Real-Time Fraud Risk Scoring through Behavioral Sequence Analysis: An Explainable Approach for online Transaction Security

Abstract

Payment fraud continues to evolve, resulting in global losses of approximately $33.83 billion last year alone. We developed a system capable of highly effective fraud detection by analyzing spending behavior over time rather than focusing solely on individual transactions. While fraudsters can imitate a single purchase, it is far more difficult for them to replicate an entire behavioral history. Our key innovation lies in a hierarchical architecture that combines modified gated recurrent units (GRUs) with gradient boosting methods, allowing each component to perform its specialized role. The proposed system achieves an area under the precision-recall curve (AUPRC) of 0.876 and a Recall@1%FPR of 0.834 on the IEEE-CIS dataset comprising 590,540 transactions. In production-level stress tests, it sustains throughput up to 12,000 transactions per second under peak load conditions (with an average daily volume of approximately 3.4 million across partner institutions) and maintains a median latency of 47.3 ms (95% CI: [45.8, 48.9] ms; p95: 67.2 ms), remaining consistently below the 50 ms operational threshold. The system's core mechanism involves a temporal-gap-aware gating module within the modified GRU encoder that captures irregular intervals between purchases, effectively distinguishing genuine consumer behavior from fraudulent activity. This encoder is integrated with an ensemble of LightGBM, XGBoost, and Random Forest models, enabling robust voting-based decision fusion. To balance interpretability and performance, the system employs selective explainability-providing gradient-based attributions and counterfactual explanations only for the 5% of transactions flagged as high-risk-thereby ensuring regulatory transparency without compromising speed. Extensive validation using real banking datasets from three financial institutions confirmed its reliability and practicality. While not flawless, the system demonstrates sufficient robustness and interpretability for large-scale deployment-an outcome that ultimately matters more than theoretical optimality when millions of dollars are at stake each day.