Pith. sign in

REVIEW 5 major objections 5 minor 36 references

TimeCF: A TimeMixer-Based Model with adaptive Convolution and Sharpness-Aware Minimization Frequency Domain Loss for long-term time seris forecasting

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TimeCF uses adaptive convolution and a sharpness-aware frequency-domain loss to achieve the best average forecast error across six datasets with 125K-275K parameters.

desk verdict A clean integration of TimeMixer, FreDF, and SAM that reports a plausible ~2% average gain, but the evaluation protocol is not auditable—no code, no seeds, and per-dataset early stopping—so the numbers need verification before being trusted. read the letter →

arxiv 2505.17532 v1 pith:2IW3XJ4T submitted 2025-05-23 cs.LG

Bin Wang , Heming Yang , Jinfang Sheng This is my paper
classification cs.LG
keywords long-termtimeseriesforecastingmulti-scaledecompositionadaptiveconvolutionfrequency-domainlosssharpness-awareminimizationchannelindependenceMixerFouriertransform
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

TimeCF is a deep learning model for long-term multivariate time series forecasting, built on the multi-scale decomposition-and-mixing architecture of TimeMixer. The paper's claim is that adding an adaptive convolution module—which applies the same kernel to differently downsampled scales so that one fixed block yields both local and global receptive fields—and a loss that combines a frequency-domain L1 term with MSE, updated with sharpness-aware minimization, produces a forecaster that beats eight established baselines on average. The supporting evidence is a comparison across six real-world datasets (Weather, ETTh1/2, ETTm1/2, Electricity) at four prediction horizons, where TimeCF reports the best total-average MSE (0.310) and MAE (0.345), about 2.2% and 1.9% better than the runner-up, with 125K-275K parameters. If correct, the paper would show that a compact channel-independent model can handle label autocorrelation and generalization without a transformer-sized parameter budget. The ablation study supports that both the convolution block and the SAMFre loss are needed for the full gain.

What carries the argument

The load-bearing machinery is the PDMC (Past Decomposable Mixing with adaptive Conv) module together with the SAMFre loss. PDMC downsamples the input into scales of length 96, 48, 24, and 12; runs the same three-layer convolution block (kernel 3, padding 1) on each scale; decomposes each into seasonal and trend parts; mixes seasons bottom-up and trends top-down; and outputs through a feed-forward network. SAMFre defines the total loss as $\alpha\|\mathrm{FFT}(pred)-\mathrm{FFT}(real)\|_1+(1-\alpha)\,\mathrm{MSE}$ and, after a threshold number of updates, replaces ordinary gradient descent with the SAM step $\hat{\epsilon}(w)=\rho\,\nabla_w\mathrm{Loss}/\|\nabla_w\mathrm{Loss}\|_2$, updating at $w+\hat{\epsilon}(w)$. The receptive-field calculation is what turns a single fixed convolution block into a multi-scale global/local feature extractor, and the frequency-domain loss is what decouples label autocorrelation in a channel-independent model.

What would settle it

Re-run TimeCF and the eight baselines on the six datasets with identical training schedules, one fixed early-stopping rule, and all hyperparameters selected on a validation split only; if TimeCF no longer has the best total-average MSE and MAE, the paper's headline comparison is an artifact of per-dataset test-time tuning.

Watch

Extended reading notes

Core claim

The central claim is that same-sized convolution kernels on downsampled scales act as a size-free receptive-field mechanism: with lookback 96, three scales of downsampling, kernel size 3, padding 1, and L stacked PDMC blocks, each time point sees 6·L neighbors, which covers the whole finest-scale sequence while covering only part of coarser scales. This gives global information where the window covers the sequence and local information elsewhere, without changing kernel size per scale. After convolution, each scale is decomposed into seasonal and trend parts, mixed bottom-up for seasons and top-down for trends, and fused through a feed-forward network. The SAMFre loss projects predictions and labels into the frequency domain, computes $\alpha\|\mathrm{FFT}(pred)-\mathrm{FFT}(real)\|_1+(1-\alpha)\,\mathrm{MSE}$, and switches to sharpness-aware gradient updates after a threshold number of steps. The paper argues this decouples label autocorrelation and flattens the loss landscape, and reports that the complete TimeCF outperforms TimeMixer and all baselines in average MSE/MAE, with ablations showing the convolution and SAMFre parts are complementary.

