Pith. sign in

REVIEW 3 major objections 10 minor 59 references

HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting

T0 review · 3 major / 10 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read HDT predicts a coarse trend of the target and then generates fine discrete tokens conditioned on it, reporting average gains of 16.7% on CRPSsum and 15.4% on NRMSEsum over prior state-of-the-art on five high-dimensional benchmarks.

desk verdict Solid VQ-forecasting method with shipped code and a genuinely new hierarchical self-conditioning idea; headline SOTA gains are weakened by test-set hyperparameter selection and missing error bars. read the letter →

arxiv 2502.08302 v1 pith:NH7RSM6E submitted 2025-02-12 cs.LG cs.AI

Shibo Feng , Peilin Zhao , Liu Liu , Pengcheng Wu , Zhiqi Shen This is my paper · ORCID
classification cs.LGcs.AI
keywords hierarchicaldiscretetransformermultivariatetimeseriesforecastingvectorquantizationprobabilisticlong-termself-conditioningtokengenerationhigh-dimensional
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's central claim is that high-dimensional multivariate time series forecasting is better posed as generating discrete tokens than as denoising continuous trajectories. The proposed Hierarchical Discrete Transformer (HDT) first predicts coarse tokens of the target's own downsampled trend and then generates the fine discrete tokens of the target conditioned on that predicted trend, so the forecast horizon is represented by a short sequence of codebook indices. On five benchmarks with up to two thousand variates, HDT reports average improvements of 16.7% in CRPSsum and 15.4% in NRMSEsum over prior state-of-the-art methods, and its inference is faster than diffusion baselines because no iterative denoising is needed. If the results hold, the practical lesson is that a model's own predicted long-term trend is a cheap and effective condition for extending forecasting length in high dimensions.

What carries the argument

The load-bearing object is the two-level discrete token hierarchy. In stage 1, two vector-quantized autoencoders quantize the target and its moving-average downsampled version, defined by $X_{\mathrm{down}} = \mathrm{AvgPool}(\mathrm{Padding}(X_{\mathrm{pred}}))$, into codebook token sequences, with $\ell^2$-normalized codebook entries and quantization along the channel dimension so each token encodes cross-variate structure. In stage 2, a context encoder and a base Transformer decoder autoregressively predict the downsampled-target tokens $s_{\mathrm{down}}$ from history, and then a frozen self-conditioned Transformer decoder generates the target tokens $s_{\mathrm{pred}}$ with cross-attention over $s_{\mathrm{down}}$ and history. The stage-1 decoder maps the sampled target tokens back to a continuous forecast, and the low-level trend prediction is what lets the high level generate longer horizons accurately.

What would settle it

Run HDT against a variant that at inference time replaces the autoregressively sampled trend tokens with the ground-truth downsampled tokens. If the oracle-trend version is barely better than randomly conditioned trend tokens, the self-condition is not carrying information; if it is much better than the model-sampled trend, exposure bias is the bottleneck and the reported gains are upper bounds of what the hierarchy can deliver.

Watch

Extended reading notes

Core claim

HDT's central claim is that discretizing the forecasting target into a compact token sequence and generating that sequence hierarchically—coarse downsampled trend tokens first, fine target tokens second—scales to hundreds or thousands of variates and extends forecasting length to 144 steps without the iterative cost of diffusion models. The discovery is that the trend tokens of the target itself, produced at the low level from history alone, act as an informative self-condition for fine-token generation: they carry future information that history alone does not, and they keep long-horizon generation accurate even when the observed history is heavily masked. Ablations support this by showing that removing the self-condition or replacing discrete tokens with continuous latents degrades performance sharply, with the gap widening as the forecasting horizon grows.

Load-bearing premise

The entire gain rests on the assumption that trend tokens sampled autoregressively at inference time—trained under teacher forcing with the true downsampled future—carry enough of the actual future trend to guide fine-token generation, and the paper does not quantify the exposure-bias gap.

