pith. sign in

arxiv: 2605.26562 · v1 · pith:IOZ7AA5Dnew · submitted 2026-05-26 · 💻 cs.LG

Beyond Holistic Models: Systematic Component-level Benchmarking of Deep Multivariate Time-Series Forecasting

Shuang Liang , Chaochuan Hou , Xu Yao , Shiping Wang , Hailiang Huang , Songqiao Han , Minqi Jiang This is my paper

Pith reviewed 2026-06-29 19:05 UTC · model grok-4.3

classification 💻 cs.LG
keywords multivariate time series forecastingcomponent-level benchmarkingperformance corpusautomated component selectiondeep learningorthogonal experimental designmodel decomposition
0
0 comments X

The pith

Systematic component selection from a performance corpus outperforms state-of-the-art holistic models in multivariate time series forecasting.

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

This paper shifts attention from ever-more-complex forecasting models to understanding their individual building blocks. It introduces TSCOMP, a benchmark that decomposes methods into fine-grained components such as series preprocessing, encoding strategies, network architectures, and optimization methods. Using constrained orthogonal experiments across many datasets, the authors assemble a corpus of over 20,000 evaluations that supports automated component selection for new data. Experiments show that this simple corpus-driven selection consistently beats existing top models.

Core claim

Deconstructing deep multivariate time-series forecasting methods into core components and using the resulting empirical performance corpus for automated selection produces models that outperform state-of-the-art holistic architectures, as shown by multi-view analyses of component effectiveness, interactions, and data characteristics.

What carries the argument

TSCOMP benchmark that applies constrained orthogonal experimental design to evaluate components and build a reusable performance corpus for zero-shot model construction.

If this is right

  • Component effectiveness varies across different model backbones and data characteristics.
  • Interactions among components become measurable through the orthogonal design.
  • Zero-shot construction of competitive models on unseen datasets is enabled by the corpus.
  • Systematic component selection surpasses manually designed complex architectures.

Where Pith is reading between the lines

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

  • Future model development may benefit from treating forecasting systems as composable modules rather than fixed end-to-end designs.
  • The benchmarking approach could be adapted to evaluate components in other sequence modeling domains.
  • Corpus quality depends on dataset diversity, so expanding the collection of evaluation datasets would strengthen the method.

Load-bearing premise

The chosen components are sufficiently independent and the experimental design captures relevant interactions without systematic bias from the selected datasets or component definitions.

What would settle it

A new multivariate time series dataset on which the corpus-driven selection method performs worse than a current state-of-the-art holistic model.

Figures

Figures reproduced from arXiv: 2605.26562 by Chaochuan Hou, Hailiang Huang, Minqi Jiang, Shiping Wang, Shuang Liang, Songqiao Han, Xu Yao.

