acceptodds
Under review as a conference paper at ICLR 2027

ExoST: Multi-System Exogenous Variable Modeling for Spatio-Temporal Forecasting

Abstract

Spatio-temporal (ST) forecasting is crucial for modeling physical world systems. However, existing ST models primarily focus on target variables within a single system, often neglecting their interactions with external systems. Although early research attempts introduced exogenous variables, they merely treated the spatial dimension as an external auxiliary feature, essentially belonging to covariate modeling within a single system. This paper first conducts a systematic study on the modeling of multi-system exogenous variables in ST forecasting. We identify two core challenges in this field: the inconsistent effects of distinct system on the target variables and the imbalance effects between historical and future information. To tackle this, we propose ExoST, a simple yet effective exogenous variable modeling general framework highly compatible with existing ST backbones that follows a "select-then-balance" paradigm. ExoST can adaptively select informative signals from multi-system exogenous variables and perform context-aware balance of historical and future exogenous information. Extensive experiments on real-world datasets demonstrate that ExoST consistently enhances the predictive performance of various backbones while maintaining its simplicity, robustness, and efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Spatio-Temporal Forecasting, Exogenous Modeling

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_EpidOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the problem that existing spatio-temporal forecasting methods primarily focus on the internal dynamics of a single system while overlooking the influence of multi-system exogenous variables. The authors identify two main challenges in multi-system exogenous variable modeling: the inconsistent effects of different exogenous variables on the prediction target and the imbalance between historical and future exogenous information. To address these challenges, the paper proposes ExoST, a general spatio-temporal forecasting framework following the "Select-then-Balance" paradigm. It employs a gated expert selector to adaptively extract informative exogenous information and utilizes dual-branch spatio-temporal encoders and a context-aware balancer to dynamically fuse historical and future information. Experiments on four datasets from the air quality and traffic forecasting domains demonstrate the effectiveness and generality of the proposed method. Overall, this paper presents a relatively simple solution for multi-system exogenous variable modeling and validates the performance advantages of the proposed framework through relatively comprehensive experiments.

Critical strengths

S1: This paper addresses the long-overlooked problem of multi-system exogenous variable modeling in spatio-temporal forecasting, overcoming the limitations of traditional methods that primarily focus on the internal dynamics of a single system. Furthermore, the paper identifies two key challenges: the inconsistent effects of different exogenous variables and the imbalance between historical and future information, providing a research direction with practical value for spatio-temporal forecasting.

S2: The proposed framework demonstrates good generality and compatibility. ExoST can be integrated with different spatio-temporal forecasting models without modifying the internal architecture of existing backbones. The authors validate the framework on various backbones, including AGCRN, GWNet, STGCN, and DCRNN, demonstrating its potential as a general framework for exogenous variable modeling.

Critical weaknesses

W1: Although the paper uses four datasets, these datasets are primarily derived from air quality and traffic monitoring data collected in Madrid in 2019 and 2022, sharing similar data structures and types of exogenous variables. Therefore, the current experiments are insufficient to demonstrate whether the proposed method can generalize to other geographical regions, different types of dynamic systems, or forecasting tasks involving more complex exogenous relationships.

W2: According to the experimental setup, meteorological variables are primarily used as future covariates, while traffic variables are mainly treated as historical covariates. This setting couples the differences in exogenous system sources with the temporal differences between historical and future information. The paper lacks controlled experiments to disentangle these two factors, making it difficult to determine whether the performance improvements arise from multi-system interaction modeling or effective fusion of different types of temporal information.

W3: The paper claims that the Selector can adaptively identify important exogenous variables. However, its implementation mainly performs expert weighting in the fused latent feature space without explicitly modeling or providing the importance of individual exogenous variables. The existing visualizations primarily illustrate the distributions of expert gating weights, which are insufficient to demonstrate that the model can accurately distinguish informative variables from irrelevant ones. Therefore, sufficient experimental evidence is still lacking to establish a clear connection between the core design motivation and the actual mechanism of the proposed module.

Questions for authors

Q1: In general, endogenous and exogenous variables from different systems often have different sampling frequencies and collection times, meaning that their timestamps may not be strictly aligned. Furthermore, the timestamps of future exogenous variables may not exactly match those of the target prediction horizon. I have a question: How does ExoST handle such differences when endogenous and exogenous variables have misaligned timestamps or different sampling frequencies, or can these variables be used directly? For future exogenous variables, can they be considered valid prediction inputs as long as their timestamps fall within the target prediction horizon, or must they satisfy specific temporal alignment and availability constraints at the time of prediction?

