acceptodds
Under review as a conference paper at ICLR 2027

Behind Zero-Shot Forecasting: How Do Time-Series Foundation Models Use Their Context?

Abstract

Time-series foundation models (TSFMs) leverage large-scale pretraining to achieve cross-domain generalization in forecasting. Among them, the zero-shot paradigm generates predictions using only the observed context, without any task-specific training or fine-tuning. This enables immediate deployment, but leaves a fundamental question open: how do these models use their context to produce forecasts? We investigate this in-context forecasting mechanism directly. Based on an architecture-agnostic framework that characterizes context use through controlled input interventions and their effects on the forecast, we establish three findings. First, TSFM forecasts rule out context parroting, indicating that the models perform inference rather than merely interpolate among values already observed in the context. Second, context influence is governed primarily by temporal position instead of phase-space distance, while similarity-based sample selection can admit a local-approximation error bound no larger than recency-based sample selection. Third, despite its influence decay with temporal distance in the majority of models, distant context can remain functionally important by providing temporal structure beyond distributional statistics. Collectively, these results reveal both the functional strengths and intrinsic limitations of current TSFMs in context utilization, providing a mechanistic basis for diagnosing zero-shot forecasting behavior and informing the design of future models and training strategies.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, foundation models, zero-shot forecasting, in-context learning, mechanistic interpretability, controlled interventions

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_cWVMOpenReview ↗
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
  1. 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.
  2. 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.
  3. 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
  1. 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.
  2. 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.
  3. 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.
  4. 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
AuthorsOpenReview ↗

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=32k=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/MSEmeanO_f=1-\mathrm{MSE}_f/\mathrm{MSE}_{\mathrm{mean}}. It assesses how effectively reference ff predicts the next observation from the context, using the context mean as a baseline (Appendix E.2.1; Table 35):

DataOlinO_{\rm lin}Global linear fit escape (%)
Lorenz, Chen, Chua0.97–1.000.8–4.5
Wind velocity0.2580.0
Sea level0.9911.0
aa index0.2820.0

On the wind velocity and aa index datasets, the lower OfO_f 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.

AuthorsOpenReview ↗

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 C(V)=k ∥V+∥2C(V)=\sqrt{k}\,\|V^+\|_2. Appendix C.1 reports the complete bound ratio: on 600 Lorenz windows with p=1p=1, it is below one in 99.8%, 99.7%, and 99.8% of windows at k=8k=8, 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:

ConditionRelative 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.

AuthorsOpenReview ↗

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

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

The paper studies how frozen time-series foundation models use their input context for zero-shot forecasting. It introduces architecture-agnostic diagnostics based on convex-hull escape, coverage dependence, block-displacement responses, and truncation/far-context replacement, calibrated against reference forecasters. Experiments evaluate 17 checkpoints on scalar Lorenz trajectories. The authors find that forecasts can escape convex historical-successor combinations, most models respond much more strongly to recent context, and distant context can still improve prediction.

Critical strengths

The paper asks concrete questions about context use and calibrates each diagnostic against forecasters with known mechanisms.

The analysis spans several model families.

Critical weaknesses

the findings concern a small collection of controlled synthetic systems without real time series.

The experiment shows that these replacements affect forecasts, but does not fully establish that temporal order specifically caused the difference.

Questions for authors

can this analysis be used for real-world time series? There seems to only be 5 different trajectories studied and all from the same system but just different starting points. that probably limits the generality of this study. only a single trajectory from Chen and a single trajectory from Chua is mentioned.

the “shuffle” experiment changes both order and numerical spread. A clean starting test would reorder the old values while keeping the values themselves unchanged. what is the intuition behind this design?

4 comments
AuthorsOpenReview ↗

Response to Reviewer UyVb

We sincerely thank you for the time and care devoted to our paper and for your clear and constructive questions. Your requests for real-world data and an intervention that holds the values fixed have directly shaped the revision. We have completed both sets of experiments and present the results below.

W1/Q1. Real-world data.

Thank you for raising this important limitation of the submission. 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; Appendices D.2 and E.2 provide the protocols and full evaluation.

