Dynamic Calibration of Decision Thresholds for Financial Anomaly Detection: Verification With Payment Platform Information and Data

Abstract

Digital payment channels have expanded quickly, reshaping transaction flows while opening new avenues for fraud. Isolation Forest (IF) remains attractive for unsupervised screening, yet deployments that rely on a fixed anomaly-score threshold deteriorate when traffic shifts or is actively manipulated. The authors present a Temporal-Attention Isolation Forest with Dynamic Calibration (TA-IFDC) that treats threshold selection as an adaptive component rather than a static post-processing step. The method monitors the evolving distribution of IF scores in streaming mode and updates the decision boundary online, while a lightweight temporal-attention module encodes short-range dependencies across consecutive transactions. Together, these pieces allow the detector to adjust to drift without sacrificing precision during stable periods.