acceptodds
Under review as a conference paper at ICLR 2027

WxFM-XL: Adapting Univariate Foundation Models to Multi-Station Weather Forecasting

Abstract

With the rise of univariate time series foundation models (e.g., Sundial, Timer), initial efforts have been made to extend them to multivariate settings. However, these models mainly focus on modeling correlations among variables. When they are applied to multi-station weather forecasting, two important factors are often overlooked: (1) the spatial information of stations, and (2) different error priors of different stations relative to the foundation model. In this paper, we propose WxFM-XL, a model for adapting univariate time series foundation models to multi-station weather forecasting. WxFM-XL introduces a cross-station error correlation prior graph to capture stationwise error priors with respect to the foundation model. Building on this, we further propose a dynamic fusion mechanism that adaptively integrates a spatial correlation graph with the error correlation prior graph. Experiments on multiple datasets demonstrate that our model outperforms state of the art baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Multi-station weather forecasting, Time series foundation models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Q2UQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to adapt univariate time-series foundation models to multi-station weather forecasting, where both spatial dependencies and foundation-model-specific forecasting errors should be considered. The authors propose WxFM-XL, which constructs frequency-specific cross-station error correlation graphs and adaptively fuses them with a spatial correlation graph to guide a lightweight graph residual adapter. Experiments on several multi-station weather datasets using Timer and Sundial demonstrate improvements over various forecasting and foundation model adaptation baselines.

Critical strengths
  1. Clear task-specific motivation. The paper identifies a meaningful limitation of applying univariate TSFMs directly to multi-station weather forecasting, where geographic relationships and station-specific prediction errors are not explicitly modeled. The observation that geographically distant stations can exhibit correlated forecasting errors provides an interesting motivation for going beyond conventional spatial proximity graphs.
  2. Reasonable and coherent adaptation framework. The integration of spatial correlations, frequency-specific error correlations, and residual correction is well aligned with the problem formulation. Freezing the pretrained foundation model while introducing a graph-based adapter is a practical strategy for exploiting existing temporal priors while learning task-specific dependencies.
  3. Competitive empirical performance. The paper evaluates the proposed approach on multiple regional and global weather datasets, with comparisons against traditional spatiotemporal forecasting models and foundation model adaptation approaches. The ablation results provide initial evidence that frequency decomposition and error correlation graphs contribute to the overall performance.
Critical weaknesses
  1. Incomplete comparison with recent multivariate time-series foundation models. The evaluation mainly considers Timer, Sundial, Moirai, Timer-XL, and AdaPTS, but omits recent multivariate-capable TSFMs such as Chronos-2. Since the paper explicitly targets the limitations of extending univariate foundation models to multivariate or multi-station forecasting, comparisons with stronger recent models that natively support multivariate dependencies are important for establishing the competitiveness and practical relevance of WxFM-XL. Without such comparisons, it is difficult to determine whether the proposed adaptation framework offers advantages over modern multivariate foundation models.
  2. Insufficient evidence for the unique value of foundation-model-specific error correlations. The paper shows that adding error correlation graphs improves performance, but does not sufficiently demonstrate that these graphs provide information beyond conventional cross-station dependencies. In particular, comparisons with graphs constructed from raw observations, foundation model predictions, or directly learned station relationships are missing. Consequently, it remains unclear whether the improvement is specifically attributable to modeling foundation-model-specific error priors or simply to introducing additional cross-station connectivity.
  3. The necessity of frequency-specific graphs and dynamic fusion is not fully validated. Removing frequency decomposition or error correlation graphs degrades performance, but these ablations change multiple aspects of the model simultaneously. Comparisons between a shared error correlation graph and separate frequency-specific graphs are needed to establish the necessity of frequency-dependent error structures. Furthermore, the proposed fusion primarily relies on learned frequency-specific coefficients rather than genuinely input-dependent dynamic relationships, making the claimed dynamic adaptation somewhat limited. Comparisons with simpler fixed fusion strategies and more expressive dynamic gating mechanisms would strengthen the justification.
  4. Limited analysis of the stability and transferability of error correlation priors. Error correlation graphs are estimated from historical training residuals, yet these relationships may vary across seasons, weather regimes, and geographic regions. The paper does not sufficiently examine whether the estimated graphs remain reliable under temporal distribution shifts or when adapting to new regions with limited labeled observations. This raises concerns about the robustness and generalizability of the proposed error-prior-based adaptation mechanism.
  5. The efficiency advantages of lightweight adaptation are not fully established. Although the foundation model is frozen, constructing cross-station error graphs and introducing graph-based residual adapters may incur additional computational and memory costs, particularly for datasets containing thousands of stations. A systematic comparison of training cost, inference latency, memory consumption, and parameter efficiency would be useful to substantiate the claimed practicality of the approach.
Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Response to Reviewer Q2UQ (Part 1 / 2): W1–W3

Thank you for your constructive feedback. In this first part, we address W1–W3 with Chronos-2 comparisons, alternative station-relation graphs, and ablations of frequency-specific graphs and fusion strategies.

