acceptodds
Under review as a conference paper at ICLR 2027

SPImpute: Structure-Preserving History Imputation for Frozen Time Series Forecasting

Abstract

Time series foundation models forecast from history, but missing intervals can change their predictions. Existing imputers often focus on reconstruction error without considering how well downstream models can use the data. We propose SPImpute to fill missing values from observed history and a given period. Its sparse quadratic reconstruction combines local differences with reference increments across periods. Decomposing the correction penalty identifies when missing intervals interact and when the penalty reduces to degree-weighted reference-value fidelity. For eligible long gaps, a same-phase candidate corrects donor levels when all available donors lie on one side of the missing position. Masked validation tasks re-estimate drift. If correction leaves the actual candidate unchanged, SPImpute retains the original procedure. Extensive experiments show its effectiveness. On nineteen series at five missing rates, SPImpute ranks first among 21 imputers under two foundation models, with mean MSE ranks of 5.39 and 6.38. MSE reductions relative to Linear, averaged equally across data sources, are 10.81%, 10.11%, and 2.69% under Sundial, TimesFM-2.5, and Chronos-2, respectively. Independent experiments confirm conditional coupling gains for repeated pulse gaps. We compare same-phase correction with the uncorrected procedure using 1,152 synthetic histories. Across 24 drift settings near the forecast origin, mean relative MSE reductions are 92.9% and 94.9% under TimesFM-2.5 and Chronos-2, respectively; all 576 interior-gap inputs remain unchanged. Median CPU imputation time is 78.66 ms.

Primary area:
learning on time series and dynamical systems
Keywords:
time series imputation, reference increments, observation constraints, frozen forecasting models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7TTYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SPImpute, a structure-preserving imputation method tailored for “frozen” time-series foundation models, which must forecast from incomplete historical inputs without adapting the predictor. The method combines a convex quadratic reconstruction that preserves observed points while aligning lag-p increments to a reference trajectory with a same-phase candidate that corrects donor levels for one-sided, long gaps using a robust median slope. A masked historical validation scheme selects among fixed quadratic weightings and the same-phase candidate without using labels or querying the predictor, and extensive experiments across 19 real datasets and synthetic studies show consistent downstream-forecasting gains, competitive runtime, and insightful attribution controls.

Critical strengths
  1. The same-phase donor correction with a robust, median lag-p slope is a simple yet thoughtful mechanism to compensate for drift in one-sided, long gaps; the trend-equivariance property for eligible coordinates is well-articulated.
  2. The selection protocol is carefully designed: historical block masking, per-mask re-estimation of reference and slope, and strict switching criteria ensure conservative adoption of the same-phase candidate while keeping the predictor frozen.
  3. Imputation that preserves structure useful to downstream frozen forecasters is an important and timely problem; SPImpute is methodologically lightweight, interpretable, and deployable, and demonstrates consistent improvements over widely used baselines.
Critical weaknesses
  1. The masked-validation scoring uses MAE for quadratic candidates but RMSE for the same-phase candidate, which could bias selection; the rationale and sensitivity are not fully justified.
  2. The same-phase correction relies on at least eight observed lag-p pairs; sensitivity to this threshold and behavior under sparse/volatile observations are not analyzed.
  3. The period is supplied by the datasets; there is no study of misspecified or unknown p, which is a common practical scenario. How sensitive is SPImpute to period misspecification? Please report a robustness study where p is perturbed or when multiple plausible periods exist, and discuss how validation behaves under wrong p
  4. Have you considered or experimented with simple auto-period estimation (e.g. STL-derived seasonal length) and then running SPImpute? Even preliminary results would strengthen applicability when p is unknown.
Questions for authors

See Weaknesses

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7TTY (Part 1 of 2)

Thank you for recognizing the one-sided drift correction and historical validation design. We added comparisons of scoring rules, observation counts, and period settings. We explain the design choices and results below.

