mythezone/LOBench-A-share-processed

LOBench: A-share data and benchmark code

Data and code for Representation Learning of Limit Order Book: A Comprehensive Study and Benchmarking, by Muyao Zhong, Yushi Lin, and Peng Yang.

Publication status: accepted for publication in IEEE Transactions on Big Data. Publication year, volume, issue, pages, and DOI will be added when available. The code release corresponds to the final manuscript revision dated 7 August 2026.

• Benchmark code and reproduction instructions
• Processed 50-asset panel
• Data format and split definitions
• Citation
• Original GitHub project

Use the code snapshot in this repository for the accepted-version configurations. The original GitHub repository may reflect an earlier stage of the project.

Repository contents

Location	Contents
deidentified_50stock_panel/asset_001.npz through asset_050.npz	Processed, alias-based ten-level A-share LOB windows
deidentified_50stock_panel/manifest.json	Feature order, labels, splits, normalization statistics, and per-asset hashes
deidentified_50stock_panel/checksums.sha256	SHA-256 checksums for the published panel and its documentation
code/	Data preparation, models, training, evaluation, configurations, tests, and reproduction guide
CITATION.bib	Temporary citation for the accepted article
Root-level *_processed.csv files	Earlier processed exports retained for existing users; the commands below use the 50-asset NPZ panel

The panel contains 50 assets and 279,815 windows, with 195,712 training, 28,024 validation, and 56,079 test windows. Every window contains 100 observations and 40 price/volume features. These counts match the accepted manuscript's 50-stock panel.

The Hub's table preview displays the legacy CSV exports. Its train label is a viewer grouping, not the accepted paper's chronological split; use the NPZ panel and manifest for that protocol.

The main protocol retains natural class frequencies. Optional balanced training indices are separate and are not the default training split. Asset tickers and absolute dates are omitted from the NPZ files. The private source-to-alias mapping is not included.

Download and quick check

Use Python 3.12. Download the code and panel without the older CSV exports:

python -m pip install huggingface_hub
hf download mythezone/LOBench-A-share-processed \
  --repo-type dataset --revision tbd-accepted-20261027 \
  --include 'README.md' 'CITATION.bib' 'code/*' 'deidentified_50stock_panel/*' \
  --local-dir LOBench-A-share-processed
cd LOBench-A-share-processed/code
python -m pip install -r requirements.txt
python -m unittest discover -s experiments/tbd_revision/tests -v
python smoke_test.py

The smoke test uses CPU, 24 windows from asset_001, one training epoch, the three primary encoders, and the three prediction settings. It also checks recovery losses and the nine supplementary architecture adapters. It checks that the package runs; it does not reproduce paper scores. Full experiments use the provided configurations and generally require a GPU.

The release tag is tbd-accepted-20261027. The code guide linked above lists full experiment commands and data prerequisites.

Reproduction scope

The public panel supports the 50-stock comparisons and the five-stock experiments that use the corresponding ten-date A-share inputs. The code includes matched direct training / frozen probing / fine-tuning, reconstruction, masked-value imputation, structural diagnostics, architecture coverage, sensitivity experiments, and aggregation scripts.

The five-stock year-end training-history experiment uses separate 2019 data with a common year-end holdout. Its full-year and short-history inputs are not the January 50-asset panel. Preparation and training configurations are included, but those inputs must be obtained separately. FI-2010 and LOBSTER inputs must also be obtained from their providers. This release does not redistribute their raw datasets, the proprietary A-share message feed, or trained checkpoints.

Original experiment directory names are retained for compatible imports. Configurations identify experiments reported in the article; other helper encoders in shared modules are not additional empirical claims.

Citation

Please cite the accepted article when using the data or code. Until the final publication record is available, use:

@unpublished{zhong_lobench_tbd,
  author = {Zhong, Muyao and Lin, Yushi and Yang, Peng},
  title = {Representation Learning of Limit Order Book: A Comprehensive Study and Benchmarking},
  note = {Accepted for publication in IEEE Transactions on Big Data. Publication year, volume, issue, pages, and DOI pending},
  url = {https://huggingface.co/datasets/mythezone/LOBench-A-share-processed}
}

Publication field	Current value
Journal	IEEE Transactions on Big Data
Status	Accepted
Publication year / volume / issue / pages	To be added
DOI	To be added

This citation will be replaced with the final journal citation when available. No DOI or publication date is assumed here.