Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Towards Time Series Generation Conditioned on Unstructured Natural Language

T0 review · 3 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 A diffusion model conditioned on BERT generates time series from free-form natural language.

desk verdict A useful new dataset and task framing, wrapped in a baseline model whose quantitative evaluation doesn't yet prove text-conditioning. read the letter →

arxiv 2506.22927 v1 pith:U6TJK2GW submitted 2025-06-28 cs.LG

Jaeyun Woo , Jiseok Lee , Brian Kenji Iwana This is my paper
classification cs.LG
keywords timeseriesgenerationtext-conditioneddiffusionmodelsBERTcross-attentiontemporalU-Netdatasetnaturallanguageprompting
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 establishes a new task: generating time series from unstructured natural language prompts, such as "Rise rapidly from the middle." The authors claim this is the first time unconstrained text has been used as the conditioning signal for time series generation. To support the claim, they build a diffusion model with a temporal U-Net conditioned on a pretrained BERT language model, and they release a dataset of 63,010 time series-description pairs covering stock prices, the UCR Archive, and synthetic curves. If the claim holds, text prompts become a usable interface for custom forecasting, data augmentation, and time series manipulation.

What carries the argument

The load-bearing mechanism is a temporal denoising U-Net inside a diffusion model: 2D convolutions are replaced by 1D temporal convolutions, and the noise-prediction loss is $\mathbb{E}_{x_0,\epsilon,t}\|\epsilon - \epsilon_\theta(x_t,t)\|^2$. The conditioning path uses a pretrained BERT, and cross-attention is computed between the query from the final BERT layer and key/value projections from U-Net feature maps at four chosen layers. Training runs for 500 diffusion steps and 30 epochs, and generation takes only a few seconds. The supporting dataset pairs each 100-step series with five GPT-4o-generated descriptions (short, medium, long, creative, resemblance), plus TRUCE-derived captions reversed to serve as prompts.

What would settle it

Take a set of test prompts and have independent human raters judge, blind to source, whether the generated series or the ground-truth series better matches the prompt. If raters do not prefer the generated series at least as often as the ground truth, the claim that the model follows natural language is not supported. Alternatively, pair semantically opposite prompts that share a similar ground truth; a prompt-following model should produce markedly different outputs while the distance metric would barely move.

Watch

Extended reading notes

Core claim

The paper's central assertion is that a diffusion model can synthesize plausible 100-step time series whose shape follows an arbitrary natural language description. The model is trained to denoise Gaussian noise conditioned on text, with cross-attention between the BERT encoder's final layer and the feature maps of a temporal denoising U-Net. On examples from the test set, the generated series track the semantic content of short, medium, long, creative, and resemblance prompts, and some generations match their prompt more closely than the ground-truth series does. The authors acknowledge failures on out-of-distribution prompts and on misleading ground-truth descriptions, but they conclude that language-conditioned time series generation is demonstrably possible.

Load-bearing premise

The quantitative evaluation assumes that the ground-truth time series paired with each description is the correct target and that smaller Euclidean or DTW distance to it measures how well the generated series follows the prompt; the authors themselves note that many descriptions are too generic or sometimes misleading, which would undermine that measure.

Editorial extensions

If this is right

  • A text prompt becomes a control signal for time series shape, enabling applications such as custom forecasting, data augmentation, and transfer learning.
  • Adding creative and resemblance descriptions to training slightly worsens literal-prompt distance scores but lets the model handle more unconventional prompts.
  • The dataset's GPT-4o labels are noisy; the authors state that some descriptions are too generic or sometimes misleading, so future cleaning or alignment should improve generation quality.
  • The method is limited to prompts within the training distribution; custom prompts too far outside it produce poor outputs.
  • Because the generators and dataset are released, other researchers can train and compare on the same pairs.

