Explainable AI in Corporate Finance: A SHAP-Based Approach to Enhancing Transparency

Abstract

The increasing adoption of machine learning models in corporate valuation has substantially improved predictive accuracy but at the cost of interpretability, a critical limitation in regulated financial environments. This study investigates whether SHAP (Shapley Additive Explanations) can systematically enhance the transparency of XGBoost based valuation models and examines whether the resulting insights extend those of classical linear regression. An empirical analysis was conducted on a cross-sectional dataset of U.S. publicly listed firms (2018), including more than 200 financial indicators. After systematic preprocessing and a hybrid feature selection procedure combining XGBoost importance, mutual information, and correlation based filtering, both an OLS regression and an XGBoost model were trained and validated. XGBoost achieved substantially higher predictive performance (R2 = 0.654 vs. 0.364), while SHAP values provided transparent global and local explanations of model decisions. Both models consistently identified earnings before tax and EV to sales as primary value drivers; however, SHAP additionally showed nonlinear effects, threshold behaviors, and context dependent interactions, particularly for EBITDA margin, share buybacks, and asset based indicators, that remain undetectable in linear models. These findings confirm that SHAP significantly enhances model transparency and generates economically meaningful insights beyond classical regression, supporting its application in auditable, regulatory compliant financial modeling.