Forecasting Equity Premia using Bayesian Dynamic Model Averaging

This paper introduces a Bayesian version for Dynamic Model Averaging for predicting aggregate stock returns. Our suggested approach simultaneously accounts for many sources of uncertainty. It is designed to handle (i) parameter instability, (ii) time-varying volatility, (iii) model uncertainty and (iv) time-varying model weights. We use our method to analyze predictability of S&P500 returns for the 1927 - 2012 period. The flexibility of the econometric setup enables us to disentangle the multitude of effects at work when generating (point and density) forecasts. A key point of our analysis is to assess which components of forecast models pay off in terms of statistical accuracy and economic value. We document that statistical and economic evaluation metrics can be in sharp contrast. While stochastic volatility emerges to be important both in terms of density forecast accuracy and economic gains, return prediction models that use economic covariates turned out to be helpful to time the market only within very limited periods of time.