acceptodds
Under review as a conference paper at ICLR 2027

Generalizable Spatio-Temporal Forecasting Beyond the Closed-World Setting

Abstract

Spatio-temporal forecasting is fundamental to many real-world systems, yet deployment environments are inherently varying: temporal distributions may evolve, spatial domains may change, and observation availability may shift from training conditions. Existing models are often tightly coupled to their training environments through node-specific representations, predefined spatial support sets, and reliance on consistently available observations, limiting generalization to changing environments. To this end, we propose VaryST, a compact spatio-temporal forecasting model trained from scratch that achieves strong generalization by dynamically constructing predictive context from available long-term observations. To accommodate temporal changes, VaryST derives node representations from long-term histories, capturing current behavior in the context of longer-term patterns. To adapt to spatial changes, it dynamically selects functionally relevant support nodes beyond predefined neighborhoods and integrates their information with recent and long-term target states and global context. To handle varying observation availability, VaryST selectively masks historical patches on which the model relies most during training, encouraging it to exploit alternative temporal and spatial dependencies for accurate forecasting when key observations become unavailable. Extensive experiments on six real-world datasets against seventeen baselines show that VaryST achieves strong in-domain forecasting accuracy, substantially improves zero- and few-shot generalization under spatial and temporal changes, and remains robust to varying observation availability.

Primary area:
learning on time series and dynamical systems
Keywords:
Spatiotemporal Forecasting, Spatiotemporal graph data, Long-term, Traffic forecasting

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_Juf3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces VaryST, a compact spatio-temporal forecasting framework designed to generalize beyond the traditional closed-world assumption. Instead of relying on static node embeddings and predefined adjacency graphs, VaryST constructs predictive contexts dynamically from available long-term observations. It extracts dynamic node representations from historical patches and retrieves functional support nodes dynamically via RBF similarity. The framework then employs Target-Centric Hierarchical Contextual Decoding, which integrates global context, recent/long-term target states, and retrieved support tokens using hop-aware relative structural encoding. Additionally, the authors propose Importance-Guided Observation Perturbation (IGOP), an attention-guided masking strategy during training to improve robustness against missing data. Extensive experiments across six real-world datasets and 17 baselines demonstrate VaryST's strong performance in in-domain, cross-domain, and missing-observation scenarios.

Critical strengths
  1. Tackling the open-world spatio-temporal forecasting problem, simultaneously addressing temporal shifts, spatial domain changes, and varying observation availability, is highly relevant and practically significant for real-world deployments.

  2. Replacing static node embeddings and fixed graph adjacency matrices with long-term observation-derived representations and dynamic functional support retrieval is a very clever and effective design for spatial and temporal generalization.

  3. The Importance-Guided Observation Perturbation (IGOP) is an intuitive and novel approach. By dynamically masking patches that the model relies on most, it forces the network to learn robust alternative evidence paths rather than overfitting to dominant signals.

  4. The empirical evaluation is highly thorough, covering 6 datasets and comparing against 17 baselines across diverse settings. The zero-shot temporal shift evaluation (PEMSD3-Stream) and cross-network spatial transfer tasks provide highly convincing evidence of the model's capabilities.

Critical weaknesses
  1. For missing data, the model explicitly applies linear interpolation before computing representations and temporal differences. This could introduce severe artifacts or noise if the missing gaps are large, contiguous blocks, potentially degrading the derived dynamic representations.

  2. The model uses shortest-path hop distances for hop-aware relative structural encoding. However, the paper evaluates on PM2.5, which is explicitly described as having "no explicit graph". It is unclear how this structural encoding is handled when no topological graph is available, or when migrating to a graph with radically different topological characteristics.

  3. While compared against 17 baselines, the paper omits recent strong generative ST models (e.g., diffusion models or advanced state-space models) which are naturally suited for handling missing data imputation and distribution shifts.

  4. The IGOP module relies on selecting KmK_m patches to mask. The paper lacks a deep sensitivity analysis on KmK_m. Masking too much might collapse training, while masking too little would fail to improve robustness.

Questions for authors
  1. Please explain explicitly how the shortest-path hop distance and Laplacian positional embeddings are handled for datasets without explicit graph topologies (such as PM2.5). Does the model rely on a fully connected assumption, distance heuristics, or are these features simply disabled?

