acceptodds
Under review as a conference paper at ICLR 2027

Diagnosing Cross-Variate Access in Time Series Forecasting: Benefits, Edge Redundancy, and Allocation

Abstract

Multivariate time-series forecasting is often framed through channel-independent or channel-dependent architectures, but this model-level distinction does not reveal which variables and forecast objectives benefit from cross-variate access, or what connectivity preserves those benefits. We introduce a controlled diagnostic framework using binary structural masks fixed throughout training and paired retraining to distinguish access benefit, budget sufficiency, and connection allocation. With input and prediction lengths both fixed at 96, we conduct primary experiments on four benchmarks using a compact iTransformer. We find that access benefits vary across datasets, receiving variables, and forecast objectives: aggregate noninferiority after removing cross-variate access can coexist with increased error for individual variables. After separately optimizing Short (steps 1–48) and Far (steps 49–96), we find stronger Short access benefits on Solar and ECL with both compact iTransformer and SAMformer. Positive access benefits also coexist with substantial edge redundancy: on Solar and ECL with compact iTransformer, randomly removing 90% of direct cross-variate edges satisfies a prespecified 1% noninferiority criterion for joint-96 mean squared error. Exploratory Solar sparsity curves further show sharply increasing Short error under extreme sparsification, while Far error remains close to the fully connected baseline. Beyond edge count, allocation also affects performance: at fixed budgets, restoring cross-variate coverage consistently reduces local error for previously isolated receivers across three real-data settings. In controlled synthetic experiments, connections to predictive-parent proxies outperform non-parent controls matched in edge count, receiver degree, and coverage. These findings support evaluating cross-variate interaction jointly in terms of forecast objectives, connection budgets, and allocation, providing empirical guidance for selective interaction mechanisms.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time-series forecasting, Cross-variate interaction, Objective-conditioned access, Structural sparsity, Receiver coverage, Predictive source relevance.

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_Srn2OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes the benefits and shortcomings of channel-dependent and channel-independent forecasting. It employs iTransformer and SAMFormer in experiments across two datasets, Solar and ECL (except 4.1 where ETTm1 and Exchange are also covered). A series of experiments analyzes types of interactions and their impact on the forecasting performance measured by MSE. Most experiments focus on the log-MSE ratio of the models using an identity attention mask (no-cross) and dense (full mixing). The main results show that short-horizon (Short) forecasts tend to benefit from cross-channel access, while long-horizon (Far) forecasts do not show much change in no-cross vs dense settings. The last set of experiments explores five arbitrary strategies of removing cross-channel edges and finds that removing up to 90% of them does not significantly hurt MSE (according to the introduced noninferiority criterion).

Overall, the paper focuses on an interesting topic and runs relevant experiments. However, the methodology, the depth of experiments, and the analyses are significantly limited, causing little to no generalizability of the observed behaviors. Moreover, the paper is hard to follow at times because there are plenty of undefined terms and vague sentences; paragraphs are loaded with numbers, and a lot of important context paramount for understanding the contents of a given section is moved to the appendix (for instance, section 6 is difficult to read without constantly looking into the appendix).

Critical strengths
  1. The paper asks interesting questions about cross-channel interactions in multivariate forecasting and analyzes them to a certain extent.
  2. The Short vs Far division is a smart idea and the results favoring channel-interactions for Short are insightful (for iTransformer).
Critical weaknesses
  1. The majority of experiments are restricted to only two datasets (Solar, ECL). It is critical to run this analysis across various datasets and domains because they exhibit diverse patterns of cross-channel interactions.
  2. Only one model type is evaluated (iTransformer; SAMFormer is based on the same architecture). This is a significant limitation of this work because iTransformer-style models compress each variate into a single token and only after that apply cross-variate attention. Multivariate forecasting is by no means dominated by only iTransformer-style models, and the analysis performed by the authors carries no generalizable information about multivariate forecasting.
  3. The fixed 96 context-horizon setup is a severe limitation. It is not unlikely that different datasets would exhibit different spurious/useful cross-variate correlations, which would impact the observed results. Furthermore, the division into Short (1-48) and Far (49-96) is also arbitrary, and some sort of sensitivity analysis on it should be performed.
  4. Analyses mostly focus on numbers and confidence intervals while omitting detailed explanations of results and the experimental setup. Moreover, some of the analyses seem to be logically flawed. For instance, in Section 4.1: “No-cross changes MSE relative to Dense by -1.31% and -1.13% with one-sided 95% upper bounds of -0.87% and - 0.51%.” This is a contradiction because the CI doesn’t cross zero, so the CIs actually suggest that the improvement of Dense vs No-Cross on ETTm1 and Exchange is also there, it's just that its magnitude is small. Moreover, I don’t see how they satisfy the 1% non-inferiority criterion since the means of the CIs are larger than -1%.

