acceptodds
Under review as a conference paper at ICLR 2027

ReLOBGen: Replayable Limit Order Book Message Generation

Abstract

We propose ReLOBGen, a method for generating limit order book (LOB) messages that are replayable by construction. Replayability is required for closed-loop market simulation, yet existing LOB message generators may produce non-replayable raw messages, i.e., messages inconsistent with the current market state. These generators therefore rely on post-hoc correction or rejection followed by resampling, which may alter the replayed message distribution or increase inference cost. ReLOBGen instead ensures replayability during generation: it selects the referenced order from the resting orders in the current LOB and then generates the remaining message fields to be consistent with that order and the market state. For realistic reference selection, ReLOBGen samples from a learned distribution over eligible resting orders, efficiently computed from cached order representations and a context-dependent query. It then enforces the consistency of the remaining fields by masking out invalid tokens. Together, these components enable efficient generation of realistic messages without post-hoc correction or resampling. In 500-message rollouts, ReLOBGen achieves 100% replayability, improves market realism, particularly for top-of-book statistics and the relative prices of LOB messages, and provides a speedup per replayed message over the LOBS5 baseline.

Primary area:
generative models
Keywords:
generative models, finance, LOB

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7UtDOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The paper describes a method for autoregressive generation of Limit Order Book messages. It restructures the LOBS5 method so that the generated tokens will be guided to produce restricted messages within message-valid borders. The process is interesting, but also restrictive as it guarantees that the messages will not have anything "unexpected", something that is happening often in the market. Experiments are using Google and Intel stocks (2025 for training, and half of January 2026 for validation and the rest of January 2026 for testing). The method is compared only with the baseline LOBS5 (and a re-arranged version of it).

Critical strengths
  • The method targets a practical problem in LOB message generation methods (specifically for the LOBS5 method).
  • Results show that the design of the method works.
  • Experiments include useful diagnostic tests.
Critical weaknesses
  • A general comment here is the actual usefulness of message generation methods. Unlike text generation, the results of which are verified for their truthfulness and informativeness and are useful any time the same/similar semantic input is needed, stochastic message generation provides plausible scenarios which become obsolete once the scenario horizon is fulfilled.
  • The claims of the paper regarding generative performance superiority require thorough comparison with directly related methods (including MarketGPT, MarS, TradeFM, LOBERT).
  • The use of the term "replayability" is arbitrary as its actual meaning is not the meaning the authors intent to use it for.
  • In Table 1, it is not clear why o[size] > 1 for "Cancel" while o[size] > 0 for "Delete".
  • The experiments use only two US stocks with 500-message rollouts and for a very limited period (10 days from January 16 to January 30, 2026).
  • The method does multiple changes in LOBS5 (it changes the token factorization, obtains direct access to the exact resting-order set, uses learned reference selection, and applies constrained decoding). Ablations for all the component-level changes would show which change contributes the most to the difference in the results between the method and its baseline.
Questions for authors
  1. Why use the term replayability instead of the term validity?
  2. Why is o[size]>1 used for Cancel and o[size]>0 for Delete?
  3. Why experiments do not include directly related methods MarketGPT, MarS, TradeFM, and LOBERT?
  4. Does ReLOBGen retain its advantage when LOBS5 and other baselines receive the same explicit order-level state information and use a validity mask?
  5. Can you report results for more equities, different liquidity and tick-size regimes, and longer rollouts?
3 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for the detailed feedback and address each point below. Changes in the revised manuscript are highlighted in blue. For the points that require additional experiments, namely a MarketGPT-style Transformer backbone on an additional stock and longer rollouts, we will provide the additional results in a follow-up response once they become available.

[Clarification] Distinguishing non-replayable messages from unexpected market events

We would like to clarify that all real LOB messages are replayable against the order book at the time they occur. Replayability concerns whether a message can be processed against that book state. Therefore, ensuring replayability only excludes messages that cannot be processed against the current book, such as cancelling a non-existent order or executing more shares than remain. It does not exclude unexpected market events.

[W1] Practical value of LOB message generation

Coupled with an LOB simulator in a closed loop, a message generator can serve as a market world model, simulating responses to an agent’s orders and providing an environment for developing and evaluating trading strategies. For example, MarS analyzes the market impact of injected TWAP orders and trains an order-execution agent in its generated environment, while M3 [1] demonstrates simulation-based stress testing and market-impact analysis. The value of these simulations for strategy development and evaluation does not expire when an individual rollout ends.

