acceptodds
Under review as a conference paper at ICLR 2027

InRelCast: In-Context Relation Learning for Time Series Forecasting with Exogenous Variables

Abstract

Exogenous variables such as weather and schedules are essential for forecasting a target series, and their values over the forecast horizon are often known or predicted in advance. Prevailing methods learn the forecast end to end from the look-back window and the future exogenous variables by fusing their representations or modeling temporal and channel correlations. However, these two inputs play different roles. The future exogenous variables specify the conditions ahead, whereas how the target responds to them depends on the current state of the series, so the forecast has to judge the current relation between the exogenous conditions and the target. The evidence for this judgment lies in the look-back window, where each transition from one period to the next pairs its conditions with the response that followed. Thus, these observed transitions act as answered examples and the future transitions, whose responses are unknown, act as new cases, which is precisely the setting of in-context learning. Nevertheless, prevailing methods encode both inputs together and leave this pairing and its use to implicit learning. Therefore, we propose InRelCast, an In-context Relation learning framework for foreCasting with exogenous variables, which infers the relation of each future transition in context from the observed transitions. Specifically, each observed transition serves as a demonstration and each future transition as a query, whose inferred relation generates low-rank parameter increments of a shared residual forecaster that corrects a baseline forecast from the target history. Since future responses are unobserved, training uses later observed transitions as pseudo-queries, whose hidden responses supervise the relations inferred for them. Experiments on multiple real-world datasets show that InRelCast outperforms state-of-the-art baselines, both in-domain and in zero-shot transfer to unseen series. The code is available at https://anonymous.4open.science/r/InRelCast.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, exogenous variables

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_7uqeOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes InRelCast for time series forecasting with exogenous variables. The method combines an endogenous forecasting backbone with a residual forecaster whose parameters are modulated by inferred relation representations. Historical conditions and observed residuals are encoded as demonstrations, from which a context encoder and relation decoder infer representations for future prediction. Reconstruction and pseudo-query objectives support this learning process.

The paper evaluates the method on several forecasting benchmarks, including cross-dataset transfer and longer-horizon forecasting. The results are promising and show improvements over a broad set of baselines. My main reservation is whether the experiments sufficiently establish the proposed mechanism (learning predictive relations from historical demonstrations) as the source of these improvements.

Critical strengths
  1. A well-motivated forecasting problem. Exploiting historical target–covariate relationships to condition future predictions is a reasonable approach when these relationships vary across series or over time.

  2. A coherent architectural design. The combination of an endogenous backbone, residual modeling, and conditional parameter modulation is technically interesting. The reconstruction and pseudo-query objectives provide explicit training signals for the proposed relation representations.

  3. Broad and encouraging empirical results. The evaluation covers multiple application domains, transfer settings, and prediction horizons. The ablations examine several meaningful design choices, including relation inference, auxiliary objectives, context encoding, and parameter modulation. The supplementary fixed-configuration experiment also provides useful evidence that the results are not entirely dependent on selecting a separate configuration for each horizon.

Critical weaknesses

W1. The experiments do not yet isolate learning from historical demonstrations from ordinary conditional prediction.

The future query already contains an endogenous representation and exogenous information. A sufficiently expressive query-conditioned predictor could therefore perform well without substantially using the historical condition–response associations stored in the demonstration memory.

The existing ablations establish that parts of the proposed architecture are useful, but do not directly distinguish these explanations. Removing relation inference or replacing inferred relations with a pooled representation changes more than access to informative demonstrations. Similarly, comparing parameter modulation against input conditioning addresses how conditioning is implemented, rather than whether the model learns from historical demonstrations.

This distinction matters because in-context relation learning is the central contribution. A capacity-comparable query-only predictor, together with an intervention that shuffles the historical responses while preserving the conditions, would provide more direct evidence.

The interpretation of the latent codes also deserves some restraint. Since historical codes receive the observed residual and are trained to reconstruct it, successful reconstruction does not by itself establish that they represent a transferable relationship. They may encode residual-specific information. This does not invalidate the predictor, but limits what can be concluded about its mechanism.

W2. The coverage of the pseudo-query objective is limited for some long-horizon configurations.

The pseudo-query task is constructed by splitting the observed history, so its maximum prediction offset is bounded by the number of historical patches. This can be substantially shorter than the actual forecasting horizon. For example, with a context length of 96 and patch length of 12, there are eight historical patches, allowing at most six pseudo-query offsets under the described construction, whereas a horizon of 720 requires 60 future patches.

The forecasting loss still supervises the full output horizon; this is therefore not an absence of long-horizon training. However, the auxiliary objective directly teaches relation prediction over a much shorter range than the range over which the inferred relations are subsequently used. Consequently, the aggregate long-horizon results do not establish how much this objective contributes at distant forecast offsets. A breakdown of its benefit by prediction offset would clarify the scope of the claimed mechanism.

