acceptodds
Under review as a conference paper at ICLR 2027

RestoreTS: Explicit Input Restoration for Robust Time-Series Forecasting

Abstract

Time series forecasting supports scientific analysis and operational planning across systems that rely on sensors. Yet sensor errors can introduce spikes and sustained level shifts into observed histories, degrading downstream forecasts. Existing approaches reduce corruption effects by changing forecasting architectures or training procedures. However, forecasting robustness alone does not provide an explicitly restored history whose corrections can be inspected. To address this need, we introduce RestoreTS, an input restoration layer that corrects suspicious intervals before an unchanged forecasting backbone. It combines Bayesian online change point detection with temporal return criteria and exposes the affected intervals and applied corrections for inspection. Across eight datasets, five forecasting backbones, four horizons, and five levels of controlled combined corruption, restoration reduces MSE in 39 of the 40 dataset–backbone pairs, while MAE decreases in all 40. Mean reductions across the five backbones range from 6.8% to 19.3% for MSE and from 4.1% to 7.8% for MAE. These results suggest that input restoration can improve corruption robustness across the tested forecasting architectures. Code is available at: https://anonymous.4open.science/r/restorets-review-B9BB/.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series forecasting, robustness to input corruption, time-series input restoration, Bayesian online change-point detection

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 2 reviews, above 27% of the others.
Reviewer_ZfSUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RestoreTS corrects potentially erroneous segments of historical data before forecasting and records the specific changes made. It is evaluated on eight datasets and five forecasting models. The experiments first introduce transient spikes and sustained value shifts, then compare forecasts before and after repair. Overall, forecasts become more accurate when using the repaired data.

Critical strengths
  1. Comparing forecasts from the same model before and after repair provides a relatively direct assessment of the repair’s impact. Improvements across multiple models also support the method’s applicability.
  2. The paper provides modification logs, input-quality analyses, and repeated experiments, and reports cases where performance deteriorates. These help clarify the method’s benefits and limitations.
Critical weaknesses
  1. Parameter selection relies on clean reference data and known corruption severity, and different settings may be used for clean and corrupted inputs. In addition, test data were examined during development. Performance under unknown corruption conditions and on new data therefore remains to be validated.
  2. On some datasets, more than half of the modifications move observations further away from the reference values. Under spike-only corruption, the average forecasting error also increases slightly. There is insufficient evidence that the method can effectively repair anomalies while preserving genuine changes.
Questions for authors
  1. How should parameters be selected when corruption severity is unknown? Can the same settings handle both clean and corrupted data?
  2. Did examining the test data influence the method’s design? Are results available for new time periods that were not used during development?
  3. In which intervals are observations made less accurate by repair primarily located, and by how much do their errors increase?
  4. What are RestoreTS’s main advantages over existing time-series repair methods?
4 comments

Thank you for the thoughtful review. Updated code is in the anonymous repository linked on line 26.

Q1/W1: Unknown Severity and the Same Settings

We apologize for the ambiguous wording. Severity is not a necessary input. Originally, validation reconstruction error selected each severity's configuration; the label retrieved it during testing. Restoration then used the observed history and selected parameters (Algorithm 1).

In the hidden-label replay, clean validation references selected configurations without severity labels. Across eight datasets with PatchTST and seed-2021 checkpoints, all choices matched Appendix A.1 and forecasts agreed to floating-point precision in the table below.

Values below average horizons and severity levels. "Both versions" means original and hidden-label versions are the same values. The results indicate that severity is not a necessary input.

DatasetBoth versions: MSE/MAEMax abs. MSE difference ↓Max abs. MAE difference ↓
ETTh10.498808/0.4761632.33e-151.22e-15
ETTh20.364613/0.4048631.22e-151.39e-15
ETTm10.476622/0.4425442.94e-152.78e-15
ETTm20.308347/0.3556322.89e-151.61e-15
Weather0.288107/0.3083861.83e-151.55e-15
AQWan0.938501/0.5668864.33e-152.55e-15
CzeLan0.402271/0.3481212.44e-157.77e-16
ZafNoo0.619361/0.5188932.33e-152.89e-15
Overall0.487079/0.4276864.33e-152.89e-15

we evaluated one fixed configuration on Weather/PatchTST. We chose one fixed configurations setting by mean validation reconstruction MSE across L0–L5 with equal condition weights. And we applied the same settings to every test under different severities.

