Pith. sign in

REVIEW 4 major objections 6 minor 37 references

MoFE-Time: Mixture of Frequency Domain Experts for Time-Series Forecasting Models

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read MoFE-Time claims that adding a frequency-domain expert cell to each mixture-of-experts route, plus pretraining on a large corpus, sets a new accuracy bar for time-series forecasting, cutting MSE by 6.95% and MAE by 6.02% on average over…

desk verdict Useful incremental architecture, but the SOTA claim is overbroad and the missing pretraining-exclusion statement undermines the comparison with non-pretrained baselines. read the letter →

arxiv 2507.06502 v1 pith:ZKMHK5NY submitted 2025-07-09 cs.LG cs.AI

Yiwen Liu , Chenyu Zhang , Junjie Song , Siqi Chen , Sun Yin , Zihan Wang , Lingming Zeng , Yuji Cao
show 1 more author
Junming Jiao
This is my paper
classification cs.LGcs.AI
keywords timeseriesforecastingMixtureofExpertsfrequencydomainanalysisFourierpretraining-finetuningRevINfoundationmodelsNEV-salesdataset
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

The paper tries to establish that a time-series forecasting model can be made more accurate by having each expert in a mixture-of-experts network learn both time-domain and frequency-domain structure, and by pretraining on a large corpus before fine-tuning on target data. On six public benchmarks, the authors report new state-of-the-art results, with average MSE and MAE of 0.2755 and 0.3226, which are 6.95% and 6.02% lower than the next-best model, Time-MoE. They also report strong results on a proprietary NEV store-traffic dataset. If the claim holds, the Frequency-Time Cell and the pretraining-finetuning recipe would be a practical improvement for forecasting periodic and non-stationary series.

What carries the argument

The Frequency-Time Cell (FTC) is the central module. It splits the incoming representation into a frequency path, which maps the segment through a linear layer and then projects it onto harmonic basis functions $e^{j\omega n}$ (sine and cosine), and a time path, a linear layer. The MoE router selects the most relevant experts; the learned frequencies are the expert parameters, and the routing weights act as amplitudes in a sum of harmonics. The combination of the harmonic combination vector with the time-domain signal vector forms the layer output. RevIN normalizes each instance and denormalizes predictions, and temporal aggregation allows variable-length context.

What would settle it

Check the composition of Time-300B and test for overlap with the six benchmarks; if any of ETTh1/2, ETTm1/2, Weather, or Exchange series, especially their test splits, are included in pretraining, the SOTA claim is undermined. Alternatively, retrain MoFE-Time on a pretraining corpus that provably excludes those series and confirm the reported error reductions persist.

Watch

Extended reading notes

Core claim

The central claim is that jointly modeling time and frequency inside each MoE expert improves forecasting accuracy over both pure time-domain MoE models and frequency-transform models. Rather than transforming the whole input with a fixed Fourier transform, MoFE-Time lets each expert learn its own set of critical harmonic frequencies; the router then picks and weights the top-k experts, so the signal is reconstructed as a weighted sum of learned harmonics. The frequency output is concatenated with a time-domain feature vector, and RevIN plus temporal aggregation handle non-stationarity and variable-length inputs. Pretrained on the Time-300B corpus and then fine-tuned, the model reports the best average MSE/MAE across ETTh1/2, ETTm1/2, Weather, and Exchange, and on the NEV-sales dataset.

Load-bearing premise

The claim rests on the evaluation datasets (ETTh1/2, ETTm1/2, Weather, and Exchange) being absent from the Time-300B pretraining corpus; if they appear there, the reported fine-tuning gains could be memorization rather than transferable forecasting skill.

Editorial extensions

If this is right

  • Averaged over the six benchmarks, MoFE-Time reports MSE 0.2755 and MAE 0.3226, beating Time-MoE by 6.95% and 6.02%; if the comparison holds, it sets a new accuracy bar for general-purpose time-series forecasters.
  • The FTC module can be added to existing MoE forecasting backbones: inference is reported equal or faster than Time-MoE at similar parameter counts (117.95M vs 113.35M), so the frequency modeling does not cost latency.
  • Ablations show the pretraining stage contributes the largest error reduction, and the FTC module improves both metrics on all six datasets; this supports the combination of external pattern knowledge with explicit periodicity modeling.
  • On NEV-sales, average MSE/MAE of 0.1956/0.3284 over prediction lengths 8, 16, and 24 indicates practical value for daily store-traffic forecasting.

