DeepLOB: Deep Convolutional Neural Networks for Limit Order Books
DeepLOB: Deep Convolutional Neural Networks for Limit Order Books
The paper introduces DeepLOB, a deep learning model combining convolutional neural networks (CNNs) with Inception modules and Long Short-Term Memory (LSTM) layers to predict short-term price movements from limit order book (LOB) data. The architecture uses convolutional filters to automatically extract spatial features from the LOB's price and volume levels, while Inception modules capture patterns at multiple timescales. LSTM layers then model temporal dependencies in the extracted features. The model is trained end-to-end on raw LOB data without handcrafted features.
DeepLOB is evaluated on two datasets: the public FI-2010 benchmark and a proprietary one-year dataset from the London Stock Exchange covering five liquid stocks. On FI-2010, it outperforms all existing methods, achieving F1 scores up to 83.4% for short horizons. On the LSE data, it maintains stable out-of-sample accuracy over three months and generalizes well to stocks not seen during training, suggesting the model learns universal LOB features. A sensitivity analysis using LIME reveals that the model focuses on order imbalances and price-volume interactions at the best bid and ask levels, consistent with financial intuition. The model's computational efficiency (forward pass in milliseconds) makes it suitable for high-frequency trading applications.
Highlights
- 1Proposes DeepLOB, a novel deep CNN-LSTM architecture for predicting price movements from limit order book data.
- 2Outperforms all existing state-of-the-art algorithms on the FI-2010 benchmark dataset.
- 3Demonstrates robust out-of-sample performance on one year of LSE data across multiple stocks.
- 4Shows transferability to instruments not seen during training, indicating extraction of universal features.
- 5Provides interpretability via LIME sensitivity analysis to identify key LOB components driving predictions.
Methods
- MConvolutional neural network (CNN) with Inception modules to capture multi-scale temporal patterns.
- MLong Short-Term Memory (LSTM) layers to model longer time dependencies.
- MEnd-to-end training with categorical cross-entropy loss and ADAM optimizer.
- MData normalization using z-score with rolling statistics to handle non-stationarity.
Results
- RAchieves F1 scores of 77.66%, 74.96%, and 76.58% for prediction horizons k=10, 50, and 100 on FI-2010 Setup 1.
- ROn FI-2010 Setup 2, achieves F1 scores of 83.40%, 72.82%, and 80.35% for k=10, 20, and 50.
- ROutperforms all compared methods including CNN-I, LSTM, and TABL variants.
- RStable out-of-sample accuracy across 3-month test period on LSE data for multiple stocks.
- RTransfer learning to unseen stocks maintains good prediction performance.
Analyze Paper
Generate insights from "DeepLOB: Deep Convolutional Neural Networks for Limit Order ...".