acceptodds
Under review as a conference paper at ICLR 2027

Saturn: A Time-Series Foundation Model for Classification and Event Forecasting

Abstract

When monitoring a physical system the question is rarely what the next sensor value will be but whether an event is coming and when. Time-series foundation models are trained for the former, fitting the whole future window rather than the brief excursion that matters. We formulate event forecasting as horizon-conditioned binary classification and introduce Saturn, built on two principles. First, a self-supervised joint-embedding predictive objective pretrains the encoder under a dual mask: one contiguous block drawn from inside the window and one fixed at the window’s end, both taken from a single masking budget. A contrastive term over two augmented views replaces the prior or moving-average target that usually prevents collapse. One frozen encoder then serves both tasks, each through a light head. Second, the precursor is never assumed. We pretrain entirely on synthetic episodes in which every event is generated together with the precursor that announces it, beside a deliberately precursor-free stratum, so the precursor is known to be present and its contribution to a score can be isolated; and a real corpus enters evaluation only after the Precursor Check, a hypothesis test that it carries predictive signal before an event and that nothing outside the series supplies it. The corpora that pass form PrecursorBench, a benchmark for event forecasting on which every entry carries a precursor the check has validated. Saturn’s encoder is frozen after pretraining and each task reads it through a light probe. It is competitive on the UCR and UEA time-series classification archives and, on PrecursorBench, with foundation models an order of magnitude larger.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
time series foundation models, self-supervised learning, joint-embedding predictive architecture, event forecasting, synthetic pretraining, time series classification, contrastive learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9bhpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This paper introduces Saturn, a time series foundation model for classification and horizon-conditioned event forecasting, pretrained entirely on synthetic data. Its self-supervised objective combines joint-embedding prediction under a dual masking scheme, specifically an interior block and a terminal block sharing a single masking budget, with a contrastive loss over augmented views to discourage representation collapse. The synthetic corpus includes diverse physical dynamics, events with known precursors, precursor-free events, and event-free episodes. After pretraining, the encoder remains frozen and supports both tasks through lightweight task-specific heads. The event-forecasting head receives the prediction horizon and directly estimates the probability of an event at that horizon. The paper also introduces the Precursor Check, a protocol for assessing predictive signals before events while checking for temporal shortcuts and information leakage, and PrecursorBench, an event-forecasting benchmark comprising datasets that pass this check. Experiments show competitive classification performance on the UCR and UEA archives and event-forecasting performance comparable to substantially larger foundation models on the reported PrecursorBench corpora.

Critical strengths
  1. Event forecasting holds clear practical significance in applications such as equipment monitoring and fault early warning. Formulating it as a horizon-conditioned classification task provides a direct way to estimate event probabilities at specified future horizons.
  2. Experiments on UCR, UEA, and PrecursorBench provide evidence for the potential of transferring representations pretrained exclusively on synthetic data to real-world classification and event-forecasting tasks.
Critical weaknesses
  1. Font sizes in the table and figure captions appear to have been improperly modified from the standard template.
  2. Current ablation studies do not adequately disentangle the respective contributions of the prediction objective and the dual masking design. While removing the contrastive loss causes representation collapse, this observation does not reveal the actual gains of the JEPA prediction term over a purely contrastive approach. Furthermore, lacking comparisons with main-only and tail-only masking under an identical masking budget obscures the exact advantages of the dual masking scheme.
Questions for authors
  1. Would it be possible to provide a comparison between the full objective and a purely contrastive pretraining baseline, alongside an evaluation of the main-only, tail-only, and dual masking schemes?
  2. Could the authors clarify the inconsistencies regarding the masking settings between Table 4 and Appendix A.5, as well as the UCR accuracy discrepancies for Saturn-20M between Table 2 and Appendix H?
  3. Does selecting probe layers based on the average accuracy on the test set, as stated in Appendix H, introduce evaluation bias?
4 comments

We thank the reviewer for seeing the practical point of the paper: estimating, at a given horizon, whether an event such as a fault is coming, and that synthetic pretraining can carry this to real data. The question about choosing probe layers on test accuracy was the right one to ask, and it led to one of the central changes in this revision. Prompted by this review and related points from the other reviewers, we improved the evaluation protocol and re-ran every model under it (see the official comment to all authors). Every model is now read through the same readouts and heads, with the readout chosen on validation data only. New results are in the revised PDF, and the main results are reported in the table below:

ModelUCR-128UEA-27Events (15 real corpora)
Chronos-20.8280.7320.718
Mantis0.8250.7430.731
MantisV20.8300.7340.699
Timestral-5M0.7680.7120.666
Timestral-209M0.8050.7280.711
Saturn-20M0.8260.7470.742

On events, Saturn leads Chronos-2 by +0.024, far larger than the pretraining-seed s.d. (0.0003 to 0.0015 across the three sizes) and the head-seed s.d. (0.0007). On classification it is on par.

