[Paper Review] The Short-Term Predictability of Returns in Order Book Markets: a Deep Learning Perspective

Lorenzo Lucchese, Mikko S. Pakkanen+1 more|arXiv (Cornell University)|Nov 24, 2022

Stock Market Forecasting Methods8 citations

TL;DR

This paper investigates short-term predictability in high-frequency order book markets using deep learning, introducing a novel volume representation of order book data. Results show that mid-price returns are predictably driven by order book dynamics across multiple stocks and horizons, with the volume representation significantly boosting model performance.

ABSTRACT

In this paper, we conduct a systematic large-scale analysis of order book-driven predictability in high-frequency returns by leveraging deep learning techniques. First, we introduce a new and robust representation of the order book, the volume representation. Next, we carry out an extensive empirical experiment to address various questions regarding predictability. We investigate if and how far ahead there is predictability, the importance of a robust data representation, the advantages of multi-horizon modeling, and the presence of universal trading patterns. We use model confidence sets, which provide a formalized statistical inference framework particularly well suited to answer these questions. Our findings show that at high frequencies predictability in mid-price returns is not just present, but ubiquitous. The performance of the deep learning models is strongly dependent on the choice of order book representation, and in this respect, the volume representation appears to have multiple practical advantages.

Motivation & Objective

• To investigate whether high-frequency returns in order book markets are predictably driven by order book data, challenging the assumption of market efficiency at ultra-high frequencies.
• To evaluate the impact of different order book data representations on deep learning model performance for return prediction.
• To assess the feasibility of multi-horizon and multi-stock generalization in deep learning models for return prediction.
• To provide a formal statistical inference framework using model confidence sets to rigorously evaluate predictive performance.

Proposed method

• Proposes a new order book representation called the 'volume representation,' which encodes queue volumes at each price level across bid and ask sides over time.
• Employs a deep learning architecture based on convolutional and recurrent layers, including an inception module with multiple temporal filter sizes to capture multi-scale dynamics.
• Uses a sequence-to-sequence model with attention mechanisms to enable multi-horizon prediction of mid-price return directions (up, down, flat) over future time steps.
• Applies model confidence sets (MCS) to formally compare predictive performance across models and representations, ensuring statistical robustness.
• Trains and evaluates models on high-frequency Nasdaq data across multiple stocks, using a standardized deep learning pipeline with shared hyperparameters.
• Integrates temporal convolutional networks and long short-term memory (LSTM) units to model both local and long-term dependencies in order book dynamics.

Experimental results

Research questions

• RQ1 Does order book data in high-frequency markets contain statistically significant predictability for future mid-price returns, and if so, how far ahead can this predictability extend?
• RQ2 How does the choice of order book representation—specifically, the proposed volume representation—impact the predictive performance of deep learning models?
• RQ3 Can a single deep learning model effectively predict returns across multiple time horizons (multi-horizon modeling), and does this improve performance compared to single-horizon models?
• RQ4 Is there evidence of universal trading patterns across different stocks, such that a single model can generalize effectively across multiple equities?

Key findings

• High-frequency returns in order book markets exhibit widespread and statistically significant predictability, with performance consistently outperforming random chance across all tested stocks and horizons.
• The volume representation of the order book significantly outperforms traditional representations in terms of predictive accuracy, demonstrating robustness and generalization across multiple stocks.
• Multi-horizon modeling using a sequence-to-sequence architecture with attention improves predictive performance compared to single-horizon models, suggesting that modeling future return trajectories jointly enhances learning.
• A single deep learning model trained on one stock generalizes reasonably well to other stocks, indicating the presence of universal, cross-asset trading patterns in order book dynamics.
• Model confidence sets confirm that models using the volume representation are statistically superior to alternatives, with p-values indicating strong confidence in the superiority of this representation.
• The performance gap between models using different representations is substantial, with the volume representation achieving a mean classification accuracy of approximately 58% across horizons, significantly above the 50% baseline.