Pith. sign in

REVIEW 3 major objections 5 minor 47 references

TimeCapsule: Solving the Jigsaw Puzzle of Long-Term Time Series Forecasting with Compressed Predictive Representations

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read TimeCapsule claims long-term multivariate forecasting reduces to one compression principle—shrink a 3D time-variate-level tensor, predict in the latent space, expand back—and reports best or near-best error on ten benchmarks.

desk verdict TimeCapsule is a well-built, honestly ablated architecture, but its SOTA claim is not established because the baselines ran with shorter tuned lookbacks while TimeCapsule used a fixed 512. read the letter →

arxiv 2504.12721 v3 pith:OIFCACSF submitted 2025-04-17 cs.LG cs.AIeess.SP

classification cs.LGcs.AIeess.SP
keywords long-termtimeseriesforecastingmultivariate3Dtensormodelingmodeproductinformationcompressionmulti-scalejoint-embeddingpredictivearchitecturerepresentationlearning
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

TimeCapsule is built on a single thesis: the many techniques that power modern long-term time series forecasters—redundancy reduction, multi-scale modelling, cross-variate attention, and linear decoding—are different faces of one operation, high-dimensional information compression. To test that thesis, the paper treats a multivariate series as a 3D tensor with time, variate, and level dimensions, applies learnable low-rank mode products to shrink each dimension, does attention in this compressed space, then decodes the future with a few MLP layers. The paper reports that this single architecture achieves the best or near-best forecast error on ten multivariate benchmarks, with its clearest margin on a traffic network dataset where it claims an average 15.8% lower mean absolute error than the second-best model. A reader should care because, if correct, it means a fixed simple recipe can replace a menu of specialised designs and make long lookbacks computationally cheap.

What carries the argument

The load-bearing object is the 3D tensor representation of the series together with the mode product: a learned matrix multiplication along one axis of a tensor after folding the other two axes, used here to compress each dimension before attention is applied. TimeCapsule's Mode-Specific Multi-head Self-Attention (MoMSA) applies vanilla self-attention to the folded matrix in the compressed space, so the attention map is itself compressed; the transform is written as the product of an expanding and a compressing matrix, allowing lossy low-rank compression rather than fixed patching. The second mechanism is residual information back: each compression step records what was lost and concatenates it into the decoder, so the final forecast is not forced to recover everything from the capsule. The third is the JEPA internal forecast, which predicts the compressed representation of the future from the compressed representation of the past, using a stop-gradient EMA target encoder to supervise the latent space.

What would settle it

Rerun every baseline and TimeCapsule under identical preprocessing and hyperparameter search, with lookback fixed at 512 for all of them rather than searched for baselines alone, and also with each model given its own best lookback; if the reported margins vanish or reverse under either protocol, the central claim of universal superiority fails.

Watch

Extended reading notes

Core claim

The central claim is that long-term multivariate forecasting is best organised as a compression problem, not an architecture-design problem. TimeCapsule encodes an observed window $X \in \mathbb{R}^{v \times t_x \times 1}$ by sequentially compressing the temporal, level, and variate axes with learnable mode-product transforms, producing a small capsule $X_3 \in \mathbb{R}^{v_c \times t_c \times l}$. A single linear layer predicts the future capsule, and three MLP blocks expand it back to the real time axis, using stored residuals $B_1, B_2, B_3$ to return detail lost in compression. A Joint-Embedding Predictive Architecture loss compares the encoded observed representation with an EMA-encoded target representation and is added to the forecast loss. In the reported experiments, the model matches or approaches the best result across all datasets and horizons, and removing the residual feedback degrades performance sharply.

Load-bearing premise

The headline comparison assumes that giving TimeCapsule a fixed 512-step lookback while letting each baseline search its own lookback is a fair protocol; if the baselines were handicapped by suboptimal windows, the reported margins could shrink or reverse.