“Specific” selects per severity; “fixed” uses one choice. Both RestoreTS columns use PatchTST. MSE/MAE average four horizons.

LevelRaw PatchTSTRestoreTS: specificRestoreTS: fixed
L00.2256/0.26360.2255/0.26340.2268/0.2633
L10.2389/0.28350.2361/0.27670.2361/0.2767
L20.2644/0.31150.2525/0.29120.2525/0.2912
L30.3052/0.33710.2784/0.30620.2784/0.3062
L40.3687/0.38200.3228/0.34160.3228/0.3416
L50.4943/0.42170.4133/0.37120.4133/0.3712

The fixed choice matches the original at L1–L5. On L0, MSE increases by 0.50% versus raw PatchTST while MAE decreases by 0.11%, which suggests the necessity of our validation selection module.

Q2/W1: Method Design and Generalization

BOCD observations on training and validation histories motivated our design: it partitions histories into runs and exposes changes between them.

Following your suggestion, we evaluated 2026 Beutenberg weather data from Jena: September 11–30 for validation and October 1–November 9 for testing. The test period occurred after the submission deadline.

We reused seed-2021 checkpoints and selected settings on new validation histories. Results average four horizons and L0–L5.

MethodMSE↓MAE↓
PatchTST0.40220.3720
PatchLGA0.36800.3651
RestoreTS + PatchTST0.35640.3503

These results support our method can generalize to an untouched new time period that were not used during the method development.

Q3/W2: Worsened Edits

Some edits worsen accuracy. Both columns use all worsened edits as the denominator and sum to 100%.

DatasetOn uncorrupted values (%)On corrupted values (%)
ETTh192.357.65
ETTh291.898.11
ETTm183.4616.54
ETTm294.085.92
Weather91.828.18
AQWan88.0911.91
CzeLan88.2411.76
ZafNoo91.678.33

Most worsened edits affect initially correct values (uncorrupted data). Appendix C, Table 12 gives improved/worsened fractions. These frequencies do not quantify error increases. Forecasting shows aggregate benefits despite harmful edits. Our study does not establish a general relationship between individual corrections and forecasting error.

Q4: Main Advantages

  1. RestoreTS shifts accepted segments by constants, preserving within-segment differences. Median/Hampel filters replace values using local medians (line 1546).
  2. SCREEN (Section 2.2, lines 115–117) and MTCSC [1] use speed constraints. RestoreTS uses departures and returns to identify temporary offsets, offering an alternative when suitable bounds are difficult to specify or estimate.
  3. Unlike EDITOR [2], RestoreTS needs no separately trained neural repair model. Clean references select validation settings; restoration needs no clean test counterpart.

Table 23 compares filters before the same PatchTST backbone. Under shift-only corruption, RestoreTS reduces MSE/MAE by 7.45%/5.97%. Median and Hampel increase MSE by 2.88% and 2.43%.

We will clarify severity-label usage in Section 3.4 and Appendix A and add the new results to Section 4. Thank you for your insightful feedback and reconsideration.

[1] Aoqian Zhang et al. Multivariate Time Series Cleaning under Speed Constraints. PACMMOD, 2024.

[2] Chenyang Li et al. EDITOR: Multi-Resolution Cleaning of Multivariate Time Series via Detect-Localize-Repair. ICDE, 2026.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer ZfSU,

Thank you again for your thoughtful review. We have uploaded a revised manuscript incorporating the additional experiments and clarifications described in our response. In particular, we clarified the usage of corruption severity, investigated the wrongly edited intervals further and conducted our method on new time periods that were not used during development.

We would appreciate your feedback on whether these updates address your main concerns. If any points remain unclear, we would be happy to provide further clarification.

