Pith. sign in

REVIEW 5 major objections 6 minor 71 references

Station weather forecasting can be sharply improved by adding a third axis—aligned history from past years at the same month, day, and hour—rather than by extending the look-back window alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 06:29 UTC pith:FAQK2TII

load-bearing objection Real architectural novelty in the period-aligned historical axis, but the headline extreme-event metric is not standard SEDI and the robustness claim rests on keeping history pristine. the 5 major comments →

arxiv 2607.13101 v1 pith:FAQK2TII submitted 2026-07-14 cs.LG cs.AI

TSSM: Triaxial State Space Model for Global Station Weather Forecasting with Temporal-Variable-Historical Modeling

Songru Yang , Zili Liu , Tao Han , Ben Fei , Fenghua Ling , Lei Bai , Chang Liu , Xiangyang Ji
show 2 more authors
Zhenwei Shi Zhengxia Zou
This is my paper
classification cs.LG cs.AI
keywords Global station weather forecastingState space modelHistorical axisPeriod-aligned dataExtreme event predictionMissing observationsTime series forecastingWeather-5K
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that global station weather forecasting is held back not by model capacity but by a missing data axis: standard windows look only at recent hours, while the strongest predictor of tomorrow's temperature at a station is often the same calendar date in past years. TSSM folds multi-year station records into a time-variable-history tensor aligned by month, day, and hour, then scans each axis with a state space model, using deviations from the historical average to flag extremes. On Weather-5K, the largest such benchmark, it reports a 10% average-accuracy gain and a 61% gain in extreme-event metrics over all tested baselines, with advantages that grow at longer horizons and under iterative forecasting. If these results hold, a roughly 5M-parameter model with no extra inputs can beat much larger foundation models and an operational physics-based system on a third of station metrics, and can tolerate heavy missingness in current observations.

Core claim

The central claim is that period-aligned history—not longer look-back—is the missing ingredient in station weather forecasting. Although weather is chaotic forward in time, series from the same month, day, and hour across years form separable patterns; a simple historical average already supports extreme-event reasoning. TSSM builds on this by stacking past years into a third dimension, scanning it with bidirectional and forward state-space modules, and causally conditioning each prediction on both current look-back and prior years' forecasts. The paper reports 10% and 61% gains in accuracy and extreme-event metrics, best or second-best on 88% of Weather-5K results, and a 103.5% gain in the

What carries the argument

The central object is the time-variable-history tensor built by folding multi-year station records onto a calendar-aligned axis (same month/day/hour). Three SSM scanning branches process it: T-Scan for short-term temporal dynamics, V-Scan for inter-variable correlations (with shuffled variable order), and H-Scan for year-to-year evolution plus anomaly patterns, defined as SSM applied to the absolute deviation from the historical average. Hierarchical patching lets coarse and fine scales share weights, and causal forecasting ensures look-back information flows forward in both time and history with no future leakage.

Load-bearing premise

The robustness claim assumes that only the current look-back window is missing data—past years' aligned records are kept complete—so real sensor outages that also erase history would remove the model's main compensating signal.

What would settle it

Corrupt the historical dimension together with the look-back window at 80% missing rates and rerun the Weather-5K evaluation: if accuracy retention falls to the baseline's ~43% or below, the robustness claim does not survive. Separately, recompute the extreme-event gains using the standard SEDI formula (including false-alarm and miss rates) rather than the hit-rate ratio in Eq. (14); if the 61% gain shrinks materially, the headline extreme-event improvement is metric-dependent.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, any station-forecasting pipeline can adopt the historical axis without new inputs, since the model uses only the station's own past records.
  • Long-horizon forecasting becomes more reliable: the paper reports a 37.5% gain at 240h and a 103.5% gain in the 48h-by-5 iterative setting, with under 4.3% added error.
  • Extreme events, normally the weakest point of averaged predictors, are captured via historical deviation scanning; SEDI-style gains reach 61% overall and over 90% at the 99.5th percentile.
  • Robustness to missing current observations follows: TSSM retains over 90% of performance at 80% missing look-back, versus under 43% for baselines.
  • Because the same structure works with week-aligned history on non-weather series (95% best or second-best), the principle may transfer to other periodic human and energy time series.

