acceptodds
Under review as a conference paper at ICLR 2027

Component-Aware Imputation Refinement for Time Series Foundation Models

Abstract

Time series foundation models generally require imputing missing historical values before forecasting. Because imputers vary across series and missingness patterns, forecast quality depends on the initial imputer. We propose component-aware imputation refinement to reduce downstream forecast error and dependence on imputer choice. Starting from an existing imputation, the method applies robust STL decomposition and constructs a local trend bridge from gap boundaries, a seasonal reference from the median of observed same-phase history, and a residual reference. Window-level structural descriptors select the correction action and magnitude, while updates are restricted to originally missing positions. In paired evaluations across 18 datasets, 16 base imputers, and four zero-shot forecasters, mean sMAPE decreases by 3.70%, and the mean within-dataset-and-forecaster standard deviation across imputers decreases by 40.51%. The results indicate improved average forecast accuracy and reduced dependence on base-imputer choice within the evaluated scope.

Primary area:
learning on time series and dynamical systems
Keywords:
time series imputation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_k4JKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a refinement module for time series imputation to improve downstream forecasting and reduce dependence on the choice of imputation method. The module adjusts imputed values using temporal structure without retraining the forecasting model. Experiments across multiple datasets and models show improvements in average forecasting performance.

Critical strengths

S1. It is interesting to improve downstream forecasting while reducing dependence on the choice of base imputation method. S2. The proposed module requires no additional training or modification of the forecasting model, and can be integrated into existing imputation pipelines.

Critical weaknesses

W1. The novelty is not sufficiently established. Prioritizing forecasting utility over reconstruction accuracy has already been explored in task-oriented imputation methods such as TOI-VSF [1]. W2. The experimental validation is insufficient, lacking component ablations and hyperparameter sensitivity analyses. The method relies on several hand-crafted rules and model-specific thresholds for TimesFM, but their selection is not sufficiently justified. [1] Hao et al., “Is Precise Recovery Necessary? A Task-Oriented Imputation Approach for Time Series Forecasting on Variable Subset,” IEEE Transactions on Knowledge and Data Engineering, 2025.

Questions for authors

See above

4 comments
AuthorsOpenReview ↗

Response to Reviewer k4JK (1 of 2): contribution relative to TOI-VSF

To W1. Thank you for drawing our attention to TOI-VSF and asking us to make the contribution more explicit. We agree that improving forecasting utility is a shared research goal. We have revised the positioning to explain the specific correction mechanism developed here and the evidence supporting it.

Our contribution is a component-aware post-imputation correction mechanism for frozen forecasters. Starting from an existing completion, we use STL to represent its trend, seasonal and residual components. A trend bridge uses the levels at the gap boundaries; a seasonal phase median draws on recurring same-phase values at originally observed positions; a weighted residual reference uses the corresponding residual values. These references supply correction directions, and statistics of the completed history set their dynamic strengths. Only originally missing entries are updated. Our correction rules and parameter settings were developed through iterative experiments guided by downstream forecasting performance and case analysis (Appendix H). Once fixed, the refinement operator requires no forecaster calls at inference.

The design uses historical temporal structure to improve the forecasting input. The references are not ground-truth components that the method seeks to reconstruct exactly: the trend bridge may be extrapolated by its strength, and seasonal and residual components move partway toward their references. Their forecasting value is assessed through paired future forecast errors. This makes the component construction and its controlled use central to our contribution, rather than the general choice to evaluate forecasting utility.

This mechanism is distinct from TOI-VSF's jointly learned imputation and forecasting. TOI-VSF addresses variables entirely absent at inference. Our correction acts on partial gaps in a completed history, requires no training or forecaster calls to compute its updates, and can follow an existing imputer while the pretrained forecaster remains frozen. The decomposition, three reference constructions and dynamic updates provide a concrete way to exploit historical structure in this setting.

Table K1. Main results, 345,600 matched cases (Sections 5.2-5.3).

