Pith. sign in

REVIEW 5 major objections 6 minor 40 references

Does Scaling Law Apply in Time Series Forecasting?

T0 review · 5 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read A linear model with a few hundred thousand parameters outperforms large-scale transformer-based forecasters on seven benchmark datasets.

desk verdict The ALinear design is a sensible lightweight follow-on to DLinear, but the paper's headline 'less than 1% of parameters' claim rests on a TimesNet parameter count that looks inflated by roughly 1000x and needs correction before the efficiency story holds. read the letter →

arxiv 2505.10172 v1 pith:VE732KPM submitted 2025-05-15 cs.LG cs.AI

Zeyan Li , Libing Chen , Yin Tang This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriesforecastingscalinglawparameterefficiencyhorizon-awaredecompositionadaptivelinearmodellong-termtrend-seasonalfrequencyattenuation
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 asks whether the scaling law that drives large language and vision models—more parameters, better performance—also applies to time series forecasting. It argues the answer is no and presents ALinear, an ultra-lightweight linear model that, in its experiments, outperforms ten established forecasting models on seven univariate datasets while using less than 1% of their parameters. The design principle is that forecasting changes character with the prediction horizon: trend stays predictable longer, while seasonal and high-frequency components become noise, so the model learns to rebalance components and attenuate high frequencies as the horizon grows. If the result holds, it would redirect research effort from parameter expansion toward horizon-aware efficiency and make accurate long-horizon forecasting feasible in resource-constrained settings.

What carries the argument

The load-bearing object is ALinear's horizon-parameterized decomposition-recombination loop. The trend component is extracted by a moving average whose kernel size depends on the forecast horizon, $\alpha(H)=\min(\max(k_1+k_2 H, w_{\min}), w_{\max})$, with learnable $k_1,k_2$; the seasonal part is the residual $X-T_H$. Each component gets its own linear projection ($\hat T_H = W_T T_H + b_T$, $\hat S_H = W_S S_H + b_S$), and the seasonal forecast is then attenuated by an exponential decay $\hat S_H(t)\exp(-\lambda t)$ with $\lambda = \delta/H$, so high-frequency content fades as the horizon grows. The two branches are recombined with horizon-dependent weights $\beta_T(H)=\sigma(v_1+v_2 H)$ and $\beta_S(H)=1-\beta_T(H)$. This single mechanism replaces attention: it lets the model put more weight on trend and suppress seasonality at long horizons, which is exactly the behavior the paper says is missing from fixed-decomposition baselines.

What would settle it

A direct replication of Table 1 using the released checkpoints would settle the claim: if ALinear's MSEs are not reproduced, or if the actual TimesNet checkpoint has tens of millions of parameters rather than 1,199.72M, the headline claim would not stand.

Watch

Extended reading notes

Core claim

The paper's central claim is that the scaling law that governs large language and vision models does not transfer to time series forecasting. The evidence is ALinear, a linear model with $2HT+2H+4$ parameters (about 297K in the reported configuration), which the authors say beats ten state-of-the-art forecasting models—including Transformers, Autoformer, Crossformer, PatchTST, and TimesNet—on seven univariate benchmarks across horizons 48 to 960, while using less than 1% of the baselines' parameters. The paper further claims the advantage grows with horizon: at 960 steps ALinear's MSE is on average 23.7% lower than the second-best model, and the model's learned trend-seasonal balance shifts with dataset characteristics rather than following a fixed schedule. It also introduces Parameter-Normalized Performance, $\mathrm{PNP}=100/(\mathrm{Metric}\cdot\log(\#\text{parameters}))$, as a parameter-aware evaluation metric, on which ALinear ranks first.

Load-bearing premise

The quantitative claim that ALinear uses less than 1% of a large model's parameters rests on the reported parameter counts for the baselines, especially TimesNet's stated 1,199.72M; if those counts are wrong, the efficiency comparison loses its basis.

