Gu-Kelly-Xiu (2020) Empirical Asset Pricing via Machine Learning

Summary

Gu, Kelly & Xiu perform a comparative analysis of machine-learning methods for the canonical problem of empirical asset pricing — measuring asset risk premiums (the cross-section of expected returns). They demonstrate large economic gains to investors from machine-learning forecasts, in some cases doubling the performance of leading regression-based strategies. The best-performing methods are trees and neural networks, whose predictive gains are traced to allowing nonlinear predictor interactions missed by other methods; all methods agree on the same dominant signals — variations on momentum, liquidity, and volatility.

Key Claims

• Comparative study. A like-for-like out-of-sample comparison across OLS, penalized linear (elastic net), dimension reduction (PCR, PLS), generalized linear (splines), random forests, gradient-boosted trees, and neural networks, on a large panel (~30,000 stocks, ~90 characteristics interacted with macro predictors and industry dummies, 1957–2016).
• Economic gains. Machine-learning forecasts generate large out-of-sample economic value, in some cases doubling the Sharpe ratio of leading regression-based strategies.
• Trees and neural networks are best. The top methods are tree ensembles and neural networks; their advantage comes from capturing nonlinear predictor interactions that linear and dimension-reduction models cannot.
• Consensus on predictors. Despite different functional forms, the methods agree on the same dominant predictive signals — momentum, liquidity, and volatility variants — assessed via predictive-$R^2$ reduction and sum-of-squared-derivative / mean-decrease-in-impurity variable importance.
• Discipline against overfitting. Regularization and out-of-sample validation are central to obtaining honest performance in the high-dimensional setting.

Concepts Introduced or Extended

• Machine Learning in Asset Pricing
• Factor Zoo

Entities Mentioned

• Shihao Gu
• Bryan Kelly
• Dacheng Xiu

Quotes

"We demonstrate large economic gains to investors using machine learning forecasts, in some cases doubling the performance of leading regression-based strategies from the literature. We identify the best-performing methods (trees and neural networks) and trace their predictive gains to allowing nonlinear predictor interactions missed by other methods."

My Take

The paper that made machine learning a first-class citizen of empirical asset pricing, and did so carefully: a controlled horse race with honest out-of-sample metrics rather than a single flashy model. Its most durable message is why the flexible methods win — nonlinear interactions among characteristics, not exotic new signals, since every method fingers the same momentum/liquidity/volatility core. That pairs it naturally with Feng-Giglio-Xiu (which prunes the factor list with valid inference) as the two complementary answers to high-dimensional return prediction. The honest caveats are the field's perennial ones — weak, time-varying predictability and black-box interpretability — but the regularize-and-validate template it set is now standard, and it seeds the latent-factor/SDF machine-learning work that followed.