Figure 1
Figure 1. Figure 1: Overview of the proposed TSCOMP framework. TSCOMP deconstructs existing SOTA models into a modular component [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Component performance distributions (standard [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Component adaptability to data characteristics. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Selection qual￾ity distribution of the meta-predictor’s top-5 recommendations across all evaluation tasks. (i) Simple models excel with sufficient samples (Fig. 4a, Tab. 5a): MLPs gain substantial performance on longer datasets as increased sample availability allows task-specific convergence to reliable patterns, whereas large models struggle, suggesting that excessive downstream adaptation may override p… view at source ↗
Figure 6
Figure 6. Figure 6: TSCOMP pipeline framework. This diagram shows the component-based design of TSCOMP. The final TSCOMP structure is formed by combining different component op￾tions [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Dimension Importance (Effect Range) (Effect Range Plots). This figure visualizes the performance distributions across [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Pipeline Stage Importance (Pipeline Importance Plots). This figure visualizes the performance distributions across [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Performance Distributions for Series Normalization (Ridgeline Plots). This figure visualizes the performance distribu [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Dataset Adaptability (Radar Charts) for Series Normalization (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Performance Distributions for Series Decomposition (Ridgeline Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Dataset Adaptability (Radar Charts) for Series Decomposition (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p018_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: Dataset Adaptability (Radar Charts) for Series Sampling/Mixing (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p018_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Performance Distributions for Channel Independence (Ridgeline Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Dataset Adaptability (Radar Charts) for Channel Independence (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 18
Figure 18. Figure 18: Dataset Adaptability (Radar Charts) for Timestamp Embedding (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p019_18.png] view at source ↗
Figure 20
Figure 20. Figure 20: Dataset Adaptability (Radar Charts) for Series Tokenization (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p020_20.png] view at source ↗
Figure 22
Figure 22. Figure 22: Dataset Adaptability (Radar Charts) for Network [PITH_FULL_IMAGE:figures/full_fig_p020_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Performance Distributions for Feature Attention Mechanisms (Ridgeline Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p020_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Dataset Adaptability (Radar Charts) for Feature Attention Mechanisms (Radar Plots). This figure visualizes the [PITH_FULL_IMAGE:figures/full_fig_p020_24.png] view at source ↗
Figure 26
Figure 26. Figure 26: Dataset Adaptability (Radar Charts) for Retrieval Augmented Generation (Radar Plots). This figure visualizes the [PITH_FULL_IMAGE:figures/full_fig_p021_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Performance Distributions for Sequence Length Configurations (Ridgeline Plots). This figure visualizes the perfor [PITH_FULL_IMAGE:figures/full_fig_p021_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Dataset Adaptability (Radar Charts) for Sequence Length Configurations (Radar Plots). This figure visualizes the [PITH_FULL_IMAGE:figures/full_fig_p021_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Performance Distributions for Loss Functions (Ridgeline Plots). This figure visualizes the performance distributions [PITH_FULL_IMAGE:figures/full_fig_p021_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Dataset Adaptability (Radar Charts) for Loss Functions (Radar Plots). This figure visualizes the performance [PITH_FULL_IMAGE:figures/full_fig_p021_30.png] view at source ↗
read the original abstract

While previous research in multivariate time series forecasting has focused on developing complex holistic models, this work advocates for a shift toward a granular, component-level understanding of their impacts. We propose TSCOMP, the first large-scale benchmark that systematically deconstructs deep forecasting methods into their core, fine-grained components--spanning series preprocessing, encoding strategies, network architectures including specific and large time-series models, and optimization methods. Using constrained orthogonal experimental design and extensive evaluations, we conduct multi-view analyses that reveal component effectiveness across different backbones, data characteristics, and their interactions. Beyond providing insights, this benchmark establishes a fine-grained performance corpus comprising over 20,000 model-dataset evaluations, which supports the learning of automated component selection, enabling zero-shot model construction on new datasets. Our experiments demonstrate that the corpus-driven approach, despite its simplicity, consistently outperforms state-of-the-art methods, validating the soundness of our evaluation design and confirming that systematic component selection surpasses manually designed complex architectures. All code and the performance corpus are publicly available at https://github.com/SUFE-AILAB/TSCOMP.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces TSCOMP, the first large-scale benchmark deconstructing deep multivariate time-series forecasting models into fine-grained components (series preprocessing, encoding strategies, network architectures including specific and large TS models, and optimization methods). Using constrained orthogonal experimental design, it conducts over 20,000 model-dataset evaluations for multi-view analyses of component effectiveness and interactions, builds a performance corpus, and shows that automated component selection from the corpus enables zero-shot construction of models that outperform state-of-the-art holistic methods on new datasets. Code and corpus are released publicly.

