Pith. sign in

REVIEW 3 major objections 5 minor 32 references

CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read CMoS claims that forecasting is best done by directly modeling stable spatial correlations between chunks, and backs the claim with state-of-the-art results at under 1% of DLinear's parameter count.

desk verdict A genuinely light linear forecasting model with a nice chunk-wise correlation-mixing idea, but the SOTA claim is a half-step ahead of the evidence. read the letter →

arxiv 2505.19090 v1 pith:76XAFQ3L submitted 2025-05-25 cs.LG stat.ML

classification cs.LGstat.ML
keywords timeseriesforecastinglightweightmodelspatialcorrelationchunk-wisemodelingmixingperiodicityinjectioninterpretabilitymultivariate
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 argues that long-term multivariate forecasting can be reduced to modeling the spatial correlations between chunks of a time series rather than learning shape embeddings or high-level representations. The proposed model, CMoS, predicts each future chunk as a softmax-weighted linear combination of past chunks using a small set of shared correlation matrices, with per-channel mixing weights derived from a smoothed window. On seven standard benchmarks, the paper reports first place in 9 of 14 metrics and second place in 3 metrics while using as little as 1% of the lightweight DLinear model's parameters. The learned correlation matrices are directly interpretable, revealing temporal dependency structures such as daily or hourly stripes in the data.

What carries the argument

The central object is the chunk-wise spatial correlation matrix, whose entry gives the learned influence of a past chunk on a future chunk. The argument runs through a mixture of K such matrices shared across all channels: per-channel convolution aggregators smooth the raw input, a shared linear weight allocator maps the smoothed representation to softmax mixing weights, and the forecast is formed as the weighted sum of the K correlation-matrix predictions. Periodicity Injection edits the first matrix by placing peaks at the dominant period divided by the chunk size, and Reversible Instance Normalization is applied before and after prediction to counter distribution shift. Together these components keep the parameter count near K times the product of the chunk counts in the lookback and horizon windows.

What would settle it

Synthesize a series whose chunk-to-chunk correlation structure is deliberately changed between training and test windows, for example a regime shift where the lag structure flips, then train CMoS on the first regime and evaluate on the second; if the fixed correlation matrices fail to track the shifted structure while an adaptive baseline succeeds, the central modeling assumption is the limiting factor.

Watch

Extended reading notes

Core claim

The central claim is that time series carry a stable and decomposable chunk-to-chunk spatial correlation structure, and that directly modeling this structure is sufficient for state-of-the-art forecasting. Formally, CMoS splits the lookback window and the forecast horizon into chunks of size S and models each predicted chunk as a linear combination of past chunks, with the coefficients forming a learned chunk-wise spatial correlation matrix. Correlation Mixing replaces a single matrix by K shared basis matrices combined through channel-specific softmax weights, so that channels with different temporal structures share parameters without paying an O($N^{2}$) cost for cross-channel dependencies. The paper proves that chunk-wise weight averaging reduces sensitivity to Gaussian noise, and it introduces Periodicity Injection, which initializes one correlation matrix with peaks at the dominant period to accelerate convergence. Empirically, the model ranks first on 9 of 14 averaged MSE and MAE metrics and second on 3, with its largest gains on datasets containing many heterogeneous channels.

Load-bearing premise

The whole result rests on the assumption that each future chunk is a fixed linear combination of past chunks and that this chunk-to-chunk correlation structure stays the same during the test period.

Editorial extensions

If this is right

  • If chunk-to-chunk correlations are stable, forecasting does not require learned shape embeddings; a linear map in chunk space is sufficient.
  • Correlation Mixing lets a shared set of matrices represent different temporal structures per channel, explaining why CMoS gains most on datasets with many heterogeneous channels.
  • Periodicity Injection speeds convergence and improves accuracy on periodic data, while the paper's Weather results show it should be disabled when the dominant period is not stable.
  • The parameter count formula shows the model stays small even as lookback and horizon grow, because chunking reduces the correlation matrices to the product of the numbers of chunks.
  • The learned matrices are interpretable and can be read directly as dependency structures, such as the previous-day or short-term reliance stripes found on the Weather dataset.

