Temporal Attention Network With Particle Swarm Optimization for High-Frequency Order Book Prediction

Predicting the short-term state of the Limit Order Book (LOB) is crucial in quantitative finance, yet challenging due to its high-dimensional and noisy nature. While deep learning models like Transformers excel at capturing temporal patterns, they often rely on suboptimal hyperparameters and lack robustness. This paper introduces PSO-Transformer, a hybrid framework for high-frequency LOB prediction. It combines a modified Particle Swarm Optimization (PSO) with adaptive inertia and a Temporal Attention Network. The framework operates through three dedicated modules: data preprocessing, PSO-based hyperparameter optimization, and temporal attention forecasting. Experiments on a large-scale LOB dataset show that PSO-Transformer outperforms state-of-the-art benchmarks, achieving an average Matthews Correlation Coefficient (MCC) of 0.612, an increase of 15.9% compared to the base line Transformer. Ablation and sensitivity studies validate each component's contribution and reveal emergent optimal model structures.