acceptodds
Under review as a conference paper at ICLR 2027

EventFM: Zero-Shot Forecasting of What Happens When

13 views

Abstract

Foundation models promise train-once zero-shot forecasting, yet event forecasting remains far less developed than regular time-series forecasting. The obstacle is structural: event streams unfold as irregular occurrences in continuous time, lacking a shared grid, fixed output space, and large interoperable corpora that make regular time-series foundation models possible. Moreover, an event forecaster must predict both what happens next and when it happens by inferring sequence-local relations among event types and arrival times. These challenges motivate EventFM, a foundation model for sequence-level zero-shot forecasting of marked temporal point processes. EventFM's event-native architecture normalizes time within each history, constructs one candidate query for each observed mark, and jointly predicts the next mark and its conditional occurrence time. To resolve the pretraining problem, we introduce the Event Structural Simulator (EventSS), an on-the-fly synthetic prior over event relations. EventSS captures trigger-response relations, overlapping workflows, delayed and cancellable consequences, recurrence, and non-additive event interactions. We pretrain EventFM on an EventSS prior and evaluate it in a strict sequence-level zero-shot setting, where each prediction uses only the observed history of the target sequence. Across five real-world benchmarks and structured synthetic scenarios, a single frozen EventFM shows strong zero-shot transfer, while targeted ablations show that both the event-native architecture and EventSS prior are important for transfer.

Primary area:
learning on time series and dynamical systems
Keywords:
temporal point process, event sequence, foundation model, zero-shot forecasting, marked temporal point process

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

The paper proposes EventFM, a foundation model for sequence-level zero-shot forecasting of marked temporal point processes (MTPPs). Given a single observed event history, the model predicts both the mark of the next event and its conditional occurrence-time distribution, without target-specific training or fine-tuning. The proposed framework has two main components. First, EventFM employs an event-native, dual-axis in-context architecture that combines time, anonymous mark identity, sequence-local mark context, and positional information. Second, the Event Structural Simulator (EventSS) generates synthetic pretraining sequences by maintaining a queue of pending event opportunities, allowing the simulation of trigger-response relationships, delayed or cancelled events, recurrence, and interactions among overlapping event processes. The paper evaluates the model on five real-world event datasets and three groups of synthetic scenarios. The authors report competitive zero-shot forecasting performance, with EventFM achieving the best time CRPS in seven of eight reported settings.

Critical strengths
  1. It's very interesting to introduce the Event Structural Simulator as an on-the-fly synthetic prior over event relations. It maintains distinct pending event instances that can create, modify, or resolve future opportunities.
  2. Sharing the prediction head across candidate marks offers a principled way to handle different vocabularies without requiring globally aligned event-type embeddings.
  3. The experiments are well designed and provide useful evidence in the field.
Critical weaknesses
  1. The evidence for cross-domain transfer is limited by benchmark coverage and baseline comparability. The paper should more clearly separate conclusions supported by real-world data from those supported by synthetic data.
  2. Quantile loss is a reasonable training objective, but good quantile loss does not automatically establish calibrated predictive distributions. The authors should provide evidence of calibration and clarify how distributional scores are computed from the finite quantile representation.
  3. EventSS is an interesting contribution, but its generative assumptions deserve closer examination. I would like to see a more direct assessment of structural coverage. These experiments would help determine whether the observed gains come from the simulator's structural semantics or from the broader diversity of its generated data.
Questions for authors
  1. It would be useful to test the model under controlled shifts in which the next event belongs to a type absent from the observed prefix.
  2. How are quantiles outside the explicitly predicted range handled?
  3. The paper reports that the synthetic-prior ablation preserves or improves some real-data results while sharply degrading other settings. This mixed pattern suggests that the prior's contribution is dataset-dependent. The analysis should identify which structural properties of the target process are associated with improvements, rather than relying primarily on aggregate rankings.
  4. What evidence suggests that EventSS covers conditional event-generation mechanisms, rather than merely the marginal statistics of real sequences? Can the authors identify target-process properties that are not represented by the prior?
5 comments

(1/3) We thank the reviewer for the thorough review and constructive questions. In response, we separated the real-data and synthetic conclusions in Section 4 and added analyses of calibration (Appendix M), of next types absent from the history (Appendix Q) and of what EventFM transfers from EventSS (Appendix R). The revision marks all changes in blue.

