acceptodds
Under review as a conference paper at ICLR 2027

GRAIL: Granger-Inspired Causal Masking Enhanced Multivariate Time Series Forecasting

Abstract

Multivariate time series forecasting is essential for a wide range of real-world applications, including finance, energy, climate analysis, and transportation. However, accurate and interpretable forecasting remains challenging due to heterogeneous temporal dynamics and complex cross-variable dependencies. Existing approaches often rely on fixed temporal modeling strategies or dense channel interactions, making it difficult to capture sparse, directional, and delayed predictive relationships among variables. In this work, we propose GRAIL, a framework that jointly models adaptive temporal patterns and lag-aware cross-channel dependencies for multivariate forecasting. The proposed method introduces a Temporal Scale Fusion module to integrate short-, base-, and long-range temporal representations through sample-level adaptive weighting. It further incorporates a Granger-inspired dependency modeling module that estimates multi-step lagged predictive relationships and constructs sparse masks to guide cross-variable information aggregation. By combining adaptive temporal representation learning with structure-aware dependency modeling, the proposed framework improves forecasting accuracy while providing enhanced interpretability. Extensive experiments on eight real-world benchmark datasets demonstrate that GRAIL achieves state-of-the-art performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Forecasting, Temporal Scale Fusion, Granger-inspired Dependency Modeling, Cross-Channel Dependencies

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_TRy7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GRAIL, a multivariate time series forecasting model that incorporates Granger-inspired causal masking. The core idea is to construct learnable masks that selectively regulate cross-variable information flow, aiming to suppress spurious dependencies and reduce overfitting to noisy correlations, thereby improving forecasting performance. The proposed masking module is modular and can be integrated into existing backbones. The authors evaluate GRAIL on several standard multivariate forecasting benchmarks and compare it with a range of existing methods. Ablation studies are also conducted to assess the contribution of the proposed causal masking module.

Critical strengths
  1. The paper addresses an important challenge in multivariate time series forecasting: distinguishing potentially useful cross-variable dependencies from spurious correlations. Incorporating Granger-inspired information into mask construction is a well-motivated approach to regulating cross-variable information flow.
  2. The proposed framework combines Temporal Scale Fusion (TSF) and the Granger-inspired causal masking (GICM). This causal masking mechanism is modular in design and could potentially be integrated into different time series forecasting backbones. This design offers flexibility and may broaden the applicability of the proposed approach beyond a single model architecture.
  3. The paper includes ablation studies to evaluate the contribution of the causal masking module. The reported results provide empirical evidence that the module can improve forecasting performance, although the extent of its benefits may vary across datasets and settings. Importantly, the work avoids overclaiming true causal discovery and only positions itself as Granger-inspired masking for prediction.
Critical weaknesses
  1. The dependency score in Eq.(15) uses lagged cosine similarity, which may capture shared trends or common drivers rather than incremental predictive contribution. Even though GICM is framed as Granger-inspired instead of true causal discovery, it remains unclear what unique benefits this scoring brings beyond standard lagged correlation. The authors should compare GICM with multi-lag Pearson correlation and prediction-error-based scores under matched sparsity and model capacity.
  2. Eq.(18) combines the lag-aware score matrix with a learnable residual term. It is unclear whether the final mask is dominated by the lag prior or the freely learned residual. Additional ablations over the mixing coefficient and analysis of score ranking shifts are required to isolate the contribution of the lag-aware scoring.
  3. Only point estimates are reported without multi-seed means, standard deviations, or confidence intervals. This is critical given small performance gaps against competing methods. Uncertainty estimates and summarized relative improvements over the strongest baseline are needed to support consistent superiority.
  4. Case studies show lag-aligned patterns but cannot quantitatively confirm that learned masks recover reliable directional dependencies. Predictive performance and structural correctness differ. Synthetic-data experiments, mask stability tests, and perturbation analysis would strengthen interpretability claims.
  5. TSF yields only modest gains, and its global aggregation routing may fail to capture variable-specific heterogeneous scale preferences. Comparisons with fixed-scale fusion and variable-specific routing, plus sensitivity analysis of the regularization coefficient, are recommended.
  6. The paper does not benchmark against recent time-series foundation models like TimesFM. The runtime and memory overhead of GICM mask estimation is also unquantified, which limits practical deployment assessment.
Questions for authors
  1. Under matched sparsity and model capacity, does GICM outperform ordinary multi-lag correlation or cosine similarity?
  2. What fraction of the final mask comes from the lag-aware score versus the learnable residual matrix?
  3. Are reported improvements statistically stable across random seeds?
  4. Can GICM recover known directional dependencies on synthetic data, and are masks stable across input windows?
  5. Does global routing outperform simpler fixed-scale fusion and variable-specific routing?
  6. How does GRAIL perform against foundation models such as TimesFM, and what extra computational overhead does GICM introduce?
Reviewer_6cwROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRAIL for multivariate time-series forecasting. It combines Temporal Scale Fusion (TSF), which adaptively fuses representations at different temporal scales, with Granger-Inspired Causal Masking (GICM), which uses lagged cross-channel scores to construct a sparse mask for channel-wise attention. The authors evaluate GRAIL on eight forecasting benchmarks, report comparisons with existing methods, and include ablations, sensitivity analyses, qualitative examples, and runtime measurements.

