Pith. sign in

REVIEW 4 major objections 3 minor 57 references

Fourier Basis Mapping: A Time-Frequency Learning Framework for Time Series Forecasting

T0 review · 4 major / 3 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 This paper claims that Fourier basis expansion — reading the DFT's real and imaginary parts as cosine and sine basis coefficients — produces time-frequency features that improve forecasting accuracy across linear, MLP, and Transformer…

desk verdict A solid extension of the authors' NeurIPS Fourier-basis work: the new synergetic model and ablations are worth refereeing, but the SOTA claim rests on single-run numbers that need multi-seed backing. read the letter →

arxiv 2507.09445 v2 pith:N2ZBGEPV submitted 2025-07-13 cs.LG cs.AIstat.ML

Runze Yang , Longbing Cao , Xin You , Kun Fang , Jianxun Li , Jie Yang This is my paper
classification cs.LGcs.AIstat.ML
keywords timeseriesforecastingFourierbasisexpansiontime-frequencyfeaturesdiscretetransformlong-termshort-termdeeplearningmultivariate
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 argues that existing Fourier-based forecasters misinterpret the discrete Fourier transform by using bare real and imaginary spectrum coefficients without the basis functions they belong to. It identifies two resulting failure modes: inconsistent starting cycles, where a phase-shifted signal cannot be mapped without amplitude and phase information, and inconsistent series length, where a frequency's meaning changes with the look-back window. The proposed Fourier Basis Mapping (FBM) multiplies the real part of the spectrum by cosine basis functions and the imaginary part by sine basis functions, producing time-frequency features that retain temporal position. The authors report that these features improve linear, MLP-based, and Transformer-based models as a plug-in, and that a synergetic variant FBM-S achieves state-of-the-art results on long and short-term forecasting benchmarks.

What carries the argument

The central object is the Fourier basis expansion: for a look-back window of length $T$, the paper builds cosine basis $C$ and sine basis $S$ as in Eq. (4), then forms time-frequency features by multiplying the real part of the DFT spectrum by $C$ and the imaginary part by $S$ and summing, so each frequency level is expanded into time-indexed sinusoidal waveforms. Because the resulting features carry a time index as well as a frequency level, the downstream mapping network operates in the time-frequency space rather than in the frequency-only space. The synergetic FBM-S model then assigns three blocks to those features: a seasonal block using a rolling window over the sinusoidal basis, a trend block using patching, centralization, and multi-scale down-sampling, and an interaction block that masks the time-frequency input to recent windows in order to capture short-lived cross-variate effects.

What would settle it

Run FBM-S and the strongest baselines (NLinear, PatchTST, TimeMixer, and iTransformer) multiple times with different seeds on the Electricity and PEMS08 benchmarks, and check whether the reported MSE advantages, some as small as a few percent relative error, survive when compared with the spread across seeds; if the confidence intervals overlap, the headline performance claim would not hold.

Watch

Extended reading notes

Core claim

The central claim is that the real and imaginary parts of DFT coefficients are best understood as coefficients of cosine and sine basis functions, and that mapping in the frequency space alone is insufficient for forecasting. The paper proves that the IDFT of a real-valued signal decomposes into $T/2+1$ cosine and $T/2-1$ sine basis functions, so each spectrum coefficient carries meaning only through its basis function; the amplitude and arctangent, not the raw real/imaginary values, encode the interpretable cycle. From this perspective, FBM constructs time-frequency features that preserve both global frequency structure and local temporal detail, and FBM-S decomposes trend, seasonal, and interaction effects into three specialized blocks that model these features separately. The reported experiments show consistent error reductions over existing time- and frequency-based methods across long-term and short-term datasets.

Load-bearing premise

The reported performance gaps are real: every MSE and MAE in Tables II, III, and IV comes from a single run without standard deviations or significance tests, so the state-of-the-art claim depends on those differences not being run-to-run noise.

