Temporal Deep Learning for Financial Fraud Detection: A Comparative Evaluation of LSTM, GRU, Transformer, and TCN on Real Transaction Datasets

Abstract

Financial anomaly detection has evolved from static classification toward sequence-aware learning that leverages the temporal structure of transaction streams. Four families of temporal neural architectures are now routinely applied in this domain: Long Short-Term Memory (LSTM) networks, Gated Recurrent Units (GRU), Transformer encoders, and Temporal Convolutional Networks (TCN). Existing empirical comparisons are typically restricted to a single dataset and a single problem granularity, leaving open the question of how these architectures rank when evaluated under a common protocol across transaction-level and node-level risk prediction. This paper presents a controlled empirical comparison of the four architectures on three real financial datasets: the ULB European credit-card transactions released by Worldline and the ULB Machine Learning Group, the IEEE-CIS card-not-present dataset contributed by Vesta Corporation, and the Elliptic Bitcoin transaction graph released by Elliptic together with the MIT-IBM Watson AI Lab. Under a unified sequence-learning interface with matched preprocessing and matched parameter budgets, the four architectures are examined using ROC-AUC, PR-AUC, F1, and recall. The Transformer attains the highest mean ROC-AUC on each of the three datasets, with GRU and TCN close behind and LSTM slightly trailing; on the smallest dataset, however, the spread is of the same order as the five-seed standard deviation, and the cross-dataset ordering should be read as an ordering of sample means under a shared sequence interface rather than as a strict apples-to-apples separation. The widest inter-architecture F1 gap emerges on the Elliptic node-level setting. By framing anomaly detection as a temporal behavioral modeling problem rather than an isolated classification task, the study highlights the capacity of sequence-aware architectures to capture evolving transaction patterns and positions temporal scoring as a mechanism to strengthen platform reliability and user trust.