W1: Benchmark coverage, baseline comparability, and transfer claims

We agree that the five real benchmarks limit the evidence for cross-domain transfer, and we now separate the conclusions supported by real data from those supported by synthetic data.

Real versus synthetic conclusions. Section 4 now reports the real and synthetic results of Table 1 separately. On the five real datasets, EventFM is in the top two in all ten accuracy and CRPS comparisons and has the best CRPS on four, while the next-best method, EVIL, is in the top two in four of these comparisons. On the synthetic groups, which use generators outside EventSS, EventFM has the best CRPS on all three. The real-data forecasting results, EventSS's coverage of real data, calibration, transfer to sequences far from EventSS and the failure modes rest on real data (Sections 4.1 and 4.2, Appendices J, M, N and O). The benefit of EventSS and of the EventFM architecture rests on both real and synthetic data (Section 4.5, Appendices L and R), and the results under controlled shifts and the forecast trajectories rest on synthetic data (Appendix K, Figure 4).

Baseline comparability. A THP-style model trained on the same EventSS data, with matched budget and inputs, is less accurate than EventFM on every real dataset and synthetic group (Appendix L). Despite the information advantage of up to 100 target-domain sequences as context, FIM-PP and EVIL in their native few-shot setting lead zero-shot EventFM by at most 5.6 accuracy points and 6% CRPS where they are ahead, and EventFM remains ahead on Retweet and Amazon (Appendix P).

Benchmark coverage. We agree that the limited diversity of public real-world benchmarks is a limitation of how MTPP models are currently evaluated. We use the same five benchmarks as FIM-PP (Berghaus et al., 2026), and we are working on adding further real datasets within the time constraints of the rebuttal period.

W2: Predictive calibration and finite-quantile scoring

Evidence of calibration. We agree that a low quantile loss does not establish calibration. The new Appendix M measures it directly, with PIT histograms and the coverage of the nine predicted quantiles, on real data and held-out EventSS sequences.

  • On sequences from the training prior, EventFM is calibrated. Coverage matches the nominal level at all nine quantiles, and the PIT histogram is flat.
  • On real data, calibration varies by dataset. Taxi and StackOverflow are close to calibrated. Retweet's upper tail is wider than predicted, Taobao's predicted delays tend to be too long, and Amazon's bimodal delays are represented only coarsely by nine quantiles.

How distributional scores are computed. Each candidate type's delay CDF is linear through the nine quantile knots, from (0, 0) to the upper endpoint q0.9+(q0.9−q0.8)q_{0.9}+(q_{0.9}-q_{0.8}). The marginal CDF mixes them with the predicted type probabilities, and its quantiles at the nine levels are the Q^α\hat Q_{\alpha} below. PIT, coverage and CRPS all use this CDF. The CRPS of Table 1 is the quantile form of CRPS on the nine levels, 29∑αρα(τ−Q^α)\frac{2}{9}\sum_{\alpha}\rho_{\alpha}(\tau-\hat Q_{\alpha}), applied in the same way to every method (for point forecasts it equals the absolute error). For EventFM, the exact CRPS of the full CDF is 3–8% lower on every dataset. Appendices B and G now state the estimator and the endpoint rule.

(2/3)

W3: Structural semantics versus generated-data diversity

Our coverage analysis (Appendix J) already includes structural properties such as type-to-type dependence and type–timing coupling, on which EventSS covers real data better than the intensity-kernel prior. However, matching summary features does not guarantee that a prior contains the mechanisms that produce them. The new Appendix R assesses structural coverage more directly, to determine whether EventSS's gains come from its structural semantics or from the broader diversity of its generated data.

Structural semantics versus diversity (Appendix R.2). The intensity-kernel prior of our w/o EventSS ablation differs from EventSS in both its generation mechanism (additive kernels rather than a queue of pending event opportunities) and the breadth of the families it samples, which controls the diversity of the generated data. We therefore trained EventFM on a 2×22\times2 design that varies the two separately.

  • A kernel prior broadened until it matches EventSS's real-data coverage still trails EventSS in type accuracy on Taxi (0.742 vs. 0.855) and on the deterministic (0.607 vs. 0.801) and mixed (0.525 vs. 0.687) groups.
  • A deliberately narrow EventSS, with two delay families, one graph family per world type and stationary sources, covers real data worse, yet it is more accurate than the broadened kernel prior in all three settings (0.869, 0.747, 0.701).