Load-bearing premise

The reported advantage assumes that the per-dataset training rounds, early stopping thresholds, and hyperparameters such as $\alpha$ and the SAM activation threshold were chosen without looking at test-set performance; if any of these were tuned on the test set, the average gains could be selection artifacts rather than real model superiority.

Editorial extensions

If this is right

  • If the average results hold, TimeCF is a parameter-efficient alternative to transformer-based forecasters on datasets with many channels, using 125K-275K parameters where the transformer baselines use millions.
  • The receptive-field mechanism implies the same fixed convolution block can be reused across datasets and lookback lengths by changing only the downsampling schedule, not the kernel sizes.
  • Because the SAMFre loss is defined independently of the PDMC architecture, it could be attached to other channel-independent forecasters and tested for the same autocorrelation-decoupling benefit.
  • The ablation pattern suggests convolution and SAMFre act as complements: convolution gathers autocorrelated multi-scale features and SAMFre then decouples that autocorrelation, so removing either leaves most of the gain unrealized.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: using the same kernel on downsampled scales is mathematically equivalent to a particular dilated receptive field without dilation rates, so a direct comparison against explicit dilated convolutions would isolate what the multi-scale mixing, rather than the receptive-field geometry, contributes.
  • Beyond the paper: the paper's ablations remove the whole SAMFre module but not its two parts separately, so the reported gain does not yet separate the effect of the frequency-domain L1 term from the effect of sharpness-aware updates; a two-factor ablation would settle that.
  • Beyond the paper: the frequency-domain term can be read as softly matching the spectral envelope of the forecast; a testable extension is to weight frequency bins by their contribution to long-horizon error, which could improve performance on non-stationary datasets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The manuscript proposes TimeCF, a multivariate long-term time series forecasting model built on the TimeMixer architecture. The model first downsamples each input channel into multiple scales, applies a shared-size convolution block (PDMC) to aggregate local and global information, decomposes each scale into seasonal and trend components, mixes these components across scales, and finally sums per-scale predictions after inverse normalization. Training uses a loss that combines MSE with an L1 frequency-domain loss derived from FreDF, together with a Sharpness-Aware Minimization update applied after a threshold number of steps. The authors evaluate TimeCF on six datasets (ETTh1, ETTh2, ETTm1, ETTm2, Weather, ECL) against eight baselines and report a total average MSE of 0.310 and MAE of 0.345, which they state is 2.2% and 1.9% better than the suboptimal model. They also report parameter counts of 125K-275K and an ablation study against TimeMixer.

Significance. If the reported results are reproducible under a fair protocol, TimeCF would be a meaningful lightweight contribution: it combines two existing ideas (frequency-domain loss and sharpness-aware minimization) with a multiscale convolution mixing module, and the reported average gains over TimeMixer, iTransformer, PatchTST, and other baselines are consistent with what a small, well-designed model can achieve. The paper also provides an explicit parameter comparison, which is a useful practical consideration. However, the central empirical claim rests on a small margin (2.2% MSE, 1.9% MAE) and the current manuscript does not provide the information needed to verify or replicate the comparison: no code, no seeds, no error bars, and several hyperparameters in the model description are left unspecified. The architecture is described clearly enough to be re-implemented, but the evaluation protocol is not auditable as written. The ablation study qualitatively supports the usefulness of the two proposed components, but again lacks statistical grounding.