W3. Architectural variation and configuration selection make some performance gains difficult to attribute.

The released configurations for the common multivariate benchmarks enable options such as endo_direct and endo_linear, and occasionally exog_linear. These introduce alternative forecasting paths, including direct horizon prediction and linear contributions that bypass the relation-inference pathway. The appendix briefly mentions optional linear mappings, but their contribution to the reported results is not separately established.

These additions may be useful modeling choices. Nevertheless, the distinction matters when improvements are attributed to the proposed relation-learning architecture. Results for a clearly specified default model, accompanied by a compact ablation of these additional paths on representative datasets, would make the evidence easier to interpret.

The tuning protocol is also described more explicitly for some experimental settings than for others. Because several baseline results are taken from prior work, clearer reporting of the proposed method’s search space and selection budget would help assess comparability. I do not regard this as evidence of unfair tuning, and the fixed-configuration results partially mitigate the concern.

W4. The wind experiments suggest an additional explanation for part of the advantage over several baselines.

Table 16 is a valuable diagnostic: explicitly adding wind speed substantially improves several competing methods. For example, the average MSE decreases from 0.961 to 0.504 for DAG and from 1.067 to 0.523 for KITE. InRelCast remains competitive and retains its advantage, which is a positive result.

However, this also suggests that an appreciable part of the original gap to these methods reflects their ability to extract a useful nonlinear feature from the wind components. The experiment therefore supports the effectiveness of the overall architecture more directly than the specific claim about learning relations from demonstrations. Repeating the context-dependence control from W1 with wind speed supplied would help separate these explanations.

Questions for authors
  1. How does InRelCast compare with a capacity-comparable query-only modulation model? How much does performance change when historical responses are shuffled across demonstration conditions?

  2. How does the benefit of the pseudo-query objective vary with forecast offset, particularly when the forecasting horizon substantially exceeds the offsets covered by this objective?

  3. Which reported configurations use the direct or linear forecasting paths, and how much do these paths contribute relative to the default relation-based architecture? Please also clarify the configuration-selection protocol across benchmark groups.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 7uqe (1/2)

We thank the reviewer for the careful reading and for the precise suggestions, which we followed in the new experiments below. All new results use the evaluation protocol of the paper, and we mark the revised parts of the manuscript in blue.

InRelCast relies on the pairing in its demonstrations. We ran the two controls the reviewer proposed. Shuffle-within permutes the responses across the demonstrations of a window while keeping their conditions, Shuffle-cross takes the responses from another window, and both are applied at test time to the trained model. Query-only removes the demonstrations and is trained with a parameter count within 0.3% of InRelCast, so the comparison holds capacity fixed. We repeated the controls with three seeds on four datasets.

DatasetSeedsShuffle-withinShuffle-crossQuery-only
DE3+32.7% ± 6.0+26.2% ± 4.7+40.1% ± 1.8
Sdwpf3+24.8% ± 0.9+22.4% ± 0.5+26.8% ± 2.8
Sdwpf zero-shot, 8 pairs1+21.6%+19.9%+17.9%
GEFCom2014 + speed1+9.0%+8.6%+11.4%

Change in MSE relative to InRelCast of the same run, positive is worse, averaged over horizons 24 and 360, mean and standard deviation over seeds where three seeds are given.

On DE and the wind datasets, breaking the pairing between conditions and responses raises the error by 8.6%--32.7%, and the capacity-matched Query-only model is 11.4%--40.1% worse, with small spread across seeds. A query-only model of the same size therefore does not recover what the demonstrations provide. The effect persists in zero-shot transfer, where the target series is never seen in training and the demonstrations are its only series-specific information. On NP, FR, BE, PJM, and Energy, the changes stay within the seed-to-seed spread. Section 4.3 now reports these controls in Table 6, and Appendix D.6 gives the complete table on all nine settings with standard deviations.

We removed the strong reading of the latent codes. We agree. A historical code receives the observed residual and is trained to reconstruct it, so a low reconstruction error alone does not show that the code represents a transferable relation. We removed this reading from Section 4.3 and now present the relation visualization only as a description of how the inferred vectors vary with the exogenous conditions. The shuffle controls above are the evidence that the relations are used.

The pseudo-query loss helps at all offsets, including those beyond its range on ETTh1-720. The concern applies to our ETTh1-720 setting, where look-back 96 and patch length 8 let the pseudo-queries cover at most 10 offsets while the horizon has 90 patches. We split the test MSE by the true forecast offset and compared InRelCast with the variant trained without Licl\mathcal{L}_{\text{icl}}.

SettingFarthest pseudo-query offset / horizon (patches)Near thirdMiddle thirdFar thirdOffsets beyond pseudo-query range
DE-36045 / 30-19.2%-17.0%-14.6%none
Sdwpfh1-36067 / 45-1.6%-0.7%-1.5%none
ETTh1-72010 / 90-16.2%-44.1%-34.8%-35.6%

Change in test MSE from adding Licl\mathcal{L}_{\text{icl}}, negative means Licl\mathcal{L}_{\text{icl}} helps.