Editorial extensions

If this is right

  • A single architecture with a weak transformer encoder and an MLP decoder can cover both datasets that need explicit multi-scale modelling and datasets where a pure MLP is nearly enough, by learning how much to compress rather than hard-coding patches.
  • Long lookbacks become cheap: because the temporal axis is compressed before self-attention, a 512-step history can be processed without quadratic attention cost, and the paper's lookback study shows accuracy keeps improving as the window grows.
  • The residual feedback is load-bearing: ablation shows that removing it causes a dramatic drop in accuracy, so any compressed-forecasting design must solve the information-recovery problem, not just the compression problem.
  • The JEPA loss is a monitoring signal as much as a training signal: it decreases even without backpropagation, and including it in training helps on some datasets while being neutral on others, suggesting latent-space prediction can guide convergence.

Reading between the lines

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

  • If the learned level-expansion matrix really decomposes the series into additive level components—the paper shows the sum of sub-level series nearly reconstructs the input—the same transform could be reused as a trainable, dataset-specific decomposition tool for other tasks, independent of forecasting.
  • The comparison protocol deserves a direct stress test: rerun every baseline under the same fixed lookback of 512, and also rerun TimeCapsule under each baseline's searched lookback, to separate the benefit of the architecture from the benefit of the longer window.
  • The JEPA target construction (zero-padding short targets, EMA-averaging long targets) is one design choice among several; comparing it against alternatives such as random chunk pooling or a learned target encoder would show whether the decreasing JEPA loss is a meaningful signal or a side effect of preprocessing.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes TimeCapsule, a multivariate long-term time series forecasting architecture that treats the input as a 3D tensor with time, variate, and level dimensions. The encoder applies mode-specific multi-head self-attention with low-rank compression transforms (MoMSA) along each dimension, and residual information is passed back to a simple MLP decoder that reconstructs the future. A Joint-Embedding Predictive Architecture (JEPA) loss, computed between the compressed representation of the input and an EMA-encoded preprocessing of the target, is added as an auxiliary objective. Experiments on ten benchmarks compare TimeCapsule with eight TFB-derived baselines and report best or near-best MSE/MAE in most settings, together with ablations for residual compensation, compression dimensions, noise, block order, positional encoding, and a classification extension.

Significance. If the empirical claims hold, TimeCapsule is a useful contribution. It unifies multi-level modeling, multi-mode dependency capture, and linear projection in a relatively simple architecture, and it demonstrates how a JEPA-style loss can be attached to a forecasting model. The paper is candid about the auxiliary and dataset-dependent effect of the JEPA term, it ships code, and it provides component-level ablations and visualizations of learned transform matrices. The main risk is not internal circularity but external validity: the headline state-of-the-art claim rests on an asymmetric lookback protocol and on single-run baseline numbers, and several margins are smaller than the model's own seed-to-seed variability reported in the appendix.

