0.17% of transactions
are fraud.
We built a system
that finds them anyway.
Cost-sensitive XGBoost (scale_pos_weight=578.55) with native SHAP game-theoretic explainability. Held-out PR-AUC 0.8424, 98.22% false-alarm reduction.
A naive model that predicts every transaction as legitimate
achieves 99.83% accuracy.
Yet it catches zero fraud.
Under extreme class imbalance, raw accuracy is mathematically meaningless.
We trained directly on the Precision-Recall operating surface,
minimized asymmetric financial friction ($5 FP vs $122.21 FN),
and engineered exact Shapley game-theoretic evidence.
Empirical Results Across Test Splits
Stratified 70/15/15 split. Evaluated on genuine precision, recall, and false-alarm costs.
+0.0520 lift over balanced baseline (0.7904)
63 of 74 fraud cases caught in test split
901 → 16 false alarms vs balanced baseline
FP per 10,000 processed transactions
Experience the Engine in Action.
Explore live payment scoring on genuine test transactions, or inspect exact game-theoretic dispute evidence.
See it score in real time
Test live probability scoring on genuine transactions and sweep the operating threshold slider.
See why it flagged something
Deconstruct exact game-theoretic feature contributions and generate audit-ready dispute summaries.
Real-time fraud intelligence designed for extreme class imbalance. Operating at cost-optimal operating points with instant Shapley game-theoretic evidence decomposition.
scale_pos_weight=578.55 · SHAP TreeExplainerHeld-Out PR-AUC 0.8424 · Precision-Recall Optimized