Reading between the lines

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

  • One could evaluate prompt-following by captioning generated series with a vision-language model and measuring whether the caption matches the prompt, sidestepping the questionable assumption that ground-truth series are the correct targets.
  • The cross-attention recipe should transfer to other structured outputs, such as graphs or trajectories, because the conditioning is modality-agnostic once text embeddings are available.
  • The same dataset could support a testable extension: training with descriptions extracted directly from numeric features rather than from rasterized images may reduce generic labels and improve controllability.
  • Since the noising process applies to raw 100-step series, scaling to longer or multivariate series would likely require a latent-space diffusion design, an implicit next step the paper does not take.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes a diffusion-based model that generates 1D time series from unstructured natural-language descriptions. The architecture combines a temporal denoising U-Net with a pretrained BERT encoder and cross-attention between the text tokens and U-Net feature maps. The authors introduce a new dataset of 63,010 (or, including TRUCE, 70,390) time-series-description pairs assembled from stock data, the UCR Archive, synthetic functions, and the reversed TRUCE captioning dataset. The central claims are that natural-language-conditioned time series generation is feasible and that this is the first such generative model for unconstrained, unstructured text. The evidence consists of qualitative examples from held-out prompts, custom prompts, and a quantitative comparison of Euclidean/DTW distances to paired ground-truth series across several training-set ablations.

Significance. If the central claim is established, the paper opens a useful new capability: text-conditioned time series generation with potential applications in forecasting, data augmentation, and modality alignment. The strongest concrete contribution is the new public dataset, which is large and spans diverse domains, and the authors provide code and data. The qualitative results, especially the custom prompts in Fig. 6, are encouraging and suggest that the model has learned at least some text-to-shape correspondence. I see no circularity in the training/evaluation setup: the GPT-4o-generated captions are a data source, the model is trained on held-out pairs, and there is no fitted constant or derivation that forces the reported numbers. However, the quantitative evaluation does not currently support the feasibility claim: the headline metric compares outputs to one arbitrary paired ground truth, the authors themselves concede that captions are sometimes generic or misleading, and there is no unconditional baseline. The novelty claim is plausible but should be qualified as 'to our knowledge' because text-to-audio and other related generative settings exist.

