3.896 3.355
9 1.383 1.833 2.262 2.821 3.250
10 1.372 1.812 2.228 2.764 3.169
11 1.363 1.796 2.201 2.718 3.106
12 1.356 1.782 2.179 2.681 3.055
13 1.350 1.771 2.160 2.650 3.012
14 1.345 1.761 2.145 2.624 2.977
15 1.341 1.753 2.131 2.602 2.947
16 1.337 1.746 2.120 2.583 2.921
17 1.333 1.740 2.110 2.567 2.898
18 1.330 1.734 2.101 2.552 2.878
19 1.328 1.729 2.093 2.529 2.861
20 1.325 1.725 2.086 2.528 2.845
21 1.323 1.721 2.080 2.518 2.831
22 1.321 1.717 2.074 2.508 2.819
23 1.319 1.714 2.069 2.500 2.807
24 1.318 1.711 2.064 2.492 2.797
25 1.316 1.708 2.060 2.485 2.787
26 1.315 1.706 2.056 2.479 2.779
27 1.314 1.703 2.052 2.473 2.771
28 1.313 1.701 2.048 2.467 2.763
29 1.311 1.699 2.045 2.462 2.756
inf. 1.282 1.645 1.960 2.326 2.576

Source: “Table of Percentage Points of the t-Distribution.”

Computed by Maxine Merrington, Biometrika, vol. 32 (1941),
p. 300. Reproduced by permission of Professor D. V. Lindley.
 APPENDIX B Tables 527

TABLE B-4 Critical Values of Chi-Square

0 xa2

df X2.995 X2.990 X2.975 X2.950 X2.900

1 0.0000393 0.0001571 0.0009821 0.0039321 0.0157908

2 0.0100251 0.0201007 0.0506356 0.102587 0.210720
3 0.0717212 0.114832 0.215795 0.351846 0.584375
4 0.206990 0.297110 0.484419 0.710721 1.063623
5 0.411740 0.554300 0.831211 1.145476 1.61031
6 0.675727 0.872085 1.237347 1.63539 2.20413
7 0.989265 1.239043 1.68987 2.16735 2.83311
8 1.344419 1.646482 2.17973 2.73264 3.48954
9 1.734926 2.087912 2.70039 3.32511 4.168216
10 2.15585 2.55821 3.24697 3.94030 4.86518
11 2.60321 3.05347 3.81575 4.57481 5.57779
12 3.07382 3.57056 4.40379 5.22603 6.30380
13 3.56503 4.10691 5.00874 5.89186 7.04150
14 4.07468 4.66043 5.62872 6.57063 7.78953
15 4.60094 5.22935 6.26214 7.26094 8.54675
16 5.14224 5.81221 6.90766 7.96164 9.31223
17 5.69724 6.40776 7.56418 8.67176 10.0852
18 6.26481 7.01491 8.23075 9.39046 10.8649
19 6.84398 7.63273 8.90655 10.1170 11.6509
20 7.43386 8.26040 9.59083 10.8508 12.4426
21 8.03366 8.89720 10.28293 11.5913 13.2396
22 8.64272 9.54249 10.9823 12.3380 14.0415
23 9.26042 10.19567 11.6885 13.0905 14.8479
24 9.88623 10.8564 12.4011 13.8484 15.6587
25 10.5197 11.5240 13.1197 14.6114 16.4734
26 11.1603 12.1981 13.8439 15.3791 17.2919
27 11.8076 12.8786 14.5733 16.1513 18.1138
28 12.4613 13.5648 15.3079 16.9279 18.9302
29 13.1211 14.2565 16.0471 17.7083 19.7677
30 13.7867 14.9535 16.7908 18.4926 20.5992
40 20.7065 22.1643 24.4331 26.5093 29.0505
50 27.9907 29.7067 32.3574 34.7642 37.6886
60 35.5347 37.4848 40.4817 43.1879 46.4589
70 43.2752 45.4418 48.7576 51.7393 55.3290
80 51.1720 53.5400 57.1532 60.3915 64.2778
90 59.1963 61.7541 65.6466 69.1260 73.2912
100 67.3276 70.0648 74.2219 77.9295 82.3581
528 APPENDIX B Tables

TABLE B-4 (Continued)

df X2.100 X2.050 X2.025 X2.010 X2.005

1 2.70554 3.84146 5.02389 6.63490 7.87944

2 4.60517 5.99147 7.37776 9.21034 10.5966
3 6.25139 7.81473 9.34840 11.3449 12.8381
4 7.77944 9.48773 11.1433 13.2767 14.8602
5 9.23635 11.0705 12.8325 15.0863 16.7496
6 10.6446 12.5916 14.4494 16.8119 18.5476
7 12.0170 14.0671 16.0128 18.4753 20.2777
8 13.3616 15.5073 17.5346 20.0902 21.9550
9 14.6837 16.9190 19.0228 21.6660 23.5893
10 15.9871 18.3070 20.4831 23.2093 25.1882
11 17.2750 19.6751 21.9200 24.7250 26.7569
12 18.5494 21.0261 23.3367 26.2170 28.2995
13 19.8119 22.3621 24.7356 27.6883 29.8194
14 21.0642 23.6848 26.1190 29.1413 31.3193
15 22.3072 24.9958 27.4884 30.5779 32.8013
16 23.5418 26.2962 28.8454 31.9999 34.2672
17 24.7690 27.5871 30.1910 33.4087 35.7185
18 25.9894 28.8693 31.5264 34.8053 37.1564
19 27.2036 30.1435 32.8523 36.1908 38.5822
20 28.4120 31.4104 34.1696 37.5662 39.9968
21 29.6151 32.6705 35.4789 38.9321 41.4010
22 30.8133 33.9244 36.7807 40.2894 42.7956
23 32.0069 35.1725 38.0757 41.6384 44.1813
24 33.1963 36.4151 39.3641 42.9798 45.5585
25 34.3816 37.6525 40.6465 44.3141 46.9278
26 35.5631 38.8852 41.9232 45.6417 48.2899
27 36.7412 40.1133 43.1944 46.9630 49.6449
28 37.9159 41.3372 44.4607 48.2782 50.9933
29 39.0875 42.5569 45.7222 49.5879 52.3356
30 40.2560 43.7729 46.9792 50.8922 53.6720
40 51.8050 55.7585 59.3417 63.6907 66.7659
50 63.1671 67.5048 71.4202 76.1539 79.4900
60 74.3970 79.0819 83.2976 88.3794 91.9517
70 85.5271 90.5312 95.0231 100.425 104.215
80 96.5782 101.879 106.629 112.329 116.321
90 107.565 113.145 118.136 124.116 128.299
100 118.498 124.342 129.561 135.807 140.169

Source: “Tables of the Percentage Points of the X2-Distribution,” Biometrika,

vol. 32 (1941), pp. 188–189, by Catherine M. Thompson. Reproduced by permission
of Professor D. V. Lindley.
 APPENDIX B Tables 529

TABLE B-5 F Distribution

For example, the F scale value for d1 = 3, d2 =10 corresponding to area .01 in right tail
is 6.55.
F value corresponding to area .05 in right tail in lightface type.
F value corresponding to area .01 in right tail in boldface type.

D2 , D1, Numerator Degrees of Freedom

Denominator
Degrees of Freedom 1 2 3 4 5 6 7 8 9 10

1 161 200 216 225 230 234 237 239 241 242
4,052 4,999 5,403 5,625 5,764 5,859 5,928 5,981 6,022 6,056
2 18.51 19.00 19.16 19.25 19.30 19.33 19.36 19.37 19.38 19.39
98.49 99.00 99.17 99.25 99.30 99.33 99.36 99.37 99.39 99.40
3 10.13 9.55 9.28 9.12 9.01 8.94 8.88 8.84 8.81 8.78
34.12 30.82 29.46 28.71 28.24 27.91 27.67 27.49 27.34 27.23
4 7.71 6.94 6.59 6.39 6.26 6.16 6.09 6.04 6.00 5.96
21.20 18.00 16.69 15.98 15.52 15.21 14.98 14.80 14.66 14.54
5 6.61 5.79 5.41 5.19 5.05 4.95 4.88 4.82 4.78 4.74
16.26 13.27 12.06 11.39 10.97 10.67 10.45 10.29 10.15 10.05
6 5.99 5.14 4.76 4.53 4.39 4.28 4.21 4.15 4.10 4.06
13.74 10.92 9.78 9.15 8.75 8.47 8.26 8.10 7.98 7.87
7 5.59 4.74 4.35 4.12 3.97 3.87 3.79 3.73 3.68 3.63
12.25 9.55 8.45 7.85 7.46 7.19 7.00 6.84 6.71 6.62
8 5.32 4.46 4.07 3.84 3.69 3.58 3.50 3.44 3.39 3.34
11.26 8.65 7.59 7.01 6.63 6.37 6.19 6.03 5.91 5.82
9 5.12 4.26 3.86 3.63 3.48 3.37 3.29 3.23 3.18 3.13
10.56 8.02 6.99 6.42 6.06 5.80 5.62 5.47 5.35 5.26
10 4.96 4.10 3.71 3.48 3.33 3.22 3.14 3.07 3.02 2.97
10.04 7.56 6.55 5.99 5.64 5.39 5.21 5.06 4.95 4.85
11 4.84 3.98 3.59 3.36 3.20 3.09 3.01 2.95 2.90 2.86
9.65 7.20 6.22 5.67 5.32 5.07 4.88 4.74 4.63 4.54
12 4.75 3.88 3.49 3.26 3.11 3.00 2.92 2.85 2.80 2.76
9.33 6.93 5.95 5.41 5.06 4.82 4.65 4.50 4.39 4.30
13 4.67 3.80 3.41 3.18 3.02 2.92 2.84 2.77 2.72 2.67
9.07 6.70 5.74 5.20 4.86 4.62 4.44 4.30 4.19 4.10
14 4.60 3.74 3.34 3.11 2.96 2.85 2.77 2.70 2.65 2.60
8.86 6.51 5.56 5.03 4.69 4.46 4.28 4.14 4.03 3.94
15 4.54 3.68 3.29 3.06 2.90 2.79 2.70 2.64 2.59 2.55
8.68 6.36 5.42 4.89 4.56 4.32 4.14 4.00 3.89 3.80
16 4.49 3.36 3.24 3.01 2.85 2.74 2.66 2.59 2.54 2.49
8.53 6.23 5.29 4.77 4.44 4.20 4.03 3.89 3.78 3.69
17 4.45 3.59 3.20 2.96 2.81 2.70 2.62 2.55 2.50 2.45
8.40 6.11 5.18 4.67 4.34 4.10 3.93 3.79 3.68 3.59
18 4.41 3.55 3.16 2.93 2.77 2.66 2.58 2.51 2.46 2.41
8.28 6.01 5.09 4.58 4.25 4.01 3.85 3.71 3.60 3.51
530 APPENDIX B Tables

TABLE B-5 F
(Continued)
Distribution

D2 , D1 , Numerator Degrees of Freedom

Denominator
Degrees of Freedom 1 2 3 4 5 6 7 8 9 10

19 4.38 3.52 3.13 2.90 2.74 2.63 2.55 2.48 2.43 2.38
8.18 5.93 5.01 4.50 4.17 3.94 3.77 3.63 3.52 3.43
20 4.35 3.49 3.10 2.87 2.71 2.60 2.52 2.45 2.40 2.35
8.10 5.85 4.94 4.43 4.10 3.87 3.71 3.56 3.45 3.37
21 4.32 3.47 3.07 2.84 2.68 2.57 2.49 2.42 2.37 2.32
8.02 5.78 4.87 4.37 4.04 3.81 3.65 3.51 3.40 3.31
22 4.30 3.44 3.05 2.82 2.66 2.55 2.47 2.40 2.35 2.30
7.94 5.72 4.82 4.31 3.99 3.76 3.59 3.45 3.35 3.26
23 4.28 3.42 3.03 2.80 2.64 2.53 2.45 2.38 2.32 2.28
7.88 5.66 4.76 4.26 3.94 3.71 3.54 3.41 3.30 3.21
24 4.26 3.40 3.01 2.78 2.62 2.51 2.43 2.36 2.30 2.26
7.82 5.61 4.72 4.22 3.90 3.67 3.50 3.36 3.25 3.17
25 4.24 3.38 2.99 2.76 2.60 2.49 2.41 2.34 2.28 2.24
7.77 5.57 4.68 4.18 3.86 3.63 3.46 3.32 3.21 3.13

Source: Abridged by permission from Statistical Methods, 7th ed., by George W. Snedecor and William C.
Cochran. Copyright © 1980 by the Iowa State University Press, Ames, Iowa.