MetricUnrefinedRefinedReduction
Mean sMAPE5.25945.01724.60%
Mean between-imputer sMAPE SD0.45710.215752.82%

The mechanism improves forecast accuracy and reduces dependence on the base imputer. The standard deviation above is computed across imputers within each dataset-forecaster cell and then averaged. These complementary benefits are evaluated across 18 datasets, 16 imputers and four frozen forecasters. Refinement also improves a forecast-selected imputer: on 10,800 pairs, sMAPE falls from 5.70697 to 5.53291, a reduction of 0.17406 [0.02877, 0.45019]. Thus the benefit remains after selecting an imputer for downstream forecasting.

The ablations in our next reply connect these gains to the proposed design: the eight-subset comparison tests the component combination, and the same-reference dynamic/fixed comparison tests correction strength. They support the mechanism beyond an overall before/after comparison.

To make the contribution explicit, we made three sets of revisions:

1. Stating the specific contributions. Section 1 identifies training-free refinement, the three-reference mechanism with dynamic strengths, and its measured benefits.

2. Distinguishing the mechanism from related work. Section 2.2 briefly introduces TOI-VSF and contrasts joint learning with frozen-forecaster refinement; Sections 3-4 specify our objective and operator.

3. Connecting the contribution to evidence. Sections 5.2-5.3 report accuracy and between-imputer dispersion. Section 5.6/Table 4 adds forecast-guided imputer selection, with its protocol in Appendix J.

We appreciate the opportunity to clarify this distinction: the shared forecasting goal motivates the problem, while the concrete operator and its measured effects establish this work's contribution.

AuthorsOpenReview ↗

Response to Reviewer k4JK (2 of 2): ablations and parameter evidence

To W2. We have taken your concern about hand-crafted rules seriously and added comparisons that separate component choice, correction strength and the original action rule. This evidence also led us to simplify the method. Below we summarize the central results and the parameter and sensitivity checks.

Each component contributes to the joint correction in the paired averages. The fixed common panel contains 18 datasets, five historical windows per dataset, all five missingness patterns at three ratios, four base imputers and four forecasters: 21,600 cases per configuration. All-three has the lowest mean sMAPE (5.32958) among eight subsets. Removing trend, seasonality or residual correction raises mean sMAPE by 0.07908, 0.04064 or 0.03839 points, respectively. Each individual 95% interval is above zero. Each active component keeps its original reference and dynamic strength.

Dynamic strengths improve over a simpler correction using the same references. Both arms in the first row below always correct all three components; only the strengths differ. The second row separately compares the original action rule (including keep-input and zero-target branches) with all-three correction under the same strength formulas.

Table K2. Action and strength controls on the same 21,600 cases.

First / secondFirst sMAPESecond sMAPEDifference [95% CI]
Dynamic / fixed (1.5, 0.5, 0.1)5.329585.37177-0.04219 [-0.07145, -0.01434]
Original action / all three5.342935.32958+0.01335 [0.00085, 0.02233]

Dynamic strengths lower overall mean error; the 345,600-case comparison also improves the mean for each forecaster. In the action-rule comparison, the dynamic formulas and shared thresholds are identical on both sides. Applying all three corrections gives lower mean error. We retain the references and dynamic strengths, remove the original action rule from the method, and report its comparison here to explain that decision.

The retained parameters have a traceable selection history. Early four-dataset, three-ratio weighted seasonal-reference searches over K={1,2,3,5}, then {3,4,5}, selected K=3; the residual reference inherited it. Trend trials compared 0.9/1.2/1.5/1.8/2.0 and favored 1.5. Case analysis then motivated boosting to 2.0 only with sufficient boundary, trend-reference and residual-reference discrepancies. The seasonal/residual formulas set response ranges and jointly improved over their predecessor on all five development forecasters across 27 datasets, supporting the formulas as a complete configuration. A later multiplier/radius search screened at window ranks 1/5 and evaluated at 9/12. Its tuned K=1 baseline was included in the subsequent 1,800-case comparison, where the adopted phase-median/K=3 configuration scored 6.82188 versus 6.83582. This selected a complete configuration. The stricter TimesFM exception came from a six-dataset, three-ratio pilot to limit aggressive trend extrapolation, followed by a 27-dataset expansion.