On ETTh1-720, Licl\mathcal{L}_{\text{icl}} reduces the error by 35.6% on the offsets that the pseudo-queries never reach. We attribute this to the offset embedding and the shared relation decoder, which let the inference learned on short offsets carry over to later queries. On the TSF-X datasets with long look-back windows, the pseudo-queries already cover the whole horizon. While writing this analysis we also found that Section 3.4 states m∈{1,…,K−2}m\in\{1,\ldots,K-2\}, whereas the code draws mm from {max⁡(1,⌊ρ(K−1)⌋),…,K−2}\{\max(1,\lfloor\rho(K-1)\rfloor),\ldots,K-2\} with ρ=0.25\rho=0.25 by default and ρ=0.1\rho=0.1 in this ETTh1-720 setting, and we corrected the text. The table is added as Appendix D.7.

AuthorsOpenReview ↗

Response to Reviewer 7uqe (2/2)

The default model has no bypass path, and it produces every TSF-X result. None of the configurations on the 12 TSF-X datasets, GEFCom2014 wind, or zero-shot transfer enables endo_direct, endo_linear, or exog_linear. Tables 1--3, the ablations, and the shuffle controls above come from the relation-based model alone, as does the ETTh1-720 offset analysis, whose configuration enables none of these paths, and our claims about learning relations from demonstrations rest on these results. The paths are optional components for some of the common multivariate benchmarks, namely ETTh1, ETTh2, ETTm2, and Electricity, where the exogenous variables are other channels of the same system and the target history carries most of the signal. On these benchmarks a direct path from the target history complements the relation pathway, which models the exogenous channels. We now list in Appendix A.2 which configurations enable each path, so that the default model and these variants are clearly separated.

The results do not depend on fine per-horizon tuning. On GEFCom2014 wind, InRelCast and every baseline are tuned with one protocol (Appendix A.2, Table 8), and in zero-shot transfer each model keeps its configuration for the source dataset. Starting from a base configuration, each run changes one hyperparameter among learning rate {1e-4, 5e-4, 1e-3}, width {64, 128, 256}, patch length {8, 12, 24}, and dropout, a final run merges the changes that lower the validation loss, and the run with the lowest validation loss is kept. On the TSF-X and common multivariate datasets, the baseline results are taken from DAG, GCGNet, and KITE where available and are otherwise obtained with the released configurations of the baselines, except GCGNet and KITE on the common multivariate datasets, which we tune with the same protocol (Appendix A.2), and the configuration of InRelCast for every dataset and horizon is released with the code. These results do not depend on fine per-horizon tuning. Table 26 shows that one configuration per dataset across both horizons still gives the best average MSE and MAE on the 12 TSF-X datasets, 0.289 and 0.306 against 0.321 and 0.333 for the second-best GCGNet, and the sensitivity study in Appendix E.2 shows that the MSE stays stable under the rank RR on all four datasets tested and under every hyperparameter on NP and GEFCom2014 zone 4.

The dependence on the demonstrations remains with wind speed supplied. We repeated the W1 controls on GEFCom2014 wind with the wind speeds at 10 m and 100 m added as exogenous variables, the setting of Table 21. Shuffle-within raises the MSE of InRelCast by 9.0%, Shuffle-cross by 8.6%, and the capacity-matched Query-only model is 11.4% worse. Supplying the nonlinear feature therefore does not remove the dependence on the demonstrations, and we attribute the remaining gap to how each wind farm converts the forecast wind into power. The results are included in Appendix D.6.

AuthorsOpenReview ↗

Follow-up before the end of the discussion

As the discussion period closes in a few hours, we would be grateful to know whether our response addresses your concerns.

The controls you proposed show that breaking the pairing in the demonstrations raises the error by 8.6% to 32.7% on DE and the wind datasets, and the capacity-matched Query-only model is 11.4% to 40.1% worse. Moreover, no TSF-X configuration enables a bypass path, so every TSF-X result comes from the relation-based model.

We are happy to clarify any remaining point.

Reviewer_LSe9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes InRelCast, an in-context relation learning framework for time-series forecasting with exogenous variables (TSF-X). The motivation is that existing representation fusion and correlation modeling methods generally process historical observations and future exogenous variables jointly within an end-to-end forecasting framework, without explicitly exploiting their distinct roles.

The paper formulates TSF-X as an in-context relation learning problem. Historical transitions provide condition-response demonstrations, while future transitions act as queries for relation inference. A context encoder and a relation decoder infer transition-specific relations. These relations dynamically adapt a shared residual forecaster through low-rank parameter updates. Historical pseudo-queries and an auxiliary ICL loss provide supervision for relation inference.