Critical strengths
  1. Relevant problem\textbf{Relevant problem}. This paper addresses heterogeneous temporal scales and potentially noisy cross-channel dependencies in multivariate forecasting.

  2. Modular formulation\textbf{Modular formulation}. TSF and GICM target different aspects of the forecasting architecture, making their intended roles reasonably easy to identify.

  3. Broad evaluation\textbf{Broad evaluation}. The experiments cover eight datasets and multiple forecast horizons, comparing several forecasting approaches.

Critical weaknesses
  1. Limited methodological novelty\textbf{Limited methodological novelty}. TSF appears to be an instance of adaptive multi-scale representation fusion, while GICM appears to use lag-based association scores to form a sparse attention mask. Both are plausible design choices, but the paper does not clearly establish what is technically new relative to existing multi-scale fusion, sparse attention, or dependency-aware channel modeling methods. The contribution currently reads more as a combination of familiar components than as a clearly novel modeling principle.

  2. Insufficient positioning against the closest prior work\textbf{Insufficient positioning against the closest prior work}. Related work enumerates method names without mechanism-level comparison. For multi-scale fusion, the paper does not explain how TSF's sample-level gating differs from TimeMixer's decomposable mixing, Pathformer's adaptive pathways, or MSGNet's multi-scale correlation learning. For dependency-aware attention, it does not compare GICM's mask construction against DUET, ARCausal, or CAIFormer at the level of what signal is used, how sparsity is induced, and what is assumed about lag structure. Without this, a reader cannot judge whether GRAIL introduces a new technique or adapts established mechanisms to forecasting. This should be addressed with both textual positioning and, ideally, a direct empirical comparison against the nearest neighbor in each family.

  3. The lagged score is not established as a Granger-style predictive test\textbf{The lagged score is not established as a Granger-style predictive test}. Eq. (15) is a positive-part, normalized lagged cross-correlation. Granger causality's defining requirement is that the source history improves prediction after controlling for the target's own history and other sources; Eq. (15) is a pairwise association that controls for neither, and involves no out-of-sample predictive-gain test. The "Granger-inspired" label therefore implies a stronger theoretical connection than the construction supports. The authors acknowledge in Sections 3.4 and D.3 that this is a predictive prior rather than interventional causality, but the naming should be either justified with a conditional or predictive-gain formulation, or replaced with more cautious terminology (e.g., "lagged predictive association prior"). As written, the central framing of the method rests on an overstated analogy.

  4. $\textbf{The ``state-of-the-art" framing should be qualified}$. GRAIL achieves best MSE in 16 of 32 and best MAE in 25 of 32 settings—competitive overall, but it does not dominate every dataset–horizon configuration, and the supplementary controls show it is beaten by simpler variants on some datasets. The advantages should be characterized narrowly (which dataset regimes, which horizons, which channel counts) rather than implying uniform superiority. (This is a framing issue rather than a methodological defect, but it interacts with W5 and W7 and should be corrected together with them.)

  5. Ablation evidence is limited and uneven\textbf{Ablation evidence is limited and uneven}. The main ablation table covers only four datasets and reports horizon-averaged values. Supplementary single-seed, single-horizon controls show that some variants outperform GRAIL. These results make it harder to attribute the reported gains reliably to the proposed modules.

