Identifying factors via automatic debiased machine learning

Abstract

Identifying risk factors that have significant explanatory power for the cross-sectional asset returns is fundamental in asset pricing. We adopt a novel automatic debiased machine learning (ADML) method proposed by Chernozhukov, Newey, and Singh (2022) to robustly estimate partial pricing effect of a certain factor controlling for a large number of confounding factors under a nonlinear stochastic discount factor (SDF) assumption. The ADML resolves biased estimation, non-robustness, and overfitting issues that are common to traditional machine learning approaches. We find that the most significant factors selected by the ADML outperform the Fama–French sparse factors and factors identified via the double-selection LASSO method under a linear factor model assumption. Out of a high-dimensional zoo of US stock market factors commonly tested in the finance literature, we identify approximately 30 to 50 factors having significant but declining pricing power in explaining the cross-section of stock returns. Our findings are robust to hyperparameter settings and choices of test assets and machine learning methods.