Towards Measuring and Modeling Geometric Structures in Time Series Forecasting via Image Modality

Mingyang Yu; Peng Chen; Xiahui Guo; Yang Shu; Zhenkai Li

REVIEW 4 major objections 6 minor 39 references

Towards Measuring and Modeling Geometric Structures in Time Series Forecasting via Image Modality

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes a geometric-structure metric TGSI and a shape-aware loss SATL that improves forecasting models on both MSE and TGSI without extra inference cost.

desk verdict SATL plausibly lowers MSE across a wide sweep of benchmarks, but the geometric claims rest on a metric validated on deformations designed to match its own construction, and the missing error bars leave the headline gains unproven. read the letter →

arxiv 2507.23253 v1 pith:X5ZLEY5S submitted 2025-07-31 cs.CV

Mingyang Yu , Xiahui Guo , Peng chen , Zhenkai Li , Yang Shu This is my paper · ORCID

classification cs.CV

keywords timeseriesforecastinggeometricstructureevaluationmetriclossfunctionimagemodalityTGSISATLstructuralsimilarity

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

The paper argues that pointwise metrics such as MSE and MAE miss the geometric structure of time series, and that forecasting quality should also be judged by shape. To make this concrete, it introduces TGSI, which turns a time series into a two-dimensional image and scores similarity by luminance and covariance, and SATL, a training loss that approximates shape-awareness directly in the time-series modality. The authors claim that models trained with SATL beat MSE-trained baselines on both MSE and TGSI across eight datasets, at no extra inference cost. If true, the practical value is a cheap, plug-in training objective that preserves temporal shape in long-horizon forecasts.

What carries the argument

The time-to-image transformation first normalizes the series, places one activated pixel per time step, and vertically expands the line by a width d (set to 100 in the experiments) to create a gradient-like probability distribution of values. TGSI then scores structural similarity as the product of luminance l(x,y) and covariance s(x,y), deliberately omitting SSIM's contrast component because the vertical expansion fixes variance along the y-axis. SATL carries the geometric information back into training as L_SATL = alpha*L_diff + beta*L_freq + gamma*L_perceptual, where L_diff compares first-order differences, L_freq matches dominant FFT coefficients and suppresses others, and L_perceptual compares features from a frozen temporal extractor trained to match an image autoencoder's latent codes.

What would settle it

Generate a suite of triples (y, x1, x2) where x1 and x2 have identical MSE to y but human raters consistently judge x1 as closer in shape; compute TGSI(d=100). If TGSI does not rank x1 above x2 for a substantial share of triples, or if changing d from 100 to 50 reverses the ranking for any triple, the claim that TGSI captures geometric structure fails.

Watch

Extended reading notes

Core claim

The central claim is that the geometric structure of a time series can be measured by embedding the series into an image and comparing image luminance and covariance, and that the same structure can be taught to a forecasting model through a loss defined entirely in the time-series modality. TGSI is defined as the product of a luminance term and a covariance term computed on vertically expanded time-series images, giving a similarity score in [-1, 1] that can distinguish cases where MSE values are identical but shapes differ. SATL combines a first-order difference loss, a frequency-domain loss based on the Fast Fourier Transform, and a perceptual feature loss from a temporal feature extractor aligned to a time-series image autoencoder, plus an MSE term for numerical accuracy. The experiments are claimed to show that models trained with SATL consistently outperform their MSE-trained counterparts on all tested datasets and metrics, with larger gains at longer prediction horizons.

Load-bearing premise

The whole method rests on the assumption that drawing a time series as a vertically smeared picture, with a smearing width of 100 pixels, preserves the shape information that matters, so that the metric and the perceptual features built from these pictures actually measure geometric structure rather than artifacts of the drawing.

Editorial extensions

If this is right

Models trained with SATL improve over MSE-trained versions on all eight datasets tested, with average MSE gains per baseline such as 6.42% for Autoformer and 5.20% on ETTh2.
SATL is plug-and-play: it requires no architectural changes, the hyperparameters are fixed for most datasets, and inference cost is unchanged because the extra machinery is used only during training.
TGSI can rank geometric similarity in cases where MSE gives identical scores, as shown by the synthetic example where TGSI(y, x1) = 0.5212 versus TGSI(y, x2) = 0.2080 while both have MSE = 0.79.
In the synthetic validation, vertical expansion with d = 100 gives near-monotonic TGSI versus controlled similarity (Pearson r = 0.95), while d = 0 fails (r = -0.1) and d = 10 loses discriminative power at low similarity.
Ablation shows that removing any SATL component degrades performance, with the MSE term being the most critical, and the advantages over MSE training grow as the forecast horizon extends from 96 to 720 steps.