Thank you for your time and consideration.

Thank you for the careful response and for running these additional experiments. The clarification about severity labels is helpful, and the results on the new time period help address my concern about generalization beyond the data used during development.

I still have two main concerns. First, parameter selection continues to rely on clean validation references. The fixed-configuration results are encouraging, but they only cover Weather with PatchTST. It is still unclear how consistently this approach would work across other datasets and models when corruption conditions are unknown.

Second, the new analysis shows where harmful edits occur, but not how much they increase the error relative to the clean reference. Since most worsened edits affect originally uncorrupted values, their magnitude matters for judging whether the method preserves valid observations. I would also appreciate a clearer explanation of the performance drop under spike-only corruption.

Overall, the additional results make the forecasting benefits more convincing, but these remaining concerns still matter to my assessment. I therefore keep my current rating.

Thank you again for taking the comments seriously and putting effort into the additional analysis.

Thank you for the follow-up. We truly appreciate your time and the care you have given to the paper. We address the two remaining concerns below.

Concern 1: Extended Fix Configuration Experiments

We acknowledge that parameter selection rely on clean validation references in our method. However, clean references are needed only during initial parameter selection in our method. Even there are new unseen time series data inputs, we solely use the initial parameter selection in the new time series data without clean references.

We conducted one new experiment: we selected one restoration method and parameter set using only the 2020 validation references and applied it unchanged to the 2026 test period. No 2026 validation clean references or severity labels were used for selection or restoration. These results in the table below demonstrate that historical reference-based selection can support later deployment without repeating that selection for the evaluated period.

BackboneRaw MSERaw MAERestoreTS MSERestoreTS MAE
PatchTST0.42270.38280.36770.3566
PatchLGA0.38260.37480.35250.3527
DLinear0.52310.40860.41880.3800
TimesNet0.98020.48200.52050.4279
TimeMixer0.43180.39160.36100.3611
iTransformer0.49050.39700.38860.3682

Following your suggestions, we extended our one fixed configuration experiments to eight datasets and all six backbones.

For each dataset, we selected one complete configuration using reconstruction MSE against clean validation references, averaged equally across L0–L5. We then used this one fixed configuration without supplying severity labels to selection or restoration.

The table below reports MAE reduction (%) versus the raw backbone. Errors are averaged equally over combined-corruption levels L1–L5 and horizons {96, 192, 336, 720} before computing the reduction. Positive values indicate improvement.

DatasetPatchTSTPatchLGADLinearTimesNetTimeMixeriTransformer
ETTh10.280.270.171.270.390.58
ETTh20.900.510.313.181.603.31
ETTm16.536.005.748.498.138.07
ETTm22.352.422.003.552.813.63
Weather8.589.349.1514.4211.939.32
AQWan6.167.495.167.266.785.38
CzeLan15.7114.467.9017.6315.9612.00
ZafNoo2.572.672.215.922.332.74

The extended experiments show consistent improvements across most tested datasets and forecasting models and these results provide broader evidence of effectiveness when corruption severity is unknown.

Concern 2: Error Magnitude & Performance Drop under Spikes

Thank you for clarifying this point. Our previous response focused on downstream forecasting errors, but your question concerned errors in the edited inputs relative to the clean reference. We apologize for this misunderstanding. Figure 1 (https://anonymous.4open.science/r/restorets-review-B9BB/rebuttal/figures/restorets_spike_failure.png) shows an ETTh1 spike-only example where 15 originally uncorrupted values at steps 141–155 are wrongly shifted. On these values, mean absolute input error increases from 0 to 3.448, and mean squared input error from 0 to 11.890. And the table below show the average magnitude of worsened edits on uncorrupted data across different datasets.

“Harmed (%)” is the percentage of originally uncorrupted values that RestoreTS changed. The error columns average only over those changed values. Their error relative to the clean reference was zero before restoration.

