acceptodds
Withdrawn from ICLR 2027

RAP: Reconstruction-Augmented Predictors for Robust Time-series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-term time-series forecasting (LTSF) supports applications in energy, traffic, and healthcare, yet state-of-the-art models remain brittle when inputs are corrupted by sensor or preprocessing noise. To resolve this problem, we introduce a reconstruction-augmented predictor (RAP) that improves robustness without corrupting or filtering training inputs by coupling the standard forecasting loss with a self-reconstruction objective on the history window. A single encoder feeds two linear heads—reconstruction and prediction—so the shared representation is explicitly regularized to preserve fine-grained, multi-scale structure. For small input noise, we relate the predictor’s Jacobian energy to noisy forecasting risk and bound the residual Jacobian of a fitted full-sequence map under explicit distributional and curvature assumptions. Controlled ablations measure the effect of reconstruction on the fitted predictor. Across representative LTSF benchmarks in high-dimensional datasets, RAP remains competitive with strong baselines in clean forecasting and degrades more gracefully under synthetic corruptions, and its advantage grows with the length of the observed context. Our measurement confirms that the reconstruction term lowers the predictor's Jacobian energy, and controlled comparisons against Gaussian-noise training, low-pass filtering and an explicit Jacobian penalty on the same backbone isolate the effect of the objective.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Robustness, Context-aware learning

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

This paper proposes RAP, a reconstruction-augmented predictor (RAP) that improves the performance of forecasting without classic data augmentation approaches such as spectral filtering or Gaussian noise. In practice, the forecaster is trained with two designated heads: a reconstruction head and a prediction head. This approach, which is generally simpler and motivated by theoretical analysis with respect to some data augmentation approaches, is shown to perform better than other prominent forecasters utilizing a transformer.

Critical strengths
  • Framing reconstruction as an alternative to augmentation, with theoretical backing, is an interesting and practically relevant approach.
  • Many claims are motivated by theory.
  • Overall, I find the paper was well structured and clearly written.
Critical weaknesses
  • Claims and analysis refer to augmentation on the lookback of the time series only: Many of the claims in secs 2.1-2.2, as well as the toy study, perturb or filter the lookback X while keeping Y clean. Generally, forecasting augmentations ([1],[2],[3]) instead transform the concatenated lookback + horizon window. This keeps the input and target pair consistent and from the same augmented distribution. (I attached here a handful of works but several more works support my observation)

[1] FrAug: Frequency Domain Augmentation for Time Series Forecasting.

[2] Chronos: Learning the Language of Time Series. (see ts mixup)

[3] Towards Diverse and Coherent Augmentation for Time-Series Forecasting.

  • Limited Novelty of reconstruction: Joint reconstruction and prediction is well established; NBEATS [4] uses backcasting, which is the same as reconstruction. PatchTST, MOMENT [5,6] have shown that reconstruction parts of the input improve forecasting accuracy. However, these approaches and elements are not fully discussed or compared in this paper. The aforementioned works are prominent works in the TSF literature; I also believe more works utilize this approach

[4] N-BEATS: Neural basis expansion analysis for interpretable time series forecasting

[5] MOMENT: A Family of Open Time-series Foundation Models

[6] A Time Series is Worth 64 Words: Long-term Forecasting with Transformers

  • Incomplete benchmarks. The evaluation section omits prominent datasets from the LTSF benchmark, namely weather, ETT, and PEMS08. Which may raise questions on the robustness of the proposed approach

  • Outdated Baselines. The newest baseline is TimeXer (Neurips 2024). The TSF forecasting literature is a fast-paced research area, while it is indeed difficult to be up to date with all models, the paper would benefit from a more recent representation of 1 or 2 models in my opinion.

  • The general claims are model-agnostic, yet the approach is only tested DLinear, iTransformer and the proposed transformer. A richer and more focused comparison would strengthen the claims of the work.

  • The work makes strong claims against augmentation, while this work only addresses spectral filtering and Gaussian noise as potential augmentation methods; however, the TSF data augmentation literature has richer and more sophisticated approaches to offer.

Questions for authors

see above

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

This paper proposes RAP, a reconstruction-augmented framework for robust time-series forecasting. It jointly optimizes forecasting and historical reconstruction through a shared encoder to improve robustness without training-time noise augmentation. The authors provide theoretical analyses relating reconstruction error, Jacobian sensitivity, and noisy forecasting risk. Experiments on six multivariate forecasting datasets demonstrate competitive clean-data performance and improved robustness against Gaussian, spike, and quantization noise, supported by ablation studies and sensitivity analyses.

Critical strengths
  1. Practical robustness improvements: RAP demonstrates improved resilience to input corruption across several forecasting benchmarks while maintaining competitive clean-data performance, addressing a relevant challenge in real-world deployment.

