Feature importance in linear models with ensemble machine learning: A study of the Fama and French five-factor model

This study explores key considerations for interpreting feature influence and importance in Machine Learning (ML) for financial models that commonly assume linearity. Simulations demonstrate that ML techniques, including Random Forest, XGBoost, and CatBoost, may produce misleading feature importance ranks when the underlying model is linear. We empirically examine the Fama–French five-factor model using U.S. monthly data from July 1964 to June 2024. While the most important factors are consistently identified, the ranks of moderately important factors vary depending on the estimation method. These results highlight the need for a critical application of ML in financial modeling when the purpose is interpretability.