LightGTS: A Lightweight General Time Series Forecasting Model

Bin Yang; Chenjuan Guo; Lujia Pan; Peng Chen; Yang Shu; Yihang Wang; Yuying Qiu; Zhongwen Rao

REVIEW 4 major objections 3 minor 1 cited by

LightGTS: A Lightweight General Time Series Forecasting Model

T0 review · 4 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read LightGTS claims a 4-million-parameter model beats foundation models ten to one hundred times larger, in zero-shot and fine-tuned settings alike, by cutting each series into patches of exactly one intrinsic period.

desk verdict The period-aligned tokenization and plug-in weight resizing are genuinely new and the ablations back them, but the abstract's 'SOTA on all 9 benchmarks' is contradicted by the paper's own tables, and the flex-resize theory has unproven steps; with corrected claims and a stronger baseline set it deserves a serious referee. read the letter →

arxiv 2506.06005 v1 pith:HLA474PH submitted 2025-06-06 cs.LG

Yihang Wang , Yuying Qiu , Peng Chen , Yang Shu , Zhongwen Rao , Lujia Pan , Bin Yang , Chenjuan Guo This is my paper

classification cs.LG

keywords timeseriesforecastingperiodicaltokenizationparalleldecodingzero-shotfoundationmodelsflexprojectionlayerlightweightmodelcycle-lengthpatching

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 claims that the main reason time series foundation models are so large is that they tokenize with fixed-length patches, which cannot represent the same real-world period consistently across datasets sampled at different rates. LightGTS instead cuts each input series into patches of exactly one intrinsic period, one full cycle, so a "daily" token means the same thing whether the data is sampled every 15 minutes or every hour. A flex projection layer resizes the embedding weights to keep these variable-length patches comparable, and a period-aligned parallel decoder seeds predictions from the last encoded period token. With 1.3 to 4 million parameters, the paper reports state-of-the-art zero-shot and fully fine-tuned accuracy on nine benchmarks, claiming a 10 to 100 times size reduction over existing foundation models while matching or exceeding their error. If right, this would mean the inductive bias of periodicity substitutes for most of the parameter mass in general time series forecasting.

What carries the argument

The argument is carried by Periodical Tokenization, whose PeriodsFinding step computes a cycle length $P$ per series, from known sampling information or from a Fast Fourier Transform, and whose periodical patching segments the series into non-overlapping patches of $P$ points, so each token spans exactly one intrinsic period and different sampling rates map to the same real-world interval. Because $P$ varies across datasets, the companion Flex Projection Layer resizes a reference weight matrix $\theta_e \in \mathbb{R}^{P^* \times D}$ on the fly using flex-resize, $\theta' = \delta^{-1}(A)^{+}\theta$, with $A$ the linear-interpolation matrix from $P^*$ to $P$ points and $\delta = \sqrt{P/P'}$ the claimed variance-compensation constant; this keeps one embedding space for all patch sizes. The second mechanism is Periodical Parallel Decoding, a non-autoregressive decoder that replicates the encoder's last token $K = \lceil F/P \rceil$ times with decaying weight $\omega(\tau) = 1/e^{\tau}$, aligning the phase of the forecast window with the history while avoiding autoregressive error accumulation. Around these two mechanisms sits a compact Transformer encoder-decoder using rotary position embeddings, an MSE loss, and channel-independent training on a multi-source corpus spanning energy, weather, health, transport, web, and economic series.

