The Promises and Pitfalls of Machine Learning for Predicting Stock Returns

Abstract

Recent research suggests that machine learning models dominate traditional linear models in predicting cross-sectional stock returns. The authors confirm this finding when predicting one-month-forward-looking returns based on a set of common stock characteristics, including predictors such as short-term reversal. Despite the statistical advantage of machine learning model predictions, the authors demonstrate that the economic gains tend to be more limited and critically dependent on the ability to take risk and implement trades efficiently. Unlike traditional models, machine learning models have been somewhat more effective over the past decade at discerning valuable predictions from cross-sectional equity characteristics.

TOPICS: Security Analysis and Valuation, big data/machine learning

Key Findings

• ▪ The authors compare a nonlinear machine learning model called gradient boosting machine (GBM) with traditional linear models in predicting cross-sectional stock returns based on well-known equity characteristics.

• ▪ They demonstrate how to rationalize the mechanics and outcome of GBM to alleviate its black-box characteristics.

• ▪ The extent to which the statistical advantage of GBM’s performance over that of linear models can be translated into economic gains depends critically on one’s ability to take risk and implement trades efficiently.