Editorial extensions

If this is right

  • If the reported numbers hold, discrete token generation becomes a practical alternative to diffusion for high-dimensional probabilistic forecasting, since HDT reaches better CRPSsum and NRMSEsum with lower inference time and memory.
  • The target's own predicted trend is the mechanism that unlocks long horizons: ablations show the gains over the no-self-condition variant widen at prediction lengths 96 and 144.
  • HDT remains largely accurate when up to nearly all of the observed history is masked on the Taxi benchmark, suggesting the trend condition supplies information that the context encoder alone does not.
  • The same architecture transfers to non-stationary high-dimensional datasets such as Hospital and COVID deaths with larger relative gains, indicating the mechanism is not limited to periodic benchmarks.
  • Sampling temperature materially changes CRPSsum, so temperature is an accuracy lever that must be tuned per dataset and per horizon rather than being a fixed diversity knob.

Reading between the lines

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

  • A natural extension is a three-level hierarchy that conditions on progressively finer downsamplings of the target, testing whether multi-scale self-conditioning compounds beyond two levels.
  • The per-dataset codebook is the paper's stated main limitation; a shared or universal codebook across datasets would be the direct next test of whether the discrete representations generalize.
  • The exposure-bias gap between teacher-forced trend tokens and autoregressively sampled trend tokens could be closed with scheduled sampling or a small diffusion prior over trend tokens, and the paper's own oracle-trend ablation would reveal how much headroom that leaves.
  • The $\ell^2$-normalized channel-wise codebook could also be applied to imputation or anomaly detection, where reconstructing missing channels is the same kind of token-generation problem.
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 / 10 minor

Summary. The paper proposes HDT, a two-stage vector-quantized generative framework for multivariate time series forecasting. In stage 1, VQGAN-style tokenizers are trained for the forecasting target and for a moving-average downsampled target, producing discrete token sequences. In stage 2, a low-level Transformer autoregressively generates the downsampled target tokens from the history, and a high-level Transformer generates the target tokens conditioned on those predicted trend tokens. The central claim is that this hierarchical discrete, self-conditioned design achieves state-of-the-art CRPSsum and NRMSEsum on Solar, Electricity, Traffic, Taxi, and Wikipedia, with average improvements of 16.7% and 15.4% over prior state-of-the-art methods, while also being more memory- and time-efficient than diffusion baselines.

Significance. The idea of using a predicted coarse trend of the target as a discrete self-condition for long-horizon high-dimensional forecasting is well motivated and practically relevant, and the paper ships code and a fairly broad set of ablations, including the quantized representation, the self-conditioning mechanism, temperature sensitivity, layer counts, and missing-data robustness. If the empirical claims hold under a clean evaluation protocol, this would be a useful contribution to probabilistic MTS forecasting, particularly because the discrete hierarchy offers a concrete efficiency advantage over diffusion models. However, the headline SOTA claim is not currently supported as stated because the main hyperparameters appear to be selected on test-set performance and because the statistical evidence for consistency is weak. The architectural contribution remains plausible, so the result is significant conditional on a properly controlled re-evaluation.