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=32k=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. To understand why extrapolation capacity may not appear as hull escape on these series, we further examined their predictive structure and defined reference-dependent observability, Of=1−MSEf/MSEmeanO_f=1-\mathrm{MSE}_f/\mathrm{MSE}_{\mathrm{mean}}. It assesses how effectively reference ff predicts the next observation from the context, using the context mean as a baseline (Appendix E.2.1; Table 35):

DataOlinO_{\rm lin}Global linear fit escape (%)
Lorenz, Chen, Chua0.97–1.000.8–4.5
Wind velocity0.2580.0
Sea level0.9911.0
aa index0.2820.0

The global linear fit escapes on the high-OfO_f series. The lower OfO_f 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. Non-escape therefore cannot establish a lack of that capacity in TSFMs. This explains our initial focus on controlled synthetic systems; the real-world results clarify the calibration rationale while corroborating the context-use findings.

The five Lorenz trajectories and the Chen and Chua checks are now complemented by these three datasets. All forecasts remain univariate and one-step. We are grateful for your suggestion to test the conclusions beyond the controlled systems.

W2/Q2. Isolating the effect of temporal order.

Thank you for suggesting a comparison that preserves the original values. We initially used rescaling to reduce discontinuities at the junction with the retained tail. Following your recommendation, we now evaluate the replacements without rescaling and include exact sample shuffling, block rearrangement, and IAAFT surrogates (Section 5.3; Table 30). Sample shuffling and block rearrangement preserve every value and all permutation-invariant context statistics; IAAFT preserves the values exactly and the spectrum approximately.

On Lorenz at k = 32 and a full-context length of 511, Table 30 reports the following relative changes in median absolute one-step error compared with the full-context baseline:

ConditionRelative 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. The contrast between sample shuffling and block rearrangement shows the importance of the within-block temporal order that the latter preserves, even though both contain exactly the same values. Complete results over all seven retained lengths appear in Appendix E.1.7.

The median boundary steps are also comparable: 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)). The different forecasting effects at comparable junction steps strengthen the evidence for temporal organization beyond the value distribution.

We hope these experiments answer your questions about the breadth of the evidence and the temporal-order interpretation. Thank you again for your careful reading and concrete suggestions; we would be glad to discuss any further questions.

AuthorsOpenReview ↗

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

Thank you for the detailed response, I have raised my score.

AuthorsOpenReview ↗

Response to Reviewer UyVb - Thank You

Thank you for your recognition of our work and rebuttal. Your constructive feedback has helped us strengthen the paper. We sincerely wish you all the best!

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

The paper studies how frozen time-series foundation models (TSFMs) leverage the observed context when making zero-shot forecasts. The authors examined 17 checkpoints from six families of TSFMs in delay coordinates on the Lorenz system, with one-step forecasts and a common context length of 511. They used controlled input interventions including convex-hull escape, coverage versus error, block perturbations, truncation, and distant-context replacement. The authors argue that convex copying cannot fully explain these models, that temporal recency generally matters more than phase-space proximity, and that useful distant context carries temporal structure. A conditional local-approximation result motivates future context selection based on phase-space similarity.

Critical strengths

The question is well motivated and moves beyond leaderboard accuracy to in-context mechanism of TSFMs. The authors find that weak sensitivity does not necessary mean the distant context is not helpful in time series forecasting. In section 5.2, perturbation responses decrease with temporal distance for 15 of the 17 ckpts. However, section 5.3 shows that removing distant context consistently hurts Chronos-T5's one-step accuracy at various retained lengths. It is informative to distinguish sensitivity to individual blocks and the benefits of full context. Comparisons among six TSFM model families on three dynamical systems are comprehensive.

Critical weaknesses
  1. I am not convinced that the far-end replacement test does not cleanly isolate temporal order from the value distribution. Section 5.3 and appendix d.11 explain that the shuffled far-end replacement context is rescaled before passed to the model. But the same rescaling procedure can generate different scale factors. Appendix b also checks pooled boundary and scale ratio, which is helpful, but such summaries cannot ensure that each window retains its distribution. In addition, the final shuffled input changes both order and value of the far end, and the model's native preprocessing is then recomputed on this modified input. It therefore leaves open how much of the response is due to order alone, which in turn affects the interpretation of the claim in Section 7 that preserving values without the temporal order is insufficient requires a control that actually preserves those values.
  2. The scope of the conclusion should be clearer. All the experiments are one-step forecasts of scalar observation from three low-dimensional chaotic systems at a fixed sampling density. This is a useful controlled setting, but it does not mean the same pattern about context-use hold in other time series scenarios.
Questions for authors
  1. Does the conclusion still hold when the original context is compared directly with Algorithm 1's exact-permutation intermediate input, before rescaling? In the same checkpoint setting, can you compare with block-shuffle/short-tail diagnostic of Zhang and Gilpin (ICLR 2025), already discussed in Section 5.3.
  2. TimesFM-1.0 shows a consistently different behavior across all three systems. Can you investigate whether this reflects a different way of using context data or other reasons. It would strengthen the interpretation. Also can you examine TimesFM-2.0 and even the 3.0 version, which is a multivariate TSFM.
  3. How dependent are the findings on the sampling interval and forecast horizon?
  4. Can you define the recovery metric in Figure 4 caption and explain what its negative values mean?
7 comments
AuthorsOpenReview ↗