What would settle it

On a strongly periodic dataset such as Solar, build period patches for the same daily cycle at two sampling resolutions, apply the paper's normalization, and compute the empirical variance ratio $\mathrm{Var}(xA)/\mathrm{Var}(x)$ for the interpolation matrix $A$; if the ratio departs from $P/P'$ to the point that $\delta = \sqrt{P/P'}$ fails to restore unit variance, or if the projected tokens from the two resolutions differ beyond residual noise, the flex-resize consistency claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a scale-invariant inductive bias, the intrinsic period, is what lets a time series model generalize across datasets, and that exploiting it makes most of the parameter mass of current foundation models unnecessary. Because an intrinsic period such as a day is the same real-world interval regardless of sampling rate, it contains 24 hourly points but 96 fifteen-minute points; the proposed Periodical Tokenization cuts each input series into non-overlapping patches of exactly one cycle length $P$, so a token always spans one full period and carries the same semantics across scales, whereas fixed patching packs different amounts of information per token and splits periods unevenly. To embed patches of varying length into a shared space, a Flex Projection Layer resizes a reference embedding weight by the flex-resize formula $\theta' = \delta^{-1}(A)^{+}\theta$, where $A$ is the linear-interpolation matrix and $\delta = \sqrt{P/P'}$ is claimed to compensate the variance change. Forecasting then uses Periodical Parallel Decoding: the last encoder token is replicated $K = \lceil F/P \rceil$ times, weighted by $1/e^{\tau}$, and all output tokens are decoded in parallel. The paper reports that the 4M-parameter LightGTS-mini reduces average MSE by roughly 30% against the strongest zero-shot baselines on nine held-out benchmarks and about 7% against six fully trained deep forecasters, with the 1.3M-parameter variant still beating most baselines; the appendix also concedes that on weakly periodic series an FFT-guessed cycle length can misalign and modestly degrade performance, though it remains competitive.

Load-bearing premise

The load-bearing premise is that the flex-resize formula of Section 3.2 keeps tokens comparable across scales: it assumes normalized input patches behave like a standard normal distribution, so a single constant $\delta = \sqrt{P/P'}$, stated as an upper bound without proof, fully compensates the variance change that linear interpolation introduces; if real patches are correlated or non-Gaussian, the token-consistency guarantee is not established.

Editorial extensions

If this is right

If the zero-shot results hold, general time series forecasting does not need billion-parameter pre-training: the 4M-parameter LightGTS-mini reports an average MSE reduction of roughly 30% over the strongest foundation-model baselines across nine held-out benchmarks, and the 1.3M-parameter tiny variant reports a 27% reduction over the same set.
In the full-shot setting the same 4M model reports about 7% lower average MSE than six state-of-the-art deep forecasting models trained on the target data, and its zero-shot predictions beat those fully trained baselines on five of the nine datasets.
The efficiency figures would put the method in a different deployment class: 4M parameters versus 67M-700M for the baselines, 0.213 GMACs versus 52.6 G for the smallest alternative, and 0.01 s per-series inference time at a 720-step horizon.
Periodical tokenization is reported to be a plug-in: applied to Timer's fixed-patch pipeline with flex-resize, it cuts zero-shot MSE on the Solar dataset by up to roughly 31% relative to Timer's own patching, suggesting the inductive bias transfers to other architectures without retraining.
Across sampling granularities from 0.25 to 4 hours on ETT data, LightGTS's zero-shot MSE stays nearly flat while Timer and Time-MoE vary substantially, which would confirm that period-aligned tokens, not patch size, supply the scale invariance.

Reading between the lines

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

A direct test I would run: pre-train LightGTS on fine-resolution series only (for example 10-minute data) and evaluate on hourly series of the same real-world processes; if the scale-invariance claim is right, the zero-shot gap should be small and should shrink further as the pre-training corpus adds more sampling rates, an experiment the paper does not perform.
Several benchmarks carry more than one intrinsic period (Electricity and Traffic are listed with both daily and weekly cycles), yet the patching selects a single cycle length; a natural extension is multi-resolution periodical patching that emits one token per detected period, which the current architecture cannot represent.
The flex-resize guarantee is derived under a Gaussian, uncorrelated patch model; on real RevIN-normalized patches, which are typically correlated and heavy-tailed, the optimal compensation constant may differ from $\sqrt{P/P'}$, and a direct variance-ratio measurement on held-out patches would show whether the formula is exact or merely close.
If the period-aligned encoder representations are as informative as the results suggest, the Transformer decoder could be replaced by a cheap linear head, pushing the parameter count well below 1M; the practical bottleneck would then be period detection, not model capacity.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: the core idea is real. Period-aligned tokenization is a clean answer to a known problem with fixed-patch foundation models, and the flex-resize layer (weights resized by delta^{-1}(A)^+) is an unusual no-learning adaptation that plausibly explains the gains. The plug-in experiments in Table 11, where flex-resize beats linear and area resize when bolted onto Timer, give the method independent support. Periodical parallel decoding is simple and the ablations in Table 4 consistently show it helping over AR and MAE decoding under periodic patching. The efficiency story is also solid: 4M parameters really is one to two orders of magnitude below MOIRAI or Chronos.

