Pith. sign in

REVIEW 4 major objections 7 minor 64 references

Dominant periods from RFFT can steer deformable convolutions so they continuously align local phases in time series, not just fold them on fixed grids.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 20:38 UTC pith:LZJOC7OF

load-bearing objection Solid multi-task architecture paper: RFFT-anchored 1-D deformable conv + Gaussian RBF is a coherent, usable package with real ablations and code; the global-period-as-anchor claim is only partially stress-tested. the 4 major comments →

arxiv 2603.21718 v3 pith:LZJOC7OF submitted 2026-03-23 eess.SP

Frequency-Guided Deformable Networks for Continuous Phase Alignment

Wangye Jiang , Haoming Yang , Jian Xu , Jingya Zhang This is my paper
classification eess.SP
keywords time series analysisdeformable convolutionRFFTGaussian RBF interpolationphase alignmentanomaly detectionshort-term forecastingfrequency-guided sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Time series carry multi-scale periodic structure that standard Transformers and rigid CNNs either ignore or treat as static. This paper argues that the right move is to treat the dominant periods extracted by a real FFT as physical anchors that set the dilation of multi-branch deformable convolutions, then let the network learn continuous sub-pixel offsets around those anchors. Because discrete FFT periods are integers and real phases are not, the authors replace ordinary linear interpolation with a smooth one-dimensional Gaussian radial-basis kernel that keeps gradients flowing even when the offset is fractional. An asymmetric cascade further routes high-energy periods through small kernels (to keep sharp peaks) and low-energy periods through large kernels (to integrate weak trends). On short-term forecasting, anomaly detection and classification benchmarks the resulting ANCHOR network matches or exceeds recent baselines, supporting the claim that explicit frequency-guided continuous phase alignment is a useful inductive bias for general time-series modelling.

Core claim

Injecting RFFT-derived dominant periods as dilation anchors into multi-branch deformable convolutions, and compensating residual quantization error with a continuously differentiable 1-D Gaussian RBF interpolator, yields continuous phase alignment that jointly captures macroscopic periodic priors and microscopic local deformations, producing competitive or best multi-task results without task-specific redesign.

What carries the argument

Frequency-Guided Deformable Module (FGDM): RFFT periods become the base dilation of 1-D deformable kernels whose continuous offsets are sampled by Gaussian RBF interpolation; orthogonal channel splits and asymmetric cascade routing then balance strong-signal and weak-feature extraction.

Load-bearing premise

A single global RFFT whose top-K energy peaks are averaged across batch and channels supplies stable enough period anchors for every local non-stationary patch the network will later sample.

What would settle it

On a controlled suite of multi-scale series whose local periods drift faster than the learned offsets can track (or whose true periods lie midway between discrete FFT bins), measure whether the continuous-offset model still reduces phase-alignment error and improves OWA/AUPRC relative to the same architecture with frozen integer dilations.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The manuscript proposes ANCHOR, a hierarchical CNN backbone for multi-task time series analysis. It extracts dominant periods via a global RFFT (energy-averaged over batch/channel, DC removed, top-K), injects those integer periods as dilation anchors into multi-branch 1D deformable convolutions, and replaces bilinear sampling with a C^∞ 1D Gaussian RBF interpolator to compensate picket-fence / sub-pixel phase error. Orthogonal channel partitioning plus an asymmetric cascade routes high-energy periods to smaller kernels and low-energy periods to larger ones. Empirically, the model reports strong or best results on M4 short-term forecasting (avg OWA 0.836), reconstruction/prediction anomaly detection (SMD/SWaT/PSM and UCR AUPRC 0.8112), and 18 UEA classification subsets, with progressive ablations (1D vs bilinear vs Gaussian; routing order; top-k; FGDM plug-in into other backbones) and a closed-form Gaussian gradient derivation.

