Real-time fraud detection
Authorization-time fraud scoring
Owner: Fraud Data Science · last validated Apr 18, 2026 · metrics illustrative on synthetic data
Outcome it answers for
Illustrative. In production this figure comes from the monitored champion-challenger result, not the card.
Top drivers (SHAP)
Disaggregated performance
| Group | AUC | Approval |
|---|---|---|
| Mass | 0.939 | 97.1% |
| Affluent | 0.945 | 97.8% |
| Private Client | 0.948 | 98.2% |
| Small Business | 0.933 | 96.4% |
Segments are business segments. This model is not a credit decision, so this is a parity check, not a fair-lending test.
Intended use
Scores each authorization for fraud probability; declines or steps up above threshold, with human review on edge cases. Not a credit decision.
Training data
18 months of authorizations with confirmed-fraud labels, temporal hold-out split. Metrics shown are illustrative on synthetic data.
Limitations
Weaker on novel card-not-present attack patterns; retraining triggers at KS drift above 0.05.
Fair treatment and decline reasons
Not a credit decision: refusals to authorize at the point of sale are not adverse action under Reg B (12 CFR 1002.2(c)(2)(iii)). No protected-class attributes are used; approval-rate parity is monitored across segments. Decline reasons go to servicing agents, never to the declined party in fraud-revealing detail.