Credit Risk Scorecard: Development, Monitoring, and Stress Testing
Predict which credit card customers will default next month, grade the portfolio, and estimate losses under a downturn.
- Approach
- Validated 30,000 credit card accounts (UCI), built behavioral features while excluding sex, marital status, and age, selected features by weight of evidence and information value, and scaled a logistic-regression scorecard to points. Added risk grades, a cutoff strategy, expected loss, PSI monitoring, and stress scenarios.
- Result
- Hold-out Gini 0.49 and KS 0.40. Grade E defaults at 53.7% vs about 10% for grades A and B; a 520 cutoff approves 71.8% of accounts and cuts their default rate from 22.1% to 12.6%. Expected loss came within about 5% of realized loss, and a severe stress scenario raised it by 50%.
Tools: Python · pandas · scikit-learn · Logistic regression · Gradient boosting · Tableau · Power BI · DAX
More detail
Business impact
Shows how a lender can turn account behavior into an explainable score, set approval cutoffs, size expected losses, and see how those losses grow in a downturn.
Technical challenges
Two utilization and payment features flipped to the wrong direction alongside the delinquency features, so they were dropped to keep the scorecard explainable. A gradient-boosting challenger scored higher (Gini 0.55) but gives no points-based reasons, so it serves as a benchmark.
Lessons learned
Explainability, stability monitoring, and fair-lending checks matter as much as raw accuracy in a credit model.