Change Control for Adaptive Machine-Learning Credit Models: A Tiered Approval Framework for Recalibration, Retraining and Redevelopment in Community Lending

Abstract

Adaptive machine-learning credit models promise to keep early-warning systems accurate as economic conditions change, but every update is also a model change that supervisors expect to be controlled, validated and documented. Recent work on model risk management for adaptive credit early-warning systems in supervised small-business lending has argued that machine-learning recalibration must be aligned with federal model-risk guidance and the NIST AI Risk Management Framework [1]. Community banks and credit unions face this challenge with small model-risk teams, so a change-control process that treats every update as a full redevelopment is impractical, while one that treats every update as routine is unsafe. This article proposes a tiered change-control framework for adaptive credit models. It classifies changes into three tiers through a simple decision tree, introduces a pre-approved recalibration envelope within which routine updates may proceed with light-touch review, scales documentation, testing, validation and approval to each tier, and defines a six-stage lifecycle with approval gates. The framework is mapped to supervisory model-risk guidance and to the govern, map, measure and manage functions of the NIST AI RMF, and it draws on the predetermined change control plans now used for adaptive medical AI. An illustrative stress scenario shows how the tiers allocate review effort. We then test the envelope empirically on real SBA small-business loans, applying its trigger, bounds and automated tests to 144 monitoring scenarios. Gradual shifts in sector mix never breached the trigger (observed-to-expected defaults 0.97 to 1.20). Under credit-cycle shifts the trigger worked, but the materiality bounds escalated 36 recalibrations to Tier 2, of which 32 needed nothing more than recalibration; no scenario required retraining. The envelope bounds should therefore distinguish large but correct recalibrations from risky ones. The framework is intended to let smaller lenders benefit from adaptive models without weakening model governance.