Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

LOB-Bench: Benchmarking Generative AI for Finance -- an Application to Limit Order Book Data

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper builds a distributional benchmark for generative models of limit order book data and reports that the autoregressive LOBS5 model beats both a parametric baseline and other generative models on most realism scores.

desk verdict Useful, honest benchmark package; the SOTA comparison is confounded by unequal training budgets and periods, so the headline ranking should be taken with salt until re-run under matched conditions. read the letter →

arxiv 2502.09172 v2 pith:MSREZ4S5 submitted 2025-02-13 cs.LG cs.CEq-fin.CPq-fin.TR

classification cs.LGcs.CEq-fin.CPq-fin.TR
keywords limitorderbookgenerativeAIbenchmarkdistributionalevaluationmarketimpactautoregressivemodelssyntheticfinancialdatastatespace
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes LOB-Bench, a benchmark that evaluates generative models of limit order book data by comparing the distributions of real and model-generated order-flow statistics. It maps high-dimensional message sequences to one-dimensional scores, then measures L1 and Wasserstein-1 distances between histograms, adding conditional scores, error-accumulation curves, an adversarial discriminator, and price-impact response functions. On Google and Intel stock data, the paper reports that LOBS5, a scaled autoregressive state-space model, achieves the lowest distributional errors overall and reproduces price-impact curves much better than the parametric baseline, while all models' errors grow with generation length. The authors argue the framework replaces qualitative stylized-facts checks with a quantitative, extensible way to rank models and guide development.

What carries the argument

The load-bearing machinery is the scoring-function framework: aggregators map messages and order book states to scalar quantities such as spread, book volumes, imbalance, inter-arrival time, and order levels; histograms of these scores are estimated with variable-width binning and compared using the L1 (total variation) and Wasserstein-1 distances. Conditional versions of the same comparison, with one score conditioned on decile buckets of another, expose how models fail in particular market contexts, and error divergence conditional on the sampling step quantifies the autoregressive trap. A trained discriminator network acts as a worst-case scoring function, and price-impact response functions from the market microstructure literature measure whether generated data responds to order-flow events the way real data does.

What would settle it

Pair each model's aggregate LOB-Bench score with the change in mid-price prediction F1 when its synthetic data is added to training: if a better benchmark score does not go with equal or better downstream F1, then distributional closeness on these statistics is not a sufficient proxy for practical quality.

Watch

Extended reading notes

Core claim

The central claim is that LOB-Bench provides a workable, distributional way to rank generative models of message-level order flow, and that under this ranking the autoregressive S5 model LOBS5 beats a parametric baseline, a conditional GAN, and RWKV sequence models on most scores. The paper further reports two systematic findings: distributional error grows with the number of sampled steps for every model, and none of the tested models' synthetic data improves a downstream mid-price classification task, though the distributional scores give a clearer separation of model quality than that task does.

Load-bearing premise

The benchmark's ranking only measures how close generated and real distributions are on its chosen statistics; the paper assumes without external validation that this closeness is what makes a model useful for downstream applications such as counterfactual simulation or strategy testing.

Editorial extensions

If this is right

  • Models can be ranked by a single aggregate distributional score with confidence intervals, enabling reproducible comparisons.
  • Error-divergence curves expose how quickly each model derails as it generates longer sequences, pointing at the autoregressive trap.
  • Discriminator scores provide a hard-to-game upper bound on realism that even the best current model fails by a clear margin.
  • The benchmark's pluggable scoring functions make it extensible to other high-dimensional event-stream domains, not only equities.
  • Current generative LOB models do not yet help downstream mid-price prediction; the paper treats this as a minimum viable bar that future models should clear.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's own F1 experiment suggests distributional closeness may not translate into practical value: a model that wins LOB-Bench could still be useless for counterfactual strategy testing until downstream correlations are demonstrated.
  • A sharper test of the benchmark would be to see whether models ranked higher on LOB-Bench also produce more realistic responses in a simulated execution task, for instance matching the square-root market impact law the authors cite as future work.
  • Because the discriminator still separates LOBS5 output with ROC 0.83, conditioning discriminator scores on perturbation level could turn adversarial measurement into a more informative continuum rather than a pass/fail.
  • The same framework could be applied to transaction data, forex quote streams, or decentralized exchange order flow, where similar realism questions arise.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper proposes LOB-Bench, an evaluation framework for generative models of limit order book message streams. The framework computes L1 and Wasserstein-1 distances between real and generated distributions of numerous LOB statistics (spread, volumes, imbalance, inter-arrival times, etc.), both unconditionally and conditionally on other scores, along with market impact response functions and an adversarially trained discriminator score. The authors benchmark five models – LOBS5 (an S5-based autoregressive model), RWKV-4/RWKV-6, the Coletta et al. (2022) conditional GAN, and the Cont et al. (2010) parametric baseline – on GOOG and INTC data, and report that LOBS5 achieves the best aggregate scores and best reproduces impact curves. The paper also reports a mid-price prediction experiment that shows no improvement when generated data is added to training.