This research direction is also studied within the broader machine learning community. MarS (ICLR 2025) develops a market simulation engine based on generative modeling of order flow, while LOB-Bench (ICML 2025) provides a systematic framework for evaluating generative LOB models. TRADES (ECAI 2025) [2] generates order flows with a diffusion model and evaluates market responses to an external trading agent. Related work includes TradeFM, presented at the ICML 2026 Workshop on Foundation Models for Structured Data, and LOBERT [3], presented at the NeurIPS 2025 Workshop on Generative AI in Finance. MarketGPT and more recent work such as M3 [1] also explore generative modeling of market microstructure dynamics.

Within this active area of research, ReLOBGen generates replayable LOB messages by construction without post-processing. In our closed-loop simulation experiments, this approach improves rollout realism and achieves a 2.7–3.6× speedup per replayed message over the LOBS5 baseline.

[1] Zhang et al. (2026). M3: A State-Event Generative Foundation Model for Market Microstructure Dynamics. arXiv:2608.19227.

[2] Berti, Prenkaj, and Velardi (2025). TRADES: Generating Realistic Market Simulations with Diffusion Models. ECAI.

[3] Linna et al. (2025). LOBERT: Generative AI Foundation Model for Limit Order Book Messages. NeurIPS Workshop on Generative AI in Finance.

[W2, Q3] Comparison with related methods

ReLOBGen is a method applied to a field-wise autoregressive generator rather than a competing backbone, and its benefits are measured against existing post-processing on the same generator. To test generality, we are training a MarketGPT-style Transformer backbone on an additional stock and will report the results once they become available. The other methods cannot be compared under matched conditions: TradeFM has no public code, MarS releases inference code and small pretrained models but no training pipeline, and both MarS and LOBERT encode each order or message as a single token, so field-wise reference selection does not apply.

[W3, Q1] The term “replayability”

We follow prior LOB simulation work, where “replay” refers to processing messages through an LOB simulator: LOBS5 calls the component that processes generated messages a “replay simulator” [4], and Nagy et al. [5] use an ABIDES-based simulator to “replay NASDAQ price-time priority limit order book markets”. We prefer “replayability” over “validity” because it directly expresses whether a message can be applied to the current order book through the simulator’s state transition without modification, which we formalize in Definition 1.

[4] Nagy et al. (2023). 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. ICAIF.

[5] Nagy, Calliess, and Zohren (2023). Asynchronous Deep Double Dueling Q-learning for trading-signal execution in limit order book markets. Frontiers in Artificial Intelligence.

AuthorsOpenReview ↗

Response (2/2)

[W4, Q2] Size conditions for cancel and delete

As defined in the first paragraph of Section 3.1, cancel denotes partial cancellation, whereas delete denotes complete removal of an order. A partial cancellation must remove at least one share while leaving at least one share outstanding, so the referenced order must have at least two shares remaining (o[size] > 1), consistent with the cancel constraint 1 ≤ X_t[size] < R_t[size] in Table 1. A deletion removes the entire remaining quantity and therefore only requires a positive remaining size (o[size] > 0).

[W5, Q5] Experimental scope

We matched the experimental setup to the LOBS5 evaluation in LOB-Bench, including the S5 backbone and the 500-message rollout length. We will additionally report results for longer rollouts, as well as on an additional stock as described in our response to W2/Q3, once available.

[W6] Contribution of individual components

We have already evaluated each component according to its purpose, either ensuring replayability or improving rollout realism, as detailed below.

  • Token factorization. Section 5.4 analyzes this change through teacher-forced NLL (Table 5), and Table 2 shows that it alone leaves replayability violations largely unchanged. We introduced reference-first factorization to enable replayability and therefore did not originally report its rollout realism separately. We have now added LOBS5 (ref.-first) to Tables 3 and 4, showing that ReLOBGen improves rollout realism over this baseline despite using the same reference-first backbone.

  • Access to the exact resting-order set. Both methods use the exact resting-order set during rollout; the only difference is that ReLOBGen uses it during generation, whereas LOBS5 uses it during post-processing. Please see our response to Q4 for a detailed discussion.

  • Reference selection and inference-time masking. These two components jointly guarantee replayability: selecting the reference from the eligible set ensures reference eligibility, and inference-time masking ensures event-order compatibility. Since neither alone guarantees replayability, we evaluate them together rather than separately. Section 4.3 reports that inference-time masking excludes only a small fraction of the pretrained model’s probability mass.

  • Learned reference selection. To evaluate the contribution of learned reference selection to rollout realism, Section 5.4 compares learned and uniform selection under the same eligible set and inference-time masking. Both configurations guarantee replayability, but learned selection clearly improves realism (Tables 3 and 4).

[Q4] Comparison under matched information and validity constraints

Both existing methods and ReLOBGen access the exact resting-order set during closed-loop market simulation, as the LOB simulator maintains it to process messages and update the order book. The difference lies in when it is used: existing methods use it to post-process generated messages (Appendix A), whereas ReLOBGen uses it during generation. ReLOBGen’s core contribution is to enable the baseline to use this information during generation through eligible reference selection and inference-time masking, thereby guaranteeing replayability. Tables 2–4 evaluate this approach against LOBS5 (ref.-first), using the same pretrained reference-first backbone, and show that ReLOBGen improves replayability, rollout realism, and computational efficiency.