DatasetHarmed (%)Mean added absolute error
ETTh10.132.660
ETTh21.461.417
ETTm10.232.350
ETTm20.671.135
Weather0.890.932
AQWan1.501.657
CzeLan0.341.032
ZafNoo0.761.702

Regard a further explanation about the performance drop under spikes, RestoreTS focuses on temporary level shifts and is less effective for isolated spikes. In Figure 1 (https://anonymous.4open.science/r/restorets-review-B9BB/rebuttal/figures/restorets_spike_failure.png), the injected spikes remain unchanged, while an initially uncorrupted interval is translated to an incorrect level. This indicated the reason why the performance degraded than the baseline model, which added more worsened regions into the inputs than the one that baseline model received. This shows that adding spikes can alter the correction decisions and trigger harmful changes to a part of clean observations. We acknowledge this weakness and the need to improve selectivity under spike-only corruption.

Thank you for your insightful and careful feedback and reconsideration.

Reviewer_x9CJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RestoreTS is an input restoration layer that combines Bayesian change point detection with temporal return criteria to correct suspicious intervals before forecasting. It leaves the forecasting backbone unchanged and records the applied corrections. Experiments across eight datasets demonstrate forecasting improvements under controlled combined corruption.

Critical strengths

The method produces corrected histories and explicit correction records, allowing interventions to be inspected and incorporated into existing forecasting pipelines without retraining.

Critical weaknesses

The paper selects a separate configuration for each dataset and supplied corruption level, whereas severity is generally unknown and may vary over time in practice. Its effectiveness without this information remains unclear.

Questions for authors

1.How would the corruption level be determined at deployment when severity is unknown or changes over time?

2.Could you evaluate a single fixed configuration across clean and corrupted inputs, and report sensitivity to incorrectly specified corruption levels?

  1. Given that validation and test histories were inspected during earlier method development, could you evaluate the frozen method on an untouched temporal period or a new dataset to assess generalization?
4 comments

Thank you for the careful and constructive review. We appreciate your recognition of explicit corrected histories. Updated code is available in the anonymous repository linked on line 26 of the submission.

Q1/Q2: Corruption Level at Deployment

We apologize for the ambiguous wording. RestoreTS can select its configuration using reconstruction error against clean validation references on the validation set without receiving a severity label. Severity is not a necessary input in our method.

Original implementation. We evaluated several corruption levels for each dataset. For each level, validation reconstruction error selected the restoration method and parameters. During testing, the severity label retrieved that saved choice. Once loaded, restoration used the observed history and selected parameters (Algorithm 1, lines 937 and 954).

Hidden-label replay. We removed severity labels from selection and restoration, selected configurations using clean validation references, and applied them to the corresponding test histories. Validation and test conditions remained matched.

Across the original eight datasets with PatchTST backbone and the same seed-2021 checkpoints, all selected configurations matched Appendix A.1. Forecast errors agreed to floating-point precision. Values below average four horizons and six severity levels (Clean L0 and Corrupted L1~L5).

DatasetOriginal MSE/MAEHidden-label MSE/MAEMax abs. MSE difference ↓Max abs. MAE difference ↓
ETTh10.498808/0.4761630.498808/0.4761632.33e-151.22e-15
ETTh20.364613/0.4048630.364613/0.4048631.22e-151.39e-15
ETTm10.476622/0.4425440.476622/0.4425442.94e-152.78e-15
ETTm20.308347/0.3556320.308347/0.3556322.89e-151.61e-15
Weather0.288107/0.3083860.288107/0.3083861.83e-151.55e-15
AQWan0.938501/0.5668860.938501/0.5668864.33e-152.55e-15
CzeLan0.402271/0.3481210.402271/0.3481212.44e-157.77e-16
ZafNoo0.619361/0.5188930.619361/0.5188932.33e-152.89e-15
Overall0.487079/0.4276860.487079/0.4276864.33e-152.89e-15

For fixed observed data and selected parameters, an incorrect severity label cannot change this implementation's output because the label is unused.

