Paper: arXiv 2509.05107
Authors: Alfred Backhouse, Kang Li, Jakob Foerster, Anisoara Calinescu, Stefan Zohren
Abstract
Simulating limit order books (LOBs) has important applications across forecasting and backtesting for financial market data. However, deep generative models struggle in this context due to the high noise and complexity of the data. Previous work uses autoregressive models, although these experience error accumulation over longer-time sequences. We introduce a novel approach, converting LOB data into a structured image format, and applying diffusion models with inpainting to generate future LOB states. This method leverages spatio-temporal inductive biases in the order book and enables parallel generation of long sequences overcoming issues with error accumulation. We also publicly contribute to LOB-Bench, the industry benchmark for LOB generative models, to allow fair comparison between models using Level-2 and Level-3 order book data (with or without message level data respectively). We show that our model achieves state-of-the-art performance on LOB-Bench, despite using lower fidelity data as input. We also show that our method prioritises coherent global structures over local, high-fidelity details, providing significant improvements over existing methods on certain metrics. Overall, our method lays a strong foundation for future research into generative diffusion approaches to LOB modelling.
Complexity vs Empirical Score
- Math Complexity: 8.0/10
- Empirical Rigor: 9.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced mathematical concepts from diffusion models and stochastic processes, and demonstrates high empirical rigor through rigorous benchmarking on LOB-Bench with Level-2 and Level-3 data, showcasing state-of-the-art performance.
Research Flowchart
flowchart TD
A["Research Goal: Generate & Forecast LOB Data<br>using Diffusion Models"] --> B["Data Preparation: LOB-Bench<br>Level-2 & Level-3 Data"]
B --> C["Methodology: Transform LOB Data<br>into Structured Image Format"]
C --> D["Spatio-temporal Diffusion Model<br>with Inpainting"]
D --> E["Parallel Generation of<br>Long Sequences"]
E --> F{"Evaluation on<br>LOB-Bench Metrics"}
F --> G["Outcomes: SOTA Performance<br>Coherent Global Structures"]
G --> H["Future Foundation for<br>Generative LOB Modelling"]
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 diffusion-based generative model with inpainting that converts limit order book data into a structured image format to generate future LOB states.
Source sentence
We introduce a novel approach, converting LOB data into a structured image format, and applying diffusion models with inpainting to generate future LOB states.
Assumes 3
- The model uses only Level‑2 order‑book data and cannot use Level‑3 message data. data
Source sentence
Unlike LOBS5, our model cannot use Level-3 message data and considers only Level-2 order book data.
- Only the top 10 price levels of the order book are needed, as they contain the majority of useful information. model
Source sentence
They use the top 10 levels of the book, claiming that 80% of the information is contained in the first level of the book and 90% of orders are never filled so more than 10 levels adds unnecessary noise
- Experiments are limited to regular trading hours. market
Source sentence
We restrict the data to regular trading hours
Data
Costs
Not applicable
Finds 4
- The model achieves state‑of‑the‑art performance on GOOG under Wasserstein loss and is competitive under L1 loss. performance
Source sentence
For GOOG, our model is competitive under L1 loss and achieves state-of-the-art performance under Wasserstein loss.
- The model performs significantly better on GOOG than on INTC, likely due to differences in tick size and volatility. performance
Source sentence
Our model performs significantly better on GOOG than INTC, both in perceived visual quality and in comparison to other models later on.
- Diffusion models can generate entire LOB sequences in parallel, allowing fast inference even with few sampling steps. performance
Source sentence
diffusion models are able to generate entire LOB sequences in parallel, perhaps even with very few sampling steps, leading to high throughput for longer prediction horizons.