Now the soft spots, in proportion. The abstract says "state-of-the-art forecasting performance on 9 real-world benchmarks in both zero-shot and full-shot settings." The paper's own tables contradict that twice. In zero-shot Electricity, MOIRAI gets 0.188 MSE against LightGTS-mini's 0.213. In full-shot Exchange, iTransformer gets 0.321 against 0.322. So the headline is factually wrong as written. That is fixable by qualifying the claim, but it needs fixing. Second, the zero-shot comparison is incomplete exactly where it matters: Electricity lacks Chronos, TimesFM, and Time-MoE; Traffic lacks MOIRAI, TimesFM, and Time-MoE. You cannot claim SOTA on a dataset when the strongest published models were not run. Third, there are no error bars or seeds anywhere, and several wins are within 1–2%, which is noise territory. Fourth, the theory in Section 3.2: the derivation assumes RevIN outputs are N(0, I), and delta = sqrt(P/P') is introduced as an "upper bound" without proof. If the actual normalization is not Gaussian, the algebra in Proposition 3.2 does not go through. The empirical evidence for flex-resize is decent, so this is a rigor gap, not a fatal flaw.

Who this is for: people working on lightweight time-series foundation models, patch-based tokenization, and cross-frequency transfer. The idea deserves serious attention; the evaluation as written does not support the headline.

Recommendation: do not desk reject. Send it to peer review, but expect a major revision — correct the claims, fill in the missing baselines, add error bars, and either prove or soften the delta claim. The core contribution is worth the referees' time.

Referee Report

4 major / 3 minor

Summary. The paper proposes LightGTS, a lightweight time series foundation model with 1.3M (tiny) and 4M (mini) parameters, built on two main ideas: periodical tokenization, which patches the input by its intrinsic cycle length and embeds patches of varying size with a 'flex-resize' projection, and periodical parallel decoding, which initializes decoder inputs by replicating the last encoder token with an exponential reweighting. The authors claim state-of-the-art zero-shot and full-shot forecasting performance on nine real-world benchmarks, with large efficiency gains over existing TSFMs. They also provide a theoretical analysis of the flex-resize operation and support the method with ablations and case studies.

Significance. If the empirical claims hold, this is a notable efficiency-accuracy contribution: a 4M-parameter model outperforming 67M-700M parameter TSFMs by large margins would be practically valuable, and the periodical tokenization concept is a sensible inductive bias for multi-source pre-training. The paper ships code and pretrained checkpoints, and the ablations in Table 4 and Table 11 do show consistent gains from periodical patching and flex-resize. However, the headline claim of state-of-the-art performance on all nine benchmarks is not supported by the paper's own tables, and the theoretical derivation of flex-resize contains a load-bearing gap. The strengths are the clear empirical demonstrations of the two proposed components and the unusually small parameter count.

major comments (4)

