Balancing Explainability and Privacy in Bank Failure Prediction: A Differentially Private Glass-Box Approach

Abstract

The increasing reliance on data-driven models for financial risk assessment has highlighted the need for both model transparency and data privacy, particularly in sensitive domains such as bank failure prediction. Traditional machine learning models often face a trade-off between explainability and privacy, where highly accurate models tend to be opaque, and privacy-preserving methods can reduce interpretability. This study proposes a novel framework that balances these requirements through a differentially private glass-box approach for bank failure prediction.The proposed system employs inherently interpretable models, such as decision trees or rule-based classifiers, combined with differential privacy mechanisms to protect sensitive financial data. By introducing controlled noise during training and inference, the model ensures that individual data points cannot be inferred while still maintaining meaningful predictive performance. The glass-box nature of the model allows stakeholders, including regulators and financial analysts, to understand the reasoning behind predictions, enhancing trust and accountability.Experimental evaluation demonstrates that the proposed approach achieves a favorable balance between prediction accuracy, interpretability, and privacy preservation, outperforming traditional black-box models with privacy constraints. The framework is particularly valuable for regulatory environments where transparency and data confidentiality are critical. Overall, this work contributes to the development of secure, explainable, and reliable financial risk prediction systems in modern banking.