Pith. sign in

REVIEW 4 major objections 5 minor 60 references

Delayformer: spatiotemporal transformation for predicting high-dimensional dynamics

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Delayformer predicts all variables of a high-dimensional system at once by converting each variable's history into a delay-embedded Hankel matrix, encoding it with a shared Vision Transformer, and decoding each variable with a linear layer.

desk verdict Useful empirical method; the cross-variable claim does not match the implementation and needs correction before publication. read the letter →

arxiv 2506.11528 v1 pith:6YROJ46I submitted 2025-06-13 cs.LG

Zijian Wang , Peng Tao , Luonan Chen This is my paper
classification cs.LG MSC 37M1062M1068T07
keywords timeseriesforecastingdelayembeddingHankelmatrixVisionTransformermultivariatespatiotemporalinformationdynamicalsystemsfoundationmodelhigh-dimensionalprediction
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

This paper introduces Delayformer, a forecasting method that predicts every variable of a high-dimensional system simultaneously by converting each variable's history into a delay-embedded Hankel matrix, processing those matrices as images through one shared Vision Transformer encoder, and decoding the shared representation with per-variable linear layers. The authors argue that forecasting system states rather than individual scalar sequences inherits the guarantees of delay embedding theory and thereby sidesteps the nonlinearity and cross-variable interaction problems that hurt current deep-learning forecasters. They report that Delayformer outperforms state-of-the-art baselines on synthetic coupled chaotic systems under noise and time-varying parameters, on short-term meteorological datasets, and on eight long-term forecasting benchmarks, while staying accurate with very small training sets and few observed variables. They also show that a larger version pretrained on a broad time-series corpus transfers zero-shot and few-shot to unseen domains, which they take as evidence the architecture can serve as a time-series foundation model.

What carries the argument

The central object is the multivariate spatiotemporal information (mvSTI) transformation: each variable $x_k$ over a window is embedded into a Hankel matrix $\mathbf{Y}_k^{\mathrm{Known}}$ whose rows are time-delayed copies of that variable, and the transformation $\Psi_k$ maps such a delay-embedded state sequence to the future states $\mathbf{X}_{\mathrm{Pred}}$ of all variables. Delayformer solves the collection of these maps with a shared Vision Transformer encoder: the Hankel matrix is treated as a two-dimensional image, sliced into patches, tokenized with position encoding, and processed by self-attention layers, so the global structure (long-range dependencies) and local structure (short-range dependencies) of the reconstructed attractor are captured together; distinct linear decoders $\mathbf{f}_k$ then project the shared latent representation to each variable's predicted future values. The theoretical load is carried by the delay embedding theorem, which guarantees that a single variable's delay coordinates reconstruct the attractor's topology, making the shared encoder's task the learning of a common dynamical-state representation across variables.

What would settle it

Estimate the box-counting dimension $d$ of the empirical attractor for a benchmark where Delayformer wins and compare it with the embedding dimension $L$ used in the experiments: if Delayformer still wins when $L \le 2d$, the delay-embedding premise is not what produces the result.

Watch

Extended reading notes

Core claim

The paper's central claim is that the multivariate spatiotemporal information (mvSTI) equation, built from delay embeddings of every observed variable, can be solved by a single shared Vision Transformer encoder followed by distinct linear decoders, turning forecasting into a map from delay-embedded state sequences to the future sequence of all original states. Under this view each variable's Hankel matrix is a topologically equivalent reconstruction of the system's attractor, so the shared encoder learns one common dynamical-state representation across variables, and each linear decoder projects that representation to its own variable's future. The authors claim that this shift from predicting raw variables to predicting states is what lets the model handle nonlinearity, interactions, noise, and limited data simultaneously; empirically they support it with top-two results on every long-term benchmark and with robustness experiments on coupled chaotic systems, meteorological data, and cross-domain transfer.

Load-bearing premise

The load-bearing premise is that the data come from a deterministic low-dimensional dynamical system whose attractor can be reconstructed from a single variable's delay embedding, and that a global inverse map from the delay-embedded state back to the full system state exists and can be learned; for noisy, non-stationary real-world data this is never verified.