[Abstract; Table 1; Table 2/Table 13] The claim that LightGTS 'achieves state-of-the-art forecasting performance on 9 real-world benchmarks in both zero-shot and full-shot settings' is contradicted by the paper's own results. In Table 1 (Zero-shot), Electricity reports LightGTS-mini MSE 0.213 while MOIRAI reports 0.188, so LightGTS is not best on that benchmark. In Table 2/Table 13 (Full-shot), Exchange reports LightGTS-mini MSE 0.322 while iTransformer reports 0.321, again not best. These are not negligible differences, and they directly undermine the unqualified SOTA claim. The abstract, Section 4.2, and Section 4.3 should be revised to state the actual ranking, e.g., 'state-of-the-art on seven of nine benchmarks' or 'competitive with state-of-the-art on all nine'.
[Section 3.2, Eq. (13)-(20)] The derivation of the flex-resize formula is not sound as presented. The paper asserts that after RevIN normalization X ~ N(0,I) and that the normalization after linear interpolation is equivalent to multiplying by a single constant delta = sqrt(P/P'), but no proof is given for the claim that a single scalar can compensate the variance change under interpolation. For linear interpolation, the covariance of xA is A^T A, which is not a scalar multiple of the identity for a general interpolation matrix A; different components have different variances, so a single scalar correction does not align the distributions. Moreover, the direction of the scaling appears inverted for upsampling: when P' > P, sqrt(P/P') < 1, which would shrink the variance further, whereas interpolation typically reduces variance relative to the original. The authors should either provide a precise derivation with the conditions under which delta is valid, or present flex-resize as a heuristic supported only by the empirical ablations in Table 11.
[Section 4.2, Table 1; Section 4.3, Table 2] No error bars, standard deviations, or multiple-seed results are reported, and several claimed wins are within 1-2% (e.g., ETTh2 zero-shot 0.348 vs 0.354, Exchange full-shot 0.322 vs 0.321). Given the central claim of state-of-the-art performance, the robustness of these margins is unclear. The authors should report at least three independent runs with standard deviations, or perform significance testing, for the main comparisons.
[Section 4.2, Table 1] The zero-shot comparison on Electricity and Traffic omits several strong TSFMs, as indicated by the dashes in Table 1: Electricity has dashes for Chronos, TimesFM, and Time-MoE, and Traffic has dashes for MOIRAI, TimesFM, and Time-MoE. While the dashes are explained by pretraining overlap, this means the 'state-of-the-art' claim on those datasets is only relative to the remaining baselines. The authors should either compare with models that are evaluated on these datasets despite the overlap (e.g., by excluding those datasets from pretraining in a controlled variant), or explicitly qualify the claim as best among models not pretrained on the target datasets.

minor comments (3)

[Throughout] There are several typos and formatting issues: 'Model Analasis' in Section 4.6, 'LightGTS-miny' in the Appendix C.2 header, 'decoing' in Table 14 caption, 'TimeMxier' in Table 2 header, and 'Resolusions' in Section 4.6. These should be corrected.
[Section 3.1.2, Eq. (8)] The reweighting function omega(tau) = 1/e^tau is introduced without explanation or ablation; the choice of exponential decay is not justified. A short justification or an ablation over alternative reweighting functions would strengthen the presentation.
[Section 4.1 / Appendix A.4] The statement 'it achieves superior performance compared to the six state-of-the-art baselines with full-data training, achieving an average MSE reduction of 7%' in Section 4.3 should specify over which datasets and how the average is computed, since the average over all datasets is not shown in Table 2 and the Exchange row contradicts the claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the flex-resize formula follows from a stated least-squares objective and the central SOTA claims are empirical, evaluated on test sets disjoint from pre-training.

full rationale

The paper's only theoretical derivation, flex-resize (Section 3.2), starts from an explicit optimization objective (Eq. 12 / Theorem 3.1) and solves it as a least-squares problem (Eq. 19) to obtain θ' = δ^{-1}(A)^+θ (Eq. 20). This is a self-contained mathematical reduction, not a restatement of the empirical claims. The δ factor is asserted as an upper bound on variance change under interpolation; this is an unproven assumption and a validity risk, but it is not circular because the result does not presuppose the performance it is used to explain. Periodical tokenization and periodical parallel decoding are justified by ablations (Tables 4-6) and evaluated on target datasets explicitly excluded from pre-training (Appendix A.1, A.2), so there is no fitted-input-called-prediction pattern. Self-citations (e.g., Qiu et al. 2024/2025a, Li et al. 2025, Pathformer) are used for evaluation conventions, benchmarks, or baselines and are not load-bearing for the central architecture or the flex-resize derivation. The abstract's 'state-of-the-art on 9 benchmarks' is contradicted by the paper's own full-shot Exchange row (0.322 vs iTransformer 0.321) and the zero-shot Electricity/Traffic comparisons have missing baselines, but these are empirical/correctness concerns, not circular reasoning. Overall, no step in the claimed derivation chain reduces by definition to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or data entities; its contributions are architectural methods (periodical tokenization, flex-resize, periodical parallel decoding). The load-bearing assumptions are the Gaussian/RevIN normalization in the flex-resize derivation and the single-period assumption for each series. Free parameters are explicit hyperparameters such as P* = 48, N = 10, K = 4, and the exponential reweighting function.