TABLE B-6 Durbin-Watson Test Bounds

Level of Significance a .05

k1 k2 k3 k4 k5

n dL dU dL dU dL dU dL dU dL dU

15 1.08 1.36 0.95 1.54 0.82 1.75 0.69 1.97 0.56 2.21
16 1.10 1.37 0.98 1.54 0.86 1.73 0.74 1.93 0.62 2.15
17 1.13 1.38 1.02 1.54 0.90 1.71 0.78 1.90 0.67 2.10
18 1.16 1.39 1.05 1.53 0.93 1.69 0.82 1.87 0.71 2.06
19 1.18 1.40 1.08 1.53 0.97 1.68 0.86 1.85 0.75 2.02
20 1.20 1.41 1.10 1.54 1.00 1.68 0.90 1.83 0.79 1.99
21 1.22 1.42 1.13 1.54 1.03 1.67 0.93 1.81 0.83 1.96
22 1.24 1.43 1.15 1.54 1.05 1.66 0.96 1.80 0.86 1.94
23 1.26 1.44 1.17 1.54 1.08 1.66 0.99 1.79 0.90 1.92
24 1.27 1.45 1.19 1.55 1.10 1.66 1.01 1.78 0.93 1.90
25 1.29 1.45 1.21 1.55 1.12 1.66 1.04 1.77 0.95 1.89
26 1.30 1.46 1.22 1.55 1.14 1.65 1.06 1.76 0.98 1.88
27 1.32 1.47 1.24 1.56 1.16 1.65 1.08 1.76 1.01 1.86
28 1.33 1.48 1.26 1.56 1.18 1.65 1.10 1.75 1.03 1.85
29 1.34 1.48 1.27 1.56 1.20 1.65 1.12 1.74 1.05 1.84
30 1.35 1.49 1.28 1.57 1.21 1.65 1.14 1.74 1.07 1.83
31 1.36 1.50 1.30 1.57 1.23 1.65 1.16 1.74 1.09 1.83
32 1.37 1.50 1.31 1.57 1.24 1.65 1.18 1.73 1.11 1.82
33 1.38 1.51 1.32 1.58 1.26 1.65 1.19 1.73 1.13 1.81
34 1.39 1.51 1.33 1.58 1.27 1.65 1.21 1.73 1.15 1.81
35 1.40 1.52 1.34 1.58 1.28 1.65 1.22 1.73 1.16 1.80
36 1.41 1.52 1.35 1.59 1.29 1.65 1.24 1.73 1.18 1.80
37 1.42 1.53 1.36 1.59 1.31 1.66 1.25 1.72 1.19 1.80
38 1.43 1.54 1.37 1.59 1.32 1.66 1.26 1.72 1.21 1.79
 APPENDIX B Tables 531

TABLE B-6 Durbin-Watson Test Bounds

(Continued)

Level of Significance a .05

k1 k2 k3 k4 k5

n dL dU dL dU dL dU dL dU dL dU

39 1.43 1.54 1.38 1.60 1.33 1.66 1.27 1.72 1.22 1.79
40 1.44 1.54 1.39 1.60 1.34 1.66 1.29 1.72 1.23 1.79
45 1.48 1.57 1.43 1.62 1.38 1.67 1.34 1.72 1.29 1.78
50 1.50 1.59 1.46 1.63 1.42 1.67 1.38 1.72 1.34 1.77
55 1.53 1.60 1.49 1.64 1.45 1.68 1.41 1.72 1.38 1.77
60 1.55 1.62 1.51 1.65 1.48 1.69 1.44 1.73 1.41 1.77
65 1.57 1.63 1.54 1.66 1.50 1.70 1.47 1.73 1.44 1.77
70 1.58 1.64 1.55 1.67 1.52 1.70 1.49 1.74 1.46 1.77
75 1.60 1.65 1.57 1.68 1.54 1.71 1.51 1.74 1.49 1.77
80 1.61 1.66 1.59 1.69 1.56 1.72 1.53 1.74 1.51 1.77
85 1.62 1.67 1.60 1.70 1.57 1.72 1.55 1.75 1.52 1.77
90 1.63 1.68 1.61 1.70 1.59 1.73 1.57 1.75 1.54 1.78
95 1.64 1.69 1.62 1.71 1.60 1.73 1.58 1.75 1.56 1.78
100 1.65 1.69 1.63 1.72 1.61 1.74 1.59 1.76 1.57 1.78
Level of Significance a =.01

k1 k2 k3 k4 k5

n dL dU dL dU dL dU dL dU dL dU

15 0.81 1.07 0.70 1.25 0.59 1.46 0.49 1.70 0.39 1.96
16 0.84 1.09 0.74 1.25 0.63 1.44 0.53 1.66 0.44 1.90
17 0.87 1.10 0.77 1.25 0.67 1.43 0.57 1.63 0.48 1.85
18 0.90 1.12 0.80 1.26 0.71 1.42 0.61 1.60 0.52 1.80
19 0.93 1.13 0.83 1.26 0.74 1.41 0.65 1.58 0.56 1.77
20 0.95 1.15 0.86 1.27 0.77 1.41 0.68 1.57 0.60 1.74
21 0.97 1.16 0.89 1.27 0.80 1.41 0.72 1.55 0.63 1.71
22 1.00 1.17 0.91 1.28 0.83 1.40 0.75 1.54 0.66 1.69
23 1.02 1.19 0.94 1.29 0.86 1.40 0.77 1.53 0.70 1.67
24 1.04 1.20 0.96 1.30 0.88 1.41 0.80 1.53 0.72 1.66
25 1.05 1.21 0.98 1.30 0.90 1.41 0.83 1.52 0.75 1.65
26 1.07 1.22 1.00 1.31 0.93 1.41 0.85 1.52 0.78 1.64
27 1.09 1.23 1.02 1.32 0.95 1.41 0.88 1.51 0.81 1.63
28 1.10 1.24 1.04 1.32 0.97 1.41 0.90 1.51 0.83 1.62
29 1.12 1.25 1.05 1.33 0.99 1.42 0.92 1.51 0.85 1.61
30 1.13 1.26 1.07 1.34 1.01 1.42 0.94 1.51 0.88 1.61
31 1.15 1.27 1.08 1.34 1.02 1.42 0.96 1.51 0.90 1.60
32 1.16 1.28 1.10 1.35 1.04 1.43 0.98 1.51 0.92 1.60
33 1.17 1.29 1.11 1.36 1.05 1.43 1.00 1.51 0.94 1.59
34 1.18 1.30 1.13 1.36 1.07 1.43 1.01 1.51 0.95 1.59
35 1.19 1.31 1.14 1.37 1.08 1.44 1.03 1.51 0.97 1.59
36 1.21 1.32 1.15 1.38 1.10 1.44 1.04 1.51 0.99 1.59
37 1.22 1.32 1.16 1.38 1.11 1.45 1.06 1.51 1.00 1.59
38 1.23 1.33 1.18 1.39 1.12 1.45 1.07 1.52 1.02 1.58
39 1.24 1.34 1.19 1.39 1.14 1.45 1.09 1.52 1.03 1.58
40 1.25 1.34 1.20 1.40 1.15 1.46 1.10 1.52 1.05 1.58
532 APPENDIX B Tables

TABLE B-6 Durbin-Watson Test Bounds

(Continued)

Level of Significance a .05

.01

k1 k2 k3 k4 k5

n dL dU dL dU dL dU dL dU dL dU

45 1.29 1.38 1.24 1.42 1.20 1.48 1.16 1.53 1.11 1.58
50 1.32 1.40 1.28 1.45 1.24 1.49 1.20 1.54 1.16 1.59
55 1.36 1.43 1.32 1.47 1.28 1.51 1.25 1.55 1.21 1.59
60 1.38 1.45 1.35 1.48 1.32 1.52 1.28 1.56 1.25 1.60
65 1.41 1.47 1.38 1.50 1.35 1.53 1.31 1.57 1.28 1.61
70 1.43 1.49 1.40 1.52 1.37 1.55 1.34 1.58 1.31 1.61
75 1.45 1.50 1.42 1.53 1.39 1.56 1.37 1.59 1.34 1.62
80 1.47 1.52 1.44 1.54 1.42 1.57 1.39 1.60 1.36 1.62
85 1.48 1.53 1.46 1.55 1.43 1.58 1.41 1.60 1.39 1.63
90 1.50 1.54 1.47 1.56 1.45 1.59 1.43 1.61 1.41 1.64
95 1.51 1.55 1.49 1.57 1.47 1.60 1.45 1.62 1.42 1.64
100 1.52 1.56 1.50 1.58 1.48 1.60 1.46 1.63 1.44 1.65

k  Number of independent variables .

Source: Reprinted, with permission, from J. Durbin and G. S. Watson, “Testing for Serial Correlation
in Least Squares Regression—II,” Biometrika, vol. 38 (1951), pp. 159–178.
Appendix C

Data Sets and Databases

This appendix contains 12 time series data sets and 3 Data Sets (Cont.)
multiple regression databases. The first database Purchasing Power Federal Minimum
involves 7 variables for 25 campgrounds, the second Total of the Dollar, Wage Rate,
database has 9 financial variables measured on each Year Population Consumer Prices Current Dollars
of 266 U.S. corporations in a recent year,1 and the
1984 236,348 0.961 3.35
third database has 13 variables related to CEO com- 1985 238,466 0.928 3.35
pensation measured on each of 50 U.S. companies in 1986 240,651 0.913 3.35
a recent year. 1987 242,804 0.880 3.35
Data sets are available on the Internet for most 1988 245,021 0.846 3.35
1989 247,342 0.807 3.35
text examples, problems, and cases. In order to
1990 249,948 0.766 3.80
access these data sets on the Internet, go to the 1991 252,639 0.734 4.25
Prentice Hall website, www.prenhall.com/hanke, and 1992 255,374 0.713 4.25
follow the download instructions. 1993 258,083 0.692 4.25
1994 260,599 0.675 4.25
1995 263,044 0.656 4.25
1996 265,463 0.638 4.75
1997 268,008 0.623 5.15
1998 270,561 0.600
Data Sets
Eastman Money Cable
Purchasing Power Federal Minimum Kodak Market Television
Total of the Dollar, Wage Rate, Year Revenue Deposits Subscribers
Year Population Consumer Prices Current Dollars
1975 5.0
1960 180,671 3.373 1.00 1976 5.4
1961 183,691 3.340 1.15 1977 6.0
1962 186,538 3.304 1.15 1978 7.0
1963 189,242 3.265 1.25 1979 8.0
1964 191,889 3.220 1.25 1980 9.7 400
1965 194,303 3.166 1.25 1981 10.3 344
1966 196,560 3.080 1.25 1982 10.8 400 21,000
1967 198,712 2.993 1.40 1983 10.2 685 25,000
1968 200,706 2.873 1.60 1984 10.6 705 30,000
1969 202,677 2.726 1.60 1985 10.6 815 32,000
1970 205,052 2.574 1.60 1986 11.5 941 37,500
1971 207,661 2.466 1.60 1987 13.3 937 41,100
1972 209,896 2.391 1.60 1988 17.0 926 44,000
1973 211,909 2.251 1.60 1989 18.4 894 47,500
1974 213,854 2.029 2.00 1990 18.9 924 50,000
1975 215,973 1.859 2.10 1991 19.4 1,045 51,000
1976 218,035 1.757 2.30 1992 20.2 1,187 53,000
1977 220,239 1.649 2.30 1993 16.3 1,219 55,000
1978 222,585 1.532 2.65 1994 13.7 1,150 57,000
1979 225,055 1.380 2.90 1995 15.3 1,135 58,000
1980 227,726 1.215 3.10 1996 16.2 1,272 60,280
1981 229,966 1.098 3.35 1997 1,397 64,050
1982 232,188 1.035 3.35 1998 64,170
1983 234,307 1.003 3.35

1We are indebted to Dr. Lynn Stephens of Eastern Washington University for providing these data values.

