Accuracy of forecasting methods

The accuracy of forecasting methods is one of the most important criteria for comparing alternative methods of forecasting. Cost, ease of application, and specific requirements of a planning situation are other factors that influence the choice of a forecasting method. It is difficult to determine which method will provide the most accurate forecast in a specific situation. However, over the years, considerable empirical evidence has been collected for both hypothetical and real world data that allow some general conclusions about the relative accuracy of various forecasting methods.

The most significant conclusion, supported by a number of studies, is that more sophisticated methods do not necessarily produce more accurate results than do simpler methods that are easier to apply and less costly to use. There is also solid support in the literature for the view that, in spite of their logical appeal, sophisticated causal models do not outperform time series models. Further, with exponential smoothing, simpler time series models often give results that compare favorably with more complex time series models.

In some important situations, more than one method of forecasting may seem appropriate. The question then is how to select a method for forecasting. Makridakis and Wikler (1983) empirically estimated the impact of the number and choice of forecasting methods on the accuracy of forecasts when the results of the methods used are simply averaged to provide the final forecast. Their main findings were as follows:

Forecasting accuracy improves as the forecasts from more methods are combined to provide the final forecast; however, the marginal impact of including an additional method decreases as the number of methods increase.

The risk of a large error in forecasting that might result from the choice of a wrong method is diminished when the results of two or more methods are combined.

Variability in forecast accuracy among different combinations of forecasting methods decreases as the number of methods increase. Thus, a practical alternative when we are unsure about the “best” method of forecasting is simply to take the average of the forecasts of two or more forecasting models.

Based on the empirical evidence reported in the literature, we can reasonably conclude that in production / inventory situations are characterized by a need for producing forecasts of thousands of items on a routine basis, exponential smoothing is the most cost effective method of forecasting. With the development of fast and user friendly computer programs, more sophisticated time series forecasting models may become practical for routine production / inventory situations.

Several other methods of forecasting are available, ranging from the relatively simple to the highly sophisticated. In this article we are trying to briefly discuss the conceptual bases of some of these methods.

Decomposition methods are extrapolative in nature, but they differ from exponential smoothing methods. The key difference is that instead of extrapolating a single pattern as is the case with exponential smoothing, each component of demand is extrapolated separately. Thus the extrapolation of the seasonal pattern, the trend pattern, and the cyclical pattern and the smoothing of randomness all take place separately. The forecast is obtained by combining these component patterns.

Box and Jenkins (1970) have proposed a framework for analyzing and modeling time series data. This framework is based on well developed statistical theory. It provides an approach for identifying patterns in data and a methodology for extrapolating these patterns into the future. The basic framework consists of three stages: identification, estimation, and diagnostic testing.

This method requires considerable expertise on the part of the analysts, and substantial analysis of the data must be performed before a forecasting model is chosen.

The Fourier series forecasting method represents time series using a mathematical function consisting of a constant term plus the sum of several sine and cosine terms. The method is useful when data have seasonal patterns.

Finally, the prediction of business cycles is important for decisions that will impact a company’s profitability for a long period of time. Measuring leading indicators is one useful technique for identifying business cycles or turning points in time series data. The idea is to identify those indicators, such as housing starts, money supply and durable new orders, whose present levels influence future business activity.