pith. sign in

arxiv: 2605.06911 · v1 · submitted 2026-05-07 · 💻 cs.LG

Dual-Scale Temporal Fusion Reveals Structured Predictability in Subseasonal-to-Seasonal Temperature Prediction

Elnaz Bashir , Jiali Wang , Lin Yan This is my paper

Pith reviewed 2026-05-11 00:48 UTC · model grok-4.3

classification 💻 cs.LG
keywords subseasonal-to-seasonal forecastingtemperature predictiondual-scale fusionpredictability structurespatial heterogeneityclimate forecastingmachine learningweather prediction
0
0 comments X

The pith

S2S temperature predictability arises from spatially shifting balance between historical climate context and recent weather evolution rather than simple lead-time decay.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that subseasonal-to-seasonal temperature forecast skill stems from a structured, multi-scale organization involving temporal components, spatial heterogeneity, and large-scale pattern coherence. It develops a dual-scale learning framework that fuses calendar-aligned historical climate data with lead-time matched recent observations through spatially adaptive weights. This fusion reveals systematic shifts: winter forecasts over high latitudes and complex terrain draw more from interannual context, while summer forecasts balance both scales across regions. Topology-aware constraints further stabilize spatial coherence of the predicted fields. If the claim holds, forecast improvement would come from explicitly modeling these interactions instead of treating predictability as a uniform function of lead time.

Core claim

S2S predictability is organized across interacting temporal components, spatial heterogeneity, and large-scale pattern coherence that can be explicitly characterized and exploited; a dual-scale framework separates calendar-aligned historical climate context from lead-time matched recent weather evolution, fuses them via spatially adaptive weights, and shows that the resulting reorganization of predictability with season and geography, rather than lead-time decay, is the primary determinant of forecast skill in the 30-to-90-day window, with topology-aware constraints improving large-scale pattern stability especially over complex terrain.

What carries the argument

Dual-scale temporal fusion framework that separates calendar-aligned historical climate context from lead-time matched recent weather evolution and combines them through spatially adaptive fusion weights, augmented by topology-aware structural constraints.

If this is right

  • Winter temperature forecasts over high latitudes and complex terrain gain more from interannual historical context than from recent weather.
  • Summer forecasts reflect a more balanced contribution from both temporal scales across the domain.
  • Topology-aware constraints improve the spatial coherence of predicted temperature fields, particularly stabilizing large-scale patterns over complex terrain.
  • Forecast skill within the subseasonal window is driven primarily by this spatially explicit reorganization of predictability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same dual-scale fusion could be tested on other variables such as precipitation or wind to check for similar structured predictability.
  • Operational forecast systems might incorporate explicit multi-scale fusion layers to exploit the seasonal and regional shifts identified here.
  • The approach suggests that regional differences in S2S utility for agriculture or energy planning arise from these varying temporal balances rather than uniform skill decay.
  • Re-training the fusion on longer observational records or reanalysis could test whether the weight patterns remain stable across decades.

Load-bearing premise

The learned fusion weights and topology-aware constraints capture genuine underlying predictability structures in the climate system instead of fitting noise or artifacts specific to the training data.

What would settle it

Apply the trained fusion model to an independent dataset or different climate model and check whether the spatially adaptive weights still exhibit the reported systematic seasonal and geographic shifts in dominance between historical context and recent evolution.

Figures

Figures reproduced from arXiv: 2605.06911 by Elnaz Bashir, Jiali Wang, Lin Yan.