Editorial extensions

If this is right

  • Because Delayformer predicts system states rather than individual scalars, it is expected to stay accurate when few variables are observed; the paper demonstrates this on the weather dataset using only 2, 5, or 10 observed variables.
  • Because the shared ViT encoder uses the same parameters for every variable while each variable keeps its own linear decoder, the method combines channel-independent training with cross-variable coupling through the learned state representation.
  • Because only linear decoders sit on top of the shared representation, the model trains with a simple MSE loss and is reported to be robust to the choice of embedding dimension $L$ and patch sizes $(p_1,p_2)$.
  • A larger Delayformer pretrained on a broad time-series corpus achieves competitive zero-shot forecasts on unseen datasets, and the paper reports that 2–5% of target-domain training data suffices for strong few-shot transfer, supporting its potential as a foundation model.
  • On the eight long-term forecasting benchmarks the paper reports that Delayformer finishes in the top two for every dataset and prediction length, so the claimed advantage is not specific to one domain.

Reading between the lines

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

  • The mvSTI formulation suggests a stress test the paper does not run: on systems whose attractor dimension is close to half the embedding dimension, the reconstruction guarantee weakens, so Delayformer's advantage over ordinary Transformers should shrink; checking this would isolate how much of the gain comes from the delay-embedding premise.
  • The Discussion's suggestion that the shared encoder's representation lives in a space where evolution is approximately linear implies the per-variable linear decoders could perhaps be replaced by one linear operator acting on the latent state to produce multi-step forecasts without retraining; the paper does not demonstrate that.
  • If the delay-embedded Hankel input is what carries the benefit, then swapping the ViT backbone for a different architecture while keeping the same input construction should preserve most of the gain; the paper only notes such flexibility as future work.
  • The zero-shot and few-shot results point toward pretraining Delayformer on larger heterogeneous time-series corpora, a scale-up the authors explicitly defer; success there would turn the state-prediction design into a candidate architecture for general time-series foundation models.
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 / 5 minor

Summary. Delayformer is a multivariate time-series forecasting method that transforms each observed variable into a delay-embedded Hankel matrix, treats that matrix as an image, patches it, and feeds the patches through a shared Vision Transformer (ViT) encoder, followed by per-variable linear decoders. The paper justifies the construction with Takens' delay-embedding theorem, introduces a multivariate spatiotemporal information (mvSTI) equation, and claims that predicting system states rather than individual variables lets the method overcome nonlinearity and cross-variable interaction problems. Experiments cover a 30-dimensional coupled Lorenz system under noise and time-varying parameters, three short-term meteorological datasets, eight long-term forecasting benchmarks, limited-data settings, and cross-domain zero-shot/few-shot forecasting.

Significance. If the central claims were fully supported, Delayformer would be a useful contribution that bridges dynamical-systems theory and Transformer-based time-series forecasting: it releases code, validates on controlled synthetic systems with noise, evaluates on a wide range of standard benchmarks, and shows robustness in limited-data settings. The cross-domain zero-shot results are also an interesting demonstration of potential as a foundation model. However, the paper's main conceptual selling point -- cross-learning interactions among variables -- is not realized by the channel-independent implementation described in Section II-D, and the theoretical derivation assumes a global inverse that Takens' theorem does not guarantee. These issues affect the interpretation of every reported result and the stated novelty, so they need to be resolved before the contribution can be assessed as stated.

