REVIEW 3 major objections 4 minor 18 references
Prospects of Imitating Trading Agents in the Stock Market
T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read Trading-agent imitation gets sizes and directions, misses cancellations
desk verdict Honest proof-of-concept, but the evaluation only checks marginal distributions, so the imitation claim is untested. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is a token-level autoregressive generative model built on the simplified structured state-space layer (S5), a sequence model that processes long sequences in linear rather than quadratic time. Each investor action is encoded as a multi-token message with a 'target' flag marking whether the action came from the imitated agent; the model masks one non-time token from the final message and learns to predict its distribution by cross-entropy. Two input branches — one for the tokenized message stream, one for sparse order book snapshots around the mid-price — are concatenated and passed through six S5 layers before a softmax outputs the vocabulary distribution. The S5 layer is
What would settle it
Using the same synthetic agent-based environment, give one agent group a strategy that depends on a market event older than 500 messages or on a price/volume value beyond the 999 cap, retrain the model, and check whether the predicted next-action distributions diverge from ground truth for that group. Alternatively, compare the model's conditional distribution of the next action given a specific market state with the empirical conditional distribution from the synthetic data; a systematic mismatch would refute the imitation claim.
Extended reading notes
Core claim
On its own terms, the paper's central claim is that a modified state-space generative architecture, originally designed for limit order book prediction, can be used to imitate trading agents by predicting distributions over their actions. Rather than forecasting the next market event, the model predicts the next message of a specific investor, conditioned on the preceding 500 messages and order book snapshots. The experiments on synthetic data show partial success: the predicted distributions of order size and direction align with the ground truth of the agent-based model, while cancellation fractions are consistently underestimated and price-return distributions are too narrow. This partial
Load-bearing premise
The model assumes that the last 500 messages and the token truncations (price differences and volumes capped at 999) preserve enough information to predict an investor's next action; if a trader's decisions depend on older market events or on extreme price/volume values, the predicted distributions will be systematically biased and the imitation claim weakens.
Editorial extensions
If this is right
- If the approach transfers to real data with identified investors, it could generate realistic synthetic trader policies for agent-based market simulators without hand-coded behavioral rules.
- Because the model conditions on the full market context through a linear-time state-space layer, it can be scaled to very long sequences, which matter for capturing the long-memory effects present in order flow.
- The specific failure modes — underestimated cancellations and narrowed returns — identify which aspects of agent behavior are not captured by the current context window and token truncation, giving a concrete target for improved architectures.
- The synthetic-data evaluation protocol (known ground truth from an agent-based model) provides a template for measuring imitation quality that is unavailable when training on anonymous real order flow.
- Successful matching of order size and direction distributions even for noise traders suggests the model reliably learns the stationary marginals of an agent's action distribution from context alone.
Reading between the lines
- A natural next test is to train the same architecture on real order flow with identifiable investor IDs and check whether the conditional action distributions match realized behavior; the paper leaves this untested.
- Varying the context length beyond 500 messages and raising the 999 cap on price/volume tokens would isolate whether the underestimated cancellations and narrow returns stem from limited memory or truncation, a decomposition the paper does not perform.
- If the model were extended to predict waiting times between a trader's actions, it could serve as a plug-in agent policy for fully synthetic market generators, not just a distributional imitator of action types and sizes.
- The narrow return distributions might be correctable by a post-hoc calibration of the predicted probabilities against empirical frequencies, an idea that is consistent with the paper's results but not explored.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes using a token-level state-space generative architecture (S5), originally developed for limit order book prediction, to imitate the behavior of individual trading agents. The model is trained on synthetic data from a heterogeneous agent-based model with 1590 agents divided into 15 strategy groups. For a target agent, the model takes as input a sequence of 500 messages and book snapshots and predicts a distribution over the tokens of the next action. Experiments report violin plots comparing model-predicted marginal distributions of action type, order direction, size, and returns to ABM ground-truth parameters. The results show that order sizes and directions roughly match, while cancellation fractions are systematically underestimated and return distributions are too narrow; the authors also report high variance. The paper is framed as an initial investigation into the prospects of such imitation.
