Identifying Factor Importance in Empirical Asset Pricing by Interpretable Machine Learning

We illustrate the use of interpretable machine learning to explain portfolio monthly returns. There are 188 financial anomalies from global-q.org in the analysis using artificial neural network models. We employ Shapley additive explanations (SHAP) and Shapley additive global importance (SAGE) to identify the importance of factors. SHAP is constructed from local observations. We use SAGE as global interpretability. These methods can show how important these factors are to explaining portfolios’ returns. The importance of the factors in explaining portfolio returns differs from the overall period to the subperiods. The top contributors can also depend on how the variable is defined in each category of factors. We also extend the analysis from monthly to daily returns, containing 900,520 observations.