ARCHIVES

Financial fraud detection in large-scale payment systems presents persistent challenges due to severe class imbalance, evolving fraud patterns, and the operational demand for both high recall and interpretability. This paper presents a complete, production-oriented fraud detection pipeline applied to the PaySim synthetic financial dataset containing 6,362,620 transactions with a fraud prevalence of 0.13%. The proposed system combines a deterministic hard-rule layer for extreme velocity blocking with a gradient-boosted tree classifier (XGBoost) trained on eight hand- crafted behavioral velocity features derived from 24-hour rolling windows, computed efficiently using cumulative- sum techniques and Numba JIT compiled kernels. Probability outputs are post-hoc calibrated via isotonic regression using the FrozenEstimator pattern. At a decision threshold of 0.20, the model achieves a ROC-AUC of 0.9994, a precision of 0.96, a recall of 0.81, and an F1-score of 0.88 on a temporally held- out test set. SHAP (SHapley Additive exPlanations) waterfall plots are integrated into a real-time Gradio inference interface, providing per-transaction explainability to analysts. A calibration analysis reveals overconfidence at low probability scores, moti- vating future work on temperature scaling. The false positive rate of 0.013% over 1.24 million legitimate test transactions demonstrates practical viability for real payment environments.