Experiments on multiple TSF-X benchmarks, wind power forecasting, and zero-shot cross-series transfer show improvements over competitive baselines. Ablation studies, relation visualizations, and efficiency analyses provide further evaluation of the framework.

Critical strengths
  1. This paper introduces a coherent methodological framework that connects historical condition-response demonstrations, in-context relation inference, and relation-conditioned forecasting. The historical pseudo-query mechanism provides supervision for relation inference when future responses are unavailable. These components fit together coherently and offer a plausible way to use historical response patterns for future forecasting.

  2. The empirical evaluation covers multiple TSF-X benchmarks, wind power forecasting and zero-shot cross-series transfer. The reported improvements over competitive baselines, while the ablation studies and efficiency analyses provide further evidence for assessing the framework.

Critical weaknesses
  1. The paper presents the formulation of TSF-X as In-Context Relation Learning as a central contribution, but its conceptual novelty appears limited. Prior work, including COSMIC (Auer et al., 2025), Chronos-2 (Ansari et al., 2025), and Baguan-TS (Yang et al., 2026), has already explored in-context forecasting with exogenous variables. In Section 2.2, the authors distinguish InRelCast by arguing that these methods do not explicitly infer relations between exogenous conditions and the target. This distinction appears relatively incremental, since existing approaches already use historical target and covariate information for in-context forecasting. InRelCast makes these relations explicit through transition-specific relation vectors that adapt the forecaster through LoRA. This design appears closer to an architectural refinement of existing ICL approaches than a fundamentally new problem formulation.

  2. The zero-shot transfer evaluation has limited diversity. The experiments are limited to transfers between different series within the same dataset. Table 3 evaluates transfers between SDWPF subsets and between wind farm zones in GEFCom2014. Within each dataset, the source and target series share the same data collection context and exogenous variable definitions. These settings may therefore offer a limited test of the proposed framework’s generalization capabilities. A stronger evaluation would examine zero-shot transfer between independent datasets with similar target variables and compatible exogenous features, such as different wind power forecasting datasets. This would help assess whether the learned relation inference mechanism generalizes beyond dataset-specific patterns without fine-tuning.

[1] Auer, A., et al. (2025). Zero-shot time series forecasting with covariates via in-context learning. arXiv preprint arXiv:2506.03128.

[2] Ansari, A. F., et al. (2025). Chronos-2: From univariate to universal forecasting. arXiv preprint arXiv:2510.15821.

[3] Yang, L., et al. (2026). Baguan-TS: A sequence-native in-context learning model for time series forecasting with covariates. In Proceedings of the International Conference on Machine Learning (ICML).

Questions for authors

The zero-shot experiments in Table 3 are conducted between different series within the same dataset. Could the authors clarify whether the proposed method has been evaluated on cross-dataset zero-shot transfer, particularly between independently collected datasets with similar target variables and compatible exogenous features (e.g., different wind power forecasting datasets)?

3 comments
AuthorsOpenReview ↗

Response to Reviewer LSe9 (1/2)

We thank the reviewer for the clear summary and for naming the closest prior work. We answer the two points below and mark all revisions of the manuscript in blue.

InRelCast differs from COSMIC, Chronos-2, and Baguan-TS in what forms an example. We agree that these models already forecast in context from the history of the target and the covariates, and we do not claim the use of ICL for TSF-X as new. Our contribution lies in what forms an example and what is inferred from it.

Unit of an exampleWhat is inferred for the futureHow the future covariates actCan the use of past pairs be tested
COSMICnone, exogenous and target patches are concatenatedthe forecast directlyas input tokensno explicit pairs
Chronos-2none, group attention across variablesthe forecast directlyas input tokensno explicit pairs
Baguan-TSa look-back segment and the segment that followedthe forecast directlyaligned with the segmentsat segment level
InRelCastone transition, the condition before a patch and the residual response that followedone relation vector per future transitionthe relation generates the weights that map them to the responseyes, by shuffling responses across conditions

Because each demonstration pairs a condition with its own response, the pairing can be broken while everything else is kept. Shuffle-within permutes the responses across the demonstrations of a window while keeping their conditions, Shuffle-cross takes the responses from another window, and both are applied at test time to the trained model. Query-only removes the demonstrations and is trained with a parameter count within 0.3% of InRelCast, so the comparison holds capacity fixed.

DatasetSeedsShuffle-withinShuffle-crossQuery-only
DE3+32.7% ± 6.0+26.2% ± 4.7+40.1% ± 1.8
Sdwpf3+24.8% ± 0.9+22.4% ± 0.5+26.8% ± 2.8
Sdwpf zero-shot, 8 pairs1+21.6%+19.9%+17.9%
GEFCom2014 + speed1+9.0%+8.6%+11.4%

Change in MSE relative to InRelCast of the same run, positive is worse, averaged over horizons 24 and 360, mean and standard deviation over seeds where three seeds are given.