Editorial extensions

If this is right

  • On the seven univariate benchmarks tested, ALinear's roughly 0.3-million-parameter model matches or beats models with millions to over a billion parameters across horizons from 48 to 960 steps.
  • The accuracy advantage grows with horizon: at 960 steps ALinear averages a 23.7% lower MSE than the second-best model.
  • Because ALinear's time and space complexity scale as the product of input length and horizon, the reported accuracy is available for real-time and resource-constrained deployment without specialized hardware.
  • The learned trend-seasonal balance is dataset-dependent, so fixed-decomposition linear baselines such as DLinear are expected to lose ground at longer horizons.
  • Under the proposed PNP metric, parameter efficiency becomes an explicit evaluation axis, and ALinear's PNP-MSE and PNP-MAE of 105 and 68.9 rank first among compared models.

Reading between the lines

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

  • Beyond the seven benchmarks, if the result transfers, the useful notion of a scaling law for forecasting would be relative to the amount of predictable signal in the data, not to raw data size; one should then see similar gains for other horizon-adaptive linear models on noisy industrial series.
  • A Fourier-domain variant that applies horizon-dependent spectral masks rather than exponential time-domain decay should reproduce ALinear's gains, which would identify frequency attenuation, not the linear projection, as the active mechanism.
  • A natural test the paper does not run is multivariate forecasting; applying the same adaptive decomposition channel-by-channel would show whether the dataset-dependent component drift persists when variables interact.
  • The PNP metric's log-parameter denominator encodes a policy choice: a tenfold parameter cut at equal error is always worth a fixed efficiency bonus, so adopting PNP as a reporting standard would change which models are called state of the art.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper challenges the applicability of scaling laws to time series forecasting by presenting ALinear, an ultra-lightweight linear model with three main components: a horizon-aware adaptive moving-average decomposition (Section 3.1), component-specific linear projections (Section 3.2), and a horizon-dependent recombination that exponentially decays the seasonal component with forecast horizon (Section 3.3). The authors report that ALinear uses only k-level parameters, outperforms large-scale baselines on seven univariate benchmarks while using less than 1% of their parameters, and introduce a Parameter-Normalized Performance (PNP) metric to compare accuracy per parameter (Section 4.2). They also present ablations, hyperparameter sensitivity results, and an analysis of component drift across horizons (Sections 4.3 and 4.4).

Significance. If the central efficiency claim survived scrutiny, this would be a practically important result: an ultra-lightweight model matching or beating much larger architectures would strengthen the case against indiscriminate parameter scaling in time series forecasting and offer a deployable baseline. The proposed horizon-dependent recombination idea and the PNP metric are useful contributions, and the paper includes an ablation study and a cross-horizon component analysis. However, the headline quantitative claims rest on baseline parameter counts that appear to be inflated by roughly three orders of magnitude for one central baseline, and on accuracy comparisons that are partly contradicted by entries in the paper's own Table 1. The manuscript also reports no error bars despite stating that experiments were repeated with five seeds. The contribution is potentially valuable but is not yet established as presented.

