A real-time analytics framework for financial crime detection using sequential and market behavior learning

This paper presents a dual-branch deep learning framework for detecting and preventing financial crimes by integrating behavioral transaction data and high-frequency trading (HFT) signals. The proposed model combines a bidirectional LSTM with attention for modeling account-level transaction sequences and a convolutional-LSTM pipeline for capturing microsecond-level trading anomalies. Experiments conducted on two real-world datasets, one containing 2512 labeled financial transactions and another with high-frequency market event data, demonstrate that the proposed approach achieves state-of-the-art performance, with an AUC of 0.939 and an F1 score of 0.861 on the transaction dataset and an AUC of 0.879 and an F1 score of 0.831 on the HFT dataset. Ablation studies confirm the importance of attention mechanisms, while robustness evaluations show stable performance under noise and feature drift. The model maintains inference latencies below 3 ms per sample and uses fewer than 80,000 trainable parameters, making it suitable for real-time deployment in financial monitoring systems.