acceptodds
Under review as a conference paper at ICLR 2027

KUKULKAN: A STATE SPACE MODEL FOR TEMPORAL POINT PROCESSES WITH INTERPRETABILITY

Abstract

We write the temporal point process as a state space model and use the parallel scan of Mamba to model the excitation kernel explicitly: the Hawkes excitation is the state of a diagonal linear system, its compensator has a closed form and its branching ratio is the steady-state gain. Through structured state space duality the same computation is an attention over past events, which places the classical kernel, the attention models and the state space models in one form. The Mamba Hawkes process and S2P2 are built on that form but give the branching ratio up in exchange for expressivity. Kukulkan keeps it and relaxes it from a number to an interval. The context selects every channel’s decay rate and the background rate freely, and the output map only through a gate between zero and one, so the likelihood stays exact and the offspring of every event lies below a number fixed at its birth. The branching ratio is partially identified, with a bound, a conditional estimate and a simulated estimate. Under an evaluation protocol that removes the quadrature bias of the standard held-out likelihood, Kukulkan matches the unconstrained neural output map on seven of nine public datasets and improves on the model with an exact branching ratio on all. On exchange order flow it improves on the unconstrained output map while bounding the offspring of every order, and its estimate of the market stays sub-critical.

Primary area:
learning on time series and dynamical systems
Keywords:
state space model, time series, event process, hawkes process

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aZ72OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper presents Kukulkan, a new approach to neural temporal point process modeling that builds on ideas from state-space architectures while keeping the interpretability of classical Hawkes processes. The main idea is to keep the excitation dynamics linear and additive, while allowing contextual information to modulate decay rates. Intuitively, Kukulkan can be viewed as a Hawkes process whose event contributions are filtered through a soft neural rejection step (although no actual rejection step is performed).

Critical strengths
  • The overall idea is creative and certainly worth exploring.
  • The connection between Hawkes processes and state-space models is interesting and useful.
  • The findings on quadrature bias and the disagreement between stated and simulated branching ratios are interesting.
Critical weaknesses
  • The biggest weakness is the presentation and the lack of narrative structure. The introduction becomes very technical too quickly, and the related work reads like a sequence of loosely connected references that are not properly contextualized. This makes it difficult to understand what is novel and what the main contribution is. There is too little motivation and explanation of why the many different design choices are made. Overall, the paper contains interesting ideas, but they are obscured by a very compressed style, weak transitions, and too little motivation for the individual design choices. In its current form, the manuscript is much harder to follow than necessary.
  • EasyTPP seems to be the wrong primary benchmark because it mainly tests next-event likelihood and prediction on standard datasets, whereas Kukulkan’s central claim concerns whether a model learns meaningful and stable excitation dynamics over longer horizons.
  • Given that interpretability is central, the evaluation should include qualitative examples showing that the learned excitation patterns are understandable and plausible.
Questions for authors
  • Does EasyTPP actually test the property Kukulkan is designed to improve?
  • Should the paper include long-horizon or free-running evaluation in the main results?
  • Can the interpretability claim be supported without qualitative examples?
1 comment
AuthorsOpenReview ↗

Enhanced readability, re-framed paper with no more architectural claim.

I totally agree with your criticism of the presentation. I rewrote the paper from the first page: it now asks one question, states three limitations of the likelihood in plain words in the introduction, and follows the same order in the method and the results. The related work has three paragraphs with one topic each.

On your questions:

  • Does EasyTPP test the property? No, and you were right to ask. This is now the main point of the paper: next-event likelihood on those benchmarks does not show whether the dynamics were learned. The revision tests models against simulated processes for which the best possible predictor can be computed exactly (Sec. 4.1).
  • Long-horizon or free-running evaluation. It is now in the main results. Each model generates 64 continuations of 512 events from each of 256 histories and is tested for calibration, and I also measure how its prediction responds to one inserted event (Secs. 4.1 and 6.1; Table 2, Fig. 2).
  • Qualitative examples. Fig. 1 shows the intensities, and Table 16 shows which kernels carry the stated ratio. They show that the learned excitation can be trusted only at short time scales, so I no longer claim interpretability beyond that (Sec. 6.2).