Reading between the lines

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

  • The reported advantage may partly come from the pretraining corpus rather than from frequency experts; a direct test would be to compare MoFE-Time against a Time-MoE variant fine-tuned with the same pretraining checkpoint and data, isolating the FTC contribution.
  • The learned frequencies per expert could be inspected for alignment with natural periods such as 24 hours or 7 days; if interpretable, the model would offer a diagnostic view of which periodicities drive predictions.
  • The same harmonic-combination mechanism could generalize to other tasks where periodicity matters, such as anomaly detection or imputation, where the router-weighted harmonics could serve as an interpretable seasonal decomposition.
  • A zero-shot transfer test across domains with different periodicity distributions would test the paper's generalization claim more sharply than fine-tuning alone.
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

4 major / 6 minor

Summary. The paper proposes MoFE-Time, a time-series forecasting model built on the Time-MoE mixture-of-experts architecture, in which selected experts are replaced by Frequency-Time Cells (FTC) that combine a linear projection, a complex-exponential frequency-style map, and a time-domain branch. The model is pretrained on the Time-300B corpus and then fine-tuned on six public benchmarks and a proprietary NEV-sales dataset. The main empirical claim, stated in the Abstract and Section 4.2, is an average improvement of 6.95% in MSE and 6.02% in MAE over Time-MoE, described as new state-of-the-art performance.

Significance. If the claimed gains are robust, the paper makes a moderate but useful contribution: it shows that frequency-domain experts inside an MoE routing framework, combined with a pretraining-finetuning paradigm, can improve on a strong billion-scale baseline, and the proprietary NEV-sales dataset is a potentially valuable real-world testbed. The paper also includes ablations and an inference-speed comparison with Time-MoE. However, the headline claim rests on an average over six benchmarks and is not yet supported by variance information, a data-contamination statement, or sufficient implementation detail; the significance of the contribution therefore depends on whether these gaps can be closed.

major comments (4)
  1. [§4.1 (Pretraining Dataset; Finetuning and Evaluation Dataset)] The pretraining corpus Time-300B is described as comprising public datasets in energy, retail, healthcare, weather, finance, transportation, and networking, while the evaluation sets are ETTh1/2, ETTm1/2, Weather, and Exchange. The paper never states that these benchmark series, or their test splits, were excluded from pretraining. Because Time-300B is a public corpus assembled from standard datasets and ETT/Weather/Exchange are standard public datasets, overlap is plausible; without an explicit exclusion statement, the 6.95%/6.02% average gain over Time-MoE could reflect memorization of the evaluation series rather than transferable time-frequency knowledge. This is load-bearing for the headline claim and must be addressed, for example with a contamination check or a precise list of constituent datasets.
  2. [§4.2, Table 2] The state-of-the-art claim is overbroad as stated. In Table 2, MoFE-Time's average MSE on ETTm2 is 0.2070, worse than TimeMixer's 0.2020, and its average MSE on Exchange is 0.2767, worse than Time-MoE's 0.2687; Section 4.2 itself concedes the ETTm2 case. Moreover, no error bars, random seeds, or significance tests are reported anywhere, so the reported average differences (for example 0.2755 vs. 0.2961 MSE) are not shown to be outside run-to-run variation. The abstract's unqualified 'new state-of-the-art performance' should be revised to describe best average performance over these six benchmarks in the reported runs, and variance evidence should be added.
  3. [§3.2, Eqs. (9)-(13)] The derivation does not connect the implemented module to the claimed learned harmonics. The text states that the model learns critical frequencies {omega_i} and amplitudes {a_i}, but Eq. (13) defines the expert output as X^k_t = exp(j X^k_f), where X^k_f is a linear projection of the input; no frequency parameters omega_i are explicitly parameterized or extracted. As written, the module is a complex exponential of a learned projection, not a Fourier transform or a sum of k learned harmonics as in Eqs. (9)-(11). The FTC may still be an effective feature map, but the conceptual claim that the network 'directly learns the Fourier transform process' is not supported by the equations and should be clarified or corrected.
  4. [§4.1 Implementation details] The paper omits several hyperparameters needed to reproduce the method and the comparisons: the number of experts k, harmonics per expert h, top-k routing count, auxiliary-loss weight alpha, number of model layers/heads/dimensions, and the exact fine-tuning protocol applied to each baseline. No code or configuration files are provided. Without these, the reported 6.95% improvement cannot be independently verified, and the comparison against Time-MoE is not fully specified.
minor comments (6)
  1. [§3.3, Eq. (14)] In Eq. (14), the surrounding text says that \hat y_t is the actual data point and y_t is the predicted value, but the equation's error terms suggest the opposite convention; the Huber threshold delta is also not specified.
  2. [§3.3, Eq. (15)] The indicator function in the definition of f_i is not defined; the reader is left to guess that it counts tokens routed to expert i.
  3. [§4.2] The text says 'we reproduce six well-known time series forecasting models' but then names only Time-MoE, TimeMixer, TimeXer, TimesNet, and AutoFormer; PatchTST appears in Table 2 but is omitted from the list.
  4. [Figure 4] The ablation bars are difficult to read because the figure lacks labeled axes and legend details, and the caption/text contains the typo 'MES' for 'MSE'.
  5. [§4.6 and Figure 6] The model name is written inconsistently as 'MoFe-Time' in Section 4.6 and Figure 6, whereas the rest of the paper uses 'MoFE-Time'.
  6. [Table 2] The row label '1th Count' is unclear; it should be defined, for example as the number of settings in which each method achieves the lowest metric.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circular step; the empirical comparison is self-contained, with only a non-circular pretraining-overlap caveat.