Editorial extensions

If this is right

  • FBM-L outperforms NLinear on every dataset and horizon, indicating that even a single linear layer benefits from mapping in time-frequency space.
  • FBM-NL and FBM-NP improve over their MLP and Transformer counterparts, showing that deeper networks help when input features carry both time and frequency information.
  • FBM-S achieves the best reported accuracy on nearly all long-term and short-term benchmarks, with the interaction block and masking contributing most on high-granularity PEMS data.
  • Patching and multi-scale down-sampling reduce the initial projection cost by a factor related to the number of patches while improving accuracy, making the time-frequency representation efficient as well as accurate.
  • The paper's stated limitation is that the input sequence length should not be too short, because the number of meaningful Fourier basis functions is bounded by that length.

Reading between the lines

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

  • A testable extension of the basis interpretation is that forecast difficulty should depend on how the look-back window divides the frequency levels: lengths that are multiples of a dominant cycle should yield more interpretable features and better forecasts, which the paper's Case II only sketches.
  • The interaction masking principle suggests that cross-variate dependencies decay quickly in time; this could be probed on other multivariate datasets by measuring how forecast accuracy degrades as the interaction mask length grows beyond a short window.
  • Because the reported MSE and MAE values come from single runs, a multi-seed evaluation would reveal whether the headline gains, some as small as a few percent relative error, are robust to run-to-run variation; the paper does not provide this evidence.
  • The same Fourier basis expansion could be transferred to anomaly detection and classification, where preserving temporal position may help localize irregularities in time rather than only in frequency.
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 / 3 minor

Summary. The paper proposes Fourier Basis Mapping (FBM), a feature-construction scheme that multiplies the real and imaginary parts of DFT coefficients by cosine and sine basis functions to produce time-frequency features, then feeds these features into linear, MLP, or Transformer backbones. It presents three plug-and-play variants (FBM-L, FBM-NL, FBM-NP) and a synergetic architecture FBM-S with separate trend, seasonal, and interaction blocks, along with techniques such as masking, centralization, patching, and multi-scale down-sampling. The paper claims that existing Fourier-based methods suffer from inconsistent starting cycles and inconsistent series length issues, and that FBM variants achieve state-of-the-art forecasting performance on both long-term and short-term tasks across twelve datasets, including M4.

Significance. The core idea is well motivated and potentially useful: expressing the standard DFT in terms of cosine and sine basis functions is a clean representational device, and the plug-and-play aspect is attractive. The cosine/sine decomposition proof in the appendix is essentially correct modulo typos, and the paper provides a wide range of benchmarks, ablations, efficiency measurements, and a public code link. However, the central contribution is an empirical performance claim, and that claim currently rests on unreplicated point estimates and on comparisons whose protocol equivalence is not fully established. If the missing uncertainty quantification and protocol controls are added, the framework could be a solid contribution; as it stands, the SOTA claim is not yet supported at the required standard.

