acceptodds
Under review as a conference paper at ICLR 2027

CoSTAR: Common-State Adaptive Refinement for Multivariate Time Series Forecasting

Abstract

Multivariate time-series forecasting is central to applications such as energy management, traffic prediction, and weather analysis, where channels exhibit both shared and channel-specific dynamics. Existing forecasters capture these dependencies through their internal architectures; however, incorporating an additional refinement signal commonly requires altering the backbone. To address this limitation, we propose CoSTAR (Common-State Adaptive Refinement), an external module that operates only through a forecaster's input and output interfaces. CoSTAR factorizes the history into a common component and a residual input, summarizes the common trajectory with four window-level statistics, and decodes a gated residual forecast correction. We evaluate CoSTAR on eight long-term benchmarks and four PEMS traffic datasets. Across paired comparisons of five backbones on six datasets, CoSTAR reduces horizon-averaged MSE in 29 of 30 backbone–dataset pairs, with one tie at the reported precision; the largest reduction is 72.4% for SparseTSF on PEMS04. Ablations, five-seed analyses, and frozen-backbone experiments further characterize the contribution and scope of the design.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, time series data, deep 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 3 reviews, above 27% of the others.
Reviewer_6BqaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes CoSTAR, a model-agnostic external module designed to improve multivariate time series forecasting without altering the underlying backbone architecture. CoSTAR operates in three stages: (1) Low-Rank State Factorization, which extracts a low-rank common component from the input and passes the residual to the backbone; (2) Statistical Context Encoding, which summarizes the latent common trajectory using four window-level statistics (mean, standard deviation, final state, and endpoint difference); and (3) Adaptive Residual Correction, which decodes this context into a gated residual forecast added to the backbone's prediction. Evaluated on 8 long-term and 4 short-term benchmarks, CoSTAR demonstrates consistent improvements, reducing horizon-averaged MSE in 29 out of 30 backbone-dataset pairs when integrated with 5 different forecasters.

Critical strengths
  1. CoSTAR operates purely via the input/output interfaces of the backbone forecaster, completely avoiding the need to access or modify hidden representations. This ensures high modularity and ease of integration with diverse architectures

  2. The module achieves broad and consistent accuracy gains. Across 30 backbone-dataset pairings, it improves horizon-averaged MSE in 29 cases, with reductions up to 72.4% for SparseTSF on PEMS04.

  3. The added computational cost scales linearly, keeping the parameter and memory overhead remarkably low. Plug-in efficiency diagnostics prove that CoSTAR introduces minimal inference latency compared to baseline networks.

  4. The framework elegantly supports scenarios where the pre-trained backbone is entirely frozen, enabling efficient task-specific adaptation. The frozen-backbone configuration successfully recovers a significant fraction of the joint-training performance improvements.

Critical weaknesses
  1. The choice of the four specific window-level statistics (mean, standard deviation, final state, endpoint difference) is highly heuristic. There is no theoretical grounding explaining why higher-order moments (e.g., skewness) or frequency-domain statistics are ignored, which limits the mathematical rigor of the context encoding.

  2. While interpreted as a "common state" extraction, the external module closely resembles an auxiliary linear residual forecaster. The ablation study does not completely dispel the possibility that the gains simply come from an ensembling effect, rather than actual cross-channel denoising.

  3. The Low-Rank State Factorization relies on strictly linear channel projections independently applied at each time step. This linear bottleneck might be insufficient to disentangle highly non-linear, complex cross-channel dependencies in real-world multivariate series.

  4. The paper promotes CoSTAR as an adapter module but only evaluates it through joint training or frozen-backbone training on the same target task. It lacks a zero-shot cross-dataset evaluation (training CoSTAR on dataset A and testing it frozen on dataset B), which is a standard benchmark for modern adaptive modules.

  5. Some recent plug-and-play methods for multivariate time series forecasting are not discussed in the paper, such as: [1] Wang Z, Jin G. Don’t Just Extract, Regulate: Contextual Residual Exchange for Non-stationary Multivariate Stock Forecasting[C]//Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2026: 253-270.

