acceptodds
Under review as a conference paper at ICLR 2027

Scaling Gains Have Shape: Structural Response in Time-Series Foundation Models

Abstract

Scaling a time-series foundation model is usually summarized by the error reduction from a smaller checkpoint to a larger one. This number identifies whether an upgrade works, but collapses the data conditions that support its gain. We introduce Structural Scaling Response (SSR): the upgrade's log-risk gain on the original series minus its gain on a matched surrogate that preserves the power spectrum while randomizing Fourier phases. Positive SSR means that the original temporal phase structure amplifies the upgrade gain; negative SSR means that it attenuates the gain. Restoring the original phases from low to high frequencies produces an SSR curve that locates the timescales at which the difference emerges. We characterize when this response vanishes and bound the risk perturbation needed to change its sign. On untouched PEMS08, all four larger checkpoints improve accuracy, with channel-aggregated error reductions of 10.02–38.54%. Their SSRs nevertheless separate: the two Toto upgrades and the Chronos upgrade are positive (, , and ), while the Moirai upgrade is negative (). Matched surrogates, alternative losses, blocked resampling, mean aggregation, and channel-resampling analyses preserve all four signs. Across 16 upgrade–dataset combinations, SSR is specific to the pair: the same upgrade changes across datasets, and one dataset can contain both signs. On PEMS04 and PEMS08, each positive SSR reaches half its endpoint after phases at periods down to 8–15 hours have returned. Temporal phase structure is therefore a second coordinate of scaling. SSR provides a training-free protocol for measuring where scaling gains arise and comparing upgrades under deployment-relevant temporal structure.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series foundation models, model scaling, temporal structure, Fourier phase, surrogate analysis, zero-shot forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hviwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Structural Scaling Response (SSR), a metric used to measure how much the metric improvement by scaling the size of a time series foundation model is attributed to the temporal structure of the evaluation data. SSR is measured by evaluating the difference between the improvement on the downstream dataset and the improvement on randomized Fourier features as the model scales up. By maintaining the power spectrum of the Fourier phases, the SSR metric can be used to distinguish between the gains from the original temporal structure and the gains that persist under phase randomization. The paper then applies SSR to many classes of time series foundation models, including Toto, Chronos, and Moirai to show the scaling behavior of each class of models.

Critical strengths
  • The question studied in the paper is interesting and novel. It does not only look at the "scaling law" of time series foundation models by analyzing the standard metrics, but also asks where the gain is coming from. I think this is a more interesting problem that will otherwise be overlooked if one only focuses on the metrics.
  • SSR provides a straightforward, training-free, and intuitive diagnostic that can be applied to existing pretrained models. It does not depend on the architecture of the model class per se, which makes it widely applicable.
  • The empirical evaluation, while not completely self-explanatory (see my comments in the weakness section), leads to interesting findings on different time series foundation model classes, which potentially suggest revealing distinctions between them.
Critical weaknesses
  • The paper offers one potential approach to analyze the attribution of gains from a larger time series foundation model via SSR. The proposed method is intuitive, and provides a reasonable way to distinguish between different sources of the gain, but it is unclear why this particular analysis should be considered as the appropriate test. In particular, the paper would benefit from a discussion of the limitations of SSR, and how it would compare to other possible evaluation metrics which are not proposed in the paper.
  • While the paper shows interesting findings related to the different classes of time series foundation models, it does not totally explain why such a distinction exists. I think the paper can be strengthened by a brief discussion or explanation of this distinction, followed by some stronger evidence beyond the SSR measurement to demonstrate this difference.
Questions for authors
  1. While the randomized Fourier feature preserves the spectral information, it would change some other properties of the time series. What makes SSR particularly informative compared with alternative ways of modifying temporal structure? Could other interventions lead to qualitatively different conclusions about the same model upgrades?
  2. Do the authors have hypotheses about which attributes, including the model architecture and training datasets, contribute to the contrasting behavior of Toto, Chronos, and Moirai?
  3. How sensitive are the findings to forecasting horizon, context length, and temporal resolution? It would be interesting to see whether the SSR scores and the identified frequency ranges remain stable under different forecasting configurations.
Reviewer_D29oOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I do not understand the paper at all. It has something to do with analyzing the different sizes of the same time series foundation model, but I don't understand anything beyond that.

Critical strengths

NONE.

Only if I could undersatnd the paper can I even judge it. I see no critical strengths as the entire paper was completely incomprehensible.

