R06 Time-Series Analysis - Answers

Question #1 of 112 Question ID: 1208528
Troy Dillard, CFA, has estimated the following equation using semiannual data: xt = 44 + 0.1×xt– 1 –
0.25×xt– 2 - 0.15×xt– 3 + et. Given the data in the table below, what is Dillard's best forecast of the second
half of 2007?
Time Value
2003: I 31
2003: II 31
2004: I 33
2004: II 33
2005: I 36
2005: II 35
2006: I 32
2006: II 33
A) 34.05.
B) 33.74.
C) 34.36.
Explanation
To get the answer, Dillard must rst make the forecast for 2007:I
E[x2007:I]= 44 + 0.1 × xt– 1 - 0.25 × xt– 2 - 0.15 × xt– 3
E[x2007:I] = 44 + 0.1×33 - 0.25×32 - 0.15×35
E[x2007:I] = 34.05
Then, use this forecast in the equation for the rst lag:
E[x2007:II] = 44 + 0.1×34.05 - 0.25×33 - 0.15×32
E[x2007:II] = 34.36
(Study Session 2, Module 6.2, LOS 6.d)
A time series x that is a random walk with a drift is best described as:
A) xt = b0 + b1xt − 1 + εt.
B) xt = xt − 1 + εt.
C) xt = b0 + b1 xt − 1.
Explanation
The best estimate of random walk for period t is the value of the series at (t − 1). If the random walk has
a drift component, this drift is added to the previous period's value of the time series to produce the
forecast.
(Study Session 2, Module 6, LOS 6.i)
Suppose you estimate the following model of residuals from an autoregressive model:
εt2 = 0.25 + 0.6ε2t-1 + µt, where ε = ε^
If the residual at time t is 0.9, the forecasted variance for time t+1 is:
A) 0.790.
B) 0.850.
C) 0.736.
Explanation
The variance at t = t + 1 is 0.25 + [0.60 (0.9)2] = 0.25 + 0.486 = 0.736. See also, ARCH models.
(Study Session 2, Module 6.5, LOS 6.m)
Which of the following statements regarding an out-of-sample forecast is least accurate?
A) Forecasting is not possible for autoregressive models with more than two lags.
B) There is more error associated with out-of-sample forecasts, as compared to in-sample

forecasts.
C) Out-of-sample forecasts are of more importance than in-sample forecasts to the analyst using
an estimated time-series model.
Explanation
Forecasts in autoregressive models are made using the chain-rule, such that the earlier forecasts are
made rst. Each later forecast depends on these earlier forecasts.
(Study Session 2, Module 6, LOS 6.g)

The regression results from tting an AR(1) to a monthly time series are presented below. What is the
mean-reverting level for the model?
Model: ΔExpt = b0 + b1ΔExpt–1 + εt
Coe cients Standard Error t-Statistic p-value
Intercept 1.3304 0.0089 112.2849 < 0.0001
Lag-1 0.1817 0.0061 30.0125 < 0.0001
A) 1.6258.
B) 0.6151.
C) 7.3220.
Explanation
The mean-reverting level is b0 / (1 − b1) = 1.3304 / (1 − 0.1817) = 1.6258.
(Study Session 2, Module 6.2, LOS 6.f)
Dianne Hart, CFA, is considering the purchase of an equity position in Book World, Inc, a leading seller of
books in the United States. Hart has obtained monthly sales data for the past seven years, and has plotted
the data points on a graph. Which of the following statements regarding Hart's analysis of the data time
series of Book World's sales is most accurate? Hart should utilize a:
A) linear model to analyze the data because the mean appears to be constant.
B) mean-reverting model to analyze the data because the time series pattern is covariance
stationary.
C) log-linear model to analyze the data because it is likely to exhibit a compound growth trend.
Explanation
A log-linear model is more appropriate when analyzing data that is growing at a compound rate. Sales
are a classic example of a type of data series that normally exhibits compound growth.
(Study Session 2, Module 6.1, LOS 6.b)
Which of the following statements regarding a mean reverting time series is least accurate?
A) If the time-series variable is x, then xt = b0 + b1xt-1.
B) If the current value of the time series is above the mean reverting level, the prediction is that
the time series will decrease.
C) If the current value of the time series is above the mean reverting level, the prediction is that
the time series will increase.
Explanation
If the current value of the time series is above the mean reverting level, the prediction is that the time
series will decrease; if the current value of the time series is below the mean reverting level, the
prediction is that the time series will increase.
Consider the estimated model xt = −6.0 + 1.1 xt − 1 + 0.3 xt − 2 + εt that is estimated over 50 periods. The
value of the time series for the 49th observation is 20 and the value of the time series for the 50th
observation is 22. What is the forecast for the 52nd observation?
A) 24.2.
B) 42
C) 27.22.
Explanation
Using the chain-rule of forecasting,
Forecasted x51 = −6.0 + 1.1(22) + 0.3(20) = 24.2.
Forecasted x52 = −6.0 + 1.1(24.2) + 0.3(22) = 27.22.
The table below shows the autocorrelations of the lagged residuals for quarterly theater ticket sales that
were estimated using the AR(1) model: ln(salest) = b0 + b1(ln salest − 1) + et. Assuming the critical t-statistic
at 5% signi cance is 2.0, which of the following is the most likely conclusion about the appropriateness of
the model? The time series:
Lagged Autocorrelations of the Log of Quarterly Theater Ticket Sales
Lag Autocorrelation Standard Error t-Statistic
1 −0.0738 0.1667 −0.44271
2 −0.1047 0.1667 −0.62807
3 −0.0252 0.1667 −0.15117
4 0.5528 0.1667 3.31614
A) contains ARCH (1) errors.

B) contains seasonality.
C) would be more appropriately described with an MA(4) model.
Explanation
The time series contains seasonality as indicated by the strong and signi cant autocorrelation of the lag-
4 residual.
(Study Session 2, Module 6.4, LOS 6.l)
Which of the following statements regarding covariance stationarity is CORRECT?
A) A time series that is covariance stationary may have residuals whose mean changes over time.
B) The estimation results of an AR model involving a time series that is not covariance stationary
are meaningless.
C) A time series may be both covariance stationary and heteroskedastic.
Explanation
Covariance stationarity requires that the expected value and the variance of the time series be constant
over time.
(Study Session 2, Module 6, LOS 6.c)

A monthly time series of changes in maintenance expenses (ΔExp) for an equipment rental company was
t to an AR(1) model over 100 months. The results of the regression and the rst twelve lagged residual
autocorrelations are shown in the tables below. Based on the information in these tables, does the model
appear to be appropriately speci ed? (Assume a 5% level of signi cance.)
Regression Results for Maintenance Expense Changes
Model: DExpt = b0 + b1DExpt–1 + et
Coe cients Standard Error t-Statistic p-value
Intercept 1.3304 0.0089 112.2849 < 0.0001
Lag-1 0.1817 0.0061 30.0125 < 0.0001
Lagged Residual Autocorrelations for Maintenance Expense Changes
Lag Autocorrelation t-Statistic Lag Autocorrelation t-Statistic
1 −0.239 −2.39 7 −0.018 −0.18
2 −0.278 −2.78 8 −0.033 −0.33
3 −0.045 −0.45 9 0.261 2.61
4 −0.033 −0.33 10 −0.060 −0.60
5 −0.180 −1.80 11 0.212 2.12
6 −0.110 −1.10 12 0.022 0.22
A) No, because several of the residual autocorrelations are signi cant.
B) Yes, because the intercept and the lag coe cient are signi cant.
C) Yes, because most of the residual autocorrelations are negative.
Explanation
At a 5% level of signi cance, the critical t-value is 1.98. Since the absolute values of several of the
residual autocorrelation's t-statistics exceed 1.98, it can be concluded that signi cant serial correlation
exists and the model should be respeci ed. The next logical step is to estimate an AR(2) model, then test
the associated residuals for autocorrelation. If no serial correlation is detected, seasonality and ARCH
behavior should be tested.
(Study Session 2, Module 6.2, LOS 6.e)
Suppose you estimate the following model of residuals from an autoregressive model:
εt2 = 0.4 + 0.80εt-12 + µt, where ε = ε^
If the residual at time t is 2.0, the forecasted variance for time t+1 is:
A) 2.0.
B) 3.2.
C) 3.6.
Explanation
The variance at t=t+1 is 0.4 + [0.80 (4.0)] = 0.4 + 3.2. = 3.6.
Consider the following estimated model:
(Salest - Sales t-1) = 30 + 1.25 (Sales t-1 - Sales t-2) + 1.1 (Sales t-4 - Sales t-5) t=1,2,.. T
and Sales for the periods 1999.1 through 2000.2:
t Period Sales
T 2000.2 $2,000
T-1 2000.1 $1,800
T-2 1999.4 $1,500
T-3 1999.3 $1,400
T-4 1999.2 $1,900
T-5 1999.1 $1,700
The forecasted Sales amount for 2000.3 is closest to:
A) $2,270.00
B) $2,625.00
C) $1,730.00
Explanation
Note that since we are forecasting 2000.3, the numbering of the "t" column has changed.
Change in sales = $30 + 1.25 ($2,000-1,800) + 1.1 ($1,400-1,900)
Change in sales = $30 + 250 - 550 = -$270
Sales = $2,000 – 270 = $1,730
(Study Session 2, Module 6.5, LOS 6.n)
Bill Johnson, CFA, has prepared data concerning revenues from sales of winter clothing made by Polar
Corporation. This data is presented (in $ millions) in the following table:
Lagged Change Seasonal Lagged

Change In Sales
In Sales Change In Sales
Quarter Sales Y Y + (−1) Y + (−4)
2013.1 182
2013.2 74 −108
2013.3 78 4 −108
2013.4 242 164 4
2014.1 194 −48 164
2014.2 79 −115 −48 −108
2014.3 90 11 −115 4
2014.4 260 170 11 w
Question #14 - 19 of 112 Question ID: 1208569
The preceding table will be used by Johnson to forecast values using:
A) a log-linear trend model with a seasonal lag.
B) a serially correlated model with a seasonal lag.
C) an autoregressive model with a seasonal lag.
Explanation
Johnson will use the table to forecast values using an autoregressive model for periods in succession
since each successive forecast relies on the forecast for the preceding period. The seasonal lag is
introduced to account for seasonal variations in the observed data.
(Study Session 2, Module 6, LOS 6.k)
The value that Johnson should enter in the table in place of "w" is:
A) −115.
B) −48.
C) 164.
Explanation
The seasonal lagged change in sales shows the change in sales from the period 4 quarters before the
current period. Sales in the year 2013 quarter 4 increased $164 million over the prior period.

