acceptodds
Under review as a conference paper at ICLR 2027

Text-Guided Multimodal Attentive Graph Learning for Time Series Forecasting with Exogenous Variables

Abstract

Time series forecasting with exogenous variables predicts endogenous targets using historical observations and associated exogenous information. Existing studies mainly rely on numerical time series, leaving textual semantics largely unexplored. Furthermore, to improve efficiency, many methods simplify inter-variable modeling and fail to capture its full capacity, particularly exogenous-endogenous relationships. To address these limitations, we propose MAGNet, a text-guided Multimodal Attentive Graph Network that integrates textual semantics into graph learning for time series forecasting with exogenous variables. MAGNet first employs an offline pretrained language model to generate contextual descriptions for exogenous and endogenous variables separately, providing global semantic representations beyond numerical observations. The resulting representations are incorporated into an asymmetric heterogeneous graph to better capture hierarchical inter-variable relationships. Building on this graph, predictive causality-aware attention mechanisms learn target-focused graph representations, improving forecasting performance. Experiments on 12 real-world datasets show that MAGNet achieves competitive short- and long-term forecasting performance compared to existing baselines, with ablation and case studies further supporting the effectiveness of the proposed framework.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting, Multimodal forecasting, Graph-based forecasting, Forecasting with exogenous variables, Exogenous

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_U4SEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the model of MAGNet, which combines the text information generated by the large language model with the graph network to predict the time series with exogenous variables. The model learned the relationships between exogenous variables and between exogenous variables and target variables, and conducted experiments on 12 datasets.

Critical strengths

None

Critical weaknesses
  1. The overall innovation of the method is relatively limited, mainly combining existing text embedding, graph networks, and attention mechanisms. Although the paper introduces the functions of these modules, it does not fully explain why they need to be designed in this way and what problems they solve compared to existing methods.

  2. The paper used graphs to model the relationship between exogenous variables and target variables, but since there is only one target variable, this process is actually very similar to using attention mechanisms to aggregate information from exogenous variables. The author did not fully explain why a graph structure is needed, nor did they provide a direct comparison with regular cross-attention. The support for the effectiveness of each module in the ablation experiment is insufficient. For example, after completely removing the text, there is almost no change in model performance, and the improvement in other modules is also very small. On some datasets, removing these modules can even achieve better results.

Questions for authors

Can you briefly explain what problems this article has solved compared to existing methods?

Can you provide a comparison with aggregation methods such as attention mechanisms?

3 comments

Thank you for your prompt review.

W1, Q1: Problem and Innovation

  • MAGNet solves time series forecasting with exogenous variables (ExoTSF) by employing text modality and heterogeneous, asymmetric graph structures.
  • ExoTSF is more coherent to the forecasting goals and dataset context. Most of the time, real-world applications only need one or two variables to be forecasted (target/endo) and consider the rest as their effectors (covariates/exo).
  • Current uses of text modality in TSF directly mix text and numbers as input, which can make input prone to mistakes given a wrong text.
  • MAGNet expresses exo-endo dependency explicitly by numerical measure, unlike other models that process it implicitly through simple attention networks.
  • We have revised Fig. 1 as follow:
    • Fig. 1 (a) [Others] → general multivariate TSF with text-infused input
      This mechanism lacks exo-endo separation (symmetrical) and text-infused
    • Fig. 1 (b) [Others] → TSF with exogenous variable with text-infused input
      This mechanism highlights exo-endo separation and text-infused
    • Fig. 1 (c) [Ours] → TSF with exogenous variable with text guidance
      This mechanism highlights exo-endo separation and text is only used to put attention in the correct numbers/variables.

MAGNet's performance compared to other models:

DatasetTermUnimodal ExoTSFMultimodal TSFUnimodal TSF
MAGNet (Ours)DAGGCGNetTimeXerT3TimeTimeCMATimeLLMTimeKANDLinearPatchTSTTFT
MAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSE
AvgShort0.39830.40900.45880.48880.42580.44010.44750.45920.45240.47060.47140.49880.47240.50010.45670.50230.43910.47000.44770.49290.44990.5054
Long0.54680.64810.56060.67030.55950.67860.58900.70660.56730.65430.60030.68810.57230.65660.56410.69270.56440.66920.59970.75920.57090.7282
All0.47250.52860.50970.57960.49270.55940.51820.58290.50980.56250.53590.59340.52240.57840.51040.59750.50180.56960.52370.62600.51040.6168

Other baselines categories:

  • Unimodal multivariate TSF → simple numerical modality with symmetrical assumption
  • Multimodal multivariate TSF → added text modality with symmetrical assumption
  • Unimodal exogenous TSF → simple numerical modality with exo-endo separation

Result: The table shows that MAGNet achieves better performance with its combined mechanism to solve exogenous TSF with text modality.