major comments (4)
  1. [Section II-D (Channel-Independence) vs. Abstract and Section II-C] The abstract and Section II-C claim that Delayformer "cross-learns states from different variables" and "exploits the information contents and their interactions of all observed variables," but the implementation in Section II-D processes each variable's tokens through the Transformer encoder blocks independently, with shared parameters and no attention or other tensor operation that mixes the N channels. The loss in Section II-D is an aggregated per-variable MSE, and the decoders {ƒ_k} are per-variable linear layers, so at inference the model cannot observe interactions between variables. The reported experiments therefore validate a weight-tied, per-variable ViT applied to Hankel images rather than the cross-variable interaction model described as central. This is an internal inconsistency, not merely a wording issue, because it directly undermines the claim that Delayformer "theoretically and computationally overcomes ... cross-interaction problems." The authors should either add a cross-variable mixing operation (e.g., an attention layer across variables) or substantially revise the theoretical narrative and novelty claims to describe a shared per-variable state-reconstruction model.
  2. [Section II-B, Eqs. (6)-(7)] The derivation of the mvSTI equation assumes the existence of a global inverse mapping Ψ_k such that Ψ_k(Θ_k^{Known}) = X^{Pred}. Takens' theorem guarantees an embedding of the attractor M into R^L for generic observation functions when L > 2d, but it does not guarantee a global inverse on all of R^L, and the theorem's hypotheses (smooth diffeomorphism, generic scalar observation, well-defined box-counting dimension d) are not verified for the noisy real-world meteorological, energy, and traffic datasets. This is a correctness risk for the paper's theoretical foundation, since Eq. (7) is presented as a consequence of Takens' theorem. A concrete remedy would be to report estimates of the relevant attractor dimension or false-nearest-neighbor statistics for the benchmark datasets, or to explicitly state the theorem as a heuristic motivation rather than a guarantee.
  3. [Table I and Section III-C] The central empirical claim that "Delayformer achieved top-two performances in all benchmarks" (Section III-C) relies on Table I, but the table is extremely hard to parse: the column grouping appears misaligned, the row labels for datasets and models are not clearly separated, and the trailing row of small integers is unexplained. Moreover, no confidence intervals, standard deviations, or statistical significance tests are reported, despite the paper reporting six-seed error bars for synthetic data (Fig. 2h) and meteorological data (Fig. 3d). Given that many per-cell differences in Table I are small (e.g., ETTh1 MSE 96: 0.379 vs. 0.386 for several baselines), the claim of consistent superiority needs either a corrected, readable table with uncertainty estimates or a multiple-run comparison to support it.
  4. [Section III-E] The hyperparameter sensitivity section contains a notational error that makes the results ambiguous: the text says "We tested m values of 12, 27, and 42 on the weather dataset" and later discusses "embedding dimension m," but the method defines the embedding dimension as L and the prediction length as m. The experiments appear to vary L, and the reported values (L=12, 27, 42) are inconsistent with the earlier claim that L=49 or L=27 in most experiments. Please correct the notation and clarify which parameter is varied in Fig. 5b and Table IV.
minor comments (5)
  1. [Throughout] The system is called both "Lorenz" and "Lorentz" (e.g., Section III-A and Fig. 2 caption); it should be Lorenz consistently.
  2. [Section III-B] The location "Wakkanani, Japan" should likely be "Wakkanai, Japan".
  3. [Section II-D and Fig. 1] There are multiple typographical issues in the notation, such as "Θₖᴊⁿᵥₒᵛ" and "𝒰𝒁ᵃ𝑣𝑧𝑒𝑛" in the bolded implementation details, and the dimension Θ_k^{Known} vs. Θ_k^{Knwown} is inconsistent; these should be cleaned up.
  4. [Section II-D] The sentence about patch count says the Hankel matrix is sliced into "L × (W-L+1)/(p1 p2) patches," but the number of patches should be (L/p1) × ((W-L+1)/p2) if p1 and p2 are patch dimensions; please clarify this formula.
  5. [Section III-F] The claim that zero-shot MSE values (e.g., 0.139 and 0.243) are "lower than the current SOTA" or "comparable to current small models" is not directly supported because the pretrained model uses a different training setup (10 encoder layers, different patches) than the models in Table I; please state the comparison conditions explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the empirical claims are tested on held-out data and standard benchmarks, and the theoretical overreach is an unproven application of Takens' theorem rather than a derivation that assumes its own conclusion.

