The Accuracy of Extrapolative Forecasting Methods: Additional Empirical Evidence

Also INSEAD Working Paper N 95/04/TM Replication is an important aspect of hard sciences in developing an objective knowledge base that the great majority of scientists can accept. In social sciences, however, replication has proved problematic. One of the exceptions has been the M-Competition whose calculations and conclusions have been replicated by many researchers. In order to extend the range under which exptrapolative methods have compared this study examines a unique set of 263 series. Unlike the M-Competition data they come from a single source containing a series of similar characteristics (the same calendar, no seasonality, little randomness and downward sloping trend). The study shows that the conclusions of the M-Competition are still valid while at the same time it is found that Robust-Trend, a new method not used in the M-Competition, outperforms all other methods. In addition, the relative performance of exponential smoothing is shown to depend on the ways its parameters are optimised. The implications of these conclusions are discussed.