acceptodds
Under review as a conference paper at ICLR 2027

Beyond Missing Values: How Imputation Changes Observed Representations in Time Series Foundation Models

Abstract

Imputation is commonly evaluated by reconstruction error at missing coordinates. In time series foundation models with input-dependent normalization, filling a gap also changes the statistics used to transform the history. It can therefore alter the representations of observed values that remain unchanged in the raw history. Reconstruction scores alone do not describe the changes actually presented to the forecaster. Assessing downstream utility requires understanding how these changes arise and under which observation conditions they are amplified. We analyze the propagation of imputation error through normalization in frozen forecasters. Preserved observations constrain the variance of every feasible completion within each statistical support used by normalization. For positive observed dispersion, we derive a uniform bound on normalized error under ordinary standardization, with a coefficient determined before imputation and sharp when at least two coordinates are missing. Experiments with three frozen forecasters separate reconstruction error, normalized input change, forecast displacement, and forecast loss. In 10% of the 190 baseline conditions, selection by reconstruction score gives forecast RMSE more than 10.14% above the best candidate. For TimesFM, perturbations of observed-history standard deviations yield mean forecast displacement above in the same scale in 454 of 518 constant-patch configurations; fixing historical statistics reduces the mean displacement in every configuration below . Imputation evaluation should report observed dispersion and completed scale within the model's statistical supports, and measure forecast loss directly.

Primary area:
learning on time series and dynamical systems
Keywords:
time series imputation, time series forecasting, error propagation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yN6qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how imputation errors propagate through normalization to alter observed-value representations, derives error bounds, and examines their effects on frozen time series foundation models.

Critical strengths
  1. Provides theoretical bounds with explicit assumptions and a tightness proof.

  2. Extensive controls, including fixed-statistics and coordinate interventions, clearly distinguish forecast displacement from forecast error.

Critical weaknesses
  1. Most amplification evidence concerns constant or low-variance segments under artificial missingness; applicability to natural missingness remains unverified.

  2. Learned imputers are trained on the current history for only 3–10 epochs, potentially affecting method rankings and quantitative conclusions about selection regret.

Questions for authors
  1. Does the mismatch between reconstruction and forecasting rankings persist with larger training budgets or pretrained imputers?

  2. How frequently does amplification occur, and how does it affect forecast error, under natural missingness and native missing-value handling?

2 comments
AuthorsOpenReview ↗

Response to Reviewer yN6q (Part 1 of 2)

Thank you for identifying training budget and natural missingness as the two issues to test further. We increased training budgets and added a pretrained imputer and controls on naturally occurring gaps.

1. More training improves reconstruction, but reconstruction and forecasting rankings remain different.

On the original 19 series, we increased SAITS's training cap from 10 epochs to 50 and 200, and the other four learned methods' caps from 3 epochs to 15 and 60. All comparisons use 30% block missingness and the original 720-step forecast horizon. Each budget selects its checkpoint by validation error using only the visible history.

The table compares the 20× and 1× budgets. Reconstruction reductions are calculated within each series and then summarized by their median. The 38 forecasting conditions consist of 19 series and two forecasters.

MethodMedian reconstruction-RMSE reduction, 20× vs. 1×Conditions with lower forecast RMSE
SAITS19.30%30/38
HELIX26.14%31/38
GPT4TS1.19%22/38
TimeMixer3.26%31/38
TOTEM8.42%23/38

Increasing training improves imputation quality: all five methods have lower seed-averaged reconstruction MAE and RMSE on all 19 series. Forecasting also improves in most conditions, but some conditions do not improve despite the lower reconstruction error. Improvements in reconstruction and forecasting therefore do not correspond one-to-one.

We also added pretrained MOMENT to the highest-budget candidate set. The following retrospective comparison uses all evaluation windows. Here, ρ\rho is the Spearman correlation between reconstruction and forecasting rankings, and excess RMSE is the additional forecast error of the reconstruction-selected method relative to the best forecasting candidate in the same set.

