Paper: arXiv 2505.22678
Authors: Jiahao Yang, Ran Fang, Ming Zhang, Jun Zhou
Abstract
In high-frequency trading (HFT), leveraging limit order books (LOB) to model stock price movements is crucial for achieving profitable outcomes. However, this task is challenging due to the high-dimensional and volatile nature of the original data. Even recent deep learning models often struggle to capture price movement patterns effectively, particularly without well-designed features. We observed that raw LOB data exhibits inherent symmetry between the ask and bid sides, and the bid-ask differences demonstrate greater stability and lower complexity compared to the original data. Building on this insight, we propose a novel approach in which leverages the Siamese architecture to enhance the performance of existing deep learning models. The core idea involves processing the ask and bid sides separately using the same module with shared parameters. We applied our Siamese-based methods to several widely used strong baselines and validated their effectiveness using data from 14 military industry stocks in the Chinese A-share market. Furthermore, we integrated multi-head attention (MHA) mechanisms with the Long Short-Term Memory (LSTM) module to investigate its role in modeling stock price movements. Our experiments used raw data and widely used Order Flow Imbalance (OFI) features as input with some strong baseline models. The results show that our method improves the performance of strong baselines in over 75$% of cases, excluding the Multi-Layer Perception (MLP) baseline, which performed poorly and is not considered practical. Furthermore, we found that Multi-Head Attention can enhance model performance, particularly over shorter forecasting horizons.
Complexity vs Empirical Score
- Math Complexity: 3.5/10
- Empirical Rigor: 6.5/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper applies established deep learning architectures (Siamese networks, LSTM, Multi-Head Attention) with minimal novel mathematical derivations, but rigorously tests on 14 real stocks with specific data (level-II LOB), benchmarks against strong baselines, and reports performance metrics over multiple horizons.
Research Flowchart
flowchart TD
A["Research Goal: Improve LOB Prediction for HFT"] --> B["Input Data: Chinese A-Share Market LOB Data"]
B --> C["Methodology: Siamese Architecture to process Ask/Bid symmetry"]
C --> D["Enhancement: Integrate Multi-Head Attention with LSTM"]
D --> E{"Computational Processing"}
E -->|"Raw Data"| F["Model Training & Evaluation"]
E -->|"OFI Features"| F
F --> G["Key Findings: >75% Baseline Improvement & MHA Benefit in Short Horizons"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes a Siamese architecture that processes ask and bid sides of LOB data with shared parameters, and integrates Multi‑Head Attention with LSTM to improve price‑movement forecasting.
Source sentence
We observed that raw LOB data exhibits inherent symmetry between the ask and bid sides, and the bid-ask differences demonstrate greater stability and lower complexity compared to the original data. Building on this insight, we propose a novel approach in which leverages the Siamese architecture to enhance the performan
Assumes 4
- LOB data has a strong symmetry between ask and bid sides that can be exploited by parameter sharing. data
Source sentence
We found that one of its most special traits is that the data structure has a strong symmetry between the ask and bid sides.
- Price movement patterns are not i.i.d. over long periods, so a rolling‑window backtesting is used. statistical
Source sentence
Since the price movement patterns of stocks are affected by the market and do not obey the assumption of independent and identically distributed (i.i.d.) distribution over a long period, we considered a rolling-window backtesting fashion...
- The three‑second LOB snapshots published by the exchange are sufficient to capture market dynamics for prediction. market
Source sentence
In the A-share market, the exchange publishes the current status of LOBs every three seconds.
- Using raw LOB or OFI features as input does not require additional manual feature engineering. model
Source sentence
No matter whether the original LOB or processed OFI is used as input, the proposed method can obtain a significant performance improvement.
Data
- 14 defense‑industry A‑share stocks order-book unknown
Source sentence
We conducted a research study on the following 14 defense‑industry‑related stocks with good liquidity, from January 6, 2021, to May 13, 2021, and all 14 stocks are listed on the A‑share market: 003026.SZ, 300864.SZ, ...
FI-2010
Costs
Not applicable
Finds 4
- The Siamese architecture improves baseline performance in over 75% of cases (excluding the poorly performing MLP baseline). performance
Source sentence
The results show that our method improves the performance of strong baselines in over 75% of cases, excluding the Multi-Layer Perception (MLP) baseline, which performed poorly and is not considered practical.
- After applying the Siamese network, MSE decreases and R² increases significantly for all baselines, demonstrating better explanatory power. predictive-accuracy
Source sentence
After the Siamese network framework is adopted, the MSE and R2 performances of baselines are significantly improved, especially for R2.
- Multi‑Head Attention further enhances model performance, particularly for short forecasting horizons. performance
Source sentence
We combined the Multi‑Head Attention (MHA) mechanism with the LSTM module and found that MHA could further enhance model performance, particularly over short forecasting horizons.