Critical weaknesses

The paper is incomprehensible as written, so it cannot be evaluated

The examples below are taken from the first few pages; the same pattern continues through Section 3, after which I could no longer follow the paper. I list them so the authors and the ACs can see the problem concretely. Specific examples:

  • Abstract is very confusing and does not serve the purpose:
  • The very third sentence is "We introduce Structural Scaling Response (SSR): the upgrade’s log-risk gain on the original series minus its gain on a matched surrogate that preserves the power spectrum while randomizing Fourier phases" What is upgrade? What is log-risk gain? It assumes that all this is defined. The abstract has to discuss the work in an abstract manner (literally).
  • I don't undersatnd the rest of the abstract too:
    • "We characterize when this response vanishes and bound the risk perturbation needed to change its sign." What does any of this mean?
    • "On untouched PEMS08, all four larger checkpoints improve accuracy, with channel-aggregated error reductions of 10.02–38.54" What is untouched PEMS08? Is it a dataset? There is no reference anywhere in the paper.
    • "Their SSRs nevertheless separate: the two Toto upgrades and the Chronos upgrade are positive (0.286, 0.209, and 0.113), while the Moirai upgrade is negative (−0.143). Matched surrogates, alternative losses, blocked resampling, mean aggregation, and channel-resampling analyses preserve all four signs." What do any of these terms mean? Why do they appear in the abstract with no context whatso ever?
    • "Across 16 upgrade–dataset combinations, SSR is specific to the pair: the same upgrade changes across datasets, and one dataset can contain both signs. On PEMS04 and PEMS08, each positive SSR reaches half its endpoint after phases at periods down to 8–15 hours have returned. Temporal phase structure is therefore a second coordinate of scaling. SSR provides a training-free protocol for measuring where scaling gains arise and comparing upgrades under deployment-relevant temporal structure." I have absolutely no clue what this refers to.

More incomprehensible sentences:

  • "It erases a second variable: the temporal phase structure of the deployment data." What is temporal phase structure? What is deployment data?

  • "Untouched PEMS08 exposes the consequence. Toto 4M→22M, Toto 22M→313M, Chronos tiny→base, and Moirai small→large improve forecasting" This is literatly the second paragraph of the intro, no context whatsoever. No citation. The rest of the introduction is incomprehensible - goes on to discuss some equations; with no context why these are discussed. Goes on to just throw jargon and terms around without defining them; I am unable to read anything

  • From the introduction itself, it is unclear to me (1) what the motivation of the work is, (2) what the authors do in the work, (3) what the results are. Until the end of the intro, there are 0 references to other papers (even Chronos,Toto are not cited ) and it keeps mentionining an untouched PEMS08. I have no idea what this means

  • Figure 2: I have no understanding of what this is - what is "phase amplifies", what is "phase attenuates"

  • All the following sections have the same problem: Sec 2 defines several random variables very ambigiously, with no context of why?

  • 4.1 is the most comprehensible but that is just because it's discussing the datasets and models evaluated, but does not discuss the experimental setting/protocol etc.

  • 4.2 is back to incomprehensible slop.

  • Section 5 onwards: I tried my best but could not undersatnd a full sentence.

  • I checked the references and there are literally 14-15 references; most of them used solely in 1-2 sections of the paper

  • Minor, but the AI use statement is at the end of the appendix and not the main text.

Questions for authors

See weaknesses. I have nothing more to ask/clarify as I did not understand anything.

Reviewer_A8tQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces Structural Scaling Response (SSR), which compares a checkpoint upgrade’s relative forecasting gain on original series with its gain on spectrum-preserving phase-randomized surrogates. Cumulative phase restoration produces response curves intended to characterize the structure associated with upgrade benefits. Experiments examine four checkpoint upgrades across four hourly mobility datasets, with additional surrogate controls and sensitivity analyses. On PEMS08, all upgrades improve original-data accuracy, while the two Toto upgrades and Chronos have positive SSR and Moirai has negative SSR.

Critical strengths
  1. A controlled and informative comparison. Holding the power spectrum fixed while changing phase organization distinguishes ordinary upgrade accuracy from sensitivity to the preserved structure. The PEMS08 result demonstrates that these quantities need not move together.
  2. Substantial robustness analysis. Marginal-preserving, cross-spectrum-preserving, and day-shuffling controls, together with alternative losses and aggregation choices, strengthen confidence that the reported sign pattern is not specific to one implementation choice.
Critical weaknesses
  1. The uncertainty procedure does not clearly target the reported estimator. Point estimates pool channels and seven surrogate realizations, whereas each bootstrap replicate selects one surrogate. For a nonlinear median, these procedures can represent different statistical functionals. Without a coverage justification or matched-estimator reanalysis, the confidence intervals and sign tests remain difficult to assess.
  2. The added contribution over existing surrogate-based forecasting analyses needs sharper substantiation. Prior work, including The Spectrum Is Not Enough, already compares forecasting benefits on original and spectrum-controlled surrogate inputs. The checkpoint-upgrade contrast and restoration curves provide a distinct empirical extension, but the paper needs to establish more clearly what substantial new conclusion follows from that extension.
Questions for authors
  1. Clarification: What estimand and coverage argument justify selecting one surrogate per bootstrap replicate when the reported point estimator pools seven? If the intended target is the pooled estimator, do the conclusions persist under resampling that recomputes that same estimator?
  2. Clarification: What conclusion about checkpoint upgrades becomes identifiable through SSR that is not obtained by applying existing paired surrogate analyses to those checkpoint pairs?
  3. Evidence needed for the shape claims: What uncertainty surrounds the reported half-restoration cutoffs, and how sensitive are they to the restoration grid and isotonic fitting procedure?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.