A Comprehensive Dynamic Bayesian Model Combination Approach to Forecasting Equity Premia

Abstract

We introduce a novel dynamic Bayesian model combination approach for predicting aggregate stock returns. Our method involves combining predictive densities in a data-adaptive fashion and simultaneously features (i) uncertainty about relevant predictor variables, (ii) parameter instability, (iii) time-varying volatility, (iv) time-varying model weights and (v) multivariate information. We analyze the predictability of monthly S&P 500 returns and disentangle which components of prediction models pay off in terms of statistical accuracy and economic utility. As a key feature of our approach, we formally address the (possibly) diminishing relevance of past information over time. The flexibility embedded in our approach enhances density forecasting accuracy and provides sizeable economic utility gains. We find predictability to be strongly tied to business cycle fluctuations and document disagreement between statistical and economic metrics of forecast performance.

Keywords: Equity premium prediction; Density forecasting; Bayesian Analysis; Forecast combination; Model uncertainty; Time-varying parameter models; Stochastic volatility; Asset allocation; Asset pricing

JEL Classification: C11, G11, G17