Identifying Factors via Automatic Debiased Machine Learning

Identifying factors that can explain cross-sectional asset returns is fundamental in asset pricing. We adopt a new automatic debiased machine learning method proposed by Chernozhukov, Newey, and Singh (2022) to robustly identify these factors and conduct statistical inference in a very high dimensional setting. Our general methodology avoids biased estimation, non-robustness, and overfitting common to traditional regularization approaches. We examine a large set of factors, and our most significant factors outperform the Fama-French factors at comparable parsimony. We consider nonlinear Stochastic Discount Factor (SDF) and other relationships and find that about 30-50 factors may explain asset returns, and predictability decreases after 2003. Linear factor models with a few factors cannot compete well. Our empirical findings are robust in both the U.S. and Chinese stock markets