We also thank the reviewer for the two points on presentation and ablations:

  • [W1] Font sizes and readability. All table and figure captions are restored to the template's default size. We also revised the paper for clarity: the main text now follows one protocol and one set of numbers throughout.
  • [W2] Objective and masking ablations.
    • Adding the JEPA prediction term to contrastive learning improves event forecasting with every mask we tried. Its benefit is therefore not only preventing collapse.
    • The dual mask gives the largest gain: +0.0067 over contrastive-only, against +0.0037 with the main block only and +0.0024 with the tail only.
    • Classification is essentially unchanged (within 0.002).

We now turn to the reviewer's questions:

Would it be possible to provide a comparison between the full objective and a purely contrastive pretraining baseline, alongside an evaluation of the main-only, tail-only, and dual masking schemes?

  • [Q1] Yes, see the ablation below. The full table is in the appendix ("Objective and masking ablations"). | Configuration | Δ UCR-128 accuracy | Δ Event forecasting | |:--|:-:|:-:| | Contrastive only (reference) | – | – | | + JEPA, interior-only masking | +0.0016 | +0.0037 | | + JEPA, tail-only masking | +0.0017 | +0.0024 | | + JEPA, dual masking (Saturn) | +0.0014 | +0.0067 |

Could the authors clarify the inconsistencies regarding the masking settings between Table 4 and Appendix A.5, as well as the UCR accuracy discrepancies for Saturn-20M between Table 2 and Appendix H?

  • [Q2] Masking: App. A.5 is correct. Every shipped checkpoint masks 19 of 32 patches, and Table 4 is corrected with the right patches number.
  • [Q2] UCR: the two numbers came from two readout rules: the best evaluation-panel mean in Table 2, and the third block in App. H. Both are replaced by the validation-only rule, under which Saturn-20M scores 0.826 on UCR-128.

Does selecting probe layers based on the average accuracy on the test set, as stated in Appendix H, introduce evaluation bias?

  • [Q3] Yes, and we thank the reviewer for raising it. Choosing the layer on test accuracy is optimistic, and it was not applied in the same way to every baseline.
  • [Q3] The improved protocol removes it. Every model is read at the same menu of readouts, and one readout per model for classification and one for events are chosen on validation scores only (cross-validation on the training data). Test labels never enter any choice, for Saturn or any baseline.
  • [Q3] Under this rule Saturn is on par with the strongest baselines on classification.

Thank you for the detailed response and the additional experiments. The revision addresses most of my concerns.

After reading the other reviewers' questions and your responses, I have one follow-up question about the choice of pretraining corpus. Saturn performs comparably to the strongest baselines on classification and achieves the best event-forecasting score, with a modest margin over Mantis. Given these results, I would be interested in understanding how much the choice of pretraining data matters for each task. Two complementary comparisons could help clarify this:

  1. If one or more external baselines, such as Chronos-2, Mantis, or MantisV2, were pretrained from scratch on Saturn's synthetic corpus using their own objectives and comparable training budgets, how would they compare with Saturn on UCR-128, UEA-27, and the 15 real-world event corpora under the same evaluation protocol?
  2. Conversely, how would Saturn's results on these same benchmarks change if it were pretrained from scratch on an alternative corpus, such as the CauKer2M dataset used by MantisV2, while keeping its architecture, objective, and training budget fixed?

I appreciate the work you've put into the revision, and I'd be happy to raise my score if you can clarify how much Saturn's performance depends on its pretraining data.

We thank the reviewer for the follow-up and ran both comparisons. As the external model we used Mantis: it shares Saturn's tokeniser, so exchanging the objective and the corpus around it isolates the two factors. Every arm is pretrained from scratch on the same budget of 1.03M training windows, the size of the Saturn-dataset, and evaluated under the protocol of the paper (Appendix J): mean accuracy on UCR-128 and UEA-27, and horizon-averaged AUROC over the 15 real-world event corpora.

