acceptodds
Withdrawn from ICLR 2027

Downsampling Time Series via Auto-Regression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High-resolution time series are essential for granular analysis but often prohibitively expensive to process. While low-resolution representations offer a pragmatic alternative, conventional direct downsampling acts as a destructive projection that irreversibly disrupts latent dynamical structures, leading to significant performance collapse in downstream tasks. We challenge this long-standing hegemony by proposing that the transition to lower resolutions should be an active, distribution-aware synthesis rather than a passive, deterministic pruning. To realize this paradigm shift, we introduce VARes, a framework that reformulates downsampling as a hierarchical auto-regressive process. Empirical results demonstrate that data obtained with VARes consistently outperforms traditionally downsampled counterparts by amplifying task-critical features and effectively denoising signals. Our work establishes a new downsampling standard, proving that our method yields data that is better suited for machine learning models on downstream time series tasks than conventional downsampling approaches.

Primary area:
learning on time series and dynamical systems
Keywords:
Downsampling, Auto-Regression

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

This paper argues that direct time-series downsampling is a destructive projection that disrupts latent dynamics and collapses downstream performance. The paper proposes VARes, an auto-regressive generative framework that treats downsampling as active, distribution-aware synthesis. Experiments show VARes-synthesized data consistently outperforms traditional downsampling baselines.

Critical strengths

Even though concerns are raised, the experiment design is relatively sound, involving quantitative analysis and qualitative analysis of a variety of tasks and aspects.

Critical weaknesses
  1. Terrible writing.
    (a) The title does not clearly represent the method this paper proposes.
    (b) There is little introduction to the method in the Abstract and Section 1.
    (c) It says the proposed method is "inspired by the Visual AutoRegressive (VAR)", but VAR is discussed neither in the Related Work nor in Appendix I.
    (d) Inconsistency in section numbering. Section 3 only has 3.1 and jumps directly to Section 4.
    (e) Much of the section is spent on motivation and failed design choices rather than a formal definition of the proposed method.

  2. I do not reckon the problem the author is trying to address is a real challenge. The author mentioned that "The explosion of high-resolution (HR) time series data has created a paradoxical challenge in modern AI: while HR sequences offer granular insights, their immense computational footprint and storage requirements often render them impractical for large-scale deployment or long-term historical analysis". Compared to the data used in LM or VLM pretraining, typical time series data are not that large-scale. I do think the high-frequency time series data are too large to be used directly.

  3. The experiment cannot fully demonstrate the effectiveness of the proposed method.
    (a) The datasets used in the experiments are far from "large-scale", which do not meet the problem the author raised at the beginning of this paper. The high-frequency time series in these datasets can be used directly without the need for downsampling.
    (b) The three backbone models used are all attention-/transformer-based models. Other types of models should be considered.

  4. The novelty of this paper does not meet the ICLR standard. It is more likely to be considered a naive combination of existing methods.

Questions for authors
  1. The method is tailored for "1D temporal data"/"univariate time series". But most of the datasets used in the experiment are multi-variate. What's the rationale for treating each channel independently and then transforming it to 2-D, but not using multi-variate time series directly as 2-D?
  2. Will the zero-padding in Delay Embedding introduce unnecessary jumps?
  3. Please refer to Weaknesses for other questions.
Reviewer_Xz3UOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the limitations of conventional time series downsampling, arguing that deterministic sampling and filtering methods may destroy temporal dynamics and degrade downstream task performance. To address this issue, the authors propose VARes, a generative downsampling framework that combines delay-embedded manifold mapping, a multi-scale VQVAE, and scale-wise autoregressive generation to synthesize low-resolution time series. Experiments across multiple datasets and downstream tasks report substantial improvements over conventional downsampling methods, while spectral and structural analyses are provided to support the claimed preservation of temporal dynamics.

Critical strengths
  1. Practically relevant research problem. The paper addresses the trade-off between reducing the computational and storage costs of high-resolution time series and preserving information important for downstream tasks. Exploring alternatives to conventional deterministic downsampling is a meaningful research direction.