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Dear Reviewer Epid,

We sincerely thank you for your time and effort in providing insightful feedback on our paper. We have carefully considered each of your comments to address your concerns (We wrote this rebuttal entirely by ourselves, word by word, without using any AI assistance. We hope that this can reciprocate your effort in taking the time to review it. Thank you!). Specifically:


Experiment-related

  • W1: Generalization beyond the Madrid benchmarks.

    • We fully understand this concern, but respectfully note it reflects a data constraint, not an evaluation choice. As stated in L286–288, standard ST benchmarks (METR-LA, PeMS-BAY) contain no real exogenous variables from heterogeneous systems (typically only time-of-day), so this setting cannot even be instantiated on them. To our knowledge AQI-19 is the only public benchmark meeting the requirement, which is itself the reason the community has neglected this problem.
    • Our evaluation nonetheless covers 4 tasks × 2 distinct target domains (air-quality NO2; traffic speed and flow intensity) × 2 years × 6 ST backbones × 3 horizons, and we curated and release three additional datasets (AQI-22, Speed-19, Intensity-22) precisely to lower this barrier for future work — we regard this as a contribution rather than a limitation.
    • To further address your concern, we additionally add three results: (i) NYC-Taxi with weather, holiday and POI features, ExoST improves every backbone, avg. ↓6.4% MAPE; (ii) two new exogenous scenarios (humidity / temperature), avg. ↓22.21% / ↓41.22% MAPE vs. w/o ExoST; (iii) industrial validation on multi-city order-demand data with 50+ exogenous variables across multiple domains (confidentiality prevents release, but it motivated this work based on industry collaboration).
  • W2: Is the benefit from multiple systems or from temporal fusion?

    • We agree with you that system identity and temporal availability are indeed two distinct dimensions; our framework is designed to handle their combination rather than treating them as identical.
    • However, given currently available data benchmarks, they typically encompass only three types of systems (each corresponding to different temporal availability); thus, Table 4 answers combinations across all groups, which in fact already isolates and ablates the influence of different systems.
    • We agree that more heterogeneous system data should be supplemented in the future; for now, however, this remains constrained by the characteristics of existing benchmarks as well as available spatio-temporal data mining and crawling.

Method-related

  • W3: Why can a latent expert module address variable relevance w/o outputting raw-variable importance scores?

    • The key distinction is predictive selection vs. feature attribution. Designed for the former, the Selector computes S(z)=∑kgk(z)(Wkz+bk)S(z) = \sum_k g_k(z) (W_k z + b_k) before spatio-temporal encoding, dynamically combining transformations rather than assigning scalar channel weights. This captures conditional effects and variable interactions to optimize the representation fed to the backbone, rather than ranking correlated sensors.
    • To verify that each module resolves the specific challenge it was designed for, rather than merely adding capacity, we added a targeted perturbation study on AQI-19:
      • Variable inconsistency (injecting irrelevant exogenous variables): the Selector recovers 11.53% MAPE.
      • Temporal imbalance (partially shuffling past/future inputs): the Balancer recovers 4.62% MAPE.
    • We also added a naive concatenation baseline requested elsewhere: AQI-19 MAE 10.50 / 10.98 / 11.36 (1/2/3-day) vs. ExoST 9.33 / 10.04 / 10.41, confirming the gain stems from the framework design rather than from merely having exogenous data available.
  • Q1: Sampling-frequency differences

    • Sharp observation, excellent question! In fact, this mainly falls into two challenges from the training stage and inference stage.
    • For the training stage, honestly, we followed the default benchmark settings and unified different sampling frequencies via standard up/downsampling in preprocessing. While heterogeneous sampling is inevitable in practice, handling it natively through dedicated inductive biases remains an exciting direction for future work.
    • For the testing stage, due to limited space, we kindly ask you to refer to our responses to reviewers F2Mx and jEsu regarding W1 & Q1. In summary, our method is quite robust to such heterogeneous sampling inputs!

Thank you for your valuable feedback, which has helped us strengthen our manuscript! We hope the above responses address your concerns, reinforce your appreciation and confidence in our work, and we look forward to further discussion.

Best regards,

Authors

The authors' responses have addressed most of my concerns satisfactorily. I have also reviewed the responses to the other reviewers and found that the authors have made substantial efforts to improve and clarify the manuscript. Therefore, I am satisfied with the revisions and will maintain my positive rating.

