DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (1808.03668v6)

Abstract: We develop a large-scale deep learning model to predict price movements from limit order book (LOB) data of cash equities. The architecture utilises convolutional filters to capture the spatial structure of the limit order books as well as LSTM modules to capture longer time dependencies. The proposed network outperforms all existing state-of-the-art algorithms on the benchmark LOB dataset [1]. In a more realistic setting, we test our model by using one year market quotes from the London Stock Exchange and the model delivers a remarkably stable out-of-sample prediction accuracy for a variety of instruments. Importantly, our model translates well to instruments which were not part of the training set, indicating the model's ability to extract universal features. In order to better understand these features and to go beyond a "black box" model, we perform a sensitivity analysis to understand the rationale behind the model predictions and reveal the components of LOBs that are most relevant. The ability to extract robust features which translate well to other instruments is an important property of our model which has many other applications.

• The paper demonstrates that combining CNN and LSTM layers for LOB data yields superior predictive performance compared to traditional methods.
• It employs convolutional layers and Inception modules to extract spatial features and capture multi-scale temporal interactions effectively.
• DeepLOB achieves higher out-of-sample accuracy with reduced parameter size and faster inference, indicating strong potential for high-frequency trading.

DeepLOB: Deep Convolutional Neural Networks for Limit Order Books

This paper presents an advanced deep learning model, DeepLOB, designed to predict future price movements from limit order book (LOB) data of cash equities. This model utilizes both Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) units to analyze spatial and temporal structures within limit order books. By combining these architectures, DeepLOB outperforms existing methods across various metrics, including accuracy, recall, precision, and F1 score.

Key Components of DeepLOB

1. CNN Architecture: The model employs convolutional layers to efficiently extract spatial features from LOB data, employing small 1x2 filters initially to summarize price and volume information, and later expanding to larger filters to integrate data across multiple levels. The use of zero-padding ensures temporal dimensions are maintained across transformations, and parameter sharing within convolutions reduces the risk of overfitting.
2. Inception Modules: Inspired by the Network-in-Network approach, Inception Modules enable the capture of interactions over multi-scale temporal windows. Distinct 1x1, 3x1, and 5x1 convolutional filters allow for dynamic feature representation, leading to improved predictive accuracy over fixed convolutional architectures.
3. LSTM Networks: Following feature extraction, LSTM units capture sequential dependencies among high-level features. Unlike traditional fully connected layers, LSTM mechanisms significantly reduce parameter size, ensuring model robustness even with extensive input sequences.

Performance Evaluation

The performance of DeepLOB is evaluated on two datasets; the FI-2010 dataset and another consisting of one-year limit order data from the London Stock Exchange (LSE). Across various setups, DeepLOB consistently exhibits superior predictive capabilities compared to traditional and neural network-based approaches. In particular, the model maintains high out-of-sample prediction accuracy and adaptability to instruments not included in the training set, suggesting the extraction of universal features within LOB data.

The paper reports remarkable numerical results under different experimental configurations, with DeepLOB obtaining average accuracies surpassing other methods by substantial margins. In addition, computational efficiency is demonstrated with reduced parameter size and faster inference times, which are particularly relevant for high-frequency trading applications.

Trading Simulation

To assess real-world applicability, DeepLOB participates in a simulated trading environment, producing consistent profits across the testing period of three months. While the model operates under simplified assumptions, such as trading at mid-prices without accounting for fees, the results indicate significant long-term profitability and variance reduction achievable through potential optimizations of trade strategies.

Interpretability

Financial models require transparency to ensure trust and reliability. The sensitivity analysis using LIME reveals that DeepLOB assigns differential significance to various LOB components compared to simpler models. Such evaluations expose active regions untouched by pooling operations, affirming DeepLOB’s detailed focus on influential features that withstand noise prominently present in financial data.

Implications and Future Directions

The demonstrated efficacy of DeepLOB posits significant implications for financial trading strategies and risk assessment models. Its ability to generalize across datasets marks a substantial leap towards universally applicable, automated feature extraction from complex LOB data. Future research may address its integration with reinforcement learning for optimized execution strategies or further use of Bayesian neural networks to incorporate predictive uncertainty.

DeepLOB represents a solid advancement in the application of deep learning to financial markets, promising forward momentum in the predictive accuracy and adaptability of machine learning models within this domain.