  2. Please provide an ablation or discussion on the choice of KmK_m (the number of masked patches in IGOP). How does varying KmK_m impact the standard forecasting performance versus the missing-data forecasting performance?

  3. Given that VaryST employs linear interpolation on masked sequences before computing differences (X = Interp(X, M)), how does this perform under extensive contiguous block missing (CM) scenarios?

1 comment

Thank you for recognizing the practical significance, methodological design, and thorough evaluation, and for your questions about experimental details, which helped us refine our experiments.

W1 / Q3. Contiguous missingness and interpolation.

Thank you for raising this concern. Table 10 in the appendix reports the complete CM results across missing rates of 10%–70%. CM is indeed more challenging than RM: at 70% missingness, VaryST’s MAE increases from 17.70 to 22.41 on PEMS04 and from 13.30 to 17.16 on PEMS08. Nevertheless, VaryST still achieves the lowest MAE, RMSE, and MAPE across all tested CM rates on both datasets. These results show that, although contiguous gaps degrade prediction accuracy, VaryST maintains its comparative advantage with the current interpolation-based preprocessing.

W2 / Q1. Structural encoding on PM2.5.

Thank you for pointing out this missing clarification. Our work primarily focuses on spatio-temporal graph forecasting, where a graph structure is available. To evaluate generalization beyond traffic, we also include the PM2.5 dataset, which provides no explicit graph. We therefore construct a 5-nearest-neighbor (5-NN) graph using Pearson correlations between station observations and compute the Laplacian positional embeddings and shortest-path hop distances on this graph. These preprocessing details have been added to Appendix B.1 (Dataset Statistics).

For cross-graph transfer, since VaryST operates inductively, we only need to recompute the Laplacian eigenvectors and shortest-path hop-distance matrix on the target graph while keeping all model parameters unchanged.

W3. Generative and state-space baselines.

Following your suggestion, we have included KITE (ICML 2026), a flow-matching-based generative forecaster, and DST-Mamba (AAAI 2025), a state-space forecaster as baselines. Compared to them, VaryST still achieves the lowest MAE, RMSE, and MAPE on all four PEMS datasets, as summarized below. The full results are reported in Table 2 of the revised manuscript (format: MAE / RMSE / MAPE).

DatasetUniSTKITEDST-MambaVaryST
PEMS0315.71 / 26.89 / 15.92%16.06 / 25.35 / 17.05%15.82 / 25.63 / 15.79%13.95 / 22.73 / 13.59%
PEMS0419.07 / 31.64 / 12.28%19.82 / 32.02 / 13.09%18.70 / 30.55 / 13.00%17.56 / 29.35 / 11.81%
PEMS0721.08 / 34.61 / 9.44%20.71 / 35.48 / 8.43%19.07 / 32.08 / 8.07%18.21 / 31.05 / 7.64%
PEMS0815.52 / 25.43 / 11.61%14.80 / 24.87 / 9.85%13.83 / 22.97 / 9.20%12.89 / 21.99 / 8.50%

W4 / Q2. Sensitivity to KmK_m

Thank you for raising this question. We evaluate Km∈{1,2,3,4,6,8,16}K_m \in \{1, 2, 3, 4, 6, 8, 16\} on PEMS08, trained at RM 70% and tested without retraining. As shown below, performance under the same missing rate (RM 70%) is insensitive to (K_m), with MAE ranging from 13.24 to 13.32. Under a different missing rate (RM 10%), increasing (K_m) from 1 to 6 reduces MAE from 16.02 to 14.61, while (K_m) values from 3 to 16 yield relatively stable performance. These results suggest that a sufficiently large (K_m) helps IGOP reduce over-reliance on dominant historical patches and improves generalization across missing rates. We have added these results to Table 16 in Appendix C.5 of the revised manuscript.

Test condition (PEMS08)Km=1K_m=1Km=2K_m=2Km=3K_m=3Km=4K_m=4Km=6K_m=6Km=8K_m=8Km=16K_m=16
RM 70% (same rate)13.2813.3013.2913.2613.2413.3213.27
RM 10% (cross rate)16.0215.2314.9314.7114.6114.7214.78
Reviewer_5usEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses spatio-temporal forecasting under temporal and spatial distribution shifts and missing observations. The authors propose a dynamic retrieval-augmented forecasting method. Experiments demonstrate performance improvements over several state-of-the-art baselines. Comparisons with more closely related methods and more targeted ablation studies would further strengthen the paper.

