In the last issue we looked at Sibyl/Runner for Windows, a new incarnation of one of the earliest forecasting programs. In this issue we continue the theme of forecasting software with a review of Forecast Pro for Windows. Originally released for the IBM-compatible arena more than five years ago, Forecast Pro eventually became one of the market leaders in the field. The Windows version came out last year, and an extended version followed. Here are my impressions.

FORECAST PRO FOR WINDOWS, EXTENDED EDITION

by Jack Yurkiewicz, Pace University

Forecast Pro, originally released as a DOS program, was one of the earliest and best "automatic" forecasting programs, and has dominated its market almost from its inception. By automatic I mean that the forecasting program reads your data and suggests the forecasting method that should be used. If you agree with the suggestion, then it finds the optimal parameters for the model. I label a program like Sibyl/Runner (reviewed last issue) a þsemiautomaticþ program. You, not the software, must choose the model, and it then finds the correct parameters.

While many users may be leery of using software as a "black forecasting box," in practice, I found the software worked quite well, almost always choosing the most appropriate model for the sample data. Even when, on some especially difficult time series, Forecast Pro suggested a method that was "sub-optimal," the model it did recommend was "off" by a statistically insignificant amount.

Overview of the Product

Forecast Pro offers just four methodologies for forecasting: simple moving averages, exponential smoothing, Box-Jenkins, and "dynamic" regression. While that list may look sparse, these methods cover most of the techniques used by practitioners and academics. For example, exponential smoothing covers the entire family: Brown's simple exponential smoothing, Holt's method for linear trend, and Winters' procedure for seasonality. You can customize the model. Thus, for example, you can specify that the data has seasonality but no trend, and choose either a multiplicative or additive model. The extended version enhances the capabilities of the standard version by including Census X-11 decomposition, allowing event adjustment models (which provides corrections to exponential smoothing models for special events such as strikes, promotions, and other occurrences), and giving cumulative forecasts and confidence limits (which can be used to compute safety stocks of items when the reorder lead time must be taken into account).

Dynamic regressive is the authors' name for what I would call a "manual-stepwise" regression procedure. To use it, the user must have a separate file for each variable. You then literally build your regression model, by clicking on the dependent variable, choosing the independent variables, and clicking on the constant term. Forecast Pro then gives you the specifics of the model. You can select the MODEL DIAGNOSE ERROR ACF sequence from the menu to get the autocorrelation function, or choose MODEL DIAGNOSE DYNAMICS to get information on how to improve the model. For example, in one of my models, the program recommended adding a term called _AUTO[-1], which was an autoregressive error term. The model I next received, based on this recommendation, was the Cochran-Orcutt model. You then examine (via the MODEL DIAGNOSE VARIABLES menu sequence) various variable diagnostics, such as the Ljung-Box test. From here you may decide to build a new model, based on the Forecast Pro's suggestions and your analysis of the variable statistics. After you get the new "improved" model, you again go through the sequence of examining, testing, and getting recommendations. Eventually you settle for your "best" regression model. While all this is certainly more time consuming than the more automated stepwise regression procedures traditionally seen in other statistical software, the flexibility, level of advice and "hand holding" given by the software, and ease-of-use, makes the process worthwhile.

Using the Program

The graphics capabilities of Forecast Pro are impressive. Using either the Graph button on the toolbar, or Graph on the menu bar, gives a time plot of your data. You can graph some or all of the values. An impressive and useful option is "Peek and Poke." With this, if you click on any point on the graph, you get its value and time period in a box superimposed on the graph.

Forecast Pro's "expert" system could not be easier to use. Clicking on either EXPERT on the menu bar or on the "brain" icon on the toolbar gives, within several seconds on a fast computer, the recommended forecasting procedure. The Expert System does a series of statistical tests, interprets the results, may perform additional tests, and eventually, based on the results, suggests the model. Figure 1 shows the recommendations Forecast Pro made for one of my sample series. If you agree with the suggestion, then clicking on MODEL on the menu bar or "" on the toolbar gives a dialog box which asks if you want the software to find the optimal parameters of the model. Figure 2 shows the model and accompanying statistics that Forecast Pro found for my sample data.

You can call for a superimposed graph of the actual and fitted values, with forecasts and confidence intervals. The Peek and Poke feature is handy here, showing values for the actual and fitted results. Figure 3 shows a partial plot of my sample data and the fitted results, along with the expected forecasts and their 95 percent prediction intervals. Notice the "Peek and Poke" box which tells you the values of one of the forecasts and the 95 percent prediction limits.

Getting the autocorrelation and partial autocorrelation is very easy. The MODEL IDENTIFY sequence on the menu bar gets you either one of these important graphs. By changing the numbers in the "difference" dialog box and clicking on REDRAW we get the graph. If you click on "automatic" then the program determines the best difference and seasonal levels for your data and draws the appropriate plot. Figure 4 shows the sample autocorrelation function that the program found. It correctly determined, for this data, that single differencing and no seasonality would be the most appropriate choice.

