Machine Learning Classification of Return on Equity from Sustainability Reporting and Corporate Governance Metrics: A SHAP-Based Explanation

The aim of this study was to develop a model that classifies companies into high or low categories based on their return on equity (RoE), the most important indicator of financial performance, using sustainability and governance-related committee reports and reports shared with the public. As a sample, the RoE, sustainability, and governance variables of all 427 companies traded on the Istanbul Stock Exchange in 2024 were used. Using a 70:30 stratified split between the training and test sets, three tree-based models (XGBoost, LightGBM, and Random Forest) were used to perform a binary classification task. The findings show that tree-based models perform only slightly better than the naive majority class rule, and therefore, have limited overall classification power. A noteworthy finding from the study is that SHAP-based explainability analysis shows that the Corporate Governance Report (IMNG), the Integrated Report (IREP) and the existence of a Sustainability Committee (ICOM) rank higher in terms of SHAP-based global importance in the High RoE classification model, although their average contributions are small and, in the case of IMNG, predominantly negative for the probability of belonging to the High RoE class. Methodologically, the article moves away from traditional econometric methods based on ESG scores, instead combining a predictive classification structure with TreeSHAP-based explanations. These findings indicate a need for reporting practices that offer deeper content, clearer evidence of governance quality, and stronger data integrity to better support investors' decision-making processes through sustainability and governance.