full rationale

Delayformer's central assertion is empirical: it outperforms baselines on synthetic and real-world benchmarks (Abstract; Section III-C). These results come from held-out test splits and standard benchmarks, and the hyperparameters L, p1, p2 are reported as robustness checks (Section III-E), not fitted to the test set; no fitted parameter is renamed as a prediction. The mvSTI equation (Section II-B, Eqs. 6-7) is stated as an inverse of the delay-embedding map and then approximated by a shared ViT encoder with per-variable linear decoders. This is an architectural ansatz, not a derivation that assumes its conclusion: the existence and accuracy of the map is learned from data, not presupposed. The main theoretical weakness is overextension rather than circularity: Takens' theorem (Section II-A) guarantees an embedding of the attractor, not a global inverse from one variable's past delay vectors to all variables' future states, so the statement 'We can derive the mvSTI equation (6) or (7) based on the delay embedding theorem' is an unsupported derivation step. That is a correctness risk, not a circular reduction. The paper also relies on the authors' own prior STI literature ([39]-[44], [57]) for background and for the STI concept, but the load-bearing mathematical premise (delay embedding) is external ([36]) and the benchmark comparisons are self-contained. One internal inconsistency should be noted: Section II-C claims a shared encoder cross-learns information across variables, while Section II-D states each variable's tokens are processed independently with shared parameters, so no cross-variable information flows at inference; this undermines the interpretive claim but does not make the empirical predictions circular. Overall, no step in the derivation chain is equivalent by construction to an input or fit.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The method rests on standard delay embedding theory, which is applied without verifying its conditions on real datasets, and on two hand-chosen hyperparameters. The most speculative step is the assumed global invertibility of the delay embedding map.

free parameters (2)
  • embedding dimension L = 49 or 27 in most experiments
    Chosen by 'empirical experience' as a value between 1 and input length, not derived from theory (Section III-E).
  • patch sizes (p1, p2) = (6,7), (24,7), (5,3) etc.
    Set based on operability and memory balance, with tests showing insensitivity (Section II-D, Section III-E).
assumptions (4)
  • domain assumption Takens' delay embedding theorem applies to the observed discrete-time multivariate data
    Used to justify that a Hankel matrix of one variable reconstructs the full system state. Invoked in Section II-A without verifying the required smooth diffeomorphism or generic observation conditions on real datasets.
  • ad hoc to paper A global inverse mapping Ψ_k exists for the embedding
    Section II-A states Φ_k ∘ Ψ_k = id, but Takens' theorem only gives a one-to-one embedding on the attractor, not a global inverse. This assumption is load-bearing for the mvSTI equation.
  • domain assumption A shared encoder can represent topologically conjugate delay embeddings across variables
    Assumed in Section II-C to justify channel-independence and the shared ViT, but no formal justification is provided.
  • domain assumption Patching the Hankel matrix preserves the local and global structure needed for forecasting
    Assumed in Section II-D based on the idea that proximate states in the reconstruction influence future states; a heuristic borrowed from image processing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Delayformer: spatiotemporal transformation for predicting high-dimensional dynamics." pith.science (2026). https://pith.science/paper/6YROJ46I

@misc{pith2026250611528,
  author       = {Pith},
  title        = {Pith review of: Delayformer: spatiotemporal transformation for predicting high-dimensional dynamics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6YROJ46I}},
  note         = {Machine review of arXiv:2506.11528}
}
read the original abstract

Predicting time-series is of great importance in various scientific and engineering fields. However, in the context of limited and noisy data, accurately predicting dynamics of all variables in a high-dimensional system is a challenging task due to their nonlinearity and also complex interactions. Current methods including deep learning approaches often perform poorly for real-world systems under such circumstances. This study introduces the Delayformer framework for simultaneously predicting dynamics of all variables, by developing a novel multivariate spatiotemporal information (mvSTI) transformation that makes each observed variable into a delay-embedded state (vector) and further cross-learns those states from different variables. From dynamical systems viewpoint, Delayformer predicts system states rather than individual variables, thus theoretically and computationally overcoming such nonlinearity and cross-interaction problems. Specifically, it first utilizes a single shared Visual Transformer (ViT) encoder to cross-represent dynamical states from observed variables in a delay embedded form and then employs distinct linear decoders for predicting next states, i.e. equivalently predicting all original variables parallelly. By leveraging the theoretical foundations of delay embedding theory and the representational capabilities of Transformers, Delayformer outperforms current state-of-the-art methods in forecasting tasks on both synthetic and real-world datasets. Furthermore, the potential of Delayformer as a foundational time-series model is demonstrated through cross-domain forecasting tasks, highlighting its broad applicability across various scenarios.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 49 canonical work pages