On DE and the wind datasets, breaking the pairing raises the error by 8.6%--32.7%, and the capacity-matched Query-only model is 11.4%--40.1% worse. The effect persists in zero-shot transfer, where the demonstrations are the only series-specific information. On NP, FR, BE, PJM, and Energy, the changes stay within the seed-to-seed spread. In the ablation of Table 5, replacing the demonstrations with joint tokens of the historical target and exogenous patches, which is close to how COSMIC and Chronos-2 read the context, is the worst variant on Sdwpf and in zero-shot transfer.

We revised Section 2.2 to state the difference in these terms, and Section 4.3 now reports these controls in Table 6.

AuthorsOpenReview ↗

Response to Reviewer LSe9 (2/2)

On GEFCom2014 wind and fev-bench, InRelCast is more accurate than Chronos-2. On GEFCom2014 wind, InRelCast reaches an MSE of 0.366, against 0.433 for Chronos-2 fine-tuned on the same zones (Table 18). We also evaluated InRelCast on the six fev-bench energy tasks whose future covariates are known, the setting of our paper. We ran InRelCast, DAG, and KITE with one tuning protocol and selected them on the validation split, and took the other models from the official fev-bench results computed on the same windows, which we verified by reproducing the results of Chronos-2.

ModelENTSO-EGFC14GFC17Solar w/ weatherEPF-DEEPF-NPGeometric mean
InRelCast0.5060.4650.5840.8870.6240.7300.618
Chronos-20.5380.5390.6101.0200.6040.8500.673
TabPFN-TS0.5370.6350.8550.8660.5800.8830.711
KITE0.5080.7500.9010.8830.5560.8260.720
DAG0.5290.9731.0350.9040.7870.9930.850
TFT0.7510.4891.6250.8220.6560.8590.807
TimesFM-2.5*0.5850.1690.1901.0511.2801.4400.575
Seasonal Naive1.1011.1991.5851.0621.7091.7611.373

MASE, lower is better, best in bold among the models without *. *fev-bench reports TimesFM-2.5 as trained on GFC14 and GFC17.

InRelCast has the lowest geometric mean MASE among the models that fev-bench does not mark as trained on these tasks, 8.2% below Chronos-2. The table is added as Appendix C.2.

Across two independent wind datasets, InRelCast has the lowest error among the trained models. We ran the transfer the reviewer suggested between two independently collected wind power datasets. GEFCom2014 wind pairs ten Australian wind farms with numerical weather predictions, and Sdwpf pairs two turbines of the public SDWPF turbine array (Zhou et al., 2024) with ERA5 weather variables, so the two differ in site, collection, and the definition of the exogenous variables. We aligned the exogenous input to wind speed, the variable both datasets share, trained each model on one dataset, and applied it to the other without any update.

DirectionInRelCastDAGKITETimeXerChronos-2 (zero-shot)
GEFCom2014 → Sdwpf0.3400.3640.3860.4140.399
Sdwpf → GEFCom20140.5950.7140.7340.7920.454

MSE averaged over horizons 24 and 360, single seed. InRelCast, DAG, KITE, and TimeXer are trained on the source dataset only, and Chronos-2 is a pretrained foundation model applied without fine-tuning.

Among the models trained on the source dataset, InRelCast has the lowest error in both directions, 6.6% and 16.7% below the next best DAG. Against zero-shot Chronos-2, InRelCast is more accurate from GEFCom2014 to Sdwpf, where it learns from five wind farms, and less accurate from Sdwpf to GEFCom2014, where it learns from a single turbine. We attribute this reversal to the size of the source, since Chronos-2 is pretrained on a large corpus. The table is added to Appendix B as a third zero-shot setting, next to the transfer across Sdwpf subsets and across GEFCom2014 zones.

AuthorsOpenReview ↗

Follow-up before the end of the discussion

As the discussion period closes in a few hours, we would be grateful to know whether our response addresses your two concerns.

The new comparison separates InRelCast from COSMIC, Chronos-2, and Baguan-TS by what forms an example, and breaking the pairing in the demonstrations raises the error by 8.6% to 32.7% on DE and the wind datasets.

In transfer between the independent GEFCom2014 wind and Sdwpf datasets, InRelCast has the lowest error among the trained models in both directions.

We are happy to clarify any remaining point.

Reviewer_3VLpOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper considers the problem of forecasting with exogenous variables. For example, temperature in a electricity forecasting problem. The authors identify that the relationship between exogenous and endogenous variables is akin to in-context learning. The authors construct tuples comprising of the history of the endogenous time series (hih_i), the corresponding exogenous variables (viv_i), and their known future (vi+1v_{i+1}) as examples. Using these examples, the model learns to predict the endogenous target, conditioning on its history, the exogenous variables, and their inter-dependencies.

Critical strengths

The authors formulate TSF-X as an in-context relation learning problem. I believe this framing is quite interesting and novel. The paper also addresses an important problem, that has received much more attention recently.

