High-Recall Deep Learning: A Gated Recurrent Unit Approach to Bank Account Fraud Detection on Imbalanced Data

Fraud detection in financial services is challenged by severe class imbalance, where fraudulent events are rare. This study provides a rigorous comparison of classical machine learning models—Logistic Regression, SVM, Random Forest, and LightGBM—and a Gated Recurrent Unit (GRU) deep learning model on a large-scale, imbalanced bank account fraud dataset. Our findings reveal a dramatic performance divide. All classical models, including the widely-used tree-based ensembles, proved ineffective at the primary detection task, with fraud recall scores below 7%.In stark contrast, the GRU network successfully identified the vast majority of fraudulent cases, achieving over 78% recall on the minority fraud class, albeit at the cost of very low precision. The results demonstrate that the GRU offers a viable, high-recall solution where classical models fail. This highlights a critical strategic choice for financial institutions: adopt a high-recall/high-alert detection framework or use conventional models that allow most fraud to go undetected. The approach is broadly applicable across financial domains—including banking and insurance—where rare-event detection is critical.