Where Pith is reading between the lines

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

  • Editorial inference: the missing-data experiment corrupts only the current year's look-back while keeping the aligned historical years intact; real sensor outages also erode the historical axis, so the reported over-90% retention is likely an upper bound on real-world robustness.
  • Editorial inference: the extreme-event metric in Eq. (14) is a hit-rate normalized by climatological event frequency rather than the standard SEDI, which also accounts for false alarms; the 61% gain may not reproduce under the standard definition.
  • Editorial inference: the paper's own ablation shows a continuous accuracy-versus-extremes trade-off controlled by historical scope, suggesting a practical dial for deploying the model in different forecast regimes.
  • Editorial inference: if aligned history is the source of the gain, then the approach should transfer to any system with strong annual or weekly periodicity—e.g., air quality, river flow, energy load—and could be stress-tested by removing entire historical years at evaluation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes TSSM, a state-space architecture for global station weather forecasting that augments a 48-hour temporal look-back with a historical axis built by aligning observations from the same month/day/hour across previous years. The pipeline reshapes station series into an H×T×V tensor, applies temporal, variable, and historical SSM scans at multiple patch scales under a causal support scheme, and evaluates on the Weather-5K dataset plus hydrology and general time-series data. The headline claims are state-of-the-art accuracy and extreme-event capture (10% and 61% average gains), stronger long-horizon and iterative forecasting (37.5% at 240h, up to 103.5% under 48h×5 iteration), and >90% performance retention under up to 80% missing observations.

Significance. The historical-axis idea is clear, the model is comparatively compact (5M parameters vs 120M for Chronos2), and the evaluation is broad, including operational ECMWF-HRES, period-aware, memory-enhanced, and foundation-model baselines. If the central claims held under standard metrics, this would be a useful contribution to station-level forecasting. However, the extreme-event metric in Eq. (14) is not the Symmetric Extremal Dependence Index and can be trivially inflated; the missingness protocol is narrower than the abstract suggests; the historical-scope ablation reveals an internal accuracy/extreme trade-off that is not transparently reflected in the headline numbers; and no error bars or significance tests are reported. These issues are load-bearing for the SOTA claim and need to be resolved.

major comments (5)
  1. [§IV-A, Eq. (14)] The metric named SEDI99.5/SEDI90 is not the standard Symmetric Extremal Dependence Index. Eq. (14) computes (1/P(extreme)) times the number of correctly detected true extremes, i.e., a base-rate-normalized probability of detection with no false-alarm control. The standard SEDI uses both hit rate H and false-alarm rate F: SEDI = (ln F − ln H − ln(1−F) + ln(1−H))/(ln F + ln H + ln(1−F) + ln(1−H)). Under Eq. (14), predicting 'extreme' at every step yields a score of 1 because all true extremes are captured and false alarms are never penalized. Since the abstract's '61% extreme-event gain,' the '>90% improvement at the 99.5th percentile,' and the 103.5% iterative gain are all based on this score, the extreme-event component of the SOTA claim is currently unsupported. Please recompute all SEDI columns using the standard SEDI, or rename the metric and add false-alarm/ROC analysis.
  2. [§III-A, Eq. (1)] Equation (1) constructs the historical tensor by folding the time axis into H equal-length chunks with integer stride T_all/H. Weather-5K spans 2014–2024 and therefore contains leap years, so T_all is not divisible by H and the index h·(T_all/H)+i+t cannot keep 'same month/day/hour' aligned across years after each leap day. This drift undermines the core period-aligned historical signal that H-Scan is designed to exploit. Please provide the calendar-aware construction actually used (e.g., leap-day removal or explicit calendar indexing), or rewrite Eq. (1) to match the implementation.
  3. [§IV-E, last paragraph] The missingness experiment keeps the historical dimension clean: the paper states 'we do not introduce observation missing in the historical dimension.' The abstract's claim of 'retains >90% performance under up to 80% missing observations' is therefore only about corruption of the current look-back window, while the model's main compensating signal—aligned history—remains complete. In real observation networks, outages also corrupt past years. Please add experiments with missingness injected into the historical dimension, or explicitly qualify the abstract claim.
  4. [All tables, especially Table I] No table reports error bars, confidence intervals, or significance tests. This is consequential because Table I shows TSSM's 24h temperature MAE (2.36) is worse than HRES (1.76), WSSM (1.75), iTransformer (1.93), Chronos2 (2.05), Pyraformer (1.85), and others, and even at 120h the temperature MAE is 2.54, only marginally better than several baselines. The claimed 10% average accuracy gain and 88% best/second-best are aggregate statements over 72 cells; without multi-seed standard deviations or paired tests, the reader cannot assess robustness. Please report standard deviations over at least 3–5 seeds or paired bootstrap intervals.
  5. [§V and §IV-G2, Tables VIII–IX] The Section V limitation states that historical information 'does not consistently improve all metrics, but instead introduces a trade-off between average accuracy and extreme event forecasting.' This is visible in Table IX: restricting the historical scope from h_st=2014 to h_st=2020 improves temperature MAE (2.54→2.03) but sharply lowers SEDI (45.6→14.8). The headline '10% accuracy and 61% extreme-event gains' is an aggregate over this trade-off, and the choice of historical scope is a free parameter. Please decompose the headline gains by variable, horizon, and historical scope, and disclose the settings used for the abstract numbers.
minor comments (6)
  1. [§III-B, Eqs. (2)/(10)] Equation (10) writes \hat{Y}_{1:H} then extracts \hat{Y}_{H-1} = Y_{1:H}[-1]; the last expression should presumably be \hat{Y}_{1:H}[-1]. Please make the notation consistent.
  2. [§IV-A vs Table I] The text says prediction horizons are 24, 48, 120, and 240 hours, but Table I and several other tables report 24, 72, 120, 240 hours. Please align the wording with the actual horizon values or add the missing 48h rows.
  3. [§IV-A, Eq. (14)] The definition of q_p_low and q_p_up is ambiguous ('lower and upper pth percentiles'). Please specify the quantile convention used, e.g., empirical quantiles of the ground-truth distribution over the evaluation set.
  4. [Figure 4] The qualitative figure lacks axis labels and units, and the color legend is embedded only in prose. A clear legend would help verify the claims about high-frequency patterns and history correction.
  5. [Reproducibility] No code, random seeds, or training/inference hyperparameters are provided. Given the model complexity and the custom benchmark, a reproducibility statement with code or a detailed configuration table is strongly recommended.
  6. [Table III] In the Solar-Energy 240h row, TSSM reports MAE 2.3e-1 and MSE 1.5e-2, which appears inconsistent with the other rows and with the relationship between MAE and MSE. Please verify this entry.