On all 86,400 TimesFM pairs, shared thresholds give 4.94106 versus 4.94338 for the original specific thresholds, a difference of -0.00232 [-0.00512, -0.00023]. We therefore use the shared thresholds for every forecaster in the revised method. On the 21,600 common cases, fixing only trend strength at 1.0/1.5 gives 5.35537/5.33894 versus 5.32958 for dynamic strengths, with seasonal/residual strengths unchanged; the two individual intervals include zero. Replacing only the residual target with zero gives 5.34654, an increase of 0.01696 [0.00329, 0.02992]. This reference-only control, reported here, supports using the same-phase residual reference.

Matched period and gap-length controls. Using the same five windows as the core ablations, half/double periods change sMAPE by +0.12605/+0.04040 on 2,160 pairs. The half-period adjusted interval is above zero. With 400 missing points, refinement-minus-base differences for 25/50/100-point gaps are -0.04570/-0.04792/-0.09875 on 680 pairs each. The interval at length 25 is below zero; the other two include zero. Appendix G/Table 22 reports all estimates.

In response to this concern, we made three sets of revisions:

1. Adding controlled experiments. Section 5.5/Tables 2-3 give common-panel controls and full-panel strength evidence; Appendices F and I specify matching and component protocols.

2. Explaining parameters and sensitivity. Section 4.3/Appendix B specifies the shared formulas, Appendix H explains their selection, and Appendix G reports period/gap sensitivity.

3. Synchronizing the simplified method. The abstract, Section 4, Figure 2, Algorithm 1 and conclusion describe the final method. These replies report the original-action, specific-threshold and zero-target controls.

Differences are sMAPE points; paired source-family intervals are 95% (period contrasts: adjusted 97.5%).

Thanks for the clarification. The results provide convincing evidence that the proposed refinement improves the final forecasting performance. However, I still have concerns about the methodological novelty.

AuthorsOpenReview ↗

Follow-up on the contribution of post-imputation refinement

Thank you for following up and for recognizing the forecasting improvements. We would like to clarify what our refinement contributes to forecasting-oriented imputation.

  1. We use historical structure to guide how an existing imputation is refined for forecasting. TOI-VSF and our work share the goal of making imputation useful for forecasting. TOI-VSF pursues this goal by jointly training the imputer and forecaster. We start from an existing imputation and use the history around each gap to construct references for its trend, seasonal, and residual components. These references give the corrections a direction. An update can move partway toward its reference or extend beyond it. During development, downstream forecast error guided our choice of rules for how far to move each component. At inference, these rules determine the correction strengths from the current historical window. Only originally missing entries are updated (Section 4).

  2. The refinement is simple to add to an existing forecasting pipeline. Users can keep both the base imputer and the pretrained forecaster they already use. The correction is computed from the completed history, the original missingness mask, and the seasonal period, without additional training or any call to the forecaster. The frozen forecaster then makes its usual single prediction from the refined history. The same procedure can follow interpolation, statistical imputation, or a learned imputer.

  3. The refinement reduces sensitivity to the choice of base imputer. We applied the same correction rules to the outputs of different imputers. In our evaluation, this lowered overall mean forecast sMAPE and narrowed the spread in mean sMAPE across imputers (Sections 5.2–5.3). In practical terms, the initial choice of imputer had less influence on average forecast accuracy after refinement.

Thank you again for sharing your perspective. Could you say a little more about which aspect of the refinement prompted your concern about methodological novelty?

Reviewer_As1yOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a post-imputation refinement method for forecasting with frozen time-series foundation models. It decomposes the completed history using STL, constructs trend, seasonal, and residual references, and applies fixed rules based on window statistics to select corrections and their strengths, modifying only originally missing entries. Experiments report lower average forecast error and reduced variation across imputers.

Critical strengths
  1. The method provides an interpretable, reusable refinement procedure that preserves observed values and can accompany different base imputers and frozen forecasters without additional model training or forecaster calls.

  2. The broad paired evaluation covers diverse datasets, imputers, forecasters, and missingness settings under matched forecasting conditions. The reported reductions in average forecast error and variation across imputers support the practical value of the overall procedure within the evaluated scope.