Significance. If the frequency-anchor + continuous-offset mechanism is genuinely responsible for the gains, the work supplies a concrete, reusable inductive bias that links classical spectral priors to modern deformable sampling, with a clean mathematical treatment of gradient flow under Gaussian RBF (useful beyond this paper). Multi-task evaluation across forecasting, anomaly detection, and classification, plus public code, is a genuine strength relative to single-task architecture papers. The contribution is incremental rather than paradigm-shifting (TimesNet-style period extraction + DCNv4-style offsets + energy routing), but the combination and the continuous-phase framing are of clear interest to the eess.SP / time-series community if the causal role of the RFFT anchors is more tightly established.

major comments (4)
  1. §2.1–2.2 and Conclusion: The central claim is that discrete RFFT periods Tk act as stable macroscopic physical anchors that learned continuous offsets Δpn + Gaussian RBF refine into true phase alignment. The extractor is a single global, batch/channel-averaged spectrum with fixed top-K and floor(L/fk). On non-stationary multi-source series (M4, industrial sets) local periods can drift inside the window faster than a static prior plus sub-pixel offsets can track; the Conclusion itself concedes fixed hyperparameters and limited dynamic perception under strong heterogeneity. Ablations (Table 6, Fig. 4, top-k sensitivity) show Gaussian helps over bilinear and k≈3 is mild, but never isolate whether the RFFT-derived Tk themselves remain accurate or helpful when local spectra change. A load-bearing control is missing: deformable+Gaussian with (i) random/fixed non-spectral dilations, (ii) per-wi
  2. Tables 2–5 and 6–8: All headline metrics are single-point estimates with no standard deviations over seeds, no repeated runs, and no statistical significance tests. Given small absolute margins on several M4 subsets and UEA datasets (and occasional second-place or mixed rankings), the claim of “best or solid” multi-task superiority is not yet statistically grounded. At minimum, report mean±std over ≥3 seeds for the main tables and the core ablations; ideally add paired tests or confidence intervals on OWA/AUPRC/F1.
  3. §2.2.3 vs §3.3.5 (Table 8): The asymmetric routing story is load-bearing for contribution 2, yet the physical motivation is not fully consistent across sections. Methodology sorts periods by spectral energy and assigns high-energy → small kernels / low-energy → large kernels; the ablation defines ANCHOR-Asc as low-frequency → narrow and high-frequency → wide, reports Asc superior, and then explains wide fields on high-frequency components to suppress false alarms. High-energy is not identical to low-frequency. Please (a) state explicitly how energy ranking maps onto the K schedule and onto frequency order, (b) reconcile the two narratives, and (c) confirm which policy is used in all main results. A short controlled swap of energy-order vs frequency-order would make the criterion falsifiable.
  4. §3.1–3.2 experimental protocol: For anomaly detection the paper mixes reconstruction-error and forecasting paradigms and deliberately strips composite criteria from Anomaly Transformer for “fairness,” while UCR uses a fixed window of 96 and no PA. These choices are defensible but under-specified for reproducibility (exact thresholding, how F1e/F1d are computed, train/val splits per UCR series). Please add a concise protocol appendix or pointer so that the strong UCR AUPRC / Delay-F1 numbers can be independently verified.
minor comments (7)
  1. Title vs acronym: the title is “Frequency-Guided Deformable Networks…” while the abstract/body name the model ANCHOR (“Adaptive Network Based on Cascaded Harmonic Offset Routing”). Align naming in title, abstract, and first mention.
  2. §2.2.2: The Gaussian gradient derivation is valuable; consider moving the long intermediate algebra to an appendix and keeping only the final mean-shift form in the main text for readability.
  3. Notation: Γ, ϕc, ϕv, DefOp, and the recursive state equation for yi are introduced densely; a short symbol table or expanded walk-through of one cascade step would help.
  4. Fig. 3 / Fig. 4 / Fig. 6: captions and axis labels are hard to read in the manuscript PDF; increase font size and ensure color-blind-safe palettes.
  5. Typos / wording: “imports a continuously differentiable…”, “denotes element-wise multiplication, denotes element-wise multiplication”, occasional doubled phrases, and “picket-fence effects” vs “picket fence effect” consistency.
  6. Related work: briefly position against other frequency-aware deformable or adaptive-kernel time-series CNNs beyond TimesNet/ModernTCN/DCNv4 to sharpen novelty.
  7. Hyperparameters: free parameters (K, σ, N, kernel schedule) are acknowledged; a short default table used for all main experiments would aid reproduction.