- Increasing the number of inference steps yields only marginal performance gains, indicating efficient sampling. performance
Source sentence
sample speed can be increased with minimal performance degradation however the model struggles with robustness and generalisability across different stocks and domains.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 34 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
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 (62)
- Language models are few-shot learners (2020)
- Limit order books (2013)
- A Continuous-Time Measurement of the Buy-Sell Pressure in a Limit Order Book Market (2004) — link
- STRATEGIC INTERACTIONS AND NEGATIVE OIL PRICES (2021) — link
- Empirical properties of asset returns: stylized facts and statistical issues (2001) — link
- Hippo: Recurrent memory with optimal polynomial projections (2020)
- Efficiently Modeling Long Sequences with Structured State Spaces (2022)
- Simplified State Space Layers for Sequence Modeling (2023) — link
and 54 more
- Attention is all you need (2017)
- Analysis and Control of Nonlinear Process Systems (2004)
- Machine Learning — Lecture 17: Recurrent Neural Networks (2025) — link
- A Study Of N-Gram And Decision Tree Letter Language Modeling Methods (1998) — link
- Diffusion-lm improves controllable text generation (2022)
- CS294-158 Deep Unsupervised Learning, Spring 2024 (2024)
- Improving Image Generation with Better Captions (2023)
- Sora: Creating Video from Text (2024)
- Sora: A review on background, technology, limitations, and opportunities of large vision models (2024) — link
- Denoising diffusion probabilistic models (2020)
- Sdxl: Improving latent diffusion models for high-resolution image synthesis (2023) — link
- Diffedit: Diffusion-based semantic image editing with mask guidance (2022) — link
- Auto-Encoding Variational Bayes (2014)
- Failure modes of variational autoencoders and their effects on downstream tasks (2020) — link
- Generative adversarial nets (2014)
- Improved techniques for training gans (2016)
- Towards realistic market simulations: a generative adversarial networks approach (2021)
- Conditional generators for limit order book environments: Explainability, challenges, and robustness — → 2306.12806
- Generating synthetic data in finance: opportunities, challenges and pitfalls (2020)
- Generating Synthetic Data in Finance: Opportunities, Challenges and Pitfalls (2020) — link
- Unlocking the power of lstm for long term time series forecasting (2025)
- Enhancing time series momentum strategies using deep neural networks (2019) — link
- MarS: a Financial Market Simulation Engine Powered by Generative Foundation Model — → 2409.07486
- JAX-LOB: A GPU-Accelerated limit order book simulator to unlock large scale reinforcement learning for trading — → 2308.13289
- Generative ai for end-to-end limit order book modelling: A token-level autoregressive generative model of message flow using a deep state space network — → 2309.00638
- LOB-Bench: Benchmarking Generative AI for Finance – an Application to Limit Order Book Data — → 2502.09172
- TRADES: Generating Realistic Market Simulations with Diffusion Models — → 2502.07071
- Conditional Image Synthesis with Diffusion Models: A Survey (2024) — link
- High-resolution image synthesis with latent diffusion models (2022)
- Photorealistic text-to-image diffusion models with deep language understanding (2022)
- Image super-resolution via iterative refinement (2021) — link
- Diffusion models beat gans on image synthesis (2021)
- Learning transferable visual models from natural language supervision (2021)
- Repaint: Inpainting using denoising diffusion probabilistic models (2022)
- Sdedit: Guided image synthesis and editing with stochastic differential equations (2021) — link
- Deeplob: Deep convolutional neural networks for limit order books (2019)
- BDLOB: Bayesian Deep Convolutional Neural Networks for Limit Order Books (2018) — link
- The information content of an open limit‐order book (2008) — link
- Axial-lob: High-frequency trading with axial attention — → 2212.01807
- Deep inception networks: A general end-to-end framework for multi-asset quantitative strategies — → 2307.05522
- Multi-Horizon Forecasting for Limit Order Books: Novel Deep Learning Approaches and Hardware Acceleration using Intelligent Processing Units (2021) — link
- Gradient-based learning applied to document recognition (1998) — link
- Overfeat: Integrated recognition, localization and detection using convolutional networks (2014)
- Deep inside convolutional networks: Visualising image classification models and saliency maps (2013) — link
- Instance normalization: The missing ingredient for fast stylization (2016) — link
- Explaining ai in finance: Past, present, prospects — → 2306.02773
- Diffusers: State-of-the-art diffusion models (2022)
- The price impact of order book events: market orders, limit orders and cancellations (2012)
- ClusterLOB: Enhancing Trading Strategies by Clustering Orders in Limit Order Books — → 2504.20349
- LOBSTER: Limit Order Book Reconstruction System (2011) — link
- Learning to simulate realistic limit order book markets from data as a world agent (2022)
- Rwkv: Reinventing rnns for the transformer era (2023) — link
- Eagle and finch: Rwkv with matrix-valued states and dynamic recurrence (2024) — link
- A stochastic model for order book dynamics (2010)
10 of 62 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.