Reading between the lines

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

  • A natural extension is to make the mixing weights depend on recent local statistics rather than only on the smoothed window, which would let the correlation structure drift across regimes while keeping the same parameter budget.
  • Because Theorem 3.2 frames chunking as variance reduction through weight averaging, the model should show its largest relative advantage on the noisiest channels; injecting synthetic Gaussian noise into a clean dataset and measuring per-channel error would test this prediction.
  • The interpretable matrices make CMoS usable as a descriptive tool: on a new dataset, the learned mappings could reveal dominant periods and lag structures without a separate analysis step.
  • The results suggest that long-term forecasting has low effective rank in chunk space, implying that further gains may come from better estimators of the mixing weights rather than from deeper networks.
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 / 5 minor

Summary. The paper proposes CMoS, a super-lightweight multivariate time series forecasting model that directly models chunk-to-chunk spatial correlations. It defines the forecast as a softmax-weighted sum of K shared correlation matrices applied to past chunks, with per-channel weights produced by a convolution and a shared linear layer (Eq. 3). A Periodicity Injection technique initializes one correlation matrix using an ACF-derived period to speed convergence. Experiments on seven long-term forecasting datasets report top-2 performance on most of 14 metrics, with parameter counts around 1% of DLinear, and the paper includes ablations, sensitivity analyses, and interpretability visualizations of the learned matrices.

Significance. If the performance claim holds, the paper offers a striking result: a model with a few hundred parameters can match or beat heavily tuned deep models on multivariate long-term forecasting benchmarks. The interpretability of the learned correlation matrices is a genuine strength, as is the disclosure of code and detailed ablations that test each component (chunking, correlation mixing, periodicity injection). The parameter-efficiency analysis in Sec. 3.6 is transparent and the claimed ratio of roughly 1% of DLinear's parameters is credible for the ETTh1 configuration. However, the paper's central 'outperforms SOTA' claim rests on very small margins that are not yet shown to be statistically robust, and the theoretical support for chunk-wise robustness is not directly connected to the actual model operation.

major comments (3)
  1. [Sec. 3.1, Theorem 3.2] Theorem 3.2 does not analyze the actual chunk-wise operation used in Eq. (3). It considers a weighted average of point-wise regression weights within a single chunk, showing that the squared norm of the averaged weight vector is no larger than the sum of squared point weights. But the model in Eq. (3) applies a full L/S × H/S matrix to a vector of S time points, and the noise sensitivity of such a mapping is σ²·||W||_F², which can be larger or smaller than the point-wise sensitivity depending on the matrix. The theorem therefore does not justify the claim that chunk-wise spatial correlation modeling is more robust to noise. Please provide a direct analysis of the chunk-wise linear map, or explicitly soften the claim to apply only to the specific averaging operation in the theorem.
  2. [Table 1 and Sec. 4.1] The headline superiority claim is not supported by the reported statistics. CMoS results are averaged over 5 seeds with standard deviations of 0.001–0.004, but all baselines are reported as point estimates from TFB or from reimplementations. Several key margins are at or below CMoS's own run-to-run variation: Electricity MSE 0.158 vs. CycleNet 0.158, Weather 0.220 vs. PatchTST 0.224, and ETTh1 0.403 vs. SparseTSF 0.406. A paired multi-seed comparison with baselines run under the same protocol is needed to establish that the ranking is stable. If such experiments are not feasible, the abstract and Sec. 4.2 should be reworded from 'outperforms existing state-of-the-art models' to 'competitive with state-of-the-art models on most benchmarks.'
  3. [Sec. 3.3, Table 1, and Appendix D.3] The use of Periodicity Injection is applied selectively to datasets with human-activity-related cycles, but Table 1 does not state which datasets used the injected variant versus the base model. Since the ablation in Table 3 shows PI hurts Weather (0.148 with PI vs. 0.144 without), the reader cannot tell whether the reported comparison is consistent. Please clearly mark in the main table or experimental setup which variant was used for each dataset, and justify why the selective application is fair relative to baselines that do not use this prior.
minor comments (5)
  1. [Table 3 caption] The word 'Periodictiy' is a typo for 'Periodicity'.
  2. [Sec. 1 and Sec. 3.2] The phrase 'we pioneer leveraging the stability and decomposability of spatial correlations' is an overclaim; previous works have used correlation-based or mixture-of-experts ideas. Suggest softening to 'we introduce'.
  3. [Eq. (3) and Sec. 3.2] The dimensions of the bias term b^k_i are not defined explicitly. It appears to be a scalar added to each output chunk, but the text should state that b^k is a vector of length H/S shared across channels for each matrix k.
  4. [Algorithm 1, Appendix E.2] The indexing in the pseudocode for Periodicity Injection is hard to follow; the condition 'i+j < L/S' and the assignment 'θ_{i, j+i} = p/L' should be explained with a small example to avoid ambiguity about which chunk-to-chunk positions are being set.
  5. [Appendix G] The paragraph 'Spatial Correlation Modeling' honestly acknowledges that the approach may not work for random-walk-like irregular series. This is useful context, but it also implies the method's applicability is narrower than the abstract suggests; a sentence in the conclusion about this limitation would improve balance.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: CMoS's forecasts are trained and evaluated on held-out data; no predicted quantity is an input by construction.

