acceptodds
Withdrawn from ICLR 2027

Snow: A Stable Signed Channel Interaction Module for High-Dimensional Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time series forecasting is crucial across fields such as economics, energy, and traffic. Modeling cross-channel dependencies in high-dimensional time series remains challenging. Current methods are commonly divided into channel-independent (CI) and channel-dependent (CD) models. CI models process each channel independently, which contributes to their robustness. However, they ignore cross-channel dependencies and become less effective as the number of channels increases. CD models explicitly model interactions among channels, yet they may struggle to represent negative cross-channel relationships, produce inconsistent predictions when the channel order is permuted, and amplify perturbations during cross-channel propagation. These limitations motivate us to explore an alternative path: instead of designing complex CD models, we equip CI models with cross-channel capabilities. Based on this motivation, we propose Snow, a plug-and-play channel interaction module. Snow adds cross-channel communication through Sample-adaptive Cluster Representations. It enables signed and low-rank channel mixing to stabilize cross-channel propagation. Snow satisfies three properties: signed-interaction compatibility, channel-permutation equivariance, and interference stability. We provide theoretical analysis of these properties and validate them empirically. Extensive experiments show that inserting Snow into CI models consistently improves forecasting performance on high-dimensional datasets, while replacing interaction modules in CD models also yields consistent gains. Code is available at: https://anonymous.4open.science/r/snow-iclr/README.md

Primary area:
learning on time series and dynamical systems
Keywords:
High-Dimensional Time-Series Forecasting, Multivariate Time-Series Forecasting, Deep Learning, Signed Channel Interaction, Channel-Permutation Equivariance, Interference Stability

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_bP5gOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Snow introduces signed, low-rank channel interactions through adaptive clusters for high-dimensional time series forecasting. Theoretical analysis characterizes its structural properties, while experiments across multiple backbones, including ablations and comparisons with related methods, demonstrate forecasting improvements.

Critical strengths

Signed routing through adaptive clusters enables low-rank channel communication with explicit control over propagation, providing a principled way to augment channel-independent models.

Critical weaknesses
  • The fixed-routing contraction guarantee offers a limited explanation of the observed robustness, since input corruption can also change the routing. The connection between this theoretical property and the empirical gains remains insufficiently established.
  • The structural guarantees largely follow from the chosen factorization, while the analysis provides limited additional insight into when these constraints are beneficial or why they improve forecasting.
Questions for authors
  • How does the fixed-routing contraction result relate to the reported experiment with noisy training inputs and clean test data, given that input corruption can also change routing?
  • Which results isolate the contribution of propagation normalization from the benefits of signed routing and adaptive clustering?
Reviewer_cJz6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Snow introduces a sample-adaptive signed low-rank propagation operator that enables cross-channel information exchange while guaranteeing permutation equivariance and non-expansive propagation.

Critical strengths

Snow demonstrates broad applicability across 8 CI and 4 CD forecasting architectures, covering diverse datasets with up to 8,600 channels. The reported improvements suggest that the module can effectively enhance existing forecasting models.

The proposed low-rank propagation operator enables sample-adaptive, signed cross-channel interactions with linear complexity in the number of channels. Its unified design is relatively simple and computationally efficient.

The paper provides formal analyses of signed-interaction compatibility, channel-permutation equivariance, and interference stability. These results offer useful theoretical insights into the structural properties of the proposed interaction mechanism.

Critical weaknesses

Tables 24 and 25 contain suspiciously identical results across multiple datasets and models, raising concerns about the reliability of the reported experiments. The authors should carefully verify these results and clarify whether the inconsistencies affect the main conclusions.

The central idea of enhancing CI models through cross-channel interactions has been extensively explored in prior work, including SOFTS, CCM, and xCPD. The paper should more clearly establish the distinctive contributions of Snow and demonstrate why its specific design provides advantages over existing alternatives.

The interference stability guarantee applies only to a fixed propagation operator and does not establish the robustness of the complete forecasting model to input perturbations. Moreover, the contraction bound can be arbitrarily close to one, making its practical significance unclear.

Questions for authors

Tables 24 and 25 report identical results for several models across different datasets. Could the authors explain these inconsistencies and clarify whether they affect the main experimental conclusions?

How does Snow fundamentally differ from closely related methods such as SOFTS, CCM, and xCPD? Could the authors provide more targeted comparisons or ablation studies to isolate the benefits of Snow's specific design?

The interference stability guarantee assumes a fixed propagation operator, while the routing matrix is input-dependent. How does this theoretical guarantee translate into robustness against perturbations in the original input time series?

Since the proposed propagation matrix is symmetric and positive semidefinite, does this restrict its ability to represent asymmetric or directional dependencies between channels? How might this limitation affect forecasting performance?

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

This paper proposes Snow, a lightweight plug-and-play channel interaction module for high-dimensional time-series forecasting (HTSF). Snow uses sample-adaptive cluster representations and signed Direction Softmax to model cross-channel dependencies through a low-rank propagation operator. The paper provides theoretical guarantees for signed interactions, permutation equivariance, and interference stability. Experiments across multiple forecasting backbones and datasets demonstrate consistent performance improvements, particularly in high-dimensional settings.

Critical strengths
  1. Snow introduces a simple and efficient low-rank channel interaction mechanism that supports signed dependencies while reducing the computational complexity of dense channel interactions.

  2. The paper provides theoretical analysis of signed interactions, permutation equivariance, and interference stability, offering useful insights into the proposed design.

  3. Extensive experiments across multiple CI and CD backbones demonstrate consistent improvements, highlighting the flexibility and practical effectiveness of Snow.

Critical weaknesses
  1. The paper lacks comparisons with closely related recent methods, particularly U-Cast, Unicorn, Falcon-X, xCPD, and MS-FLOW. These methods also explore efficient or structured cross-channel interactions, making direct comparisons important for establishing the novelty and effectiveness of Snow.

  2. The theoretical stability guarantee only applies to the linear propagation operator rather than the complete module. Moreover, the contraction bound may become weak when the number of channels is large, limiting its practical significance.

  3. The motivation for applying rFFT along the hidden feature dimension is insufficient. Additionally, more detailed evaluations of permutation robustness, noise sensitivity, and computational overhead would strengthen the experimental evidence.

  4. Almost all experimental results are in the appendix.

Questions for authors
  1. What motivates applying rFFT along the hidden feature dimension rather than the temporal dimension? How does it compare with simpler MLP-based routing?

  2. How tight is the theoretical contraction bound in practice? Does the complete Snow module maintain stability under input perturbations?

  3. How does Snow compare with U-Cast and other recent channel interaction methods under the same backbones and comparable computational budgets?

  4. How sensitive is Snow to the number of clusters (K) and other hyperparameters, particularly as the number of channels increases?

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

Related papers

Discussion (0)

Sign in to comment.