Why do pooled forecasts do better than individual forecasts ex post?

Pooled forecasts frequently outperform individual forecasts of economic time series. This paper shows that the introduction of model uncertainty into the formation of expectations can account for the regularity. We conjecture that agents learn in a Bayesian way, using an optimally designed combination of forecasts to form expectations. When these expectations alter the ex-post realization of the data generating mechanism the pooled forecast may dominate the best individual device.