- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper investigates how pretrained time-series foundation models (TSFMs) utilize historical context to generate zero-shot forecasts, focusing on whether their predictions reflect context parroting, phase-space similarity, or temporal recency. The authors develop an architecture-agnostic diagnostic framework based on controlled input interventions, geometric criteria, and local approximation theory. Experiments across multiple TSFM families reveal that the models can extrapolate beyond retrieved historical continuations, predominantly exhibit recency-biased context utilization, and may benefit from distant context through temporal structure rather than distributional statistics alone.
Critical strengths
- Important and well-motivated research question. Rather than proposing another forecasting architecture, the paper investigates the largely unexplored mechanisms underlying zero-shot context utilization in TSFMs. The three research questions are logically connected and provide a clear progression from identifying forecasting behavior to understanding context weighting and the functional role of distant observations.
- Principled and architecture-agnostic analysis. The proposed framework combines geometric criteria, local approximation theory, and controlled input interventions. Importantly, the diagnostic measures are calibrated using reference forecasters with known mechanisms, which strengthens the interpretability and credibility of the empirical findings.
- Interesting insights into TSFM limitations. The finding that many TSFMs exhibit stronger dependence on temporal recency than phase-space similarity is particularly valuable. The additional observation that distant context can remain useful despite weak local sensitivity provides a nuanced understanding of context utilization and suggests directions for improving future TSFM architectures.
Critical weaknesses
- Limited generalizability beyond controlled dynamical systems. The empirical analysis primarily relies on low-dimensional chaotic systems, such as Lorenz, Chua, and Chen, under a fixed context length and one-step forecasting setting. While these systems facilitate controlled mechanistic analysis, they differ substantially from real-world time series characterized by stochasticity, non-stationarity, external covariates, and heterogeneous temporal patterns. It remains unclear whether the identified recency bias and phase-space utilization patterns hold in practical forecasting scenarios or across different prediction horizons.
- The diagnostic criteria do not uniquely identify the underlying mechanisms. Escaping the convex hull of retrieved continuations rules out the specific convex-combination form of context parroting, but does not necessarily demonstrate sophisticated dynamical inference. Similarly, weak dependence on phase-space coverage or strong sensitivity to recent observations does not conclusively establish that models primarily rely on temporal position, as other factors such as normalization, architectural biases, and pretrained representations may produce similar behaviors. The paper would benefit from a clearer distinction between the mechanisms its criteria can rule out and those they can positively identify.
- The theoretical results have limited implications for practical TSFM forecasting. The comparison between phase-space-based and recency-based selection concerns local approximation error bounds under specific smoothness and conditioning assumptions. However, these guarantees do not directly apply to the nonlinear inference mechanisms implemented by pretrained TSFMs, nor do they establish that incorporating phase-space retrieval would improve their predictions. Therefore, the connection between the theoretical findings and practical model improvements remains incomplete.
- The role of distant temporal structure is not fully isolated. The far-context replacement experiments suggest that temporal organization matters beyond marginal distributional statistics. However, shuffled or noise-based replacements may also introduce unnatural discontinuities, alter local autocorrelation, or induce distribution shifts relative to the model's pretraining data. More carefully controlled alternatives, such as block shuffling or phase-randomized surrogates, would help distinguish genuine temporal-structure utilization from sensitivity to out-of-distribution perturbations.
Questions for authors
See weakness.
5 comments
Response to Reviewer cWVM - Q3 & Q4
W3/Q3. Connecting the theorem to forecasting improvements.
Thank you for raising this point; we agree that the connection between approximation bounds and forecasting gains deserves a clearer explanation. We have clarified the normalized design matrix underlying the stability factor (Lemma 1; Appendix C.1) and explain the connection to the forecasting comparison in Section 5.2 and Appendix D.8 (Table 7).
The bound depends on both the neighborhood radius and the stability factor . Appendix C.1 reports the complete bound ratio: on 600 Lorenz windows with , it is below one in 99.8%, 99.7%, and 99.8% of windows at , 16, and 32.
Across five Lorenz trajectories, the phase-space local-linear reference has mean normalized error 0.0509, versus 0.1050 for the best trailing polynomial in a grid varying length and order: a 51.5% reduction. Although these results do not constitute a performance guarantee for TSFMs, they motivate future models that exploit phase-space similarity alongside temporal recency. As discussed in Section 8, a natural direction for future work is to test whether explicitly retrieving, selecting, or upweighting phase-space-similar samples translates this advantage into lower forecasting error.
W4/Q4. Exact permutations, block rearrangement, and spectral surrogates.
We are grateful for this concrete suggestion. We have added both block rearrangement and spectral surrogates, and now evaluate all far-context replacements without rescaling (Section 5.3). Implementation details and boundary comparisons are provided in Appendices D.11 and B (A6), respectively.