Critical weaknesses

The paper is quite hard to follow, and it took me multiple passes of paragraphs to understand what the authors were trying to do. Figure 2 as it stands, with its caption is hard to follow. A few things are also unintuitive, for example, equation 4, and why we should modulate ψ\psi using generated parameter increments. The methods section is also littered with symbols, which one loses track of, after a while (ψ,u,v,v′,z,A,B,g,R,W,r,E,D,xˇ,x,...\psi, u, v, v', z, A, B, g, R, W, r, E, \mathcal{D}, \check{x}, x, ...) to name a few. It appears that parts of the paper were written by AI, as is apparent from sequences of words which add little value (e.g., generated parameter increments).

I am also concerned that the experimental settings do not inspire confidence. First, the value of exogenous forecasting, over vanilla multi-variate modeling is unclear. The only difference I see whether a subset of time series are available at test-time. It is unclear why the authors chose not to use benchmarks such as fev-bench or GIFT-Eval, each of which has a share of vetted, widely-used multivariate time series. Instead the time series datasets are fairly new, and sourced from a set of papers with the same set of authors. This in and of itself is not a bad thing, but the absence of other widely used benchmarks is concerning. Their absence is also not explained. Secondly, recent work [1, 2] has shown that not all multivariate time series benefit from multivariate modeling. Therefore, I wonder if these datasets are appropriate to demonstrate the benefits of exogenous forecasting. I do not think that the ablation studies themselves answer this question.

On baselines, some inaccurate claims are concerning. The authors use TiDE as an approach for exogenous forecasting, but the term "exogenous" is not mentioned in the paper. Meanwhile, Chronos 2 is supposed to natively support exogenous forecasting, but it is not compared. I would also encourage the authors to move away from the transformer-heavy baselines, to linear methods such as NBEATs.

Moreover, the datasets for forecasting are known to have issues. For example, in Exchange, repeating the last value can beat most forecasters. Most recent work use fev-bench, GIFT-Eval etc. for a much better read on forecasting performance.

Finally, some experimental settings are not motivated, for example: why evaluate zero-shot transfer at all?

References

  1. Potosnak, Willa, et al. "MICA: Multivariate Infini Compressive Attention for Time Series Forecasting." arXiv preprint arXiv:2604.06473 (2026).
  2. Żukowska, Nina, et al. "Towards long-context time series foundation models." arXiv preprint arXiv:2409.13530 (2024).
Questions for authors
  • Line 195: “We treat the window as a sequence of transitions i→i+1i \rightarrow i + 1,” Why can’t these transitions be from i to j, where j≥i+1j \geq i + 1?
  • Line 209: It is unclear what hih_i is, and what xˇ\check{x} is. Where is hih_i coming from?
  • Line 220: What is the goal of the condition and demonstration networks?
  • Line 229: What is the form of the memory?
  • Line 258: I am not sure what “the relation modulates ψ\psi through generated parameter increments”. Equation 5 does not make sense.
  • Line 327: What about the standard deviation? How do we know that the differences are significant?
  • Line 373: Zero-shot transfer: Why do these results matter? Motivate these experiments.
5 comments
AuthorsOpenReview ↗

Response to Reviewer 3VLp (1/3)

We thank the reviewer for the detailed reading. We revised the method section so that the text answers each of the reviewer's questions, added fev-bench and the requested baselines, and mark all revisions in blue.

A running example, a trace figure, and a notation table for Section 3. Following the reviewer's questions, we made three additions to Section 3. The overview paragraph now explains the method with one running example before any equation, namely that each transition spans one patch, that the K−1K-1 observed transitions are demonstrations, and that the JJ future transitions are queries. A new Figure 3 traces one window with K=3K=3 and J=2J=2 through the four steps, from the conditions and demonstrations to the relations and the generated weights of Eqs. 4 and 5, and Figure 2 now uses the same module names as the text. A notation table (Table 9, Appendix A.3) lists every symbol with its shape, and we tightened the wording of Sections 3 and 4, including the phrase the reviewer quoted.

In short, InRelCast works in three steps. A causal Transformer forecasts the target from its own history, and the residual of each observed patch is paired with its exogenous condition to form a demonstration. For each future patch, attention over these demonstrations infers a relation vector from the future exogenous condition. The relation then generates low-rank weight increments of a small forecaster, which adds the exogenous effect to the history-based forecast.

Q1. One-patch transitions keep demonstrations and queries comparable. Every observed and future transition spans one patch, so a demonstration and a query describe the same kind of step and the decoder can compare them directly. A transition from ii to jj with j>i+1j>i+1 would make the demonstrations depend on the gap j−ij-i, and the history holds few pairs for long gaps. Longer-range dependence still enters through the context encoder, which attends over all demonstrations, and through the relation decoder, which attends across all JJ queries and embeds the offset of each one.