Circularity Check

0 steps flagged

No significant circularity: empirical forecasting benchmark; metric and missing-data caveats are correctness risks, not circularity.

full rationale

The paper is an empirical architecture paper with no derivation-to-fit circularity. Its core claims rest on training TSSM with MSE on station data and evaluating MAE/MSE and extreme-event metrics on held-out data; the historical-axis reshaping (Eq. 1) is an input representation, not a fitted parameter renamed as a prediction. The 'SOTA' comparison includes externally published baselines retrained on the same benchmark, and the self-citations (Weather-5K dataset [27], WSSM [19]) are benchmark/comparison references rather than load-bearing justifications of the model design. No uniqueness theorem, ansatz, or fitted value from prior author work is invoked to force the architecture. I therefore do not identify any circular step. Two caveats are non-circular: Eq. (14) is a base-rate-normalized hit rate rather than the standard SEDI (a metric-validity/correctness risk), and Sec. IV-E states historical observations are not masked in the missing-data experiment (a scope limitation). These affect the strength of the claims but do not make the derivation equivalent to its inputs.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical entities (particles, forces, fields) are invented. The “historical axis” is a re-indexing of existing observations, not an independent entity.

free parameters (5)
  • Look-back window length T_o = 48 hours
    All models use a 48h look-back; chosen for comparison, not derived.
  • Historical depth H = 9-10 years (Weather-5K 2014-2024)
    Ablation Tab. IX shows changing h_st changes accuracy vs SEDI tradeoff; default is full history.
  • Hierarchical patch sizes / levels = not stated; default between 1 and 3 levels
    Eq. 12 and Tab. VIII show performance changes with level count; exact patch sizes are not reported.
  • Latent dimension d and number of TSSM blocks N
    Model capacity; no values or sensitivity analysis are provided.
  • Extreme-event thresholds p = 99.5 / 90.0
    Chosen percentiles for SEDI-like metric; results are sensitive to threshold choice, no robustness shown.
axioms (6)
  • domain assumption Same month/day/hour aligned observations across years are consistent and separable predictors of future weather
    Motivating observation in Sec. I and Fig. 1; not proven statistically beyond selected examples.
  • domain assumption Historical records are complete in missingness experiments
    Sec. IV-E explicitly keeps history uncorrupted; required for the >90% retention claim.
  • ad hoc to paper Anomaly magnitude |value - historical average| flags extreme events
    Eq. 6 builds H-Scan features from absolute deviations; useful heuristic but not validated as an extreme-event detector.
  • domain assumption Weather-5K is quality-controlled and representative of global station weather
    Sec. IV-A claims integration with HadISD/ICOADS; no independent audit in this paper.
  • standard math State space scanning provides sufficient temporal/variable/historical dependency modeling
    Relies on Mamba/SSM literature [15], [60]; accepted in the field, not proven here.
  • domain assumption Previous years' future windows Y0:H-1 are available at inference and are legitimate inputs
    Sec. III-B; true for archived data but means the method needs multi-year complete archives.