major comments (3)
  1. [Implementation Details; Effect of Temperature Levels in Inference; Table 4] The evaluation protocol for the headline result is not controlled. The Implementation Details state only that "We sample 100 times to report metrics on the test set," and no validation split or model-selection rule is described. Section "Effect of Temperature Levels in Inference" shows that temperature is searched over {1.0, 2.0, 3.0, 5.0, 8.0} ({1.0, 1.5, 2.0, 3.0, 6.0} for Solar) and that the choice matters materially, and Table 4 selects the number of self-conditioned layers per dataset and prediction length. If the HDT rows in Table 1 are the best over these grids evaluated on the test set, the reported average gains of 16.7% CRPSsum and 15.4% NRMSEsum are post-hoc selections rather than predictive performance, and the baselines are not given the same test-set tuning latitude. Please specify a temporal validation split, select all hyperparameters on validation data only, and report both validation-selected and test-set numbers; if this cannot be done, the SOTA claim should be restated accordingly.
  2. [Table 1; Table 12] Table 1 reports point estimates without error bars, while Table 12 shows that the advantage over the strongest baselines is often within one standard deviation. For example, Solar-48 CRPSsum is 0.329 +/- 0.004 for HDT versus 0.328 +/- 0.006 for MG-TSD and 0.324 +/- 0.005 for TSDiff, and Wikipedia-48 CRPSsum is 0.073 +/- 0.004 for HDT versus 0.063 +/- 0.003 for VQ-TR. The claim of "consistent state-of-the-art performance" is therefore not statistically established. Add standard deviations to the main table and report paired significance tests (across seeds or test windows) for the average improvements, or weaken the consistency claim accordingly.
  3. [Table 2] The continuous baseline in Table 2 does not support the conclusion that discretization is responsible for the gains. The C-Transformer, which uses continuous latents from stage 1 in the same stage-2 pipeline, reports CRPSsum 0.327 on Electricity-48 and 0.861 on Taxi-48, versus 0.025 and 0.166 for HDT; these values are implausibly poor for a Transformer operating on learned continuous representations and suggest an uncontrolled difference in training, capacity, normalization, or input distribution rather than a clean test of the quantization hypothesis. The explanation that continuous z acts as an "infinitely large codebook" is not a quantitative mechanism. Please provide a controlled continuous baseline with matched architecture, training budget, and hyperparameters, and report the reconstruction quality of the continuous latents, before attributing the gains to discretization.
minor comments (10)
  1. [Main results] The term "CRPM sum" appears twice in the probabilistic forecasting paragraph and should be "CRPSsum."
  2. [Table 2 caption] The caption contains the typo "C-Trasformer" and should read "C-Transformer."
  3. [Effect of Number of Layers in Eqn. (13)] The text refers to "our SDT" when describing the layer-count experiment; this should be "our HDT."
  4. [Table 6 caption] The caption contains the typo "whih" and should read "which."
  5. [Table 1, 1stCount row] The "1stCount" row is unclear: the counts cannot be reconciled with the 14 settings shown, and the row should be explained or removed.
  6. [Appendix B] The paper lists total time steps for each dataset but never states the train/validation/test split sizes; please add this information, particularly because the evaluation protocol is in question.
  7. [Algorithm 2, line 7] Please clarify whether the sdown used to train the high-level decoder is sampled, and if so with what temperature, or generated greedily; this detail is needed to reproduce and understand the self-conditioning training.
  8. [Equation (1)] The token sequence lengths s_d and s_p are used in stage 2 but are not defined in terms of tau and the downsampling/pooling parameters; please define them explicitly.
  9. [Equations (14)-(15)] The paper does not state how the deterministic point forecast is computed from the 100 generated samples; please specify whether the median, mean, or another summary is used for NRMSEsum.
  10. [Figure 4 and Appendix D] The efficiency comparison is presented only as a figure; include a table with numerical memory usage and per-sample inference time so the claimed advantages can be verified quantitatively.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported SOTA gains are selected from test-set hyperparameter tuning rather than predicted by a fixed model.

  1. fitted input called prediction [Implementation Details; 'Effect of Temperature Levels in Inference'; 'Effect of Number of Layers in Eqn. (13)'; Tables 4 and 7]
    "We sample 100 times to report metrics on the test set. ... During our experiments, we observed that sampling temperature is a crucial hyperparameter in a probabilistic setting. As shown in Figure 2, tests on the Traffic and Taxi datasets revealed significant differences in results with varying temperatures. ... we report the CRPSsum and NRMSEsum results of our SDT with different number of layers (e.g.2, 3, 4, 5) in Table 4."