major comments (5)
  1. [4.1 and 3.5] The experimental protocol is not auditable as written. Section 4.1 states that TimeCF 'sets different training rounds and early stopping thresholds according to the size of different data sets,' while Section 3.5 leaves the loss weight alpha in Eq. (10), the SAM neighborhood size rho in Eq. (11), and the SAM activation threshold unspecified. Similarly, Section 3.2 and 3.3 do not report the pooling window d, the embedding dimension D, the number of scales k, or the PDMC stack count L. No code or seed list is provided. Because the main claim is a 0.310 vs 0.317 MSE margin (2.2%) and 0.345 vs 0.352 MAE margin (1.9%), run-to-run variation and early-stopping choices in long-term forecasting can easily account for the difference. I request a fixed protocol, full hyperparameter disclosure, multiple seeds with mean and standard deviation, and either code release or a complete configuration listing.
  2. [4.2] The sentence 'The average values of MSE and MAE increased by 2.2% and 1.9% compared with the suboptimal model' states the direction incorrectly: lower MSE/MAE is better, and Table 1 shows that TimeCF's values are lower, not higher. This should be rewritten as a relative improvement or decrease. The sentence should also clarify how the 'Total AVG' is computed, since it appears to be an unweighted mean of six dataset averages and this should be stated explicitly.
  3. [4.1 and 4.2] The comparison is not uniform across baselines: Section 4.2 says that FreTS results are taken from the original paper, while all other results are from local experiments. This mixes hardware, training protocols, and possibly data splits. Given that the reported gains over some baselines are small, the FreTS comparison should either be reproduced under the same protocol or clearly marked as an external reference without being included in the aggregate ranking.
  4. [4.3] The ablation study in Table 2 reports only single numbers with no variance or significance testing, yet the text claims that the improvements 'are not significant' for some variants and that both modules are 'indispensable parts of the TimeCF model.' These are statistical claims that cannot be supported by single runs. At minimum, the authors should either provide multiple seeds with error bars or explicitly state that no significance testing was performed and temper the 'indispensable' conclusion accordingly.
  5. [3.3 and 3.5] The symbol alpha is used for two different quantities: the residual weight in Eq. (3) and the loss combination weight in Eq. (10). No value is given for either, and the SAM update in Eq. (11) introduces rho without stating its value. This makes the model impossible to reproduce from the paper alone. Please use distinct symbols and report the numerical values used in the experiments.
minor comments (5)
  1. [Title and throughout] The title contains a typo: 'time seris' should be 'time series'. There are several other typographical errors in the text, including 'imporving' in the introduction and 'roposes' in Section 2.1.
  2. [Table 1] The row labeled '1st Times' is not explained in the text; please state that it counts the number of horizons/datasets where each model achieves the best result. Also, the description 'bold red' and 'bold black' is not visible in a black-and-white printout; consider using symbols or a clear notation.
  3. [Table 3] The parameter counts for SparseTSF (0.041K-0.581K, i.e., 41-581 parameters) appear inconsistent with the SparseTSF paper's claim of roughly 1K parameters. Please clarify the units, the condition (lookback/prediction length), and whether these counts are measured or taken from the original paper.
  4. [Section 3.2] The RevIN normalization is used but not cited; please add the reference to the RevIN paper (Kim et al., 2022).
  5. [Figure 1] Figure 1 includes symbols such as 'L' and 'F(.)' that are not all defined in the caption or the surrounding text; a clearer caption would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: TimeCF is an empirical architecture combining cited external components (TimeMixer, FreDF, SAM); none of its equations forces the reported accuracy by construction.

full rationale

The claimed derivation chain is architectural rather than mathematical: Section 3.2-3.4 defines preprocessing, PDMC convolution-mixing, and output prediction through Eqs. (1)-(9), and Section 3.5 defines the SAMFre loss as a weighted combination of frequency-domain L1 and MSE in Eq. (10), with standard SAM updates in Eqs. (11)-(13). No equation is defined in terms of the target evaluation metric, and no learned or fitted parameter is renamed as a prediction. The cited bases (TimeMixer, FreDF, SAM) are prior external work, not self-citations by the present authors, so there is no load-bearing self-citation chain. The central claims in Section 4.2 ('TimeCF has shown good performance on most datasets' and the 2.2%/1.9% average improvements) are empirical statements supported by Table 1; they are not entailed by the model definition. The main weakness is auditability, not circularity: Section 4.1 says 'the test of the TimeCF model we proposed sets different training rounds and early stopping thresholds according to the size of different data sets to improve test efficiency,' and Section 3.5 leaves alpha, rho, and the SAM-activation threshold unspecified, so the comparison protocol is not fully fixed and the margin could be affected by protocol choices. That is a correctness or reproducibility risk, which I flag as such, but it does not make the result equivalent to its inputs by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The model's load-bearing content is imported from TimeMixer, FreDF, and SAM; the new pieces are architectural arrangements and a loss combination. The central free parameters are undisclosed, weakening the empirical claim.