full rationale

MoFE-Time's central claims are empirical: benchmark MSE/MAE comparisons and an architecture proposal. The FTC equations (Eqs. 9-13) define a parameterized harmonic representation and are not derived from the prediction targets; the autoregressive and auxiliary losses (Eqs. 14-16) are standard supervised objectives. The paper invokes Time-MoE [23] as prior architecture and corpus, but that work is by a different author group, so no self-citation chain is load-bearing. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported, and no result reduces to its input by construction. The one caveat is that Section 4.1 never states that ETTh/ETTm/Weather/Exchange (or their test splits) are excluded from Time-300B; if they were included, the fine-tuning gains could reflect memorization. That would be data contamination, a soundness issue, not circularity by the paper's equations or citations. Accordingly, no circular steps are listed.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central performance claim rests on a standard Fourier representation, an unproven parameterization of expert frequencies, and an untested assumption that the pretraining corpus does not leak evaluation data. The model's hyperparameters (k, h, top-k, auxiliary loss weight) are free choices that are not reported, making the method difficult to audit or reproduce.

free parameters (4)
  • Number of experts k = not reported
    The MoE architecture requires choosing the number of experts; the paper never states k, although the reconstruction in Eq. 9-11 depends on it.
  • Harmonics per expert h = not reported
    Each expert is claimed to learn h distinct critical frequencies; h is not reported anywhere in the paper.
  • Top-k routing count = not reported
    The router selects a subset of experts, but the top-k value is not disclosed, making the sparse activation pattern impossible to reproduce.
  • Auxiliary loss weight alpha = not reported
    Eq. 16 uses alpha to balance forecasting and load-balancing losses; the chosen value is not given.
assumptions (3)
  • standard math A discrete non-periodic signal can be represented as a finite sum of k harmonics with learned frequencies and amplitudes (Eq. 9-11).
    This is the Fourier representation used to justify the FTC design; it is standard signal processing.
  • ad hoc to paper Exponentiating a linear projection of the input gives a valid learned harmonic basis e^{j X_f} (Eq. 13).
    The paper provides no proof that this parameterization learns the claimed critical frequencies; it is an architectural choice.
  • domain assumption The pretraining corpus Time-300B is disjoint from the evaluation datasets.
    Section 4.1 lists Time-300B as containing public energy, weather, and finance data, but does not state that ETTh/ETTm/Weather/Exchange are excluded from pretraining.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MoFE-Time: Mixture of Frequency Domain Experts for Time-Series Forecasting Models." pith.science (2026). https://pith.science/paper/ZKMHK5NY

@misc{pith2026250706502,
  author       = {Pith},
  title        = {Pith review of: MoFE-Time: Mixture of Frequency Domain Experts for Time-Series Forecasting Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKMHK5NY}},
  note         = {Machine review of arXiv:2507.06502}
}
read the original abstract

