Pith. sign in

REVIEW 2 major objections 5 minor 32 references

Forecasting works better when each future point soft-routes among three simple linear bases instead of one heavy backbone.

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 02:17 UTC pith:6YPWCIQL

load-bearing objection Solid methods paper: three simple linear bases plus a channel×horizon×phase gate deliver best average MSE/MAE on the usual suite, with real ablations and interpretable weights; period/phase assumption is the main scope limit. the 2 major comments →

arxiv 2607.09537 v1 pith:6YPWCIQL submitted 2026-07-10 cs.LG

GatedLinear: Adaptive Routing of Complementary Linear Bases for Time Series Forecasting

Qitai Tan , Ruiwen Gu , Yilin Su , Mo Li , Xu Lin , Xiao-Ping Zhang This is my paper
classification cs.LG
keywords time series forecastinglinear basesadaptive routingmixture of expertsphase-aligned recurrencedifference-based forecastingchannel-horizon-phase gatelightweight models
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.

Real multivariate series mix smooth trends, local drift, and strict phase-aligned cycles in different proportions by variable and horizon. Most deep models still push every series through one fixed mechanism—attention, spectral filters, or a single linear map—so accuracy suffers when the series’s dominant regime does not match that bias. GatedLinear instead treats forecasting as adaptive soft routing over three complementary linear bases: direct trend–seasonal projection, difference-based cumulative increments from the last observation, and a phase-aligned historical recurrence template plus residual. A Tri-Factorized Fusion Gate builds the mixing weights from three lookups—channel preference, horizon offset, and future-phase index—so each forecast point can emphasize the right regime without deep modules. On eight standard benchmarks the model records the best average MSE and MAE among recent linear, frequency, and Transformer baselines, while remaining parameter-light and producing readable gate maps that track known temporal traits.

Core claim

Point-wise soft routing among three complementary linear bases—global trend–seasonal projection, difference-based incremental evolution, and phase-aligned recurrence—controlled by a gate factorized over channel, horizon, and future phase, yields state-of-the-art or highly competitive average accuracy on standard multivariate forecasting benchmarks with a far smaller parameter footprint and explicitly interpretable branch weights.

What carries the argument

Tri-Factorized Fusion Gate: for each future step, channel, and basis the logit is the sum of a channel base tendency, a horizon-channel offset, and a phase-indexed bias (from decoder time marks or horizon modulo period); a temperature-softened softmax then produces the convex combination of the three linear forecasts.

Load-bearing premise

The series must have a usable base period and meaningful future phase indices so the phase-aligned basis and phase-indexed gate remain reliable; without them the third basis and fine-grained routing lose their footing.

What would settle it

On a controlled suite of weakly periodic or drifting-period series (or with phase marks deliberately corrupted), if GatedLinear’s average error no longer beats single-basis linear and recent Transformer baselines, and the learned phase weights become uncorrelated with seasonality traits, the central claim fails.

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

2 major / 5 minor

Summary. The paper proposes GatedLinear, a lightweight multivariate forecasting model that treats prediction as soft routing over three complementary linear bases: a global trend–seasonal projection (Eqs. 3–4), a difference-based incremental integrator from the last observation (Eqs. 5–6), and a phase-aligned recurrence template plus residual (Eqs. 7–8). Mixing is performed by a Tri-Factorized Fusion Gate whose logits factor into channel preference, horizon offset, and future-phase bias (Eqs. 9–10), yielding point-wise convex weights (Eq. 2). On eight standard benchmarks with L=336 and H∈{48,96,192,336}, the method reports the best average MSE/MAE and many per-dataset wins against recent linear, frequency, and Transformer baselines (Tables 1 and 4). Ablations (Table 2), gate–trait Spearman analysis (Fig. 6), efficiency comparison (Fig. 7), and an MMSE argument for the difference basis under integrated random walk with drift (Appendix C) support the mechanism story. Limitations of channel independence, period/phase dependence, and linear expressiveness are stated in Appendix B.