Critical weaknesses
  1. The broad paired evaluation supports the overall refinement procedure, but the incremental value of the selective gate and feature-dependent correction strengths over a simpler refinement using the same STL component references remains unclear.
Questions for authors
  1. Could the authors clarify what evidence supports the added value of selective correction and feature-dependent strengths over simpler STL-based refinement? If available, a comparison using the same component references but always correcting all three components with fixed strengths would help assess their contribution.
1 comment
AuthorsOpenReview ↗

Response to Reviewer As1y: contribution beyond fixed-strength refinement

To W1 and Q1. We appreciate your recognition of the reusable refinement procedure and your precise suggestion to compare it with fixed-strength refinement using the same STL references. This comparison is particularly helpful because it separates the benefit of constructing the references from the benefit of deciding how strongly to use them. We have implemented that control and examined the original action rule separately.

With identical references and all three components active, dynamic strengths reduce mean sMAPE by 1.23% relative to fixed (1.5, 0.5, 0.1). The result is based on all 345,600 matched cases, and each forecaster's mean improves.

For direct comparison with the component ablations, Table A1 uses the same fixed common panel. Manuscript Table 3 reports the full-panel dynamic/fixed comparison; Appendix F specifies its protocol.

Table A1. Matched controls on the common 21,600-case panel.

First / secondFirst sMAPESecond sMAPEDifference [95% CI]
Dynamic / fixed (1.5, 0.5, 0.1)5.329585.37177-0.04219 [-0.07145, -0.01434]
Original action / all three5.342935.32958+0.01335 [0.00085, 0.02233]

The fixed common panel contains 18 datasets, five historical windows per dataset, all five missingness patterns at three ratios, four base imputers and four forecasters: 21,600 cases per configuration.

In the first row, both arms use the same trend bridge, seasonal phase median and residual reference. The fixed arm replaces only the correction strengths, so the difference directly addresses your question about feature-dependent magnitudes. In the retained 345,600-case comparison, dynamic strengths also reduce mean between-imputer standard deviation from 0.2661 to 0.2157. They therefore improve both mean forecast accuracy and consistency across base imputers in this comparison.

Your question also led us to simplify selective correction. With the same dynamic formulas and shared thresholds on both sides, always applying all three references lowers mean sMAPE by 0.01335 relative to the original action rule. That rule includes keep-input and zero-residual-shrinkage branches. We remove it from the revised method and report the comparison here as the basis for that change. The final procedure applies the three original reference corrections with dynamic strengths and shared thresholds.

The component comparison further supports this choice. On the same 21,600 cases, all-three has the lowest mean sMAPE among eight subsets (5.32958). Removing trend, seasonality or residual correction raises mean sMAPE by 0.07908, 0.04064 or 0.03839 points, respectively. Each individual 95% interval is above zero. These results distinguish the supported component and strength contributions from the original selective action rule.

Your proposed comparison led to three sets of revisions:

1. Adding matched design controls. Section 5.5/Table 2 reports the same-reference dynamic/fixed comparison, original action policy, eight subsets and fixed-trend controls on the complete common panel.

2. Providing broader evidence and protocols. Table 3 gives the 345,600-case strength comparison; Appendices F and I specify its protocol and the common cases. Table A1 above directly answers the action and strength questions.

3. Updating the method throughout. The abstract, Section 4, Figure 2, Algorithm 1 and conclusion consistently describe the final all-three procedure with dynamic strengths and shared thresholds.

These changes connect your suggested control to both the evidence and the simplified design. Differences are sMAPE points with paired source-family 95% intervals.

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

This paper studies the problem of post-imputation correction for time series foundation models. It applies robust STL decomposition to the imputation results and then computes correction values according to predefined rules to correct the original imputations. The authors conduct a relatively large number of experiments to demonstrate the effectiveness of the method.

Critical strengths

1, The research problem is practically meaningful, and the proposed method can be easily integrated into existing pipelines. 2, The experiments are relatively thorough.

Critical weaknesses

1, The method contains many heuristic operations, and there is no corresponding ablation study. 2, The hyperparameter selection process is unclear, and no concrete method for choosing hyperparameters is provided. 3, The comparison with task-aware imputation methods is insufficient (e.g., the method proposed in Section 2.2 of the paper).

