acceptodds
Under review as a conference paper at ICLR 2027

SCOPE:STRUCTURE- AND IMPORTANCE-AWARE PROJECTION FOR TIME-SERIES IMPUTATION

Abstract

Contiguous gaps corrupt the historical inputs of time-series foundation models (TSFMs), yet lower reconstruction error at missing points does not necessarily yield better forecasts. We present SCOPE, a structure- and importance-aware projection method that requires no dedicated imputation-model training. SCOPE estimates six structural targets through Gaussian-process posterior sampling and observation-based statistics, and distinguishes downstream-use importance from structural adjustment leverage. It first reconstructs trend, seasonal, and residual components, then selectively applies a constrained structural correction that preserves observations and limits adjustment magnitude. When the target model is accessible, an optional model-aware extension provides further alignment with a fallback to the model-unaware result. Across 27 public datasets, four TSFMs, and nine block-missingness configurations, model-unaware SCOPE achieves the best combined average rank against deep baselines under two training budgets and against classical reference methods. It reduces average downstream MAE by 6.97% relative to linear interpolation and 10.43% relative to STL-Kalman. Structural-fidelity and per-unit-adjustment analyses further support the complementary roles of component-wise initialization and constrained projection.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series foundation models, imputation, structural preservation, importance- weighted projection

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_Cqb8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose the SCOPE: model-agnostic time-series imputation for fixing incomplete historical inputs of frozen time-series foundation models (TSFM) as the main motivation. The key argument here is that pointwise accuracy may not ensure preservation of key temporal structures for downstream forecasting.

The proposed SCOPE consists of component-wise numerical initialization and structure-aware refinement. The component M3 will provide information for trend seasonality residual. Then M1 provides estimation of 6 structural targets based on the observed data. M2 will estimate positional importance structural leverage. M4 will do a bounded optimization to refine the initialized values. An unnecessary modification can be prevented by a gate that is based on calibration, with a fallback mechanism. Also, an optional M5 is to provide model-aware alignment.

The experiments are done with 27 datasets. For main TSFMs, 4. Also 9 combinations of missing rate and contiguous block length are tested. As a result, SCOPE shows a good average rank to classical/deep imputation counterparts. They also show a 6.97% decrease in downstream MAE from linear interpolation.

Ablation studies indicate that the component-wise initialization brings the most contribution to the improvement for forecasting. For structural projection, average 6-dim structural error is decreased by 5.49%. Also, a paired study is conducted for block length 50. With a prespecified 0.5% MAE range, average forecasting non-inferiority is shown.

Critical strengths
  1. Large empirical experiments across forecasting models and missingness settings.

27 datasets from different applications, 4 frozen TSFMs, 9 missingness settings.

SCOPE vs. both classical imputation and various deep baselines, with 2 budget levels.

Also direct comparison against native/default missing handling. Shows strong wins for 3 TSFMs, with the caveat of stronger native performance for Chronos-2.

Results by datasets. Module ablations. Structural faithfulness analysis. Sensitivity.

All this demonstrates that the numerical reconstruction has a practical role.

  1. The clear disentanglement of numerical recovery and structural refinement.

The paper very clearly separates the component-wise initialization and structural projection.

The authors clearly admits that the main forecasting improvement comes from the M3. The M1, M2, M4 add much lesser downstream improvement.

For structural refinement experiment, the end sequences are compared to clean historical references that are not used for optimization. Moreover, structural projection is constrained to be observation consistent and with bounded change with a fallback to original initialization.

Critically, the paper conducts formal non-inferiority analysis instead of taking a small forecasting difference (close to zero) as a statistically proven improvement.

This design and reporting helps to enhance the credibility.

Critical weaknesses
  1. The separate methodological and practical contribution of structural projection is not clear enough.

The key novelty that the paper proposes is an importance-aware structural projection. Which utilizes structural targets from observations, positional importance, and bounded optimization.

But the results show that most of the downstream forecasting improvement comes from component-wise initialization.

Dropping M3 alone increases the average forecasting MAE by 7.24%. While the drop in M1/M2/M4 causes much less change. More importantly, for direct pair-wise comparison between M3 and complete projected approach, the average MAE improvement is only 0.0198%. And 90% CI includes zero.

