You no longer need to be a statistician to generate sophisticated business forecasts
George Stewart
Software-based business forecasting is not new, but it has been limited to a select group of professionals who understand the methodologies and terminologies of forecasting. Now a new class of forecasting software incorporates the decision strategies used by statisticians while guiding nontechnical businesspeople through a forecasting session much as a statistical consultant would.
The four programs featured in this review--Forecast Pro for Windows, SmartForecast, Solo Statistical Software, and Autocast II--generate forecasts based on real-world data known as time series. A time series is a set of observations taken at regular intervals; for example, a retail-store chain's weekly sales, a family's monthly electr
icity usage, or the number of individuals applying for unemployment benefits each quarter. Using such data, the forecasting programs predict what will happen in the near future: expected sales, next month's electricity usage, or future unemployment figures.
In this review, I evaluate products that focus primarily on time-series forecasting and on making forecasting available to the lay person. I have not included large, general-purpose statistical software.
The Nuts and Bolts
The first step in the statistical forecasting process involves collecting the time-series data. Depending on the circumstances, you may need as few as five or six annual observations or as many as several thousand, taken at regular intervals. The data may already exist in a computer database or spreadsheet, or it can be entered manually into the program itself.
All the programs accept univariate data (i.e., data involving a single variable). Some of the programs accept multivariate data. An example of multivariate
data would be a time series showing average daily electricity usage and average daily temperature. Multivariate data can treat one independent variable as a leading indicator for another, dependent, variable.
The second step in forecasting is to analyze the time series. The programs apply a battery of statistical tests to the data, providing you with an array of graphical and numerical information. This information helps you determine whether the data contains certain characteristics relevant to forecasting: fluctuation around a constant level, a rising or falling trend, a seasonal pattern, or a nonseasonal cycle. You might also discover irregularities in the data that could have an impact on the forecast's reliability.
Using information from the data analysis, you select a mathematical model for the time series. Forecasting software selects or recommends the model to use and then implements the selected model. You never have to deal directly with the formulas.
With the model selected, y
ou generate a forecast for a limited number of observation periods. Forecasts are typically limited to one cycle of observation periods (i.e., 12 months, four quarters, or 52 weeks into the future). However, where the time series consists of annual observations made over several decades, the forecast horizon might include up to 10 annual predictions. Forecast validity decreases rapidly as the forecast is extended beyond one cycle of observation periods.
A final, essential step in the process is attaching levels of confidence to the forecast. In addition to predicting specific numbers, the forecast must estimate the probability that the actual data will be above or below this prediction. Typically, the forecast specifies a 95 percent confidence level associated with each prediction (i.e., a range within which the actual data is expected to fall with a 95 percent probability).
For practical purposes, this confidence level is more important than the actual predicted value. For example, to avoid run
ning out of stock, a purchasing agent using a forecast to predict inventory requirements may find it prudent to order sufficient quantities to satisfy the 95 percent confidence level, rather than simply ordering the most likely forecast.
In evaluating and comparing the programs, I have considered the following characteristics: variety of forecasting methodologies offered; ease of use; quality of output, both numerical and graphical; degree of expert assistance provided; and flexibility or ability to customize the forecasting methods. I ran all the programs on a 66-MHz 486 Gateway system (the 486 has an on-chip math coprocessor). Even when processing a time series of several hundred observations, results came back almost instantaneously.
FORECAST PRO FOR WINDOWS
Forecast Pro for Windows, from Business Forecast Systems, is the only reviewed package that runs as a full Windows application, so it provides some built-in advantages--such as DDE links between Forecast Pro and your other Windows appl
ications. Three of Forecast Pro's supported methodologies are univariate (i.e., they look only at the behavior of a single variable). Dynamic regression calculates the relationship between a dependent variable and its historical values as well as the historical values of one or more independent variables.
Forecast Pro's audit-trail window operates as a notepad for automatic storage of the statistical data and expert advice generated by the program during the entire session. The notepad can be edited, stored, or printed. This important feature has no exact counterpart in the other programs.