Q2: Fixed Configuration across Clean and Corrupted Inputs

Following your suggestion, we evaluated one fixed configuration on Weather/PatchTST. We chose one set of fixed configurations by mean validation reconstruction MSE across L0–L5 with equal condition weights. And we applied the fixed configurations to every test under different severities.

Below, “specific” selects separately for each severity, and “fixed” uses the single selected configuration. Both RestoreTS columns use PatchTST. Values are averaged forecasting MSE/MAE, using identical checkpoints.

LevelRaw PatchTSTRestoreTS: specificRestoreTS: fixed
L00.2256/0.26360.2255/0.26340.2268/0.2633
L10.2389/0.28350.2361/0.27670.2361/0.2767
L20.2644/0.31150.2525/0.29120.2525/0.2912
L30.3052/0.33710.2784/0.30620.2784/0.3062
L40.3687/0.38200.3228/0.34160.3228/0.3416
L50.4943/0.42170.4133/0.37120.4133/0.3712

The fixed configuration matches the original at L1–L5. On clean L0, MSE increases by 0.50% versus raw PatchTST, while MAE decreases by 0.11%. This indicates the necessity of our validation selection module.

Q3: Assess Generalization

Following your suggestion, we evaluated new 2026 data from Beutenberg weather station in Jena, Germany. We used September 11–30 for validation and October 1–November 9 for testing. The test period occurred after the submission deadline and could not have been inspected during pre-submission development.

We reused the saved seed-2021 forecasting checkpoints and selected restoration settings using the new validation histories. MSE/MAE average four horizons and six conditions (L0–L5).

MethodMSE↓MAE↓
PatchTST0.40220.3720
PatchLGA0.36800.3651
RestoreTS + PatchTST0.35640.3503

These results support transfer to an untouched time period, which indicates that our RestoreTS can generalize to unseen time series data.

We will clarify severity-label usage in Section 3.4 and Appendix A and add the new results to Section 4. Thank you for your valuable feedback and reconsideration.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer x9CJ,

Thank you again for your careful review. We have uploaded a revised manuscript incorporating the additional experiments and clarifications described in our response. In particular, we clarified the usage of severity level inputs, added one fixed-configuration experiment and conducted our method on untouched temporal periods.

We would appreciate your feedback on whether these updates address your main concerns. If any points remain unclear, we would be happy to provide further clarification.

Thank you for your time and consideration.

Thank you for the clarifications and additional experiments. The fixed-configuration experiment on Weather/PatchTST addresses part of my concern about severity-dependent tuning, and the evaluation on an untouched temporal period helps address my concern about prior inspection of test data.

However, the hidden-label replay still selects configurations using clean validation references under corruption conditions matched to those at test time. Removing the severity label therefore leaves open whether configuration selection remains effective when deployment corruption differs from validation conditions or changes over time. The fixed-configuration results are encouraging, but their coverage of one dataset and backbone provides limited evidence for robustness in this broader setting.

As this deployment concern remains central to my assessment, I will maintain my current rating.

Thank you for the follow-up. We truly appreciate your time and the care you gave to our paper. We address your deployment concerns below.

Thanks for clarifying that your concern is how well a configuration transfers when deployment conditions differ from validation. We extended the fixed-configuration study to eight datasets and six backbones. For each dataset, we selected one fixed configuration using reconstruction MSE against clean validation references, averaged equally across L0–L5. We then used this one fixed configuration without supplying severity labels to selection/ restoration and matched severity level test, which is one fixed configuration for various severity levels cases.

The table below reports MAE reduction (%) versus the raw backbone. Errors are averaged equally over combined-corruption levels L1–L5 and horizons {96, 192, 336, 720} before computing the reduction. Positive values indicate improvement.

DatasetPatchTSTPatchLGADLinearTimesNetTimeMixeriTransformer
ETTh10.280.270.171.270.390.58
ETTh20.900.510.313.181.603.31
ETTm16.536.005.748.498.138.07
ETTm22.352.422.003.552.813.63
Weather8.589.349.1514.4211.939.32
AQWan6.167.495.167.266.785.38
CzeLan15.7114.467.9017.6315.9612.00
ZafNoo2.572.672.215.922.332.74

The extended experiments show consistent improvements across most tested datasets and forecasting models and these results indicate that our method configuration selection remains effective when deployment corruption differs from validation conditions or changes over time.It also provided broader evidence for robustness in such broader setting.

