Stock market anomalies and machine learning across the globe

Abstract

We identify the characteristics and specifications that drive the out-of-sample performance of machine-learning models across an international data sample of nearly 1.9 billion stock-month-anomaly observations from 1980 to 2019. We demonstrate significant monthly value-weighted (long-short) returns of around 1.8–2.2%, and a vast majority of tested models outperform a linear combination of predictors (our baseline factor benchmark) by a substantial margin. Composite predictors based on machine learning have long-short portfolio returns that remain significant even with transaction costs up to 300 basis points. By comparing 46 variations of machine-learning models, we find that the models with the highest return predictability apply a feed-forward neural network or composite predictors, with extending rolling windows, including elastic net as a feature reduction, and using percent ranked returns as a target. The results of our nonlinear models are significant across several classical asset pricing models and uncover market inefficiencies that challenge current asset pricing theories in international markets.