1. Candidate scoring

We use RMSE, which is more sensitive to large errors, to evaluate the same-phase candidate and impose a stricter condition for replacing the default quadratic reconstruction. If selected, this candidate replaces the quadratic repair at every missing position. RMSE allows large deviations at even a few positions to affect this decision. The same-phase candidate must also outperform every quadratic candidate and satisfy the 3% switching threshold. Replacement therefore requires sufficient evidence from historical validation.

We compared the original mixed scoring with two alternatives: MAE for all candidates and RMSE for all candidates. The candidates, validation blocks, and switching threshold remained unchanged. The experiment used 19 series from 10 sources under BM30. Each predictor evaluated 903 histories; Sundial and TimesFM-2.5 shared masks, while Chronos-2 used independent masks. Forecast MSE changes relative to mixed scoring are:

Scoring ruleSundialTimesFM-2.5Chronos-2
MAE for all candidates0.16% increase0.31% decrease0.55% increase
RMSE for all candidates0.25% increase0.62% decrease0.16% increase

In the main mask panel, same-phase selections increased from 141 with mixed scoring to 185 with uniform MAE and 182 with uniform RMSE. The corresponding counts in the independent mask panel were 142, 172, and 173. Mixed scoring thus selects the same-phase candidate more conservatively.

Under both uniform rules, the absolute change in mean MSE was below 0.62%, and the direction varied across predictors. We retain mixed scoring to control candidate switching; the comparisons show that this choice has a small effect on mean forecasting performance.

2. The minimum of eight observed pairs

Eight is the minimum number of observed pairs required to enable drift estimation, preventing correction based on very few observed changes. The slope is computed from all valid pairs. Both endpoints must be observed and exactly one period apart. Once the minimum is met, the algorithm computes the rate of change for each pair and takes their median.

We recounted the pairs in 1,806 actual histories and their 5,418 internal masked-validation tasks:

Mask panelValid pairs in the original incomplete historyValid pairs after masking a validation block
Main822 to 2,015774 to 1,941
Independent765 to 2,052673 to 1,986

Even the smallest validation task retained 673 pairs. Changing the threshold from 8 to 4, 16, or 32 therefore uses exactly the same pairs to estimate the slope in every task, leaving candidate selection and imputation unchanged. All these estimates are supported by hundreds of observed pairs.

When fewer pairs are available than the threshold requires, the algorithm sets the slope to zero and continues with the uncorrected same-phase candidate and quadratic reconstruction. Taking the median reduces the influence of individual extreme pairwise changes. Each historical validation task also re-estimates the slope and checks the error after correction. The threshold controls whether correction is enabled when observations are scarce.

AuthorsOpenReview ↗

Response to Reviewer 7TTY (Part 2 of 2)

3. Sensitivity to an incorrectly specified period

The effect of period misspecification depends on how the period is changed.

We tested p−1p-1, p+1p+1, p/2p/2, and 2p2p on the same incomplete histories, keeping all other settings unchanged. The following forecast MSE changes use the supplied period pp as the reference:

Input periodSundialTimesFM-2.5Chronos-2
p−1p-12.41% increase3.62% increase0.46% decrease
p+1p+12.35% increase1.15% increase1.85% decrease
p/2p/211.79% increase16.13% increase2.26% increase
2p2p0.83% increase1.10% decrease0.92% decrease

The absolute change in mean error was at most 3.62% for a one-step period shift and 1.10% for a doubled period. Halving the period caused larger increases: 11.79%, 16.13%, and 2.26% under Sundial, TimesFM-2.5, and Chronos-2, respectively.

For a signal repeating every pp steps, positions separated by 2p2p still share a phase, whereas positions separated by p/2p/2 may lie in different parts of the waveform. For example, a position that should use another peak as its reference may instead use a trough. Doubled and halved periods therefore need separate evaluation.