pith-pipeline@v1.3.0-alltime-deepseek · 33642 in / 15500 out tokens · 154924 ms · 2026-08-02T06:29:34.671702+00:00 · methodology

0 comments
read the original abstract

Global Station Weather Forecasting (GSWF) is pivotal for localized and extreme weather prediction over key regions. Despite efforts to exploit look-back windows, existing methods show limited accuracy gains and struggle with extreme events and error accumulation. These limitations stem from overreliance on short-term patterns, which are insufficient to capture chaotic weather dynamics, especially under partial observations. To address this problem, we propose a novel Triaxial State Space Model (TSSM) with a history-enhanced Temporal-VariableHistorical paradigm, which incorporates period-aligned historical weather data to compensate for long-term, large-scale periodic, and full-window weather patterns beyond the temporal lookback window. Specifically, TSSM stacks historical samples into period-aligned batches, where forecasting is causally supported by historical and current observations. Temporal, variable, and historical scanning are designed to capture axial temporal dependencies, variable correlations, and historical evolution. This structure is hierarchically shared to model seasonal to extreme events while alleviating misalignment across historical patterns. TSSM achieves SOTA performance on Weather-5K, the largest station weather dataset to date, with 10% and 61% gains in accuracy and extreme event metrics, and obtains 95% best or second-best results on human-involved datasets. Its advantages are more pronounced in long-horizon and iterative forecasting, reaching a 37.5% gain at 240h and up to 103.5% under a 48h times 5 iterative setting. Moreover, TSSM retains > 90% performance under up to 80% missing observations, compared with < 43% for baselines, demonstrating robustness and practical potential for reliable GSWF in global in-situ observation networks.

Figures

Figures reproduced from arXiv: 2607.13101 by Ben Fei, Chang Liu, Fenghua Ling, Lei Bai, Songru Yang, Tao Han, Xiangyang Ji, Zhengxia Zou, Zhenwei Shi, Zili Liu.

Figure 1
Figure 1. Figure 1: Consistent and separable historical weather patterns across aligned [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our extension of GSWF to the historical dimension through (a) data structure, (b) modeling strategy, (c) forecasting framework designed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Detailed architecture of the TSSM including (a) model architecture and forecasting pipeline, (b) triaxial reorganized data preprocessing, and (c) design [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: It can be observed in gray boxes that our method [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of 120h forecasting results across all variables, together with our corresponding results under noised look-back window. The blue solid [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Robustness comparison under observation missingness. The left panels [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of historical scope. Both curves are computed relative to the [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

71 extracted references · 9 linked inside Pith

  1. [1]

    [Online]

    WMO,Manual on the Global Observing System, World Meteorological Organization, Geneva, Switzerland, 2018. [Online]. Available: https: //library.wmo.int/doc num.php?explnum id=1000000000067

    2018

  2. [2]

    Interpretable weather forecasting for worldwide stations with a unified deep model,

    H. Wu, H. Zhou, M. Long, and J. Wang, “Interpretable weather forecasting for worldwide stations with a unified deep model,”Nature Machine Intelligence, vol. 5, no. 6, pp. 602–611, 2023

    2023

  3. [4]

    How climate change affects extreme weather events,

    P. Stott, “How climate change affects extreme weather events,”Science, vol. 352, no. 6293, pp. 1517–1518, 2016

    2016

  4. [5]

    On the foundations of earth foundation models,

    X. X. Zhu, Z. Xiong, Y . Wang, A. J. Stewart, K. Heidler, Y . Wang, Z. Yuan, T. Dujardin, Q. Xu, and Y . Shi, “On the foundations of earth foundation models,”Communications Earth & Environment, 2026

    2026

  5. [6]

    Are transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 9, 2023, pp. 11 121–11 128

    2023

  6. [7]

    Cyclenet: Enhancing time series forecasting through modeling periodic patterns,

    S. Lin, W. Lin, X. Hu, W. Wu, R. Mo, and H. Zhong, “Cyclenet: Enhancing time series forecasting through modeling periodic patterns,” inThirty-eighth Conference on Neural Information Processing Systems, 2024

    2024

  7. [8]

    Deriving accurate surface meteorological states at arbitrary locations via observation-guided continous neural field modeling,

    Z. Liu, H. Chen, L. Bai, W. Li, K. Chen, Z. Wang, W. Ouyang, Z. Zou, and Z. Shi, “Deriving accurate surface meteorological states at arbitrary locations via observation-guided continous neural field modeling,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  8. [9]