Thank you for your attention; we greatly appreciate it.

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

This paper investigates how to effectively leverage covariates to improve spatiotemporal forecasting in multi-system environments. To address this challenge, the authors propose ExoST, a framework that employs gated projection to select informative features along the spatial dimension and forecasting branches to integrate historical and future information. Experiments on four spatiotemporal datasets demonstrate the effectiveness of the proposed framework, while ablation studies further validate the contributions of its key components.

Critical strengths
  1. The proposed ExoST framework can insert multiple spatiotemporal backbones.
  2. Achieves state-of-the-art results on real-world datasets; designed experiments demonstrate robustness.
Critical weaknesses

Table 1 claims ExoST to be efficient, but in practice, there is no significant lead in parameter size and training time, and inference latency lacks supporting evidence.

Questions for authors
  1. The text mentions that efficiency is superior to other models, but lacks validation of prediction efficiency.

  2. The headers of Tables 4, 11, and 12 are somewhat inconsistent, and there are minor contradictions between the captions and the text.

  3. Why choose a prediction duration of 1/2/3 days? Is there a need for shorter or longer prediction periods?

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Dear Reviewer BnhK,

We sincerely thank you for your time and effort in providing feedback on our paper. Although we feel there may be a few misunderstandings, we deeply appreciate that your review appears to be genuinely human-written without AI assistance, which feels especially precious in this era. We have carefully considered each of your comments to address your concerns (We wrote this rebuttal entirely by ourselves, word by word, without using any AI assistance. We hope that this can reciprocate your effort in taking the time to review it. Thank you!). Specifically:


Experiment-related

  • W1 & Q1: Inference latency and computational overhead
    • First, we believe there are some misunderstandings here. You may have overlooked the experimental results in Section 5.6. Figure 12 shows that our method achieves a leading efficiency advantage among the methods that fall within the feasible performance range.

    • Second, regarding Table 1, it serves as an overarching comparison across different categories of methods. Specifically, statistical and TS-based methods require invoking model predictions independently for each node, which naturally makes them less efficient than ST-based methods. Meanwhile, considering that specialized neural networks in current ST-based models typically maintain around 10M parameters, LLM-based methods, which introduce language models with 10B–100B+ parameters, are naturally inefficient by comparison. Serving as a plug-and-play module, our ExoST introduces only two lightweight module, allowing it to effectively leverage existing ST-based models and enjoy comparable parallel efficiency.

    • Third, we agree that the introduced inference cost warrant further evaluation. We measured the computational costs on the AQI-19 dataset (single A6000 GPU, averaged across the entire test set over 5 runs), with the results summarized in the table below. The actual runtime latency increase remains limited—a worthwhile trade-off given the significant performance gains (53.20%–57.40%). | Backbone | Inference Time (s) w/o →\rightarrow w/ ExoST| MAPE (%) w/o →\rightarrow w/ ExoST| | --- | --- | --- | | AGCRN | 1.9 →\rightarrow 3.4 | 60.71 →\rightarrow 38.57 | | DCRNN | 6.1 →\rightarrow 9.7 | 68.45 →\rightarrow 44.68 |

    • Furthermore, we provide an additional comparison of inference performance against baselines. It is worth noting that the MAPE of most competing methods exceeds 40%, placing them in a practically unusable regime. In contrast, our method maintains competitive inference latency while achieving viable, strong predictive performance. Also, we would like to point out that inference latency is currently not a critical bottleneck in ST forecasting: because the overall parameter counts of neural networks for this task are relatively compact, total inference times typically range from 10s to 100s, which easily satisfies the requirements of minute-level and hour-level forecasting systems. | Model | Type | Inference Time (s) | MAE | RMSE | MAPE (%) | | --- | --- | --- | --- | --- | --- | | TiDE | TS-based | 105.6 | 14.73 | 20.73 | 93.01 | | TimeXer | TS-based | 50.4 | 13.11 | 20.20 | 64.48 | | ChronosX | FM-based | 86.4 | 12.46 | 18.37 | 55.86 | | MAGCRN | ST-based | 4.2 | 9.82 | 15.18 | 38.66 | | ExoST | ST-based | 3.4 | 9.33 | 14.24 | 38.57 |


Presentation-related

  • Q2: Inconsistent headers and captions in Tables 4, 11, and 12

    • Thank you for your careful inspection. We confirm that there is a typo here: the captions for Tables 4, 11, and 12 are accurate, whereas the letter "U" inside Table 12 should actually be "D".
  • Q3: Why evaluate one, two, and three days?