AuthorsOpenReview ↗

Additional experimental results

[W2, Q3] Extended evaluation across backbones and stocks

Our original evaluation matches the LOBS5 evaluation in LOB-Bench in the number of stocks (two: GOOG and INTC) and rollout length (500 messages). We now extend the evaluation to the MarketGPT Transformer backbone on GOOG and an additional stock, JPM, retaining its model size with minor modifications to the token ordering and context length. For each stock, we compare ReLOBGen with a reference-first baseline using the same pretrained backbone and report the results in Appendix F.

The gains observed with the LOBS5 state-space backbone extend to the MarketGPT Transformer. On both stocks, ReLOBGen achieves 100% replayability and improves rollout realism, with a 2.3–3.1× speedup per replayed message. ReLOBGen achieves lower unconditional L1L_1 point estimates than MarketGPT on 19 of 22 metrics for GOOG and 21 of 22 for JPM.

MarketGPT backbone results on GOOG and JPM. C / R denotes correction and rejection rates, respectively. Runtime is measured in milliseconds per replayed message. Event TV is the total variation distance between generated and observed event-type distributions, in percentage points. Columns from Overall to Trades report unconditional L1L_1 distances for the LOB-Bench metrics. Lower is better throughout. See Appendix F for detailed results.

StockMethodPost-processing (C / R)Runtime (ms)Event TV (pp)OverallStateTimesVolumesDepthsLevelsTrades
GOOGMarketGPT (ref.-first)26.7% / 37.6%78226.40.210.160.120.150.280.130.32
GOOGReLOBGen0.0% / 0.0%2541.80.090.050.110.080.110.090.09
JPMMarketGPT (ref.-first)38.5% / 37.9%51448.40.350.310.400.210.470.230.45
JPMReLOBGen0.0% / 0.0%2252.50.090.100.130.130.080.090.06

[W5, Q5] Longer rollouts

We report 100 rollouts of 5,000 messages each from the GOOG test period in Appendix G. ReLOBGen maintains 100% replayability without correction or rejection throughout these rollouts, while realism metrics generally degrade as the horizon increases. To address the mismatch in mid-price displacement from the fixed anchor between training and longer rollouts, we also test updating the anchor to the current mid-price every 100 generated messages, which yields a modest improvement in rollout realism.

[W1] Use in interactive market simulation

We additionally report a market-order intervention experiment in Appendix H, demonstrating how ReLOBGen can be used to examine market responses to external orders while maintaining replayability. This provides a concrete example of its use as an interactive market simulation environment.

Reviewer_vceHOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LOB message generators have a range of potential applications in finance. Existing LOB message generators may produce messages that cannot be replayed against the current market state. ReLOBGen solves this problem by proposing a new method for LOB message generation that ensures 100% replayability by construction. First, the method leverages the same model architecture and training loss as past work of LOBS5, but they reorder the LOB message data so that the reference order comes before the event order. Then, rather than using the base model to predict the reference order and event order, the method uses a learned query-key retrieval mechanism composed of two MLP heads, which ensures that an eligible reference order is selected. Finally, for predicting the event order, the method masks out ineligible tokens. Experiments validate that the method ensures 100% replayability by construction, increasing generation throughput by 2.6-3.7x. They further demonstrate improved rollout realism against selected baselines.

Critical strengths

The design of the method to ensure 100% replayability is elegant, as it reframes the objective of predicting a valid reference order autoregressively to using query key selection on the eligible existing reference orders.

Their analysis on the effect of reference-first factorization with respect to NLL is interesting, and they explain the tradeoff in NLL convincingly.

Critical weaknesses

I would like to see more results for rollout realism. The only baseline being compared to ReLOBGen is LOBS5, the method that they built directly upon. They also show results for an ablation, ref.-first + uniform, which is suboptimal by design, so this comparison feels insignificant. Instead, I would like to see rollout realism results for the trained ref-first S5 backbone model πθ\pi_\theta, without the pϕp_\phi scaffolding and using the LOBS5 post-processing method instead. As I understand it, there are already replayability results for this setup, so it should be simple enough to obtain the rollout realism results. This would illuminate the necessity of the pϕp_\phi dense retrieval mechanism. I would also like to see a result for at least one other method from past work (with a slight preference for a transformer model). I would be happy to increase my score, provided there are additional baselines.