free parameters (6)
  • alpha = not reported
    Weight in Eq. 10 between FFT L1 loss and MSE; chosen by hand, no value or sensitivity analysis given.
  • pooling window d = not reported
    Eq. 1 downsampling window length; defines the multi-scale sequence lengths, no value is given.
  • embedding dimension D = not reported
    Dimension of token and temporal embeddings in Eq. 2; architectural hyperparameter not disclosed.
  • PDMC stack count L = not reported
    Number of repeated PDMC blocks, stated as L >= 2 in Section 3.3; the exact value is not specified.
  • SAM activation threshold = not reported
    Training step after which SAM perturbation in Eqs. 11-13 is used; the threshold is mentioned but never specified.
  • training rounds and early stopping per dataset = varies by dataset
    Section 4.1 says different training rounds and early stopping thresholds are set according to dataset size 'to improve test efficiency'; this per-dataset choice can affect the reported results.
assumptions (5)
  • domain assumption Multi-scale time series decomposition followed by seasonal-trend mixing improves forecasting accuracy.
    Taken from TimeMixer; TimeCF inherits this prior knowledge without testing alternatives.
  • domain assumption Frequency-domain L1 loss computed on the FFT of predictions and targets reduces label autocorrelation and improves generalization under channel-independent modeling.
    Adapted from FreDF; the paper provides no proof or ablation isolating the autocorrelation mechanism; it is an imported assumption.
  • ad hoc to paper Sharpness-aware minimization improves generalization when applied after a fixed number of training steps.
    SAM is a published method, but the thresholded application schedule is introduced here without justification or sensitivity analysis.
  • standard math Three convolutional layers with kernel size 3 and padding 1 yield a receptive field of at least 6 neighboring points per scale, and identical kernels at different scales provide distinct local versus global receptive fields.
    Receptive field arithmetic in Section 3.3 is standard, but the claim that this 'adaptively aggregates information of different scales' is asserted, not proven.
  • domain assumption The benchmark dataset splits and normalization are standard and comparable across models.
    The paper relies on standard ETT, Weather, and ECL evaluation protocols but does not describe the splits or exactly how baselines were run; different splits or protocols would change the comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TimeCF: A TimeMixer-Based Model with adaptive Convolution and Sharpness-Aware Minimization Frequency Domain Loss for long-term time seris forecasting." pith.science (2026). https://pith.science/paper/2IW3XJ4T

@misc{pith2026250517532,
  author       = {Pith},
  title        = {Pith review of: TimeCF: A TimeMixer-Based Model with adaptive Convolution and Sharpness-Aware Minimization Frequency Domain Loss for long-term time seris forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2IW3XJ4T}},
  note         = {Machine review of arXiv:2505.17532}
}
read the original abstract

Recent studies have shown that by introducing prior knowledge, multi-scale analysis of complex and non-stationary time series in real environments can achieve good results in the field of long-term forecasting. However, affected by channel-independent methods, models based on multi-scale analysis may produce suboptimal prediction results due to the autocorrelation between time series labels, which in turn affects the generalization ability of the model. To address this challenge, we are inspired by the idea of sharpness-aware minimization and the recently proposed FreDF method and design a deep learning model TimeCF for long-term time series forecasting based on the TimeMixer, combined with our designed adaptive convolution information aggregation module and Sharpness-Aware Minimization Frequency Domain Loss (SAMFre). Specifically, TimeCF first decomposes the original time series into sequences of different scales. Next, the same-sized convolution modules are used to adaptively aggregate information of different scales on sequences of different scales. Then, decomposing each sequence into season and trend parts and the two parts are mixed at different scales through bottom-up and top-down methods respectively. Finally, different scales are aggregated through a Feed-Forward Network. What's more, extensive experimental results on different real-world datasets show that our proposed TimeCF has excellent performance in the field of long-term forecasting.