    • In short, we follow the default settings of existing benchmarks, which is well-justified. Specifically, existing datasets collect one observation per hour; therefore, a 24-hour target window covers a complete diurnal cycle rather than selecting a short interval that might miss traffic peaks or changing meteorological conditions. Shorter prediction horizons are suited for immediate response, whereas longer horizons support advance planning.
    • Moreover, ExoST is not constrained by an equal-length assumption. When the forecast horizon exceeds the lookback window, we demonstrate (see Table 2) that rolling forecasting can be achieved without additional training, exhibiting strong robustness. For longer prediction horizons, we can naturally achieve long-term forecasting via similar rolling predictions; meanwhile, short-term predictions are inherently encompassed within long-term predictions (i.e., when predicting 24 future steps, you naturally predict the 3, 6, and 12-step horizons as well).

Thank you again for your constructive suggestions, which have helped us substantially sharpen the positioning of this work. If possible, we kindly request that you reconsider the score, and we would be very happy to discuss any remaining questions.

Best regards,

Authors

Thank you for your detailed response and additional experiments. While some concerns have been addressed, the new results show that ExoST introduces additional inference overhead, and the efficiency evaluation remains limited to a single dataset. Therefore, I believe the evidence is still insufficient to support the claimed efficiency advantages or change my overall assessment. I will maintain my original score.

AuthorsOpenReview ↗

Some heartfelt words

We appreciate you responding as a human, but we also deeply regret that reviewers who do not use AI are so harsh and terrible. Ask yourself, have you truly shown fair respect and conducted a fair review? Or have you reviewed with malicious bias? What a terrible time this is for human research. Thank you for your feedback.

Here are some honest thoughts from me that I really want to share with you:

From 2020 to now, I have provided over 400 review services. Following past courtesy and restraint, I should respect the efforts of all reviewers and demonstrate good professionalism. However, by the end of 2025, seeing that the entire research community has completely failed and collapsed, I feel so deeply distressed. Since mid-2026, I have decided to stop participating in these reviews. I don't know if you are a new researcher or a new reviewer forced to take on review tasks, but I hope you can maintain a respectful and kind heart in the future. Thank you!

Although we have emphasized that inference latency is not currently a critical bottleneck in ST forecasting, since the overall parameter size of neural networks for this task is relatively small, and the total inference time typically ranges from 10 to 100 seconds, which easily meets the requirements of minute-level and hour-level prediction systems. We have, in order to be responsible and respectful of your review time, supplemented the comparison of inference times on the remaining datasets, and the results show that our method is competitive in both performance and efficiency.


SPEED-19

ModelTypeInference Time (s)MAERMSEMRE (%)
TiDETS-based108.26.7044.4717.44
TimeXerTS-based48.15.0738.8413.20
ChronosXFM-based89.539.20139.33102.00
MAGCRNST-based4.53.8731.8310.08
ExoSTST-based3.63.7931.739.88

AQI-22

ModelTypeInference Time (s)MAERMSEMRE (%)
TiDETS-based102.910.5114.6756.27
TimeXerTS-based53.69.6514.9851.64
ChronosXFM-based84.18.8713.2347.49
MAGCRNST-based4.07.4111.4339.66
ExoSTST-based3.27.3311.1439.23

INTENSITY-22

ModelTypeInference Time (s)MAERMSEMRE (%)
TiDETS-based109.4150.18217.0554.43
TimeXerTS-based52.8128.05191.7246.41
ChronosXFM-based88.7142.06216.7551.35
MAGCRNST-based4.437.4768.5213.58
ExoSTST-based3.731.9156.4911.57

No matter what your thoughts may be going forward, we respect your comments, and we leave this fairness assessment to the AC and the community. Thank you once again for your time in reviewing. Be kind!

Best,

Authors

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

The paper proposes ExoST, a framework for incorporating multi-system exogenous variables into spatio-temporal forecasting models. It follows a “select-then-balance” design: conditional embeddings and gated experts process historical and future exogenous inputs, two independently parameterized ST encoders model the resulting streams, and a context-aware balancer combines their outputs. Experiments cover four air-quality and traffic forecasting tasks using data from 2019 and 2022. The paper evaluates several ST backbones, reports results over five runs, and includes ablations, exogenous-input corruption experiments, and efficiency comparisons.