While the authors correctly point out that the result shows average forecasting non-inferiority/equivalence for their specified margin. But it does beg the question on the practical merit of introducing all these extra estimation/optimization components.

Even the stated 5.49% improvement in terms of structural fidelity is not trivial. But it's measured for the same structural dimensions which are used for the optimization. But compared to independently held-out clean references.

Further assessment (e.g. with separate metrics of temporal structure or downstream task-based evaluation) would help make the case that these benefits reflect an interesting behavior distinct from the proxy-optimized representation.

Interestingly, also residual autocorrelation is somewhat worse (and 19 out of 97 activated inputs have lower clean-ref structure fidelity) suggesting some non-uniform reliability of the refinement.

Last, to better establish the need of downstream-aware imputation, it'd be helpful to more explicitly contrast to existing task-aware imputation approaches (e.g. Wang et al. NeurIPS 2024).

  1. Still quite unknown are the practical efficiency, baseline fairness and robustness of the refinement.

The compute analysis omits the precomputed SCOPE gating calibration, but includes per-input compute budget for the deep baselines. With Gaussian process fitting, sensitivity estimation and iterative optimization included, the reported runtimes don't give us a complete picture of the deployment cost of your proposed framework.

Further, the baseline comparison to deep imputation models depends on training/tuning choices. Two budget for training does give us a nice piece of evidence; but is insufficient to support that these baselines have achieved representative performance for the intended deployment environment. More explanation on model selection. Maybe other well-known baselines like SAITS?

The forecast non-inferiority for structural projection is shown for block 50 only, for fixed imputations and gates. The range of equivalence for the subset of activated inputs is close to the pre-defined margin. The supplemental experiments reveal non-negligible sensitivity wrt imputation seeds.

So this does not yet show robust forecast preservation for the larger missing blocks used in the main benchmark.

And focusing on artificially created sequential holes in the retained historical data. To make this more useful, it should also include missing block assessments in the vicinity of the forecast origination point, especially endpoint missingness (unusable future observations for interpolation).

Questions for authors

Q1. Independence of structural projection.

Are there any M3 vs SCOPE comparisons that you can make for all 9 missingness scenarios (e.g., multiple indep. imputation seeds)?

Are you able to show improvements to independent structural characteristics (e.g., season phase, trend turning points, other downstream tasks)? In addition to the 6 optimized structural parameters by M4.

What makes you sure that this structural improvement is practically useful, given that downstream forecasting performance is statistically identical as the M3 initialization?

This is crucial to understanding the methodological impact of your paper.

Q2. Verifying importance of position/leveraging structure.

How well do your model independent importance of position scores agree with the sensitivity of forecasting for the considered TSFMs?

Could you evaluate the proposed importance distributions vs straightforward recency weights, random position weights, uniform reg. under otherwise the same projection goals?

With such small M2 impact on forecasting it is critical to see if the proposed importance / leverage ration has a consistent benefit over simpler approaches.

Q3. Complete computational load and base training settings.

Could you provide the full computational cost for GP fitting, two block gate calibration, structural sensitivity analysis, projection optimization (end-to-end)?

For the deep imputation baselines, how do you choose the training budget, hyper-parameters?

Is their original training preserved? How would this change with more established imputation techniques such as SAITS?

If the calibration outputs are reusable over multiple inference calls, then amortized cost analysis would also be useful.

Q4. Robustness of the method in challenging missingness patterns:

Is the forecast non-inferiority property of structural projection still holding when the block sizes get longer (e.g., 100 and 150) and the seeds of imputation change?

How does SCOPE fare if the entire missing block is contiguous just before the origin of the forecast (i.e. we don't have any right-side observations to interpolate?)

This is key to understanding how useful the method can be in sensor outage scenarios or other patterns relevant to deployment.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Cqb8(Part 1 of 4)

Overall response. Thank you for the detailed questions about the projection and its use in practice. We now report all nine missingness settings, with separate seed checks and measurements outside the projection objective. The clearest improvement is in recovery of hidden historical values. At the 0.5% margin, only 50-step internal gaps meet average forecast non-inferiority in the three-origin aggregates. We also test simpler position weights, report the full cost with and without calibration reuse, and compare SAITS at both epoch caps on the same inputs.

