acceptodds
Under review as a conference paper at ICLR 2027

NMask: A Lightweight Network via Masked Alignment for Time Series Forecasting with Exogenous Variables

Abstract

Time series forecasting with exogenous variables is crucial in many real-world applications, as exogenous information often provides complementary signals beyond the target series itself. In particular, future exogenous variables, when available, offer informative priors about future dynamics and can substantially improve forecasting performance. However, existing methods often treat future exogenous variables as auxiliary features to be fused with historical observations, leading to heavy architectures and inefficient interaction modeling. In this paper, we revisit this problem from a new perspective and propose NMask, a lightweight network that explicitly leverages the inherent temporal correspondence between future exogenous variables and future prediction targets. Through a Masked Alignment mechanism, NMask associates learnable tokens with future exogenous information for future endogenous forecasting, while a Target-Conditioned Exogenous Attention module captures delayed effects of exogenous variables by modeling short-term lags and long-term temporal dependencies. Extensive experiments on multiple datasets demonstrate that NMask consistently outperforms existing forecasting methods with exogenous variables, with significantly fewer parameters and lower computational cost. Code is available at https://anonymous.4open.science/r/NMask.

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

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uP8pOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

NMask forecasts a target series from its history and historical and future covariates. It aligns learned future target tokens with covariate patches, uses target-conditioned attention over local windows and pooled summaries, and decodes each future token through a shared patch head. The paper reports results on twelve energy-related datasets and also studies a variant without future covariates at inference.

Critical strengths
  1. The architecture offers a compact way to connect future covariates to their corresponding prediction positions. A shared token-to-patch head avoids a large flattened forecasting head, while the local window and pooled summaries provide paths for delayed effects. The reported NP efficiency comparison is encouraging, though it establishes efficiency in that setting rather than at arbitrary covariate dimensionality.

  2. The evaluation covers twelve datasets and reports both horizon-averaged and individual-horizon results. NMask achieves the lowest average MSE and MAE on all twelve datasets in Table 1. Table 2 also tests the interaction module and several decoding choices, and the anonymous code helps recover details omitted from the text.

Critical weaknesses
  1. The information settings need a more precise account. Appendix A.1 lists issued EPF forecasts alongside generation measurements, hydropower precipitation/inflow and ERA5 meteorological variables. The paper does not establish that every future input path was available at the forecast origin, especially for the 360-step experiments. A comparison conditional on known future covariates is valid, but its practical interpretation depends on this distinction. The no-future variant also needs a complete training objective: the released code uses future covariates as auxiliary training targets while replacing them with learned tokens in the predictor. This is not automatically leakage, but the paper should distinguish training supervision from inference inputs and show how much the auxiliary task contributes.

  2. The evidence does not yet isolate the proposed alignment mechanism or fully specify it. TiDE already connects each horizon representation to that time step's covariates, so the novelty claim needs to distinguish this existing correspondence from NMask's token alignment. Table 2 replaces alignment with a different MLP fusion path, changing more than alignment alone. Its reported NMask numbers match the 360-step results, but the horizon is not specified in the caption. Equations (11)–(13) also omit the gate parameterization and fusion Q/K/V definitions, although these can be recovered from the code. Finally, the broad short- and long-horizon claim should reflect Table 5: DAG has lower short-horizon MSE on several datasets, including PJM and Rapel. These do not negate the average gains, but they limit the current mechanism and performance conclusions.

Reference: Das et al. Long-term Forecasting with TiDE: Time-series Dense Encoder. TMLR, 2023. Section 4.2.

Questions for authors
  1. Which future inputs are forecasts issued before each origin, and how is the 360-step path obtained? For the no-future variant, what supervision, loss weights and validation/test inputs were used, and what is the effect of removing the auxiliary covariate loss?

  2. Can you specify the TCEA gates, fusion and W/R settings, confirm the horizon used in Table 2, and clarify whether the alignment ablation preserves the same attention, head and parameter budget? Which result isolates the new alignment beyond an existing same-time covariate decoder?

