acceptodds
Withdrawn from ICLR 2027

Zero-Shot Time-Series Reconstruction from Statistical Summary

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Retaining raw time-series data is often impractical under resource constraints, leaving only coarse statistical summaries available. However, when raw observations are required for downstream analysis, reconstructing fine-grained time series from summaries alone remains inherently limited. Meanwhile, deep generative and predictive models can synthesize realistic temporal patterns, but suffer from severe distribution shift when applied to unseen time series, as they are only capable of interpolating from learned data. This study proposes zero-shot time-series reconstruction from statistical summary with generative and predictive models. Statistical summary can arise from many situations where the full retention of time series is impractical, but a quick summary is feasible. We augment existing generative and predictive models with the statistical summary to enable zero-shot time-series generation. Experimental results show that the proposed method effectively reconstructs original data from statistical summary with both unseen time-range and unseen time-series generations. This allows to train a generator/predictor only once with a generic time series and use the same generator/predictor for the reconstruction of all other datasets from their statistical summaries alone.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, reconstruction, zero-shot learning, extrapolation, statistical summary

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VFbUOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies time-series reconstruction when raw measurements have been discarded but blockwise statistical summaries remain available. It combines existing generative or predictive models with either mean–variance summaries or IDEALEM-based representations. Summary-derived information is supplied during inference, followed by a merging procedure that assigns summary-derived values according to the ranks of the model outputs. Experiments consider reconstruction of later portions of a series and transfer from an ETTh1 source dataset to other datasets.

Critical strengths
  • This is a practically motivated problem. Reconstructing approximate sequences from retained summaries is relevant when storing raw measurements is impractical. The paper explicitly uses information that remains available under these constraints.

  • The approach is a simple and reusable procedure. The rank-based merging operation can be applied to several model families without redesigning their architectures. The evaluation includes multiple backbones, two summary types, and a relevant summary-only baseline.

Critical weaknesses
  • The assumption supporting transfer of temporal ordering is neither clearly stated nor sufficiently validated. The merging operation discards the model output’s numerical values and retains its ordering. Consequently, the learned component is useful only insofar as it produces an ordering informative about the target sequence. It is unclear why a model trained on one domain should provide such an ordering in another domain with potentially different dynamics. Matching the target block’s distribution does not correct misplaced peaks, transitions, or oscillations. The paper does not characterize when this assumption holds or provide direct evidence that the transferred ordering captures target-specific dynamics. Appendix C identifies conditions for exact reconstruction when both the ordering and sorted values are correct, but does not establish when a source-trained model will satisfy those conditions. Using ETTh1 as the sole source also provides limited evidence for the broader claims concerning generic source datasets. This distinction matters because summaries for every 24-point block already reveal local levels and variability throughout the target sequence. Low pointwise reconstruction error can therefore reflect information retained in the summaries rather than successful transfer of dynamics. The contribution of learned ordering should be directly isolated and evaluated.

  • Another concern is that the results do not establish a sufficiently reliable benefit over reconstruction from summaries alone. The paper includes a no-model baseline in Table 7. However, that baseline is competitive with, and sometimes better than, the full method. For Gold Price with mean–variance summaries, summary-only MSE is 112.85, compared with 119.84 and 129.81 for the proposed LLMTime and TTM configurations in extrapolation. In zero-shot reconstruction, LLMTime achieves 112.68, almost identical to the baseline, while TTM achieves 120.01. For bike rental with IDEALEM, summary-only MSE is 30,457, outperforming all listed full-method zero-shot results in Table 6. Some configurations do improve over summary-only reconstruction. Nevertheless, the paper does not adequately explain this variation or establish how a beneficial backbone would be selected when raw target data are unavailable. The large improvements over vanilla models primarily establish the usefulness of providing target summaries; they do not isolate the contribution of the learned ordering. Moreover, the gold-price vanilla-merging result in Table 14 is numerically better than the best full-method result, raising questions about the necessity of summary conditioning before merging.

Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

Thank you for the feedback

  1. We ask that the proposed methodology is viewed from a different perspective: What would you do to reconstruct target time series when you’re only given with statistical summary that was not originally meant to recover original time series? We only expect the predicted/generated temporal pattern to be similar to the original pattern at most. For summary statistics, unless variance is zero, we cannot guarantee an identical recovery of the same value. For IDEALEM-style summary, unless similarity threshold by Kolmogorov—Smirnov test enforces zero distributional distance for exchangeability, we cannot guarantee an identical recovery either. We note that even forecasting algorithms such as LLMTime and TTM, which claim to be zero shot forecasting, cannot predict well on their own in our setting of extrapolation and zero shot learning, as shown in Tbls. 1 and 4.

  2. We selected ETTh1 since their choice as zero-shot baseline yielded best overall results. If we have to use diferent sources every time for the reconstuction of different dataset, it wouldn’t be called zero shot anymore.

  3. Tbl. 7, along with Tbls. 2, 3, 5, and 6 , establish the contribution of learned temporal patterns by generators/predictors. Here we are not championing a specific backbone nor tired to one specific backbone. While there are some exceptions, our results overall outperform statistical summary only resutls. Bike Rental dataset, as discussed in Sects. 4.2 through 5, is where generators/predictors have difficulty to match the temporal patterns of recurring peaks and valleys.

  4. We note that vanilla-merging sometimes outperforms full method presented in Tbls. 2, 3, 5, and 6. And we feel the need for improving the current learning of temporal patterns by backbones.

Reviewer_DM6vOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates zero-shot time series reconstruction from statistical summaries when the original time series observations are no longer available. The proposed framework combines existing generative or predictive models with target-series statistical information, followed by a rank-based merging procedure. Specifically, the method uses model-generated sequences to provide temporal ordering and statistical summaries to provide the marginal values of each reconstructed block. Experiments are conducted on several datasets under extrapolation and cross-dataset zero-shot reconstruction settings, using several generative and forecasting backbones. The reported results show substantial improvements in reconstruction accuracy when target statistical summaries and rank-based merging are incorporated into existing models.

Critical strengths
  1. Reconstructing temporal data when only compressed statistical summaries are retained is an interesting and potentially useful problem.
  2. The proposed procedure can be applied to different generative and predictive architectures without redesigning their training objectives.
  3. The method is easy to follow and reproduce
Critical weaknesses
  1. Concerns regarding the claimed zero-shot generation setting. I have some concerns regarding the paper's claimed zero-shot generation. The proposed method receives block-wise statistical information computed directly from the time-series segment being reconstructed. In the extrapolation setting, this includes statistics from the future target interval, which are unavailable to conventional forecasting models, and it creates a substantially different information setting from standard forecasting or zero-shot generation. Consequently, the substantial performance improvements over the original models cannot necessarily be interpreted as evidence of improved generalization to unseen temporal dynamics.

  2. The main merging algorithm preserves the rank ordering of a model-generated sequence while replacing its values with sorted samples obtained from the target statistical summary. This is closely related to established rank-based rearrangement and empirical quantile-mapping procedures, such as Schaake Shuffle [1] and ensemble copula coupling [2], where marginal information and rank-based dependency structures are combined through reordering. Please add to compare

  3. Unclear experimental settings and lack of controlled comparisons. I found that the authors provide coarse experimental settings in the appendix, the evaluation protocol remains unclear and difficult to follow. In particular, the main manuscript lacks a clear description of the experimental configurations, making it challenging to interpret the reported results and assess the validity of the claims. The authors should consider conducting controlled comparisons across different models under consistent historical context lengths and prediction horizons to ensure a fair evaluation. Without a clearly defined and standardized experimental protocol, the reported performance improvements are insufficient to convincingly substantiate the paper's claims.

[1]. Clark, Martyn, et al. "The Schaake shuffle: A method for reconstructing space–time variability in forecasted precipitation and temperature fields." Journal of Hydrometeorology 5.1 (2004): 243-262.

[2]. Schefzik, Roman, Thordis L. Thorarinsdottir, and Tilmann Gneiting. "Uncertainty quantification in complex simulation models using ensemble copula coupling." (2013): 616-640.

Questions for authors

Please refer to the critical weaknesses

1 comment
AuthorsOpenReview ↗

Thank you for the feedback

  1. We call our methodology zero-shot reconstruction, since the trained backbones are not tied to target time series. We are not improving the backbone models’ generalization performance, but providing a way to reconstruct target time series from given statistical summary alone.