Figure 1
Figure 1. Figure 1: Conceptual overview of a dual-scale S2S temperature forecasting architecture. Historical climate context and recent weather evolution are processed by parallel generators and combined through spatially adaptive temporal fusion, producing regionally varying fusion weights. A lead-conditioned refinement network improves coherence of the fused forecast. Topology-aware structural descriptors are incorporated w… view at source ↗
Figure 2
Figure 2. Figure 2: Dual-scale formulation and forecast characteristics. (A) Conceptual schematic illustration of the dual-scale fusion framework, combining calendar-aligned interannual context and lead-time–matched recent dynamics. A spatially adaptive fusion weight λ is predicted by the model to modulate their relative contributions and produce the final forecast at lead time τ. (B) Forecast error and spatial correlation me… view at source ↗
Figure 3
Figure 3. Figure 3: Seasonal spatial structure and distributional consistency of S2S temperature prediction. Seasonal mean near-surface temperature over CONUS comparing ground truth (GT), model prediction (Pred), and spatial root-mean-square error (RMSE) averaged across lead times from 30–90 days (5-day intervals). Winter (DJF) and summer (JJA) are shown in separate rows. GT and Pred share identical color scales within each s… view at source ↗
Figure 4
Figure 4. Figure 4: Seasonal reorganization of the learned fusion weight λ and its dependence on predictability regimes. (A) Spatial distribution of λ during winter (DJF). (B) Spatial distribution of λ during summer (JJA). (C) Seasonal difference, ∆λ = λwinter −λsummer. Winter exhibits enhanced λ over inland and higher-latitude regions, indicating increased reliance on interannual variability under reduced subseasonal predict… view at source ↗
Figure 5
Figure 5. Figure 5: Topology-aware modeling improves spatial coherence and predictive accuracy. The first row (A–C) shows the annual mean temperature fields over the evaluation period, comparing the topology-aware prediction (A), the non-topological baseline (B), and the ground truth (C). The second row (D–F) presents the corresponding spatial RMSE for the topology-aware model (D), the baseline model (E), and their difference… view at source ↗
read the original abstract

Subseasonal-to-seasonal (S2S) temperature forecasts, spanning several weeks to a few months, are critically needed in agriculture practice, energy planning, and extreme-weather induced risk management, yet their reliability varies substantially across seasons and regions. Forecast skill is often attributed primarily to lead time, but this perspective does not fully explain the spatiotemporal patterns of predictability. Here we show that S2S predictability is organized across interacting temporal components, spatial heterogeneity, and large-scale pattern coherence, and that this structure can be explicitly characterized and exploited. We develop a dual-scale learning framework that separates calendar-aligned historical climate context from lead-time matched recent weather evolution, combining them through spatially adaptive fusion to enable stable temperature forecasts across the 30 to 90-day window. The learned fusion weights reveal that the balance between these two temporal scales shifts systematically with season and geography: during winter, interannual context dominates over high latitudes and complex terrain where forecast is the most difficult, while summer predictions reflect a more balanced temporal contribution across the domain. This spatially explicit reorganization of predictability, rather than simple lead-time decay, emerges as the primary determinant of forecast skill within the subseasonal window. Topology-aware structural constraints further improve spatial coherence of predicted temperature fields, stabilizing large-scale pattern organization particularly over complex terrain. These results reframe S2S predictability as a structured, multi-scale phenomenon, providing a more interpretable foundation for improving forecast systems and informing their use in practice.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces a dual-scale temporal fusion framework for subseasonal-to-seasonal (S2S) temperature prediction. It separates calendar-aligned historical climate context from lead-time-matched recent weather evolution, fuses them via spatially adaptive weights, and adds topology-aware structural constraints. The central claim is that the learned weights reveal a season- and geography-dependent reorganization of predictability (e.g., winter high-latitude dominance of interannual context) that, rather than simple lead-time decay, is the primary determinant of skill in the 30–90-day window, with the constraints further improving spatial coherence.

Significance. If the empirical claims are supported by rigorous controls, the work could reframe S2S predictability as an explicitly structured, multi-scale phenomenon rather than a monotonic function of lead time. The interpretability of the fusion weights and the emphasis on spatial coherence over complex terrain would be useful for applications in agriculture and energy planning. However, the current text supplies no quantitative skill scores, baseline comparisons, error bars, cross-validation details, or ablation studies, so the significance cannot yet be assessed.