Questions for authors

Clarifications

  1. Besides increased implementation complexity, is there a reason for not testing joint training between piϕpi_\phi and pϕp_\phi? Perhaps this could further improve performance?
  2. My understanding of the method is that it incorporates information on the eligible resting-order set when predicting MtM_t, so this additional data would naturally improve rollout realism. LOBS5 simply does not have access to this information, which seems to partly explain its poorer performance. Is there related work on methods that do incorporate eligible resting-order information?
  3. Why are replayabilty scores shown for LOBS5 ref.-first, but the rollout realism scores are not?
5 comments

We thank the reviewer for the detailed feedback and address each point below. Changes in the revised manuscript are highlighted in blue. We will additionally evaluate a MarketGPT-style Transformer backbone on an additional stock and provide the additional results in a follow-up response once they become available.

[W1, Q3] Role of ablations and rollout realism of reference-first LOBS5

We have added rollout-realism results for reference-first LOBS5 to Tables 3 and 4 in the revised manuscript. These results show that ReLOBGen improves rollout realism over both reference-first and reference-last LOBS5, while sharing the same backbone with the reference-first baseline.

Our ablations are designed to evaluate replayability and rollout realism separately. For replayability, Table 2 compares token orderings and evaluates the combined use of eligible reference selection and inference-time masking, which jointly guarantee replayability. The uniform-selection ablation evaluates the contribution of learned selection to rollout realism: both uniform and learned selection use the same eligible set and inference-time masking and thus guarantee replayability, but learned selection improves rollout realism.

Reference-first LOBS5 vs. ReLOBGen. C / R denotes correction / rejection rates. Runtime is per replayed message; Overall–Trades report unconditional L1L_1 distances. Lower is better.

StockMethodPost-processing (C / R)Runtime (ms)Event TV (pp)OverallStateTimesVolumesDepthsLevelsTrades
GOOGLOBS5 (Ref-first)11.7% / 27.6%5918.70.320.400.150.260.400.200.46
GOOGReLOBGen0.0% / 0.0%2276.20.170.240.150.180.200.110.18
INTCLOBS5 (Ref-first)11.5% / 13.2%10584.20.260.150.270.280.330.160.30
INTCReLOBGen0.0% / 0.0%3412.20.110.050.210.160.070.040.15

[W2] Comparison with additional baselines

We will train a MarketGPT-style Transformer backbone on an additional stock and report rollout-realism results for both the baseline and ReLOBGen once they become available.

[Q1] Joint training of the selector and backbone

We train the backbone first and then train the selector while keeping the backbone fixed, allowing us to evaluate the contribution of the proposed method without changes to the backbone parameters. Joint training, as suggested by the reviewer, may further improve overall performance.

[Q2] Access to eligible resting-order information

Both existing methods and ReLOBGen access the exact resting-order set during closed-loop market simulation, as the LOB simulator maintains it to process messages and update the order book. The difference lies in when it is used: existing methods use it to post-process generated messages (Appendix A), whereas ReLOBGen uses it during generation. Enabling its use during generation to ensure replayability is one of our core contributions.

AuthorsOpenReview ↗

Additional experimental results

[W2] Extended evaluation across backbones and stocks

We now extend the evaluation to the MarketGPT Transformer backbone on GOOG and an additional stock, JPM, retaining its model size with minor modifications to the token ordering and context length. For each stock, we compare ReLOBGen with a reference-first baseline using the same pretrained backbone and report the results in Appendix F.

The gains observed with the LOBS5 state-space backbone extend to the MarketGPT Transformer. On both stocks, ReLOBGen achieves 100% replayability and improves rollout realism, with a 2.3–3.1× speedup per replayed message. ReLOBGen achieves lower unconditional L1L_1 point estimates than MarketGPT on 19 of 22 metrics for GOOG and 21 of 22 for JPM.

MarketGPT backbone results on GOOG and JPM. C / R denotes correction and rejection rates, respectively. Runtime is measured in milliseconds per replayed message. Event TV is the total variation distance between generated and observed event-type distributions, in percentage points. Columns from Overall to Trades report unconditional L1L_1 distances for the LOB-Bench metrics. Lower is better throughout. See Appendix F for detailed results.

StockMethodPost-processing (C / R)Runtime (ms)Event TV (pp)OverallStateTimesVolumesDepthsLevelsTrades
GOOGMarketGPT (ref.-first)26.7% / 37.6%78226.40.210.160.120.150.280.130.32
GOOGReLOBGen0.0% / 0.0%2541.80.090.050.110.080.110.090.09
JPMMarketGPT (ref.-first)38.5% / 37.9%51448.40.350.310.400.210.470.230.45
JPMReLOBGen0.0% / 0.0%2252.50.090.100.130.130.080.090.06
AuthorsOpenReview ↗

Request for feedback on our responses

Dear Reviewer vceH,

Thank you again for your review. Following your suggestion, we have added both requested comparisons:

  • Reference-first LOBS5 (Tables 3–4). With the same pretrained backbone, ReLOBGen guarantees replayability, improves rollout realism, and runs faster per replayed message than reference-first LOBS5 with LOBS5 post-processing.
  • MarketGPT Transformer (Appendix F). The same conclusion holds on a Transformer backbone evaluated on GOOG and an additional stock, JPM: with the same pretrained backbone, ReLOBGen guarantees replayability, improves rollout realism, and runs faster per replayed message than MarketGPT on both stocks.