Significance. The benchmark is a timely and potentially useful tool for the community: it is open-source, flexible, includes standard financial statistics, and provides quantitative distributional scores rather than qualitative stylized facts. The authors include a bin-size sensitivity analysis and make code and generated data available. However, the headline comparative claim ('autoregressive GenAI approach beats traditional model classes') is not supported by the experiments as designed, because the models are trained on different time periods and with very different compute budgets. The paper's own downstream validation does not show practical benefit of the generated data. The benchmark's core machinery is sound, but the evaluation protocol and the strength of the claims need substantial revision.

major comments (3)
  1. [Section 6 and Appendices B–C] The model comparison is confounded by non-uniform training windows and compute budgets. LOBS5 and the RWKV models are trained on all of 2022, while the Coletta model is trained on three days in January 2019 and tested on the next three days (as reported in Section 6). Since the benchmark scores are distributional distances to January 2023 test data, a model trained on 2022 has a large in-domain advantage. Additionally, LOBS5 uses 30.4 L40 days of training (Appendix B) versus 10 L40S days for all four RWKV models (Appendix C). The abstract and Section 6 claim that LOBS5 is state-of-the-art and that the autoregressive GenAI approach beats traditional model classes, but this conflates model class with training data recency and compute. A controlled comparison (e.g., retraining Coletta on 2022 data, or at least matching the test window) or substantially weakened claims are required.
  2. [Section 6, Figure 7] The impact response analysis omits Coletta and RWKV curves because they 'quickly diverge due to error accumulation in inference.' This selective reporting removes two of the five models from a headline metric and cannot be used as evidence for the superiority of LOBS5 on this metric. If divergence occurs, it should be quantified and included (e.g., as a separate 'derailment' score, consistent with the benchmark's own divergence analysis in Figure 18), or the omission criterion should be specified in advance.
  3. [Section 4.3 and Figure 8] The mid-price prediction experiment, the paper's only external validation, shows that adding generated data from any model never improves F1 scores and often degrades them. The paper acknowledges this, but it undercuts the claim that the benchmark measures 'realism' in a sense that is useful for downstream applications. The discriminator ROC of 0.83 for LOBS5 (Section 4.2) further indicates that even the best model is clearly distinguishable from real data. The paper should either provide a different external validation (e.g., matching stylized facts, execution cost in a simulator) or explicitly state that the benchmark measures distributional fidelity on a fixed set of statistics, not fitness for downstream use.
minor comments (7)
  1. [Section 3] The sentence 'This pre-processed dataset containst orders' contains a typo; it should be 'contains'.
  2. [Equation (6)] The symbol p_t is used for the mid-price while the paper also uses p for probability densities; please disambiguate the notation.
  3. [Appendix D] The word 'halfing' should be 'halving'.
  4. [Figure 3 caption] The model name 'coletta' should be capitalized consistently as 'Coletta'.
  5. [Section 4.2] The description of the discriminator input dimension is unclear: n=10 gives D=40, but the encoding Psi_i is described as mapping to R^3; please reconcile these numbers.
  6. [Section 4.1] The statement that event lag times are 'distributed uniformly on a logarithmic scale' should be rephrased as 'log-uniformly' or 'uniformly in log-space' for clarity.
  7. [Figure 4] The figure reports mean, median, and IQM alongside bootstrapped confidence intervals; please clarify how the CIs are computed for the median and IQM summaries.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain: the benchmark ranks models against external real-data statistics, so LOBS5's leading score is an empirical result rather than a construct of the scoring equations.

