Pith. sign in

REVIEW 3 major objections 3 minor 38 references

Inference-time wrappers adapt frozen time-series foundation models to regional drought forecasting, cutting MSE by up to 26% without touching backbone weights.

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

2026-08-01 17:45 UTC pith:6IMOSMZO

load-bearing objection The wrapper idea is sensible and the empirical claim is plausible, but the evaluation protocol is under-specified in exactly the place that matters, so the headline gains are not yet trustworthy. the 3 major comments →

arxiv 2607.17511 v1 pith:6IMOSMZO submitted 2026-07-20 cs.LG

Lightweight Wrappers for Adapting Time Series Foundation Models to Regional 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 time series foundation modelsinference-time adaptationdrought forecastingSPEImulti-resolution residual correctionmoving block bootstrapbagging ensemblingblack-box model adaptation
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 claims that frozen time-series foundation models, which cannot be fine-tuned because their weights are proprietary or computing budgets are tight, can still be adapted to regional drought forecasting through two lightweight wrappers operating entirely at inference time. The first wrapper, SMR2, decomposes the input into multi-resolution temporal views, learns stride-specific residual corrections via ridge regression, and ensembles the corrected forecasts. The second, MBB, resamples the backbone's validation residuals in temporally coherent blocks, rebuilds perturbed histories, and averages the resulting forecasts. Across three South Australian sites and three foundation model backbones, the wrappers improve every backbone, with up to 26% MSE reduction; cross-region tests on three other climate regimes show 16–41% reductions. The authors frame the approach as bagging-style: diverse input views, a shared frozen backbone, and aggregation.

Core claim

The paper's central claim is that systematic regional bias left in the residuals of a frozen time-series foundation model is learnable and correctable without any parameter update. SMR2 operationalizes this by building per-stride mean/std views of the context window, asking the frozen backbone to forecast each view, fitting a lightweight ridge corrector on the stride-specific residuals, and adaptively combining the corrected forecasts using validation R² or inverse-MSE weights. MBB operationalizes variance stabilization by block-bootstrapping the backbone's validation residuals to create many perturbed target histories, forecasting each with the same frozen backbone, and averaging. The autho

What carries the argument

SMR2 (Stationarity aware multi-resolution Residual): decomposes the context window into stride-s segments, computing per-segment mean and standard deviation to form a length-L multi-resolution view; a ridge corrector (λ=0.1, q=6 recent residuals) predicts the residual for each stride; adaptive weights are selected by validation R² (falling back to inverse MSE) and fused into a final forecast. MBB (Moving Block Bootstrap): constructs overlapping blocks of length ⌊n^{1/3}⌋ from the backbone's validation residuals, samples blocks with replacement to create K=100 perturbed target histories, and averages the frozen backbone forecasts on each perturbation. Both wrappers are bagging-style ensembles

Load-bearing premise

The reported gains assume that wrapper and baseline forecasts are scored over the same time steps; if wrappers drop early or high-variance test points that baselines keep, the headline MSE reductions are not a controlled comparison.

What would settle it

Recompute every metric on exactly the same test timestamps for all methods; if the implied target variance for the same location differs across rows of Table 1 (as the reported MSE and R² suggest it does), or if gains shrink markedly under a common scoring window, the central claim is not a controlled comparison.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Regional agencies can improve proprietary foundation models without needing weight access or fine-tuning compute, making adaptation feasible under realistic deployment constraints.
  • The wrapper logic is backbone-agnostic: it should transfer to any frozen time-series model that can consume a context window, with gains depending on the structure of that model's residuals.
  • Cross-region results suggest the approach does not overfit to South Australia; similar gains appear in tropical, arid, and Mediterranean climates.
  • Both wrappers add only a small number of extra forward passes (|S| strides for SMR2, K bootstrap samples for MBB), keeping inference latency near that of the frozen model.

Where Pith is reading between the lines

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

  • The theoretical risk bound is Jensen-based and covariance-agnostic, so the paper does not establish that the R²-based weighting is optimal; a head-to-head against uniform stride weights across more sites would test whether the adaptive weighting actually drives the improvement.
  • The reported gains may come mostly from bias correction of systematic regional residuals rather than from variance reduction; decomposing the MSE drop into bias and variance components would reveal which wrapper mechanism contributes more.
  • The evaluation's 'scored time steps' wording (Appendix D) leaves open the possibility that wrappers and baselines were evaluated over different test windows; a re-analysis on a common timestamp set would confirm whether the 26% reduction is a fully controlled comparison.