Reading between the lines

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

Beyond the paper's claims, TGSI's numerical values depend on arbitrary rendering choices such as image height, expansion width, and normalization, so its scores are not directly comparable across different transformation settings; a principled version might define shape similarity in a way that is invariant to these choices.
The perceptual feature loss is trained on images generated by the same transformation used for TGSI, so an open question is whether the shape awareness transfers to other time-to-image encodings such as spectrograms or recurrence plots; testing this would reveal whether the representation or the loss itself carries the geometry.
Because none of SATL's three components is an exact gradient of TGSI, the bridge between the non-differentiable metric and the trainable loss is indirect; a direct test would be to verify that minimizing SATL monotonically raises TGSI on held-out sequences across datasets.
If the improvements persist when the perceptual extractor is frozen and only the difference and frequency terms are used, the practical recipe for structure-aware training may be simpler than the full three-part loss.

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 SATL loss package appears to genuinely lower MSE across a wide set of forecasting benchmarks, and that result is worth taking seriously. But the paper pairs this with a geometric metric (TGSI) whose validation is circular, and the absence of error bars or code makes it impossible to tell whether the gains are real or noise.

What's new: TGSI is essentially SSIM applied to a hand-crafted image representation of a time series (vertical expansion with decaying intensity). SATL is a three-part loss: first-difference MSE, an FFT-based frequency loss with top-k selection, and a perceptual feature loss from an aligned temporal/image encoder. Each component is familiar, but the combination is new, and the paper shows it can be bolted onto four different base models without changing inference cost. The experimental sweep is unusually wide: eight datasets, four models, retrained baselines, and an ablation showing all three components contribute. That is solid workmanship.

Where it gets soft. Section 5.5 tunes the vertical expansion width d to make TGSI monotonic against synthetic deformations, then uses that same d=100 to define the metric and to evaluate every model. The perceptual loss in Section 4.3 trains an image autoencoder on exactly the same transformation and aligns a temporal feature extractor to it, so L_perceptual is a differentiable proxy for TGSI itself. Under those conditions, SATL-trained models improving TGSI is partly by construction. The paper never compares TGSI to DTW, shape-based distances, or human judgments, so there is no independent evidence that TGSI measures anything a practitioner would call shape. The MSE gains are small in several cells (0.23% TGSI improvement on Exchange), and no significance tests or multiple seeds are reported. The loss hyperparameters and the expansion width are free parameters, with no sensitivity analysis except the single d sweep.

Proportionately: the core empirical claim—SATL improves MSE—is plausible and consistently signed, but it is not rigorously established. The geometric-structure claim is weaker because the metric's validation is self-referential. This does not make the loss useless; first-difference and frequency losses are reasonable regularizers. But the paper oversells the geometric interpretation.

Who should read it: forecasting practitioners looking for a plug-and-play training loss with a wide empirical comparison. For a reviewer: yes, this deserves a serious referee, but with major requests: multi-seed variances, code release, a TGSI validation that does not use the same images the loss is trained on, and a comparison with DTW/soft-DTW. I would not cite it in my own work until those are addressed.

Referee Report

4 major / 6 minor

Summary. The paper proposes TGSI, an image-based evaluation metric intended to quantify the geometric structure similarity of time series, and SATL, a multi-component loss (first-order difference, frequency-domain, and perceptual feature) designed to train forecasting models to preserve such structure. The authors evaluate four base models (TimeMixer, PatchTST, TimesNet, Autoformer) on eight benchmark datasets, reporting that SATL training improves MSE, MAE, and TGSI relative to MSE training at no extra inference cost. The paper also includes a synthetic validation of TGSI and an ablation study of SATL components.