major comments (4)
  1. [Section V-C; Tables II, III, IV] All MSE/MAE numbers in Tables II-IV are single-run point estimates, with no standard deviations, no multiple seeds, and no significance tests; Section V-C in fact states 'We use the same random seed'. Several headline advantages are tiny, for example the M4 average SMAPE of 11.555 versus TimeMixer's 11.723 in Table IV, and the Electricity horizon-720 MSE of 0.195 that exactly ties iTransformer in Table II. Because the paper's SOTA claim is load-bearing, the authors should report mean and standard deviation over at least three seeds and provide significance tests (or at least identify which differences are stable across seeds).
  2. [Section V-C; Table XIV; Table IV footnote] Baseline protocol comparability is not established. The main experiments use a 0.65/0.15/0.2 train/validation/test split, yet Table XIV shows that FBM-S is uniformly better under a 0.7/0.1/0.2 split, e.g., Electricity horizon-96 MSE 0.124 versus 0.127 and Traffic horizon-96 MSE 0.346 versus 0.357. Table IV's footnote also states that some M4 results are directly copied from TimeMixer. The authors should either retrain all baselines under the identical split, tuning budget, and evaluation protocol, or explicitly justify why the copied numbers are protocol-equivalent. The sentence in Section VI-D that the authors 'will consider use a larger training set in future work' is an admission that the current split is not the one that gives their best results.
  3. [Section VI-B; Table II] The statement in Section VI-B that FBM-S 'achieves SOTA performance all the time' is contradicted by the paper's own Table II: on Exchange horizon-192, FBM-S has MSE 0.194 versus CrossGNN's 0.188; on Exchange horizon-720, FBM-S has MSE 0.963 versus CrossGNN's 0.931; and on ETTm2 horizon-336, FBM-S has MSE 0.273 versus TimeMixer's 0.272. The SOTA claim should be qualified to an average or a subset of settings, or these exceptions should be explained.
  4. [Section III, Case I] The claim that 'no mathematical solution exists to establish such a mapping' when the starting cycle differs is mathematically overstated. For a fixed frequency, the coefficient pair (a,b) transforms under a phase shift by a simple linear rotation, e.g., [a_Y; b_Y] = R(Delta_phi)[a_X; b_X], and indeed Eq. (3) itself exhibits that rotation. The genuine point is that coefficients are less interpretable without the basis functions, not that the mapping is impossible. This overstatement should be corrected because it is used to motivate the entire framework.
minor comments (3)
  1. [Appendix, Eq. (7)] In Eq. (7), the term 'cos(pi T)' should be 'cos(pi n)', and the notation for H[T-k] should be made consistent with the Hermitian-symmetry convention used earlier; the same typo-like issue appears in Eq. (5), where 'pi/2' should evidently be 'T/2'.
  2. [Section VI-D] There are several wording and typographical errors, including 'bankbone' for 'backbone' and 'may not has' for 'may not have'; these should be corrected in a final pass.
  3. [Table I and Section IV-C] The hyperparameters in Table I (h1, h2, h3, P, K, C1, C2, lr) are said to be explained in Section VI-C, but several of them are not defined precisely in the main text before they are used in the ablations; please add explicit definitions at first use.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FBM's time-frequency features are a self-contained linear encoding of the input, and the SOTA claim is tested against external baselines.

full rationale

The paper's claimed derivation chain is self-contained. The time-frequency features are constructed from the input by DFT followed by multiplication with cosine/sine basis functions (Eqs. (1)-(4), with the cosine/sine representation proved in the Appendix); summing the per-frequency components recovers X, so the representation is a linear encoding of the input, not a fitted target. The downstream FBM variants (FBM-L, FBM-NL, FBM-NP, FBM-S) are standard trainable mappings evaluated against external baselines in Tables II-IV; no parameter is fitted to a subset of the evaluation data and then reported as a predicted quantity. The self-citations are disclosed and non-load-bearing: [51] is cited for priority as the conference version, and [53] is cited only for the auxiliary timestamp block ('The timestamp block is directly copied from [53]'), which is not the source of the FBM claim. The lack of standard deviations and significance tests in Tables II-IV is a statistical-robustness concern about the SOTA claim, not evidence that any result reduces to its inputs by construction. The acknowledged limitation that short input lengths bound the meaningful Fourier basis functions is a stated scope restriction, not a circular step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central contribution is an empirical deep-learning architecture, so the main free parameters are per-dataset hyperparameters tuned by validation. The mathematical foundation is standard DFT mathematics. The domain assumption of separable trend, seasonal, and interaction effects, and the ad hoc assumption of short-lived interactions, are not independently verified beyond the paper's own ablations. No new physical or formal entities are introduced.