  2. We reviewed both papers to see if there are any similarity to our merging algorithm, though we developed ours independently. The ensemble copula coupling paper [2] is not even remotely close to our algorithm. They used rank-based ordering simply to compute a joint CDF (cumulative distribution function) from sampling from marginal ditribution. The Schaake shuffle [1] on the other hand is conceptually similar. However, they are not pursuing temporal continuity within a block like we do; rather they are randomly selecting historical observations and matching the order of ensemble to make more sense of them. We will include the reference and discuss it in our paper.

  3. The suggested consistent historical context lengths and prediction horizons to ensure a fair evaluation would make sense if we were proposing a new time-series forecasting algorithm. However, we are proposing reconstruction from given statistical summary and our methodology is noninvasive and not tied to a specific backbone, not to mention we employ both generators and predictors. As noted in Appendix D, we follow default settings of backbones whenever possible. Our standardzed experimental protocol is to accurately reconstruct target time series from given summary regardless of backbones.

Reviewer_QV11OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies reconstruction of time series when raw target observations are unavailable but blockwise statistical summaries have been retained. It considers two summary representations: blockwise means and variances, and IDEALEM-based distributional summaries. These summaries are introduced during inference into existing generative or predictive models. A subsequent merging operation transfers the model output’s within-block rank ordering to values obtained from the statistical summary.

Experiments evaluate reconstruction of a held-out temporal segment and cross-dataset reconstruction using ETTh1 as the source dataset. The study covers six model families and seven univariate datasets, with comparisons against unassisted models, summary-only reconstruction, and selected merging ablations. The results demonstrate substantial improvements over models without target summaries, although the incremental benefit over simpler summary-informed alternatives is less consistent.

Critical strengths

The paper addresses a useful reconstruction setting with an accessible implementation. Combining retained summaries with existing models could be valuable when raw observations have been discarded. The merging procedure is simple and can be applied without changing the model architecture.

The experiments include controls that help expose where improvements originate. In addition to multiple model families, the paper reports summary-only reconstruction, block-size sensitivity, repeated runs, and before-merging versus vanilla-merging ablations. These comparisons are consequential because the supplied target summaries already contain substantial reconstruction information.

Critical weaknesses
  1. The incremental contribution of the learned model is not sufficiently established under equal access to target information. The large improvements over vanilla models compare methods that receive target-block summaries with methods that do not. Those results establish the usefulness of the summaries, but do not isolate the proposed model-assisted reconstruction mechanism.

The existing summary-only baseline makes this concern concrete. For Gold Price, its MSE is 112.85 in Table 7, compared with 108.88 for the best summary-statistics-assisted model in Table 5. For Bike Rental with IDEALEM, summary-only reconstruction achieves 30457, outperforming the best complete method’s 33465 in Table 6. Appendix F provides useful vanilla-merging controls, but only for three selected datasets. Broader comparisons against simple summary-informed reconstruction and ordering templates are needed to establish when the learned temporal prior provides a meaningful advantage.

  1. The claims about recovering the original temporal trajectory exceed the demonstrated identifiability and temporal validation. Blockwise means and variances do not determine within-block ordering: many different trajectories share exactly the same summaries. Algorithm 1 obtains that ordering from a model trained on other observations, but the paper does not establish when this ordering should match an unseen target.

Theorem C.1 characterizes exact reconstruction when both the ordering and sorted target values are already correct; it does not show that these conditions follow from the available summaries or cross-dataset training. Consequently, low pointwise error on the selected datasets does not establish general recovery of fine-grained temporal structure. Evaluation of autocorrelation, spectral structure, turning points, and downstream utility would help distinguish recovery of temporal information from correction of local levels and scales.

Questions for authors

Assessment-changing evidence: Can you extend the summary-only and vanilla-merging comparisons to all datasets and include block-mean reconstruction, interpolation between block means, and simple temporal-ordering templates? Please evaluate both reconstruction error and summary consistency to isolate the contribution of the learned model.

Clarification: What assumptions make the target’s temporal ordering recoverable from the retained summaries? Please clarify what Theorem C.1 guarantees about unseen targets, and whether the intended output is an estimate of the original realization or a plausible summary-consistent realization.