Writing/formatting:

  1. L194, L200: “receivers” is a vague term that hasn’t been clearly defined. Presumably, receivers mean series/channels/variates, but this is not clear.
  2. I think it would improve the formatting and clarity if standard terms like joint-96, Dense, No-cross, Far, and Short were set in a distinct font.
  3. Undefined
    1. compact iTransformer.
    2. Random-95%.
    3. test worlds
    4. Random mask
    5. original-zero receiver
    6. improvements in all 20 families
    7. DGP
    8. n = 5, seeds 2–6 (Figure 3) // there are 6 seeds in this interval but n=5
Questions for authors
  1. What does “20 paired Dense-No-cross training blocks for each of four datasets” precisely refer to? (L183).
  2. How was instance normalization applied during training? Specifically, were model predictions inverse-scaled back to the original scale for loss computation?
Reviewer_DeKpOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies when cross-variate information helps multivariate forecasting and how much connectivity is needed to retain its benefits. Fixed attention masks and paired retraining separate the effects of access, connection budget, and allocation. Experiments on four benchmarks, mainly with a compact iTransformer, show that access benefits vary across datasets, receiving variables, and separately optimized forecast intervals. Random sparse masks can preserve aggregate performance despite local losses. Coverage-repair experiments and synthetic parent-versus-decoy comparisons examine allocation beyond edge count.

Critical strengths
  1. The intervention design makes access comparisons interpretable. Section 3 fixes the backbone and training protocol, preserves self-access, and pairs initialization and data order. Separate Short/Far training goes beyond post hoc horizon slicing and shows that aggregate accuracy can conceal substantial differences across forecast objectives and receivers.

  2. The direct-edge redundancy result is supported beyond one dataset: 90% removal meets the stated 1% aggregate noninferiority criterion on Solar and ECL (Tables C1 and C4). Receiver-level losses and remaining indirect reachability are also reported, helping readers understand what aggregate noninferiority does and does not establish.

Critical weaknesses
  1. Robustness to the training budget is incompletely established. Matching training settings is valuable, but Dense, sparse, and self-only attention can converge differently. Appendix A.3 retains ETTm1 despite 13/20 Dense runs reaching the 10-epoch cap; only 2/20 choose the last checkpoint, which is reassuring but does not establish stability under longer training. Solar's stated gates likewise permit some cap hits without giving complete per-mask convergence summaries. This does not prove undertraining, but it leaves uncertainty about whether some apparent access or sparsity effects reflect the particular training budget.

  2. The practical benefit of coverage repair is less clear than the improvement for previously isolated receivers. Coverage-1 transfers edges from other receivers, yet Table D3 reports all originally nonzero receivers together, combining donors with untouched receivers. This can dilute concentrated donor costs. The paper acknowledges a trade-off: Solar/SAMformer improves isolated receivers by 18.98%, while the overall MSE point estimate increases by 0.28%, with its interval including zero. Donor-specific results are needed to judge whether repair removes a local vulnerability or redistributes it.

  3. The synthetic source-relevance contrast is not specified sufficiently to verify its interpretation. Section 6.3 and Appendix D.4 define receiver-to-parent or non-parent proxy connections, but do not state the complete masks for proxy query rows or other receiver connections. In a multilayer model these choices determine indirect access to parent information. Matching direct edge count, receiver degree, and coverage alone does not resolve this ambiguity, so the claimed isolation of source identity needs clarification.

Questions for authors
  1. Beyond the cap-hit counts in Appendix A.3, is the ETTm1 Dense-No-cross result stable under a longer epoch cap? Please provide a small paired extension if available, and per-mask stopping summaries for the primary Solar comparisons. This would clarify whether the conclusions persist when training-budget sensitivity is checked directly.

  2. For Table D3, how many receivers lose edges under Coverage-1, and what are their paired MSE changes, separately from untouched nonzero receivers? These existing-checkpoint results would clarify the cost of repairing isolated receivers, particularly in the Solar/SAMformer setting.

  3. What are the complete Oracle and Decoy masks, including proxy query rows and any receiver-to-receiver edges? Are all connections outside the specified receiver-to-proxy edges identical or self-only, and can any permitted multilayer route supply parent information to Decoy? This clarification is necessary to assess the source-relevance contrast.

Reviewer_29QqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors study the benefits of cross-variate access in multivariate time series forecasting. They design a controlled experimental framework using binary structural masks to enable paired comparisons and different configurations of connections.

The main conclusions are that:

  • Cross-variate benefits vary across datasets, receiving variables, and forecasting horizons.
  • In particular, cross-variate access can be very impactful for short-term forecasting.
  • Most direct connections may be redundant. They show that removing 90% of connections can preserve performance, while most variables remain reachable within one or two attention layers.

Experiments are conducted on 4 datasets and 2 models.

Critical strengths
  • A lot of attention and care were given to the experimental setup and statistical analysis. The experiments are carefully designed, which is really appreciated.
  • The experimental sections follow naturally from one another and are well justified, making the paper easy to read.
Critical weaknesses

The paper reads more like a carefully designed empirical analysis, with limited novelty in terms of methodology or theoretical insights. I think it provides an interesting starting point, but would benefit from a further step building on these findings to propose or validate a new approach.

Questions for authors

Building on this work, could the authors consider a practical method for designing cross-variate connections and evaluate its benefits on a variety of real datasets? This would help demonstrate the analysis's practical implications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.