Combination of Forecasts An Extension.pdfVIP

MANAGEMENT SaENCE Vol. 33. No. 3, March 1987 PnmedinVSA. COMBINATION OF FORECASTS: AN EXTENSION* SUNIL GUPTA AND PETER C. WILTON Graduate School of Business, Columbia University, New York, New York 10027 School of Business Administration, University of California, 350 Barrows Hail, Berkeley, California 94720 Existing approaches to combining multiple forecasts generally ofTer either theoretical richness or empirical robustness, but not both together. In this paper, we propose a new method for combining forecasts which attempts to overcome this imbalance. The method allows easy inclusion of relevant subjective and empirical information about the forecasts, while providing weights which are: (i) intuitively meaningful, and (ii) not dependent upon large numbers of observations of prior forecast accuracy. Results of a simulation experiment show the method to be highly robust, and significantly superior to existing approaches under many conditions. (PLURAL ANALYSIS; WEIGHTING METHODS; OUTPERFORMANCE; ODDS- MATRIX) Introduction In many decision making situations, information is obtained from a variety of sources. Examples are subjective judgments from various experts, information from secondary documents, forecasts from different methods or models, or some combina- tion of these. Multiple sources are consulted because there is a belief that more infor- mation leads to better decisions. When these models are in disagreement, the decision maker faces the problem of either identifying the best model, or combining the forecasts fiom the various models into a single assessment. Since identifying the best model, a priori, is rarely feasible (Clemen and Winkler 1986, Makridakis and Winkler 1983, Schnaars 1984), considerable research attention has recently been focussed upon the utility of combining the model forecasts (see Mahmoud 1984 for an extensive review). Numerous approaches are available. The simplest is to take an (equally weighted) arithmetic average of the individual fo