Pith. sign in

REVIEW 3 major objections 5 minor 60 references

An inference-time wrapper refines frozen time-series foundation models for drought forecasting, cutting one-month SPEI error by up to 18.9 percent without retraining the model.

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 ↗

A residual-guided, coarse-to-fine inference wrapper consistently improves frozen time-series foundation models for monthly drought-index forecasting, cutting one-month-ahead MSE by up to 18.9%.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Inference-time residual-guided multi-resolution refinement is a genuinely new wrapper that consistently improves frozen TSFMs for drought forecasting, but the key residual-transfer assumption is untested and the theory is conditional; worth serious refereeing. the 3 major comments →

arxiv 2607.17507 v1 pith:EWI255WM submitted 2026-07-20 cs.LG

Residual-Guided Multi-Resolution Refinement of Foundation Models: A Case Study in Drought Forecasting

Wentao Gao , Jiuyong Li , Lin Liu , Thuc Duy Le , Jixue Liu , Yanchang Zhao , Yun Chen This is my paper
classification cs.LG
keywords drought forecastingSPEItime series foundation modelsinference-time adaptationmulti-resolution refinementresidual correctionfrozen backbonecoarse-to-fine forecasting
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.

The reading

The paper tries to establish that a frozen time-series foundation model can be made materially better at regional drought forecasting by wrapping it in an inference-time, residual-guided, coarse-to-fine refinement loop, with no weight updates. RGMR projects the input history into five temporal resolutions, asks the frozen model for a forecast at each, learns lightweight per-level predictors of the model's systematic short-window errors, and at test time applies those predicted residuals with adaptive weights as it moves from coarse to fine. The load-bearing evidence is that this wrapper lowers one-month-ahead SPEI mean squared error by up to 18.9 percent across three South Australian sites when applied to TimesFM, and produces consistent gains across the other two tested backbones, three outside-region sites, temperature, and horizons up to six months. If the paper is right, anyone with access to a frozen model's forward pass can improve regional drought forecasts without retraining. The cost is modest: a few extra forward passes and a lightweight closed-form correction per forecast origin.

Core claim

On the paper's own terms, the central claim is that single-pass inference is the bottleneck, not the foundation model's capacity. RGMR operationalizes the climatologist's workflow—look at broad patterns, isolate scales, diagnose residual error, refine—as a plug-in wrapper: block-average the input to a stride, run the frozen model, learn a Ridge regressor that maps simple features (lagged targets, rolling mean/std, linear trend, residual history) to the errors the short-window proposal makes, then at inference combine the long-window proposal at each stride with the previous refined forecast plus an adaptively gated predicted residual. The reported result is that the wrapper consistently lowe

What carries the argument

The carrying object is the residual recursion by(k) = α(k) ¯y(k) + (1−α(k)) ( by(k−1) + η(k) ω(k−1) ⊙ bR(k−1) ). The residual predictor gφ(k−1) is a closed-form Ridge regressor trained on 12-month-window residuals, and the adaptive weight ω(k−1) is a clipped sigmoid of how far the predicted residual exceeds a validation-selected threshold. The projection P_r = U_r ∘ D_r (block average then repeat) creates the five strides {12,6,3,2,1} months. Theorem 4.1 shows that if residual-prediction noise has bounded second moment, each refinement level multiplies the previous expected squared error by a factor below one plus an additive noise floor; Corollary 4.2 gives a finest-level no-harm envelope.

Load-bearing premise

The correction works only if the errors learned from one-year calibration windows are the same systematic errors the long-window proposals make at test time; if those error patterns diverge, the wrapper can inject bias instead of removing it.

What would settle it

Take a stationary synthetic SPEI-like series with no annual cycle and a single regime; train the residual predictors on a 12-month window and evaluate RGMR on a held-out span dominated by a strong climate-regime shift (e.g., a sharp ENSO phase transition). If the wrapper's MSE is not below the frozen baseline in that regime, the short-to-long-window residual-transfer assumption fails.