W1: Comparison with Chronos-2.

We added zero-shot and full fine-tuning comparisons with Chronos-2 on the Global dataset (24h; MSE; seed 2021). Both WxFM-XL variants achieve lower MSE than zero-shot and fine-tuned Chronos-2 on both variables, supporting their competitiveness against a recent multivariate foundation model.

MethodTemp-24Wind-24
Chronos-2 zero-shot8.1753.947
Chronos-2 full-FT7.5823.712
WxFM-XL (Timer)7.5793.591
WxFM-XL (Sundial)7.5183.660

W2: Comparison with alternative station-relation graphs.

We added comparisons with graphs constructed from (1) raw observations, (2) foundation-model predictions, and (3) directly learned station relationships, together with a geographic-proximity baseline. Results use WxFM-XL (Timer) on the Global dataset (24h; MSE; seed 2021).

The foundation-model-specific error correlation prior graphs achieve the lowest MSE on both variables. These comparisons support attributing the improvements to modeling foundation-model-specific error priors, beyond simply introducing additional cross-station connectivity.

Graph constructionTemp-24Wind-24
Geographic proximity only7.6683.638
Raw observations: correlation graph7.6113.657
Foundation-model predictions: Timer correlation graph7.6003.630
Directly learned station relationships: sparse graph7.5913.674
Foundation-model-specific error correlation prior graphs (ours)7.5793.591

W3: Frequency-specific error graphs and fusion strategies.

We compare frequency-specific error correlation prior graphs with a shared full-frequency error graph, retaining frequency decomposition to isolate the effect of graph construction. We additionally compare our learnable fusion with fixed coefficients and prediction-conditioned dynamic gating. Results use WxFM-XL (Timer) on the Global dataset (24h; MSE; seed 2021).

Prediction-conditioned dynamic gating generates fusion weights from the current foundation-model forecasts, allowing them to vary across samples and providing a more expressive fusion mechanism.

Frequency-specific graphs outperform the shared graph, supporting the value of modeling frequency-dependent error structures. Our learnable fusion also achieves lower MSE than fixed coefficients and prediction-conditioned dynamic gating on both variables. These results support learning a separate fusion weight for each frequency band to effectively integrate spatial relations and error correlation priors.

ConfigurationTemp-24Wind-24
Shared full-frequency error graph7.6333.721
Fixed fusion coefficients7.6173.647
Prediction-conditioned dynamic gating7.5973.768
Frequency-specific graphs + learnable fusion (ours)7.5793.591
AuthorsOpenReview ↗

Response to Reviewer Q2UQ (Part 2 / 2): W4–W5

In this second part, we address W4–W5 by evaluating robustness across seasons and weather regimes, adaptation under sparse regional station coverage, and computational efficiency.

W4: Robustness across seasons, weather regimes, and geographic regions.

We supplemented the evaluation using WxFM-XL (Timer) on the Global dataset (24h; MSE; seed 2021).

Across seasons. We compared the static training graph with seasonally re-estimated graphs, each constructed from all observed foundation-model forecast errors within a seasonal test block and used to evaluate that block. The August–December period covers summer/autumn/winter in the Northern Hemisphere and winter/spring/summer in the Southern Hemisphere.

Season-specific re-estimation changes overall MSE by only +0.52% for temperature and +0.02% for wind. The small wind improvements in two blocks show limited benefits from seasonal re-estimation, supporting the robustness of the static training graph under the observed seasonal changes.

Season: Northern / Southern Hemisphere (months)Temp: staticTemp: re-estimatedWind: staticWind: re-estimated
Summer / Winter (August)8.77248.85802.92552.9342
Autumn / Spring (September–November)7.05217.07393.63243.6320
Winter / Summer (December)8.46118.52413.94833.9471

Across weather regimes. We evaluated cold/normal/hot temperature and calm/normal/strong wind conditions, using thresholds determined from training data. Here, the re-estimated graph is periodically updated using recent foundation-model forecast errors whose corresponding observations are available before each prediction, whereas the static training graph remains unchanged.