533
534 APPENDIX C Data Sets and Databases

Household Major Appliances, Industry Shipments, Freezers (1,000s)

Year 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003

January 100 109 110 90 84 91 89 98 95 90 78

February 97 100 73 77 82 89 88 79 75 91 121
March 115 94 88 80 105 82 85 87 100 128 130
April 111 97 90 99 97 108 89 90 114 120 119
May 113 118 112 106 106 112 106 103 109 132 112
June 136 134 136 140 136 126 110 126 142 176 158
July 148 153 149 159 137 156 125 155 154 205 167
August 134 133 131 122 138 152 139 136 138 166 156
September 117 106 108 110 109 128 121 123 132 135 137
October 92 87 100 94 87 110 93 101 118 160
November 78 80 66 72 93 99 84 97 106 138
December 82 70 75 72 86 94 90 100 125 136

Beer Production (millions of barrels)

Year 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003

January 14.77 14.15 15.50 15.71 15.60 15.80 15.88 16.46 16.27 15.65 15.36
February 14.53 14.75 14.55 15.21 15.63 15.85 15.29 15.74 15.17 16.10 15.78
March 16.78 17.72 16.76 16.51 17.66 17.12 17.57 17.97 16.08 18.06 17.41
April 18.42 16.81 17.97 17.99 17.42 17.73 17.30 17.48 17.23 18.00 17.44
May 18.17 18.74 18.86 18.67 17.44 18.31 18.41 18.10 18.90 18.89 18.87
June 18.47 18.47 18.23 18.65 18.58 18.58 18.82 18.58 19.16 18.95 18.96
July 18.50 19.12 18.59 18.33 18.09 18.17 18.28 18.25 19.88 18.34 18.51
August 18.27 17.59 17.71 17.06 16.81 17.72 18.88 18.96 18.63 17.55
September 15.71 14.58 14.54 15.26 15.82 15.78 15.28 16.08 16.11 15.66
October 15.41 15.14 14.36 15.62 15.50 15.61 15.82 16.62 16.65 16.15
November 13.62 13.06 13.12 13.53 13.81 14.02 14.78 15.44 14.47 14.43
December 12.46 12.89 13.13 13.97 13.69 13.22 13.45 13.97 13.64 14.32

New Prescriptions (1,000s)

Year 2000 2001 2002 2003 2004 2005 2006

January 154 200 223 346 518 613 628

February 96 118 104 261 404 392 308
March 73 90 107 224 300 273 324
April 49 79 85 141 210 322 248
May 36 78 75 148 196 189 272
June 59 91 99 145 186 257 634
July 95 167 135 223 247 324 299
August 169 169 121 272 343 404 424
September 210 289 335 445 464 677 548
October 278 347 460 560 680 858 372
November 298 375 488 612 711 895 876
December 245 203 326 467 610 664 676
 APPENDIX C Data Sets and Databases 535
Seattle, Washington, Daily Bus Ridership for 146 Days on the Pike Street Route (read across the table)

350 339 351 364 369 331 331 340 346 341 357 398 381 367 383 375
353 361 375 371 373 366 382 429 406 403 429 425 427 409 402 409
419 404 429 463 428 449 444 467 474 463 432 453 462 456 474 514
489 475 492 525 527 533 527 522 526 513 564 599 572 587 599 601
611 620 579 582 592 581 630 663 638 631 645 682 601 595 521 521
516 496 538 575 537 534 542 538 547 540 526 548 555 545 594 643
625 616 640 625 637 634 621 641 654 649 662 699 672 704 700 711
715 718 652 664 695 704 733 772 716 712 732 755 761 748 748 750
744 731 782 810 777 816 840 868 872 811 810 762 634 626 649 697
657 549

Monthly Occupancy Statistics for Motel Nine

Year 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

January 563 635 647 676 748 795 843 778 895 875
February 599 639 658 748 773 788 847 856 856 993
March 669 712 713 811 814 890 942 939 893 977
April 598 622 688 729 767 797 804 813 875 969
May 580 621 724 701 729 751 840 783 835 872
June 668 676 707 790 749 821 872 828 935 1006
July 499 501 629 594 681 692 656 657 833 832
August 215 220 238 231 241 291 370 310 300 346
September 556 561 613 617 680 727 742 780 791 850
October 587 603 730 691 708 868 847 860 900 914
November 546 626 735 701 694 812 732 780 782 869
December 571 606 652 706 772 800 899 808 880 994

Wood Pulp Production (1,000s of short tons)

Year 2002 2003 2004 2005 2006

January 5,276 5,466 5,480 5,579 5,681

February 4,919 4,821 4,988 5,033 5,254
March 5,227 5,307 5,272 5,319 5,525
April 5,003 5,193 5,254 5,285 5,489
May 5,004 5,087 5,156 5,188 5,320
June 4,949 5,102 5,175 5,197 5,417
July 5,219 5,399 5,490 5,604 5,711
August 5,265 5,301 5,366 5,357 5,419
September 5,007 5,063 5,200 5,133 5,318
October 5,079 5,295 5,410 5,483 5,409
November 4,974 5,021 5,123 5,186 5,310
December 5,282 4,991 5,428 5,472 5,458
536 APPENDIX C Data Sets and Databases

Higher Education
Price Index, Number of
Year 1983  100 U.S. Farms (1,000s)

1970 39.5
42.1
44.3
46.7
49.9
1975 54.3 2,521
57.9 2,497
61.7 2,456
65.8 2,436
70.6 2,437
1980 77.5 2,440
85.9 2,440
94.0 2,407
100.0 2,379
104.7 2,334
1985 110.5 2,293
115.6 2,250
120.3 2,213
125.8 2,197
133.1 2,171
1990 140.8 2,140
148.3 2,105
153.1 2,094
158.2 2,040

Average Wellhead Price, Moody’s Electric Utility

Year Natural Gas (cents per Common Stocks Annual
million cubic feet) Average Price

1972 18.6 80.20

21.6 71.21
30.4 48.26
1975 44.5 51.25
58.0 60.10
79.0 67.55
90.5 63.54
117.8 60.28
1980 158.8 54.80
198.2 55.41
245.7 63.56
259.3 74.04
265.5 71.16
1985 248.0 87.24
194.2 111.11
166.5 105.90
169.0 97.99
169.0 110.45
1990 172.0 112.61
164.0 126.97
174.0
203.0
 APPENDIX C Data Sets and Databases 537
Thousands of Freight Car Loadings, Forest Products

Year 2002 2003 2004 2005 2006

January 17.6 16.9 16.1 15.6 13.2

February 19.0 18.5 15.8 16.7 14.5
March 19.1 18.6 16.0 17.2 14.7
April 18.7 17.9 16.6 16.7 14.1
May 19.1 18.0 16.2 16.3 15.2
June 19.9 17.4 15.8 16.4 15.3
July 18.3 16.4 16.6 13.9 14.9
August 19.1 16.5 15.8 15.9 14.0
September 18.6 16.7 15.6 15.5 15.3
October 18.8 17.0 15.6 15.2 14.4
November 16.7 12.4 14.4 13.9 14.2
December 17.3 16.9 14.3 14.0 15.0

Industrial Price/Earnings Ratio

Year Quarter P/E Year Quarter P/E

1999 1 17.56 2003 1 15.80

2 18.09 2 17.00
3 16.47 3 14.80
4 18.73 4 15.90
2000 1 22.16 2004 1 18.80
2 21.80 2 20.40
3 20.82 3 23.70
4 14.09 4 29.20
2001 1 13.61 2005 1 28.40
2 13.29 2 26.80
3 12.50 3 26.00
4 12.35 4 26.00
2002 1 12.46
2 13.30
3 14.73
4 15.40

DATABASES
Campsites
A group of businesspeople in Spokane is planning to develop a number of campsites in
the state. One of their problems is to decide the amount to be charged per day.
According to their observations, the fee depends on a number of variables such as
whether a swimming pool is accessible and the size of the campground. In an effort to
be objective, the following information was collected on 25 campgrounds in
Washington (from the Rand McNally Campground Guide for Washington) and a com-
puter analysis completed. The variables analyzed were
Y = Daily fee 1FEE2
X1 = Number of acres 1ACRES2
X2 = Number of sites 1SITES2
X3 = Whether there were flush toilets or not 1TOILET2
X4 = Whether there was a swimming pool or not 1POOL2
X5 = Whether there were electric hookups or not 1HOOKUP2
X6 = Number of additional recreational facilities 1ADD2
538 APPENDIX C Data Sets and Databases

Site Y X1 X2 X3 X4 X5 X6

1 7.00 40 32 0 0 0 2
2 8.50 20 47 1 0 1 2
3 9.00 45 18 1 1 1 1
4 8.00 110 32 1 0 1 3
5 8.00 30 54 1 0 1 2
6 7.00 50 30 1 0 1 3
7 7.75 35 30 1 0 1 2
8 8.00 18 40 1 0 1 1
9 8.50 23 60 1 1 1 1
10 8.50 9 60 1 0 1 3
11 9.00 52 50 1 1 1 2
12 7.00 25 21 0 0 1 1
13 9.00 250 30 1 0 1 2
14 8.50 140 70 1 1 1 2
15 9.00 120 80 1 1 1 1
16 7.50 60 50 1 1 1 2
17 8.50 120 35 1 0 1 2
18 9.00 173 25 1 1 1 2
19 8.00 100 75 1 0 1 2
20 9.50 134 35 1 1 1 1
21 7.50 114 120 0 1 1 2
22 7.50 2 17 0 0 1 2
23 7.50 32 15 0 1 0 3
24 9.00 25 30 1 1 1 2
25 7.50 66 100 1 0 1 2

Financial Data for Selected U.S. Companies

Cost of Labor and Research and

Capital Intangibles Goods Related Advertising Development
Company Sales Employees Expenditures Expenditures Sold Expense Expense Expense
Number ($ millions) (1,000s) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions)

1 3,221.8008 42.0000 147.9000 30.6000 2,285.2007 599.7998 118.3000 28.0000

2 1,690.6001 20.9050 93.0000 29.1000 1,057.2002 343.2000 114.9000 8.9000
3 2,197.2764 39.0000 66.8670 55.8600 1,387.0679 661.3997 95.5680 11.1820
4 2,357.8206 23.3000 59.5560 69.6080 1,743.7952 25.6320 51.9170 8.5000
5 8,129.0000 35.0000 297.0000 29.0000 7,423.0000 1,178.0000 12.8000 9.2530
6 11,851.0000 23.0000 394.0000 20.0000 10,942.0000 2,556.0000 11.6530 14.6000
7 323.8606 3.9000 2.5900 4.2880 233.5300 22.8350 3.5290 30.7320
8 660.4856 8.3780 10.9840 3.3720 582.2649 25.6250 44.9990 64.8730
9 4,351.1601 50.9120 102.7080 217.0920 4,156.8671 12.8360 66.2640 8.7790
10 985.8357 5.5000 16.6010 29.5900 874.1287 19.5000 112.3860 18.3650
11 3,802.5581 39.6000 206.1020 157.3520 2,997.2703 518.0000 139.7290 16.4130
12 2,576.0464 22.6000 50.6690 47.0790 1,885.9053 349.4910 48.8170 9.5000
13 106.0160 28.0000 1.3120 42.0000 84.6590 35.5550 22.9370 8.7330
14 5,669.8945 46.8810 103.0000 31.1000 4,424.3007 785.0000 141.3000 18.5000
15 319.6570 2.8940 4.5770 2.2090 246.6980 42.8370 87.0000 1.1000
16 511.7217 10.1000 19.5600 27.0000 286.2288 48.9990 1.8700 23.6520
17 884.6189 22.8010 58.0940 33.0000 467.4436 36.5000 16.0350 29.6320
18 166.3750 2.3000 3.9510 5.2890 111.0310 31.0000 4.0230 38.5420
19 59.1310 18.0000 1.1400 14.5000 43.7430 26.3210 90.3250 56.9820
20 136.6970 3.1000 2.0090 18.4930 105.3300 15.8880 46.3000 8.6330
21 767.8799 8.1000 37.4250 18.0560 519.3948 112.1350 21.8470 2.7860
22 61.3280 1.1390 1.3880 26.3250 35.2020 17.3140 2.4270 88.5230
23 445.6387 5.8000 18.9780 12.6000 213.2880 12.1000 62.8060 1.4600
 APPENDIX C Data Sets and Databases 539
Financial Data for Selected U.S.
U. S.Companies
Companies(Cont.)
Cost of Labor and Research and
Capital Intangibles Goods Related Advertising Development
Company Sales Employees Expenditures Expenditures Sold Expense Expense Expense
Number ($ millions) (1,000s) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions)

