Machine Learning Models for Cross-Sectional Alpha Prediction

Summary

The report compares machine learning approaches with traditional linear methods for predicting stock-level excess returns from alpha factors. Rather than predict excess return directly, it forecasts return dispersion with an AR(1) model and predicts cross-sectional standardized returns from factors. The study uses 51 alpha factors, monthly cross-sectional training, and an out-of-sample period spanning 2009 through 2018; it evaluates forecasts using out-of-sample R-squared and Diebold–Mariano tests.

The reported findings favor nonlinear models overall, particularly gradient-boosted trees and random forests, while dimensionality reduction and Elastic Net regularization improve linear models. Averaging forecasts across models also performs better than relying on one model. However, forecast accuracy does not always translate into portfolio returns under stock-weight constraints. The report says technical-factor preference can raise turnover, making trading-cost control important; its findings are tied to the tested data, factors, and portfolio setup, and the supplied text does not include detailed tables or numerical performance results.

Key ideas

• The study predicts return dispersion and cross-sectional standardized returns as separate components.
• It compares 17 models using 51 alpha factors and out-of-sample forecast tests.
• Nonlinear tree models and simple combinations of model forecasts are reported as strong approaches.
• Factor preprocessing and regularization can improve linear prediction accuracy.
• Portfolio constraints and elevated turnover can weaken the link between forecast accuracy and realized returns.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.