Critical strengths
  1. The framework addresses a practical limitation of forecasting models that use only the target system’s history. Its modular design makes it possible to incorporate exogenous information without redesigning the underlying ST encoder. Testing the framework with six different backbones is useful evidence of its applicability beyond a single architecture.

  2. The experiments consider both forecasting accuracy and imperfect exogenous inputs. Reporting five-run statistics, multiple error metrics, and comparisons across forecasting durations makes the evaluation more informative. The air-quality results are particularly encouraging: on AQI-19, ExoST reduces one-day MAE from 9.82 for MAGCRN to 9.33.

Critical weaknesses
  1. The availability of future exogenous information is not sufficiently established. The motivation and method describe future meteorological inputs as forecasts from upstream systems, but Appendix B.1 only states that meteorological variables serve as future covariates. It does not identify the forecasting source, issue times, lead times, or whether these inputs are actually recorded future observations. This distinction is important: using realized weather would evaluate an oracle-covariate setting rather than the operational scenario described in the paper. Table 4 shows that future covariates account for a large part of the improvement, making their provenance central to the conclusions. The two- and three-day rolling experiments also need to specify which observations and exogenous forecasts become available at each rollout step.

  2. The ablations do not cleanly isolate the proposed mechanisms. ExoST introduces two independently parameterized encoders, gated expert projections, and a residual fusion path. In Appendix C.3.2, the shared-parameter variant also replaces the proposed balancer with simple averaging, while the fixed-weight variant omits the residual term retained by ExoST. These comparisons therefore change more than the property they are intended to test. The reported advantages cannot be attributed solely to independent encoders or context-dependent weighting. Moreover, the final fusion is equivalent to �0��0�, so neither branch can be directly suppressed by the balancer. This deserves discussion given the motivation of handling unreliable exogenous information. The universality claims also need qualification: Table 6 includes cases where adding ExoST worsens individual metrics.

Questions for authors
  1. Are the future meteorological inputs archived forecasts available at each prediction origin, or realized observations from future timestamps? Please specify their source and temporal availability. If realized observations are used, how does performance change with realistic forecast inputs?

  2. How are the two- and three-day rolling forecasts constructed? Are predicted target values fed back into the model, and are new target observations or updated exogenous forecasts introduced during the rollout?

  3. Can the authors compare shared and independent encoders while keeping the selector, balancer, and residual connection unchanged? Similarly, please compare fixed and dynamic fusion weights under the same residual formulation and with comparable model capacity.

  4. How should the residual fusion be interpreted when one branch is unreliable? Appendix A.2 claims that the formulation can approximate arbitrary fusion functions, but its bounded positive coefficients appear to restrict this claim for fixed branch outputs. Please clarify the scope of the analysis.

  5. In the corruption experiments, are missingness and noise introduced during training, testing, or both? Including MAGCRN or a simple concatenation baseline under the same corruption protocol would help establish whether robustness comes specifically from ExoST.

3 comments
AuthorsOpenReview ↗

Rebuttal by Authors

Dear Reviewer jEsu,

We sincerely thank you for your constructive feedback and the effort invested in reviewing our work. We have carefully addressed each of your concerns below: (We wrote this rebuttal entirely by ourselves, word by word, without using any AI assistance. We hope that this can reciprocate your effort in taking the time to review it. Thank you!). Specifically:


Problem-related

  • W1 & Q1 & Q2: Availability of future exogenous inputs
    • Excellent observation! You have raised a question similar to that of Reviewer F2Mx. Specifically in deployment, we distinguish among three scenarios: (a) deterministic exogenous (calendars, holidays): always available; (b) predictable exogenous (weather): obtained from upstream NWP/forecasting systems, which is standard practice and the protocol used by prior work (DAG, TimeXer, etc.); (c) unavailable exogenous values: zero or persistence imputation.
    • For the first two, we follow the default benchmark and baseline settings. For the third, although no benchmark previously existed, our RQ3 study (Section 5.4, Tables 9–10) is precisely a simulation experiment designed for this. ExoST degrades gracefully rather than failing.
    • Two observations can be drawn: even with only 20% exogenous inputs, ExoST still outperforms all baselines in Table 1 with full access to exogenous variables, and still outperforms the backbone without ExoST by 19.4% in MAE. Under Speed-19, this effect is almost negligible (3.81 MAE vs. 3.79 under 80% masking).
    • Furthermore, this work actually originated from an industrial scenario. We have collaborated with industry partners to validate the effectiveness of ExoST in order volume scenarios across multiple cities (involving over 50 cross-domain exogenous variables), where future covariates were derived from different non-oracular expert forecasting system. Experiments demonstrated exceptional performance improvements (average MAPE improvement of 22.21% / 41.22%). Although the original business data cannot be disclosed due to confidentiality agreements, this fully demonstrates the generalization capability of our method.
AQI-19 (1-day)MAERMSEMAPE (%)
ExoST, unmasked9.3314.2438.57
ExoST, 80% zero-masking10.9917.1047.36
ExoST, 80% random-masking11.1317.6743.96
Backbone w/o ExoST (full data)13.6320.4360.71
Best baseline w/ full exogenous (NBEATSx)11.6318.4456.54