Significance. If the empirical average holds under the stated protocol, the work is a useful contribution to efficient time-series forecasting: it shows that adaptive selection among a small set of interpretable linear mechanisms can match or beat heavier single-backbone models while remaining parameter-light and producing readable routing patterns. Strengths include a clean theoretical motivation for the difference basis (Appendix C), systematic ablations of both bases and gate factors (Table 2), gate–trait correlations that align with intended semantics (Fig. 6), and an explicit efficiency comparison (Fig. 7). The design is reproducible in principle via the described architecture, public benchmarks, and planned code release. The main scientific value is the demonstration that structured, point-wise mechanism routing can be as important as architectural scale for heterogeneous series.

major comments (2)
  1. The strongest claim (Abstract; §4.2; Table 1 average) depends on a usable base period P and meaningful future phase indices for both the phase-aligned basis (Eqs. 7–8) and the phase-indexed gate logits (Eq. 9). Appendix B correctly notes that this weakens under weakly periodic series, drifting periods, holidays, or irregular sampling, and the fallback is horizon mod P. The main results suite is dominated by datasets with clear daily/weekly structure (ETT, Electricity, Traffic, Weather). Without at least one controlled weakly periodic or period-drift experiment (or a clear scope statement that the average-SOTA claim is restricted to series with stable P), the generality of the fine-grained routing claim remains under-supported relative to the abstract wording.
  2. Table 1 / Traffic and Appendix B: under the channel-independent design, GatedLinear is not best on Traffic (high-C, 862 sensors), where iTransformer and TQNet win. The paper acknowledges this, but the abstract’s “state-of-the-art or highly competitive” framing and the average-first narrative should more carefully separate average gains on the eight-dataset suite from the known limitation when cross-variable dependence is central. A short quantitative discussion of when channel independence is expected to lose would strengthen the central claim without changing the method.
minor comments (5)
  1. Checklist item 7 and experimental reporting: numbers are averaged over three seeds but no standard deviations, confidence intervals, or significance tests appear in Tables 1–2 or 4. Adding seed-level variability for the main average would improve statistical transparency.
  2. Figure 4–6 and §4.4: gate visualizations are informative; a brief note on how weights are aggregated (already in Appendix D) in the main text would help readers interpret the heatmaps without flipping to the appendix.
  3. Notation: Decomp (Eq. 3) and the moving-average kernel size (Appendix E.4, kernel 25) should be stated once in the main methodology for self-contained reading.
  4. Related work (Appendix A) is thorough; a short pointer in the main introduction to mixture-of-experts / Pathformer-style adaptive routing would better situate the Tri-Factorized Gate relative to existing adaptive pathways.
  5. Broader impacts are marked N/A in the checklist; a one-sentence note that societal impact depends on downstream deployment domains would align with common journal practice for methodological forecasting papers.

Circularity Check

0 steps flagged

No significant circularity: forecasts and SOTA claims rest on external benchmarks and standard MMSE derivation, not on self-defined or fitted-as-prediction loops.

full rationale

GatedLinear’s derivation chain is constructive and externally evaluated. The three bases (Eqs. 3–8) are explicit linear constructions (moving-average decomp + affine maps; first differences + cumulative sum from XL; phase template + residual), and the Tri-Factorized Fusion Gate (Eqs. 9–10) learns soft weights from training loss—none of these quantities is algebraically forced to equal the reported test MSE/MAE. Appendix C’s optimality argument for the difference basis is a standard MMSE derivation under an integrated random walk with drift; it shows structural identity to E[xT+h|x≤T]=xT+hc and shift-invariance of differencing, not a tautology that re-labels fitted parameters as predictions. Main accuracy claims (Table 1, ablations Table 2) are measured on public benchmarks against independently published baselines. The only author-overlapping citation (SynTSBench [20]) is used as motivational evidence of pattern heterogeneity, not as a uniqueness theorem or load-bearing premise that forces the architecture or the accuracy numbers. Gate–trait Spearman correlations (Fig. 6) are post-hoc diagnostics of learned weights, not inputs that define the metrics. No self-definitional loop, fitted-input-as-prediction, uniqueness import, or renaming of a known result as a first-principles derivation is present.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on standard forecasting protocol choices, a small set of hand-set hyperparameters (temperature, periods, cycle count, kernel size), the modeling assumption that three linear mechanisms plus soft routing suffice, and the period/phase indexing scheme. No physical constants or machine-checked lemmas are involved; free parameters are ordinary ML knobs, and invented entities are architectural components validated only inside this paper’s experiments.