24 2,259.6316 16.0270 228.7270 27.3350 1,696.3772 421.8057 116.5990 9.6000

25 624.8040 8.7000 86.4030 2.8080 408.4707 168.0200 33.4700 9.4440
26 329.9578 4.0000 14.9460 8.3710 225.0410 20.9850 12.9790 32.0000
27 308.7327 2.1070 14.8080 43.5920 239.1300 36.5000 18.1220 1.8510
28 598.9507 5.0000 39.7150 27.8920 481.9436 45.0000 39.8230 .7500
29 172.7920 1.5760 1.6590 23.5420 118.7090 48.2000 7.9090 26.3330
30 910.8406 7.0000 14.4610 5.5880 677.2527 7.0000 58.2130 1.8000
31 142.1830 1.6000 5.5880 72.5190 126.9660 1.6000 2.7310 57.2710
32 425.0828 6.8330 72.5190 31.8030 256.2837 6.8330 12.1440 44.1550
33 4,337.9140 36.1000 306.0220 101.4290 2,344.1631 36.1000 270.2576 16.1100
34 271.0080 2.0780 27.1230 6.5030 134.3790 35.7730 20.2540 87.4830
35 209.4520 2.9980 14.4690 14.6060 176.4890 2.0780 1.8970 714.9990
36 62.4180 3.8000 3.7390 7.6680 34.4700 2.9980 44.0500 121.3300
37 4,300.0000 95.5000 412.2886 157.6030 2,108.5503 5.1000 257.6807 11.6440
38 390.6829 5.1000 30.8480 10.8550 225.1080 6.3000 18.3780 33.4770
39 270.0127 6.3000 40.0340 22.4540 189.8000 2.0000 4.9080 43.7430
40 97.9660 2.0000 6.9940 5.2500 64.5920 31.9700 2.5900 18.9700
41 66.4090 12.5261 3.7570 1.0090 57.2310 33.2000 59.1300 14.9460
42 56.5550 3.9000 1.6240 6.9940 44.0550 53.5000 19.5600 1.6590
43 3,267.9551 31.9790 502.0398 45.6140 2,517.7566 754.8977 3.9510 57.7210
44 2,745.7439 43.9680 251.0340 16.1110 1,638.7969 45.0000 161.2000 108.1480
45 2,609.0000 33.2000 248.0001 10.0000 1,874.0000 564.0000 18.0000 83.0000
46 1,677.6016 11.6440 284.6089 87.4830 1,185.9717 24.4530 6.4840 36.1310
47 6,887.6210 53.5000 1,075.1719 84.0390 4,721.9570 1,375.7996 44.0700 231.4690
48 10,584.1990 132.1400 714.2002 22.6000 7,353.5000 3,204.2688 93.4000 377.1001
49 2,912.7644 45.8540 195.2680 45.6430 2,189.5293 879.6548 14.9460 66.0560
50 4,309.5820 66.8000 275.3079 67.3120 2,913.9036 993.3997 1.6590 40.5470
51 1,946.4766 24.4530 121.3300 6.2920 1,403.4976 546.0508 35.2020 40.0810
52 9,254.1171 151.2000 1,431.0906 121.3300 6,187.7851 2,125.2012 95.9510 334.8057
53 5,018.6914 62.8510 479.8997 1.6240 3,478.0989 1,318.0999 9.2530 144.3000
54 1,510.7798 15.3000 207.9320 63.5190 1,157.2117 13.9700 27.6660 39.7150
55 1,560.0750 22.7000 162.5190 61.9380 1,188.9126 18.4340 19.3190 24.7010
56 2,794.0000 37.4000 256.0999 7.3000 1,928.4988 780.7996 18.3650 70.1000
57 921.3689 13.9700 61.9380 18.4340 597.7000 45.1640 19.2020 22.6500
58 1,253.5430 13.0580 66.4310 13.9700 806.6758 236.5000 32.0000 48.6510
59 1,328.1138 13.1160 201.1960 31.2730 851.8938 1.1550 31.2730 33.5620
60 1,314.6299 27.3460 36.9330 43.0750 569.7327 6.4690 174.4610 42.1160
61 7,869.6914 113.3710 687.7998 90.2000 5,580.5976 1,931.5005 76.5000 155.9000
62 73.0550 7.8240 26.5680 20.6650 38.9980 22.8990 43.0750 99.8430
63 108.5090 87.4350 5.6630 37.3860 77.1740 36.9990 90.2000 1.6500
64 1,422.4507 16.5000 100.4700 69.8820 1,060.5420 305.7000 6.3970 25.4520
65 87.4350 7.6550 8.5150 15.3750 51.3970 11.3940 69.8820 2.7200
66 7.8240 9.5280 26.6950 7.7640 6.7860 20.5720 4.2100 52.1780
67 868.7107 15.3400 42.4040 1.2120 686.0518 200.4850 10.4000 22.7240
68 137.3950 2.8750 14.1080 9.7470 112.2350 30.7620 83.1580 1.9000
69 753.8848 6.5480 24.2870 4.2120 596.5076 13.4000 88.8250 6.4200
70 1,445.0166 27.0030 84.1490 99.9080 786.8777 1.9360 39.8650 76.1870
71 3,062.6316 49.6190 67.6310 83.1580 1,446.5227 668.9910 243.0450 74.5240
72 2,450.4285 32.6000 81.9220 88.8250 906.9639 6.7120 423.2698 90.5730
73 141.2580 1.3040 4.5050 6.7300 95.1540 3.7000 9.9040 9.7580
540 APPENDIX C Data Sets and Databases

Financial Data for Selected U.S.

U. S.Companies
Companies (Cont.)

Cost of Labor and Research and

Capital Intangibles Goods Related Advertising Development
Company Sales Employees Expenditures Expenditures Sold Expense Expense Expense
Number ($ millions) (1,000s) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions)

74 6.8030 5.1000 9.5230 1.4590 2.3980 12.2490 .7230 11.9490

75 1,852.0896 25.4000 89.5500 57.7900 672.7947 4.5070 28.4910 148.0770
76 365.7217 4.9030 17.0620 16.7160 217.5420 3.4720 6.7300 11.8950
77 1,981.4397 28.7000 155.8530 141.2700 668.7720 634.0596 55.2940 161.3500
78 2,362.1326 40.7000 110.1000 99.8430 1,055.4187 11.3940 75.7000 113.1280
79 357.0696 5.5500 12.6430 52.1780 141.2700 2.1330 36.8860 18.9510
80 220.3790 3.7000 10.7860 9.7580 67.1220 20.5720 7.1610 6.2610
81 1,082.4927 17.9000 51.3360 52.1780 310.7820 315.8997 114.9660 65.6910
82 848.3799 17.1000 41.2990 11.9490 386.0066 16.0000 40.6150 61.6940
83 1,112.0386 16.5890 74.5790 44.6610 378.7710 7.3000 91.2150 77.3130
84 1,515.8816 37.0000 108.0460 52.3290 758.5320 469.9229 74.5950 61.8300
85 1,328.5508 19.9200 44.6810 6.2850 566.2200 323.7090 36.9560 115.5890
86 2,878.4956 58.0000 182.2670 348.1426 1,247.2339 1.1500 391.6277 85.3970
87 4,312.0507 56.6000 169.2950 66.9970 2,672.3262 6.4600 260.3870 37.6540
88 54.3250 37.3860 1.0660 2.8130 26.5960 4.7670 0.7520 44.6610
89 122.9470 57.1720 13.7480 7.5620 94.6720 17.6580 1.4590 3.8670
90 2,014.7056 31.0000 74.7910 0.0000 700.4778 503.6768 45.0900 21.1460
91 969.8328 18.5170 40.8340 54.2710 448.5286 9.4450 91.2690 8.5670
92 45.3670 8.3500 1.6430 7.0670 15.7310 2.1230 5.1820 52.3290
93 255.1320 3.3000 10.6420 20.2520 131.6750 12.2220 42.5670 6.2850
94 1,710.4700 31.7000 91.5640 54.7540 752.5889 530.2456 239.9010 42.0600
95 365.8809 3.4800 20.0140 6.7300 177.5500 25.8740 16.7100 23.7910
96 33.2650 2.0870 1.5120 4.4840 19.7100 19.7100 1.1550 2.8890
97 53.7460 .5250 2.0870 42.2810 16.1820 16.1800 7.6770 19.7100
98 52.8760 1.1420 2.4190 1.2160 27.1500 27.1500 6.4690 16.1820
99 9.6630 2.4190 12.7460 7.9670 5.6960 5.6950 .4570 27.1500
100 1,451.6687 29.0000 86.6820 97.2690 505.8267 36.1200 137.7250 30.7620
101 321.3638 4.9110 13.1180 11.0840 268.0159 57.2600 1.1110 13.4000
102 156.4580 2.3500 4.5670 3.8620 114.1930 6.4800 4.7670 5.6960
103 52.1870 .8650 1.5100 20.6490 36.5130 59.3250 18.0150 1.9360
104 447.2100 7.7670 12.7460 41.7940 280.3218 26.8120 9.4400 505.8267
105 86.8170 1.1000 1.2810 19.3850 57.2600 26.6950 2.1230 1.3420
106 1,132.3499 18.0150 16.8570 1.6970 785.0718 36.9240 25.8740 6.7120
107 217.4120 3.2000 4.4840 10.5440 142.6020 57.5790 3.2520 3.7000
108 7.7640 86.6820 1.2810 7.2210 6.4800 9.5280 20.8580 268.0159
109 1,581.8760 20.8580 142.2810 5.8820 1,280.1670 359.0999 1.1000 12.2490
110 201.4650 1.1000 7.9670 1.3370 169.2630 57.5700 7.4000 114.1930
111 198.9010 .9110 9.7470 .4290 164.1940 73.9670 1.8400 36.5130
112 1,497.0076 7.4000 131.9400 6.0210 1,098.2969 99.4080 5.1000 280.3280
113 153.2290 1.8400 11.0840 3.4390 59.2350 9.2800 8.3500 4.5070
114 367.9246 5.1000 20.6490 11.2110 230.1690 73.9670 1.2110 11.3940
115 494.4136 8.3500 19.3850 3.1490 342.9849 6.4690 3.8030 2.1000
116 52.4550 1.2120 7.2210 7.0620 26.8120 4.7670 4.0510 57.2600
117 37.3860 .8200 1.3370 44.3730 26.6950 9.4400 5.6000 785.0718
118 57.7120 13.1190 3.4390 .7160 36.9240 2.1230 7.5620 .9510
119 586.4766 3.8030 44.3730 34.2780 391.3706 25.8740 2.8100 20.5720
120 476.2078 4.0510 34.2780 30.2360 244.7830 99.9080 5.8820 16.0000
121 15.3570 4.5671 16.8570 53.2830 9.5280 29.0000 6.0200 142.6020
122 393.6016 5.6000 30.2360 2.8890 265.3079 9.2800 11.2110 7.3000
123 4,701.1210 7.5620 353.2830 48.6920 3,707.6846 4.9110 3.1490 6.4800
 APPENDIX C Data Sets and Databases 541
Financial Data for Selected U.S.
U. S.Companies
Companies (Cont.)

Cost of Labor and Research and

Capital Intangibles Goods Related Advertising Development
Company Sales Employees Expenditures Expenditures Sold Expense Expense Expense
Number ($ millions) (1,000s) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions)

124 1,167.8340 2.8100 48.6920 8.4580 1,017.6038 2.3500 7.0620 59.2350

125 12,298.3980 50.7000 1,221.8008 10.4000 9,285.7109 1,016.5000 13.1160 64.6000
126 439.4727 1.9020 65.1100 39.8650 263.8108 51.1480 27.3460 31.2730
127 29,127.0030 108.7000 1,897.0005 9.9040 20,032.0000 78.7700 16.5000 86.0000
128 1,993.6624 8.0000 43.4190 45.7820 1,755.5662 3.5730 31.1370 43.0750
129 4,660.8945 18.1000 636.1238 28.4900 3,675.6895 440.7996 3.4000 11.6000
130 976.4578 8.8280 14.8590 55.2940 879.3516 91.8000 15.3440 90.2000
131 3,834.9324 6.6610 316.7156 68.2690 3,557.4734 7.4050 2.8250 69.8820
132 9,535.7382 42.7800 1,107.3838 75.7000 7,075.1875 971.0000 6.5480 29.7730
133 657.7776 1.2640 56.1460 36.8860 565.0176 14.4700 27.0030 4.2120
134 100.4570 43.0750 44.0680 7.1610 72.7830 22.0310 49.6110 83.1580
135 60,334.5110 130.0000 4,186.9296 40.6150 45,999.0070 3,405.0000 32.6000 290.0000
136 2,150.0000 90.2110 311.7000 91.2150 1,460.7996 57.4030 1.3040 25.1000
137 18,069.0000 58.3000 1,680.0000 74.5900 13,442.0000 1,345.0000 25.4000 88.8250
138 109.7380 69.8870 32.2560 36.9560 97.0130 2.5200 4.9030 6.7300
139 592.7710 3.2520 123.7680 3.8770 420.3206 67.3300 28.7000 1.4590
140 4,642.3945 14.3280 353.5999 33.5620 4,085.0989 324.0000 40.7000 25.0000
141 2,072.4412 11.1480 270.1846 42.1160 1,640.8118 1.2400 5.5500 4.9810
142 4,509.3828 13.3540 502.2720 1.6500 2,838.0845 236.4540 2.0370 12.8000
143 34,736.0030 207.7000 1,760.7100 2.7200 26,053.9060 20.9400 3.7000 16.7160
144 1,191.0337 4.2070 255.6150 1.9000 865.6477 82.6730 .2670 99.8430
145 312.7300 4.2120 76.5000 6.4200 452.4130 17.0050 17.9000 52.1780
146 1,553.1077 9.1500 343.9539 23.6410 988.8760 185.6600 12.5840 9.7580
147 6,997.7734 30.0080 956.1719 11.2330 4,886.8125 720.5000 17.1000 58.2460
148 513.1880 5.1420 41.9800 41.9800 375.3599 25.0200 11.3330 11.9490
149 28,085.0030 94.8000 2,913.0000 32.5600 20,632.0000 2,344.0000 89.0000 231.0000
150 11,062.8980 34.9740 1,774.3904 43.0250 8,259.7656 1,051.0000 16.5890 114.0000
151 23,232.4060 37.5750 1,049.6729 90.2110 19,964.6050 994.0000 37.0000 89.7370
152 14,961.5000 47.0110 1,744.0364 69.8870 10,046.0000 1,126.7310 19.9200 80.3250
153 5,197.7070 24.1450 762.2510 4.2120 3,336.7566 431.9976 7.7130 15.0520
154 7,428.2343 33.7210 601.1216 6.7310 5,714.3085 9.7320 58.0000 21.0000
155 28,607.5030 67.8410 1,344.3777 10.4000 24,787.6050 1,572.7996 56.6000 52.0000
156 87.6100 6.7310 12.7120 39.8650 74.5510 31.5580 31.0000 44.6610
157 1,165.6736 3.5310 26.6780 9.9040 1,035.7129 6.6000 18.5170 2.4490
158 567.3650 1.5420 97.4910 28.9400 480.5110 23.5230 3.300 52.3290
159 5,954.9414 16.2970 732.0000 55.9240 4,540.4609 444.8997 31.7000 18.5000
160 368.0940 2.3150 15.0860 2.7160 319.4939 10.6050 3.4800 6.2850
161 751.7327 6.2550 51.1060 13.5380 606.8318 3.5230 6.8000 9.9000
162 895.4087 10.9000 145.5140 9.3840 681.9656 26.3250 39.0000 30.6000
163 1,063.2908 16.1790 51.1480 25.7670 746.2820 12.6000 16.6980 14.6320
164 1,306.0867 19.3970 78.7700 2.7490 1,021.4856 435.2998 23.3000 13.2830
165 140.4440 1.9190 3.5730 55.8600 122.3210 27.3350 35.0000 29.1000
166 4,357.2812 52.1400 110.4470 12.0830 3,540.9612 1,235.0000 3.9000 55.8600
167 263.9048 3.7000 7.4050 27.2080 203.3440 2.8080 8.3780 3.2500
168 6,184.8945 94.5000 398.2000 69.6080 5,224.0000 2,550.0000 50.9120 37.1000
169 257.6509 3.3640 14.4730 7.5700 190.4190 8.3710 5.5000 69.6080
170 50.5150 52.5350 29.1000 29.0000 18.0560 43.5920 39.6000 29.0000
171 419.6470 4.3020 22.0310 20.0000 341.5906 135.6000 22.6000 20.0000
172 1,227.4490 20.0000 57.4030 4.2880 999.7520 27.8920 28.0000 9.0000
173 779.3450 8.8000 22.0670 3.3700 678.4258 229.1270 46.8810 4.2880
542 APPENDIX C Data Sets and Databases

Financial Data for Selected U.S.

