Forecasting of Jump Arrivals in Stock Prices: New Attention-based Network Architecture using Limit Order Book Data
Forecasting of Jump Arrivals in Stock Prices: New Attention-based Network Architecture using Limit Order Book Data
This paper investigates the predictability of return jumps in stock prices using high-frequency limit order book (LOB) data and advanced neural network models. The authors propose a novel CNN-LSTM-Attention architecture that combines convolutional layers for feature extraction, LSTM for temporal memory, and an attention mechanism to focus on the most informative time steps. Using NASDAQ TotalView-ITCH data for five liquid stocks (AAPL, FB, GOOG, MSFT, INTC) over 360 days, they detect jumps via the Lee-Mykland test and extract 144 features from the LOB, including raw prices/volumes, spreads, derivatives, and order intensities. The models are trained to predict whether a jump will occur in the next minute. Results show that the CNN-LSTM-Attention model outperforms standard MLP, CNN, and LSTM baselines in terms of F1-score and AUC. The attention mechanism provides interpretability, indicating that time-sensitive features like order intensities are often more important than raw LOB data. The study also finds that LOB data improves jump prediction for most stocks, but the effect varies by stock, suggesting path-dependence in LOB markets is stock-specific. The findings have implications for market makers and liquidity providers who can benefit from anticipating jumps to manage adverse selection risk.
Highlights
- 1Proposes a novel CNN-LSTM-Attention network for predicting jump arrivals in stock prices using high-frequency limit order book data.
- 2Demonstrates that limit order book data improves jump prediction performance, with stock-specific effects.
- 3Shows that the proposed attention-based model outperforms MLP, CNN, and LSTM baselines.
- 4Provides interpretability via attention weights, revealing the importance of LOB features for jump prediction.
Methods
- MLee-Mykland jump detection test to identify jump locations in high-frequency mid-price data.
- MFeature extraction from limit order book data including raw prices/volumes, time-insensitive statistics, and time-sensitive derivatives/intensities.
- MNeural network architectures: Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and a novel CNN-LSTM-Attention model.
- MAttention mechanism to focus on important temporal features and improve prediction performance.
Results
- RThe proposed CNN-LSTM-Attention model achieves the best performance across all stocks, outperforming MLP, CNN, and LSTM.
- RLimit order book data improves jump prediction performance for most stocks, but the effect is stock-specific (e.g., clear improvement for AAPL, marginal for GOOG).
- RJumps are more frequent in the morning hours and around 2 PM, with an average of about 3 jumps per day per stock.
- RThe attention mechanism reveals that time-sensitive features (e.g., order intensities) are often more important than raw LOB data for jump prediction.
Analyze Paper
Generate insights from "Forecasting of Jump Arrivals in Stock Prices: New Attention-...".