Re-estimation does not improve forecasting accuracy in any evaluated weather condition. These results support the robustness of the static training graph and suggest that frequent re-estimation is unnecessary under the observed weather changes.

Variable / conditionWindowsStatic training graphRe-estimated graph
Temperature / cold1810.772011.0133
Temperature / normal3,0177.51077.6319
Temperature / hot2838.10578.1897
Wind / calm4912.85672.8852
Wind / normal2,6703.62923.6771
Wind / strong1575.23585.3428

Across geographic regions. We further evaluated regional adaptation on an Australia-region subset of the Global dataset, containing 90 geographically dispersed stations with a median nearest-neighbor distance of 108.62 km. We kept the pretrained Timer frozen, constructed regional spatial and error correlation prior graphs using training-period data, and trained a newly initialized lightweight adapter.

WxFM-XL reduces temperature and wind MSE by 23.08% and 8.24%, respectively, compared with Timer alone. These results demonstrate effective adaptation under sparse station coverage by reusing pretrained temporal knowledge and learning region-specific residual corrections, without retraining the foundation model.

MethodStationsTemp-24Wind-24
Timer9010.5575.256
WxFM-XL (Timer)908.1214.822

W5: Computational efficiency and practical feasibility.

Parameter efficiency. We added empirical comparisons of trainable parameters. WxFM-XL keeps Timer frozen and trains only 0.493M parameters, reducing trainable parameters by 99.4% relative to full fine-tuning.

Adaptation methodTrainable parameters
Timer full-FT84.142M
Timer LoRA0.524M
WxFM-XL (Timer)0.493M

Computational cost. We measured graph-construction time, training time, inference latency and peak memory across increasing station counts. Profiling on the Global dataset uses a 96h input and forecast horizon with a fixed top-k neighborhood size of 36.

Error-graph construction is performed offline and includes residual preparation, correlation computation and top-k selection. Training cost is measured per batch, and inference latency includes both the foundation model and adapter. These measurements demonstrate practical operation at thousands of stations.

DatasetStationsError graph (s)Training (ms/batch)FM+adapter inference (ms/window)Training GPU peak (GiB)Graph CPU peak (GiB)
Global2500.64046.115.40.631.38
Global5001.20050.420.90.871.57
Global1,0002.62545.726.91.361.94
Global2,0006.08247.239.52.442.67
Global3,85019.01793.577.24.824.11

The authors’ response strengthens the evaluation. The shared-graph comparison and the runtime measurements at 3,850 stations address my concerns about the graph design and practical cost. The new comparisons with Chronos-2 and alternative graphs are useful, though they cover one dataset, one forecast horizon, and one random seed. My overall assessment remains positive, and I retain my rating of 3 (Weak acceptance).

We are deeply thankful for the reviewer's timely acknowledgement and feedback. We are glad that our responses have addressed your concerns. Thank you again for your appreciation of our work and insightful comments.

Reviewer_Ef5TOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WxFM-XL, a residual correction framework that adapts univariate time series foundation models to joint forecasting across multiple weather stations. A key idea is to exploit correlations between the foundation model’s prediction errors at different stations. Specifically, the framework constructs frequency-specific error correlation graphs and integrates them with a geographic proximity graph to guide a graph-based residual adapter. The adapter predicts corrections that are added to the original forecasts from the frozen foundation model to produce the final predictions. Experiments across multiple weather station datasets and foundation model backbones demonstrate the effectiveness of the proposed framework.

Critical strengths
  1. The research problem is clearly defined and has practical value. The paper addresses the difficulty of directly exploiting cross-station information with univariate time series foundation models. It proposes an adaptation framework that enables existing models to perform joint forecasting across multiple weather stations, providing a clear application-driven motivation.

  2. The framework is modular and reuses pretrained models. The method keeps the foundation model frozen and refines its forecasts through an additional graph residual adapter, without updating the foundation model’s parameters. Evaluations with both Timer and Sundial demonstrate that the framework can accommodate different foundation models.

