Forecasting stock prices: Deep Learning and Full Order Book Dynamics

This paper examines the effectiveness of deep learning models in short-term stock price forecasting using full order book (FOB) data alongside traditional OHLCV features. Historical prices and order book data for 39 stocks over one month were collected, cleaned, and normalized to build three datasets. Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and hybrid CNN-LSTM models were developed with optimized hyperparameters and trained to predict the next 100 five-minute intervals based on the preceding 100 intervals. Evaluation metrics included mean squared error, mean absolute error, and directional accuracy. Results show that the CNN-LSTM model combining OHLCV and FOB data outperforms other configurations, achieving the lowest prediction errors. Sensitivity analysis using Sobol indices indicates that aggregate liquidity variables and recent time steps are the most influential predictors. These findings demonstrate that integrating granular microstructure data significantly improves forecasting accuracy, providing valuable insights for the design of data-driven trading strategies that capitalize on short-term market dynamics.