Circularity Check

0 steps flagged

No circularity: RFFT periods are data-derived inductive biases injected into deformable sampling; model is trained end-to-end and evaluated on external public benchmarks with no claim reducing to a fitted free parameter or self-citation by construction.

full rationale

The paper's derivation chain is architectural and empirical, not a closed mathematical prediction. Section 2.1 extracts dominant periods Tk = floor(L/fk) via a single global RFFT (energy-averaged, DC zeroed, top-K) directly from the input tensor X; these are then used as fixed dilation anchors in the FGDM sampling equation pn = p0 + Tk · n + Δpn (Sec. 2.2.1). The continuous offsets Δpn and Gaussian RBF weights are learned by standard gradient descent on task losses; the Gaussian gradient derivation (Sec. 2.2.2) is an independent calculus identity showing C∞ flow versus bilinear truncation, not a tautology. Asymmetric routing and channel partitioning (Sec. 2.2.3) are design choices whose complexity reduction is algebraic, not circular. All quantitative claims (M4 OWA/SMAPE, UCR AUPRC, UEA accuracies, ablations in Tables 6–8 and Figs. 4–6) are obtained by training on and scoring against external public benchmarks under controlled protocols; no free parameter is fitted to a target quantity and then re-presented as a prediction of that quantity. There are no load-bearing self-citations of uniqueness theorems or prior author results that force the architecture; references are to standard external literature (TimesNet, DCNv4, etc.). The conclusion itself flags the fixed-hyperparameter limitation under extreme non-stationarity, confirming the claims remain falsifiable rather than definitional. Consequently the strongest claim (synergistic macro-period / micro-phase modeling yielding multi-task gains) stands or falls on external evidence and does not reduce to its inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 2 invented entities

The central empirical claim rests on standard signal-processing and deep-learning primitives plus a handful of architectural free parameters (top-k, Gaussian width σ, kernel-size schedule, channel-partition count N) that are chosen by validation or fixed by convention. No new physical constants or unobserved particles are postulated; the “physical anchors” are simply the RFFT periods of the observed window.

free parameters (3)
  • top-K frequency count = typically 3
    Number of dominant periods retained after RFFT; sensitivity analysis shows k=3 is preferred but is a free hyper-parameter fitted on validation performance.
  • Gaussian RBF scale σ
    Controls kernel width of the continuous interpolator; treated as a tunable hyper-parameter without theoretical derivation of an optimal value.
  • kernel-size schedule K and partition count N
    Monotonically increasing receptive-field sizes and number of orthogonal channel groups; chosen by architecture search / convention rather than derived.
axioms (3)
  • domain assumption Dominant spectral peaks of a finite RFFT window supply useful macroscopic period priors for non-stationary real-world series.
    Invoked in §2.1; standard in period-aware models (TimesNet et al.) but not universally true for aperiodic or rapidly drifting signals.
  • standard math A C^∞ Gaussian radial-basis interpolator yields more stable sub-pixel offset gradients than C^0 bilinear interpolation.
    Derived in §2.2.2; mathematically correct under the stated smoothness assumptions.
  • ad hoc to paper High-energy periods should be routed to small kernels and low-energy periods to large kernels (asymmetric energy-compensating routing).
    Design choice introduced in §2.2.3 and ablated in Table 8; empirically helpful but not forced by first principles.
invented entities (2)
  • Frequency-Guided Deformable Module (FGDM) no independent evidence
    purpose: Plug-and-play block that injects RFFT periods into deformable dilation and performs progressive orthogonal cascade routing.
    Core architectural invention of the paper; independent evidence is limited to the empirical tables and ablations inside this work.
  • Asymmetric energy-compensating routing no independent evidence
    purpose: Dynamically balance strong-signal sharpness versus weak-signal integration via channel partitioning and kernel-size assignment.
    New routing heuristic; validated only by the paper’s own ascending/descending ablation.