major comments (3)
  1. [§5.1.2 / Table 1 caption / Appendix C.1] The headline comparison is asymmetric: TimeCapsule is evaluated with a fixed lookback of 512 for all datasets except ILI, while every baseline uses lookback lengths searched for the best performance. The caption does not report the actual lookbacks selected for the baselines; if those are shorter than 512, Table 1 conflates the architecture with access to a longer input context. Appendix C.1 (Fig. 8) shows that TimeCapsule's error consistently decreases as the lookback grows from 96 to 720 for ETTh1, Weather, PEMS04, and Traffic. Since many margins are tiny (e.g., Weather 96 MSE is 0.141 vs. 0.142 for the two TimeCapsule columns), a single lookback choice could reverse several rankings. Please re-run the baselines at T=512 under the TFB protocol, or run TimeCapsule at each baseline's searched lookback, and report both configurations side by side.
  2. [§4.2.3, Eq. (16)] The target preprocessing for the JEPA loss is a heuristic that is neither derived nor validated. Zero-padding when t_y < t_x and EMA-averaging chunks when t_y > t_x produce a target tensor of length t_x, but no evidence is given that this target is a meaningful prediction target in the representation space; for t_y > t_x, the EMA mixes future points across chunks in a way that may not correspond to any actual input-like sequence. Moreover, the JEPA loss itself is never defined: the text gives 'Loss(Enc_x(X), sg(Enc_y(Y))' with no closing parenthesis and no specification of the distance used. Because the paper's own ablation in §5.2.2 shows the JEPA effect is small and dataset-dependent, this missing definition and missing validation matter. Please specify the loss explicitly and compare the proposed padding/EMA scheme with simpler alternatives, such as using the last t_x points of the target without preprocessing.
  3. [Appendix B / Table 5] The robustness analysis reports standard deviations only for TimeCapsule, and only for four datasets, while all baseline numbers in Table 1 are point values taken from TFB. Given margins at the third decimal place, this is insufficient to substantiate the claim in §5.1.3 that TimeCapsule 'consistently achieves or approaches the best performance.' For example, the Weather 96 entry 0.141 in Table 1 is within one standard deviation of the 0.143±0.001 reported in Table 5, so a second-run difference of 0.002 could change the ranking. Please provide baseline error bars or a sensitivity analysis, at least for the settings where TimeCapsule's margin over the second-best model is smaller than the variability of either method.
minor comments (5)
  1. [§3, Definition 1] The definitions of Fold and Unfold appear to be swapped relative to standard tensor terminology: unfolding maps a tensor to a matrix and folding maps a matrix back to a tensor. Please correct the terminology or state the convention explicitly to avoid confusion in Eqs. (11)–(15).
  2. [§3 and §4.1] The text uses both 'mode production' and 'mode product'; these should be unified to 'mode product' throughout.
  3. [Appendix C.1] The text states that the lookback window ranges from 96 to 512, but the horizontal axis of Fig. 8 goes to 720. Align the text and figure.
  4. [Table 1] The row labeled '1st Count' at the bottom of the table is not explained in the caption. Please specify whether it counts best MSE, best MAE, or best on either metric, and over which set of settings.
  5. [§5.1.2] The hyperparameter selection is reported only as ranges (batch size 32–128, learning rate 1e-4 to 2e-3). Please list the per-dataset selected values for reproducibility, since the appendix reports multiple seeds but the main table is for a single seed (2021).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TimeCapsule's reported gains are empirical benchmark results against external TFB baselines, not quantities derived by construction from fitted parameters or self-citations.

full rationale

TimeCapsule is an empirical architecture paper. Its chain of reasoning is architectural: a 3D tensor encoding, mode-product compression, residual information back, a JEPA auxiliary loss, and a supervised forecasting objective. No equation in the paper reduces the forecast to a fitted constant or to the paper's own inputs. The compression dimensions are hyperparameters chosen from a small grid and reported in ablations; they are not renamed predictions. The JEPA target uses an EMA of the input encoder, but the paper explicitly and candidly shows the JEPA effect is small and dataset-dependent (Section 5.2.2, Fig. 5), so it is not load-bearing. There are no self-citations by the authors: the cited works (RevIn, iTransformer, JEPA, TFB) are independent and external. The only notable concern is the asymmetric lookback protocol: TimeCapsule uses a fixed T=512 while baselines use searched lookbacks (Table 1 caption), and Appendix C.1 shows longer lookbacks improve TimeCapsule. That is a fairness/correctness concern about benchmark comparability, not a circularity: the reported numbers are not derived by construction from the baseline numbers, and no fitted parameter is relabeled as a prediction. Because no specific reduction from output to input can be exhibited, the appropriate score is 0.