Compared directly with SPImpute without periodic guidance, using p/2p/2 reduced mean MSE by 5.81%, 8.86%, and 6.90%, respectively. The other three perturbations also retained average gains over the version without periodic guidance. Period errors reduced some forecasting gains, while periodic guidance continued to help.

With p/2p/2 in the main panel, the selected candidate's validation score worsened in 525 of 903 histories, and same-phase selections decreased from 141 to 99. Historical validation thus reduces the use of same-phase candidates with less favorable scores and reselects the imputation under the supplied period.

These comparisons show smaller average effects for slight period shifts and doubled periods, and a larger effect for halved periods. Reliable period information yields lower mean forecasting error, while approximate periods can retain gains over using no period.

4. Estimating the period from observed history

SPImpute can first estimate a period from observed history and then perform imputation. With historical period estimation, all three predictors achieved lower mean forecasting error than with periodic guidance removed entirely.

The original paper takes pp as an input and uses the dataset-supplied period to compare imputation strategies under the same period information. When the period is unknown, we can identify a repetition length in the observed history and pass it to the same SPImpute algorithm, leaving the subsequent imputation procedure unchanged.

Our added estimator first removes a linear trend from the observed history, then examines repetition using pairs with both endpoints observed. A correlation peak satisfying fixed criteria supplies the period; if none is found, SPImpute runs without periodic guidance. The estimator receives only the observed history, with no dataset name, sampling frequency, original period, or future values.

We compared supplied periods, automatically estimated periods, and no period:

ComparisonSundialTimesFM-2.5Chronos-2
Estimated vs. supplied period10.75% increase9.40% increase2.79% increase
Estimated vs. no period7.26% decrease13.11% decrease6.44% decrease

When reliable period information is available, using it directly avoids estimation and gives lower mean error. When it is unavailable, historical estimation reduces mean MSE relative to using no period by 7.26%, 13.11%, and 6.44%. It therefore provides a usable imputation approach for inputs without period metadata.

For traffic, the supplied period is 24 hours, but automatic estimation selected the 168-hour weekly period in 42 of 45 histories in each mask panel. Daily and weekly variation can coexist in the same history, and the period used for imputation affects subsequent forecasts. Supplied periods and historical estimation support use with known and unknown periods, respectively.

We have revised the manuscript in response to your comments; the revised text is marked in blue, with purple margin labels identifying changes addressing your questions.

Thank you to the authors for the thorough rebuttal and for incorporating the new experimental results and discussions into the revised manuscript.

As most of my previous concerns have been resolved with empirical evidence, I am raising my rating from 2 to 3.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for reviewing our response and revised manuscript, and for your positive feedback on the additional experiments. Your suggestions on candidate scoring, period misspecification, and automatic period estimation helped us strengthen the experimental evaluation and explain the rationale and practical use of our method more clearly. These comments have helped us improve the paper. We appreciate the time and care you have devoted to reviewing our work.

Reviewer_B4ufOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SPImpute fills missing historical values before forecasting with frozen time-series models. Given a seasonal period, it combines observation-constrained quadratic reconstruction, cross-period reference increments, and drift-corrected same-phase estimates. Masked historical validation selects candidates without querying the forecaster. Experiments on 19 real-world series, three forecasting models, and controlled synthetic data report improved average forecasting performance and examine when the proposed mechanisms are beneficial.

Critical strengths
  1. The mathematical analysis clearly characterizes when the reference-increment penalty introduces coupling between missing positions and when it reduces to degree-weighted reference-value fidelity. This makes the proposed mechanism interpretable and establishes useful boundaries on its claimed contribution.
Critical weaknesses
  1. The comparisons do not sufficiently isolate the benefit of the method from its information and context advantages. SPImpute explicitly uses a supplied period, whereas many baselines do not incorporate this prior. Moreover, neural baselines use 96-step windows, while SPImpute accesses histories of up to 2,880 steps and some datasets have periods of 288. These settings make it difficult to attribute the reported gains to the proposed mechanism rather than additional periodic information or longer context. Period-aware, context-matched comparisons are needed.