pith-pipeline@v1.1.0-grok45 · 24083 in / 2966 out tokens · 32355 ms · 2026-07-13T20:38:40.537084+00:00 · methodology

0 comments
read the original abstract

The core of time series analysis lies in effectively modeling the physical laws within complex signals. Existing Transformer and Convolution Neural Network (CNN) architectures are often constrained by insufficient temporal inductive bias, restricted frequency extraction capabilities, or weak local phase alignment. To this end, this paper proposes Adaptive Network Based on Cascaded Harmonic Offset Routing (ANCHOR), an Adaptive Network based on Cascaded Harmonic Offset Routing. The model utilizes the Real Fast Fourier Transform (RFFT) to extract explicit dominant periods, injecting them as physical anchors into the dilation operators of multi-branch deformable convolutions. This guides the adaptive optimization of sampling locations in the time domain, achieving synergistic modeling of macroscopic periodic priors and microscopic geometric deformations. Furthermore, to address the quantization errors and picket-fence effects introduced by the discrete RFFT, this paper imports a continuously differentiable 1D Gaussian Radial Basis Function interpolation operator to replace traditional linear interpolation. This maintains the differentiability of the interpolation process and enhances the accuracy of sub-pixel phase compensation. Additionally, ANCHOR introduces an asymmetric routing mechanism and orthogonal channel partitioning to dynamically balance the extraction weights between high-energy strong signals and low-energy weak features. Multi-task benchmark experiments demonstrate that ANCHOR achieves the best or solid performance in short-term forecasting, anomaly detection, and time series classification tasks. Code is available at https://github.com/Jwy-EE/Anchor_pub

Figures

Figures reproduced from arXiv: 2603.21718 by Haoming Yang, Jian Xu, Jingya Zhang, Wangye Jiang.

Figure 1
Figure 1. Figure 1: Structure diagram of FGDM For a given local one-dimensional feature sequence x ∈ R L, under standard convolution, the relative sampling grid point set for a convolution kernel of size Sk is G = {−⌊Sk/2⌋, ..., ⌊Sk/2⌋}. After injecting the physical period Tk, the actual continuous spatial absolute coordinate pn of the n-th sampling point (n ∈ G) is reconstructed as: pn = p0 + Tk · n + ∆pn, where ∆pn is the c… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the ANCHOR model [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of M4-Hourly short-term forecasting results given by models [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparative results of quantization fragment compensation capability [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity Analysis of Top-k Frequency Components [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Accuracy-Throughput Trade-off Across Architectures [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗

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

64 extracted references · 8 linked inside Pith

  1. [1]

    Foundation models for time series analysis: A tutorial and survey

    Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6555–6565, 2024. 19

    2024

  2. [2]

    Rethinking the role of llms in time series forecasting.arXiv preprint arXiv:2602.14744, 2026

    Xin Qiu, Junlong Tong, Yirong Sun, Yunpu Ma, Wei Zhang, and Xiaoyu Shen. Rethinking the role of llms in time series forecasting.arXiv preprint arXiv:2602.14744, 2026

    arXiv 2026

  3. [3]

    Timexer: Empowering transformers for time series forecasting with exogenous variables

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Guo Qin, Haoran Zhang, Yong Liu, Yunzhong Qiu, Jianmin Wang, and Mingsheng Long. Timexer: Empowering transformers for time series forecasting with exogenous variables. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume...

