Assignment Forecasting

Overview: Your mission on this assignment is to forecast the next three years of quarterly sales at
Gap Inc., a well-known chain of shopping-mall clothing stores. This exercise should help you to
(i) get some practice with basic seasonal time series forecasting models, (ii) see how forecasts for
the future are affected by model assumptions, and (iii) see how different models respond to changes
in the pattern of the data.
In particular, you should first try fitting most of the “default” seasonal models that Statgraphics
uses when seasonality is present. Then, based on everything you know (or can find out without too
much trouble) about what is currently happening at the Gap, pick a model that you think gives the
most plausible forecasts and confidence intervals for the next three years.

How to proceed:
(a) The data can be obtained from the Thomson Research database via the Fuqua library web site,
which can be reached from the “Other Fuqua Databases” link on the course home page. (You will
need to use the VPN or proxy server if doing this from off-campus. Alternatively you could use
the Thomson One Banker database, although its server seems to get overloaded at some times of
day.) Use the company search field in the upper left of the Thomson Research web page, enter
“Gap Inc” (without the quotes), and click the “Go” button. (Alternatively you can search on the
ticker symbol, which is GPS.) This should take you to the “Overview” page for Gap Inc. Look
under “Spreadsheet Financials” for the “SEC 10-Q History” report. (Be sure to choose 10-Q
history and not one of the other options. If you are using the Thomson One Banker database
instead, you first need to click the “Filings” list from the overview page to get to the page with the
spreadsheet financials link.) Click the 10-Q history link and you should get a prompt to save the
data in an Excel spreadsheet file. This spreadsheet contains (among other things) a 10-year history
of net sales up to the quarter ending 5/5/2007. The net sales figures are found in the “Quarterly
Income” section near the bottom of the sheet. You can get an approximate figure for net sales in
the 2nd quarter (which ended on 8/4/2007) from a August 23 press release that is available on the
www.GapInc.com web site. See “About GAP Inc… Media… Press Releases…. 8/23/07… Gap
Inc. reports second quarter earnings per share..”. This will lead to a pdf document, and the net sales
figure for the quarter ending 8/4/2007 is reported in the third paragraph on the first page.

(b) Now create a new workbook containing only the net sales data and quarter ending dates in
column-wise format—i.e., dates in one column and net sales in another column, with variable
names in the first row (e.g., DATE and NETSALES). To do this, go to the Quarterly Income
section at the bottom of the 10Q history spreadsheet, highlight and copy the two rows entitled
“FISCAL QUARTER ENDING” and “NET SALES” (including the row labels), and use the Paste-
Special command with the “Transpose” box checked to paste them to a new worksheet in column-
wise rather than row-wise format. Change the column headings to DATE and NETSALES, and
use the Data/Sort command to sort the data in ascending order of dates. Your worksheet should
like the picture below, with the final entry for 5/5/2007 in row 41, and you should then manually
add the figure for the quarter ending 8/4/2007 in row 42.

Save the new Excel workbook, close it, and then open it in Statgraphics. You should see the
variables appear there, pretty much as they looked in Excel. The file will initially be “readonly”,
so to get write-access to it, right-click on the datasheet, select “Databook properties,” and uncheck
the read-only box for sheet A. (The numbers on the datasheet should change from gray to black at
this point.) Next, in column 3, use the right-mouse-button “Modify Column” option to create a
new variable called TIME whose values are generated by the following formula: COUNT(1,41,1).
This will create a series of TIME numbers from 1 to 41 in steps of size 1.