Significance. If the central claims are established, the paper would offer a useful plug-and-play training loss and a new perspective on evaluating time series forecasts beyond point-wise accuracy. The experimental coverage is broad (4 architectures, 8 datasets, multiple horizons), and the authors re-train all baselines themselves rather than citing numbers from prior papers. However, the manuscript currently does not convincingly validate TGSI as an independent measure of geometric structure, and the reported improvements lack statistical support. The main strengths—consistent directional improvements across settings and a zero-inference-cost loss—are real, but they are not sufficient to support the stronger claims about geometric fidelity. I regard the issues below as load-bearing and addressable in a major revision.

major comments (4)

[§5.5 and §3.1] The validation of TGSI is circular. In Section 5.5, the vertical expansion width d=100 is selected specifically because it yields near-monotonic TGSI scores on the synthetic deformations T1–T3, and then the same d=100 is used without further discussion in the definition of the metric (Section 3.1) and in every evaluation reported in Table 1. This makes the observed monotonicity a property of the construction, not an independent confirmation that TGSI captures geometric structure. The paper should validate TGSI against external criteria—for example, human perceptual judgments, or established shape-aware time series distances such as DTW or shape-based distance—on held-out deformation types that were not used to tune d, and it should report the sensitivity of the main conclusions to the choice of d.
[§4.3 and Table 1] The perceptual feature loss in Section 4.3 is trained on the exact same time-series-to-image transformation (with the same d=100 and the same image encoder) that defines TGSI. Consequently, SATL is effectively optimizing a differentiable proxy of the TGSI representation itself, so the TGSI gains in Table 1 are partly a consequence of optimizing the same representation rather than independent evidence that SATL preserves geometric structure. To support the claim that SATL improves geometric fidelity, the authors should show that SATL also improves independent structural or shape-aware metrics, or that the perceptual features generalize across different image representations of the same time series.
[§5.2 and Table 1] The comparative results are presented as single runs with no error bars, no multiple seeds, and no significance tests. Several of the reported differences are very small—for example, 0.23% TGSI improvement on Exchange and 1.46% on Weather, and 0.48% on ILI—so the claim that SATL-trained models "consistently achieve superior performance" is not statistically supported. The authors should report means and standard deviations over at least three to five seeds and perform paired significance tests (e.g., Wilcoxon signed-rank or paired t-test) across prediction horizons for both MSE and TGSI.
[§4.2, Eq. (7)] The frequency-domain loss Lnoise penalizes all non-dominant frequency components of the prediction toward zero, on the assumption that the corresponding components in the ground truth are noise. This assumption is not justified for datasets that may contain meaningful high-frequency or irregular structure (e.g., ILI or Weather), and the paper does not analyze the spectral content of the benchmark datasets. The ablation in Table 2 shows only that removing the frequency component changes performance; it does not establish that the specific noise-suppression strategy is valid. The authors should either justify the noise assumption with spectral evidence or compare against a variant that preserves non-dominant frequencies without suppressing them.

minor comments (6)

[Abstract and §3] The metric is called "time series Geometric Structure Index" in the Abstract but "Temporal Geometric Structural Index" in Section 3; please standardize the name and abbreviation throughout.
[§5.2] There is a typo: "MSE-trianed" should be "MSE-trained."
[§5.1] The hyperparameter description says the loss weights are fixed for "the majority of datasets" and selected on the validation set for the rest, but it does not state which datasets fall into which category. Please list the specific datasets and the chosen hyperparameter values.
[§5.3] The comparison against TILDE-Q in Figure 7 does not specify how the TILDE-Q hyperparameters were selected or whether the model was tuned fairly. Please add experimental details for all compared loss functions.
[§5.5 and Figure 6] The Pearson correlation r is reported only for d=0 and d=100; for completeness and comparability, report r for d=10 as well, and clarify the axis labels in the fitted curve figure.
[§5.1] The paper states that all baselines were re-trained by the authors, which is good practice, but the source of the ground-truth TGSI values and the exact normalization used in the image transform are not fully specified. Please provide a precise algorithmic description of the transformation and downscaling steps.