    2024

  4. [4]

    Times- net: Temporal 2d-variation modeling for general time series analysis

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Times- net: Temporal 2d-variation modeling for general time series analysis. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  5. [5]

    Themis: Unlocking pretrained knowledge with foundation model embeddings for anomaly detection in time series.arXiv e-prints, pages arXiv–2510, 2025

    Yadav Mahesh Lorik, Kaushik Sarveswaran, Nagaraj Sundaramahalingam, and Aravin- dakumar Venugopalan. Themis: Unlocking pretrained knowledge with foundation model embeddings for anomaly detection in time series.arXiv e-prints, pages arXiv–2510, 2025

    2025

  6. [6]

    A survey of network anomaly detection techniques.Journal of network and computer applications, 60:19–31, 2016

    Mohiuddin Ahmed, Abdun Naser Mahmood, and Jiankun Hu. A survey of network anomaly detection techniques.Journal of network and computer applications, 60:19–31, 2016

    2016

  7. [7]

    Opentslm: Time-series language models for reasoning over multivariate medical text-and time-series data.arXiv preprint arXiv:2510.02410, 2025

    Patrick Langer, Thomas Kaar, Max Rosenblattl, Maxwell A Xu, Winnie Chow, Martin Maritsch, Robert Jakob, Ning Wang, Juncheng Liu, Aradhana Verma, et al. Opentslm: Time-series language models for reasoning over multivariate medical text-and time-series data.arXiv preprint arXiv:2510.02410, 2025

    arXiv 2025

  8. [8]

    Components of a new research resource for complex physiologic signals.Phys- ioBank, PhysioToolkit, and Physionet, 2000

    Ary L Goldberger, Lu´ ıs Amaral, Leon Glass, Jeffrey M Hausdorff, Plamen Ch Ivanov, Roger G Mark, Joseph E Mietus, George B Moody, Chung-Kang Peng, and H Eugene Stanley. Components of a new research resource for complex physiologic signals.Phys- ioBank, PhysioToolkit, and Physionet, 2000

    2000

  9. [9]

    Units: A unified multi-task time series model

    Shanghua Gao, Teddy Koker, Owen Queen, Thomas Hartvigsen, Theodoros Tsiligkaridis, and Marinka Zitnik. Units: A unified multi-task time series model. In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors,Advances in Neural Information Processing Systems, volume 37, pages 140589–140631. Curran Asso- ciates, Inc., 2024

    2024

  10. [10]

    John Wiley & Sons, 2015

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

    2015

  11. [11]

    Forecasting seasonals and trends by exponentially weighted moving aver- ages.International journal of forecasting, 20(1):5–10, 2004

    Charles C Holt. Forecasting seasonals and trends by exponentially weighted moving aver- ages.International journal of forecasting, 20(1):5–10, 2004

    2004

  12. [12]

    Dlinear makes efficient long-term predictions

    Chaoqun Su. Dlinear makes efficient long-term predictions. InProceedings of ACM Con- ference (Baidu KDD Cup), 2022

    2022

  13. [13]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    Robust speech recognition via large-scale weak supervision

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine Mcleavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In Andreas Krause, 20 Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scar- lett, editors,Proceedings of the 40th International Conference on Machine Learning, volume 202 ofPr...

    2023

  16. [16]

    Seamlessm4t: Massively multilingual & multimodal machine translation.arXiv preprint arXiv:2308.11596, 2023

    Lo¨ ıc Barrault, Yu-An Chung, Mariano Cora Meglioli, David Dale, Ning Dong, Paul- Ambroise Duquenne, Hady Elsahar, Hongyu Gong, Kevin Heffernan, John Hoffman, et al. Seamlessm4t: Massively multilingual & multimodal machine translation.arXiv preprint arXiv:2308.11596, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4195–4205, October 2023

    2023

  18. [18]

    Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015–4026, October 2023

    2023

  19. [19]

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

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. volume 35, pages 11106–11115, May 2021

    2021

  20. [20]

    Autoformer: Decomposi- tion transformers with auto-correlation for long-term series forecasting

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposi- tion transformers with auto-correlation for long-term series forecasting. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 22419–22430. Curran Associates, Inc., 2021

    2021

  21. [21]

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

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  22. [22]

    Timegpt-1.arXiv preprint arXiv:2310.03589, 2023

    Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco. Timegpt-1.arXiv preprint arXiv:2310.03589, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Maddix, Hao Wang, Michael W

    Abdul Fatir Ansari, Lorenzo Stella, Ali Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Ma- honey, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Bernie Wang. Chronos: Learning the lan...