W2, Q2: MAGNet vs Attention Mechanism

  • MAGNet is not a simple attention mechanism as it applies two-step graph learning:
    1. Exo-Exo GAT: Exogenous nodes learn from one another through correlation-based graph attention network
    2. Exo-Endo GAT: Endogenous node learn from the enriched-exogenous nodes through a final dependency-based graph attention network
  • A simple attention mechanism cannot handle the difference between exo-exo correlation and exo-endo dependency relationship explicitly.

To better support, we extended MAGNet's ablation table to more MAGNet's variants:

  • w/o Graph → DLinear with matched ensemble and configurations as if the graph is there
  • w/ ExoMLP → replacing the graph system with simple exo-aware MLP
  • w/ Cross-attn → replacing the graph system with simple cross-attention network
Fullw/o Textw/o Exow/o Graphw/ Exo MLPw/ Cross-attn
MAE (Overall)0.47250.47690.49280.49430.52680.4858
MSE (Overall)0.52860.53670.56840.56240.59860.5461

Result: Full MAGNet performs best on both overall MAE and MSE, showing that the graph network is important.

Thank you for conducting additional experiments. These results have addressed my concerns to some extent, but they are still insufficient to convince me of the technological innovation of this method.

I still believe that the core design of MAGNet has significant similarities with existing attention aggregation mechanisms. Especially, it is still debatable whether the cross attention comparison currently provided is sufficient to demonstrate the unique advantages of graph structures. MAGNet itself uses two-stage attention aggregation of Exo-Exo and Exo-Endo, so it may not be entirely fair compared to a simple cross-attention module.

The current supplementary experiments have shown that MAGNet has certain performance advantages, but it has not yet been fully demonstrated that its core design has sufficient clear technological innovation compared to existing deep learning modules. So, I decide to keep my original score at 2.

Thank you for your careful follow-up! We understand your skepticism regarding MAGNet's technical innovation and how it differs from usual aggregation methods. We hope this comment clarifies your concerns.

Concern 1: Technical Innovation

What makes MAGNet stand out

  • Explicit exo-exo and exo-endo relationships. MAGNet infers these relationships through correlation and dependency metrics. Usual self-attention or cross-attention networks have to infer them implicitly from the same input. MAGNet's inter-node weights complement such mechanisms and improve accuracy.
  • Controlled message passing. The correlation and dependency weights are used directly in graph learning to control how much information one node passes to another. This matters because not all exo variables benefit the prediction of future endo patches equally.
  • Text as guidance, not as a mixer. This is one of MAGNet's core contributions, separate from the graph attention network. Current models tend to merge text and numerical data into a shared embedding and forecast from it, so a flawed text can directly corrupt the predicted values. In MAGNet, text only guides the attention.
  • Text as global information. Many recent models use patching, which emphasizes local information within each patch. In MAGNet, text describes the overall pattern and statistics of the whole series, so every patch is informed uniformly about the context outside the patch itself.

What we do and do not claim

  • MAGNet uses attention. MAGNet revolves around attention mechanisms (a graph attention network). It is a technique we adopt and extend to fit the ExoTSF problem.
  • Our main contribution is not the attention network itself. MAGNet aims to solve ExoTSF with a text modality. Our contribution lies in the text-guidance mechanism and the exo-endo separation, not in improving existing attention networks.

PatchTST [1] is a comparable case. It applies attention in a manner similar to Vision Transformers (ViT) [2], yet its core contribution is showing that patching is effective for time series forecasting.

Empirical support for these points

  • Removing the exo-exo graph while keeping the SIG prior (w/ SIG-NoExo) or replacing the graph with a homogeneous one (w/ HomGraph) both lower accuracy compared with full MAGNet.
  • In the shuffled-exo study, shuffling MAGNet's signal raises MAE by about 20-29% in short-term settings for non-exo backbones (DLinear, PatchTST, TFT), so the signal is not redundant.

Concern 2: Fairness of the Ablation Comparison

We agree that the single cross-attention variant is not a fully matched counterpart to MAGNet's two-stage design, and we do not claim that our graph attention network is superior to every attention-based aggregator. Our claim is narrower: for ExoTSF, explicit exo-exo and exo-endo dependency priors, exo-endo separation and text guidance improve forecasting over models that treat exogenous variables implicitly.

