TransactionGuard XGB+LGB
Trained on 590,000 real transactions from the IEEE-CIS Kaggle competition. Uses an XGBoost + LightGBM ensemble with time-aware train/val split to prevent leakage. Features include PCA-transformed velocity signals (V1–V28), amount patterns, card metadata, and engineered time features. SHAP explainer returns per-prediction risk factors in real time.
Testing Lab
Enter values to run a single prediction
~0.95+
Ensemble AUC
XGB + LGB
Models
590K
Training rows
F1-tuned
Threshold
Security Report
Fill in the fields and run analysis to see the prediction result
Architecture Notes
Two independently trained gradient-boosted tree ensembles (500 estimators each, max_depth=6, lr=0.05, early stopping on validation AUC). Final score = simple average of both probabilities. Threshold tuned on F1 to handle severe class imbalance (3.5% fraud rate). SHAP TreeExplainer provides local feature attribution per transaction.