Limit Order Book (LOB) Sequence Classification

Experiments and data preparation for LOB sequence classification using pre-split walk-forward folds and a reference training notebook. Data are stored as LZMA-compressed whitespace matrices (.txt.xz) and loaded into NumPy for sliding-window construction.

Layout

• data/ - compressed training/testing folds (see data/README.md).
• notebooks/ - training and evaluation notebook (see notebooks/README.md).
• scripts/ - dataset utilities (see scripts/README.md).
• url.txt - dataset and reference links.

Quickstart

1. Decompress the dataset files:

   bash scripts/unxz_data.sh

2. Open the notebook:

   jupyter lab notebooks/train.ipynb

Data model (as used by the notebook)

• Each fold is a 2D array data with shape (R, T) loaded via np.loadtxt.
• Columns are timesteps; rows are variables.
• Features: X = data[:num_features, :].T (default num_features=144).
• Labels: y = data[-horizon, :][seq_size - 1:] - 1.
• Sliding window sample i spans timesteps [i, i+seq_size-1] and uses y[i].
• Horizon mapping: 1 -> 100, 2 -> 50, 3 -> 30, 4 -> 20, 5 -> 10 ticks.

See data/README.md for details on file names and folders.

Dependencies

The notebook uses standard scientific Python libraries. Install what you need for your environment, for example:

• Python 3.9+
• numpy, pandas, matplotlib, scikit-learn, tqdm
• torch
• jupyter (lab or notebook)
• unxz (from xz-utils) for decompression

References

url.txt lists the dataset source and related papers or data format references.