The two findings you found interesting, the quadrature bias and the disagreement about the branching ratio, are now what the paper is about.

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

This paper investigated a key problem in existing neural MTPPs: the compensator cannot be computed in closed form, and numerical estimates of the integral (and the likelihood) could lead to bias in both training and evaluation. Additionally, the branching ratio and interpretability of these neural MTPPs are lost in exchange for expressivity.

This paper is motivated by analyzing the quadrature bias due to the modeling choices of existing neural MTPPs. It then proposes a new model Kukulkan to maintain the linear output map that preserves interpretability via the branching ratio. Instead of applying Softplus that leads to an intractable compensator, Kukulkan is parameterized with non-negative matrices B,CB, C to ensure the non-negativity of intensities. Comprehensive experiments demonstrate that Kukulkan sacrifices some expressivity to maintain the branching ratio, but greatly improves its flexibility when compared with other baseline models such as parametric Hawkes processes.

Critical strengths
  1. The paper is technically solid and very well-motivated in addressing the problem of an intractable compensator for many neural MTPPs. Sections 2-4 show great understanding and in-depth analysis of the MTPP literature, e.g., the quadrature bias in neural MTPPs caused by the nonlinear head (usually Softplus activation) to ensure non-negative intensities. Both the overview of key design choices in autoregressive neural MTPPs and the analysis of their implications will greatly contribute to the MTPP literature and the community.

  2. The proposed model Kukulkan is another contribution that retains the branching ratio in the setting of a linear output map. The experimental setup is well-documented in both the main text and the appendix, where the linear Hawkes baselines set a great scale for interpreting NLL results. From the results, Kukulkan is much quicker to train while maintaining comparable NLL and retains interpretability through the estimated branching ratio.

Critical weaknesses

Both weaknesses are suggestions for further improvements rather than reasons for potential rejection.

  1. The paper is a bit dense, partially due to a lot of great discussions, but it would be great to elaborate more on some details to further improve clarity. For example, it would be good to also define Kukulkan in a set of equations in the main text, in addition to Table 1 (though the summary is great and places Kukulkan in the context of related work) and Figure 2. There could also be more discussion on how the gate is parameterized (via a single parameter or a function) and the three estimates of the branching ratio.

  2. The code is currently not available, but the authors commit to releasing it when the paper gets accepted, and the paper provides sufficient experimental details. I would encourage the authors to open-source their code and preprocessing pipelines, as this will greatly improve the reproducibility of MTPP models and transparency in the MTPP field.

Questions for authors

Clarification questions:

  • The EasyTPP repo supports both trapezoidal rule and unbiased Monte Carlo estimates of the compensator [1], so the comment on "...placed on a fixed grid of equally spaced positions..." is not precise (lines 130-131). There are other discussions in Sec. 3 around the fixed-grid integration points; please double-check and update as needed.
  • Both "mark" and "type" are used in the paper, e.g., mjm_j in line 193. Is "type" intended to refer to the realization of a "mark"? It would be great to clarify it and also define "mark" in Sec. 3.
  • Related to the above, what does eje_j refer to in Figure 2? It says "type, mark, covariates," but it is not defined in the main text.
  • In Appendix B, there are ties identified for some datasets in Table 4; how does this change the evaluation (if any)?
  • Does Kukulkan support event-level attribution or only focus on aggregate properties like the branching ratio? My understanding is the latter, and I do not think a model needs to cover both aspects, but it would be great to make the distinction explicit about the scope of interpretability.

Minor issue:

  • Some hyperlinks seem to be broken, e.g., "Section 3" in lines 56 and 317, "equation 2" in line 231.