  2. An interesting generative perspective on downsampling. The proposed framework introduces multi-scale autoregressive generation into time series resolution reduction. The combination of delay embedding, channel-independent processing, and hierarchical generation offers an alternative to traditional signal-processing-based downsampling.

  3. Broad downstream evaluation. The paper evaluates several downsampling strategies across different datasets and forecasting/imputation architectures. Additional spectral, structural fidelity, and computational analyses provide potentially useful perspectives on the behavior and practical trade-offs of the proposed framework.

Critical weaknesses
  1. Conceptual ambiguity in the definition and objective of generative downsampling. The paper argues that conventional downsampling destroys temporal information and proposes distribution-aware generation as an alternative. However, generating a statistically plausible low-resolution sequence is fundamentally different from preserving information associated with a particular high-resolution input. Even if the generated sequence exhibits similar spectral or distributional properties, it may alter the timing, phase, or local dynamics of important events in the original sequence. The paper does not sufficiently distinguish generative synthesis, denoising, task-aware representation learning, and downsampling, making it unclear what properties the proposed method is specifically intended to preserve.

  2. Significant concerns about the fairness and clarity of the downstream evaluation protocol. VARes benefits from a separately trained generative model, whereas most competing approaches are deterministic sampling or filtering operators with no learned parameters. More importantly, the paper reports exceptionally large, sometimes orders-of-magnitude improvements in downstream MSE without sufficiently clarifying the temporal alignment between generated sequences and prediction targets, or the exact information accessible during the generation process. For forecasting and imputation, it is essential to establish that the generative model does not access future prediction targets or ground-truth values at masked positions. Although the reported results do not establish that information leakage has occurred, the current experimental description is insufficient to rule out such concerns or to fully assess whether the methods are evaluated under comparable information constraints.

  3. Insufficient comparisons with learning-based alternatives and inadequate ablation of the core architecture. The main comparisons focus on conventional operators such as uniform sampling, weighted moving averages, wavelet transforms, and Butterworth filtering. Since these methods cannot leverage learned distributional priors, their comparison with a separately trained generative model does not establish the necessity or superiority of the proposed autoregressive architecture. Stronger learning-based baselines, such as autoencoder-based compression, multi-scale VQVAE reconstruction, or non-autoregressive generative downsampling, are needed. Moreover, the paper does not sufficiently isolate the contributions of delay embedding, multi-scale quantization, and scale-wise autoregression, leaving the source of the observed improvements unclear.

  4. Insufficient evidence for the claimed preservation of dynamical structures and high-frequency information. The authors interpret increased high-frequency spectral energy as evidence that VARes preserves or amplifies task-critical temporal dynamics. However, stronger high-frequency components do not necessarily imply better preservation of the original signal and may instead reflect generated artifacts or unrealistic fluctuations. Although the paper reports Pearson correlation, DTW, and PSD cosine similarity, some lower-resolution configurations exhibit relatively modest correlations with the original signals. These observations do not fully support the stronger claims that underlying dynamical invariants are preserved. Additional analyses of spectral phase consistency, local event alignment, and trajectory-specific structural fidelity would be necessary.

  5. The computational overhead is not adequately justified relative to the original motivation. The paper motivates downsampling partly by the high computational and storage costs of high-resolution time series, yet VARes requires a relatively expensive two-stage training procedure and autoregressive generation. Although the appendix reports training and inference costs, the paper does not provide a comprehensive end-to-end comparison against directly training downstream models on high-resolution data or using inexpensive conventional downsampling. Without such an analysis, it remains unclear whether the proposed framework provides a favorable overall efficiency–accuracy trade-off in practical applications.

  6. Several claims appear stronger than what the experiments directly establish. The paper presents VARes as a new downsampling standard and attributes its downstream improvements to the preservation of latent dynamical invariants. However, the current evidence primarily establishes that the complete framework performs better than selected deterministic baselines under the reported evaluation settings. The experiments do not sufficiently isolate the proposed causal mechanism or establish its general applicability. The claims should therefore be moderated and more closely aligned with the available empirical evidence.