As a prominent data modality task, time series forecasting plays a pivotal role in diverse applications. With the remarkable advancements in Large Language Models (LLMs), the adoption of LLMs as the foundational architecture for time series modeling has gained significant attention. Although existing models achieve some success, they rarely both model time and frequency characteristics in a pretraining-finetuning paradigm leading to suboptimal performance in predictions of complex time series, which requires both modeling periodicity and prior pattern knowledge of signals. We propose MoFE-Time, an innovative time series forecasting model that integrates time and frequency domain features within a Mixture of Experts (MoE) network. Moreover, we use the pretraining-finetuning paradigm as our training framework to effectively transfer prior pattern knowledge across pretraining and finetuning datasets with different periodicity distributions. Our method introduces both frequency and time cells as experts after attention modules and leverages the MoE routing mechanism to construct multidimensional sparse representations of input signals. In experiments on six public benchmarks, MoFE-Time has achieved new state-of-the-art performance, reducing MSE and MAE by 6.95% and 6.02% compared to the representative methods Time-MoE. Beyond the existing evaluation benchmarks, we have developed a proprietary dataset, NEV-sales, derived from real-world business scenarios. Our method achieves outstanding results on this dataset, underscoring the effectiveness of the MoFE-Time model in practical commercial applications.

Figures

Figures reproduced from arXiv: 2507.06502 by the authors.

Figure 1
Figure 1. Proposed architecture. Our network is based on the successful Time-MoE[ [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Details of the FTC module. We incorporate a Frequency-Time Cell (FTC) within the MoE framework as a specialized [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Time series forecasting cases from ETTh1 by different models with context-predict length = 512-96. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The ablation study of pretrain , RevIN and FTC in MoFE-Time. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The visualization of frequency features between MoFE-Time and MoFE-Time w/o FTC. In the first column, we [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of Inference Speed per Data Point be [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 15 canonical work pages

  1. [1]

    Rafal A Angryk, Petrus C Martens, Berkay Aydin, Dustin Kempton, Sushant S Mahajan, Sunitha Basodi, Azim Ahmadzadeh, Xumin Cai, Soukaina Fi- lali Boubrahimi, Shah Muhammad Hamdi, et al. 2020. Multivariate time series dataset for space weather data analytics. Scientific data 7, 1 (2020), 227

    work page 2020

  2. [2]

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebas- tian Pineda Arango, Shubham Kapoor, et al. 2024. Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815 (2024)

    arXiv 2024

  3. [3]

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung. 2015. Time series analysis: forecasting and control . John Wiley & Sons

    2015

  4. [4]

    Ching Chang, Wei-Yao Wang, Wen-Chih Peng, and Tien-Fu Chen. 2025. LLM4TS: Aligning Pre-Trained LLMs as Data-Efficient Time-Series Forecasters.ACM Trans. Intell. Syst. Technol. (Feb. 2025). https://doi.org/10.1145/3719207 Just Accepted

    doi:10.1145/3719207 2025

  5. [5]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2023. A decoder-only foundation model for time-series forecasting. arXiv preprint arXiv:2310.10688 (2023)

    arXiv 2023

  6. [6]

    Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. 2019. Unified language model pre- training for natural language understanding and generation. Advances in neural information processing systems 32 (2019)

    work page 2019

  7. [7]

    Javier Duoandikoetxea. 2024. Fourier analysis. Vol. 29. American Mathematical Society

    work page 2024

  8. [8]

    Jeffrey L Elman. 1990. Finding structure in time. Cognitive science 14, 2 (1990), 179–211

    1990

Show all 37 references
  1. [9]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. 2024. Moment: A family of open time-series foundation models. arXiv preprint arXiv:2402.03885 (2024)

    2024 arXiv

  2. [10]

    Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory.Neural computation 9, 8 (1997), 1735–1780

    1997

  3. [11]

    Ming Jin, Qingsong Wen, Yuxuan Liang, Chaoli Zhang, Siqiao Xue, Xue Wang, James Zhang, Yi Wang, Haifeng Chen, Xiaoli Li, et al. 2023. Large models for time series and spatio-temporal data: A survey and outlook. arXiv preprint arXiv:2310.10196 (2023)

    2023 arXiv

  4. [12]

    Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. 2021. Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift. In International Conference on Learning Representations. https://openreview.net/forum?id=...

    2021

  5. [13]

    Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. 2017. Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks. CoRR abs/1703.07015 (2017). arXiv:1703.07015 http://arxiv.org/abs/1703.07015

    2017 arXiv

  6. [14]

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. 2024. Foundation models for time series analysis: A tutorial and survey. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining . 6555–6565

    2024

  7. [15]

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Cheng- gang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  8. [16]

    Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zhao. 2024. Timecma: Towards llm-empowered time series forecasting via cross-modality alignment. arXiv preprint arXiv:2406.01638 (2024)

    2024 arXiv

  9. [17]