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

If this is right

  • A user who cannot update a foundation model's weights can still get materially better regional drought forecasts by adding an inference-time correction layer.
  • The correction is architecture-agnostic in the tested sense: the same wrapper lowers MSE for three different frozen backbones, so it can be layered onto a model chosen for other reasons.
  • Because no weights are touched, the wrapper can be added to operational rolling-origin forecasting pipelines and could be combined with test-time adaptation methods that do update parameters.
  • The gains carry over to temperature forecasting, to three regions outside South Australia, and to horizons of 3 and 6 months, so the mechanism appears to correct general systematic output error rather than a SPEI-specific quirk.
  • Per-level contraction and a finest-level no-harm condition imply that error reduction is stable rather than oscillatory as refinement proceeds, up to a bounded noise floor.

Where Pith is reading between the lines

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

  • Editorial inference: because RGMR only needs the backbone's forward interface, the same recipe should transfer to API-only foundation models where weights are never exposed; that is a direct, testable consequence of treating the wrapper as black-box.
  • Editorial inference: the residual features are handcrafted (lagged values, rolling statistics, linear trend, residual history), so a natural stress test is whether a nonlinear or learned residual predictor preserves the contraction bound or whether the simplicity of Ridge is load-bearing.
  • Editorial inference: the one-annual-cycle calibration window (12 months) assumes one year of data exposes the errors that matter; in regimes with multi-year climate modes, a regime-conditioned or longer calibration window may be needed to keep residual transfer valid.
  • Editorial inference: the paper reports point forecasts of an index; the same coarse-to-fine residual correction could be applied to probabilistic forecasts or ensemble members, and whether contraction holds at the quantile level is an open question.
Share X Bluesky LinkedIn Reddit HN

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

3 major / 5 minor

Summary. The paper proposes RGMR, an inference-time wrapper for frozen time-series foundation models (TSFMs). RGMR projects the input into multiple resolutions, obtains base forecasts from a frozen backbone at each resolution, and refines them coarse-to-fine by adding predicted residuals from lightweight Ridge predictors. The residual predictors are trained offline on short-window proposals (Lshort=12) and applied at inference to long-window proposals. The paper evaluates RGMR on one-month-ahead SPEI forecasting at three South Australian sites and three additional regions, reporting consistent MSE reductions over three TSFM backbones (TimesFM, TimeGPT, TabPFN), with up to 18.9% reduction for TimesFM. A contraction-style theorem and a no-harm corollary are provided, along with ablation and sensitivity studies.

Significance. The empirical results, if statistically reliable and attributable to the residual correction, indicate a useful recipe for improving frozen TSFMs in regional climate forecasting without weight updates. The paper includes reproducible code, a careful rolling-origin protocol, and a clean ablation chain showing that multi-resolution averaging and residual weighting both contribute. The theoretical analysis is correct as a conditional bound but is not sufficient on its own to guarantee the observed gains; the empirical claims carry the paper.