References

[1] Nie et al. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. ICLR 2023.

[2] Dosovitskiy et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. ICLR 2021.

Reviewer_Rz1kOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2(+1)
Read review
Paper summary

This paper proposes MAGNet, a text-guided multimodal attentive graph network for time series forecasting with exogenous variables. The method distinguishes between endogenous target variables and exogenous predictors and constructs a heterogeneous graph to capture their relationships. Specifically, exogenous–exogenous relationships are modeled through cosine similarity, while exogenous–endogenous relationships are modeled using a neural Granger causality (NGC) mechanism. Textual descriptions generated offline by a pretrained language model are encoded into semantic representations and incorporated into graph attention. The resulting endogenous graph embeddings are integrated with conventional forecasting backbones.

The authors evaluate the method on twelve datasets covering electricity prices, energy, hydropower, and wind power forecasting. They report improvements over several forecasting baselines, particularly when MAGNet is combined with DLinear, and provide ablation studies and visualizations to support the proposed architecture.

Critical strengths

Relevant problem formulation. Forecasting a target variable using multiple exogenous time series is practically important. Explicitly distinguishing exogenous and endogenous variables is a reasonable modeling choice, particularly when different variables have different roles in the forecasting task.

Intuitive architectural design. The proposed two-stage graph aggregation, which first models exogenous–exogenous interactions and then aggregates information toward the endogenous variable, provides a potentially useful inductive bias. Incorporating text-derived contextual information may also be beneficial when variable descriptions contain meaningful domain knowledge.

Extensive empirical evaluation. The paper reports experiments across twelve datasets, multiple forecasting horizons, and several backbones, with five random seeds. In particular, the improvements achieved by MAGNet+DLinear suggest that the proposed augmentation framework has some empirical potential.

Critical weaknesses

W1. A fundamental issue in the proposed neural Granger causality mechanism.

The paper defines the causal edge score as the difference between the restricted and unrestricted prediction losses and subsequently applies a softmax over the candidate exogenous variables. However, since the restricted loss is common to all candidates, it cancels exactly under the softmax: softmaxi(Lrest−Lunr,i)=softmaxi(−Lunr,i).softmax_i(L_{rest} − L_{unr,i}) = softmax_i(−L_{unr,i}). Consequently, the restricted model does not affect the resulting attention weights. More importantly, Appendix C explicitly states that the restricted and unrestricted networks are trained only through the final forecasting objective, without auxiliary prediction losses. The restricted model therefore receives no gradient through the proposed mechanism.

This is a fundamental methodological concern because the proposed NGC scores no longer represent improvements in predictive performance relative to a restricted model. Instead, they effectively act as learned attention biases derived from the unrestricted networks' losses. The claimed Granger-causal interpretation is therefore unsupported. Even setting causality aside, just for prediction, it acts somewhat like self-ablation.

W2. The predictive-causality formulation is dubious even independently of W1.

The restricted and unrestricted networks appear to reconstruct the current endogenous patch representation, rather than predict an unseen future outcome from historical observations. Algorithm 2 explicitly uses the current endogenous embedding both as input and as the loss target. It is unclear how this corresponds to the temporal prediction problem required for Granger causality.

To be more concrete, just see how Eq (6) is estimated. In Section 3.2, the MSE difference is supposed to measure the improvement in prediction resulting from including an exogenous variable. That interpretation requires the two networks to be trained to predict their targets under their respective information sets. However, in Appendix C, the two losses are not directly minimized. The reported Lrest−Lunr,IL_{\mathrm{rest}}-L_{\mathrm{unr},I} is not an estimated reduction in the predictive performance, i.e., all information explained by the full minus all information that can be explained by only the reduced. And, as in W1, LrestL_{\mathrm{rest}} cancels anyway.

Eq (1) is hence not touched by this pipeline. And meanwhile, Eq (1) is not properly formulated, as the LHS and RHS are both random vectors, and if the inequality means holding a.s., I think they hardly do. A plausible fix may be adding expectations on both sides, which also somewhat agrees with my previous practical interpretation on NGC.

These issues substantially weaken the methodological foundation of the proposed predictive-causality-aware graph.

W3. Insufficient evidence for the necessity of the heterogeneous graph architecture.

According to Table 13, seven of the twelve datasets contain only two exogenous variables. Consequently, the exogenous–exogenous graph contains just one distinct inter-variable pair in these datasets, limiting the complexity of the relational structures being modeled.