Questions for authors
  1. Could the authors run an ablation comparing the chosen four statistics (mean, standard deviation, final state, endpoint difference) against alternatives (e.g., max/min pooling, median, or higher-order moments) to empirically justify this heuristic selection?

  2. Provide further discussion or a specific experimental baseline to isolate the "common state" extraction from a naive residual linear ensemble. For instance, what happens if XcommonX_{common} is replaced with purely uncoupled linear projections per channel without a low-rank bottleneck rr?

  3. Given the positioning as an adapter module, providing a zero-shot cross-dataset evaluation (e.g., training CoSTAR+Backbone on Weather and inferring on Solar) would strongly bolster the claims about the robustness of the learned common state representation.

Reviewer_UN3cOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes CoSTAR, a plug-in module that refines any multivariate forecaster through its input and output only

  1. A learned low-rank projection across channels extracts a "common" component
  2. Part of this component is subtracted from the input before it reaches the backbone
  3. Four statistics of the component (mean, std, last value, endpoint change) are decoded by small MLPs into a correction. A learned gate scales the correction before it is added to the backbone's forecast

Attached to five backbones, CoSTAR is reported to lower MSE in 29 of 30 backbone–dataset pairs (up to 72.4%), with competitive results on eight long-term benchmarks and four PEMS datasets, plus ablations, frozen-backbone, sensitivity, seed, and efficiency studies

Critical strengths
  1. Simple, backbone-agnostic module. CoSTAR works only through the forecaster's input and output, so it can wrap any model. It can be trained jointly with the backbone, or with the backbone frozen while keeping most of the gain (Table 7). This makes it practical for refining existing models
  2. Extensive experiments. The paper reports:
    • results on 12 datasets
    • a paired comparison of 5 backbones on 6 datasets
    • component ablations on 4 backbones
    • frozen-backbone adaptation
    • sensitivity to rank, gates, and lookback length
    • five-seed runs
    • efficiency measurements
    • full horizon-wise results in the appendix