Questions for authors

1, How are all the thresholds, coefficient formulas, and TimesFM-specific configurations determined? 2, Can the authors add key ablations and simpler correction baselines? It would be helpful to compare at least seasonal-only correction, removing the residual reference, removing gating, and fixing the trend coefficient to 1.0, 1.5, and the current dynamic configuration. What are the gains of the full method relative to each of these variants? 3, Is the method sensitive to the period and the missing length?

6 comments
AuthorsOpenReview ↗

Response to Reviewer Eiis (1 of 5): component, action and strength ablations

To W1 and Q2. Thank you for recognizing the practical relevance of the problem and for specifying the ablations needed to assess the heuristic choices. We agree that overall improvement alone does not establish the contribution of each design. Your suggestions prompted us to separate three questions: which component corrections contribute, whether the action rule helps, and whether dynamic strengths improve over fixed strengths. This reply addresses those questions; the next reply continues Q2 with the residual-target and trend-strength controls.

The three-component combination has the lowest average error among all eight subsets. Each active component retains its original STL reference and dynamic strength, with the action gate disabled. Thus the comparison changes which component updates are applied to the same base completion, without changing their references or coefficient rules.

Table E1. Eight subsets on the same 21,600 cases (Section 5.5, Appendix I).

ConfigurationMean sMAPEDifference from all-three [95% CI]
Unrefined5.50091+0.17133 [0.04711, 0.31052]
Trend only5.41500+0.08542 [0.02980, 0.14871]
Seasonal only5.45117+0.12159 [0.02751, 0.25414]
Residual only5.45410+0.12452 [0.02818, 0.24932]
Trend + seasonal5.36796+0.03839 [0.01645, 0.06424]
Trend + residual5.37022+0.04064 [0.01306, 0.07301]
Seasonal + residual5.40865+0.07908 [0.00134, 0.20127]
All three, dynamic5.32958Reference

The fixed common panel contains 18 datasets, five historical windows per dataset, all five missingness patterns at three ratios, four base imputers and four forecasters: 21,600 cases per configuration.

The seasonal-only row directly answers your requested simpler baseline. Removing trend, seasonality or residual correction raises mean sMAPE by 0.07908, 0.04064 or 0.03839 points, respectively. Each individual 95% interval is above zero. The trend-only and trend-plus-seasonal rows also test whether the full gain can be attributed to trend correction alone. Each comparison uses the same base inputs and evaluation cases.

Dynamic strengths add value when the references and active components are fixed. To distinguish this effect from action selection, we separately compare the original action rule with all-three correction, and dynamic with fixed strengths under all-three correction.

Table E2. Action and strength controls on the same 21,600 cases.

First / secondFirst sMAPESecond sMAPEDifference [95% CI]
Original action / all three5.342935.32958+0.01335 [0.00085, 0.02233]
Dynamic / fixed (1.5, 0.5, 0.1)5.329585.37177-0.04219 [-0.07145, -0.01434]

In the dynamic/fixed comparison, both arms use the same three references and differ only in correction strengths. The lower mean error supports dynamic strengths beyond reference construction. The 345,600-case comparison in manuscript Table 3 also favors dynamic strengths, by 0.06260 points, with lower mean error for each forecaster. In the action comparison, both arms use the same dynamic formulas and shared thresholds; only the original action rule is switched on or off. That rule includes keep-input and residual-to-zero branches. Applying all three reference corrections gives lower average error.

We have used this evidence to simplify the method. The revised main method retains the three references and original dynamic strengths and applies all three corrections. The original action-rule comparison is reported here to explain this revision. This revision directly addresses your concern about unnecessary heuristic operations while preserving the parts supported by the paired controls.

In response to this comment, we made three sets of revisions:

1. Adding the requested core controls. Section 5.5 and Table 2 present the eight subsets, original action policy, fixed strengths and fixed trend coefficients on the same complete panel.

2. Documenting the supporting comparisons. Appendix I specifies the common cases and broader component evidence. Section 5.5/Table 3 reports the full-panel dynamic/fixed comparison; Appendix F gives its protocol.