free parameters (3)

Reference patch size P* = 48
Hyperparameter in the Flex Projection Layer (Section 4.6); the ablation shows insensitivity across 24 to 192, but 48 is chosen by hand for all experiments.
Decoder reweighting function omega(tau) = 1/e^tau
Hand-chosen exponential decay in periodical parallel decoding (Eq. 8); no derivation or external justification is provided.
Pretraining token counts N and K = N=10, K=4
Number of historical tokens and prediction tokens during pretraining (Appendix A.4), set empirically.

assumptions (4)

ad hoc to paper After RevIN normalization, input patches follow X ~ N(0,I), and variance change under interpolation is uniformly compensated by delta = sqrt(P/P').
Invoked in Proposition 3.2 and Eq. (14)-(20); the delta formula is asserted as an upper bound without proof, and the model's actual normalization is not verified.
domain assumption Each time series has one scalar intrinsic period P that defines the patch size.
Periodical patching assumes a single cycle length per series, but datasets such as Traffic and Electricity have both daily and weekly periods, so an arbitrary choice is required (Table 8).
standard math Linear interpolation of a patch is exactly x' = x A for a fixed matrix A.
Definitional statement in Section 3.1.1 used to set up the optimization problem.
ad hoc to paper The last encoder token consolidates historical periodicity and is the best decoder initialization.
Asserted in Section 3.1.2 and supported only by the ablation in Table 6; there is no theoretical derivation of why the last token is optimal.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LightGTS: A Lightweight General Time Series Forecasting Model." pith.science (2026). https://pith.science/paper/HLA474PH

@misc{pith2026250606005,
  author       = {Pith},
  title        = {Pith review of: LightGTS: A Lightweight General Time Series Forecasting Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLA474PH}},
  note         = {Machine review of arXiv:2506.06005}
}

read the original abstract

Existing works on general time series forecasting build foundation models with heavy model parameters through large-scale multi-source pre-training. These models achieve superior generalization ability across various datasets at the cost of significant computational burdens and limitations in resource-constrained scenarios. This paper introduces LightGTS, a lightweight general time series forecasting model designed from the perspective of consistent periodical modeling. To handle diverse scales and intrinsic periods in multi-source pre-training, we introduce Periodical Tokenization, which extracts consistent periodic patterns across different datasets with varying scales. To better utilize the periodicity in the decoding process, we further introduce Periodical Parallel Decoding, which leverages historical tokens to improve forecasting. Based on the two techniques above which fully leverage the inductive bias of periods inherent in time series, LightGTS uses a lightweight model to achieve outstanding performance on general time series forecasting. It achieves state-of-the-art forecasting performance on 9 real-world benchmarks in both zero-shot and full-shot settings with much better efficiency compared with existing time series foundation models.

Figures

Figures reproduced from arXiv: 2506.06005 by the authors.

**Figure 1.** Comparison of model sizes and average zero-shot performance across seven benchmark datasets between LightGTS and the state-of-the-art TSFMs. datasets, several Time Series Foundation Models (TSFM) have emerged (Liu et al., 2024; Woo et al., 2024a; Shi et al., 2024), demonstrating promising potential. However, the generalization capability of existing TSFMs largely depends on massive pre-training data and large model… view at source ↗

**Figure 2.** (a) The illustration of three tokenization methods (Point embedding, Fixed patch embedding, Periodical patch embedding). For example, ETTh2 and ETTm2 share the same daily intrinsic period, but their cycle lengths differ due to differences in scale. (b) The case study of training on ETTh1 and testing on datasets with different scales and cycle lengths, all three tokenization methods recognized the intrinsic period on… view at source ↗

**Figure 3.** LightGTS architecture. It is worth noting when prior knowledge of the input series is available, the cycle length can be inferred using information such as the sampling rate. In the absence of prior knowledge, the cycle length can also be deduced using methods such as Fast Fourier transform (FFT) (Wu et al., 2022). After obtaining the cycle length 𝑃, the input series x is segmented into non-overlapping period patch… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Comparisons of robustness across different sampling granularities between LightGTS, Timer, and Time-MoE in the zero-shot setting. Representation Learning of the Periodical Tokenization As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Similarity of token representation across various sampling granularities in Fixed (Timer) versus Periodical (LightGTS) Tokenization. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Comparisons of the model performance between different Periodical Finding methods in zero-shot setting. The SOTA baselines refer to the best-performing baseline results for each dataset. 𝐹𝑆 means the seasonality strength for each dataset. 14 [PITH_FULL_IMAGE:figures/f…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Towards Measuring and Modeling Geometric Structures in Time Series Forecasting via Image Modality
cs.CV 2025-07 conditional novelty 4.0 of 10

