The Short-Term Predictability of Returns in Order Book Markets: A Deep Learning Perspective, December 2022#
2. Affiliation:#
Department of Mathematics, Imperial College London (Lorenzo Lucchese and Almut E. D. Veraart)
3. Keywords:#
deep learning, high-frequency finance, model confidence sets, order book, predictability
4. Urls:#
arXiv:2211.13777v2 [q-fin.CP] (Url), Github: None
Summary:
(1): The background of this article is to study the predictability of returns in high-frequency finance with order book data and leverage deep learning techniques.
(2): Past methods in literature have either focused on short predictability horizons or have not considered model uncertainty. However, this paper proposes using model confidence sets to address model uncertainty and investigates multi-horizon predictability. The approach is motivated by the ubiquity of predictability in mid-price returns at high frequencies in financial markets.
(3): The research methodology proposed in this paper includes introducing a new and robust representation of the order book called the volume representation and using model confidence sets to address model uncertainty. Deep learning techniques are also leveraged to learn nonlinear features from the order book data.
(4): The methods in this paper achieve high prediction accuracy for mid-price returns up to 500ms ahead. The performance supports the goal of improving predictability and implies potential profitable trading strategies.
Methods:
(1): The article proposes a methodology to study the predictability of returns in high-frequency finance with order book data using deep learning techniques. The approach investigates multi-horizon predictability and addresses model uncertainty by introducing model confidence sets. The novel representation of the order book called the volume representation is also introduced to model the complex and high-dimensional data.
(2): The methodology uses deep learning techniques, specifically convolutional neural networks (CNNs), to learn non-linear features directly from the volume representation of the order book data.
(3): Model confidence sets are then used to address model uncertainty and obtain prediction intervals for the mid-price returns. The article uses various statistical tests to evaluate the validity of the model confidence sets and obtain meaningful predictions.
(4): The methodology achieves high prediction accuracy for mid-price returns up to 500 ms ahead, supporting the goal of improving predictability and implying potential profitable trading strategies. Ultimately, the proposed methodology provides a new perspective for short-term predictability in order book markets with deep learning techniques and model uncertainty analysis.
7. Conclusion:#
(1): This article is significant in its exploration of the predictability of returns in high-frequency finance using deep learning techniques and model uncertainty analysis. The novel approach investigated multi-horizon predictability and used model confidence sets to address model uncertainty, ultimately achieving high prediction accuracy for mid-price returns up to 500 ms ahead.
(2): Innovation point: This article introduces a novel representation of the order book data called the volume representation and uses deep learning techniques to learn non-linear features directly from the data.
(3): Performance: The proposed methodology achieves high prediction accuracy for mid-price returns up to 500 ms ahead, supporting the potential profitability of the trading strategies.
(4): Workload: The methodology requires a significant amount of computational resources and expertise in deep learning techniques and statistical analysis, potentially limiting its accessibility to a broader audience.