- The performance improvement holds whether the input is raw LOB data or the OFI feature set. performance
Source sentence
No matter whether the original LOB or processed OFI is used as input, the proposed method can obtain a significant performance improvement.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 46 related papers). Coverage is partial, so this is neither replication nor refutation.
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (57)
- Enhancing stock movement prediction with market index and curriculum learning (2023)
- Curriculum learning in deep neural networks for financial forecasting (2019)
- A Survey on Curriculum Learning. (2021)
- An efficient loss function and deep learning approach for ranking stock returns in the absence of prior knowledge (2024)
- Stock selection via spatiotemporal hypergraph attention network: A learning to rank approach (2021)
- Essential tensor learning for multimodal information-driven stock movement prediction (2023)
- A stock rank prediction method combining industry attributes and price data of stocks (2023)
- Enhancing stock ranking forecasting by modeling returns with heteroscedastic Gaussian Distribution (2025)
and 49 more
- DR-GAT: Dynamic routing graph attention network for stock recommendation (2024)
- Forecasting stock price movement: New evidence from a novel hybrid deep learning model (2022)
- Separating the predictable part of returns with CNN-GRU-attention from inputs to predict stock returns (2024)
- Forecasting stock price movement: New evidence from a novel hybrid deep learning model (2022)
- Stock price prediction based on hybrid CNN-LSTM model (2024)
- High Frequency Trading: Overview of Recent Developments (2016)
- High-Frequency Trading and its Impact on Market Liquidity: A Review of Literature (2020)
- Benchmark Dataset for Short-Term Market Prediction of Limit Order Book in China Markets (2021)
- Deep LOB trading: Half a second please! (2022)
- High-frequency trading in a limit order book (2008)
- A Dynamic Model of the Limit Order Book (2008)
- Statistical Modeling of High-Frequency Financial Data (2011)
- Tracking Order Book Price Densities with Kalman Filters (2011)
- Modelling high-frequency limit order book dynamics with support vector machines (2015)
- Analyzing Performance of High Frequency Currency Rates Prediction Model Using Linear Kernel SVR on Historical Data (2017)
- Machine learning techniques for price change forecast using the limit order book data (2015)
- A Decision Support System for Trading in Apple Futures Market Using Predictions Fusion (2021)
- Order book mid-price movement inference by CatBoost classifier from convolutional feature maps (2021)
- Mid-price prediction based on machine learning methods with technical and quantitative indicators (2018)
- Machine Learning for Forecasting Mid-Price Movements Using Limit Order Book Data (2018)
- Universal features of price formation in financial markets: perspectives from deep learning (2018)
- Forecasting Stock Prices from the Limit Order Book Using Convolutional Neural Networks (2017)
- DeepLOB: Deep Convolutional Neural Networks for Limit Order Books (2019)
- Sequence Classification of the Limit Order Book Using Recurrent Neural Networks (2017)
- Stock Price Forecasting with Deep Learning: A Comparative Study (2020)
- The price impact of order book events (2014)
- Deep LOB trading: Half a second please! (2023)
- Stock Price Manipulation Detection Using Deep Unsupervised Learning: The Case of Thailand (2021)
- Temporal Bag-of-Features Learning for Predicting Mid Price Movements Using High Frequency Limit Order Book Data (2020)
- An Optimal Limit Order Book Prediction Analysis Based on Deep Learning and Pigeon-Inspired Optimizer (2021)
- Deep order flow imbalance: Extracting alpha at multiple horizons from the limit order book (2023)
- Multi-level order-flow imbalance in a limit order book (2018)
- The relevance of features to limit order book learning (2022)
- Attention is All you Need (2017)
- Deep learning for limit order books (2019)
- Order imbalance based strategy in high frequency trading (2015)
- Multi-level order-flow imbalance in a limit order book (2018)
- Are transformers effective for time series forecasting? (2023)
- Adam: a method for stochastic optimization (2014)
- Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks (2019)
- LOBSTER: Limit Order Book Reconstruction System (2011)
- Deep attentive survival analysis in limit order books: Estimating fill probabilities with convolutional-transformers — → 2306.05479
- The short-term predictability of returns in order book markets: A deep learning perspective (2024)
- HLOB–Information persistence and structure in limit order books — → 2405.18938
- Deep Learning with Reinforcement Learning on Order Books. (2024)
- Position Attention Mechanism-based Ensemble Network for Trend Prediction of Limit Order Book (2024)
- Deep Hawkes process for high-frequency market making (2024)
- Lob-based deep learning models for stock price trend prediction: a benchmark study — → 2308.01915
- A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023)
3 of 57 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.