Method-related

  • W2 & Q3 & Q4: Do the ablations work? What about the fusion mechanism?
    • We respectfully point out that what you mentioned actually refers to the strategy study in lines 459–466. In fact, we also conducted a related ablation study in lines 424–431, and the results demonstrate the effectiveness of our design.
    • We understand your concern regarding the parameter and performance efficiency of ExoST. First, we must point out that parameters for different fusion methods are inherently method-specific, making it difficult to guarantee an exact, apples-to-apples parameter comparison across every baseline.
    • To further alleviate your concerns, we added a comprehensive comparison across six backbones, including parameter counts, FLOPs, and training time. On average, ExoST introduces only +8.05MB parameters, -31.92% FLOPs, and +14.53 min. training time, while ↓\downarrow MAPE by 30%.
    • Furthermore, we respectfully point out that our final fusion mechanism is not equivalent to the formulation you described; we have already provided a detailed formalization and theoretical analysis in Appx. A.2.

Experiment-related

  • Q5: Corruption protocol and the evidence for robustness.
    • Thank you for your comment. We inject noise strictly during the inference stage to simulate sensor signal anomalies and corruption encountered during real-world system deployment.
    • As per your request, we have added a comparison baseline that directly concatenates exogenous variables (Concat) on the AQI-19 dataset. As shown in the below, ExoST outperforms the simple Concatstrategy by 5.20% in MAE, demonstrating that ExoST's performance gains stem from the superiority of its dedicated "Selection and Balancing" framework design.
ConcatExoSTConcatExoSTConcatExoST
1-day2-day3-day
MAE10.50±0.169.33±0.1310.98±0.5310.04±0.1911.36±0.7210.41±0.12
RMSE16.21±0.2714.24±0.1816.78±0.6515.12±0.2617.13±0.8915.51±0.16
MAPE(%)41.96%±0.0138.57±0.3244.48%±0.0242.62±0.5247.10%±0.0344.65±0.08
MRE(%)40.61%±0.0136.09±0.4942.51%±0.0238.84±0.7544.02%±0.0340.34±0.47

Thank you again for your constructive suggestions, which have helped us substantially sharpen the positioning of this work. If possible, we kindly request that you reconsider the score, and we would be very happy to discuss any remaining questions.

Best regards,

Authors

Thank you for clarifying that noise is added only at inference and for including the Concat baseline. Two concerns remain. It is still unclear whether the benchmark experiments use weather forecasts available at prediction time or realized observations, and what information is available during the multi-day rollouts. The new comparisons also leave open how much of the gain comes from dynamic weighting versus other architectural choices. Given these remaining concerns, I am keeping my original rating.

AuthorsOpenReview ↗

Direct Response to Reviewer jEsu's Concerns

Dear Reviewer jEsu,

Thank you for taking the time and effort to follow up. However, we hope you can avoid using AI to skim through our responses and commenting based on preconceived notions. Instead, please respect our time and replies! (If we have misunderstood you, we are willing to apologize, but currently the community is flooded with such a frustrating culture! We have to say this!) Although we already addressed your concerns in our previous response, we provide a direct reply here once again.


Baseline covariates and information available during multi-day prediction

  • We use the existing standard data benchmark AQI-19. We reiterate that for future meteorological inputs, this benchmark uses real recorded values from the Madrid benchmark, obtained from upstream NWP/forecasting systems. This is standard practice and the protocol used in prior works (DAG, TimeXer, etc.). Therefore, these are not archived forecasts issued at each prediction start time. They are extracted from meteorological columns and aligned with the target window.
  • For the multi-day rolling forecasting process, we assume you have a basic understanding of machine learning and autoregressive forecasting models. Therefore, we actually follow this approach: we input only the historical 24 steps to predict the next 24 steps (day 1); then use the predicted future 24 steps as new input to predict steps 24–48 (day 2), and so on. Throughout this process, the trained checkpoint remains unchanged. The predicted target values are not used for training, but only as input for the next step. Similarly, all exogenous variables are available and extracted from corresponding recorded values; there is no data or future information leakage involved.