  6. Critique on the Efficiency Claims and Experimental Support\textbf{Critique on the Efficiency Claims and Experimental Support}. This paper claims GRAIL is computationally efficient, but the empirical evaluation does not adequately support this. i) Lack of Scaling Analysis on High-Variable Datasets. The core advantage of the sparse masking mechanism should be most consequential in high-dimensional settings where dense cross-variable interactions become prohibitive, yet all profiling is confined to low-variable datasets (ETT, N=7; Weather, N=21). The high-variable benchmarks that motivate the method (Electricity N=321, Solar N=137, Traffic N=862) are never measured, so Figure 5 cannot be extrapolated to the regime the paper targets. ii) Theoretical and Implementation Discrepancies. The paper acknowledges that channel attention retains quadratic complexity in the number of variables and that the sparse mask is not implemented with specialized sparse kernels. Consistently, D.2 shows GRAIL is slower than its own dense-attention and base-only controls (4.01 vs. 2.91 ms on ETTh2; 4.46 vs. 2.69 ms on Weather), while Figure 5 compares only against external baselines. It therefore remains unclear whether the "sparse" design translates into any empirical speedup; runtime/memory as a function of N, and the channel-count range over which GRAIL is practical, should be reported.

Questions for authors

Refer to weaknesses.

Reviewer_KDa6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GRAIL, a multivariate time-series forecasting framework combining two components: (1) Temporal Scale Fusion (TSF), which adaptively combines representations obtained from short, base, and long moving-average scales; and (2) a Granger-Inspired Causal Mask (GICM), which computes lagged cross-channel similarity scores, sparsifies them, and uses the resulting mask to guide channel-wise attention. Experiments cover eight forecasting benchmarks, four prediction horizons, several ablations, sensitivity studies, qualitative examples, and computational profiling.

The forecasting results are competitive, particularly in MAE. However, I have substantial concerns about the technical interpretation and specification of the main GICM component, inconsistencies between the mathematical description and the reported implementation, and the lack of statistical and structural validation needed to support the main claims. I therefore recommend rejection in the current form.

Critical strengths
  1. Broad forecasting evaluation. The main experiment covers eight datasets and four prediction horizons and includes recent channel-dependent and causal-aware baselines.

  2. Competitive predictive performance. GRAIL obtains the best reported MAE in 25 of 32 settings and the best MSE in 16 settings. The results are particularly strong on Weather and Solar-Energy.

  3. Relevant modeling problem. Jointly adapting temporal scales and selecting delayed cross-channel information is well motivated for multivariate forecasting.

  4. Useful supplementary analysis. The appendix documents data splits, optimization settings, sensitivity to lag count and input length, internal computational costs, and several negative results. I appreciate that the authors explicitly state that the visualizations do not demonstrate interventional causality and that the sparse mask is not implemented with sparse kernels.

Critical weaknesses

1. The proposed score does not establish Granger-style incremental predictability

Equation (15) is a positive, uncentered lagged cosine similarity between two observed sequences. Standard Granger predictability asks whether the past of a source improves prediction of a target after conditioning on the target's own past and, in a multivariate setting, other available histories. GICM does not compare restricted and unrestricted predictors, condition on other histories, measure an improvement in prediction error, or conduct a statistical test. Consequently, its score cannot distinguish a predictive link from common trends, seasonality, autocorrelation, or common causes.

Related concerns are that negative associations are discarded even though they may be predictive, cosine similarity is invariant to amplitude rescaling despite the stated motivation for using raw observations, and all channel pairs share the same lag weights πℓ\pi_\ell. The latter appears inconsistent with the claim that the method captures variable-specific response delays. The authors appropriately avoid claiming interventional causality in parts of the text, but terms such as “causal mask,” “causal graph,” and “causal correction” still overstate what is identified by the method. As currently defined, this component is more accurately described as a lagged positive-similarity prior.

2. Key parts of the method are not defined

The paper does not specify how the adaptive residual matrix RR in Eq. (18) is generated. It also does not define Gate⁡(Pij,τi)\operatorname{Gate}(P_{ij}, \tau_i) in Eq. (21), including its forward function, gradient estimator, stochastic behavior, and inference-time operation. This is especially important because Appendix D states that the large-channel mask branch can be stochastic during inference. Several consequential parameters, including the initialization or constraints for ρ\rho, α\alpha, β\beta, and λbal\lambda_{\mathrm{bal}}, are also missing.

