acceptodds
Under review as a conference paper at ICLR 2027

FSC: Bridging Periodic Patterns and Local Dynamics in Multivariate Time Series Forecasting

Abstract

Multivariate time series forecasting is widely used in domains such as traffic, energy, and finance, where effective forecasting requires capturing both stable periodic patterns and rapidly evolving local dynamics. However, under distribution shifts, periodic priors learned from historical data can become mismatched with fast local dynamics, leading to prediction lag. To address this issue, we introduce Fast-to-Slow Correction (FSC), a lightweight and plug-and-play framework that can be seamlessly integrated into existing forecasting models. FSC explicitly decouples fast and slow features via parallel pathways, and employs a learnable refinement module to correct slow periodic features using fast local dynamics as adaptive references. This design effectively mitigates prediction lag in periodic modeling without modifying the backbone architecture. Extensive experiments on 12 real-world datasets show that FSC achieves state-of-the-art results on 9 datasets, with an average 2.51% MSE reduction over strong periodic forecasting methods. FSC also achieves these improvements with minimal computational overhead, introducing only 1.26M additional parameters.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Periodic Modeling, Distribution Shifts, Fast-to-Slow Correction

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hFHsOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

FSC addresses the mismatch between learned periodic priors and current local dynamics in multivariate time-series forecasting. It constructs separate periodic and input-dependent representations, then uses their difference to generate a residual correction to the periodic representation. Adaptive channel encoding and an orthogonality regularizer support this design. The paper evaluates FSC on 12 benchmarks, including Traffic and four PEMS datasets, reports improvements when incorporated into three forecasting backbones, and provides ablations, repeated-run results, efficiency measurements, and a bias–noise analysis.

Critical strengths
  1. A useful and interpretable approach to adapting periodic priors. Conditioning a learned periodic representation on current observations addresses a practical weakness of static periodic templates, especially when the observation window contains only part of a cycle. The residual correction provides a simple way to retain historical structure while adapting to local changes. Appendix D offers a meaningful bias–noise rationale for balancing a stable but biased prior against a more current but noisy signal.

  2. Broad evidence of practical utility across datasets and architectures. The evaluation includes high-dimensional energy and traffic datasets, multiple horizons, and FSC-augmented iTransformer, PatchTST, and DLinear models. The displayed errors indicate useful improvements across these backbone–dataset combinations even after correcting the percentage calculations discussed below. Four-seed results on three datasets and measured training costs provide further support.

Critical weaknesses
  1. The temporal-lag interpretation is stronger than the direct evidence. A difference between learned representations can reflect general prediction mismatch, including amplitude or level changes, without identifying a temporal delay. Correlations involving learned templates and independently projected features do not by themselves establish train–test distribution shift or reduced timing error. The bias–noise analysis supports idealized correction under shared-state and error-correlation assumptions, but does not establish that the learned representations satisfy those assumptions.

  2. The necessity of the specific correction design remains uncertain. The additive alternative matches FSC's reported average MSE on Weather and PEMS08 and is close on the other ablation datasets. Removing the orthogonality loss also leaves average MSE unchanged at the displayed precision on three of four datasets. These results support the usefulness of the overall framework more clearly than the necessity of subtraction-based lag modeling or orthogonal decoupling. Repeated-run comparisons of these close controls would strengthen the claimed attribution.

  3. The plug-in evaluation needs a more precise specification. The paper does not sufficiently explain where FSC is attached to each backbone, how training and tuning budgets are matched, or whether the gains persist against a similarly sized periodic-feature fusion module. Complete optimization settings are also missing.

Questions for authors
  1. What evidence distinguishes correction of temporal delay from general residual adaptation? How are the learned periodic templates constrained or calibrated so that their Pearson correlation with observations can be interpreted as distribution mismatch?

  2. Are the differences between full FSC, additive fusion, and removal of the orthogonality loss consistent across repeated runs with matched capacity and training budgets?