Questions for authors
  1. Conditional versus pairwise predictive relationships: Why does the proposed NGC calculation consider only the endogenous variable and one exogenous variable at a time? In terms of standard causal inference, how does the method distinguish a variable's incremental predictive contribution from spurious importance caused by correlations with other exogenous variables?

  2. Could the authors compare MAGNet against appropriately matched DLinear ensembles, nonlinear DLinear variants, and simpler exogenous fusion mechanisms under comparable model capacity and hyperparameter tuning?

  3. Given that seven datasets contain only two exogenous variables, could the authors provide experiments on higher-dimensional exogenous systems or synthetic data with known dependency structures to establish the advantage of the proposed graph modeling?

  4. Granger causality formulation: Why do the restricted and unrestricted networks predict the current endogenous patch embedding rather than a future target from historical observations? How is the temporal ordering, which is the fundamental assumption for Granger-genre causal theory, enforced, particularly for the first patch?

7 comments

Thank you very much for the helpful critic.

W1, Q1: NGC Fundamental Problem

We extend NGC with sparse inputs following the same Granger causality concept. The Sparse Input Granger (SIG) process:

  • Pre-whitening step → removing chances of unrelated series looking related.

    x~t=xt−xt−s (t>s);x~t=0 (t≤s),\tilde{x}_t = x_t - x_{t-s}\ (t > s);\quad \tilde{x}_t = 0\ (t \le s),

    where ss is the number of steps per cycle.

  • cMLP prediction step → predict endo patch given all lagged exo patch (lag-0 used as nuisance only)

    Y^p=Wout⋅ReLU(∑v,kW1[v,k]⋅x~v,p−k+∑jW0[j]⋅x~j,p+b),\hat{Y}_p = W_{out} \cdot \mathrm{ReLU}\Big(\sum_{v,k} W_1[v,k] \cdot \tilde{x}_{v,p-k} + \sum_{j} W_0[j] \cdot \tilde{x}_{j,p} + b\Big),

    where W1W_1 and W0W_0 are the learned weights of lagged and current patches; W1W_1 covers exo and endo, W0W_0 only exo.

  • Loss function → group lasso loss function detached from the forecasting loss

    LossG=MSE(Y^p,Yp)+λ(∑∥W1[v,k]∥F+∑∥W0[j]∥F).\mathrm{Loss}_G = \mathrm{MSE}(\hat{Y}_p, Y_p) + \lambda\Big(\sum \Vert W_1[v,k]\Vert _F + \sum \Vert W_0[j]\Vert _F\Big).

    with λ\lambda the penalty score.

  • Weight assignment → after training, weight each exo's dependency on endo

    sj=∑k∥W1[j,k]∥F,sendo=∑k∥W1[endo,k]∥F,s_j = \sum_k \Vert W_1[j,k]\Vert _F,\quad s_{\text{endo}} = \sum_k \Vert W_1[\text{endo},k]\Vert _F,

    wj=sjsj+sendo.w_j = \frac{s_j}{s_j + s_{\text{endo}}}.

Further description:

  • No subtraction or cancellation of term
  • Conditional (not pairwise) dependency for better inference

Possible limitation: the metric is static (same for all patches); it shows how strongly an exo predicts an endo, while the graph-learning attention decides its relevance per patch.

Result: MAGNet is better on most short-term horizons, e.g. NP (by 1.9%), Sdwpfm1 (by 1.3%), and Energy (by 1.2%), while long-term results are more mixed.

W2, Q4: Causality Formulation

We have made several revisions:

  • Problem 1: Reconstruction, not prediction
    Lag-0 endo patch is strictly excluded from the cMLP input; the endo patch is predicted from exo/endo past and exo present only.

  • Problem 2: Loss minimization
    A group lasso loss is now minimized directly for an endo patch and its exo patches, resolving the previous concern.

  • Problem 3: Eq. 1 formulation
    We have modified Eq. 1:

    E∥Yt+1−E[Yt+1∣Y≤t,Z≤t]∥2 > E∥Yt+1−E[Yt+1∣Y≤t,Z≤t,X≤t]∥2.\mathbb{E}\big\Vert Y_{t+1} - \mathbb{E}[Y_{t+1} \mid Y_{\le t}, Z_{\le t}]\big\Vert ^2 \ >\ \mathbb{E}\big\Vert Y_{t+1} - \mathbb{E}[Y_{t+1} \mid Y_{\le t}, Z_{\le t}, X_{\le t}]\big\Vert ^2.

    where YY is the target, XX the additional exo variable, and ZZ the other exo variable, highlighting the conditional dependency in SIG.