    2024

  24. [24]

    Unified training of universal time series forecasting transformers

    Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. InForty-first International Conference on Machine Learning, 2024

    2024

  25. [25]

    Scinet: Time series modeling and forecasting with sample convolution and interaction

    Minhao LIU, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia LAI, Lingna Ma, and Qiang Xu. Scinet: Time series modeling and forecasting with sample convolution and interaction. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 5816–5828. Curran Associates, Inc., 2022

    2022

  26. [26]

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

    Huiqiang Wang, Jian Peng, Feihu Huang, Jince Wang, Junhui Chen, and Yifei Xiao. MICN: Multi-scale local and global context modeling for long-term series forecasting. InThe Eleventh International Conference on Learning Representations, 2023. 21

    2023

  27. [27]

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

    Luo donghao and wang xue. ModernTCN: A modern pure convolution structure for general time series analysis. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  28. [28]

    Are transformers effective for time series forecasting? volume 37, pages 11121–11128, Jun

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? volume 37, pages 11121–11128, Jun. 2023

    2023

  29. [29]

    Why do transformers fail to forecast time series in-context?arXiv preprint arXiv:2510.09776, 2025

    Yufa Zhou, Yixiao Wang, Surbhi Goel, and Anru R Zhang. Why do transformers fail to forecast time series in-context?arXiv preprint arXiv:2510.09776, 2025

    arXiv 2025

  30. [30]

    Transformers in time series: A survey.arXiv preprint arXiv:2202.07125, 2022

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. Transformers in time series: A survey.arXiv preprint arXiv:2202.07125, 2022

    Pith/arXiv arXiv 2022

  31. [31]

    Yuwen Xiong, Zhiqi Li, Yuntao Chen, Feng Wang, Xizhou Zhu, Jiapeng Luo, Wenhai Wang, Tong Lu, Hongsheng Li, Yu Qiao, Lewei Lu, Jie Zhou, and Jifeng Dai. Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5652–5661, 2024

    2024

  32. [32]

    Uniconvnet: Expanding effective receptive field while maintaining asymptotically gaussian distribution for convnets of any scale.ArXiv, abs/2508.09000, 2025

    Yuhao Wang and Wei Xi. Uniconvnet: Expanding effective receptive field while maintaining asymptotically gaussian distribution for convnets of any scale.ArXiv, abs/2508.09000, 2025

    Pith/arXiv arXiv 2025

  33. [33]

    M4 dataset.https://github.com/M4Competition/M4-methods/ tree/master/Dataset, 2018

    Spyros Makridakis. M4 dataset.https://github.com/M4Competition/M4-methods/ tree/master/Dataset, 2018

    2018

  34. [34]

    Semantic-enhanced time-series forecasting via large language models.ArXiv, abs/2508.07697, 2025

    Hao Liu, Chun Yang, Xiaoxing Zhang, and Xiaobin Zhu. Semantic-enhanced time-series forecasting via large language models.ArXiv, abs/2508.07697, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Timemixer++: A general time series pattern machine for universal predictive analysis

    Shiyu Wang, Jiawei LI, Xiaoming Shi, Zhou Ye, Baichuan Mo, Wenze Lin, Ju Shengtong, Zhixuan Chu, and Ming Jin. Timemixer++: A general time series pattern machine for universal predictive analysis. InThe Thirteenth International Conference on Learning Rep- resentations, 2025

    2025

  36. [36]

    Context-alignment: Activating and enhancing llm capabilities in time series.arXiv preprint arXiv:2501.03747, 2025