Critical strengths
  1. Meaningful motivation and problem formulation. Forecasting under temporal and spatial distribution shifts and missing observations is an important problem with practical relevance.
  2. Targeted methodological design. The proposed method is designed to address these challenges. In particular, dynamically retrieving relevant support nodes based on current observations is an interesting approach.
  3. Promising experimental results. The reported improvements over strong baselines, including STOP in the temporal generalization experiments, provide evidence of the method’s effectiveness.
Critical weaknesses
  1. Insufficiently established novelty. The methodological novelty and substantive contributions relative to existing generalizable spatio-temporal forecasting methods, such as UniST and STONE, are not clearly established.
  2. Limited comparisons with closely related methods. The generalization experiments would benefit from comparisons with additional relevant state-of-the-art methods under comparable evaluation settings.
  3. Incomplete ablation studies. The current ablations do not sufficiently isolate the contribution of dynamic retrieval. Comparisons against fixed graph neighbors and randomly selected support nodes, using the same historical context length and number of support nodes, would help clarify its effectiveness.
Questions for authors
  1. Could you add comparison to STONE?
  2. Could you add ablation experiments mentioned in Weakness 3?
2 comments

Thank you for recognizing our motivation, retrieval design, and generalization results, and for suggesting baselines and experiments that helped us strengthen the paper.

W1. Novelty relative to UniST and STONE.

First, we would like to clarify how VaryST differs from UniST and STONE. UniST generates adaptive prompts from learned spatio-temporal memory pools, while STONE learns invariant node dependencies through graph interventions. Our central contribution is the idea of dynamically constructing target-specific predictive contexts by retrieving support nodes using representations derived from available long-term observations. Neither UniST nor STONE explores this idea. We clarified this distinction in Section 2.1 of the revised manuscript.

W2 / Q1. Additional baseline comparisons.

Following your suggestion, we expand our baseline comparisons to include STONE (KDD 2024), UniST (KDD 2024), KITE (ICML 2026), and DST-Mamba (AAAI 2025). In standard in-domain forecasting, VaryST still achieves the lowest MAE, RMSE, and MAPE on all four PEMS datasets, as summarized below. The full results are reported in Table 2 of the revised manuscript (format: MAE / RMSE / MAPE).

DatasetUniSTKITEDST-MambaVaryST
PEMS0315.71 / 26.89 / 15.92%16.06 / 25.35 / 17.05%15.82 / 25.63 / 15.79%13.95 / 22.73 / 13.59%
PEMS0419.07 / 31.64 / 12.28%19.82 / 32.02 / 13.09%18.70 / 30.55 / 13.00%17.56 / 29.35 / 11.81%
PEMS0721.08 / 34.61 / 9.44%20.71 / 35.48 / 8.43%19.07 / 32.08 / 8.07%18.21 / 31.05 / 7.64%
PEMS0815.52 / 25.43 / 11.61%14.80 / 24.87 / 9.85%13.83 / 22.97 / 9.20%12.89 / 21.99 / 8.50%

For zero-shot temporal generalization on PEMSD3-Stream, models trained on 2011 data are evaluated on each year from 2012 to 2017 without parameter updates. VaryST continues to achieve the best performance among all compared methods. Detailed year-by-year comparisons are reported in Tables 6 and 7 of the revised manuscript.

Model (six-year average)MAERMSEMAPE (%)
STOP13.0122.4416.91
STD-PLM14.4025.0618.55
STONE14.9824.7123.63
VaryST10.4918.2214.05

W3 / Q2. Expanded ablation studies of the dynamic support retrieval module.

Following your suggestion and related feedback from another reviewer, we added no-support, random-support, fixed graph-hop, and k-means-based variants. The k-means variant clusters node patterns and uses cluster centroids as supporting information for broader pattern coverage. All other settings are fixed, varying only the support-selection strategy. The results on PEMS03 are shown below.