    The final test-set numbers in Table 1 are produced after selecting sampling temperature and self-condition layer count from grids evaluated on the same test set (Figure 2; Table 4; Table 7 gives only ranges, with no validation split or selection protocol). The reported headline gains (16.7% CRPSsum, 15.4% NRMSEsum) are therefore the best over hyperparameters chosen using the test labels, i.e., a selected statistic rather than a prediction of a fixed model. Because baselines are not given the same test-set tuning latitude, the claimed comparison is not controlled and the central SOTA claim reduces, for the reported numbers, to test-set selection.

full rationale

The model's internal derivation chain is not circular: the low-level trend tokens are predicted from history at inference (Algorithm 3), and the high-level decoder is trained with low-level outputs (Algorithm 2), so conditioning on a predicted coarse target is a genuine hierarchical forecast rather than a definitional use of the target. Stage-1 tokenization and reconstruction are standard VQGAN losses (Eqs. 2-8), and no load-bearing self-citation or imported uniqueness theorem is present. However, the empirical central claim—state-of-the-art test-set CRPSsum/NRMSEsum—is contaminated by test-set hyperparameter selection: temperature and layer counts are varied per dataset/length (Figure 2, Table 4) with the ranges in Table 7, and the final metrics are reported on the same test set with no validation split. The reported improvements are thus partially forced by construction (minimum over a grid on the test labels). This is a partial, not total, circularity because the architectural comparison and benchmark numbers are externally defined; a proper validation-then-test protocol could in principle preserve or overturn the ranking.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard deep learning components and three domain assumptions: teacher-forced training transfers to sampled conditioning, the moving-average trend is a useful self-condition, and VQ codebooks preserve forecast-relevant information. No invented entities or free physical parameters are introduced.

free parameters (2)
  • Sampling temperature (inference) = Chosen per dataset and prediction length from {1.0, 1.5, 2.0, 3.0, 6.0} or {1.0, 2.0, 3.0, 5.0, 8.0}
    Table 7 lists temperature grids per dataset; Figure 2 shows CRPSsum across temperatures, implying the reported results use the best test-set temperature. This is a fitted value that directly influences the headline CRPSsum numbers.
  • Number of self-conditioned Transformer layers = 2, 3, 4, or 5; best per dataset and horizon (e.g., 4 for Taxi 96, 2 for Traffic 48 in Table 4)
    Table 4 sweeps layers per dataset and prediction length and reports the best. The paper fixes base decoder layers at 3 but varies self-cond layers, making the architecture choice data-dependent.
assumptions (4)
  • domain assumption Teacher forcing: during stage-2 training the low-level and high-level decoders are conditioned on the true downsampled targets and true target tokens, respectively (Algorithms 1-2).
    The model's training objective assumes that conditioning on ground-truth coarse tokens transfers to inference, where coarse tokens are sampled from the low-level model; exposure bias is not analyzed.
  • domain assumption The moving-average downsampling in Eqn. (1) captures the long-term trend of the target and is a useful self-condition.
    The AvgPool with padding operation is adopted from Autoformer and is asserted to highlight long-term patterns; no alternative trend extraction is compared.
  • domain assumption The VQGAN codebooks learned in stage 1 preserve enough information for forecasting, not merely reconstruction.
    The paper validates discrete reparametrization indirectly via the C-Transformer ablation, but there is no measure of how much forecast-relevant information is lost in quantization.
  • standard math Standard Transformer and convolution architecture components behave as described in prior literature.
    Cross-attention, LayerNorm, and Conv1d operations are standard; the paper does not derive new math.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting." pith.science (2026). https://pith.science/paper/NH7RSM6E

@misc{pith2026250208302,
  author       = {Pith},
  title        = {Pith review of: HDT: Hierarchical Discrete Transformer for Multivariate Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NH7RSM6E}},
  note         = {Machine review of arXiv:2502.08302}
}
read the original abstract

