Change an applicant's credit file and the scorecard re-scores it instantly: points, probability of default, an illustrative decision, and the adverse-action reasons a lender would have to disclose.
| Model | AUC | Gini | KS |
|---|
Decision bands are illustrative: approve below 8% PD, refer to an underwriter from 8% to 20%, decline above 20%. The model was trained on a synthetic loan book that mirrors public consumer-lending data, with an out-of-time split at July 2024. Points come from a WoE-binned logistic regression with monotonic bad rates; reason codes rank attributes by points lost against their best bin. Source code: github.com/martinogie/credit-risk-scoring by Martins Ogiekpolor, built with AI-assisted coding. Next step for this book: the Credit Loss Reserve (CECL and IFRS 9 allowance).