Critical weaknesses
  1. The task novelty and methodological contributions are insufficiently established. The paper essentially addresses forecast post-processing rather than station forecasting. Existing works, such as Graph Neural Networks and Spatial Information Learning for Post-Processing Ensemble Weather Forecasts (ICML 2024 Workshop), have already explored GNN-and-station-based correction of weather forecasts. The main difference appears to be applying this paradigm from numerical systems to time-series foundation models, with limited evidence of fundamental methodological novelty.

  2. The rationale for frequency decomposition and the description of its ablation study are insufficiently detailed. The authors divide the frequency spectrum into three approximately equal bands (low, mid, and high) but do not explain why this partition is chosen or evaluate how different numbers of bands or partitioning schemes affect the results.

  3. The term “dynamic fusion” is not fully consistent with the actual formulation. The error graphs are computed offline from the training set, the spatial graph is determined by fixed station locations, and the fusion coefficients and node gates are learned parameters. According to the equations in the paper, the graphs and their weights do not vary with input samples or weather conditions during inference. Therefore, the use of the term “dynamic fusion” warrants reconsideration.

  4. The evaluation of efficiency and generalization remains limited. Although the paper provides a theoretical complexity analysis, it lacks empirical comparisons of parameter counts, training time, inference latency, and GPU memory usage, leaving the claim that the method is lightweight insufficiently supported. Furthermore, the experiments primarily evaluate forecasting future observations at a fixed set of stations. The stability of the error graphs across different seasons or weather conditions, as well as the method’s applicability to newly added stations and cross-region transfer, has not been evaluated.

  5. In terms of presentation, the contribution list is rather generic and does not clearly highlight the most distinctive technical aspects of the work. The contributions on page 2 are summarized primarily as an adaptation framework, dynamic fusion, and experimental improvements over baselines. However, outperforming baselines is an empirical result, while the more distinctive aspects—foundation-model-specific error correlations, frequency-specific graph construction, and cross-station residual correction—are not clearly articulated in the contribution list. This weakens the reader’s initial impression of the paper’s novelty.

  6. Section 4.1(2) retains only positive Pearson correlations and zeroes out non-positive weights, implicitly assuming stationwise forecast errors are never negatively correlated. This discards potentially useful complementary error structure. Please explicitly justify the positive-only constraint if it is an engineering choice to avoid sign cancellation in graph propagation.

  7. The paper lacks direct visualization of the correction effect. While the figures illustrate graph structures and error correlations, a representative comparison of the original forecast error and the corrected forecast error across baselines in Fig. 1 format would help clarify the practical benefits of the proposed method.

Questions for authors
  1. The error correlation graph is constructed using Pearson correlation coefficients, which capture co-variation in errors across stations but do not directly reflect each station’s mean bias or error magnitude. Nevertheless, the adapter’s predicted corrections are added to the foundation model’s forecasts as numerical residuals. Could the authors clarify what information in the adapter’s inputs or training enables it to learn the bias and error magnitude that are not directly encoded in the correlation graph?

  2. The spatial graph uses Euclidean distances computed from standardized longitude, latitude, and elevation. Could the authors justify this distance definition for globally distributed stations? Specifically, how does it handle longitude wraparound at ±180° and the fact that the same longitude difference corresponds to different physical distances at different latitudes? Furthermore, does assigning equal weights to separately standardized longitude, latitude, and elevation adequately reflect the different effects of horizontal distance and elevation differences on meteorological relationships between stations?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Ef5T (Part 1): W1–W3

Thank you for your constructive feedback. In this first part, we address W1–W3 by clarifying our methodological contribution, presenting frequency-partition sensitivity experiments, and explaining the dynamic fusion mechanism.

W1: Distinctive methodological contribution.

Our goal is to exploit pretrained foundation-model knowledge for multi-station weather forecasting. The cited work uses geographically defined station relations, which cannot robustly leverage the pretrained foundation-model knowledge for performance improvement (see the experiment below). Our key contribution is the finding that frequency-specific error correlation prior graphs can effectively leverage this knowledge to improve forecasting performance. To validate this, we conducted an ablation study on the Global dataset. We compare train-from-scratch and pretrained Sundial backbones using either the spatial correlation graph alone or its combination with the error correlation prior graphs on the Global dataset (24h; MSE; seed 2021).

According to the results:

(1) For both scratch and pretrained Sundial, combining the error correlation prior graphs with the spatial correlation graph outperforms using the spatial correlation graph alone, supporting the benefit of modeling foundation-model-specific error correlations beyond geographic proximity.