Candidate setMean ρ\rhoMedian excess RMSE90th percentile of excess RMSE
14 non-learned methods0.6172.49%8.64%
19 methods, 1× budget0.5851.56%7.64%
19 methods, 5× budget0.5852.07%7.87%
19 methods, 20× budget0.5982.41%7.85%
20× budget + MOMENT0.5992.48%7.85%

At 20× budget and after adding MOMENT, median excess RMSE remains 2.41% and 2.48%, respectively. Ranking differences also occur among the methods requiring no training. Thus, increasing the training budget and adding a pretrained method do not eliminate the ranking difference in this evaluation. Improved reconstruction still needs to be assessed against the target forecasting task.

Selecting by historical forecasting performance improves average performance on later data. We selected an imputer on earlier windows and evaluated it on later windows. At 20× budget, reconstruction-based selection produced 2.83% higher later RMSE on average than forecast-based selection. Historical forecast error therefore provides a useful basis for choosing an imputer.

AuthorsOpenReview ↗

Response to Reviewer yN6q (Part 2 of 2)

2. Small perturbations can produce large responses at natural gaps, while actual forecast error can increase or decrease.

We retained the original gaps in household-power and Beijing air-quality data and used 512-step histories to forecast the next 128 steps. After imputation, we perturbed only missing values in the first 32-point segment containing a gap. The perturbation RMS was 10−510^{-5} times the observed-history standard deviation. A large response was recorded when forecast RMS displacement, averaged over three directions, exceeded 1% of that standard deviation. The table reports event frequencies among gap-containing configurations:

DataGap-containing configurationsTimesFM eventsToto events
Household power272 (7.41%)2 (7.41%)
Beijing, single station2051 (0.49%)1 (0.49%)
Beijing, multiple stations, evaluated per channel12,587110 (0.87%)107 (0.85%)

Perturbation amplification also occurs at natural gaps and is concentrated in statistical segments that become constant after imputation. Across these gap-containing configurations, the large-response frequencies are 0.88% for TimesFM and 0.86% for Toto. The results connect sensitivity to the dispersion within the normalization support and identify constant segments as a practical condition to inspect.

Forecast changes can also reduce error. Among TimesFM's 113 large responses, RMSE increased in 50 and decreased in 63; both directions were also observed for Toto. The magnitude of a forecast change measures sensitivity, while forecast quality is evaluated against the observed future values.

The effect of observed-position representations also occurs with positive observed variance. For SAITS, we examined 775 histories whose first statistical support containing an actual imputation error has β>0\beta>0. After equal weighting across data sources, replacing only the observed representations changed RMSE by more than 1%, in either direction, in 79.15% of cases for TimesFM and 72.92% for Toto. Actual imputation errors can therefore affect forecasting through these representations in positive-β\beta supports as well.

For native missing-value handling, we compared Toto's native mask input with linear imputation using windows whose last 32 steps were observed in all input channels, as required by the interface. The native path directly receives the missingness mask. Its pooled RMSE was 0.21% lower on household power and 1.72% higher on the single-station Beijing data. These processing options should therefore also be selected using historical performance on the target forecasting task.

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

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

This paper investigates how imputation affects frozen time-series foundation models beyond reconstruction error. The authors derive an observation-dependent lower bound on the variance of feasible completions and a sharp bound on normalized input differences under ordinary standardization. Experiments with three forecasting models, 19 imputation implementations, and multiple datasets show that small imputation perturbations can substantially alter normalized representations, including those of unchanged observed values, and consequently affect forecasts. The paper further demonstrates that forecast displacement does not necessarily translate into increased forecast error.

Critical strengths
  1. Important problem formulation. The paper highlights a relevant but underexplored issue: reconstruction quality alone may not reliably reflect the downstream utility of imputation for time-series foundation models.
  2. Useful theoretical analysis. The observation-dependent variance bound and the sharp normalized-error bound provide a principled explanation of when imputation-induced changes may be amplified.
Critical weaknesses
  1. Limited generalizability. The experiments primarily concern three frozen univariate forecasters and artificial block missingness. Natural or value-dependent missingness, multivariate forecasting, and models with native missing-value handling remain insufficiently investigated. The two additional datasets provide useful evidence but do not establish broad generality.
  2. Limited practical resolution. The paper establishes that reconstruction-based selection can be suboptimal, but it does not provide a validated general-purpose strategy for selecting imputers. The practical recommendations therefore remain primarily diagnostic.
Questions for authors
  1. Can the authors more clearly isolate the effects of input normalization and output-scale restoration, particularly in cases where the latter dominates?
  2. How well does the proposed analysis generalize to naturally occurring missingness, multivariate forecasting, and models with native missing-value handling?
  3. Can the authors clarify the extent to which the internal representation interventions correspond to realizable imputed histories, and distinguish conclusions supported by the theoretical bounds from those supported only by empirical observations?
3 comments
AuthorsOpenReview ↗