To begin working with data, you click on the Tableau tool. The data tableau is where you specify the variables (univariate time series) you want to forecast or analyze. Available variables are displayed in a list box. If you select more than one variable, the first variable selected in a given tableau is treated as the dependent variable, and subsequent variables are treated as independent. After selecting one
or more variables, you specify the portion of the time series to be considered and the number of forecasts desired. The program allows you to "hold out" a given number of observations from the end of the time series; you can then generate a forecast for the hold-out period and compare the forecast with what actually occurred in the period.
The next step is to view the data graphically. You can look at a plot of all the variables selected in the tableau. The graphs help you identify characteristics of your variables and relationships between them. You can also adjust the data by pointing to a location on the graph and dragging the value up or down. This shortcut lets you get rid of irregularities in the data that you do not want included in the forecast--for example, you might want to smooth out a "blip" caused by a freak occurrence.
To pick a forecast methodology, you select the Expert tool, which automatically performs a battery of data exploration tests, identifies the characteristics of the
series, and recommends a forecasting model. The Forecast tool generates the forecast according to the selected model. You can then copy the forecast data to the audit trail and plot the forecast, along with the confidence limits.
You can save forecasts and recall them later for comparison with subsequent forecasts using different models. For instance, having forecast the variable KWH (kilowatt-hours) using Box-Jenkins, you might then forecast KWH using exponential smoothing and compare the two forecasts by recalling the Box-Jenkins model.
When you specify your own model (rather than relying on the Expert recommendation), Forecast Pro gives you plenty of flexibility in combining the applicable parameters. For example, under exponential smoothing, you can select simple, Holt, Winters, or a Custom model. The Custom model lets you specify the trend (none, linear, or damped) and seasonality (none, multiplicative, or additive).
Forecast Pro also includes a batch mode for forecasts (univariate o
nly) of up to 50 time series without need for your intervention. Dynamic regression is not available in batch mode.
Forecast Pro offers only rudimentary facilities for entering and editing time series. For instance, if you inadvertently omit an observation from the middle of a series, you can't insert the number into the series; you must retype the series beginning with the omitted value. You'll probably choose to enter your data from a spreadsheet, editor, or database application.
Overall, the program and its documentation are designed to make forecasting an interesting and enjoyable process for the total beginner. At the same time, the program has the power and flexibility a more experienced forecaster would require.
SmartForecast
SmartForecast opens with a spreadsheet window and a menu. Each column in the spreadsheet corresponds to a variable, and each row to an observation. You can enter data directly into the spreadsheet or load it from a disk file. As with Forecast Pro, this prog
ram automatically labels the rows according to the time of the observation. Data entry and editing under SmartForecast are particularly convenient. For example, you can insert a value in the middle of a time series, and you can define new variables by transformations on existing variables.
Data analysis is straightforward under SmartForecast via the Explore mode. For instance, the Correlate option works on a single variable to show autocorrelations (e.g., seasonality), or on two variables to determine whether one variable may be a leading indicator of the other. Results of the computations can be copied into a new variable in the program's built-in spreadsheet. The upper and lower confidence limits, the forecast, and the fit (i.e., values of the model during the base period) can be saved individually to the spreadsheet. You can output the tabular results to a printer, but, regrettably, you cannot save them to disk (compare this with Forecast Pro, which keeps all the statistical reports in an audit trai
l).
Having completed the analysis phase, SmartForecast selects a model or lets you choose your own. As with Forecast Pro, you can specify the applicable parameters for each methodology chosen. The Multiseries option produces automatic forecasts of up to 60 related variables and a total of these forecasts. This is a convenient feature for forecasting sales of an entire product line, for instance.
The Eyeball option lets you adjust forecasts manually using "business judgment" or "management overrides." The feature is intended to give you greater flexibility and control, but it seems ill-conceived. The program does not document the manual changes made in the forecast graph, and a person looking at the forecast after the fact could get the mistaken impression that the graph was generated by objective methods.
