Using Machine Learning to Capture Nonlinear Equity Factor Effects

Summary

This study tests whether machine learning can explain stock returns left unexplained by a conventional linear equity factor model. It uses 22 style factor exposures to predict standardized stock specific returns, then evaluates boosted trees, random forests, and neural networks in a walk forward setup. Models are trained on five years of data and retrained at set intervals; their predictions are also averaged into ensembles. The study examines feature importance, partial dependence, and interactions to interpret the resulting signals.

The reported findings identify momentum and liquidity as influential inputs, with interactions among factors also contributing to predictions. Averaging models improves robustness, and the resulting machine learning factor has low average correlation with established style factors while adding explanatory power in the authors’ tests. The analysis covers global equities from 1995 to 2020, with a later period used for out of sample evaluation. Results are specific to the data, factor definitions, and validation design; the authors also emphasize the risk of fitting noise in low signal to noise return data.

Key ideas

• The model predicts standardized stock specific returns to focus on nonlinear effects omitted by a linear factor model.
• The study compares boosted trees, random forests, and neural networks using a walk forward training and evaluation process.
• Ensembling predictions improves stability and reduces sensitivity to model complexity and retraining dates.
• Momentum and liquidity rank highly in feature analysis, and factor interactions can materially affect predictions.
• The reported performance is tied to the study’s sample, equity model, and validation choices, and may not generalize.

Tags

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