free parameters (5)
  • Per-dataset hyperparameters (h1, h2, h3, P, K, C1, C2, lr, multi-scale) = Varies: h1/h2 in 128 to 1440, h3 in 256 to 512, P=14, K=3 or 4, C1 in 24 to 96, C2 in 12 to 48, lr 1e-5 to 5e-4
    Listed in Table I and selected by validation performance and ablations (Tables VII to XIII), not derived from theory.
  • C1 input interaction mask length = 24 for PEMS, ECL, WTH, ETTm datasets
    Chosen from Table IX ablation on PEMS08, where C1=24 gave the best results; applied broadly to other datasets.
  • C2 output interaction mask length = Usually set to the forecast horizon L, or 24/48 for long-term datasets
    From Table X ablation on Electricity; best performance with small C2 values. The paper concludes interactions do not last long.
  • Patch count P = 14
    Section VI-C states 14 patches are used across FBM-NP and FBM-S trend block; chosen for efficiency and performance.
  • Learning rate per dataset = 1e-5 to 5e-4
    Table I. The paper states smaller LR is required to learn time-frequency features, but exact values are tuned per dataset.
assumptions (4)
  • standard math DFT/IDFT orthogonality and Hermitian symmetry for real-valued signals (Eq. 2 and Appendix A)
    Used to rewrite the IDFT as a sum of cosine and sine basis functions with coefficients given by the real and imaginary parts of the spectrum.
  • domain assumption Real-world time series exhibit separable trend, seasonal, and interaction effects aligned with Fourier frequencies (Section VI-G)
    Motivates the three-block decomposition of FBM-S. The paper provides descriptive frequency-spectrum plots but no statistical test of separability.
  • ad hoc to paper Cross-channel interaction effects occur mainly over short recent windows (Section IV-C, 'inspired by this observation')
    Justifies the C1 and C2 masks in the interaction block. The paper gives an intuitive traffic example but no independent evidence beyond its own ablations.
  • ad hoc to paper The inconsistent starting cycles and inconsistent series length issues are genuine obstacles; 'no mathematical solution' exists for mapping phase-shifted coefficients without basis functions (Section III)
    Asserted without proof. A complex linear rotation can represent phase shifts, so the strong form of the claim is doubtful. This assumption supports the motivation but not necessarily the empirical method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fourier Basis Mapping: A Time-Frequency Learning Framework for Time Series Forecasting." pith.science (2026). https://pith.science/paper/N2ZBGEPV

@misc{pith2026250709445,
  author       = {Pith},
  title        = {Pith review of: Fourier Basis Mapping: A Time-Frequency Learning Framework for Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2ZBGEPV}},
  note         = {Machine review of arXiv:2507.09445}
}
read the original abstract

The integration of Fourier transform and deep learning opens new avenues for time series forecasting. We reconsider the Fourier transform from a basis functions perspective. Specifically, the real and imaginary parts of the frequency components can be regarded as the coefficients of cosine and sine basis functions at tiered frequency levels, respectively. We find that existing Fourier-based methods face inconsistent starting cycles and inconsistent series length issues. They fail to interpret frequency components precisely and overlook temporal information. Accordingly, the novel Fourier Basis Mapping (FBM) method addresses these issues by integrating time-frequency features through Fourier basis expansion and mapping in the time-frequency space. Our approach extracts explicit frequency features while preserving temporal characteristics. FBM supports plug-and-play integration with various types of neural networks by only adjusting the first initial projection layer for better performance. First, we propose FBM-L, FBM-NL, and FBM-NP to enhance linear, MLP-based, and Transformer-based models, respectively, demonstrating the effectiveness of time-frequency features. Next, we propose a synergetic model architecture, termed FBM-S, which decomposes the seasonal, trend, and interaction effects into three separate blocks, each designed to model time-frequency features in a specialized manner. Finally, we introduce several techniques tailored for time-frequency features, including interaction masking, centralization, patching, rolling window projection, and multi-scale down-sampling. The results are validated on diverse real-world datasets for both long-term and short-term forecasting tasks with SOTA performance.

Figures

Figures reproduced from arXiv: 2507.09445 by the authors.

Figure 1
Figure 1. Comparison of Existing Fourier-based Methods with Our Approach Fourier Basis Mapping (FBM). Existing methods [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Two Issues of Existing Fourier-based TSF Models: Inconsistent Starting Cycles Issue and Inconsistent Series Length [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Architecture of the Fourier Basis Mapping (FBM). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FBM-S: A Three-Block Architecture for Trend, Seasonality, and Interaction. In the seasonal block, a rolling window [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visualization of the Weights W0, W10, W20, W30, and W40 of FBM-L on the Electricity Dataset. Each Wi represents the influence of time-frequency features on the i-th time step of the predicted output Yˆ . The x-axis denotes time, and the y-axis denotes frequency. 0 100 …
Figure 6
Figure 6. Figure 6: Forecasting Performance Visualization: Two Case Studies on PEMS08. (a) and (b) form the first case, and (c) and (d) [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Frequency Spectrum Distribution of the Last Seven Dimensions on the Twelve Datasets. Here, the green bars represent [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: The Layers of FBM-L, FBM-NL, and FBM-NP in [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 38 canonical work pages

