Model uncertainty and expected return proxies

Over the last two decades, alternative expected return proxies have been proposed with substantially lower variation than realized returns. This helped to reduce parameter uncertainty and to identify many seemingly robust relations between expected returns and variables of interest, which would have gone unnoticed with the use of realized returns. In this study, I argue that these findings could be spurious due to the ignorance of model uncertainty: because a researcher does not know which of the many proposed proxies is measured with the least error, any inference conditional on only one proxy can lead to overconfident decisions. As a solution, I introduce a Bayesian model averaging (BMA) framework to directly incorporate model uncertainty into the statistical analysis. I employ this approach to three examples from the implied cost of capital (ICC) literature and show that the incorporation of model uncertainty can severely widen the coverage regions, thereby leveling the playing field between realized returns and alternative expected return proxies.