To Q1 (W1) — nine settings and independent measures. The nine rate-length settings use 2,916 forecast pairs from 27 datasets and four frozen backbones at three origins. The pooled gain over M3 is -0.22% (95% CI [-0.70%, +0.14%]). At the middle origin each setting has 108 matched forecasts per GP seed (0, 1, 2), with masks and gates held fixed. All 27 seed estimates and intervals are reported (Appendix G.2, Tables 36-38).

The original six-dimensional structural error fell 5.49%, with 19 of 97 activated inputs worsening and a slight deterioration in residual autocorrelation (Section 4.4, Table 3). Independent tests show lower masked-point MAE for internal and terminal gaps (reductions of 0.0056 and 0.0128 observed standard deviations, Holm p=0.0118p=0.0118 and p=0.0323p=0.0323). Internal phase and turning-point intervals span zero, and NAB detection shows no consistent improvement (Appendix G.3, Tables 41-42). This supports a use when recovered historical values matter. Section 2 distinguishes the pseudo-gap calibration signal from the downstream gain estimates of Wang et al. (2024b).

AuthorsOpenReview ↗

Response to Reviewer Cqb8(Part 2 of 4)

To Q2 — positional weighting. At 20% missingness and length 100 we perturbed 12 missing positions by 0.10.1 observed standard deviations in each direction and measured forecast sensitivity by central differences. Median within-series Spearman correlations range from -0.08 to +0.23 for the use-importance score and +0.26 to +0.51 for recency. The composite cost is often constant (four defined correlations per model). In 972 forced-projection pairs with the target, initialization, gap basis and solver fixed, the original weighting yields gains of -0.53%, -0.38% and -0.40% over recency, shuffled recency and uniform weighting. All 95% intervals span zero. This comparison tests the combined weights, and a separate leverage ablation is still needed (Appendix G.5, Tables 44-45).

For leverage alone, the median correlations with forecast sensitivity are +0.01 to +0.15 across the four backbones. Recency has a higher median correlation than the use-importance score for each backbone (Table 44).

AuthorsOpenReview ↗

Response to Reviewer Cqb8(Part 3 of 4)

To Q3 (W2a-W2b) — full costs and SAITS. Across seven settings, SCOPE's end-to-end median is 1.163-3.142 s, including calibration, completion and forecasting. For 1,536-step histories and 50-step gaps, calibration medians for GP fitting, sampling, M2 and M4 are 0.766, 0.025, 0.069 and 0.558 s. The median paired overhead over M3 is 2.773 s. With reusable calibration, per-call medians are 1.198 s over 10 calls and 1.084 s over 100 (Appendix G.6, Tables 46-47).

The original deep baselines use fixed univariate PyPOTS adapters, with 96-step windows, stride 48 and batch size up to 32. These are benchmark configurations. The larger budget triples each epoch cap with architecture and other settings fixed (Table 48). The exploratory Sundial-SAITS comparison uses the same 126 pairs at both caps, across 42 datasets at the middle origin and 20% missingness (Appendix G.7, Table 49). Both confidence intervals span zero.

Table C1. SCOPE forecast gain over SAITS on Sundial (revision Table 49).

Epoch capGain [95% CI]Wins / losses
10+2.76% [-2.93%, +8.66%]72 / 54
30+2.49% [-2.65%, +7.86%]72 / 54
AuthorsOpenReview ↗

Response to Reviewer Cqb8(Part 4 of 4)

To Q4 (W2c-W2d) — long and terminal gaps. At the 0.5% margin, only the 50-step internal gap meets average non-inferiority in the three-origin aggregates (90% CI [-0.40%, +0.25%]). At that length, all three middle-origin GP seeds meet it when pooled across rates. Internal gaps of 100 and 150 steps have lower endpoints of -0.69% and -0.86%. For terminal gaps, the imputer sees only the left history. Lower endpoints at lengths 50, 100 and 150 are -1.90%, -2.24% and -0.54%, respectively, so none of the terminal aggregates meets the margin. Historical terminal-point reconstruction still improves (Section 4.6, Table 4, Appendix G.2-G.3, Tables 37-39 and 41).

Reviewer_VqbFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SCOPE, a time-series imputation method designed for downstream forecasting. It first produces an initial imputation, estimates several structural properties and data importance scores, and then projects the imputed values toward these targets. The method is evaluated on 27 datasets, four pretrained forecasting models, and nine missing-data settings.