At matched coverage, the kernel prior therefore misses structure that EventSS covers, and where EventSS helps most, its gains come from its structural semantics rather than from the diversity of its generated data. A shuffle control confirms that EventFM uses conditional structure learned from EventSS on real data (Appendix R.1 and our answer to Q4). We would welcome suggestions if the reviewer had another assessment in mind.

Q1: Next-event types absent from the observed prefix

EventFM predicts only types that occur in the history. Type identities are local to a sequence, so a history carries no information about a type that has not yet occurred, and anticipating such a type is a separate problem, type cold start, which needs information from outside the sequence. We therefore measure how much EventFM's forecast suffers when the next type is absent and how quickly it incorporates a new type once it appears. The new Appendix Q uses the five real datasets and 2,000 held-out EventSS sequences.

  • Next type absent from the history. After 32 observed events, this happens for 5–6% of forecasts on Taxi, Taobao, StackOverflow and Amazon, and for 0.3% on Retweet. On these forecasts, EventFM's type accuracy is zero by construction, but its CRPS is still 6–28% lower than that of mean/mode, whose timing forecast ignores types, on every dataset.
  • After the type has been observed. EventFM incorporates the type without retraining. After a single observation, EventFM's accuracy for the type rises from zero to 8–47%, and it keeps rising with every further observation. Its timing is within 3% of that of a typical forecast at the same history length after two observations on every dataset except Retweet, which reaches it after five.

EventFM therefore forecasts the timing of events whose type it has not yet seen, and it learns a new type in context from a few observations. Using information from outside the sequence for type cold start is an exciting extension for future work.

Q2: Handling quantiles outside the predicted range

Below the 0.1-quantile, each candidate CDF continues linearly down to zero delay. Above the 0.9-quantile, it continues with the slope of the last segment until it reaches 1 at q0.9+(q0.9−q0.8)q_{0.9}+(q_{0.9}-q_{0.8}), and longer delays receive zero probability. Appendix G now states this rule. On real data, 0.9–13.6% of true delays lie beyond this endpoint, but the scores are insensitive to it. Moving the endpoint to q0.9+c(q0.9−q0.8)q_{0.9}+c(q_{0.9}-q_{0.8}) with c∈{0.5,2,4}c\in\lbrace 0.5,2,4\rbrace changes the CRPS of Table 1 by at most 0.7% (Appendix M).

(3/3)

Q3: Dataset-dependent prior gains and target-process structure

We agree that the contribution of EventSS depends on the dataset. The new Appendix R.4, summarized in Section 4.5, identifies which structural properties of the target process are associated with its improvements, property by property rather than by aggregate ranking. We pair each structural property measured by our feature space with the EventSS component that generates it, and ask which datasets show it and whether EventFM improves on them, relative to the w/o EventSS model, to EventFM trained without that component and to mean/mode.

  • Predictable type transitions. This property is strongest on Taxi, where EventFM is 67 points more accurate than the w/o EventSS model and 9.6 points more accurate than without Markov worlds. Within every dataset, the accuracy gain over the w/o EventSS model grows with the property.
  • Slowing activity. This property is strongest on Retweet, where EventFM's CRPS is 3.2% lower than that of the w/o EventSS model and 9.5% lower than without decaying root sources. Within every dataset, the CRPS gain over the w/o EventSS model grows with slowdown.
  • Other properties. Same-type repetition, type–timing coupling, periodic activity and heavy-tailed gaps show no clear association. This does not rule out an effect, as each is clearly present in only one or two benchmarks. Type–timing coupling appears only on Taxi, where predictable transitions dominate the gain. The intensity-kernel prior also represents same-type repetition.

The w/o EventSS ablation therefore loses most where the next type is predictable, as on Taxi, or where activity slows, as on Retweet. The synthetic scenarios, whose structure is known by construction, provide complementary evidence (our answer to W3).

Q4: Conditional mechanism coverage and missing process properties

Conditional mechanisms versus marginal statistics. A new control in Appendix R.1 retrains EventFM on EventSS with the conditional structure removed and the marginal statistics kept. Within each generated sequence, we randomly reassign the event types (keeping type counts and timestamps), or both the types and the gaps.