full rationale

The paper's central claim (autoregressive GenAI beats traditional model classes; LOBS5 is state-of-the-art) is produced by computing L1 and Wasserstein-1 distances between distributions of explicit score functions applied to real LOBSTER data and to model-generated data (Section 4, Eqs. 1-4). None of the scoring functions is defined in terms of the LOBS5 model, no parameter is fitted to the test-period real data and then renamed a prediction, and no uniqueness theorem from the authors' prior work is invoked to forbid alternative model choices. The LOBS5 architecture is inherited from Nagy et al. (2023) and inference uses the authors' JAX-LOB simulator, but the benchmark comparison is external to those citations: the winning model must still match real hold-out data. The main weaknesses are experimental-fairness confounds rather than circularity: Section 6 states LOBS5 was trained on 'the entire year of 2022' while 'The Coletta model is trained on three days from January 2019 and tested on three subsequent days', and Appendix B/C disclose 30.4 L40 days for LOBS5 vs 10 L40S days for all four RWKV models; the Coletta INTC results and Coletta/RWKV impact curves are omitted because of divergence. These choices make the headline comparison uncontrolled, but they do not make the derivation equivalent to its inputs. Likewise, the benchmark's own evidence (discriminator ROC 0.83; Figure 8 showing generated data does not improve mid-price F1) is a validity caveat, not a circularity. Accordingly, no specific circular step can be quoted; the score reflects only the minor self-referential element that the benchmark's designers are also the developers of the winning model.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The benchmark introduces no fitted constants; its numbers depend on hand-chosen evaluation design (deciles, lag grid, book depth). The central claim rests on the domain assumption that distributional closeness on these statistics is what matters, which the paper does not independently validate.

free parameters (3)
  • Number of deciles for conditional evaluation = 10
    Chosen by hand for conditional distribution bucketing; affects the granularity of conditional scores and thus the benchmark ranking.
  • Lag times for impact response functions = log-uniform between 1 and 200 ticks
    Chosen to match Eisler et al. (2012); affects the response curve comparison.
  • Order book depth for discriminator input = n = 10 levels
    Chosen for the discriminator state representation (dimension 40); affects adversarial scores.
assumptions (3)
  • domain assumption Distributional similarity between real and generated data on the selected scoring functions is a valid proxy for the realism and usefulness of generative LOB models.
    The entire benchmark framework depends on this premise; Section 4 defines the evaluation this way without external validation.
  • domain assumption The LOBSTER message data and the JAX-LOB simulator faithfully represent true market dynamics, including order splitting behavior.
    LOBS5 inference uses JAX-LOB (Frey et al., 2023), and the paper notes JAX-LOB merges partial fills into single updates, which affects generated data and response functions (Section 6).
  • standard math Standard statistical tools (Freedman-Diaconis binning, bootstrap confidence intervals, Wasserstein-1 under mean-variance normalization) are appropriate for the data scales in question.
    These are standard methods; the paper checks bin-size sensitivity in Appendix D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LOB-Bench: Benchmarking Generative AI for Finance -- an Application to Limit Order Book Data." pith.science (2026). https://pith.science/paper/MSREZ4S5

@misc{pith2026250209172,
  author       = {Pith},
  title        = {Pith review of: LOB-Bench: Benchmarking Generative AI for Finance -- an Application to Limit Order Book Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MSREZ4S5}},
  note         = {Machine review of arXiv:2502.09172}
}
read the original abstract

While financial data presents one of the most challenging and interesting sequence modelling tasks due to high noise, heavy tails, and strategic interactions, progress in this area has been hindered by the lack of consensus on quantitative evaluation paradigms. To address this, we present LOB-Bench, a benchmark, implemented in python, designed to evaluate the quality and realism of generative message-by-order data for limit order books (LOB) in the LOBSTER format. Our framework measures distributional differences in conditional and unconditional statistics between generated and real LOB data, supporting flexible multivariate statistical evaluation. The benchmark also includes features commonly used LOB statistics such as spread, order book volumes, order imbalance, and message inter-arrival times, along with scores from a trained discriminator network. Lastly, LOB-Bench contains "market impact metrics", i.e. the cross-correlations and price response functions for specific events in the data. We benchmark generative autoregressive state-space models, a (C)GAN, as well as a parametric LOB model and find that the autoregressive GenAI approach beats traditional model classes.

Figures

Figures reproduced from arXiv: 2502.09172 by the authors.