  1. [1]

    A machine learning model that outperforms conventional global subseasonal forecast models,

    L. Chen et al., "A machine learning model that outperforms conventional global subseasonal forecast models," Nature Communications, vol. 15, no. 1, p. 6425, 2024

    work page 2024

  2. [2]

    Forecasting Andean rainfall and crop yield from the influence of El Niñ o on Pleiades visibility,

    B. S. Orlove, J. C. Chiang, and M. A. Cane, "Forecasting Andean rainfall and crop yield from the influence of El Niñ o on Pleiades visibility," Nature, vol. 403, no. 6765, pp. 68-71, 2000

    work page 2000

  3. [3]

    Self -organizing maps of typhoon tracks allow for flood forecasts up to two days in advance,

    L.-C. Chang, F. -J. Chang, S. -N. Yang, F. -H. Tsai, T. -H. Chang, and E. E. Herricks, "Self -organizing maps of typhoon tracks allow for flood forecasts up to two days in advance," Nature Communications, vol. 11, no. 1, p. 1983, 2020

    work page 1983

  4. [4]

    Stock price prediction using LSTM, RNN and CNN -sliding window model,

    S. Selvin, R. Vinayakumar, E. Gopalakrishnan, V. K. Menon, and K. Soman, "Stock price prediction using LSTM, RNN and CNN -sliding window model," in 2017 international conference on advances in computing, communications and informatics (icacci), 2017: IEEE, pp. 1643-1647

    work page 2017

  5. [5]

    Prediction of net energy consumption based on economic indicators (GNP and GDP) in Turkey,

    A. Sö zen and E. Arcaklioglu, "Prediction of net energy consumption based on economic indicators (GNP and GDP) in Turkey," Energy policy, vol. 35, no. 10, pp. 4981- 4992, 2007

    work page 2007

  6. [6]

    Baroreflex sensitivity and heart -rate variability in prediction of total cardiac mortality after myocardial infarction,

    M. T. La Rovere, J. T. Bigger, F. I. Marcus, A. Mortara, and P. J. Schwartz, "Baroreflex sensitivity and heart -rate variability in prediction of total cardiac mortality after myocardial infarction," The Lancet, vol. 351, no. 9101, pp. 478-484, 1998

    work page 1998

  7. [7]

    P. N. Ramkumar et al. , "Artificial intelligence and arthroplasty at a single institution: real-world applications of machine learning to big data, value -based care, mobile health, and remote patient monitoring," The Journal of arthroplasty, vol. 34, no. 10, pp. 2204-2209, 2019

    work page 2019

  8. [8]

    Machine learning based early warning system enables accurate mortality risk prediction for COVID-19,

    Y. Gao et al. , "Machine learning based early warning system enables accurate mortality risk prediction for COVID-19," Nature communications, vol. 11, no. 1, p. 5033, 2020

    work page 2020

Show all 60 references
  1. [9]

    Machine learning-based fault diagnosis for single- and multi-faults in induction motors using measured stator currents and vibration signals,

    M. Z. Ali, M. N. S. K. Shabbir, X. Liang, Y. Zhang, and T. Hu, "Machine learning-based fault diagnosis for single- and multi-faults in induction motors using measured stator currents and vibration signals," IEEE Transactions on Industry Applications, vol. 55, no. 3, pp. 2378-2...