Comparison with PINFDiT: Please discuss the relationship to PINFDiT: Energy-Based Physics-Informed Diffusion Transformers for General-purpose Time Series Tasks, particularly its inference-time constraint correction. Where feasible, could you compare rank-based merging against a PINFDiT-based reconstruction baseline under identical target summaries and source/target splits? A statistical-summary energy adaptation should be explicitly identified as an adaptation, and physical constraints should be used only where justified. Reconstruction quality, temporal fidelity, summary consistency, and inference cost would be informative comparison axes.

Numerical clarification: In Table 13, Gold Price under Before Merging/Statistics has MSE 452.33 and MAE 184.79. For the same errors and scale, MAE cannot exceed MSE\sqrt{\mathrm{MSE}}, approximately 21.27 here. Is this a typographical error, or were different normalizations used? Please reconcile the reported metrics.

2 comments
AuthorsOpenReview ↗

Thank you for the feedback

  1. We observe that in general the proposed method shown in Tbls. 2, 3, 5, and 6 outperform summary-only results in Tbl. 7. For the said example (Gold Price), there’s about 3.5% improvement; but in another case of ETTh1 MSE, about 26% improvement. Bike Rental dataset, as discussed in Sects. 4.2 through 5, is where IDEALEM-style summary works better and generators/predictors have difficulty to match the temporal patterns of recurring peaks and valleys.

  2. The problem of exactly predicting temporal ordering from learned generators/predictors is virtually impossible. If it were known already, there would be no more research on time series prediction. The only way to exactly reconstruct temporal pattern & values is to use existing lossless compression techniques for floating point values. It should be noted that those compression techniques require their own encoding method, which is quite different from our scenario where statistical summary is already given.

  3. Thm. C.1 states exact reconstruction condition. In reality, statistical summary and temporal pattern can never be identical to unseen target. And we believe our assumption in this theoretical treatment is in line with most other theorems. As discussed earlier, we only expect the predicted/generated temporal pattern to be similar to the original pattern at most. For summary statistics, unless variance is zero, we cannot guarantee an identical recovery of the same value. For IDEALEM-style summary, unless similarity threshold by Kolmogorov—Smirnov test enforces zero distributional distance for exchangeability, we cannot guarantee an identical recovery either. In short, Thm. C.1 is about an ideal scenario, whereas the performance of proposed method can only be measured by real-world datasets.

  4. Along with pointwise errors in the manuscript, we can provide the following cross correlation results that show the similarity of reconstruction and ground truth. We show the best average results of cross correlation in the following (The closer to 1.0, the better.):

[Best cross correlation results for extrapolation]

Data/TypeVanillaStatisticsIDEALEM
L1MAG-0.00690.92850.9247
L2MAG0.00630.93510.9045
L3MAG0.02060.93470.8977
Gold Price0.00540.99590.9919
LIC0.70070.99270.9853
Bike Rental0.06620.12600.2744
ETTh10.06950.84610.7856

[Best cross correlation results for zero-shot learning]

Data/TypeVanillaStatisticsIDEALEM
L1MAG0.00140.92340.9140
L2MAG0.14780.92440.9023
L3MAG0.08100.92200.8983
Gold Price0.00920.99490.9910
LIC0.60500.99380.9843
Bike Rental0.08450.20550.2728
  1. We reviewed PINFDiT and noticed that while generalist-to-specialist inference is somehow similar at high level, their Langevin correction is completely different from our merging algorithm. They enforce governing PDE during the correction step in an indirect way, which yields an optimizatoin problem including KL divergence. To solve the problem, they suggest sampling from Boltzmann energy distribution.

  2. Tbl. 13 is not a typograhical error; we verified again and confirm that every value reported in the paper is correct. The reason they look inconsistent is we selected best-performing models corresponding to the best results from main tables 1 through 6. For the said example, we look at Tbl. 2, Gold Price, where Informer shows best MSE and DoppelGANger shows best MAE. Informer’s output for Before Merging/Statistics are as follows: MSE 452.33, MAE 14.741, SMAPE 1.9133. DoppelGANger’s output are as follows: MSE 50046, MAE 184.79, SMAPE 20.226.

  3. We ran experiments on block-mean reconstruction and interpolation between block means, and found that these results are comparable to the statistical summary only results in Tbl. 7. We feel the need for improving the current learning of temporal patterns by backbones.

Thanks for your response. Your response has successfully addressed part of my concerns, so the reviewers decided to maintain the current score.

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

Related papers

Discussion (0)

Sign in to comment.