major comments (3)
  1. [Results (weight interpretation and skill attribution)] The manuscript reports that fusion weights shift systematically with season and geography and that this reorganization is the primary driver of skill, yet no ablation is described that fixes the fusion weights to be spatially uniform or lead-time-only. Without such a control, it is impossible to isolate whether the spatially adaptive component explains skill beyond baseline temporal decay or added parameter flexibility (see the skeptic note on variance decomposition).
  2. [Methods (dual-scale fusion) and Results (weight analysis)] The interpretation that the learned weights 'reveal' underlying predictability structure is circular: the same optimization that produces the weights is also used to evaluate forecast skill. No independent validation (e.g., out-of-sample weight transfer or perturbation experiments) is reported to show that the observed winter/summer patterns are not an epiphenomenon of fitting.
  3. [Results (topology constraints)] Topology-aware constraints are stated to improve spatial coherence, but no quantitative metric (e.g., spatial autocorrelation, pattern correlation scores, or comparison to an unconstrained baseline) is provided to support this claim or to show that the improvement is not simply due to regularization.
minor comments (2)
  1. [Abstract and Introduction] The abstract and introduction repeatedly use 'primary determinant' without defining the term quantitatively; a variance-decomposition or relative-importance analysis should be added.
  2. [Methods] Notation for the two temporal scales (historical context vs. recent evolution) and the fusion weights should be introduced with explicit equations early in the methods section to avoid ambiguity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. These have prompted us to strengthen the manuscript with additional experiments and quantitative analyses. We address each major comment below and have incorporated revisions accordingly. We have also added overall skill scores, baseline comparisons, and cross-validation details to the revised text to address the broader concerns raised in the summary.

read point-by-point responses

  1. Referee: [Results (weight interpretation and skill attribution)] The manuscript reports that fusion weights shift systematically with season and geography and that this reorganization is the primary driver of skill, yet no ablation is described that fixes the fusion weights to be spatially uniform or lead-time-only. Without such a control, it is impossible to isolate whether the spatially adaptive component explains skill beyond baseline temporal decay or added parameter flexibility (see the skeptic note on variance decomposition).

    Authors: We agree that an explicit ablation is required to isolate the contribution of spatial adaptivity. In the revised manuscript we have added a dedicated ablation study (new Section 4.3 and Figure 5) that compares three controlled variants: (i) spatially uniform fusion weights, (ii) lead-time-only weights that are spatially constant, and (iii) the full spatially adaptive model. Skill differences are quantified via RMSE and anomaly correlation over the 30–90-day window, with error bars obtained from 10-fold cross-validation. The adaptive model improves RMSE by 12–18 % relative to the uniform case in winter high latitudes and complex terrain, while the lead-time-only variant captures only part of the gain. A variance decomposition (now included) attributes approximately 35 % of the explained skill variance to the spatially adaptive term beyond temporal decay and parameter count, directly addressing the skeptic note. revision: yes

  2. Referee: [Methods (dual-scale fusion) and Results (weight analysis)] The interpretation that the learned weights 'reveal' underlying predictability structure is circular: the same optimization that produces the weights is also used to evaluate forecast skill. No independent validation (e.g., out-of-sample weight transfer or perturbation experiments) is reported to show that the observed winter/summer patterns are not an epiphenomenon of fitting.

    Authors: We acknowledge the circularity concern and have added two independent validation protocols. First, we train on 1980–2010 and transfer the learned fusion weights (without retraining) to the fully held-out 2011–2020 period; the transferred weights preserve the reported winter high-latitude dominance and summer balance, and yield skill scores within 3 % of the in-sample model. Second, we perform perturbation experiments in which winter and summer weight maps are swapped across the test set; this degrades skill by 8–14 % relative to the original weights, demonstrating that the patterns are not merely fitting artifacts. These results are now reported in Section 4.4 with accompanying figures. revision: yes

  3. Referee: [Results (topology constraints)] Topology-aware constraints are stated to improve spatial coherence, but no quantitative metric (e.g., spatial autocorrelation, pattern correlation scores, or comparison to an unconstrained baseline) is provided to support this claim or to show that the improvement is not simply due to regularization.

    Authors: We have revised the results to include quantitative spatial-coherence metrics. We now report Moran’s I (global and local) and pattern-correlation scores against ERA5 reanalysis for both the topology-constrained and unconstrained models. The constrained model increases global Moran’s I by 0.22 (p < 0.01) and raises pattern correlation by 0.09–0.14 over complex terrain. To separate the effect from generic regularization, we compare against an L2-regularized unconstrained baseline with matched parameter count; the topology constraints still yield statistically higher coherence scores, indicating a benefit beyond simple regularization. These metrics and the associated statistical tests are added to Section 4.2 and Table 3. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain.