Ablation study on dynamic weighting

  • We are quite puzzled about what exactly you are questioning, or whether you have understood our method. As described in the paper, our method is a general plug-and-play framework, not a specific model structure design. In Section 5.3 and Figure 6, we demonstrate that our method achieves further performance improvements on six existing model architectures * four datasets.
  • Regarding dynamic weighting, we have already emphasized in the ablation study of RQ4 and Figure 7 the performance difference when the selection module is completely removed. In addition, in the strategy study of RQ4 and Figure 11, we present comparisons of various dynamic fusion strategies we could think of, and as per your request, we have added the performance comparison of a direct concatenation scheme. The experimental results clearly demonstrate the performance superiority of our method design. What do you mean by "other architectural choices"? If you have any, please specify. We believe this might be a misunderstanding. We hope to clarify this issue, that is, we are not a model-specific design, but a general framework solution.

We hope these points address your concerns, and we thank you again for your time and comments. This is a terrible time for the human research community (maybe my last participation); thank you for the little attention you have paid, and if you are willing to reconsider, thank you!

Best,

Authors

Reviewer_F2MxOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ExoST, a plug-in framework for spatio-temporal forecasting that incorporates heterogeneous exogenous information through a “select-then-balance” strategy. It employs a gated expert-based Selector to identify relevant external signals and a context-aware Balancer to adaptively integrate historical and future exogenous information without modifying the underlying forecasting backbone. Experiments on four benchmarks with six backbones demonstrate consistent improvements, supported by ablation, robustness, and efficiency analyses.

Critical strengths

Strength 1: Important and practical problem. Air quality and traffic forecasts often depend on factors outside the target system. The paper treats these factors as a central modeling problem rather than an afterthought. That focus is useful and has clear practical value.

Strength 2: Modular design and useful empirical coverage. ExoST works with several existing backbones, and its evaluation includes ablations and input corruption tests. In Table 4, combining past, future, and calendar features gives an MAE of 9.33 on the one-day AQI-19 task, compared with 9.97 for future features alone. Together, these results suggest that different input types can complement one another. Its plug-in design should also make the approach easier to reuse.

Critical weaknesses

Weakness 1: Unclear availability of future exogenous inputs. A key uncertainty is where the future exogenous signals come from. Are they recorded future values or forecasts available when the prediction is made? Actual future observations would usually be unavailable when making a real forecast. Using them could make the reported errors too optimistic. Table 4 shows a strong benefit from future features, which makes this point especially important. Appendix D.2 notes the availability issue but does not settle the experimental protocol. Please clarify when each input becomes available. If ground-truth future values are used, the results should be described as an oracle setting. A test with realistic forecasts would help.

Weakness 2: Limited isolation of the new technical contribution. Selector resembles existing mixture-of-experts methods, while Balancer draws on familiar gated-fusion designs. Much of the novelty may therefore lie in how these ideas are combined for multi-system forecasting. That can still be valuable, but the paper should show what the proposed modules add beyond extra capacity. Comparisons with similarly sized MoE, attention, and gated-fusion baselines would be more convincing.

Questions for authors

Q1 : Are the future exogenous inputs actual future observations or forecasts available at the prediction time? If actual observations are used, could you report results with inputs that would be available in practice?

Q2: Appendix C.3.2 already compares ExoST with several fusion methods. Could the authors clarify whether these baselines use comparable parameter counts and training budgets? For the Selector, how does the proposed design compare with a standard MoE module under the same settings?

Q3 : Table 4 already studies the contributions of past, future, and calendar inputs. Since these input types are closely linked to different external systems, how can we separate the effect of each system from the effect of temporal availability? Could the authors clarify this distinction or provide a more targeted ablation?

Q4 : Table 3 evaluates robustness using random masking and Gaussian noise. However, real forecast errors may have temporal correlations or systematic bias. How does ExoST perform under these more structured errors? In particular, does it still outperform the baselines when future inputs become less reliable?

