2023 INFORMS Annual Meeting Poster Session

Abstract

This study introduces an innovative framework to interpret the behaviors of firm characteristics in predicting expected returns through machine learning models, directly addressing the challenges of transparency and interpretability. Our approach utilizes the Local Interpretable Model-Agnostic Explanations (LIME) to evaluate firm characteristics based on their statistical significance and behaviors—linearity, independence, insignificance, and interaction—offering a novel perspective on their predictive roles. Empirical findings demonstrate a complex interplay among these behaviors, with interaction effects playing a pivotal role, thus challenging the traditional emphasis on linear and independent influences in asset pricing models. Our research provides new insights into the mechanisms of machine learning predictions in asset pricing, paving the way for further exploration into the economic rationale behind data-driven findings and enhancing understanding of complex asset pricing dynamics.

Zequn Li

PhD Candidate

My research interests include empirical asset pricing and interpretable machine learning.