Figures

Figures reproduced from arXiv: 2505.17532 by the authors.

Figure 1
Figure 1. TimeCF Architecture regarded as the input instance of TimeCF. The overall structure of TimeCF is shown in [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 29 canonical work pages

  1. [1]

    , author Mohtar, W.H.M.W

    author Afan, H.A. , author Mohtar, W.H.M.W. , author Khaleel, F. , author Kamel, A.H. , author Mansoor, S.S. , author Alsultani, R. , author Ahmed, A.N. , author Sherif, M. , author El-Shafie, A. , year 2024 . title Data-driven water quality prediction for wastewater treatment plants . journal Heliyon volume 10 . note Publisher: Elsevier

    work page 2024

  2. [2]

    , author Mehmood, R

    author Alkhayat, G. , author Mehmood, R. , year 2021 . title A review and taxonomy of wind and solar energy forecasting methods based on deep learning . journal Energy and AI volume 4 , pages 100060 . note Publisher: Elsevier

    work page 2021

  3. [3]

    , author Xie, L

    author Bi, K. , author Xie, L. , author Zhang, H. , author Chen, X. , author Gu, X. , author Tian, Q. , year 2023 . title Accurate medium-range global weather forecasting with 3D neural networks . journal Nature volume 619 , pages 533--538 . note Publisher: Nature Publishing Group

    work page 2023

  4. [4]

    , author Olivares, K.G

    author Challu, C. , author Olivares, K.G. , author Oreshkin, B.N. , author Ramírez, F.G. , author Canseco, M.M. , author Dubrawski, A. , year 2023 . title NHITS : Neural Hierarchical Interpolation for Time Series Forecasting , in: booktitle AAAI , pp. pages 6989--6997 . https://doi.org/10.1609/aaai.v37i6.25854

    doi:10.1609/aaai.v37i6.25854 2023

  5. [5]

    , author Gulcehre, C

    author Chung, J. , author Gulcehre, C. , author Cho, K. , author Bengio, Y. , year 2014 . title Empirical evaluation of gated recurrent neural networks on sequence modeling , in: booktitle NIPS 2014 Workshop on Deep Learning , December 2014

    work page 2014

  6. [6]

    , author xue, w

    author donghao, L. , author xue, w. , year 2024 . title ModernTCN : A Modern Pure Convolution Structure for General Time Series Analysis , in: booktitle The Twelfth International Conference on Learning Representations . https://openreview.net/forum?id=vpJMJerXHU

    work page 2024

  7. [7]

    , author Kohen, E

    author Farhi, N. , author Kohen, E. , author Mamane, H. , author Shavitt, Y. , year 2021 . title Prediction of wastewater treatment quality using LSTM neural network . journal Environmental Technology & Innovation volume 23 , pages 101632 . note Publisher: Elsevier

    work page 2021

  8. [8]

    , author Kleiner, A

    author Foret, P. , author Kleiner, A. , author Mobahi, H. , author Neyshabur, B. , year 2021 . title Sharpness-aware Minimization for Efficiently Improving Generalization , in: booktitle International Conference on Learning Representations . https://openreview.net/forum?id=6Tm1mposlrM

    work page 2021

Show all 36 references
  1. [9]

    , author Dao, T

    author Gu, A. , author Dao, T. , year 2023 . title Mamba: Linear -time sequence modeling with selective state spaces . journal arXiv preprint arXiv:2312.00752

    2023 arXiv

  2. [10]

    , author Schmidhuber, J

    author Hochreiter, S. , author Schmidhuber, J. , year 1997 . title Long short-term memory . journal Neural computation volume 9 , pages 1735--1780 . note Publisher: MIT press

    1997

  3. [11]

    , author Zhao, Z

    author Huang, S. , author Zhao, Z. , author Li, C. , author Bai, L. , year 2025 . title TimeKAN : KAN -based Frequency Decomposition Learning Architecture for Long -term Time Series Forecasting . journal arXiv preprint arXiv:2502.06910

