Adaptive Lifecycle Management of Credit Risk Models Under Distributional Shift: A Governance-Aware Framework

Credit risk models are estimated from historical data but are deployed in environments that evolve over time. Changes in borrower characteristics, lending policy, pricing, portfolio composition, and macroeconomic conditions can alter both model inputs and the reliability of predicted probabilities. This thesis develops a governance-aware lifecycle framework for credit risk modeling under temporal distribution shift, treating model development as a continuing process of validation, monitoring, and controlled adaptation rather than as a one-time classification exercise. The empirical study uses Lending Club accepted-loan data from 2007–2018 and applies a common contractual horizon, explicit maturity and leakage controls, and strictly chronological data partitions. The framework combines model-family comparison, probability calibration, decision-threshold analysis, statistical uncertainty assessment, matured and label-free monitoring, explainability, challenger evaluation, macroeconomic sensitivity, and reproducibility controls within a single workflow. The results show that greater algorithmic complexity does not automatically justify model replacement. Discrimination, calibration, and population stability can evolve differently over time, while monitoring can reveal persistent distributional change before repayment outcomes become fully observable. Updating a model with more recent data can also be more valuable than simply adopting a more complex algorithm. The proposed framework therefore supports evidence-based model adaptation through investigation, validation, challenger review, and human approval rather than automatic model replacement.