Peer-Reviewed Theory Does Not Help Predict the Cross-section of Stock Returns

Abstract

To examine whether theory helps predict the cross-section of returns, we combine text analysis of publications with out-of-sample tests. Based on the original texts, only 18% of predictors are attributed to risk-based theory. 58% are attributed to mispricing, and 24% have uncertain origins. Post-publication, risk-based predictability decays by 65%, compared to 50% for non-risk predictors. Out-of-sample, risk-based predictors fail to outperform data-mined accounting predictors that are matched on in-sample summary statistics. Published and data-mined returns rise before in-sample periods end and fall out-of-sample at similar rates. Overall, peer-reviewed research adds little information about future mean returns above naive back testing.

Keywords: Cross-Section of Returns,Return Predictability, Machine Learning, Big Data