Tizian Otto, University of Hamburg (visiting Stanford University)
This paper evaluates the predictive performance of machine learning techniques in estimating time-varying market betas of U.S. stocks. Compared to established estimators, machine learning-based approaches outperform from both a statistical and an economic perspective. They provide the lowest forecast errors and lead to truly ex-post market-neutral portfolios. Among the different techniques, random forests perform the best overall. Moreover, the inherent model complexity is strongly time-varying. Historical betas, as well as turnover and size signals, are the most important predictors. Compared to linear regressions, interactions and nonlinear effects substantially enhance predictive performance.