free parameters (5)
  • fusion temperature τ = 0.8
    Controls sharpness of branch softmax; default 0.8 chosen by validation; sensitivity shown but still a free hyperparameter that affects routing hardness.
  • base period P per dataset = dataset-specific (e.g. 24, 96, 168, 144)
    Sets phase template length and phase-indexed gate table size; fixed from frequency (24/96/168/144) or validation-selected; load-bearing for phase basis.
  • number of phase cycles K = 2–3 (dataset-dependent)
    How many recent complete cycles form the recurrence template; target 2–3, clipped by window; sensitivity study shows degradation at large K.
  • moving-average kernel for Decomp = 25
    Trend–seasonal split window size set to 25; affects Global basis inputs.
  • learning rate / training schedule = lr=0.005, batch=128
    Adam lr 0.005, batch 128, ≤30 epochs, early-stop patience 20; standard but chosen for reported numbers.
axioms (4)
  • domain assumption Channel-independent temporal linear maps shared across variables are an adequate inductive bias for the evaluated multivariate benchmarks.
    Stated in §3 and Limitations B; Traffic results where cross-variate models win show the assumption is not universal.
  • domain assumption Future phase indices from decoder time marks (or horizon mod P) are meaningful selectors for both the recurrence basis and the gate.
    §3.1–3.2 and Appendix B; core to the claimed point-wise phase-aware routing.
  • standard math Under an integrated random walk with drift, the MMSE predictor is last level plus cumulative drift, motivating the difference basis.
    Appendix C derivation; standard ARIMA(0,1,0)+drift argument.
  • ad hoc to paper Convex soft combination of the three linear bases is expressive enough for the heterogeneity of the eight benchmarks.
    Design choice of the method; Limitations B admits nonlinear/regime-shift cases may exceed the pool.
invented entities (2)
  • Tri-Factorized Fusion Gate (channel base + horizon offset + phase-indexed bias logits) no independent evidence
    purpose: Produce point-wise, channel-wise, phase-aware mixing weights over the three bases without a heavy neural router.
    Architectural invention of the paper; evidence is internal ablations and interpretability plots, not an external independent measurement.
  • Tri-basis linear pool (Global Trend-Seasonal, Difference-Based Incremental, Phase-Aligned Recurrence) no independent evidence
    purpose: Cover smooth extrapolation, nonstationary drift, and exact cyclic reuse as complementary mechanisms.
    Composition of known linear operations into a fixed three-way pool is the paper’s mechanism library; validated only on the reported benchmarks.

pith-pipeline@v1.1.0-grok45 · 29205 in / 3272 out tokens · 39379 ms · 2026-07-13T02:17:50.925736+00:00 · methodology

0 comments
read the original abstract

Time series forecasting requires models to capture diverse, often mutually exclusive, temporal dynamics, from smooth trend continuation to nonstationary drift and strict phase-aligned recurrence. While recent deep learning models have improved accuracy, they typically force these diverse patterns through a single computational backbone governed by fixed algorithmic inductive biases (e.g., self-attention or spectral filtering). This single-mechanism approach often struggles with the profound heterogeneity of real-world series, where different variables and forecast horizons necessitate fundamentally different predictive treatments. To address this, we propose GatedLinear: a lightweight framework that frames forecasting as the adaptive routing of complementary linear bases. GatedLinear leverages a pool of three specialized mechanisms: a global trend-seasonal basis for smooth projection, a difference-based incremental basis for nonstationary drift, and a phase-aligned recurrence basis for explicit cyclic reuse. To dynamically orchestrate these distinct behaviors, we introduce a Tri-Factorized Fusion Gate that disentangles routing decisions into channel-specific preferences, horizon-aware offsets, and phase-indexed biases derived from known future time marks. This design allows the model to perform highly granular, point-wise soft routing across different predictive regimes without stacking computationally heavy neural modules. Experiments on standard benchmarks show that our method achieves state-of-the-art or highly competitive accuracy against recent complex foundational models, while offering explicitly interpretable routing patterns and operating with a substantially smaller parameter footprint.

Figures

Figures reproduced from arXiv: 2607.09537 by Mo Li, Qitai Tan, Ruiwen Gu, Xiao-Ping Zhang, Xu Lin, Yilin Su.