full rationale

The paper's derivation chain is self-contained in the relevant sense. Eq. (2) defines the forecast as a linear combination of past chunks, Eq. (3) adds a softmax-weighted mixture of shared correlation matrices, and the model is trained with MSE on standard train/validation/test splits. The reported Table 1 results are ordinary empirical evaluations against held-out future values, not quantities derived from the model's own assumptions. The only data-derived hyperparameter is the period obtained by AutoCorrelation Function for Periodicity Injection, but this is a disclosed initialization heuristic applied before training; it does not define the evaluation target, and the ablation in Table 3 shows that the performance gain from periodicity injection is measured, not assumed. Theorem 3.2 is a simple inequality showing that averaging point weights reduces Gaussian noise sensitivity; it is not used to derive the empirical results and is not circular. The paper invokes no load-bearing self-citation: baselines are taken from TFB or reimplemented from public source code, and no uniqueness theorem or prior result by the same authors is used to force the model choice. Concerns that several reported wins are within a few thousandths of baseline MSE are evaluation-robustness concerns, not circularity. Accordingly, no circular step can be exhibited with a quote and a specific reduction, and the correct score is 0.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The model's core is a linear map with shared correlation matrices. The only new conceptual objects are the chunking operation and the mixing mechanism, which are parameterizations, not new entities. The model's performance depends on several hyperparameters that are selected per dataset, and on the ACF-derived period for the optional injection mechanism.

free parameters (6)
  • chunk size S = per dataset, searched over {2,4,8,24}
    Controls the granularity of chunk-wise correlation; the paper notes that a chunk size that is not a common divisor of the period hurts performance, so it must be tuned.
  • number of correlation matrices K = per dataset, searched over {2,4,8}
    Number of basis correlation matrices in the mixture; directly controls model capacity and performance.
  • convolution kernel size c = not reported per dataset (e.g., 8 in the ETTh1 example)
    Smoothing window for the per-channel aggregator; must be chosen alongside the lookback length.
  • lookback window L = searched over {96,336,720}
    Standard lookback choice; evaluated by grid search.
  • learning rate = searched over {2e-5, 5e-5, 8e-5, 8e-4}
    Training hyperparameter.
  • dominant period p from ACF = dataset-specific, e.g., 168 for Electricity and Traffic, 24 for ETTh, 96 for ETTm, 144 for Weather
    Used to initialize the first correlation matrix in Periodicity Injection; estimated from the target dataset.
assumptions (5)
  • domain assumption Future chunks are linear combinations of past chunk vectors (Eq. 2)
    The entire model is a linear map between chunks; no nonlinearity is used.
  • domain assumption Chunk-to-chunk spatial correlations are stable/translation-equivariant over time (Sec. 1, Fig. 1)
    The paper motivates the model by claiming the correlations remain similar as windows slide; if false, a fixed weight matrix cannot predict.
  • domain assumption A small set of K shared correlation matrices can represent the diverse channel-specific correlations (Sec. 3.2)
    Correlation Mixing assumes decomposability of correlations; the number of channels N can be much larger than K.
  • domain assumption The ACF-dominant period is a reliable summary of the series' periodic structure (Sec. 3.3)
    Periodicity Injection places peaks at p/S intervals based on ACF; a wrong or unstable period degrades performance, as acknowledged for Weather.
  • domain assumption Instance Normalization statistics mean/variance of the input window are appropriate for the output window (Sec. 3.4)
    RevIN assumes the input window's mean and variance can be transferred back to the prediction; this is a cited technique but still an assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations." pith.science (2026). https://pith.science/paper/76XAFQ3L