    2019

  2. [10]

    On-line building energy optimization using deep reinforcement learning,

    E. Mocanu et al., "On-line building energy optimization using deep reinforcement learning," IEEE transactions on smart grid, vol. 10, no. 4, pp. 3698-3708, 2018

    2018

  3. [11]

    A review on time series data mining,

    T.-c. Fu, "A review on time series data mining," Engineering Applications of Artificial Intelligence, vol. 24, no. 1, pp. 164-181, 2011

    2011

  4. [12]

    Deep learning for time series classification: a review,

    H. Ismail Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P. -A. Muller, "Deep learning for time series classification: a review," Data mining and knowledge discovery, vol. 33, no. 4, pp. 917-963, 2019

    2019

  5. [13]

    J. D. Hamilton, Time series analysis. Princeton university press, 2020

    2020

  6. [14]

    Nearly efficient estimation of time series models with predetermined, but not exogenous, instruments,

    F. Hayashi and C. Sims, "Nearly efficient estimation of time series models with predetermined, but not exogenous, instruments," Econometrica: Journal of the Econometric Society, pp. 783-798, 1983

    1983

  7. [15]

    Vector autoregressive models,

    H. Lü tkepohl, "Vector autoregressive models," in Handbook of research methods and applications in empirical macroeconomics : Edward Elgar Publishing, 2013, pp. 139-164

    2013

  8. [16]

    A training algorithm for optimal margin classifiers,

    B. E. Boser, I. M. Guyon, and V. N. Vapnik, "A training algorithm for optimal margin classifiers," in Proceedings of the fifth annual workshop on Computational learning theory, 1992, pp. 144-152

    1992

  9. [17]

    Natural language processing,

    K. Chowdhary and K. Chowdhary, "Natural language processing," Fundamentals of artificial intelligence, pp. 603-649, 2020

    2020

  10. [18]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016 2016, pp. 770-778

    2016

  11. [19]

    Recurrent neural networks and robust time series prediction,

    J. T. Connor, R. D. Martin, and L. E. Atlas, "Recurrent neural networks and robust time series prediction," IEEE transactions on neural networks, vol. 5, no. 2, pp. 240-254, 1994

    1994

  12. [20]

    Stock price pattern recognition-a recurrent neural network approach,

    K.-i. Kamijo and T. Tanigawa, "Stock price pattern recognition-a recurrent neural network approach," in 1990 IJCNN international joint conference on neural networks, 1990: IEEE, pp. 215-221

    1990

  13. [21]

    Temporal convolutional networks for action segmentation and detection,

    C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager, "Temporal convolutional networks for action segmentation and detection," 2017, pp. 156-165

    2017

  14. [22]

    Are transformers effective for time series forecasting?,

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, "Are transformers effective for time series forecasting?," in Proceedings of the AAAI conference on artificial intelligence , 2023, vol. 37, no. 9, pp. 11121-11128. 13

    2023

  15. [23]

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

    H. Zhou et al., "Informer: Beyond efficient transformer for long sequence time-series forecasting," in Proceedings of the AAAI conference on artificial intelligence , 2021, vol. 35, no. 12, pp. 11106-11115

    2021

  16. [24]

    Revisiting long -term time series forecasting: An investigation on linear mapping,

    Z. Li, S. Qi, Y. Li, and Z. Xu, "Revisiting long -term time series forecasting: An investigation on linear mapping," arXiv preprint arXiv:2305.10721, 2023

    2023 arXiv

  17. [25]

    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," arXiv preprint arXiv:2211.14730, 2022

    2022 arXiv

  18. [26]

    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 The eleventh international conference on learning representations, 2023

    2023

  19. [27]

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting,

    Y. Liu et al. , "iTransformer: Inverted Transformers Are Effective for Time Series Forecasting," in The Twelfth International Conference on Learning Representations

  20. [28]

    From Similarity to Superiority: Channel Clustering for Time Series Forecasting,

    J. Chen et al., "From Similarity to Superiority: Channel Clustering for Time Series Forecasting," arXiv preprint arXiv:2404.01340, 2024

    2024 arXiv

  21. [29]

    The capacity and robustness trade-off: Revisiting the channel independent strategy for multivariate time series forecasting,

    L. Han, H. -J. Ye, and D. -C. Zhan, "The capacity and robustness trade-off: Revisiting the channel independent strategy for multivariate time series forecasting," IEEE Transactions on Knowledge and Data Engineering, 2024

    2024

  22. [30]