Critical strengths
  1. The experimental coverage is broad. The paper includes multiple forecasting models, datasets, missing ratios, and block lengths.

  2. The authors evaluate structural preservation in addition to standard imputation and forecasting errors. The paper also contains useful ablations and confidence intervals.

Critical weaknesses
  1. The submission shows a lack of care and respect for the review process. Text overlaps in the main figure, and the supplementary material is a disorganized collection of duplicated files, outdated results, and hard-coded paths. It is unnecessarily difficult for reviewers to identify and verify the final experiments.

  2. The reported numbers do not support the main claim. The complete projection improves MAE over M3 by only 0.0198%, without statistical significance. Most of the improvement comes from M3 rather than the proposed projection.

  3. The evaluation protocol is not convincing. The final code uses a benchmark-specific gate, and the reported mean improvement is 6.97%, while the median is only 0.29%.

Questions for authors

Can the authors evaluate the complete SCOPE method against M3 alone on entirely new datasets that were not used to construct the gate or select its threshold? A clear and statistically significant improvement in this setting would materially affect my assessment.

4 comments
AuthorsOpenReview ↗

Response to Reviewer VqbF(Part 1 of 2)

Overall response. Thank you for asking us to test the gate on independent data and clarify how the evidence can be checked. In the revised evaluation, we set aside each forecast target before masking, calibration and imputation. We keep the 0.25% threshold fixed on 23 independent datasets. Under the primary metric, SCOPE has a higher mean forecast MAE than M3. On those same history inputs, masked-point reconstruction MAE is 0.93% lower. Section 4.5 reports the two outcomes separately. Figure 2 and Algorithm 1 show how to follow the method, and Table 50 points to the evidence for each review item.

To Q1 (W2-W3) — independent fixed-gate test. We compare SCOPE and M3 on 23 datasets excluded from both gate development and threshold selection, using the original 0.25% threshold and two pseudo-gaps per visible prefix. The two methods share the mask, origin, forecast target and frozen backbone settings. Across 7,344 paired forecasts the mean gain relative to M3 is -0.71% (95% CI [-1.96%, -0.04%], two-sided source-family sign-flip p=0.0275p=0.0275), indicating higher MAE for SCOPE under the primary metric. With the symmetric denominator the estimate is -0.19% (95% CI [-0.64%, +0.07%]). The latest three-dataset batch's 150-step subgroup gives +0.47% (95% CI [-0.17%, +1.42%]) over 324 pairs, an exploratory result (Section 4.5, Appendix G.1, Table 34).

Across the same 1,836 histories, standardized MAE at deliberately hidden points falls from 0.4984 to 0.4937 (0.93% reduction). The paired decrease is 0.0046 visible-history standard deviations (95% CI [0.0026, 0.0069], p=0.0002p=0.0002). The revision gives separate conclusions for future prediction and recovering missing historical values.

Table V1. Pooled external evaluation (Section 4.5 and Table 34).

OutcomePairs or inputsEffect [95% CI]
Forecast gain, M3 denominator7,344-0.71% [-1.96%, -0.04%]
Forecast gain, symmetric denominator7,344-0.19% [-0.64%, +0.07%]
Reconstruction MAE decrease (SD units)1,8360.0046 [0.0026, 0.0069]

Comparison and time boundary. The original 6.97% mean and 0.29% median compare the full pipeline with linear interpolation. The 972 configurations contain 548 wins, 180 losses and 244 ties. In the matched length-50 comparison with M3, the gain is 0.0198% and the 90% interval spans zero. The revision puts M3 at the center of numerical recovery. The historical pipeline imputed the full masked series before the rolling split, so visible entries after some origins could be used. The new evaluation withholds each target before masking, calibration and imputation. Changes to origin sampling, context length and masks also contribute to differences between the two protocols (Sections 4.1-4.2, Appendix G).

AuthorsOpenReview ↗

Response to Reviewer VqbF(Part 2 of 2)

To W1 — figure and evidence. Figure 2 gives the workflow. Algorithm 1 shows where the gate is evaluated and when the method keeps the M3 fill. The review labels in the margin and the index in Appendix G.8 (Table 50) point to the corresponding passages and results. For example, the index leads directly to Section 4.5 and Table 34 for the independent validation.

