Bridging Data Silos in Corporate Governance: A Hierarchical Stacking Ensemble With Federated Dynamic Aggregation

Accurate Environmental, Social, and Governance (ESG) forecasting is pivotal for modern sustainable finance, yet it remains hampered by data heterogeneity, outlier distortions, and strict privacy regulations that inhibit cross‐institutional data sharing. This study addresses these limitations by proposing a comprehensive framework that synergizes advanced ensemble learning with privacy‐preserving federated architectures. First, a centralized “HSE‐ESG” is developed, utilizing a stacking ensemble of LightGBM, XGBoost, and multilayer perceptrons, synthesized by a Bayesian Ridge metalearner with a unique feature passthrough mechanism to capture complex nonlinear dependencies efficiently. Subsequently, to mitigate data leakage risks and address data silos, the framework transitions to “Fed‐ESGNet,” employing a novel Federated Dynamic Weighted Averaging (FedDWA) algorithm that aggregates client updates based on local validation performance rather than traditional sample volume weighting. Empirical analysis using a longitudinal Refinitiv dataset (2010–2022) reveals that the centralized model achieves a remarkable coefficient of determination ( R 2 ) of 0.9854, while the federated approach maintains near‐parity performance R 2 = 0.9799 despite data fragmentation, significantly outperforming standalone baselines. Furthermore, the integration of Explainable AI (XAI) techniques, specifically SHAP and LIME, guarantees granular decision‐making transparency, establishing a robust, scalable, and interpretable paradigm for secure ESG governance assessment.