    Foundation models for time series analysis: A tutorial and survey,

    Y. Liang et al. , "Foundation models for time series analysis: A tutorial and survey," in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 6555-6565

    2024

  23. [31]

    Complex network from time series based on phase space reconstruction,

    Z. Gao and N. Jin, "Complex network from time series based on phase space reconstruction," Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. 19, no. 3, 2009

    2009

  24. [32]

    Determining Lyapunov exponents from a time series,

    A. Wolf, J. B. Swift, H. L. Swinney, and J. A. Vastano, "Determining Lyapunov exponents from a time series," Physica D: nonlinear phenomena, vol. 16, no. 3, pp. 285- 317, 1985

    1985

  25. [33]

    Ergodic theory of chaos and strange attractors,

    J.-P. Eckmann and D. Ruelle, "Ergodic theory of chaos and strange attractors," Reviews of modern physics, vol. 57, no. 3, p. 617, 1985

    1985

  26. [34]

    Y. A. Kuznetsov, I. A. Kuznetsov, and Y. Kuznetsov, Elements of applied bifurcation theory. Springer, 1998

    1998

  27. [35]

    Latent ordinary differential equations for irregularly -sampled time series,

    Y. Rubanova, R. T. Chen, and D. K. Duvenaud, "Latent ordinary differential equations for irregularly -sampled time series," Advances in neural information processing systems, vol. 32, 2019

    2019

  28. [36]

    Detecting strange attractors in turbulence,

    F. Takens, "Detecting strange attractors in turbulence," in Dynamical Systems and Turbulence, Warwick 1980: proceedings of a symposium held at the University of Warwick 1979/80, 2006: Springer, pp. 366-381

    1980

  29. [37]

    Nonlinear dynamics, delay times, and embedding windows,

    H. S. Kim, R. Eykholt, and J. Salas, "Nonlinear dynamics, delay times, and embedding windows," Physica D: Nonlinear Phenomena, vol. 127, no. 1-2, pp. 48-60, 1999

    1999

  30. [38]

    The dimension of chaotic attractors,

    J. D. Farmer, E. Ott, and J. A. Yorke, "The dimension of chaotic attractors," Physica D: Nonlinear Phenomena, vol. 7, no. 1-3, pp. 153-180, 1983

    1983

  31. [39]

    Randomly distributed embedding making short -term high-dimensional data predictable,

    H. Ma, S. Leng, K. Aihara, W. Lin, and L. Chen, "Randomly distributed embedding making short -term high-dimensional data predictable," Proceedings of the National Academy of Sciences, vol. 115, no. 43, pp. E9994-E10002, 2018

    2018

  32. [40]

    Predicting future dynamics from short - term time series using an Anticipated Learning Machine,

    C. Chen et al. , "Predicting future dynamics from short - term time series using an Anticipated Learning Machine," National Science Review, vol. 7, no. 6, pp. 1079 -1091, 2020

    2020

  33. [41]

    Autoreservoir computing for multistep ahead prediction based on the spatiotemporal information transformation,

    P. Chen, R. Liu, K. Aihara, and L. Chen, "Autoreservoir computing for multistep ahead prediction based on the spatiotemporal information transformation," Nature communications, vol. 11, no. 1, p. 4568, 2020

    2020

  34. [42]

    Spatiotemporal Transformer Neural Network for Time - Series Forecasting,

    Y. You, L. Zhang, P. Tao, S. Liu, and L. Chen, "Spatiotemporal Transformer Neural Network for Time - Series Forecasting," Entropy, vol. 24, no. 11, p. 1651, 2022

    2022

  35. [43]

    Predicting time series by data -driven spatiotemporal information transformation,

    P. Tao, X. Hao, J. Cheng, and L. Chen, "Predicting time series by data -driven spatiotemporal information transformation," Information Sciences, vol. 622, pp. 859 - 872, 2023/04/01/ 2023, doi: https://doi.org/10.1016/j.ins.2022.11.159

    2023 doi

  36. [44]