Reference: [1] https://github.com/ant-research/EasyTemporalPointProcess/blob/main/easy_tpp/model/basemodel.py#L150-L153

1 comment
AuthorsOpenReview ↗

Narrowed the discussion down into core questions, uploaded the code

To be honest, I should tell you first that the revision no longer proposes Kukulkan, which you valued as a contribution. The new experiments show that a stated branching ratio is determined by the data only in part (Sec. 6.2), so I did not feel I could keep presenting the gate and the three estimates as interpretable. The closed-form output remains, but only as a tool for the analysis in Sec. 4.2. The analysis you found useful is now the subject of the paper.

On your questions:

  • EasyTPP. You are right that the option exists, and the paper now describes it precisely (App. B.3). use_mc_samples changes the weights (mean instead of trapezoid), but the sample points come from make_dtime_loss_samples, which uses torch.linspace under both settings (version 0.3.0 and the current main branch). So the points are equally spaced either way. The repository of the S2P2 authors does draw uniform random points, and I use that as the unbiased comparison: with random points neither error appears (Table 1).
  • Mark and type. The revision uses "type" only and defines it in Sec. 3. The figure with the undefined symbol is removed.
  • Ties. An interval of length zero contributes its event term and no compensator. Removing the ties by jitter changes the NLL by at most 0.002 nats on Retweet and StackOverflow (App. B.2).
  • Attribution. Aggregate only, and Sec. 6.2 shows that even the aggregate is determined only in part.
  • The hyperlinks are fixed, and the code is attached as supplementary material, as you encouraged.
Reviewer_Y9s2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Kukulkan which is a temporal point process that has the basic structure of a Hawkes process and allows dynamics to depend on context. It represents excitation as a state space recurrence, with decay rates which depend on the context and background intensity. It restricts how the excitation state enters the intensity using a gate between 0 and 1. This means the model cannot recover a single classical branching ratio, but it can bound the number and provide some estimates. The paper identifies bias in the numerical integration commonly used to evaluate certain neural TPPs and proposes an evaluation procedure that avoids this issue. Empirically, Kukulkan is competitive on standard datasets and performs better on the exchange order experiment.

Critical strengths
  • The paper is nicely written and well motivated
  • It's a nice compromise between flexibility and pure Hawkes interpretability
  • Offspring bound is simple and potentially useful, particularly for applications such as market microstructure
  • It uses exact likelihood which avoids numerical integration at inference
Critical weaknesses
  • The paper has limited architectural novelty. The Hawkes+SSM connection are not new, as also pointed out by the authors themselves. The main modeling contribution seems to be the constrained output gate and the resulting offspring bound
  • The claimed interpretability is somewhat weaker than for a classical Hawkes model
  • Predictive improvements on the public benchmarks are modest. The main justification for the model is then interpretability rather than substantially better performance
Questions for authors
  • How meaningful is the proposed offspring upper bound in practice? I'm not sure if there are many use cases that need to know this
  • What prevents the model from moving explanatory mass arbitrarily between background intensity and excitation?
1 comment
AuthorsOpenReview ↗

Architectural claim dropped, focused more on interpretations and assessments.

I completely agree with your point on novelty. The revision no longer claims an architecture, and the gate and the offspring bound are removed.

Your second question, what prevents the model from moving explanatory mass between background and excitation, turned out to be the most important question for this paper. To be honest, the answer is that nothing prevents it, and this is now one of the three main results. The likelihood only sees the intensity as a whole, not how it is split. A model with Hawkes kernels and a learned background states a branching ratio of 0.71, 0.81, 0.86, 0.91 or 0.99 for a true value of 0.70, depending only on the slowest kernel it is allowed, and the likelihoods of these models are within 0.001 nats of each other (Sec. 6.2, Table 4). A larger model or more data does not fix this. What stays stable is the part of the ratio at short time scales (0.64 against the true 0.70), both in simulation and on order flow, so the paper now recommends reporting that part.

For the same reason I no longer present the bound as an interpretable quantity, and the revision makes no claim about predictive performance. I also agree that the interpretability is weaker than for a classical Hawkes model; Sec. 6.2 now says how far it goes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.