  1. [1]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  2. [2]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014

    arXiv 2014

  3. [3]

    Deep state space models for time series forecasting,

    S. S. Rangapuram, M. W. Seeger, J. Gasthaus, L. Stella, Y . Wang, and T. Januschowski, “Deep state space models for time series forecasting,” Neural Information Processing Systems Conference , vol. 31, 2018

    work page 2018

  4. [4]

    Dilated recurrent neural networks,

    S. Chang, Y . Zhang, W. Han, M. Yu, X. Guo, W. Tan, X. Cui, M. Wit- brock, M. A. Hasegawa-Johnson, and T. S. Huang, “Dilated recurrent neural networks,” Neural Information Processing Systems Conference , vol. 30, 2017

    work page 2017

  5. [5]

    Impact of covid-19 pandemic on electricity demand in the uk based on multivariate time series forecasting with bidirectional long short term memory,

    X. Liu and Z. Lin, “Impact of covid-19 pandemic on electricity demand in the uk based on multivariate time series forecasting with bidirectional long short term memory,” Energy, vol. 227, p. 120455, 2021

    work page 2021

  6. [6]

    Spatiotemporal attention for multivariate time series prediction and interpretation,

    T. Gangopadhyay, S. Y . Tan, Z. Jiang, R. Meng, and S. Sarkar, “Spatiotemporal attention for multivariate time series prediction and interpretation,” in IEEE International Conference on Acoustics, Speech and Signal Processing . IEEE, 2021, pp. 3560–3564

    work page 2021

  7. [7]

    WITRAN: Water- wave information transmission and recurrent acceleration network for long-range time series forecasting,

    Y . Jia, Y . Lin, X. Hao, Y . Lin, S. Guo, and H. Wan, “WITRAN: Water- wave information transmission and recurrent acceleration network for long-range time series forecasting,” in Neural Information Processing Systems Conference, 2023

    work page 2023

  8. [8]

    DeepAR: Probabilistic forecasting with autoregressive recurrent networks,

    D. Salinas, V . Flunkert, J. Gasthaus, and T. Januschowski, “DeepAR: Probabilistic forecasting with autoregressive recurrent networks,” Inter- national Journal of Forecasting , vol. 36, no. 3, pp. 1181–1191, 2020

    work page 2020

Show all 57 references
  1. [9]

    FiLM: Frequency improved legendre memory model for long-term time series forecasting,

    T. Zhou, Z. Ma, Q. Wen, L. Sun, T. Yao, W. Yin, R. Jin et al. , “FiLM: Frequency improved legendre memory model for long-term time series forecasting,” Neural Information Processing Systems Conference , vol. 35, pp. 12 677–12 690, 2022

    2022

  2. [10]

    ModernTCN: A modern pure convolution structure for general time series analysis,

    D. Luo and X. Wang, “ModernTCN: A modern pure convolution structure for general time series analysis,” in International Conference on Learning Representations , 2024

    2024

  3. [11]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” arXiv preprint arXiv:1803.01271, 2018

    2018 arXiv

  4. [12]

    SCINet: Time series modeling and forecasting with sample convolution and interaction,

    M. Liu, A. Zeng, M. Chen, Z. Xu, Q. Lai, L. Ma, and Q. Xu, “SCINet: Time series modeling and forecasting with sample convolution and interaction,” Neural Information Processing Systems Conference, vol. 35, pp. 5816–5828, 2022

    2022

  5. [13]

    MICN: Multi-scale local and global context modeling for long-term series forecasting,

    H. Wang, J. Peng, F. Huang, J. Wang, J. Chen, and Y . Xiao, “MICN: Multi-scale local and global context modeling for long-term series forecasting,” in International Conference on Learning Representations , 2022

    2022

  6. [14]

    Unsupervised scalable representation learning for multivariate time series,

    J.-Y . Franceschi, A. Dieuleveut, and M. Jaggi, “Unsupervised scalable representation learning for multivariate time series,” Neural Information Processing Systems Conference, vol. 32, 2019

    2019

  7. [15]

    Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting,

    R. Sen, H.-F. Yu, and I. S. Dhillon, “Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting,” Neural Information Processing Systems Conference , vol. 32, 2019

    2019

  8. [16]

    TimesNet: Temporal 2d-variation modeling for general time series analysis,

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: Temporal 2d-variation modeling for general time series analysis,” in International Conference on Learning Representations , 2022

    2022

  9. [17]

    Timemixer: Decomposable multiscale mixing for time series forecasting,

    S. Wang, H. Wu, X. Shi, T. Hu, H. Luo, L. Ma, J. Y . Zhang, and J. Zhou, “Timemixer: Decomposable multiscale mixing for time series forecasting,” arXiv preprint arXiv:2405.14616 , 2024

    2024 arXiv

  10. [18]

    Unlocking the power of patch: Patch-based mlp for long-term time series forecasting,

    P. Tang and W. Zhang, “Unlocking the power of patch: Patch-based mlp for long-term time series forecasting,” in AAAI, vol. 39, no. 12, 2025, pp. 12 640–12 648

    2025

  11. [19]

    Tsmixer: An all-mlp architecture for time series forecasting,

    S.-A. Chen, C.-L. Li, N. Yoder, S. O. Arik, and T. Pfister, “Tsmixer: An all-mlp architecture for time series forecasting,” arXiv preprint arXiv:2303.06053, 2023

    2023 arXiv

  12. [20]

    N-BEATS: Neural basis expansion analysis for interpretable time series forecasting,

    B. N. Oreshkin, D. Carpov, N. Chapados, and Y . Bengio, “N-BEATS: Neural basis expansion analysis for interpretable time series forecasting,” in International Conference on Learning Representations , 2019

    2019

  13. [21]

    N-HiTS: Neural hierarchical interpolation for time series forecasting,

    C. Challu, K. G. Olivares, B. N. Oreshkin, F. G. Ramirez, M. M. Canseco, and A. Dubrawski, “N-HiTS: Neural hierarchical interpolation for time series forecasting,” in AAAI Conference on Artificial Intelli- gence, vol. 37, no. 6, 2023, pp. 6989–6997

    2023

  14. [22]

    Frequency-domain mlps are more effective learners in time series forecasting,

    K. Yi, Q. Zhang, W. Fan, S. Wang, P. Wang, H. He, N. An, D. Lian, L. Cao, and Z. Niu, “Frequency-domain mlps are more effective learners in time series forecasting,” in Neural Information Processing Systems Conference, 2023

    2023

  15. [23]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Neural Information Processing Systems Conference , vol. 30, 2017

    2017

  16. [24]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,” in International Conference on Learning Representations , 2022

    2022

  17. [25]

    BasisFormer: Attention-based time series forecasting with learnable and interpretable basis,

    Z. Ni, H. Yu, S. Liu, J. Li, and W. Lin, “BasisFormer: Attention-based time series forecasting with learnable and interpretable basis,” in Neural Information Processing Systems Conference , 2023

    2023

  18. [26]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in AAAI Conference on Artificial Intelligence , vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  19. [27]

    Autoformer: Decomposition trans- formers with auto-correlation for long-term series forecasting,

    H. Wu, J. Xu, J. Wang, and M. Long, “Autoformer: Decomposition trans- formers with auto-correlation for long-term series forecasting,”Advances in Neural Information Processing Systems , vol. 34, pp. 22 419–22 430, 2021

    2021

  20. [28]

    Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting,

    S. Liu, H. Yu, C. Liao, J. Li, W. Lin, A. X. Liu, and S. Dustdar, “Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting,” inInternational conference on learning representations, 2021

    2021

  21. [29]

    Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,

    Y . Zhang and J. Yan, “Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,” in International Conference on Learning Representations , 2022

    2022

  22. [30]

    InParformer: evolutionary decomposition transformers with interactive parallel at- tention for long-term time series forecasting,

    H. Cao, Z. Huang, T. Yao, J. Wang, H. He, and Y . Wang, “InParformer: evolutionary decomposition transformers with interactive parallel at- tention for long-term time series forecasting,” in AAAI Conference on Artificial Intelligence, vol. 37, no. 6, 2023, pp. 6906–6915