major comments (3)
  1. [§5.1, Table 1] The quantitative evaluation cannot establish that generation is text-conditioned because every column in Table 1 conditions on text. The ablations 'w/o Creative', 'w/o Resembles', 'w/o both', and 'w/o TRUCE' all still receive language input, so the reported ED/DTW values could plausibly be achieved by a model that ignores the prompt and samples from the dataset marginal. Given the authors' own admission in §5.2 that captions are sometimes 'too generic or misleading', this is not a merely hypothetical risk. To make the central claim load-bearing, the evaluation needs an unconditional baseline obtained by dropping or fixing the text conditioning, plus a dataset-marginal or nearest-neighbor baseline. In addition, a direct prompt-fidelity check would help, for example a held-out classifier that predicts which prompt generated a given series, or blinded human/LLM alignment ratings on a fixed random sample of prompts.
  2. [§5.1, Table 1 and §5.2, Fig. 9] The primary metric—Euclidean and DTW distance to the paired ground-truth series—does not measure whether the generated series follows the text prompt. A text-following output can legitimately differ from one arbitrary ground-truth realization, and the authors explicitly show a case where the generated series matches the prompt better than the ground truth (Fig. 9). Thus a prompt-ignoring model could score well on this metric, and a correctly conditioned model could score poorly. The metric is therefore invalid as evidence for text conditioning unless the evaluation is restricted to pairs whose captions are verified to uniquely determine a target. The authors should either clean/filter the dataset to such pairs, or replace the metric with prompt-alignment measures that do not presuppose a unique correct series.
  3. [§4.3, Figs. 4-9] After the quantitative metric concern above, the remaining positive evidence for the feasibility claim consists of a small number of selected qualitative examples, including the custom prompts in Fig. 6. No selection protocol is reported, no number of generated samples per prompt is given, no random seeds are provided, and there are no error bars or repeated-sampling statistics. As a result, the reader cannot assess how representative these successes are. A systematic evaluation protocol—for example, a random or fixed set of 20-30 prompts per description type, repeated sampling from the diffusion model, and blinded human ratings of prompt-output alignment—should be added. The reported quantitative differences in Table 1 are also small (e.g., Short ED 30.19 for the proposed model vs. 29.08 for w/o both) and lack significance tests, so the claim that the all-data model is best is not currently supported by the table.
minor comments (6)
  1. [Abstract, §4.1.2, Conclusion] The reported dataset total is inconsistent: 7,380 TRUCE pairs plus 32,790 stock, 28,490 UCR, and 1,730 synthetic pairs sum to 70,390, while the paper repeatedly states 63,010. The 63,010 figure appears to count only the labeled dataset and excludes TRUCE; please reconcile the totals across the abstract, methodology, and conclusion.
  2. [§5.1] The text defines 'Euclidean distance' as the pairwise L1 distance between elements; this is actually the Manhattan/L1 distance, not Euclidean. Please correct the terminology or the definition.
  3. [§2.2 and References] 'As far as the applicants know' should be 'as far as the authors know', and 'Waswani et al.' should be 'Vaswani et al.'; the 'Reassemble' typo in §5.1 and 'denosing' in §3.1 should also be fixed.
  4. [§1, Contributions] The claim that this is 'the first time' time series are generated from unstructured text should be qualified as 'to our knowledge', since related text-conditioned audio/speech generation also involves generating sequential signals from language.
  5. [§4.1.2] For the stock dataset, it is not stated whether overlapping windows from the same ticker are confined to a single split; please specify the splitting procedure to rule out train/test leakage from overlapping windows.
  6. [§3.2, Eq. (4)] The notation in Eq. (4) says QLM is the input to the final transformer layer, and that KDU and VDU come from the U-Net; please clarify whether the query is actually the output hidden state of the language model, and state whether BERT is frozen or fine-tuned during training.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is an empirical training/evaluation pipeline with external data and held-out pairs, not a self-referential reduction.

full rationale

The paper's central claim is that time series generation conditioned on unstructured natural language is possible, supported by a diffusion model with BERT conditioning and a newly constructed dataset of 63,010 time series-description pairs. No load-bearing step reduces to its own inputs by construction. The GPT-4o-generated descriptions are an external data-generation tool, not a parameter fitted to the evaluation target; the model is trained on one split and evaluated on a time-series-independent test split. The quantitative metric (Euclidean and DTW distance to the paired ground-truth time series) has validity weaknesses, as the authors themselves note in Sec. 5.2 when they say some descriptions are 'too generic or sometimes misleading' and show examples where the generated series matches the prompt better than the ground truth. However, this is an evaluation-validity concern, not circularity: the reported distances do not make the feasibility claim true by definition, and the absence of an unconditional baseline is a missing comparison, not a circular reduction. There are no self-citations used as load-bearing evidence, no imported uniqueness theorem, and no fitted input renamed as a prediction. The paper is self-contained as an empirical demonstration, so the circularity score is 0.

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

The central claim rests on data construction choices (UCR caps, stock subsampling, interpolation length) and on trust in the semantic alignment of BERT and GPT-4o labels. No new physical or mathematical entities are introduced. The evaluation also assumes ground-truth distance is a faithful proxy for text fidelity, an assumption the paper itself weakens.