  • Types. With types reassigned, accuracy falls to near the mean/mode level on Taxi (0.855 to 0.234, mean/mode 0.191), Taobao (0.596 to 0.514, mean/mode 0.515) and all synthetic groups, which removes over 90% of EventFM's advantage over mean/mode. Reassigning types in the real test sequences has the same effect on the frozen EventFM.
  • Timing. With types and gaps reassigned, CRPS is clearly worse on the deterministic and mixed groups and on Retweet. It is similar elsewhere, where timing forecasts rely mainly on the gap distribution of the history.

EventSS therefore covers the conditional mechanisms behind type forecasts on Taxi and Taobao, and the conditional timing of Retweet and the structured scenarios.

Target-process properties not represented by the prior. Appendices N and O identify two, both in the waiting-time distribution, and our answers to R1.W1, R2.W1 and R2.Q1 discuss them in more detail. The test sequences farthest from EventSS on Retweet, Taobao and Amazon typically have heavier-tailed gaps than 95% of EventSS samples, and on Taobao EventFM's CRPS moves from parity with FIM-PP on the nearest sequences to 8% worse on these. The worst-forecast Amazon sequences have more bimodal gaps than 99% of EventSS sequences, and EventFM, which predicts a single mode, underestimates these delays.

We hope these additions address the reviewer's concerns, and we would be glad to answer any further questions.

Reviewer_1mwcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work studies sequence‑level zero‑shot marked temporal point‑process forecasting: given only a single observed event history, frozen model jointly predicts next event type and occurrence time. Authors propose two core components: (1) EventFM, an event‑native PFN‑style transformer with time/mark normalization, candidate‑query prediction and dual‑axis attention; (2) EventSS, an on‑the‑fly synthetic simulator built on pending‑opportunity queues for pretraining, which can express trigger‑response, cancellation, overlapping workflows beyond additive Hawkes‑style intensity kernels. Evaluations cover 5 real‑world EasyTPP benchmarks plus multiple synthetic out‑of‑distribution scenarios, with heavy ablation, sensitivity analysis, theoretical bounds in appendices.

Critical strengths
  1. Zero‑shot foundation‑model setup for marked event sequences fills an important gap; real event data is often fragmented, proprietary, so synthetic‑pretrain paradigm is practically meaningful.
  2. Queue‑based simulation can model instance‑level trigger/cancellation that standard additive intensity‑kernel/Hawkes generators cannot capture. UMAP and feature‑space coverage experiments show EventSS better covers real‑world event manifold.
  3. Comprehensive ablations for architecture modules and EventSS components, hyper‑parameter sensitivity, distribution shift tests, both next‑event and long‑horizon autoregressive rollout. Good comparison against existing zero‑shot/in‑context TPP baselines.
Critical weaknesses
  1. All training data comes purely from EventSS synthetic worlds. Even though feature‑space coverage improves over Hawkes‑kernel prior, there is no guarantee real‑world event dynamics lie near EventSS support. The paper should more explicitly discuss real‑data failure modes.
  2. FIM‑PP and EVIL are originally designed for few‑shot, given multiple target‑domain context sequences. Here authors adapt them to strict single‑history zero‑shot. While this is reasonable, this is not their native setup; results may underestimate their true capability. Should add more discussion about this evaluation caveat.
Questions for authors
  1. EventSS is a very complex simulator. Could you tease apart: how much of EventFM’s gain comes from richer distribution coverage vs particular structural mechanisms inside EventSS? Can you give more intuition which real‑world scenarios would most suffer from EventSS misspecification?
  2. FIM‑PP / EVIL are repurposed for strict single‑history zero‑shot. Could you quantify how much performance drop comes from this deviation from their original few‑shot setup?
2 comments

We thank the reviewer for the detailed review and for recognizing the practical value of synthetic pretraining and the breadth of our ablations. In response, we added a discussion of real-data failure modes (Appendix N), an evaluation on sequences far from EventSS (Appendix O), a few-shot evaluation of FIM-PP and EVIL (Appendix P) and a study that separates coverage from mechanism (Appendix R.2). The revision marks all changes in blue.