Generative models have gained significant attention in multivariate time series forecasting (MTS), particularly due to their ability to generate high-fidelity samples. Forecasting the probability distribution of multivariate time series is a challenging yet practical task. Although some recent attempts have been made to handle this task, two major challenges persist: 1) some existing generative methods underperform in high-dimensional multivariate time series forecasting, which is hard to scale to higher dimensions; 2) the inherent high-dimensional multivariate attributes constrain the forecasting lengths of existing generative models. In this paper, we point out that discrete token representations can model high-dimensional MTS with faster inference time, and forecasting the target with long-term trends of itself can extend the forecasting length with high accuracy. Motivated by this, we propose a vector quantized framework called Hierarchical Discrete Transformer (HDT) that models time series into discrete token representations with l2 normalization enhanced vector quantized strategy, in which we transform the MTS forecasting into discrete tokens generation. To address the limitations of generative models in long-term forecasting, we propose a hierarchical discrete Transformer. This model captures the discrete long-term trend of the target at the low level and leverages this trend as a condition to generate the discrete representation of the target at the high level that introduces the features of the target itself to extend the forecasting length in high-dimensional MTS. Extensive experiments on five popular MTS datasets verify the effectiveness of our proposed method.

Figures

Figures reproduced from arXiv: 2502.08302 by the authors.

Figure 1
Figure 1. An illustration of our proposed HDT is provided. In stage 1, the model generates discrete downsampled targets and [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Performance of HDT with different temperature levels of different prediction lengths in Traffic and Taxi datasets. The [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Probabilistic and deterministic performance of HDT and HDT-variants on different prediction length and datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Model memory usage and time efficiency comparison under input-96-predict-48, 96 of Traffic and Taxi, respectively. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Comparison of prediction intervals with TiemGrad and MG [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: The forecasting results of 16 samples from the Solar dataset with input-96-predict-48. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: The forecasting results of 16 samples from the Electricity dataset with input-96-predict-48. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The forecasting results of 16 samples from the Traffic dataset with input-96-predict-96. [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: The forecasting results of 16 samples from the Taxi dataset with input-96-predict-96. [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: The forecasting results of 16 samples from the Wikipedia dataset with input-96-predict-96. [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: The forecasting results of 16 samples from the hospital dataset with input-24-predict-48. [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 33 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Alcaraz, J. M. L.; and Strodthoff, N. 2022. Diffusion-based time series imputation and forecasting with structured state space models. arXiv preprint arXiv:2208.09399

    arXiv 2022

  4. [4]

    Bian, Y.; Ju, X.; Li, J.; Xu, Z.; Cheng, D.; and Xu, Q. 2024. Multi-patch prediction: Adapting llms for time series representation learning. arXiv preprint arXiv:2402.04852

    arXiv 2024

  5. [5]

    Brophy, E.; Wang, Z.; She, Q.; and Ward, T. 2023. Generative adversarial networks in time series: A systematic literature review. ACM Computing Surveys, 55(10): 1--31

    work page 2023

  6. [6]

    T.; Rubinstein, M.; et al

    Chang, H.; Zhang, H.; Barber, J.; Maschinot, A.; Lezama, J.; Jiang, L.; Yang, M.-H.; Murphy, K.; Freeman, W. T.; Rubinstein, M.; et al. 2023. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704

    arXiv 2023

  7. [7]

    Chen, X.; Mishra, N.; Rohaninejad, M.; and Abbeel, P. 2018. Pixelsnail: An improved autoregressive generative model. In International Conference on Machine Learning, 864--872. PMLR

    work page 2018

  8. [8]

    Desai, A.; Freeman, C.; Wang, Z.; and Beaver, I. 2021. Timevae: A variational auto-encoder for multivariate time series generation. arXiv preprint arXiv:2111.08095

    arXiv 2021

Show all 59 references
  1. [9]

    Dinh, L.; Sohl-Dickstein, J.; and Bengio, S. 2016. Density estimation using real nvp. arXiv preprint arXiv:1605.08803

    2016 arXiv

  2. [10]