Thank you again for your thoughtful feedback and we would appreciate it if you can take another reconsideration.

Comments on the paper

3 comments

Thank you for the thoughtful and insightful review. Updated code is in the anonymous repository linked on line 26.

W1: Severity-agnostic Evaluation

We apologize for the ambiguous wording. RestoreTS can select its configuration using validation dataset without severity inputs. Originally, validation reconstruction error selected each severity's configuration; the label retrieved it during testing. Restoration then used the observed history and selected parameters (Algorithm 1).

In the hidden-label replay, clean validation references selected configurations without severity labels. Across eight datasets with PatchTST and seed-2021 checkpoints, all choices matched Appendix A.1 and forecasts agreed to floating-point precision in the table below.

Values below average horizons and severity levels. "Both versions" means original and hidden-label versions are the same numerical values. The results indicate that severity is not necessary input.

DatasetBoth versions: MSE/MAEMax abs. MSE difference ↓Max abs. MAE difference ↓
ETTh10.498808/0.4761632.33e-151.22e-15
ETTh20.364613/0.4048631.22e-151.39e-15
ETTm10.476622/0.4425442.94e-152.78e-15
ETTm20.308347/0.3556322.89e-151.61e-15
Weather0.288107/0.3083861.83e-151.55e-15
AQWan0.938501/0.5668864.33e-152.55e-15
CzeLan0.402271/0.3481212.44e-157.77e-16
ZafNoo0.619361/0.5188932.33e-152.89e-15
Overall0.487079/0.4276864.33e-152.89e-15

Regarding the concern about the lack of clean references at fault sites, our method only needs clean reference on the validation sets and, in practice, such references could come from historical data curated by domain experts.

W2: Spike-only Corruption & SAPFLUXNET Motivation

Thanks for pointing out this important weakness. We acknowledge that RestoreTS methods focus on temporary level shifts. Our experiments show larger benefits under shift-only and combined corruption, while spike-only performance remains a limitation. During the method development, we have tried specific solutions to handle spikes, but when we transfered such solutions to combined corruption data, the performance degraded than baseline models (Under ETTm1, MSE increased 0.558039 -> 0.634202 from raw PatchTST to the spike-specific solution). We can view this as trade-off, which motivated the current model design can have better performance under combined corruption.

Regarding the concern about SAPFLUXNET motivation, we would like to clarify that SAPFLUXNET motivates our emphasis on explicit corrections and documented changes to observations. RestoreTS does not aim to reproduce its whole expert quality-control workflow.

W3: Sigficance Tests, Seed Studies and Generalization Assessment

Following your valuable and practical suggestions, we added the significance test, conducted seed studies covering all six backbones and assessed generalization on unseen 2026 weather dataset, which indicate that our method performance is not due to chance.

We acknowledge that our RestoreTS provides forecasting benefits across multiple datasets and backbones, although the size and statistical support of those benefits vary across datasets. The table below reports MSE reductions (%) from adding RestoreTS to each backbone, averaged over three training seeds, four horizons, and combined corruption L1–L5. * indicates p<0.05. The results show the gains and uncertainty vary across datasets.

DatasetRestoreTS+PatchTST MSE reduction (%)p
ETTh10.37*0.0470
ETTh21.470.3062
ETTm110.83*0.0001
ETTm23.02*0.0001
Weather13.94*0.0001
AQWan5.54*0.0001
CzeLan19.89*0.0001
ZafNoo6.260.0599
Nominally significant gains6/8—

We thank the reviewer for highlighting this limitation. We acknowledge the smaller, statistically inconclusive gains on some datasets and will investigate ways to improve performance on these datasets in future work.

