Explainable Federated Early-Warning for Commercial Banks: Privacy-Preserving Financial-Distress Prediction and Decision Support in Bangladesh

Bank supervisors need early-warning models that learn from multiple institutions without centralizing confidential bank records. This paper evaluates an explainable cross-silo federated learning framework for one-year-ahead bank-distress prediction in Bangladesh. The compiled panel contains 780 bankyear records for 57 banks over 1997–2025; after forward target construction, 614 observations remain for 2013-2024, with 452 training and 162 strictly out-of-time test observations. Distress
is defined as a next-year breach of either a 10% NPL or 10% CAR threshold. Seventeen CAMELS-derived predictors are winsorized and standardized, and five unbalanced silos train classweighted logistic models aggregated by sample-weighted FedAvg. Centralized Logistic Regression, Random Forest, XGBoost, and Explainable Boosting Machine provide reference baselines, while FedProx, clustered personalization, and differentially private FedAvg test robustness and privacy. FedAvg reaches ROC-AUC 0.9685 versus 0.9714 centrally, retaining 99.7% of reference discrimination without moving raw records. At an RDP-accounted operating point of ϵ ≈ 3.52, δ = 10−5
, DP-FedAvg reaches ROCAUC 0.926. SHAP identifies NPL ratio, ROE, log Z-score, ROA,
and loan-to-assets as leading drivers; independent CAMELS ablation confirms asset quality, earnings, and capital as the most consequential groups. A two-year horizon loses substantial threshold-dependent performance, and a supervisory watchlist
translates probabilities into risk bands and drivers. The results support privacy-aware collaborative bank-risk screening while clearly distinguishing retrospective federation emulation from live deployment.