major comments (3)
  1. [§4.3, §4.5, App. D] The residual predictors g(k) are fit on short-window residuals eR(k) = y − e y(k) with Lshort=12, but at inference they are applied to long-window proposals. Theorem 4.1 assumes the prediction error e(k−1) is zero-mean; the paper states (App. B.2) that this 'is a standard simplifying assumption' and absorbs the short-vs-long window mismatch into e. This transfer assumption is never tested. In addition, the feature vector z(k) at inference is built from residual history of final refined predictions by(K) (App. D), while during calibration the residual-history feature is built from short-window proposals, creating a feature distribution shift. The paper should report a direct check of the transfer assumption — e.g., comparing the distributions of training residuals and inference-time residuals on validation, or retraining the residual predictors on long-window residuals and showing that th
  2. [§5.2, App. C] The test sets are small. With T=444 months and a 70/10/20 chronological split, the test span has ~89 months; for H=1 this yields ~89 non-overlapping rolling origins per site. No confidence intervals, standard errors, or significance tests are reported, so the observed differences (e.g., TimesFM MSE 0.391 → 0.318 at Location 1) may be within sampling variability. Please report bootstrap confidence intervals for MSE differences or paired tests over rolling origins. The cross-region results in Table 10 are also reported only for TimesFM, which weakens the 'architecture-agnostic' generalization claim.
  3. [§4.5, Eq. (8), Eq. (14)] Theorem 4.1 is a conditional bound: E||R(k)||² ≤ c E||R(k−1)||² + B_k, with c<1. Since B_k contains terms proportional to the proposal error and the residual-prediction noise, this does not imply that the refined forecast improves on the unrefined one; the noise floor can lie above the baseline error. Corollary 4.2's no-harm condition (9) is also not empirically verified. The paper should state this limitation explicitly and, ideally, estimate the empirical noise floor from validation to show that the contraction is operative in the evaluated regime.
minor comments (5)
  1. [§4.5] Equation (8) uses B_k before its definition; define it in the main text or explicitly refer to App. B.
  2. [Table 3] Report cumulative MSE per refinement level, since the per-level sample-wise improvement ratio of 48.5% at the finest level (k=5) does not convey whether the final step reduces aggregate MSE.
  3. [App. C] The practice of retaining the last daily SPEI-30 value in each calendar month changes the temporal aggregation; please justify this or discuss its effect.
  4. [Figure 2] The residual-history feature (block (iv) of z(k)) is not depicted; clarify how it is generated at training vs. inference.
  5. [Notation] The symbol φ appears in g(k)_φ but is not defined in the notation table.

Circularity Check

0 steps flagged

No circularity found: RGMR's central improvement is supported by a held-out rolling-origin evaluation, and the contraction theorem is conditional on an explicit zero-mean residual-error assumption rather than an equation-level reduction to its inputs.

full rationale

The central empirical claim—that RGMR lowers test-set MSE of frozen TSFMs—is not forced by construction. Residual predictors g(k) are fitted offline on training/validation residuals eR(k)=y−ey(k) from short-window proposals (Eq. 2, Alg. 1), then frozen and applied to long-window proposals at inference; test targets are never used for fitting, threshold selection, or the residual-history feature (Sec. 4.3, App. C, App. D). Tables 1, 2, 10, and 11 compare against held-out test periods, so the improvement is an independent empirical result rather than a renamed fit. The theoretical part is explicitly conditional: Theorem 4.1 assumes bR(k−1)=R(k−1)+e(k−1) with E[e]=0 and bounded second moment, and the paper states that e absorbs both Ridge prediction noise and the short-vs-long-window distribution shift (Sec. 4.5, App. B.1). That is an unverified modeling assumption, not a circular definition; the appendix even labels it a 'standard simplifying assumption' (App. B.2). No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggling citation is present. The main weakness—transfer of short-window residual behavior to long-window inference—is a correctness/robustness risk, not a circularity. Therefore no specific circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

8 free parameters · 5 axioms · 0 invented entities

The method uses no new physical entities. The central free parameters are design choices and validation-selected hyperparameters: the resolution scale, mixing constants, step size, weighting constants, δ(k), λ(k), feature lags, and short-window length. The key domain assumption is the transfer of short-window residuals to long-window inference, which the theory absorbs into an error term rather than testing directly.

