- 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
- 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.
- Sharing the prediction head across candidate marks offers a principled way to handle different vocabularies without requiring globally aligned event-type embeddings.
- The experiments are well designed and provide useful evidence in the field.
Critical weaknesses
- 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.
- 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.
- 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
- 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.
- How are quantiles outside the explicitly predicted range handled?
- 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.
- 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
(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 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 , 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 with 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.
(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.
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). The marginal CDF mixes them with the predicted type probabilities, and its quantiles at the nine levels are the Q^α below. PIT, coverage and CRPS all use this CDF. The CRPS of Table 1 is the quantile form of CRPS on the nine levels, 92∑αρα(τ−Q^α), 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.