@misc{pith2026250519090,
  author       = {Pith},
  title        = {Pith review of: CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/76XAFQ3L}},
  note         = {Machine review of arXiv:2505.19090}
}
read the original abstract

Recent advances in lightweight time series forecasting models suggest the inherent simplicity of time series forecasting tasks. In this paper, we present CMoS, a super-lightweight time series forecasting model. Instead of learning the embedding of the shapes, CMoS directly models the spatial correlations between different time series chunks. Additionally, we introduce a Correlation Mixing technique that enables the model to capture diverse spatial correlations with minimal parameters, and an optional Periodicity Injection technique to ensure faster convergence. Despite utilizing as low as 1% of the lightweight model DLinear's parameters count, experimental results demonstrate that CMoS outperforms existing state-of-the-art models across multiple datasets. Furthermore, the learned weights of CMoS exhibit great interpretability, providing practitioners with valuable insights into temporal structures within specific application scenarios.

Figures

Figures reproduced from arXiv: 2505.19090 by the authors.

Figure 1
Figure 1. , for the subsequences of the sliding window, while their shapes vary over time, as long as their relative po￾sitions are consistent, their dependencies, referred to as chunk-to-chunk spatial correlations, often maintain stable and reveal the temporal regularity of the system. This phe￾nomenon carries both analytical and practical implications: from an analytical perspective, it reveals that there’s an in￾herent pro… view at source ↗
Figure 2
Figure 2. The spatial correlations of multiple time series can be represented by the combination of fewer sub-correlations. In this paper, we pioneer leveraging the stability and decom￾posability of spatial correlations to build a super-lightweight forecasting model for multivariate time series. Specifically, we propose CMoS, a Chunk-wise Mixture of Spatial cor￾relations architecture to predict multivariate time series. The a… view at source ↗
Figure 3
Figure 3. CMoS Architecture. how specific segment influence the final prediction, limiting the interpretability of these methods. In contrast, the proposed chunking technique splits both historical and future series, and instead of learning the high￾level representations, chunk-based CMoS focus on directly modeling of the spatial correlation between historical and future segments. which is quite interpretable. Our further exp… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: For time series data with significant periodicity, since the initialized parameters already incorporate most periodic correlations, the model’s learning burden is reduced. By providing a strong inductive bias towards periodicity, this op￾eration not only accelerates co…
Figure 5
Figure 5. Figure 5: Test loss during training process on different datasets with horizon= 96. PI is the abbreviation of Periodicity Injection. The model with Periodicity Injection exhibits faster loss reduction. nomena and is influenced by more random and complex factors. As a result, the…
Figure 6
Figure 6. Figure 6: Comparison of the prediction performance and parameter count between CMoS and other baselines on Electricity dataset with horizon= 192. A key advantage of CMoS lies in its ability to achieve su￾perior predictive performance with remarkably few param￾eters. As demonstra…
Figure 7
Figure 7. Figure 7: Visualization of the learned spatial correlation matrices (mappings) on the Weather dataset. The chunk size is set to 4. The visualization analysis reveals that each spatial correla￾tion mapping captures and emphasizes distinct patterns of temporal dependencies, which …
Figure 8
Figure 8. Figure 8: Raw time series and their corresponding averaged map￾ping allocation proportion. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: The logical similarity between CMOS circuits and CMoS model. B. More Information about Channel Strategy B.1. Classification of Channel Strategies We summarize the channel strategies that existing methods applied as follows (and shown in [PITH_FULL_IMAGE:figures/full_f…
Figure 10
Figure 10. Figure 10: Illustration of different channel strategies [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: b demonstrates the prediction results with different numbers of spatial correlation matrices. For the Electricity dataset which contains over 100 channels, when the number is small (1 and 2), the model fails to capture sufficient temporal dependencies, leading to subo…
Figure 12
Figure 12. Figure 12: Visualizations of all channels within a long time window in the Weather dataset. 1 2 3 4 Map Index 0.00 0.25 0.50 0.75 1.00 Proportion Channel 1 1 2 3 4 Map Index 0.00 0.25 0.50 0.75 1.00 Proportion Channel 2 1 2 3 4 Map Index 0.00 0.25 0.50 0.75 1.00 Proportion Chann…
Figure 13
Figure 13. Figure 13: Visualizations of mapping allocation situation on all channels. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 19 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    Fundamental limitations of foundational forecasting models: The need for multimodality and rigorous evaluation