Imagine that Johnson prepares a change-in-sales regression analysis model with seasonality, which
includes the following:
Coe cients
Intercept −6.032
Lag 1 0.017
Lag 4 0.983
Based on the model, expected sales in the rst quarter of 2015 will be closest to:
A) 190.
B) 155.
C) 210.
Explanation
Substituting the 1-period lagged data from 2014.4 and the 4-period lagged data from 2014.1 into the
model formula, change in sales is predicted to be −6.032 + (0.017 × 170) + (0.983 × −48) = −50.326.
Expected sales are 260 + (−50.326) = 209.674.
Johnson's model was most likely designed to incorporates correction for:
A) heteroskedasticity of model residuals.
B) cointegration in the time series.
C) nonstationarity in time series data.
Explanation
Johnson's model transforms raw sales data by rst di erencing it and then modeling change in sales.
This is most likely an adjustment to make the data stationary for use in an AR model.
To test for covariance-stationarity in the data, Johnson would most likely use a:
A) Durbin-Watson test.
B) Dickey-Fuller test.
C) t-test.
Explanation
The Dickey-Fuller test for unit roots could be used to test whether the data is covariance non-
stationarity. The Durbin-Watson test is used for detecting serial correlation in the residuals of trend
models but cannot be used in AR models. A t-test is used to test for residual autocorrelation in AR
models.
The presence of conditional heteroskedasticity of residuals in Johnson's model is would most likely to lead
to:
A) invalid estimates of regression coe cients, but the standard errors will still be valid.
B) invalid standard errors of regression coe cients and invalid statistical tests.
C) invalid standard errors of regression coe cients, but statistical tests will still be valid.
Explanation
The presence of conditional heteroskedasticity may leads to incorrect estimates of standard errors of
regression coe cients and hence invalid tests of signi cance of the coe cients.
Winston Collier, CFA, has been asked by his supervisor to develop a model for predicting the warranty
expense incurred by Premier Snowplow Manufacturing Company in servicing its plows. Three years ago,
major design changes were made on newly manufactured plows in an e ort to reduce warranty expense.
Premier warrants its snowplows for 4 years or 18,000 miles, whichever comes rst. Warranty expense is
higher in winter months, but some of Premier's customers defer maintenance issues that are not essential
to keeping the machines functioning to spring or summer seasons. The data that Collier will analyze is in
the following table (in $ millions):
Seasonal Lagged
Change in Lagged Change
Warranty Change in
Quarter Warranty in Warranty
Expense Warranty
Expense yt Expense yt-1
Expense yt-4
2002.1 103
2002.2 52 -51
2002.3 32 -20 -51
2002.4 68 +36 -20
2003.1 91 +23 +36
2003.2 44 -47 +23 -51
2003.3 30 -14 -47 -20
2003.4 60 +30 -14 +36
2004.1 77 +17 +30 +23
2004.2 38 -39 +17 -47

2004.3 29 -9 -39 -14
2004.4 53 +24 -9 +30
Winston submits the following results to his supervisor. The rst is the estimation of a trend model for the
period 2002:1 to 2004:4. The model is below. The standard errors are in parentheses.
(Warranty expense)t = 74.1 - 2.7* t + et
R-squared = 16.2%
(14.37) (1.97)
Winston also submits the following results for an autoregressive model on the di erences in the expense
over the period 2004:2 to 2004:4. The model is below where "y" represents the change in expense as
de ned in the table above. The standard errors are in parentheses.
yt = -0.7 - 0.07* yt-1 + 0.83* yt-4 + et
R-squared = 99.98%
(0.643) (0.0222) (0.0186)
After receiving the output, Collier's supervisor asks him to compute moving averages of the sales data.
Collier's supervisors would probably not want to use the results from the trend model for all of the
following reasons EXCEPT:
A) the slope coe cient is not signi cant.
B) the model is a linear trend model and log-linear models are always superior.
C) it does not give insights into the underlying dynamics of the movement of the dependent
variable.
Explanation
Linear trend models are not always inferior to log-linear models. To determine which speci cation is
better would require more analysis such as a graph of the data over time. As for the other possible
answers, Collier can see that the slope coe cient is not signi cant because the t-statistic is
1.37=2.7/1.97. Also, regressing a variable on a simple time trend only describes the movement over
time, and does not address the underlying dynamics of the dependent variable.

For this question only, assume that Winston also ran an AR(1) model with the following results:
yt = −0.9 − 0.23* yt −1 + et
R-squared = 78.3%
(0.823) (0.0222)
The mean reverting level of this model is closest to:
A) 0.77.
B) −0.73.
C) 1.16.
Explanation
The mean reverting level is X1 = bo/(1 − b1)
X1 = −0.9/[1 − (−0.23)] = −0.73
Based upon the output provided by Collier to his supervisor and without any further calculations, in a
comparison of the two equations' explanatory power of warranty expense it can be concluded that:
A) the two equations are equally useful in explaining warranty expense.
B) the information provided is not su cient to determine which equation has greater
explanatory power.
C) the autoregressive model on the rst di erenced data has more explanatory power for
warranty expense.
Explanation
Although the R-squared values would suggest that the autoregressive model has more explanatory
power, there are a few problems. First, the models have di erent sample periods and di erent numbers
of explanatory variables. Second, the actual input data is di erent. To assess the explanatory power of
warranty expense, as opposed to the rst di erenced values, we must transform the tted values of the
rst-di erenced data back to the original level data to assess the explanatory power for the warranty
expense.
Based on the autoregressive model, expected warranty expense in the rst quarter of 2005 will be closest
to:
A) $60 million.
B) $51 million.
C) $65 million.
Explanation
Substituting the 1-period lagged data from 2004.4 and the 4-period lagged data from 2004.1 into the
model formula, change in warranty expense is predicted to be higher than 2004.4.
11.73 =-0.7 - 0.07*24 + 0.83*17.
The expected warranty expense is (53 + 11.73) = $64.73 million.
Based upon the results, is there a seasonality component in the data?
A) Yes, because the coe cient on yt is small compared to its standard error.
B) No, because the slope coe cients in the autoregressive model have opposite signs.
C) Yes, because the coe cient on yt-4 is large compared to its standard error.
Explanation
The coe cient on the 4th lag tests the seasonality component. The t-ratio is 44.6. Even using
Chebychev's inequality, this would be signi cant. Neither of the other answers are correct or relate to
the seasonality of the data.
Collier most likely chose to use rst-di erenced data in the autoregressive model:
A) because the time trend was signi cant.
B) in order to avoid problems associated with unit roots.
C) to increase the explanatory power.
Explanation
Time series with unit roots are very common in economic and nancial models, and unit roots cause
problems in assessing the model. Fortunately, a time series with a unit root may be transformed to
achieve covariance stationarity using the rst-di erencing process. Although the explanatory power of
the model was high (but note the small sample size), a model using rst-di erenced data often has less
explanatory power. The time trend was not signi cant, so that was not a possible answer.

Diem Le is analyzing the nancial statements of McDowell Manufacturing. He has modeled the time series
of McDowell's gross margin over the last 15 years. The output is shown below. Assume 5% signi cance
level for all statistical tests.
Autoregressive Model
Gross Margin – McDowell Manufacturing
Quarterly Data: 1st Quarter 1985 to 4th Quarter 2000
Regression Statistics
R-squared 0.767
Standard error of forecast 0.049
Observations 64
Durbin-Watson 1.923 (not statistically signi cant)
Coe cient Standard Error t-statistic
Constant 0.155 0.052 ?????
Lag 1 0.240 0.031 ?????
Lag 4 0.168 0.038 ?????
Autocorrelation of Residuals
Lag Autocorrelation Standard Error t-statistic
1 0.015 0.129 ?????
2 -0.101 0.129 ?????
3 -0.007 0.129 ?????
4 0.095 0.129 ?????
Partial List of Recent Observations
Quarter Observation
4th Quarter 2002 0.250
1st Quarter 2003 0.260
2nd Quarter 2003 0.220
3rd Quarter 2003 0.200
4th Quarter 2003 0.240
Abbreviated Table of the Student's t-distribution (One-Tailed Probabilities)
df p = 0.10 p = 0.05 p = 0.025 p = 0.01 p = 0.005
50 1.299 1.676 2.009 2.403 2.678
60 1.296 1.671 2.000 2.390 2.660
70 1.294 1.667 1.994 2.381 2.648

This model is best described as:
A) an ARMA(2) model.
B) an MA(2) model.
C) an AR(1) model with a seasonal lag.
Explanation
This is an autoregressive AR(1) model with a seasonal lag. Remember that an AR model regresses a
dependent variable against one or more lagged values of itself.
Which of the following can Le conclude from the regression? The time series process:
A) Does not include a seasonality factor and and has signi cant explanatory power.
B) Does not include a seasonality factor and has insigni cant explanatory power.
C) includes a seasonality factor, has signi cant explanatory power.
Explanation
The gross margin in the current quarter is related to the gross margin four quarters (one year) earlier.
To determine whether there is a seasonality factor, we need to test the coe cient on lag 4. The t-
statistic for the coe cients is calculated as the coe cient divided by the standard error with 61 degrees
of freedom (64 observations less three coe cient estimates). The critical t-value for a signi cance level
of 5% is about 2.000 (from the table). The computed t-statistic for lag 4 is 0.168/0.038 = 4.421. This is
greater than the critical value at even alpha = 0.005, so it is statistically signi cant. This suggests an
annual seasonal factor.
The process has signi cant explanatory power since both slope coe cients are signi cant and the
coe cient of determination is 0.767.
Le can conclude that the model is:
A) properly speci ed because the Durbin-Watson statistic is not signi cant.
B) properly speci ed because there is no evidence of autocorrelation in the residuals.
C) not properly speci ed because there is evidence of autocorrelation in the residuals and the
Durbin-Watson statistic is not signi cant.
Explanation
The Durbin-Watson test is not an appropriate test statistic in an AR model, so we cannot use it to test
for autocorrelation in the residuals. However, we can test whether each of the four lagged residuals
autocorrelations is statistically signi cant. The t-test to accomplish this is equal to the autocorrelation
divided by the standard error with 61 degrees of freedom (64 observations less 3 coe cient estimates).
The critical t-value for a signi cance level of 5% is about 2.000 from the table. The appropriate t-
statistics are:
Lag 1 = 0.015/0.129 = 0.116

