Prediction with Misspecified Models

Abstract

The assumption that one of a set of prediction models is a literal description of reality formally underlies many formal econometric methods, including Bayesian model averaging and most approaches to model selection. Prediction pooling does not invoke this assumption and leads to predictions that improve on those based on Bayesian model averaging, as assessed by the log predictive score. The paper shows that the improvement is substantial using a pool consisting of a dynamic stochastic general equilibrium model, a vector autoregression, and a dynamic factor model, in conjunction with standard US postwar quarterly macroeconomic time series.