Q2 and Q3. hih_i comes from the causal Transformer, and the two networks build the conditions and the demonstrations. The state hih_i is the output of the causal Transformer at patch ii, which depends only on the target patches up to ii (Eq. 1, Section 3.1). The condition network C\mathcal{C} encodes what is known before a response is observed, namely hih_i and the exogenous patches viv_i and vi+1v_{i+1}, and it is shared by demonstrations and queries so that both lie in one space. The demonstration network D\mathcal{D} binds the encoded response to its condition and yields one vector per observed transition. Both are two-layer MLPs over concatenated inputs, and Section 3.1 now states their roles in these words.

Q4. The memory keeps one vector per demonstration. The memory M∈R(K−1)×dmodelM\in\mathbb{R}^{(K-1)\times d_{\text{model}}} keeps one vector per demonstration after self-attention over all demonstrations, and the relation decoder cross-attends to it (Section 3.2).

Eqs. 4 and 5. The relation changes the mapping from conditions to response. Eq. 4 forecasts each future patch as the history-based baseline plus a residual Δyj\Delta y_j predicted by Ψ\Psi from the exogenous condition of transition jj. Eq. 5 gives transition jj its own weight W(j)=W+gjRB(rj)A(rj)W^{(j)}=W+\frac{g_j}{R}B(r_j)A(r_j), where a hypernetwork maps the relation rjr_j to a rank-RR increment as in LoRA and a gate gjg_j computed from rjr_j scales it. The same wind forecast can lead to different power outputs in different weather regimes, so the relation should change the mapping from conditions to response rather than add one more input. Adding rjr_j to the input tokens of Ψ\Psi instead is less accurate on all six datasets of Table 20, with an average MSE of 0.229 against 0.211. Section 3.3 now gives this reading after each equation, and Figure 3 (d) shows it for one window.

AuthorsOpenReview ↗

Response to Reviewer 3VLp (2/3)

InRelCast leads the fev-bench energy tasks among models not trained on them. We evaluated InRelCast on the six fev-bench energy tasks whose future covariates are known, the setting of our paper. We ran InRelCast, DAG, and KITE with one tuning protocol and selected them on the validation split, and took the other models, including TFT, from the official fev-bench results computed on the same windows, which we verified by reproducing the results of Chronos-2.

ModelENTSO-EGFC14GFC17Solar w/ weatherEPF-DEEPF-NPGeometric mean
InRelCast0.5060.4650.5840.8870.6240.7300.618
Chronos-20.5380.5390.6101.0200.6040.8500.673
TabPFN-TS0.5370.6350.8550.8660.5800.8830.711
KITE0.5080.7500.9010.8830.5560.8260.720
DAG0.5290.9731.0350.9040.7870.9930.850
TFT0.7510.4891.6250.8220.6560.8590.807
TimesFM-2.5*0.5850.1690.1901.0511.2801.4400.575
Seasonal Naive1.1011.1991.5851.0621.7091.7611.373

MASE, lower is better, best in bold among the models without *. *fev-bench reports TimesFM-2.5 as trained on GFC14 and GFC17.

InRelCast has the lowest geometric mean MASE among the models that fev-bench does not mark as trained on these tasks, 8.2% below Chronos-2. GIFT-Eval provides almost no covariates that are known over the forecast horizon, so it tests a different task, and we chose fev-bench as the external benchmark. The table is added as Appendix C.2.

Most datasets are public benchmarks that DAG reused. Most of our datasets are public benchmarks that DAG reused rather than created. NP, PJM, BE, FR, and DE are the public electricity price benchmark of Lago et al. (2021), which fev-bench also includes, and GEFCom2014 wind is a public forecasting competition. The four Sdwpf subsets were taken by DAG from two turbines of the public SDWPF dataset (Zhou et al., 2024), and only Energy, Colbun, and Rapel were collected by DAG itself. Appendix A.1 now gives the original source of each dataset.

Future exogenous values reduce the MSE by a further 30.9%--52.9%. We agree that multivariate inputs do not help on every dataset, so we compared three input levels with the same architecture.

DatasetTarget only+ past exogenous+ future exogenous
NP0.4580.4210.267
DE0.6550.6030.328
Energy0.1840.1520.105
Sdwpf0.8160.7630.359
GEFCom20141.151–0.366

MSE of InRelCast averaged over horizons 24 and 360. The last two columns are the results of Tables 11, 1, and 2, and Target only is a single-seed run with the configuration of the past-only setting. The paper evaluates GEFCom2014 wind only with future exogenous variables, so it has no past-only entry.

Past exogenous variables reduce the MSE by 6.5%--17.4%, and their future values reduce it by a further 30.9%--52.9%. TSF-X therefore differs from multivariate forecasting in the information it uses, namely exogenous values that are known over the horizon, such as day-ahead load forecasts and numerical weather predictions. The table is added as Appendix D.8.