  2. The central mechanism has limited demonstrated benefit on real data. In the matched BM30 experiments (Table F2), SPImpute differs from degree-weighted reference-value guidance by less than 0.1% under all three forecasters, without significant improvements. Simpler matched controls are also competitive. Although synthetic experiments establish benefits under specific conditions, the real-data evidence does not clearly establish the practical contribution of increment coupling beyond reference construction and same-phase estimation.

Questions for authors
  1. Evaluation clarification and targeted validation: Which baselines explicitly receive the supplied period? How much of SPImpute’s advantage remains against period-aware baselines using context windows that cover at least one or two periods, particularly when p>96p>96?

  2. Contribution clarification: Given the negligible differences from matched reference-value guidance in Table F2, what evidence establishes an independent real-data benefit from increment coupling? If its benefits are confined to particular missingness geometries, how frequently do those conditions occur in the evaluated real histories?

5 comments
AuthorsOpenReview ↗

Response to Reviewer B4uf (Part 1 of 2)

Thank you for your questions about period information, input windows, and increment coupling. We added neural-baseline comparisons with longer windows and examined cross-period connections in naturally occurring missingness.

1. The external baseline STL-Kalman receives the same supplied period as SPImpute. The internal weighted-reference-value, periodic-consistency, and same-phase controls also use this period. Neural baselines train on 96-step segments extracted from the history. When the period exceeds 96 steps, an individual segment cannot cover a full period. We therefore extended the neural baseline's window to one and two periods to evaluate the effect on forecasting.

We reran TimeMixer, included in the original paper, on all six series with periods greater than 96 steps. weather has period 144; the other five have period 288. For each masked history, we used windows of 96 steps, one period, and two periods. For period 288, these correspond to 96, 288, and 576 steps. Observed values, missing positions, random seeds, and settings such as model depth and hidden width were fixed. Each window setting used 59 training segments per epoch, three epochs, and six parameter updates in total. Longer segments still came from the same visible history and covered its beginning and end. Some TimeMixer layers gain parameters as the window grows; the matched quantities are the number of training segments, epochs, and parameter updates.

This experiment used contiguous missing blocks at a 30% missing rate. The six series represent five sources. Sundial and TimesFM-2.5 shared 288 masked histories; Chronos-2 used another 288 histories with independent masks. Each completed history was passed to the same frozen forecaster to evaluate forecast MSE over the next 720 time points. TimeMixer's changes relative to its original 96-step window are:

TimeMixer windowSundialTimesFM-2.5Chronos-2
One period (pp)1.70% lower0.65% lower1.24% higher
Two periods (2p2p)2.49% lower1.61% lower3.36% lower

The two-period window reduces mean error under all three forecasters. We further compare SPImpute with these TimeMixer settings and with STL-Kalman using the same period. Both STL-Kalman and SPImpute can use the full history of up to 2,880 steps before the forecast origin. The following values are SPImpute's forecast MSE reductions relative to each method:

ComparatorSundialTimesFM-2.5Chronos-2
TimeMixer, 96 steps13.95%12.09%17.23%
TimeMixer, one period12.88%11.53%18.54%
TimeMixer, two periods12.34%10.98%15.06%
STL-Kalman, same period and full history6.77%2.42%11.10%

Relative to TimeMixer with a two-period window, SPImpute reduces mean error by 12.34%, 10.98%, and 15.06%, with improvements on 3/5, 4/5, and 4/5 sources, respectively. Against STL-Kalman with the same period and full history, improvements occur on 5/5, 4/5, and 4/5 sources. SPImpute retains lower mean forecast error after extending TimeMixer's window and supplying STL-Kalman with the same period and full history. Thus, the gains of the complete imputation procedure remain after strengthening the baselines' period information and context.

AuthorsOpenReview ↗

Response to Reviewer B4uf (Part 2 of 2)

2. Increment coupling is mainly useful when missingness recurs across periods while the series level changes. Missing positions one period apart then need to adjust together to recover the changes between them.

