acceptodds
Under review as a conference paper at ICLR 2027

SS2P2: A STABLE STATE-SPACE POINT PROCESS FOR LIMIT-ORDER-BOOK SIMULATION

Abstract

The limit order book simulation and prediction are important for financial market analysis and market making strategies. The existing State-Space Point Process (S2P2) model predicts the next LOB event accurately but suffers from instability when rolled out which leads to drift or explosion in the simulation. We decouple the output heads of the S2P2 model to create a Stable State-Space Point Process (SS2P2) model. The new model incorporates a softmin-bounded rate head and a soft-max mark head, ensuring stability and non-explosiveness in the point-process sense. Decoupling the heads gives us a hard closed-form intensity ceiling that bounds the event count and prevents explosion, while the soft-max mark head keeps the rate neutral at a fixed state so simulation stays stable. On Coinbase BTC, ETH, and SOL order flow against six baselines and three seeds each, we attain the best NLL and type accuracy on two of three assets (SOL NLL 0.27 vs. 0.38) and tie NHP on the third. We pass verified rate calibration on all nine checkpoints where the uncapped S2P2 fails on six, and lead five of twelve stylized facts in closed-loop roll-outs. The SS2P2 model provides a reliable and accurate framework for simulating limit order book events, making it a practical foundation for market-making world models and agent training.

Primary area:
learning on time series and dynamical systems
Keywords:
marked temporal point processes, state-space models, generative models, limit order book simulation, market simulation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_qLsXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes an extension of S2P2 Limit Order Book simulator. It modifies the S2P2 architecture by factorizing the decoder into a total-rate head and a categorical mark head. Experiments are conducted on data of three cryptocurrencies for seven days. A number of experiments showcasing different properties of the methods are provided.

Critical strengths
  • The paper targets a weakness of neural point-process models, i.e., that a good teacher-forced prediction does not imply stable or realistic simulation.
  • Experiments include different aspects relevant to LOB simulators, which are prediction, closed-loop simulation, rate calibration, and generation efficiency.
Critical weaknesses
  • The dataset used for the experiments contains only seven days for three cryptocurrencies. This is very limited for supporting the claims for stable LOB simulator.
  • The data is split on a per-day basis (70/15/15% for train/validation/test). This means that the training set contains information about the validation and test sets.
  • Experimental comparisons are not done in a fair manner. S2P2 is substantially modified from its published version before it is tested and compared with the modified version SS2P2 presented in this paper. Also, the number of epochs used in the two papers is very different. The experimental protocol for the S2P2 is not followed. Is the clearly bad performance of S2P2 due to not sufficient training? A different protocol is applied also for SAHP.
Questions for authors
  1. How would the proposed method compare with the related works in a fair experimental setup (including the input received by each method, the training process, and parameter count)?
  2. Why did the experimental setup not use unbiased data splits?
  3. How would the results be if the experiments were conducted on multiple assets and for longer periods?
Reviewer_u8C3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a method to generate the limit order book using a point process model. However, for the evaluation and generalizability of the proposed model, I have several concerns. In particular, the paper misses some key recent baselines like Mars and price-related evaluation metrics.

Critical strengths
  1. Clear motivation. This paper deals with the question of how to simulate/generate the limit order book, which is important to many markets, e.g., the stock market and crypto market. With rigorous validation and further experiments on its generalizability, I believe that this paper would have great influence.

  2. Encouraging predictive performances. This proposed method, as reported in the manuscript, has achieved the best mean NLL and event-type accuracy on ETH and SOL.

Critical weaknesses
  • Missing comparisons with key baselines. Although this paper has identified MarS as a closely related method, it does not include MarS as a baseline for comparison.
  • Lack of price-level evaluation. As far as I am concerned, in both stock and crypto markets, price is an essential indicator that emerges from order-book interactions and provides traders with one of the most direct signals of market conditions. Therefore evaluating price dynamics is indispensable for assessing whether a market simulator reproduces realistic market behaviour.
  • Broad framing but limited experimental validation. This work frames itself as an approach to limited-order-book simulation, while the experiments are limited to cryto markets. I believe the evaluation should also include stock markets, which likewise rely on limit order books.

Li, Junjie, et al. "Mars: a financial market simulation engine powered by generative foundation model." International Conference on Learning Representations. Vol. 2025. 2025.

Questions for authors
  • What is the performance of the proposed method on stock markets?
  • Can the proposed method reproduce the price dynamics accurately?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.