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 / 3 minor

Summary. The paper proposes two inference-time wrappers, SMR2 and MBB, for adapting frozen time-series foundation models (TabPFN, TimeGPT, TimesFM) to one-month-ahead regional SPEI-30 forecasting. SMR2 builds multi-resolution stride views, learns stride-specific ridge residual correctors, and adaptively ensembles corrected forecasts; MBB block-bootstraps validation residuals to create perturbed target histories and averages the corresponding frozen-backbone forecasts. The empirical claim is that these wrappers improve MSE over the frozen backbones at three South Australian sites and in three additional regions, with up to 26% MSE reduction.

Significance. If the central empirical claim is valid, the framework is practically valuable: it adapts black-box or proprietary TSFMs without weight access, fine-tuning, or large compute, and the paper includes reproducible details (code repository, algorithms, ablations) and cross-region experiments. The idea of bagging-style inference-time wrappers is simple and appealing. However, the reported evaluation is not currently trustworthy as a controlled comparison: the caption of Table 1 and Appendix D do not define the 'scored time steps', and the internal R2/MSE inconsistencies in Table 1 suggest that wrapper and baseline metrics may be computed over different test points. This directly affects the headline quantitative claim.

major comments (3)
  1. [Table 1 / Appendix D] The caption says 'Per method metrics are computed over their scored time steps (see Appendix D)', but Appendix D never specifies which time steps are scored. For TimesFM at Location 1, the implied test variance MSE/(1-R2) equals 0.897 for the frozen baseline, 0.825 for +SMR2, and 0.981 for +MBB. For the same target series and the same test period, this quantity must be identical across rows unless the R2 denominator differs or the scored windows differ. The latter appears likely. Please report all metrics on the identical test window and define the R2 denominator explicitly. Until then, the claimed 'up to 26% MSE reduction' is not a controlled comparison.
  2. [Method, SMR2 Eqs. (5)-(8), Algorithms 1-2] Equation (7) uses the q most recent stride residuals as features for the corrector, but the paper does not state how these residuals are obtained during the test period: are they recomputed at each test origin using the full context and the observed y_t, or taken from the calibration set only? This affects the first few test forecasts and therefore the 'scored time steps' used in Table 1. Please specify the exact residual-update procedure and the index of the first test point scored by each method.
  3. [Generalization / Table 3] The cross-region claim that 'SMR2 reduces MSE in every region, by roughly 16 to 41%' rests on a single test split with no uncertainty quantification. Given the high autocorrelation of monthly SPEI and the small number of test points per region, standard errors or confidence intervals (e.g., across origins or sub-periods) are needed to support the 'every region' generalization claim.
minor comments (3)
  1. [Appendix D] Please define the evaluation metrics explicitly, especially R2 (denominator: test variance, training variance, or baseline variance) and what constitutes a 'scored time step' for each method.
  2. [MBB, Eq. (13)] The perturbed history is formed by adding resampled residuals to the model's own rolling one-step predictions, not to the observed history. This is a nontrivial design choice; please explain why it is preferable to perturbing the observed history and how it preserves the temporal dependence that motivates block resampling.
  3. [Figure 5] The figure caption lists wrappers and baselines but does not draw attention to specific features of the wrapper forecasts; please add a sentence describing what the reader should observe.

Circularity Check

0 steps flagged

No circular derivation: SMR2 and MBB are standard supervised/resampling wrappers on a frozen backbone, and the theoretical bounds are explicit assumption-light inequalities; the headline empirical claim is a held-out comparison.

full rationale

The load-bearing claim is empirical: SMR2/MBB improve frozen TSFMs on held-out SPEI test periods. The wrapper mechanisms do not reduce to their inputs by construction. SMR2 fits a ridge residual corrector on calibration residuals R^(s) (Eq. 5, T_train) and applies it unchanged at inference (Eq. 7), which is ordinary supervised learning rather than circularity. MBB resamples validation residuals (T_val) to form perturbed histories and averages frozen-backbone forecasts (Eqs. 13-15); no test target or fitted constant is embedded in the perturbation scheme. The theoretical results are explicit and assumption-light: Proposition 1 is Jensen's inequality plus a bias-variance identity (Appendix A), and Eq. 19 is the standard variance of an average; neither asserts the empirical gain, so they cannot smuggle the conclusion in. There are no author self-citations used as load-bearing support; the bootstrap consistency results are cited to Künsch, Lahiri, Politis-Romano, and the normalization idea is attributed to an external NeurIPS paper. One non-circular caveat belongs in correctness risk, not circularity: Table 1's caption says metrics are computed over 'their scored time steps' but Appendix D does not define which steps are scored, and the MSE/R2 pairs for TimesFM/TimeGPT imply different implied test variances across rows for the same location; if the scored windows differ, the reported gains would not be a controlled comparison. That is an evaluation-protocol concern, not a derivation that reduces to its inputs.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 1 invented entities

