Beta Estimation Under Infrequent Trading: A Machine Learning Approach

ABSTRACT

When shares are traded infrequently, beta estimates are often severely biased. We find that machine learning methods significantly improve forecasts of the conventional beta proxy in this infrequently traded market. They generate superior beta forecasts, statistically and economically outperforming the traditional model used by practitioners. Among machine learning models, neural network estimation performs best, improving precision by up to 43% over the practitioner counterpart. Market‐neutral momentum and turnover strategies using machine‐learning betas outperform traditional betas. Machine‐learning beta estimation also provides more accurate beta forecasts during economic crises when predicting beta is particularly difficult. These findings offer practitioners and regulators more accurate forecasts of the conventional beta proxy, a key input to pricing systematic risk and estimating the cost of equity.