TiDE uses the future covariates of our setting. TiDE models dynamic covariates whose values are known over the horizon, which are the future exogenous variables of our setting. We now use the term covariates when describing TiDE in the introduction.

InRelCast outperforms fine-tuned Chronos-2 on GEFCom2014 wind. Appendix C.1 (Table 18) compares InRelCast with Chronos-2 on GEFCom2014 wind on the same test windows. InRelCast reaches an MSE of 0.366, against 0.433 for Chronos-2 fine-tuned with its official procedure and 0.659 for zero-shot Chronos-2. Chronos-2 is also in the fev-bench table above, and Section 4.2 now refers to both comparisons.

AuthorsOpenReview ↗

Response to Reviewer 3VLp (3/3)

InRelCast outperforms NBEATSx on all five datasets. The paper already includes CrossLinear, TiDE, and TimeKAN, and we added NBEATSx, which receives both historical and future exogenous variables and is tuned with the protocol of Appendix A.2.

DatasetNBEATSx MSEInRelCast MSE
NP0.3850.267
DE0.4080.328
Energy0.2170.105
Sdwpf0.7390.359
GEFCom20140.9460.366

InRelCast reduces the MSE of NBEATSx by 19.6%--61.3% on these datasets. NBEATSx is added to Appendix B.

The last-value forecast is strong on Exchange only at short horizons. The eight common multivariate datasets, including Exchange, follow DAG and serve as a robustness check in Table 4, and the main evidence comes from the TSF-X datasets, GEFCom2014 wind, and fev-bench. We agree that the last-value forecast is hard to beat on Exchange at short horizons. We evaluated it without training under the same split, look-back window, and normalized metrics as Table 4. At horizon 96 it reaches an MSE of 0.088, below the 0.093 of InRelCast, and only DAG is lower with 0.087 among the eleven models, as expected for a series close to a random walk. However, its error grows quickly with the horizon, so its average MSE over the four horizons is 0.411, against 0.190 for InRelCast, and only GCGNet, TimeKAN, and PatchTST have a higher average.

MSE on Exchange96192336720Avg
Last value0.0880.1870.3701.0010.411
InRelCast0.0930.1440.1710.3510.190

We added this naive baseline to Table 15 in the revised version.

Zero-shot transfer matches new sites and tests the central claim. A new wind farm or region has no history long enough to train a model, so the forecaster has to be applied to it directly. Zero-shot transfer also tests the central claim of the paper, since InRelCast infers the relations from the demonstrations of each window without updating any parameter, so it should adapt to an unseen series through its own history. Section 4.2 now opens the zero-shot paragraph with this motivation.

The gains exceed the seed-to-seed standard deviation. All main results of the paper are means over three seeds, and Table 25 (Appendix E.1) gives the standard deviations. On the TSF-X and GEFCom2014 datasets in that table, the standard deviation of InRelCast is at most 0.017 MSE. In the 12 of these 16 settings where InRelCast is best, its gap to the second-best model is 0.010--0.103 MSE, which exceeds the standard deviation in every setting and is more than twice it in 11. Section 4.1 now points to Table 25.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Summary of revisions. We thank all reviewers for their constructive comments. We uploaded a revised version in which every new or changed passage is marked in blue. Table bodies are not marked, so blue numbers in the result tables still denote the second-best model, and new tables are marked by blue captions. The method and the experimental protocol are unchanged, table and figure numbers below refer to the revised version, and the main revisions are as follows.

  • Method presentation (Reviewer 3VLp). Section 3 now opens with a running example, a new Figure 3 traces one window with K=3K=3 and J=2J=2 through all steps, Figure 2 uses the module names of the text, and Appendix A.3 adds a notation table.

  • Dependence on the demonstrations (Reviewers 7uqe and LSe9). Section 4.3 adds the shuffle controls and a capacity-matched Query-only model (Table 6), and Appendix D.6 gives the complete results, with standard deviations over three seeds on four datasets.

  • External benchmark and in-context forecasters (Reviewers 3VLp and LSe9). Appendix C.2 adds the six fev-bench energy tasks with known covariates, and Appendix C.1 compares InRelCast with fine-tuned and zero-shot Chronos-2.

  • Zero-shot transfer across independent datasets (Reviewer LSe9). Appendix B adds transfer between GEFCom2014 wind and Sdwpf in both directions (Table 16).

  • Baselines and diagnostics (Reviewers 3VLp and 7uqe). Appendix B adds NBEATSx (Table 17) and the last-value forecast on Exchange (Table 15). Appendix D.7 splits the benefit of the pseudo-query loss by forecast offset, Appendix D.8 compares three levels of exogenous input, and Appendix A.2 lists the optional bypass paths and the configurations that enable them.

  • Datasets and evaluation (Reviewer 3VLp). Appendix A.1 gives the original source of each dataset, Section 4.2 motivates zero-shot transfer, and Section 4.1 points to the standard deviations in Table 25.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.