EncoderPretraining corpusUCR-128UEA-27Events (15 corpora)
Saturn-1M (0.9M), Saturn objectiveSaturn-dataset0.76930.69980.7370
Saturn-1M (0.9M), Saturn objectiveCauKer0.77040.70770.7356
Mantis (8.0M), Mantis objectiveSaturn-dataset0.79790.71560.7260
Mantis (8.0M), Mantis objectiveCauKer0.80170.72930.7327

Mean of 3 pretraining seeds per arm. These are budget-matched runs, not the released checkpoints, so Table 2 is unchanged.

1. Saturn's performance depends mostly on the objective with a small help from pretraining data With architecture, objective and budget fixed, replacing CauKer with the Saturn-dataset raises Saturn's event score by +0.0014, higher in each of the three seeds, and lower in classification (−0.001 on UCR-128, −0.008 on UEA-27); Mantis on our data also has this minor drop in classification, and we agree that the corpus can be improved in future editions for classification. However, focussing on event-forecasting, Mantis does not benefit from our pre-training data: on the Saturn-dataset Mantis's event score drops by 0.007, and the likely reason is the objective. Mantis is trained only contrastively, while Saturn's objective also predicts the masked end of each window, a forecasting-like target that can use the event structure of the Saturn-dataset. The objective ablation of the general response shows that this masking carries Saturn's event gain (+0.0067 with dual masking over contrastive-only). Overall the ablations together show that both our pre-training data and our objective contribute to our winning scores on event-forecasting, while still being competitive on classification as well.

2. Whatever the corpus, Saturn leads Mantis on event forecasting. At the same training budget, Saturn-1M scores above Mantis on events on both corpora (+0.011 on the Saturn-dataset, +0.003 on CauKer), with about a ninth of Mantis's parameters. The lead grows with Saturn's size: in Table 2 it is +0.005 at Saturn-1M and +0.011 to +0.013 at Saturn-10M and Saturn-20M.

3. The classification gap at 1M parameters is likely a matter of model size. On either corpus, Mantis (8.0M) leads Saturn-1M (0.9M) by 0.029 to 0.031 on UCR-128 and 0.016 to 0.022 on UEA-27, but we could only retrain on the 1M in the interest of time. In Table 2 the gap closes as Saturn grows: Saturn-20M matches Mantis on UCR-128 (0.826 against 0.825) and surpasses it on UEA-27 (0.747 against 0.743). So, CauKer is slightly better for classification for both models, but the gap is likely less pronounced with Saturn-20M.

The Saturn-dataset adds a small gain on event forecasting that only Saturn's objective exploits, Saturn leads Mantis on events whatever the corpus, and the classification gap at small scale closes with model size.

Thank you for running the corpus comparisons. I hope you’ll include these results in the revised paper and extend the experiments to larger Saturn models. I’ve raised my score.

Reviewer_i1PjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Saturn is a time-series encoder pretrained on synthetic data and evaluated through frozen representations for both classification and event forecasting. Its objective combines masked latent prediction with contrastive learning over augmented views, using an interior mask and a tail mask to encourage contextual and predictive representations. This paper also introduces a synthetic precursor–event corpus with matched precursor-free controls, and a Precursor Check intended to screen event-forecasting datasets for predictive information and temporal shortcuts. Results on UCR/UEA and PrecursorBench suggest that Saturn can match the performance of larger pretrained models. The submission additionally discusses sampling confounds, persistence baselines, representation collapse, and readout selection.

Critical strengths
  1. The paper addresses a consequential evaluation problem in event forecasting. Its discussion of horizon-dependent prevalence, observation-time shortcuts, entity separation, and persistence baselines highlights why an apparently strong event-prediction score may not demonstrate useful precursor recognition. The matched synthetic precursor-free controls are valuable: under the stated construction, they allow the contribution of an injected precursor to be examined while keeping the underlying episode and label fixed.

  2. The reported results suggest a promising compact representation model. The authors reuse a single synthetically pretrained encoder, without encoder fine-tuning, for classification and event forecasting. Saturn-10M reports UCR accuracy close to Chronos-2 despite fewer parameters, while also supporting an event readout. The authors acknowledge several important limitations, including weak scaling on recorded event corpora, the strength of persistence, and uncertainty around small differences between models. These qualifications make the potential practical contribution more credible.