4 comments
AuthorsOpenReview ↗

Response to Reviewer uP8p (Part I)

We would like to sincerely thank Reviewer uP8p for providing a detailed review and insightful comments. We have revised our paper accordingly.

Reply to W1 & Q1:

(1) The availability of future covariates needs further clarification.

Thank you for your valuable comment! The datasets, baselines, and experimental settings in our paper follow DAG [1] and GCGNet [2] to ensure fair and consistent comparisons. In the original experiments, we use the ground-truth values of future exogenous variables to simulate their availability. We have added the relevant references in lines 358–360 of the revised paper‾\underline{\text{in lines 358–360 of the revised paper}} (highlighted in blue).

Considering that future exogenous variables may not be fully available in real-world scenarios, we further use TiRex-2 [3], a model supporting multivariate forecasting, to predict future exogenous variables using only historical data available before each prediction origin. These predictions cover the entire forecasting horizon and are used as future exogenous inputs to models. NMask achieves the best overall performance. The average results are presented below, with the complete results reported in Table 7 of the revised paper‾\underline{\text{in Table 7 of the revised paper}}:

DatasetsNMaskDAGGCGNetTimeXerTiDE
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
NP0.4220.3640.4290.3760.4300.3780.4370.3830.4620.412
PJM0.1230.2080.1260.2100.1320.2150.1420.2280.1570.252
BE0.4420.2780.4570.2950.4580.3050.4710.3020.5420.344
FR0.4300.2340.4340.2490.4380.2440.4500.2460.4950.287
DE0.5980.4780.6070.4880.6120.4980.6540.5060.6360.514
Energy0.1590.3090.1610.3090.1630.3090.1690.3240.1610.310
Sdwpfm10.7050.5660.7050.5960.7070.5820.7060.6050.7080.629
Sdwpfm20.7730.6090.7830.6510.8020.6300.7980.6480.8120.677
Sdwpfh10.7380.6140.7420.6380.7410.6230.7420.6370.8000.693
Sdwpfh20.8420.6640.8670.6910.8820.6910.8880.7160.9160.743
Colbun0.1140.1480.1160.1530.1180.1600.1290.2150.1840.236
Rapel0.2510.2850.2570.2860.2630.2850.3410.3590.3230.352

[1] DAG: A dual correlation network for time series forecasting with exogenous variables. ICML, 2026.

[2] GCGNet: Graph-consistent generative network for time series forecasting with exogenous variables. ICLR, 2026.

[3] TiRex-2: Generalizing TiRex to Multivariate Data and Streaming. NeurIPS 2026.

(2) Further clarification of the model variant without future exogenous variables.

For the model variant without future exogenous variables, we use two L1 losses during training: a prediction loss between the predicted and ground-truth future endogenous variables, and an auxiliary reconstruction loss between the predicted and ground-truth future exogenous variables. The weight of the auxiliary reconstruction loss is set to 0.5 by default. During validation and testing, only historical endogenous and exogenous variables are used as inputs, without access to ground-truth future exogenous variables. We also provide pseudocode in Algorithm 1‾\underline{\text{pseudocode in Algorithm 1}} to further clarify the implementation.

(3) The impact of the auxiliary exogenous reconstruction loss on model performance.

We conduct an additional ablation study by removing the auxiliary covariate reconstruction loss while keeping all other components unchanged, as shown in the table. These results demonstrate that the auxiliary reconstruction loss improves the representation learning of future exogenous tokens and enhances forecasting performance when future exogenous variables are unavailable.

DatasetsNPPJMDEEnergyAverage
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(a) Remove Auxiliary Covariate Loss0.6470.4940.1850.2620.8010.5750.2370.3920.4680.431
NMask (without future)0.6380.4870.1800.2590.7690.5570.1970.3470.4460.413
AuthorsOpenReview ↗

Response to Reviewer uP8p (Part II)

Reply to W2:

(1) The novelty of the Masked Alignment mechanism proposed in NMask compared with TiDE.

Thank you for your valuable comment! We acknowledge that TiDE already exploits temporal correspondence. The key distinction of NMask lies not in establishing temporal correspondence itself, but in how this correspondence is leveraged to learn future endogenous representations.

