Decomposition Models

Decomposition procedures are used in time series to describe the trend and seasonal factors in
a time series. More extensive decompositions might also include long-run cycles, holiday
effects, day of week effects and so on. Here, well only consider trend and seasonal
decompositions.

One of the main objectives for a decomposition is to estimate seasonal effects that can be
used to create and present seasonally adjusted values. A seasonally adjusted value removes
the seasonal effect from a value so that trends can be seen more clearly. For instance, in
many regions of the U.S. unemployment tends to decrease in the summer due to increased
employment in agricultural areas. Thus a drop in the unemployment rate in June compared to
May doesnt necessarily indicate that theres a trend toward lower unemployment in the
country. To see whether there is a real trend, we should adjust for the fact that unemployment
is always lower in June than in May.

Basic Structures

The following two structures are considered for basic decomposition models:

1. Additive: xt = Trend + Seasonal + Random

2. Multiplicative: xt = Trend * Seasonal * Random

The Random term is often called Irregular in software for decompositions.

How to Choose Between Additive and Multiplicative Decompositions

The additive model is useful when the seasonal variation is relatively constant over
time.

The multiplicative model is useful when the seasonal variation increases over time.

Example 1: In Lesson 1.1, we looked at quarterly beer production in Australia. The seasonal
variation looked to be about the same magnitude across time, so an additive decomposition
might be good. Heres the time series plot:
Example 2: Weve seen at least one example so far in the course where a multiplicative
decomposition would be good the quarterly earnings data for the Johnson and Johnson
Corporations. The seasonal variation increases as we move across time. A multiplicative
decomposition could be useful. Heres the plot of the data:

Basic Steps in Decomposition

1. The first step is to estimate the trend. Two different approaches could be used for this
(with many variations of each).

One approach is to estimate the trend with a smoothing procedure such as moving
averages. (See Lesson 5.2 for more on that.) With this approach no equation is used to
describe trend.

The second approach is to model the trend with a regression equation.

2. The second step is to de-trend the series. For an additive decomposition, this is done by
subtracting the trend estimates from the series. For a multiplicative decomposition, this is
done by dividing the series by the trend values.

3. Next, seasonal factors are estimated using the de-trended series. For monthly data, this
entails estimating an effect for each month of the year. For quarterly data, this entails
estimating an effect for each quarter. The simplest method for estimating these effects is to
average the de-trended values for a specific season. For instance, to get a seasonal effect for
January, we average the de-trended values for all Januarys in the series, and so on. (Minitab
uses medians rather than means, by the way.)

The seasonal effects are usually adjusted so that they average to 0 for an additive
decomposition or they average to 1 for a multiplicative decomposition.

4. The final step is to determine the random (irregular) component.

For the additive model, random = series trend seasonal.

For the multiplicative model, random = series / (trend*seasonal)

The random component could be analyzed for such things as mean absolute size, or mean
squared size (variance), or possibly even for whether the component is actually random or
might be modeled with an ARIMA model.

A few programs iterate through the steps 1 to 3. For example, after step 3 we could use the
seasonal factors to de-seasonalize the series and then return to step 1 to estimate the trend
based on the de-seasonalized series. Minitab does this (and estimates the trend with a straight
line in the iteration.

Decomposition in R

The basic command is decompose.

For an additive model decompose(name of series, type ="additive").
For a multiplicative decomposition decompose(name of series, type
="multiplicative").

Important first step: As a preliminary you have to use a ts command to define the seasonal
span for a series.

For quarterly data, it might be name of series = ts(name of series, freq

=4).
For monthly data, it might be name of series = ts(name of series, freq =
12).

You can plot the elements of the decomposition by putting the decompose command as an
argument of a plot command.

As an example, plot (decompose(earnings, type="multiplicative")).

Another way to plot is to store the results of the decomposition into a named object and then
plot the object. As an example,

decompearn = decompose (earnings, type = "multiplicative")

plot (decompearn)

To see all elements of a stored object, simply type its name. For instance, entering
decompearn will show all elements of the decomposition in the example above.

When the decomposition is stored in an object, you also have access to the various elements
of the decomposition. For instance, in the example just given, decompearn$figure
contains the seasonal effects values for four quarters.

You could print the seasonal figures simply by entering decompearn$figure. You
could plot them using plot(decompearn$figure).

Example 1 Continued: Additive Decomposition for Beer Production

The following commands produced the graph and numerical output that follows for the beer
Australian beer production series.

beerprod = scan("beerprod.dat")
beerprod = ts(beerprod, freq = 4)
decompbeer = decompose (beerprod, type="additive")
plot (decompbeer)
decompbeer

The plot shows the observed series, the smoothed trend line, the seasonal pattern and the
random part of the series. Note that the seasonal pattern is a regularly repeating pattern.
Heres the numerical output:

$seasonal
Qtr1 Qtr2 Qtr3 Qtr4
1 7.896324 -40.678676 -24.650735 57.433088
2 7.896324 -40.678676 -24.650735 57.433088
3 7.896324 -40.678676 -24.650735 57.433088
4 7.896324 -40.678676 -24.650735 57.433088
5 7.896324 -40.678676 -24.650735 57.433088
same rows down to 18