3. Simplifying the method consistently. The abstract, Section 4, Figure 2, Algorithm 1 and conclusion now describe all-three dynamic refinement with shared thresholds. Table E2 above explains the removal of the original action rule.

All differences above are sMAPE points with paired source-family 95% intervals.

AuthorsOpenReview ↗

Response to Reviewer Eiis (2 of 5): residual reference and trend strength

Continuing Q2. Your request to remove the residual reference raises a useful distinction: does residual correction help at all, and does its same-phase target help beyond simply shrinking residuals toward zero? We address these with separate controls, then report the requested trend-strength comparison.

The residual-target control changes only the reference target. Removing residual correction altogether is the trend-plus-seasonal row in Table E1. The comparison below instead retains all three corrections and all three original per-case dynamic coefficients, computed with the original references; only the residual target becomes zero. This avoids mixing a target change with a change in correction strength.

Table E3. Residual-target control on the same 21,600 cases as Tables E1-E2.

ConfigurationMean sMAPEDifference from all-three [95% CI]
All three, dynamic5.32958Reference
Residual target set to zero5.34654+0.01696 [0.00329, 0.02992]

The zero-target minus same-phase difference is +0.01696 [0.00329, 0.02992] sMAPE points. The positive paired difference supports retaining the same-phase residual reference. Tables E1 and E3 thus distinguish the effect of applying residual correction from the effect of using its same-phase reference.

The requested trend-strength comparison now uses the full common panel and the method's actual seasonal/residual coefficients. We fix only the trend coefficient at 1.0 or 1.5. The seasonal and residual strengths remain dynamic and unchanged, and all three configurations use the same STL references with all three components enabled.

Table E4. Trend-strength control on the same 21,600 cases.

ConfigurationMean sMAPEDifference from all-three [95% CI]
Trend coefficient fixed at 1.05.35537+0.02580 [-0.00038, 0.06386]
Trend coefficient fixed at 1.55.33894+0.00936 [-0.00008, 0.02171]
All three, dynamic5.32958Reference

The dynamic trend rule has the lowest mean among these three choices; the individual intervals for both contrasts changing only trend strength include zero. These controls isolate the trend-strength rule in the revised method. Table E2 separately evaluates the complete dynamic coefficient set against the fixed tuple (1.5, 0.5, 0.1).

For these controls, we made two sets of manuscript revisions:

1. Reporting the requested comparisons together. Table 2 includes removal of residual correction and trend coefficients fixed at 1.0/1.5 on the same common panel.

2. Specifying what each comparison changes. Appendix H explains the trend-strength controls, and Appendix I documents the component protocol. These descriptions identify the references and coefficients held fixed.

Table E3 in this reply additionally tests the same-phase residual reference against zero-target shrinkage. Differences are variant minus all-three in sMAPE points, with individual paired source-family 95% intervals.

AuthorsOpenReview ↗

Response to Reviewer Eiis (3 of 5): parameter choices and thresholds

To W2 and Q1. Thank you for asking how the parameters were chosen. We revisited the original candidate trials and code to explain their selection, beyond specifying how to use them.

K=3 originated in a neighborhood search. Early distance-weighted seasonal-reference trials covered four datasets and three missing ratios. At strength 0.5, K=1/2/3/5 gave mean relative sMAPE changes of -0.030/-0.032/-0.621/-0.599%; a follow-up comparing K=3/4/5 at strengths 0.7/0.9 also selected K=3. Here, historical percentages average per-case changes from base imputation, with negative values better. K=3 means up to three cycles on each side. The residual reference inherited this radius.

Trend strength 1.5 and its conditional boost came from successive comparisons. On 12 development cases, strengths 0.9/1.2/1.5/1.8/2.0 gave -7.30/-7.94/-8.11/-7.01/-6.01% in the same historical measure, supporting 1.5. Continuous-strength alternatives did not improve that development mean. Subsequent case diagnostics motivated a boost to 2.0 only when three conditions hold: boundary jump >=0.05 requires a discontinuity; trend-reference discrepancy >=0.03 avoids amplifying nearly unchanged references; residual-reference discrepancy >=0.035 screens out unsupported strong extrapolation. These thresholds arose through successive candidate comparisons and case analysis, not an exhaustive grid.