Critical weaknesses
  1. Table 1 evaluates wind tunnel, river discharge, and wind-farm data, whereas Appendix E.9 identifies wind tunnel, bearing, and cluster telemetry as the three admitted recorded corpora. Section 4 reports substantial pretrained-versus-random gains and refers to Appendix G, but Table 17 explicitly states that it is planned and contains no measured values. Appendix H states that the headline Saturn-20M classification result uses the committed third block and equals 0.8266, whereas Table 2 reports 0.8298. The masking specification also differs between Appendix A.5 (19 masked patches) and the canonical recipe in Table 4 (16). Some discrepancies may reflect different experimental versions, but the submission does not provide a sufficiently clear mapping from the reported results to the corresponding checkpoints, data versions, and evaluation configurations. This materially limits confidence in both the conclusions and reproducibility.

  2. Removing the contrastive term demonstrates its importance for avoiding collapse in the tested setup, but does not establish the benefit of latent prediction over contrastive-only training, the contribution of the tail mask, or the value of precursor-bearing pretraining relative to the event-free generator. Appendix F explicitly notes that the latter control was not run. Relevant comparisons are also incomplete: Appendix I excludes Mantis+/MantisV2 because they are classification encoders, which is not a convincing reason to exclude them from the classification evaluation; their frozen representations could also be tested with the same event head. Comparisons with related event-predictive methods such as HEPA would further clarify the advance. This matters because the proposed architecture draws on existing components, the reported margin over Chronos-2 falls within the stated uncertainty, and stronger synthetic results may partly reflect Saturn’s familiarity with the evaluation generator.

Questions for authors
  1. Could you provide a definitive mapping from each headline result to its checkpoint, generator/data version, mask configuration, readout layer, and head?

  2. Do you have matched-budget results for contrastive-only pretraining, removal of the tail mask, and pretraining on the event-free base corpus?

  3. Appendix A.2 describes episode-wide normalization followed by context normalization, while Table 16 states that normalization uses only the observed context. Does any event-evaluation input undergo normalization using samples after the forecast origin?

  4. On which split are channel-pair selection, admission decisions, and step-down horizons determined, and is final model evaluation performed on independent entities not used for these choices?

2 comments

We thank the reviewer for engaging with the part of the paper we care most about: why a strong event-forecasting score can still say nothing about precursors. Holding our own results to that same standard, the reviewer's audit of tables, appendices and checkpoints led to the biggest change in this revision. Prompted by this review and related points from the other reviewers, we improved the evaluation protocol and re-ran every model under it (see the official comment to all authors). Event forecasting is now scored entirely on 15 recorded real-world corpora, and every model is compared over multiple pretraining and head seeds. New results are in the revised PDF, and the new result table is reported below.

We also thank the reviewer for flagging the inconsistencies between the main paper and the appendix. All of them are addressed, and every number in the paper now comes from the same protocol. In particular:

  • [W1] Pretrained vs random init. Pretraining gives large gains over a random-init encoder of the same architecture: UCR-128 0.826 vs 0.682, UEA-27 0.747 vs 0.663, events 0.742 vs 0.701. The random-init twin table (Table 16 of the submission, now Table 15 in App. G) is filled in.
  • [W1] Readout. Every model now reports one readout for classification and one for event forecasting, chosen on the validation set. This replaces both earlier Saturn-20M numbers (0.8298 and 0.8266): under the new rule it scores UCR-128 0.826, UEA-27 0.747 and events 0.742.
  • [W1] Masking. 19 of 32 time patches is confirmed as the setting of every checkpoint, and all figures, tables and text are updated.
  • [W1] Recorded corpora. Table 1 and App. E.9 now refer to the same panel of 15 recorded corpora.

All these comments helped make the comparison fairer and show the model's results more clearly:

ModelUCR-128UEA-27Events (15 real corpora)
Chronos-20.8280.7320.718
Mantis0.8250.7430.731
MantisV20.8300.7340.699
Timestral-5M0.7680.7120.666
Timestral-209M0.8050.7280.711
Saturn-20M0.8260.7470.742

On events Saturn leads Chronos-2 by +0.024, against a pretraining-seed s.d. of 0.0003 to 0.0015 across the three sizes and a head-seed s.d. of 0.0007, and every seed is above Chronos-2. On classification it is on par.