    Bergmeir, C. Fundamental limitations of foundational forecasting models: The need for multimodality and rigorous evaluation. Technical report, Department of Computer Science and Artificial Intelligence, University of Granada, Vancouver, Canada, 2024

  3. [3]

    Box, G. E. and Pierce, D. A. Distribution of residual autocorrelations in autoregressive-integrated moving average time series models. Journal of the American statistical Association, 65 0 (332): 0 1509--1526, 1970

  4. [4]

    O., Yoder, N

    Chen, S.-A., Li, C.-L., Arik, S. O., Yoder, N. C., and Pfister, T. TSM ixer: An all- MLP architecture for time series forecast-ing. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id=wbpxTuXgm0

  5. [5]

    K., Sen, R., and Yu, R

    Das, A., Kong, W., Leach, A., Mathur, S. K., Sen, R., and Yu, R. Long-term forecasting with ti DE : Time-series dense encoder. Transactions on Machine Learning Research, 2023. ISSN 2835-8856. URL https://openreview.net/forum?id=pCbC3aQB5W

  6. [6]

    Moment: A family of open time-series foundation models

    Goswami, M., Szafer, K., Choudhry, A., Cai, Y., Li, S., and Dubrawski, A. Moment: A family of open time-series foundation models. In International Conference on Machine Learning, 2024

  7. [7]

    SOFTS : Efficient multivariate time series forecasting with series-core fusion

    Han, L., Chen, X.-Y., Ye, H.-J., and Zhan, D.-C. SOFTS : Efficient multivariate time series forecasting with series-core fusion. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=89AUi5L1uA

  8. [8]

    Temporal convolutional neural (tcn) network for an effective weather forecasting using time-series data from the local weather station

    Hewage, P., Behera, A., Trovati, M., Pereira, E., Ghahremani, M., Palmieri, F., and Liu, Y. Temporal convolutional neural (tcn) network for an effective weather forecasting using time-series data from the local weather station. Soft Computing, 24: 0 16453--16482, 2020

Show all 32 references
  1. [9]

    A., Jordan, M

    Jacobs, R. A., Jordan, M. I., Nowlan, S. J., and Hinton, G. E. Adaptive mixtures of local experts. Neural Computation, 3 0 (1): 0 79--87, 1991. doi:10.1162/neco.1991.3.1.79

  2. [10]

    Reversible instance normalization for accurate time-series forecasting against distribution shift

    Kim, T., Kim, J., Tae, Y., Park, C., Choi, J.-H., and Choo, J. Reversible instance normalization for accurate time-series forecasting against distribution shift. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=cGDAkQo1C0p

  3. [11]

    Segrnn: Segment recurrent neural network for long-term time series forecasting

    Lin, S., Lin, W., Wu, W., Zhao, F., Mo, R., and Zhang, H. Segrnn: Segment recurrent neural network for long-term time series forecasting. arXiv preprint arXiv:2308.11200, 2023

  4. [12]

    Cyclenet: Enhancing time series forecasting through modeling periodic patterns

    Lin, S., Lin, W., Hu, X., Wu, W., Mo, R., and Zhong, H. Cyclenet: Enhancing time series forecasting through modeling periodic patterns. In Thirty-eighth Conference on Neural Information Processing Systems, 2024 a

  5. [13]

    Sparsetsf: modeling long-term time series forecasting with 1k parameters

    Lin, S., Lin, W., Wu, W., Chen, H., and Yang, J. Sparsetsf: modeling long-term time series forecasting with 1k parameters. In Proceedings of the 41st International Conference on Machine Learning, ICML'24. JMLR.org, 2024 b

  6. [14]

    SCIN et: Time series modeling and forecasting with sample convolution and interaction

    LIU, M., Zeng, A., Chen, M., Xu, Z., LAI, Q., Ma, L., and Xu, Q. SCIN et: Time series modeling and forecasting with sample convolution and interaction. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems, 2022. URL htt...