    2025 arXiv

  4. [12]

    , author Zhang, B

    author Huang, X. , author Zhang, B. , author Feng, S. , author Ye, Y. , author Li, X. , year 2023 . title Interpretable local flow attention for multi-step traffic flow prediction . journal Neural networks volume 161 , pages 25--38 . note Publisher: Elsevier

    2023

  5. [13]

    , author Zhang, Y

    author Huo, G. , author Zhang, Y. , author Wang, B. , author Gao, J. , author Hu, Y. , author Yin, B. , year 2023 . title Hierarchical spatio–temporal graph convolutional networks and transformer network for traffic flow forecasting . journal IEEE Transactions on Intelligent T...

    2023

  6. [14]

    , author Odonnat, A

    author Ilbert, R. , author Odonnat, A. , author Feofanov, V. , author Virmaux, A. , author Paolo, G. , author Palpanas, T. , author Redko, I. , year 2024 . title SAMformer : Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness - Aware Minimization ...

    2024

  7. [15]

    , author Jian, C

    author Li, L. , author Jian, C. , author Wan, F. , author Geng, D. , author Fang, Z. , author Chen, L. , author Gao, Y. , author Jiang, W. , author Zhu, J. , year 2024 . title LagCNN : A Fast yet Effective Model for Multivariate Long -term Time Series Forecasting , in: booktit...

    2024

  8. [16]

    , author Zhou, P

    author Li, T. , author Zhou, P. , author He, Z. , author Cheng, X. , author Huang, X. , year 2024 . title Friendly Sharpness - Aware Minimization , in: booktitle Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition ( CVPR ) , pp. pages 5631--5640

    2024

  9. [17]

    , author Yang, E

    author Liang, X. , author Yang, E. , author Deng, C. , author Yang, Y. , year 2024 . title CrossFormer : Cross - Modal Representation Learning via Heterogeneous Graph Transformer . journal ACM Trans. Multim. Comput. Commun. Appl. volume 20 , pages 380:1--380:21 . https://doi.o...

    2024 doi

  10. [18]

    , author Lin, W

    author Lin, S. , author Lin, W. , author Wu, W. , author Chen, H. , author Yang, J. , year 2024 . title SparseTSF : Modeling Long -term Time Series Forecasting with *1k* Parameters , in: booktitle Forty-first International Conference on Machine Learning . https://openreview.ne...

    2024

  11. [19]

    , author Hu, T

    author Liu, Y. , author Hu, T. , author Zhang, H. , author Wu, H. , author Wang, S. , author Ma, L. , author Long, M. , year 2024 . title iTransformer : Inverted Transformers Are Effective for Time Series Forecasting , in: booktitle The Twelfth International Conference on Lear...

    2024

  12. [20]

    , author Nguyen, N.H

    author Nie, Y. , author Nguyen, N.H. , author Sinthong, P. , author Kalagnanam, J. , year 2023 . title A Time Series is Worth 64 Words : Long -term Forecasting with Transformers , in: booktitle The Eleventh International Conference on Learning Representations . https://openrev...

    2023

  13. [21]

    , author Flunkert, V

    author Salinas, D. , author Flunkert, V. , author Gasthaus, J. , author Januschowski, T. , year 2020 . title DeepAR : Probabilistic forecasting with autoregressive recurrent networks . journal International journal of forecasting volume 36 , pages 1181--1191 . note Publisher: Elsevier

    2020

  14. [22]

    , author Dharrao, D.S

    author Sonkavde, G. , author Dharrao, D.S. , author Bongale, A.M. , author Deokate, S.T. , author Doreswamy, D. , author Bhat, S.K. , year 2023 . title Forecasting stock market prices using machine learning and deep learning models: A systematic review, performance analysis an...