major comments (5)
  1. [§1, Figure 1(b), §4.2] The reported TimesNet parameter count of 1199.72M appears to be inflated by roughly 1000x. The official TimesNet implementation for univariate long-term forecasting typically uses about 1M parameters, not 1.2B. This number is load-bearing: the abstract and Section 4.1 claim 'less than 1% of their parameters', and Figure 4's PNP scores use log(#parameters) in Eq. (10). With ALinear reported at 186K-297K parameters, the ratio to a ~1.2M-parameter TimesNet would be roughly 16-25%, not below 1%. Please recompute all baseline parameter counts using a consistent counting methodology (including whether per-variable univariate models, embeddings, and projection layers are included) and update the quantitative efficiency claims accordingly.
  2. [§4.1, Table 1] The claim that ALinear 'consistently outperforms existing methods' is contradicted by Table 1 itself. For example, at ETTm1 with pred_len=48, TimesNet (MSE 0.019), iTransformer (0.020), PatchTST (0.020), and FEDformer (0.021) all outperform ALinear (0.036); similar counterexamples appear at ETTm2 pred_len=48 and weather pred_len=48. The statement in Section 4.1 that ALinear achieves optimal results in 71.4% of scenarios needs a precise definition of 'scenario' and must be re-audited against the table. The headline conclusion and the abstract's 'consistently outperforms' formulation should be qualified to reflect the actual comparison outcomes.
  3. [§3.3, Eq. (5)] The text describes the exponential decay as a 'learnable spectral filter' and a 'progressive frequency attenuation strategy', but the decay rate λ=δ/H is fixed by hyperparameter δ and the horizon H; no parameter in Eq. (5) is learned. This mischaracterizes the mechanism. In addition, the ablation 'w/o adaptive' in Table 2 removes the adaptive recombination of Eq. (6), not the decay itself, so the attribution of gains to 'progressive frequency decay' is not directly established by the reported ablations.
  4. [§3.5, Eq. (8), Figure 1(b)] The parameter counts are internally inconsistent. Equation (8) gives Θ_ALinear = 2HT + 2H + 4, which for T=96 and H=960 equals 186,244 parameters, while Figure 1(b) reports ALinear as 297.02K parameters. This discrepancy must be resolved before the parameter-efficiency comparison can be assessed, and it also affects the claim that ALinear uses 'k-level' parameters.
  5. [§4.1, Experimental setup] The paper states that all experiments were repeated five times with different random seeds, but Table 1 reports only mean MSE and MAE values with no standard deviations, confidence intervals, or significance tests. Several comparisons are very close (e.g., ETTh1 pred_len=48 ALinear MSE 0.042 vs PatchTST 0.042; ETTh2 pred_len=336 ALinear 0.223 vs Transformer 0.223). Without error bars, the claim of consistent superiority over baselines is not statistically supported.
minor comments (6)
  1. [Abstract, throughout] The model name is written inconsistently as 'Alinear' in the abstract and 'ALinear' elsewhere; please standardize the spelling.
  2. [Figure 2] Figure 2 contains literal '???' placeholders in the kernel size boxes, which suggests the figure is unfinished and should be completed before publication.
  3. [Figure 1(b)] The parameter labels contain typographical errors (e.g., 'FEDformer-16,83M' and '16,95M') and the horizontal axis is partially garbled; please clean up the figure and its captions.
  4. [§3.5, Eq. (8)] The parameter count in Eq. (8) counts projection matrices as 2HT + 2H but does not explicitly list k1, k2, v1, v2 in the 'decomp.' and 'recomb.' terms; please clarify exactly which parameters are included in each term.
  5. [§4.3, Figure 5] The y-axis of Figure 5 is labeled only as 'Value' without units or a definition; please indicate what quantity is plotted (e.g., mean seasonal component magnitude).
  6. [Reproducibility] The manuscript does not state whether code or trained models will be released; for a paper whose central claim is an empirical efficiency comparison, providing the implementation and random seeds is important for verification.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity: learned adaptive weights are presented as validating the adaptive design; the central benchmark claim is otherwise independent.

  1. fitted input called prediction [Abstract and Section 4.3 (RQ3), with Eq. (6) in Section 3.3]
    "This dataset-dependent behavior validates our hypothesis that the balance between trend and seasonality requires dynamic adaptation rather than a fixed decomposition strategy."

    The 'dataset-dependent behavior' is the learned value of βT(H)=σ(v1+v2·H) and βS(H)=1−βT(H) from Eq. (6), where v1 and v2 are free parameters trained to minimize MSE on the same datasets. Because the model explicitly parameterizes component weights as a learnable function of horizon H, observing that these fitted weights drift with horizon and dataset is a property of the parameterization, not an independent confirmation of the necessity of adaptivity. The non-circular support is the 'w/o adaptive' ablation in Table 2; the RQ3 narrative itself treats fitted parameters as validation of the design.

full rationale

ALinear's prediction equations (Eqs. 2-6) define a concrete linear/decomposition model with learnable parameters, and the reported MSE/MAE results on ETT, Exchange, Traffic, and Weather are external benchmark comparisons. Thus the central claim—competitive accuracy with very few parameters—does not reduce to its inputs by construction. The only genuinely circular element is Section 4.3's use of the learned adaptive weights to validate the adaptive mechanism: since βT(H) and α(H) are explicitly fit as functions of H, their observed drift is largely a consequence of the chosen parameterization and cannot by itself prove that adaptivity is necessary; the ablation study is the proper test of that claim. Reference [14] is a prior paper by the first author, but it is used only as supporting context for linear decomposition models and is not load-bearing for the main result. The disputed TimesNet parameter count (reported as ~1.2B versus typical ~1M) is a factual/correctness concern about the 'less than 1%' claim, not a circularity in the derivation chain. Overall, the circularity is minor and confined to the internal-validation narrative.

