- 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
-
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.
-
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
-
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.
-
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
-
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?
-
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
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 . NMask consistently achieves lower forecasting errors than TiDE across all datasets.
| Datasets | NMask | TiDE | ||
|---|---|---|---|---|
| Metrics | MSE | MAE | MSE | MAE |
| NP | 0.308 | 0.320 | 0.443 | 0.400 |
| PJM | 0.085 | 0.178 | 0.142 | 0.246 |
| BE | 0.410 | 0.278 | 0.498 | 0.325 |
| FR | 0.380 | 0.216 | 0.484 | 0.281 |
| DE | 0.325 | 0.354 | 0.499 | 0.447 |
| Energy | 0.086 | 0.228 | 0.153 | 0.302 |
| Sdwpfm1 | 0.375 | 0.399 | 0.483 | 0.507 |
| Sdwpfm2 | 0.424 | 0.432 | 0.486 | 0.516 |
| Sdwpfh1 | 0.365 | 0.411 | 0.453 | 0.508 |
| Sdwpfh2 | 0.413 | 0.456 | 0.599 | 0.583 |
| Colbun | 0.097 | 0.147 | 0.164 | 0.227 |
| Rapel | 0.223 | 0.260 | 0.320 | 0.351 |
(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.
| Datasets | NP | PJM | DE | Energy | Average | |||||
|---|---|---|---|---|---|---|---|---|---|---|
| Metrics | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE |
| (a) Remove temporal alignment constraint | 0.472 | 0.431 | 0.120 | 0.216 | 0.423 | 0.417 | 0.134 | 0.288 | 0.287 | 0.338 |
| NMask (ours) | 0.454 | 0.419 | 0.110 | 0.209 | 0.400 | 0.395 | 0.114 | 0.268 | 0.270 | 0.323 |
In addition, the prediction horizon used in Table 2 is 360. We have explicitly specified this in the table caption (highlighted in blue) .
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 (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 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: denotes the local window size, including the current position, while denotes the number of global summaries obtained through adaptive average pooling. We use and 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 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.
| Datasets | NP | PJM | DE | Energy | Average | |||||
|---|---|---|---|---|---|---|---|---|---|---|
| Metrics | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE | MSE | MAE |
| (a) Replace Masked Alignment | 0.465 | 0.428 | 0.120 | 0.215 | 0.419 | 0.408 | 0.122 | 0.277 | 0.282 | 0.332 |
| NMask (ours) | 0.454 | 0.419 | 0.110 | 0.209 | 0.400 | 0.395 | 0.114 | 0.268 | 0.270 | 0.323 |
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
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 (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:
[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 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.