Lag 2 = -0.101/0.129 = -0.783
Lag 3 = -0.007/0.129 = -0.054
Lag 4 = 0.095/0.129 = 0.736
None of these are statically signi cant, so we can conclude that there is no evidence of autocorrelation
in the residuals, and therefore the AR model is properly speci ed.
What is the 95% con dence interval for the gross margin in the rst quarter of 2004?
A) 0.168 to 0.240.
B) 0.158 to 0.354.
C) 0.197 to 0.305.
Explanation
The forecast for the following quarter is 0.155 + 0.240(0.240) + 0.168(0.260) = 0.256. Since the standard
error is 0.049 and the corresponding t-statistic is 2, we can be 95% con dent that the gross margin will
be within 0.256 – 2 × (0.049) and 0.256 + 2 × (0.049) or 0.158 to 0.354.
With respect to heteroskedasticity in the model, we can de nitively say:
A) an ARCH process exists because the autocorrelation coe cients of the residuals have
di erent signs.
B) nothing.
C) heteroskedasticity is not a problem because the DW statistic is not signi cant.
Explanation
None of the information in the problem provides information concerning heteroskedasticity. Note that
heteroskedasticity occurs when the variance of the error terms is not constant. When heteroskedasticity
is present in a time series, the residuals appear to come from di erent distributions (model seems to t
better in some time periods than others).

Using the provided information, the forecast for the 2nd quarter of 2004 is:
A) 0.253.
B) 0.192.
C) 0.250.
Explanation
To get the 2nd quarter forecast, we use the one period forecast for the 1st quarter of 2004, which is
0.155 + 0.240(0.240) + 0.168(0.260) = 0.256. The 4th lag for the 2nd quarter is 0.22. Thus the forecast for
the 2nd quarter is 0.155 + 0.240(0.256) + 0.168(0.220) = 0.253.
Albert Morris, CFA, is evaluating the results of an estimation of the number of wireless phone minutes
used on a quarterly basis within the territory of Car-tel International, Inc. Some of the information is
presented below (in billions of minutes):
Wireless Phone Minutes (WPM)t = bo + b1 WPMt-1 + εt
ANOVA Degrees of Freedom Sum of Squares Mean Square
Regression 1 7,212.641 7,212.641
Error 26 3,102.410 119.324
Total 27 10,315.051
Standard Error of the

Coe cients Coe cient
Coe cient
Intercept -8.0237 2.9023
WPMt-1 1.0926 0.0673
The variance of the residuals from one time period within the time series is not dependent on the variance
of the residuals in another time period.
Morris also models the monthly revenue of Car-tel using data over 96 monthly observations. The model is
shown below:
Sales (CAD$ millions) = b0 + b1 Salest−1 + εt
Standard Error of the

Coe cients Coe cient
Coe cient
Intercept 43.2 12.32
Salest−1 0.8867 0.4122

The value for WPM this period is 544 billion. Using the results of the model, the forecast Wireless Phone
Minutes three periods in the future is:
A) 683.18.
B) 691.30.
C) 586.35.
Explanation
The one-period forecast is −8.023 + (1.0926 × 544) = 586.35.
The two-period forecast is then −8.023 + (1.0926 × 586.35) = 632.62.
Finally, the three-period forecast is −8.023 + (1.0926 × 632.62) = 683.18.
(Study Session 2, Module 6.1, LOS 6.a)
The R-squared for the WPM model is closest to:
A) 70%.
B) 97%.
C) 33%.
Explanation
R-squared = SSR/SST = 7,212.641/10,315.051 = 70%.
The WPM model was speci ed as a(n):
A) Autoregressive (AR) Model.
B) Moving Average (MA) Model.
C) Autoregressive (AR) Model with a seasonal lag.
Explanation
The model is speci ed as an AR Model, but there is no seasonal lag. No moving averages are employed
in the estimation of the model.

Question ID: 1208497
Question #35 - 37 of 112
Based upon the information provided, Morris would most likely get more meaningful statistical results by:
A) rst di erencing the data.
B) adding more lags to the model.
C) doing nothing. No information provided suggests that any of these will improve the
speci cation.
Explanation
Since the slope coe cient is greater than one, the process may not be covariance stationary (we would
have to test this to be de nitive). A common technique to correct for this is to rst di erence the
variable to perform the following regression: Δ(WPM)t = bo + b1 Δ(WPM)t-1 + εt.
The mean reverting level of monthly sales is closest to:
A) 381.29 million.
B) 8.83 million.
C) 43.2 million.
Explanation
b0 43.2
MRL = = = 381.29 million
1 − b 1 1 − 0.8867
Morris concludes that the current price of Car-tel stock is consistent with single stage constant growth
model (with g=3%). Based on this information, the sales model is most likely:
A) Incorrectly speci ed and taking the natural log of the data would be an appropriate remedy.
B) Correctly speci ed.
C) Incorrectly speci ed and rst di erencing the data would be an appropriate remedy.
Explanation
If constant growth rate is an appropriate model for Car-tel, its dividends (as well as earnings and
revenues) will grow at a constant rate. In such a case, the time series needs to be adjusted by taking the
natural log of the time series. First di erencing would remove the trending component of a covariance
non-stationary time series but would not be appropriate for transforming an exponentially growing time
series.

Alexis Popov, CFA, is analyzing monthly data. Popov has estimated the model xt = b0 + b1 × xt-1 + b2 × xt-2 +
et. The researcher nds that the residuals have a signi cant ARCH process. The best solution to this is to:
A) re-estimate the model using only an AR(1) speci cation.
B) re-estimate the model using a seasonal lag.
C) re-estimate the model with generalized least squares.
Explanation
If the residuals have an ARCH process, then the correct remedy is generalized least squares which will
allow Popov to better interpret the results.
(Study Session 2, Module 6, LOS 6.o)
David Brice, CFA, has used an AR(1) model to forecast the next period's interest rate to be 0.08. The AR(1)
has a positive slope coe cient. If the interest rate is a mean reverting process with an unconditional
mean, a.k.a., mean reverting level, equal to 0.09, then which of the following could be his forecast for two
periods ahead?
A) 0.113.
B) 0.072.
C) 0.081.
Explanation
As Brice makes more distant forecasts, each forecast will be closer to the unconditional mean. So, the
two period forecast would be between 0.08 and 0.09, and 0.081 is the only possible answer.

William Zox, an analyst for Opal Mountain Capital Management, uses two di erent models to forecast
changes in the in ation rate in the United Kingdom. Both models were constructed using U.K. in ation
data from 1988-2002. In order to compare the forecasting accuracy of the models, Zox collected actual
U.K. in ation data from 2004-2005, and compared the actual data to what each model predicted. The rst
model is an AR(1) model that was found to have an average squared error of 10.429 over the 12 month
period. The second model is an AR(2) model that was found to have an average squared error of 11.642
over the 12 month period. Zox then computed the root mean squared error for each model to use as a
basis of comparison. Based on the results of his analysis, which model should Zox conclude is the most
accurate?
A) Model 1 because it has an RMSE of 3.23.
B) Model 2 because it has an RMSE of 3.41.
C) Model 1 because it has an RMSE of 5.21.
Explanation
The root mean squared error (RMSE) criterion is used to compare the accuracy of autoregressive models
in forecasting out-of-sample values. To determine which model will more accurately forecast future
values, we calculate the square root of the mean squared error. The model with the smallest RMSE is
the preferred model. The RMSE for Model 1 is √10.429 = 3.23, while the RMSE for Model 2 is √11.642 =
3.41. Since Model 1 has the lowest RMSE, that is the one Zox should conclude is the most accurate.
Alexis Popov, CFA, wants to estimate how sales have grown from one quarter to the next on average. The
most direct way for Popov to estimate this would be:
A) a linear trend model.
B) an AR(1) model.
C) an AR(1) model with a seasonal lag.
Explanation
If the goal is to simply estimate the dollar change from one period to the next, the most direct way is to
estimate xt = b0 + b1 × (Trend) + et, where Trend is simply 1, 2, 3, ....T. The model predicts a change by
the value b1 from one period to the next.
Clara Holmes, CFA, is attempting to model the importation of an herbal tea into the United States which
last year was $ 54 million. She gathers 24 years of annual data, which is in millions of in ation-adjusted
dollars.
She computes the following equation:
(Tea Imports)t = 3.8836 + 0.9288 × (Tea Imports)t − 1 + et
t-statistics (0.9328) (9.0025)

R2 = 0.7942
Adj. R2 = 0.7844
SE = 3.0892
N = 23
Holmes and her colleague, John Briars, CFA, discuss the implication of the model and how they might
improve it. Holmes is fairly satis ed with the results because, as she says "the model explains 78.44
percent of the variation in the dependent variable." Briars says the model actually explains more than
that.
Briars asks about the Durbin-Watson statistic. Holmes said that she did not compute it, so Briars reruns
the model and computes its value to be 2.1073. Briars says "now we know serial correlation is not a
problem." Holmes counters by saying "rerunning the model and computing the Durbin-Watson statistic
was unnecessary because serial correlation is never a problem in this type of time-series model."
Briars and Holmes decide to ask their company's statistician about the consequences of serial correlation.
Based on what Briars and Holmes tell the statistician, the statistician informs them that serial correlation
will only a ect the standard errors and the coe cients are still unbiased. The statistician suggests that
they employ the Hansen method, which corrects the standard errors for both serial correlation and
heteroskedasticity.
Given the information from the statistician, Briars and Holmes decide to use the estimated coe cients to
make some inferences. Holmes says the results do not look good for the future of tea imports because the
coe cient on (Tea Import)t − 1 is less than one. This means the process is mean reverting. Using the
coe cients in the output, says Holmes, "we know that whenever tea imports are higher than 41.810, the
next year they will tend to fall. Whenever the tea imports are less than 41.810, then they will tend to rise in
the following year." Briars agrees with the general assertion that the results suggest that imports will not
grow in the long run and tend to revert to a long-run mean, but he says the actual long-run mean is
54.545. Briars then computes the forecast of imports three years into the future.
With respect to the statements made by Holmes and Briars concerning serial correlation and the
importance of the Durbin-Watson statistic:
A) they were both incorrect.
B) Holmes was correct and Briars was incorrect.
C) Briars was correct and Holmes was incorrect.
Explanation
Briars was incorrect because the DW statistic is not appropriate for testing serial correlation in an
autoregressive model of this sort. Holmes was incorrect because serial correlation can certainly be a
problem in such a model. They need to analyze the residuals and compute autocorrelation coe cients
of the residuals to better determine if serial correlation is a problem.