  2. Controlled empirical validation: Comparisons between reconstruction-augmented and prediction-only training on identical architectures help isolate the contribution of the auxiliary objective. Additional Jacobian sensitivity measurements and theoretical analysis provide supporting evidence for the proposed robustness mechanism.

Critical weaknesses

Critical Weaknesses

  1. Limited methodological novelty: Joint reconstruction and forecasting objectives have already been extensively explored in prior work (Laptev et al., 2018; Liu et al., 2025; Zhao et al., 2026). RAP largely reuses this established multi-task learning strategy without introducing a sufficiently distinctive optimization mechanism. Its incremental contribution appears limited for ICLR.

  2. Insufficient theoretical and empirical justification: The proposed Jacobian bound does not establish that reconstruction necessarily reduces forecasting sensitivity. Moreover, robustness improvements are inconsistent across corruption types, and multi-seed evaluation under noisy conditions remains insufficient. These limitations weaken the claimed robustness guarantees.

References

[1] Laptev, N., Yu, J., & Rajagopal, R. (2018). Reconstruction and Regression Loss for Time-Series Transfer Learning. SIGKDD MiLeTS Workshop.

[2] Liu, K., Duan, Z., & Chen, C. (2025). Paeformer: Patch-Wise Representation Learning with Autoencoder for Multivariate Time Series Forecasting. ECAI 2025. https://doi.org/10.3233/FAIA251135

[3] Zhao, Y., Xie, Z., Wang, Z., et al. (2026). RecMamba: Reconstruction-Augmented Dual-Path Mamba for Time Series Forecasting. Knowledge-Based Systems, 331, 114801. https://doi.org/10.1016/j.knosys.2025.114801

Questions for authors

Novelty: How does RAP fundamentally differ from existing joint reconstruction–forecasting methods, particularly Paeformer (2025) and RecMamba (2026)? What is the specific methodological contribution beyond applying an established auxiliary reconstruction objective?

Theoretical justification: Theorem 4.3 bounds Jacobian sensitivity using the full-sequence reconstruction and prediction loss, but does not establish that adding reconstruction necessarily reduces forecasting sensitivity. Can the authors provide a direct theoretical comparison between reconstruction-augmented and prediction-only training?

Robustness and reproducibility: Why are the robustness gains inconsistent across corruption types and reconstruction weights? Can the authors provide multi-seed results with statistical significance tests under noisy conditions to establish whether the reported improvements are reliable?

Reviewer_5f3DOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Reconstruction-Augmented Predictor (RAP) to improve the robustness of long-term time-series forecasting through an auxiliary reconstruction objective. Theoretical analysis and extensive experiments demonstrate that RAP reduces prediction sensitivity to input noise while maintaining competitive forecasting accuracy, particularly with longer historical contexts.

Critical strengths
  1. Extensive experiments demonstrate the effectiveness of RAP.
  2. The proposed method is supported by theoretical analysis.
  3. Analyzing robustness from the perspective of Jacobian sensitivity is an interesting approach.
Critical weaknesses
  1. The technical novelty of this paper is limited. The core idea consists of combining a shared encoder, a reconstruction head, and a forecasting head, which essentially follows the well-established paradigm of multi-task learning with auxiliary reconstruction supervision.
  2. The paper lacks direct comparisons with existing models specifically designed for robust time-series forecasting [1,2]. Most of the selected baselines are time-series forecasting models rather than robustness-oriented approaches.
  3. The authors assume that noise primarily affects the input features, while overlooking the possibility of noise in the target labels. This assumption may not hold in real-world applications, where both input observations and target labels can be corrupted.
  4. The tasks and challenges described by the authors are closely related to those investigated in prior work [1]. However, the paper fails to adequately discuss the limitations of this existing work or clarify the specific advantages of the proposed method over it.
  5. Although the paper primarily focuses on robustness under noisy conditions, a substantial portion of the main experimental evaluation is conducted on clean datasets. This creates a disconnect between the stated motivation and the experimental validation, weakening the overall coherence and justification of the study. A more comprehensive evaluation under common missing-data scenarios, such as random missingness and block missingness, would be more appropriate and convincing.
  6. The motivation of the proposed method is not sufficiently clear, and the overall writing and presentation need substantial improvement. Furthermore, all the compared baselines were proposed in or before 2024, raising concerns about the timeliness and competitiveness of the experimental evaluation. When compared with more recent state-of-the-art methods, the reported performance improvements appear marginal.

[1] SEER,ICML2026

[2] Merlin,KDD2025

Questions for authors

See weakness

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

Related papers

Discussion (0)

Sign in to comment.