TiDE concatenates future time-step representations with their corresponding exogenous variables during decoding and fuses them through an MLP. In contrast, NMask represents future endogenous variables as learnable tokens, allowing them to model temporal dependencies with historical endogenous variables during encoding and interact with temporally aligned future exogenous tokens through Masked Alignment. Unlike TiDE, which directly fuses all exogenous features, Masked Alignment enables future endogenous tokens to adaptively attend to more informative exogenous information based on their specific needs.

The average forecasting results of NMask and TiDE are presented below, with the complete results reported in Table 1 of the paper‾\underline{\text{in Table 1 of the paper}}. NMask consistently achieves lower forecasting errors than TiDE across all datasets.

DatasetsNMaskTiDE
MetricsMSEMAEMSEMAE
NP0.3080.3200.4430.400
PJM0.0850.1780.1420.246
BE0.4100.2780.4980.325
FR0.3800.2160.4840.281
DE0.3250.3540.4990.447
Energy0.0860.2280.1530.302
Sdwpfm10.3750.3990.4830.507
Sdwpfm20.4240.4320.4860.516
Sdwpfh10.3650.4110.4530.508
Sdwpfh20.4130.4560.5990.583
Colbun0.0970.1470.1640.227
Rapel0.2230.2600.3200.351

(2) The ablation study in Table 2 does not isolate the contribution of temporal alignment, and the prediction horizon of 360 is not specified.

Thank you for your valuable comment! The existing MLP fusion ablation mainly aims to evaluate the effectiveness of the overall interaction module. To further isolate the contribution of temporal alignment, we conduct an additional controlled ablation study by removing only the temporal alignment constraint while keeping the attention architecture, shared prediction head, and parameter count unchanged. The ablation results demonstrate the importance of explicit temporal alignment, as NMask consistently outperforms the "Remove temporal alignment constraint" variant.

DatasetsNPPJMDEEnergyAverage
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(a) Remove temporal alignment constraint0.4720.4310.1200.2160.4230.4170.1340.2880.2870.338
NMask (ours)0.4540.4190.1100.2090.4000.3950.1140.2680.2700.323

In addition, the prediction horizon used in Table 2 is 360. We have explicitly specified this in the table caption in lines 379 of the revised paper‾\underline{\text{in lines 379 of the revised paper}} (highlighted in blue) .

AuthorsOpenReview ↗

Response to Reviewer uP8p (Part III)

(3) Equations (11)–(13) lack details on the parameterization of the TCEA gating mechanism and the definitions of Q/K/V.

Thank you for your valuable comment! We have added the parameterization of the TCEA gating mechanism and the specific definitions of Q/K/V in in lines 287-300 of the revised paper‾\underline{\text{in lines 287-300 of the revised paper}} (highlighted in blue).

Specifically, TCEA consists of two stages: intra-variable aggregation and inter-variable fusion.

In the intra-variable aggregation stage, the linear projection of the endogenous patch representation Ht\mathbf{H}_t serves as the Query, while the linear projections of exogenous patch representations serve as the Keys and Values. Attention is applied separately to the local historical window and global summaries to obtain the historical and global summary contexts.

The gating mechanism uses two independent learnable linear layers followed by sigmoid functions. Each gate takes the concatenation of the endogenous patch representation, the Value projection of the current exogenous patch, and the corresponding context representation to compute a gating weight for either the historical or global summary context. The gate parameters are shared across exogenous variables and time positions. The two contexts are then combined with the current exogenous representation through gated weighted summation.

In the inter-variable fusion stage, the linear projection of the endogenous patch representation again serves as the Query, while the linear projections of the aggregated representations of different exogenous variables serve as the Keys and Values. Multi-head cross-attention is then used to adaptively fuse information from different exogenous variables.

(4) NMask underperforms DAG in short-term forecasting on PJM and Rapel.

Thank you for pointing this out! This difference may be related to the dependency structures of the datasets. In PJM, the two load-related covariates are highly correlated, resulting in relatively homogeneous exogenous information. In Rapel, water level variations are influenced by historical states and the cumulative effects of rainfall and inflow. These characteristics may favor DAG's inductive bias toward explicitly modeling and transferring historical dependencies.