full rationale

The paper trains a dual-scale fusion model on historical and recent data to generate S2S temperature forecasts, then inspects the resulting learned fusion weights to describe seasonal and geographic shifts in temporal-scale balance. Forecast skill is measured independently via prediction error on held-out data, and the claim that weight-derived reorganization (rather than lead-time decay alone) is the primary skill determinant rests on observed correlations between weight patterns and skill maps. This does not reduce to a self-definitional equivalence, a fitted parameter renamed as prediction, or a self-citation chain; the weight patterns are an output of optimization, not an input assumption, and no uniqueness theorem or ansatz is smuggled in. The derivation remains self-contained against external benchmarks of forecast accuracy.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that S2S predictability decomposes cleanly into two temporal scales whose balance varies spatially and seasonally, plus standard neural-network fitting assumptions.

free parameters (2)
  • spatially adaptive fusion weights
    Learned parameters that determine the relative contribution of historical context versus recent evolution at each location and season.
  • topology constraint hyperparameters
    Parameters controlling the strength of structural constraints that enforce spatial coherence in the predicted fields.
axioms (2)
  • domain assumption S2S temperature predictability can be usefully separated into calendar-aligned historical climate context and lead-time-matched recent weather evolution
    This decomposition is the foundational design choice of the dual-scale framework.
  • ad hoc to paper Spatially adaptive fusion of the two scales plus topology constraints will improve forecast stability and coherence
    Core modeling assumption required for the claimed performance gains.

pith-pipeline@v0.9.0 · 5566 in / 1461 out tokens · 82567 ms · 2026-05-11T00:48:30.466577+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

26 extracted references · 26 canonical work pages

  1. [1]

    J.et al.Advances in the application and utility of subseasonal-to-seasonal predictions.Bull

    White, C. J.et al.Advances in the application and utility of subseasonal-to-seasonal predictions.Bull. Am. Meteorol. Soc. 103, E1448–E1472 (2022)

    work page 2022

  2. [2]

    J., Kirtman, B

    Becker, E. J., Kirtman, B. P., L’Heureux, M., Muñoz, Á. G. & Pegion, K. A decade of the north american multimodel ensemble (nmme): Research, application, and future directions.Bull. Am. Meteorol. Soc.103, E973 – E995, DOI: 10.1175/BAMS-D-20-0327.1 (2022)

    work page doi:10.1175/bams-d-20-0327.1 2022

  3. [3]

    The potential for skill across the range of the seamless weather-climate prediction problem: a stimulus for our science.Q

    Hoskins, B. The potential for skill across the range of the seamless weather-climate prediction problem: a stimulus for our science.Q. J. Royal Meteorol. Soc.139, 573–584 (2013). 8/10

    work page 2013

  4. [4]

    & Robertson, A

    Vitart, F. & Robertson, A. W. The sub-seasonal to seasonal prediction project (s2s) and the prediction of extreme events. npj climate atmospheric science1, 3 (2018)

    work page 2018

  5. [5]

    Madden—julian oscillation prediction and teleconnections in the s2s database.Q

    Vitart, F. Madden—julian oscillation prediction and teleconnections in the s2s database.Q. J. Royal Meteorol. Soc.143, 2210–2220 (2017)

    work page 2017

  6. [6]

    W., Vitart, F

    Robertson, A. W., Vitart, F. & Camargo, S. J. Subseasonal to seasonal prediction of weather to climate with application to tropical cyclones.J. Geophys. Res. Atmospheres125, e2018JD029375 (2020)

    work page 2020

  7. [7]

    & Vitart, F.Sub-seasonal to seasonal prediction: the gap between weather and climate forecasting(Elsevier, 2018)

    Robertson, A. & Vitart, F.Sub-seasonal to seasonal prediction: the gap between weather and climate forecasting(Elsevier, 2018)

    work page 2018

  8. [8]

    & McPherson, R

    Klemm, T. & McPherson, R. A. Assessing decision timing and seasonal climate forecast needs of winter wheat producers in the south-central united states.J. Appl. Meteorol. Climatol.57, 2129–2140 (2018)

    work page 2018

  9. [9]

    & Mackey, L

    Hwang, J., Orenstein, P., Cohen, J., Pfeiffer, K. & Mackey, L. Improving subseasonal forecasting in the western us with machine learning. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, 2325–2335 (2019)

    work page 2019

  10. [10]

    Sengupta, A.et al.Advances in subseasonal to seasonal prediction relevant to water management in the western united states.Bull. Am. Meteorol. Soc.103, E2168–E2175 (2022). 11.Zscheischler, J.et al.Future climate risk from compound events.Nat. climate change8, 469–477 (2018)

    work page 2022

  11. [11]

