Learning, Forecasting and Structural Breaks

The literature on structural breaks focuses on ex post identification of break points that may have occurred in the past. While this question is important, a more challenging problem facing econometricians is to provide forecasts when the data generating process is unstable. The purpose of this paper is to provide a general methodology for forecasting in the presence of model instability. We make no assumptions on the number of break points or the law of motion governing parameter changes. Our approach makes use of Bayesian methods of model comparison and learning in order to provide an optimal predictive density from which forecasts can be derived. Estimates for the posterior probability that a break occurred at a particular point in the sample are generated as a byproduct of our procedure. We discuss the importance of using priors that accurately reflect the econometrician’s opinions as to what constitutes a plausible forecast. Several applications to macroeconomic time-series data demonstrate the usefulness of our procedure.