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

The model relies on standard decomposition assumptions and a hand-specified exponential decay form. The learnable parameters are the kernel scaling, recombination weights, and projection matrices; the decay rate and kernel bounds are hyperparameters. No new physical or mathematical entities are introduced.

free parameters (5)
  • k1, k2 (kernel size scaling)
    Learned via gradient descent to make the moving-average window size depend on forecast horizon (Eq. 3). No fitted values are reported in the paper.
  • v1, v2 (recombination weights)
    Learned to balance trend and seasonal components via sigmoid of v1 + v2*H (Eq. 6).
  • Projection weights W_T, W_S and biases b_T, b_S
    Linear regression weights learned on each benchmark; they constitute the bulk of the parameter count (2HT + 2H) in Eq. 8.
  • delta (decay hyperparameter)
    Hyperparameter chosen by hand for short, long, and ultra-long horizons; controls the exponential seasonal decay rate in Eq. 5.
  • w_min, w_max (kernel bounds)
    Hyperparameters bounding the adaptive kernel size in Eq. 3; no values are reported.
assumptions (4)
  • domain assumption Time series can be separated into trend and seasonal components whose predictability differs across horizons.
    Section 1 and Section 3.1 assume the decomposition is meaningful and that trend is more predictable at long horizons, citing Hyndman and Athanasopoulos.
  • domain assumption A moving average with a horizon-dependent window extracts the trend component.
    Eq. 2 defines the trend as the moving average output; no formal justification is given beyond standard practice.
  • ad hoc to paper Exponential decay of seasonal components as the horizon increases is a valid model of predictability loss.
    Eq. 5 introduces exp(-lambda*t) without derivation; it is the paper's chosen functional form, and lambda = delta/H is itself a hyperparameter.
  • domain assumption The chosen benchmark datasets and evaluation protocol are representative of time series forecasting.
    Section 4 uses seven datasets and a fixed input length of 96; claims of general superiority rest on this selection.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Does Scaling Law Apply in Time Series Forecasting?." pith.science (2026). https://pith.science/paper/VE732KPM

@misc{pith2026250510172,
  author       = {Pith},
  title        = {Pith review of: Does Scaling Law Apply in Time Series Forecasting?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VE732KPM}},
  note         = {Machine review of arXiv:2505.10172}
}
read the original abstract

Rapid expansion of model size has emerged as a key challenge in time series forecasting. From early Transformer with tens of megabytes to recent architectures like TimesNet with thousands of megabytes, performance gains have often come at the cost of exponentially increasing parameter counts. But is this scaling truly necessary? To question the applicability of the scaling law in time series forecasting, we propose Alinear, an ultra-lightweight forecasting model that achieves competitive performance using only k-level parameters. We introduce a horizon-aware adaptive decomposition mechanism that dynamically rebalances component emphasis across different forecast lengths, alongside a progressive frequency attenuation strategy that achieves stable prediction in various forecasting horizons without incurring the computational overhead of attention mechanisms. Extensive experiments on seven benchmark datasets demonstrate that Alinear consistently outperforms large-scale models while using less than 1% of their parameters, maintaining strong accuracy across both short and ultra-long forecasting horizons. Moreover, to more fairly evaluate model efficiency, we propose a new parameter-aware evaluation metric that highlights the superiority of ALinear under constrained model budgets. Our analysis reveals that the relative importance of trend and seasonal components varies depending on data characteristics rather than following a fixed pattern, validating the necessity of our adaptive design. This work challenges the prevailing belief that larger models are inherently better and suggests a paradigm shift toward more efficient time series modeling.

Figures

Figures reproduced from arXiv: 2505.10172 by the authors.

