A Design Science Approach to Predicting ESG Performance Using Ensemble Machine Learning

Environmental, Social, and Governance (ESG) metrics have become a corner-stone to sustainable finance, yet their measurement and predictability remain constrained by data heterogeneity, methodological divergence, and disclosure bias. This study develops a comprehensive ESG prediction framework grounded in the Design Science Research paradigm, integrating advanced machine learning techniques with rigorous data preprocessing, feature selection, and temporal validation. Using firm-level data from Refinitiv and Bloomberg, the analysis distinguishes between ESG composite performance and disclosure-based ro-bustness, addressing a critical gap in the literature. Ensemble learning models, including Random Forest and XGBoost, are evaluated alongside deep learning architectures using multiple sampling strategies and rolling-window validation. The results demonstrate that ESG performance is moderately forecastable, with ensemble methods consistently outperforming neural networks in structured datasets. In contrast, disclosure robustness exhibits lower predictability, re-flecting its dependence on discretionary strategic reporting and institutional factors. The findings highlight the importance of data quality, model selection, and validation design in ESG analytics, while emphasizing the limitations of deep learning in tabular financial contexts. The integration of explainable arti-ficial intelligence further enhances interpretability by identifying key predictors of ESG outcomes. Overall, the study contributes to the literature by providing a robust, interpretable, and methodologically rigorous framework for ESG predic-tion, with implications for investors, regulators, and corporate decision-making.