We have also responded to your three clarification questions in our initial response. We would appreciate your feedback on whether our responses address your concerns. If they do, we would be grateful if you would consider updating your assessment. We are happy to answer any remaining questions.

Best regards,

The Authors

Reviewer_JETgOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses replayability in closed-loop LOB simulation. Autoregressive message generators (LOBS5, MarketGPT, MarS) often emit cancel/delete/execute messages that reference non-existent orders or violate size/price constraints, so they rely on heuristic correction or rejection-and-resample. ReLOBGen makes three changes. (i) It reorders the LOBS5 token layout to event → reference → event order, and encodes the reference by the order's current attributes (remaining size, price relative to the current mid). (ii) It replaces autoregressive generation of the eight reference tokens with a learned categorical distribution over the eligible resting orders. Each order is scored by the dot product between a query (from the frozen base model's hidden state after the event tokens) and a cached per-order key; keys are encoded relative to a fixed mid-price anchor so that they can be reused across steps. (iii) It masks invalid tokens in the event-order fields and fills in fields whose valid support is a single token. Replayability then holds by construction. Experiments use Databento ITCH data for GOOG and INTC (train 2025, test January 2026), a 35M-parameter S5 backbone, and 1,000 rollouts of 500 messages per stock. ReLOBGen has 0% replayability violations, against 25–40% raw violations for LOBS5. It is 2.7–3.6× faster per replayed message. It has lower LOB-Bench unconditional L1/W1 distances on most metrics (overall L1 0.27→0.17 on GOOG, 0.22→0.11 on INTC) and lower average conditional and market-impact discrepancies. An ablation shows that learned selection beats uniform selection over the same eligible set.

Critical strengths
  • The right design for a real bottleneck. Appendix D shows that with a 500-message context only 60% (GOOG) and 29% (INTC) of the current top-10 resting orders ever appear in the history. A generator that sees only messages and the aggregated book therefore cannot be expected to name the order it cancels or executes. ReLOBGen solves this directly: the simulator already maintains the exact order-level book, so the model retrieves the reference from it, with a learned selector that scores each resting order by its price, remaining size and submission time. This uses information that is legitimately available in any closed-loop simulation and that prior generators leave unused. Keys are cached and anchored to a fixed mid-price, so retrieval stays cheap. Definition 1 and Table 1 give a clean decomposition into reference eligibility and event-order compatibility. Appendix A documents the ad-hoc post-processing of LOBS5, MarketGPT and MarS, which is useful to the community on its own.
  • Large, carefully reported gains. The paper shows 100% replayability and fewer forward passes (17 → 7–8 for non-add events), giving a 2.7–3.6× speedup. LOB-Bench improvements are consistent across unconditional, conditional and market-impact evaluations, and most are statistically significant under 99% bootstrap CIs (16/18 and 18/20 metrics with lower L1 point estimates are significant). Market-impact response curves move from largely flat to close to the real ones. The "Ref.-first + uniform" ablation, which picks an eligible resting order uniformly at random, is also 100% replayable, but its overall L1 on INTC is 0.30, worse than LOBS5 (0.22); ReLOBGen reaches 0.11. Validity alone does not produce realism; the learned selector does. The event-wise breakdowns and the INIT-vs-exact initialization comparison make the evaluation transparent.
Critical weaknesses

Neither weakness is fatal; both can be addressed in revision.

  1. The distributional argument should be stated precisely. The paper says that post-hoc correction and rejection "may alter the replayed message distribution", which implies that ReLOBGen does not. But whole-message rejection-and-resample samples exactly from π_θ(M | valid). Token-level masking with local renormalization generally differs from that conditional (cf. Park et al., "Grammar-Aligned Decoding", NeurIPS 2024). And the reference in ReLOBGen comes from a separately trained selector rather than from π_θ. ReLOBGen therefore samples a different, well-defined distribution, π_θ(E) · p_φ(R | h_t, C_t) · π_θ^mask(X | E, R), rather than an unaltered one. I think a stronger argument is available and the paper should make it. Rejection also suppresses event types whose references the model cannot name: LOBS5 under-generates deletes (36.9% vs 46.6% real on GOOG, Table 14). ReLOBGen keeps the model's event-type distribution and delegates "which order" to a component that can actually see the book. Stating this explicitly, ideally with a short quantitative comparison, would make the motivation rigorous.