(2) With only the geographic proximity graph, pretrained Sundial performs worse than scratch Sundial on temperature. Thus, using a pretrained backbone does not always improve forecasting when the adapter relies only on geographic relations. After adding our error correlation prior graphs, the temperature ranking reverses, and pretrained Sundial performs better on both temperature and wind, achieving the best overall results. This suggests that our error correlation priors help the adapter make better use of pretrained temporal knowledge.

Graph configurationBackboneTemp-24Wind-24
Spatial onlySundial-scratch + adapter8.1384.132
Spatial onlySundial-pretrained + adapter8.7323.995
Spatial + error correlationSundial-scratch + adapter8.0663.938
Spatial + error correlationSundial-pretrained + adapter (WxFM-XL)7.5183.660

W2: Frequency partitioning and sensitivity.

We added sensitivity experiments for WxFM-XL (Timer), comparing different band counts and partition schemes for temperature and wind forecasting on the Global dataset (24h; seed 2021).

Performance is similar across one to three equal-width bands, and more bands do not consistently improve accuracy. Two equal bands perform best for temperature, while three unequal bands perform best for wind. Three equal bands remain within 0.04% and 0.34% of the respective best MSEs, providing a simple, reproducible configuration shared across variables without variable-specific boundary tuning.

PartitionTemp-24Wind-24
No decomposition7.6533.689
2 equal bands7.5763.603
3 equal bands7.5793.591
6 equal bands7.8353.663
3 unequal bands7.6363.579

W3: Clarification of dynamic fusion.

Our fusion learns to combine the spatial correlation graph with the frequency-specific error correlation prior graphs. Here, “dynamic” refers to this learned, frequency-specific combination; the resulting graphs remain fixed during inference. The learned weights allow each frequency band to balance geographic proximity and foundation-model-specific error dependencies, guiding residual correction according to its error structure.

AuthorsOpenReview ↗

Response to Reviewer Ef5T (Part 2): W4

W4: Efficiency and generalization.

Parameter efficiency. We added empirical parameter and computational-cost measurements. WxFM-XL trains approximately 0.493M parameters while keeping Timer frozen, reducing trainable parameters by 99.4% compared with full fine-tuning.

MethodTrainable parameters
Timer full-FT84.142M
Timer LoRA0.524M
WxFM-XL (Timer)0.493M

Computational cost. Measurements on the Global dataset use a 96h input and forecast horizon, with a fixed top-k neighborhood size of 36. Graph construction is performed offline; training time is measured per batch, and inference latency includes both Timer and the adapter. These measurements demonstrate practical operation at thousands of stations.

StationsError graph (s)Training (ms/batch)FM+adapter inference (ms/window)Training GPU peak (GiB)Graph CPU peak (GiB)
5001.20050.420.90.871.57
2,0006.08247.239.52.442.67
3,85019.01793.577.24.824.11

Performance on seasonal changes. We evaluated WxFM-XL (Timer) on the Global dataset (24h; seed 2021). The August–December test period covers summer/autumn/winter in the Northern Hemisphere and winter/spring/summer in the Southern Hemisphere. For each seasonal block, we re-estimated a graph from all observed forecast errors in that block and held it fixed throughout the block.

Season-specific re-estimation changes overall MSE by only +0.52% for temperature and +0.02% for wind. Wind forecasting shows marginal improvements in the autumn/spring and winter/summer blocks, supporting the robustness of the static training graph under the observed seasonal changes.

Season: Northern / Southern Hemisphere (months)Temp: staticTemp: re-estimatedWind: staticWind: re-estimated
Summer / Winter (August)8.77248.85802.92552.9342
Autumn / Spring (September–November)7.05217.07393.63243.6320
Winter / Summer (December)8.46118.52413.94833.9471

Performance on weather regimes. We evaluated cold/normal/hot temperature and calm/normal/strong wind conditions, using thresholds determined exclusively from training data. Here, re-estimated graphs use recently observed forecast errors available before each prediction.

Re-estimating the graph does not improve forecasting accuracy in any evaluated weather condition. These results support the robustness of the static training graph and suggest that frequent re-estimation is unnecessary under the observed weather changes.

Variable / conditionWindowsStatic training graphRe-estimated graph
Temperature / cold1810.772011.0133
Temperature / normal3,0177.51077.6319
Temperature / hot2838.10578.1897
Wind / calm4912.85672.8852
Wind / normal2,6703.62923.6771
Wind / strong1575.23585.3428

Applicability to newly added stations. We evaluated station-network expansion using WxFM-XL (Timer) on the Global dataset (24h; MSE; seed 2021). The three configurations are:

(1) Training on 2,000 stations: train the adapter using only the original 2,000 geographically distributed stations.