    2023

  23. [31]

    An encoder–decoder architecture with fourier attention for chaotic time series multi-step prediction,

    K. Fu, H. Li, and X. Shi, “An encoder–decoder architecture with fourier attention for chaotic time series multi-step prediction,” Applied Soft Computing, p. 111409, 2024

    2024

  24. [32]

    Temporal chain network with intuitive attention mechanism for long-term series forecasting,

    Z. Zhang, Y . Han, B. Ma, M. Liu, and Z. Geng, “Temporal chain network with intuitive attention mechanism for long-term series forecasting,” IEEE Transactions on Instrumentation and Measurement , 2023

    2023

  25. [33]

    Tempo: Prompt-based generative pre-trained transformer for time series forecasting,

    D. Cao, F. Jia, S. O. Arik, T. Pfister, Y . Zheng, W. Ye, and Y . Liu, “Tempo: Prompt-based generative pre-trained transformer for time series forecasting,” arXiv preprint arXiv:2310.04948 , 2023

    2023 arXiv

  26. [34]

    Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting,

    P. Chen, Y . Zhang, Y . Cheng, Y . Shu, Y . Wang, Q. Wen, B. Yang, and C. Guo, “Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting,” arXiv preprint arXiv:2402.05956 , 2024

    2024 arXiv

  27. [35]

    iTrans- former: Inverted transformers are effective for time series forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “iTrans- former: Inverted transformers are effective for time series forecasting,” in International Conference on Learning Representations , 2024

    2024

  28. [36]

    Duet: Dual clustering enhanced multivariate time series forecasting,

    X. Qiu, X. Wu, Y . Lin, C. Guo, J. Hu, and B. Yang, “Duet: Dual clustering enhanced multivariate time series forecasting,” arXiv preprint arXiv:2412.10859, 2024

    2024 arXiv

  29. [37]

    Timer-xl: Long- context transformers for unified time series forecasting,

    Y . Liu, G. Qin, X. Huang, J. Wang, and M. Long, “Timer-xl: Long- context transformers for unified time series forecasting,” arXiv preprint arXiv:2410.04803, 2024

    2024 arXiv

  30. [38]

    Sepformer-based models: More efficient models for long sequence time-series forecasting,

    J. Fan, Z. Wang, D. Sun, and H. Wu, “Sepformer-based models: More efficient models for long sequence time-series forecasting,” IEEE Transactions on Emerging Topics in Computing , 2022

    2022

  31. [39]

    Fedformer: Frequency enhanced decomposed transformer for long-term series fore- casting,

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency enhanced decomposed transformer for long-term series fore- casting,” in International Conference on Machine Learning . PMLR, 2022, pp. 27 268–27 286

    2022

  32. [40]

    CrossGNN: Confronting noisy multivariate time series via cross interaction refinement,

    Q. Huang, L. Shen, R. Zhang, S. Ding, B. Wang, Z. Zhou, and Y . Wang, “CrossGNN: Confronting noisy multivariate time series via cross interaction refinement,” in Neural Information Processing Systems Conference, 2023

    2023

  33. [41]

    FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective,

    K. Yi, Q. Zhang, W. Fan, H. He, L. Hu, P. Wang, N. An, L. Cao, and Z. Niu, “FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective,” arXiv preprint arXiv:2311.06190, 2023

    2023 arXiv

  34. [42]

    Graph deep learning for time series forecasting,

    A. Cini, I. Marisca, D. Zambon, and C. Alippi, “Graph deep learning for time series forecasting,” arXiv preprint arXiv:2310.15978 , 2023

    2023 arXiv

  35. [43]

    Adaptive dependency learning graph neural networks,

    A. Sriramulu, N. Fourrier, and C. Bergmeir, “Adaptive dependency learning graph neural networks,” Information Sciences , vol. 625, pp. 700–714, 2023

    2023

  36. [44]

    Multivariate time series forecasting with dynamic graph neural odes,

    M. Jin, Y . Zheng, Y .-F. Li, S. Chen, B. Yang, and S. Pan, “Multivariate time series forecasting with dynamic graph neural odes,” IEEE Trans- actions on Knowledge and Data Engineering , vol. 35, no. 9, pp. 9168– 9180, 2022. IEEE TRANSACTIONS ON XXXXXXXXX 17

    2022

  37. [45]

    Mixmamba: Time series modeling with adaptive expertise,

    K. Alkilane, Y . He, and D.-H. Lee, “Mixmamba: Time series modeling with adaptive expertise,” Information Fusion, vol. 112, p. 102589, 2024

    2024

  38. [46]

    Is mamba effective for time series forecasting?