(b) As the first stage of your analysis, you should go to the Describe/Time_Series/
Seasonal_Decomposition procedure in order to get plots and printouts of the seasonal indices that
will be used to perform multiplicative seasonal adjustment, as well as to see what the data look
like in seasonally adjusted terms. Specify that the data are QUARTERLY with a starting date of
Q2/97 and seasonality of 4. (Note: you do not have to save any of the results from the Seasonal
Decomposition procedure in order to perform seasonal adjustment later as part of your forecasting
model. When you specify “Multiplicative Seasonal Adjustment” as part of your forecasting model,
the seasonal indices will automatically be computed and applied for you. However, the forecasting
procedure does not print out the seasonal indices that were used, so to get this information—
assuming you are interested in seeing it—you must use the Seasonal Decomposition procedure
separately.) Copy and paste the table of seasonal indices to a slide in your presentation, as well as
graph of the seasonally adjusted data. What has been happening over the last few years, in
seasonally adjusted terms? (Note: unfortunately the Seasonal Decomposition procedure does not
have an additive seasonal adjustment option, whereas the User-specified Forecasting procedure
does have this option. Hence, if you end up fitting a forecasting model that uses additive seasonal
adjustment in conjunction with a log transformation, you cannot get a printout of the additive
indices that were estimated. However, the net effect of using additive seasonal adjustment on
logged data is almost equivalent to using multiplicative seasonal adjustment on unlogged data, so
the multiplicative seasonal indices will accurately reflect the estimated seasonal pattern in the
original data for both types of models.)
(c) Next, go to the Forecast/User-Specified-Model procedure to analyze the same series. Hold out
12 observations for validation and generate 12 forecasts, and look initially at the following five
models:
A. Random walk with drift (w/ multiplicative seasonal adjustment)
B. Holt’s linear exponential smoothing (w/ multiplicative seasonal adjustment)
C. Brown’s simple exponential smoothing (w/ multiplicative seasonal adjustment)
D. Brown’s linear exponential smoothing (w/ multiplicative seasonal adjustment)
E. Winter’s seasonal smoothing

(Note that these are not exactly the five “default” models that you get when you first run the
procedure: model B has been changed from linear trend model to Holt’s linear exponential
smoothing.) Which model appears best in the estimation period? Which appears best in the
validation period? Look at the time sequence plots of the data and forecasts—do the models differ
significantly in their predictions for the next three years? Pay close attention to the trends that are
projected as well as the appearance of the confidence intervals for the different models— do they
appear realistic, particularly when forecasting more than a few quarters into the future? Would you
be comfortable presenting them in the boardroom? Also, look at the plots of residuals versus time
and residual autocorrelations: is the error variance roughly constant over time, and is there
significant autocorrelation? Copy and paste the Model Comparison report in which 12
observations were held out for validation onto one of your slides.

(d) You may wish to try to “fine tune” the models by using appropriate data transformations. For
example, you may want to add a log transformation to some of the models and/or add a trend to
Model C via the inflation adjustment option. (As discussed in video clip #10, you can estimate the
inflation rate that should be used for this purpose by fitting a random walk with drift model with a
log transformation—the drift term will be interpretable as the average percentage growth per
period over the history of the series. However, keep in mind that this approach assumes a constant
rate of growth over the whole history of the series. There might be reasons for using some other
assumed growth rate for predicting what will happen in the near future.) Do not change the basic
model when doing your fine-tuning: model A should still be a random walk with drift, model B
should still be Holt’s model, and so on. If you use a log transformation, remember to change the
type of seasonal adjustment from multiplicative to additive. However, do NOT use a log
transformation in conjunction with model E—the Winter’s model—because that model has built-
in multiplicative seasonal adjustment and there is no additive option for it in Statgraphics. See how
your fine tuning affects the appearance of the plots of forecasts and confidence intervals and the
plots of residuals versus time, as well as the error statistics. In interpreting the residual plots, one
issue to consider is whether there is any trend in the residuals, which would reflect an inability of
the model to respond to any systematic changes in the trend in the original data. Also pay close
attention to the issue of whether the variance of the residuals is roughly constant over time, and
think about the implications for calculating realistic confidence intervals from the model. When
you have converged on a “final” set of five models, copy and paste the Model Comparison report.