    Transductive lstm for time-series prediction: An application to weather forecasting,

    Z. Karevan and J. A. Suykens, “Transductive lstm for time-series prediction: An application to weather forecasting,”Neural Networks, vol. 125, pp. 1–9, 2020

    2020

  9. [10]

    Recurrent neural net- works for time series forecasting: Current status and future directions,

    H. Hewamalage, C. Bergmeir, and K. Bandara, “Recurrent neural net- works for time series forecasting: Current status and future directions,” International Journal of Forecasting, vol. 37, no. 1, pp. 388–427, 2021

    2021

  10. [11]

    Fedformer: Frequency enhanced decomposed transformer for long-term series fore- casting,

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency enhanced decomposed transformer for long-term series fore- casting,” inInternational conference on machine learning. PMLR, 2022, pp. 27 268–27 286

    2022

  11. [12]

    itrans- former: Inverted transformers are effective for time series forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “itrans- former: Inverted transformers are effective for time series forecasting,” arXiv preprint arXiv:2310.06625, 2023

    Pith/arXiv arXiv 2023

  12. [13]

    Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting,

    S. Liu, H. Yu, C. Liao, J. Li, W. Lin, A. X. Liu, and S. Dustdar, “Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting,” inInternational conference on learning representations, 2021

    2021

  13. [14]

    Timepro: Efficient multivariate long-term time series forecasting with variable-and time-aware hyper- state,

    X. Ma, Z. Ni, S. Xiao, and X. Chen, “Timepro: Efficient multivariate long-term time series forecasting with variable-and time-aware hyper- state,”arXiv preprint arXiv:2505.20774, 2025. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    Pith/arXiv arXiv 2025

  14. [15]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  15. [16]

    Mambads: Near-surface meteorological field downscaling with topog- raphy constrained selective state space modeling,

    Z. Liu, H. Chen, L. Bai, W. Li, W. Ouyang, Z. Zou, and Z. Shi, “Mambads: Near-surface meteorological field downscaling with topog- raphy constrained selective state space modeling,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  16. [17]

    Regional weather variable predictions by machine learning with near-surface observational and atmospheric numerical data,

    Y . Huanget al., “Regional weather variable predictions by machine learning with near-surface observational and atmospheric numerical data,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  17. [18]

    The quiet revolution of numerical weather prediction,

    P. Bauer, A. Thorpe, and G. Brunet, “The quiet revolution of numerical weather prediction,”Nature, vol. 525, no. 7567, pp. 47–55, 2015

    2015

  18. [19]

    Wssm: Geographic-enhanced hierarchical state-space model for global station weather forecast,

    S. Yang, Z. Liu, Z. Shi, and Z. Zou, “Wssm: Geographic-enhanced hierarchical state-space model for global station weather forecast,”arXiv preprint arXiv:2501.11238, 2025

    Pith/arXiv arXiv 2025

  19. [20]

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,

    H. Wu, J. Xu, J. Wang, and M. Long, “Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting,” in Advances in Neural Information Processing Systems, 2021

    2021

  20. [21]

    G. E. P. Box, G. M. Jenkins, G. C. Reinsel, and G. M. Ljung,Time Series Analysis: Forecasting and Control, 5th ed. Hoboken, NJ: John Wiley & Sons, 2015

    2015

  21. [22]

    Distribution of the estimators for autoregressive time series with a unit root,

    D. A. Dickey and W. A. Fuller, “Distribution of the estimators for autoregressive time series with a unit root,”Journal of the American Statistical Association, vol. 74, no. 366, pp. 427–431, 1979

    1979

  22. [23]

    Stationarity is dead: Whither water management?

    P. C. D. Milly, J. Betancourt, M. Falkenmark, R. M. Hirsch, Z. W. Kundzewicz, D. P. Lettenmaier, and R. J. Stouffer, “Stationarity is dead: Whither water management?”Science, vol. 319, no. 5863, pp. 573–574, 2008

    2008

  23. [24]

    Indication of a universal persistence law governing atmospheric variability,

    E. Koscielny-Bunde, A. Bunde, S. Havlin, H. E. Roman, Y . Goldreich, and H.-J. Schellnhuber, “Indication of a universal persistence law governing atmospheric variability,”Physical Review Letters, vol. 81, no. 3, pp. 729–732, 1998

    1998

  24. [25]

    Atmospheric predictability as revealed by naturally occurring analogues,

    E. N. Lorenz, “Atmospheric predictability as revealed by naturally occurring analogues,”Journal of the Atmospheric Sciences, vol. 26, no. 4, pp. 636–646, 1969

    1969

  25. [26]

    Temperature prediction using the missing data refinement model based on a long short-term memory neural network,

    I. Park, H.-S. Kim, J.-Y . Lee, J. Kim, C.-H. Song, and H.-J. Kim, “Temperature prediction using the missing data refinement model based on a long short-term memory neural network,”Atmosphere, vol. 10, no. 11, p. 718, 2019

    2019

  26. [27]

    How far are today’s time- series models from real-world weather forecasting applications?