Assumptions & free parameters 3 free parameters · 6 assumptions · 1 invented entities

The central claim rests on a set of learned compression transforms and hyperparameters (compression dimensions, JEPA loss weight, depth). The most fragile assumptions are that low-rank compression is sufficient, that the level dimension acts as a valid decomposition basis, and that the JEPA target preprocessing (zero-padding or EMA of the future series) yields a meaningful representation. No entirely new physical entity is introduced, but the level dimension is a new latent axis.

free parameters (3)
  • compression dimensions per dataset (t_c, l_c, v_c) = selected from {1,4,8,32}; default (4,8,4)
    Hyperparameters chosen per dataset from a small grid; they directly control the information bottleneck and are central to the reported performance.
  • JEPA loss weight = not explicitly reported
    The auxiliary JEPA loss is added to the final loss 'by default' without specifying its weight; this affects training and the reported results.
  • number of tunnel blocks = 0 to 2
    Model depth is tuned per dataset; not fixed in the paper, and the choice affects the encoder capacity.
assumptions (6)
  • domain assumption Mode-k product with low-rank factors C_T E_T compresses a time series while preserving information needed for forecasting.
    The central compression principle (Section 4.2.1, Eq. 12) assumes the learned low-rank transform is a sufficient bottleneck; the paper supports this empirically via ablations (Table 3) but offers no theoretical guarantee.
  • domain assumption The level dimension, expanded by a learned linear map, serves as a multi-scale or decomposition basis.
    Eq. 3 and Appendix D assume the level dimension can capture multiscale structure; Appendix D shows post-hoc linear decomposition, but this is not established a priori.
  • ad hoc to paper Zero-padding or EMA-splitting the target (Eq. 16) yields a target representation comparable to the input representation for the JEPA loss.
    Required to define the auxiliary loss when t_y < t_x or t_y > t_x; the paper does not justify that padding zeros or averaging chunks preserves the predictive signal.
  • domain assumption Reversible instance normalization does not alter the stationarity of the series.
    Stated in Section 4.2.3 without proof; the paper claims it only handles distribution shifts. This is questionable but not central to the forecasting results.
  • standard math Standard tensor algebra operations (folding, unfolding, mode products) are correctly defined and computed.
    Definition 1 and Section 4.2.1 rely on tensor mode products from [13]; however, the paper uses nonstandard naming (Fold for unfolding and Unfold for folding), which is confusing but the underlying algebra is standard.
  • domain assumption Benchmark data splits and baseline results from TFB are accurate and comparable.
    Section 5.1.3 relies on TFB benchmark numbers; any error or inconsistent evaluation protocol in TFB propagates to the SOTA comparison.
invented entities (1)
  • Level dimension (L)
    purpose: Adds a third tensor axis so multi-scale and decomposition features can be learned inside the representation space without explicit patching or downsampling.
    The paper introduces a new latent dimension not present in the input; it is learned and shown post-hoc (Appendix D) to yield multi-scale decomposition, but there is no independent falsifiable prediction outside the model itself.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TimeCapsule: Solving the Jigsaw Puzzle of Long-Term Time Series Forecasting with Compressed Predictive Representations." pith.science (2026). https://pith.science/paper/OIFCACSF

@misc{pith2026250412721,
  author       = {Pith},
  title        = {Pith review of: TimeCapsule: Solving the Jigsaw Puzzle of Long-Term Time Series Forecasting with Compressed Predictive Representations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OIFCACSF}},
  note         = {Machine review of arXiv:2504.12721}
}
read the original abstract

Recent deep learning models for Long-term Time Series Forecasting (LTSF) often emphasize complex, handcrafted designs, while simpler architectures like linear models or MLPs have often outperformed these intricate solutions. In this paper, we revisit and organize the core ideas behind several key techniques, such as redundancy reduction and multi-scale modeling, which are frequently employed in advanced LTSF models. Our goal is to streamline these ideas for more efficient deep learning utilization. To this end, we introduce TimeCapsule, a model built around the principle of high-dimensional information compression that unifies these techniques in a generalized yet simplified framework. Specifically, we model time series as a 3D tensor, incorporating temporal, variate, and level dimensions, and leverage mode production to capture multi-mode dependencies while achieving dimensionality compression. We propose an internal forecast within the compressed representation domain, supported by the Joint-Embedding Predictive Architecture (JEPA), to monitor the learning of predictive representations. Extensive experiments on challenging benchmarks demonstrate the versatility of our method, showing that TimeCapsule can achieve state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2504.12721 by the authors.