free parameters (8)
  • Resolution scale R = {12,6,3,2,1} months
    Chosen by hand to match climatological scales and fixed across all sites; sensitivity analysis tests nearby variants but there is no principled selection mechanism.
  • Mixing schedule constants = 0.3, 0.5 in α(k)=0.3+0.5(1-r_k/max(R))
    Hand-set coarse-to-fine ramp; the paper does not tune these constants.
  • Step size η(k) = 1.0
    Fixed in all reported experiments; theory only constrains η∈(0,2]. Not independently ablated.
  • Weighting constants γ, ε_min = γ=3.0, ε_min=1e-3
    Fixed defaults for sigmoid sharpness and weight floor.
  • Validation-selected residual threshold δ(k) = quantile on grid {0.60,...,0.90}
    Selected per level on validation; controls when corrections are applied.
  • Ridge penalty λ(k) = grid {1e-4,...,1e2}
    Selected by time-ordered validation on the log grid.
  • Feature hyperparameters p_lag, q = p_lag=6, q=12
    Lags and trend-window length are fixed defaults, not tuned on test data.
  • Short window L_short = 12 months
    Chosen as one annual cycle to expose short-window errors; fixed across all sites.
axioms (5)
  • domain assumption Short-window residual patterns transfer to long-window inference proposals; any distribution shift is zero-mean bounded noise.
    RGMR trains residual predictors on Lshort=12 projections (§4.3 Eq. 2) but applies them to full-context proposals at inference; Theorem 4.1 folds the transfer problem into e(k−1).
  • domain assumption Frozen TSFM fθ can ingest the multivariate projected inputs and repeated forward passes at different strides improve or at least do not destroy forecast information.
    The method assumes a standard forward interface for all backbones and that projection operators preserve the information needed for forecasting.
  • standard math Residual-prediction error e(k−1) and proposal error u(k) have finite second moments; e(k−1) is zero-mean.
    These probability assumptions are stated in Theorem 4.1 and are not empirically validated in the paper.
  • ad hoc to paper The fixed resolution scale R matches the dominant SPEI variability scales; nearby variants in the sensitivity table cover the relevant design space.
    The scale is chosen from climatological practice; Table 5 probes only nearby variants and no automatic scale-selection criterion is provided.
  • domain assumption Public data sources (SPEI-GD, NCEP-NCAR Reanalysis 1, climate indices) are accurate and the split protocol prevents leakage.
    The empirical claim depends on data quality and on no future information entering preprocessing or residual features; the paper asserts this but cannot prove it.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Residual-Guided Multi-Resolution Refinement of Foundation Models: A Case Study in Drought Forecasting." pith.science (2026). https://pith.science/paper/EWI255WM

@misc{pith2026260717507,
  author       = {Pith},
  title        = {Pith review of: Residual-Guided Multi-Resolution Refinement of Foundation Models: A Case Study in Drought Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWI255WM}},
  note         = {Machine review of arXiv:2607.17507}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Regional climate prediction presents unique challenges for time series foundation models, which typically process temporal patterns through single-pass inference. Expert climatologists, in contrast, employ multi-scale temporal analysis and iterative refinement based on systematic error diagnosis. We present RGMR (Residual-Guided Multi-Resolution Refinement), an inference-time framework that adapts pre-trained foundation models to perform structured coarse-to-fine refinement for climate forecasting without updating backbone parameters. Applied to drought forecasting using the Standardized Precipitation Evapotranspiration Index (SPEI), RGMR is architecture-agnostic across the three TSFM backbones evaluated per site (TimesFM, TimeGPT, TabPFN) and consistently lowers test-set MSE on three South Australian sites and three additional regions outside South Australia. Applied to TimesFM, the wrapper reduces one-month-ahead SPEI MSE by up to 18.9\% across the three South Australian sites (mean reduction $\approx$18.7\%). Overall, RGMR provides a practical route for deploying frozen TSFMs in regional climate forecasting workflows.

Figures

Figures reproduced from arXiv: 2607.17507 by Jiuyong Li, Jixue Liu, Lin Liu, Thuc Duy Le, Wentao Gao, Yanchang Zhao, Yun Chen.