Figure 1
Figure 1. Schematic of LOB-Bench methodology for con￾ditional distributional evaluation 1 arXiv:2502.09172v2 [cs.LG] 16 Jun 2025 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Schematic of the LOB. Not immediately ex￾ecutable orders are placed in First-In, First-Out (FIFO) queues at the specified limit price. Sell limit orders are placed on the Ask side (red) and Buy limit orders on the Bid side (blue). “stylized facts” from the literature, “impact” or the famous “square-root law” (Toth et al. ´ , 2016; Brokmann et al., 2015; Almgren et al., 2005b). However, most of these metrics are unqu… view at source ↗
Figure 3
Figure 3. Model comparison spider plot: the LOBS5 model beats the baseline and coletta model on almost all scores. Note: the radial axis is inverted by plotting the negative loss (larger is better). Based on the evaluation framework outlined in section 4, we develop a Python benchmark package, allowing for a convenient and comprehensive evaluation of generated LOB data. The benchmark is highly customizable, as scoring functio… view at source ↗
Figures from the paper (19 more)
Figure 4
Figure 4. Figure 4: Model score summaries (lower is better). The LOBS5 model achieves the lowest overall scores. Coletta beats the baseline on the Wasserstein metric, but not for L1. Error bars are bootstrapped 99% CIs. The benchmark also evaluates model response functions (6) in aggregat…
Figure 5
Figure 5. Figure 5: LOBS5 results – (left): histogram matching of unconditional score distributions for real and generated data. (right): error accumulation – the further out the prediction horizon, the worse is the model performance – an important model characteristic to measure. bootstr…
Figure 6
Figure 6. Figure 6: L1 distance between real and generated data histograms (incl. 99% CIs). baseline performs well on LOB depth and level-related scores, and much worse on time and volume metrics. LOBS5 dominates L1 loss for GOOG. execution messages and additional resting limit orders, th…
Figure 7
Figure 7. Figure 7: Comparison of tick-normalized mid-price impact response functions for different event types between real and generated data-sets. Shaded regions are 99% confidence intervals. Compared are the LOBS5 and stochastic baseline models. In contrast to the baseline, LOBS5 repr…
Figure 8
Figure 8. Figure 8: Comparison of F1-scores for the MLP classifi￾cation models trained on historical (real) data only (solid lines), and trained on both real data and data generated (dashed lines) for a subset of the generated models. The error bars depict the 95% confidence intervals boo…
Figure 9
Figure 9. Figure 9: Test set (2023 data) loss curves for the LOBS5 model, measuring the mean [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Training loss curves for RWKV model training. The y-axis represents the average negative log likelihood of [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: L1 Divergence Scores with half and double regular bin size. Large bin size deviations (halfing or doubling) [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: L1 and Wasserstein-1 errors of generated unconditional distributions for easy comparison between Alphabet [PITH_FULL_IMAGE:figures/full_fig_p015_12.png]
Figure 13
Figure 13. Figure 13: LOBS5 - histograms comparing score distributions for real (blue) and generated (orange) LOB data for Alphabet (GOOG) and Intel (INTC) stocks. Overall, the generative LOBS5 model evaluated here, adapted from Nagy et al. (2023), does a good job in matching data along va…
Figure 14
Figure 14. Figure 14: baseline - histograms comparing score distributions for real (blue) and generated (orange) LOB data for Alphabet (GOOG) and Intel (INTC) stocks. The (Cont et al., 2010) model does a decent job matching some of the scores, particularly discrete ones, such as depths and…
Figure 15
Figure 15. Figure 15: rwkv4 - histograms comparing score distributions for real (blue) and generated (orange) LOB data for Alphabet (GOOG) and Intel (INTC) stocks. The model produces volatile data with larger spreads, missing correct order levels, leading to difficulty matching book volume…
Figure 16
Figure 16. Figure 16: rwkv6 - histograms comparing score distributions for real (blue) and generated (orange) LOB data for Alphabet (GOOG) and Intel (INTC) stocks. The model has similar shortcomings to RWKV 4 (wrong price levels, mismatched book volumes etc.) due to tokenization of raw dat…
Figure 17
Figure 17. Figure 17: coletta - GOOG - histograms comparing score distributions for real (blue) and generated (orange) LOB data for Alphabet (GOOG) and Intel (INTC) stocks. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_17.png]
Figure 18
Figure 18. Figure 18: L1 error divergence: comparing the L1 errors of score distributions of real data with generated data distributions [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]
Figure 19
Figure 19. Figure 19: Histograms of conditional score distributions for real (blue) and generated (orange) data for the Alphabet stock [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]
Figure 20
Figure 20. Figure 20: Histograms of total book volume conditional [PITH_FULL_IMAGE:figures/full_fig_p023_20.png]
Figure 22
Figure 22. Figure 22: LOBS5 - Histogram of logit scores for real and generated sequences on held-out test data (GOOG). Matching [PITH_FULL_IMAGE:figures/full_fig_p023_22.png]
Figure 23
Figure 23. Figure 23: Comparison of impact response functions for different event types between real and generated data-sets, tick [PITH_FULL_IMAGE:figures/full_fig_p024_23.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Prospects of Imitating Trading Agents in the Stock Market

