Deep limit order book forecasting: a microstructural guide

Area of Science:

• Quantitative Finance
• Machine Learning
• Financial Market Microstructure

Background:

• High-frequency trading relies on accurate price prediction.
• Limit Order Book (LOB) data offers granular insights into market dynamics.
• Assessing deep learning model performance in LOB data requires specialized metrics.

Purpose of the Study:

• To explore the predictability of high-frequency Limit Order Book mid-price changes using deep learning.
• To introduce LOBFrame, an open-source tool for processing LOB data and evaluating deep learning models.
• To propose an innovative framework for assessing the practical utility of LOB price predictions.

Main Methods:

• Utilized cutting-edge deep learning methodologies.
• Developed and released LOBFrame for large-scale LOB data processing.
• Proposed a new operational framework focusing on transaction completion probability.

Main Results:

• Deep learning model efficacy is influenced by stock microstructural characteristics.
• High forecasting power does not directly translate to actionable trading signals.
• Traditional machine learning metrics are inadequate for LOB forecasting assessment.

Conclusions:

• Deep learning shows potential in LOB mid-price prediction, but practical application requires careful evaluation.
• The proposed framework enhances the assessment of prediction practicality beyond standard metrics.
• Academics and practitioners can leverage these findings for informed decisions on deep learning in LOB analysis.