Capital AI

Step 4 of 5

Model Training

Train and compare three complementary scorers — a supervised gradient booster, an unsupervised anomaly detector, and a transparent rules engine — then review combined evaluation metrics.

XGBoost

Supervised gradient booster

Trained

AUC

0.962

Precision

0.871

Recall

0.793

0.830

Fraud probability output0.00 – 1.00

Feature importance

Isolation Forest

Unsupervised anomaly detector

Trained

AUC vs labeled fraud

0.884

Anomaly score distribution

Isolation percentile histogram

Rules Engine

Transparent deterministic rules

Trained

Rules

Triggered 24h

1,847

Coverage

64%

Top rules · editable thresholds

R-001
High amount > $5,000
412 hits
Threshold$
R-014
Velocity: 5+ txns / 10 min
287 hits
Thresholdtxn
R-022
Geo velocity > 800 km/h
198 hits
Thresholdkm/h
R-007
New device + amount > $300
164 hits
Threshold$
R-019
IP country ≠ card BIN
141 hits
Thresholdflag
R-031
Merchant chargeback > 2%
92 hits
Threshold%

Combined evaluation

Hold-out validation dashboard

n = 184,302 · stratified split

Fraud detection rate

79.4%

False positives

287

False negatives

509

True positives

1,961

True negatives

181,545

ROC Curve

True positive rate vs false positive rate

XGBoost Iso Forest Rules

Precision–Recall Curve

Precision across recall thresholds

XGBoost Iso Forest Rules

Confusion Matrix

Counts at the calibrated decision threshold

Pred. Legit

Pred. Fraud

Actual Legit

181,545

287

Actual Fraud

509

1,961

Accuracy 99.57% · Recall 79.39%