ViT-LOB: Efficient Vision Transformer for StockPrice Trend Prediction Using Limit Order Books

Predicting stock price trends in High-frequency trading (HFT) demands utmost time sensitivity and resource efficiency. Previous research has stacked attention mechanisms with Convolutional Neural Networks (CNNs) to enhance predictive performance. However, such stacked complex network structures exhibit a certain degree of redundancy, resulting in excessive memory consumption and protracted training and inference time, posing challenges for deploying the model on edge intelligence devices. In this paper, we introduce ViT-LOB, a lightweight deep learning model exclusively reliant on transformers and attention mechanisms for forecasting stock price trends using Limit Order Book (LOB) data. Through evaluation on the FI-2010 dataset, our model substantially achieved a minimum reduction of 54% in inference time and over 90% reduction in memory consumption, delivering noteworthy results compared to state-of-the-art methodologies.

Recommended citation:

@INPROCEEDINGS{10547868, author={Liu, Zhihang and Sham, Chiu-Wing and Ma, Longyu and Fu, Chong}, booktitle={2024 10th International Conference on Applied System Innovation (ICASI)}, title={ViT-LOB: Efficient Vision Transformer for StockPrice Trend Prediction Using Limit Order Books}, year={2024}, volume={}, number={}, pages={436-438}, keywords={Performance evaluation;Training;Technological innovation;Sensitivity;Memory management;Redundancy;Predictive models;Deep Learning;Vision Transformer;High-Frequency Trading;Limit Order Book}, doi={10.1109/ICASI60819.2024.10547868}}