  • Problem 4: First patches handling
    Patches before KK lack full K-lag history and are padded with the first patch; the loss is computed only after the lag, so padded patches are never training targets.

W3, Q3: Higher-dimensional Exogenous Datasets

We constructed 5 synthetic datasets with 5, 20, and 50 exogenous variables.

Dataset#Ex.Ex. DescriptionsEn. Description
Syn553 direct, 1 chain, 1 proxy, 0 irrelevantEq. 11 w/ β\beta = 0.284 (50% exo share)
Syn20203 direct, 1 chain, 4 proxies, 12 irrelevantEq. 11 w/ β\beta = 0.249 (50% exo effects)
Syn50503 direct, 1 chain, 10 proxies, 36 irrelevantEq. 11 w/ β\beta = 0.288 (50% exo effects)
Syn20weak20Same as Syn20Eq. 11 w/ β\beta = 0.127 (20% exo effects)
Syn20null20Same as Syn20Eq. 11 w/ β\beta = 0 (0% exo effects)

where chain means the variable is correlated to the direct variable, not target, and proxy means the variable is direct plus significant noise. Eq. 11 is as follow:

Yt=0.6Yt−1+0.3sin⁡(2πt24)+β(tanh⁡(X(1,t−24))+0.5(X(2,t−30)2−1)+X(3,t−36))+noise,(11)Y_t = 0.6Y_{t-1} + 0.3\sin\left(\frac{2\pi t}{24}\right) + \beta\left(\tanh(X_{(1,t-24)}) + 0.5\left(X^2_{(2,t-30)} - 1\right) + X_{(3,t-36)}\right) + \text{noise}, \quad (11)

Synthetic dataset result:

DatasetUnimodal ExoTSFMultimodal TSFUnimodal TSF
MAGNet (Ours)GCGNetTimeCMADLinearPatchTST
MAEMSEMAEMSEMAEMSEMAEMSEMAEMSE
Avg0.59430.57390.63600.65670.62980.63750.64000.66050.66310.7074

The full table can be seen as Table 1 in Section 5.2.

Result:

  • MAGNet improves more on short-term synthetic settings
  • Long-term results are less positive, yet MAGNet remains best on average.

Q2: MAGNet vs DLinear ensemble

We extended MAGNet's ablation with w/o Graph (DLinear with matched ensemble and configurations), w/ ExoMLP (graph replaced by an exo-aware MLP), and w/ Cross-attn (graph replaced by simple cross-attention).

Fullw/o Textw/o Exow/o Graphw/ Exo MLPw/ Cross-attn
MAE (Overall)0.47250.47690.49280.49430.52680.4858
MSE (Overall)0.52860.53670.56840.56240.59860.5461

Result: Full MAGNet performs best on both overall MAE and MSE.

Reviewer_Rz1kOpenReview ↗

Assessment of the revised manuscript

We deeply appreciate the authors taking my comments into serious consideration and offering a substantially revised manuscript. Given the substantial revision, I would first appreciate it if authors could post a general comment to actually summarize what has been changed. That would help reviewers in later discussions

By replacing the original NGC construction with separately trained SIG weights, my first and fundamental criticism W1 is decidely resolved.

As for W2, my judgment is that it has substantially improved, although there are still remaining conceptual limitations. Using group Lasso weight to quantify conditional predictive improvement is sometimes dangerous because it could be sensitive to the strength of penalty, and selecting this penalty inevitably double-uses the data. If you use cross-validation to select the hyperparameter, every now and then, the causality could be confounded, especially the revised model now introduces contemporaneous exogenous inputs (c.f. Eq (5) of the original). In this sense, Q1 is not sufficiently discussed, which is actually different from W1, although I would be less tempted to raise Q1 if this were the original submission. In terms of predictive performance, however, I think that is a good try.

That said, it's worth noting that contemporaneous exogenous variable indeed poses many risks on models from the simplest VARX to NGC, and even to conditional-independence-based methods like PCMCI, risks such as mediator blocking and intra-period simultaneity. This would involve many further discussions spanning ML and statistics, from Granger and Neyman-Rubin to Pearl to even the literature today, and the authors should add more (even informal) assumptions to strengthen this part if Granger-causality claim is still wanted to be maintained. In light of the difficulties, and given this being less relevant, I lower my confidence to 3, also because of the group-lasso part.

We have also seen a substantial and direct refinement on the experimental side. But now I think the ablation to be added would be the comparison with a homogeneous graph using the same parameter budget, or a model retaining the SIG prior but removing the exo-exo graph. Overall, the completeness of the experiments is largely improved, but as a standalone manuscript, I think it is still important to isolate the architectural novelty, which is still a weakness.

