Interpretable ESG risk analysis using decision trees and SHAP in applied machine learning

ESG classifications are commonly used for screening, but dependence on rating-provider data and repeated firm observations can lead to overly optimistic estimates. We examine 2,313 firm-year records for 488 Fortune 500 companies (2017–2021) using decision trees, SHapley Additive exPlanations (SHAP) and four comparator models. The outcome is a relative high-risk indicator derived from the upper quartile of Bloomberg Risk_score. We re-evaluate all models under identical five-fold company-grouped partitions, fitting the 75th-percentile threshold, missing-value imputation and encoding within the training fold only. Mean grouped ROC-AUC is 0.723 (SD 0.065) for the decision tree, versus 0.809 (SD 0.049) for random forest and 0.806 (SD 0.043) for logistic regression. At the 0.5 threshold, the decision tree attains accuracy 0.692, below the 0.749 always-low rule, while sensitivity is 0.709. Identifying 80% of high-risk cases requires reviewing a mean 56.2% of observations within validation folds. Removing two aggregate Bloomberg predictors reduces tree AUC to 0.709; a broad exclusion reduces it to 0.653. SHAP makes model contributions auditable but does not independently validate ESG harms. The results support explanation-led exploratory analysis, not operational risk decisions without independent outcome validation and calibrated review costs.