Stock Return Prediction with Fully Flexible Models and Coefficients

Abstract

We evaluate stock return predictability using a fully flexible Bayesian framework, which explicitly allows for different degrees of time-variation in coefficients and in forecasting models. We believe that asset return predictability can evolve quickly or slowly, based upon market conditions, and we should account for this. Our approach has superior out-of-sample predictive performance compared to the historical mean, from a statistical and economic perspective. We also find that our model statistically dominates its nested models, including models in which parameters evolve at a constant rate. By decomposing sources of prediction uncertainty into five parts, we find that our fully flexible approach more precisely identifies time-variation in coefficients and in forecasting models, leading to mitigation of estimation risk and forecasting improvements. Finally, we relate predictability to the business cycle.