Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Representation Learning of Limit Order Book: A Comprehensive Study and Benchmarking

Muyao Zhong, Yushi Lin, Peng Yang

TL;DR

This work tackles the lack of standardized evaluation for LOB representations, addressing challenges of strong autocorrelation, cross-feature constraints, and feature-scale disparity. It introduces LOBench, a standardized benchmark with real China A-share LOB data, unified preprocessing, and three downstream tasks to enable fair comparisons. Through reconstruction-focused experiments across nine architectures and transferability tests, the study demonstrates that robust LOB representations can be decoupled from task-specific objectives and still yield strong performance in prediction and imputation, with notable transferability when using a frozen encoder approach. The framework provides reproducible guidelines, a rigorous evaluation protocol, and a pathway toward more reusable, generalizable LOB analyses across markets and tasks.

Abstract

The Limit Order Book (LOB), the mostly fundamental data of the financial market, provides a fine-grained view of market dynamics while poses significant challenges in dealing with the esteemed deep models due to its strong autocorrelation, cross-feature constrains, and feature scale disparity. Existing approaches often tightly couple representation learning with specific downstream tasks in an end-to-end manner, failed to analyze the learned representations individually and explicitly, limiting their reusability and generalization. This paper conducts the first systematic comparative study of LOB representation learning, aiming to identify the effective way of extracting transferable, compact features that capture essential LOB properties. We introduce LOBench, a standardized benchmark with real China A-share market data, offering curated datasets, unified preprocessing, consistent evaluation metrics, and strong baselines. Extensive experiments validate the sufficiency and necessity of LOB representations for various downstream tasks and highlight their advantages over both the traditional task-specific end-to-end models and the advanced representation learning models for general time series. Our work establishes a reproducible framework and provides clear guidelines for future research. Datasets and code will be publicly available at https://github.com/financial-simulation-lab/LOBench.

Representation Learning of Limit Order Book: A Comprehensive Study and Benchmarking

TL;DR

This work tackles the lack of standardized evaluation for LOB representations, addressing challenges of strong autocorrelation, cross-feature constraints, and feature-scale disparity. It introduces LOBench, a standardized benchmark with real China A-share LOB data, unified preprocessing, and three downstream tasks to enable fair comparisons. Through reconstruction-focused experiments across nine architectures and transferability tests, the study demonstrates that robust LOB representations can be decoupled from task-specific objectives and still yield strong performance in prediction and imputation, with notable transferability when using a frozen encoder approach. The framework provides reproducible guidelines, a rigorous evaluation protocol, and a pathway toward more reusable, generalizable LOB analyses across markets and tasks.

Abstract

The Limit Order Book (LOB), the mostly fundamental data of the financial market, provides a fine-grained view of market dynamics while poses significant challenges in dealing with the esteemed deep models due to its strong autocorrelation, cross-feature constrains, and feature scale disparity. Existing approaches often tightly couple representation learning with specific downstream tasks in an end-to-end manner, failed to analyze the learned representations individually and explicitly, limiting their reusability and generalization. This paper conducts the first systematic comparative study of LOB representation learning, aiming to identify the effective way of extracting transferable, compact features that capture essential LOB properties. We introduce LOBench, a standardized benchmark with real China A-share market data, offering curated datasets, unified preprocessing, consistent evaluation metrics, and strong baselines. Extensive experiments validate the sufficiency and necessity of LOB representations for various downstream tasks and highlight their advantages over both the traditional task-specific end-to-end models and the advanced representation learning models for general time series. Our work establishes a reproducible framework and provides clear guidelines for future research. Datasets and code will be publicly available at https://github.com/financial-simulation-lab/LOBench.
Paper Structure (33 sections, 19 equations, 9 figures, 6 tables)

This paper contains 33 sections, 19 equations, 9 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Representation Learning
  4. Related Surveys
  5. LOBench Framework
  6. Dataset: China A-share Market
  7. Data Preprocessing
  8. Selecting
  9. Temporal Alignment and Cleaning
  10. Normalizing
  11. Segmenting
  12. Labeling
  13. Dataloader Construction
  14. Baseline Models
  15. Downstream Tasks
  16. ...and 18 more sections

Figures (9)

  • Figure 1: A snapshot of LOB and the trading mechanism. Red blocks represent ask (sell) orders, while green blocks denote bid (buy) orders. Both ask and sell orders are sorted descending by price. Orders executed at the best bid and ask prices are highlighted in blue, while canceled orders are masked by slashes.
  • Figure 2: The basic pipeline of representation learning for MTS.
  • Figure 3: The commonly used feature-wise z-score normalization breaks the price-level constraint of LOB, while the Global Z-score does not.
  • Figure 4: The commonly-used feature-wise z-score normalization introduced large deviations among different levels, while the Global Z-score does not.
  • Figure 5: The Data Preprocessing Pipeline
  • ...and 4 more figures