XGBoost
SMOTE
Random Forest
10,000 Txns
1.12% Fraud Rate
MODEL ACTIVE
--:--:-- IST
Financial Crime Detection · ML Dashboard
Kishore U. · github.com/ukishore33 · linkedin.com/in/kishore-techie  |  XGBoost · ROC-AUC 1.00 · PR-AUC 1.00 · F1 1.00 · SMOTE Oversampling
Total Txns
10,000
30-day simulation
Fraud Cases
112
1.12% fraud rate
Fraud Amount
₹2.87Cr
Avg ₹2.56L/case
ROC-AUC
1.00
XGBoost
PR-AUC
1.00
Imbalanced metric
F1 Score
1.00
Precision + Recall
SMOTE Ratio
88x
90 → 7,910 fraud
False Neg.
0
No missed fraud
Fraud vs Legit by Transaction Type 10,000 txns
Fraud Distribution by Hour
Alert Tier Distribution
XGBoost Feature Importance · Top Drivers
SMOTE — Oversampling for Class Imbalance imblearn
1.12%
Original fraud rate
50 / 50
Post-SMOTE split
88×
Synthetic samples added
Why SMOTE? With only 1.12% fraud (112/10,000), a naïve model predicts all-legit and achieves 98.88% accuracy while catching zero fraud. SMOTE (Synthetic Minority Oversampling Technique) generates synthetic fraud samples using k-nearest-neighbours interpolation — forcing the model to learn genuine fraud decision boundaries rather than the majority-class shortcut.
Model Performance Comparison
▶ XGBOOST (PRIMARY)
Accuracy1.0000
Precision1.0000
Recall1.0000
F1 Score1.0000
ROC-AUC1.0000
PR-AUC (imbalance-aware)1.0000
▶ RANDOM FOREST (BASELINE)
ROC-AUC1.0000
PR-AUC1.0000
F1 Score1.0000
5-Fold CV AUC (mean ± std)1.00 ± 0.00
Confusion Matrix · XGBoost · Test Set (2,000 txns)
1,978
TRUE NEGATIVE
0
FALSE POSITIVE
0
FALSE NEGATIVE
22
TRUE POSITIVE
Fraud Alert Queue · ML-Detected Transactions 112 fraud cases
Filter:
Txn IDAccountType Amount (₹)Orig Bal BeforeOrig Bal AfterDest Bal After Fraud ScoreRisk Tier NightRound AmtDest ZeroedOrig Drained