free parameters (3)
  • UCR samples per dataset = 50
    Hand-chosen cap on how many time series are taken from each UCR dataset to reduce redundancy and bias; affects dataset composition and model training.
  • Stock data fraction = 1/10 (6,558 series)
    Random subsample of gathered stock series, chosen to limit dataset size; could influence diversity and performance.
  • TRUCE interpolation length = 100 steps from 12
    Linear interpolation chosen to match the model output length; may distort the shape semantics described by the original TRUCE captions.
assumptions (4)
  • domain assumption BERT semantic representations are aligned well enough with time series shape concepts for cross-attention conditioning to work.
    The entire method depends on this; Section 3.2 introduces cross-attention between BERT and the U-Net without any validation that BERT features capture shape properties like 'sharp peak' or 'steady rise'.
  • domain assumption GPT-4o's descriptions of rasterized line plots are accurate and informative labels for the numeric time series.
    The large labeled dataset is built from GPT-4o outputs in the image domain, and the authors acknowledge in Section 5.2 that some descriptions are generic or misleading.
  • domain assumption Linear interpolation from 12 to 100 time steps preserves the semantic meaning of the TRUCE captions.
    TRUCE captions were written for 12-step series; Section 4.1.1 interpolates them to 100 steps, which could flatten or exaggerate features, changing the text-shape correspondence.
  • domain assumption Euclidean and DTW distance to the paired ground truth is a valid measure of how well a generated series follows its text prompt.
    Table 1 uses this metric, but Fig. 9 and Section 5.2 show cases where the prompt matches the generated series better than the ground truth, undermining the premise.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Time Series Generation Conditioned on Unstructured Natural Language." pith.science (2026). https://pith.science/paper/U6TJK2GW

@misc{pith2026250622927,
  author       = {Pith},
  title        = {Pith review of: Towards Time Series Generation Conditioned on Unstructured Natural Language},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U6TJK2GW}},
  note         = {Machine review of arXiv:2506.22927}
}
read the original abstract

Generative Artificial Intelligence (AI) has rapidly become a powerful tool, capable of generating various types of data, such as images and text. However, despite the significant advancement of generative AI, time series generative AI remains underdeveloped, even though the application of time series is essential in finance, climate, and numerous fields. In this research, we propose a novel method of generating time series conditioned on unstructured natural language descriptions. We use a diffusion model combined with a language model to generate time series from the text. Through the proposed method, we demonstrate that time series generation based on natural language is possible. The proposed method can provide various applications such as custom forecasting, time series manipulation, data augmentation, and transfer learning. Furthermore, we construct and propose a new public dataset for time series generation, consisting of 63,010 time series-description pairs.

Figures

Figures reproduced from arXiv: 2506.22927 by the authors.

Figure 1
Figure 1. Time series generation conditioned on text. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The proposed time series diffusion model [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An example from the LargeKitchenAppli￾ances dataset and the corresponding descriptions Finally, the GPT-4o was prompted with the phrase, "Describe this time series with a short, medium, and long description. Make sure to de￾scribe the overall trends and changes in direction of the line. Also, a creative description and a descrip￾tion of what it resembles." and provided the ras￾terized time series and the expected fo… view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Example of generated time series with their [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 4
Figure 4. Figure 4: Examples of generated time series from Short, [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: Examples of generated time series from cus [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 9
Figure 9. Figure 9: Examples of poor descriptions from the La [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 26 canonical work pages

  1. [1]

    Beatrice Acciaio, Stephan Eckstein, and Songyan Hou. 2024. Time-causal vae: Robust financial time series generator. arXiv preprint arXiv:2411.02947

    arXiv 2024

  2. [2]

    Yifan Bao, Yihao Ang, Qiang Huang, Anthony KH Tung, and Zhiyong Huang. 2024. Towards controllable time series generation. arXiv preprint arXiv:2403.03698

    arXiv 2024

  3. [3]

    James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, and 1 others. 2023. Improving image generation with better captions. Computer Science, 2(3):8

    work page 2023

  4. [4]

    Per-Erik Danielsson. 1980. Euclidean distance mapping. Computer Graphics and image processing, 14(3):227--248

    work page 1980

  5. [5]

    Hoang Anh Dau, Eamonn Keogh, Kaveh Kamgar, Chin-Chia Michael Yeh, Yan Zhu, Shaghayegh Gharghabi, Chotirat Ann Ratanamahatana, Yanping, Bing Hu, Nurjahan Begum, Anthony Bagnall, Abdullah Mueen, Gustavo Batista, and Hexagon-ML. 2018. The ucr time series classification archive. Https://www.cs.ucr.edu/ eamonn/ time\_series\_data\_2018/

    work page 2018

  6. [6]

