TCAE: A Temporal Contextual Autoencoder for Unsupervised Financial Fraud Detection

Abstract Financial fraud detection remains a significant challenge due to the rarity of fraudulent events, their evolving behaviour, and the inherent temporal complexity of large-scale transaction streams. Traditional rule-based systems and supervised learning models often suffer from severe class imbalance and limited generalization to unseen fraud patterns. To address these limitations, this study proposes the Temporal Contextual Autoencoder (TCAE), an unsupervised deep learning framework that models normal transaction behaviour through joint temporal–contextual representation learning. The architecture integrates Long Short-Term Memory (LSTM) units with a deterministic self-attention mechanism to capture long-range dependencies, local fluctuations, and contextual relevance within sequential financial data. Trained exclusively on legitimate transactions, TCAE reconstructs normal sequences with high fidelity while producing elevated reconstruction errors for anomalous patterns. Experiments on a benchmark credit card fraud dataset demonstrate that TCAE achieves state-of-the-art performance, attaining an AUC-ROC of 0.961 and an F1-score of 0.932, surpassing Isolation Forest, One-Class SVM, a standard autoencoder, and an LSTM autoencoder. These findings highlight the effectiveness of temporal–contextual modelling for unsupervised fraud detection and establish TCAE as a robust and reproducible approach for identifying anomalous financial transactions.