Figure 1
Figure 1. We visualize the ETTh1 dataset as an example to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Categorization of advanced LTSF models into four [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the TimeCapsule model. The original time series is transformed into a 3D representation by adding [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (21 more)
Figure 4
Figure 4. Figure 4: We visualize ETTh1 dataset as a typical example to [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The top figures show the variation trend of JEPA [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visualizations of 512-720 prediction on PEMS04 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Efficiency comparisons in terms of FLOPs (GB) and parameter counts (MB) with the latest advanced models on the [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Study on varying lookback windows. We set the length of window [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Representaion of ETTm2 in the encoder. We select [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 10
Figure 10. Figure 10: Representaion of ETTm2 in the decoder. We select [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Visualization of T (temporal)-transform matrices, i.e., [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Visualization of V (variable)-transform matrices, i.e., [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]
Figure 13
Figure 13. Figure 13: Multil-level property of PEMS04 series in the representation space by applying the learnd level-expansion matrix. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]
Figure 14
Figure 14. Figure 14: Multil-level property of ETTm2 series in the representation space by applying the learnd level-expansion matrix. [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]
Figure 15
Figure 15. Figure 15: Multil-level property of ETTm2 series in the normalized time space by applying the learnd level-expansion matrix. [PITH_FULL_IMAGE:figures/full_fig_p018_15.png]
Figure 16
Figure 16. Figure 16: Multil-level property of PEMS04 series in the normalized time space by applying the learnd level-expansion matrix. [PITH_FULL_IMAGE:figures/full_fig_p018_16.png]
Figure 17
Figure 17. Figure 17: A summation of the sub-level series of the ETTm2 series in the normalized time space demonstrates that it approaches [PITH_FULL_IMAGE:figures/full_fig_p019_17.png]
Figure 18
Figure 18. Figure 18: A summation of the sub-level series of the ETTm2 series in the normalized time space demonstrates that it approaches [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]
Figure 19
Figure 19. Figure 19: We visualize ETTh1 dataset as a typical example to illustrate 512-96 predictions of different components (trend, [PITH_FULL_IMAGE:figures/full_fig_p021_19.png]
Figure 20
Figure 20. Figure 20: We visualize ETTh1 dataset as a typical example to illustrate 512-96 predictions of different components (trend, [PITH_FULL_IMAGE:figures/full_fig_p022_20.png]
Figure 21
Figure 21. Figure 21: Prediction on the PEMS04 dataset, with lookback window 512 and forecast length 96. [PITH_FULL_IMAGE:figures/full_fig_p023_21.png]
Figure 22
Figure 22. Figure 22: Prediction on the Solar dataset, with lookback window 512 and forecast length 96. [PITH_FULL_IMAGE:figures/full_fig_p023_22.png]
Figure 23
Figure 23. Figure 23: Prediction on the Traffic dataset, with lookback window 512 and forecast length 96. [PITH_FULL_IMAGE:figures/full_fig_p024_23.png]
Figure 24
Figure 24. Figure 24: Prediction on the Electricity dataset, with lookback window 512 and forecast length 96. [PITH_FULL_IMAGE:figures/full_fig_p024_24.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 19 canonical work pages

  1. [1]

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. 2023. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 15619–15629

  2. [2]

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. 2024. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471 (2024)

  3. [3]

    George EP Box and David A Pierce. 1970. Distribution of residual autocorrelations in autoregressive-integrated moving average time series models. Journal of the American statistical Association 65, 332 (1970), 1509–1526

  4. [4]