U. S.Companies
Companies (Cont.)

Cost of Labor and Research and

Capital Intangibles Goods Related Advertising Development
Company Sales Employees Expenditures Expenditures Sold Expense Expense Expense
Number ($ millions) (1,000s) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions)

174 72.1760 1.3000 2.5210 29.5900 50.9650 24.8290 2.8940 3.3700

175 3,248.0076 36.0620 263.6167 19.4460 2,710.3455 974.3379 10.1000 29.5900
176 921.1270 12.6590 67.3340 10.5250 771.0059 23.5420 22.8010 157.3520
177 711.9827 12.5120 133.3850 45.0790 653.8069 351.4700 2.3000 45.0790
178 72.4110 1.0250 1.2400 42.0000 60.0820 93.0000 18.0000 42.0000
179 297.5686 4.1520 20.9420 .8990 248.7160 123.1000 3.1000 31.1000
180 677.8489 6.0700 17.0050 31.1000 613.3047 169.2570 8.1000 2.2090
181 582.6238 1.4000 25.0290 2.2090 474.3450 66.8670 1.1390 27.0000
182 3,750.4109 38.1700 120.8280 27.0000 3,240.7886 1,132.6216 5.8000 33.0000
183 88.8070 2.3330 9.7320 7.2110 66.6540 59.5500 16.0270 5.2890
184 306.9397 2.8000 31.5880 33.0000 220.4980 2.5900 8.7000 14.5000
185 331.7366 5.2000 6.6000 11.0250 295.3848 10.9840 4.0000 18.4930
186 546.9500 8.9000 23.5230 5.2890 439.0479 16.0010 2.1000 18.0560
187 7.5910 30.6000 7.8900 14.5000 5.0480 50.6690 5.0000 26.3250
188 3,479.4573 41.3940 170.3720 18.4930 3,100.5391 1,177.5999 1.5760 15.0500
189 485.6138 6.6580 58.6750 3.5250 335.3318 42.0000 93.0000 1.4320
190 123.2280 2.0450 10.6050 5.2550 96.6630 20.9000 66.8670 12.6000
191 488.2327 4.6500 20.4800 1.1111 402.8457 402.8400 77.0101 22.2426
192 100.7820 1.7030 2.4430 1.6800 88.7960 4.0000 21.0000 28.3032
193 165.7970 4.7660 3.2790 88.0003 120.1080 2.0000 4.0008 18.2022
194 274.8440 3.5500 21.7900 2.9530 213.1860 3.0000 3.7521 24.2628
195 11,049.5000 166.8480 667.7998 55.5000 9,955.3984 4,485.1953 22.0007 52.5000
196 1,154.8477 14.4190 32.2360 4.0800 1,037.4727 424.4556 21.1234 30.3234
197 578.7107 11.4920 26.3000 8.0141 433.8230 1.0111 12.3456 5.3300
198 124.5440 1.8000 4.6280 1.9850 101.5300 23.6630 78.9101 36.3840
199 3,711.2029 63.4000 303.3838 4.5720 2,729.9280 22.0222 91.0111 33.0000
200 124.8600 2.0000 5.2240 2.3200 79.7770 51.0000 21.3141 2.6500
201 2,466.0000 26.8650 161.7000 2.0202 2,028.7996 18.4021 3.2000 14.9000
202 2,829.2991 36.2000 156.8000 27.1000 2,261.0000 930.2000 51.1617 25.2000
203 814.8196 14.8000 48.5520 16.0111 622.9507 204.9000 18.1920 1.4150
204 4,051.7996 46.0000 349.5999 2.6000 3,036.5999 1,215.2996 21.2223 56.6000
205 67.0390 28.0000 3.5010 2.5170 54.9070 66.5620 24.2526 42.4446
206 240.5670 4.0000 5.5670 1.3220 184.1350 61.6900 2.5860 3.0470
207 45.2140 2.0000 1.4110 18.1010 38.0970 62.3201 27.2829 48.5052
208 69.9520 81.0000 33.3333 8.0033 65.4570 52.3302 30.3132 54.5658
209 54.5490 1.1270 1.7720 17.7200 42.5990 42.4444 33.3435 60.6264
210 317.4480 5.7840 12.6650 11.0330 254.1990 80.1010 36.3738 66.6870
211 847.9927 24.0000 85.0240 19.7930 664.9578 34.1021 39.4041 10.4000
212 467.9546 4.8450 13.1650 2.3810 400.5806 4.0999 42.4344 1.0011
213 126.6750 14.0007 7.7490 14.1441 109.6830 50.6410 45.4647 1.0022
214 85.7290 49.0000 2.1610 49.4949 72.8400 9.9901 48.4950 1.0033
215 680.7666 8.2200 19.2340 77.7878 578.8528 9.8175 51.5253 1.3090
216 211.3230 1.5670 4.8350 15.6180 171.4130 65.0000 54.5556 1.8201
217 254.3030 3.1000 2.7620 2.3570 205.8410 42.4381 57.5859 2.0880
218 1,396.8108 29.4160 79.9820 28.2626 1,000.2886 3.8107 16.1580 3.4510
219 3,981.0000 52.9000 188.3000 70.3000 3,120.5999 1,085.7996 75.8000 37.5000
220 3,943.0990 56.5320 259.5000 49.9000 3,352.3008 1,275.7002 60.6162 42.3000
 APPENDIX C Data Sets and Databases 543
Financial Data for Selected U.S.
U. S.Companies
Companies (Cont.)

Cost of Labor and Research and

Capital Intangibles Goods Related Advertising Development
Company Sales Employees Expenditures Expenditures Sold Expense Expense Expense
Number ($ millions) (1,000s) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions) ($ millions)

221 1,260.2349 17.2880 103.0320 11.4810 1,055.9436 12.0000 63.6465 2.1133

222 973.2527 9.8850 25.4530 5.5580 848.7227 4.0877 66.6768 3.3210
223 19.9060 18.0002 5.6666 1.4100 16.5170 3.3876 69.7071 4.2242
224 66.8260 1.3200 6.1110 88.1388 48.9480 4.5222 72.7374 5.6210
225 178.7460 2.1980 5.5430 138.0000 138.5690 43.4350 75.7677 6.2155
226 26.7510 1.0560 8.8888 211.0113 17.9930 18.1111 78.7980 7.2102
227 20.5750 43.1111 7.7777 82.1003 13.9720 14.2222 81.8283 8.9712
228 51.5960 18.5216 1.6940 1.1620 38.8190 88.9922 81.0077 24.2601
229 106.1150 2.6000 4.6850 9.9210 64.0500 12.4243 77.0222 23.2810
230 8.5160 14.2421 12.0818 12.1402 5.9500 7.8246 22.4443 24.8588
231 308.8916 5.7000 15.8370 13.1402 144.7340 42.4444 47.8877 2.7060
232 753. 8069 16.8750 37.4620 3.6210 491.1160 210.0050 16.4370 4.9340
233 41.2960 1.1080 2.5820 12.1213 28.1320 81.8118 12.5456 24.5293
234 145.6840 3.4190 13.3250 1.0087 105.1630 51.7100 51.8196 1.8480
235 51.3130 1.0000 1.5700 8.0025 35.9730 43.4400 21.4231 59.6085
236 21.4070 12.5358 18.7842 5.5554 12.9550 12.8888 37.8286 64.8821
237 585.6597 8.2000 56.0530 80.9960 359.8350 77.9999 13.6920 8.9610
238 516.7239 10.3000 17.9320 9.3610 376.4170 1.1007 5.6670 5.6000
239 316.8147 7.0000 3.9360 12.1314 267.2456 2.0008 86.8686 76.7686
240 509.7000 10.0000 27.0360 15.1617 375.3457 179.9240 85.8686 3.6080
241 341.3887 7.1270 7.1570 8.1819 287.6907 9.0007 86.8888 86.7795
242 33.0660 1.0872 1.9540 9.2021 24.0720 12.7210 83.1111 95.9594
243 200.5920 4.0000 5.3320 20.0290 153.5480 7.6660 82.2222 94.9388
244 184.5810 4.0500 7.2780 10.3570 142.7160 8.7770 22.6665 1.0790
245 217.7520 4.0880 7.3840 10.1146 179.1020 78.3910 44.6621 89.9012
246 386.8118 7.4040 18.4880 47.1213 302.5586 2.9990 18.1716 3.8620
247 69.1530 12.1212 1.6190 48.1415 54.4310 11.3410 15.1413 13.8125
248 81.4650 1.6220 4.1890 16.4950 70.5080 4.4555 12.1110 47.8552
249 329.5518 6.0040 12.2520 8.0540 269.6377 12.1417 9.8765 51.9185
250 36.3870 133.0000 12.7246 51.5355 27.7690 21.8283 4.3210 54.3321
251 344.7937 7.5000 24.7400 57.5982 205.0610 92.9395 8.1234 4.8200
252 22.8030 84.1000 2.1060 83.4952 10.6830 96.9899 5.6788 43.8388
253 196.3030 5.4660 5.9730 99.9242 142.1520 97.9294 12.4582 2.2710
254 31.5660 13.7124 8.1264 10.1115 22.3750 95.0092 14.5220 66.7274
255 108.8580 1.7000 1.2870 92.4445 45.9130 92.6666 1.4330 53.5422
256 83.6260 1.2320 4.1220 55.6677 45.0950 92.5555 13.5620 22.5673
257 390.8726 6.1660 17.3310 40.5880 296.8577 58.2130 18.0000 10.0000
258 363.9839 7.0160 11.2700 11.5610 234.6320 2.7310 6.4860 86.0000
259 52.2620 .4420 5.1030 1.1500 43.5110 12.1440 44.0700 16.0000
260 228.6110 5.6500 1.8370 41.5600 161.4700 20.5400 14.9460 3.0730
261 60.8250 1.5000 1.4910 45.3100 41.6820 1.8970 1.6590 30.7300
262 16.6890 40.5000 57.6000 9.8540 9.8450 18.3700 35.2020 63.5300
263 39.8290 62.1000 3.9900 1.5610 32.6580 4.9080 9.2530 30.0000
264 28.9020 93.4000 1.1040 36.5000 23.1410 2.5900 18.3650 10.0000
265 8.7410 27.0000 55.6000 32.1400 6.3700 59.3100 27.6600 56.6660
266 61.9446 7.0000 35.0470 43.2000 432.3777 160.6660 19.2020 6.3530
 544
APPENDIX C
Executive (CEO) Compensation Data

Data Sets and Databases

Other Total Market
Salary Bonus Compens. Compens. Professional Tenure Exper. Valuation Pct Firm Profits Firm Sales
($1,000s) ($1,000s) ($1,000s) ($1,000s) Age Education* Background (years) (years) ($ millions) Own ($ millions) ($ millions)