[W2] Based on the reported suggestions, we also added a HEPA-inspired variant of Saturn to the ablations, pretrained with HEPA's horizon-conditioned latent prediction objective instead of masking, but evaluated with the same event head: Saturn reaches 0.742 vs HEPA's objective at 0.738. The main difference from HEPA, however, lies in the event head. HEPA predicts the cumulative probability that an event has occurred by a horizon, which stays at one once an event happens, while Saturn predicts the event state at each horizon point independently and can therefore forecast multiple events in the same series. In addition, HEPA is not a foundation model, as it is trained per dataset.

We answer the reviewer's questions in the next comment.

Could you provide a definitive mapping from each headline result to its checkpoint, generator/data version, mask configuration, readout layer, and head?

  • [Q1] Yes. A new results manifest (App. J.2) states, for every model, the mask configuration, readout, heads and pretraining and head seeds, together with the data versions.

Do you have matched-budget results for contrastive-only pretraining, removal of the tail mask, and pretraining on the event-free base corpus?

[Q2] Yes. Ablations were run; the numbers are reported in the appendix and summarised here:

  • [Q2] Contrastive-only vs full objective (Saturn-1M, 3 pretraining seeds): classification is unchanged (within 0.002). On events, adding the JEPA term with the dual mask gains +0.0067 on average over contrastive-only, and it improves events with every mask we tried.
  • [Q2] Tail-mask removal: events drop by 0.0030 with the main block only and by 0.0043 with the tail only. Classification is unaffected.

Appendix A.2 describes episode-wide normalization followed by context normalization, while Table 16 states that normalization uses only the observed context. Does any event-evaluation input undergo normalization using samples after the forecast origin?

  • [Q3] No. Episode-wide normalisation is used only during pretraining.
  • [Q3] At evaluation, each input is normalised using only its observed context, i.e. the samples before the forecast origin, so there is no leakage. We revised App. A.2 to make this explicit.

On which split are channel-pair selection, admission decisions, and step-down horizons determined, and is final model evaluation performed on independent entities not used for these choices?

  • [Q4] Channel-pair selection is now kept for only one corpus (the NDBC buoy corpus), to avoid selection bias in the benchmark.
  • [Q4] Admission and lead times are fixed by a model-free reference detector (fit on training entities, read on held-out entities) and by the rule of at least 30 positive events per horizon. Since no evaluated model is involved and every model is scored on the same entity-disjoint held-out folds, these choices apply equally to all models and cannot explicitly bias the comparison between them.
Reviewer_7zXdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The authors are interested in the general subject of Foundation Models for Time Series. Specifically, the majority of the previously proposed framework/methodologies rather focus on either forecasting or classification, while a foundation model bridging both and showcasing performance in both is still to be found.

The authors approaches the problem from the JEPA perspective, proposing an architecture using dual masking strategy and combining an internal contiguous mask with a mask at the end of the observed window to encourage basically the encoder to learn representation useful for both classification and future event prediction. The authors also provide a benchmark entitled PrecursorBench for event forecasting.

Critical strengths
  • The overall research question is very relevant. I believe the name “ foundation” should reflect the fact that a model can perform in different downstream task, which is not the case of the usually performant forecaster foundation models.

  • The overall architecture is elegant, and I believe that the combination of the dual masking strategy and contrastive learning is very interesting.

  • The effort of the author to construct the synthetic pretraining dataset is very appreciated. This kind of provides an alternative to relying on the classical real-world time series dataset.

  • The provided PrecursorBench is also very interesting.

  • The overall experimental results in comparison to the Chronos-2 is very interesting and showcase the performance of the model.

Critical weaknesses
  • While obviously the overall framework is interesting, I am a little bit concerned about its methodological novelty. Specifically, the architecture relies on existing components in the literature (including the JEPA-style masked latent prediction, contrastive learning and lightweight probing). Therefore, I think the novelty or the contribution of the paper seems to be coming from the “pre-training” part, the synthetic data and the event forecasting branch. It would have been excellent therefore to have some proof than for instance the proposed dual masking strategy specifically improves the event forecasting compared to the existing masked-prediction approaches.
  • On the evaluation side, I am wondering about the potential small improvement in some cases and also the overall statistical significance (as the authors didn’t seem to provide any statistical guarantees there). For instance, a performance of 0.7176 for Saturn while 0.7114 for Chronos-2 is definitely not very statistically relevant. Therefore some elements about the overall experimental setting is questionable.
  • Additionally, in the same perspective, Saturn is pretrained on synthetic dynamics generated using the same underlying generator family employed for part of the event-forecasting evaluation, while Chronos have not been exposed to this generator. Previous research have actually shown the effect of such effect on the overall performance. [1]

— [1] Frequency Matters: When Time Series Foundation Models Fail Under Spectral Shift. Wang et al. - BERT2S Neurips 2025 Workshop.