Figure 1
Figure 1. Figure 1: Comparison of temporal processing approaches in cli￾mate forecasting. Multi-scale decomposition analysis (middle row) versus single forward pass processing (bottom row) for complex climate pattern prediction. Madaan et al., 2023), while operational climate systems re￾quire efficient inference that avoids gradient updates to the foundation backbone (Bauer et al., 2015). To address this setting, we propose R… view at source ↗
Figure 2
Figure 2. Figure 2: Two-stage workflow of RGMR. Stage I: Offline calibration learns level-wise residual correction modules from short-window projected inputs. At each resolution level, the frozen foundation model produces a short-window base forecast, and the observed forecast error is used to train a lightweight residual correction module. Feature vectors and training residuals are used to fit the residual predictors and sel… view at source ↗
Figure 3
Figure 3. Figure 3: RGMR forecast quality and drought-event detection at a representative South Australia site. Left: monthly SPEI timeline with ground truth (black) and RGMR prediction (red); shaded bands indicate SPEI categories ranging from drought conditions at the bottom of the panel to wet conditions at the top. Middle: predicted vs. actual SPEI with y=x (dashed); Pearson r=0.83; point color encodes SPEI. Right: confusi… view at source ↗
Figure 4
Figure 4. Figure 4: South Australia study-area context. The rainfall-decile map is used to illustrate regional rainfall heterogeneity and the 2024 drought context. 5. Experiments We evaluate whether RGMR improves frozen foundation models for monthly-to-seasonal SPEI forecasting without parameter updates. The study targets overall accuracy against strong baselines, the effect of multi-resolution and residual refinement, the ro… view at source ↗
Figure 5
Figure 5. Figure 5: One–month–ahead SPEI at Location (−26.125, 129.125) (appendix view). Each panel overlays ground truth (blue) and model prediction (orange) under the same leakage–safe protocol as the main paper (standardization fit on train; identical splits; H=1). TimesFM (foundation-scale) shows the tightest phase alignment on mild–moderate variability; linear/frequency families (e.g., DLinear, FiLM) reduce variance but … 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

60 extracted references · 1 canonical work pages

  1. [1]

    Vicente-Serrano, Sergio M. and Dom. Long-Term Variability and Trends in Meteorological Droughts in. International Journal of Climatology , volume=. 2020 , doi=

    2020

  2. [2]

    2025 , howpublished =

    2025

  3. [3]

    Current Climate Change Reports , volume=

    Climate Change and Drought: a Perspective on Drought Indices , author=. Current Climate Change Reports , volume=. 2018 , publisher=

    2018

  4. [4]

    Sustainability , volume=

    Drought Prediction: A Comprehensive Review of Different Drought Prediction Models and Adopted Technologies , author=. Sustainability , volume=. 2023 , issn=

    2023

  5. [5]

    2025 , doi=

    Tan, Xuwei and Zhao, Qian and Liu, Yanlan and Zhang, Xueru , booktitle=. 2025 , doi=

    2025

  6. [6]

    Neural Computation , volume=

    Long short-term memory , author=. Neural Computation , volume=. 1997 , publisher=

    1997

  7. [7]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=

    Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=. 2021 , doi=

    2021

  8. [8]

    Science , volume=

    Learning Skillful Medium-Range Global Weather Forecasting , author=. Science , volume=. 2023 , doi=

    2023

  9. [9]

    Accurate Medium-Range Global Weather Forecasting with

    Bi, Kaifeng and Xie, Lingxi and Zhang, Hengheng and Chen, Xin and Gu, Xiaotao and Tian, Qi , journal=. Accurate Medium-Range Global Weather Forecasting with. 2023 , publisher=

    2023

  10. [10]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Scaleformer: Iterative Multi-scale Refining Transformers for Time Series Forecasting , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  11. [11]

    2024 , eprint=

    Minusformer: Improving Time Series Forecasting by Progressively Learning Residuals , author=. 2024 , eprint=

    2024

  12. [12]

    2310.08278 , archivePrefix=

    Rasul, Kashif and Ashok, Arjun and Williams, Andrew Robert and others , year=. 2310.08278 , archivePrefix=

    Pith/arXiv arXiv

  13. [13]

    Transactions on Machine Learning Research , year=

    Chronos: Learning the Language of Time Series , author=. Transactions on Machine Learning Research , year=

