Benchmark Dataset for Mid-Price Forecasting of Limit Order Book Data with Machine Learning Methods

1. Introduction

• The dataset is derived from NASDAQ OMX Nordic ITCH messages, and its complete market-wide history covers ten trading days.The paper also supplies a day-based anchored forward cross-validation protocol for evaluating prediction methods.Source: §1 · p. 2
• The benchmark combines approximately 4,000,000 trading events across five stocks and ten consecutive days with five mid-price prediction tasks and baseline regression models.The authors present it as the first publicly available dataset for LOB-based HFT method evaluation.Source: §1 · pp. 1–3
• The benchmark addresses the absence of detailed public datasets for evaluating HFT methods using limit-order-book information.The authors intend the dataset and baselines to serve as a testbed for future work.Source: §1 · p. 2
• The paper connects LOB analysis to identifying stable-market circumstances for liquidity providers, model selection by speculative traders, and possible future spoofing analysis.The authors also identify supervisors and regulators among the research's relevant audiences.Source: §1 · p. 3
• The authors provide ridge-regression and RBF neural-network baselines, while explicitly limiting the work to machine-learning metric prediction rather than trading strategies.They describe the paper as an empirical exploration of HFT and machine-learning challenges.Source: §1 · p. 3
• The NASDAQ Nordic setting avoids the liquidity fragmentation found across multiple U.S. exchanges, which complicates comparisons between limit order books.The Helsinki Exchange is described as a pure limit order market where market makers have a limited role.Source: §1 · p. 3

2. Machine Learning for HFT and LOB

• The review frames mid-price prediction as a traditional time-series task that nonetheless has challenges distinctive to HFT.It situates regression, neural networks, and other machine-learning methods within this broader problem setting.Source: §2 · pp. 3–4

2.1. Regression Analysis

• Regression studies apply linear, logistic, and other regression methods to price movements, volume, transaction costs, liquidity, order-book information, and trading activity.Examples include LASSO for variable selection and autoregressive models for price and order-book analysis.Source: §2.1 · pp. 4–5

2.2. Neural Networks

• Neural-network research spans feedforward, recurrent, and other architectures for price prediction, order-book modeling, quote-stuffing detection, currency forecasting, and trading applications.The reviewed work also reports input-volume limitations for a neural-network approach.Source: §2.2 · pp. 5–6

2.3. Maximum Margin and Reinforcement Learning

• Maximum-margin studies use SVMs for mid-price and spread-crossing prediction, while reinforcement-learning work addresses market making, trade execution, and trader or strategy identification.Reviewed approaches include bagged decision trees alongside SVMs and inverse reinforcement learning.Source: §2.3 · p. 6

2.4. Additional Methods for HFT and LOB

• Other reviewed approaches model order flow, trading opportunities, order placement, trade types, microstructure noise, and market dynamics.Examples include stochastic order-flow models, Gaussian classifiers, queueing systems, clustering, fuzzy systems, and Bayesian methods.Source: §2.4 · pp. 6–7

3.2. Limit Order and Message Books

• The market-wide ITCH feed records submissions, trades, cancellations, timestamps, and administrative messages; add and modify messages are used to process limit order books.Source: §3.1 · pp. 7–8, §3.2 · p. 9

3.1. Data Description

• Five Helsinki-listed stocks were collected from NASDAQ OMX Nordic ITCH over ten days, retaining 10:30–18:00 events to avoid opening- and closing-period dynamics.Source: §3.1 · pp. 7–9

3.3. Data Availability and Distribution

• The research community receives normalized features, which prevent reconstruction of the original NASDAQ data.Source: §3.3 · p. 11

3.4. Experimental Protocol

• Anchored forward cross-validation adds one training day per fold and rolls a one-day test window; accuracy, recall, precision, and F1 are averaged across folds with standard deviations.F1 is emphasized because the classes are unbalanced.Source: §3.4 · p. 11
• Event-based sampling groups ten consecutive events per representation, avoiding irregular elapsed-time gaps and yielding 394,337 144-dimensional samples.The features cover raw 10-level bid/ask data, limit order book state, and time-aware information-edge features.Source: §3.4 · pp. 11–12
• Three dataset variants use z-score, min-max, or decimal-precision normalization.Source: §3.4 · pp. 12–13
• Labels cover 1, 2, 3, 5, and 10 future events, assigning upward movement at ≥0.002, stationary movement from -0.00199 to 0.00199, and downward movement at ≤-0.002.Source: §3.4 · pp. 13–14