    Elizabeth Bradley. 1999. Time-series analysis. Intelligent data analysis: An introduction (1999), 167–194

  5. [5]

    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, Vol. 37. 6989–6997

  6. [6]

    Si-An Chen, Chun-Liang Li, Nate Yoder, Sercan O Arik, and Tomas Pfister. 2023. Tsmixer: An all-mlp architecture for time series forecasting. arXiv preprint arXiv:2303.06053 (2023)

  7. [7]

    Abhimanyu Das, Weihao Kong, Andrew Leach, Shaan Mathur, Rajat Sen, and Rose Yu. 2023. Long-term forecasting with tide: Time-series dense encoder.arXiv preprint arXiv:2304.08424 (2023)

  8. [8]

    Marco AR Ferreira, Mike West, Herbert KH Lee, and David M Higdon. 2006. Multi-Scale and Hidden Resolution Time Series Models. Bayesian Analysis 1, 4 (2006), 947–968

Show all 47 references
  1. [9]

    Abanoub M Girgis, Alvaro Valcarce, and Mehdi Bennis. 2024. Time-Series JEPA for Predictive Remote Control under Capacity-Limited Networks. arXiv preprint arXiv:2406.04853 (2024)

  2. [10]

    Chenjuan Guo, Bin Yang, Jilin Hu, Christian S Jensen, and Lu Chen. 2020. Context- aware, preference-based vehicle routing. The VLDB Journal 29 (2020), 1149–1170

  3. [11]

    Michael J Kane, Natalie Price, Matthew Scotch, and Peter Rabinowitz. 2014. Comparison of ARIMA and Random Forest time series models for prediction of avian influenza H5N1 outbreaks. BMC bioinformatics 15 (2014), 1–9

  4. [12]

    Zahra Karevan and Johan AK Suykens. 2020. Transductive LSTM for time-series prediction: An application to weather forecasting. Neural Networks 125 (2020), 1–9

  5. [13]

    Eric Kernfeld, Misha Kilmer, and Shuchin Aeron. 2015. Tensor–tensor products with invertible linear transforms. Linear Algebra Appl. 485 (2015), 545–570

  6. [14]

    Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. 2020. Supervised contrastive learning. Advances in neural information processing systems 33 (2020), 18661– 18673

  7. [15]

    Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. 2021. Reversible instance normalization for accurate time-series forecasting against distribution shift. In International Conference on Learning Representations

  8. [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. 95–104

  9. [17]

    Yann LeCun. 2022. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review 62, 1 (2022), 1–62

  10. [18]

    Minhao Liu, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, and Qiang Xu. 2022. Scinet: Time series modeling and forecasting with sample convolution and interaction. Advances in Neural Information Processing Systems 35 (2022), 5816–5828

  11. [19]

    Shizhan Liu, Hang Yu, Cong Liao, Jianguo Li, Weiyao Lin, Alex X Liu, and Schahram Dustdar. 2021. Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. In International conference on learning representations

  12. [20]

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. 2024. itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625 (2024)

  13. [21]

    Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Non-stationary transformers: Exploring the stationarity in time series forecasting. Advances in Neural Information Processing Systems 35 (2022), 9881–9893

  14. [22]

    Zhiding Liu, Mingyue Cheng, Zhi Li, Zhenya Huang, Qi Liu, Yanhu Xie, and Enhong Chen. 2024. Adaptive normalization for non-stationary time series fore- casting: A temporal slice perspective. Advances in Neural Information Processing Systems 36 (2024)

  15. [23]

    I Loshchilov. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017)

  16. [24]

    Gregory P Meyer. 2021. An alternative probabilistic interpretation of the huber loss. In Proceedings of the ieee/cvf conference on computer vision and pattern recognition. 5261–5269

  17. [25]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In The Eleventh International Conference on Learning Representations

  18. [26]

    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)

  19. [27]

    Xiangfei Qiu, Jilin Hu, Lekui Zhou, Xingjian Wu, Junyang Du, Buang Zhang, Chenjuan Guo, Aoying Zhou, Christian S Jensen, Zhenli Sheng, et al. 2024. Tfb: Towards comprehensive and fair benchmarking of time series forecasting meth- ods. arXiv preprint arXiv:2403.20150 (2024)

  20. [28]

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. 2020. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019. Applied soft computing 90 (2020), 106181. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Yihang Lu, Yangyang...

