Stock Asset Pricing Based on the Integration of Multifactor Models and Machine Learning

Abstract

This study examines whether combining traditional multifactor asset-pricing theory with machine-learning methods can improve the explanation and prediction of U.S. stock returns. The sample comprises ten large U.S.-listed firms across six sectors, with SPY as the market proxy. Monthly returns cover August 2023 through June 2026, providing 35 observations for the preliminary market-model analysis, while the underlying dataset also contains approximately three years of daily prices for the proposed machine-learning extension. Preliminary estimates reveal substantial heterogeneity: market betas range from 1.945 for NVIDIA to 0.117 for Johnson & Johnson, while R-squared ranges from 49.6% for Amazon to 0.6% for Johnson & Johnson. These differences indicate that a single market factor does not explain return variation equally across securities. The paper therefore develops a framework that uses the Fama-French five-factor model as an economically interpretable benchmark and Random Forest and Extreme Gradient Boosting (XGBoost) as nonlinear extensions. Factor selection and functional form are distinct: traditional factors capture economically meaningful risks, while machine learning tests nonlinear, interactive, and time-varying effects. However, additional complexity is justified only when it produces stable out-of-sample improvement. The study consequently treats interpretability, chronological validation, and resistance to overfitting as necessary conditions for accepting a hybrid asset-pricing model.