    T. Han, S. Guo, W. Xu, and L. Bai, “How far are today’s time- series models from real-world weather forecasting applications?”arXiv preprint arXiv:2406.14399, 2024

    Pith/arXiv arXiv 2024

  27. [28]

    Mcann: A mixture clustering-based attention neural network for multivariate time series forecasting,

    Y . Li and D. C. Anastasiu, “Mcann: A mixture clustering-based attention neural network for multivariate time series forecasting,” inBayLearn 2025, 2025

    2025

  28. [29]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  29. [30]

    Modeling long- and short-term temporal patterns with deep neural networks,

    G. Lai, W.-C. Chang, Y . Yang, and H. Liu, “Modeling long- and short-term temporal patterns with deep neural networks,” inThe 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, 2018, pp. 95–104

    2018

  30. [31]

    Accurate medium-range global weather forecasting with 3d neural networks,

    K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, and Q. Tian, “Accurate medium-range global weather forecasting with 3d neural networks,” Nature, vol. 619, no. 7970, pp. 533–538, 2023

    2023

  31. [32]

    Learning skillful medium-range global weather forecasting,

    R. Lam, A. Sanchez-Gonzalez, M. Willson, P. Wirnsberger, M. For- tunato, F. Alet, S. Ravuri, T. Ewalds, Z. Eaton-Rosen, W. Huet al., “Learning skillful medium-range global weather forecasting,”Science, vol. 382, no. 6677, pp. 1416–1421, 2023

    2023

  32. [33]

    Gencast: Diffusion-based ensemble forecasting for medium-range weather,

    DeepMind, “Gencast: Diffusion-based ensemble forecasting for medium-range weather,”arXiv preprint, 2024

    2024

  33. [34]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators,

    J. Pathak, S. Subramanian, P. Harrington, S. Raja, A. Chattopad- hyay, M. Mardani, T. Kurth, D. Hall, Z. Li, K. Azizzadenesheli, and A. Anandkumar, “Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators,”arXiv preprint arXiv:2202.11214, 2022

    Pith/arXiv arXiv 2022

  34. [35]

    Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast,

    L. Chen, X. Zhong, F. Zhang, Y . Cheng, Y . Xu, Y . Qi, H. Li, M. Tang, R. Gao, M. Wanget al., “Fuxi: A cascade machine learning forecasting system for 15-day global weather forecast,”npj Climate and Atmo- spheric Science, vol. 6, no. 1, p. 190, 2023

    2023

  35. [36]

    Skillful twelve hour pre- cipitation forecasts with large context neural networks,

    L. Espeholt, S. Agrawal, S. Byrneet al., “Skillful twelve hour pre- cipitation forecasts with large context neural networks,”arXiv preprint arXiv:2210.12065, 2022

    Pith/arXiv arXiv 2022

  36. [37]

    The era5 global reanalysis,

    H. Hersbach, B. Bell, P. Berrisford, S. Hirahara, A. Hor ´anyi, J. Mu ˜noz- Sabater, J. Nicolas, C. Peubey, R. Radu, D. Schepers, A. Simmons, C. Soci, S. Abdalla, X. Abellan, G. Balsamo, P. Bechtold, G. Bia- vati, J.-R. Bidlot, M. Bonavita, G. De Chiara, P. Dahlgren, D. Dee, M. Diamantakis, R. Dragani, J. Flemming, R. Forbes, M. Fuentes, A. Geer, L. Haimb...

    1999

  37. [38]

    Skillful radar-based heavy rainfall nowcasting using task-segmented generative adversarial network,

    Y . Wanget al., “Skillful radar-based heavy rainfall nowcasting using task-segmented generative adversarial network,”IEEE Transactions on Geoscience and Remote Sensing, 2023

    2023

  38. [39]

    A short-long term sequence learning network for pre- cipitation nowcasting,

    Z. Liet al., “A short-long term sequence learning network for pre- cipitation nowcasting,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  39. [40]

    Coordinate-transformed dynamic mode decomposition for short-term rainfall forecasting,

    Q. Yuet al., “Coordinate-transformed dynamic mode decomposition for short-term rainfall forecasting,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  40. [41]