Circularity Check

2 steps flagged · score 4.0 of 10

TGSI is validated and optimized on the same time-series-to-image representation, so the reported TGSI gains are partly built into the method rather than independently demonstrated.

fitted input called prediction [Section 5.5 (Validation of TGSI Metric) and Section 5.1 (Settings)]
"The evaluation tests three vertical expansion configurations (d = 0, d = 10, and d = 100) to analyze the structural encoding capability. ... Most significantly, full expansion (d = 100) achieves near-perfect monotonicity (r = 0.95) with clear differentiation across all similarity levels, confirming that adequate vertical expansion is essential for capturing temporal geometric structures."

The metric's free parameter d is selected in Section 5.5 to make TGSI monotonic on synthetic sequences whose 'similarity level p' is defined by amplitude scaling, constant offset, and noise injection. The same fitted value d=100 is then fixed in Section 5.1 for all TGSI evaluations in Table 1 and Figure 7. The validation therefore does not test TGSI against an external or independent notion of geometric structure; it tunes the image expansion until the metric ranks the designer's own deformations as desired, and then uses that tuning as if the metric had been independently established.
self definitional [Section 3.2 (Definition and Analysis of TGSI), Section 4.3.1 (Stage One), Section 4.3.3 (Perceptual Feature Loss)]
"The TGSI metric quantifies the structural similarity between two time series images, with values ranging from -1 to 1... Since there is a significant semantic gap between natural images and time series images, we train the autoencoder specifically on time series images instead of using pre-trained image models. ... This alignment enables the temporal feature extractor to also capture image features."

The perceptual loss L_perceptual (Eq. 12) optimizes the temporal features f_time(x) to match f_enc(img(x)), the autoencoder features of the very same image representation img(x) on which TGSI (Eqs. 1-3) is computed. Because the autoencoder is trained on the identical time-to-image transformation used by TGSI, SATL's objective directly pushes predictions toward the ground truth in the representation that TGSI scores. Reporting that SATL-trained models have higher TGSI is therefore partly a self-consistency result of optimizing the same image representation, not an independent demonstration of better geometric structure.

full rationale

The MSE improvements of SATL are independent, externally benchmarked results, and the paper re-trains baselines under fixed settings; there is no load-bearing self-citation chain or uniqueness argument. However, the geometric-structure contribution is partially circular: Section 5.5 tunes the vertical expansion parameter d to make TGSI monotonic on synthetic deformations, and Section 4.3 trains the perceptual feature loss on the same time-series images that TGSI evaluates. The TGSI gains in Table 1 are thus measured with a metric whose own parameter was fitted to the authors' deformations and whose training signal is aligned to the same image modality. This is not a complete reduction because TGSI uses pixel luminance/covariance while L_perceptual uses learned latent features, and the MSE improvements stand on their own, but the 'geometric structure' evidence is not independent. Score 4 reflects partial circularity in the central TGSI/SATL evaluation, not in the numerical forecasting results.

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

The paper introduces a new measurement philosophy (geometric structure in time series via images) but no new physical entities. The central claim relies on several free parameters, most notably the vertical expansion width d, the loss weights, and the top-k frequency set. The axioms are domain assumptions about the validity of the image representation and the learned features.

free parameters (4)

Loss weights alpha, beta, gamma, delta = 0.2, 0.2, 0.1, 0.5 (tuned per dataset on validation for 'remaining datasets')
The overall loss is a weighted sum of the three SATL components plus MSE; the weights are chosen, and for some datasets they are explicitly selected according to validation performance, which is a form of fitting.
Vertical expansion width d = 100
The time-to-image transformation expands each time point vertically by d pixels. Section 5.5 shows that d materially changes TGSI behavior and that d=100 was selected to achieve near-monotonic correlation with the authors' synthetic similarity levels.
Image height and downscaling factor = height = 200; downscaling factor unspecified
The image height is fixed at 200 pixels, and the covariance computation uses a downscaled image, but the paper does not specify the downscaling factor. These choices affect TGSI values.
Top-k frequency selection = k proportional to sequence length, exact value not given
In the frequency-domain loss, the set of dominant frequencies is chosen as the top-k from the ground truth spectrum. The paper states k is proportional to sequence length but does not give the constant, leaving a free choice.