Reviewer_u9ooOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FSC retrieves a periodic template for the current window, encodes the template and the observed series with separate linear maps, and adds a correction derived from their difference. The main results evaluate this construction as a standalone forecaster on eight long-horizon datasets and four PEMS datasets. Its use as a module on existing backbones is reported separately.

Critical strengths
  • The correction is specified explicitly. The lag is the difference between the fast and slow features, refined by a two-layer MLP, and added back to the slow representation through a linear map.
  • The evaluation follows the standard long-horizon protocol, with input length 96 and both MSE and MAE. On Electricity and the ETT datasets, FSC improves over the reported periodic baselines.
  • Section 4.4 evaluates FSC as an added module on iTransformer, PatchTST, and DLinear, and reports the memory overhead together with the change in MSE.
  • The ablation compares subtraction with addition, removes the orthogonal penalty, and substitutes an alternative temporal encoding, so the effect of each design choice can be measured.
Critical weaknesses
  • Equations 3–7 define a linear forecasting head on the corrected features. The plug-and-play setting is the backbone study in Figure 3, whereas the main tables evaluate the standalone model. The abstract presents the former as the proposed method.
  • Several gains are small, and some comparisons reverse. On Traffic, FSC obtains an MSE of 0.462, against 0.409 for SOFTS. In Table 4, Re-Time reaches 0.416 MSE on ETTh1, slightly below FSC at 0.417, and on Weather several variants match FSC at 0.238. These results do not support the claim that removing a component consistently degrades accuracy.
  • The discussion connecting Pearson correlation to the reported gains still contains unresolved placeholder text, including “[Dataset 2, e.g., Solar]” and “[Value, e.g., -0.056]”. The same passage refers to the model as FSR. In its current form, this analysis cannot be reviewed.
Questions for authors
  • Please give the exact definition of the 2.51% average MSE reduction, including which baselines are used and whether datasets on which FSC is worse, such as Traffic, are included.
  • Please report the period used for each dataset and state whether it was selected on the test split.
  • Please replace the placeholder discussion and report whether a lower Pearson correlation corresponds to a larger MSE reduction on each dataset.
Reviewer_BRAjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Fast-to-Slow Correction (FSC), a lightweight module intended to correct periodic representations when recent dynamics diverge from historical periodic priors. FSC creates a slow feature from a CycleNet-style periodic embedding and a fast feature from the current input augmented with Adaptive Channel Encoding. Their difference is refined by a residual MLP, projected into a correction, and added to the slow feature; an orthogonality penalty encourages complementary information. Experiments cover eight long-horizon and four PEMS short-horizon datasets, with best reported performance on nine of 12 datasets and a 2.51% average MSE improvement over strong periodic methods. Ablations examine the correction, channel encoding, refinement, orthogonal loss, and alternative lag constructions. Plug-in studies add FSC to three backbones, while additional analyses report four-seed robustness on three datasets, representation correlations, and computational cost.

Critical strengths
  1. The correction idea is simple: compare recent input features with periodic features and use their difference to update the prediction. The equations make the main computation easy to follow.
  2. The paper evaluates 12 datasets and tests several alternatives to its correction and channel encoding. These comparisons provide useful evidence about where the full model helps.
  3. Adding FSC improves three different backbones on the selected datasets. Memory measurements and four-seed results on three datasets add practical context to the accuracy results.
Critical weaknesses
  1. The main claim concerns changing data patterns, but the experiments do not vary the size or type of change in a controlled way. Standard forecasting results support accuracy gains without establishing why those gains occur.
  2. Low Pearson correlation does not by itself establish time lag or distribution shift. Likewise, lower correlation between learned features does not show that they represent separate fast and slow processes. The mechanism explanation therefore goes beyond the evidence.
  3. FSC requires a predefined period, yet the experiments do not establish how it behaves when that period is incorrect or changes over time. This is a key limitation for a method motivated by unreliable periodic patterns.
Questions for authors

See Weakness

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.