    Yuxiao Hu, Qian Li, Dongxiao Zhang, Jinyue Yan, and Yuntian Chen. Context-alignment: Activating and enhancing llm capabilities in time series.arXiv preprint arXiv:2501.03747, 2025

    arXiv 2025

  37. [37]

    Time- VLM: Exploring multimodal vision-language models for augmented time series forecasting

    Siru Zhong, Weilin Ruan, Ming Jin, Huan Li, Qingsong Wen, and Yuxuan Liang. Time- VLM: Exploring multimodal vision-language models for augmented time series forecasting. InForty-second International Conference on Machine Learning, 2025

    2025

  38. [38]

    Autotimes: Autoregressive time series forecasters via large language models.Advances in Neural Infor- mation Processing Systems, 37:122154–122184, 2024

    Yong Liu, Guo Qin, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Autotimes: Autoregressive time series forecasters via large language models.Advances in Neural Infor- mation Processing Systems, 37:122154–122184, 2024

    2024

  39. [39]

    S2ip-llm: Semantic space informed prompt learning with llm for time series forecast- ing

    Zijie Pan, Yushan Jiang, Sahil Garg, Anderson Schneider, Yuriy Nevmyvaka, and Dongjin Song. S2ip-llm: Semantic space informed prompt learning with llm for time series forecast- ing. InInternational Conference on Machine Learning, 2024

    2024

  40. [40]

    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. Time-LLM: Time series forecasting by reprogramming large language models. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  41. [41]

    One fits all: Power general time series analysis by pretrained lm.Advances in Neural Information Processing Systems 36, 2023

    Tian Zhou, Peisong Niu, Xue Wang, Liang Sun, and Rong Jin. One fits all: Power general time series analysis by pretrained lm.Advances in Neural Information Processing Systems 36, 2023. 22

    2023

  42. [42]

    itransformer: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  43. [43]

    Robust anomaly detection for multivariate time series through stochastic recurrent neural network

    Ya Su, Youjian Zhao, Chenhao Niu, Rong Liu, Wei Sun, and Dan Pei. Robust anomaly detection for multivariate time series through stochastic recurrent neural network. InPro- ceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2828–2837, 2019

    2019

  44. [44]

    Swat: A water treatment testbed for research and training on ics security

    Aditya P Mathur and Nils Ole Tippenhauer. Swat: A water treatment testbed for research and training on ics security. In2016 international workshop on cyber-physical systems for smart water networks (CySWater), pages 31–36. IEEE, 2016

    2016

  45. [45]

    Practical approach to asyn- chronous multivariate time series anomaly detection and localization

    Ahmed Abdulaal, Zhuanghua Liu, and Tomer Lancewicki. Practical approach to asyn- chronous multivariate time series anomaly detection and localization. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, pages 2485–2494, 2021

    2021

  46. [46]

    Anomaly transformer: Time series anomaly detection with association discrepancy

    Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. InInternational Conference on Learning Representations, 2022

    2022

  47. [47]

    KAN-AD: Time series anomaly detection with kolmogorov–arnold networks

    Quan Zhou, Changhua Pei, Fei Sun, HanJing, Zhengwei Gao, Haiming Zhang, Gaogang Xie, Dan Pei, and Jianhui li. KAN-AD: Time series anomaly detection with kolmogorov–arnold networks. InForty-second International Conference on Machine Learning, 2025

    2025

  48. [48]

    Non-stationary transform- ers: Exploring the stationarity in time series forecasting.Advances in neural information processing systems, 35:9881–9893, 2022

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Non-stationary transform- ers: Exploring the stationarity in time series forecasting.Advances in neural information processing systems, 35:9881–9893, 2022

    2022

  49. [49]

    Time-series anomaly detection service at mi- crosoft

    Hansheng Ren, Bixiong Xu, Yujing Wang, Chao Yi, Congrui Huang, Xiaoyu Kou, Tony Xing, Mao Yang, Jie Tong, and Qi Zhang. Time-series anomaly detection service at mi- crosoft. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 3009–3017, 2019

    2019

  50. [50]