With respect to the statement that the company's statistician made concerning the consequences of serial
correlation, assuming the company's statistician is competent, we would most likely deduce that Holmes
and Briars did not tell the statistician:
A) the model’s speci cation.
B) the value of the Durbin-Watson statistic.
C) the sample size.
Explanation
Serial correlation will bias the standard errors. It can also bias the coe cient estimates in an
autoregressive model of this type. Thus, Briars and Holmes probably did not tell the statistician the
model is an AR(1) speci cation.
The statistician's statement concerning the bene ts of the Hansen method is:
A) correct, because the Hansen method adjusts for problems associated with both serial
correlation and heteroskedasticity.
B) not correct, because the Hansen method only adjusts for problems associated with serial
correlation but not heteroskedasticity.
C) not correct, because the Hansen method only adjusts for problems associated with
heteroskedasticity but not serial correlation.
Explanation
The statistician is correct because the Hansen method adjusts for problems associated with both serial
correlation and heteroskedasticity.
Using the model's results, Briar's forecast for three years into the future is:
A) $54.108 million.
B) $54.543 million.
C) $47.151 million.
Explanation
Briars' forecasts for the next three years would be:
year one: 3.8836 + 0.9288 × 54 = 54.0388
year two: 3.8836 + 0.9288 × (54.0388) = 54.0748
year three: 3.8836 + 0.9288 × (54.0748) = 54.1083
With respect to the comments of Holmes and Briars concerning the mean reversion of the import data,
the long-run mean value that:
A) Briars computes is not correct, and his conclusion is probably not accurate.
B) Briars computes is correct.
C) Briars computes is not correct, but his conclusion is probably accurate.
Explanation
Briars has computed a value that would be correct if the results of the model were reliable. The long-run
mean would be 3.8836 / (1 − 0.9288)= 54.5450.
Given the nature of their analysis, the most likely potential problem that Briars and Holmes need to
investigate is:
A) autocorrelation.
B) unit root.
C) multicollinearity.
Explanation
Multicollinearity cannot be a problem because there is only one independent variable. For a time series
AR model, autocorrelation is a bigger worry. The model may have been misspeci ed leading to
statistically signi cant autocorrelations. Unit root does not seem to be a problem given the value of
b1<1.

The table below shows the autocorrelations of the lagged residuals for the rst di erences of the natural
logarithm of quarterly motorcycle sales that were t to the AR(1) model: (ln salest − ln salest − 1) = b0 +
b1(ln salest − 1 − ln salest − 2) + εt. The critical t-statistic at 5% signi cance is 2.0, which means that there is
signi cant autocorrelation for the lag-4 residual, indicating the presence of seasonality. Assuming the time
series is covariance stationary, which of the following models is most likely to CORRECT for this apparent
seasonality?
Lagged Autocorrelations of First Di erences in the Log of Motorcycle Sales
1 −0.0738 0.1667 −0.44271
2 −0.1047 0.1667 −0.62807
3 −0.0252 0.1667 −0.15117
4 0.5528 0.1667 3.31614
A) ln salest = b0 + b1(ln salest − 1) − b2(ln salest − 4) + εt.
B) (ln salest − ln salest − 4) = b0 + b1(ln salest − 1 − ln salest − 2) + εt.
C) (ln salest − ln salest − 1) = b0 + b1(ln salest − 1 − ln salest − 2) + b2(ln salest − 4 − ln salest − 5) + εt.
Explanation
Seasonality is taken into account in an autoregressive model by adding a seasonal lag variable that
corresponds to the seasonality. In the case of a rst-di erenced quarterly time series, the seasonal lag
variable is the rst di erence for the fourth time period. Recognizing that the model is t to the rst
di erences of the natural logarithm of the time series, the seasonal adjustment variable is (ln salest − 4 −
ln salest − 5).
Housing industry analyst Elaine Smith has been assigned the task of forecasting housing foreclosures.
Speci cally, Smith is asked to forecast the percentage of outstanding mortgages that will be foreclosed
upon in the coming quarter. Smith decides to employ multiple linear regression and time series analysis.
Besides constructing a forecast for the foreclosure percentage, Smith wants to address the following two
questions:
Research Is the foreclosure percentage signi cantly a ected by short-term

Question 1: interest rates?
Research Is the foreclosure percentage signi cantly a ected by

Question 2: government intervention policies?
Smith contends that adjustable rate mortgages often are used by higher risk borrowers and that their
homes are at higher risk of foreclosure. Therefore, Smith decides to use short-term interest rates as one
of the independent variables to test Research Question 1.
To measure the e ects of government intervention in Research Question 2, Smith uses a dummy variable
that equals 1 whenever the Federal government intervened with a scal policy stimulus package that
exceeded 2% of the annual Gross Domestic Product. Smith sets the dummy variable equal to 1 for four
quarters starting with the quarter in which the policy is enacted and extending through the following 3
quarters. Otherwise, the dummy variable equals zero.
Smith uses quarterly data over the past 5 years to derive her regression. Smith's regression equation is
provided in Exhibit 1:
Exhibit 1: Foreclosure Share Regression Equation
foreclosure share = b0 + b1(ΔINT) + b2(STIM) + b3(CRISIS) + ε
where:
Foreclosure = the percentage of all outstanding mortgages foreclosed upon during

share the quarter
= the quarterly change in the 1-year Treasury bill rate (e.g., ΔINT = 2 for
ΔINT
a two percentage point increase in interest rates)
STIM = 1 for quarters in which a Federal scal stimulus package was in place
= 1 for quarters in which the median house price is one standard

CRISIS
deviation below its 5-year moving average
The results of Smith's regression are provided in Exhibit 2:
Exhibit 2: Foreclosure Share Regression Results
Variable Coe cient t-statistic
Intercept 3.00 2.40
ΔINT 1.00 2.22
STIM -2.50 -2.10
CRISIS 4.00 2.35
The ANOVA results from Smith's regression are provided in Exhibit 3:
Exhibit 3: Foreclosure Share Regression Equation ANOVA Table
Degrees of
Source Sum of Squares Mean Sum of Squares
Freedom
Regression 3 15 5.0000
Error 16 5 0.3125
Total 19 20
Smith expresses the following concerns about the test statistics derived in her regression:
If my regression errors exhibit conditional heteroskedasticity, my t-

Concern 1:
statistics will be underestimated.
If my independent variables are correlated with each other, my F-

Concern 2:
statistic will be overestimated.
Before completing her analysis, Smith runs a regression of the changes in foreclosure share on its lagged
value. The following regression results and autocorrelations were derived using quarterly data over the
past 5 years ( Exhibit 4 and Exhibit 5, respectively):

Exhibit 4. Lagged Regression Results
Δ foreclosure sharet = 0.05 + 0.25(Δ foreclosure sharet– 1)
Exhibit 5. Autocorrelation Analysis
Lag Autocorrelation t-statistic
1 0.05 0.22
2 -0.35 -1.53
3 0.25 1.09
4 0.10 0.44
Exhibit 6 provides critical values for the Student's t-Distribution
Exhibit 6: Critical Values for Student's t-Distribution
Area in Both Tails Combined
Degrees of Freedom 20% 10% 5% 1%
16 1.337 1.746 2.120 2.921
17 1.333 1.740 2.110 2.898
18 1.330 1.734 2.101 2.878
19 1.328 1.729 2.093 2.861
20 1.325 1.725 2.086 2.845
Using a 1% signi cance level, which of the following is closest to the lower bound of the lower con dence
interval of the ΔINT slope coe cient?
A) −0.296
B) −0.045
C) −0.316
Explanation
The appropriate con dence interval associated with a 1% signi cance level is the 99% con dence level,
which equals;
slope coe cient ± critical t-statistic (1% signi cance level) × coe cient standard error
The standard error is not explicitly provided in this question, but it can be derived by knowing the
formula for the t-statistic:
COEF
t-statistic =
STD ERROR
From Exhibit 1, the ΔINT slope coe cient estimate equals 1.0, and its t-statistic equals 2.22. Therefore,
solving for the standard error, we derive:
COEF 1.00
STD ERROR for the ΔINT slope estimate = = = 0.450
t-statistic 2.22
The critical value for the 1% signi cance level is found down the 1% column in the t-tables provided in
Exhibit 6. The appropriate degrees of freedom for the con dence interval equals n − k − 1 = 20 − 3 − 1 =
16 (k is the number of slope estimates = 3). Therefore, the critical value for the 99% con dence interval
(or 1% signi cance level) equals 2.921.
So, the 99% con dence interval for the ΔINT slope coe cient is:
1.00 ± 2.921(0.450): lower bound equals 1 − 1.316 and upper bound 1 + 1.316
or (−0.316, 2.316).
Based on her regression results in Exhibit 2, using a 5% level of signi cance, Smith should conclude that:
A) both stimulus packages and housing crises have signi cant e ects on foreclosure
percentages.
B) stimulus packages have signi cant e ects on foreclosure percentages, but housing crises do
not have signi cant e ects on foreclosure percentages.
C) stimulus packages do not have signi cant e ects on foreclosure percentages, but housing
crises do have signi cant e ects on foreclosure percentages.
Explanation
The appropriate test statistic for tests of signi cance on individual slope coe cient estimates is the t-
statistic, which is provided in Exhibit 2 for each regression coe cient estimate. The reported t-statistic
equals -2.10 for the STIM slope estimate and equals 2.35 for the CRISIS slope estimate. The critical t-
statistic for the 5% signi cance level equals 2.12 (16 degrees of freedom, 5% level of signi cance).
Therefore, the slope estimate for STIM is not statistically signi cant (the reported t-statistic, -2.10, is not
large enough). In contrast, the slope estimate for CRISIS is statistically signi cant (the reported t-
statistic, 2.35, exceeds the 5% signi cance level critical value).

