Uncertainty-Aware Credit-Risk Prediction with Structured Bayesian Temporal Modeling

Credit-risk models may become unreliable when longitudinal observations are incomplete and predictive uncertainty is ignored. We propose a Structured Dynamic Bayesian Network and Bayesian Neural Network (Structured DBN + BNN) combining borrower-conditioned latent-state transitions, differentiable filtering, complementary GRU and direct-input encoders, and variational Bayesian discrimination. Training-time masking improves robustness to missing evidence, while reconstruction regularizes the latent states. At inference, Monte Carlo sampling, validation-based temperature scaling, predictive entropy, and mutual information support calibrated prediction and uncertainty-aware triage. On the Default of Credit Card Clients dataset, ten paired seeded evaluations show competitive clean-data discrimination, with an AUROC of 0.7828. Under 30% MCAR missingness, the model achieves an AUROC of 0.7623 and achieves a higher mean AUROC than the principal baselines drawn from recurrent, Bayesian, and Transformer model families; similar robustness is observed under MAR, MNAR, and recent-month removal. At 10% selective coverage, entropy-based ranking reduces the mean error to 3.51%. These findings show that the framework’s main benefit is robust, uncertainty-aware prediction when temporal evidence is incomplete, rather than uniformly superior discrimination on complete data.