acceptodds
Under review as a conference paper at ICLR 2027

COMPASS: Channel-wise Calibration of Periodic Priors for Multivariate Time Series Forecasting

Abstract

Multivariate time series forecasting requires effectively leveraging both historical periodic patterns and current dynamics. Existing periodic forecasting methods typically employ a uniform strategy to incorporate historical periodic priors across channels, overlooking the channel-wise differences between historical priors and current dynamics. Such uniform fusion can introduce mismatched periodic information into channels where historical priors differ more from current dynamics. To address this issue, we propose COMPASS, a channel-wise periodic prior adaptation framework for multivariate time series forecasting. COMPASS focuses on the channel-wise inconsistency between historical periodic priors and current dynamics and first performs channel-wise relationship calibration to quantify this inconsistency and adaptively adjust their contributions across channels. To further exploit useful periodic information, COMPASS employs dynamic-guided prior enhancement, which aggregates periodic information across channels according to the current dynamics. The enhanced periodic prior is then adaptively fused with current dynamics for forecasting. Extensive experiments on 12 public forecasting benchmarks demonstrate the effectiveness of COMPASS, reducing MSE by an average of 2.23% over the strongest competing methods. COMPASS achieves the best performance on 6 of 8 long-term forecasting datasets and all 4 short-term forecasting datasets.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Periodic Prior Calibration, Channel-wise Modeling, Time Series 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 3 reviews, above 27% of the others.
Reviewer_zrCtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes COMPASS for multivariate time series forecasting. It builds on CycleNet’s learnable periodic embeddings and adjusts their contribution separately for each channel. The Channel-wise Relationship Calibration module compares the cross-channel correlation structures of the periodic prior and current input to generate calibration weights. Dynamic-guided Prior Enhancement then uses current features to aggregate periodic information across channels, followed by adaptive fusion and forecasting. Experiments cover eight long-term and four short-term forecasting datasets. The paper reports an average MSE reduction of 2.23% over the strongest competing methods, with additional ablations, comparisons across random seeds, and efficiency measurements.

Critical strengths
  1. The paper addresses a reasonable limitation of periodic forecasting, as historical periodic patterns may be useful for some channels and less useful for others. The proposed architecture provides an explicit way to adjust their contribution and retrieve relevant periodic information from other channels. This is a useful extension of a learned periodic-prior model.

  2. The evaluation covers 12 datasets and several forecasting horizons, including comparisons with recent periodic forecasting methods. The improvement on PEMS08 is substantial, and the experiments across four random seeds on three datasets provide additional support for the forecasting results. Reporting parameter counts and training time also helps assess the cost of introducing cross-channel modeling.

Critical weaknesses
  1. The claimed decoupling effect of CRC is not explained by the operation as written. Equation (6) multiplies each channel of the prior and input by positive scalar weights. For a fixed input window, Pearson correlation is invariant to such scaling: $ \operatorname{corr}(s_i\pi_i,(1-s_i)x_i) =\operatorname{corr}(\pi_i,x_i). $ The same applies to cross-channel correlations under positive channel-wise scaling. Figures 6 and 8 nevertheless show reduced Pearson correlation after CRC and interpret this as evidence of reduced redundancy. If the measurements include the learned projections, subsequent FFNs, or pooling across windows with different weights, the correlations may change, but the effect cannot then be attributed to the scalar calibration alone. This matters because decoupling the prior and current dynamics is a central explanation of the method. Appendix D also assumes that larger relationship mismatch indicates greater prior uncertainty, which does not establish or empirically validate this connection.

  2. The evidence for the specific calibration design is weaker than the overall forecasting results suggest. In Table 3, replacing channel-wise scores with a shared score gives nearly identical results on ETTh1 and identical MSE at the reported precision on Solar and PEMS03. Cosine-based calibration also matches the full model on several datasets. These comparisons need uncertainty estimates to establish the benefit of the proposed correlation-discrepancy measure. More importantly, Table 7 labels its final dataset as PEMS08 but reports COMPASS MSE 0.086 and MAE 0.185, matching PEMS03 rather than the PEMS08 results in Tables 2 and 6. This leaves the ablation evidence for the dataset with the largest improvement unclear. The reported gains are also uneven. For example, PEMS08 improves substantially, while Traffic MSE is 0.455 compared with 0.409 for SOFTS.