Significance. If the central empirical claims hold after addressing methodological details, the work could meaningfully shift the field from end-to-end holistic model design toward systematic, corpus-driven component selection, offering both practical gains in forecasting performance and new insights into component interactions across data regimes. The public release of the 20k+ evaluation corpus and code is a clear strength that supports reproducibility and follow-on research.

major comments (3)
  1. [Abstract] Abstract: the claim that the corpus-driven approach 'consistently outperforms state-of-the-art methods' is load-bearing for the central contribution, yet the abstract (and by extension the manuscript) provides no details on component definitions, the precise values or ranges tested in each category, the statistical testing protocol, or controls against post-hoc selection bias, so the outperformance result cannot be verified.
  2. [Abstract] Abstract (experimental design paragraph): the constrained orthogonal design is asserted to capture interactions and support reliable zero-shot selection, but no evidence or validation is given that the design accounts for higher-order interactions (e.g., specific normalization paired with temporal encoders or large TS models) that may be dataset-dependent; if such interactions exist, the learned selector's predictions will deviate from actual performance on unseen data.
  3. [Abstract / Experiments] The manuscript does not report the number of datasets, the exact component enumeration, or the learning procedure for the automated selector (e.g., which regression or ranking model is trained on the corpus), all of which are required to assess whether the 20k evaluations suffice to generalize beyond the corpus.
minor comments (2)
  1. [Abstract] The GitHub link for code and corpus is a positive step for reproducibility; ensure the released materials include the exact component definitions and experimental scripts used to generate the corpus.
  2. [Abstract] Clarify the phrasing 'specific and large time-series models' when listing architecture components to avoid ambiguity in the taxonomy.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and indicate where revisions will be made to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the corpus-driven approach 'consistently outperforms state-of-the-art methods' is load-bearing for the central contribution, yet the abstract (and by extension the manuscript) provides no details on component definitions, the precise values or ranges tested in each category, the statistical testing protocol, or controls against post-hoc selection bias, so the outperformance result cannot be verified.

    Authors: We agree that the abstract is too high-level and does not supply these specifics, which hinders immediate verification of the central claim. We will revise the abstract to briefly describe the four component categories, note the use of statistical significance testing, and reference the orthogonal design as a safeguard against selection bias. Full definitions and ranges will remain in the methods section but will be cross-referenced from the abstract. revision: yes

  2. Referee: [Abstract] Abstract (experimental design paragraph): the constrained orthogonal design is asserted to capture interactions and support reliable zero-shot selection, but no evidence or validation is given that the design accounts for higher-order interactions (e.g., specific normalization paired with temporal encoders or large TS models) that may be dataset-dependent; if such interactions exist, the learned selector's predictions will deviate from actual performance on unseen data.

    Authors: The constrained orthogonal design is intended to support estimation of main effects and selected two-factor interactions while remaining computationally tractable; the multi-view analyses in the paper examine several such interactions across data regimes. The empirical success of the selector on held-out datasets provides indirect support for generalization. We acknowledge, however, that an explicit check for higher-order interactions on unseen data is not reported, and we will add a short discussion of this limitation together with any additional validation that can be performed within the existing corpus. revision: partial

  3. Referee: [Abstract / Experiments] The manuscript does not report the number of datasets, the exact component enumeration, or the learning procedure for the automated selector (e.g., which regression or ranking model is trained on the corpus), all of which are required to assess whether the 20k evaluations suffice to generalize beyond the corpus.

    Authors: This observation is correct; these quantities are not stated with sufficient clarity or prominence. We will add an explicit summary (including the number of datasets, a table or enumeration of all tested components, and the precise training procedure for the selector) in a new or expanded subsection of the experimental setup. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical corpus built from independent evaluations; outperformance claim tested on new datasets

full rationale

The paper constructs a performance corpus via 20k+ evaluations under constrained orthogonal design, then shows corpus-driven component selection outperforms SOTA on unseen datasets. This is a standard empirical pipeline with no self-definitional reductions (no X defined in terms of Y that is then 'predicted'), no fitted parameters renamed as predictions, and no load-bearing self-citations or uniqueness theorems. The central result is externally falsifiable against prior models and held-out data, satisfying the criteria for a self-contained derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that forecasting models can be meaningfully decomposed into independent components whose effects can be isolated via orthogonal experimental design. No free parameters or invented entities are introduced in the abstract; the work is purely empirical benchmarking.

