Privacy-Preserving ESG Risk Prediction via Federated Hybrid Neural-Boosting Ensembles

The rapid evolution of sustainability legislation, including the CSDDD and CSRD, has created a complex dual mandate for modern enterprises. Organizations are now required to provide granular transparency into their supply chains while simultaneously protecting proprietary operational data and trade secrets. This regulatory landscape exacerbates a critical financial problem: pervasive information gaps in ESG data hinder accurate risk assessment, especially for banks and investors who lack access to confidential supply chain details. Existing centralized reporting models are inadequate, as they either compromise data privacy or fail to deliver the cross-institutional insights needed for compliance. This paper addresses this challenge by introducing a privacy-preserving framework that combines a Hybrid Neural-Boosting Stack (HNB-Stack) with a decentralized Federated Voting-Ensemble System (Fed-VES). We employ a rigorous hybrid feature selection process to isolate high-value predictors from over 400 variables, which then inform a stacked ensemble architecture comprising LightGBM, XGBoost, and Multi-Layer Perceptrons. The results demonstrate that the HNB-Stack achieves superior predictive fidelity with a centralized <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> of 0.9835, significantly outperforming traditional baselines. Most notably, the federated implementation maintains this high accuracy across five isolated institutional clients with minimal performance degradation, successfully proving that collaborative intelligence does not require centralizing sensitive data. By integrating SHAP and LIME for post-hoc interpretability, the framework also ensures that AI-driven risk assessments remain transparent and auditable. This research ultimately provides a scalable solution for navigating the regulatory landscape, enabling secure and reliable ESG reporting without compromising data sovereignty.