(2) Expansion and continued training: expand configuration (1) to 3,850 stations by preserving shared adapter weights and original station gates, initializing new station gates, and rebuilding the graphs using training-period data. Then continue adapter training using all 3,850 stations.

(3) Direct training on 3,850 stations: train a new adapter using all stations from the outset as a reference.

The expanded model effectively adapts to the additional 1,850 stations while largely preserving performance at the original stations: temperature MSE improves from 7.358 to 7.262, while wind MSE changes by only +0.54%. This demonstrates that WxFM-XL can accommodate new stations through continued adapter training without substantial loss of accuracy at existing stations.

WxFM-XL configurationTemp: original 2,000Temp: new 1,850Temp: all 3,850Wind: original 2,000Wind: new 1,850Wind: all 3,850
Trained on 2,000 stations7.358——3.670——
Expanded to 3,850 with continued adapter training7.2627.5227.3873.6903.5373.616
Trained directly on 3,850 stations (reference)7.4597.7097.5793.6663.5103.591

Cross-region adaptation. For a new region, WxFM-XL can retain the frozen foundation model, reconstruct the spatial and error correlation prior graphs using target-station coordinates, observations and foundation-model residuals, and train a new lightweight adapter. This avoids retraining the large foundation model and keeps regional adaptation lightweight.

AuthorsOpenReview ↗

Response to Reviewer Ef5T (Part 3): W5–W7 and Q1

Here, we address W5–W7 and Q1 regarding contribution presentation, positive-only correlations, correction visualizations, and how the adapter learns correction magnitudes.

W5: Presentation of contributions.

Thank you for this constructive suggestion. We agree that the contribution list should foreground the distinctive technical ideas. We will revise it to make these contributions clearer:

  • Foundation-model-specific error priors, modeled through frequency-specific error correlation prior graphs to capture forecast error dependencies beyond geographic proximity.

  • A dynamic fusion mechanism that adaptively integrates the spatial correlation graph with frequency-specific error correlation prior graphs.

  • A lightweight graph residual adapter that performs cross-station residual correction while preserving the strong temporal priors of the frozen foundation model.

W6: Positive-only error correlations.

The error correlation prior graphs guide which stations exchange information and their relative aggregation weights. The Graph Residual Adapter learns the numerical residual corrections, which can be either positive or negative. Retaining only positive correlations between stationwise foundation-model forecast errors therefore does not prevent the adapter from learning corrections in either direction.

We use positive correlations to connect stations whose forecast errors tend to change in the same direction. To assess the effect of excluding negative correlations, we added a signed variant that propagates positive and negative edges through separate channels before fusion.

Results for WxFM-XL (Timer) on the Global dataset (24h; MSE; seed 2021) show that the signed variant does not improve either task, supporting positive-only propagation as a simple and effective choice in these settings. We will clarify the distinction between correlation signs and residual correction signs in the paper.

PropagationTemp-24Wind-24
Positive-only (ours)7.5793.591
Positive/negative two-channel7.5913.683

W7: Visualization of forecast correction.

Thank you for this helpful suggestion. We have added a figure to the Visualization Study on page 10 of the main text, combining regional station-error maps with representative single-station forecasting cases to directly illustrate the correction effects. Additional cases comparing observations, the original foundation-model forecasts, and WxFM-XL are provided on page 23 of the appendix.

Q1: Learning correction magnitudes from correlation graphs.

The error correlation prior graphs specify which stations exchange information and their relative aggregation weights, rather than directly encoding mean bias or error magnitude. The residual corrections are predicted by the Graph Residual Adapter. This adapter receives the frequency-decomposed foundation-model forecasts as input and learns the bias and error magnitude through supervision from the corresponding observations.

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

This paper proposes WxFM-XL, a framework for adapting univariate time-series foundation models to multi-station weather forecasting. It combines spatial relations with frequency-specific error correlation priors through learnable graph fusion and uses lightweight graph residual adapters to refine forecasts from frozen backbones. Experiments on regional and global weather datasets report improvements over the evaluated baselines.