Figure 1
Figure 1. Comparison of model’s parameter efficiency on ETTm1 dataset when prediction length [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. ALinear model. The model separates the input time series into trend and seasonal [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Efficiency analysis and forward pass algorithm of ALinear [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Parameter-Normalized Performance (PNP) comparison across prediction lengths. The bar [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Dataset-dependent evolution of trend (dark blue) and seasonal (light blue) components across prediction horizons. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Hyper-parameter sensitivity analysis. The experiments exhibit the model’s stability under [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 18 canonical work pages

  1. [1]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017

    2017

  2. [2]

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

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

    work page 2021

  3. [3]

    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. 22 419– 22 430, 2021

    work page 2021

  4. [4]

    Finding structure in time,

    J. L. Elman, “Finding structure in time,” Cognitive science, vol. 14, no. 2, pp. 179–211, 1990

    1990

  5. [5]

    Long short-term memory,

    A. Graves and A. Graves, “Long short-term memory,”Supervised sequence labelling with recurrent neural networks, pp. 37–45, 2012

    2012

  6. [6]

    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. PMLR, 2022, pp. 27 268–27 286

    work page 2022

  7. [7]

    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

    arXiv 2022

  8. [8]

    Are transformers effective for time series forecasting?

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

    work page 2023

Show all 40 references
  1. [9]

    Scaling laws for neural language models,

    J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei, “Scaling laws for neural language models,” arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  2. [10]

    Scaling vision transformers,

    X. Zhai, A. Kolesnikov, N. Houlsby, and L. Beyer, “Scaling vision transformers,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 12 104–12 113

    2022

  3. [11]

    R. J. Hyndman and G. Athanasopoulos, Forecasting: Principles and Practice. OTexts, 2018

    2018

  4. [12]

    Time series analysis,

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

    2020

  5. [13]

    D. B. Percival and A. T. Walden,Wavelet methods for time series analysis. Cambridge university press, 2000, vol. 4

    2000

  6. [14]

    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

  7. [15]

    Time-moe: Billion-scale time series foundation models with mixture of experts,

    X. Shi, S. Wang, Y . Nie, D. Li, Z. Ye, Q. Wen, and M. Jin, “Time-moe: Billion-scale time series foundation models with mixture of experts,” arXiv preprint arXiv:2409.16040, 2024

    2024 arXiv

  8. [16]

    Deep learning for time series forecasting: Tutorial and literature survey,

    K. Benidis, S. S. Rangapuram, V . Flunkert, B. Wang, D. Maddix, C. Turkmen, J. Gasthaus, M. Bohlke- Schneider, D. Salinas, L. Stella et al., “Deep learning for time series forecasting: Tutorial and literature survey,”Proceedings of the IEEE, vol. 110, no. 7, pp. 1108–1135, 2022

    2022

  9. [17]

    Energy and policy considerations for modern deep learning research,

    E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for modern deep learning research,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, 2020, pp. 13 693–13 696

    2020

  10. [18]

    Green ai,

    R. Schwartz, J. Dodge, N. A. Smith, and O. Etzioni, “Green ai,” Communications of the ACM, vol. 63, no. 12, pp. 54–63, 2020

    2020

  11. [19]

    Faith: Frequency-domain attention in two horizons for time series forecasting,

    R. Li, M. Jiang, Q. Liu, K. Wang, K. Feng, Y . Sun, and X. Zhou, “Faith: Frequency-domain attention in two horizons for time series forecasting,” Knowledge-Based Systems, vol. 309, p. 112790, 2025

    2025

  12. [20]

    Stl: A seasonal-trend decomposition procedure based on loess,

    C. RB, “Stl: A seasonal-trend decomposition procedure based on loess,” J Off Stat, vol. 6, pp. 3–73, 1990

    1990

  13. [21]

    Forecasting at scale,

    S. J. Taylor and B. Letham, “Forecasting at scale,”The American Statistician, vol. 72, no. 1, pp. 37–45, 2018

    2018

  14. [22]

