Having learned forecasting via the mainframe version of Sibyl/Runner many years ago, I migrated to the DOS version (called PC/Sibyl) in 1989. So when the vendor announced a Windows version several months ago, he piqued my curiosity. Here is what I found.

SIBYL/RUNNER FOR WINDOWS

by Jack Yurkiewicz, Pace University

Sibyl/Runner, the classic forecasting program, first appeared as a mainframe program in 1973. Offering a collection of univariate and multivariate forecasting techniques, the product later became available for the personal computer as PC/Sibyl. Just recently, the Windows(TM) version became available, and it is ``not your father's'' Sibyl/Runner anymore. The graphical interface makes this version much easier to use, and the pixel graphs are a significant improvement over the character plots seen in earlier versions.

Overview of the Product

Sibyl/Runner is a program for analyzing and forecasting time-series data and estimating relationships between variables. It includes tools to help discover the seasonality, trend, and cycle in time series data, and gives expert advice, based on the M (Makridakis) Competition, for the appropriate forecasting model to choose for your data. The program offers a large collection of forecasting methods. They are:

• Single exponential smoothing
• Quadratic exponential smoothing
• Adaptive response rate smoothing
• Winters' method
• Holt's method
• Dampened trend 3-parameter smoothing
• Simple moving average
• Linear moving average
• Straight line trend extrapolation
• S-curve trend extrapolation
• Exponential growth trend extrapolation
• Box-Jenkins ARIMA
• Classical decomposition
• Census II decomposition
• Generalized adaptive filtering
• Harrison's harmonic smoothing
• Multiple regression

Using the Program

Sibyl/Runner allows you to enter data at the keyboard, using a spreadsheet format, or it can import ASCII data files. It allows up to 1000 observations and 20 variables (for multiple regression).

Once you have entered the data, you then must go to the ``setup'' procedure. Figure 1 shows the setup dialog box. You specify the seasonality, if any, from a set of choices (monthly, quarterly, etc.). If your data does not fit one of the seasonal options, then you can specify it manually. You must also tell the program how many of the observations to use, and the number of forecasts to make. If your data is seasonal and you want to deseasonalize it, you can optionally let Sibyl/Runner do it via classical decomposition, or you can enter your own seasonal indices as a separate variable. The explanatory variables' section is used only if you have multivariate data and want multiple regression analysis. By default, Sibyl/Runner will estimate the model's parameters, but you can bypass this and use your own. When you save the data, the software saves this accompanying dialog box also, so that you do not have to respecify these details the next time you want to analyze this data set.

After the data setup, Sibyl/Runner brings you to the Analyze module, which gives you guidance on which model is appropriate for the data. This module has several sections. First it gives you a time plot of the data. The resulting graph is attractive (see Figure 2), but you cannot specify dates on the horizontal axis. The Summary Analysis section investigates the data and returns with a description of the trend, seasonality (if appropriate), and randomness components. The Autocorrelation Analysis section is especially good. Taking the tone of a statistics teacher, the program gives a series of screens showing the autocorrelation and partial autocorrelation functions (tables and plots) along with a running commentary offering advice about their interpretation. Eventually, Sibyl/Runner advises you about the stationarity of the data and its seasonality. Figure 3 shows one screen of the autocorrelation analysis. Finally the program offers the Seasonal Factors and M-Competition Advice sections. The former analyzes the data and gives the seasonal indices. The latter proposes an appropriate set of possible forecasting techniques to try, with accompanying rationale based on the Makridakis Competition (in which various experts forecast 111 time series).

The Forecasting module allows you to choose the model to make the forecasts. A brief description of each technique is given in the Help section. After choosing a model, Sibyl/Runner brings you to the Forecast section, where you get the model parameters and the aggregate error statistics (MSE, MAD, Theil's U-Statistic, McLaughlin's Batting Average, and others). The software can give tables and graphs of the actual versus the fitted values and the forecasts. If you have tried a few forecasting models, you can easily compare the results from each. Sibyl/Runner will keep track of up to twenty runs and you can get a table showing side-by-side comparisons of each analysis, making it easier to decide which one is most appropriate.

Ease of Learning and of Use

Sibyl/Runner is very easy to learn. All the menus and dialog boxes are intuitive, and I had no trouble typing in data, going through the setup, choosing a technique, and making a forecast. However, I found importing data from other applications a bit awkward. Sibyl/Runner will not directly import spreadsheet files from Excel or Lotus because it reads only ASCII files. Even if you have data in ASCII format and you choose ``Import Data File'' from the menu, the succeeding sub-menu asks you to specify if it is ``Lotus.PRN'' file, or a ``.DAT'' file, or a ``.TXT'' file. If your data file does not have one of these extensions, then Sibyl/Runner will not even display it as a potential file to import.

I found the program easy to use. Sibyl/Runner is what I call a ``semiautomatic'' forecasting program (you choose the model and it optimizes the parameters, compared with an ``automatic'' program that analyzes the data, picks the appropriate model, and then optimizes the parameters). As mentioned, the autocorrelation advice is especially helpful. However, the advice given by the M-Competition, shown in Table 1 for one of my data files, while useful, can be confusing to novices or practitioners who are not familiar with that competition. Neither the help portion of the program nor the documentation discuss this matter.

Sibyl/Runner always gave me accurate results for my sample data. Unfortunately, I could not, at times, compare Sibyl/Runner's answers with those found by other products because of Sibyl/Runner's different choice of initial conditions. For example, in Brown's simple exponential smoothing, most forecasting texts recommend the user call the first observation the initial smoothed value. This is both intuitive and practical. Instead, Sibyl/Runner uses the average of the first three observations, resulting in a final forecast different from that found by these other sources. The optimal smoothing parameters that Sibyl/Runner found were always accurate and precise.

Documentation

Sibyl/Runner's documentation consists of a sparse 60-page manual. Well-written, it describes the program in a colloquial style. Much of the documentation is also found in the Help section of the software. However, the book could include more details. As mentioned, the M-Competition results should be explained. Also the book takes the approach that the user understands forecasting, and thus gives no guidance or instructions on the different techniques available. The data transformation section, for example, allows ten forms of moving averages but does not explain when you should or should not use each one. The documentation could also benefit by having a tutorial chapter.

Support and Pricing

Registered users get life-time free (but not toll-free) technical support. The single unit price of Sibyl/Runner is $495 for the first copy and additional copies are discounted 50%. There are substantial academic discounts and a student version for $40.

Conclusions

You should look elsewhere if you know little about forecasting and want an ``automatic'' product acting like a ``black box'' (it chooses the ``correct'' model and makes the forecast). However, I recommend it highly to those who know something about forecasting and want an easy-to-learn-and-use product that gives some guidance in choosing the appropriate model, has good graphics, and yields accurate results. For those people who were weaned on forecasting using the mainframe or even DOS version of Sibyl/Runner, they should get this upgrade because it is easier to use and has superior graphics.

Lincoln Systems Corporation
P.O. Box 391
Westford, MA 01886
508-692-3910

If you are interested in writing a software review for a future issue of , please call Professor Jack Yurkiewicz at Pace University (212) 346-1908, or e-mail: yurk@pacevm.dac. pace.edu.