Significance. If the imitation claim were substantiated, this line of work would be relevant to data-driven financial simulation and to extending agent-based models with learned individual behavior. The paper has two clear strengths: (i) the synthetic benchmark provides external ground-truth parameters, so there is no circularity in the evaluation; and (ii) the authors are candid about observed mismatches and the preliminary nature of the results. The significance is currently conditional, however, because the reported validation compares only unconditional marginals and provides no evidence that the model actually uses the context to make agent-specific predictions. A context-free model that emits empirical marginal frequencies could pass the current checks. Thus the paper is a useful proof-of-concept of the architecture, but it does not yet establish the central claim of imitation.
major comments (3)
- [Sections 3 and 4] The central claim is that the model imitates an individual investor's behavior by conditioning on the previous 500 messages and book snapshots. However, all reported evaluation metrics in Section 3 are unconditional marginal distributions: the model's per-token probabilities are averaged over all test samples to produce the violin plots in Figs. 2-4. For noise traders, whose actions are by construction independent of market state, a model that ignores the context entirely and outputs the empirical marginal frequencies would reproduce the matched action-type, size, and direction distributions and could plausibly generate the too-narrow return distribution. The paper provides no conditional likelihood, no calibration check, and no ablation with shuffled or truncated context. Section 4 explicitly postpones conditional evaluation ('A curious next step would be to focus more on the evaluation
- [Section 3, Figs. 2-4] The evaluation is qualitative and lacks uncertainty quantification. The authors themselves state in Section 4 that 'high variability' is observed, yet Figs. 2-4 show only averaged distributions without error bars, confidence intervals, or per-sample dispersion. This makes it impossible to assess whether differences between model and ground truth (e.g., the underestimation of cancellations in Fig. 2 and the too-narrow returns in Fig. 3) are systematic or within sampling noise. Quantitative metrics such as KL divergence, expected calibration error, or per-sample distributional distance, along with their bootstrap intervals, are needed. Without these, the claimed 'results mostly align with expected values' is not supported in a statistically meaningful way.
- [Section 2.2] The training and evaluation protocol masks one token of the last message at a time and predicts a distribution for that token. Consequently, the reported results are per-token marginals: action type, direction, size, and returns are each considered independently. An imitation model that is intended to generate investor behavior should produce coherent multi-token actions, and the current evaluation does not validate the joint distribution over (direction, size, type) for a single action. For example, the model could match each marginal while never generating an action whose tokens are mutually consistent. To support the imitation claim, the paper should either demonstrate joint fidelity (e.g., by sampling full actions and comparing joint statistics) or clearly restrict the claim to per-token marginals.
minor comments (4)
- [Section 2.1] The sentence 'As a result, we have 23 tokens: price is described by two tokens...' is structurally unclear. It should state separately the number of tokens per field and the total vocabulary size. Also, the tokenization of time and time-difference fields is described in a run-on way.
- [Figures 2-4] The captions would benefit from explicitly stating that red crosses/dashed lines are ground-truth ABM values. In Fig. 2, the right panel labels 'Chartists' but actually shows two separate models (momentum and mean reversion); this should be made clearer.
- [Section 2.2] The description of training says sequences are truncated to 500 messages and 500 book snapshots. Given that the paper later notes high variance, a short paragraph justifying this context length (e.g., what fraction of target-agent actions occur within 500 messages of a previous action of the same agent) would be useful.
- [Section 4] The sentence 'Even with only 500 messages, the number of possible configurations far exceeds the number of data points' is ambiguous: '500 messages' is the sequence length, not the training set size. Please rephrase to distinguish context length from dataset size.
Circularity Check
No circularity: the generative model is trained on synthetic data and tested against the agent-based model's known ground-truth distributions; no predicted quantity is defined as, or fitted to, the input.
full rationale
The paper does not derive its claimed predictions from the quantities it predicts. The model is a trained state-space network that takes message/book contexts and outputs token distributions; evaluation compares these model outputs to the fixed ab initio parametrization of the three agent types used to generate the data (Figures 2-4). No parameter of the ABM is fitted to the model, and no reported distribution is defined in terms of the model parameters; the ground truths (e.g., red crosses for noise trader action fractions) are simulation inputs, not fitted outputs. The only self-citation that plays a substantive role is [13], which supplies the synthetic environment. That is a benchmark/data source rather than an unverified premise used to forbid alternatives; the paper's claim is about a model trained on and evaluated against that environment. The adopted architecture is cited from non-overlapping prior work [11,14,15], and no uniqueness theorem or ansatz is imported from the authors' own prior papers to force the modeling choice. Section 4 states that conditional predictions are a 'curious next step' and that the paper reports only unconditional marginals; this is an evaluation gap. A context-free model might also reproduce unconditional marginals, but that underdetermination is an external-validity/correctness concern, not a circular reduction of the prediction to its input. Thus no circular step is exhibited.