  14. [14]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , series=

    Timer: Generative Pre-trained Transformers Are Large Time Series Models , author=. Proceedings of the 41st International Conference on Machine Learning (ICML) , series=. 2024 , url=

    2024

  15. [15]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Attention Is All You Need , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  16. [16]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , series=

    A Decoder-only Foundation Model for Time-series Forecasting , author=. Proceedings of the 41st International Conference on Machine Learning (ICML) , series=. 2024 , url=

    2024

  17. [17]

    The Emergence of Numerical Weather Prediction:

    Lynch, Peter , year=. The Emergence of Numerical Weather Prediction:

  18. [18]

    Nature , volume=

    The Quiet Revolution of Numerical Weather Prediction , author=. Nature , volume=. 2015 , publisher=

    2015

  19. [19]

    Nature Climate Change , volume=

    Global Warming and Changes in Drought , author=. Nature Climate Change , volume=. 2014 , doi=

    2014

  20. [20]

    2003 , publisher=

    Forecast Verification: A Practitioner's Guide in Atmospheric Science , author=. 2003 , publisher=

    2003

  21. [21]

    Palmer, T. N. , journal=. Nonlinear Dynamics and Climate Change:. 1998 , doi=

    1998

  22. [22]

    Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

    Can deep learning beat numerical weather prediction? , author=. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=. 2021 , doi=

    2021

  23. [23]

    International Journal of Forecasting , volume=

    Temporal Fusion Transformers for Interpretable Multi-horizon Time Series Forecasting , author=. International Journal of Forecasting , volume=. 2021 , doi=

    2021

  24. [24]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  25. [25]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Large Language Models Are Zero-Shot Reasoners , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  26. [26]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  27. [27]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  28. [28]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  29. [29]

    and Beguer

    Vicente-Serrano, Sergio M. and Beguer. A Multiscalar Drought Index Sensitive to Global Warming:. Journal of Climate , volume=. 2010 , doi=

    2010

  30. [30]

    and Carpov, Dmitri and Chapados, Nicolas and Bengio, Yoshua , booktitle=

    Oreshkin, Boris N. and Carpov, Dmitri and Chapados, Nicolas and Bengio, Yoshua , booktitle=. 2020 , url=

    2020

  31. [31]

    Proceedings of the 37th International Conference on Machine Learning (ICML) , series=

    Test-time Training with Self-supervision for Generalization under Distribution Shifts , author=. Proceedings of the 37th International Conference on Machine Learning (ICML) , series=. 2020 , url=

    2020

  32. [32]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Language Models are Few-Shot Learners , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  33. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=

    Meta-Learning Framework with Applications to Zero-Shot Time-Series Forecasting , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=. 2021 , doi=

    2021

  34. [34]

    2310.03589 , archivePrefix=

    Garza, Azul and Challu, Cristian and Mergenthaler-Canseco, Max , year=. 2310.03589 , archivePrefix=

    Pith/arXiv arXiv

  35. [35]

    2003 , isbn=

    Fixed Point Theory , author=. 2003 , isbn=

    2003

  36. [36]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Deep Equilibrium Models , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , volume=

  37. [37]

    Gao, Zhihan and Shi, Xingjian and Wang, Hao and others , booktitle=

  38. [38]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , year=

    Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , year=

  39. [39]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  40. [40]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=

    Are Transformers Effective for Time Series Forecasting? , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=. 2023 , doi=

    2023

  41. [41]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  42. [42]

    TimesNet: Temporal 2

    Wu, Haixu and Hu, Tengge and Liu, Yong and Zhou, Hang and Wang, Jianmin and Long, Mingsheng , booktitle=. TimesNet: Temporal 2. 2023 , url=

    2023

  43. [43]

    and Pfister, Tomas , journal=

    Chen, Si-An and Li, Chun-Liang and Yoder, Nate and Arik, Sercan O. and Pfister, Tomas , journal=

  44. [44]