173 275 5 453 64 2 3 40 26 54.5 3.14 91 872

1,441 429 78 1,948 55 1 1 23 23 7.6 0.55 145 1,237
1,646 0 89 1,735 47 2 7 5 5 21.7 0.52 -47 1,712
294 325 24 643 65 1 3 29 23 8.9 0.89 44 1,681
1,254 105 102 1,461 63 1 6 23 8 3.6 0.05 201 5,673
325 25 7 357 54 2 5 20 1 0.5 0.06 71 1,117
658 0 11 669 61 2 5 2 2 0.7 0.05 -187 1,475
1,723 289 82 2,094 63 1 3 41 8 5.9 0.04 1,166 10,818
504 69 24 597 57 2 5 27 13 1.4 0.03 377 2,686
822 38 29 889 56 2 1 5 5 0.7 0.03 224 2,201
374 129 11 514 57 2 4 3 3 4.1 0.17 79 661
447 11 8 466 48 2 1 17 1 0.2 0.01 189 1,539
2,781 0 52 2,833 50 1 3 4 4 11.7 0.39 -332 11,663
128 282 17 427 54 1 8 31 15 71.4 10.09 55 2,366
1,782 0 74 1,856 60 2 7 33 3 0.4 0.03 -507 4,864
1,137 423 92 1,652 60 2 1 34 14 11.5 0.06 856 14,949
761 20 1 782 49 2 9 18 9 1.3 0.07 14 5,061
505 0 108 613 56 2 9 8 1 0.1 0.02 -29 1,929
976 448 64 1,488 58 1 3 9 8 9.4 0.56 126 2,643
434 12 1 447 50 2 9 5 1 0.3 0.03 54 1,084
1,010 687 55 1,752 63 1 7 14 14 534.2 3.99 249 5,137
956 1,452 89 2,497 64 0 1 28 28 221.1 3.98 91 844
700 37 31 768 60 2 9 30 8 0.7 0.02 322 2,097
1,813 489 40 2,342 71 1 4 46 34 9.6 0.83 99 835
3,396 0 13 3,409 64 0 9 30 30 29.4 0.31 -99 12,021
 2,108 38 98 2,244 64 2 5 41 5 4.0 0.04 30 4,451
597 0 4 601 59 2 4 35 5 0.1 0.12 -85 1,911
616 862 76 1,554 61 1 8 41 17 30.6 2.23 82 1,435
237 221 4 462 61 2 5 25 11 16.8 1.03 27 1,314
571 0 16 587 55 2 4 5 5 1.6 0.17 -76 2,301
269 391 28 688 54 2 9 28 28 1,689.0 34.04 317 3,277
721 101 71 893 60 2 5 36 15 2.0 0.04 417 4,444
328 238 34 600 60 0 7 42 1 85.6 17.66 43 1,214
538 25 7 570 60 1 4 3 3 0.2 0.21 49 804
741 104 9 854 62 1 7 30 3 2.6 0.17 81 669
607 380 47 1,034 51 1 6 23 3 7.0 0.83 82 578
1,044 107 36 1,187 55 2 8 2 1 3.1 1.21 10 1,214
2,409 1,487 143 4,039 55 1 7 32 17 35.2 0.29 715 12,651
287 198 32 517 59 1 8 37 21 181.0 6.70 136 3,180
567 15 34 616 51 2 1 16 9 0.3 0.01 237 2,754
682 0 2 684 62 2 1 36 2 0.8 0.01 -1,086 12,794

APPENDIX C Data Sets and Databases

1,226 174 56 1,456 52 1 4 30 10 6.0 0.17 98 4,439
952 80 56 1,088 45 2 7 11 11 2.6 0.17 48 415
432 0 12 444 50 2 7 25 3 0.1 0.01 -50 1,569
1,085 440 97 1,622 57 1 8 29 5 9.7 0.19 347 9,886
1,009 117 0 1,126 64 1 3 28 14 3.4 0.21 63 2,545
1,711 182 134 2,027 62 2 2 25 5 3.4 0.04 806 8,379
408 183 9 600 52 2 7 12 2 4.2 0.26 10 21,351
543 13 2 558 62 2 5 34 12 0.2 0.01 265 2,359
278 209 74 561 61 2 6 24 24 15.4 0.95 52 695
*0 = no college degree; 1 = undergraduate degree; 2 = graduate degree

545
This page intentionally left blank
 Index

A Autoregressive (AR) models, 357–358, Bayesian Information Criterion (BIC)

367, 404–405, 445 and, 431–432
AAA Washington, 210–212, 271–274, application to management, 444–445 moving average (MA) model in,
328–330, 389–391, 469–471 autocorrelation, 400–405, 407 405–406
Ace Manufacturing, 269 Box-Jenkins technique and, seasonal analysis and, 432–442
Adaptive filtering, 80 404–405 model building strategy and,
Additive components model, 167, 176, Autoregressive moving average 407–411
179, 187, 192 (ARIMA) models, 78, 79, 399, Ljung-Box (modified Box-Pierce)
Akaike Information Criterion (AIC), 407, 445 statistic and, 410
431, 446 advantages of, 443–444 Box-Pierce statistic, 69, 429
Alcam Electronics, 53–54 Akaike information criteria (AIC), Box plot, 19–21
All-possible-regressions procedure, 431–432 Burns, Arthur F., 184
290–292 application to management, 444–445 Busby Associates, 514–517
Alomega Food Stores, 8–9, 56, 99, Bayesian information criterion (BIC), Business Activity Index for Spokane
212–213, 392–393, 497–498 431–432 County, 379–382
Alpha. See Significance level best practices for parameter Business cycle indicators, 184–185, 192
Alternative regression equation selection, 430 Business forecasting
slope, 40, 46 characteristics of, 399 acceptance by management, 9, 80
American Economic Association, 7 differencing and, 408 adequacy of forecasting techniques,
American Statistical Association, 7 disadvantages of, 443–444 84–86
Analysis (in a time series), 185 forecasting with, 411 applications to management, 44, 86,
Anderson-Darling test, 43 implementing model building and, 135–136, 189–190, 250, 252,
Anderson-Parente, 485 407–430 310–311, 367, 444–445
Applications to management, 44, 86, model checking, 410–411 averaging methods for, 78, 107,
135–136, 189–190, 250, 252, model estimation, 409–410 111–119, 134, 135 136
310–311, 367, 444–445 model identification, 407–409 choice and evaluation of method
ANOVA (Analysis of Variance), 231 seasonal analysis, 432–442 for, 4, 76–81, 107, 124–125,
Armstrong, J. Scott, 486 Average. See Mean 126, 136, 175, 180, 503–504,
Assumptions in regression analysis, 230, Averaging methods of forecasting, 78, 505–506
239, 339, 489 107, 111–119, 134, 135 136 combining forecasts, 486–488
Assumptions of past–future similarity, 4 correlation and, 34, 44, 221
Autocorrelation, 64–67, 87, 339–358, 367 cost, 6, 80, 504, 509, 511
Box-Jenkins technique and,
B data in, 5, 61–86
400–407 B and E datalinks, 7 data manipulation and cleaning
coefficient formula for, 65, 87 Bayesian Information Criterion (BIC) and, 5
and correlograms, 22, 67, 87 (Schwarz), 431, 446 decision theory and, 493–495
definition of, 64, 87 Bayes’ Theorem, 493–495, 496 Delphi method in, 483–485
Durbin-Watson test for, 343–347, 367 Bell and Hillmer, 179 error measurement in, 81–84, 88, 107,
first order serial correlation and, Bernstein, Peter, 1, 2, 444 122–126, 226–227, 309–310,
340–343 Beta. See Hypothesis testing 410–411
formulas for, 65, 87 Binomial distribution, 24, 44, 45 evaluation of, 5–6
generalized differences and, 347, Binomial equation, 24, 45 expected value and, 23, 45, 492, 496
354–357, 368 Boundary Electronics, 513–514 future of, 512–513
nonlinear least squares approach and, Box-Jenkins technique, 78, 399–445 growth curve procedure, 79, 80, 174,
355–357 advantages and disadvantages of, 246–250
partial, 400–403, 407 443–444 history of, 1–2
regression of time series data and, Akaike Information Criterion (AIC) independent vs. dependent variables,
339–358 and, 431–432 221, 281–282. See also
second order, 66 application to management, Autocorrelation; Collinearity;
solutions for, 347–358 444–445 Serial correlation
standard error of autocorrelation autoregressive (AR) model in, 404 intermediate, 79–80, 504
coefficient, 68, 87 autoregressive moving average intuitive. See Business forecasting,
Autocorrelation coefficient, 65, 87 (ARIMA) model in, 407 judgmental aspects of

