An Explainable, Privacy-Preserving AI Framework for Credit Risk in Nepalese Cooperative Banks: A Federated Survival Graph Bayesian Approach

Abstract

Nepal’s cooperative banking sector is in serious trouble. By mid-July 2025, non-performing loans across the country’s banking and financial institutions had climbed to Rs 258.19 billion, and the cooperative sector accounts for a large share of this problem. Forty cooperatives alone embezzled Rs 87.89 billion from over 60,000 depositors, many of them farmers, housewives, retirees, and daily wage laborers. The National Cooperative Bank Limited reported a capital adequacy ratio of just 0.82 percent, which points to technical insolvency rather than mere distress. The International Monetary Fund has warned that further cooperative failures could threaten the stability of Nepal’s entire banking system. The paper proposes a new theoretical architecture for credit risk management in Nepalese cooperative banks, which we call the Federated Survival Graph Bayesian Framework (FSGBF). Conventional binary classification approaches predict whether a borrower will default, but they cannot tell us when. Our framework addresses this through Cox proportional hazards survival analysis. It captures network contagion effects using Graph Neural Networks that map borrower interdependencies within cooperative communities. Cross-institutional learning is enabled through Federated Learning with Differential Privacy, so that 31,450 primary cooperatives can collectively strengthen their models without sharing sensitive member data. Epistemic uncertainty is quantified through Bayesian Neural Networks with Monte Carlo Dropout, which helps distinguish between data noise and genuine model ignorance. Causal explanations come through interventional Shapley values, which identify which features correlate with default and which features actually cause changes in default risk when manipulated. The paper proves the consistency of the Cox partial likelihood estimator, derives the Breslow baseline hazard estimator, and establishes the convergence of GraphSAGE under neighborhood sampling. We also prove the FedAvg convergence theorem under convexity and bounded gradient assumptions, derive the Renyi differential privacy composition rule, and show that Monte Carlo Dropout approximates variational inference with a Bernoulli variational posterior. The four Shapley axioms for the interventional (causal) Shapley value are verified, and the bias-variance decomposition for the composite FSGBF predictor is established. Drawing on Nepal Rastra Bank’s Financial Stability Report FY 2024/25, the National Cooperative Federation of Nepal’s 2025 sectoral snapshot, parliamentary investigation findings, and the IMF’s July 2024 country report, we situate this mathematical framework within Nepal’s institutional reality. The sector has 7.4 million members, 56 percent of them women, holding Rs 847 billion in deposits. Yet it operates under split oversight, where 25,000 of 32,000 cooperatives are financial intermediaries with minimal real-time supervision. The paper argues that without predictive, network-aware, privacy-preserving, uncertainty-quantified, and causally explainable credit risk management, Nepal’s cooperative movement will collapse under the weight of its own opacity, and millions of livelihoods will go with it.