    Sand: Streaming subsequence anomaly detection.Proc

    Paul Boniol, John Paparrizos, Themis Palpanas, and Michael J Franklin. Sand: Streaming subsequence anomaly detection.Proc. VLDB Endow., 14(10):1717–1729, 2021

    2021

  51. [51]

    Jennings

    Shreshth Tuli, Giuliano Casale, and Nicholas R. Jennings. Tranad: Deep transformer networks for anomaly detection in multivariate time series data.Proc. VLDB Endow., 15:1201–1214, 2022

    2022

  52. [52]

    Lof: identify- ing density-based local outliers

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and J¨ org Sander. Lof: identify- ing density-based local outliers. InProceedings of the 2000 ACM SIGMOD international conference on Management of data, pages 93–104, 2000

    2000

  53. [53]

    FITS: Modeling time series with$10k$parameters

    Zhijian Xu, Ailing Zeng, and Qiang Xu. FITS: Modeling time series with$10k$parameters. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  54. [54]

    Revisiting vae for unsupervised time series anomaly detection: A frequency perspective

    Zexin Wang, Changhua Pei, Minghua Ma, Xin Wang, Zhihan Li, Dan Pei, Saravan Rajmo- han, Dongmei Zhang, Qingwei Lin, Haiming Zhang, et al. Revisiting vae for unsupervised time series anomaly detection: A frequency perspective. InProceedings of the ACM web conference 2024, pages 3096–3105, 2024

    2024

  55. [55]

    Shroff, and Puneet Agarwal

    Pankaj Malhotra, Lovekesh Vig, Gautam M. Shroff, and Puneet Agarwal. Long short term memory networks for anomaly detection in time series. InThe European Symposium on Artificial Neural Networks, 2015. 23

    2015

  56. [56]

    Hou, and Max Tegmark

    Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljacic, Thomas Y. Hou, and Max Tegmark. KAN: Kolmogorov–arnold networks. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  57. [57]

    Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress.IEEE transactions on knowledge and data engineering, 35(3):2421–2429, 2021

    Renjie Wu and Eamonn J Keogh. Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress.IEEE transactions on knowledge and data engineering, 35(3):2421–2429, 2021

    2021

  58. [58]

    The uea multivariate time series classifi- cation archive, 2018.arXiv preprint arXiv:1811.00075, 2018

    Anthony Bagnall, Hoang Anh Dau, Jason Lines, Michael Flynn, James Large, Aaron Bostrom, Paul Southam, and Eamonn Keogh. The uea multivariate time series classifi- cation archive, 2018.arXiv preprint arXiv:1811.00075, 2018

    Pith/arXiv arXiv 2018

  59. [59]

    A transformer-based framework for multivariate time series representation learn- ing

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

    2021

  60. [60]

    Mambasl: Exploring single-layer mamba for time series classification

    Yoo-Min Jung and Leekyung Kim. Mambasl: Exploring single-layer mamba for time series classification. InThe Fourteenth International Conference on Learning Representations, 2026

    2026

  61. [61]

    Tscmamba: Mamba meets multi-view learning for time series classification.Information Fusion, 120:103079, 2025

    Md Atik Ahamed and Qiang Cheng. Tscmamba: Mamba meets multi-view learning for time series classification.Information Fusion, 120:103079, 2025

    2025

  62. [62]

    Shedding light on time series classification using interpretability gated networks

    Yunshi Wen, Tengfei Ma, Ronny Luss, Debarun Bhattacharjya, Achille Fokoue, and Anak Agung Julius. Shedding light on time series classification using interpretability gated networks. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  63. [63]

    Tslanet: Rethinking transformers for time series representation learning

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, and Xiaoli Li. Tslanet: Rethinking transformers for time series representation learning. InInternational Conference on Machine Learning, 2024

    2024

  64. [64]

    FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. InProceed- ings of the 39th International Conference on Machine Learning (ICML), volume 162, pages 27268–27286. PMLR, 2022. 24

    2022