The standard error of estimate for Smith's regression is closest to:
A) 0.56
B) 0.16
C) 0.53
Explanation
The formula for the Standard Error of the Estimate (SEE) is:
−−−−−−−−−−− −−−−
 
 SSE  5
 
SEE = = = 0.56
⎷ ⎷
n - k - 1 16
The SEE equals the standard deviation of the regression residuals. A low SEE implies a high R2.
Is Smith correct or incorrect regarding Concerns 1 and 2?
A) Incorrect on both Concerns.
B) Correct on both Concerns.
C) Only correct on one concern and incorrect on the other.
Explanation
Smith's Concern 1 is incorrect. Heteroskedasticity is a violation of a regression assumption, and refers to

regression error variance that is not constant over all observations in the regression. Conditional
heteroskedasticity is a case in which the error variance is related to the magnitudes of the independent
variables (the error variance is "conditional" on the independent variables). The consequence of
conditional heteroskedasticity is that the standard errors will be too low, which, in turn, causes the t-
statistics to be too high. Smith's Concern 2 also is not correct. Multicollinearity refers to independent
variables that are correlated with each other. Multicollinearity causes standard errors for the regression
coe cients to be too high, which, in turn, causes the t-statistics to be too low. However, contrary to
Smith's concern, multicollinearity has no e ect on the F-statistic.
The most recent change in foreclosure share was +1 percent. Smith decides to base her analysis on the
data and methods provided in Exhibit 4 and Exhibit 5, and determines that the two-step ahead forecast for
the change in foreclosure share (in percent) is 0.125, and that the mean reverting value for the change in
foreclosure share (in percent) is 0.071. Is Smith correct?
A) Smith is correct on the mean-reverting level for forecast of change in foreclosure share only.
B) Smith is correct on both the forecast and the mean reverting level.
C) Smith is correct on the two-step ahead forecast for change in foreclosure share only.
Explanation
Forecasts are derived by substituting the appropriate value for the period t-1 lagged value.
ΔForeclosure Sharet = 0.05 + 0.25(ΔForeclosure Sharet-1)
= 0.05 + 0.25(1) = 0.30
So, the one-step ahead forecast equals 0.30%. The two-step ahead (%) forecast is derived by substituting
0.30 into the equation.
ΔForeclosure Sharet+1 = 0.05 + 0.25(0.30) = 0.125
Therefore, the two-step ahead forecast equals 0.125%.
b0 0.05
mean reverting level = = = 0.067
(1 - b ) (1 - 0.25)
1
Assume for this question that Smith nds that the foreclosure share series has a unit root. Under these
conditions, she can most reliably regress foreclosure share against the change in interest rates (ΔINT) if:
A) ΔINT does not have unit root.
B) ΔINT has unit root and is not cointegrated with foreclosure share.
C) ΔINT has unit root and is cointegrated with foreclosure share.
Explanation
The error terms in the regressions for choices A, B, and C will be nonstationary. Therefore, some of the
regression assumptions will be violated and the regression results are unreliable. If, however, both
series are nonstationary (which will happen if each has unit root), but cointegrated, then the error term
will be covariance stationary and the regression results are reliable.
Which of the following statements regarding seasonality is least accurate?
A) Not correcting for seasonality when, in fact, seasonality exists in the time series results in a
violation of an assumption of linear regression.
B) The presence of seasonality makes it impossible to forecast using a time-series model.
C) A time series that is rst di erenced can be adjusted for seasonality by incorporating the rst-
di erenced value for the previous year's corresponding period.
Explanation
The goal of a time series model is to identify factors that can be predicted. Seasonality in a time series
refers to patterns that repeat at regular intervals. When a time series exhibits seasonality, seasonal lags
should be included in the model in order to increase its predictive ability.
To qualify as a covariance stationary process, which of the following does not have to be true?
A) Covariance(xt, xt-1) = Covariance(xt, xt-2).
B) E[xt] = E[xt+1].
C) Covariance(xt, xt-2) = Covariance(xt, xt+2).
Explanation
If a series is covariance stationary then the unconditional mean is constant across periods. The
unconditional mean or expected value is the same from period to period: E[xt] = E[xt+1]. The covariance
between any two observations equal distance apart will be equal, e.g., the t and t-2 observations with
the t and t+2 observations. The one relationship that does not have to be true is the covariance between
the t and t-1 observations equaling that of the t and t-2 observations.
Suppose that the time series designated as Y is mean reverting. If Yt+1 = 0.2 + 0.6 Yt, the best prediction of
Yt+1 is:
A) 0.8.
B) 0.3.
C) 0.5.
Explanation
The prediction is Yt+1 = b0 / (1-b1) = 0.2 / (1-0.6) = 0.5
Yolanda Seerveld is an analyst studying the growth of sales of a new restaurant chain called Very Vegan.
The increase in the public's awareness of healthful eating habits has had a very positive e ect on Very
Vegan's business. Seerveld has gathered quarterly data for the restaurant's sales for the past three years.
Over the twelve periods, sales grew from $17.2 million in the rst quarter to $106.3 million in the last
quarter. Because Very Vegan has experienced growth of more than 500% over the three years, the
Seerveld suspects an exponential growth model may be more appropriate than a simple linear trend
model. However, she begins by estimating the simple linear trend model:
(sales)t = α + β × (Trend)t + εt
Where the Trend is 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12.
Multiple R 0.952640
R2 0.907523
Adjusted R2 0.898275
Standard Error 8.135514
Observations 12
1st order autocorrelation coe cient of the residuals: −0.075
ANOVA
df SS
Regression 1 6495.203
Residual 10 661.8659
Total 11 7157.069
Coe cients Standard Error
Intercept 10.0015 5.0071
Trend 6.7400 0.6803
The analyst then estimates the following model:
(natural logarithm of sales)t = α + β × (Trend)t + εt
Multiple R 0.952028
R2 0.906357
Adjusted R2 0.896992
Standard Error 0.166686
Observations 12
1st order autocorrelation coe cient of the residuals: −0.348
ANOVA
df SS
Regression 1 2.6892
Residual 10 0.2778
Total 11 2.9670
Coe cients Standard Error
Intercept 2.9803 0.1026
Trend 0.1371 0.0140
Seerveld compares the results based upon the output statistics and conducts two-tailed tests at a 5% level
of signi cance. One concern is the possible problem of autocorrelation, and Seerveld makes an
assessment based upon the rst-order autocorrelation coe cient of the residuals that is listed in each set
of output. Another concern is the stationarity of the data. Finally, the analyst composes a forecast based
on each equation for the quarter following the end of the sample.
Are either of the slope coe cients statistically signi cant?
A) The simple trend regression is, but not the log-linear trend regression.
B) Yes, both are signi cant.
C) The simple trend regression is not, but the log-linear trend regression is.
Explanation
The respective t-statistics are 6.7400 / 0.6803 = 9.9074 and 0.1371 / 0.0140 = 9.7929. For 10 degrees of
freedom, the critical t-value for a two-tailed test at a 5% level of signi cance is 2.228, so both slope
coe cients are statistically signi cant.
Based upon the output, which equation explains the cause for variation of Very Vegan's sales over the
sample period?
A) The cause cannot be determined using the given information.
B) Both the simple linear trend and the log-linear trend have equal explanatory power.
C) The simple linear trend.
Explanation
To actually determine the explanatory power for sales itself, tted values for the log-linear trend would
have to be determined and then compared to the original data. The given information does not allow
for such a comparison.

With respect to the possible problems of autocorrelation and nonstationarity, using the log-linear
transformation appears to have:
A) improved the results for nonstationarity but not autocorrelation.
B) improved the results for autocorrelation but not nonstationarity.
C) not improved the results for either possible problems.
Explanation
The fact that there is a signi cant trend for both equations indicates that the data is not stationary in
either case. As for autocorrelation, the analyst really cannot test it using the Durbin-Watson test
because there are fewer than 15 observations, which is the lower limit of the DW table. Looking at the
rst-order autocorrelation coe cient, however, we see that it increased (in absolute value terms) for the
log-linear equation. If anything, therefore, the problem became more severe.
The primary limitation of both models is that:
A) regression is not appropriate for estimating the relationship.
B) the results are di cult to interpret.
C) each uses only one explanatory variable.
Explanation
The main problem with a trend model is that it uses only one variable so the underlying dynamics are
really not adequately addressed. A strength of the models is that the results are easy to interpret. The
levels of many economic variables such as the sales of a rm, prices, and gross domestic product (GDP)
have a signi cant time trend, and a regression is an appropriate tool for measuring that trend.
Using the simple linear trend model, the forecast of sales for Very Vegan for the rst out-of-sample period
is:
A) $97.6 million.
B) $123.0 million.
C) $113.0 million.
Explanation
The forecast is 10.0015 + (13 × 6.7400) = 97.62.

Using the log-linear trend model, the forecast of sales for Very Vegan for the rst out-of-sample period is:
A) $117.0 million.
B) $109.4 million.
C) $121.2 million.
Explanation
The forecast is e2.9803 + (13 × 0.1371) = 117.01.
The model xt = b0 + b1 xt-1 + b2 xt-2 + b3 xt-3 + b4 xt-4 + εt is:
A) a moving average model, MA(4).
B) an autoregressive model, AR(4).
C) an autoregressive conditional heteroskedastic model, ARCH.
Explanation
This is an autoregressive model (i.e., lagged dependent variable as independent variables) of order p=4
(that is, 4 lags).
Marvin Greene is interested in modeling the sales of the retail industry. He collected data on aggregate
sales and found the following:
Salest = 0.345 + 1.0 Salest-1
The standard error of the slope coe cient is 0.15, and the number of observations is 60. Given a level of
signi cance of 5%, which of the following can we NOT conclude about this model?
A) The model has a unit root.
B) The slope on lagged sales is not signi cantly di erent from one.
C) The model is covariance stationary.
Explanation
The test of whether the slope is di erent from one indicates failure to reject the null H0: b1=1 (t-critical
with df = 58 is approximately 2.000, t-calculated = (1.0 - 1.0)/0.15 = 0.0). This is a 2-tailed test and we
cannot reject the null since 0.0 is not greater than 2.000. This model is nonstationary because the 1.0
coe cient on Salest-1 is a unit root. Any time series that has a unit root is not covariance stationary
which can be corrected through the rst-di erencing process.
The main reason why nancial and time series intrinsically exhibit some form of nonstationarity is that:
A) serial correlation, a contributing factor to nonstationarity, is always present to a certain