Regarding the concern about inspection during method development, we evaluated 2026 Beutenberg weather data from Jena: September 11–30 for validation and October 1–November 9 for testing. The test period occurred after the submission deadline. Table below is untouched 2026 Weather period experiments results. This indicates that RestoreTS's forecasting benefits can extend to this previously unseen data with statistical evidence support.

PipelineMSE/MAE reduction (%)ΔMSEMSE p
RestoreTS+PatchTST13.01/6.820.055020.0006
RestoreTS+PatchLGA7.84/5.870.030000.0001

We will clarify severity-label usage in Section 3.4 and Appendix A, put the extended seed studies Appendix D and add the new results to Section 4.

Thank you for your insightful and detailed feedback. We appreciate your positive assessment of our work and your constructive suggestions for improvement.

AuthorsOpenReview ↗

Follow-up on Rebuttal

Dear Reviewer oURg,

Thank you again for your insightful review and positive comments. We have uploaded a revised manuscript incorporating the additional experiments and clarifications described in our response. Particularly, we clarified the usage of corruption severity input, analyzed the experiments performance results with more statistical support and acknowledged the weakness on the spikes.

We would appreciate your feedback on whether these updates address your main concerns. If any points remain unclear, we would be happy to provide further clarification.

Thank you for your time and effort in reviewing our paper.

AuthorsOpenReview ↗

Final Summary of Rebuttal

We sincerely thank all reviewers and chairs for the time and thoughtful engagement throughout the discussion. We would like to provide a concise summary of the rebuttal discussion.

1. Summary of reviewer feedback

Reviewer x9CJ highlighted the value of explicit corrected histories and correction records, which allow users to inspect edits and use RestoreTS with existing forecasting models without retraining. The reviewer asked whether RestoreTS could handle unknown or changing corruption severity, use one configuration across clean and corrupted inputs, and transfer to an untouched test period. We clarified how severity labels were used, expanded the fixed-configuration evaluation to eight datasets and six backbones, and evaluated transfer to the 2026 Weather period.

Reviewer ZfSU highlighted that the corrected histories allow direct assessment of repairs and that RestoreTS improved forecasts across several models. The reviewer asked about configuration selection without severity labels, how RestoreTS compares with existing repair methods, and where and by how much harmful edits increase error. We clarified the comparison with prior work, expanded the fixed-configuration evaluation, and measured the error introduced on originally uncorrupted values against the clean reference. We also illustrated how spike-only corruption can trigger an incorrect interval correction.

Reviewer oURg asked how reliable the forecasting gains were across datasets and training seeds, especially given the smaller gains on ETTh1 and ETTh2. The reviewer also asked about prior inspection of benchmark data and the lack of evaluation on real flagged intervals in SAPFLUXNET. We expanded the original benchmark evaluation to three training seeds per backbone, added paired statistical tests, and evaluated the method on an untouched Weather test period. We clarified that the SAPFLUXNET-derived experiments use synthetic corruption and that evaluation on real flagged intervals remains open.

2. How we addressed the central concerns

  • Corruption severity and fixed configurations. We clarified that the original benchmark used severity labels to retrieve saved configurations. In the expanded evaluation, we selected one configuration per dataset using reconstruction error against clean validation references and applied it unchanged across clean and corrupted inputs, four horizons, and six backbones. No severity labels were supplied at test time. On corrupted inputs, average MSE improved in 46 of 48 dataset–model pairs and MAE improved in all 48. On clean inputs, the largest average MSE increase was 2.18%.

  • Transfer to an untouched test period. We applied the exact Weather configuration saved from 2020 validation to the 2026 Weather test period without retuning. Across the six backbones, average MSE on corrupted inputs fell by 7.9–46.9% relative to the same models using raw inputs.

  • Magnitude of harmful edits. We measured error relative to the clean reference on originally uncorrupted values within corrupted test histories. RestoreTS changed 0.13–1.50% of these values across datasets. Among the changed values, mean absolute input error rose from zero to 0.932–2.660 in training-standardized units.

We again thank the reviewers and chairs for the constructive discussion!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.