The seasonal/residual formulas encode specified response ranges, then were tested jointly. Original code sets active seasonal strength from 0.4 for noisy structure to 0.8 for clear seasonality, and zero for negligible seasonal amplitude. Residual strength starts at 0.1, adds up to 0.15 for seasonal support, and subtracts 0.2 times the normalized boundary jump, capped at 0.1. The 0.01 amplitude guards, clipping and two-decimal rounding complete this fixed empirical design. The resulting configuration improved the historical mean over its predecessor on all five development forecasters (27 datasets x three ratios), by 0.12-0.30 percentage points. This supports the complete formula; it does not imply that every internal constant was independently optimized.

The later comparison explains retaining K=3 and the unscaled formulas. A predecessor-v46 search used five datasets, three imputers and four forecasters, screening at window ranks 1/5 and evaluating at 9/12 by macro-average sMAPE. It held thresholds fixed and searched trend multipliers {0.5,0.75,1,1.25}, seasonal/residual multipliers {0.5,1,1.5}, and K={1,3,5}. Its selected (1.25,1,1.5), K=1 reached 6.83582 versus default v46's 6.87202. Crucially, the subsequent 1,800-case mechanism comparison included this tuned baseline: the phase-median configuration with K=3 and unscaled formulas achieved 6.82188 and was selected. This was an exploratory comparison of complete configurations, not an isolated K=3-versus-K=1 test for the final reference.

The original TimesFM exception also has a recorded origin. Early trials compared boosts 1.8/2.0 and threshold tuples (0.05,0.03,0.035), (0.08,0.03,0.05) and (0.10,0.05,0.06). A six-dataset, three-ratio pilot favored the last tuple with boost 2.0 for both TimesFM versions. It was intended to limit aggressive trend extrapolation and was retained in a 27-dataset expansion.

The broader paired evidence supports a simpler, shared threshold configuration. We change only the TimesFM trend-enhancement thresholds, keeping the three references and other coefficient rules fixed.

Table E5. Threshold comparison on the common panel's 5,400 TimesFM pairs.

Thresholds: boundary / trend / residualMean sMAPE
Original TimesFM: (0.10, 0.05, 0.06)5.32427
Shared: (0.05, 0.03, 0.035)5.32259

Shared minus specific is -0.00168 [-0.00301, -0.00012] sMAPE points. On all 86,400 TimesFM pairs, the corresponding means are 4.94106 and 4.94338, a difference of -0.00232 [-0.00512, -0.00023]. Both are paired source-family 95% intervals. We therefore adopt the shared thresholds in the revised method and synchronize the main results, core ablations and sensitivity experiments to that configuration. The historical exception and its control are reported here to explain the decision.

To explain both selection and use, we made three sets of revisions:

1. Specifying the configuration. Section 4.3 defines the history descriptors; Appendix B (Table 5) gives the shared formulas, thresholds and rounding conventions.

2. Explaining the selection history. Appendix H documents the candidate trials and case analysis behind the neighborhood radius, trend settings and dynamic formulas.

3. Testing the retained choices. Section 5.5/Tables 2-3 report common-panel controls and the full-panel strength comparison; Appendices F and H give the matching protocol and fixed-trend definitions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response

We thank the reviewers for recognizing the practical value of post-imputation refinement and for identifying the comparisons needed to assess its design. Your comments helped us distinguish the contribution of the component references from that of the action rule and correction strengths, and to clarify the relationship to forecasting-oriented imputation.

We have revised the manuscript throughout the relevant sections, including the abstract, contributions, related work, method, experimental analysis and conclusion. In the marked manuscript, revisions use red for Eiis, blue for As1y and green for k4JK, with margin annotations identifying related concerns. Actual changes and additions are marked; unchanged text remains black.

The revisions center on the following findings and the changes they motivate. Reported intervals are paired source-family bootstrap confidence intervals (CIs).

(1) Three-component refinement improves both forecasting accuracy and consistency across imputers. On 345,600 matched cases across 18 datasets, 16 imputers and four frozen forecasters, mean sMAPE decreases from 5.2594 to 5.0172 (4.60%). Mean between-imputer standard deviation decreases from 0.4571 to 0.2157 (52.82%). The operator preserves observed values, requires no training and makes no forecaster calls when computing the correction.

