AI era in Finance Symposium

Our study introduces a novel framework to interpret machine learning asset pricing models through the Local Interpretable Model-agnostic Explanations (LIME) method. This methodology illuminates how the inclusion of LIME local coefficients, representing the interaction among characteristics within ML models, modifies the relationship between a firm characteristic and stock returns. The empirical results underscore the significance of incorporating moderation effects into portfolio analysis. Our results present that certain firm characteristics exhibit varying long-short portfolio performance across LIME groups, suggesting their predictive power is specific to certain asset segments. These findings deepen our understanding of the complexities in cross-sectional stock returns, uncovering the detailed dynamics between firm characteristics and their return effects, and distinguishing our research from existing studies.