This breadth makes the claims easy to check

Critical weaknesses
  1. Evaluation and baselines are unreliable.
    • The evaluation does not test the plug-in claim.
      • CoSTAR is claimed as a plug-in module, but Tables 1–2 rank backbone+CoSTAR against standalone models
      • Other plug-ins are compared only on PEMS (GTR, HN-MVTS) or in a single efficiency plot (CCM, LIFT). There is no plug-in comparison on the long-term benchmarks
      • Even there, several of these reproductions do worse than their own backbones (HN-MVTS 1.002 vs DLinear alone 0.385 on PEMS04, DLinear+CCM 1.247 vs DLinear alone 0.466 on PEMS07), so these reproductions look broken
    • The baselines are unreliable.
      • Some paired baselines are worse than the same model elsewhere in the paper. For example, PatchTST/Electricity scores 0.212 in Table 12 but 0.190 in Table 1, and PatchTST+CoSTAR scores 0.203
      • iTransformer on PEMS is worse than its published results (0.152 / 0.120 / 0.176 / 0.196 vs the published 0.113 / 0.111 / 0.101 / 0.150). Even with CoSTAR, it is worse than the published plain model on PEMS03 and PEMS08
      • The 72.4% headline gain is on SparseTSF, which is worse than DLinear on every PEMS dataset
      • Some models (xPatch, ...) are only used in short-term forecasting experiment (Table 2), and not used in long-term forecasting experiment (Table 1), although they were designed for long-term forecasting
    • Inconsistent reporting.
      • In Tables 1–2, backbone model differs (VPNet and iTransformer) with no motivation for such decision
      • Training curves don't match the tables. In Figure 8, iTransformer stays above 1.15 at every epoch on PEMS07 (H = 96), but Table 12 reports 0.322
      • In Figure 10, the DLinear/Electricity curves stay below the baseline values given in the tables
      • Five-seed means don't match the single runs: 0.121 ± 0.003 vs 0.078 (PatchTST+CoSTAR, PEMS07, H = 24)
      • No baseline variance or significance tests are reported
  2. Gains are not attributed to the design, and novelty is limited.
    • Gains are small over strong backbones.
      • Since CoSTAR is an add-on, the decoder adds 2rHN parameters: ≈1M on PEMS04 at H = 96 and ≈20M on Traffic at H = 720, vs ≈19k initial parameters for DLinear (H = 96). The paper never reports parameter counts
      • CoSTAR has the best MSE on only 1 of 8 long-term datasets
    • Limited novelty.
      • The common/specific split is a learned cross-channel factor model
      • There is no control with matched parameter count
      • SOFTS, the closest prior work, is not compared
Questions for authors
  1. Which training settings were used for each backbone? A.4 says all runs use lr = 1e-4, but it also says the original authors' recommended settings were used. Why is iTransformer below its published PEMS results, and how was SparseTSF set up for 5-minute data?
  2. Why does the CoSTAR backbone change between Tables 1 and 2, and how were the baseline sets chosen?
  3. Which runs produced Tables 12 and 14 and Figs. 8–10? Are the entries single runs, best-of-k, or seed means? Please reconcile the mismatches above
  4. How were CCM, HN-MVTS and GTR configured, given that they underperform their own backbones by up to 2.7x?
  5. Please report parameter counts and FLOPs with and without CoSTAR. Please also explain the zero memory overhead and the negative latency overheads in Tables 22–23
  6. A plug-in comparison would help this work: the same backbones and datasets with +CCM, +LIFT, +GTR, +HN-MVTS and +CoSTAR
Reviewer_5sxTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • CoSTAR is an external module for multivariate time series forecasting that refines a backbone’s inputs and outputs without modifying its internal architecture.
  • Low-Rank State Factorization constructs a common component and subtracts a learned fraction from the backbone input. Statistical Context Encoding summarizes the latent common trajectory using its mean, standard deviation, final state, and endpoint difference, and Adaptive Residual Correction generates a gated forecast correction.
  • The paper evaluates multiple backbones on forecasting benchmarks and examines performance and applicability through component ablations, five-seed analyses, and frozen-backbone experiments.
Critical strengths
  1. CoSTAR operates solely through input and output interfaces, making it compatible with different backbones. Improvements with frozen backbones further support its utility as an external refinement module.
  2. Paired comparisons across multiple backbones and datasets, together with component ablations, provide empirical support for the method and demonstrate improvements across diverse settings.
Critical weaknesses

1. The comparisons do not sufficiently isolate the benefits of the proposed design: Without a correction MLP matched in additional parameter count, it is difficult to distinguish gains from increased capacity from those attributable to the proposed common-state refinement design. The paper also lacks a discussion of, and direct comparison with, CycleNet’s Residual Cycle Forecasting (RCF) [1], which removes a learned periodic component from the input and restores the corresponding future periodic component at the output.

2. Some design choices lack direct validation: The paper does not directly compare reprojecting the reconstructed common component with using the initial latent representation. Moreover, the two learned global scalars determine a single output correction scale through the product of their sigmoid-transformed values, but no ablation establishes an advantage over a single scalar in accuracy or training stability.

3. The contribution of each statistic remains unclear: Although different summarization methods are compared, no ablation isolates the contribution of each statistic, despite the statement in Section 3.3 that these contributions are evaluated.


Reference

[1] LIN, Shengsheng, et al. Cyclenet: Enhancing time series forecasting through modeling periodic patterns. Advances in Neural Information Processing Systems, 2024, 37: 106315-106345.

Questions for authors
  1. Could the authors compare CoSTAR with a correction MLP matched in additional parameter count and with RCF under the same backbone and training protocol? These comparisons would help clarify the additional benefits of the proposed common-state refinement design.
  2. What is the benefit of reprojection over directly using the initial latent representation? With the initial effective correction scale matched, do two learned global scalars improve accuracy or training stability over a single scalar?
  3. Could the authors provide leave-one-statistic-out ablations to establish the contribution of each of the four statistics?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.