axioms (1)
  • domain assumption Forecasting models can be decomposed into separable components (preprocessing, encoding, architecture, optimization) whose individual contributions can be measured independently.
    Invoked in the description of TSCOMP and the experimental design; if components interact strongly in ways the design does not capture, the corpus and selection method lose validity.

pith-pipeline@v0.9.1-grok · 5738 in / 1279 out tokens · 37403 ms · 2026-06-29T19:05:33.154794+00:00 · methodology

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

75 extracted references · 7 canonical work pages · 3 internal anchors

  1. [1]

    Mustafa Abdallah, Ryan Rossi, Kanak Mahadik, Sungchul Kim, Handong Zhao, and Saurabh Bagchi. 2022. Autoforecast: Automatic time-series forecasting model selection. InProceedings of the 31st ACM International Conference on Information & Knowledge Management. 5–14

    2022

  2. [2]

    2009.Statistical methods for forecasting

    Bovas Abraham and Johannes Ledolter. 2009.Statistical methods for forecasting. John Wiley & Sons

    2009

  3. [3]

    Francisco Martinez Alvarez, Alicia Troncoso, Jose C Riquelme, and Jesus S Aguilar Ruiz. 2010. Energy time series forecasting based on pattern sequence similarity. IEEE Transactions on Knowledge and Data Engineering23, 8 (2010), 1230–1243

    2010

  4. [4]

    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 Lan- guage of Time Series.Transactions on Machine Learning Research2024 (2024)

    2024

  5. [5]

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  6. [6]

    Marília Barandas, Duarte Folgado, Letícia Fernandes, Sara Santos, Mariana Abreu, Patrícia Bota, Hui Liu, Tanja Schultz, and Hugo Gamboa. 2020. TSFEL: Time Series Feature Extraction Library.SoftwareX11 (2020), 100456

    2020

  7. [7]

    C Bui, N Pham, A Vo, A Tran, A Nguyen, and T Le. 2018. Time series forecasting for healthcare diagnosis and prognostics with the focus on cardiovascular diseases. In6th International Conference on the Development of Biomedical Engineering in Vietnam (BME6) 6. Springer, 809–818

    2018

  8. [8]

    Colin Catlin. 2020. AutoTS: Automated Time Series Forecasting for Python. https://github.com/winedarksea/AutoTS

    2020

  9. [9]

    Ching Chang, Wen-Chih Peng, and Tien-Fu Chen. 2023. Llm4ts: Two-stage fine-tuning for time-series forecasting with pre-trained llms.CoRR(2023)

    2023

  10. [10]

    Peng Chen, Yingying ZHANG, Yunyao Cheng, Yang Shu, Yihang Wang, Qingsong Wen, Bin Yang, and Chenjuan Guo. 2024. Pathformer: Multi-scale Transformers with Adaptive Pathways for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    2024

  11. [11]

    Si-An Chen, Chun-Liang Li, Sercan O Arik, Nathanael Christian Yoder, and Tomas Pfister. 2023. TSMixer: An All-MLP Architecture for Time Series Forecast-ing. Transactions on Machine Learning Research(2023)

    2023

  12. [12]

    Razvan-Gabriel Cirstea, Bin Yang, Chenjuan Guo, Tung Kieu, and Shirui Pan

  13. [13]

    In2022 IEEE 38th International Conference on Data Engineering (ICDE)

    Towards spatio-temporal aware traffic time series forecasting. In2022 IEEE 38th International Conference on Data Engineering (ICDE). IEEE, 2900–2913

  14. [14]

    Chirag Deb, Fan Zhang, Junjing Yang, Siew Eang Lee, and Kwok Wei Shah. 2017. A review on time series forecasting techniques for building energy consumption. Renewable and Sustainable Energy Reviews74 (2017), 902–924

    2017

  15. [15]

    Wei Fan, Pengyang Wang, Dongkun Wang, Dongjie Wang, Yuanchun Zhou, and Yanjie Fu. 2023. Dish-ts: a general paradigm for alleviating distribution shift in time series forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 7522–7529

    2023

  16. [16]