Support strategy (PEMS03)MAERMSEMAPE (%)
No support14.3323.4613.81
Random supports14.0622.9313.62
Fixed graph-hop supports14.1423.1613.64
K-means centroids13.9922.8513.33
Dynamic RBF retrieval13.9522.7313.59

All support-based variants outperform the no-support model, confirming the predictive value of support information. Dynamic RBF retrieval further outperforms random and fixed graph-hop supports across all three metrics, supporting the value of context-dependent selection. It achieves the lowest MAE (13.95) and RMSE (22.73), while k-means centroids achieve the lowest MAPE (13.33%). These results support our use of RBF retrieval as the default, while also indicating that broader pattern coverage can be beneficial. Our contribution centers on the dynamic retrieval mechanism; exploring selection strategies that combine similarity and complementary information is a promising future direction. We incorporated detailed results and discussion in Table 14, Appendix C.4 of the revised manuscript.

Dear Reviewer,

We would like to follow up on our rebuttal. When you have a chance, we would appreciate hearing whether our responses have addressed your concerns and whether there are any points you would like us to clarify. Thank you again for your time and feedback.

Reviewer_4hznOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents VaryST, a compact spatio-temporal forecasting framework designed for varying deployment environments. It constructs node representations from long-term observations, retrieves target-specific support nodes by representation similarity, and decodes a hierarchical context containing recent and long-term target states, retrieved supports, global context, and relative graph structure. The paper also introduces Importance-Guided Observation Perturbation (IGOP), which masks highly attended historical patches during training. Experiments on six real-world datasets evaluate standard forecasting, temporal and cross-network zero-/few-shot transfer, robustness to missing rates and patterns, ablations, and inference efficiency.

Critical strengths
  • Clear Problem Formulation: Addresses an important deployment challenge with a coherent unifying perspective: temporal shifts, spatial changes, and observation loss all make a forecaster overly dependent on its training environment.

  • Sensible Model Design: Constructing node representations and predictive contexts from currently available observations, rather than fixed node identities, provides a sound inductive design for transfer to changing networks.

  • Extensive Evaluation: The empirical evaluation is broad and the reported performance is consistently strong—covering six datasets, standard forecasting, six future years of zero-shot evaluation, all twelve transfer directions among four PEMS networks, multiple few-shot budgets, unseen missing rates and patterns, component ablations, hyperparameter sweeps, and efficiency analysis.

Critical weaknesses
  • The core dynamic support retrieval mechanism is not directly isolated, and its selection criterion lacks sufficient validation.
  • Similarity between long-term node representations does not necessarily translate to incremental predictive value once conditioned on the target history and global token; highly similar nodes can be redundant, whereas less similar nodes with leading or complementary signals may offer greater utility.
  • Current ablations compare RBF retrieval only against cosine retrieval and vary KK, omitting critical baselines such as no-support, random, fixed, graph-neighbor, lag-aware, or directly learned relevance controls. As a result, the performance gains cannot yet be confidently attributed to dynamic functional support selection.
  • The experiments do not establish that IGOP benefits specifically from importance-guided masking, as it is only compared against its complete removal rather than random patch masking or ordinary block masking under a matched masking budget.
  • VaryST receives the availability mask natively, whereas the reported baselines only receive linearly interpolated inputs; therefore, the advantage under missing observations cannot be cleanly attributed to IGOP without a mask-aware or otherwise matched baseline control.
Questions for authors

See Weaknesses.

1 comment

Thank you for your positive assessment and constructive questions, which helped us clarify our contributions and strengthen the experimental evidence.

W1. Value of dynamic functional support node retrieval.

Your comments helped us recognize that the original presentation and ablations did not sufficiently highlight the value of dynamic support retrieval.

First, our central contribution is the idea of dynamically retrieving support nodes based on the current context, using node representations derived from available long-term histories. RBF and cosine are alternative implementations of the similarity criterion within this framework.

Second, we considered your point that less similar nodes may provide complementary information. We tested a k-means-based approach that clusters node patterns and uses cluster centroids as supporting information for broader pattern coverage. Following related feedback from another reviewer, we also added no-support, random-support, and fixed graph-hop variants. In these comparisons, all other settings are held fixed, varying only the support-selection strategy. The results on PEMS03 are shown below.