(e) Before proceeding to generate your “final” forecasts from the various models, go back to the
Data Input panel and set the number of points held out for validation to zero. (When producing
real forecasts for the future, you obviously want to use all of the most recent data to estimate the
model!) Now, for each of the five models, copy and paste the following three plots onto your
slides: (i) the time series plot of forecasts and actual values, (ii) the time series plot of residuals,
and (iii) the autocorrelation plot of residuals. For each model, you should be able to fit all 3 onto
one slide if the panes in Statgraphics from which they are copied are about three times as wide as
they are tall.

(f) Another interesting plot to draw is a side-by-side plot of the forecasts of all five models over
the next 12 quarters. To do this, use the “Save results” button on the toolbar to save the forecasts
of each model, in turn, to the original datasheet, i.e., sheet A. (In other words, go back to the
“Analysis options” panel, select model A, then use the “Save results” button to save the forecasts
under a unique name such as FORECASTSA. Then repeat this process to change to model B and
save its forecasts under another name such as FORECASTSB, etc. Note that you will have to click
the “A” radio button to save the variables to sheet A, because the default option will be sheet B.)
Then go look at the datasheet to make sure they are there—the forecast columns should extend 12
rows below the end of the data. Extend the TIME variable by modifying the COUNT formula to
count up to 53 instead of 41, then use the “Plot/Scatterplot/Multiple X-Y Plot” procedure to plot
the five forecast variables side-by-side versus TIME. How similar are the five models in their one-
step-ahead forecasts for the past quarters? How similar are they in their extrapolations into the
future? (You may want to use the “Select” field to make two versions of this plot—one that shows
only the 41 past quarters and another that shows the 12 future quarters. You can do this by using
the selection criteria TIME41, respectively, assuming that the TIME variable has been extended
to row 53. To the extent that the models differ, can you explain the differences in terms of their
respective underlying assumptions about trends and/or seasonal patterns?

(g) What would the 3-year-ahead predictions of the various models have looked like if the
forecasts had been generated at the end of the 2nd quarter 2004, i.e., using only the 29 rows of data
only up to 7/31/04 and how well would they have agreed with the actual last three years? A simple
way to test this is as follows. First, highlight the sales data and use the copy and paste commands
to create a new column containing exactly the same sales data. Use the right-mouse-button
“Modify column” command to assign the new column the name NETSALES04. Then, at the
bottom of the NETSALES04 column delete all the data values after 7/31/04 (leaving only the first
29 values). Thus, NETSALES04 has 12 fewer quarters of data than the original variable. Go back
to the Forecasting procedure and re-run it by substituting NETSALES04 for NETSALES in the
input dialog box, leaving everything else the same. For a particularly dramatic comparison, use
the right-mouse button Graphics Options to “hold” the Yaxis scale of the forecast plot before doing
this, so that the new forecast plots are drawn with the same scale as the original one. Go back and
look at the plot of the seasonally adjusted data, and then look at the forecast plots for each of the
five models and see if they jibe with your intuition about how the different models react to changes
in the trend in the data. Which models are more stable and/or more sensible in changing economic
times? Which forecasts would have been more credible in the boardroom? (For that matter, what
HAS happened in the boardroom at Gap Inc. lately?) You can also save the forecasts from the
truncated data sets to the spreadsheet and generate a multiple X-Y plot comparing them to the
actual values for the last three years.

(h) From your five “final” models, choose the one that you consider the most accurate and most
credible for predicting the next three years, i.e., your “best guess”. For this model, copy and paste
the Analysis Summary report and the lower part of the “Forecast Table”, which shows the forecasts
and 50% confidence limits for the next 12 quarters, onto your slides. (Note that you can highlight
just the bottom part of Forecast Table before copying, and only the highlighted rows will be
copied.) The forecasts and confidence limits will be formatted in “scientific” notation, so that
$1,234,567 would be displayed as 1.234567E6, where the E6 means times-10- to-the-6th power,
i.e. millions. Admittedly this a bit geeky for the boardroom. For a more 4 businesslike presentation,
you might want to do the following: copy and paste the selected rows of the Forecast Table report
to an Excel worksheet. Then reformat the numbers there, say, in currency format with no decimal
places, and then paste the table from here onto your slides. (I recommend you use Paste
Special/Picture when copying the table from Excel. Another issue you may want to consider is
how to appropriately round off the displayed numbers so that you don’t show more digits than are
really meaningful for forecasts.) Finally, you should also copy and paste the Model Comparison
report, showing the statistics of the expanded estimation period with no validation period.