1 comment
AuthorsOpenReview ↗

Rebuttal by Authors

Dear Reviewer F2Mx,

We sincerely thank you for your time and effort in providing insightful feedback on our paper. We have carefully considered each of your comments to address your concerns (We wrote this rebuttal entirely by ourselves, word by word, without using any AI assistance. We hope that this can reciprocate your effort in taking the time to review it. Thank you!). Specifically:


Problem-related

  • W1 & Q1: Availability of future exogenous inputs
    • Excellent observation! Specifically in deployment, we distinguish among three scenarios: (a) deterministic exogenous (calendars, holidays): always available; (b) predictable exogenous (weather): obtained from upstream NWP/forecasting systems, which is standard practice and the protocol used by prior work (DAG, TimeXer, etc.); (c) unavailable exogenous values: zero or persistence imputation.
    • For the first two, we follow the default benchmark and baseline settings. For the third, although no benchmark previously existed, our RQ3 study (Section 5.4, Tables 9–10) is precisely a simulation experiment designed for this. ExoST degrades gracefully rather than failing.
    • Two observations can be drawn: even with only 20% exogenous inputs, ExoST still outperforms all baselines in Table 1 with full access to exogenous variables, and still outperforms the backbone without ExoST by 19.4% in MAE. Under Speed-19, this effect is almost negligible (3.81 MAE vs. 3.79 under 80% masking).
AQI-19 (1-day)MAERMSEMAPE (%)
ExoST, unmasked9.3314.2438.57
ExoST, 80% zero-masking10.9917.1047.36
ExoST, 80% random-masking11.1317.6743.96
Backbone w/o ExoST (full data)13.6320.4360.71
Best baseline w/ full exogenous (NBEATSx)11.6318.4456.54

Method-related

  • W2 & Q2: Contribution beyond MoE, gated fusion, and additional capacity
    • We agree with you; in fact, we do not claim to propose any fancy or novel new modules and architectures. Recent research in spatio-temporal domains often introduces seemingly novel yet overly complex architectures, but yields limited gains on saturated benchmarks.
    • As stated in our contributions, our work is the first to systematically study multi-system exogenous variable modeling in ST forecasting, identify key challenges, and propose a simple yet scalable framework achieving significant improvements (an average of 22.24% ↓\downarrow MAPE across dataset/backbone combinations). We believe this contribution is timely and practical rather than superficially pursuit-driven.
    • We also understand your concerns regarding parameter and performance efficiency; however, we must point out that parameters for different fusion methods are method-specific, making it difficult to guarantee an exact, apples-to-apples parameter comparison across every method.
    • To further alleviate your concerns, we added a comprehensive comparison across six backbones, including parameter counts, FLOPs, and training time. On average, ExoST introduces only +8.05MB parameters, -31.92% FLOPs, and +14.53 min. training time, while ↓\downarrow MAPE by 30%.

Experiment-related

  • Q3: System identity versus temporal availability

    • We agree with you that system identity and temporal availability are indeed two distinct dimensions; our framework is designed to handle their combination rather than treating them as identical.
    • However, given currently available data benchmarks, they typically encompass only three types of systems (each corresponding to different temporal availability); thus, Table 4 answers combinations across all groups, which in fact already isolates and ablates the influence of different systems.
    • We agree that more heterogeneous system data should be supplemented in the future; for now, however, this remains constrained by the characteristics of existing benchmarks as well as available spatio-temporal data mining and crawling.
  • Q4: Structured errors and decreasing future-input reliability

    • Thank you for your comments. To further alleviate your concerns, we additionally conducted a targeted perturbation experiment to verify whether the model genuinely resolves variable inconsistency (injecting irrelevant exogenous variables) and temporal imbalance (partially shuffling and exchanging past/future exogenous inputs):
      • Variable inconsistency: ExoST achieves 11.53% ↑\uparrow w/ Selector and 7.57% ↑\uparrow w/ Balancer.
      • Temporal imbalance: ExoST achieves 3.91% ↑\uparrow w/ Selector and 4.62% ↑\uparrow w/ Balancer.

Thank you for your valuable feedback, which has helped us strengthen our manuscript! We hope the above responses address your concerns, reinforce your appreciation and confidence in our work, and we look forward to further discussion.

Best regards,

Authors

Comments on the paper

1 comment
Area_Chair_Vj8TOpenReview ↗

Discussion Reminder

Dear Reviewers,

Thank you for your efforts so far. This is a friendly reminder to review the authors' rebuttal and respond to their comments as soon as possible, before the discussion deadline (November 18 AoE). Please also indicate whether your concerns have been addressed. Thank you for your time and cooperation!

Best regards,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.