degree in most nancial and time series.
B) most nancial and time series have a natural tendency to revert toward their means.
C) most nancial and economic relationships are dynamic and the estimated regression
coe cients can vary greatly between periods.
Explanation
Because all nancial and time series relationships are dynamic, regression coe cients can vary widely
from period to period. Therefore, nancial and time series will always exhibit some amount of instability
or nonstationarity.
(Study Session 2, Module 6, LOS 6.h)
The model xt = b0 + b1 xt − 1 + b2 xt − 2 + b3 xt −12 + εt is an autoregressive model of type:
A) AR(12).
B) AR(1).
C) AR(2).
Explanation
The b1xt − 1 and b2xt − 2 lag terms make this an autoregressive model of order p = 2 with a seasonal lag.
The b3xt −12 term is a seasonal term which does not transform the model to AR(12).
Alexis Popov, CFA, has estimated the following speci cation: xt = b0 + b1 × xt-1 + et. Which of the following
would most likely lead Popov to want to change the model's speci cation?
A) b0 < 0.
B) Correlation(et, et-2) is signi cantly di erent from zero.
C) Correlation(et, et-1) is not signi cantly di erent from zero.
Explanation
If correlation(et, et-2) is not zero, then the model su ers from 2nd order serial correlation. Popov may
wish to try an AR(2) model. Both of the other conditions are acceptable in an AR(1) model.
Which of the following is NOT a requirement for a series to be covariance stationary? The:
A) covariance of the time series with itself (lead or lag) must be constant.
B) time series must have a positive trend.
C) expected value of the time series is constant over time.
Explanation
For a time series to be covariance stationary: 1) the series must have an expected value that is constant
and nite in all periods, 2) the series must have a variance that is constant and nite in all periods, and
3) the covariance of the time series with itself for a xed number of periods in the past or future must
be constant and nite in all periods.
Trend models can be useful tools in the evaluation of a time series of data. However, there are limitations
to their usage. Trend models are not appropriate when which of the following violations of the linear
regression assumptions is present?
A) Serial correlation.
B) Model misspeci cation.
C) Heteroskedasticity.
Explanation
One of the primary assumptions of linear regression is that the residual terms are not correlated with
each other. If serial correlation, also called autocorrelation, is present, then trend models are not an
appropriate analysis tool.

Consider the estimated model xt = -6.0 + 1.1 xt-1 + 0.3 xt-2 + εt that is estimated over 50 periods. The value
of the time series for the 49th observation is 20 and the value of the time series for the 50th observation is
22. What is the forecast for the 51st observation?
A) 24.2.
B) 23
C) 30.2.
Explanation
Forecasted x51 = -6.0 + 1.1 (22) + 0.3 (20) = 24.2.
Troy Dillard, CFA, has estimated the following equation using quarterly data: xt = 93 - 0.5×xt– 1 + 0.1×xt– 4 +
et. Given the data in the table below, what is Dillard's best estimate of the rst quarter of 2007?
Time Value
2005: I 62
2005: II 62
2005: III 66
2005: IV 66
2006: I 72
2006: II 70
2006: III 64
2006: IV 66
A) 66.40.
B) 66.60.
C) 67.20.
Explanation
To get the answer, Dillard will use the data for 2006: IV and 2006: I, xt– 1 = 66 and xt– 4 = 72 respectively:
E[x2007:I] = 93– 0.5×xt– 1 + 0.1×xt– 4
E[x2007:I] = 93– 0.5×66 + 0.1×72
E[x2007:I] = 67.20

Suppose that the following time-series model is found to have a unit root:
Salest = b0 + b1 Sales t-1+ εt
What is the speci cation of the model if rst di erences are used?
A) (Salest - Salest-1)= b0 + b1 (Sales t-1 - Sales t-2) + εt.
B) Salest = b0 + b1 Sales t-1 + b2 Sales t-2 + εt.
C) Salest = b1 Sales t-1+ εt.
Explanation
Estimation with rst di erences requires calculating the change in the variable from period to period.
(Study Session 2, Module 6, LOS 6.j)
Jason Cranfell, CFA, has hypothesised that sales of luxury cars have grown at a constant rate over the past
15 years.
Which of the following models is most appropriate for modelling these data?
A) ln(LucCarSales) = b0 + b1(t) + et
B) LuxCarSales = b0 + b1(t) + et
C) LuxCarSalest = b0 + b1LuxCarSales(t-1) + et
Explanation
Whenever the rate of change is constant over time, the appropriate model is a log-linear trend model.
The other two choices are a linear trend model and an autoregressive model.

After discussing the above matter with a colleague, Cranwell nally decides to use an autoregressive
model of order one i.e. AR(1) for the above data. Below is a summary of the ndings of the model:
b0 0.4563
b1 0.6874
Standard error 0.3745
R-squared 0.7548
Durbin Watson 1.23
F 12.63
Observations 180
Calculate the mean reverting level of the series:
A) 1.46
B) 1.66
C) 1.26
Explanation
The formula for the mean reverting level is b0/(1-b1) = 0.4563/(1-0.6874)=1.46
Cranwell is aware that the Dickey Fuller test can be used to discover whether a model has a unit root. He is
also aware that the test would use a revised set of critical t-values. What would it mean to Bert to reject
the null of the Dickey Fuller test (Ho: g = 0) ?
A) There is no unit root
B) There is a unit root but the model can be used if covariance-stationary
C) There is a unit root and the model cannot be used in its current form
Explanation
The null hypothesis of g = 0 actually means that b1 – 1 = 0 , meaning that b1 = 1. Since we have rejected
the null, we can conclude that the model has no unit root.
Cranwell would also like to test for serial correlation in his AR(1) model. To do this, Cranwell should:
A) use the provided Durbin Watson statistic and compare it to a critical value.
B) determine if the series has a nite and constant covariance between leading and lagged terms
of itself.
C) use a t-test on the residual autocorrelations over several lags.
Explanation
To test for serial correlation in an AR model, test for the signi cance of residual autocorrelations over
di erent lags. The goal is for all t-statistics to lack statistical signi cance. The Durbin-Watson test is used
with trend models; it is not appropriate for testing for serial correlation of the error terms in an
autoregressive model. Constant and nite unconditional variance is not an indicator of serial correlation
but rather is one of the requirements of covariance stationarity.
When using the root mean squared error (RMSE) criterion to evaluate the predictive power of the model,
which of the following is the most appropriate statement?
A) Use the model with the lowest RMSE calculated using the in-sample data.
B) Use the model with the lowest RMSE calculated using the out-of-sample data.
C) Use the model with the highest RMSE calculated using the in-sample data.
Explanation
RMSE is a measure of error hence the lower the better. It should be calculated on the out-of-sample
data i.e. the data not directly used in the development of the model. This measure thus indicates the
predictive power of our model.
If Cranwell suspects that seasonality may be present in his AR model, he would most correctly:
A) test for the signi cance of the slope coe cients.
B) examine the t-statistics of the residual lag autocorrelations.
C) use the Durbin Watson statistic.
Explanation
Seasonality in monthly and quarterly data is apparent in the high (statistically signi cant) t-statistics of
the residual lag autocorrelations for Lag 12 and Lag 4 respectively. To correct for that, the analyst
should incorporate the appropriate lag in his/her AR model.

Modeling the trend in a time series of a variable that grows at a constant rate with continuous
compounding is best done with:
A) simple linear regression.
B) a moving average model.
C) a log-linear transformation of the time series.
Explanation
The log-linear transformation of a series that grows at a constant rate with continuous compounding
(exponential growth) will cause the transformed series to be linear.
Rhonda Wilson, CFA, is analyzing sales data for the TUV Corp, a current equity holding in her portfolio. She
observes that sales for TUV Corp. have grown at a steadily increasing rate over the past ten years due to
the successful introduction of some new products. Wilson anticipates that TUV will continue this pattern of
success. Which of the following models is most appropriate in her analysis of sales for TUV Corp?
A) A log-linear trend model, because the data series can be graphed using a straight, upward-
sloping line.
B) A linear trend model, because the data series is equally distributed above and below the line
and the mean is constant.
C) A log-linear trend model, because the data series exhibits a predictable, exponential growth
trend.
Explanation
The log-linear trend model is the preferred method for a data series that exhibits a trend or for which
the residuals are predictable. In this example, sales grew at an exponential, or increasing rate, rather
than a steady rate.
Which of the following statements regarding the instability of time-series models is most accurate? Models
estimated with:
A) a greater number of independent variables are usually more stable than those with a smaller
number.
B) shorter time series are usually more stable than those with longer time series.
C) longer time series are usually more stable than those with shorter time series.
Explanation
Those models with a shorter time series are usually more stable because there is less opportunity for
variance in the estimated regression coe cients between the di erent time periods.
Which of the following is a seasonally adjusted model?
A) (Salest - Sales t-1)= b0 + b1 (Sales t-1 - Sales t-2) + b2 (Sales t-4 - Sales t-5) + εt.
B) Salest = b0 + b1 Sales t-1 + b2 Sales t-2 + εt.
C) Salest = b1 Sales t-1+ εt.
Explanation
This model is a seasonal AR with rst di erencing.
David Brice, CFA, has tried to use an AR(1) model to predict a given exchange rate. Brice has concluded the
exchange rate follows a random walk without a drift. The current value of the exchange rate is 2.2. Under
these conditions, which of the following would be least likely?
A) The residuals of the forecasting model are autocorrelated.
B) The forecast for next period is 2.2.
C) The process is not covariance stationary.
Explanation
The one-period forecast of a random walk model without drift is E(xt+1) = E(xt + et ) = xt + 0, so the
forecast is simply xt = 2.2. For a random walk process, the variance changes with the value of the
observation. However, the error term et = xt - xt-1 is not autocorrelated.
An AR(1) autoregressive time series model:

A) cannot be used to test for a unit root.
B) can be used to test for a unit root, which exists if the slope coe cient is less than one.
C) can be used to test for a unit root, which exists if the slope coe cient equals one.
Explanation
If you estimate the following model xt = b0 + b1 × xt-1 + et and get b1 = 1, then the process has a unit
root and is nonstationary.
An analyst modeled the time series of annual earnings per share in the specialty department store
industry as an AR(3) process. Upon examination of the residuals from this model, she found that there is a
signi cant autocorrelation for the residuals of this model. This indicates that she needs to:
A) switch models to a moving average model.
B) revise the model to include at least another lag of the dependent variable.
C) alter the model to an ARCH model.
Explanation
She should estimate an AR(4) model, and then re-examine the autocorrelations of the residuals.
Given an AR(1) process represented by xt+1 = b0 + b1×xt + et, the process would not be a random walk if:
A) b1 = 1.
B) the long run mean is b0 / (1-b1).
C) E(et)=0.
Explanation
For a random walk, the long-run mean is unde ned. The slope coe cient is one, b1=1, and that is what
makes the long-run mean unde ned: mean = b0/(1-b1).