Finally and most importantly, [AC/SAC please read this part!], as a matter of procedure,

  • the central causal-scoring mechanism, optimization procedure (the pipeline, not just some details or minor lines), and source of edge weights are all changed, on top of the expanded evaluations.
  • On the other hand, the overall forecasting task, the mechanism related to text guidance and graph attention, and the basic idea remain related to, and recognizable from, the original manuscript.

So obviously, this is not entirely a new submission, but I think it really warrants a new round of review, and frankly, this would take too much time at this stage. And now new concerns may arise: for example, in terms of novelty, the sparse Granger-learning mechanism (which is established) combined with the multi-modal covariate information still lacks compelling evidence that it solves problems that simpler alternatives in the ML literature cannot address comparably well. This concern is suppressed before, but now this question stands out---and I think I'm not being hypercritical because if the main selling point is prediction, then modern competitors (like LLMs itself) must be included as competitors (approx. confidence 2~3), and if the selling point is causal discovery, my abovementioned concern tells why I would give at most a 2 with confidence 4.

Conclusion

My final judgement is a score of 2 with confidence 3 if all the revisions are accepted by AC, and 1 with confidence 4 if AC decides to exercise the right to "ignore changes that are significantly different from the original paper", a scenario definitely applicable here. Currently my attitude is the latter, so for now I do not change my rating in the system.

Thank you very much for the thoughtful comments.

General Summary of Changes

What has changed:

  • To better capture dependency measures from exo towards endo, we extend the prior Neural Granger Causality (NGC) to apply sparse input and loss function: Sparse Input Granger (SIG).
  • Additional independent training and group lasso loss function needed to accommodate the extension.
  • We added higher-dimensional synthetic datasets, as requested.
  • We added more ablation studies by replacing MAGNet's graph structure with simpler exo-aware structures (exoMLP, cross-attention network), as requested.
  • We added more experiments to support our claim of MAGNet's robustness as its own model and as a plug-and-play model.

What has not changed:

  • MAGNet remains solving the same problem: ExoTSF with text modality.
  • MAGNet's main framework and flow remains the same (offline text generation, graph construction, graph learning, and backbone forecasting).
  • MAGNet's use and purpose of graph learning to highlight exo-endo remains the same.
  • MAGNet's use of text modality as a guidance and not a fusion remains the same.
  • MAGNet's general forecasting loss and evaluation metric remains the same.
  • MAGNet's pool of comparison baselines and benchmarks (beside the requested synthetic additions) remains the same.
  • The extended dependency metric, SIG, hinges on the same concept of "Granger causality" with the prior dependency metric, NGC.

Disclaimer: SIG is not an entirely new dependency measure as it is simply a more rigorous extension of NGC that we choose to employ after receiving this reviewer's main criticism of NGC's flaws. Both SIG and NGC originate from the same concept of Granger causality, which stance remains the same through revisions.

In short, we understand the concern. We want to highlight that the changes are made as a response directly to each reviewer request. We leave the procedural decision to the AC.

Concern 1: Group Lasso Penalty Score

We conduct a sensitivity experiment on KK (amount of lag patches) and λ\lambda (penalty score) given the loss function. Results shown below is the average of dataset representatives (NP, Energy, Rapel, Sdwpfm1, and Syn20). Each cell shows MAE / MSE.

KK \ λ\lambda0.0010.010.11.0
10.4376 / 0.44600.4444 / 0.45430.4376 / 0.44660.4368 / 0.4459
20.4353 / 0.44390.4341 / 0.44130.4386 / 0.44600.4355 / 0.4474
30.4345 / 0.44050.4355 / 0.44040.4404 / 0.45080.4358 / 0.4445
50.4355 / 0.44390.4354 / 0.44340.4349 / 0.44180.4334 / 0.4416

Table result:

  • Both the choice of KK and λ\lambda simply vary over 2.5% (MAE) and 3.2% (MSE).
  • Our current implementation (KK = 3 and λ\lambda = 0.01) is within the best average MSE and 0.5% of the best average MAE.
  • Changing values of KK and λ\lambda within a reasonable threshold do not affect the performance significantly.

