The quality of a decision often depends on the quality of a forecast. Forecasting models are an integral part of many DSS. One can build a forecast model or one may use preprogrammed software packages.
The major use of forecasting is to predict the value of variables at some time in the future. The future time period of interest depends on "when" we want to evaluate the results. For example, in an investment decision we may be interested in prices and income a year from today, while in a capital investment decision we may be interested in projected prices and income during the next five years. Generally speaking, we distinguish between two types of forecasts: (a) short run, where the forecast is used mainly in deterministic models, and (b) long run, where the forecast is used in both deterministic and probabilistic models.
Many types of forecasting models exist, but forecasting remains an extremely difficult task (cf., Makridakis and Wheelwright, 1982). What is going to happen in the future depends on many factors that are uncontrollable. Furthermore, data availability, accuracy, cost, and the time required to make a forecast play an important role in choosing a forecasting method. We can select forecasting methods based on convenience, popularity, expert advice, and guidelines from prior research. In general the last two approaches should be used in building Forecasting DSS.
The best Web resource on Forecasting Models and Methods is the Forecasting Principles site (hops.wharton.upenn.edu/forecast/) maintained by J. Scott Armstrong. It provides a comprehensive review of our knowledge about forecasting. The site also provides: evidence showing the relevance of forecasting principles to a given problem, expert judgment about the applicability of forecasting principles, sources of data and forecasts, details about how to use forecasting methods, and guidance to locating the most recent research findings.
Forecasting methods can be grouped in several ways. One classification scheme distinguishes between formal forecasting techniques and informal approaches such as intuition, expert opinions, spur-of-the-moment guesses, and seat-of-the-pants predictions.
The following paragraphs review the more formal and analytical methods that have been used in building Forecasting DSS. The methods reviewed include naïve extrapolation, judgment methods, moving averages, exponential smoothing, time series extrapolation, and regression and econometric models. Each method is discussed briefly and major issues associated with using the methods are summarized. According to Scott Armstrong, given enough data, quantitative methods are more accurate than judgmental methods. He notes that when large changes are expected, causal methods are more accurate than naive methods. Also, simple methods are preferable to complex methods; they are easier to understand, less expensive, and seldom less accurate.
Naïve Extrapolation. This technique involves collecting data and developing a chart or graph of the data. The user extrapolates or estimates the data for future time periods. This technique is easy to update and minimal quantitative knowledge is needed. It is easy and inexpensive to implement using a spreadsheet. However it provides limited accuracy.
Judgment Methods. Judgment methods are based on subjective estimates and expert opinion, rather than on hard data. They are often used for long-range forecasts, especially where external factors may play a significant role. They also are used where historical data are very limited or nonexistent. A group DSS could be used with a judgment method like the Delphi technique to obtain judgments. The results are not necessarily accurate, but the experts may be the best source of forecast information.
Moving Average. This type of forecast uses an average of historical values that "moves" or includes the new period in each succeeding forecast. It is for short-run forecasts and the results are easy to manipulate and test. Overall, a Forecasting DSS built using a moving average model will be easy to understand and inexpensive.
Exponential Smoothing. The historical data is mathematically altered to better reflect the forecasterís assumptions about the future of the variable being forecast. This model is similar to the moving average model, but it is harder to explain. A short-term forecast based on exponential smoothing is often acceptable.
Time-series Extrapolation. A time series is a set of values for a business or economic variable measured at successive intervals of time. For example, quarterly sales of a firm make up a time series. Managers use time-series analysis in decision-making because they believe that knowledge of past behavior of the time series might help understand the behavior of the series in the future. In managerial planning we often assume that history will repeat itself and that past tendencies will continue. Time-series analysis efforts conclude with the development of a time-series forecasting model that can then be used to predict future events. Both moving average and exponential smoothing use a time series of data.
Regression and Econometric Models. Association or causal forecasting methods use data analysis tools like linear and multiple regression to find data associations and, if possible, cause and effect relationships. Causal methods are more powerful than time-series methods, but they are also more complex. Their complexity comes from two sources: First, they include more variables, some of which are external to the situation. Second, they use sophisticated statistical techniques for evaluating variables. Causal approaches are most appropriate for intermediate term (3-5 year) forecasting. An econometric model using simultaneous equations for a supply-demand system is x demand= f (x price, yield, etc...) and x supply = f (x price, production inputs prices, etc...) Econometric Resources on the Internet (www.oswego.edu/~kane/econometrics/) by John Kane contains links to a variety of resources. You can work with Fairmodel, a macroeconometric model of the USA economy, to forecast the economy and do policy analysis (fairmodel.econ.yale.edu).
In general, subjective forecasting methods are used in those cases where quantitative methods are inappropriate or cannot be used. Time pressure, lack of data, or lack of money may prevent the use of quantitative models. Complexity of historical data may also inhibit its use. Model-Driven DSS primarily incorporate quantitative methods and often use multiple forecasting models.