Response to Reviewer DvAk (Part 1 of 3)

Thank you for recognizing the importance of the question and the theoretical analysis. Your questions concern the effects of the different paths and whether the findings hold under practical missingness. We added controls addressing both issues.

1. Input normalization and output restoration both affect forecasting, and their relative effects vary across forecasters.

We held the completed input fixed and separately replaced the input-normalization or output-restoration statistics with those of the corresponding complete history. The experiment covers 897 histories from 19 series. The table reports the absolute forecast-RMSE change for each path, divided by the window's observed-history standard deviation ss and then averaged with equal source weights:

Forecaster and imputerInput-path change / ssOutput-path change / ss
TimesFM + Linear0.04420.0197
TimesFM + SAITS0.04400.0264
TimesFM + MOMENT0.05870.0402
Toto + Linear0.02460.0197
Toto + SAITS0.02880.0273
Toto + MOMENT0.04230.0394

The input effect is larger for TimesFM, while the two effects are closer for Toto. Regarding the output-dominant cases you specifically asked about, output restoration had the larger absolute effect in 22 of the 30 source–imputer combinations in the original Sundial experiment. Thus, the relative size of the input-normalization and output-restoration effects depends on the forecaster; neither is dominant across all the tested models.

AuthorsOpenReview ↗

Response to Reviewer DvAk (Part 2 of 3)

2. Natural gaps also exhibit imputation sensitivity, and joint multivariate forecasting permits cross-variable propagation.

Small changes in imputed values also produced large forecast changes in the original gaps of the household-power and Beijing air-quality data. Under the original perturbation criterion, large responses occurred in 0.88% and 0.86% of the 12,819 gap-containing histories for TimesFM and Toto, respectively. They were concentrated in statistical segments that became constant after linear imputation. These segments therefore deserve particular attention when assessing sensitivity under natural missingness.

Imputation effects can propagate to other variables in joint multivariate forecasting. We jointly supplied six pollutants to Toto and slightly changed only one variable's imputed gap values. Of 45,003 channel-pair and forecast-origin comparisons, 177 showed a cross-variable response, spanning 9 stations. The receiving variable's forecast remained unchanged when it was predicted alone. This shows that joint modeling allows an imputation change to affect forecasts of an unmodified variable.

Native missing-value handling should also be evaluated on the target forecasting task. We compared Toto's native mask input with linear imputation using paired windows whose last 32 steps were observed in every input channel, as required by the interface. Relative to linear imputation, the native-mask path's pooled RMSE was 0.21% lower on household power and 1.72% higher on the single-station Beijing data. These paths can therefore be treated as candidates and selected using historical forecasting performance.

For value-dependent missingness, the variance lower bound is determined by the available observations and the requirement that a completion preserve them. It does not require independence between missingness and values. The observation-based constraint thus applies under different missingness mechanisms, using the actual support over which the model computes its statistics.

AuthorsOpenReview ↗

Response to Reviewer DvAk (Part 3 of 3)

3. The actual completions are valid model inputs; the representation controls identify the forecasting effect of originally observed positions.

All three imputers preserve the original observations, so their completed histories can be supplied directly to the models. The internal control replaces only the observed-position representations, leaving missing-position representations and output restoration fixed. This isolates the effect of the observed representations. The resulting mixed representation need not correspond to another raw completion.

The control confirms that changes in these representations affect forecast error even when the raw observations are unchanged. With high-budget SAITS, replacing only the observed representations changed forecast RMSE by more than 1%, in either direction, in 78.88% of cases for TimesFM and 72.25% for Toto after equal weighting across data sources.

The theory establishes an observation-determined lower bound on completion variance and, for positive β\beta under ordinary standardization, bounds the change in normalized input. The experiments establish that representation changes caused by actual imputations propagate to forecast error. Together, they explain why imputation can affect forecasting through positions whose raw observed values it preserves.

