How does machine learning credit scoring compare with traditional logistic regression?

Question

AcadiFi · Accepted Answer

Gradient-boosted tree models (XGBoost, LightGBM, CatBoost) and neural networks typically achieve 2 to 8 Gini points of lift over well-tuned logistic scorecards on the same data, with the biggest gains on complex high-dimensional datasets including bureau attributes, transactional, and alternative data. However, the tradeoffs for FRM-relevant model risk are significant.

Interpretability is the first concern. Logistic scorecards produce transparent points tables suitable for reason code generation under the Equal Credit Opportunity Act (Regulation B) and Fair Credit Reporting Act adverse action rules. ML models require SHAP or LIME explanations, which are post-hoc approximations and may not map cleanly to the four required reason codes.

Stability is the second. Tree ensembles can overfit to sample-specific interactions and degrade faster under population drift than monotonic logistic scorecards. Monitoring via PSI and feature drift becomes more critical.

Fairness and disparate impact testing — measuring approval rate ratios across protected classes — is harder with ML because fairness metrics can diverge from accuracy metrics. Third, regulatory scrutiny: SR 11-7 in the US and PRA SS1/23 in the UK explicitly require conceptual soundness, which examiners find harder to evaluate on deep models.

Practical recommendation for a firm like Northglade Bank: deploy ML as a challenger model with human-in-the-loop overrides, run side-by-side for 12 to 18 months, and build a parallel interpretability framework before migrating the champion.

How does machine learning credit scoring compare with traditional logistic regression?

Master Part I with our FRM Course

Related Questions

Practice Questions