    q-fin.CP 2025-08 conditional novelty 4.0 of 10

    A state-space generative model trained on synthetic limit order book data can partially imitate trading agent behavior, matching some action distributions while underestimating cancellations, with results limited by h...

Reference graph

Works this paper leans on

23 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [6]

    Defazio, A

    URL http://github.com/google-deepmind. Defazio, A. and Mishchenko, K. Learning-rate-free learning by d-adaptation. The 40th International Conference on Machine Learning (ICML 2023),

  2. [9]

    3https://lobsterdata.com/info/DataStructure.php 13 LOB-Bench: Benchmarking Generative AI for Finance C

    With this larger model, we also successfully removed the explicit error correction mechanism, which originally rejected semantically incorrectly generated messages, as error rates could be sufficiently reduced by scaling the model. 3https://lobsterdata.com/info/DataStructure.php 13 LOB-Bench: Benchmarking Generative AI for Finance C. RWKV Training Details...

  3. [10]

    D. Sensitivity of Divergence Metrics to the Bin Size We adopt a dynamic bin size determined by the Freedman–Diaconis (FD) rule, which is specifically designed to adapt to the underlying data distribution. As a result, we do not anticipate significant sensitivity to the choice of bin size. This choice is further supported by a theoretical convergence prope...

  4. [11]

    org/abs/2406.11903v1

    URL https://arxiv. org/abs/2406.11903v1. Ntakaris, A., Magris, M., Kanniainen, J., Gabbouj, M., and Iosifidis, A. Benchmark dataset for mid-price forecasting of limit order book data with machine learning methods. Journal of Forecasting, 37(8):852–866,

  5. [13]

    Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M., Grella, M., et al

    URL https://ora.ox.ac.uk/objects/uuid: 929fa3fe-4e5f-4cef-ad9f-03eb40110818 . Peng, B., Alcaide, E., Anthony, Q., Albalak, A., Arcadinho, S., Biderman, S., Cao, H., Cheng, X., Chung, M., Grella, M., et al. Rwkv: Reinventing rnns for the transformer era. arXiv preprint arXiv:2305.13048,

  6. [16]

    The square-root impact law also holds for option markets

    T´oth, B., Eisler, Z., and Bouchaud, J.-P. The square-root impact law also holds for option markets. Wilmott, 2016 (85):70–73,

  7. [18]

    Parden, can you repeat that? defending against jailbreaks via repetition

    Zhang, Z., Zhang, Q., and Foerster, J. Parden, can you repeat that? defending against jailbreaks via repetition. arXiv preprint arXiv:2405.07932,

  8. [19]

    12 LOB-Bench: Benchmarking Generative AI for Finance A. Benchmark Code The benchmark code can be found on GitHub at https://github.com/peernagy/lob bench, additional resources can be found on the project website https://lobbench.github.io/. The benchmark suite provides a convenient API functionality to evaluate model data for a range of scoring functions ...

Show all 23 references
  1. [21]

    We divide each dataset into chunks of 16384 tokens, and randomly shuffle these chunks for training

    trained on GOOG 2017 messages, resulting in datasets of 5.5 billion tokens for INTC 2022 (corresponding to 276 million messages) and 7.5 billion tokens for GOOG 2022 (corresponding to 380 million messages). We divide each dataset into chunks of 16384 tokens, and randomly shuff...

  2. [22]

    For stability, we clipped the maximum global gradient norm to 1.0 (Pascanu et al., 2013)

    in Optax (DeepMind et al., 2020), without scaling the learning rate or using any learning rate schedulers. For stability, we clipped the maximum global gradient norm to 1.0 (Pascanu et al., 2013). In total, training all 4 of our RWKV models (2 model architectures, each over 2 ...

