Paper: arXiv 2509.00982

Authors: Mateusz Wilinski, Juho Kanniainen

Abstract

In this work we show how generative tools, which were successfully applied to limit order book data, can be utilized for the task of imitating trading agents. To this end, we propose a modified generative architecture based on the state-space model, and apply it to limit order book data with identified investors. The model is trained on synthetic data, generated from a heterogeneous agent-based model. Finally, we compare model’s predicted distribution over different aspects of investors’ actions, with the ground truths known from the agent-based model.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 2.5/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced mathematical concepts including state-space models (S5 architecture) and generative modeling, but relies entirely on synthetic data from an agent-based model rather than real market data or backtesting, with results presented as theoretical distributions rather than performance metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: Imitate Trading Agents in Stock Market"] --> B["Data Input: Synthetic LOB Data<br>(Heterogeneous Agent-Based Model)"]
  B --> C["Methodology: Generative Model<br>Modified State-Space Architecture"]
  C --> D["Computational Process: Model Training"]
  D --> E["Outcome: Predicted<br>Action Distributions"]
  E --> F["Evaluation: Comparison with<br>Ground Truth (Agent-Based Model)"]
  F --> G["Key Finding: Demonstrated<br>Generative Tools for<br>Agent Imitation"]

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 modified generative architecture based on a state‑space model to imitate trading agents using limit order book data.

Source sentence
To this end, we propose a modified generative architecture based on the state-space model, and apply it to limit order book data with identified investors.

Assumes 4

  • The synthetic environment contains 1590 agents divided into 15 groups with distinct trading strategies. model
    Source sentence
    The setting is the same as in the mentioned article, which means there are 1590 agents divided into 15 groups.
  • Only three types of messages (market orders, limit orders, cancellations) are recorded; executions are not recorded. data
    Source sentence
    In the end, we only have three types of messages: (i) market orders, (ii) limit orders, and (iii) cancellations.
  • The model is trained on synthetic data generated from a heterogeneous agent‑based model. data
    Source sentence
    The model is trained on synthetic data, generated from a heterogeneous agent-based model.
  • Estimating conditional probabilities is difficult because the number of possible configurations far exceeds the number of data points. statistical
    Source sentence
    Even with only 500 messages, the number of possible configurations far exceeds the number of data points, which is a problem for estimating conditional probabilities.

Data

  • Synthetic limit order book data (environment from [13]) synthetic unknown
    Source sentence
    The data used for both training and evaluation is generated using the synthetic environment first introduced in [13].

Costs

Not applicable The paper reports no trading or execution result.

Finds 2

  • The predicted distributions of order size for noise traders and chartists align well with the true distributions, though with high variability. predictive-accuracy
    Source sentence
    In all three cases the results nicely align with the true distributions. Although, as before, high variability is observed.
  • High variability is observed in estimated probabilities due to limited message count relative to configuration space. negative
    Source sentence
    We observe high variability in estimated probabilities.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 10 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 (20)

  1. Financial markets as nonlinear adaptive evolutionarysystems (2001)
  2. Agent-based modeling (2012)
  3. 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
  4. ABIDES: Towards high-fidelity multi-agent market simulation (2020)
  5. Fast Agent-Based Simulation Framework with Applications to Reinforcement Learning and the Study of Trading Latency Effects (2021)
  6. JAX-LOB: A GPU-Accelerated limit order book simulator to unlock large scale reinforcement learning for trading — → 2308.13289
  7. Black-box Bayesian inference for economic agent-based models (2022) — link
  8. A simulation analysis of the microstructure ofdouble auction markets (2002)
and 12 more
  1. An empirical behavioral model of liquidity and volatility (2008)
  2. Data-driven economic agent-based models (2024) — link
  3. On learning agent-based models from data (2023)
  4. Calibrating Agent-based Models to Microdata with Graph Neural Networks
  5. MarketGPT: Developing a Pre-trained transformer (GPT) for Modeling Financial Time Series — → 2411.16585
  6. Classifying and Clustering Trading Agents — → 2505.21662
  7. A generative model of a limit order book using recurrent neural networks (2023)
  8. LOB-Bench: Benchmarking Generative AI for Finance–an Application to Limit Order Book Data — → 2502.09172
  9. Simplified state space layers for sequence modeling (2022) — link
  10. Attention is all you need (2017)
  11. Time series analysis by state space methods (2012)
  12. Adam: A method for stochastic optimization (2014) — link

5 of 20 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.