Casting Doubt on the Predictability of Stock Returns in Real Time: Bayesian Model Averaging using Realistic Priors*

Abstract

Previous studies have identified several variables that would have predicted future stock returns, though other studies suggest these results may be due to data snooping. To guard against data snooping, researchers have suggested use of Bayesian model averaging (BMA) to account for the uncertainty about prediction models. In common with other researchers, I find evidence of predictability during time periods when a hypothetical investor uses BMA with no restrictions on what variables may be included in the model. However, when the hypothetical investor is limited to using only variables whose predictive ability would have been known at the time of the forecast, predictability disappears. Moreover, predictability also disappears when data are updated through 2010, even without constraints on variable use. The results cast doubt on whether stock returns were ever predictable in real time and also suggest that returns may no longer be predictable even if real-time constraints are removed.