DC-Guard: A Risk-Tiered Data-Centric Framework for Governing Dataset Shift in Credit-Risk Models — Empirical Validation on Fannie Mae Loans

Abstract

Machine-learning credit-default models degrade when production data diverge from training distributions. We introduce DC-Guard, a governance framework that maps Population-Stability-Index (PSI) and model-level drift (AUC, ECE) to Green / Amber / Red risk tiers and prescribes auditable actions (monitor, diagnose, retrain, restrict). We first ground the policy on the canonical UCI credit-card data set; we then validate it end-to-end on 2.1 million Fannie Mae single-family loans originated 2020–2023. On the live portfolio DC-Guard triggered exactly one red alert (Sep-2022) when PSI ≥ 0.2 and AUC dropped 0.058, prompting human-in-the-loop review and selective automation freeze until performance recovered. No false-positive retraining occurred. All code and data are publicly available.