(2) Common-case ablations support the three components and dynamic strengths. The fixed common panel contains 18 datasets, five historical windows per dataset, all five missingness patterns at three ratios, four base imputers and four forecasters: 21,600 cases per configuration. All-three has the lowest mean error among eight subsets. Removing trend, seasonality or residual correction raises mean sMAPE by 0.07908, 0.04064 or 0.03839 points; each individual 95% interval is above zero. Dynamic strengths improve over fixed (1.5, 0.5, 0.1) by 0.04219 points. Guided by the controls, we simplify the method to three-component dynamic refinement with the same thresholds for every forecaster. The main-text table includes the eight subsets, the original action-rule comparison, fixed-strength baseline and fixed trend coefficients 1.0/1.5. TimesFM-specific thresholds and residual zero-target comparisons are reported in the replies below.

(3) Refinement adds value after forecast-guided imputer selection. On 10,800 pairs, correcting the base imputer selected using prior forecast losses reduces sMAPE from 5.70697 to 5.53291. This addresses whether refinement still helps after choosing an imputer for the downstream task. We also report the parameter trials, residual-target and trend-strength controls, and period and gap-length sensitivity requested by the reviewers.

We made three sets of manuscript revisions in response to the reviews:

1. Clarifying the contribution and method. Sections 1 and 2.2 distinguish our correction mechanism from task-aware imputation. The abstract, Section 4, Figure 2, Algorithm 1 and conclusion consistently describe all-three dynamic refinement with shared thresholds.

2. Strengthening the experimental evidence. Sections 5.2-5.3 update accuracy and between-imputer dispersion. Section 5.5/Tables 2-3 present the core ablations and full-panel strength comparison; Section 5.6/Table 4 reports forecast-guided imputer selection. Appendices F, I and J retain the protocols and component details.

3. Explaining parameters and sensitivity. Appendix B specifies the formulas and numerical conventions, Appendix H explains their selection, and Appendix G reports matched period and gap-length comparisons. The replies below connect each concern to the evidence and revision.

AuthorsOpenReview ↗

Final summary of revisions for all reviewers

We thank the reviewers for the questions and suggestions that helped shape the revised manuscript. Having considered the feedback from all three reviewers, we summarize below the three main revisions and findings in the final manuscript.

  1. How forecasting guides post-imputation refinement. The revised introduction clarifies what the component references contribute. Starting from an existing imputation, we use the surrounding history to define a correction direction for each component. An update can move partway toward its reference or, for trend, extend beyond it. Downstream forecast error guided the design of the rules governing how far to move. At inference, the current history determines the correction strengths, so the refinement can follow an existing imputer and requires no additional training or forecaster calls. (Sections 1 and 4)

  2. What the component and strength comparisons show. On the same 21,600 cases, applying trend, seasonal, and residual corrections together gave the lowest mean forecast sMAPE among all eight component subsets. Removing any one component increased mean error. We then isolated the contribution of the dynamic strengths on all 345,600 cases: both variants applied all three corrections using identical references. Replacing fixed strengths of (1.5, 0.5, 0.1) with the window-dependent strengths reduced mean sMAPE from 5.0798 to 5.0172, with a lower mean for each of the four frozen forecasters. (Section 5.5, Tables 2–3)

  3. Sensitivity to the seasonal period and gap length. We tested both questions on the same five historical windows per dataset used in the core ablations. The configured seasonal period gave lower mean forecast sMAPE (5.1932) than either half that period (5.3193) or twice that period (5.2336). For gap length, we held the total number of missing points at 400 and varied how they were distributed across gaps. Refinement lowered mean sMAPE relative to the same base imputation at all three gap lengths tested, by 0.0457, 0.0479, and 0.0987 points for gaps of 25, 50, and 100 points, respectively. (Appendix G, Table 22)

We appreciate the reviewers’ time and the opportunity to address these questions. The marked manuscript uses reviewer-specific colors and margin labels to help readers locate the corresponding changes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.