Concern 2: More Ablation Studies

We conduct more ablation studies as requested: w/ HomGraph and w/ SIG-NoExo

GroupMetricFullw/ HomGraphw/ SIG-NoExow/o Graph
Overall (3 seeds)MAE0.45050.4540 (+0.8%)0.4532 (+0.6%)0.4943 (+9.7%)
MSE0.47780.4837 (+1.2%)0.4810 (+0.7%)0.5624 (+17.7%)

Table result:

  • Full MAGNet still performs the best against w/ HomGraph and w/ SIG-NoExo
  • We note that the difference is not huge. It falls between Full and w/o Graph, which makes sense because w/ HomGraph and w/ SIG-NoExo take off the graph partially, while w/o Graph takes off the graph fully.

Concern 3: Contemporaneous Exogenous Inputs

  • We admit the caution of contemporaneous exogenous inputs affecting the lagged patches and ultimately the endo prediction in SIG.
  • We understand that our attempt in applying lag-0 as nuisance only reduces those chances, but not preventing it from happening entirely.

We conduct another ablation study to support this: we remove lag-0 patches at all from the dependency measure.

GroupExo weight rise without lag-0: mean per cellLag-1 shareMAE (with → without)MSE (with → without)
Overall (3 seeds)+52%0.37 → 0.430.4505 → 0.4505 (0.00%)0.4778 → 0.4781 (+0.06%)
Short+60%0.40 → 0.470.3946 → 0.3943 (−0.10%)0.3907 → 0.3890 (−0.43%)
Long+42%0.34 → 0.380.5162 → 0.5161 (−0.00%)0.5799 → 0.5822 (+0.39%)

Table result:

  • Removing lag-0 as nuisance rises the dependency weight of exo hugely with extra weights concentrated in lag-1 (nearest patch to present) → same-time exo leaks
  • Lag-0 existence does not change the forecasting MSE or MAE, which makes sense since its loss computation is detached from the overall forecasting loss.

Conclusion: Adding lag-0 as nuisance helps reduce contemporaneous exo input pitfalls with no cost in accuracy.

Reviewer_kaYfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a text-guided Multimodal Attentive Graph Network (named MAGNet) for time series forecasting with exogenous variables. The proposed MAGNet comprises three modules: Context-aware Text Generation, Exo-endo Graph Construction, and Causality-aware Graph Learning. The first module uses an offline pretrained language model to generate contextual descriptions that capture global semantic information about exogenous and endogenous variables. The remaining two modules employ graph learning to model intricate inter-variable dependencies, thus enhacing the forecasting performance with exogenous variables. Extensive experiments on 12 real-world datasets show that MAGNet outperforms existing forecasting baselines, including those using exogenous variables and multimodal data.

Critical strengths
  1. This paper addresses the problem of text-guided time series forecasting with exogenous variables, which has practical value in real-world applications. Figure 1 clearly shows how MAGNet differs from existing methods.

  2. The paper is well organized, with clear writing and well-designed figures and tables that help readers understand its main ideas.

  3. The authors provide anonymous source code for MAGNet, supporting the reproducibility of the experimental results.

Critical weaknesses
  1. The text modality relies on prompt design and a BERT encoder, yet the impact of varying prompts and language models on forecasting performance remains unexplored.

  2. Although Neural Granger Causality (NGC) is employed to model exogenous variable relationships (Section 6), the paper lacks a substantive analysis of how these weights capture useful variable dependencies or why NGC is particularly effective for this task.

  3. Graph learning has been widely studied for modeling relationships among time series variables. Recent studies have also explored hypergraph-based approaches, such as [R1]. However, the paper lacks discussion of these hypergraph-based methods and comparisons with related baselines.

  4. The coverage of multimodal time series modeling is limited to text, neglecting recent image-based studies [R2] in both the related work and conclusion, thereby restricting the scope of recent advancements addressed.

[R1] Ada-MSHyper: adaptive multi-scale hypergraph transformer for time series forecasting. Advances in Neural Information Processing Systems, 2024.

[R2] Multi-modal view enhanced large vision models for long-term time series forecasting. Advances in Neural Information Processing Systems, 2025.

Questions for authors
  1. Given the evolution of language models beyond BERT to newer open-source architectures such as GPT-3.5, Qwen and DeepSeek, could the authors examine how varying language model backbones impact MAGNet's forecasting performance?

  2. Table 2 demonstrates MAGNet's utility as a plug-and-play module; however, its integration with unimodal ExoTSF methods (e.g., DAG, GCGNet, and TimeXer) improves short-term forecasting while degrading long-term performance. Could the authors elucidate the underlying reasons for this discrepancy?

  3. As graph construction and learning may introduce computational overhead, could the authors analyze the impact of the Exo-endo Graph Construction and Causality-aware Graph Learning modules on MAGNet's overall training time?