    Abhyuday Desai, Cynthia Freeman, Zuhui Wang, and Ian Beaver. 2021. Timevae: A variational auto-encoder for multivariate time series generation. arXiv preprint arXiv:2111.08095

    arXiv 2021

  7. [7]

    Jacob Devlin. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

  8. [8]

    Manqing Dong, Hao Huang, and Longbing Cao. 2024. Can llms serve as time series anomaly detectors? arXiv preprint arXiv:2408.03475

    arXiv 2024

Show all 45 references
  1. [9]

    Mohammad Reza Eskandari Nasab, Shah Muhammad Hamdi, and Soukaina Filali Boubrahimi. 2024. Seriesgan: Time series generation via adversarial and autoregressive learning. In BigData, pages 860--869

    2024

  2. [10]

    Crist \'o bal Esteban, Stephanie L Hyland, and Gunnar R \"a tsch. 2017. Real-valued (medical) time series generation with recurrent conditional gans. arXiv preprint arXiv:1706.02633

    2017 arXiv

  3. [11]

    Sylvia Fr \:u hwirth-Schnatter. 1994. https://doi.org/10.1111/j.1467-9892.1994.tb00184.x Data augmentation and dynamic linear models . J. Time Series Anal., 15(2):183--202

    1994

  4. [12]

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

    2023 arXiv

  5. [13]

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. 2020. Generative adversarial networks. Communications of the ACM, 63(11):139--144

    2020

  6. [14]

    Nate Gruver, Marc Finzi, Shikai Qiu, and Andrew G Wilson. 2023. Large language models are zero-shot time series forecasters. NeurIPS, 36:19622--19635

    2023

  7. [15]

    Shota Haradal, Hideaki Hayashi, and Seiichi Uchida. 2018. https://doi.org/10.1109/embc.2018.8512396 Biosignal data augmentation based on generative adversarial networks . In EMBC

    2018

  8. [16]

    Ramin Hasibi, Matin Shokri, and Mehdi Dehghan. 2019. Augmentation scheme for dealing with imbalanced network traffic classification using deep learning. arXiv preprint arXiv:1901.00204

    2019 arXiv

  9. [17]

    Dan Hendrycks and Kevin Gimpel. 2016. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415

    2016 arXiv

  10. [18]

    Harsh Jhamtani and Taylor Berg-Kirkpatrick. 2021. Truth-conditional captioning of time series data. arXiv preprint arXiv:2110.01839

    2021 arXiv

  11. [19]

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, and 1 others. 2023. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728

    2023 arXiv

  12. [20]

    Woojeong Jin, Rahul Khanna, Suji Kim, Dong-Ho Lee, Fred Morstatter, Aram Galstyan, and Xiang Ren. 2020. Forecastqa: A question answering challenge for event forecasting with temporal text data. arXiv preprint arXiv:2005.00792

    2020 arXiv

  13. [21]

    Hyndman, and Feng Li

    Yanfei Kang, Rob J. Hyndman, and Feng Li. 2020. https://doi.org/10.1002/sam.11461 GRATIS : GeneRAting TIme series with diverse and controllable characteristics . Stat. Anal. and Data Mining: The ASA Data Sci. J

    2020 doi

  14. [22]

    D. P. Kingma and J. Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  15. [23]

    Diederik P Kingma. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  16. [24]

    Kushal Lakhotia, Eugene Kharitonov, Wei-Ning Hsu, Yossi Adi, Adam Polyak, Benjamin Bolte, Tu-Anh Nguyen, Jade Copet, Alexei Baevski, Abdelrahman Mohamed, and 1 others. 2021. On generative spoken language modeling from raw audio. Trans. Assoc. Comp. Ling., 9:1336--1354

    2021

  17. [25]