SmartForecast is a comprehensive forecasting program with an intuitive keyboard interface (i.e., single-key command menus and submenus). You can maneuver quickly through the program to g
enerate automatic or customized forecasts. However, it lacks the design elegance of Forecast Pro. For example, there's no convenient way to export tabular statistical results to disk for further incorporation into word processing documents. The program also lacks a convenient way of holding out data to verify forecasting accuracy. (You can, however, perform hold-outs indirectly by copying the time series into a new variable and inserting "missing" values for the period to be held out.)
The Batch Processing Edition of SmartForecast lets you make forecasts involving hundreds of individual data files and thousands of time series, using automatic model selection or a model you specify in advance.
SOLO STATISTICAL SOFTWARE
This DOS-based package from BMDP Statistical Software consists of a Base System plus optional modules, including the Forecasting and Time Series module reviewed here. The Base System offers simple moving averages and exponential smoothing. The FTS module adds Box-Jenkins and har
monic regression methodologies.
From Solo's Transfer menu, you select the Base System functions as well as add-on modules such as FTS. The program contains a spreadsheet with editing capabilities, such as selective copying and erasing of observations. Unlike Forecast Pro and SmartForecast, this program does not have a facility for automatically labeling observations (rows) by the date of observation; you must do this manually. And even if you take the trouble to label rows according to the time, dates do not appear in the output; the program simply references the time variable by the row number in which the data appears. Solo does include a number of transformation functions and keyboard macros to automate repetitive tasks.
To begin data analysis in Solo, you select the Time Series menu. Three of the options facilitate manual analysis without expert recommendations. Another option performs classical forecasting (i.e., linear trend, single and double exponential smoothing, season plus trend multi
plicative, and season plus trend additive). In addition to specifying a forecasting method, you can specify the values of the smoothing constants. Only the Box-Jenkins forecasts are accompanied by explicit confidence limits.
The forecast output appears as a graph, where observations are discrete points and the forecast is shown as a continuous line. This makes it somewhat difficult to visually gauge the goodness of fit between the model and the historical data. On the plus side, Solo lets you designate a report file to contain all statistical output from the program.
Compared to Forecast Pro and SmartForecast, Solo is difficult to learn and requires a greater level of knowledge about statistics.
AUTOCAST II
Autocast II, a DOS-based program from Delphus, provides single-variable forecasting based on exponential smoothing. It is the simplest to use of the four programs reviewed. And because of the general applicability of exponential smoothing, the program is fit for a wide range of uses
.
For inputting data, Autocast II includes a single-variable "worksheet" for entering a time series. You can also import data created with Lotus 1-2-3 or other programs, so long as the data is properly formatted. Associated with each Autocast II data file is a short description of the data, the season length, and the beginning date of the time series. Therefore, it is not necessary to manually specify the dates of each individual observation, as with Solo. Autocast II's data editor also lets you insert and delete records, which is handy when you inadvertently omit an observation in the middle of a series.
After you have entered or loaded a time series, you are ready to have Autocast II prepare an initial automatic forecast, based on its own expert estimation of the most appropriate exponential smoothing model to use. It is appropriate to go directly to automatic forecasting in this program, skipping the model-selection stage, since all the modeling is done with exponential smoothing.
On t
he graph, historical data appears as a solid line, and forecast and modeled values appear as discrete points. Confidence limits are shown with solid lines. The use of discrete points to show the modeled data makes comparison with historical data a little difficult (as with Solo).
For refinement of the model or possible selection of an alternative, you use the Analysis menu. You compare the plot to a set of forecast profiles (graphs of the profiles appear in the manual). Such a comparison could suggest, for example, that a particular time series exhibits multiple seasonality with a linear trend.
For further assistance in selecting a model, you can examine variance, autocorrelation, and the use of transformations to increase the stationarity of the data (i.e., remove long-term trend effects). The instructions for evaluating this information are simple and to the point. The variance and autocorrelation options do not change anything, but they give you information useful in making a manual model sel
ection. The autocorrelation option lets you perform indicated transformations on the data. Multiple transformations are allowed.
