Forecasting the Mid-price Movements with High-Frequency LOB: A Dual-Stage Temporal Attention-Based Deep Learning Architecture

Effectively forecasting the stock mid-price movements based on Limit Order Book (LOB) data is crucial for issuing the right trade instructions in an automated trading market. The LOB data contains much valuable information which reflects the microstructure of the stock market and investor trading behavior. However, there are bear several challenges when utilize LOB to built forecasting model, such as dealing with very large amounts of data with high-dimensional, high-frequency, and underlying nonlinear dependencies relationship. In this paper, we propose a novel deep learning architecture tailored to meet the needs of high-frequency mid-price movements forecasting using complex LOB. Specifically, we introduce dual-stage temporal attention mechanism to repeatedly highlight the most valuable time-dimension information. In the first stage, we adaptively assign the temporal attention weights for input data to emphasize the degree of importance of different moments. In the second stage, we adaptively assign the temporal attention weights for hidden states across all time steps to emphasize the degree of importance of different states. Moreover, the architecture utilizes stacked GRU network to enhance the ability of representation learning. Extensive experiments on two datasets demonstrate the validity and superiority of the proposed architecture. Furthermore, by feature importance analysis we find that the time-sensitive feature set is more important than other feature sets, where the derivation of price and volume is the most important feature in all features. The results provide evidence for further understanding the influence factors of the stock market and give a reference for processing financial time series data with high-frequency characteristics.