    Shengzhong Mao, Chaoli Zhang, Yichi Song, Jindong Wang, Xiao-Jun Zeng, Zenglin Xu, and Qingsong Wen. 2024. Time series analysis for education: Meth- ods, applications, and future directions. arXiv preprint arXiv:2408.13960 (2024)

    2024 arXiv

  10. [18]

    Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam

    Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In International Conference on Learning Representations

    2023

  11. [19]

    Yuqi Nie, Yaxuan Kong, Xiaowen Dong, John M Mulvey, H Vincent Poor, Qing- song Wen, and Stefan Zohren. 2024. A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges. arXiv preprint arXiv:2406.11903 (2024)

    2024 arXiv

  12. [20]

    Alan V Oppenheim. 1999. Discrete-time signal processing . Pearson Education India

    1999

  13. [21]

    Kashif Rasul, Arjun Ashok, Andrew Robert Williams, Arian Khorasani, George Adamopoulos, Rishika Bhagwatkar, Marin Biloš, Hena Ghonia, Nadhir Hassen, Anderson Schneider, et al. 2023. Lag-llama: Towards foundation models for time series forecasting. In R0-FoMo: Robustness of Few...

    2023

  14. [22]

    Javier Selva, Anders S Johansen, Sergio Escalera, Kamal Nasrollahi, Thomas B Moeslund, and Albert Clapés. 2023. Video transformers: A survey. IEEE Transac- tions on Pattern Analysis and Machine Intelligence 45, 11 (2023), 12922–12943

    2023

  15. [23]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2024. Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts. arXiv preprint arXiv:2409.16040 (2024)

    2024 arXiv

  16. [24]

    Steven W Smith et al. 1997. The scientist and engineer’s guide to digital signal processing

    1997

  17. [25]

    1971.Introduction to Fourier analysis on Euclidean spaces

    Elias M Stein and Guido Weiss. 1971.Introduction to Fourier analysis on Euclidean spaces. Vol. 1. Princeton university press

    1971

  18. [26]

    Shiyu Wang, Jiawei Li, Xiaoming Shi, Zhou Ye, Baichuan Mo, Wenze Lin, Shengtong Ju, Zhixuan Chu, and Ming Jin. 2024. TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis. arXiv preprint arXiv:2410.16032 (2024)

    2024 arXiv

  19. [27]

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Guo Qin, Haoran Zhang, Yong Liu, Yunzhong Qiu, Jianmin Wang, and Mingsheng Long. 2024. TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables. In The Thirty-eighth Annual Conference on Neural Information Proc...

    2024

  20. [28]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. arXiv:2210.02186 [cs.LG] https://arxiv.org/abs/2210.02186

    2023 arXiv

  21. [29]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. In Advances in Neural Information Processing Systems , A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan (Eds.)

    2021

  22. [30]

    Zhijian Xu, Ailing Zeng, and Qiang Xu. 2023. FITS: Modeling time series with 10𝑘 parameters. arXiv preprint arXiv:2307.03756 (2023)

    2023 arXiv

  23. [31]

    Yiyuan Yang, Chaoli Zhang, Tian Zhou, Qingsong Wen, and Liang Sun. 2023. Dcdetector: Dual attention contrastive representation learning for time series anomaly detection. In Proceedings of the 29th ACM SIGKDD Conference on Knowl- edge Discovery and Data Mining . 3033–3045

    2023

  24. [32]

    Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. 2022. Ts2vec: Towards universal representation of time series. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 36. 8980–8987

    2022

  25. [33]

    George Zerveas, Srideepika Jayaraman, Dhaval Patel, Anuradha Bhamidipaty, and Carsten Eickhoff. 2021. A transformer-based framework for multivariate time series representation learning. In Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining . 2114–2124

    2021

  26. [34]

    Kexin Zhang, Qingsong Wen, Chaoli Zhang, Rongyao Cai, Ming Jin, Yong Liu, James Y Zhang, Yuxuan Liang, Guansong Pang, Dongjin Song, et al. 2024. Self- supervised learning for time series analysis: Taxonomy, progress, and prospects. IEEE Transactions on Pattern Analysis and Mac...

    2024

  27. [35]

    Xiang Zhang, Ziyuan Zhao, Theodoros Tsiligkaridis, and Marinka Zitnik. 2022. Self-supervised contrastive pre-training for time series via time-frequency consis- tency. Advances in Neural Information Processing Systems 35 (2022), 3988–4003

    2022

  28. [36]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond efficient transformer for long se- quence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115

    2021

  29. [37]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International conference on machine learning. PMLR, 27268–27286

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.