    Matthias Feurer, Aaron Klein, Katharina Eggensperger, Jost Springenberg, Manuel Blum, and Frank Hutter. 2015. Efficient and Robust Automated Machine Learning. InAdvances in Neural Information Processing Systems 28 (2015). 2962–2970

    2015

  17. [17]

    Raphael Fischer and Amal Saadallah. 2024. AutoXPCR: Automated multi- objective model selection for time series forecasting. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 806–815

    2024

  18. [18]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. 2024. MOMENT: A Family of Open Time-series Foundation Models. InForty-first International Conference on Machine Learning

    2024

  19. [19]

    Albert Gu and Tri Dao. 2023. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752(2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  20. [20]

    Sungwon Han, Seungeon Lee, Meeyoung Cha, Sercan O Arik, and Jinsung Yoon

  21. [21]

    InForty-second International Conference on Machine Learning

    Retrieval Augmented Time Series Forecasting. InForty-second International Conference on Machine Learning

  22. [22]

    Anggit Dwi Hartanto, Yanuar Nur Kholik, and Yoga Pristyanto. 2023. Stock price time series data forecasting using the light gradient boosting machine (LightGBM) model.JOIV: International Journal on Informatics Visualization7, 4 (2023), 2270–2279

    2023

  23. [23]

    Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. 2023. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. InThe Eleventh International Conference on Learning Representations

    2023

  24. [24]

    Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and Qingsong Wen. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. InThe Twelfth International Conference on Learning Representations

    2024

  25. [25]

    Shruti Kaushik, Abhinav Choudhury, Pankaj Kumar Sheron, Nataraj Dasgupta, Sayee Natarajan, Larry A Pickett, and Varun Dutt. 2020. AI in healthcare: time- series forecasting using statistical, neural, and ensemble architectures.Frontiers in big data3 (2020), 4

    2020

  26. [27]

    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. InInternational conference on learning representations

    2021

  27. [28]

    Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. 2020. Reformer: The efficient transformer.arXiv preprint arXiv:2001.04451(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  28. [29]

    Josef Kraus et al . 2025. xLSTM-Mixer: Long Short-Term Memory Mixing for Time Series Forecasting. InICLR

    2025

  29. [30]

    Dilfira Kudrat, Zongxia Xie, Yanru Sun, Tianyu Jia, and Qinghua Hu. 2025. Patch- wise Structural Loss for Time Series Forecasting. InInternational Conference on Machine Learning. PMLR, 31841–31859. KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Shuang Liang et al

    2025

  30. [31]

    Shiyang Li, Xiaoyong Jin, Yao Xuan, Xiyou Zhou, Wenhu Chen, Yu-Xiang Wang, and Xifeng Yan. 2019. Enhancing the locality and breaking the memory bottle- neck of transformer on time series forecasting.Advances in neural information processing systems32 (2019)

    2019

  31. [32]

    Shengsheng Lin, Weiwei Lin, Wentai Wu, Feiyu Zhao, Ruichao Mo, and Haotong Zhang. 2025. Segrnn: Segment recurrent neural network for long-term time series forecasting.IEEE Internet of Things Journal(2025)

    2025

  32. [33]

    Minhao Liu, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, and Qiang Xu. 2022. Scinet: Time series modeling and forecasting with sample convolution and interaction.Advances in Neural Information Processing Systems 35 (2022), 5816–5828

    2022

  33. [34]

    Liu, and Schahram Dustdar

    Shizhan Liu, Hang Yu, Cong Liao, Jianguo Li, Weiyao Lin, Alex X. Liu, and Schahram Dustdar. 2022. Pyraformer: Low-Complexity Pyramidal Attention for Long-Range Time Series Modeling and Forecasting. (2022)

    2022

  34. [35]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    2024

  35. [36]

    Yong Liu, Chenyu Li, Jianmin Wang, and Mingsheng Long. 2023. Koopa: Learning non-stationary time series dynamics with koopman predictors.Advances in neural information processing systems36 (2023), 12271–12290