We constructed synthetic pulse series with short gaps aligned at the same positions over six consecutive periods and different levels of drift. The same-phase branch was disabled. Both reconstruction methods used the same reference trajectory, local smoothing terms, and candidate-selection rules; only the cross-terms linking missing positions changed. The table reports forecast MSE reductions for the increment objective relative to the matched reference-value objective with these cross-terms removed:

Missingness conditionTimesFM-2.5Chronos-2
Aligned short gaps across six periods, moderate drift13.13%9.86%
Aligned short gaps across six periods, stronger drift21.27%23.65%

Retaining the cross-terms reduces mean MSE by 9.86%–23.65% across these four comparisons. This repeated-gap experiment shows that allowing missing values one period apart to adjust together can improve subsequent forecasts of drifting signals.

To address the frequency of such configurations, we examined the original PM2.5 records with naturally occurring missing values from 12 Beijing air-quality monitoring stations in the UCI dataset (Chen, 2017), which contains hourly records spanning 2013–2017. At a daily lag of 24 hours, the original data contain 2,048 pairs of positions that are both missing, involving 2,912 missing values, or 33.32% of all original missing values. Partitioning each station's records from its starting time into nonoverlapping, full-length 2,880-hour segments yields 144 segments. Of these, 89 contain such cross-day missingness connections, or 61.81%. Moreover, 501 of the 2,048 pairs connect distinct contiguous gaps, and all 12 stations exhibit this pattern.

Field observations provide concrete examples of the underlying collection failures. A U.S. Geological Survey field report by Stannard et al. (2013, p. 16) documents 13 nighttime data outages caused by insufficient battery storage; collection resumed when solar power became available during the day. This produced repeated missingness at similar times across daily cycles. The same report records a 13.1-day continuous outage caused by a communication failure. Such a long gap also contains many pairs of positions one day apart that are both missing.

These original-data statistics and field records establish a practical context for the cross-period missingness connections addressed by increment coupling.

Song Chen. Beijing Multi-Site Air Quality [Dataset]. UCI Machine Learning Repository, 2017. doi: 10.24432/C5RK5G.

David I. Stannard, Marshall W. Gannett, Danial J. Polette, Jason M. Cameron, M. Scott Waibel, and J. Mark Spears. Evapotranspiration from marsh and open-water sites at Upper Klamath Lake, Oregon, 2008–2010. Scientific Investigations Report 2013–5014, U.S. Geological Survey, Reston, VA, 2013. doi: 10.3133/sir20135014.

We have revised the manuscript in response to your comments; the revised text is marked in blue, with green margin labels identifying changes addressing your questions.

AuthorsOpenReview ↗

Follow-up on our response to Reviewer B4uf

Thank you again for your suggestions on the fairness of the comparisons and the role of increment coupling. We have added baseline comparisons with longer windows and an analysis of missingness recurring across periods. Please let us know if any of these additions need further explanation or if you have other questions. We would be happy to provide further details and discuss any remaining concerns. We look forward to your feedback.

Reviewer_zm3kOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

For the setting where a frozen time series foundation model receives a history with missing segments, the paper proposes SPImpute, which treats observed values as equality constraints and solves a sparse positive-definite linear system combining first-/second-order difference smoothing with a lag-pp periodic increment penalty; this is complemented by same-phase donor candidates with one-sided drift correction and masked-validation selection, supported by a graph Dirichlet-energy analysis and an affine-equivariance property, and validated on 19 series with 3 frozen forecasters.