(i) [This should not take long.] Go back to the Thomson Research database and download
the 10-Q history for American Eagle Outfitters (ticker symbol AEO), which is one of
Gap’s key competitors although a much smaller company. Repeat the same steps as in
part (b) before to download the spreadsheet financials into an Excel file and then create
a new Excel file with (only) the date and net sales information in the first two columns,
sorted in ascending order of dates. Use the same names (e.g., DATE and NETSALES)
that you used before. Save this file under a new name (e.g., AEO.XLS). You will not
need to search for the second-quarter sales figure from another source: the Q2/07 value
is already in the history file, although the Q2/97 value is missing. Now go back to your
Gap statfolio in Statgraphics, where your final models for Gap are set up. Use the
File/Open/Open Data source command to open the new Excel file with the AEO data,
and answer “Yes” when asked whether you want to replace the data in datasheet A.
You should now see exactly the same models fitted to the AEO data. The only thing
 you will need to change is the starting date on the model input panel in the Forecasting
and Seasonal Decomposition procedures to Q3/97 rather than Q2/97. As before, first
go to the Seasonal Decomposition procedure as before and look at the estimated
seasonal indices and the plot of the seasonally adjusted data. How well has American
Eagle been doing in comparison to Gap? (Note: the seasonal indices will be offset by
one period because the AEO data begins in Q3 rather than Q4. Thus, AEO’s index in
season 1 actually corresponds to Gap’s index in season 2. The “high” season for both
companies is the fourth quarter, which is season 3 in the Gap data and season 2 in the
AEO data.) Now go to the forecasting procedure, look at the model comparison table
and forecast plots, and see which model you think fits the AEO data best.

What to hand in: Prepare a Powerpoint presentation illustrated by the reports and graphs from
your models, including your comments on the comparative performance of the different
models and your best guess as to what is really going to happen to sales at Gap Inc. over the
next 12 quarters if recent historical trends continue, including confidence limits that you
consider to be realistic. The first few slides of your presentation should clearly show the
forecasts and confidence limits of your final chosen model in tabular and/or graphical format,
together with a short explanation of why this model was chosen. As in the last assignment, you
may wish to consult public news sources for information about the history of GAP Inc. and its
industry (e.g., how the outlook and strategy for the future has changed over the last few years)
that might shed light on the patterns you see in the data and help to justify the assumptions that
are made concerning the appropriate model type and relevant data history (especially insofar
as trend estimation is concerned). Next, include the slides showing the key plots from all the
models that were tested, with brief 5 comments on the goodness of fit of each model. Finally,
include one or two slides (only) discussing AEO’s performance over the same period in
comparison to Gap’s. In particular, include a slide that shows (only) the following three plots
from the AEO model you think is best: the seasonally adjusted data, the forecasts and their
confidence intervals, and the residual time series plot. Submit your Powerpoint file along with
your Statgraphics data and statfolio files. P.S. Gap stores are classified as “Family clothing
stores: NAICS 44814”. If you are interested in how the rest of the industry has been doing, you
can compare Gap’s quarterly sales to the retail sales (NSA) of all family clothing stores as
shown in Economagic, converting the latter from monthly to quarterly. You will note that Gap
has a significant share of that market. (The “Transform this series” option in Economagic will
do a monthly-to-quarterly transformation, although the quarter-ending dates are not the same
as used by Gap Inc. You could also do your own transformation in Excel.)