These omissions prevent independent reproduction and make the interpretation of GICM unclear. If the learned residual RR dominates the explicit score GG, the final mask need not retain meaningful lag information. The paper does not report the learned α\alpha, remove RR in an ablation, or otherwise quantify how much of the final mask is attributable to Eq. (15).

There is also a concern with the claimed adaptive threshold. Since each row of PP sums to one, Eq. (20) reduces to

τi=ρ1−PiiN−1.\tau_i = \rho \frac{1-P_{ii}}{N-1}.

Thus, apart from ρ\rho, the threshold depends only on the self-entry and does not adapt to the shape or concentration of the non-self dependency distribution.

3. The TSF equations and implementation description are inconsistent

Equations (9)–(13) aggregate statistics over both time and channels and define one set of scale weights per sample. Appendix A.5 instead states that routing is evaluated for each window-channel pair, which describes a materially different model. The notation and implementation need to be reconciled.

Moreover, the router operates after RevIN. Under standard instance normalization, the temporal mean and variance of each normalized window-channel pair are approximately fixed at zero and one. If routing is channel-wise as stated in the appendix, s1s_1 and s2s_2 therefore carry little or no sample-specific information, leaving the average absolute difference s3s_3 as the primary informative statistic. This apparent degeneration is neither discussed nor ablated.

Finally, Eq. (14) constrains only the batch-average use of each branch. It does not prevent individual samples from receiving highly concentrated routing weights, contrary to the accompanying explanation.

4. The empirical evidence does not establish robustness

All reported results use a single seed. Many main-table improvements are only 0.001–0.005, so their magnitude cannot be assessed without variance estimates or paired statistical comparisons.

The supplementary controls also weaken the conclusion that both proposed modules consistently improve forecasting. In Table A6, all five controls obtain lower MSE than GRAIL on ETTh2. On Weather, dense attention and the single-lag variant outperform GRAIL in both MSE and MAE. Appendix C further reports no uniform per-step advantage for GRAIL in the diagnostic examples. These experiments use an MSE-based training and selection protocol, whereas the main experiments use MAE, so they are not direct contradictions; however, the change of protocol prevents them from validating the main ablation results. The large difference between the main and supplementary results for nominally similar settings—for example, Weather at L=96L=96, K=3K=3, and H=96H=96—also deserves a clearer account.

5. The evidence for structural interpretability and novelty is limited

The qualitative analysis shows that selected variables can have lagged correlations with a target, but this largely verifies the statistic used to select them. There is no experiment on a synthetic VAR or nonlinear dynamical system with known edges and delays. Consequently, the paper does not measure edge precision/recall, direction recovery, lag recovery, or resistance to confounding. Attention and correlation visualizations alone cannot validate the correctness of the learned dependency structure.

The individual design ideas also have substantial precedent: adaptive multi-scale routing is closely related to models such as Pathformer, while combining temporal heterogeneity modeling with sparse channel selection is closely related to DUET. The use of a lagged similarity prior may still constitute a useful engineering contribution, but its added value is not sufficiently isolated. In particular, the main comparison omits some highly relevant temporal and channel modeling methods discussed by the paper, including Pathformer, TimeMixer, and SOFTS. The paper also does not explain the search and selection procedure for the dataset- and horizon-specific temporal window parameter MM, making baseline fairness difficult to assess.

Questions for authors
  1. What are the exact definitions and architectures of RR and Gate⁡\operatorname{Gate}? Please specify gradient propagation, stochastic operations, inference behavior, and all associated hyperparameters and initializations.

  2. Is TSF routing performed once per sample or once per window-channel pair? Please provide corrected equations matching the implementation and explain how the mean and standard-deviation router features remain informative after RevIN.

  3. What values does the learned mixture coefficient α\alpha take across datasets and horizons? How does performance and structural recovery change when RR is removed?

  4. Can the authors evaluate GICM on synthetic multivariate systems with known directed edges and delays, reporting edge F1, direction accuracy, and lag-recovery error? How does it behave under common causes, strong autocorrelation, negative effects, or shared trends?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.