    Dong, E.; Du, H.; and Gardner, L. 2020. An interactive web-based dashboard to track COVID-19 in real time. The Lancet infectious diseases, 20(5): 533--534

    2020

  3. [11]

    Esser, P.; Rombach, R.; and Ommer, B. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12873--12883

    2021

  4. [12]

    Fan, X.; Wu, Y.; Xu, C.; Huang, Y.; Liu, W.; and Bian, J. 2024. MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process. arXiv preprint arXiv:2403.05751

    2024 arXiv

  5. [13]

    Feng, S.; Miao, C.; Xu, K.; Wu, J.; Wu, P.; Zhang, Y.; and Zhao, P. 2023. Multi-scale attention flow for probabilistic time series forecasting. IEEE Transactions on Knowledge and Data Engineering

    2023

  6. [14]

    Feng, S.; Miao, C.; Zhang, Z.; and Zhao, P. 2024. Latent diffusion transformer for probabilistic time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, 11979--11987

    2024

  7. [15]

    Feng, S.; Xu, C.; Zuo, Y.; Chen, G.; Lin, F.; and XiaHou, J. 2022. Relation-aware dynamic attributed graph attention network for stocks recommendation. Pattern Recognition, 121: 108119

    2022

  8. [16]

    S.; Salinas, D.; Flunkert, V.; and Januschowski, T

    Gasthaus, J.; Benidis, K.; Wang, Y.; Rangapuram, S. S.; Salinas, D.; Flunkert, V.; and Januschowski, T. 2019. Probabilistic forecasting with spline quantile function RNNs. In The 22nd international conference on artificial intelligence and statistics, 1901--1910. PMLR

    2019

  9. [17]

    Gray, R. 1984. Vector quantization. IEEE Assp Magazine, 1(2): 4--29

    1984

  10. [18]

    Han, X.; Zheng, H.; and Zhou, M. 2022. Card: Classification and regression diffusion models. Advances in Neural Information Processing Systems, 35: 18100--18115

    2022

  11. [19]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33: 6840--6851

    2020

  12. [20]

    B.; Ord, J

    Hyndman, R.; Koehler, A. B.; Ord, J. K.; and Snyder, R. D. 2008. Forecasting with exponential smoothing: the state space approach. Springer Science & Business Media

    2008

  13. [21]

    Isola, P.; Zhu, J.-Y.; Zhou, T.; and Efros, A. A. 2017. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1125--1134

    2017

  14. [22]

    Y.; Wen, Q.; Zambon, D.; Alippi, C.; Webb, G

    Jin, M.; Koh, H. Y.; Wen, Q.; Zambon, D.; Alippi, C.; Webb, G. I.; King, I.; and Pan, S. 2023. A survey on graph neural networks for time series: Forecasting, classification, imputation, and anomaly detection. arXiv preprint arXiv:2307.03759

    2023 arXiv

  15. [23]

    Kim, T.; Kim, J.; Tae, Y.; Park, C.; Choi, J.-H.; and Choo, J. 2021. Reversible instance normalization for accurate time-series forecasting against distribution shift. In International Conference on Learning Representations

    2021

  16. [24]

    P.; and Welling, M

    Kingma, D. P.; and Welling, M. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114

    2013 arXiv

  17. [25]

    F.; Bohlke-Schneider, M.; Zschiegner, J.; Wang, H.; and Wang, Y

    Kollovieh, M.; Ansari, A. F.; Bohlke-Schneider, M.; Zschiegner, J.; Wang, H.; and Wang, Y. 2023. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. arXiv preprint arXiv:2307.11494

    2023 arXiv

  18. [26]

    F.; Bohlke-Schneider, M.; Zschiegner, J.; Wang, H.; and Wang, Y

    Kollovieh, M.; Ansari, A. F.; Bohlke-Schneider, M.; Zschiegner, J.; Wang, H.; and Wang, Y. B. 2024. Predict, refine, synthesize: Self-guiding diffusion models for probabilistic time series forecasting. Advances in Neural Information Processing Systems, 36