W1: EventSS support and real-data failure modes

We agree that there is no guarantee that real-world dynamics lie near EventSS's support, and some real sequences do fall outside it. Our answer to R1.W1 shows that EventFM's transfer degrades gracefully away from EventSS, both on real sequences far from it (Appendix O) and under synthetic shifts outside its training prior (Appendix K.2). To identify where it does fail on real data, a new Appendix N analyses the worst-forecast sequences. We examined the ten worst-forecast test sequences of each dataset and kept a cause only if synthetic sequences with that property reproduced the same error. The three most common causes are the following.

  • Heavy-tailed waiting times (Retweet). EventFM overestimates these delays, as does FIM-PP.
  • Bimodal waiting times (Amazon). About 31% of Amazon delays fall in a short mode that EventFM, which predicts a single mode, does not capture (Appendix M). FIM-PP and EVIL have 30–49% higher quantile loss than EventFM on these sequences.
  • Next type absent from the history (Amazon, StackOverflow, Taobao). EventFM only predicts types that occur in the history. Such targets are uncommon, at 5–6% of next events after 32 observed events (Appendix Q).

W2: Evaluation caveat for FIM-PP and EVIL

We agree that the zero-shot evaluation of FIM-PP and EVIL deserves more discussion, since it differs from the few-shot setting they were designed for. We now also state this caveat in Section 4, next to Table 1, and it points to a new few-shot evaluation of both models (Appendix P), which we discuss in our answer to Q2.

Q1: Distribution coverage, structural mechanisms, and misspecification

Coverage versus structural mechanisms. The w/o EventSS ablation changes coverage and mechanism together. In the new Appendix R.2, we train EventFM on two further priors that each change one factor. The first keeps the additive intensity kernel but broadens its families until it covers real data as well as EventSS. The second keeps EventSS's queue mechanism but narrows its families, so it covers real data worse.

  • Type accuracy comes from the mechanism where type order is informative. On Taxi and the deterministic and mixed scenarios, matching coverage alone recovers 60–83% of EventFM's accuracy gain over the w/o EventSS model. The mechanism alone, with worse coverage, recovers at least 90%. Where type order carries little information, on Retweet, StackOverflow and Amazon, breadth matters more than the mechanism.
  • Real-data timing depends little on the mechanism. Kernel priors match or beat EventSS's CRPS on Taxi, Taobao and StackOverflow. Broader families lower CRPS by about 10% on Retweet and Amazon.
  • Structured timing needs both. On the deterministic scenarios, CRPS is 0.361 for the w/o EventSS prior, 0.274 with broad families only, 0.240 with the mechanism only and 0.145 with both.

Which real-world scenarios suffer most from misspecification. Appendices O and N point to waiting-time distributions that EventSS rarely generates. Heavy-tailed gaps characterize the test sequences farthest from EventSS on Retweet, Taobao and Amazon. On Taobao, EventFM's CRPS moves from parity with FIM-PP on the nearest sequences to 8% worse on these. Amazon's waiting times are bimodal, and EventFM underestimates delays on its worst-forecast Amazon sequences, although FIM-PP and EVIL suffer from the same issue there.

Q2: Performance drop from adapting few-shot baselines

We ran FIM-PP and EVIL in their native few-shot mode (new Appendix P), with k∈{1,5,20,100}k\in\lbrace 1,5,20,100\rbrace sequences from the target dataset's training split as context and the protocol of Table 1. This shows how much each baseline benefits from target-domain information that the zero-shot setting does not provide.

  • EVIL changes little. At k=100k=100 its accuracy rises by 0.025 on average and its CRPS falls by 4%.
  • FIM-PP benefits more on some datasets. At k=100k=100 its accuracy rises by 0.11–0.16 on Taxi, Taobao and StackOverflow, but it loses 0.03–0.09 on Retweet and Amazon at every kk.

Even with 100 target-domain sequences, the best few-shot baseline stays close to zero-shot EventFM, which uses only the query history. Where it is ahead, it leads by at most 5.6 accuracy points and 6% CRPS, and EventFM remains ahead on Retweet and Amazon, by 3.6 and 1.9 accuracy points and by 7% and 17% CRPS.

We hope these additions address the reviewer's concerns, and we would be glad to answer any further questions.

