acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Cross-Channel Importance In Time-Series Forecasting

Abstract

Cross-channel modeling is central to multivariate time-series forecasting, yet channels that are statistically related, predictively useful, and actually used by a trained forecaster are often treated as if they defined the same notion of importance. We show that they need not coincide. Cross-channel dependency structures change substantially across future offsets, and horizon-adaptive source selection improves a controlled Ridge predictor in 21 of 32 dataset–prediction-length conditions, with a mean gain of . This selected-set signal also transfers to a matched nonlinear predictor. Yet imposing the same horizon-specific source logic on iTransformer yields only 11 of 20 wins and a mean gain of , with little alignment between controlled and neural gains. Functional interventions further show that strong forecasters use cross-channel information, while their source-reliance rankings agree little with controlled utility or with one another across iTransformer, TimesNet, and a cross-channel TimeMixer. As a constructive consequence, bounded post-hoc support improves a frozen channel-independent forecaster in 12 of 16 dataset–horizon conditions, with a positive aggregate bootstrap interval. Cross-channel importance should therefore be interpreted relative to the forecasting mechanism and question that define it: related useful used.

Primary area:
learning on time series and dynamical systems
Keywords:
multivariate time-series forecasting, cross-channel modeling, cross-channel dependency, predictive utility, functional reliance, horizon adaptation

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

This paper discuses channel modeling in time-series forecasting. The key claims is that channels are often treated as if they have the same importance in modeling. However, channel dependency changes across future offsets, and an adaptive selection strategy would be beneficial. Authors decompose dependency, controlled utility, and functional reliance, and demonstrate that functional source reliance is dependent on forecaster models.

Critical strengths

The authors investigate spatial/channell dependency modeling, which is a critical topic in time-series forecasting.

The authors provide a comprehensive and thorough literature review from three aspects in time-series forecasting, including forecasting backbones and channel modeling, adaptive cross-channel structure, and feature importance and intervention.

Critical weaknesses

The contribution of is paper is not very clearly introduced. The limitation of existing studies should be more carefully analyzed and justified, with substantial discussion with literature support and quantitive/qualitative analysis.

Experiments results should report in similar protocol (metric system, datasets, statistics). For example, the CI should be reported in the involved tables and figures, not only table 4. The experiment setup should be detailed in the main paper, to make the paper self-contained.

The structure of this paper is not very clear and different from existing papers. Referring to standard structures (possibly using subsections, research questions) could improve the clarity of this paper.

Theoretical justification is not available. While not necessary for application-directed papers, an insightful theoretical analysis could improve the rigor of the statement in the paper and help the authors (readers) to understand the research questions.

The quality of tables and figures should be improved. Some figures and tables would be not very informative but takes large space. For example, figure 2 (a) contains only 4 numbers but takes 10+ lines (left panel). Figure 2 (b) is a table, instead of a figure.

Questions for authors

How does the referred cross-channel dependency relate to cross-time dependency? That is, is there unique challenges for modeling cross-channel dependency, given the rich literature on modeling time dependency (self-attention, mlp, gnn, etc)?

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

This paper distinguishes three notions of cross-channel importance: observed dependency, controlled predictive utility, and functional reliance, and shows that these notions need not coincide. The authors find that dependency structures vary across future offsets and that horizon-specific source utility can improve prediction under controlled settings. However, the resulting source rankings do not reliably translate into better neural topologies or identify the sources on which trained forecasters rely.

Critical strengths

1.The paper addresses an important topic, as effectively using multivariate information is crucial for multivariate time-series forecasting.

2.The paper’s insights may inform the development of future forecasting methods that make more effective use of multivariate information.

Critical weaknesses

1.I strongly recommend that the authors carefully revise the manuscript to improve its clarity and precision. Several terms, variables, and functions are insufficiently defined, making the formulation and analysis difficult to follow. Specific examples are provided in the questions below.

2.Effectively using multivariate information is a central challenge in multivariate time-series forecasting. Prior work has already shown that incorporating information from other variables does not necessarily improve forecasting performance [1,2]. In my view, the key challenge is how to use this information effectively. The paper should explain more clearly what new understanding its distinction between dependency, predictive utility, and functional reliance provides beyond this existing observation, and how this understanding can guide the development of better forecasting methods.

