[Paper Review] BDLOB: Bayesian Deep Convolutional Neural Networks for Limit Order Books

Zihao Zhang, Stefan Zohren+1 more|arXiv (Cornell University)|Nov 25, 2018

Stock Market Forecasting Methods6 references3 citations

TL;DR

This paper proposes BDLOB, a Bayesian deep convolutional neural network that uses dropout variational inference to predict short-term price movements from limit order book (LOB) data. By leveraging uncertainty estimates from Monte Carlo dropout, BDLOB improves predictive performance and enables risk-aware trading strategies with better profitability and reduced risk compared to non-Bayesian models.

ABSTRACT

We showcase how dropout variational inference can be applied to a large-scale deep learning model that predicts price movements from limit order books (LOBs), the canonical data source representing trading and pricing movements. We demonstrate that uncertainty information derived from posterior predictive distributions can be utilised for position sizing, avoiding unnecessary trades and improving profits. Further, we test our models by using millions of observations across several instruments and markets from the London Stock Exchange. Our results suggest that those Bayesian techniques not only deliver uncertainty information that can be used for trading but also improve predictive performance as stochastic regularisers. To the best of our knowledge, we are the first to apply Bayesian networks to LOBs.

Motivation & Objective

• To integrate Bayesian deep learning into limit order book (LOB) prediction for enhanced uncertainty quantification.
• To evaluate whether dropout-based variational inference improves predictive accuracy and generalization in LOB modeling.
• To design and test trading strategies that use model uncertainty for dynamic position sizing and trade timing.
• To demonstrate that uncertainty-aware strategies outperform standard approaches in terms of profitability and risk-adjusted returns.
• To establish the first application of Bayesian neural networks to limit order book data.

Proposed method

• The model uses a deep convolutional architecture (DeepLOB) with Inception modules and LSTMs to extract temporal and spatial features from raw LOB data.
• Dropout variational inference is applied via Monte Carlo dropout, where multiple forward passes with dropped neurons estimate posterior predictive distributions.
• A dropout rate of 0.2 is applied after the Inception module, with grid search used to calibrate uncertainty estimates.
• Concrete Dropout is also tested on convolutional layers, but all rates converge to zero due to large dataset size and parameter sharing.
• Trading strategies are designed using softmax probabilities and uncertainty-aware thresholds to control position sizing and entry/exit decisions.
• The Bayesian trading strategy uses hyperparameters α (confidence threshold) and β1, β2 (risk controls) to balance profit and risk, with α=0.7 and β1=0.1, β2=0.9 selected via empirical analysis.

Experimental results

Research questions

• RQ1 Can dropout variational inference improve predictive performance in deep learning models for limit order book data?
• RQ2 How can uncertainty estimates from Bayesian deep learning be leveraged to enhance trading strategy design?
• RQ3 Does incorporating uncertainty into position sizing lead to higher risk-adjusted returns compared to standard classification-based strategies?
• RQ4 How do different hyperparameters in the Bayesian trading strategy affect profitability and risk exposure?
• RQ5 Is the proposed Bayesian approach the first successful application of Bayesian neural networks to limit order book prediction?

Key findings

• BDLOB achieves an F1 score of 0.60 and AUC of 0.811, outperforming the baseline CNN (F1: 0.48, AUC: 0.672) and DeepLOB5 (F1: 0.58, AUC: 0.803).
• The Bayesian trading strategy using uncertainty information yields higher normalized daily profits and better Downward Deviation Ratio (DDR) than non-Bayesian strategies.
• BDLOB shows narrower interquartile range and fewer outliers in daily accuracy, indicating more consistent performance across test stocks.
• A higher α value (e.g., 0.8 or 0.9) leads to larger risk or lower profits due to fewer trades and higher sensitivity to prediction errors.
• The choice of β2 significantly affects risk and profit: lower β2 increases profits but also risk, while higher β2 reduces exposure to bad trades but may miss profitable opportunities.
• The model demonstrates that dropout acts as a stochastic regularizer, improving generalization and predictive performance on large-scale LOB data.