    2023

  15. [23]

    , author Shazeer, N

    author Vaswani, A. , author Shazeer, N. , author Parmar, N. , author Uszkoreit, J. , author Jones, L. , author Gomez, A.N. , author Kaiser, . , author Polosukhin, I. , year 2017 . title Attention is all you need . journal Advances in neural information processing systems volume 30

    2017

  16. [24]

    , author Pan, L

    author Wang, H. , author Pan, L. , author Chen, Z. , author Yang, D. , author Zhang, S. , author Yang, Y. , author Liu, X. , author Li, H. , author Tao, D. , year 2024 . title Fredf: Learning to forecast in frequency domain . journal arXiv preprint arXiv:2402.02399

    2024 arXiv

  17. [25]

    , author Peng, J

    author Wang, H. , author Peng, J. , author Huang, F. , author Wang, J. , author Chen, J. , author Xiao, Y. , year 2023 . title MICN : Multi -scale Local and Global Context Modeling for Long -term Series Forecasting , in: booktitle The Eleventh International Conference on Learn...

    2023

  18. [26]

    , author Wu, H

    author Wang, S. , author Wu, H. , author Shi, X. , author Hu, T. , author Luo, H. , author Ma, L. , author Zhang, J.Y. , author Zhou, J. , year 2024 . title TimeMixer : Decomposable Multiscale Mixing for Time Series Forecasting , in: booktitle ICLR . https://openreview.net/for...

    2024

  19. [27]

    , author Hu, T

    author Wu, H. , author Hu, T. , author Liu, Y. , author Zhou, H. , author Wang, J. , author Long, M. , year 2023 . title TimesNet : Temporal 2D - Variation Modeling for General Time Series Analysis , in: booktitle The Eleventh International Conference on Learning Representatio...

    2023

  20. [28]

    , author Xu, J

    author Wu, H. , author Xu, J. , author Wang, J. , author Long, M. , year 2021 . title Autoformer: Decomposition Transformers with Auto - Correlation for Long - Term Series Forecasting , in: editor Beygelzimer, A. , editor Dauphin, Y. , editor Liang, P. , editor Vaughan, J.W. (...

    2021

  21. [29]

    , author Zeng, A

    author Xu, Z. , author Zeng, A. , author Xu, Q. , year 2024 . title FITS : Modeling Time Series with 10k Parameters , in: booktitle The Twelfth International Conference on Learning Representations . https://openreview.net/forum?id=bWcnvZ3qMb

    2024

  22. [30]

    , author Zhang, Q

    author Yi, K. , author Zhang, Q. , author Fan, W. , author Wang, S. , author Wang, P. , author He, H. , author An, N. , author Lian, D. , author Cao, L. , author Niu, Z. , year 2023 . title Frequency-domain mlps are more effective learners in time series forecasting . journal ...

    2023

  23. [31]

    , author Cao, X

    author Yin, L. , author Cao, X. , author Liu, D. , year 2023 . title Weighted fully-connected regression networks for one-day-ahead hourly photovoltaic power forecasting . journal Applied Energy volume 332 , pages 120527 . note Publisher: Elsevier

    2023

  24. [32]

    , author Jiang, J

    author Yue, Y. , author Jiang, J. , author Ye, Z. , author Gao, N. , author Liu, Y. , author Zhang, K. , year 2023 . title Sharpness- Aware Minimization Revisited : Weighted Sharpness as a Regularization Term , in: booktitle KDD , pp. pages 3185--3194 . https://doi.org/10.1145...

    2023

  25. [33]

    , author Chen, M

    author Zeng, A. , author Chen, M. , author Zhang, L. , author Xu, Q. , year 2023 . title Are transformers effective for time series forecasting? , in: booktitle Proceedings of the AAAI conference on artificial intelligence , pp. pages 11121--11128 . note Issue: 9

    2023

  26. [34]

    , year 2003

    author Zhang, G.P. , year 2003 . title Time series forecasting using a hybrid ARIMA and neural network model . journal Neurocomputing volume 50 , pages 159--175 . note Publisher: Elsevier

    2003

  27. [35]

    , author Zhang, S

    author Zhou, H. , author Zhang, S. , author Peng, J. , author Zhang, S. , author Li, J. , author Xiong, H. , author Zhang, W. , year 2021 . title Informer: Beyond efficient transformer for long sequence time-series forecasting , in: booktitle Proceedings of the AAAI conference...

    2021

  28. [36]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.