547
548 Index
Business forecasting (continued) Alomega Food Stores, 56, 99, Computers
judgmental aspects of, 1, 3, 5, 7, 212–213, 392–393, 497–498, cost considerations, 80, 511
9–10, 115, 165–166, 246–247, 519–520 data accumulation and, 61
306–307, 430–431, 481–495, The Bond Market, 324–327 organization of forecasting function
503 Boundary Electronics, 513–514 and, 80, 511
long term, 80, 189–190, 504 Busby Associates, 514–517 technological shift, 1, 80, 495, 511–512
macroeconomic forecasting, 3–4 Business Activity Index for Spokane Computer simulation, 495
management of, 6, 503–513 County, 379–382 Confidence interval (interval estimate),
management information systems Butcher Products, Inc., 268–269 29, 46
(MIS) and, 511–512 City of College Station, 463–466 formula for mean in, 30, 46
micro forecast, 4 Company of Your Choice, 378 in regression, 228–229, 253,
multiple regression and, 281–311 Consumer Credit Counseling, 10–11, 291, 313
need for, 1, 2, 6, 80, 135–136, 165, 367, 98–99, 148, 206–207, 270–271, Consultants, forecasting by, 511
444, 503 388–389, 460, 517–518 Consumer Credit Counseling, 10–11,
neural networks and, 488–490 Fantasy Baseball (A), 330–334 98–99, 148, 206–207, 270–271,
personal judgment in. See Business Fantasy Baseball (B), 334–336 388–389, 460, 517–518
forecasting, judgmental Five-Year Revenue Projection for Continuous distribution, 25
aspects of Downtown Radiology, Continuous random variable, 23, 44
politics of, 9 149–154 Correlation, 34–39, 44, 221
problem formulation and data Golden Gardens Restaurant, 497 application to management, 44
collection in, 4–6 Lydia E. Pinkham Medicine Company, causation vs., 39
process of, 4–6, 503–513 461–463, 498–500 coefficient of, 37–39, 40, 44, 46,
quantitative vs. qualitative, 2, 3, 6, Mr. Tux, 10, 54–55, 97–98, 147, 202–206, 235, 289
79–80, 135–136, 503 270, 385–387, 459–460, linear relationship in, 39
regression analysis and, 34, 44, 80, 518–519 matrix, 282
221–252, 281–311, 339–367, Murphy Brothers Furniture, 94–96, scatter diagrams of, 34–37
503–504, 505 148, 207–209 Correlation matrix, 282
residuals in, 82, 87, 223, 239–241, Restaurant Sales, 383–385, 457–459 Correlogram, 22, 67, 87
410, 430 Small Engine Doctor, 201–202 Costs of forecasting, 6, 80, 504, 511
responsibility for, 510–511 Solar Alternative Company, 145–146 Cross-sectional data, 19, 61, 62, 87
rules for, 9 Southwest Medical Center, 158–159, Curved relationship, 35, 36
scenario writing in, 485–486 214, 394–395, 476–477, 520 Cyclical component, 63, 79, 87
short-term, 79–80, 185, 504 Surtido Cookies, 100, 159, 213–214, Cyclical-Irregular, 180, 193
steps in, 4–6, 509–510 393–394, 474–476 Cyclical variations, 166, 180, 185
subjective. See Business forecasting, Tiger Transport, 266–267
judgmental aspects of
technological, 481, 491
UPS Air Finance Division, 466–469
Web Retailer, 154–157, 471–474
D
time horizons, 79, 80 Causation vs. correlation, 38–39 Data
time series analysis and, 19, 62–64, Census II, 187–189 accumulation of, 5, 61
80, 165–192, 339–367, 381 Central limit theorem, 27 accuracy, 61
Business forecasting methods Central tendency, measures of, 18, 44 consistency, 61
autoregressive. See Autoregressive Chi square, 69, 527–528 cross-sectional, 19, 61, 62, 87
moving average (ARIMA) City of College Station, 463–466 cyclical, 63, 79
models Classical decomposition, 79, 80, 166–185 horizontal, 62
exponential methods for, 80, 107, Cleman, Robert, 487 nonstationary, 72, 109, 365, 407–408
111, 119–135, 136 Coefficient, autocorrelation. See random, 69–71
model building and evaluation, 5 Autocorrelation coefficient relevancy, 61
model selection criteria, 431–432 Coefficient of correlation, 37–39, 40, 44, reliability, 61
naïve methods for, 80, 108–111, 136 46, 235, 289 reliance on existing databases, 512
seasonal models, 130–134, 166–168, coefficient of determination and, seasonal, 63–64, 76, 79, 166–167
361–364, 432–442 235 stationary, 72, 78, 365, 407–408
smoothing methods for. See formula for, 38, 46 timeliness, 61
Exponential smoothing Coefficient of determination (in time series, 62–64, 87, 165
methods simple linear regression), trend in, 63, 72–75, 87
summary of, 80, 505–506 234–236, 239, 252 types of, 61–62
types of, 2–3 vs. coefficient of correlation, 235 Decision theory, 491–495
Business indicators, 184–185 formula for, 235, 254 Decomposition in time series analysis,
Business Week, 310 relation to F statistic, 239, 254 166–185
Butcher Products, Inc., 268–269 Coincident indicators, 184 Census II, method for, 187–189
Cointegrated relationship, 366, 367 classical, 79, 80, 166–185
C Cointegrated time series, 365, 367 problems with, 166
Collinearity, 281–282, 297–300, 312 Decomposition of variability, 230–234,
Case studies Combining forecasts, 486–488 287
AAA Washington, 210–212, benefits of, 488 Deflation, price. See Price deflation
271–274, 328–330, 389–391, simple average, 488, 496 Degrees of freedom, 17, 28–29, 44
469–471 weighted average, 488, 496 Delphi method, 483–485
Ace Manufacturing, 269 Computer examples. See Excel; Minitab advantage of, 485
Alcam Electronics, 53–54 Computer forecasting packages, 6–7 key features of, 483
 Index 549
Descriptive statistics, 15–22, 44 Fitted regression line, 223, 253 L
Deseasonalizing data, 179–184 500-Year Delta: What Happens After
Differencing, 72, 350–354, 354–357, What Comes Next (Taylor), 491 Lagged variables, 357. See also
368, 408 Five-Year Revenue Projection for Autoregressive Models
Discrete random variable, 22–23, 44–45 Downtown Radiology, 149–154 Lagging indicators, 184
Distributions Forecasting. See Business Forecasting Large sample 95% prediction interval,
binomial, 24, 44, 45, 523–524 Forecasting, need for, 1, 2, 6, 80, 135–136, 228, 253, 291, 313
chi-square, 527–528 165, 367, 444, 503 Leading indicators, 80, 184
F, 529–530 Fortune, 310 Least squares criterion, 223. See also
normal, 25, 45, 525 Fortune 500, 19 Method of least squares
probability, 22–26 F statistic, 238, 239, 254, 288–290, 312 Least squares line, 223
sampling, 26–29 multiple regression, for, 288–289 Level of confidence, 29
t, 28, 526 relation to coefficient of Linear exponential smoothing, 80
Dot plots, 19 determination, 239, 254, 290, 313 Linear relationship, 37–39, 221
Double moving average, 116–119, 137 Future Shock (Toffler), 491 Ljung (and Box), 69, 84–85, 87, 130
Dummy variables, 293–297, 312, 361 Long-term forecasts, 80, 189, 504, 505–506
Durbin-Watson test, 343–347, 367, 368, G
530–532
“Garbage in, Garbage out” (GIGO), 61
M
Gates, Bill, 490 Macroeconomic forecasting, 3–4, 63, 180
E Georgoff and Murdick, 483 Macro forecast, 3, 3–4
“Early Warning Signals for the Gibson, Rowan, 490 MAD. See Mean absolute deviation
Economy” (Moore and Global Paradox (Naisbitt), 491 Mahoud, E., 487
Shiskin), 185 Golden Gardens Restaurant, 497 Makridakis, Spyros, 6, 81, 187, 482
Econometric forecasting, 364–365 Gompertz growth model, 80, 174 Managers and forecasting. See Business
Econometric models, 79, 80, 364–365 Good fit, 37, 39 forecasting, applications to
Economic (business) indicators, Growth curves, 80, 174, 246–250 management
184–185, 192 Martin, Chuck, 490
Eight Asian Megatrends That Are H Mean (statistical), 15
Reshaping the World (Naisbitt), sample, of a, 15
491 Heteroscedasticity, 230, 358–361 stationary, 62
Error allowance, 29 Hibon, M., 81 Mean absolute deviation (MAD ), 79,
Errors, forecasting, 81–84, 87–88, 107, Histogram, 21–22, 239 88, 173
124–126, 225, 226–227, 307–309, Holt’s smoothing model, 126–130, 132, Mean absolute percentage error
410 137 (MAPE), 83, 88, 173
Estimation, 29–30, 44 Horizontal data, 62 Mean percentage error (MPE ), 83, 88
Excel examples Hypothesis testing, 30–32, 42–43, Mean squared deviation (MSD), 173
autocorrelation coefficients, 103–104 236–239, 289, 291 Mean squared error (MSE ), 82, 88, 231,
exponential smoothing, 161–163 232, 287
exponential trend model, 217–219 I Measures of central tendency. See Mean
first order autoregressive model, Median, 18
Imperfect negative linear relationship, 35
296–297 Megatrends (Naisbitt), 491
Independent variable, 221, 226, 252, 281
mean and standard deviation, 58–59 Megatrends 2000 (Naisbitt and
Index number (for seasonal variation),
multiple regression, 337–338 Aburdene), 491
176, 192
simple linear regression, 277–278 Method of least squares, 39, 46, 169
Indicator variables. See dummy
Expected value, 23, 45, 492, 496 regression analysis in, 223, 253, 284
variables
Exponential smoothing methods, 79, 80, Methods, forecasting. See Business
Intermediate (medium-term) forecasts,
107, 119–135, 136, 444, 506 forecasting
79, 80, 504, 505–506
adjusted for trend (Holt’s method), Methods for selecting multiple regression
International Journal of Forecasting, 486
126–130, 132, 137 equation, 300–307
Interquartile range, 18
adjusted for trend and seasonal Microcomputers. See Computers
Interval estimate, 29, 30, 45, 46
variation (Winters’ method), Micro–macro continuum, 3
Interval prediction, 228
130–135, 137 Minitab examples
Intuitive judgment. See Business
application to management, 135–136 ARIMA, 478–480
Forecasting, judgmental
short-term forecasting and, 134 autocorrelation and differencing,
aspects of
tracking, 124–126 101–102
Irregular component (in time series
decomposition of time series, 214–217
analysis), 167, 180
descriptive statistics, 56–58
F moving average and exponential
Fantasy Baseball (A), 330–334 J smoothing, 159–161
Fantasy Baseball (B), 334–336 Judgmental forecasting. See Business regression analysis, 274–277,
Federal Reserve Board, 187 forecasting, judgmental stepwise regression, 336–337
Federal Reserve Bulletin, 165 aspects of time series regression, 395–396
Fildes, R., 81 Mr. Tux, 10, 54–55, 97–98, 147, 202–206,
First differencing, 351, 368 270, 385–387, 459–460, 518–519
First order serial correlation, 340, 367
K Mitchell, Wesley, 184
Fitted (estimated) regression equation, Kamp, Di, 490 Model building, 5, 407–411
224, 253, 284, 312 k th order autocorrelation, 65, 87 Model implementation, 5
550 Index
Monthly Labor Review, 165 Net Future: The 7 Cybertrends That Random series, 68
Moore, Geoffrey H., 185 Will Drive Your Business, Random variable, 22, 23
Morgenthau, Henry, 184 Create New Wealth, and Define expected value of, 23
Moving average (MA) models, 78, 79, Your Future (Martin), 490 Range, 18
80, 113–119, 136, 137, 405–406 Neural networks (and forecasting), Ratio-to-moving-average method,
disadvantages of, 113 488–490 176–179
MSE. See Mean squared error NeuralWare, Inc., 489 Regression analysis, 39–42, 79, 80,
M3-IJF competition, 81 Nonstationary data, 72, 109, 365, 221–254, 281–313, 339–368
Multicollinearity, 281, 297–300 407–408 all-possible-regressions procedure in,
Multiple regression, 79, 80, 281–313, Normal distribution, 25, 45, 525 302–304
312, 505 Normal scores plot, 42 application to management, 251, 252,
adjusted correlation coefficient for, Normality, assessment of, 42–43 310–311
290, 313 Null hypothesis, 30–31 assumptions in, 230, 239, 339, 489
all-possible-regression procedure in, autocorrelation and, 240, 254, 339–358
302–304 O coefficient of correlation and, 235,
ANOVA table, 288 289, 313
application to management, 310–311 Outliers, 20 coefficient of determination in,
coefficient of determination, 289, 312 Overfitting, 309–310, 312 234–236, 239, 254, 289–290,
correlation coefficient, 289, 313 312, 313
correlation matrix and, 282 P decomposition of variance, 230–234,
definition of, 281 253, 287, 312
Parente, 485
degrees of freedom in, 287, 312 fitted regression function, 46, 223, 252,
Partial or net regression coefficients,
dummy variables in, 293–297, 312 253, 284, 312
285, 286, 312
econometric forecasting and, 310–311, forecasting Y, 227–230, 291, 313
Past-future similarity, assumptions of, 4
364–365 F statistic, 238, 239, 254, 289, 312
Pearl-Reed logistic model, 174
equation, selection of “best,” 300–307 hypothesis testing in, 236–238, 254, 289,
Perfect negative linear relationship, 35
F statistic for, 289, 312 291, 312, 313
Perfect positive linear relationship, 35
function of, 283–284, 312 interval prediction, 228, 253, 291, 313
Personal (subjective, intuitive) judgment.
inference for models, 286–291 method of least squares, 39, 46, 223,
See Business forecasting,
interpretation of, 285–286 253, 284
judgmental aspects of
models for, 283–285 point prediction equation, 227, 291
Plug-compatible, 489
multicollinearity and, 281–282, 297–300 population regression equation,
Point estimate, 29, 30, 45
overfitting in, 309–310, 312 224–225, 283, 312
Point prediction, 227
partial or net regression coefficients in, predictor variables in, 243, 281, 339
Politics of forecasting, 9
285, 286, 312 residuals in, 239–241, 307–309
Population (statistical), 15, 37
population multiple regression sample regression line in, 226
Positive relationship, 35
function, 283, 312 serial correlation and, 230
Prediction
predictor variables in, 281–282, simple linear regression, 221–254
interval, 228, 253, 291, 313
290–291 simple linear vs. multiple regression,
point, 227, 291
regression coefficients and, 285–286 283
Predictor variables, 243, 281, 339
residuals and, 307–309 standard error of estimate for,
Price deflation, 191, 192
selection of equation in, 300–307 226–227, 252, 253, 287–288,
Price index, 190–191
t statistic for, 291, 313 312
Pricewaterhouse Coopers, 491
vs. simple regression, 283 standard error of the forecast for,
Principle of parsimony, 409, 445
standard error of estimate in, 287–288, 228, 252, 253, 291
Prob value. See p -value,
312 standard error of the regression
Probability distributions, 22–26
stepwise regression procedure in, coefficient for, 237, 254,
Product life cycle, 171–172
304–307, 312 291, 297
Publications on technological change
variance inflation factor, 297, 313 stepwise regression and, 304–307
(books), 490–491
Multiplicative model, 167, 176, 192 straight line regression model of, 226
p-value, 32–33, 45
Murphy Brothers Furniture, 94–96, 148, sum of squares error (SSE), 231, 232,
207–209 234, 235, 253, 287, 289, 312
Q sum of squares regression (SSR), 231,
N Q statistic, 69, 87 234, 235, 253, 287, 289, 312
Quadratic exponential smoothing, 80 sum of squares total (SST ), 231, 232,
Naisbitt, John Qualitative forecasting. See Business 234, 235, 253, 287, 289, 312
Eight Asian Megatrends That forecasting, judgmental time series data, with, 339–368
Are Reshaping the World aspects of t statistic for, 237, 254, 291, 313
(audio tape), 491 Quantitative vs. qualitative, 2, 3, 4, 79–80, variable transformations and, 239,
Global Paradox, 491 135, 503 243–246
Megatrends, 491 Quartiles, 18 Regression line, 222–226
Megatrends 2000 (with Patricia Quenouille, M. H., 68 Reliability (of data), 61
Aburdene), 491 Residuals, 82, 87, 180, 223, 239–241,
Naive methods, 78, 80, 108–111, 136 253, 307–309, 343–344,
formulas for, 108–110, 136
R 410
National Bureau of Economic Research Rand Corporation and Delphi method, multiple regression, in, 307–309
(NBER), 184, 185 483–485 time series, in, 343–344, 410
Negative relationship, 35, 37 Random model, 69, 87 Restaurant Sales, 383–385, 457–459
 Index 551
Rethinking the Future: Rethinking Standard error of the forecast, 228, 252, Trend, 62, 63, 72–75, 87, 109,
Business, Principles, 253, 291 168–175, 192
Competition, Control and Stationary time series, 72, 78, 365, definition of, 63, 87
Complexity, Leadership, 407–408 exponential, 173, 192
Markets and the World (Gibson, Statistical Abstract of the United States, linear, 169, 192
ed., Toffler, and Toffler), 490 165 quadratic, 172, 192
Road Ahead, The (Gates), 490 Statistical Applications for the Sciences Trend-cycle, 166, 169
Root mean square error (RMSE ), 83, 88 (SAS), 7 t statistic, 28, 46, 237, 254, 291, 313
Statistical inference, 15, 29–34 21st Century Manager: Future-Focused
S Statistics (descriptive), 15–19 Skills for the Next Millennium
Stepwise regression procedure, 304–307 (Kamp), 490
Sample, 15, 29, 37, 44 Straight-line regression model, 226
correlation coefficient of, 38, 46 Surtido Cookies, 100, 159, 213–214, U
mean of, 15 393–394, 474–476
standard deviation of, 16 Survey of Current Business, The, Unobserved components model, 167
variance of, 16 165, 184 UPS Air Finance Division, 466–469
Sampling distributions, 26–29, 45
Scatter diagrams, 22, 34–37, 45, 61
T V
Scenario writing, 485–486
S-curve fitting, 80 Taylor, Jim, 491 Variables
Seasonal adjustment, 167, 179–180, 193 t distribution, 28, 526 continuous random, 23, 44
Seasonal component of time series, 76, Technology Forecast, 491 dependent, 221, 226, 252, 281
87, 166–167, 175–180 Technological forecasting, 481, 491 discrete random, 22–23, 44–45
and Box-Jenkins technique, 432–442 The Bond Market, 324–327 dummy, 293–297, 312, 361
exponential smoothing adjusted for, The Lydia E. Pinkham Medicine endogenous, 364, 365
130–135 Company, 461–463, 498–500 exogenous, 364, 365
moving average and, 116 Third Wave, The (Toffler), 491 independent, 221, 226, 252, 281, 271,
time series analysis and, 63–64, 110, Tiger Transport, 266–267 274
166–167 Time horizon, 79, 80 predictor, 243, 281, 339
Serial correlation, 339–358 Time series analysis, 19, 62–64, 79, 80, random, 22, 23
autoregressive models and, 357–358 339–368, 399–446 Variable transformations, 239, 243–246
Durbin-Watson test for, 343–347 autocorrelated errors and, 354–358 Variance, 16
solutions for, 347–358 co-integration in, 365–367 heteroscedasticity and, 230, 358–361
Shiskin, Julius, 185, 187 decomposition of, 166–185
Short-term forecasts, 79, 80, 185, 504 economic forecasting and, 364–365 W
Significance level, 31 first-order autoregressive model,
Wall Street Journal, 310
Significance probability, 33, 45 357–358, 404
Web Retailer, 154–157, 471–474
Simple average, 111–113, 136, 137 forecasting of, 174–175, 185–187,
White noise model, 69
Simple exponential smoothing, 361–364, 411
Winters’ model, 130–135, 137
119–124, 137 generalized differences in, 354–357
World Wide Web, 7, 165, 513
Simple linear regression, 221–254 log-linear regression model and, 368
Simple moving average, 113–116, 137 model specification error in, 348–349
Simple regression analysis application regression of and first-order serial X
to management, 250, 252 correlation, 340–343 X-12 ARIMA, 187–189
Slope, in linear relationship, 39, 223, 253 regression with differences, 350–357 X –Y relationship, 224, 225 See also
Small Engine Doctor, 201–202 seasonal model, 361, 368, 332–342 Regression analysis
Smoothing. See Exponential smoothing solutions to autocorrelation,
Solar Alternative Company, 145–146 347–358
Southwest Medical Center, 158–159, trend in, 168–175 Y
214, 394–395, 476–477, 520 Toffler, Alvin Y, forecasting of, 227–230, 291, 313
Standard deviation, 16 Future Shock, 491 Y -intercept, 39, 46, 223, 253
Standard error of autocorrelation Third Wave, The, 491
coefficient, 68, 87 Tracking, 124–126 Z
Standard error of the estimate, 226–227, Tracking signal, 124, 125, 136
252, 253, 287–288, 312 Tree diagram, 493 Z-scores, 25, 38
Areas for Standard Normal Probability Distribution