    2023

  36. [37]

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary transformers: Exploring the stationarity in time series forecasting.Advances in neural information processing systems35 (2022), 9881–9893

    2022

  37. [38]

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary Transformers: Rethinking the Stationarity in Time Series Forecasting. InNeurIPS

    2022

  38. [39]

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024. Timer: Generative Pre-trained Transformers Are Large Time Series Models. InForty-first International Conference on Machine Learning

    2024

  39. [40]

    Zhining Liu, Ze Yang, Xiao Lin, Ruihong Qiu, Tianxin Wei, Yada Zhu, Hendrik Hamann, Jingrui He, and Hanghang Tong. 2025. Breaking Silos: Adaptive Model Fusion Unlocks Better Time Series Forecasting. InInternational Conference on Machine Learning

    2025

  40. [41]

    Ricardo P Masini, Marcelo C Medeiros, and Eduardo F Mendes. 2023. Machine learning advances for time series forecasting.Journal of economic surveys37, 1 (2023), 76–111

    2023

  41. [42]

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

    2023

  42. [43]

    Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S Jensen, Zhenli Sheng, et al. 2024. TFB: Towards Comprehensive and Fair Benchmarking of Time Series Forecasting Methods.Proceedings of the VLDB Endowment17, 9 (2024), 2363–2377

    2024

  43. [44]

    Xiangfei Qiu, Xingjian Wu, Hanyin Cheng, Xvyuan Liu, Chenjuan Guo, Jilin Hu, and Bin Yang. 2026. Dbloss: Decomposition-based loss function for time series forecasting.Advances in Neural Information Processing Systems38 (2026), 27741–27768

    2026

  44. [45]

    Xiangfei Qiu, Xingjian Wu, Yan Lin, Chenjuan Guo, Jilin Hu, and Bin Yang. 2025. DUET: Dual Clustering Enhanced Multivariate Time Series Forecasting(KDD ’25). Association for Computing Machinery, New York, NY, USA, 1185–1196. doi:10.1145/3690624.3709325

    work page doi:10.1145/3690624.3709325 2025

  45. [46]

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. 2020. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019.Applied soft computing90 (2020), 106181

    2020

  46. [47]

    Zezhi Shao, Fei Wang, Yongjun Xu, Wei Wei, Chengqing Yu, Zhao Zhang, Di Yao, Tao Sun, Guangyin Jin, Xin Cao, et al. 2024. Exploring progress in multi- variate time series forecasting: Comprehensive benchmarking and heterogeneity analysis.IEEE Transactions on Knowledge and Data Engineering(2024)

    2024

  47. [48]

    Oleksandr Shchur, Ali Caner Turkmen, Nick Erickson, Huibin Shen, Alexander Shirkov, Tony Hu, and Bernie Wang. 2023. AutoGluon–TimeSeries: AutoML for probabilistic time series forecasting. InInternational Conference on Automated Machine Learning. PMLR, 9–1

    2023

  48. [49]

    Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. 2025. Time-moe: Billion-scale time series foundation models with mix- ture of experts. InInternational conference on learning representations, Vol. 2025. 34635–34667

    2025

  49. [50]

    Mingtian Tan, Mike Merrill, Vinayak Gupta, Tim Althoff, and Tom Hartvigsen

  50. [51]

    Are language models actually useful for time series forecasting?Advances in Neural Information Processing Systems37 (2024), 60162–60191

    2024

  51. [52]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  52. [53]

    Hao Wang, Lichen Pan, Yuan Shen, Zhichao Chen, Degui Yang, Yifei Yang, Sen Zhang, Xinggao Liu, Haoxuan Li, and Dacheng Tao. 2025. Fredf: Learning to forecast in the frequency domain. InInternational Conference on Learning Representations, Vol. 2025. 6893–6922

    2025

  53. [54]

    Huiqiang Wang, Jian Peng, Feihu Huang, Jince Wang, Junhui Chen, and Yifei Xiao

  54. [55]