Critical strengths
  1. A concise formulation with well-defined properties. Casting imputation as "observed equality constraints + quadratic energy minimization" yields a unique solution that can be computed efficiently with sparse direct solvers while exactly preserving observed values. The decomposition in Eq. 3 reduces the difference between "reference increments" and "reference values" guidance to cross terms on missing–missing edges: without such edges the two are exactly equivalent; with them, multiple gaps move jointly. This gives an interpretable and testable mechanistic criterion, which is more convincing than simply asserting that "increments are better."
  2. Training-free and forecaster-query-free. The method does not rely on pretrained imputation weights or queries to the forecaster, so it can serve as a front-end module for any frozen TSFM. The authors are restrained in stating the scope of Property 1 (it holds only on one-sided coordinates and says nothing about candidate selection or forecasting performance), and the discussion of limitations is fairly honest.
Critical weaknesses
  1. The "downstream-forecasting-oriented" motivation is disconnected from the method; the problem setting never actually enters the method.
    • The central argument of the introduction is that reconstruction error does not accurately reflect the value of imputation for downstream forecasting. Yet every design choice in SPImpute is independent of the downstream forecaster: the quadratic energy is a pure signal prior, and weight/candidate selection uses reconstruction MAE/RMSE on masked blocks (Eq. 6). In other words, SPImpute is essentially a generic period-aware imputer; "frozen forecasting" is only the evaluation scenario, not a constraint on or source of signal for the method design.
    • The paper provides no argument for why "preserving increment structure" should improve foundation-model forecasts. For example, which input statistics are foundation models sensitive to? How do over-smoothing or level misalignment affect attention and tokenization? If the "downstream-oriented" positioning is to be maintained, the method should incorporate forecaster-related criteria, or at least provide a theoretical or mechanistic argument; otherwise, the paper should be repositioned as "periodic-structure-preserving imputation."
  2. Limited methodological novelty, with the distinctiveness of the core concept undermined by Eq. 3 itself. In addition, the method relies on many hand-crafted rules without a unifying principle.
    • First-/second-order difference smoothing corresponds to Whittaker/HP smoothing and trend filtering; seasonal difference penalties, constrained reconstruction guided by reference gradients (the authors themselves cite Poisson image editing), and STL-style trend-plus-phase references are all well-established ideas. Combining them into a single quadratic objective is an engineering-level integration.
    • More importantly, Eq. 3 shows that without missing–missing periodic edges, e.g., a single gap shorter than the period, or gaps whose spacing is not pp, increment fidelity is exactly equivalent to degree-weighted value fidelity. This is a very common missingness structure. Hence the paper's headline "reference increments" only make a difference under specific missingness geometries, which makes it a rather thin core contribution.
    • Same-phase candidates, the one-sided test, the median slope, "fall back to the original pipeline if the candidate is unchanged", 8 fixed weight configurations, and various thresholds and scoring coefficients (0.05, 0.01, 3%) together form a rule system lacking a unifying modeling principle. For instance, one-sided drift correction could be naturally absorbed into the quadratic objective as a trend term instead of being a separate candidate followed by validation-based switching; likewise, the reference trajectory rr already contains a trend estimate, and the relation between it and the separate β^\hat\beta correction is not discussed.
    • The theoretical results (the graph-energy expansion in Eq. 3, positive definiteness, the translation equivariance of the median in C.2, the spectral bound in C.3) are standard results obtainable in one or two lines, offering limited guidance for method design.
Questions for authors

Clarifications (answers may affect my rating):

  1. Where exactly is "frozen-forecaster orientation" reflected in the method? If the downstream model were replaced by an arbitrary analysis task, would any part of SPImpute's design change?
  2. Why not incorporate one-sided drift correction directly into the quadratic objective as a linear trend term (or explicitly model the trend in the reference trajectory rr), instead of designing a separate candidate plus validation-based switching? What are the principled trade-offs between the two designs?
  3. Eq. 3 shows that without missing–missing edges, increment fidelity is equivalent to value fidelity. Please argue for the frequency and importance of such edges in realistic missingness structures, to support "reference increments" as a core contribution.
  4. The selection criterion is still masked-block reconstruction error. How is this fundamentally different from the "reconstruction-driven selection" criticized in the introduction?
  5. A brief experimental question: are the average gains concentrated in a few strongly periodic sources (e.g., traffic, Australia_Solar)? Please report results with these two sources removed.

