acceptodds
Under review as a conference paper at ICLR 2027

Do Time Series Foundation Models Use Imputed Values?

Abstract

Forecasting pipelines impute missing histories for time series foundation models. Do better reconstructions improve forecasts? The answer depends on both computational access and learned use. We probe eighteen checkpoints from eleven families through their missingness interfaces, paths combining fills with observation indicators. Ten of the eighteen checkpoints block position-specific fill content on their declared paths. In Chronos-Bolt, masking retains dependence on the context mean and standard deviation but removes this content, explaining why even exact fills recover little forecast accuracy under scattered and block missingness. More generally, an interface bound characterises an imputer’s access to the forecast, which is distinct from learned use in paired pretraining experiments. Gap exposure can improve accuracy without responding to fill quality, whereas repair-aware training makes retained content useful. Our retrofit passes fills with observation flags and learns across gap mechanisms and repair qualities. After 5,000 continued-pretraining steps, it reduces zero- and linear-fill degradation relative to complete-context error across three families, with little change on complete histories. On nine GIFT-Eval subsets with recorded gaps, the retrofit lowers median MASE by up to 18% in high-missingness windows. Our code is available at https://anonymous.4open.science/r/hole-in-input-repro-2E97/README.md.

Primary area:
foundation or frontier models, including LLMs
Keywords:
time series foundation models, time series, missing values

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vNvxOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether external imputations reach a time series foundation model and whether the model has learned to use them. It characterizes fill access through surviving preprocessing statistics, audits eighteen checkpoints from eleven families, and separates interface access from training exposure using paired experiments. A retrofit retains fills with observation flags and adapts released models for 5,000 steps. Across three model families, it reduces degradation under zero and linear fills; on recorded GIFT-Eval gaps, improvements concentrate in high-missingness windows. The main contribution is the distinction between access to repaired values, robustness to gaps, and learned use of repairs.

Critical strengths
  1. The computational diagnosis is precise and experimentally supported. The analysis distinguishes position-specific content from surviving normalization statistics, with a rank-at-most-two restriction for the explicit-mask Chronos-Bolt path. Jacobian, permutation, and oracle-fill interventions test this mechanism directly. This provides an actionable explanation for why better imputation can fail to improve forecasting.
  2. The within-family controls make the access-versus-learning distinction convincing. The within-family experiments use paired training seeds and show that augmentation can improve robustness while fill response remains flat. The three-family retrofit and recorded-gap evaluation extend the finding beyond one model, while the reported negative results help delimit the practical benefit.
Critical weaknesses
  1. The practical comparison does not yet establish when the retrofit should replace native missing-value handling. The recorded-gap evaluation emphasizes linear-fill comparisons, while the input-convention audit reports NativeNaN winning 26 of 36 declaration cells. The recorded-gap evaluation also reports a 5% worse pooled median MASE for Chronos-Bolt. A paired recorded-gap comparison against validation-selected native conventions is needed to substantiate a deployment recommendation.
  2. The adapter transfer summary uses evaluation-dependent selection. The transfer summary reports the better normalization variant in each cell, and the closure plot takes a pointwise envelope across variants. This does not measure one frozen adapter configuration on unseen datasets. Report both variants separately and a selection rule fixed before held-out evaluation, including paired absolute errors alongside own-clean closure.
Questions for authors
  1. On the same recorded-gap windows used in the reported evaluation, how does each retrofit compare with its native NaN path or a convention selected using separate validation data? Please report per-dataset changes and paired uncertainty, especially for the two high-missingness strata.
  2. What are the held-out results for each normalization variant separately, and was either variant selected without examining held-out forecast errors? Please distinguish this fixed-choice result from the best-variant closure reported in the transfer summary.
1 comment
AuthorsOpenReview ↗

Kind reminder before the end of the discussion period

Dear Reviewer vNvx,

We believe the native-handling comparisons, validation-based input selection, fixed-normalization results, and absolute-error analyses have addressed your concerns. We also revised our deployment recommendation to clarify when native handling is preferable.

If you agree that your concerns have been addressed, we would be grateful if you would consider updating your review and rating; if any issues remain unresolved, please let us know before the discussion period closes so that we can address them promptly.