[1] Are transformers effective for time series forecasting?

[2] Univariate vs multivariate time series forecasting with transformers

Questions for authors

1.In line 41, are “predictor” and “forecaster” used interchangeably, or is a distinction intended? Please clarify any intended distinction and use the terminology consistently throughout the manuscript.

2.The notation in Eq. (1) requires clarification. Although Xt∈RL×C\mathbf{X}_t \in \mathbb{R}^{L \times C} is defined earlier, the scalar quantities xj(t)x_j(t) and xi(t+h)x_i(t+h) are introduced without explicit definitions. Please define these quantities and clarify their relationship to the previously introduced notation.

3.In lines 125–130, terms such as “target-history baseline” and “controlled predictor” are not sufficiently specified in the main text.

4.Several aspects of the notation in Eq. (2) require clarification. The function PijP_{ij} explicitly depends on HH and ℓ\ell, but this dependence is left implicit on the right-hand side. Please explain how these quantities enter the loss evaluation. In addition, Li(⋅)L_i(\cdot) is not explicitly defined, while LL was previously used to denote the input history length. Please define LiL_i and consider using distinct notation to avoid ambiguity.

5.Equation (3) uses EH\mathbb{E}_H without specifying the underlying probability distribution or an explicit averaging rule.

6.Equation (5) similarly leaves the dependence on HH and ℓ\ell implicit on the right-hand side. Please make this dependence clear and explicitly introduce the forecasting function F\mathcal{F} before using it.

7.Although Section 5 summarizes several numerical findings, the detailed experimental comparisons are presented in the appendix. Including a key comparison table or figure in the main text would make the evidence supporting this section easier to assess.

Reviewer_rx8XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates three notions of cross-channel importance in multivariate time-series forecasting: observed dependency, predictive utility, and functional reliance. Through experiments with Ridge regression and neural forecasting models, the authors show that these importance measures do not necessarily agree and that horizon-specific source selection does not consistently improve strong neural forecasters. The paper also introduces a bounded post-hoc support module to incorporate cross-channel information into a frozen channel-independent forecaster, achieving modest forecasting improvements.

Critical strengths
  1. The paper investigates an interesting question about the relationship between statistical dependency, predictive utility, and model reliance in multivariate time-series forecasting.

  2. The experimental analysis is relatively comprehensive, covering multiple datasets and forecasting architectures. The additional robustness checks and controlled comparisons help support the empirical observations.

Critical weaknesses
  1. Questionable measurement of functional reliance. The proposed EPI measures channel importance by replacing a source channel with the corresponding channel from another sample. However, this may introduce unrealistic cross-channel combinations and distribution shifts. The resulting performance degradation could reflect sensitivity to the perturbation rather than the model's actual reliance on that channel. Although the authors examine the stability of EPI, it remains unclear whether the conclusions would hold under alternative interventions, such as masking or conditional imputation.

  2. Limited novelty and practical contribution. The distinction between statistical dependency, predictive utility, and model reliance is closely related to existing feature-importance research, while the proposed measurements mainly rely on conventional techniques. Furthermore, the final bounded-support module is essentially a residual cross-channel attention adapter and does not directly use the importance rankings studied earlier. Its forecasting improvements are also relatively small, making the practical implications of the analysis less convincing.

Questions for authors
  1. Have the authors considered alternative interventions, such as mean masking or conditional imputation, to verify whether the functional reliance rankings remain consistent? How can the effects of distribution shift introduced by donor replacement be distinguished from genuine model reliance?

  2. In the iTransformer experiments, the sparse attention mask is applied only to the final encoder block, while earlier layers still perform cross-channel interactions. Could this affect the conclusion that horizon-specific predictive utility does not transfer to neural forecasting models?

  3. Since the final bounded-support adapter does not directly use the proposed importance rankings, what specific insight from the preceding analysis motivates its design beyond simply introducing additional cross-channel interactions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.