Frank Batchelder and Miriam Yenkin are analysts for Bishop Econometrics. Batchelder and Yenkin are
discussing the models they use to forecast changes in China's GDP and how they can compare the
forecasting accuracy of each model. Batchelder states, "The root mean squared error (RMSE) criterion is
typically used to evaluate the in-sample forecast accuracy of autoregressive models." Yenkin replies, "If we
use the RMSE criterion, the model with the largest RMSE is the one we should judge as the most accurate."
With regard to their statements about using the RMSE criterion:
A) Batchelder is incorrect; Yenkin is correct.
B) Batchelder is correct; Yenkin is incorrect.
C) Batchelder is incorrect; Yenkin is incorrect.
Explanation
The root mean squared error (RMSE) criterion is used to compare the accuracy of autoregressive models
in forecasting out-of-sample values (not in-sample values). Batchelder is incorrect. Out-of-sample
forecast accuracy is important because the future is always out of sample, and therefore out-of-sample
performance of a model is critical for evaluating real world performance.
Yenkin is also incorrect. The RMSE criterion takes the square root of the average squared errors from
each model. The model with the smallest RMSE is judged the most accurate.
The regression results from tting an AR(1) model to the rst-di erences in enrollment growth rates at a
large university includes a Durbin-Watson statistic of 1.58. The number of quarterly observations in the
time series is 60. At 5% signi cance, the critical values for the Durbin-Watson statistic are dl = 1.55 and du
= 1.62. Which of the following is the most accurate interpretation of the DW statistic for the model?
A) Since DW > dl, the null hypothesis of no serial correlation is rejected.
B) Since dl < DW < du, the results of the DW test are inconclusive.
C) The Durbin-Watson statistic cannot be used with AR(1) models.
Explanation
The Durbin-Watson statistic is not useful when testing for serial correlation in an autoregressive model
where one of the independent variables is a lagged value of the dependent variable. The existence of
serial correlation in an AR model is determined by examining the autocorrelations of the residuals.

David Wellington, CFA, has estimated the following log-linear trend model: LN(xt) = b0 + b1t + εt. Using six
years of quarterly observations, 2001:I to 2006:IV, Wellington gets the following estimated equation: LN(xt)
= 1.4 + 0.02t. The rst out-of-sample forecast of xt for 2007:I is closest to:
A) 6.69.
B) 1.88.
C) 4.14.
Explanation
Wellington's out-of-sample forecast of LN(xt) is 1.9 = 1.4 + 0.02 × 25, and e1.9 = 6.69. (Six years of
quarterly observations, at 4 per year, takes us up to t = 24. The rst time period after that is t = 25.)
Which of the following statements regarding unit roots in a time series is least accurate?
A) Even if a time series has a unit root, the predictions from the estimated model are valid.
B) A time series that is a random walk has a unit root.
C) A time series with a unit root is not covariance stationary.
Explanation
The presence of a unit root means that the least squares regression procedure that we have been using
to estimate an AR(1) model cannot be used without transforming the data rst.
A time series with a unit root will follow a random walk process. Since a time series that follows a
random walk is not covariance stationary, modeling such a time series in an AR model can lead to
incorrect statistical conclusions, and decisions made on the basis of these conclusions may be wrong.
Unit roots are most likely to occur in time series that trend over time or have a seasonal element.
One choice a researcher can use to test for nonstationarity is to use a:
A) Dickey-Fuller test, which uses a modi ed t-statistic.
B) Breusch-Pagan test, which uses a modi ed t-statistic.
C) Dickey-Fuller test, which uses a modi ed χ2 statistic.
Explanation
The Dickey-Fuller test estimates the equation (xt – xt-1) = b0 + (b1 - 1) * xt-1 + et and tests if H0: (b1 – 1) =
0. Using a modi ed t-test, if it is found that (b1– 1) is not signi cantly di erent from zero, then it is
concluded that b1 must be equal to 1.0 and the series has a unit root.
The primary concern when deciding upon a time series sample period is which of the following factors?
A) The length of the sample time period.
B) The total number of observations.
C) Current underlying economic and market conditions.
Explanation
There will always be a tradeo between the increase statistical reliability of a longer time period and the
increased stability of estimated regression coe cients with shorter time periods. Therefore, the
underlying economic environment should be the deciding factor when selecting a time series sample
period.
A time series that has a unit root can be transformed into a time series without a unit root through:
A) calculating moving average of the residuals.
B) mean reversion.
C) rst di erencing.
Explanation
First di erencing a series that has a unit root creates a time series that does not have a unit root.
Barry Phillips, CFA, has estimated an AR(1) relationship (xt = b0 + b1 × xt-1 + et) and got the following result:
xt+1 = 0.5 + 1.0xt + et. Phillips should:
A) not rst di erence the data because b1 − b0 = 1.0 − 0.5 = 0.5 < 1.
B) rst di erence the data because b1 = 1.

C) not rst di erence the data because b0 = 0.5 < 1.
Explanation
The condition b1 = 1 means that the series has a unit root and is not stationary. The correct way to
transform the data in such an instance is to rst di erence the data.
Which of the following is least likely a consequence of a model containing ARCH(1) errors? The:
A) model's speci cation can be corrected by adding an additional lag variable.
B) regression parameters will be incorrect.
C) variance of the errors can be predicted.
Explanation
The presence of autoregressive conditional heteroskedasticity (ARCH) indicates that the variance of the
error terms is not constant. This is a violation of the regression assumptions upon which time series
models are based. The addition of another lag variable to a model is not a means for correcting for
ARCH (1) errors.
Vikas Rathod, an enrolled candidate for the CFA Level II examination, has decided to perform a calendar
test to examine whether there is any abnormal return associated with investments and disinvestments
made in blue-chip stocks on particular days of the week. As a proxy for blue-chips, he has decided to use
the S&P 500 index. The analysis will involve the use of dummy variables and is based on the past 780
trading days. Here are selected ndings of his study:
RSS 0.0039
SSE 0.9534
SST 0.9573
R-squared 0.004
SEE 0.035
Jessica Jones, CFA, a friend of Rathod, overhears that he is interested in regression analysis and warns him
that whenever heteroskedasticity is present in multiple regression this could undermine the regression
results. She mentions that one easy way to spot conditional heteroskedasticity is through a scatter plot,
but she adds that there is a more formal test. Unfortunately, she can't quite remember its name. Jessica
believes that heteroskedasticity can be recti ed using White-corrected standard errors. Her son Jonathan
who has also taken part in the discussion, hears this comment and argues that White correction would
typically reduce the number of Type I errors in nancial data.

How many dummy variables should Rathod use?
A) Four
B) Six
C) Five
Explanation
There are 5 trading days in a week, but we should use (n − 1) or 4 dummies in order to ensure no
violations of regression analysis occur.
What is most likely represented by the intercept of the regression?
A) The return on a particular trading day.
B) The intercept is not a driver of returns, only the independent variables.
C) The drift of a random walk.
Explanation
The omitted variable is represented by the intercept. So, if we have four variables to represent Monday
through Thursday, the intercept would represent returns on Friday.
What can be said of the overall explanatory power of the model at the 5% signi cance?
A) There is no value to calendar trading.
B) There is value to calendar trading.
C) The coe cient of determination for the above regression is signi cantly higher than the
standard error of the estimate, and therefore there is value to calendar trading.
Explanation
This question calls for a computation of the F-stat. F = (0.0039/4)/(0.9534/(780−4−1) = 0.79. The critical F
is somewhere between 2.37 and 2.45 so we fail to reject the Null that all the coe cients are equal to
zero.

The test mentioned by Jessica is known as the:
A) Durbin-Watson, which is a two-tailed test
B) Breusch-Pagan, which is a one-tailed test
C) Breusch-Pagan, which is a two-tailed test
Explanation
The Breusch-Pagan is used to detect conditional heteroskedasticity and it is a one-tailed test. This is
because we are only concerned about large values in the residuals coe cient of determination.
Are Jessica and her son Jonathan, correct in terms of the method used to correct for heteroskedasticity
and the likely e ects?
A) Neither is correct
B) One is correct
C) Both are correct
Explanation
Jessica is correct. White-corrected standard errors are also known as robust standard errors. Jonathan is
correct because White-corrected errors are higher than the biased errors leading to lower computed t-
statistics and therefore less frequent rejection of the Null Hypothesis (remember incorrectly rejecting a
true Null is Type I error).
Assuming the a1 term of an ARCH(1) model is signi cant, the following can be forecast:
A) The square of the error term.
B) A signi cant a1 implies that the ARCH framework cannot be used.
C) The variance of the error term.
Explanation
A Model is ARCH(1) if the coe cient a1 is signi cant. It will allow for the estimation of the variance of the
error term.