    InThe eleventh international conference on learning representations

    Micn: Multi-scale local and global context modeling for long-term series forecasting. InThe eleventh international conference on learning representations

  55. [56]

    Zhang, and JUN ZHOU

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y. Zhang, and JUN ZHOU. 2024. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. InThe Twelfth International Conference on Learning Representations

    2024

  56. [57]

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Chen Wang, Mingsheng Long, and Jianmin Wang. 2026. Deep time series models: A comprehensive survey and benchmark.IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    2026

  57. [58]

    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. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  58. [59]

    Qingsong Wen, Liang Sun, Fan Yang, Xiaomin Song, Jingkun Gao, Xue Wang, and Huan Xu. 2021. Time Series Data Augmentation for Deep Learning: A Survey. In Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI). 4653–4660

    2021

  59. [60]

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. 2023. Transformers in time series: a survey. InProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence. 6778–6786

    2023

  60. [61]

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven C. H. Hoi. 2022. ETSformer: Exponential Smoothing Transformers for Time-series Forecasting.arXiv preprint arXiv:2202.01381(2022)

    work page arXiv 2022

  61. [62]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2023. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InThe Eleventh International Conference on Learning Representations

    2023

  62. [63]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Fore- casting. InNeurIPS

    2021

  63. [64]

    Peter T Yamak, Li Yujian, and Pius K Gadosey. 2019. A comparison between arima, lstm, and gru for time series forecasting. InProceedings of the 2019 2nd international conference on algorithms, computing and artificial intelligence. 49–55

    2019

  64. [65]

    Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, and Zhendong Niu. 2023. Frequency-domain mlps are more effective learners in time series forecasting.Advances in Neural Information Processing Systems36 (2023), 76656–76679

    2023

  65. [66]

    Yi Yin and Pengjian Shang. 2016. Forecasting traffic time series with multivariate predicting method.Appl. Math. Comput.291 (2016), 266–278

    2016

  66. [67]

    Wenzhen Yue et al. 2025. OLinear: A Linear Model for Time Series Forecasting in Orthogonally Transformed Domain. InNeurIPS

    2025

  67. [68]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting?. InProceedings of the AAAI conference on artificial intelligence, Vol. 37. 11121–11128

    2023

  68. [69]

    G Peter Zhang. 2003. Time series forecasting using a hybrid ARIMA and neural network model.Neurocomputing50 (2003), 159–175

    2003

  69. [70]

    Tianping Zhang, Yizhuo Zhang, Wei Cao, Jiang Bian, Xiaohan Yi, Shun Zheng, and Jian Li. 2022. Less Is More: Fast Multivariate Time Series Forecasting with Light Sampling-oriented MLP Structures. arXiv:2207.01186 [cs.LG] https://arxiv. org/abs/2207.01186

    work page arXiv 2022

  70. [71]

    Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. InThe eleventh international conference on learning representations

    2023

  71. [72]

    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. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115

    2021

  72. [73]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, Rong Jin, et al. 2022. Film: Frequency improved legendre memory model for long-term time series forecasting.Advances in neural information processing systems35 (2022), 12677–12690

    2022

  73. [74]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. 2022. FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. InICML

    2022

  74. [75]

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al . 2023. One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems36 (2023), 43322–43355. A Datasets We conduct extensive evaluations on 13 standard long-term fore- casting benchmarks: four ETT variants (ETTh1, ETTh2, ETTm1, ETTm2), Electricity (abb...

    2023

  75. [76]

    Table 18: Full results for the short-term forecasting task in the M4 dataset

    [34] [11] [61] [32] [36] [57] [70] [31] [66] [59] [68] [26] [49] [69] Metric MSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAEMSE MAE ETTh1 96 0.371 0.3960.403 0.4160.483 0.4930.396 0.4080.707 0.6310.534 0.4990.496 0.4810.377 0.4160.467 0.4570.449 0.4520.438 0.4500.920 0.7290.834 0.6640.852 0.7230.411 0.428...

    work page arXiv 2026