    Improved precipitation nowcasting through a deep learning model based on three-dimensional cloud structures,

    M. Chenet al., “Improved precipitation nowcasting through a deep learning model based on three-dimensional cloud structures,”IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  41. [42]

    A deep learning-based precipitation nowcasting model fusing gnss-pwv and radar echo observations,

    H. Liuet al., “A deep learning-based precipitation nowcasting model fusing gnss-pwv and radar echo observations,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  42. [43]

    Rsg-gan: A gan-based precipitation nowcasting model integrating radar qpe, goes-16 swd, and gnss ztds,

    R. Zhanget al., “Rsg-gan: A gan-based precipitation nowcasting model integrating radar qpe, goes-16 swd, and gnss ztds,”IEEE Transactions on Geoscience and Remote Sensing, 2025

    2025

  43. [44]

    Unveiling the role of weighted loss functions in deep learning-based nowcasting of extreme rainfall events,

    Q. Liuet al., “Unveiling the role of weighted loss functions in deep learning-based nowcasting of extreme rainfall events,”IEEE Transac- tions on Geoscience and Remote Sensing, 2025

    2025

  44. [45]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” inInterna- tional Conference on Learning Representations, 2023

    2023

  45. [46]

    Timesnet: Temporal 2d-variation modeling for general time series analysis,

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis,” in International Conference on Learning Representations, 2023

    2023

  46. [47]

    Earthformer: Exploring space-time transformers for earth system fore- casting,

    Z. Gao, X. Shi, H. Wang, Y . Zhu, Y . Wang, M. Li, and D.-Y . Yeung, “Earthformer: Exploring space-time transformers for earth system fore- casting,” inAdvances in Neural Information Processing Systems, 2022

    2022

  47. [48]

    Softs: Efficient multi- variate time series forecasting with series-core fusion,

    L. Han, X.-Y . Chen, H.-J. Ye, and D.-C. Zhan, “Softs: Efficient multi- variate time series forecasting with series-core fusion,” inAdvances in Neural Information Processing Systems, 2024

    2024

  48. [49]

    Non-stationary transformers: Exploring the stationarity in time series forecasting,

    Y . Liu, H. Wu, J. Wang, and M. Long, “Non-stationary transformers: Exploring the stationarity in time series forecasting,” inAdvances in Neural Information Processing Systems, 2022

    2022

  49. [50]

    Time-ffm: Towards lm- empowered federated foundation model for time series forecasting,

    Q. Liu, X. Liu, C. Liu, Q. Wen, and Y . Liang, “Time-ffm: Towards lm- empowered federated foundation model for time series forecasting,” in Advances in Neural Information Processing Systems, 2024

    2024

  50. [51]

    Mofo: Empowering long-term time series forecasting with periodic pat- tern modeling,

    J. Ma, B. Wang, Q. Huang, G. Wang, P. Wang, Z. Zhou, and Y . Wang, “Mofo: Empowering long-term time series forecasting with periodic pat- tern modeling,” inAdvances in Neural Information Processing Systems, 2025

    2025

  51. [52]

    Climax: A foundation model for weather and climate,

    T. Nguyen, R. Shah, H. Bansal, T. Arcomano, S. Madireddy, R. Maulik, K. Kashinathet al., “Climax: A foundation model for weather and climate,” inInternational Conference on Machine Learning, 2023

    2023

  52. [53]

    Skilful precipitation nowcasting using deep generative models of radar,

    S. Ravuri, K. Lenc, M. Willson, D. Kangin, R. Lam, P. Mirowski, M. Fitzsimons, M. Athanassiadou, S. Kashem, S. Madgeet al., “Skilful precipitation nowcasting using deep generative models of radar,”Nature, vol. 597, no. 7878, pp. 672–677, 2021

    2021

  53. [54]

    From news to forecast: Integrating event analysis in llm-based time series forecasting with reflection,

    X. Wang, M. Feng, J. Qiu, J. Gu, and J. Zhao, “From news to forecast: Integrating event analysis in llm-based time series forecasting with reflection,” inAdvances in Neural Information Processing Systems, 2024

    2024

  54. [55]

    Corrformer: Correlation-aware transformer for multivariate time series forecasting,

    R. Wu, X. Wang, W. Chen, and X. Li, “Corrformer: Correlation-aware transformer for multivariate time series forecasting,” inAdvances in Neural Information Processing Systems, 2023

    2023

  55. [56]

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting,

    Y . Zhang, J. Yan, W. Jin, X. Wang, S. Yang, and H. Zha, “Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting,” inInternational Conference on Learning Represen- tations, 2023

    2023

  56. [57]