Thank you again for your constructive feedback. We would welcome your thoughts on the additional results we provided in response to your questions:

  • Real-data failure modes, identified on the worst-forecast sequences and reproduced on synthetic sequences (W1; Appendix N)
  • The evaluation caveat for FIM-PP and EVIL, now stated next to Table 1 (W2; Section 4)
  • Coverage versus structural mechanisms, separated with two new priors (Q1; Appendix R.2), and the scenarios most affected by misspecification (Q1; Appendices N and O)
  • FIM-PP and EVIL in their native few-shot setting, with up to 100 target-domain sequences as context (Q2; Appendix P)

Do these revisions and additional results address your concerns? If any aspect remains unclear, we would be happy to discuss it before the discussion period closes tomorrow (Nov 18). If your concerns are resolved, we would be grateful if you would consider updating your assessment.

Best regards,
The Authors

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

EventFM is a transformer-based event forecaster trained entirely on the EventSS simulator, enabling prediction from a single sequence’s history without target-dataset training.

Critical strengths
  • The synthetic-only training and frozen-model evaluation provide meaningful evidence of transfer to real event sequences without target-specific training.
  • The paper is clearly presented and includes useful architecture ablations, simulator ablations, and sensitivity analyses.
Critical weaknesses
  • EventSS relies on manually chosen graph families, event rules, and delay distributions, leaving uncertainty about how reliably EventFM transfers to processes outside these assumptions.
  • The evaluation lacks a simpler event model trained on the same synthetic data, making it difficult to separate the benefits of EventSS from those of the proposed architecture.
Questions for authors
  • How sensitive is real-data performance to removing entire graph or event-rule families from the synthetic training prior?
  • How does a THP-style model trained on the same EventSS data compare under similar training budgets and input representations?
2 comments

We thank the reviewer for the careful review and for recognizing that synthetic-only training with frozen-model evaluation provides meaningful evidence of transfer. In our response, we added a THP-style baseline trained on EventSS (Appendix L), an evaluation on real sequences far from EventSS (Appendix O) and a finer ablation of EventSS's components (Appendix R.3). The revision marks all changes in blue.

W1: Transfer beyond EventSS assumptions

We agree that transfer to processes outside EventSS's assumptions cannot be guaranteed. Appendix I makes this explicit as a misspecification term that longer contexts cannot remove. We therefore measured how reliably EventFM transfers in practice to sequences away from the prior.

Transfer to data away from the prior. On real data (new Appendix O), we compare the test sequences farthest from EventSS in the feature space of Appendix J with the nearest ones. EventFM's accuracy falls with distance, but so does that of the strongest zero-shot baseline, which does not use EventSS. Comparing against this baseline separates misspecification from sequences that are simply harder. Relative to it, EventFM's accuracy changes by less than one point on four of five datasets, and on the farthest sequences EventFM keeps the lowest CRPS of all zero-shot models on every dataset except Taobao. On synthetic data (Appendix K.2), EventFM stays ahead of the naive baselines on delay-law families not used in training and on worlds with twice as many event types as typical training worlds. Although misspecification cannot be ruled out in principle, in practice EventFM still transfers to data away from the prior as well as the baselines do, or better. Taobao timing is the main exception, where EventFM moves from parity with FIM-PP to an 8% deficit on the farthest sequences.

Reliance on manually chosen graph families, event rules and delay distributions. The submission already varies these choices (Appendices B.6.2 and B.6.3), and the new Appendix R.3 removes them one at a time. No single removal changes real-data accuracy by more than 0.8 points (details in our answer to Q1).

W2/Q2: Same-data baseline to separate simulator and architecture gains

We added this baseline (new Appendix L): a Transformer Hawkes Process (THP; Zuo et al., 2020) style model trained on the same EventSS data as EventFM, with matched training budget, input features, objective and prediction set.

With the prior held fixed, the THP-style model is less accurate than EventFM on every real dataset and synthetic group. The gap is largest on Taxi (0.571 vs. 0.855) and on the deterministic and mixed scenarios (0.355 and 0.343 vs. 0.801 and 0.687), where its CRPS is also clearly worse. Conversely, the w/o EventSS ablation in Table 1 keeps the EventFM architecture but replaces the prior, and it also has much lower accuracy on these settings. Both EventSS and EventFM's architecture are therefore needed for maximizing event forecasting performance.

Q1: Sensitivity to removing graph or event-rule families

Real-data performance is not sensitive to removing a single graph family or event rule, but it is sensitive to removing a whole world type or many families at once (new Appendix R.3). We trained 16 EventFM models, each without one EventSS component, such as a graph family, an event rule, a delay family, a root-source type, a state mechanism or a world type.