Support strategy (PEMS03)MAERMSEMAPE (%)
No support14.3323.4613.81
Random supports14.0622.9313.62
Fixed graph-hop supports14.1423.1613.64
K-means centroids13.9922.8513.33
Dynamic RBF retrieval13.9522.7313.59

All support-based variants outperform the no-support model, confirming the predictive value of support information. Dynamic RBF retrieval further outperforms random and fixed graph-hop supports across all three metrics, supporting the value of context-dependent selection. It achieves the lowest MAE (13.95) and RMSE (22.73), while k-means centroids achieve the lowest MAPE (13.33%). These results support our use of RBF retrieval as the default, while also indicating that broader pattern coverage can be beneficial. Our contribution centers on the dynamic retrieval mechanism; exploring selection strategies that combine similarity and complementary information is a promising future direction.

We also appreciate your suggestions to explore lag-aware retrieval and directly learned relevance. Our current implementation uses discrete top-K selection, so learning the relevance criterion end-to-end would require additional changes to the retrieval mechanism. For the lag-aware variant, we would welcome clarification on the formulation you envision so that we can design a comparison that directly addresses your suggestion.

We have incorporated detailed results and discussion in Table 14, Appendix C.4 of the revised manuscript.

W2. Further investigation into the source of IGOP’s gains.

First, following your suggestion, we compare IGOP with random patch masking and block masking at the same masking budget of three patches per instance. On PEMS08, when trained at 30% point missingness and evaluated at 70% without retraining, IGOP achieves an MAE of 13.86, versus 13.91 for random patch masking and 13.96 for block masking (refer to Table 17 in the revised manuscript).

Second, to ensure a fairer comparison with baselines that receive only linearly interpolated inputs, we evaluate VaryST without the explicit availability-mask input, retaining IGOP and keeping all other settings fixed. As shown below, without this input, VaryST still outperforms the best baseline across all ten tested settings, showing that its advantage does not depend on access to additional missingness information.

Setting (PEMS08)w/ mask MAEw/o mask MAEBest baseline MAE
Same-rate: RM 30% → 30%12.9113.0413.61
Same-rate: CM 30% → 30%14.3013.8614.73
Cross-rate: RM 30% → 10%13.0513.0313.55
Cross-rate: RM 30% → 50%13.0713.2013.76
Cross-rate: RM 30% → 70%13.8313.8614.18
Cross-rate: CM 30% → 10%13.8413.3814.05
Cross-rate: CM 30% → 50%16.4215.6017.30
Cross-rate: CM 30% → 70%27.7624.9327.43
Cross-pattern: RM 30% → CM 30%15.0014.9215.33
Cross-pattern: CM 30% → RM 30%14.1413.3814.02

Comments on the paper

1 comment

We thank the reviewers for recognizing the novelty, practical significance, and strong empirical results of our work. The constructive feedback has helped us clarify our contributions and strengthen the experimental evidence. Below, we summarize the key updates in our rebuttal addressing the reviewers’ shared concerns.

  1. Additional baseline comparisons. Clarified our differences from UniST (KDD 2024) and STONE (KDD 2024), and added comparisons with UniST, STONE, KITE (ICML 2026), and DST-Mamba (AAAI 2025) on standard in-domain forecasting and zero-shot temporal generalization, where VaryST remains the best-performing method.
  2. Effectiveness of Dynamic Support Retrieval. We conducted controlled comparisons with no-support, random-support, fixed graph-hop, and k-means-based variants. The results validate the effectiveness of context-dependent support selection over static or random strategies, providing direct empirical evidence for the core innovative idea underlying VaryST.
  3. Sources of IGOP's Effectiveness and Sensitivity to KmK_m. We compared IGOP with alternative masking strategies, analyzed the sensitivity to KmK_m, and evaluated VaryST without explicit availability-mask inputs. The results confirm the effectiveness of importance-guided masking, the benefits of appropriately increasing KmK_m for cross-rate generalization, and that VaryST's superiority does not rely on additional missingness information.
  4. Clarification of Structural Encoding and Implementation Details. We clarified the construction of the correlation-based 5-NN graph for PM2.5 and the inductive computation of structural encodings on unseen graphs without updating model parameters, improving the reproducibility and clarity of our evaluation protocol.

All new results and clarifications have been incorporated into the revised manuscript (highlighted in red) and detailed in our individual responses below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.