Transaction-Sequence Modeling for Early Detection of Coordinated Payment Fraud

Abstract

Coordinated payment fraud is difficult to detect from individual transactions because suspicious behavior often emerges only through sequences of interactions among multiple accounts. This paper proposes a transaction-sequence modeling approach for identifying coordinated fraud before a complete fraudulent pattern has formed. The method represents account activity through temporal transaction sequences and constructs interaction features describing repeated counterparties, transaction intervals, amount changes, and short-term fund circulation. A temporal encoder is used to capture behavioral evolution, while an interaction aggregation module incorporates information from related accounts without requiring a complete transaction graph to be reconstructed for every prediction. The model is evaluated on payment transaction data containing both isolated fraudulent activities and coordinated fraud patterns. Experimental results demonstrate improved precision-recall performance compared with transaction-level classifiers and static graph-based baselines, particularly when only partial fraud sequences are available. Case analysis further shows that abnormal transaction timing and repeated fund transfers contribute substantially to early detection.