    2024 , url=

    Liu, Yong and Hu, Tengge and Zhang, Haoran and Wu, Haixu and Wang, Shiyu and Ma, Lintao and Long, Mingsheng , booktitle=. 2024 , url=

    2024

  45. [45]

    Zhou, Tian and Ma, Ziqing and Wang, Xue and Wen, Qingsong and Sun, Liang and Yao, Tao and Yin, Wotao and Jin, Rong , booktitle=

  46. [46]

    Proceedings of the International Conference on Learning Representations (ICLR) , year=

    Pyraformer: Low-Complexity Pyramidal Attention for Long-Range Time Series Modeling and Forecasting , author=. Proceedings of the International Conference on Learning Representations (ICLR) , year=

  47. [47]

    2023 , url=

    Wang, Huiqiang and Peng, Jian and Huang, Feihu and Wang, Jince and Chen, Junhui and Xiao, Yifei , booktitle=. 2023 , url=

    2023

  48. [48]

    Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , year=

    Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting , author=. Proceedings of Advances in Neural Information Processing Systems (NeurIPS) , year=

  49. [49]

    and Zhou, Jun , booktitle=

    Wang, Shiyu and Wu, Haixu and Shi, Xiaoming and Hu, Tengge and Luo, Huakun and Ma, Lintao and Zhang, James Y. and Zhou, Jun , booktitle=. 2024 , url=

    2024

  50. [50]

    2024 , howpublished =

    Time-Series-Library: A Library for Advanced Deep Time Series Models , author =. 2024 , howpublished =

    2024

  51. [51]

    Kalnay, Eugenia and Kanamitsu, Masao and Kistler, Robert and others , journal=. The. 1996 , doi=

    1996

  52. [52]

    , journal=

    Trenberth, Kevin E. , journal=. The Definition of. 1997 , doi=

    1997

  53. [53]

    Saji, N. H. and Goswami, B. N. and Vinayachandran, P. N. and Yamagata, T. , journal=. A Dipole Mode in the Tropical. 1999 , doi=

    1999

  54. [54]

    , journal=

    Marshall, Gareth J. , journal=. Trends in the. 2003 , doi=

    2003

  55. [55]

    Technometrics , volume=

    Ridge Regression: Biased Estimation for Nonorthogonal Problems , author=. Technometrics , volume=. 1970 , doi=

    1970

  56. [56]

    Computational Statistics & Data Analysis , volume=

    A note on the validity of cross-validation for evaluating autoregressive time series prediction , author=. Computational Statistics & Data Analysis , volume=. 2018 , doi=

    2018

  57. [57]

    Science , volume=

    Global Increase in the Occurrence and Impact of Multiyear Droughts , author=. Science , volume=. 2025 , doi=

    2025

  58. [58]

    and Stockdale, Timothy N

    Johnson, Stephanie J. and Stockdale, Timothy N. and Ferranti, Laura and Balmaseda, Magdalena A. and Molteni, Franco and Magnusson, Linus and Tietsche, Steffen and Decremer, Damien and Weisheimer, Antje and Balsamo, Gianpaolo and Keeley, Sarah P. E. and Mogensen, Kristian and Zuo, Hao and Monge-Sanz, Beatriz M. , journal=. 2019 , doi=

    2019

  59. [59]

    and Min, Dughong and Infanti, Johnna M

    Kirtman, Ben P. and Min, Dughong and Infanti, Johnna M. and Kinter III, James L. and Paolino, Daniel A. and Zhang, Qin and van den Dool, Huug and Saha, Suranjana and Mendez, Malaqu. The. Bulletin of the American Meteorological Society , volume=. 2014 , doi=

    2014

  60. [60]

    Scientific Data , year =

    Liu, Xuebang and Yu, Shuying and Yang, Zhiwei and Dong, Jianquan and Peng, Jian , title =. Scientific Data , year =. doi:10.1038/s41597-024-03047-z , url =

    work page doi:10.1038/s41597-024-03047-z

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.