  • Single families and rules. No graph family or event rule changes real-data accuracy by more than 0.7 points or CRPS by more than 1.5%. The other single components have similarly small effects, except decaying root sources (exogenous arrivals with declining rate). Without them, Retweet CRPS rises by 10.6%. Retweet is the only benchmark whose activity clearly slows down within a sequence.
  • Whole world types. Without Markov worlds, Taxi type accuracy falls by 9.6 points. Without structural worlds, accuracy falls by 0.9–3.3 points on all five datasets. CRPS rises on every dataset without Markov worlds, but only on Retweet without structural worlds. Neither world type alone is best for both accuracy and timing, so the two complement each other.
  • Many families at once. A deliberately narrow EventSS (Appendix R.2) is 0.9–2.9 points less accurate on four of the five datasets. Its CRPS rises by 11.7% on Retweet, 11.6% on Amazon and 4.8% on Taobao, and by 66% on the deterministic scenarios.

Single components have small effects because they overlap, as other graphs and motifs still produce branching or merging. The benchmarks respond to large changes of the prior, but no single specific family drives the results. This is in line with the design goal of EventSS, which is to cover a broad variety of event mechanisms.

We hope these additions address the reviewer's concerns, and we would be glad to answer any further questions.

Thank you again for your constructive feedback. We would welcome your thoughts on the additional results we provided in response to your questions:

  • Transfer away from the training prior, on real sequences far from EventSS and under synthetic shifts outside it (W1; Appendices O and K.2)
  • A THP-style baseline trained on the same EventSS data, which separates the gains of the simulator and the architecture (W2/Q2; Appendix L)
  • Sensitivity to removing single EventSS components, whole world types or many families at once (Q1; Appendices R.2 and R.3)

Do these revisions and additional results address your concerns? If any aspect remains unclear, we would be happy to discuss it before the discussion period closes tomorrow (Nov 18). If your concerns are resolved, we would be grateful if you would consider updating your assessment.

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Discussion Summary

Dear Area Chair,

All three reviewers now recommend acceptance, with Reviewer ZddA raising their score from 2 to 3 after our responses. Since the other reviewers have yet to reply to our responses, we would be grateful if you could invite them to let us know of any remaining concerns before the discussion closes. We believe our responses and the revision address all the points they raised.

To summarize the discussion so far, the reviewers initially:

  • recognized that a zero-shot foundation model for marked event sequences addresses an important gap (1mwc);
  • highlighted the meaningful evidence of zero-shot transfer from synthetic pretraining (mEhG, 1mwc), the novel queue-based simulator EventSS (1mwc, ZddA) and its broader coverage of real data (1mwc), and the principled vocabulary-agnostic prediction head of the EventFM architecture (ZddA);
  • acknowledged the well-designed experiments, comprehensive ablations, thorough sensitivity analyses and comparisons to baselines, which provide useful evidence in the field (mEhG, 1mwc, ZddA).

In response to their feedback, we have substantially strengthened the revised manuscript through:

New ablations

  • A THP-style model trained on the same EventSS data, separating the gains of the EventSS simulator from the EventFM architecture (Appendix L; mEhG, ZddA)
  • Controls that remove conditional structure from EventSS while keeping its marginal statistics (Appendix R.1; ZddA)
  • Priors that vary EventSS's generation mechanism and breadth separately (Appendix R.2; 1mwc, ZddA)
  • Removal of individual EventSS components (Appendix R.3; mEhG)

New analysis

  • Calibration of the predicted time distribution (Appendix M; ZddA)
  • Real-data failure modes (Appendix N; 1mwc)
  • Transfer to real sequences far from the EventSS prior (Appendix O; mEhG, 1mwc)
  • Forecasts with unobserved or infrequently observed next-event types (Appendix Q; ZddA)
  • Structural properties of the target process associated with EventSS's gains (Appendix R.4; ZddA)

Baselines and datasets

  • FIM-PP and EVIL in their native few-shot setting (Appendix P; 1mwc, ZddA)
  • Four additional real-world datasets (follow-up to ZddA)

Enhanced clarity

  • Separate real-data and synthetic conclusions, and the evaluation caveat for FIM-PP and EVIL (Section 4; 1mwc, ZddA)
  • How distributional scores are computed from the nine predicted quantiles (Appendices B and G; ZddA)

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.