acceptodds
Under review as a conference paper at ICLR 2027

Same Error, Different Consequences: Information Interfaces between Deterministic Completion and Stochastic Time-Series Generation

Abstract

Deterministic completion is often used as an intermediate step before stochastic generation for irregular time series, and its quality is typically assessed by pointwise reconstruction errors such as MAE or RMSE. However, two completions with the same reconstruction error need not be equally useful as conditions for a downstream generator. This raises a basic question: is completion accuracy sufficient to characterize downstream conditioning utility? We study this question through the information interface between deterministic completion and stochastic generation. To separate how much a completion is wrong from how it is wrong, we construct controlled interventions that match aggregate missing-region RMSE while changing the temporal organization of completion errors. We then measure how these interventions alter the relative utility of transmitting the full completion versus a coarse temporal representation to the generator. Across four time-series datasets and two downstream generator architectures, we find that aggregate completion RMSE alone is insufficient to predict downstream behavior. Reorganizing learned completion errors with RMS matching can substantially change generation quality and the relative preference between full and coarse interfaces. Importantly, the direction of this effect is not universal: sensor datasets and a prospectively preregistered financial-domain test exhibit different, and in some cases opposite, interface shifts. These results show that completion accuracy and downstream conditioning utility are distinct quantities, and suggest that the information interface between deterministic completion and stochastic generation should be treated as a modeling problem in its own right rather than as a passive handoff of the completed sequence.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series imputation, irregular time series, stochastic time-series generation, diffusion models, missing data, conditioning interfaces, completion error, downstream utility

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_2MrhOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies whether deterministic completion accuracy, measured by aggregate missing-region RMSE, is sufficient to characterize the usefulness of a completion as a condition for a downstream stochastic time-series generator. The authors compare two interfaces: Full, which passes the dense completion directly to the generator, and Avg16, which applies a width-16 temporal moving average. They construct residual interventions by temporally shuffling completion residuals, reapplying the observation mask, and rescaling the residuals to match the original pooled missing-region RMS. The downstream effect is measured through the change in the Full-versus-Avg16 CRPS gap.

Critical strengths

The paper identifies a real limitation of evaluating an intermediate completion only through pointwise reconstruction metrics. In a completion-to-generation pipeline, a completion may be close to the target while still being difficult or misleading for a finite downstream model to use.

Critical weaknesses

The paper structure is disorganized. For example, the motivation, problem formulation, and organization of evidence are not sufficiently connected in intro. After briefly introducing the completion-to-generation setting, the introduction moves directly to the Full/Avg16 interfaces and their comparison statistics. The discussion of how the closest prior work handles conditioning, what its evaluations leave unresolved, and how the proposed comparisons address that gap appears only afterward and remains brief. Consequently, when these definitions are introduced, readers have not yet been given sufficient motivation for why comparing dense and smoothed conditions is an appropriate way to answer the research question.

The paper also needs to distinguish more clearly among three claims: aggregate RMSE is insufficient to characterize downstream utility; the temporal assignment of residuals affects interface utility; and the direction of this effect depends on the domain. These claims require different levels of experimental control. In particular, the in-mask permutation in Appendix A.2.5 preserves MAE and per-window RMSE, and the negative mean shifts observed on Stock under the primary intervention become positive under this stricter control. This result materially affects the interpretation of the prominently discussed Stock reversal, yet it is not adequately integrated into the main argument. I recommend establishing the prior findings and specific research gap before introducing the experimental statistics, explaining how the formulation and interventions address that gap, and organizing the results around the evidence supporting each claim.