    X-L Meng. 1999. https://doi.org/10.1093/biomet/86.2.301 Seeking efficient data augmentation schemes via conditional and marginal augmentation . Biometrika, 86(2):301--320

    1999 doi

  18. [26]

    Andreas P Mentzelopoulos, Dixia Fan, Themistoklis P Sapsis, and Michael S Triantafyllou. 2024. Variational autoencoders and transformers for multivariate time-series generative modeling and forecasting: Applications to vortex-induced vibrations. Ocean Engineering, 310:118639

    2024

  19. [27]

    Raktim Mitra and Adam L MacLean. 2021. Rvagene: generative modeling of gene expression time series data. Bioinformatics, 37(19):3252--3262

    2021

  20. [28]

    Paul Mooney. 2023. Stock market data (NASDAQ, NYSE, S&P500) . Https://www.kaggle.com/datasets/ paultimothymooney/stock-market-data

    2023

  21. [29]

    OpenAI . 2024. Hello gpt-4o. Https://openai.com/index/hello-gpt-4o/

    2024

  22. [30]

    Andrew Philpot, Jose Luis Ambite, and Eduard Hovy. 2002. Dgrc askcal: Natural language question answering for energy time series. In ACM International Conference Proceeding Series, volume 129, pages 1--7

    2002

  23. [31]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In ICML, pages 8821--8831

    2021

  24. [32]

    Julius Richter, Simon Welker, Jean-Marie Lemercier, Bunlong Lay, and Timo Gerkmann. 2023. Speech enhancement and dereverberation with diffusion-based generative models. IEEE/ACM Trans. Audio, Speech, and Language Processing, 31:2351--2364

    2023

  25. [33]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684--10695

    2022

  26. [34]

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention, pages 234--241

    2015

  27. [35]

    Hiroaki Sakoe and Seibi Chiba. 2003. Dynamic programming algorithm optimization for spoken word recognition. IEEE Trans. Acoustics, Speech, and Signal Processing, 26(1):43--49

    2003

  28. [36]

    Slawek Smyl and Karthik Kuber. 2016. Data preprocessing and augmentation for multiple short time series forecasting with recurrent neural networks. In ISF

    2016

  29. [37]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

    2020 arXiv

  30. [38]

    Tanner and Wing Hung Wong

    Martin A. Tanner and Wing Hung Wong. 1987. https://doi.org/10.1080/01621459.1987.10478458 The calculation of posterior distributions by data augmentation . J. American Stat. Assoc., 82(398):528--540

    1987

  31. [39]

    A Waswani, N Shazeer, N Parmar, J Uszkoreit, L Jones, A Gomez, L Kaiser, and I Polosukhin. 2017. Attention is all you need. In NeurIPS

    2017

  32. [40]

    Magnus Wiese, Robert Knobloch, Ralf Korn, and Peter Kretschmer. 2020. Quant gans: deep generation of financial time series. Quantitative Finance, 20(9):1419--1440

    2020

  33. [41]

    Yuxin Wu and Kaiming He. 2018. Group normalization. In ECCV, pages 3--19

    2018

  34. [42]

    Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. 2023. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56(4):1--39

    2023

  35. [43]

    Yiyuan Yang, Ming Jin, Haomin Wen, Chaoli Zhang, Yuxuan Liang, Lintao Ma, Yi Wang, Chenghao Liu, Bin Yang, Zenglin Xu, and 1 others. 2024. A survey on diffusion models for time series and spatio-temporal data. arXiv preprint arXiv:2404.18886

    2024

  36. [44]

    Jinsung Yoon, Daniel Jarrett, and Mihaela Van der Schaar. 2019. Time-series generative adversarial networks. NeurIPS, 32

    2019

  37. [45]

    Liangwei Nathan Zheng, Chang George Dong, Wei Emma Zhang, Lin Yue, Miao Xu, Olaf Maennel, and Weitong Chen. 2024. Revisited large language model for time series analysis through modality alignment. arXiv preprint arXiv:2410.12326

    2024 arXiv

Pith tools

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