    G.et al.Flash droughts present a new challenge for subseasonal-to-seasonal prediction.Nat

    Pendergrass, A. G.et al.Flash droughts present a new challenge for subseasonal-to-seasonal prediction.Nat. Clim. Chang. 10, 191–199 (2020). 13.Leutbecher, M. & Palmer, T. N. Ensemble forecasting.J. computational physics227, 3515–3539 (2008). 14.Palmer, T. N.et al.Stochastic parametrization and model uncertainty.ECMWF Reading, UK(2009). 15.Mouatadid, S.et ...

    work page 2020

  12. [12]

    Rasp, S.et al.Weatherbench: a benchmark data set for data-driven weather forecasting.J. Adv. Model. Earth Syst.12, e2020MS002203 (2020)

    work page 2020

  13. [13]

    Rasp, S.et al.Weatherbench 2: A benchmark for the next generation of data-driven global weather models.J. Adv. Model. Earth Syst.16, e2023MS004019 (2024). 20.Extreme weather bench. https://github.com/brightbandtech/extremeweatherbench (2024)

    work page 2024

  14. [14]

    Chen, L.et al.A machine learning model that outperforms conventional global subseasonal forecast models.Nat. Commun. 15, 6425 (2024)

    work page 2024

  15. [15]

    InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 72–85 (2025)

    Hatanpää, V .et al.Aeris: Argonne earth systems model for reliable and skillful predictions. InProceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, 72–85 (2025)

    work page 2025

  16. [16]

    & Kotamarthi, R

    Stock, J., Arcomano, T. & Kotamarthi, R. Swift: An autoregressive consistency model for efficient weather forecasting. arXiv preprint arXiv:2509.25631(2025)

    work page arXiv 2025

  17. [17]

    Wheeler, M. C. & Hendon, H. H. An all-season real-time multivariate mjo index: Development of an index for monitoring and prediction.Mon. weather review132, 1917–1932 (2004). 25.Zhang, C. Madden-julian oscillation.Rev. Geophys.43(2005)

    work page 1917

  18. [18]

    & Waliser, D

    Kim, H., Vitart, F. & Waliser, D. E. Prediction of the madden–julian oscillation: A review.J. Clim.31, 9425–9443 (2018)

    work page 2018

  19. [19]

    I.et al.The role of the stratosphere in subseasonal to seasonal prediction: 1

    Domeisen, D. I.et al.The role of the stratosphere in subseasonal to seasonal prediction: 1. predictability of the stratosphere. J. Geophys. Res. Atmospheres125, e2019JD030920 (2020). 28.Hersbach, H.et al.The ERA5 global reanalysis.Q. journal royal meteorological society146, 1999–2049 (2020)

    work page 2020

  20. [20]

    P.et al.The era-interim reanalysis: Configuration and performance of the data assimilation system.Q

    Dee, D. P.et al.The era-interim reanalysis: Configuration and performance of the data assimilation system.Q. J. royal meteorological society137, 553–597 (2011)

    work page 2011

  21. [21]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, 770–778 (2016)

    work page 2016

  22. [22]

    & Brox, T

    Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical image computing and computer-assisted intervention, 234–241 (Springer, 2015). 9/10

    work page 2015

  23. [23]

    Topology-based visualization techniques for scientific data exploration.IEEE Comput

    Yan, L. Topology-based visualization techniques for scientific data exploration.IEEE Comput. Graph. Appl.45, 89–98 (2025)

    work page 2025

  24. [24]

    InComputer Graphics Forum, vol

    Yan, L.et al.Scalar field comparison with topological descriptors: Properties and applications for scientific visualization. InComputer Graphics Forum, vol. 40, 599–633 (Wiley Online Library, 2021)

    work page 2021

  25. [25]

    C., Sheikh, H

    Wang, Z., Bovik, A. C., Sheikh, H. R. & Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 600–612 (2004). 35.Goodfellow, I. J.et al.Generative adversarial nets.Adv. neural information processing systems27(2014). 36.Edelsbrunner, H., Harer, J.et al.Persistent homology-a surve...

    work page 2004

  26. [26]

    & Harer, J

    Cohen-Steiner, D., Edelsbrunner, H. & Harer, J. Stability of persistence diagrams. InProceedings of the twenty-first annual symposium on Computational geometry, 263–271 (2005). 10/10

    work page 2005