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

    B. N. Oreshkin, D. Carpov, N. Chapados, and Y . Bengio, “N-beats: Neural basis expansion analysis for interpretable time series forecasting,” arXiv preprint arXiv:1905.10437, 2019

    1905 arXiv

  15. [23]

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

    T. Zhou, Z. Ma, Q. Wen, L. Sun, T. Yao, W. Yin, R. Jinet al., “Film: Frequency improved legendre memory model for long-term time series forecasting,” Advances in neural information processing systems, vol. 35, pp. 12 677–12 690, 2022. 10

    2022

  16. [24]

    Statistical and machine learning forecasting methods: Concerns and ways forward,

    S. Makridakis, E. Spiliotis, and V . Assimakopoulos, “Statistical and machine learning forecasting methods: Concerns and ways forward,” PloS one, vol. 13, no. 3, p. e0194889, 2018

    2018

  17. [25]

    A comparison of direct and iterated multistep ar methods for forecasting macroeconomic time series,

    M. Marcellino, J. H. Stock, and M. W. Watson, “A comparison of direct and iterated multistep ar methods for forecasting macroeconomic time series,” Journal of econometrics, vol. 135, no. 1-2, pp. 499–526, 2006

    2006

  18. [26]

    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

  19. [27]

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

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

    2018 arXiv

  20. [28]

    Reformer: The efficient transformer,

    N. Kitaev, Ł. Kaiser, and A. Levskaya, “Reformer: The efficient transformer,” arXiv preprint arXiv:2001.04451, 2020

    2001 arXiv

  21. [29]

    Longformer: The long-document transformer,

    I. Beltagy, M. E. Peters, and A. Cohan, “Longformer: The long-document transformer,” arXiv preprint arXiv:2004.05150, 2020

    2004 arXiv

  22. [30]

    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

  23. [31]

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

    K. Yi, Q. Zhang, W. Fan, S. Wang, P. Wang, H. He, N. An, D. Lian, L. Cao, and Z. Niu, “Frequency-domain mlps are more effective learners in time series forecasting,”Advances in Neural Information Processing Systems, vol. 36, pp. 76 656–76 679, 2023

    2023

  24. [32]

    Micn: Multi-scale local and global con- text modeling for long-term series forecasting,

    H. Wang, J. Peng, F. Huang, J. Wang, J. Chen, and Y . Xiao, “Micn: Multi-scale local and global con- text modeling for long-term series forecasting,” in The eleventh international conference on learning representations, 2023

    2023

  25. [33]

    Time-llm: Time series forecasting by reprogramming large language models,

    M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Panet al., “Time-llm: Time series forecasting by reprogramming large language models,” arXiv preprint arXiv:2310.01728, 2023

    2023 arXiv

  26. [34]

    Timegpt-1,

    A. Garza, C. Challu, and M. Mergenthaler-Canseco, “Timegpt-1,”arXiv preprint arXiv:2310.03589, 2023

    2023 arXiv

  27. [35]

    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,” in Forty-first International Conference on Machine Learning, 2024

    2024

  28. [36]

    Chronos: Learning the language of time series,

    A. F. Ansari, L. Stella, C. Turkmen, X. Zhang, P. Mercado, H. Shen, O. Shchur, S. S. Rangapu- ram, S. P. Arango, S. Kapoor et al., “Chronos: Learning the language of time series,” arXiv preprint arXiv:2403.07815, 2024

    2024 arXiv

  29. [37]

    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

  30. [38]

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

    K. Rasul, A. Ashok, A. R. Williams, A. Khorasani, G. Adamopoulos, R. Bhagwatkar, M. Biloš, H. Ghonia, N. Hassen, A. Schneider et al., “Lag-llama: Towards foundation models for time series forecasting,” in R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundati...

    2023

  31. [39]

    itransformer: Inverted transformers are effective for time series forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “itransformer: Inverted transformers are effective for time series forecasting,”arXiv preprint arXiv:2310.06625, 2023

    2023 arXiv

  32. [40]

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

    S. Liu, H. Yu, C. Liao, J. Luo, W. Wang, D. Chen, and R. Xie, “Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting,” in International Conference on Learning Representations, 2022. 11

    2022

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.