    2024

  19. [27]

    Lai, G.; Chang, W.-C.; Yang, Y.; and Liu, H. 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, 95--104

    2018

  20. [28]

    Lee, D.; Malacarne, S.; and Aune, E. 2023. Vector Quantized Time Series Generation with a Bidirectional Prior Model. arXiv preprint arXiv:2303.04743

    2023 arXiv

  21. [29]

    Li, Y.; Chen, W.; Hu, X.; Chen, B.; Zhou, M.; et al. 2024. Transformer-Modulated Diffusion Models for Probabilistic Multivariate Time Series Forecasting. In The Twelfth International Conference on Learning Representations

    2024

  22. [30]

    Li, Y.; Lu, X.; Wang, Y.; and Dou, D. 2022. Generative time series forecasting with diffusion, denoise, and disentanglement. Advances in Neural Information Processing Systems, 35: 23009--23022

    2022

  23. [31]

    C.; Zhao, P.; and Sun, J

    Liu, C.; Hoi, S. C.; Zhao, P.; and Sun, J. 2016. Online arima algorithms for time series prediction. In Proceedings of the AAAI conference on artificial intelligence

    2016

  24. [32]

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

    2023 arXiv

  25. [33]

    Liu, Y.; Wu, H.; Wang, J.; and Long, M. 2022. Non-stationary transformers: Exploring the stationarity in time series forecasting. Advances in Neural Information Processing Systems, 35: 9881--9893

    2022

  26. [34]

    E.; and Winkler, R

    Matheson, J. E.; and Winkler, R. L. 1976. Scoring rules for continuous probability distributions. Management science, 22(10): 1087--1096

    1976

  27. [35]

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

    Nie, Y.; Nguyen, N. H.; Sinthong, P.; and Kalagnanam, J. 2022. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730

    2022 arXiv

  28. [36]

    Papamakarios, G.; Pavlakou, T.; and Murray, I. 2017. Masked autoregressive flow for density estimation. Advances in neural information processing systems, 30

    2017

  29. [37]

    Qiu, M.; Zhao, P.; Zhang, K.; Huang, J.; Shi, X.; Wang, X.; and Chu, W. 2017. A short-term rainfall prediction model using multi-task convolutional neural networks. In 2017 IEEE international conference on data mining (ICDM), 395--404. IEEE

    2017

  30. [38]

    Ramesh, A.; Pavlov, M.; Goh, G.; Gray, S.; Voss, C.; Radford, A.; Chen, M.; and Sutskever, I. 2021. Zero-shot text-to-image generation. In International Conference on Machine Learning, 8821--8831. PMLR

    2021

  31. [39]

    Rao, H.; Li, Y.; and Miao, C. 2022. Revisiting k-reciprocal distance re-ranking for skeleton-based person re-identification. IEEE Signal Processing Letters, 29: 2103--2107

    2022

  32. [40]

    Rao, H.; Xu, S.; Hu, X.; Cheng, J.; and Hu, B. 2021. Multi-Level Graph Encoding with Structural-Collaborative Relation Learning for Skeleton-Based Person Re-Identification. In Zhou, Z.-H., ed., Proceedings of the Thirtieth International Joint Conference on Artificial Intellige...

    2021

  33. [41]

    Rasul, K.; Bennett, A.; Vicente, P.; Gupta, U.; Ghonia, H.; Schneider, A.; and Nevmyvaka, Y. 2023. VQ-TR: Vector Quantized Attention for Time Series Forecasting. In The Twelfth International Conference on Learning Representations

    2023

  34. [42]

    Rasul, K.; Bennett, A.; Vicente, P.; Gupta, U.; Ghonia, H.; Schneider, A.; and Nevmyvaka, Y. 2024. VQ-TR: Vector Quantized Attention for Time Series Forecasting. In The Twelfth International Conference on Learning Representations