Critical strengths
  1. The framework integrates frequency-specific error correlation priors with geographic relations through learnable graph fusion. Its lightweight graph residual adapters enable cross-station corrections while preserving the temporal priors of frozen foundation models.
  2. The visualizations in Figure 1 provide a clear empirical motivation for the design. They reveal correlated forecast errors across geographically distant stations and distinct graph structures across frequency bands, supporting the use of error-based connections beyond geographic proximity.
  3. The evaluation covers regional and global datasets, multiple forecasting horizons, and two foundation-model backbones, complemented by component ablations and temporal patch-size comparisons. The graph-structure analysis in Figure 4 further illustrates how frequency-specific error priors complement spatial relations.
Critical weaknesses
  1. The adaptation strategy needs stronger comparisons with alternative approaches. Timer and Sundial are evaluated only in zero-shot settings; full fine-tuning and LoRA baselines under comparable downstream training budgets would help establish the value of the proposed adapters. Although Timer-XL and Moirai are included, comparisons with newer multivariate foundation models such as Chronos-2 [1] would strengthen the evaluation. The paper also lacks a discussion of its relationship to weather models such as Aurora [2] and Pangu-Weather [3], accounting for differences in their inputs and grid-based forecasting settings.
  2. The experiments do not isolate the benefit of foundation-model pretraining from that of the residual correction framework. Existing ablations remove adapter components but do not apply the same correction framework to a conventional forecasting model trained from scratch. Comparing pretrained and trained-from-scratch versions of the same backbone architecture, each with and without the adapter, would clarify whether the gains specifically depend on pretrained temporal knowledge or primarily reflect the residual correction mechanism.
  3. Practical scalability to larger station networks remains insufficiently validated. The experiments already include 3,570- and 3,850-station datasets, and Appendix A.4 provides a complexity analysis, but no empirical scaling study is reported. Additional evaluation on WEATHER-5K [4] and the station datasets considered in MIGN [5], with graph-construction time, training/inference time, and peak memory measured as station count increases, would strengthen the scalability claim. This is particularly relevant because offline error-correlation construction costs O(N2)O(N^2), despite the adapter’s linear online complexity.

References:

[1] Abdul Fatir Ansari et al. Chronos-2: From Univariate to Universal Forecasting. 2025.

[2] Cristian Bodnar et al. A foundation model for the Earth system. Nature, 2025.

[3] Kaifeng Bi et al. Accurate medium-range global weather forecasting with 3D neural networks. Nature, 2023.

[4] Tao Han et al. Benchmarking Physics-Informed Time-Series Models for Operational Global Station Weather Forecasting. arXiv:2406.14399.

[5] Zinan Zheng, Yang Liu, and Jia Li. Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting. NeurIPS, 2025.

Questions for authors
  1. How robust are the training-derived error correlation graphs to distribution shifts, such as seasonal changes or changing weather regimes? Would these shifts require re-estimating the graphs, and if so, what observations would be needed?
3 comments
AuthorsOpenReview ↗

Response to Reviewer nBxc (Part 1 / 2): W1–W3

Thank you for your constructive feedback. In this first part, we address W1–W3 regarding additional baselines, the contributions of pretraining and residual correction, and empirical scalability. Our response to Q1 follows in Part 2.

W1: Additional adaptation baselines and weather foundation models.

Additional baselines. We added full fine-tuning and LoRA baselines for Timer and Sundial, together with zero-shot and fine-tuned Chronos-2. Results on the Global dataset are shown below (MSE; seed 2021).Both WxFM-XL variants achieve lower MSE than their respective full-FT/LoRA baselines and Chronos-2 full-FT in both settings. This supports the effectiveness of our adaptation strategy beyond comparison with zero-shot backbones.

MethodTemp-24Wind-24
Timer full-FT7.9203.775
Timer LoRA8.0233.796
Sundial full-FT10.6615.177
Sundial LoRA11.4095.412
Chronos-2 zero-shot8.1753.947
Chronos-2 full-FT7.5823.712
WxFM-XL (Timer)7.5793.591
WxFM-XL (Sundial)7.5183.660

Relationship to Aurora and Pangu-Weather. The key differences are as follows:

(1) Forecasting task. Aurora and Pangu-Weather predict global gridded atmospheric fields, whereas WxFM-XL predicts hourly observations at discrete weather stations.

(2) Input requirements. Aurora and Pangu-Weather require multiple surface and upper-air atmospheric variables, whereas WxFM-XL uses historical observations of the target variable at the stations.

(3) Spatial resolution. The spatial resolutions of Aurora and Pangu-Weather can be insufficient to capture local differences between densely located stations, which interpolation alone cannot recover.