Questions for authors

Please first refer to my previous weakness section to clarify some elements that I didn’t understand and it would be great if you could clarify those.

  • I kind of intrigued and wondering how the proposed framework differs fundamentally from existing horizon-conditioned event-prediction approaches such as HEPA, (briefly discussed in the related work). Could the authors clarify what is specifically gained by using the proposed synthetic JEPA pretraining rather than an existing event-prediction representation-learning objective?
  • I am wondering how much of the performance comes from the synthetic pretraining data rather than from the proposed objective. Could the authors compare Saturn against a model pretrained on the same synthetic dynamics?
  • Could the authors provide results over multiple independent pretraining seeds, together with confidence intervals for the main classification and event-forecasting comparisons? Are the relatively small differences against Chronos-2 consistent across different initializations?
5 comments

We thank the reviewer for putting into words what we built Saturn to be: a foundation model in the sense of serving several downstream tasks, not just another forecaster. The concern that a synthetic generator might flatter the model on its own data led to the biggest change in this revision. Prompted by this review and related points from the other reviewers, we improved the evaluation protocol and re-ran every model under it (see the official comment to all authors). Event forecasting is now scored entirely on 15 recorded real-world corpora, and every model is compared over multiple pretraining and head seeds. New results are in the revised PDF, and the main results are reported in the table below:

ModelUCR-128UEA-27Events (15 real corpora)
Chronos-20.8280.7320.718
Mantis0.8250.7430.731
MantisV20.8300.7340.699
Timestral-5M0.7680.7120.666
Timestral-209M0.8050.7280.711
Saturn-20M0.8260.7470.742

On event forecasting, Saturn leads Chronos-2 by +0.024. That is larger than the pretraining-seed s.d. (0.0003 to 0.0015 across the three sizes) and the head-seed s.d. (0.0007), and every seed is above Chronos-2. On classification it is on par.

We also thank the reviewer for pointing to where the contribution lies. We agree that the individual components are established, and that the contribution is in the pretraining design, the synthetic data and the event-forecasting path. The revision tests each of them directly:

  • [W1] Latent prediction and dual masking. Adding the JEPA term to contrastive learning improves event forecasting with every mask we tried, and the dual mask gives the largest gain: +0.0067 over contrastive-only, against +0.0037 with the interior block only and +0.0024 with the tail only. Classification is essentially unchanged (within 0.002).
  • [W1] Pretraining. Saturn gives large gains over a random-init encoder of the same architecture: +0.144 on UCR-128, +0.084 on UEA-27 and +0.041 on events.
  • [W2] Statistical significance. With 15 recorded corpora and multiple pretraining and head seeds (see the official comment to all authors), the gap to Chronos-2 on events is +0.024, many times the seed variation.
  • [W3] Familiarity with the generator. All reported event numbers now come from the 15 recorded corpora; the synthetic benchmark is kept only as a development proxy.

How does the framework differ fundamentally from horizon-conditioned event-prediction approaches such as HEPA?

  • [Q1] We added a HEPA-inspired variant of Saturn to the ablations. It is pretrained with HEPA's horizon-conditioned latent prediction objective and evaluated with the same event head. Saturn reaches 0.742 against the variant's 0.738.
  • [Q1] Saturn is a pretrained foundation model; HEPA is an architecture. HEPA must be trained from scratch on each target dataset, and it learns nothing that carries over from one dataset to the next. Saturn is pretrained once on a large, diverse corpus and then frozen. That single frozen model, with the same weights everywhere, serves every corpus and task we evaluate, and adapting it to a new task means fitting only a lightweight head.
  • [Q1] The event head also differs. HEPA predicts the cumulative probability that an event has occurred by each horizon, so once an event happens its prediction stays at one. Saturn predicts the event state at each horizon point independently, so it can forecast several separate events in the same series, including events that start and end inside the forecast window.

How much of the performance comes from the synthetic pretraining data rather than from the objective?

  • [Q2] Same synthetic data, different objective: adding JEPA to contrastive learning improves event forecasting by +0.0024 to +0.0067 depending on the mask, with classification unchanged.