4. Existing Datasets Described in the Literature

• Prior datasets span public, academic-request, and subscription access, with subscription sources including Kercheval & Zhang (2015), Li et al. (2016), and Sirignano (2016).The paper lists Dukascopy and truefx as public sources and several studies as request-only sources.Source: §4 · p. 14
• Most listed datasets use millisecond resolution, but equities coverage of this type was absent in the authors’ review; truefx instead offers FX tick data for currency pairs.The described truefx files contain timestamps, bid and ask prices, and about 200,000 events per day.Source: §4 · p. 14, Table 5 · p. 14
• Request-only studies vary in accessibility and detail: some provide timestamps and prices without event details or feature vectors, while other ITCH data are not directly accessible for baseline testing.The paper says its dataset supplies full limit-order-book depth, ready for use and comparison with baselines.Source: §4 · pp. 14–15
• Kercheval & Zhang (2015) used 4,000 samples, while the paper notes that its own database contains 100,000 daily events for one stock and that some cited datasets are not public.The review also includes TAQ and Tel-Aviv datasets used outside machine-learning methods and requiring subscriptions.Source: §4 · p. 15

5. Baselines

• The baseline suite uses ridge regression and SLFN regression models on the data representations described in Section 3.4.The authors provide these baselines for their new high-frequency limit-order-book dataset.Source: §5 · p. 15

5.1. Ridge Regression (RR)

• Ridge regression learns a linear mapping from 144-dimensional event vectors to three-class target vectors, using a ridge-regularized fit.The paper notes that one solution form is computationally expensive for the dataset size and recommends the alternative form.Source: §5.1 · pp. 15–16

5.2. SLFN Network-based Nonlinear Regression

• Because kernel-based regression is too computationally intensive, the SLFN uses K-means prototypes as hidden weights and maps inputs through radial-basis-function features.It fits output weights from hidden-layer representations and classifies each test sample by its maximal network output.Source: §5.2 · pp. 16–17

6. Results

• The first experiments excluded the auction period and covered five label scenarios under raw data and three normalization setups.The evaluated metrics were accuracy, precision, recall, and F1 score.Source: §6 · p. 18
• For label 5, SLFN F1 was 33% unnormalized, 46% with Z-score, 43% with min-max, and 43% with decimal precision; RR recorded 40%, 43%, 42%, and 40%, respectively.These values are reported across the unfiltered and three normalized representations.Source: Table 6 · p. 18, Table 7 · p. 18, Table 8 · p. 19
• F1 was higher for label 5 than label 1 regardless of decoding method, with about 43% versus 27% under unfiltered, min-max, and decimal-precision conditions.The authors interpret label 5 as a prediction further into the future than label 1.Source: §6 · pp. 19–20
• RR performed better on raw data and Z-score than SLFN in F1, while SLFN was similar or better with min-max and decimal-precision normalization.The paper attributes possible differences to RR’s ridge-parameter complexity control and SLFN’s greater flexibility and overfitting risk.Source: §6 · p. 20

• For both regression methods, average out-of-sample F1 was approximately 46%, which the authors describe as promising evidence that machine learning can predict mid-price movement.Source: §7 · p. 20
• The benchmark uses NASDAQ ITCH data for five stocks across ten consecutive trading days and defines five mid-price movement tasks at 1-, 2-, 3-, 5-, and 10-step horizons.Source: §7 · p. 20
• Potential applications include predicting market stability for liquidity provision, forecasting market movements for speculative trading, and identifying order-book spoofing.Source: §7 · p. 20
• The dataset has no spoofing-activity information, though its large corpus may support analysis of stock-market patterns as normal or abnormal.Source: §7 · p. 20