    Spatiotemporal information conversion machine for time -series forecasting,

    H. Peng, P. Chen, R. Liu, and L. Chen, "Spatiotemporal information conversion machine for time -series forecasting," Fundamental Research, 2022

    2022

  37. [45]

    Predicting multiple observations in complex systems through low -dimensional embeddings,

    T. Wu et al., "Predicting multiple observations in complex systems through low -dimensional embeddings," Nature Communications, vol. 15, no. 1, p. 2242, 2024

    2024

  38. [46]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy et al., "An image is worth 16x16 words: Transformers for image recognition at scale," arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  39. [47]

    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," arXiv preprint arXiv:2210.02186, 2022

    2022 arXiv

  40. [48]

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

    H. Wu, J. Xu, J. Wang, and M. Long, "Autoformer: Decomposition transformers with auto -correlation for long-term series forecasting," Advances in neural information processing systems, vol. 34, pp. 22419-22430, 2021

    2021

  41. [49]

    Modeling long-and short -term temporal patterns with deep neural networks,

    G. Lai, W. -C. Chang, Y. Yang, and H. Liu, "Modeling long-and short -term temporal patterns with deep neural networks," in The 41st international ACM SIGIR conference on research & development in information retrieval, 2018, pp. 95-104

    2018

  42. [50]

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

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, "Fedformer: Frequency enhanced decomposed transformer for long -term series forecasting," in International conference on machine learning , 2022: PMLR, pp. 27268-27286

    2022

  43. [51]

    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

    2022

  44. [52]

    Long-term forecasting with tide: Time -series dense encoder,

    A. Das, W. Kong, A. Leach, S. Mathur, R. Sen, and R. Yu, "Long-term forecasting with tide: Time -series dense encoder," arXiv preprint arXiv:2304.08424, 2023

    2023 arXiv

  45. [53]

    Lag-llama: Towards foundation models for time series forecasting,

    K. Rasul et al., "Lag-llama: Towards foundation models for time series forecasting," arXiv preprint arXiv:2310.08278, 2023

    2023 arXiv

  46. [54]

    TimeGPT -1,

    A. Garza and M. Mergenthaler -Canseco, "TimeGPT -1," arXiv preprint arXiv:2310.03589, 2023. 14

    2023 arXiv

  47. [55]

    Timer: Generative Pre -trained Transformers Are Large Time Series Models,

    Y. Liu, H. Zhang, C. Li, X. Huang, J. Wang, and M. Long, "Timer: Generative Pre -trained Transformers Are Large Time Series Models," in Forty-first International Conference on Machine Learning

  48. [56]

    Hamiltonian Systems and Transformation in Hilbert Space,

    B. O. Koopman, "Hamiltonian Systems and Transformation in Hilbert Space," Proceedings of the National Academy of Sciences, vol. 17, no. 5, pp. 315-318, 1931, doi: doi:10.1073/pnas.17.5.315

    1931 doi

  49. [57]

    Predicting Time Series from Short -Term High -Dimensional Data,

    H. Ma, T. Zhou, K. Aihara, and L. Chen, "Predicting Time Series from Short -Term High -Dimensional Data," International Journal of Bifurcation and Chaos, vol. 24, no. 12, p. 1430033, 2014, doi: 10.1142/s021812741430033x

    2014 doi

  50. [58]

    Cvt: Introducing convolutions to vision transformers,

    H. Wu et al. , "Cvt: Introducing convolutions to vision transformers," in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 22- 31

    2021

  51. [59]

    A decoder -only foundation model for time -series forecasting,

    A. Das, W. Kong, R. Sen, and Y. Zhou, "A decoder -only foundation model for time -series forecasting," arXiv preprint arXiv:2310.10688, 2023

    2023 arXiv

  52. [60]

    Moment: A family of open time -series foundation models,

    M. Goswami, K. Szafer, A. Choudhry, Y. Cai, S. Li, and A. Dubrawski, "Moment: A family of open time -series foundation models," arXiv preprint arXiv:2402.03885, 2024. Zijian Wang received his B.S. degree in physics from Xiamen University in 2021. He is currently a Ph.D. candid...

    2024 arXiv

Pith tools

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