The evidence archive identifies the saved records by analysis and uses relative paths. From the extracted directory, python verify.py recomputes the main aggregate estimates from the saved scores and checks the 27 middle-origin seed estimates and their 90% intervals. The README lists the required packages and maps each analysis to its records.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer VqbF

Thank you again for your detailed comments and suggestions. In our response and revised manuscript, we have evaluated the unchanged gate on 23 datasets excluded from gate development and threshold selection, clarified the evaluation protocol, and reported historical reconstruction and future forecasting outcomes separately. We have also organized the supporting evidence to make the results easier to verify. These revisions clarify both the contribution and the limitations of the proposed approach. Please let us know if any explanations or results would benefit from further clarification. We would be happy to address your remaining concerns and look forward to your feedback.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer VqbF

To address your concern about the mean and median, Section 4.5 (p. 9) now reports both for the independent SCOPE–M3 comparison. With the gate threshold fixed at 0.25%, the mean relative reduction in forecast MAE was −0.71% across 22 equally weighted source families (95% bootstrap CI, −1.96% to −0.04%). The median family estimate was −0.03%. Ten families favored SCOPE, eleven favored M3, and one showed no difference.

Solar 10 Minutes had the largest loss (−13.24%). In the sensitivity analysis below Table 34 (Appendix G.1, p. 38), omitting this source moved the mean to −0.12% across the other 21 families. The magnitude of the average loss is therefore sensitive to this source, while the point estimate remains negative when it is omitted. The primary estimate includes all 22 families. We’d welcome your view on whether this clarifies how representative the pooled result is across new sources.

Reviewer_Rc4vOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCOPE, a training-free time-series imputation framework designed to improve downstream forecasting with pretrained time-series foundation models. The method combines component-wise statistical reconstruction based on trend, seasonality, and residual decomposition (M3) with additional structure-aware refinement modules (M1, M2, and M4). The authors evaluate SCOPE across 27 datasets and four time-series foundation models, reporting improvements over both classical and deep learning-based imputation methods.

While the empirical results are interesting, I believe the paper is not appropriately positioned, and the main claimed methodological contributions are insufficiently supported by the experiments.

Critical strengths
  • The paper provides extensive evaluations across 27 datasets, four time-series foundation models, and multiple missingness settings.
  • The finding that classical statistical decomposition can outperform several deep learning-based imputation methods in downstream forecasting is interesting and potentially valuable.
  • The method does not require training a dedicated imputation model, making it relatively straightforward to apply to existing forecasting pipelines.
Critical weaknesses
  1. Inappropriate positioning of the main contribution

I believe the most interesting finding of this paper is that classical statistical imputation methods can outperform several deep learning-based approaches in downstream time-series forecasting. In particular, the component-wise initialization (M3), based on established STL/MSTL decomposition and interpolation techniques, appears to account for the vast majority of the reported forecasting improvements.

However, the paper is primarily positioned around a complex structure- and importance-aware projection framework, rather than systematically investigating why these relatively simple statistical techniques outperform learned imputers.

I think the paper would be substantially more compelling if it focused on this empirical observation and provided a rigorous analysis of when and why statistical imputation is preferable to deep learning-based approaches. For example, a detailed comparison of standalone M3 with carefully tuned classical decomposition methods, along with sensitivity analyses across dataset characteristics and missingness patterns, would provide more meaningful insights.

  1. Limited contribution of the proposed structural projection

The ablation results suggest that almost all forecasting improvements come from M3, while the additional modules (M1, M2, and M4) contribute very little to downstream prediction accuracy.

In particular, the full SCOPE framework improves forecasting MAE by only 0.0198% over M3 in the paired evaluation, without a statistically significant improvement. Although the proposed projection reduces structural error by 5.49%, this does not translate into a meaningful improvement in forecasting performance.

This raises questions about whether the additional complexity of Gaussian process-based structural estimation, positional importance weighting, and constrained optimization is justified. The paper would benefit from a clearer demonstration of the practical advantages of these modules beyond improving a small set of handcrafted structural metrics.

Overall, the experiments appear to support the effectiveness of classical component-wise imputation much more strongly than the necessity of the proposed structural projection framework.

(I am giving score 1 because the paper is not positioned well and not focusing on the core finding in my opinion, not because the paper has significant flaws)

Questions for authors
  1. Given that M3 accounts for most of the forecasting improvements, why is the paper primarily positioned around structure-aware projection rather than the effectiveness of classical statistical imputation?

