Probabilistic Risk Aggregation Models for Adaptive Financial Fraud Detection in Enterprise Systems

Abstract

Financial fraud detection in enterprise environments presents persistent challenges due to evolving attack patterns, high transaction volumes, and the inherent limitations of static rule-based systems. This paper proposes a Bayesian risk aggregation framework that fuses behavioral, transactional, and network-topological signals into a unified posterior fraud probability estimate. The framework employs conjugate Gaussian-Gaussian updating to incorporate confirmed fraud labels incrementally, enabling continuous model adaptation without full retraining. An adaptive drift detection mechanism monitors rolling classification performance and triggers weight recalibration when statistically significant concept drift is detected. The decision engine applies a three-zone routing architecture — automatic block, human review, and pass-through — with per-signal log-odds attribution to satisfy explainability requirements under applicable regulatory standards. Experimental evaluation against published benchmark systems demonstrates that the proposed framework achieves competitive detection performance while maintaining interpretability and operational adaptability. The architecture is designed for production deployment in enterprise transaction processing environments and addresses key limitations of existing approaches including static thresholds, single-modality detection, and opaque decision logic.