assumptions (4)

domain assumption The vertical-expansion image representation, with d=100, faithfully captures the geometric structure of time series so that luminance and covariance of the images quantify shape similarity.
This assumption underlies both TGSI and the perceptual loss training. The validation in Section 5.5 is the only evidence, and it is a self-consistency check on synthetic data.
domain assumption An autoencoder trained on time series images from the forecasting datasets learns features that are meaningful proxies for geometric structure.
The perceptual feature loss relies on these features. The paper does not evaluate the features against any external shape similarity ground truth.
ad hoc to paper Non-dominant frequency components of the ground truth are noise and can be penalized toward zero in the forecast.
The frequency loss in Eq. (7) suppresses all frequencies outside the top-k set. This may eliminate legitimate high-frequency dynamics and is a modeling assumption specific to this loss.
ad hoc to paper SSIM-style luminance and covariance (without contrast) are sufficient to measure structural similarity of time series images.
The paper drops the contrast term based on arguments about the fixed vertical intensity decay and coupling with luminance, but this is a design choice rather than a proven fact.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Towards Measuring and Modeling Geometric Structures in Time Series Forecasting via Image Modality." pith.science (2026). https://pith.science/paper/X5ZLEY5S

@misc{pith2026250723253,
  author       = {Pith},
  title        = {Pith review of: Towards Measuring and Modeling Geometric Structures in Time Series Forecasting via Image Modality},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X5ZLEY5S}},
  note         = {Machine review of arXiv:2507.23253}
}

read the original abstract

Time Series forecasting is critical in diverse domains such as weather forecasting, financial investment, and traffic management. While traditional numerical metrics like mean squared error (MSE) can quantify point-wise accuracy, they fail to evaluate the geometric structure of time series data, which is essential to understand temporal dynamics. To address this issue, we propose the time series Geometric Structure Index (TGSI), a novel evaluation metric that transforms time series into images to leverage their inherent two-dimensional geometric representations. However, since the image transformation process is non-differentiable, TGSI cannot be directly integrated as a training loss. We further introduce the Shape-Aware Temporal Loss (SATL), a multi-component loss function operating in the time series modality to bridge this gap and enhance structure modeling during training. SATL combines three components: a first-order difference loss that measures structural consistency through the MSE between first-order differences, a frequency domain loss that captures essential periodic patterns using the Fast Fourier Transform while minimizing noise, and a perceptual feature loss that measures geometric structure difference in time-series by aligning temporal features with geometric structure features through a pre-trained temporal feature extractor and time-series image autoencoder. Experiments across multiple datasets demonstrate that models trained with SATL achieve superior performance in both MSE and the proposed TGSI metrics compared to baseline methods, without additional computational cost during inference.

Figures

Figures reproduced from arXiv: 2507.23253 by the authors.

**Figure 2.** Comparison of time series y with 𝑥1 (left) and 𝑥2 (right). Both have MSE = 0.79, but 𝑥1 shows a closer geometric resemblance to y. been the main focus for evaluation, the shape of the data is often neglected, even though it is crucial for comprehending how the data changes over time. In current time series forecasting tasks, metrics like Mean Squared Error (MSE) and Mean Absolute Error (MAE) are commonly used [13, 2… view at source ↗

**Figure 3.** (a) time series to image transformation. (b) Left of dashed line: Visualization of conversion results with TGSI and its [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Overall process of computing the geometric perception loss function. Yellow indicates features, gray represents model [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: Visualization of time series forecasting tasks on the ETTh2 dataset for the 96_336 task. Each column is divided into [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Fitted curve of TGSI versus sample similarity for varying vertical extension lengths [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Comparison of different loss functions the mean value across all three deformation types: TGSI(𝑝) = 1 3 Í3 𝑖=1 TGSI(𝑦, T𝑖(𝑥)). The evaluation tests three vertical expansion configurations (𝑑 = 0, 𝑑 = 10, and 𝑑 = 100) to analyze the structural encoding capability. The r…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 28 canonical work pages

[1]

Abubakar Abid and James Y Zou. 2018. Learning a warping distance from unla- beled time series using sequence autoencoders. Advances in neural information processing systems (2018)

work page 2018
[2]

Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. 2023. Nhits: Neural hierarchi- cal interpolation for time series forecasting. InProceedings of the AAAI conference on artificial intelligence

work page 2023
[3]