Questions for authors
  1. Which tensors and axes are used to compute the correlations in Figures 6 and 8? Are these correlations computed within individual windows or after pooling multiple windows? Please explain how the reported reduction arises given the invariance of Pearson correlation to positive scalar weighting in Eq. (6).

  2. What evidence shows that the relationship discrepancy �0��0� measures the usefulness of the periodic prior for a particular channel? For example, does a larger �1��1� predict a larger forecasting error when relying on that prior? The uncertainty relationship in Appendix D is currently an assumption.

  3. Does the final block of Table 7 correspond to PEMS03 or PEMS08? Please provide the correct PEMS08 ablations and results across multiple seeds for the full model, shared-score calibration, and cosine-based calibration.

  4. Please clarify the implementation and training settings. In particular, Eq. (9) uses values derived from the refined prior �2��2�, whereas Algorithm 1 uses �3��3�. Is �4��4� constrained to remain positive as required by Appendix D, and how were the model configurations and baseline hyperparameters selected?

Reviewer_nHd4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes COMPASS, a channel-wise periodic prior calibration framework for multivariate time series forecasting. The method compares the cross-channel correlation structures of historical periodic priors and current observations to estimate channel-specific consistency and adaptively adjust the contribution of periodic information. It then employs dynamic-guided cross-channel aggregation to enhance the periodic representations, followed by adaptive fusion to generate forecasts.

The authors conduct comparative experiments on 12 public forecasting benchmarks and provide component ablations, hyperparameter sensitivity studies, and mechanism analyses. The paper reports that COMPASS outperforms existing forecasting methods on several datasets.

Critical strengths

S1. A meaningful research problem. The paper investigates channel-wise inconsistencies between historical periodic priors and current dynamics, offering a potentially useful perspective on periodic information utilization in time series forecasting.

S2. A coherent architectural framework. COMPASS integrates channel-wise relationship calibration, dynamic-guided periodic prior enhancement, and adaptive fusion into a unified forecasting framework.

S3. Broad experimental coverage. The paper evaluates the proposed method on 12 public forecasting datasets and includes component ablations, model-dimension sensitivity analyses, and mechanism investigations.

Critical weaknesses

W1. Insufficient transparency regarding hyperparameter tuning undermines the fairness of experimental comparisons. The paper does not adequately report the hyperparameter search spaces, search strategies, or model-selection procedures used for the compared methods. It also remains unclear whether the baselines underwent reasonable hyperparameter searches. Merely stating that a common experimental setup was followed does not establish that all methods received fair tuning opportunities. Therefore, the current experiments cannot adequately rule out the influence of differences in hyperparameter configurations, making the reported performance advantages insufficiently supported.

W2. The incremental methodological novelty is not sufficiently established. COMPASS primarily uses correlation differences to compute channel-wise calibration weights and combines them with attention-based aggregation and adaptive feature fusion. These mechanisms are closely related to existing techniques. The paper does not sufficiently clarify the substantive technical distinctions between its proposed approach and prior adaptive periodic modeling or feature fusion methods. As a result, the methodological novelty remains somewhat limited.

W3. The effectiveness of the core calibration mechanism requires more direct validation. The proposed method measures the consistency between periodic priors and current dynamics through differences in their cross-channel correlation matrices. However, differences in correlation structures do not necessarily correspond to the actual predictive value of historical periodic information. Although the paper provides ablation experiments, several alternative mechanisms achieve results close to those of the complete model. Therefore, the current evidence does not sufficiently establish that the proposed consistency measurement and calibration mechanism are the primary drivers of the reported improvements.

Questions for authors

Q1. Were hyperparameter searches conducted for COMPASS and all baseline models? If so, could the authors provide the search spaces, search strategies, numbers of trials, and model-selection criteria for each method? If not, how do the authors establish the fairness of the main performance comparisons?

