acceptodds
Under review as a conference paper at ICLR 2027

CCo-Patch: Continuous Correction with Temporal Scaffolding for Patch-based Forecasting

Abstract

Patch-based Transformer models such as PatchTST have become a dominant paradigm for long-term time series forecasting, but patchification introduces an inter-patch continuity gap: local patterns inside each patch are explicitly encoded, while the transitions between neighboring patches must be inferred implicitly, which can result in incoherent predictions across patch boundaries. We propose CCo-Patch, a continuous correction framework that augments a patch-based backbone with a continuous correction branch (hereafter, the Correction branch). The Correction branch extracts a causal low-frequency scaffold—a reference sequence extracted from the original input—through FFT-based filtering, and fuses it with the main branch via a learnable coefficient, with patch-wise mean-centering and Huber optimization further improving robustness. Across seven long-term forecasting benchmarks and four prediction horizons, CCo-Patch achieves the best MSE in 17 out of 28 settings, competitive with or better than the ICLR 2026 PMDformer on most datasets and horizons, with clear gains on Weather, ECL, and ETTh2, and comparable performance on ETTm2. Continuity-oriented evaluation and a patch-granularity analysis further suggest that CCo-Patch improves cross-patch continuity and that this advantage is robust across patch lengths. Code is available in the supplementary material.

Primary area:
learning on time series and dynamical systems
Keywords:
Long-term Time Series Forecasting, Patch-based Forecasting, Temporal Continuity, Continuous Correction, PatchTST

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yc7AOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes CCo-Patch, a time series forecasting framework that augments PatchTST with an auxiliary correction branch to address the claimed inter-patch continuity gap. The correction branch extracts a low-frequency temporal scaffold through FFT-based filtering and generates an auxiliary forecast using an MLP, which is combined with the backbone prediction through a learnable coefficient. The framework additionally incorporates patch-wise mean-centering and Huber loss, and is evaluated on seven long-term forecasting benchmarks.

Critical strengths

The paper is well writen and easy to follow.

Critical weaknesses

The central contribution is not established.

The fundamental hypothesis is not validated. The paper assumes that patchification introduces an inter-patch continuity gap, but provides no direct evidence for this claim and whether it will damage forecasting. Patchification does not lead to any information loss, and a strong enough backbone (e.g. PatchTST) typically can restore the continuity structure that patchification may hurt. The authors neither demonstrate that such a continuity problem systematically exists nor establish that existing patch-based backbones fail to capture these relationships.

The methodological novelty is unclear. The proposed method primarily combines existing techniques, including frequency filtering followed by an MLP, Huber loss, and patch-wise mean-centering. More importantly, the so-called continuous correction pathway is entirely independent of the patch structure and contains no mechanism specifically designed to model or enforce inter-patch continuity. The two branches operate in parallel and interact only through additive fusion of their predictions. Consequently, the proposed architecture is essentially an ensemble of two forecasting pathways, rather than a method that explicitly addresses the claimed inter-patch continuity gap.

The empirical improvements are limited and do not substantiate the claimed contribution. Improvements over competitive baselines are often marginal or inconsistent. The improvement over PatchTST is still marginal.

Questions for authors

See Critical Weaknesses.

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

The paper studies long-term time series forecasting with patch-based Transformers such as PatchTST. Its motivation is that patching encodes patterns inside each patch explicitly, while transitions between neighboring patches are only learned implicitly. The authors argue that this can make forecasts less coherent across patch boundaries. To address this, CCo-Patch adds a lightweight correction branch next to the PatchTST backbone. This branch extracts a smooth, low-frequency reference signal from the input with a sliding FFT filter. An MLP maps it to the forecast horizon, and the result is added to the main forecast through a learnable weight. The method also mean-centers each input patch and trains with a Huber loss. The authors evaluate it on standard long-term forecasting benchmarks against recent baselines, and introduce two metrics to measure forecast continuity. The model is simple and easy to follow the idea.

Critical strengths

S1. A good question, studied with transparent experiments. The task people study is clear. The paper asks whether forecasts are temporally coherent, not only accurate point-wise, a property that standard LTSF evaluation rarely examines. The ablation separates the three components and reports unfavorable cases openly, such as the MAE cost of the correction branch alone and the better ETTm2 result without mean-centering. Main results are averaged over five seeds. The paper further includes a δ sensitivity study, an efficiency table and a patch-length sweep.

S2. A simple, lightweight and reproducible add-on. On Weather the method adds 0.13M to 0.21M parameters and 13% to 19% inference time. The correction branch is channel-shared, so its size does not grow with the number of variables. The code is released with one script per dataset, and the δ values and Gaussian centers in the scripts match Table 6. Most ablation variants can be reproduced with three command-line switches.

Critical weaknesses

W1. Limited novelty. The work is an incremental improvement on a mature backbone. Each component has close precedents: patch mean-centering in PMDformer, a smooth trend branch beside a patch branch in xPatch, plug-in branches for PatchTST in CycleNet, and a robust training loss. The paper does not cite or compare with several of these works, so its specific contribution is unclear.

W2. The core claim is weakly grounded. The paper names an inter-patch continuity gap but never defines where it appears in the forecast, since the flatten head outputs the whole horizon at once. There is no derivation or analytical framework that explains why a low-frequency scaffold should close this gap. The mechanism also differs from the narrative. On Weather and ETTh2 the selected filter is band-pass rather than a trend extractor, and removing the FFT filter barely changes the results.