Suggestions:

  • The main text should substantially compress implementation details (moving them to the appendix) and use the space to explain the design principles and to analyze the mechanism by which increment structure affects the forecaster.
  • Unknown or time-varying periods and multivariate dependencies are key extensions for practical use.
6 comments
AuthorsOpenReview ↗

Response to Reviewer zm3k (Part 1 of 3)

Thank you for your positive assessment of the formulation and theoretical analysis, and for your questions about the frozen-forecasting setting, design choices, and experimental results. We explain the roles of one-sided drift correction and historical validation, and provide statistics on naturally occurring missingness and results excluding traffic and Australia_Solar. We address each question below.

1. The frozen-forecaster setting determines the information available during imputation and the relationship between the imputer and the prediction model.

In forecasting, a model receives an observed time series and predicts subsequent time points. Imputation therefore uses only the visible history that will be supplied to the forecaster, its missing positions, and its period, all before the forecast origin. Retrospective completion of recorded data may allow additional observations after a gap. At a forecast origin, subsequent observations have not yet become available and cannot be used. We also treat the downstream forecaster as a fixed, pretrained model. SPImpute neither reads its internal parameters nor calls it to select an imputation. We first complete the history and then pass it to the forecaster. This procedure can be used with any time series foundation model.

If the downstream task changes to an analysis task whose model also takes a complete time series, SPImpute remains applicable. The task determines which observations are available. If it permits a longer observed interval, the algorithm can impute that interval.

2. The trend in the reference trajectory and the one-sided drift correction serve two different imputation candidates. SPImpute constructs several candidate completions and selects one. Quadratic reconstruction uses the reference trajectory rr, local smoothness, and periodic relationships to infer missing values. The same-phase candidate instead directly uses observations at corresponding positions in nearby preceding and following periods. For daily data, these can be observations at the same time on the previous or following day.

The one-sided correction addresses a bias in the latter candidate. Suppose a missing position has observations only one and two periods earlier, while the series level is rising. Averaging these observations carries the earlier, lower level into the current position. We estimate the rate of increase or decrease from observed pairs one period apart and translate the donors to the target time. In the paper's example, donors at i−pi-p and i−2pi-2p have an average time 1.5p1.5p earlier than the target, so their mean requires a correction of 1.5pβ^1.5p\hat\beta.

When selected, the same-phase candidate directly replaces the quadratic reconstruction. Changing only the trend in rr therefore cannot change this same-phase average. Correcting the candidate addresses this bias directly and makes its scope explicit: adjust positions whose donors all lie on one side, while retaining the original values where donors lie on both sides. Incorporating the correction into the quadratic objective would additionally require a weight relative to the smoothing terms.

We constructed 24 synthetic conditions with drift and gaps near the forecast origin, comparing the complete method against the same method with only the one-sided correction removed. With correction, mean relative forecast MSE decreased by 14.6%, 92.9%, and 94.9% under Sundial, TimesFM-2.5, and Chronos-2, respectively. For another 576 histories with interior gaps and same-phase observations on both sides, all imputations and forecasts remained unchanged.

AuthorsOpenReview ↗

Response to Reviewer zm3k (Part 2 of 3)

3. Increment coupling is mainly useful when missingness recurs across periods while the series level changes. Missing positions one period apart then need to adjust together to recover the changes between them.

To address the frequency of such configurations, we examined the original PM2.5 records with naturally occurring missing values from 12 Beijing air-quality monitoring stations in the UCI dataset (Chen, 2017), which contains hourly records spanning 2013–2017. At a daily lag of 24 hours, the original data contain 2,048 pairs of positions that are both missing, involving 2,912 missing values, or 33.32% of all original missing values. Partitioning each station's records from its starting time into nonoverlapping, full-length 2,880-hour segments yields 144 segments. Of these, 89 contain such cross-day missingness connections, or 61.81%. Moreover, 501 of the 2,048 pairs connect distinct contiguous gaps, and all 12 stations exhibit this pattern.

