Dissecting the Factor Zoo : A Correlation-Robust Machine Learning Approach

This paper sheds light on a new perspective of the "factor zoo enigma", in which factor correlation prevails and distorts inferences from standard approaches such as Fama-MacBech regression and the LASSO shrinkage method. Subsequently, I consider a newly developed machine learning method to dissect this chaotic factor zoo: the Ordered-Weighted-LASSO (OWL) estimator, which circumvents complications from correlations and can identify correlated factors while shrinking off useless/redundant ones. I derive the grouping property which quantifies the condition when correlated factors will be identified. Then I further develop the asymptotic properties, including the oracle inequality (error bounds) and convergence rate for the OWL estimator. Monte Carlo experiments show that OWL outperforms LASSO, adaptive LASSO and Elastic Net in various settings, particularly when factors are highly correlated. Empirically, I apply the OWL estimator on 80 anomaly factors to find useful factors that jointly contribute to the cross-section of stock returns, and find that liquidity, momentum and profitability related factors are primary to drive asset prices. Out-of-sample analysis confirms superior performance of OWL-hedged portfolios compared with LASSO, Elastic-Net and Fama-MacBeth regression