    Z. Wang, F. Kong, S. Feng, M. Wang, X. Yang, H. Zhao, D. Wang, and Y . Zhang, “Is mamba effective for time series forecasting?” Neurocom- puting, vol. 619, p. 129178, 2025

    2025

  39. [47]

    Cmamba: Channel correlation enhanced state space models for multivariate time series forecasting,

    C. Zeng, Z. Liu, G. Zheng, and L. Kong, “Cmamba: Channel correlation enhanced state space models for multivariate time series forecasting,” arXiv preprint arXiv:2406.05316 , 2024

    2024 arXiv

  40. [48]

    Are transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” in AAAI Conference on Artificial Intelligence , vol. 37, no. 9, 2023, pp. 11 121–11 128

    2023

  41. [49]

    FITS: Modeling time series with 10k parameters,

    Z. Xu, A. Zeng, and Q. Xu, “FITS: Modeling time series with 10k parameters,” arXiv preprint arXiv:2307.03756 , 2023

    2023 arXiv

  42. [50]

    FL-Net: A multi-scale cross-decomposition network with frequency external attention for long-term time series forecasting,

    S. Huang and Y . Liu, “FL-Net: A multi-scale cross-decomposition network with frequency external attention for long-term time series forecasting,” Knowledge-Based Systems, p. 111473, 2024

    2024

  43. [51]

    Rethinking fourier transform from a basis functions perspective for long-term time series forecasting,

    R. Yang, L. Cao, J. Y ANG et al. , “Rethinking fourier transform from a basis functions perspective for long-term time series forecasting,” Advances in Neural Information Processing Systems , vol. 37, pp. 8515– 8540, 2024

    2024

  44. [52]

    En- hancing the locality and breaking the memory bottleneck of transformer on time series forecasting,

    S. Li, X. Jin, Y . Xuan, X. Zhou, W. Chen, Y .-X. Wang, and X. Yan, “En- hancing the locality and breaking the memory bottleneck of transformer on time series forecasting,” Neural Information Processing Systems Conference, vol. 32, 2019

    2019

  45. [53]

    Variational hierarchical n-beats model for long-term time-series forecasting,

    R. Yang, L. Cao, J. Li, and J. Yang, “Variational hierarchical n-beats model for long-term time-series forecasting,” IEEE Transactions on Neural Networks and Learning Systems , 2025

    2025

  46. [54]

    Reversible instance normalization for accurate time-series forecasting against distri- bution shift,

    T. Kim, J. Kim, Y . Tae, C. Park, J.-H. Choi, and J. Choo, “Reversible instance normalization for accurate time-series forecasting against distri- bution shift,” in International Conference on Learning Representations , 2021

    2021

  47. [55]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in MICCAI. Springer, 2015, pp. 234–241

    2015

  48. [56]

    Less is more: Fast multivariate time series forecasting with light sampling- oriented mlp structures. arxiv 2022,

    T. Zhang, Y . Zhang, W. Cao, J. Bian, X. Yi, S. Zheng, and J. Li, “Less is more: Fast multivariate time series forecasting with light sampling- oriented mlp structures. arxiv 2022,” arXiv preprint arXiv:2207.01186 , 2022

    2022 arXiv

  49. [57]

    Non-stationary transformers: Exploring the stationarity in time series forecasting,

    Y . Liu, H. Wu, J. Wang, and M. Long, “Non-stationary transformers: Exploring the stationarity in time series forecasting,” Advances in neural information processing systems , vol. 35, pp. 9881–9893, 2022. Runze Yang received the B.S. and M.S. degrees from the Faculty of Mathe...

    2022

Pith tools

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