In contrast, NMask's advantages depend on the heterogeneity of exogenous information and its temporal correspondence with the target variable, which may explain its better short-term forecasting performance on other datasets.

Reply to Q2:

(1) Can you specify the TCEA gates, fusion, and W/R settings, and confirm the prediction horizon used in Table 2?

Thank you for your valuable comment! The details of the TCEA module and the clarification of the prediction horizon used in Table 2 have been provided in W2.

W/R settings: WW denotes the local window size, including the current position, while RR denotes the number of global summaries obtained through adaptive average pooling. We use W=3W=3 and R=2R=2 by default.

(2) Does the alignment ablation preserve the same attention architecture, prediction head, and parameter budget?

The original alignment ablation did not preserve the same prediction head and parameter budget. To address this concern, we have added a controlled ablation experiment, as described in W2-2.

(3) Does NMask’s temporal alignment mechanism outperform existing same-time covariate decoders?

We will conduct an additional comparison with an existing same-time covariate decoding approach. Specifically, the encoder first generates representations for the future SS endogenous patches, which then interact with future exogenous patches at the corresponding time positions. Finally, predictions are generated through a shared prediction head.

By keeping the attention architecture, shared prediction head, and model parameter count as consistent as possible, we replace only Masked Alignment to further isolate and evaluate the independent contribution of the proposed alignment mechanism. The ablation results demonstrate that NMask’s temporal alignment mechanism is more effective than existing same-time covariate decoders.

DatasetsNPPJMDEEnergyAverage
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(a) Replace Masked Alignment0.4650.4280.1200.2150.4190.4080.1220.2770.2820.332
NMask (ours)0.4540.4190.1100.2090.4000.3950.1140.2680.2700.323

Dear Reviewer,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If it does, we kindly request that you reconsider the score. If you have any additional suggestions, we are more than willing to engage in further discussions and make necessary improvements to the paper. Thanks again for dedicating your time to enhancing our paper!

Kind Regards,

Authors

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

This paper addresses the problem of time series forecasting with exogenous variables, with a particular focus on the interaction between future exogenous variables and future prediction targets. It proposes NMask, which aligns learnable future target tokens with future exogenous patches through Masked Alignment. The TCEA module is introduced to capture the lag effects of exogenous variables, and a shared patch-wise predictor is employed to reduce predictor complexity. Experiments demonstrate that NMask achieves good results.

Critical strengths
  1. The motivation is clear and valuable. This paper argues that future exogenous variables naturally have a clear temporal correspondence with future prediction targets. The examples provided are insightful, and the methods are designed around this motivation, making it easy to follow.
  2. The core technology is easy to understand and intuitive. The example in the Introduction clearly illustrates how it differs from other methods.
Critical weaknesses
  1. The experiments are not sufficiently solid, and certain details regarding the experimental setup need to be clarified, which could undermine the author's claims.
  2. There are a few presentation issues need to be resolved.
Questions for authors
  1. This paper compares 9 baselines, but the experiment section claims that 10 were compared.
  2. This paper does not specify whether each dataset provides future exogenous variables that were already available at the time of prediction or ground-truth future exogenous variables observed retrospectively; the latter should not be considered available.
  3. When future exogenous variables are unavailable, the auxiliary loss and exogenous channel attention module used are not described in detail.
  4. Eq.1-3 appear to be missing a right parenthesis.
  5. The paper states that the encoder consists of L stacked blocks, but Eq.7–8 do not include the output from the previous layer. Furthermore, the second part of Figure 3 does not show the output from the previous layer feeding into the next layer. As it stands, the figure appears to show two series inputs passing through a single encoder layer and then merging into a single series.
  6. Section 5.2 MODEL ANALYSES is an empty section.
7 comments
AuthorsOpenReview ↗

Response to Reviewer TUAv (Part I)

We would like to sincerely thank Reviewer TUAv for providing a detailed review and insightful comments. We have revised our paper accordingly.