Figure 1
Figure 1. Figure 1: Trait heterogeneity in time series forecasting benchmarks. (a) Different datasets exhibit [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison between PatchTST and GatedLinear on data with distinct temporal [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed GatedLinear framework. The input sequence is processed by [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Learned fusion gate weights across horizons and variables. (a) Different datasets allocate [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Channel-wise branch-weight preferences across different datasets. Each point denotes a [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spearman correlation heatmap between channel-level temporal traits and learned branch preferences. The computation of Spearman corre￾lations is provided in Appendix D [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Forecasting visualization examples on eight benchmark datasets with input length [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Forecasting visualizations with weighted tri-branch contributions on ECL and Traffic, using [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Performance comparison of GatedLinear under different input lengths across benchmark [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Parameter sensitivity study on the ETTh1 dataset. The left panel varies the fusion [PITH_FULL_IMAGE:figures/full_fig_p022_11.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

32 extracted references · 6 linked inside Pith

  1. [1]

    G. E. P. Box, G. M. Jenkins, G. C. Reinsel, et al.Time series analysis: forecasting and control. John Wiley & Sons, 2015.arXiv:2311.10122

    Pith/arXiv arXiv 2015

  2. [2]

    Vector autoregression and causality: a theoretical overview and simulation study.Econometric reviews, 13(2):259–285, 1994

    Hiro Y Toda and Peter CB Phillips. Vector autoregression and causality: a theoretical overview and simulation study.Econometric reviews, 13(2):259–285, 1994

    1994

  3. [3]

    Springer Science & Business Media, 2008

    Rob Hyndman, Anne B Koehler, J Keith Ord, and Ralph D Snyder.Forecasting with exponential smoothing: the state space approach. Springer Science & Business Media, 2008

    2008

  4. [4]

    Deepar: Probabilistic forecasting with autoregressive recurrent networks.International journal of forecasting, 36(3): 1181–1191, 2020

    David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. Deepar: Probabilistic forecasting with autoregressive recurrent networks.International journal of forecasting, 36(3): 1181–1191, 2020

    2020

  5. [5]

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

    Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long-and short-term temporal patterns with deep neural networks. InThe 41st international ACM SIGIR conference on research & development in information retrieval, pages 95–104, 2018

    2018

  6. [6]

    Segrnn: Segment recurrent neural network for long-term time series forecasting.IEEE Internet of Things Journal, 2025

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

    2025

  7. [7]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271, 2018

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling.arXiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  8. [8]

    Scinet: Time series modeling and forecasting with sample convolution and interaction.Advances in neural information processing systems, 35:5816–5828, 2022

    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.Advances in neural information processing systems, 35:5816–5828, 2022

    2022

  9. [9]

    Are transformers effective for time series forecasting? InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128, 2023

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

    2023

  10. [10]

    Timemixer: Decomposable multiscale mixing for time series forecasting

    Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and JUN ZHOU. Timemixer: Decomposable multiscale mixing for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024

    2024

  11. [11]

    Bridging past and future: Distribution-aware alignment for time series forecasting.arXiv preprint arXiv:2509.14181, 2025

    Yifan Hu, Jie Yang, Tian Zhou, Peiyuan Liu, Yujin Tang, Rong Jin, and Liang Sun. Bridging past and future: Distribution-aware alignment for time series forecasting.arXiv preprint arXiv:2509.14181, 2025

    arXiv 2025

  12. [12]

    Autoformer: Decomposition trans- formers with auto-correlation for long-term series forecasting.Advances in neural information processing systems, 34:22419–22430, 2021

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition trans- formers with auto-correlation for long-term series forecasting.Advances in neural information processing systems, 34:22419–22430, 2021

    2021

  13. [13]

    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

  14. [14]

    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

  15. [15]

    Phaseformer: From patches to phases for efficient and effective time series forecasting.arXiv preprint arXiv:2510.04134, 2025

    Yiming Niu, Jinliang Deng, and Yongxin Tong. Phaseformer: From patches to phases for efficient and effective time series forecasting.arXiv preprint arXiv:2510.04134, 2025

    arXiv 2025

  16. [16]

    Filternet: Harnessing frequency filters for time series forecasting.Advances in Neural Information Processing Systems, 37:55115–55140, 2024

