Hollstein, Fabian

This paper evaluates the predictive performance of machine learning techniques in estimating time-varying betas of US stocks. Compared to established estimators, tree-based models and neural networks outperform from both a statistical and an economic perspective. Random forests perform the best overall. Machine learning-based estimators provide the lowest forecast errors. Moreover, unlike traditional approaches, they lead to truly ex-post market-neutral portfolios. The inherent model complexity is strongly time-varying. The most important predictors are various historical betas as well as fundamental turnover and size signals. Compared to linear regressions, interactions and nonlinear effects enhance the predictive performance substantially.

Machine learning-based market beta estimators outperform established benchmark models both statistically and economically. Analyzing the predictability of time-varying market betas of U.S. stocks, we show that machine learning-based estimators produce the lowest forecasting and hedging errors. They also help create better market-neutral anomaly strategies and minimum variance portfolios. Among the various techniques, random forests perform best overall. Model complexity is highly time-varying. Historical betas, turnover, and size are the most important predictors. Compared to linear regressions, allowing for nonlinearity and interactions significantly improves the predictive performance.