    2024

  35. [43]

    Rasul, K.; Seward, C.; Schuster, I.; and Vollgraf, R. 2021. Autoregressive denoising diffusion models for multivariate probabilistic time series forecasting. In International Conference on Machine Learning, 8857--8868. PMLR

    2021

  36. [44]

    Rasul, K.; Sheikh, A.-S.; Schuster, I.; Bergmann, U.; and Vollgraf, R. 2020. Multivariate probabilistic time series forecasting via conditioned normalizing flows. arXiv preprint arXiv:2002.06103

    2020 arXiv

  37. [45]

    Razavi, A.; Van den Oord, A.; and Vinyals, O. 2019. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32

    2019

  38. [46]

    Salinas, D.; Bohlke-Schneider, M.; Callot, L.; Medico, R.; and Gasthaus, J. 2019. High-dimensional multivariate forecasting with low-rank gaussian copula processes. Advances in neural information processing systems, 32

    2019

  39. [47]

    Salinas, D.; Flunkert, V.; Gasthaus, J.; and Januschowski, T. 2020. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting, 36(3): 1181--1191

    2020

  40. [48]

    B.; Gudelek, M

    Sezer, O. B.; Gudelek, M. U.; and Ozbayoglu, A. M. 2020. Financial time series forecasting with deep learning: A systematic literature review: 2005--2019. Applied soft computing, 90: 106181

    2020

  41. [49]

    Shen, L.; and Kwok, J. 2023. Non-autoregressive Conditional Diffusion Models for Time Series Prediction. arXiv preprint arXiv:2306.05043

    2023 arXiv

  42. [50]

    Tashiro, Y.; Song, J.; Song, Y.; and Ermon, S. 2021. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems, 34: 24804--24816

    2021

  43. [51]

    Van Den Oord, A.; Vinyals, O.; et al. 2017. Neural discrete representation learning. Advances in neural information processing systems, 30

    2017

  44. [52]

    Woo, G.; Liu, C.; Kumar, A.; Xiong, C.; Savarese, S.; and Sahoo, D. 2024. Unified training of universal time series forecasting transformers. arXiv preprint arXiv:2402.02592

    2024 arXiv

  45. [53]

    Wu, H.; Xu, J.; Wang, J.; and Long, M. 2021. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in Neural Information Processing Systems, 34: 22419--22430

    2021

  46. [54]

    Wu, S.; Xiao, X.; Ding, Q.; Zhao, P.; Wei, Y.; and Huang, J. 2020. Adversarial sparse transformer for time series forecasting. Advances in neural information processing systems, 33: 17105--17115

    2020

  47. [55]

    Yoon, J.; Jarrett, D.; and Van der Schaar, M. 2019. Time-series generative adversarial networks. Advances in neural information processing systems, 32

    2019

  48. [56]

    Y.; Zhang, H.; Pang, R.; Qin, J.; Ku, A.; Xu, Y.; Baldridge, J.; and Wu, Y

    Yu, J.; Li, X.; Koh, J. Y.; Zhang, H.; Pang, R.; Qin, J.; Ku, A.; Xu, Y.; Baldridge, J.; and Wu, Y. 2021. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627

    2021 arXiv

  49. [57]

    Zheng, C.; Vuong, T.-L.; Cai, J.; and Phung, D. 2022. Movq: Modulating quantized vectors for high-fidelity image generation. Advances in Neural Information Processing Systems, 35: 23412--23425

    2022

  50. [58]

    Zhicheng, C.; SHIBO, F.; Zhang, Z.; Xiao, X.; Gao, X.; and Zhao, P. 2024. SDformer: Similarity-driven Discrete Transformer For Time Series Generation. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  51. [59]

    Zhou, T.; Niu, P.; Sun, L.; Jin, R.; et al. 2023. One fits all: Power general time series analysis by pretrained lm. Advances in neural information processing systems, 36: 43322--43355

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.