$trend
Qtr1 Qtr2 Qtr3 Qtr4
1 NA NA 255.3250 254.4125
2 257.4500 260.1000 262.8375 264.6875
3 265.4125 264.6500 262.4625 260.4000
4 261.2625 262.9875 266.1875 269.2375
5 270.5125 271.4625 272.1750 274.0125
6 274.3750 277.4500 278.9750 279.1750
7 282.9000 285.2875 287.9375 290.3875
and so on to the 18th row

$random
Qtr1 Qtr2 Qtr3 Qtr4
1 NA NA -3.77426471 -3.44558824
2 -3.34632353 8.47867647 -2.08676471 -1.72058824
3 -1.40882353 8.82867647 -0.81176471 -4.43308824
4 7.75882353 4.49117647 8.36323529 -12.37058824
5 7.69117647 -4.28382353 12.87573529 -20.04558824
6 12.42867647 -4.17132353 2.87573529 2.59191176
7 -11.69632353 5.19117647 6.51323529 -2.12058824
and so on to the 18th row

$figure
[1] 7.896324 -40.678676 -24.650735 57.433088

Using the Seasonal Values

The elements of $figure are the effects for the four quarters. Note that the seasonal effect
values are repeated each year (row) in the $seasonal object at the top of this page.

The seasonal values are used to seasonally adjust future values. Suppose for example that the
next quarter 4 seasonal value past the end of the series has the value 535. The quarter 4
seasonal effect is 57.433088, or about 57.43. Thus for this future value, the de-seasonalized
or seasonally adjusted value = 535 57.43 = 477.57.

How the Trend Values Were Calculated

The trend values were determined as centered moving averages of span 4 (because there
are four quarters per year). Heres how the centered moving average for time = 3 would be
calculated.

Average the observed data values at times 1 to 4: [Math Processing Error]

Average the values at times 2 to 5: [Math Processing Error]

Then average those two averages:

[Math Processing Error]

More generally, the centered moving average smoother for time t (with 4 quarters) is

[Math Processing Error]

Following are the first 8 values in the observed series. The smoothed trend value for time 3
in the series (Qtr 3 of year 1) is 255.325 and the smoothed trend value for time 4 is
254.4125. Use the data below to verify these values (and your understanding of the
procedure).

Qtr1 Qtr2 Qtr3 Qtr4

1 284.4 212.8 226.9 308.4
2 262.0 227.9 236.1 320.4

For monthly data the centered moving average smoother for time t will be

[Math Processing Error]

Example 1 Continued: Multiplicative Decomposition for Beer Production

The following two commands will do a multiplicative decomposition of the beer production
series and print the seasonal effects.

decombeermult = decompose (beerprod, type = "multiplicative")

decombeermult$figure

The seasonal (quarterly) effects are:

1.0237877 0.8753662 0.9233315 1.1775147

To seasonally adjust a value, divide the observed value of the series by the seasonal factors.
For example if a future quarter 4 value is 535, the seasonally adjusted value = 535/1.1775147
= 454.34677.

Lowess Seasonal and Trend Decomposition

A lowess smoother essentially replaces values with a locally weighted robust regression
estimate of the value. The R command stl does an additive decomposition in which a
lowess smoother is used to estimate the trend and (potentially) the seasonal effects as well.
There are several parameters that can be adjusted, but the default does a fairly good job.

For our beer production example, the following command works:

stl (beerprod, "periodic")

The periodic parameter essentially causes the seasonal effects to be estimated in the usual
way, as averages of de-trended values. The alternative to this is a s.window = some odd
number of lags, which uses lowess smoothing procedures to estimate the seasonal effects
based on a number of years = the specified s.window value. When you do this, the
seasonal effects will change as you move through the series.

Heres a piece of the result of the stl(beerprod, "periodic") command. Note

that the word remainder is used rather than random. (Maybe its not really random!)

seasonal trend remainder

1 Q1 8.06289 267.3569 8.9802489
1 Q2 -41.58529 261.8710 -7.4856917
1 Q3 -24.68456 257.1444 -5.5598227
1 Q4 58.20698 253.8595 -3.6664533
2 Q1 8.06289 257.4133 -3.4761521
2 Q2 -41.58529 260.6083 8.8769848
2 Q3 -24.68456 262.9967 -2.2121517
2 Q4 58.20698 264.2030 -2.0099463
3 Q1 8.06289 265.5992 -1.7621096
3 Q2 -41.58529 265.2844 9.1008543
3 Q3 -24.68456 262.6567 -0.9721191
3 Q4 58.20698 259.6218 -4.4287360
and so on for years 4 to 16

17 Q1 8.06289 417.3459 -6.2088201

17 Q2 -41.58529 420.2610 -1.9757578
17 Q3 -24.68456 428.1381 -10.6535322
17 Q4 58.20698 435.8692 12.0238431
18 Q1 8.06289 441.0740 9.2631448
18 Q2 -41.58529 447.1144 -18.1291496
18 Q3 -24.68456 453.0243 -1.4397012
18 Q4 58.20698 459.3832 7.4098529

The additive seasonal effects are 8.06289, -41.58529, -24.68456, 58.20698.

These arent much different than what we got from the additive decompose. Those seasonal
values were

$figure

[1] 7.896324 -40.678676 -24.650735 57.433088

The command plot(stl(beerprod, "periodic")) gave the following plot.