3 comments

Thank you very much for your insightful review.

W1, Q1: Further text-based ablation studies

We construct three ablation tables: text encoder variation, text LLM generator variation, and text prompt variation. The result is averaged over 5 dataset representatives: NP, Energy, Rapel, Sdwpfm1, and Syn20.

  • Text encoder ablation → fixed LLM generator (Llama-3B) and fixed original prompt
BERT-fullBERTMiniLMRoBERTaMPNetBGE
Average MAE0.43240.43370.43060.43430.43710.4324
  • Text LLM generator ablation → fixed text encoder (BERT-full) and fixed original prompt
Llama-3.2-1BLlama-3.2-3BLlama-3.1-8BQwen2.5-3BQwen2.5-7BDeepSeek-7B
Average MAE0.43320.43240.43340.43290.43490.4339
  • Text prompt ablation → fixed text encoder (BERT-full) and fixed LLM generator (Llama-3B)
OursP1 (w/o Statistics)P2 (w/o Dependency)P3 (w/ only Metadata)
Average MAE0.43240.43320.43280.4346

Result:

  • MAGNet performance doesn't differ much given different text encoder and LLM generators.
  • LLM models with larger parameters don't guarantee a better performance.
  • Full text prompt shows better performance compared to its variants: w/o statistics, w/o dependency questions, and w/ only metadata.

All three full tables can be seen in Appendix F.

Q2: Plug-and-Play Discrepancy

Table 2 shows how MAGNet is better-equipped with non-exogenous backbones, especially on short-term settings. We conduct another experiment (Table 4) where we compare the MAE of MAGNet with shuffled exo inputs and without.

$\Delta$$\text{MAE}$ = MAE\text{MAE}{shuf\text{shuf}}$ $ - MAE\text{MAE}{ori\text{ori}}

DAGGCGNetTimeXerTimeKANDLinearPatchTSTTFT
Exo-aware?YesYesYesNoNoNoNo
Avg Δ\DeltaMAE, short+7.5%+1.3%+10.4%+22.8%+21.7%+20.3%+29.1%
Avg Δ\DeltaMAE, long+0.4%+0.4%+0.2%+2.8%+1.1%+0.2%+9.5%
Avg Δ\DeltaMAE, all+3.4%+0.8%+4.5%+11.3%+9.8%+8.5%+17.8%

Result:

  • Δ\DeltaMAE is bigger on non-exo backbones (DLinear, PatchTST, etc.) compared to exo-aware backbones.
  • Bigger Δ\DeltaMAE shows that shuffled MAGNet brings detrimental results to overall model performance, which means MAGNet signal is not considered redundant, vice versa.
  • Short-term performance shows a more prominent result compared to long-term due to general task difficulty, yet the pattern remains.

Q3: MAGNet's Training Time

We conduct an experiment (Section 6.3) to determine how much training time is taken for each main component of MAGNet.

DatasetData processingGraph constructionGraph learningForecastingOthersTotal / run
Real35.814.38.86.835.038 min 51 s
Synthetic40367~31416 min 16 s

Result:

  • Graph construction and learning takes a small percentage of the training time (below 15% and 9%, respectively) given its major role in the model.
  • Exception: synthetic datasets spend more time on graph construction (36%) due to a major increase in the number of exo variables (up to 50 variables), resulting in denser graphs.

W3, W4: Related Work

We have added both hypergraphs and image-based modality into our related work (Section 2.2, 2.3). We describe how hypergraphs added unnecessary complexities and image-based modality is more difficult to implement for ExoTSF problems with no guarantee of its helpfulness.

W2: NGC Significance

We decided to extend NGC with sparse inputs following the same Granger causality concept. This extension, Sparse Input Granger (SIG) follows these properties that NGC doesn't possess:

  • Strict K-lag history used (no future exo leakage)
  • Conditional dependency of endo towards all exo (not pairwise)
  • Current patch (lag-0) only used as nuisance, not weight assignment
  • Model (cMLP) trained over its own group loss function detached from the forecasting loss

With these properties, SIG can better capture the one-way dependency of lagged exo patches towards the current endo patch.

Thank you for the thoughtful and detailed response. The explanations have addressed the questions raised in my initial review and helped me gain a clearer understanding of the paper’s approach and contributions. In light of these clarifications, my previous concerns have been resolved, and my overall assessment of the work is more positive. I have therefore decided to raise my score and recommend acceptance of the paper.

Thank you very much for your support! I am happy to hear that your previous concerns have been resolved. Thank you very much for the score and the recommendation!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.