Barry Phillips, CFA, is analyzing quarterly data. He has estimated an AR(1) relationship (xt = b0 + b1 × xt-1 +
et) and wants to test for seasonality. To do this he would want to see if which of the following statistics is
signi cantly di erent from zero?
A) Correlation(et, et-1).
B) Correlation(et, et-5).
C) Correlation(et, et-4).
Explanation
Although seasonality can make the other correlations signi cant, the focus should be on correlation(et,
et-4) because the 4th lag is the value that corresponds to the same season as the predicted variable in
the analysis of quarterly data.
Consider the following estimated model:
(Salest - Sales t-1)= 100 - 1.5 (Sales t-1 - Sales t-2) + 1.2 (Sales t-4 - Sales t-5) t=1,2,.. T
and Sales for the periods 1999.1 through 2000.2:
t Period Sales
T 2000.2 $1,000
T-1 2000.1 $900
T-2 1999.4 $1,200
T-3 1999.3 $1,400
T-4 1999.2 $1,000
T-5 1999.1 $800
The forecasted Sales amount for 2000.3 is closest to:
A) $1,730.00
B) $730.00
C) $1,430.00
Explanation
Change in sales = $100 - 1.5 ($1,000-900) + 1.2 ($1,400-1,000)
Change in sales = $100 - 150 + 480 =$430
Sales = $1,000 + 430 = $1,430
The data below yields the following AR(1) speci cation: xt = 0.9 – 0.55xt-1 + Et , and the indicated tted
values and residuals.
Time xt tted values residuals
1 1 - -
2 -1 0.35 -1.35
3 2 1.45 0.55
4 -1 -0.2 -0.8
5 0 1.45 -1.45
6 2 0.9 1.1
7 0 -0.2 0.2
8 1 0.9 0.1
9 2 0.35 1.65
The following sets of data are ordered from earliest to latest. To test for ARCH, the researcher should
regress:
A) (-1.35, 0.55, -0.8, -1.45, 1.1, 0.2, 0.1, 1.65) on (0.35, 1.45, -0.2, 1.45, 0.9, -0.2, 0.9, 0.35)
B) (1, 4, 1, 0, 4, 0, 1, 4) on (1, 1, 4, 1, 0, 4, 0, 1)
C) (1.8225, 0.3025, 0.64, 2.1025, 1.21, 0.04, 0.01) on (0.3025, 0.64, 2.1025, 1.21, 0.04, 0.01, 2.7225).
Explanation
The test for ARCH is based on a regression of the squared residuals on their lagged values. The squared
residuals are (1.8225, 0.3025, 0.64, 2.1025, 1.21, 0.04, 0.01, 2.7225). So, (1.8225, 0.3025, 0.64, 2.1025,
1.21, 0.04, 0.01) is regressed on (0.3025, 0.64, 2.1025, 1.21, 0.04, 0.01, 2.7225). If coe cient a1 in:
2 2
ε̂ t = a 0 + a 1 ε̂ t-1 + μt
is statistically di erent from zero, the time series exhibits ARCH(1).

In the time series model: yt=b0 + b1 t + εt, t=1,2,...,T, the:
A) disturbance terms are autocorrelated.
B) disturbance term is mean-reverting.
C) change in the dependent variable per time period is b1.
Explanation
The slope is the change in the dependent variable per unit of time. The intercept is the estimate of the
value of the dependent variable before the time series begins. The disturbance term should be
independent and identically distributed. There is no reason to expect the disturbance term to be mean-
reverting, and if the residuals are autocorrelated, the research should correct for that problem.
Barry Phillips, CFA, has the following time series observations from earliest to latest: (5, 6, 5, 7, 6, 6, 8, 8, 9,
11). Phillips transforms the series so that he will estimate an autoregressive process on the following data
(1, -1, 2, -1, 0, 2, 0, 1, 2). The transformation Phillips employed is called:
A) beta drift.
B) moving average.
C) rst di erencing.
Explanation
Phillips obviously rst di erenced the data because the 1=6-5, -1=5-6, .... 1 = 9 - 9, 2 = 11 - 9.

The table below includes the rst eight residual autocorrelations from tting the rst di erenced time
series of the absenteeism rates (ABS) at a manufacturing rm with the model ΔABSt = b0 + b1ΔABSt-1 + εt.
Based on the results in the table, which of the following statements most accurately describes the
appropriateness of the speci cation of the model, ΔABSt = b0 + b1ΔABSt-1 + εt?
Lagged Autocorrelations of the Residuals of the First Di erences in

Absenteeism Rates
1 −0.0738 0.1667 −0.44271
2 −0.1047 0.1667 −0.62807
3 −0.0252 0.1667 −0.15117
4 −0.0157 0.1667 −0.09418
5 −0.1262 0.1667 −0.75705
6 0.0768 0.1667 0.46071
7 0.0038 0.1667 0.02280
8 −0.0188 0.1667 −0.11278
A) The negative values for the autocorrelations indicate that the model does not t the time
series.
B) The Durbin-Watson statistic is needed to determine the presence of signi cant correlation of
the residuals.
C) The low values for the t-statistics indicate that the model ts the time series.
Explanation
The t-statistics are all very small, indicating that none of the autocorrelations are signi cantly di erent
than zero. Based on these results, the model appears to be appropriately speci ed. The error terms,
however, should still be checked for heteroskedasticity.
An analyst wants to model quarterly sales data using an autoregressive model. She has found that an
AR(1) model with a seasonal lag has signi cant slope coe cients. She also nds that when a second and
third seasonal lag are added to the model, all slope coe cients are signi cant too. Based on this, the best
model to use would most likely be an:
A) AR(1) model with 3 seasonal lags.
B) AR(1) model with no seasonal lags.
C) ARCH(1).
Explanation
She has found that all the slope coe cients are signi cant in the model xt = b0 + b1xt– 1 + b2xt– 4 + et.
She then nds that all the slope coe cients are signi cant in the model xt = b0 + b1xt– 1 + b2xt– 2 + b3xt–
3 + b4xt– 4 + et. Thus, the nal model should be used rather than any other model that uses a subset of
the regressors.
The procedure for determining the structure of an autoregressive model is:
A) estimate an autoregressive model (e.g., an AR(1) model), calculate the autocorrelations for the
model's residuals, test whether the autocorrelations are di erent from zero, and revise the
d l if h i i l i
B) estimate an autoregressive model (for example, an AR(1) model), calculate the
autocorrelations for the model's residuals, test whether the autocorrelations are di erent
f d dd A l f h i i l i
C) test autocorrelations of the residuals for a simple trend model, and specify the number of
signi cant lags.
Explanation
The procedure is iterative: continually test for autocorrelations in the residuals and stop adding lags
when the autocorrelations of the residuals are eliminated. Even if several of the residuals exhibit
autocorrelation, the lags should be added one at a time.
Which of the following statements regarding time series analysis is least accurate?
A) If a time series is a random walk, rst di erencing will result in covariance stationarity.
B) An autoregressive model with two lags is equivalent to a moving-average model with two lags.
C) We cannot use an AR(1) model on a time series that consists of a random walk.
Explanation
An autoregression model regresses a dependent variable against one or more lagged values of itself
whereas a moving average is an average of successive observations in a time series. A moving average
model can have lagged terms but these are lagged values of the residual.

Consider the estimated AR(2) model, xt = 2.5 + 3.0 xt-1 + 1.5 xt-2 + εt t=1,2,...50. Making a prediction for
values of x for 1 ≤ t ≤ 50 is referred to as:
A) requires more information to answer the question.
B) an in-sample forecast.
C) an out-of-sample forecast.
Explanation
An in-sample (a.k.a. within-sample) forecast is made within the bounds of the data used to estimate the
model. An out-of-sample forecast is for values of the independent variable that are outside of those
used to estimate the model.

R06 Time-Series Analysis - Answers

Uploaded by

Copyright:

Available Formats

R06 Time-Series Analysis - Answers

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

R06 Time-Series Analysis - Answers

Uploaded by

Copyright:

Available Formats

Question #1 of 112 Question ID: 1208528

E[x2007:I]= 44 + 0.1 × xt– 1 - 0.25 × xt– 2 - 0.15 × xt– 3

E[x2007:I] = 44 + 0.1×33 - 0.25×32 - 0.15×35

E[x2007:II] = 44 + 0.1×34.05 - 0.25×33 - 0.15×32

(Study Session 2, Module 6.2, LOS 6.d)

Question #2 of 112 Question ID: 1208552

(Study Session 2, Module 6, LOS 6.i)

Question #3 of 112 Question ID: 1208581

εt2 = 0.25 + 0.6ε2t-1 + µt, where ε = ε^

(Study Session 2, Module 6.5, LOS 6.m)

Question #4 of 112 Question ID: 1208544

Which of the following statements regarding an out-of-sample forecast is least accurate?

B) There is more error associated with out-of-sample forecasts, as compared to in-sample

(Study Session 2, Module 6, LOS 6.g)

Question #5 of 112 Question ID: 1208540

Model: ΔExpt = b0 + b1ΔExpt–1 + εt

Coe cients Standard Error t-Statistic p-value

Intercept 1.3304 0.0089 112.2849 < 0.0001

Lag-1 0.1817 0.0061 30.0125 < 0.0001

The mean-reverting level is b0 / (1 − b1) = 1.3304 / (1 − 0.1817) = 1.6258.

(Study Session 2, Module 6.2, LOS 6.f)

Question #6 of 112 Question ID: 1208507

(Study Session 2, Module 6.1, LOS 6.b)

Question #7 of 112 Question ID: 1208539

A) If the time-series variable is x, then xt = b0 + b1xt-1.

(Study Session 2, Module 6.2, LOS 6.f)

Question #8 of 112 Question ID: 1208530

observation is 22. What is the forecast for the 52nd observation?

Using the chain-rule of forecasting,

Forecasted x51 = −6.0 + 1.1(22) + 0.3(20) = 24.2.

Forecasted x52 = −6.0 + 1.1(24.2) + 0.3(22) = 27.22.

(Study Session 2, Module 6.2, LOS 6.d)

Question #9 of 112 Question ID: 1208577

Lagged Autocorrelations of the Log of Quarterly Theater Ticket Sales

Lag Autocorrelation Standard Error t-Statistic

1 −0.0738 0.1667 −0.44271

2 −0.1047 0.1667 −0.62807

3 −0.0252 0.1667 −0.15117

4 0.5528 0.1667 3.31614

A) contains ARCH (1) errors.

C) would be more appropriately described with an MA(4) model.

(Study Session 2, Module 6.4, LOS 6.l)

Question #10 of 112 Question ID: 1208518

Which of the following statements regarding covariance stationarity is CORRECT?

C) A time series may be both covariance stationary and heteroskedastic.

(Study Session 2, Module 6, LOS 6.c)

Question #11 of 112 Question ID: 1208538

Regression Results for Maintenance Expense Changes

Model: DExpt = b0 + b1DExpt–1 + et

Coe cients Standard Error t-Statistic p-value

Intercept 1.3304 0.0089 112.2849 < 0.0001

Lag-1 0.1817 0.0061 30.0125 < 0.0001

Lagged Residual Autocorrelations for Maintenance Expense Changes

Lag Autocorrelation t-Statistic Lag Autocorrelation t-Statistic

1 −0.239 −2.39 7 −0.018 −0.18

2 −0.278 −2.78 8 −0.033 −0.33

3 −0.045 −0.45 9 0.261 2.61

4 −0.033 −0.33 10 −0.060 −0.60

5 −0.180 −1.80 11 0.212 2.12

6 −0.110 −1.10 12 0.022 0.22