LiT: limit order book transformer.

Area of Science:

• Artificial Intelligence
• Computational Finance
• Deep Learning

Background:

• Transformer architecture shows success in NLP and computer vision.
• Limited application of transformers for limit order book (LOB) forecasting.
• Need for models capturing spatial and temporal dependencies in LOB data.

Purpose of the Study:

• Introduce Limit Order Book Transformer (LiT), a novel deep learning architecture.
• Forecast short-term market movements using high-frequency limit order book data.
• Address limitations in capturing spatial and temporal dependencies in LOB forecasting.

Main Methods:

• Leverage structured patches and transformer-based self-attention.
• Model spatial and temporal features in market microstructure dynamics.
• Utilize high-frequency limit order book data for forecasting.

Main Results:

• LiT consistently outperforms traditional machine learning methods.
• LiT surpasses state-of-the-art deep learning baselines.
• LiT demonstrates robust performance under distributional shifts via fine-tuning.

Conclusions:

• LiT is a novel and effective deep learning architecture for LOB forecasting.
• LiT offers a practical solution for dynamic financial environments.
• The model successfully captures complex market microstructure dynamics.