Response to Reviewer CLFM - Q1 & Q2

We are deeply grateful for your thorough and thoughtful review, and for appreciating our finding that distant context can remain useful despite weak sensitivity to individual blocks. Your questions have helped us strengthen the temporal-order experiment, broaden the evaluation, and align the TimesFM-1.0 evaluation with its officially recommended settings. Thank you for the time and attention you devoted to these questions.

W1/Q1. Isolating the effect of temporal order and comparing with block rearrangement.

Thank you for proposing this direct comparison. We originally used rescaling to reduce discontinuities at the junction with the retained tail. Your suggestion allows us to test the role of temporal order with the values held exactly fixed. Section 5.3 now directly evaluates the unrescaled intermediate context ypermy^{\mathrm{perm}} from Algorithm 1 in the original submission. We also compare with Zhang & Gilpin's (ICLR 2025) block rearrangement on the same checkpoints and prediction origins, retaining the final kk samples and rearranging the preceding context in blocks of length kk.

Sample shuffling and block rearrangement preserve every original value and all permutation-invariant context statistics, including the mean, standard deviation, and mean absolute value. Proposition 4 therefore applies exactly, with each model retaining its native preprocessing. Table 30 reports the following relative changes in median absolute one-step error on Lorenz at L=511L=511 and k=32k=32, compared with the full-context baseline:

ConditionRelative 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. The mild effect of block rearrangement agrees with Zhang & Gilpin; sample shuffling additionally reveals the importance of the within-block order that their intervention preserves. IAAFT provides a further control that preserves the values exactly and the spectrum approximately. Complete results over all seven retained lengths appear in Appendix E.1.7.

We also report the boundary comparison in Appendix B (A6). The normalized median junction steps are 6.96, 7.05, and 6.77 for same-system replacement, block rearrangement, and sample shuffling, pooled over the seven retained lengths and five trajectories. Their forecasting effects differ substantially despite comparable junction steps. Your suggested comparison thus strengthens the temporal-order interpretation while connecting it directly to the earlier result for block rearrangement.

Q2. The TimesFM-1.0 anomaly and the other TimesFM versions.

We are especially grateful for this question, which prompted us to look more closely at the anomaly. In our original shared wrapper, we disabled positional embedding in both TimesFM-1.0 and TimesFM-2.0 to use a common configuration. Rechecking the officially recommended settings revealed that the versions differ: TimesFM-1.0 uses positional embedding by default, whereas TimesFM-2.0 does not. We have enabled positional embedding in TimesFM-1.0 and rerun all its experiments on the synthetic systems and real-world data. All models now use the officially recommended architecture settings for their checkpoints, including the version-specific positional embeddings (Appendix A.3).

On the five Lorenz trajectories, TimesFM-1.0 with positional embedding enabled has an escape rate of 2.94%, a far-end response of 0.307 relative to the uniform reference, and a 376% error increase under sample shuffling at k=32k=32. It no longer shows the anomalous pattern, and we have removed that interpretation. We sincerely thank you for the careful reading that prompted us to recheck the settings and rerun the evaluation.

We also appreciate the suggestion to examine later versions. Actually, TimesFM-2.0, 2.5, and 3.0 were all evaluated in the original submission, and are retained in the revised comparison. Their results are reported in Table 8. The revised Appendix A.3 explicitly distinguishes the configurations of the different versions, including the officially recommended positional-embedding settings for TimesFM-1.0 (enabled) and TimesFM-2.0 (disabled). All three escape the hull on Lorenz (1.12%–1.66%) and have far-end responses below the uniform reference (0.045–0.237; Table 8). We hope the updated evaluation and these comparisons resolve your concern.

AuthorsOpenReview ↗

Response to Reviewer CLFM - Q3 & Q4

W2/Q3. Real-world scope, sampling interval, and forecast horizon.