Q2. What is the central technical novelty of the proposed channel-wise calibration mechanism compared with existing adaptive periodic information fusion approaches? Could the authors clarify its substantive differences and independent contributions?

Q3. Why should differences in cross-channel correlation structures accurately reflect the predictive usefulness of historical periodic priors? Could the authors provide more direct evidence demonstrating the relationship between the proposed consistency scores and actual forecasting gains?

Reviewer_G7zcOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes COMPASS, a multivariate forecasting framework intended to calibrate periodic priors separately for different channels. It constructs a learnable periodic prior following CycleNet, compares the cross-channel Pearson-correlation matrices of the prior and current input, and converts their discrepancy into a channel-wise calibration score. A dynamics-guided cross-channel attention module then enhances the calibrated prior, followed by adaptive fusion with the current representation. Results on eight long-term and four short-term forecasting datasets show generally competitive performance. While channel-dependent use of periodic information is a relevant problem, the proposed discrepancy does not directly measure prior–current consistency, and the ablation results provide limited evidence that the specific calibration mechanism accounts for the reported gains.

Critical strengths
  1. The paper identifies a reasonable limitation of uniform periodic-prior utilization. Different variables can exhibit different degrees of periodic stability, so applying the same prior strength to every channel can introduce misleading information.
  2. The overall architecture is straightforward and computationally manageable. Periodic calibration, cross-channel enhancement, and adaptive fusion form a coherent pipeline that can be trained end to end without an excessively large forecasting backbone.
  3. The experiments cover both long- and short-term forecasting. Evaluation on twelve benchmarks, together with component ablations and efficiency comparisons, provides a reasonably broad empirical picture.
Critical weaknesses
  1. The central CRC discrepancy does not directly measure consistency between a channel’s periodic prior and its current trajectory. It compares two within-source cross-channel correlation matrices: one computed from the periodic prior and one from the current input. Similar cross-channel relationships do not imply that the corresponding prior is temporally or numerically aligned with the current dynamics. For example, temporally shifted or otherwise incorrect priors may preserve essentially the same inter-channel correlation structure and therefore receive high consistency scores.
  2. The gate is not guaranteed to have the monotonic interpretation claimed by the paper. The consistency score uses a learnable scaling parameter without an explicit non-negativity constraint. If this parameter becomes negative, a larger relationship discrepancy produces a larger prior weight, directly contradicting the stated behavior that inconsistent channels should rely less on the periodic prior.Although the appendix explicitly assumes this condition, the main method does not state whether or how it is enforced.
  3. The ablation results do not demonstrate that the proposed components are consistently necessary. Removing DPE produces identical or nearly identical results on several datasets, removing CRC has little effect in some cases, and alternative similarity or calibration variants frequently match the full model. These results weaken the claim that the specific relationship-calibration and dynamics-guided enhancement mechanisms are responsible for the overall performance.
  4. COMPASS adopts the periodic embedding mechanism of CycleNet and assumes a predefined period length PP. However, the paper does not report the value of PP used for each dataset or explain whether these values are inherited from CycleNet, determined from dataset frequency, or selected through validation. Since the periodic prior and all subsequent calibration operations depend directly on PP, this missing specification affects the reproducibility of the method and leaves its dependence on prior knowledge unclear.
  5. The CRC submodules in Figure 2 are labeled (a), (b), (c), (e), and (f), with module (d) missing. Several mathematical symbols are also inconsistent across the figure and the main text. For example, Equation (3) denotes the relationship matrices by RπR_\pi and RxR_x, whereas Figure 2 uses different notation, and the enhanced prior is denoted by πe\pi^e in Equation (10) but by πaug\pi^{aug} in the figure.
Questions for authors

See Weaknesses.

1 comment

Thank you for the authors' response, which has addressed most of my concerns. While I have already assigned a positive score, the manuscript still contains several minor issues that require further revision. Therefore, I will maintain my original score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.