  21. [29]

    Zezhi Shao, Fei Wang, Yongjun Xu, Wei Wei, Chengqing Yu, Zhao Zhang, Di Yao, Tao Sun, Guangyin Jin, Xin Cao, et al. 2024. Exploring progress in multi- variate time series forecasting: Comprehensive benchmarking and heterogeneity analysis. IEEE Transactions on Knowledge and Dat...

  22. [30]

    Hiro Y Toda and Peter CB Phillips. 1994. Vector autoregression and causality: a theoretical overview and simulation study. Econometric reviews 13, 2 (1994), 259–285

  23. [31]

    James Townsend, Tom Bird, and David Barber. 2019. Practical lossless compres- sion with latent variables using bits back coding. arXiv preprint arXiv:1901.04866 (2019)

  24. [32]

    Stijn Verdenius, Andrea Zerio, and Roy LM Wang. 2024. LaT-PFN: A Joint Embedding Predictive Architecture for In-context Time-series Forecasting. arXiv preprint arXiv:2405.10093 (2024)

  25. [33]

    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)

  26. [34]

    Thomas Wang, Adam Roberts, Daniel Hesslow, Teven Le Scao, Hyung Won Chung, Iz Beltagy, Julien Launay, and Colin Raffel. 2022. What language model architecture and pretraining objective works best for zero-shot generalization?. In International Conference on Machine Learning . ...

  27. [35]

    Qingsong Wen, Tian Zhou, Chaoli Zhang, Weiqi Chen, Ziqing Ma, Junchi Yan, and Liang Sun. 2022. Transformers in time series: A survey. arXiv preprint arXiv:2202.07125 (2022)

  28. [36]

    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)

  29. [37]

    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 34 (2021), 22419–22430

  30. [38]

    Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Fran- cis EH Tay, Jiashi Feng, and Shuicheng Yan. 2021. Tokens-to-token vit: Training vision transformers from scratch on imagenet. In Proceedings of the IEEE/CVF international conference on computer vision ...

  31. [39]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting?. In Proceedings of the AAAI conference on artificial intelligence, Vol. 37. 11121–11128

  32. [40]

    Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. In The eleventh international conference on learning representations

  33. [41]

    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 Proceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115

  34. [42]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, Rong Jin, et al. 2022. Film: Frequency improved legendre memory model for long-term time series forecasting. Advances in neural information processing systems 35 (2022), 12677–12690

  35. [43]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International conference on machine learning . PMLR, 27268–27286. TimeCapsule: Long-Term Time Series Forecasti...

  36. [44]

    The Solar dataset [16] captures 10-minute intervals of solar power production from 137 PV plants in 2006

    records weekly patient counts and influenza-like illness (ILI) ratios. The Solar dataset [16] captures 10-minute intervals of solar power production from 137 PV plants in 2006. Traffic [37] dataset records the hourly road occupancy rates from 862 sensors on San Francisco freew...

  37. [45]

    Do the transforms retain and differentiate variable and level information?

  38. [46]

    What do the transform matrices reveal?

  39. [47]

    We select X2∈ R4×8×7 at the end of the L-TransBlock

    What kind of decomposition strategies has TimeCapsule learnt? D.1 Do these transforms retain and distinguish the information of variables and levels ? 0 10 20 30 40 50 0 1 2 3 -0.15 -0.10 -0.05 0.00 0.05 0.10 0.15 0.20 Visualization of Representation in the Middle of the Encod...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.