acceptodds
Under review as a conference paper at ICLR 2027

DrafTS: Time-Aware Decomposition with Residual Correction for Time Series Modeling

Abstract

Real-world time series contain evolving underlying dynamics with irregular variations that lack stable temporal patterns and are often referred to as noise. Existing methods address this mixture by filtering frequencies or suppressing noisy observations. They either miss temporal evolution or risk suppressing useful dynamics. We propose DrafTS, a model-agnostic framework that aims to reduce noise while preserving evolving dynamics through time-aware **D**ecomposition with **R**esidu**A**l correction **F**or **T**ime **S**eries. DrafTS uses features derived from instantaneous amplitude and frequency to guide decomposition into a primary component intended to capture underlying dynamics. A task-specific backbone models the primary component, while a lightweight correction module uses residual information to correct the backbone output. Across four time series modeling tasks, DrafTS improves six diverse backbones, demonstrating its effectiveness. Code is at <https://anonymous.4open.science/r/DrafTS-A14F/>.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series Analysis, Deep learning, Representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_dZbMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposed a decomposition-based method for time series modelling. The paper applys multiple decomposition approaches prior to backbone model to extract and expose more feature for the model to dymanically select. In particular, the framework uses Empirical Mode Decomposition (EMD) and the Hilbert transform to break the input down into Intrinsic Mode Functions (IMFs) and extract instantaneous frequency and amplitude. IMF guide a learnable mask. This mask dynamically splits the time series into a primary component designed to capture structural dynamics and a residual component containing noise. Then, A standard, task-specific backbone model processes the primary component to generate a base prediction and a lightweight residual-aware correction (RC) module processes the residual to recover any structural dynamics that the decomposition stage missed. However, the paper presentation is not very clear and hard to follow, notation system is a bit messy. Technical component such as time-aware mask and loss function needs more details.

Critical strengths

Removing noise can also remove useful information, so keeping the residual and using it to improve the prediction is a reasonable idea.

The framework works with existing models. It can be added to different backbones without changing their internal architecture, which makes it potentially useful in practice.

The paper tests several backbones across forecasting, classification, and anomaly detection, and includes ablations to examine the contributions of decomposition and residual correction.

Critical weaknesses
  1. The main idea seems fairly incremental. It combines existing EMD and Hilbert analysis with a simple gate and residual correction. Since the mask is trained through task losses, it is unclear whether it actually separates noise from useful signals or simply divides the input between two models. Comparisons with simpler IMF weights and clearer evidence of noise removal would help.

  2. The correction module is basically a standard reservoir RNN with fixed internal weights and a trainable output layer. Its residual awareness mainly comes from taking the residual as input. The experiments show that this choice can work, but it is still unclear what is new about the module.

  3. The losses need a clearer explanation. In Eq 12, the correction and fused losses have the same value, but the detach operation changes where gradients flow. Please clarify which components each loss updates, especially whether the correction loss also trains the mask.

  4. Table 6 is not clear what "Raw" model is. The context in Qualitative analysis only said backbone.

  5. Over comlicated notatino system. I can see many footnote in equation, and some variable dimensionality is not consistent. Some presentation problems:

  6. The notation changes between sections.The main text uses NN for the number of variables and CC for the number of classes, but the appendix uses CC for variables instead. Keeping this consistent would make the paper easier to follow.

  7. Some window details are unclear. How are the window boundaries rounded, and what happens near the start and end of the sequence? These details should be explained.

  8. The two residuals are easy to mix up. One is what remains after EMD, and the other is what the mask sends to the residual branch. It would help to explain directly how both fit into the final decomposition.

Questions for authors

The time-aware mask uses features with clear signal-processing meanings, but its training objective only rewards better task performance. This does not necessarily teach it to distinguish noise from useful dynamics. can the authors show stronger evidence that the learned mask actually supports the proposed denoising interpretation?

1 comment

Many of my concerns still remain. I will keep my original judgment

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

This paper proposes DrafTS, a model-agnostic framework that uses EMD, Hilbert transform, and a learnable mask to decompose time series into primary and residual components. A backbone models the primary component, while a lightweight reservoir-based module uses the residual to correct predictions.

Critical strengths

The framework is evaluated extensively across four tasks, six backbones, and diverse datasets, demonstrating generally consistent performance improvements.

Critical weaknesses

The paper lacks sufficient justification for its core assumption that EMD, Hilbert transform, and a learnable mask can effectively separate underlying dynamics from noise. There is no guarantee that the primary component captures meaningful dynamics while the residual represents noise. The observed performance gains may instead stem from increased model capacity rather than effective signal-noise decomposition.

Questions for authors

See weaknesses.

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

This work propose the method DrafTS, which is a model-agnostic plugin component that aims to reduce noise while preserving temporal dynamics. It is achieved with dynamic-aware decomposition with residual correction, which aims to both track major component dynamics but also exploits dynamics left in the residual. The proposed method is added to many backbone methods, and evaluated on many datasets to show its robustness.

Critical strengths

Overall, extensive experimental evaluation. The proposed method can be seamlessly integrated into existing approaches, improving performance in most cases while introducing only a small number of parameters that are interpretable.

Critical weaknesses

The main concerns are as follows:

  1. Limited performance gains. The improvements appear modest across many baseline models and datasets.
  2. Modest methodological novelty. The proposed approach builds on well-established ideas from prior work, including FEDformer and TimeMixer. While it offers a meaningful extension, the conceptual novelty appears limited.
  3. Discrepancies in baseline results. Some reported baseline results appear worse than those published in the original papers. For example, TimeMixer reports an average error of 0.240 on Weather, whereas this paper reports 0.283 for the baseline and 0.241 after adding DrafTS. Could the authors clearly document and explain these discrepancies?
Questions for authors

NA

Reviewer_qd9xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

DrafTS splits a time series into a primary component and a residual. It uses empirical mode decomposition and Hilbert features to learn how much of each oscillatory component to retain at each time step. The backbone works on the primary component, and a small reservoir model uses the residual to correct its output. The paper tests the approach with six backbones across short and long term forecasting, classification, and anomaly detection, with additional ablations on noisy Lorenz63 data.

Critical strengths
  1. DrafTS can be added to existing backbones without changing their architecture. The gains are modest in many cases, but they show up across several backbones and tasks.
  2. I found the Raw + RC comparison in Table 17 useful. With TimeFilter, the full method beats the version that feeds the raw input to the correction branch on both MSE and MAE in all six Lorenz63 settings.
Critical weaknesses
  1. How much does the mask gain from varying over time? Table 17 leaves this open. I would compare it with a learned, time-constant weight for each IMF and use the clean Lorenz63 histories to check how much of the underlying signal each version recovers.
  2. I had trouble matching two parts of the method to the linked code. Equation (4) weights the amplitude changes, while the code takes an unweighted local RMS. For M4, the paper describes sMAPE, while the linked correction branch uses squared error. Please clarify which version was used for the reported results.

Update after the author response The authors added the comparisons I asked for and cleared up my questions about Eq. (4) and the M4 loss. I raised presentation to 3.

Questions for authors

No additional questions.

Comments on the paper

1 comment
Area_Chair_7wrNOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions. Thank you all for your time and contributions to a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.