Questions for authors
  1. What is the primary claim of the paper?
  2. Why is the in-mask control not included in the main paper, given that it materially changes the Stock result?
Reviewer_YAPJOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether deterministic time-series completion accuracy is sufficient to characterize its utility for downstream stochastic generation. The authors compare two interfaces: the full completed sequence and a temporally smoothed representation. They introduce residual interventions that preserve aggregate completion RMSE while modifying the temporal organization of errors. The approach is evaluated across several time-series datasets using Convolution and Transformer based diffusion generators. The results show that similar completion RMSE can lead to different downstream generation quality, with varying effects across datasets and architectures. The paper further examines these differences through additional residual controls and sensitivity analyses.

Critical strengths
  • The paper studies the relationship between deterministic completion accuracy and downstream stochastic generation quality, highlighting limitations of evaluating completion solely through reconstruction error.
  • The experimental setup includes RMSE-matched residual interventions, multiple generator architectures, and seed-wise evaluations to investigate the effect of completion errors on downstream generation.
  • The authors report contrasting results across datasets and include additional controls to examine the sensitivity of their findings.
Critical weaknesses
  • The primary intervention preserves aggregate RMSE but changes other error statistics, including MAE and per-window errors. This makes it difficult to isolate the contribution of temporal error organization.
  • The experiments are limited to one completer architecture and primarily a single missingness pattern. Furthermore, the paper does not provide a practical method for selecting an appropriate conditioning interface based on its findings.
  • Appendix shows the learned completer performs substantially worse than linear interpolation on the stock dataset. This raises questions about whether the observed cross-domain differences are influenced by the poor quality of the completion model.
Questions for authors
  1. Since the stock-domain reversal disappears under the MAE-preserving intervention, how much of the observed effect can be attributed specifically to temporal error organization?
  2. Would the cross-domain findings persist with a stronger completion model, particularly on the stock dataset where linear interpolation substantially outperforms the learned completer?
  3. Given that the observed interface shifts do not consistently change which interface performs better, what practical conclusions can be drawn about selecting Full versus Avg16 conditioning?
Reviewer_bFZKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether pointwise completion accuracy (e.g., RMSE) is enough to judge how useful a deterministic completion is as a condition for a downstream stochastic generator. It tests this with interventions that keep aggregate RMSE fixed while reorganizing completion errors in time.

Critical strengths
  1. good problem defintion, the detemernistic completion and stochastic generation should be decoupled to study.

  2. good results on 4 experiments

  3. Careful experimental hygiene. Completion pairs are frozen and hash-checked, which avoids mask/completion mismatch.

Critical weaknesses
  1. The primary intervention leaks ground truth, which likely explains most of the effect. The primary intervention shuffles the full 96-step residual and then remasks it. With a contiguous 48-step block, about half of the missing positions receive a residual that came from an observed position, and that residual is exactly zero. The rest of the missing residuals are scaled up by about √2. The paper’s own numbers confirm this almost exactly: The rescaling factor is 1.4129 ≈ √2 (A.3, step 3). Shuffled MAE is 0.2931 ≈ 0.4152 × 0.5 × √2 = 0.2936. So the “time-shuffled” completion equals the ground truth at roughly 50% of missing coordinates, scattered at random, and is noisier at the rest. This changes the information content of the condition, not just the temporal organization of its error.

The authors should make the in-mask permutation the primary intervention, add an order-preserving “in-place sparsification” control and report the fraction of exact-zero residuals to separate temporal-order effects from ground-truth leakage, and estimate a null distribution of S from identical-condition runs with different generator seeds.

  1. The downstream generators may be undertrained, which weakens the finite-learner interpretation Under the aligned/Full condition, the generator’s sample-mean MAE is worse than simply returning the completion: ETTh1: 0.624 versus 0.414 (Table 19 against Table 5). Weather96: about 0.53 versus 0.19. Stock: 4.40 versus 1.90. A deterministic point forecast at the completion would have CRPS equal to its MAE, about 0.41 on ETTh1. That beats the generator’s CRPS of about 0.50. Training is 5,120 steps, which takes 3 to 9 minutes. This must get fixed.
Questions for authors

I think the underlying setting of the paper is not so fair to study the main focus of this paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.