  2. Evaluation scope is narrower than the motivation. The paper motivates the work with closed-loop simulation for training and evaluating trading strategies, but:

    • no experiment injects external agent orders;
    • rollouts last 500 messages, and the evidence on error accumulation is described only as "suggesting potential";
    • the evaluation uses two stocks, one backbone, one baseline and one training seed.

    Since ReLOBGen never stalls, longer rollouts are cheap and would strengthen the stability claim. A small agent-in-the-loop experiment would show that the method serves its stated purpose.

Minor suggestions:

  • Say explicitly in the discussion that the selector exploits order-level state that the baselines do not use. This is a strength, and making it explicit clarifies where the gains come from.
  • On GOOG, ReLOBGen still generates three times as many executes as real data (9.2% vs 3.0%).
  • On INTC, the Times group is worse than LOBS5 (L1 0.21 vs 0.14).
  • The symbol C_t denotes both the eligible set (Sec. 3) and coverage (App. D).
Questions for authors
  • Can you state the sampling distribution of ReLOBGen explicitly and compare it with rejection sampling on the same reference-first model? For example, how do event-type marginals differ between the two? Please also report the selector's held-out NLL and top-k accuracy, compared with uniform selection (log|C_t|).
  • In an agent-in-the-loop setting, are agent orders included in C_t, so that the model could cancel an agent's order? Could you show one experiment with injected agent orders, e.g., a market-order or TWAP intervention?
  • How do the realism metrics evolve over longer rollouts (e.g., 5k or more messages)?
  • What causes the over-generation of executes on GOOG and the weaker Times metrics on INTC?
3 comments
AuthorsOpenReview ↗

Response (1/2)

We thank the reviewer for the detailed feedback and address each point below. Changes in the revised manuscript are highlighted in blue. We are conducting additional experiments and will report the results in a separate follow-up response once they become available.

[W1, Q1] Precise distributional argument

Thank you for requesting a more precise distributional argument. Our statement in the Introduction that “Such corrections may alter message semantics and shift the distribution of replayed messages” refers to heuristic correction, such as relaxed reference matching and price/size modification.

Following your suggestion, we have added Appendix E to explicitly compare sampling with post-processing and sampling with ReLOBGen. We formulate the target distribution for closed-loop simulation as p(Mt∣Ht−1,Ot−1)p(M_t\mid\mathcal H_{t-1},\mathcal O_{t-1}), incorporating the current resting-order set alongside the finite model context.

Rejection sampling. Under a simplified rule that rejects only references absent from the current resting-order set, the accepted event distribution is

πθrej(Et=e∣Ht−1,Ot−1)∝πθ(Et=e∣Ht−1)(1−r(e)),\pi_\theta^{\mathrm{rej}}(E_t=e\mid\mathcal H_{t-1},\mathcal O_{t-1}) \propto \pi_\theta(E_t=e\mid\mathcal H_{t-1})(1-r(e)),

where r(e)r(e) is the event-specific rejection rate and r(add)=0r(\mathrm{add})=0. Thus, the total non-add probability decreases when non-add reference rejection occurs, even if the backbone perfectly learns the history-conditioned distribution.

ReLOBGen. For non-add events, assuming limited additional dependence on the remaining order-level state for event and event-order generation, the target admits the following approximate reference-first factorization:

p(Mt∣Ht−1,Ot−1)≈p(Et∣Ht−1) p(Rt∣Ht−1,Ot−1,Et) p(Xt∣Ht−1,Et,Rt).p(M_t\mid\mathcal H_{t-1},\mathcal O_{t-1}) \approx p(E_t\mid\mathcal H_{t-1})\,p(R_t\mid\mathcal H_{t-1},\mathcal O_{t-1},E_t)\,p(X_t\mid\mathcal H_{t-1},E_t,R_t).

ReLOBGen directly implements this sampling structure: it samples the event from πθmask(Et∣Ht−1)\pi_\theta^{\mathrm{mask}}(E_t\mid\mathcal H_{t-1}), selects the reference from pϕ(Rt∣Ht−1,Ot−1,Et)p_\phi(R_t\mid\mathcal H_{t-1},\mathcal O_{t-1},E_t), and samples the event order from πθmask(Xt∣Ht−1,Et,Rt)\pi_\theta^{\mathrm{mask}}(X_t\mid\mathcal H_{t-1},E_t,R_t).

[Q1] Selector performance

We evaluate the selector on 10,000 held-out samples and compare it with uniform selection over the same eligible resting-order sets. As shown below, learned selection achieves lower NLL and higher top-k accuracy on both stocks.