A new image-based similarity metric (TGSI) and a three-part training loss (SATL) that together aim to improve the geometric fidelity of time series forecasts.

Reference graph

Works this paper leans on

30 extracted references · 14 canonical work pages · cited by 1 Pith paper

[1]

F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S

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

arXiv
[5]

I., Hyndman, R

Godahewa, R., Bergmeir, C., Webb, G. I., Hyndman, R. J., and Montero-Manso, P. Monash time series forecasting archive. arXiv preprint arXiv:2105.06643, 2021a. Godahewa, R., Bergmeir, C., Webb, G. I., Hyndman, R. J., and Montero-Manso, P. Monash time series forecasting archive. arXiv preprint arXiv:2105.06643, 2021b. Goswami, M., Szafer, K., Choudhry, A., ...

arXiv
[6]

O., and Socher, R

Gu, J., Bradbury, J., Xiong, C., Li, V. O., and Socher, R. Non-autoregressive neural machine translation. arXiv preprint arXiv:1711.02281,

arXiv
[8]

Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

arXiv
[9]

Code repositories for baselines. Model Types Models Code Repositories Foundation model Timer https://github.com/thuml/Large-Time-Series-Model MOIRAI https://github.com/redoules/moirai Chronos https://github.com/amazon-science/chronos-forecasting TimesFM https://github.com/google-research/timesfm Time-MoE https://github.com/Time-MoE/Time-MoE Small Model PD...

arXiv
[10]

Drop Last

Downstream Forecasting In downstream forecasting, we configure the model to perform periodical patching based on the cycle length, tailored to the characteristics of each dataset. The number of historical tokens is set to 𝑁 = 10, and the model is tasked with making predictions for target lengths of 96, 192, 336, and 720, respectively. The “Drop Last” issu...

work page 2024
[12]

As shown in Table 10, we include nine representative real-world datasets, demonstrating that LightGTS achieves state-of-the-art forecasting performance. C.2. Full-shot Forecasting Table 13 provides the comprehensive results for in-distribution forecasting, showcasing the performance of LightGTS-miny in both zero-shot and full-shot settings, as well as oth...

work page 2025
[13]

Lower MSE or MAE values indicate better predictions

The results of LightGTS-miny in zero-shot and full-shot setting and other baselines in full-shot setting. Lower MSE or MAE values indicate better predictions. Red: the best, Blue: the 2nd best. Models LightGTS (zero-shot) LightGTS (full-shot) PDF iTransformer Pathformer FITS TimeMixer PatchTST Metric MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE...

work page 2025

Show all 30 references

[14]

S., and Yang, B

Qiu, X., Li, Z., Qiu, W., Hu, S., Zhou, L., Wu, X., Li, Z., Guo, C., Zhou, A., Sheng, Z., Hu, J., Jensen, C. S., and Yang, B. Tab: Unified benchmarking of time series anomaly detection methods. In Proc. VLDB Endow., 2025a. Qiu, X., Wu, X., Lin, Y., Guo, C., Hu, J., and Yang, B...

arXiv
[17]

K 2vae: A koopman-kalman enhanced variational autoencoder for probabilistic time series forecasting

Wu, X., Qiu, X., Gao, H., Hu, J., Yang, B., and Guo, C. K 2vae: A koopman-kalman enhanced variational autoencoder for probabilistic time series forecasting. In ICML, 2025a. Wu, X., Qiu, X., Li, Z., Wang, Y., Hu, J., Guo, C., Xiong, H., and Yang, B. Catch: Channel-aware multiva...

arXiv
[18]

Implementation Details A.1

11 Submission and Formatting Instructions for ICML 2025 A. Implementation Details A.1. Pre-training Datasets We incorporate a diverse range of multi-source datasets for pre-training, which include portions from the Monash (Godahewa et al., 2021b), UEA (Bagnall et al., 2018), a...