Field observations provide concrete examples of the underlying collection failures. A U.S. Geological Survey field report by Stannard et al. (2013, p. 16) documents 13 nighttime data outages caused by insufficient battery storage; collection resumed when solar power became available during the day. This produced repeated missingness at similar times across daily cycles. The same report records a 13.1-day continuous outage caused by a communication failure. Such a long gap also contains many pairs of positions one day apart that are both missing.

These original-data statistics and field records establish a practical context for the cross-period missingness connections addressed by increment coupling.

Song Chen. Beijing Multi-Site Air Quality [Dataset]. UCI Machine Learning Repository, 2017. doi: 10.24432/C5RK5G.

David I. Stannard, Marshall W. Gannett, Danial J. Polette, Jason M. Cameron, M. Scott Waibel, and J. Mark Spears. Evapotranspiration from marsh and open-water sites at Upper Klamath Lake, Oregon, 2008–2010. Scientific Investigations Report 2013–5014, U.S. Geological Survey, Reston, VA, 2013. doi: 10.3133/sir20135014.

AuthorsOpenReview ↗

Response to Reviewer zm3k (Part 3 of 3)

4. Equation (6) selects an imputation using visible history. We temporarily hide a short block of known observations, let the candidates reconstruct it, and compare their results with the observed values. After each masking step, the reference trajectory, same-phase donors, and drift slope are recomputed from the remaining observations. Alongside reconstruction error at the hidden positions, the score considers jumps at gap boundaries and excessive local variation.

The Introduction discusses the relationship between imputation accuracy and downstream forecasting. We use historical validation to select a completion and forecast error over the next 720 time points to evaluate its downstream value. Imputation uses only available observations, without reading future labels or querying the forecaster's internal parameters. This selection procedure is therefore directly applicable to forecasting.

5. SPImpute improves forecasting across multiple sources beyond traffic and Australia_Solar. We address this through source-exclusion results and results from five additional sources. The table reports forecast MSE reductions relative to linear interpolation: positive values indicate lower error, and negative values indicate higher error. We first compute results within each series, average series belonging to the same source, and then weight sources equally.

Sources includedSundialTimesFM-2.5Chronos-2
All 10 sources10.81%10.11%2.69%
Excluding traffic5.52%3.57%-0.47%
Excluding Australia_Solar8.97%7.51%0.33%
Excluding both: 8 sources remain2.79%-0.17%-3.53%

After excluding both sources, Sundial improves on 7 of the remaining 8 sources, with mean forecast MSE reduced by 2.79%. TimesFM-2.5 improves on 5 sources; its mean MSE increases by 0.17%, remaining close to parity. Chronos-2 improves on 2 sources, with mean MSE increasing by 3.53%. Sundial and TimesFM-2.5 improve on most remaining sources. Thus, gains occur across multiple sources, while their average magnitude varies with the data and forecaster.

The ablation experiments provide further direct evidence using EWELD electricity load, Finland traffic, Northeast China wind, Singapore air quality, and the electricity dataset. The table reports SPImpute's forecast MSE reductions relative to linear interpolation, with EWELD load also reported as a specific example.

Additional-source comparisonSundialTimesFM-2.5Chronos-2
Five-source average16.61%19.86%12.93%
EWELD electricity load34.35%41.70%39.16%

The average reductions are positive under all three forecasters. EWELD load yields reductions of 34.35%, 41.70%, and 39.16%, respectively, providing direct evidence of gains on electricity load beyond the two specified sources. traffic and Australia_Solar increase the mean improvement in the main evaluation, and the method's forecasting gains are also supported by other sources.

We have revised the manuscript in response to your comments; the revised text is marked in blue, with red margin labels identifying changes addressing your questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.