Reviewer_Ymc4OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper examines whether better imputation improves time series foundation model forecasts. It distinguishes whether input interfaces preserve imputed values from whether models learn to use them, supported by an interface bound and controlled experiments. The proposed retrofit retains fills with observation flags and trains across missingness patterns and repair qualities. Experiments on synthetic and recorded missingness show reduced forecasting degradation across three model families, including up to 18% lower median MASE on high-missingness GIFT-Eval windows.

Critical strengths
  1. The paper distinguishes access to imputed values from the learned ability to use them, explaining why even perfect imputation may fail to improve forecasts.
  2. Matched-window interventions and paired training controls help isolate interface and training effects. Evaluation across multiple model families and both synthetic and recorded missingness strengthens the findings’ practical relevance.
Critical weaknesses

Limited diversity in benchmark-level adapter training. The adapter is trained only on ETTh1, ETTm1, and Weather, which have relatively few channels. Electricity and Traffic are representative benchmarks for high-dimensional multivariate time series, with 321 and 862 channels, respectively, yet they are reserved for transfer evaluation using only 24 channels each. The paper does not justify this selection or examine whether broader training coverage changes the findings.

Questions for authors

Could you justify why benchmark-level adapter training focuses on ETT and Weather, which have relatively few channels? Or could you also provide results using the representative high-dimensional benchmarks such as Electricity and Traffic?

2 comments
AuthorsOpenReview ↗

Kind reminder before the end of the discussion period

Dear Reviewer Ymc4,

We believe our full-channel Electricity and Traffic evaluations have addressed your concern about channel diversity and subsampling. The full-channel results remain consistent with the original findings, and the additional absolute-error analyses further support them.

If you agree that your concern has been addressed, we would be grateful if you would consider updating your review and rating; if any issues remain unresolved, please let us know before the discussion period closes so that we can address them promptly.

Thank you for your response. I found the original experimental evaluation thorough, and your clarifications and additional results have addressed my main concern. I have no further questions and will maintain my original score.

Reviewer_wG6xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether time series foundation models effectively use imputed values for forecasting. The authors analyze 18 checkpoints from 11 families, distinguish computational access from learned use, and establish a theoretical bound on fill access. They propose a repair-aware retrofit and validate its effectiveness on synthetic and real-world missing data, showing improvements particularly under high missingness.

Critical strengths

S1: The paper identifies an important issue in TSFMs by distinguishing computational access to imputed values from their learned use. S2: The experiments are comprehensive, covering multiple model families, missingness mechanisms, real-world datasets, and training ablations.

Critical weaknesses

W1: Although Section 4.5 includes controlled pretraining experiments, the final CPT evaluation lacks a baseline with the same missingness augmentation but the original content-masking interface. Since missingness augmentation alone can improve forecasting, the additional benefit of restoring fill access remains unclear. W2: The released code and results show that block masks can overlap, resulting in an actual missing rate of about 0.51 when the nominal rate is 0.7, compared with 0.7 for other mechanisms. This inconsistency with the definition of r affects the fairness of cross-mechanism comparisons, such as Figure 3.

Questions for authors

Q1: Appendix D.7.5 identifies a decoder attention mask that restricts fully missing patches. Does the final Chronos-Bolt retrofit address this restriction, allowing their imputed values to affect forecasts beyond normalization statistics? Q2: How does the retrofit compare with 5,000-step missingness-augmented CPT that retains the original content-masking interface under otherwise identical settings? Q3: Do the cross-mechanism conclusions involving block missingness (e.g., Figure 3) still hold when the actual missing rates are matched? Q4: Can the authors evaluate the final CPT models with realistic learned imputers (e.g., SAITS and BRITS) across multiple datasets to demonstrate that the proposed method improves forecasting beyond oracle or synthetic blended fills?

1 comment
AuthorsOpenReview ↗

Kind reminder before the end of the discussion period

Dear Reviewer wG6x,

We believe our matched CPT, exact-missingness controls, decoder analysis, and SAITS/BRITS evaluations have addressed your concerns. These results are included in the revised PDF.

If you agree that your concerns have been addressed, we would be grateful if you would consider updating your review and rating. If any issues remain unresolved, please let us know before the discussion period closes so that we can address them promptly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.