2025
[19]

The complete list of pre-training datasets is shown in Table

time series datasets, as well as additional classic datasets (Zhang et al., 2017; Wang et al., 2024b; Liu et al., 2022; McCracken & Ng, 2016; Taieb et al., 2012). The complete list of pre-training datasets is shown in Table

2017
[20]

Saugeen River Flow Daily 23741 Monash (Godahewa et al., 2021b) Sunspot Daily 73924 Monash (Godahewa et al., 2021b) Weather Daily 43032000 Monash (Godahewa et al., 2021b) KDD Cup 2018 Daily 2942364 Monash(Godahewa et al., 2021b) US Births Daily 7305 Monash (Godahewa et al., 202...

2018
[21]

12 Submission and Formatting Instructions for ICML 2025 A.2. Evaluation Datasets We use the following 9 multivariate time-series datasets for downstream forecasting task: ETT datasets1 contain 7 variates collected from two different electric transformers from July 2016 to July

2025
[22]

Electricity 2 contains the electricity consumption of 321 customers from July 2016 to July 2019, recorded hourly

It consists of four subsets, of which ETTh1/ETTh2 are recorded hourly and ETTm1/ETTm2 are recorded every 15 minutes. Electricity 2 contains the electricity consumption of 321 customers from July 2016 to July 2019, recorded hourly. Solar3 collects production from 137 PV plants ...

2016
[23]

The statistics of evaluation datasets. Dataset Domain # Frequency# Timestamps# Split # Dims# Intrinsic Period# Cycle Length ETTh1 Energy 1 hour 14400 6:2:2 7 Daily 24 ETTh2 Energy 1 hour 14400 6:2:2 7 Daily 24 ETTm1 Energy 15 mins 57600 6:2:2 7 Daily 96 ETTm2 Energy 15 mins 57...

2024
[26]

Models Encoder LayersDecoder LayersModel Dim.FFN Dim.Parameters LightGTS-tiny 1 1 256 512 1.3M LightGTS-mini 3 3 256 512 4M B

Detailed model configurations of LightGTS and corresponding parameter counts. Models Encoder LayersDecoder LayersModel Dim.FFN Dim.Parameters LightGTS-tiny 1 1 256 512 1.3M LightGTS-mini 3 3 256 512 4M B. More Results and Analysis B.1. Analysis of the Periods Finding In this e...

1910
[27]

The SOTA baselines refer to the best-performing baseline results for each dataset.𝐹𝑆 means the seasonality strength for each dataset

Comparisons of the model performance between different Periodical Finding methods in zero-shot setting. The SOTA baselines refer to the best-performing baseline results for each dataset.𝐹𝑆 means the seasonality strength for each dataset. 14 Submission and Formatting Instructio...

2025
[29]

Lower MSE or MAE values indicate better predictions

Full results of zero-shot forecasting experiments. Lower MSE or MAE values indicate better predictions. A dash (’-’) denotes datasets included in the model’s pretraining and therefore excluded from testing. Red: the best, Blue: the 2nd best. Models LightGTS-tiny LightGTS-mini ...

2025
[2012]

Y., and Zhou, J

Wang, S., Wu, H., Shi, X., Hu, T., Luo, H., Ma, L., Zhang, J. Y., and Zhou, J. Timemixer: Decomposable multiscale mixing for time series forecasting. In ICLR, 2024a. Wang, Y., Han, Y., Wang, H., and Zhang, X. Contrast everything: A hierarchical contrastive framework for medica...

arXiv
[2014]

Crafting papers on machine learning

Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Confer- ence on Machine Learning (ICML 2000), pp. 1207–1216, Stanford, CA,

2000
[2016]

H., Sinthong, P., and Kalagnanam, J

Nie, Y., Nguyen, N. H., Sinthong, P., and Kalagnanam, J. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730,

arXiv
[2017]

Y., Shi, X., Chen, P.-Y., Liang, Y., Li, Y.-F., Pan, S., et al

9 Submission and Formatting Instructions for ICML 2025 Jin, M., Wang, S., Ma, L., Chu, Z., Zhang, J. Y., Shi, X., Chen, P.-Y., Liang, Y., Li, Y.-F., Pan, S., et al. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728,