Questions for authors

Q1. Evaluation protocol and information accessibility. Could the authors clarify exactly which portions of the original high-resolution sequences are accessible to VARes when generating low-resolution inputs for forecasting and imputation? Are the physical timestamps, input coverage, and prediction targets identical across all downsampling methods? A detailed explanation of the train/validation/test split and the masking procedure would help rule out potential information leakage and establish the comparability of the reported results.

Q2. Necessity of autoregressive generation. Could the authors provide comparisons with learning-based downsampling or compression methods under comparable training data and computational budgets? In particular, how does VARes compare with using the multi-scale VQVAE alone, a conventional autoencoder, or a non-autoregressive multi-scale generator? These comparisons are important for understanding whether the improvements specifically originate from scale-wise autoregressive generation.

Q3. Preservation of input-specific information. How do the authors distinguish distributional plausibility from the preservation of information in a particular input sequence? Can the authors provide evaluations of event alignment, spectral phase consistency, and trajectory-specific reconstruction quality, together with comparisons against the original high-resolution data under a consistent downstream evaluation protocol?

Reviewer_bRy6OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes VARes, which reformulates the time series data down-sampling problem as a distribution-aware synthesis process, instead of conventional fixed down-sampling methods. VARes adapts visual autoregressive modeling to time series data through a combination of delay embedding, channel flattening, multi-scale VQVAE, and scale-wise autoregressive generation. Conducted experiments across 4 datasets and 3 backbone time series models, aiming to show that synthesized low resolution sequences preserve information and patterns required by various downstream tasks for better modeling results.

Critical strengths

S1. Potentially meaningful research direction.

  • This paper explores a new paradigm for time series downsampling problem, reformulating it from deterministic filtering methods to distribution-aware synthesis process. It is potentially interesting in that it learns task specific low resolution representations, improving efficiency while preserving critical data patterns.
Critical weaknesses

W1. Insufficient novelty and design justifications.

  • The paper introduces VARes, adapting existing visual autoregressive modeling method to time series data, but does not sufficiently establish the methodological contribution specific to downsampling process. The mapping between latent scales and temporal resolution remains unclear, as well as the missing of key component ablation studies. In particular, the necessity of autoregressive generation over simply learned reconstruction is not illustrated.

W2. Unclear evaluation protocols and insufficiently comparisons.

  • The paper does not provide clear evaluation and experimental settings. Specifically, Figure 1 does not distinguish the impact of downsampling from changes in data observation availability, weakening the motivation. More importantly, forecasting and imputation experiments in Section 4 do not sufficiently establish comparability of different data construction strategies, particularly in terms of data availability, evaluation targets, and temporal alignment. All these issues make it difficult to assess whether the reported improvements stem from better information preservation.

W3. Limited evidence for structural preservation and comparative advantages.

  • The paper attributes its improvements to preserving dynamical structures and high frequency dependencies, but does not clearly define which temporal properties are preserved or provide direct evidence supporting these claims. Moreover, the comparisons focus mainly on deterministic downsampling methods, without evaluating learning-based downsampling model alternatives.
Questions for authors

Q1. Methodological clarity.

  • How are the intermediate latent scales mapped to the target temporal resolutions in VARes? Please clarify the actual downsampling procedure and explain why autoregressive generation is necessary beyond simpler learned reconstruction.

Q2. Design validation.

  • What are the individual contributions of delay embedding, channel-independent flattening, and autoregressive generation? Please provide component-level ablations to justify these design choices.

Q3. Experimental validity and comparative evaluation.

  • How are data observation availability, temporal alignment, and evaluation targets controlled across different downsampling strategies? In particular, how does Figure 1 distinguish the effects of downsampling from reduced observation availability, and why are learning-based alternatives and more recent forecasting models not included in the main comparisons?

Q4. Presentation and result writing issues.

  • Figure 2 (p. 3) is not referenced in the main text, and its caption lacks a meaningful explanation of the workflow.
  • Table 1 (p. 4) introduces DS, PS, and C-FID without defining these metrics at first use.
  • Section 4.2 (p. 7) incorrectly claims a 607.5% reduction in forecasting error, which is mathematically impossible for non-negative MSE values.

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

Related papers

Discussion (0)

Sign in to comment.