Paper: arXiv 2409.02277
Authors: Jiwon Jung, Kiseop Lee
Abstract
Managing high-frequency data in a limit order book (LOB) is a complex task that often exceeds the capabilities of conventional time-series forecasting models. Accurately predicting the entire multi-level LOB, beyond just the mid-price, is essential for understanding high-frequency market dynamics. However, this task is challenging due to the complex interdependencies among compound attributes within each dimension, such as order types, features, and levels. In this study, we explore advanced multidimensional sequence-to-sequence models to forecast the entire multi-level LOB, including order prices and volumes. Our main contribution is the development of a compound multivariate embedding method designed to capture the complex relationships between spatiotemporal features. Empirical results show that our method outperforms other multivariate forecasting methods, achieving the lowest forecasting error while preserving the ordinal structure of the LOB.
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- Empirical Rigor: 4.0/10
- Quadrant: Lab Rats — high math complexity, limited empirical testing
Why this score: The paper introduces advanced attention-based architectures and spatiotemporal embedding techniques, indicating high mathematical complexity. However, while it claims empirical validation with forecasting error comparisons, it lacks specific backtest performance metrics, real-world implementation details, or code/dataset links, placing it on the lower end of empirical rigor.
Research Flowchart
flowchart TD A["Research Goal<br>Predict entire multi-level LOB<br>(prices & volumes)"] --> B["Data Input<br>High-Frequency<br>Limit Order Book Data"] B --> C["Methodology<br>Compound Multivariate Embedding"] C --> D["Computational Process<br>Spatiotemporal<br>Sequence-to-Sequence Model"] D --> E["Outcomes<br>Lowest Forecasting Error<br>Preserved Ordinal Structure"]
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 an attention‑based forecasting model for the full limit order book that uses a novel compound multivariate embedding to capture spatiotemporal relationships.
Source sentence
Our main contribution is the development of a compound multivariate embedding method designed to capture the complex relationships between spatiotemporal features.
Assumes 6
- The number of price/volume levels is fixed at five for both bid and ask sides. data
Source sentence
To maintain consistency, we fix the number of levels, K, at 5 for both bid and ask orders in our experiments.
- Timestamps are standardized to equal 5‑second intervals. data
Source sentence
We standardized the timestamps to equal intervals of 5 seconds during data cleansing.
- Percent‑change transformation makes price series stationary for training. model
Source sentence
In our experiment, we employ percent-change transformation of prices to facilitate training.
- Min‑max scaling is applied variable‑wise to prevent large‑fluctuation variables from dominating the model. model
Source sentence
During training, we also apply min-max scaling to price and volume on a variable-wise basis to ensure that variables with large fluctuations do not disproportionately influence the model.
Show 2 more
- Linear‑attention (Performer) approximates full attention with linear complexity. model
Source sentence
In our experiment, we use the Performer method (Choromanski et al. 2020) to enhance computational efficiency, which approximates attention with linear space and time complexity.
- The ordinal price ordering across LOB levels must be respected during training. model
Source sentence
For any fixed i-th LOB snapshot, the multi-level price structure is given by: pa k1i < pa k2i for all k1 < k2, ...
Data
Costs
Not applicable
Finds 2
- Combining percent‑change and min‑max scaling yields the lowest forecasting error (MSE = 0.0464, MAE = 0.1260) among all tested transformations. performance
Source sentence
the combined approach, using both percent-change and min-max scaling, achieves the lowest error rates with an MSE of 0.0464 and an MAE of 0.1260, outperforming each transformation used individually.
- Overall, the method consistently outperforms existing multivariate forecasting approaches, achieving the lowest forecasting errors while preserving the LOB’s ordinal structure. performance
Source sentence
empirical results demonstrate that our approach consistently outperforms existing multivariate forecasting methods, achieving the lowest forecasting errors while maintaining the ordinal structure of the LOB.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 12 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 (40)
- Attention in natural language processing (2020)
- Fine-tune BERT with sparse self-attention mechanism (2019)
- Context-aware self-attention networks for natural language processing (2021)
- Attention-based models for speech recognition (2015)
- End-to-end attention-based large vocabulary speech recognition (2016)
- Improved training of end-to-end attention models for speech recognition (2018) — link
- Temporal pattern attention for multivariate time series forecasting (2019)
- Multi-horizon time series forecasting with temporal attention learning (2019)
and 32 more
- Multivariate time series forecasting via attention-based encoder–decoder framework (2020)
- Deeplob: Deep convolutional neural networks for limit order books (2019)
- Using deep learning for price prediction by exploiting stationary limit order book features (2020)
- Sequence classification of the limit order book using recurrent neural networks (2018)
- BERT : Pre-training of Deep Bidirectional Transformers for Language Understanding (2019) — link
- Long- Range Transformers for Dynamic Spatiotemporal Forecasting (2022) — link
- Time2Vec : Learning a Vector Representation of Time (2019) — link
- A Survey of Transformers (2021) — link
- Transformer- Based Deep Learning Model for Stock Price Prediction : A Case Study on Bangladesh Stock Market (2022) — link
- Efficient Transformers : A Survey (2022) — link
- Attention is all you need (2017)
- Transformers for Limit Order Books (2020) — link
- Deep Transformer Models for Time Series Forecasting : The Influenza Prevalence Case (2020) — link
- Predicting Stock Price Changes Based on the Limit Order Book : A Survey (2022) — link
- Are Transformers Effective for Time Series Forecasting ? (2022) — link
- Deep Learning for Market by Order Data (2021) — link
- Self- Attentive Hawkes Process
- Transformer-Based Attention Network for Stock Movement Prediction (2022) — link
- Lobster: Limit order book reconstruction system (2011)
- Informer: Beyond efficient transformer for long sequence time-series forecasting (2021)
- Non-stationary transformers: Rethinking the stationarity in time series forecasting (2022) — link
- Reversible instance normalization for accurate time-series forecasting against distribution shift (2021)
- Rethinking attention with performers (2020) — link
- Long short-term memory (1997)
- Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods (2018)
- Modelling high-frequency limit order book dynamics with support vector machines (2015)
- Deep learning for limit order books (2019)
- Deep attentive survival analysis in limit order books: Estimating fill probabilities with convolutional-transformers — → 2306.05479
- ViT-LOB: Efficient Vision Transformer for StockPrice Trend Prediction Using Limit Order Books (2024)
- mWDN-Transformer: Utilizing Cyclical Patterns for Short-Term Forecasting of Limit Order Book in China Markets (2024)
- Reinforcement learning for optimized trade execution (2006)
- Limit order books (2016)
1 of 40 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.