Ease of Learning and of Use

I found data entry to be the weakest aspect of the software. Forecast Pro permits you to enter data at the keyboard, or it can import data from Lotus (WK1 format), Excel (XLS format), or ASCII (called UNV) files. When trying to import Excel 5.0 files, I frequently received a General Protection Fault error message, and was unceremoniously exited to Windows' Program Manager. Lotus or ASCII file importation went smoother, but still is cumbersome because of Forecast Pro's strict rules for the layout of the spreadsheet. Every variable needs special specific header information. This header must have six items: the variable name, a description of the variable, the starting year, the starting period, the number of periods per year, and the number of periods per seasonal cycle. You must also tell Forecast Pro if the data is arranged by row (horizontal) or column (vertical). You do this by putting the word "horizontal" or "vertical" respectively in cell A1. You must not put any other data in the A column or in the first row. Figure 5 shows how you must arrange the spreadsheet (shown here for my multiple regression data) before you can import it. In this spreadsheet, the actual observations start in row eight; the data in rows one through seven is the specific header information.

If you are an infrequent user of the software and routinely keep your data in a spreadsheet, you will undoubtedly forget these rules (I did) and will have to resort to the manual for help. It would be better if the program just read your spreadsheet with the time series and, after it is imported, added the needed header information based on your responses in a dialog box. You would then save that file in Forecast Pro's UNV format for future use.

Data entry at the keyboard is also awkward. The usual sequence of clicking on FILE and expecting then to click on NEW does not work; there is no NEW command. However, there is something called TABLEAU, which then leads to a series of confusing options. For example, NEW is a choice, yet you cannot enter new data here. I eventually found that the EDIT menu led to an EDIT VARIABLE dialog box, which allowed me to type in the obligatory header information, and then finally the data. The interface there is disappointing also because you do not get the usual spreadsheet layout for you to enter the data. Finally, only square root and logarithmic transformations are allowed, which may be insufficient for many users.

Retrieving a file can be awkward for experienced Windows users. Clicking on FILE and expecting to click next on OPEN does not work. The OPEN option is "greyed" (inaccessible); it is available if you specify beforehand, via the OPTIONS menu, that you want to retrieve a Lotus (WK1) or Excel (XLS) files. However, getting a file by using the toolbar is easier than with the menu bar. Clicking on the large blue "Yt" button lets you choose the file to retrieve.

In summary, Forecast Pro's nonstandard Windows interface makes getting data into the program much harder than it should be. Data entry and manipulation can only be rated as "fair." However, the other components of the software, from the expert system which identifies the model and finds the appropriate parameters, to the clear and flexible graphics capabilities, rates an "excellent" in ease of learning and of use.

Documentation

The documentation for the standard version of Forecast Pro is a 212-page wire-bound book, consisting of three main sections. The first, Getting Started, gives lucid instructions on how to install the software, and is highlighted by "A Quick Overview of Forecasting," which discusses forecasting methodologies and gives some forecasting tips. The second section, a tutorial, covers clearly all the features of the software. Each method (exponential smoothing, Box-Jenkins, and dynamic regression) gets its own chapter. The last part of the book, a reference section, goes into details about all the software commands. There is also a statistical reference component, which is an excellent mini-textbook explaining the various statistical forecasting proce- dures that the software uses. The book even includes a good glossary of forecasting terms and an extensive bibliography. The extended version of the software adds a second 50-page manual that covers, in the same format as the standard book, those features pertinent to the extended software.

The on-line help system works similarly to many Windows' products. It is easy to use, clearly written, and extensive. Combining this with the manuals, Forecast Pro rates an "excellent" for documentation.

Support and Pricing

Registered users get free (but not toll-free) technical support for one year. The single unit price of Forecast Pro is $495. The company suggests you call for an academic price schedule. Sadly, there is no student version.

Conclusions

Seeing why Forecast Pro has become a market leader in the field of statistical forecasting is not hard. Professional forecasters can use the product manually and get all the statistical output they need to validate their models. Novices and people not interested in the statistical details can use Forecast Pro in the automatic mode and trust the software to find an appropriate model and get forecasts. For most data sets I tried this works well, and even on those "messy" time series for which I found a better model than the program suggested, the recommendations (based on comparing statistics) were very close to my own choices. In short, novices can rely on the expert system.

Compared with Sibyl/Runner (reviewed in DECISION LINE, Vol. 26, No. 1), Forecast Pro has fewer forecasting procedures, but offers an expert system that Sibyl/Runner does not match. The graphs in Forecast Pro are superior also, permitting dates on the horizontal axis, and offering the excellent "Peek and Poke" feature to read points on the graph. However, the user interface is far clumsier in Forecast Pro. Sibyl/Runner's data entry procedures are much easier to use. I like and use them both, and both are permanent fixtures on my hard disk. Considering how limited and precious that storage area is, that is high praise indeed.

Business Forecast Systems
68 Leonard Street
Belmont, MA 02178
617-484-5050

If you are interested in writing a software review for a future issue of Decision Line, please contact:

Professor Jack Yurkiewicz
Pace University
212-346-1908
e-mail: yurk@pacevm.dac. pace.edu