The use of prior information in forecast combination *

Simple averages often, but not always, outperform more sophisticated “optimal” forecast composites. We used Bayesian shrinkage techniques to allow the incorporation of prior information into the estimation of combining weights; the estimated combining weights were coaxed or “shrunken” toward equality but were not forced to be exactly equal. The least-squares and prior (i.e., arithmetic average) weights then emerged as polar cases for the posterior mean; the exact location depended on prior precision, which was estimated from the data. In a simple example involving U.S. GNP forecasts, a large amount of shrinkage was found to be optimal.