Forecasting and Conditional Projection Using Realistic Prior Distributions

Abstract

This paper develops a forecasting procedure based on a Bayesian method for estimating vector autoregressions. We apply the procedure to 10 macroeconomic variables and show that it produces more accurate out-of-sample forecasts than univariate equations do. Although cross-variable responses are damped by the prior, our estimates capture considerable interaction among the variables. We provide unconditional forecasts as of 1982:12 and 1983:3. We also describe how a model such as this can be used to make conditional projections and analyze policy alternatives. As an example, we analyze a Congressional Budget Office forecast made in 1982:12. While no automatic casual interpretations arise from models like ours, such models provide a detailed characterization of the dynamic statistical interdependence of a set of economic variables. That information may help evaluate casual hypotheses without containing any such hypotheses.