W3. The experiments do not show that the method is effective. The gains over PatchTST are small, and the margins to PMDformer are mostly within 0.005 with no variance reported. The ablation does not isolate the claimed mechanism, and much of the gain could come from adding a second forecaster and a tuned Huber loss. The continuity metrics do not measure patch boundaries, and TRE can improve simply because forecasts are smoother. The evaluation also drops part of the test set, and baselines may not be compared on the same footing.

Questions for authors
  1. Could the authors clarify where the inter-patch continuity gap appears in the forecast? The backbone uses a flatten head that outputs the whole horizon at once, so it is not clear to me which forecast positions correspond to patch boundaries.
  2. I would appreciate more explanation of why a low-frequency scaffold should improve cross-patch continuity. A short analysis or intuition about the mechanism would make the motivation much easier to follow.
  3. For Weather and ETTh2, the selected Gaussian centers seem to suppress most of the low-frequency (DC) component, so the filter acts more like a band-pass filter. Could the authors explain how this fits the trend-based motivation? It would also help to discuss why the variant without the FFT filter performs almost the same.
  4. Could the authors help separate the effect of the scaffold from simply adding a second forecaster? A comparison with PatchTST plus an MLP branch on the raw input, trained with the same Huber loss and tuning, would be very informative.
  5. Could the authors say a bit more about the continuity metrics? With a window stride of 8, IWCE compares points that are not adjacent in time. TRE is about 1 for a constant forecast, so it may also reward smoother forecasts. A boundary-specific measure or a smoothed-baseline control would make these results more convincing.
  6. Could the authors clarify the evaluation setup? In particular: whether all models are evaluated on the same test windows, where the baseline numbers come from, and the standard deviations over the five seeds. A discussion of closely related work such as xPatch, HDMixer, CycleNet and the patch-mean decoupling in PMDformer would also be helpful.
Reviewer_FaHSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies patch-based long-term time series forecasting and focuses on the continuity issue introduced by patchification. The authors argue that patch-based models such as PatchTST encode local patterns within each patch, but transitions between neighboring patches are only learned implicitly. To address this, the paper proposes CCo-Patch, which adds a continuous correction branch to a PatchTST-style backbone. This branch extracts a causal low-frequency temporal scaffold from the input using an FFT-based filter, maps it to the prediction horizon with a lightweight MLP, and fuses it with the main branch through a learnable coefficient. The model also uses patch-wise mean-centering and Huber loss.

Critical strengths

The motivation is reasonable and relevant for patch-based forecasting. Patchification is widely used in recent time series models, but the paper points out a concrete issue that is not always directly evaluated, namely the consistency of transitions across patch boundaries. I found this framing useful, especially because the proposed method does not require changing the patching scheme or replacing the backbone.

The experimental section is fairly complete. The authors compare against several strong baselines, report results across multiple datasets and horizons, and include ablations for the correction branch, mean-centering, Huber loss, and the FFT filter. The continuity analysis with IWCE and TRE is also helpful because it tries to evaluate the main claim of the paper beyond standard MSE and MAE. The gains over PatchTST are fairly consistent, and the method seems lightweight enough to be practical.

Critical weaknesses

The novelty is somewhat incremental. The method combines several familiar ideas, including low-frequency filtering, an auxiliary MLP branch, residual-style fusion, mean-centering, and Huber loss. The paper presents them in a coherent way for patch-based forecasting, but the technical contribution is not very deep. The main reason I lean positive is that the combination is well motivated and empirically useful, rather than because the method is conceptually very new.

Some design choices need stronger justification. The FFT filter uses dataset-specific Gaussian centers selected by grid search, and Huber loss also has dataset-specific tuning. This makes it less clear how much of the improvement comes from the proposed continuity mechanism versus careful per-dataset tuning. The ablation table helps, but I would still like to see a more standardized setting or a clearer discussion of how these hyperparameters should be chosen in practice for a new dataset.

The continuity evaluation is interesting but not yet fully convincing. IWCE and TRE are reasonable attempts to measure temporal consistency, but the connection between these metrics and the actual patch boundary problem could be explained more carefully. For example, TRE measures first-order temporal differences over the reconstructed sequence, which is useful, but it is not specific to patch boundaries. The paper would be stronger if it separated boundary-specific effects from general smoothness or low-frequency bias more explicitly.

Questions for authors

How sensitive is the method to the dataset-specific selection of Gaussian centers in the FFT filter? If the same centers were used across all datasets, or if a simpler fixed low-pass filter were used, would the main improvements remain?

Can the authors provide more detail on the validation protocol for selecting both the Huber parameter and the spectral filter centers? Since these are tuned per dataset, it would be helpful to know whether the tuning budget is comparable to that used for the baselines.

Does the correction branch ever oversmooth predictions in datasets or variables where high-frequency changes are important? The paper briefly mentions this limitation, but a more concrete failure case or analysis would help clarify when the method should not be expected to work well.

Have the authors tried adding the same correction branch to another patch-based backbone besides PatchTST? Since the paper argues that the mechanism is orthogonal to the backbone, even one additional backbone experiment would make this claim much stronger.

Comments on the paper

1 comment

I sincerely appreciate the authors' thoughtful rebuttal and the additional efforts made to address my questions. While the new clarifications are helpful and resolve several of my minor points, I still have reservations regarding the marginal technical novelty of the proposed approach. Given the bar for ICLR conf, I feel the overall contribution to the community would benefit from further development before publication. Therefore, I will maintain my original score, though I encourage the authors to continue refining this promising direction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.