W2: Separating pretraining from residual correction.

We compare pretrained and trained-from-scratch versions of Sundial, with and without our adapter. Each predictor uses error graphs constructed from its own training residuals. Results on the Global dataset are reported below (MSE; seed 2021).

According to the results in the table below:

(1) Sundial-scratch + adapter outperforms Sundial-scratch on both variables, supporting the effectiveness of the residual correction mechanism beyond pretrained models.

(2) Sundial-pretrained + adapter outperforms Sundial-scratch + adapter on both variables, supporting the additional value of pretrained temporal knowledge.

MethodTemp-24Wind-24
Sundial-scratch11.0895.228
Sundial-pretrained10.2664.914
Sundial-scratch + adapter8.0663.938
Sundial-pretrained + adapter (WxFM-XL)7.5183.660

W3: Empirical scalability.

We report graph-construction time, training/inference latency, and peak memory for WxFM-XL (Timer) across increasing station counts. Measurements on the Global dataset use a 96-hour input and prediction horizon with a fixed neighborhood size of 36.

GPU peak refers to adapter training. Error-graph timing includes residual preparation, correlation computation, and top-k selection. At 5,672 stations, complete offline preparation, including backbone forecast generation and geographic graph construction, takes 64.01 seconds.

The results show manageable graph-construction time, training/inference latency, and memory usage at thousands of stations, supporting practical scalability despite the quadratic cost of offline graph construction.

DatasetStationsError graph (s)Training (ms/batch)FM+adapter inference (ms/window)GPU peak (GiB)Graph CPU peak (GiB)
Global2500.64046.115.40.631.38
Global5001.20050.420.90.871.57
Global1,0002.62545.726.91.361.94
Global2,0006.08247.239.52.442.67
Global3,85019.01793.577.24.824.11
WEATHER-5K5,67235.887165.4123.17.646.59
AuthorsOpenReview ↗

Response to Reviewer nBxc (Part 2 / 2): Q1

In this second part, we address Q1 by evaluating the robustness of the error correlation prior graphs under seasonal changes and different weather regimes, and examining whether graph re-estimation improves forecasting accuracy.

Q1: Robustness of error graphs under seasonal changes and weather regimes.

We evaluated error-graph strategies using WxFM-XL (Timer) on the Global dataset (24h; MSE; seed 2021):

(1) Static training graph: retain the error correlation prior graph constructed from training residuals.

(2) Re-estimated graph: estimate a graph separately for each seasonal block in the seasonal analysis, or periodically estimate a recent graph in the weather-regime analysis.

Performance on seasonal changes. The August–December test period covers summer, autumn and winter in the Northern Hemisphere, and winter, spring and summer in the Southern Hemisphere. For each seasonal block, we re-estimated a graph from all observed forecast errors in that block and held it fixed throughout the block.

Season-specific re-estimation changes overall MSE by only +0.52% for temperature and +0.02% for wind. Wind forecasting shows small improvements in the autumn/spring and winter/summer blocks, but their magnitudes are marginal, supporting the robustness of the static training graph under the observed seasonal changes.

Season: Northern / Southern Hemisphere (months)Temp: staticTemp: re-estimatedWind: staticWind: re-estimated
Summer / Winter (August)8.77248.85802.92552.9342
Autumn / Spring (September–November)7.05217.07393.63243.6320
Winter / Summer (December)8.46118.52413.94833.9471

Performance on weather regimes. We evaluated cold/normal/hot temperature and calm/normal/strong wind conditions, using thresholds determined exclusively from training data. The re-estimated graphs are constructed from recently observed foundation-model forecast errors, using only observations available before each prediction.

Re-estimating the graph does not improve forecasting accuracy in any of the evaluated weather conditions. These results support the robustness of the static training graph and suggest that frequent re-estimation is unnecessary under the observed weather changes.

Variable / conditionWindowsStatic training graphRe-estimated graph
Temperature / cold1810.772011.0133
Temperature / normal3,0177.51077.6319
Temperature / hot2838.10578.1897
Wind / calm4912.85672.8852
Wind / normal2,6703.62923.6771
Wind / strong1575.23585.3428
AuthorsOpenReview ↗

Looking forward to further feedback

Thank you again for your valuable feedback. We have revised the manuscript and provided additional experiments in response to your suggestions. We hope our responses have satisfactorily addressed your concerns and would greatly appreciate your further feedback.

If you have any additional questions or comments, we would be happy to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.