  3. [2001]

    doi: 10.1080/713665670

    ISSN 1469-7688, 1469-7696. doi: 10.1080/713665670. URL http://www.tandfonline.com/doi/abs/ 10.1080/713665670. Cont, R., Stoikov, S., and Talreja, R. A stochastic model for order book dynamics. Operations research, 58(3): 549–563,

  4. [2002]

    doi: 10.1088/1469-7688/2/ 5/303

    ISSN 1469-7688, 1469-7696. doi: 10.1088/1469-7688/2/ 5/303. URL http://www.tandfonline.com/ doi/abs/10.1088/1469-7688/2/5/303. Coletta, A., Moulin, A., Vyetrenko, S., and Balch, T. Learn- ing to simulate realistic limit order book markets from data as a world agent. In Proceed...

  5. [2010]

    The price impact of order book events

    Cont, R., Kukanov, A., and Stoikov, S. The price impact of order book events. JOURNAL OF FINANCIAL ECONO- METRICS (Winter 2014), 12(1):47–88,

  6. [2011]

    Kingma, D

    doi: https://doi.org/10.2139/ssrn.1977207. Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

  7. [2013]

    Paulin, J

    URL https: //arxiv.org/abs/1211.5063. Paulin, J. Understanding flash crash contagion and systemic risk: a calibrated agent-based approach. http://purl.org/dc/dcmitype/Text, University of Oxford, January

  8. [2016]

    doi: 10.18653/v1/P16-1162

    Association for Compu- tational Linguistics. doi: 10.18653/v1/P16-1162. URL https://aclanthology.org/P16-1162/. Spooner, T., Fearnley, J., Savani, R., and Koukorinis, A. Market making via reinforcement learning. In Proceed- ings of the 17th International Conference on Autonomo...

  9. [2017]

    org/abs/1706.03762

    URL http://arxiv. org/abs/1706.03762. arXiv:1706.03762. Vyetrenko, S., Byrd, D., Petosa, N., Mahfouz, M., Dervovic, D., Veloso, M., and Balch, T. Get real: realism metrics for robust limit order book market simulations. In Pro- ceedings of the First ACM International Conferenc...

  10. [2018]

    doi: 10.1007/s10588-017-9258-0

    ISSN 1572-9346. doi: 10.1007/s10588-017-9258-0. URL https://doi. org/10.1007/s10588-017-9258-0 . Murphy, K. P.Machine learning: a probabilistic perspective. MIT press,

  11. [2019]

    arXiv:1905.03554

    URL http:// arxiv.org/abs/1905.03554. arXiv:1905.03554. Lecun, Y . and Bengio, Y . Convolutional networks for im- ages, speech, and time-series. In The handbook of brain theory and neural networks. MIT Press,

  12. [2020]

    Dslob: a synthetic limit order book dataset for bench- marking forecasting algorithms under distributional shift

    Cao, D., El-Laham, Y ., Trinh, L., Vyetrenko, S., and Liu, Y . Dslob: a synthetic limit order book dataset for bench- marking forecasting algorithms under distributional shift. arXiv preprint arXiv:2211.11513,

  13. [2023]

    Eagle and finch: Rwkv with matrix- valued states and dynamic recurrence

    Peng, B., Goldstein, D., Anthony, Q., Albalak, A., Al- caide, E., Biderman, S., Cheah, E., Du, X., Ferdinan, T., Hou, H., et al. Eagle and finch: Rwkv with matrix- valued states and dynamic recurrence. arXiv preprint arXiv:2404.05892,

  14. [2024]

    Llacay, B

    URL https: //arxiv.org/abs/2402.02370v2. Llacay, B. and Peffer, G. Using realistic trading strate- gies in an agent-based stock market model. Computa- tional and Mathematical Organization Theory , 24(3): 308–350, September

  15. [6266]

    URL https://www.worldscientific.com/doi/ abs/10.1142/S2382626615500070

    doi: 10.1142/S2382626615500070. URL https://www.worldscientific.com/doi/ abs/10.1142/S2382626615500070. Publisher: World Scientific Publishing Co. Byrd, D., Hybinette, M., and Balch, T. H. Abides: Towards high-fidelity multi-agent market simulation. In Proceed- ings of the 202...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.