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

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

The paper is difficult to read in terms of language and flow. The examples provided in the paper are not easy to follow. I found Figures 1 and 2 to be of little value; instead, they were confusing.

From my reading, the core premise of the paper is that imputation methods, despite yielding the same reconstruction errors, can lead to different prediction errors when used in forecasting, or, more generally, variations in downstream task performance due to the normalization step applied to the imputed data stream.

The authors present two propositions that, to me, seem straightforward. Because I am not familiar with the literature, I have to assume that there is some novelty, although the authors do not explicitly state what is new.

The empirical analysis, on the other hand, presents four findings that, to me, seem detached from the theoretical results. As a result, I was unsure of the overall takeaways from the work.

With regard to the experimental design, I do not think the experiments were actually "designed." Rather, it appeared that a collection of univariate time series datasets and a range of methods were used, but the choices did not seem intentional. The analysis did not appear to be qualitative. I would have considered at least one multivariate case with asynchronous missing data.

In my opinion, the literature review was lacking, and consequently, the paper was not clearly positioned within the existing literature to establish the novelty of the work, in my opinion.

Critical strengths

None

Critical weaknesses

The paper does not clearly establish what is new relative to the existing literature. The two theoretical propositions appear straightforward, and the authors do not explain whether their contribution lies in the propositions themselves, the interpretation of their implications, or the empirical findings. The insufficient literature review compounds this problem.

The paper presents two theoretical propositions and four empirical findings, but their relationship is unclear. The experiments do not appear to be designed to test, validate, or extend the theoretical results. Consequently, in my opinion, the paper lacks a coherent argument connecting its theoretical premise to its empirical conclusions.

The experiments appear to involve applying multiple methods to a collection of univariate time series datasets without a clear rationale for the selection of datasets, methods, or evaluation conditions. In particular, the absence of multivariate experiments with asynchronous missingness limits the investigation of the proposed phenomenon in more complex settings. The concern is not simply the absence of additional experiments, but the lack of an intentional experimental design that isolates and investigates the mechanisms underlying the paper's central claim.

Although the paper identifies four empirical findings, it does not clearly explain how they advance understanding of the relationship between imputation accuracy and downstream predictive performance. Specifically, it remains unclear what generalizable insights the reader should draw from the empirical results taken together with the theoretical results.

The language and logical flow make the paper difficult to follow. The examples are not sufficiently intuitive, and Figures 1 and 2 appear to introduce confusion rather than clarify the underlying concepts. This makes it difficult to assess the theoretical arguments and their significance.

Questions for authors
  1. What are the specific novel contributions of this work relative to the existing literature? The two propositions appear relatively straightforward. What are the nontrivial implications of these results?

  2. How do the four empirical findings relate to the two theoretical propositions? Are the experiments intended to validate the propositions, investigate their implications, or establish independent empirical observations? The connection between the theoretical and empirical components of the paper is currently unclear.

  3. What motivated the selection of datasets, imputation methods, and forecasting models? Were the experiments designed to investigate specific hypotheses arising from the theoretical analysis? If so, what are those hypotheses, and how does each experiment address them?

  4. Why is the empirical evaluation restricted to univariate time series? Would the conclusions hold for multivariate time series with asynchronous missingness across variables?

  5. To what extent are the observed differences in downstream prediction performance specifically attributable to the normalization step? Can the authors isolate the effect of normalization from other factors that may influence forecasting performance?

  6. What are the main takeaways from the theoretical and empirical results considered together? In particular, do the findings provide actionable guidance for selecting imputation methods for downstream forecasting tasks, or are they primarily intended to demonstrate limitations of reconstruction error as an evaluation metric?

8 comments
AuthorsOpenReview ↗

Response to Reviewer eDFE (Part 1 of 6)

Thank you for your specific questions about the contribution, the connection between theory and experiments, and the figures. We added controls using actual imputation errors, experiments with natural missingness and joint multivariate forecasting, and method-selection experiments with larger training budgets. These results directly test the proposed error-propagation mechanism.

1. We derive a constraint computable before imputation and use it to explain how imputation errors can affect forecasting through originally observed positions.