0 z

For z  1.93, shaded area is .4732 out of total area of 1.

z .00 .01 .02 .03 .04 .05 .06 .07 .08 .09

0.0 .0000 .0040 .0080 .0120 .0160 .0199 .0239 .0279 .0319 .0359
0.1 .0398 .0438 .0478 .0517 .0557 .0596 .0636 .0675 .0714 .0753
0.2 .0793 .0832 .0871 .0910 .0948 .0987 .1026 .1064 .1103 .1141
0.3 .1179 .1217 .1255 .1293 .1331 .1368 .1406 .1443 .1480 .1517
0.4 .1554 .1591 .1628 .1664 .1700 .1736 .1772 .1808 .1844 .1879
0.5 .1915 .1950 .1985 .2019 .2054 .2088 .2123 .2157 .2190 .2224
0.6 .2257 .2291 .2324 .2357 .2389 .2422 .2454 .2486 .2518 .2549
0.7 .2580 .2612 .2642 .2673 .2704 .2734 .2764 .2794 .2823 .2852
0.8 .2881 .2910 .2939 .2967 .2995 .3023 .3051 .3078 .3106 .3133
0.9 .3159 .3186 .3212 .3238 .3264 .3289 .3315 .3340 .3365 .3389
1.0 .3413 .3438 .3461 .3485 .3508 .3531 .3554 .3577 .3599 .3621
1.1 .3643 .3665 .3686 .3708 .3729 .3749 .3770 .3790 .3810 .3830
1.2 .3849 .3869 .3888 .3907 .3925 .3944 .3962 .3980 .3997 .4015
1.3 .4032 .4049 .4066 .4082 .4099 .4115 .4131 .4147 .4162 .4177
1.4 .4192 .4207 .4222 .4236 .4251 .4265 .4279 .4292 .4306 .4319
1.5 .4332 .4345 .4357 .4370 .4382 .4394 .4406 .4418 .4429 .4441
1.6 .4452 .4463 .4474 .4484 .4495 .4505 .4515 .4525 .4535 .4545
1.7 .4554 .4564 .4573 .4582 .4591 .4599 .4608 .4616 .4625 .4633
1.8 .4641 .4649 .4656 .4664 .4671 .4678 .4686 .4693 .4699 .4706
1.9 .4713 .4719 .4726 .4732 .4738 .4744 .4750 .4756 .4761 .4767
2.0 .4772 .4778 .4783 .4788 .4793 .4798 .4803 .4808 .4812 .4817
2.1 .4821 .4826 .4830 .4834 .4838 .4842 .4846 .4850 .4854 .4857
2.2 .4861 .4864 .4868 .4871 .4875 .4878 .4881 .4884 .4887 .4890
2.3 .4893 .4896 .4898 .4901 .4904 .4906 .4909 .4911 .4913 .4916
2.4 .4918 .4920 .4922 .4925 .4927 .4929 .4931 .4932 .4934 .4936
2.5 .4938 .4940 .4941 .4943 .4945 .4946 .4948 .4949 .4951 .4952
2.6 .4953 .4955 .4956 .4957 .4959 .4960 .4961 .4962 .4963 .4964
2.7 .4965 .4966 .4967 .4968 .4969 .4970 .4971 .4972 .4973 .4974
2.8 .49740 .4975 .4976 .4977 .4977 .4978 .4979 .4979 .4898 .4981
2.9 .4981 .4982 .4982 .4983 .4984 .4984 .4985 .4985 .4986 .4986
3.0 .49865 .4987 .4987 .4988 .4989 .4989 .4989 .4989 .4990 .4990
4.0 .4999683
This Software Product was reproduced by Publisher under a special arrangement with Decisioneering, Inc.
For this reason, Publisher is responsible for the product warranty. If your diskette is defective, please return
it to Publisher, who will arrange for its replacement. PLEASE DO NOT RETURN IT TO OR CONTACT
DECISIONEERING, INC. FOR SOFTWARE SUPPORT. This Software Product is provided as part of an
integrated part of this academic textbook. No support is provided by the Publisher or Decisioneering, Inc.
To the extent of any inconsistencies between this statement and the End User License Agreement which
accompanies this Software Product, this statement shall govern.

1
 LICENSE AGREEMENT AND LIMITED WARRANTY
READ THIS LICENSE CAREFULLY BEFORE USING THIS PACKAGE. BY USING THIS PACKAGE, YOU ARE AGREEING TO THE
TERMS AND CONDITIONS OF THIS LICENSE. IF YOU DO NOT AGREE, DO NOT USE THE PACKAGE. PROMPTLY RETURN THE
UNUSED PACKAGE AND ALL ACCOMPANYING ITEMS TO THE PLACE YOU OBTAINED IT. THESE TERMS APPLY TO ALL
LICENSED SOFTWARE ON THE DISK EXCEPT THAT THE TERMS FOR USE OF ANY SHAREWARE OR FREEWARE ON THE
DISKETTES ARE AS SET FORTH IN THE ELECTRONIC LICENSE LOCATED ON THE DISK.

1. GRANT OF LICENSE and OWNERSHIP: The enclosed computer programs and data (“Software”) are licensed, not sold, to you by Prentice-Hall,
Inc. (“We” or the “Company”) in consideration of your purchase or adoption of the accompanying Company textbooks and/or other materials, and your
agreement to these terms.We reserve any rights not granted to you. You own only the disk(s) but we and/or licensors own the Software itself. This license
allows individuals who have purchased the accompanying Company textbook to use and display their copy of the Software on a single computer (i.e., with
a single CPU) at a single location for academic use only, so long as you comply with the terms of this Agreement. You may make one copy for back up, or
transfer your copy to another CPU, provided that the Software is usable on only one computer. This license allows instructors at educational institutions
[only] who have adopted the accompanying Company textbook to install, use and display the enclosed copy of the Software on individual computers in the
computer lab designated for use by any students of a course requiring the accompanying Company textbook and only for as long as such textbook is a
required text for such course, at a single campus or branch or geographic location of an educational institution, for academic use only, so long as you
comply with the terms of this Agreement.

Notwithstanding the foregoing, the above license grant shall not apply to that portion of the Software currently entitled Crystal Ball Professional 2000. Use
of Crystal Ball Professional 2000 is governed by the electronic End User License Agreement for Crystal Ball Professional 2000 which is included on this
disk, which must be agreed to prior to its use. Also notwithstanding the foregoing, the above terms concerning computer lab usage will not apply to that
portion of the Software currently entitled Crystal Ball Professional 2000. Use of Crystal Ball Professional 2000 in a computer lab is expressly prohibited by
this license. For details on Crystal Ball Professional 2000, including lab licenses for such product, please see the electronic End User License Agreement
for Crystal Ball Professional 2000, which is included on this disk. Should an academic institution wish to obtain a Lab License, please contact
Decisioneering at 800-289-2550.

2. RESTRICTIONS: You may not transfer or distribute the Software or documentation to anyone else. You may not copy the documentation or the
Software. You may not reverse engineer, disassemble, decompile, modify, adapt, translate, or create derivative works based on the Software or the
Documentation. You may be held legally responsible for any copying or copyright infringement which is caused by your failure to abide by the terms of
these restrictions.

3. TERMINATION: This license is effective until terminated. This license will terminate automatically without notice from the Company if you fail to
comply with any provisions or limitations of this license. Upon termination, you shall destroy the Documentation and all copies of the Software. All
provisions of this Agreement as to limitation and disclaimer of warranties, limitation of liability, remedies or damages, and our ownership rights shall
survive termination.

4. LIMITED WARRANTY AND DISCLAIMER OF WARRANTY: Company warrants that for a period of 60 days from the date you purchase this
Software (or purchase or adopt the accompanying textbook), the Software, when properly installed and used in accordance with the Documentation, will
operate in substantial conformity with the description of the Software set forth in the Documentation, and that for a period of 30 days the disk(s) on which
the Software is delivered shall be free from defects in materials and workmanship under normal use. The Company does not warrant that the Software will
meet your requirements or that the operation of the Software will be uninterrupted or error-free. Your only remedy and the Company’s only obligation
under these limited warranties is, at the Company’s option, return of the disk for a refund of any amounts paid for it by you or replacement of the disk.
THIS LIMITED WARRANTY IS THE ONLY WARRANTY PROVIDED BY THE COMPANY AND ITS LICENSORS, AND THE COMPANY
AND ITS LICENSORS DISCLAIM ALL OTHER WARRANTIES, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION, THE
IMPLIEDWARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE COMPANY DOES NOT
WARRANT, GUARANTEE, OR MAKE ANY REPRESENTATION REGARDING THE ACCURACY, RELIABILITY, CURRENTNESS, USE,
OR RESULTS OF USE, OF THE SOFTWARE.

5. LIMITATION OF REMEDIES AND DAMAGES: IN NO EVENT, SHALL THE COMPANY OR ITS EMPLOYEES,AGENTS, LICENSORS,
OR CONTRACTORS BE LIABLE FOR ANY INCIDENTAL, INDIRECT, SPECIAL, OR CONSEQUENTIAL DAMAGES ARISING OUT OF
OR IN CONNECTION WITH THIS LICENSE OR THE SOFTWARE, INCLUDING FOR LOSS OF USE, LOSS OF DATA, LOSS OF INCOME OR
PROFIT, OR OTHER LOSSES, SUSTAINED AS A RESULT OF INJURY TO ANY PERSON, OR LOSS OF OR DAMAGE TO PROPERTY,
OR CLAIMS OF THIRD PARTIES, EVEN IF THE COMPANY OR AN AUTHORIZED REPRESENTATIVE OF THE COMPANY HAS BEEN
ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. IN NO EVENT SHALL THE LIABILITY OF THE COMPANY FOR DAMAGES
WITH RESPECT TO THE SOFTWARE EXCEED THE AMOUNTS ACTUALLY PAID BY YOU, IF ANY, FOR THE SOFTWARE OR THE
ACCOMPANYING TEXTBOOK. SOME JURISDICTIONS DO NOT ALLOW THE LIMITATION OF LIABILITY IN CERTAIN
CIRCUMSTANCES,THE ABOVE LIMITATIONS MAY NOT ALWAYS APPLY.

6. GENERAL: THIS AGREEMENT SHALL BE CONSTRUED IN ACCORDANCE WITH THE LAWS OF THE UNITED STATES OF
AMERICA AND THE STATE OF NEW YORK, APPLICABLE TO CONTRACTS MADE IN NEW YORK, AND SHALL BENEFIT THE
COMPANY, ITS AFFILIATES AND ASSIGNEES. This Agreement is the complete and exclusive statement of the agreement between you and the
Company and supersedes all proposals, prior agreements, oral or written, and any other communications between you and the company or any of its
representatives relating to the subject matter. If you are a U.S. Government user, this Software is licensed with “restricted rights” as set forth in
subparagraphs (a)-(d) of the Commercial Computer-Restricted Rights clause at FAR 52.227-19 or in subparagraphs (c)(1)(ii) of the Rights in Technical
Data and Computer Software clause at DFARS 252.227-7013, and similar clauses, as applicable.

Should you have any questions concerning this agreement or if you wish to contact the Company for any reason, please contact in writing:

Director, Media Production

Pearson Education
1 Lake Street
Upper Saddle River, NJ 07458