Table R.1. Held-out selector performance. Learned and uniform selection over the same eligible resting-order sets. Lower NLL and higher top-k accuracy are better.

ConfigurationNLL ↓Top-1 ↑Top-5 ↑Top-10 ↑
GOOG-uniform3.63.0%15.0%29.9%
GOOG-learned3.016.4%53.1%72.5%
INTC-uniform4.51.2%6.1%12.2%
INTC-learned3.017.1%56.3%75.7%

[W2, Q2, Q3] Additional experiments

We are extending our evaluation to include an agent-in-the-loop setting, longer rollouts, and a MarketGPT-style Transformer backbone on GOOG and JPM. We will report the results in a separate follow-up response once they become available.

[S1] Stating exploitation of order-level state

Thank you for the suggestion. We now explicitly state in the final paragraph of the Introduction that ReLOBGen exploits the order-level state already maintained by the LOB simulator during closed-loop simulation, information that existing approaches consult only during post-processing.

AuthorsOpenReview ↗

Response (2/2)

[S2, Q4] Over-generated execution

We acknowledge that ReLOBGen still over-generates execute events on GOOG. Our analysis suggests that this discrepancy is amplified during rollout, rather than simply reflecting the model’s learned execute probability. The increase during rollout is associated with both more frequent initiation and greater persistence of execute runs. We detail these findings below.

  • Execute probability. The model assigns a mean execute probability of 3.4% on 10,000 real test contexts, compared with an actual execute frequency of 2.6% in the same samples. The higher execute frequency in the training sample (3.6%) may partly explain this difference.
  • Rollout progression (Table R.2). Execute frequency rises from 4.5% in the first 10 messages to 13.4% in the final 10 messages, while matched real continuations remain around 3%.
  • Execution onset and persistence (Table R.3). Execute runs both start more frequently and persist more strongly as rollouts progress. The backbone also overestimates execute continuation probabilities in real test contexts (2.7–4.6 percentage points after 1–5 consecutive executions), which may contribute to the increased persistence during rollout.

Table R.2. Execute frequency over rollout progression. Averages over 1,000 GOOG rollouts and matched real continuations, shown for selected 10-message intervals.

Message position1–10101–110201–210301–310401–410491–500
Real2.8%3.1%2.7%3.5%3.1%3.4%
ReLOBGen4.5%7.3%8.5%9.5%11.4%13.4%

Table R.3. Execute transitions during rollout. Observed conditional probabilities across 1,000 matched GOOG rollouts, shown for selected message-position intervals. ee denotes execute.

MethodTransition2–10491–500
Real¬e→e\neg e\rightarrow e1.7%1.8%
Reale→ee\rightarrow e42.1%48.8%
ReLOBGen¬e→e\neg e\rightarrow e2.2%5.7%
ReLOBGene→ee\rightarrow e54.1%61.9%

[S3, Q4] Times metrics for INTC

The weaker performance in the Times group on INTC primarily stems from time-to-cancel (TTC). Our analysis indicates that ReLOBGen generally has higher TTC than the real data on INTC. INTC’s larger resting-order set may increase the probability of sampling older orders, resulting in higher TTC. To mitigate this, we test top-10 and top-1 sampling on INTC and observe that this improves Times metric from 0.21 to 0.16 and 0.13, respectively, but increase discrepancies in Depths and Levels. The reference-first baseline also shows weaker performance in the Times group, with an L1 distance of 0.27 compared with ReLOBGen’s 0.21.

Table R.4. Rollout realism on INTC under different reference-sampling strategies. L1 distances by metric group; lower is better.

MethodOverallStateTimesVolumesDepthsLevelsTrades
LOBS5 (ref.-last)0.220.140.140.200.320.170.27
LOBS5 (ref.-first)0.260.150.270.280.330.160.30
Ref.-first + uniform0.300.100.350.270.360.340.32
ReLOBGen0.110.050.210.160.070.040.15
ReLOBGen (top-10)0.120.040.160.140.120.130.12
ReLOBGen (top-1)0.180.130.130.160.230.250.11

[S4] Duplicate notation

Thank you for the feedback. We have changed the coverage notation in Appendix D from Ct(L)C_t(L) to coveraget(L)\mathrm{coverage}_t(L) to distinguish it from the eligible resting-order set CtC_t.

AuthorsOpenReview ↗

Additional experimental results

[W2] Extended evaluation across backbones and stocks

Our original evaluation matches the LOBS5 evaluation in LOB-Bench in the number of stocks (two: GOOG and INTC) and rollout length (500 messages). We now extend the evaluation to the MarketGPT Transformer backbone on GOOG and an additional stock, JPM, retaining its model size with minor modifications to the token ordering and context length. For each stock, we compare ReLOBGen with a reference-first baseline using the same pretrained backbone and report the results in Appendix F.