  7. [15]

    itransformer: Inverted transformers are effective for time series forecasting

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., and Long, M. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024 a . URL https://openreview.net/forum?id=JePfAI8fah

  8. [16]

    Timer: Generative pre-trained transformers are large time series models

    Liu, Y., Zhang, H., Li, C., Huang, X., Wang, J., and Long, M. Timer: Generative pre-trained transformers are large time series models. In Forty-first International Conference on Machine Learning, 2024 b

  9. [17]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

  10. [18]

    Time series analysis

    Madsen, H. Time series analysis. Chapman and Hall/CRC, 2007

  11. [19]

    Nguyen, N., Sinthong, P., and Kalagnanam, J

    Nie, Y., H. Nguyen, N., Sinthong, P., and Kalagnanam, J. A time series is worth 64 words: Long-term forecasting with transformers. In International Conference on Learning Representations, 2023

  12. [20]

    N., Carpov, D., Chapados, N., and Bengio, Y

    Oreshkin, B. N., Carpov, D., Chapados, N., and Bengio, Y. N-beats: Neural basis expansion analysis for interpretable time series forecasting. In International Conference on Learning Representations, 2020. URL https://openreview.net/forum?id=r1ecqn4YwB

  13. [21]

    PyTorch: an imperative style, high-performance deep learning library

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., K\" o pf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., and Chintala, S. PyTorch: an im...

  14. [22]

    S., Sheng, Z., and Yang, B

    Qiu, X., Hu, J., Zhou, L., Wu, X., Du, J., Zhang, B., Guo, C., Zhou, A., Jensen, C. S., Sheng, Z., and Yang, B. Tfb: Towards comprehensive and fair benchmarking of time series forecasting methods. Proc. VLDB Endow. , 17 0 (9): 0 2363--2377, 2024

  15. [23]

    Y., and ZHOU, J

    Wang, S., Wu, H., Shi, X., Hu, T., Luo, H., Ma, L., Zhang, J. Y., and ZHOU, J. Timemixer: Decomposable multiscale mixing for time series forecasting. In International Conference on Learning Representations (ICLR), 2024

  16. [24]

    Unified training of universal time series forecasting transformers

    Woo, G., Liu, C., Kumar, A., Xiong, C., Savarese, S., and Sahoo, D. Unified training of universal time series forecasting transformers. In Forty-first International Conference on Machine Learning, 2024

  17. [25]

    Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting

    Wu, H., Xu, J., Wang, J., and Long, M. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in neural information processing systems, 34: 0 22419--22430, 2021

  18. [26]

    Timesnet: Temporal 2d-variation modeling for general time series analysis

    Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., and Long, M. Timesnet: Temporal 2d-variation modeling for general time series analysis. In International Conference on Learning Representations, 2023

  19. [27]

    FITS : Modeling time series with \ 10k\ parameters

    Xu, Z., Zeng, A., and Xu, Q. FITS : Modeling time series with \ 10k\ parameters. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=bWcnvZ3qMb

  20. [28]

    Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 11121--11128, 2023

    Zeng, A., Chen, M., Zhang, L., and Xu, Q. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 11121--11128, 2023

  21. [29]

    Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures

    Zhang, T., Zhang, Y., Cao, W., Bian, J., Yi, X., Zheng, S., and Li, J. Less is more: Fast multivariate time series forecasting with light sampling-oriented mlp structures. arXiv preprint arXiv:2207.01186, 2022

  22. [30]

    and Yan, J

    Zhang, Y. and Yan, J. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=vSVLM2j9eie

  23. [31]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., and Zhang, W. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 11106--11115, 2021

  24. [32]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting

    Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., and Jin, R. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International conference on machine learning, pp.\ 27268--27286. PMLR, 2022

Pith tools

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