Mouxiang Chen, Lefei Shen, Zhuo Li, Xiaoyun Joy Wang, Jianling Sun, and Chenghao Liu. 2024. Visionts: Visual masked autoencoders are free-lunch zero- shot time series forecasters. arXiv preprint arXiv:2408.17253 (2024)

arXiv 2024
[4]

Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qing- song Wen, Bin Yang, and Chenjuan Guo. 2024. Pathformer: Multi-scale trans- formers with adaptive pathways for time series forecasting. arXiv preprint arXiv:2402.05956 (2024)

arXiv 2024
[5]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv 2014. arXiv preprint arXiv:1412.3555 (2014)

arXiv 2014
[6]

Razvan-Gabriel Cirstea, Bin Yang, and Chenjuan Guo. 2019. Graph attention recurrent neural networks for correlated time series forecasting.MileTS19@ KDD (2019)

work page 2019
[7]

James W Cooley and John W Tukey. 1965. An algorithm for the machine calcula- tion of complex Fourier series. Mathematics of computation (1965)

work page 1965
[8]

Marco Cuturi and Mathieu Blondel. 2017. Soft-dtw: a differentiable loss function for time-series. In International conference on machine learning

work page 2017

Show all 39 references

[9]

Philippe Esling and Carlos Agon. 2012. Time-series data mining.ACM Computing Surveys (CSUR) (2012)

2012
[10]

Clive William John Granger and Paul Newbold. 2014. Forecasting economic time series. Academic press

2014
[11]

Nima Hatami, Yann Gavet, and Johan Debayle. 2018. Classification of time- series images using deep convolutional neural networks. In Tenth international conference on machine vision (ICMV 2017)

2018
[12]

Rob J Hyndman and Yeasmin Khandakar. 2008. Automatic time series forecasting: the forecast package for R. Journal of statistical software (2008)

2008
[13]

Aryan Jadon, Avinash Patil, and Shruti Jadon. 2024. A comprehensive survey of regression-based loss functions for time series forecasting. In International Conference on Data Management, Analytics & Innovation

2024
[14]

Eamonn Keogh. 2003. Efficiently finding arbitrarily scaled patterns in massive time series databases. In European Conference on Principles of Data Mining and Knowledge Discovery

2003
[15]

Tung Kieu, Bin Yang, Chenjuan Guo, Razvan-Gabriel Cirstea, Yan Zhao, Yale Song, and Christian S Jensen. 2022. Anomaly detection in time series with robust variational quasi-recurrent autoencoders. In 2022 IEEE 38th international conference on data engineering (ICDE)

2022
[16]

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

2018
[17]

Vincent Le Guen and Nicolas Thome. 2019. Shape and time distortion loss for training deep time series forecasting models. Advances in neural information processing systems (2019)

2019
[18]

Hyunwook Lee, Chunggi Lee, Hongkyu Lim, and Sungahn Ko. 2022. TILDE-Q: a transformation invariant loss function for time-series forecasting. arXiv preprint arXiv:2210.15050 (2022)

2022 arXiv
[19]

Xixi Li, Yanfei Kang, and Feng Li. 2020. Forecasting with time series imaging. Expert Systems with Applications (2020)

2020
[20]

Bryan Lim and Stefan Zohren. 2021. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A (2021)

2021
[21]

Arthur Mensch and Mathieu Blondel. 2018. Differentiable dynamic programming for structured prediction and attention. In International Conference on Machine Learning

2018
[22]

Umberto Michelucci. 2022. An introduction to autoencoders. arXiv preprint arXiv:2201.03898 (2022)

2022 arXiv
[23]

Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730 (2022)

2022 arXiv
[24]

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

2019 arXiv
[25]

Thanawin Rakthanmanon, Bilson Campana, Abdullah Mueen, Gustavo Batista, Brandon Westover, Qiang Zhu, Jesin Zakaria, and Eamonn Keogh. 2012. Searching and mining trillions of time series subsequences under dynamic time warping. In Proceedings of the 18th ACM SIGKDD internationa...

