FI-2010 Limit Order Book Benchmark Dataset

FI-2010 is the reference benchmark for predicting mid-price movement from limit order book microstructure. Each record carries 144 features and five classification labels, shipped in three normalizations (z-score, min-max, decimal precision), with variants that include or exclude auction periods and nine folds of a day-based anchored cross-validation protocol the authors specify so results across papers stay comparable. The paper calls it “the first publicly available benchmark dataset of high-frequency limit order markets for mid-price prediction” (arXiv:1705.03233); it appeared in the Journal of Forecasting 37(8):852–866 (2018).

Getting the canonical copy is awkward. The Fairdata deposit is a single 1.86 GB BenchmarkDatasets.zip, and its download service requires a token issued through the Etsin web interface, so there is no documented stable direct-file URL for a headless script. In practice a headless workflow falls back to a GitHub mirror — the DeepLOB reference implementation carries a 56 MB data.zip — but those downstream repositories ship no license file of their own. The CC BY 4.0 term is the depositor’s declaration in the Fairdata record; because the deposit is a zip rather than a code repository, there is no LICENSE file to read.

The dataset is frozen rather than maintained: its metadata was last modified in 2017 and the file was frozen in 2018. It nonetheless remains the common baseline in deep-learning order book work, with Crossref counting 116 works that cite the paper. One naming note — “FI-2010” is the literature’s shorthand, not the repository’s; the Fairdata record is titled Benchmark Dataset for Mid-Price Forecasting of Limit Order Book Data with Machine Learning Methods.