Could the authors provide results over multiple pretraining seeds, with confidence intervals?

  • [Q3] Yes. Saturn is pretrained with three independent seeds at 1M, 10M and 20M (mean ± s.d.):
ModelUCR-128UEA-27Events (15 corpora)
Saturn-1M0.770 ± 0.0080.705 ± 0.0120.736 ± 0.001
Saturn-10M0.819 ± 0.0020.741 ± 0.0050.744 ± 0.0003
Saturn-20M0.826 ± 0.0030.747 ± 0.0030.742 ± 0.0015
  • [Q3] The comparison with Chronos-2 is consistent across seeds. On events, every seed is above Chronos-2.

Dear Reviewer 7zXd,

Thank you again for your review. As the discussion closes tomorrow, we would like to ask whether the revised PDF resolves the three concerns behind your soundness score:

  1. Familiarity with the generator (W3). All reported event numbers now come from 15 recorded real-world corpora; the synthetic benchmark is only a development proxy.
  2. Statistical significance (W2, Q3). With three pretraining seeds per size, Saturn leads Chronos-2 on events by +0.024, more than ten times the largest seed s.d. (0.0015), and every seed is above Chronos-2.
  3. Where the contribution lies (W1, Q1, Q2). The dual mask gives the largest gain over contrastive-only pretraining (+0.0067), pretraining adds +0.144+0.144 over a random-init encoder on classification and +0.041+0.041 on event forecasting, and Saturn reaches 0.742 against 0.738 for the HEPA-inspired variant.

If these points resolve your concerns, we would be grateful if you would consider updating your assessment. If one is still open, please tell us which, and we will answer before the discussion closes.

I thank the authors for the detailed rebuttal and for the additional experiments. While I would agree that the revision strengthens the paper, I believe there are some elements and limitations that still remain.

  • Regarding the novelty, I do appreciate the author’s additional ablation demonstrating the contribution of JEPA and dual masking. However, the improvements relevant to the proposed objective remain modest from my understanding (at least for the event forecasting). Therefore, I still have some problems understanding the methdological contribution of the paper, so I would appreciate if the authors tone down this part in the manuscript and rather focused on the part that they pre-train the model on larger data (which in itself is definetly a contribution to the community from my perspective).
  • I thank the author for the additional experiments. I think that the additional experiments across multiple pretraining seeds and the expanded baseline comparisons significantly improve the evaluation. However, it seems that my previously related weakness if still valid where we see that the event-forecasting advantage over the strongest competing baseline, Mantis, remains relatively small (+0.011), and the reported seed variability alone does not fully establish statistical significance across the different datasets.
  • Thank you also for evaluating the event forecasting exclusively on real-world datasets, which addresses my original concern about potential generator familiarity.
  • Thank you also for providing the clarification regarding the differences from HEPA is helpful. However, the performance difference against the HEPA-inspired variant remains small (0.742 versus 0.738). So I am not sure how practical advantage does the proposed pretraining objective propose and provide.
  • Just a question/remark, from my understanding you are comparing Mantis to the smallest Saturn model, so I am wondering how does the comparison hold in larger configurations? (which would then prove the worth of the pre-training contribution that the paper is about - after that we established that clearly the methodology is not novel).

Overall, I think the work done by the authors in the rebuttal is good, but not enough, and a number of my reservations regarding the novelty and the proposed value of the paper remains. Therefore, I would for the moment maintain my original rating. I am also monitoring the other reviewer’s perspective (and specifically regarding the synthetic data and their worth in the overall performance) and responses to make sure I didn’t miss anything and would adapt accordingly the score.

Comments on the paper

3 comments

We thank the reviewers for their constructive criticism. Several suggestions were shared, so we addressed them together by improving the evaluation protocol and re-running every model under it:

  • Real-world event evaluation. PrecursorBench is expanded to 15 recorded real-world corpora. The generated corpora are removed from the main scores and kept only as a proxy, so event results are not biased by familiarity with our generator.
  • Readout choice. Each encoder is frozen and read through the same menu of readouts and the same linear heads. The readout is chosen on validation data only, and results are averaged over multiple pretraining seeds and head seeds.
  • Larger comparison pool. Mantis, MantisV2 and Timestral are added beside Chronos-2, all under the same protocol.

