The Black-Box MSME Scorecard That Stopped Working
Pragati Finance, an NBFC with ₹3,200 crore of MSME assets, runs a vendor-supplied machine-learning scorecard for unsecured working-capital loans of ₹2–15 lakh. Inputs are GST return regularity, UPI merchant inflows, telecom recharge behaviour and bureau data. At validation in January the model showed AUC 0.81 and KS 0.34; by August AUC had fallen to 0.74. The Population Stability Index for the GST-turnover band feature is 0.28 against a flag threshold of 0.10, and the share of borrowers with less than 12 months of GST history has risen from 18% to 41% after the sales team widened sourcing to newer merchants. Approval rates in two eastern districts have dropped 9 percentage points. Delinquency at 90+ days in the affected segment is 4.1% against 2.2% portfolio-wide. The credit head wants to switch the model off and revert to the manual scorecard. The vendor refuses to disclose feature weights, calling them proprietary. The audit committee wants a model risk note before the next disbursal cycle.
1. Distinguish between population shift, concept drift and a data-quality problem in this evidence. What should be done before the next disbursal cycle?
Read the worked answer
The evidence points first to population shift (covariate shift), not yet to concept drift. PSI of 0.28 on GST-turnover band, plus the jump in under-12-month GST filers from 18% to 41%, says the input distribution moved because sourcing changed. The AUC fall is exactly what you expect when a model built on established filers is applied to new ones. Concept drift - a change in the relationship between inputs and default - can only be concluded by re-running the model on the new segment's realised outcomes and comparing predicted against observed bad rates by decile. Do that test before concluding anything about the algorithm. A data-quality check is also mandatory: newer merchants often file nil returns, so turnover bands may misclassify them rather than the behaviour having changed.
Before the next cycle: do not switch off. Reverting to an unvalidated manual scorecard is a governance regression and removes the only consistent decision standard. Instead throttle - cap the share of new-to-GST borrowers per branch, apply a conservative cut-off overlay for that segment, and route cases above ₹8 lakh to manual review with documented reasons. Commission a two-week revalidation: segment-wise Gini/KS, calibration tables, and observed-versus-expected 90+ dpd at the live cut-off. If realised delinquency at the current cut-off is 4.1% against a 2.2% pricing assumption, either re-price or decline the segment. Put the overlay, the reasoning and an expiry date in front of the model risk committee.
2. The vendor will not disclose feature weights. Can the NBFC keep using the model? What must it require to do so?
Read the worked answer
It can keep using it only with compensating controls, and the current position is not defensible. The NBFC is the lender of record: the decision, the decline and the reason are its responsibility, not the vendor's. It must be able to defend a rejection to the borrower under fair-practices expectations, to internal audit, to the regulator, and potentially to a court. 'Proprietary' cannot be a shield for a decision taken in the lender's name.
Minimum requirements to continue:
- Full model documentation to supervisory standard: development population, observation and performance windows, feature list with direction and monotonicity, calibration, known limitations.
- A per-decision reason code or attribution output generated at scoring time - the vendor need not publish weights, but it must return the top adverse factors for each applicant.
- Independent validation access: either the NBFC's validator sees the model logic under NDA, or a third-party validator does; plus audit and inspection rights.
- Model change notification - no silent recalibration - with a defined notice period and revalidation trigger.
- Performance and liability terms: accuracy warranties, a service level for retraining, and liability if the model materially underperforms.
Parallel action: build a challenger model on the NBFC's own data. If the vendor refuses reason codes and validation access, the challenger becomes champion and the vendor score survives only as one input among several.
3. Design the quarterly monitoring pack the audit committee should see. What would make you escalate?
Read the worked answer
The pack should cover four layers, each with a threshold and an owner.
Discrimination and stability: AUC/Gini and KS against the validation baseline; PSI per feature and overall; characteristic stability for bureau score, GST band and UPI inflow. Escalate on PSI above 0.25 for two consecutive quarters or AUC drop above 0.05.
Calibration and outcomes: predicted versus observed 90+ dpd by score band and by segment, vintage-level delinquency, and the realised loss against the pricing assumption. A widening gap here is the earliest true signal of concept drift.
Decision quality: approval rate by channel, branch and district; override rate and the delinquency of overrides versus model-approved cases - an override population performing worse than the model is the most common hidden loss driver; rejection reason distribution.
Fairness and conduct: approval rates and cut-offs across region, gender and age band, with disparate-impact ratios computed and explained; complaints citing automated rejection; data-feed completeness and latency.
Escalate immediately if realised delinquency in a segment exceeds the pricing assumption by more than 50 basis points, if a data feed fails for more than one cycle, if any disparate-impact ratio falls below 0.80 without a documented justification, or if the vendor changes the model without notice. Each escalation must state the trigger, the interim control and who owns the fix.