Reply to W1 & Q1: This paper compares 9 baselines, but the experiment section claims that 10 were compared.

Thank you for pointing this out! This was a typo, as we actually compared 9 baseline models. We have corrected it in lines 362 of the revised paper‾\underline{\text{in lines 362 of the revised paper}} (highlighted in blue).

Reply to Q2: Please refer to Response (Part II).

Reply to Q3: When future exogenous variables are unavailable, the auxiliary loss and exogenous channel attention module are not described in detail.

Thank you for your valuable suggestion! We have added descriptions of the auxiliary reconstruction loss and exogenous channel attention module in lines 468-473 of the revised paper‾\underline{\text{in lines 468-473 of the revised paper}} (highlighted in blue). We also provide pseudocode in Algorithm 1‾\underline{\text{pseudocode in Algorithm 1}} to further clarify the implementation. A brief description is provided below:

When future exogenous variables are unavailable, we replace them with learnable future exogenous tokens. To utilize future exogenous information during training, we introduce a future exogenous reconstruction loss to enhance the representation learning of these tokens. To improve reconstruction, we also add an exogenous Channel Attention layer after TCEA to capture dependencies among exogenous variables. Importantly, Channel Attention is used only for auxiliary reconstruction, while TCEA still interacts with exogenous representations before Channel Attention, leaving the original endogenous-exogenous interaction unchanged. The ground-truth future exogenous variables are used only for auxiliary supervision during training and are not used as prediction inputs.

Reply to W2 & Q4: Eq. 1–3 appear to be missing a right parenthesis.

Thank you for pointing this out! The missing right parentheses in Equations (1)–(3) were formatting errors, which we have corrected in lines 211, 215, 240 of the revised paper‾\underline{\text{in lines 211, 215, 240 of the revised paper}}.

Reply to Q5: Equations (7)–(8) and Figure 3 do not clearly illustrate the stacked encoder structure and information flow between layers.

Thank you for pointing this out! We have revised the relevant equations in lines 267-279 of the revised paper‾\underline{\text{in lines 267-279 of the revised paper}} (highlighted in blue) to clarify the information flow between encoder layers. Specifically, H~(ℓ)\widetilde{\mathbf{H}}^{(\ell)} and Z~(ℓ)\widetilde{\mathbf{Z}}^{(\ell)} denote the endogenous and exogenous representations after FFN processing at layer ℓ\ell, respectively. They serve as the final outputs of the current layer and are passed to layer ℓ+1\ell+1. We have also updated Figure 3 accordingly.

Reply to Q6: Section 5.2 MODEL ANALYSES is an empty section.

Thank you for pointing this out! Section 5.2, "MODEL ANALYSES," was mistakenly left as an empty section due to a formatting oversight. We have removed it from the revised paper.

AuthorsOpenReview ↗

Response to Reviewer TUAv (Part II)

Reply to Q2: The availability of future covariates needs further clarification.

Thank you for your valuable comment! The datasets, baselines, and experimental settings in our paper follow DAG [1] and GCGNet [2] to ensure fair and consistent comparisons. In the original experiments, we use the ground-truth values of future exogenous variables to simulate their availability. We have added the relevant references in lines 358–360 of the revised paper‾\underline{\text{in lines 358–360 of the revised paper}} (highlighted in blue).

Considering that future exogenous variables may not be fully available in real-world scenarios, we further use TiRex-2 [3], a model supporting multivariate forecasting, to predict future exogenous variables using only historical data available before each prediction origin. These predictions cover the entire forecasting horizon and are used as future exogenous inputs to models. NMask achieves the best overall performance. The average results are presented below, with the complete results reported in Table 7 of the revised paper‾\underline{\text{in Table 7 of the revised paper}}:

DatasetsNMaskDAGGCGNetTimeXerTiDE
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
NP0.4220.3640.4290.3760.4300.3780.4370.3830.4620.412
PJM0.1230.2080.1260.2100.1320.2150.1420.2280.1570.252
BE0.4420.2780.4570.2950.4580.3050.4710.3020.5420.344
FR0.4300.2340.4340.2490.4380.2440.4500.2460.4950.287
DE0.5980.4780.6070.4880.6120.4980.6540.5060.6360.514
Energy0.1590.3090.1610.3090.1630.3090.1690.3240.1610.310
Sdwpfm10.7050.5660.7050.5960.7070.5820.7060.6050.7080.629
Sdwpfm20.7730.6090.7830.6510.8020.6300.7980.6480.8120.677
Sdwpfh10.7380.6140.7420.6380.7410.6230.7420.6370.8000.693
Sdwpfh20.8420.6640.8670.6910.8820.6910.8880.7160.9160.743
Colbun0.1140.1480.1160.1530.1180.1600.1290.2150.1840.236
Rapel0.2510.2850.2570.2860.2630.2850.3410.3590.3230.352

[1] DAG: A dual correlation network for time series forecasting with exogenous variables. ICML, 2026.

[2] GCGNet: Graph-consistent generative network for time series forecasting with exogenous variables. ICLR, 2026.

[3] TiRex-2: Generalizing TiRex to Multivariate Data and Streaming. NeurIPS 2026.

Dear Reviewer,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If it does, we kindly request that you reconsider the score. If you have any additional suggestions, we are more than willing to engage in further discussions and make necessary improvements to the paper. Thanks again for dedicating your time to enhancing our paper!

Kind Regards,

Authors

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

The paper proposes NMask for forecasting with known future exogenous variables. It introduces learnable tokens for future target patches, aligns these tokens with the corresponding exogenous patches, and uses target-conditioned attention to incorporate local lagged information and global exogenous summaries. A shared patch-wise prediction head reduces the number of decoder parameters. Experiments cover 12 datasets, two forecasting horizons, component ablations, and a setting without future covariates.

Critical strengths
  1. The architecture provides a sensible way to preserve the temporal correspondence between future covariates and forecast positions. The shared prediction head also gives this design a practical benefit: Figure 4 reports approximately 0.10M parameters and competitive runtime. The combination of accuracy and model size is more compelling than either result alone.

  2. The forecasting results are strong across the reported benchmark. NMask achieves the best horizon-averaged MSE and MAE on all 12 datasets in Table 1, although it does not win every individual horizon. The ablations also show that both exogenous attention and the prediction strategy matter, providing useful evidence beyond comparisons with external baselines.

Critical weaknesses
  1. The availability of future covariates needs a more precise justification. Appendix A.1 includes ERA5 meteorological variables, generation measurements, precipitation, and inflow, but does not clearly distinguish observations from forecasts available at each prediction origin. This is particularly consequential for the 360-step setting. Giving every method the same inputs makes the architectural comparison fair, but does not establish that the reported gains carry over to operational forecasting.

  2. The ablations do not cleanly isolate the claimed contribution of temporal alignment. Replacing Masked Alignment with MLP fusion changes the information-processing structure, while the alternative prediction heads also change parameter sharing and capacity. These comparisons support the overall design, but leave open how much of the gain comes specifically from alignment rather than the shared decoder or other architectural differences.

Questions for authors
  1. For each dataset, are future covariates realized observations, archived forecasts, or values generated by another model? For archived forecasts, were they available at the prediction origin for the entire evaluated horizon?
  2. Does the alignment advantage remain when the shared prediction head and model capacity are held fixed, and only aligned exogenous interaction is replaced by unrestricted cross-attention?
  3. On datasets using realized future measurements, how much of the advantage remains with forecast-quality covariates, including errors that increase with the forecast horizon?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Futp (Part I)

We would like to sincerely thank Reviewer Futp for providing a detailed review and insightful comments. We have revised our paper accordingly.

Reply to W1 & Q1: Availability of future exogenous variables.

Thank you for your valuable comment! The datasets, baselines, and experimental settings in our paper follow DAG [1] and GCGNet [2] to ensure fair and consistent comparisons. In the original experiments, we use the ground-truth values of future exogenous variables to simulate their availability. We have added the relevant references in lines 358–360 of the revised paper‾\underline{\text{in lines 358–360 of the revised paper}} (highlighted in blue).

