Credit Underwriting Model Drift vs. a Fixed Validation Queue
You lead AI risk management at a mid-size bank. The SME credit-limit model (logistic-regression scorecard plus a gradient-boosted challenger blended 70/30) processes about 12,000 applications a month. Monthly monitoring shows the population stability index on the income-volatility feature at 0.31 against a 0.20 trigger, and the approval rate for applicants in two geographic clusters fell from 41% to 29% in one quarter while the rest of the book moved two points. The model owner wants to retrain immediately on the last nine months of decisions and redeploy within two weeks. Independent validation has a three-week backlog, the annual model review is due in six weeks, and a supervisory exam begins in four months. Adverse-action reason codes were last revised eighteen months ago and are generated from the top three logistic coefficients only. A product manager argues the drop reflects a genuine change in small-business creditworthiness and that pausing applications would cost roughly 900 bookings a quarter.
1. Within the first five business days, what actions do you take and in what order? Justify why retraining is not the first action.
Read the worked answer
Day 1–2 - establish materiality before touching the model. Pull the decline cohort and decompose it: is the approval-rate gap explained by score distribution shift, by threshold movement, or by a change in the applicant mix in those clusters? Compute the gap holding score constant. If similarly-scored applicants are being declined at different rates across clusters, the driver is not creditworthiness and the product manager's explanation fails on its own evidence. Document this as a potential discriminatory-outcome indicator requiring escalation, not a performance issue.
Day 2–3 - classify severity. Run the incident through the model risk tier: a tier-1 lending model with a live outcome gap and stale reason codes is at minimum a high-severity model risk event. That triggers notification to the model risk committee and the compliance function, which the retraining proposal does not.
Day 3–5 - containment before remediation. Options in ascending cost: (a) hold thresholds, (b) apply a temporary manual-review overlay for the affected clusters with a documented override rationale, (c) suspend automated declines in the clusters and route to human underwriters. Option (c) is usually correct because it removes the harm while preserving volume - the 900-booking cost estimate assumes a full pause, which overstates the real cost and should not drive the decision.
Why retraining is not first. Retraining on the last nine months learns from decisions the current model produced. If the model has been systematically declining a segment, the recent labels encode that pattern and the retrained model reproduces or amplifies it. Retraining also consumes the two weeks in which the harm continues, and it forecloses the diagnostic that tells you whether this is drift or discrimination. Retraining answers 'the model is stale'; the evidence so far does not yet establish that.
2. The model owner insists the cause is a genuine shift in small-business creditworthiness. What evidence would you require to accept that claim, and what evidence would refute it?
Read the worked answer
Acceptance requires three independent lines of evidence that converge, because any single one is confounded.
1. Externally observed outcomes, not model outputs. Compare realized delinquency and default in the two clusters against matched peers (same industry, revenue band, tenure) in unaffected regions over the same window. If the clusters are not defaulting more, the model's decline decision has no predictive basis. This is the decisive test and it uses ground truth the model never saw.
2. Feature-level attribution against a benchmark. The income-volatility PSI of 0.31 tells you the input distribution moved, not why. Check whether the shift is in the underlying economic variable or in how it is sourced - a change in a data vendor's coverage, a reclassification of business types, or a new reporting lag produces identical PSI. If the raw economic variable is stable and only the derived feature moved, the cause is data pipeline, not creditworthiness.
3. Score-conditional comparison. If applicants in the clusters and outside them with near-identical scores are approved at materially different rates, the outcome gap is not score-driven and the creditworthiness story collapses.
What would refute the claim: stable or improving realized default in the clusters, a data-provenance change coinciding with the PSI jump, or a gap that disappears once score is held constant. What would support it: worsening realized default in the clusters and matched peers, no provenance change, and approval rates that track score consistently everywhere - in which case the correct response is a documented recalibration, still subject to validation, plus an explanation to the affected market rather than a quiet retrain.
3. Draft the core recommendation for the model risk committee, including what you would tell the committee not to approve and the residual risk you are accepting.
Read the worked answer
Recommendation. (1) Immediately route automated declines in the two clusters to human review, effective within 5 business days, with a documented override field captured for later analysis. (2) Commission a full diagnostic - score-conditional approval analysis, realized-default comparison, feature provenance review - due in 15 business days. (3) Freeze the current model version and its thresholds; no parameter changes until the diagnostic closes. (4) Rebuild adverse-action reason codes from the full model rather than the top three coefficients, since the blended challenger already makes the current codes inaccurate independent of this incident.
What not to approve. Retraining-and-redeploy in two weeks. It compresses validation, it predates the diagnostic, and its training labels come from the disputed decisions - it converts an open question into a permanent bias in the production model. Also reject 'no action pending the annual review': the annual review is six weeks away and the supervisory exam is four months away; a six-week delay puts the remediation inside the exam window with no evidence of timely detection.
Residual risk being accepted. Manual review adds an estimated 3–6 days to decision time for roughly 700 applications a quarter in the affected clusters and creates reviewer-capacity risk; it also introduces human inconsistency that must itself be monitored. There is a real possibility the diagnostic finds a legitimate shift, in which case the manual overlay was an unnecessary cost - that cost is smaller than the alternative if the finding goes the other way. Record the acceptance with an owner, an expiry date of 90 days, and the specific evidence that would lift it.