The gains observed with the LOBS5 state-space backbone extend to the MarketGPT Transformer. On both stocks, ReLOBGen achieves 100% replayability and improves rollout realism, with a 2.3–3.1× speedup per replayed message. ReLOBGen achieves lower unconditional L1L_1 point estimates than MarketGPT on 19 of 22 metrics for GOOG and 21 of 22 for JPM.

Appendix F also reports reference-selection performance with the MarketGPT backbone. The learned selector achieves lower NLL and higher top-1, top-5, and top-10 accuracy than uniform selection on both GOOG and JPM.

MarketGPT backbone results on GOOG and JPM. C / R denotes correction and rejection rates, respectively. Runtime is measured in milliseconds per replayed message. Event TV is the total variation distance between generated and observed event-type distributions, in percentage points. Columns from Overall to Trades report unconditional L1L_1 distances for the LOB-Bench metrics. Lower is better throughout. See Appendix F for detailed results.

StockMethodPost-processing (C / R)Runtime (ms)Event TV (pp)OverallStateTimesVolumesDepthsLevelsTrades
GOOGMarketGPT (ref.-first)26.7% / 37.6%78226.40.210.160.120.150.280.130.32
GOOGReLOBGen0.0% / 0.0%2541.80.090.050.110.080.110.090.09
JPMMarketGPT (ref.-first)38.5% / 37.9%51448.40.350.310.400.210.470.230.45
JPMReLOBGen0.0% / 0.0%2252.50.090.100.130.130.080.090.06

[W2, Q3] Longer rollouts

We report 100 rollouts of 5,000 messages each from the GOOG test period in Appendix G. ReLOBGen maintains 100% replayability without correction or rejection throughout these rollouts, while realism metrics generally degrade as the horizon increases. To address the mismatch in mid-price displacement from the fixed anchor between training and longer rollouts, we also test updating the anchor to the current mid-price every 100 generated messages, which yields a modest improvement in rollout realism.

[W2, Q2] Agent-in-the-loop

Whether agent orders are included in CtC_t depends on the intended simulation setting. To preserve the agent’s control over its orders, we can exclude them from the eligible sets for background cancel and delete events while keeping them available for execution in the shared order book. Both designs are possible, but this separation is useful for developing and evaluating trading strategies, as illustrated by MarS’s intervention experiments.

Following the reviewer’s request, we conduct repeated market-order intervention experiments and report the results in Appendix H. In our intervention experiment, the agent submits market orders, so it leaves no resting order for the model to cancel. ReLOBGen maintains 100% replayability under intervention. The mid-price response is in the direction of the injected orders, and price differences tend to widen between interventions. This widening tendency appears to weaken later in the rollout after the final intervention.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of author responses and additional results

We thank the Area Chair and reviewers for their constructive feedback. We summarize our responses to the main concerns below.

  • Limited evaluation scope (JETg, vceH, 7UtD). We added experiments with a MarketGPT Transformer backbone on GOOG and an additional asset, JPM, and report the results in Appendix F. On both stocks, ReLOBGen achieves 100% replayability without correction or rejection, improves overall rollout realism, and delivers a 2.3–3.1× speedup per replayed message over the baseline using the same pretrained backbone.

  • Longer rollouts and agent-in-the-loop evaluation (JETg, 7UtD). We tested ReLOBGen on 5,000-message rollouts (Appendix G) and under repeated market-order interventions (Appendix H), observing 100% replayability in both experiments. Buy and sell interventions shift the mid-price in the expected directions relative to paired rollouts without intervention.

  • The use of order-level information (vceH, 7UtD). We clarified that both the baselines and ReLOBGen have access to the same exact resting-order set during closed-loop rollouts. They differ in when they use this information. The baselines use it during post-processing, whereas ReLOBGen uses it during generation to guarantee replayability by construction. Enabling this generation-time guarantee is one of our core contributions.

  • Component roles and ablations (vceH, 7UtD). We clarified which ablations assess replayability and which assess rollout realism, and where each is reported. We also added the reference-first LOBS5 baseline to Tables 3–4, showing that ReLOBGen improves rollout realism with the same pretrained backbone.

  • Sampling distributions of LOBS5 and ReLOBGen (JETg). We clarified that our statement about distribution shift refers to heuristic corrections. Following the reviewer’s suggestion, we added Appendix E to analyze how reference rejection reweights event-type probabilities and to describe ReLOBGen’s sampling factorization given the message history and current order book.

AuthorsOpenReview ↗

Follow-up on Our Responses

Dear Reviewers,

Thank you again for your thoughtful feedback. As we are now past the midpoint of the discussion period, we would appreciate your feedback on whether our responses address your concerns. Please let us know if there are any remaining questions or points we can clarify.

Thank you for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.