Year: 2025
Type: article
Abstract: While the transformer architecture has demonstrated strong success in natural language processing and computer vision, its application to limit order book forecasting, particularly in capturing spatial and temporal dependencies, remains limited. In this work, we introduce Limit Order Book Transformer (LiT), a novel deep learning architecture for forecasting short-term market movements using high-frequency limit order book data. Unlike previous approaches that rely on convolutional layers, LiT leverages structured patches and transformer-based self-attention to model spatial and temporal features in market microstructure dynamics. We evaluate LiT on multiple LOB datasets across different prediction horizons, LiT consistently outperforms traditional machine learning methods and state-of-the-art deep learning baselines. Furthermore, we show that LiT maintains robust performance under distributional shifts via fine-tuning, making it a practical solution for fast-paced and dynamic financial environments.
Authors: Yue Xiao, Carmine Ventre, Yuhan Wang, Haochen Li, Yuxi Huan, Buhong Liu
Language: English
Field: Decision Sciences
Domain: Social Sciences
Open Access status: gold
Estimated APC paid: $2,353
Funder: UK Research and Innovation
Award: G301723715
OpenAlex ID: W4415115238