On Lorenz at L = 511 and k = 32, Table 30 reports the following relative changes in median absolute one-step error compared with the full-context baseline:
| Condition | Relative change in one-step error |
|---|---|
| Truncation | −18% to +74% |
| Same-system passage | −6% to +23% |
| Block rearrangement | −17% to +22% |
| IAAFT | −24% to +80% |
| Sample shuffling (exact permutation) | +163% to +750% |
| Gaussian replacement | +176% to +788% |
These are ranges across the 16 checkpoints after averaging each checkpoint's relative change over five trajectories. Sample shuffling is worse than the full context, same-system passage, and block rearrangement in all 80 checkpoint–trajectory pairs. IAAFT, which preserves the original values exactly and the spectrum approximately (median amplitude-spectrum residual 1.1%), also produces much smaller error increases than sample shuffling.
The median boundary steps are 6.96, 7.05, and 6.77 times a typical original step for same-system replacement, block rearrangement, and sample shuffling, respectively, pooled over the seven retained lengths and five trajectories (Appendix B (A6)). Sample shuffling and block rearrangement contain exactly the same values; their substantially different forecasting effects at comparable junction steps strengthen the evidence for the importance of within-block temporal order beyond the value distribution.
Thank you again for a thoughtful review that has directly improved the paper. We hope the new experiments and explanations address your concerns, and would be happy to discuss any further questions.
A Gentle Follow-up on Our Rebuttal
Dear Reviewer,
We hope this message finds you well. As the discussion period is nearing its end with less than three days remaining, we would greatly appreciate the opportunity to hear your thoughts on our rebuttal and revised manuscript. Should any questions remain, please let us know; we would be very happy to provide further clarification. If our responses have addressed your concerns, we would be grateful if you could update your assessment accordingly.
We understand how demanding this period can be and sincerely appreciate the time and care you have devoted to reviewing our paper. Thank you again for your constructive feedback, which has helped us strengthen the work, and for your continued consideration.
Best regards,
The Authors
Response to Reviewer cWVM - Q1 & Q2
We are sincerely grateful for your insightful review and positive assessment of our work. We especially appreciate your assessment of our research question as important and well motivated, and your recognition of the significance of our work. Your suggestions have guided the new experiments and helped us explain their interpretation more clearly.
W1/Q1. Scope: simulated systems, sampling density, and noise.
Thank you for encouraging us to test the findings beyond the controlled setting. We added three real-world time-series datasets spanning different domains, sampling intervals, and inherent noise levels: vertical wind velocity (atmospheric turbulence, 1 second), sea-level height (oceanography, 1 hour), and the geomagnetic aa index (geophysics, 1 day). Section 7 and Table 2 summarize the results, with protocols and full results in Appendices D.2 and E.2.
The results confirm the main context-use findings: all 16 checkpoints show stronger near-end than far-end influence on all three datasets, and the sea-level dataset reproduces hull escape (0.3%–2.0%) and increased error under sample shuffling (115%–580% at k=32). On the wind velocity dataset and most windows of the aa index dataset, however, forecasts remain within the hull, while the one-sided criterion is inconclusive as noted in Appendices B (A2) and C.2 of the original manuscript. The reference calibration in W2 below explains how to interpret these results. These additions extend the evidence across domains and physical time scales within the univariate one-step setting.
W2/Q2. What the criteria establish about mechanism.
Thank you for highlighting the need to distinguish more clearly what the criteria exclude and what they establish about mechanism. We have made the one-sided interpretation, already stated in Appendices B (A2) and C.2 of the original manuscript, more explicit in Section 7 and expanded the reference calibration in Appendix E.2.1 (Table 35). The paper's distinction is concrete: parroting produces a convex combination of retrieved continuations; inference extends beyond that class. A hull escape excludes every such convex combination on that window. This establishes an extrapolation capability under the framework, without equating it with accurate dynamical reconstruction. A forecast inside the hull can arise from either mechanism.
To explain why extrapolation capacity may not appear as hull escape, we further examined the real-world series and defined reference-dependent observability, Of=1−MSEf/MSEmean. It assesses how effectively reference f predicts the next observation from the context, using the context mean as a baseline (Appendix E.2.1; Table 35):
On the wind velocity and aa index datasets, the lower Of values indicate weaker accessible predictive structure, i.e., harder to obtain accurate one-step predictions based on the context. The global linear fit retains its capacity to extrapolate, but that capacity does not translate into observable convex-hull escape on these series. Conversely, the phase-space local-linear reference escapes on 5.6% of windows in the aa index dataset while predicting worse than the context mean. These controls distinguish extrapolation capacity, observed escape, and forecasting skill.
To address the normalization explanation, we added a block-swap probe that preserves every context value and permutation-invariant normalization statistic (Appendix E.1.6; Table 24). On Lorenz, for example, the median normalized response across checkpoints is 0.344 at offset 0 and 0.020 in the far field; every checkpoint's far-field response is nonzero, whereas the uniform average's is zero. Changes in these statistics therefore cannot explain the order sensitivity or the strong near-versus-far contrast.
The measured allocation pattern may arise from architectural biases. The framework identifies context dependence through interventions and calibrated alternatives, without requiring a unique internal implementation; Chronos-T5 activation patching adds a complementary structural intervention (Appendix F). We thank you for encouraging us to make this distinction clearer.