Unsupervised Deep Learning for Sequential Anomaly Detection in Financial Transactions Using LSTM and CNN Autoencoders
Main Article Content
Anomaly detection in banking transactions is a critical task, particularly in scenarios where labeled fraud data is unavailable. This study proposes an unsupervised deep learning approach for detecting anomalous transaction patterns using sequential modeling with LSTM and CNN autoencoders. The transaction data is preprocessed and transformed into fixed-length sequences to capture temporal and behavioral patterns of individual accounts. Both models are trained to learn normal transaction behavior and identify anomalies based on reconstruction error. Experimental results show that the CNN autoencoder achieves lower reconstruction loss and detects a higher number of anomalies, indicating strong capability in capturing local patterns, while the LSTM autoencoder demonstrates a more selective detection behavior by modeling temporal dependencies. The comparison between the two models reveals partial agreement, where consistently detected anomalies represent high-confidence deviations from normal patterns. Statistical analysis further confirms that anomalous sequences exhibit significantly higher reconstruction errors than normal sequences. The findings highlight the effectiveness of sequence-based unsupervised deep learning methods for anomaly detection in financial transaction data and suggest that combining multiple architectures can improve detection robustness in real-world applications.