The central claim depends on a handful of hand-chosen hyperparameters (strides, ridge penalty, lag q, K) and on the domain assumption that frozen backbones produce useful forecasts from modified input views. The only clearly invented entity is the latent residual decomposition of Assumption 1, which is motivation rather than load-bearing evidence.

free parameters (7)
  • temporal stride set S = S={1,2,4,6,8}
    Selected to align with dominant spectral components of monthly SPEI; value controls the diversity of multi-resolution views and hence the ensemble behavior.
  • ridge penalty lambda_ridge = 0.1
    Fixed hyperparameter of the SMR2 residual corrector; not derived from first principles.
  • residual lag q = 6
    Number of recent residuals fed to the ridge corrector; determines how much autoregressive structure can be exploited in the correction.
  • stabilizer epsilon = 1e-6
    Added to inverse-MSE fallback weights to avoid division by zero.
  • MBB bootstrap sample count K = 100
    Number of bootstrap forecasts averaged in MBB; larger K reduces variance at higher compute cost.
  • context lengths L and L_train = not reported
    Full and reduced context lengths are defined but no numeric values are given; they determine how many calibration residuals are available and which test steps are scored.
  • SMR2 ridge corrector coefficients = learned during calibration
    The residual correction depends on coefficients fit on calibration residuals; the method's gains hinge on their generalization to test data.
axioms (4)
  • domain assumption The frozen backbone is a fixed mapping f_base_theta and its forecasts from modified input views are meaningful.
    The wrapper design assumes the backbone cannot be modified and that feeding it multi-resolution mean/std views or perturbed histories produces useful forecasts. See Eqs. (1), (6), and (14).
  • ad hoc to paper The backbone residual decomposes into stride-specific predictable components plus noise (Assumption 1, Eq. 16).
    Postulated latent components r_t^(s) are not observed and have no independent falsifiable handle; the paper itself states these are not observed residual sequences.
  • standard math MBB consistency requires the residual series to be stationary and mixing.
    The paper invokes Künsch, Lahiri, and Politis-Romano for block bootstrap consistency; if residuals are nonstationary under regime shifts, the variance argument may fail. The authors acknowledge this in the limitations section.
  • domain assumption Baseline and wrapper predictions are scored over comparable time steps.
    Table 1 says metrics are computed over each method's scored time steps; if these windows differ, the reported comparisons are not controlled. Internal R2/MSE inconsistencies at Location 1 suggest this assumption may be violated.
invented entities (1)
  • Latent stride-specific residual components r_t^(s) no independent evidence
    purpose: Theoretical decomposition of the backbone residual into resolution-specific predictable parts, used to motivate SMR2's multi-resolution correction.
    These components are never observed, have no independent measurement, and are not needed for the empirical evaluation; they are an interpretative overlay rather than a tested scientific claim.

pith-pipeline@v1.3.0-alltime-deepseek · 16976 in / 17947 out tokens · 146492 ms · 2026-08-01T17:45:07.865688+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Lightweight Wrappers for Adapting Time Series Foundation Models to Regional Drought Forecasting." pith.science (2026). https://pith.science/paper/6IMOSMZO

@misc{pith2026260717511,
  author       = {Pith},
  title        = {Pith review of: Lightweight Wrappers for Adapting Time Series Foundation Models to Regional Drought Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IMOSMZO}},
  note         = {Machine review of arXiv:2607.17511}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large \emph{Time Series Foundation Models} (TSFMs) demonstrate strong zero-shot forecasting capabilities across diverse domains. However, their application to regional climate forecasting faces practical challenges: model weights are often proprietary, local training records are limited, and computational budgets are constrained, making traditional fine-tuning approaches infeasible. To address these constraints, we introduce a lightweight, black-box adaptation framework (requiring no access to backbone parameters and no backbone fine-tuning) that enhances frozen TSFMs at inference time through two plug-and-play wrappers: \textbf{SMR\textsuperscript{2}} (Stationarity aware multi-resolution Residual), which decomposes the input into multi-resolution temporal views, learns stride specific residual corrections that capture regional dynamics, then adaptively ensembles them into a single forecast, and \textbf{MBB} (Moving Block Bootstrap), which preserves temporal dependencies through block resampling and ensembles over temporally coherent residual perturbations to stabilize the point forecast. Both wrappers instantiate the same bagging style principle: they build diverse views of the input or its residuals, forecast each with the same frozen backbone, and aggregate, so all adaptation comes from inference time ensembling rather than any weight update. Evaluated on one month ahead Standardized Precipitation Evapotranspiration Index (SPEI) prediction across multiple sites in South Australia, our framework consistently improves forecasting performance across several backbone models, demonstrating up to 26\% mean squared error (MSE) reduction over the corresponding frozen backbone while enabling practical deployment in resource constrained regional forecasting systems.

