A compact implementation of the convolutional and recurrent architecture proposed in DeepLOB for short-horizon limit-order-book forecasting.
The model receives a rolling window of bid and ask prices and volumes, then predicts the next mid-price direction. The implementation uses chronological splits and train-only normalization to avoid look-ahead leakage.
Obtain the FI-2010 files separately, then provide a local numeric training file:
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python run.py --data /path/to/Train_Dst_NoAuction_ZScore_CF_7.txt --epochs 3
python -m unittest -vInput rows should contain 40 book features followed by an integer class label. Labels may also use the set -1, 0, 1.
The repository preserves the core architectural idea with smaller windows, width, and training budgets. It is a learning-oriented replication and does not reproduce the paper's reported benchmark configuration or score.