A Novel DWT and Attention-Based GRU-CNN Model for Transaction Throughput Prediction in Financial Trading Systems

Accurate prediction of throughput, measured in Transactions per Second (TPS), is critical for financial trading systems, as it ensures system stability and scalability under high transaction volumes. We address this challenge by introducing a hybrid model that combines Discrete Wavelet Transform (DWT), Convolutional Neural Networks (CNN), and a Gated Recurrent Unit with an Attention Mechanism (GRU-AM). Our approach decomposes the nonlinear, non-stationary TPS data into high- and low-frequency components using DWT, allowing CNN and GRU-AM models to capture dynamic and nonlinear features effectively. Experimental results on real-world trading system data demonstrate the method's effectiveness. Compared to a variety of models, including ARIMA, CNN, and GRU, our method improves prediction accuracy by 2% to 20%. These results underscore the model's enhanced accuracy and efficiency, providing robust support for TPS prediction in financial trading systems.