The program makes it so easy to modify the forecasting model that you can quickly lose track of the current model, so it's essential that you print a model-fitting report with each forecast. Unfortunately, the program does not allow you to store the model-fitting reports to disk. As a further means of testing the validity of a model, Autocast II lets you hold back part of the historical data, build a model using the remaining data, and compare the forecast with what actually occurred. As with Forecast Pro and SmartForecast, Autocast II offers a batch-processing module to allow automatic forecasting of multiple time series.
If you are looking for an easy-to-use forecasting program and your data does not require Box-Jenkins methodology, Autocast II is a good choice.
Caveat Forecasters
All four programs reviewed here offer expert-system forecasting for nonexper
ts. With the exception of Solo, all the programs are well suited to the lay person who wants to do statistical forecasting of time series without becoming too deeply involved in statistics.
Forecast Pro wraps all the essential methodologies and forecasting features into an elegant Windows interface. It is the only program reviewed that provides an efficient audit trail and DDE links to Windows spreadsheets and word processors. Forecast Pro strikes an optimal mix between the ease-of-use features that a novice forecaster seeks and the power and flexibility an experienced forecaster requires.
The Facts
Autocast II 1.50 $349
Delphus, Inc.
103 Washington St., Suite 348
Morristown, NJ 07960
(201) 267-9269
fax: (201) 285-1228
Forecast Pro for Windows
Standard edition $595
Extended edition $995
Business Forecast Systems, Inc.
68 Leonard St.
Belmont, MA 02178
(617) 484-5050
fax: (617) 484-9219
SmartForecast
Standard edition 3.0 $595
Smart Software, Inc.
4 Hill Rd.
Belmont, MA 02178
(617) 489-2743
fax: (617) 489-2748
Solo Statistical Software
Base System (includes graphics) $195
Time Series and Forecasting module $95
BMDP Statistical Software, Inc.
1440 Sepulveda Blvd., Suite 386
Los Angeles, CA 90025
(310) 479-7799
fax: (310) 312-0161
Forecast Pro for Windows
Runs under Windows 3.x, NT, or OS/2 2.x
Supports up to 50 variables
(Extended edition supports 100 variables)
Methodologies
--Simple moving averages
--Exponential smoothing
--Box-Jenkins
--Dynamic regression
Forecast Pro wraps all the essential methodologies and forecasting features into an elegant Windows interface. It provides an efficient audit trail and DDE links to Windows spreadsheets and word processors. Shown here is a graph of the forecasted values along with confidence intervals (displayed on the right).
SmartForecast
Runs under DOS 3.0 or higher
Supports up to 60 variables
(260 observations per variable)
Methodologies
--Simple moving averages
--Exponential smoothing
--Box-Jenkins
--Dynamic regression
SmartForecast is a comprehensive forecasting program with an intuitive keyboard interface, but it lacks the design elegance of Forecast Pro. With its graphical display of multivariate data, SmartForecast highlights data dependencies.
Solo
Base System 4.0 with Time Series and Forecasting module
Runs on DOS 3.x
Supports up to 500 variables (32,000 observations total)
Methodologies
--Simple moving averages
--Exponential smoothing
--Box-Jenkins
--Harmonic regression
Solo graphs observations as discrete points and the forecast as a continuous line. This makes it somewhat difficult to gauge the goodness of fit between the model and the historical data. Compared to Forecas
t Pro and SmartForecast, Solo is difficult to learn and requires a greater level of knowledge about statistics.
Autocast II
Runs on DOS 3.x
Univariate (one variable only)
Supports up to 500 observations
Methodologies
--Exponential smoothing
--4CAST/2 from Delphus, a separate product not reviewed here,
adds other methodologies, including Box-Jenkins
Autocast II provides single-variable forecasting based on exponential smoothing. It can display data, a graphical depiction of the forecast model, and confidence intervals.
George Stewart is a former BYTE editor living in Hancock, New Hampshire. He has a B.A. in mathematics and a degree in law. He can be reached on the Internet or BIX c/o
editors@bix.com
.
Forecasting the Future
