Model Averaging in Markov-Switching Models: Predicting National Recessions with Regional Data

This paper introduces new weighting schemes for model averaging when one is interested in combining discrete forecasts from competing Markov-switching models. In particular, we extend two existing classes of combination schemes - Bayesian (static) model averaging and dynamic model averaging - so as to explicitly reflect the objective of forecasting a discrete outcome. Both simulation and empirical exercises show that our new combination schemes outperform competing combination schemes in terms of forecasting accuracy. In the empirical application, we estimate and forecast U.S. business cycle turning points with state-level employment data. We find that forecasts obtained with our best combination scheme provide timely updates of the U.S. business cycles.