    Multivariate probabilistic time series fore- casting with correlated errors,

    V . Z. Zheng and L. Sun, “Multivariate probabilistic time series fore- casting with correlated errors,” inAdvances in Neural Information Processing Systems, 2024

    2024

  57. [58]

    A decoder-only foundation model for time-series forecasting,

    A. Das, W. Kong, R. Sen, and Y . Zhou, “A decoder-only foundation model for time-series forecasting,”arXiv preprint arXiv:2310.10688, 2023

    Pith/arXiv arXiv 2023

  58. [59]

    Chronos: Learning the language of time series,

    A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. Arango, S. Kapoor, D. C. Maddix et al., “Chronos: Learning the language of time series,”Transactions on Machine Learning Research, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 15

    2024

  59. [60]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. Re, “Efficiently modeling long sequences with structured state spaces,” inInternational Conference on Learning Representations, 2022

    2022

  60. [61]

    Transformers are ssms: Generalized models and ef- ficient algorithms through structured state space duality,

    T. Dao and A. Gu, “Transformers are ssms: Generalized models and ef- ficient algorithms through structured state space duality,” inProceedings of the 41st International Conference on Machine Learning, 2024

    2024

  61. [62]

    Phaseformer: From patches to phases for efficient and effective time series forecasting,

    Y . Niu, J. Deng, and Y . Tong, “Phaseformer: From patches to phases for efficient and effective time series forecasting,” inInternational Conference on Learning Representations, 2026

    2026

  62. [63]

    Diffusion-ts: Interpretable diffusion for general time series generation,

    X. Yuan and Y . Qiao, “Diffusion-ts: Interpretable diffusion for general time series generation,” inInternational Conference on Learning Rep- resentations, 2024

    2024

  63. [64]

    Temporal query network for efficient multivariate time series forecasting,

    S. Lin, H. Chen, H. Wu, C. Qiu, and W. Lin, “Temporal query network for efficient multivariate time series forecasting,” inProceedings of the 42nd International Conference on Machine Learning, 2025, pp. 37 797– 37 814

    2025

  64. [65]

    Mgru: A multi-scale gated recurrent unit with long-memory structure for mid-to-long term time series forecasting,

    H. Yang, X. Li, Z. Zhang, and L. Li, “Mgru: A multi-scale gated recurrent unit with long-memory structure for mid-to-long term time series forecasting,”International Journal of Forecasting, 2024

    2024

  65. [66]

    xlstm-mixer: Multivariate time series forecasting by mixing via scalar memories,

    M. Kraus, F. Divo, D. S. Dhami, and K. Kersting, “xlstm-mixer: Multivariate time series forecasting by mixing via scalar memories,” in The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  66. [67]

    Memcast: Memory-driven time series forecasting with experience- conditioned reasoning,

    X. Tao, M. Cheng, Z. Guo, S. Yu, Y . Liu, Q. Liu, and S. Wang, “Memcast: Memory-driven time series forecasting with experience- conditioned reasoning,”arXiv preprint arXiv:2602.03164, 2026

    Pith/arXiv arXiv 2026

  67. [68]

    Merlin: Multi-view representation learning for robust multivari- ate time series forecasting with unfixed missing rates,

    C. Yu, F. Wang, C. Yang, Z. Shao, T. Sun, T. Qian, W. Wei, Z. An, and Y . Xu, “Merlin: Multi-view representation learning for robust multivari- ate time series forecasting with unfixed missing rates,” inProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2025, pp. 3633–3644

    2025

  68. [69]

    IFS Documentation CY49R1 – Part V: Ensemble Prediction System,

    European Centre for Medium-Range Weather Forecasts, “IFS Documentation CY49R1 – Part V: Ensemble Prediction System,” https://www.ecmwf.int/en/elibrary/ 81373-ifs-documentation-cy49r1-part-v-ensemble-prediction-system, 2024

    2024

  69. [70]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” in International Conference on Learning Representations, 2023

    2023

  70. [71]

    Chronos-2: A local-global framework for foundation models of time series,

    A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapuram, S. Arango, S. Kapoor, D. C. Maddix et al., “Chronos-2: A local-global framework for foundation models of time series,”arXiv preprint arXiv:2505.19497, 2025

    arXiv 2025

  71. [72]

    Sparsetsf: Modeling long-term time series forecasting with 1k parameters,

    S. Lin, W. Lin, W. Wu, H. Chen, and J. Yang, “Sparsetsf: Modeling long-term time series forecasting with 1k parameters,” inProceedings of the 41st International Conference on Machine Learning, 2024, pp. 30 211–30 226

    2024