Imputation is commonly evaluated by reconstruction metrics such as MAE and MSE. A smaller distance between imputed and true values does not necessarily make the completed history more useful for a downstream task. Wang et al. (NeurIPS 2024), in Task-Oriented Time Series Imputation Evaluation via Generalized Representers, address this distinction by estimating downstream gains and combining imputation strategies. RevIN removes and restores instance statistics to address distribution shift, while Ahmed et al. compare normalization methods and training-loss scales in time-series foundation models. We hold the forecaster's parameters fixed and study how different completions preserving the same observations change the representations received by the model, and which observed-data conditions constrain those changes. This provides an observation-based constraint and a mechanism analysis of imputation's effect on forecasting.

Within the history used to compute normalization statistics, the observed variance multiplied by the observed fraction is a lower bound on the variance of every completion preserving those observations. Let β\beta be the square root of this bound. When β>0\beta>0, the difference after ordinary standardization is at most 1/β1/\beta times the difference between the raw completions. With at least two missing positions, feasible completion differences can approach this factor arbitrarily closely. A larger β\beta imposes a stronger restriction; a smaller β\beta permits greater amplification. Before imputing, we can therefore quantify how strongly the available observations constrain changes in the normalized input.

The propositions also identify the relevant unit of analysis: the actual history over which the model computes its statistics. The same physical gap can have different observed-variance constraints under different statistical supports. This connects the imputer, normalization settings, and forecaster, and motivates the controls that fix statistics, change their support, or isolate coordinate groups.

AuthorsOpenReview ↗

Response to Reviewer eDFE (Part 2 of 6)

2. The propositions constrain input changes; the experiments test whether those changes affect forecasting.

Imputation changes only missing positions. Normalization, however, computes statistics using the completed history, so the imputed values also change the numerical representations of observed values supplied to the model. This links the propositions to the experiments. We first measure the normalized-input changes, then hold the other components fixed and alter only the representations at originally observed positions.

We used 30% contiguous missingness on 19 series, obtaining 897 masked histories, and ran 128-step forecasts from the actual outputs of Linear, high-budget SAITS, and MOMENT. For each completed history, we replaced only the observed-position representations with their counterparts from the complete history, leaving missing-position representations and output restoration unchanged. The table reports the proportions of cases, after equal weighting across data sources, in which forecast RMSE increased or decreased by more than 1% relative to the forecast before replacement:

ImputerTimesFMToto
Linear73.96%66.67%
SAITS 20×78.88%72.25%
MOMENT81.58%79.65%

Although the raw observations remain unchanged, the normalization-induced changes in their representations affect forecasting. This also occurs when the statistical support contains observations and has β>0\beta>0. For SAITS, among the 775 histories whose first cumulative support containing an actual imputation error has positive β\beta, the corresponding proportions remain 79.15% for TimesFM and 72.92% for Toto.

These experiments test whether the propagation mechanism identified by the propositions occurs in practice and measure its forecasting consequences.

AuthorsOpenReview ↗

Response to Reviewer eDFE (Part 3 of 6)

3. The datasets and method families support the comparison between reconstruction and forecasting; the controls isolate the reasons for their differences.

The 19 series come from 10 sources spanning energy, environmental measurements, computing workloads, and traffic, providing coverage across several domains. The 19 imputers include simple rules, statistical methods, and learned methods. We measure reconstruction and forecast errors on the same histories to test whether more accurate reconstruction also gives more accurate forecasts.

TimesFM and Sundial provide the forecasting comparison. TimesFM and Toto allow separate intervention on internal statistics and representations, making them suitable for the mechanism controls. Each control addresses a specific question: keeping pointwise error magnitudes fixed while changing the error pattern tests whether reconstruction metrics are sufficient; keeping the gap fixed while changing the history used for normalization tests the statistical support; replacing only observed-position representations tests their effect on forecasting; and increasing training budgets tests the concern about undertrained learned imputers.

For the example and figures, consider the history [0, ?, 2][0,\,?,\,2]. Filling the gap with 1 gives the first observation, 0, a standardized value of approximately −1.225-1.225. Filling it with 2 changes the same observation's standardized value to approximately −1.414-1.414. Its raw value is unchanged, but the model receives a different number. Figures 1 and 2 explain this change and its subsequent effects.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.