Considering that future exogenous variables may not be fully available in real-world scenarios, we further use TiRex-2 [3], a model supporting multivariate forecasting, to predict future exogenous variables using only historical data available before each prediction origin. These predictions cover the entire forecasting horizon and are used as future exogenous inputs to models. NMask achieves the best overall performance. The average results are presented below, with the complete results reported in Table 7 of the revised paper‾\underline{\text{in Table 7 of the revised paper}}:

DatasetsNMaskDAGGCGNetTimeXerTiDE
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
NP0.4220.3640.4290.3760.4300.3780.4370.3830.4620.412
PJM0.1230.2080.1260.2100.1320.2150.1420.2280.1570.252
BE0.4420.2780.4570.2950.4580.3050.4710.3020.5420.344
FR0.4300.2340.4340.2490.4380.2440.4500.2460.4950.287
DE0.5980.4780.6070.4880.6120.4980.6540.5060.6360.514
Energy0.1590.3090.1610.3090.1630.3090.1690.3240.1610.310
Sdwpfm10.7050.5660.7050.5960.7070.5820.7060.6050.7080.629
Sdwpfm20.7730.6090.7830.6510.8020.6300.7980.6480.8120.677
Sdwpfh10.7380.6140.7420.6380.7410.6230.7420.6370.8000.693
Sdwpfh20.8420.6640.8670.6910.8820.6910.8880.7160.9160.743
Colbun0.1140.1480.1160.1530.1180.1600.1290.2150.1840.236
Rapel0.2510.2850.2570.2860.2630.2850.3410.3590.3230.352

[1] DAG: A dual correlation network for time series forecasting with exogenous variables. ICML, 2026.

[2] GCGNet: Graph-consistent generative network for time series forecasting with exogenous variables. ICLR, 2026.

[3] TiRex-2: Generalizing TiRex to Multivariate Data and Streaming. NeurIPS 2026.

AuthorsOpenReview ↗

Response to Reviewer Futp (Part II)

Reply to W2 & Q2: The existing ablation studies do not fully isolate the contribution of temporal alignment.

Thank you for your valuable suggestion! We agree that the existing ablation studies mainly demonstrate the effectiveness of the overall module design, but do not fully isolate the contribution of temporal alignment. To address this concern, we conduct an additional controlled ablation study. Specifically, we replace Masked Alignment with unrestricted Cross-Attention, which allows each endogenous token to attend to exogenous tokens at all time positions, while keeping the attention architecture, shared prediction head, and model parameter count unchanged. The ablation results demonstrate the importance of explicit temporal alignment, as NMask consistently outperforms the unrestricted Cross-Attention variant.

DatasetsNPPJMDEEnergyAverage
MetricsMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
(a) Unrestricted Cross-Attention0.4720.4310.1200.2160.4230.4170.1340.2880.2870.338
NMask (ours)0.4540.4190.1100.2090.4000.3950.1140.2680.2700.323

Reply to Q3: Can NMask maintain its performance advantage when future exogenous variables contain prediction errors?

Thank you for your valuable suggestion! We further use TiRex-2, a model supporting multivariate forecasting, to predict future exogenous variables using only historical data available before each prediction origin. The results have shown in W1.

AuthorsOpenReview ↗

Waiting for response

Dear Reviewer Futp,

Since the End of the Rebuttal comes soon, we would like to inquire if our response addresses your primary concerns. If it does, we kindly request that you reconsider the score. If you have any additional suggestions, we are more than willing to engage in further discussions and make necessary improvements to the paper. Thanks again for dedicating your time to enhancing our paper!

Kind Regards,

Authors

Dear Reviewer Futp,

Sorry to bother you again. We sincerely appreciate the effort you have put into reviewing our paper during this busy period, as well as the recognition of the strengths of our work. We fully respect your current evaluation of our work.

The rebuttal is ending soon. However, based on the scoring standards of past ICLRs, we find that our current score of 2.67 is at the borderline level. Therefore, we kindly ask if you could consider raising the score once again, so that we would have the opportunity to present our work at the conference. We would be deeply grateful. Thank you again for taking the time to review and comment on our paper.

Kind Regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.