2025 arXiv
[2018]

arXiv preprint arXiv:1811.00075,

arXiv
[2019]

Units: Building a unified time series model

Gao, S., Koker, T., Queen, O., Hartvigsen, T., Tsiligkaridis, T., and Zitnik, M. Units: Building a unified time series model. arXiv preprint arXiv:2403.00131,

arXiv
[2021]

Wu, X., Zhang, D., Zhang, M., Guo, C., Yang, B., and Jensen, C. S. Autocts+: Joint neural architecture and hy- perparameter search for correlated time series forecasting. 10 Submission and Formatting Instructions for ICML 2025 Proceedings of the ACM on Management of Data, 1(1): 1–26,

2025
[2022]

itransformer: Inverted transformers are effective for time series forecasting

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

arXiv
[2023]

Timer: Transformers for time series analysis at scale

Liu, Y., Zhang, H., Li, C., Huang, X., Wang, J., and Long, M. Timer: Transformers for time series analysis at scale. arXiv preprint arXiv:2402.02368,

arXiv
[2024]

A decoder- only foundation model for time-series forecasting

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

arXiv
[2025]

Cyclenet: Enhancing time series forecasting through mod- eling periodic patterns

Lin, S., Lin, W., Hu, X., Wu, W., Mo, R., and Zhong, H. Cyclenet: Enhancing time series forecasting through mod- eling periodic patterns. CoRR, abs/2409.18479, 2024a. Lin, S., Lin, W., Wu, W., Chen, H., and Yang, J. Sparsetsf: Modeling long-term time series forecasting with *1...

arXiv

Pith tools

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

[1] [1]

F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Shchur, O., Rangapuram, S

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

arXiv

[2] [5]

I., Hyndman, R

Godahewa, R., Bergmeir, C., Webb, G. I., Hyndman, R. J., and Montero-Manso, P. Monash time series forecasting archive. arXiv preprint arXiv:2105.06643, 2021a. Godahewa, R., Bergmeir, C., Webb, G. I., Hyndman, R. J., and Montero-Manso, P. Monash time series forecasting archive. arXiv preprint arXiv:2105.06643, 2021b. Goswami, M., Szafer, K., Choudhry, A., ...

arXiv

[3] [6]

O., and Socher, R

Gu, J., Bradbury, J., Xiong, C., Li, V. O., and Socher, R. Non-autoregressive neural machine translation. arXiv preprint arXiv:1711.02281,

arXiv

[4] [8]

Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,

arXiv

[5] [9]

Code repositories for baselines. Model Types Models Code Repositories Foundation model Timer https://github.com/thuml/Large-Time-Series-Model MOIRAI https://github.com/redoules/moirai Chronos https://github.com/amazon-science/chronos-forecasting TimesFM https://github.com/google-research/timesfm Time-MoE https://github.com/Time-MoE/Time-MoE Small Model PD...

arXiv

[6] [10]

Drop Last

Downstream Forecasting In downstream forecasting, we configure the model to perform periodical patching based on the cycle length, tailored to the characteristics of each dataset. The number of historical tokens is set to 𝑁 = 10, and the model is tasked with making predictions for target lengths of 96, 192, 336, and 720, respectively. The “Drop Last” issu...

work page 2024

[7] [12]

As shown in Table 10, we include nine representative real-world datasets, demonstrating that LightGTS achieves state-of-the-art forecasting performance. C.2. Full-shot Forecasting Table 13 provides the comprehensive results for in-distribution forecasting, showcasing the performance of LightGTS-miny in both zero-shot and full-shot settings, as well as oth...

work page 2025

[8] [13]

Lower MSE or MAE values indicate better predictions

The results of LightGTS-miny in zero-shot and full-shot setting and other baselines in full-shot setting. Lower MSE or MAE values indicate better predictions. Red: the best, Blue: the 2nd best. Models LightGTS (zero-shot) LightGTS (full-shot) PDF iTransformer Pathformer FITS TimeMixer PatchTST Metric MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE MSE MAE...

work page 2025