    Kun Yi, Jingru Fei, Qi Zhang, Hui He, Shufeng Hao, Defu Lian, and Wei Fan. Filternet: Harnessing frequency filters for time series forecasting.Advances in Neural Information Processing Systems, 37:55115–55140, 2024

    2024

  17. [17]

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

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. InThe Eleventh International Conference on Learning Representations, 2023. 10

    2023

  18. [18]

    Freqcycle: A multi-scale time-frequency analysis method for time series forecasting

    Boya Zhang, Shuaijie Yin, Huiwen Zhu, and Xing He. Freqcycle: A multi-scale time-frequency analysis method for time series forecasting. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 28159–28167, 2026

    2026

  19. [19]

    Temporal query net- work for efficient multivariate time series forecasting

    Shengsheng Lin, Haojun Chen, Haijie Wu, Chunyun Qiu, and Weiwei Lin. Temporal query net- work for efficient multivariate time series forecasting. InForty-second International Conference on Machine Learning, 2025

    2025

  20. [20]

    Syntsbench: Rethinking temporal pattern learning in deep learning models for time series

    Qitai Tan, Yiyun Chen, Mo Li, Ruiwen Gu, Yilin Su, and Xiao-Ping Zhang. Syntsbench: Rethinking temporal pattern learning in deep learning models for time series. InThe Thirty- ninth Annual Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2025

    2025

  21. [21]

    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. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 11106–11115, 2021

    2021

  22. [22]

    Mixlinear: Extreme low resource multivariate time series forecasting with 0.1 k parameters.arXiv preprint arXiv:2410.02081, 2024

    Aitian Ma, Dongsheng Luo, and Mo Sha. Mixlinear: Extreme low resource multivariate time series forecasting with 0.1 k parameters.arXiv preprint arXiv:2410.02081, 2024

    arXiv 2024

  23. [23]

    N-beats: Neural basis expansion analysis for interpretable time series forecasting.arXiv preprint arXiv:1905.10437, 2019

    Boris N Oreshkin, Dmitri Carpov, Nicolas Chapados, and Yoshua Bengio. N-beats: Neural basis expansion analysis for interpretable time series forecasting.arXiv preprint arXiv:1905.10437, 2019

    Pith/arXiv arXiv 1905

  24. [24]

    Nhits: Neural hierarchical interpolation for time series fore- casting

    Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. Nhits: Neural hierarchical interpolation for time series fore- casting. InProceedings of the AAAI conference on artificial intelligence, volume 37, pages 6989–6997, 2023

    2023

  25. [25]

    Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053, 2023

    Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O Arik, and Tomas Pfister. Tsmixer: An all-mlp architecture for time series forecasting.arXiv preprint arXiv:2303.06053, 2023

    Pith/arXiv arXiv 2023

  26. [26]

    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. InInternational conference on machine learning, pages 27268–27286. PMLR, 2022

    2022

  27. [27]

    Cyclenet: Enhancing time series forecasting through modeling periodic patterns.Advances in Neural Information Processing Systems, 37:106315–106345, 2024

    Shengsheng Lin, Weiwei Lin, Xinyi Hu, Wentai Wu, Ruichao Mo, and Haocheng Zhong. Cyclenet: Enhancing time series forecasting through modeling periodic patterns.Advances in Neural Information Processing Systems, 37:106315–106345, 2024

    2024

  28. [28]

    Tfb: Towards comprehensive and fair benchmarking of time series forecasting methods.arXiv preprint arXiv:2403.20150, 2024

    Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S Jensen, Zhenli Sheng, et al. Tfb: Towards comprehensive and fair benchmarking of time series forecasting methods.arXiv preprint arXiv:2403.20150, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Adaptive mixtures of local experts.Neural computation, 3(1):79–87, 1991

    Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts.Neural computation, 3(1):79–87, 1991

    1991

  30. [30]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Pith/arXiv arXiv 2017

  31. [31]

    Limitations

    Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qingsong Wen, Bin Yang, and Chenjuan Guo. Pathformer: Multi-scale transformers with adaptive pathways for time series forecasting. InICLR 2024: The Twelfth International Conference on Learning Representations., 2024. 11 A Additional Related Work Efficient linear and MLP-based forecasting.Long...

    2024

  32. [32]

    IRB or equivalent approval is therefore not applicable to the described study

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...