  2. Can the authors provide a more rigorous comparison between standalone M3 and carefully tuned STL/MSTL-based or other classical statistical imputation methods? Under what dataset characteristics and missingness patterns does M3 perform particularly well?

  3. What is the practical benefit of M1, M2, and M4 when their combined forecasting improvement over M3 is negligible? Are there downstream tasks where preserving the proposed structural metrics yields more substantial benefits?

  4. How much additional computational overhead does the full structural projection introduce compared with M3 alone, including calibration and Gaussian process fitting?

6 comments
AuthorsOpenReview ↗

Response to Reviewer Rc4v(Part 1 of 3)

Overall response. Thank you for asking what projection adds once M3 has filled the gaps. The abstract and discussion now present M3 as the main recovery step and projection as optional refinement. The revision adds comparisons with tuned STL and MSTL, and uses held-out historical values to measure reconstruction directly. The clearest gain from projection is in recovering those missing values. In the matched forecast comparison at gap length 50, the gain over M3 is close to zero. We also report the full timing, including gate calibration and GP fitting.

To Q1 and Q3 (W1-W2) — contribution and practical use. M3 provides the main numerical recovery. It reconstructs trend, seasonality and residuals, using period detection and STL/MSTL or a trend-residual fallback. In the original ablation, removing M3 raised forecast MAE by 7.24%. The matched length-50 projection gain was 0.0198% (90% CI [-0.1578%, +0.1787%]). The abstract, introduction, Section 4.4 and conclusion now describe refinement as optional.

When completed histories are used downstream, the new held-out historical scores provide a direct measure of recovery. Projection lowers standardized reconstruction MAE by 0.0056 for internal gaps (95% CI [0.0023, 0.0092], Holm p=0.0118p=0.0118) and 0.0128 for terminal gaps ([0.0042, 0.0228], Holm p=0.0323p=0.0323). Units are visible-history standard deviations. Internal phase and turning-point intervals span zero. For terminal gaps, turning-point F1 changes by -0.0111 (95% CI [-0.0232, -0.0012], Holm p=0.2168p=0.2168). The offline NAB detector shows no consistent improvement. The measured benefit is therefore historical point recovery, with further downstream utility to be tested (Appendix G.3, Tables 41-42).

AuthorsOpenReview ↗

Response to Reviewer Rc4v(Part 2 of 3)

To Q2 (W1) — tuned classical controls. We paired M3 with each of five classical baselines on 4,536 forecasts across 42 datasets. STL and MSTL are selected by recovery MAE on two pseudo-gaps in the visible history. All methods share masks, origins and frozen forecasters. Against linear interpolation, the gains at lengths 50 and 150 are +0.58% and -0.58%, and both intervals include zero. Relative to tuned MSTL, the gains in the low- and high-seasonality strata are -1.76% and -2.39%. For length 150, the exploratory MSTL contrast is -2.46% (95% CI [-4.12%, -0.96%], Holm p=0.0300p=0.0300 within its eight strata). These comparisons make tuned MSTL a strong reference for long gaps (Appendix G.4, Table 43).

Table R1. Pooled forecast MAE gain of M3 (revision Table 43).

ComparatorGain [95% CI]Holm p
Linear-0.03% [-1.05%, +0.98%]0.9558
Tuned STL-1.24% [-2.73%, +0.17%]0.4187
Tuned MSTL-1.81% [-3.58%, -0.16%]0.2435
AuthorsOpenReview ↗

Response to Reviewer Rc4v(Part 3 of 3)

To Q4 — total overhead. We measured 112 serial runs on eight datasets, including gate calibration, completion and one Chronos-2 forecast. At a 1,536-step history and 50-step gap, median total time is 0.024 s for M3 and 2.797 s for SCOPE. The median paired overhead is 2.773 s. Median gate time is 1.950 s. GP fitting takes a median 0.766 s within calibration and 1.383 s within actual projection when activated. Total latency at the 95th percentile is 18.543 s. Timing starts with the model loaded and the input in memory (Appendix G.6, Tables 46-47).

If the same calibration can be reused, the estimated median per-call time is 1.198 s over 10 calls and 1.084 s over 100 calls. The estimate spreads calibration across the calls and still counts completion and forecasting on every call (Appendix G.6, Table 46).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.