Figures

Figures reproduced from arXiv: 2607.17511 by Jiuyong Li, Jixue Liu, Lin Liu, Thuc Duy Le, Wentao Gao, Yanchang Zhao, Yun Chen.

Figure 1
Figure 1. Figure 1: Regional forecasting limitations across three model [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed inference time wrappers for frozen TSFMs. SMR [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Block bootstrap ensembling motivation. (a) Input [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Predicted and observed SPEI time series for one month ahead forecasting at Location 1. Panels (a) and (e) show our [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Study Area in South Australia Experimental Setup Dataset. We evaluate on monthly SPEI-30 prediction across 3 South Australian locations over 1982–2018. The target is the SPEI-30 vari￾able from the global daily SPEI-GD database [19] (0.25◦ spatial resolution, natively spanning 1982–2021), converted to a monthly cadence and restricted to 1982–2018. The dataset comprises 39 variables: the target SPEI plus 𝑑 =… 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

38 extracted references · 4 linked inside Pith

  1. [1]

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271(2018)

    Pith/arXiv arXiv 2018

  2. [2]

    Santiago Beguería, Sergio M Vicente-Serrano, Fergus Reig, et al. 2014. Standard- ized precipitation evapotranspiration index (SPEI) revisited: parameter fitting, evapotranspiration models, tools, datasets and drought monitoring.International Journal of Climatology34, 10 (2014), 3001–3023

    2014

  3. [3]

    Mariana Belgiu and Lucian Drăguţ. 2016. Random forest in remote sensing: A review of applications and future directions.ISPRS Journal of Photogrammetry and Remote Sensing114 (2016), 24–31

    2016

  4. [4]

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, et al. 2023. Accurate medium-range global weather forecasting with 3D neural networks.Nature619 (2023), 533–538

    2023

  5. [5]

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, et al. 2021. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258(2021)

    Pith/arXiv arXiv 2021

  6. [6]

    George E. P. Box, Gwilym M. Jenkins, Gregory C. Reinsel, et al. 2015.Time Series Analysis: Forecasting and Control(fifth ed.). John Wiley & Sons, Hoboken, NJ

    2015

  7. [7]

    Si-An Chen, Chun-Liang Li, Nate Yoder, et al . 2023. TSMixer: An All-MLP Architecture for Time Series Forecasting.Transactions on Machine Learning Research(2023)

    2023

  8. [8]

    Abhimanyu Das, Weihao Kong, Rajat Sen, et al. 2024. A decoder-only foundation model for time-series forecasting. InProceedings of the International Conference on Machine Learning (Proceedings of Machine Learning Research, Vol. 235). PMLR, 10148–10167

    2024

  9. [9]

    Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco. 2023. TimeGPT-1. arXiv preprint arXiv:2310.03589(2023)

    Pith/arXiv arXiv 2023

  10. [10]

    Qiang Guo, Lexin Fang, Ren Wang, et al. 2025. Multivariate Time Series Forecast- ing Using Multiscale Recurrent Networks With Scale Attention and Cross-Scale Guidance.IEEE Transactions on Neural Networks and Learning Systems36, 1 (2025), 540–554

    2025

  11. [11]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.Neural Computation9, 8 (1997), 1735–1780

    1997

  12. [12]

    Shi Bin Hoo, Samuel Müller, David Salinas, et al . 2025. From Tables to Time: Extending TabPFN-v2 to Time Series Forecasting.arXiv preprint arXiv:2501.02945 (2025)

    arXiv 2025

  13. [13]

    Eugenia Kalnay, Masao Kanamitsu, Robert Kistler, et al. 1996. The NCEP/NCAR 40-year reanalysis project.Bulletin of the American Meteorological Society77, 3 (1996), 437–471

    1996

  14. [14]

    Philip Kokic, Huidong Jin, and Steven Crimp. 2013. Improved point scale climate projections using a block bootstrap simulation and quantile matching method. Climate Dynamics41, 3 (2013), 853–866

    2013

  15. [15]

    Hans R Künsch. 1989. The jackknife and the bootstrap for general stationary observations.The Annals of Statistics17, 3 (1989), 1217–1241

    1989

  16. [16]

    Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. 2023. FourCastNet: Accelerating Global High-Resolution Weather Forecasting Using Adaptive Fourier Neural Operators. InProceedings of the Plat- form for Advanced Scientific Computing Conference. Associa...

    arXiv 2023

  17. [17]

    Soumendra Nath Lahiri. 1993. On the moving block bootstrap under long range dependence.Statistics & Probability Letters18, 5 (1993), 405–413

    1993

  18. [18]

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, et al . 2023. Learning skillful medium-range global weather forecasting.Science382, 6677 (2023), 1416– 1421

    2023

  19. [19]

    Xuebang Liu, Shuying Yu, Zhiwei Yang, et al. 2024. The first global multi-timescale daily SPEI dataset from 1982 to 2021.Scientific Data11 (2024), 223

    2024

  20. [20]

    Yong Liu, Tengge Hu, Haoran Zhang, et al. 2024. iTransformer: Inverted Trans- formers Are Effective for Time Series Forecasting. InProceedings of the Interna- tional Conference on Learning Representations

    2024

  21. [21]

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting. InAdvances in Neural Information Processing Systems (NeurIPS)

    2022

  22. [22]

    Ashok K Mishra and Vijay P Singh. 2010. A review of drought concepts.Journal of Hydrology391, 1–2 (2010), 202–216

    2010

  23. [23]

    Nguyen, Phanwadee Sinthong, et al

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, et al. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. InProceedings of the International Conference on Learning Representations

    2023

  24. [24]

    Boris N Oreshkin, Dmitri Carpov, Nicolas Chapados, et al . 2020. N-BEATS: Neural Basis Expansion Analysis for Interpretable Time Series Forecasting. In Proceedings of the International Conference on Learning Representations

    2020

  25. [25]

    2000.Wavelet Methods for Time Series Analysis

    Donald B Percival and Andrew T Walden. 2000.Wavelet Methods for Time Series Analysis. Cambridge University Press, Cambridge, UK

    2000

  26. [26]

    Dimitris N Politis and Joseph P Romano. 1994. The stationary bootstrap.J. Amer. Statist. Assoc.89, 428 (1994), 1303–1313

    1994

  27. [27]

    Md Mamunur Rashid and Simon Beecham. 2019. Characterization of meteoro- logical droughts across South Australia.Meteorological Applications26, 4 (2019), 556–568

    2019

  28. [28]

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, et al . 2023. Lag-Llama: Towards Foundation Models for Probabilistic Time Series Forecasting.arXiv preprint arXiv:2310.08278(2023)

    Pith/arXiv arXiv 2023

  29. [29]

    N. H. Saji, B. N. Goswami, P. N. Vinayachandran, et al. 1999. A dipole mode in the tropical Indian Ocean.Nature401, 6751 (1999), 360–363

    1999

  30. [30]

    Trenberth

    Kevin E. Trenberth. 1997. The definition of El Niño.Bulletin of the American Meteorological Society78, 12 (1997), 2771–2777

    1997

  31. [31]

    Tsinghua Machine Learning Group. 2024. Time-Series-Library: A Library for Advanced Deep Time Series Models. https://github.com/thuml/Time-Series- Library. Accessed: 2026-05-07

    2024

  32. [32]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, et al. 2017. Attention is all you need. InProceedings of Advances in Neural Information Processing Systems, Vol. 30. 5998–6008

    2017

  33. [33]

    Sergio M Vicente-Serrano, Santiago Beguería, and Juan I López-Moreno. 2010. A multiscalar drought index sensitive to global warming: the standardized precipi- tation evapotranspiration index.Journal of Climate23, 7 (2010), 1696–1718

    2010

  34. [34]

    Haixu Wu, Tengge Hu, Yong Liu, et al. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InProceedings of the International Conference on Learning Representations

    2023

  35. [35]

    Haixu Wu, Jiehui Xu, Jianmin Wang, et al. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting. InProceed- ings of Advances in Neural Information Processing Systems, Vol. 34. 22419–22430

    2021

  36. [36]

    Ailing Zeng, Muxi Chen, Lei Zhang, et al . 2023. Are Transformers Effective for Time Series Forecasting?. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 11121–11128

    2023

  37. [37]

    Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. InProceedings of the International Conference on Learning Representations

    2023

  38. [38]

    Tian Zhou, Ziqing Ma, Xue Wang, et al. 2022. FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting. InProceedings of Advances in Neural Information Processing Systems, Vol. 35. 12677–12690. Lightweight Wrappers for Adapting Time Series Foundation Models to Regional Drought Forecasting KDD ’26, August 09–13, 2026, Jeju Island...

    2022