The short-term predictability of returns in order book markets: A deep learning perspective

Lorenzo Lucchese, Mikko S. Pakkanen, Almut E. D. VeraartInternational Journal of Forecasting 2022

TL;DR

This paper investigates the predictability of high-frequency stock returns using deep learning models and a novel "volume representation" of order book data. They find that short-term predictability is ubiquitous, and crucially, that the way order book data is represented significantly impacts model performance, with their proposed volume representation showing strong advantages.

What Problem Does This Paper Address?

High-frequency trading (HFT) relies on the ability to predict short-term price movements. While it's known that order book data contains information about future price changes, effectively extracting this information has been a challenge. Traditional methods often rely on handcrafted features, which may not capture the full complexity of the order book. This paper aims to systematically explore the extent of short-term predictability in order book markets using deep learning, focusing on how different data representations and model architectures influence predictive power. They specifically investigate how far into the future predictability exists, the importance of robust data representations, the benefits of multi-horizon forecasting, and whether trading patterns are universal across different stocks.

Methods

The authors employ deep learning architectures, specifically focusing on the deepLOB model and its variants. A key contribution is the introduction of a new order book data representation called "volume representation" (deepVOL). This representation aims to be more robust and compatible with deep learning models compared to traditional "level-based" representations. They also explore order flow (deepOF) as another input feature.

To address their research questions, they utilize a model confidence set (MCS) framework. This statistical inference procedure allows for a rigorous comparison of multiple models and benchmarks, accounting for parameter estimation error. They frame the prediction task as a classification problem, predicting whether the mid-price will go down, stay the same, or go up within a certain horizon. They also investigate multi-horizon forecasting using a sequence-to-sequence decoder architecture.

Datasets

The study uses the LOBSTER dataset, which reconstructs historical order book data from Nasdaq's TotalView-ITCH feed. They analyze one year of data (January 2019 - January 2020) for 115 Nasdaq tickers, selecting a subset of 10 stocks with diverse liquidity characteristics for their experiments.

Key Insights and Ideas

  • Ubiquitous Predictability: Short-term predictability in mid-price returns is not just present but widespread at high frequencies, extending up to 50-300 order book updates ahead.
  • Representation Matters: The choice of how order book data is represented is critical for model performance. Traditional "level-based" representations are outperformed by more robust methods.
  • Volume Representation Shines: The proposed "volume representation" (deepVOL) demonstrates significant advantages, particularly when using universal models (trained across multiple stocks). This representation is more robust to small perturbations and better suited for deep learning architectures.
  • L2 Granularity is Key: Using L2 order book data (all bid/ask prices and volumes) provides a substantial performance boost over L1 data (only the best bid/ask). However, L3 data (including queue information) doesn't consistently offer further improvements, suggesting it might be excessively granular for this task.
  • Order Flow and Volume Outperform: Models using order flow or volume features significantly outperform those using raw order book data.
  • Universal Trading Patterns Exist: The study provides evidence for common trading patterns across different stocks, suggesting that universal models can be effective.

Key Results

  • Predictability was identified for most stocks up to 50 order book events ahead at the 99% confidence level.
  • Models using L2 data were consistently placed in the set of superior models 85-90% of the time when predictability was identified, compared to only 5-11% for L1 data models.
  • DeepVOL models (using volume representation) with L2 or L3 data achieved high performance, being in the superior model set 84-86% of the time at the 99% confidence level.
  • Multi-horizon models using seq2seq decoders generally outperformed their single-horizon counterparts.
  • Universal models trained on a subset of stocks showed strong performance on "out-of-sample" stocks, indicating common trading patterns.

Limitations and Future Work

  • Tradability: The paper acknowledges that the identified predictability is based on mid-price returns, which are not directly tradable. Future work could explore more tradable return definitions (e.g., bid-ask returns).
  • Infrastructure Latency: The study assumes instantaneous access to order book information. In practice, latencies can affect trading. While the authors argue these effects are negligible for their definition of returns, practical applications would need to account for them.
  • Stock Selection: The experiments were conducted on a subset of 10 stocks due to computational constraints. A more comprehensive study with a larger set of stocks could provide stronger evidence for universal patterns.
  • Feature Engineering vs. Deep Learning: While deep learning models show strong performance, further research is needed to definitively determine if their expressivity is superior to carefully engineered features, especially in applications where prediction speed is critical.
  • Market Structure: The study focused on Nasdaq-traded stocks. Exploring other market structures could reveal differences in predictability.
  • Model Tuning: The authors did not perform extensive hyperparameter tuning, suggesting that further optimization could potentially improve model performance.

Summary updated

Abstract

This paper uses deep learning techniques to conduct a systematic large-scale analysis of order book-driven predictability in high-frequency returns. First, we introduce a new and robust representation of the order book, the volume representation. Next, we conduct 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 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.
BibTeX
@article{lucchese2022short,
  author = {Lorenzo Lucchese and Mikko S. Pakkanen and Almut E. D. Veraart},
  title = {{The short-term predictability of returns in order book markets: A deep learning perspective}},
  journal = {International Journal of Forecasting},
  year = {2022},
}
Figure