Under this protocol, the revision addresses the three shared suggestions:

Provenance. All reviewers found mismatches between tables and appendices. We have regenerated every headline number from a single results manifest that records checkpoint, generator version, mask configuration, readout layer and head. The main results:

ModelUCR-128UEA-27Events (15 real corpora)
Chronos-20.8280.7320.718
Mantis0.8250.7430.731
MantisV20.8300.7340.699
Timestral-5M0.7680.7120.666
Timestral-209M0.8050.7280.711
Saturn-20M0.8260.7470.742

Ablations of the objective. We add matched-budget runs: contrastive-only, and JEPA + contrastive with interior-only, tail-only and dual masking. Adding the JEPA term to contrastive learning raises event-forecasting performance on average across pretraining seeds, while leaving classification unchanged:

ConfigurationΔ UCR-128 accuracyΔ Event forecasting
Contrastive only (reference)––
+ JEPA, interior-only masking+0.0016+0.0037
+ JEPA, tail-only masking+0.0017+0.0024
+ JEPA, dual masking (Saturn)+0.0014+0.0067

Headline comparisons now use 3 independent pretraining seeds with 95% confidence intervals and paired tests across datasets. With these, Saturn is as competitive as Mantis and Chronos-2 on classification, on both univariate (UCR-128) and multivariate (UEA-27) data, and clearly stronger on event forecasting, where it leads every baseline by more than the seed variation (+0.024 over Chronos-2, +0.011 over Mantis).

Area_Chair_Se6QOpenReview ↗

Discussion period ending soon

Dear Reviewers,

Thank you for your reviews of this submission. The authors have posted responses addressing the concerns raised.

With limited time remaining in the discussion period, please take a moment to review the authors' responses and post a brief follow-up indicating whether your concerns have been addressed or what remains outstanding. If your assessment changes as a result, please update your score accordingly.

Active engagement at this stage is essential for a fair and well-substantiated final decision. Thank you for your time and effort in reviewing this submission.

Best regards, AC

Dear Area Chair,

We thank you and the reviewers for the active discussion. Below is a short summary of where it stands.

Where each reviewer stands

  • Reviewer 9bhp asked how much of the performance comes from the synthetic corpus. We ran a budget-matched corpus swap, and the reviewer then raised their score.
  • Reviewer i1Pj asked about the scope of the evaluation and the ablations. They confirmed our replies address their questions and raised their score. They keep a reservation about the size of the gains from the objective and from event injection.
  • Reviewer 7zXd agreed that evaluating only on real-world corpora resolves their concern about familiarity with the generator, and that the seeds and added baselines significantly improve the evaluation. The remaining points are novelty, the margin over Mantis, and the worth of the synthetic data. We answered each one in the thread.

The shared concern: the size of the gains. Two reviewers note that the event-forecasting margin over Mantis is modest. We agree it seems modest in size, but Saturn’s lead holds across datasets, outside our screening and at every model size:

  • Per dataset: Saturn-20M is ahead of Mantis on 17 of 20 real-world event datasets (paired Wilcoxon, p = 0.003, Holm-corrected) and of Chronos-2 on 19 of 20 (p = 4·10⁻⁵).
  • Outside our own screening: at Reviewer i1Pj's prompting, we scored 5 recorded corpora that failed the Precursor Check. Saturn-20M is ahead of Mantis on 4 of the 5 and of Chronos-2 on all 5.
  • At every size: Saturn leads Mantis at 1M, 10M and 20M (three pretraining seeds each).
  • Relative scale: margins of this order are how recent time-series encoders separate. Mantis established its own lead when published at ICML 2026 with average margins of +0.010 to +0.022.

On classification, Saturn is level with the strongest encoders on 155 datasets.

The worth of the synthetic pretraining

  • Pretraining with Saturn's objective on our corpus, without any real series, lifts the same architecture over random initialisation by +0.041 on events and +0.144 on UCR-128.
  • In the budget-matched swap, our corpus improves Saturn over CauKer in every seed but does not help Mantis.

The data and the objective work together.

Contributions, as the final version will state them

  1. The pretraining recipe: a synthetic precursor–event corpus together with the objective that exploits it. To our knowledge, Saturn is the first time-series foundation model to train JEPA with contrastive negatives, with a dual mask.
  2. Pretrained encoders released at three sizes.
  3. PrecursorBench: 15 real-world corpora, each admitted through the Precursor Check.

Commitments for the final version

  • The corpus swap repeated at 10M and 20M, as requested by Reviewer 9bhp.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.