Paper: arXiv 2309.00638

Authors: Peer Nagy, Sascha Frey, Silvia Sapora, Kang Li, Anisoara Calinescu, Stefan Zohren, Jakob Foerster

Abstract

Developing a generative model of realistic order flow in financial markets is a challenging open problem, with numerous applications for market participants. Addressing this, we propose the first end-to-end autoregressive generative model that generates tokenized limit order book (LOB) messages. These messages are interpreted by a Jax-LOB simulator, which updates the LOB state. To handle long sequences efficiently, the model employs simplified structured state-space layers to process sequences of order book states and tokenized messages. Using LOBSTER data of NASDAQ equity LOBs, we develop a custom tokenizer for message data, converting groups of successive digits to tokens, similar to tokenization in large language models. Out-of-sample results show promising performance in approximating the data distribution, as evidenced by low model perplexity. Furthermore, the mid-price returns calculated from the generated order flow exhibit a significant correlation with the data, indicating impressive conditional forecast performance. Due to the granularity of generated data, and the accuracy of the model, it offers new application areas for future work beyond forecasting, e.g. acting as a world model in high-frequency financial reinforcement learning applications. Overall, our results invite the use and extension of the model in the direction of autoregressive large financial models for the generation of high-frequency financial data and we commit to open-sourcing our code to facilitate future research.

Complexity vs Empirical Score

  • Math Complexity: 8.5/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper employs advanced state-space models (S5) and detailed statistical metrics (perplexity, correlation), while also using real LOBSTER data and a custom Jax simulator for backtesting.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Generative model for realistic<br>limit order book (LOB) message flow"] --> B["Methodology<br>Token-level autoregressive generative model<br>with structured state-space network"]
  B --> C["Data & Inputs<br>NASDAQ LOBSTER data"]
  C --> D["Computation<br>Custom tokenizer &<br>Jax-LOB simulator"]
  D --> E["Findings & Outcomes<br>Low perplexity & high correlation<br>for mid-price returns"]
  E --> F["Implications<br>Foundation for large<br>financial world models"]

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 the first end-to-end autoregressive generative model that generates tokenized limit order book (LOB) messages.

Source sentence
we propose the first end-to-end autoregressive generative model that generates tokenized limit order book (LOB) messages.

Assumes 3

  • Data is limited to regular NASDAQ trading times on working days between 09:30 and 16:00 US East Coast time. market
    Source sentence
    We only use data during regular NASDAQ trading times, on working days between 09:30 and 16:00 US East Coast time.
  • Only the first four LOBSTER message types are used, ignoring hidden orders and trading halts. data
    Source sentence
    we only use the first 4 types, ignoring hidden orders, which do not affect the visible dynamics of the LOB, and trading halts.
  • Prices are truncated to the range [-999, 999] ticks around the mid-price and order sizes to a maximum of 9999, affecting less than 0.1% of the data. data
    Source sentence
    In practice, these thresholds affect less than 0.1% of the data.

Data

  • LOBSTER · LOBSTER order-book unknown
    NASDAQ equity LOBs (e.g., GOOG, INTC)
    Source sentence
    Using LOBSTER data of NASDAQ equity LOBs

Recognised sources: LOBSTER

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • The model achieves per-token perplexity of 3.63 (std.err. 0.0047) for GOOG and 4.04 (std.err. 0.0043) for INTC. performance
    Source sentence
    Calculating overall per-token perplexity results in a score of 3.63 (std.err. 0.0047) for GOOG and 4.04 (std.err. 0.0043) for INTC.
  • Generated mid-price returns show a Pearson correlation of approximately 0.1 for GOOG and 0.2 for INTC, remaining statistically significant up to 80–100 messages ahead. predictive-accuracy
    Source sentence
    Figure 8 shows a sustained positive correlation for both GOOG (𝜌≈0.1) and INTC (𝜌≈0.2). For INTC, the stock with less liquidity, correlations remain statistically significant at least 100 messages into the future.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 18 related papers). Coverage is partial, so this is neither replication nor refutation.

Cited by 40 papers in the archive (35 by independent authors): DiffLOB: Diffusion Models for Counterfactual Generation in … (used as a baseline); Microstructure Modes -- Disentangling the Joint Dynamics of … (uses its method or data); Limit Order Book Simulation and Trade Evaluation with … (uses its method or data); Limit Order Book Simulations: A Review (uses its method or data); MarketGPT: Developing a Pre-trained transformer (GPT) for … (uses its method or data)

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.