2012
[26]

David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. 1986. Learning representations by back-propagating errors. nature (1986)

1986
[27]

Shiyu Wang, Haixu Wu, Xiaoming Shi, Tengge Hu, Huakun Luo, Lintao Ma, James Y Zhang, and Jun Zhou. 2024. Timemixer: Decomposable multiscale mixing for time series forecasting. arXiv preprint arXiv:2405.14616 (2024)

2024 arXiv
[28]

Yihang Wang, Yuying Qiu, Peng Chen, Yang Shu, Zhongwen Rao, Lujia Pan, Bin Yang, and Chenjuan Guo. 2025. LightGTS: A Lightweight General Time Series Forecasting Model. arXiv preprint arXiv:2506.06005 (2025)

2025 arXiv
[29]

Yihang Wang, Yuying Qiu, Peng Chen, Kai Zhao, Yang Shu, Zhongwen Rao, Lujia Pan, Bin Yang, and Chenjuan Guo. 2025. Towards a General Time Series Forecast- ing Model with Unified Representation and Adaptive Transfer. In International Conference on Machine Learning

2025
[30]

Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing (2004)

2004
[31]

Ruofeng Wen, Kari Torkkola, Balakrishnan Narayanaswamy, and Dhruv Madeka. 2017. A multi-horizon quantile recurrent forecaster. arXiv preprint arXiv:1711.11053 (2017)

2017 arXiv
[32]

Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2022. Timesnet: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.02186 (2022)

2022 arXiv
[33]

Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. Advances in neural information processing systems (2021)

2021
[34]

Haixu Wu, Hang Zhou, Mingsheng Long, and Jianmin Wang. 2023. Interpretable weather forecasting for worldwide stations with a unified deep model. Nature Machine Intelligence (2023)

2023
[35]

Xiongxiao Xu, Haoran Wang, Yueqing Liang, Philip S Yu, Yue Zhao, and Kai Shu
[36]

Luoxiao Yang, Yun Wang, Xinqi Fan, Israel Cohen, Jingdong Chen, Yue Zhao, and Zijun Zhang. 2024. Vitime: A visual intelligence-based foundation model for time series forecasting. arXiv preprint arXiv:2407.07311 (2024)

2024
[37]

Xueyan Yin, Genze Wu, Jinze Wei, Yanming Shen, Heng Qi, and Baocai Yin. 2021. Deep learning on traffic prediction: Methods, analysis, and future directions. IEEE Transactions on Intelligent Transportation Systems (2021)

2021
[38]

Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond Efficient Transformer for Long Se- quence Time-Series Forecasting. In The Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Virtual Conference

2021
[2025]

Can Multimodal LLMs Perform Time Series Anomaly Detection? arXiv preprint arXiv:2502.17812 (2025)

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Abubakar Abid and James Y Zou. 2018. Learning a warping distance from unla- beled time series using sequence autoencoders. Advances in neural information processing systems (2018)

work page 2018

[2] [2]

Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. 2023. Nhits: Neural hierarchi- cal interpolation for time series forecasting. InProceedings of the AAAI conference on artificial intelligence

work page 2023

[3] [3]

Mouxiang Chen, Lefei Shen, Zhuo Li, Xiaoyun Joy Wang, Jianling Sun, and Chenghao Liu. 2024. Visionts: Visual masked autoencoders are free-lunch zero- shot time series forecasters. arXiv preprint arXiv:2408.17253 (2024)

arXiv 2024

[4] [4]

Peng Chen, Yingying Zhang, Yunyao Cheng, Yang Shu, Yihang Wang, Qing- song Wen, Bin Yang, and Chenjuan Guo. 2024. Pathformer: Multi-scale trans- formers with adaptive pathways for time series forecasting. arXiv preprint arXiv:2402.05956 (2024)

arXiv 2024

[5] [5]

Junyoung Chung, Caglar Gulcehre, KyungHyun Cho, and Yoshua Bengio. 2014. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv 2014. arXiv preprint arXiv:1412.3555 (2014)

arXiv 2014

[6] [6]

Razvan-Gabriel Cirstea, Bin Yang, and Chenjuan Guo. 2019. Graph attention recurrent neural networks for correlated time series forecasting.MileTS19@ KDD (2019)

work page 2019

[7] [7]

James W Cooley and John W Tukey. 1965. An algorithm for the machine calcula- tion of complex Fourier series. Mathematics of computation (1965)

work page 1965

[8] [8]

Marco Cuturi and Mathieu Blondel. 2017. Soft-dtw: a differentiable loss function for time-series. In International conference on machine learning

work page 2017