Machine Learning in Empirical Asset Pricing Models

Abstract

Although machine learning has achieved great success in computer science, its performance in the canonical problem of asset pricing in finance is yet to be fully investigated. To compare machine learning techniques and traditional models, we use 8 macroeconomic predictors and 102 firm characteristics to predict stock returns in a monthly basis. It is shown that the neural network outperforms others: Specifically, when building bottom-up portfolios based on the predicted stock-level returns for both buy-and-hold and long-short strategies, XGBoost and neural networks produce portfolios with the highest Sharpe ratios. Limitations and challenges in using machine learning techniques in empirical asset pricing models are also discussed.