Assumptions & free parameters
free parameters (1)
- training hyperparameters =
epochs=50, seq_len=500, embed_dim=128, latent_dim=256, layers=6
assumptions (3)
- domain assumption Truncation at ±999 for price differences and volumes loses negligible information
- domain assumption A 500-message window captures the market context relevant for an agent's next action
- domain assumption The agent-based model's parameters provide the correct ground-truth distributions
Cite this review
Pith. "Pith review of Prospects of Imitating Trading Agents in the Stock Market." pith.science (2026). https://pith.science/paper/XZSVRGUR
@misc{pith2026250900982,
author = {Pith},
title = {Pith review of: Prospects of Imitating Trading Agents in the Stock Market},
year = {2026},
howpublished = {\url{https://pith.science/paper/XZSVRGUR}},
note = {Machine review of arXiv:2509.00982}
}
read the original 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.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[13]
Classifying and Clustering Trading Agents
Mateusz Wilinski, Anubha Goel, Alexandros Iosifidis, and Juho Kanniainen. Classifying and clustering trading agents. arXiv preprint arXiv:2505.21662, 2025
work page Pith review arXiv 2025
-
[11]
Peer Nagy, Sascha Frey, Silvia Sapora, Kang Li, Anisoara Calinescu, Stefan Zohren, and Jakob Foerster. 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. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 91–99, 2023
work page 2023
-
[1]
Financial markets as nonlinear adaptive evolutionarysystems
Cars H Hommes. Financial markets as nonlinear adaptive evolutionarysystems. Quantitative Finance, 1(1):149, 2001
work page 2001
-
[2]
Dirk Helbing. Agent-based modeling. In Social self-organization: Agent-based simulations and experiments to study emergent social behavior, pages 25–70. Springer, 2012
work page 2012
-
[3]
Data-driven economic agent-based models
Marco Pangallo and R Maria del Rio-Chanona. Data-driven economic agent-based models. arXiv preprint arXiv:2412.16591, 2024
arXiv 2024
-
[4]
An empirical behavioral model of liquidity and volatility
Szabolcs Mike and J Doyne Farmer. An empirical behavioral model of liquidity and volatility. Journal of Economic Dynamics and Control, 32(1):200–234, 2008
work page 2008
-
[5]
On learning agent-based models from data
Corrado Monti, Marco Pangallo, Gianmarco De Francisci Morales, and Francesco Bonchi. On learning agent-based models from data. Scientific Reports, 13(1):9268, 2023
work page 2023
-
[6]
Calibrating agent-based models to microdata with graph neural networks
Joel Dyer, Patrick Cannon, J Doyne Farmer, and Sebastian M Schmon. Calibrating agent-based models to microdata with graph neural networks. In ICML 2022 Workshop AI for Agent-Based Modelling
work page 2022
Show all 18 references
-
[7]
Abides: Towards high-fidelity multi- agent market simulation
David Byrd, Maria Hybinette, and Tucker Hybinette Balch. Abides: Towards high-fidelity multi- agent market simulation. In Proceedings of the 2020 ACM SIGSIM Conference on Principles of Advanced Discrete Simulation, pages 11–22, 2020
2020
-
[8]
Fast agent-based simulation framework with applications to reinforcement learning and the study of trading latency effects
Peter Belcak, Jan-Peter Calliess, and Stefan Zohren. Fast agent-based simulation framework with applications to reinforcement learning and the study of trading latency effects. In International Workshop on Multi-Agent Systems and Agent-Based Simulation, pages 42–56. Springer, 2021
2021
-
[9]
Jax-lob: A gpu-accelerated limit order book simulator to unlock large scale reinforcement learning for trading
Sascha Yves Frey, Kang Li, Peer Nagy, Silvia Sapora, Christopher Lu, Stefan Zohren, Jakob Foerster, and Anisoara Calinescu. Jax-lob: A gpu-accelerated limit order book simulator to unlock large scale reinforcement learning for trading. In Proceedings of the Fourth ACM Internat...
2023
-
[10]
A generative model of a limit order book using recurrent neural networks
Hanna Hultin, Henrik Hult, Alexandre Proutiere, Samuel Samama, and Ala Tarighati. A generative model of a limit order book using recurrent neural networks. Quantitative Finance, 23(6):931–958, 2023
2023
-
[12]
Marketgpt: Developing a pre-trained transformer (gpt) for modeling financial time series
Aaron Wheeler and Jeffrey D Varner. Marketgpt: Developing a pre-trained transformer (gpt) for modeling financial time series. arXiv preprint arXiv:2411.16585, 2024
2024 arXiv
-
[14]
Lob-bench: Benchmarking generative ai for finance–an application to limit order book data
Peer Nagy, Sascha Frey, Kang Li, Bidipta Sarkar, Svitlana Vyetrenko, Stefan Zohren, Ani Cali- nescu, and Jakob Foerster. Lob-bench: Benchmarking generative ai for finance–an application to limit order book data. arXiv preprint arXiv:2502.09172, 2025
2025 arXiv
-
[15]
Simplified state space layers for sequence modeling
Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022
2022 arXiv
-
[16]
Time series analysis by state space methods
James Durbin and Siem Jan Koopman. Time series analysis by state space methods. Oxford university press, 2012
2012
-
[17]
Attention is all you need
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017
2017
-
[18]
Adam: A method for stochastic optimization
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5
2014 arXiv
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.