Two-Stage Model Averaging: An Application to Forecasting Stock Returns

Abstract

This paper develops two-stage model averaging (2SMA), an extension of model averaging. 2SMA allows researchers to incorporate economic theory or prior economic information into their forecasts. By using prior economic information, 2SMA leverages models known to perform well while diversifying across peripheral forecasts. 2SMA is an easy-to-implement, general framework that can be easily combined with other model averaging techniques. In an application, we develop two-stage Bayesian model averaging (2SBMA) and two-stage equal-weighted averaging (2SEWA). 2SBMA and 2SEWA are the two-stage model averaging extensions of dynamic Bayesian model averaging and equal-weighted averaging. Using these techniques, we forecast stock returns. Our results indicate that the 2SBMA and the 2SEWA forecasts statistically outperform the benchmark random-walk plus drift over the sample period. The 2SBMA and 2SEWA forecasts also both beat their traditional model averaging counterparts.