Thank you for raising this important question about the scope of the findings. 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 the full evaluation in Appendix 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=32k=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. To understand why extrapolation capacity may not appear as hull escape on these series, we further examined their predictive structure and defined reference-dependent observability, Of=1−MSEf/MSEmeanO_f=1-\mathrm{MSE}_f/\mathrm{MSE}_{\mathrm{mean}}. It assesses how effectively reference ff predicts the next observation from the context, using the context mean as a baseline (Appendix E.2.1; Table 35):

DataOlinO_{\rm lin}Global linear fit escape (%)
Lorenz, Chen, Chua0.97–1.000.8–4.5
Wind velocity0.2580.0
Sea level0.9911.0
aa index0.2820.0

The global linear fit escapes on the high-OfO_f series. The lower OfO_f 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. Non-escape therefore cannot establish a lack of that capacity in TSFMs. This explains our initial focus on controlled synthetic systems; the real-world results clarify the calibration rationale while corroborating the context-use findings.

We retain the univariate one-step setting, so these datasets test physical horizons of one second, one hour, and one day. Longer-horizon forecasting concerns the recovery of entire dynamical trajectories, while multivariate forecasting also involves how influence is allocated across variables. These extensions call for additional diagnostics and are left to future work. Thank you for encouraging us to broaden the evaluation and clarify its scope.

Q4. Recovery and negative values.

Thank you for pointing out this omission, and we apologize that the caption did not explain the metric. The Figure 4 caption now defines recovery as

R=Etrunc−EreplEtrunc−Efull,\mathcal{R}=\frac{E_{\mathrm{trunc}}-E_{\mathrm{repl}}}{E_{\mathrm{trunc}}-E_{\mathrm{full}}},

where EE is the median absolute one-step error. Thus R=1\mathcal{R}=1 means that replacement matches the full context, R=0\mathcal{R}=0 matches truncation, and R<0\mathcal{R}<0 means replacement performs worse than truncation. As detailed in Appendix D.11, recovery is computed only when truncation causes a noticeable increase in error (more than 10% of the full-context error), to avoid an unstable ratio when the truncation cost is close to zero. The caption now makes the negative values directly interpretable; the complete error comparisons are in Appendix E.1.7.

AuthorsOpenReview ↗

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

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers and the Area Chair

We sincerely thank all three reviewers and the Area Chair for their time, their careful reading, and their constructive comments. We are grateful for the reviewers' recognition of the research question and the calibrated diagnostic framework, and especially for the concrete suggestions that have helped us improve the experiments and their interpretation. The main concerns raised by all three reviewers center on two questions, both of which we have addressed in the revision:

  1. Isolating the effect of temporal order. Following the reviewers' suggestions, we removed rescaling and added exact sample shuffling, block rearrangement, and IAAFT surrogates. On Lorenz at k=32k=32, shuffling individual samples raises error by 163%–750%, whereas rearranging the same values in blocks changes it by −17% to +22%. This contrast strengthens the evidence for temporal structure beyond the value distribution (Section 5.3).
  2. Extending the analysis to real-world data. 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 equivalent amplitude antipodal (aa) index (geophysics, 1 day). These experiments confirm the main context-use findings beyond the synthetic setting, while reference calibration clarifies why non-escape on some datasets is inconclusive (Section 7; Appendix E.2).

The revised study covers 16 checkpoints from five families; Mamba4Cast was removed from the comparison because its released weights were deleted by the developers and are therefore no longer publicly available for the new experiments. Following the suggestions of Reviewer CLFM, we also reran all TimesFM-1.0 experiments with positional embedding enabled, which complies with the officially recommended settings (Appendix A.3). Code have been updated in the anonymous repository to support reproducibility. We hope the new experiments and explanations address the concerns raised. Thank you again for the care and thought you put into these reviews. We are pleased to welcome any further questions during the discussion.

AuthorsOpenReview ↗

Summary of the Rebuttal and Discussion

Dear Chairs and Reviewers,

As the discussion period draws to a close, we would like to thank all three reviewers for their careful reading and constructive engagement, and the Area Chair for overseeing the review process. All three reviewers responded positively to our rebuttal and revised manuscript. Reviewers CLFM and UyVb each raised their rating from 2 to 3, and Reviewer cWVM confirmed that they would keep their positive score of 3.

These exchanges have made our arguments more rigorous, sharpened our conclusions, and broadened the applicability of our framework. We are sincerely grateful to the reviewers for the questions and suggestions that led to these improvements.

Best regards,
The Authors of Paper 35432

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.