Pith. sign in

REVIEW 3 major objections 4 minor 37 references

DC-Mamber: A Dual Channel Prediction Model based on Mamba and Linear Transformer for Multivariate Time Series Forecasting

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read DC-Mamber claims that a dual-channel design pairing Mamba and a linear Transformer can capture both local and global temporal patterns at once, and reports average forecast error reductions of 4.2% in MSE and 4.9% in MAE against eleven…

desk verdict New dual-channel architecture is a legitimate incremental contribution, but the empirical comparison is too inconsistent and unfairly benchmarked to support the SOTA claim. read the letter →

arxiv 2507.04381 v1 pith:MIMY5WO4 submitted 2025-07-06 cs.AI

Bing Fan , Shusen Ma , Yun-Bo Zhao , Yu Kang This is my paper
classification cs.AI
keywords multivariatetimeseriesforecastingchannel-independentchannel-mixingMambalinearTransformerdual-channelmodelselectivestatespacemodelsattention
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

DC-Mamber is a proposed forecasting architecture that runs the same input through two parallel feature channels: a channel-independent path based on Mamba that models each variable's local temporal dynamics, and a channel-mixing path based on a linear Transformer that models global dependencies across time steps. The paper's central claim is that this decoupled design lets one model capture both kinds of temporal structure that single-strategy models miss, without giving up linear-time processing. On eight public datasets across traffic, electricity, weather, solar, and energy domains, the authors report average reductions of 4.2% in MSE and 4.9% in MAE against eleven state-of-the-art baselines. If the comparison is fair, the result suggests that pairing each data-processing strategy with the architecture naturally suited to it is a broadly effective design principle for multivariate forecasting.

What carries the argument

The load-bearing object is the dual-channel token embedding plus encoder pairing. T-Embedding keeps the input as $L$ time-step tokens and feeds them to a T-Encoder whose linear attention projects Key and Value matrices to a low dimension $k$ (Linformer-style), reducing self-attention from quadratic to linear cost while modeling global temporal dependencies. V-Embedding transposes the input to $V$ per-variable tokens and feeds them to a V-Encoder built on Bi-Mamba, a bidirectional selective state-space model whose recurrence and convolutional modes capture local within-variable dynamics in linear time. A Feature-fusion layer linearly projects the temporal feature map onto the variable dimension, concatenates the two maps along the feature dimension, and passes them through an MLP, so the two decoupled representations are recombined before the final projection.

What would settle it

Run DC-Mamber and all eleven baselines under one codebase with identical data splits, normalization, look-back window (96), forecast horizons, and hyperparameter search, and re-measure average MSE and MAE on the eight datasets; the central claim fails if the average reductions do not reproduce the reported 4.2% and 4.9% against these controlled baselines.

Watch

Extended reading notes

Core claim

On the paper's own terms, DC-Mamber establishes that channel-independent and channel-mixing tokenizations can be combined in one model with dedicated encoders: a V-Encoder of bidirectional Mamba layers over per-variable tokens, and a T-Encoder of linear attention over per-timestamp tokens. The V-Encoder captures local intra-variable patterns, the T-Encoder captures global cross-timestep dependencies, and a linear feature-fusion layer merges the two representations before projection to the forecast horizon. The authors argue that this arrangement is not incidental: swapping which encoder receives which tokenization degrades performance, and removing either channel degrades performance, so the alignment of strategy to architecture carries the accuracy gains. They further report that the model beats all eleven baselines on average across the eight datasets, with the largest gains on the high-dimensional PEMS traffic datasets.

Load-bearing premise

The paper's evidence for its accuracy claim assumes that the baseline numbers copied from the published iTransformer results were obtained under the same data preprocessing, normalization, input length, and training conditions as DC-Mamber, because only S-Mamba was reproduced with official code in this study.

Editorial extensions

If this is right

  • If the reported 4.2% MSE and 4.9% MAE average reductions hold under controlled comparison, DC-Mamber would be a practical accuracy upgrade over current Transformer and Mamba baselines on the eight benchmark datasets tested.
  • The architecture keeps each channel at linear complexity, so the claimed gains would not come at the cost of the quadratic scaling that limits Transformers on long sequences or many variables.
  • The ablation results imply that both the dual-channel structure and the specific strategy-architecture assignment matter, and that simply mixing both strategies in one architecture without the dedicated pairing is not enough.
  • The larger improvements on the PEMS traffic datasets suggest the fusion of local and global dependencies matters most for high-dimensional, spatially coupled series, guiding where the model is most likely to pay off.

Reading between the lines

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

  • One extension implied by the design is a learned gating mechanism in the Feature-fusion layer, which could weight the temporal and variable channels by dataset characteristics instead of merging them through a fixed MLP.
  • The strategy-architecture pairing (channel-independent with SSM, channel-mixing with attention) could be tested in other hybrid sequence models, such as language or audio modeling, where the same distinction between per-token and global context arises.
  • On datasets with very many variables, the linear-scaling V-Encoder suggests DC-Mamber should degrade more gracefully than channel-mixing Transformers, a scaling comparison the paper does not report.
  • A direct test of whether the alignment principle, rather than raw parameter count, drives the gains would be to run the same dual-channel model with both encoders swapped back to their original roles while controlling total capacity.
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 / 4 minor

Summary. The paper proposes DC-Mamber, a dual-channel multivariate time series forecasting model that combines a channel-independent Mamba-based encoder (V-Encoder) with a channel-mixing linear Transformer encoder (T-Encoder), followed by a feature fusion layer. The authors claim that this design captures both local intra-variable and global temporal dependencies while maintaining linear time complexity. The central empirical claim, stated in the abstract and introduction, is that DC-Mamber achieves superior accuracy over state-of-the-art baselines on eight public datasets, with an average reduction of 4.2% in MSE and 4.9% in MAE. The paper reports experiments comparing against 11 baselines, ablation studies, and hyperparameter sensitivity analyses.

Significance. If the performance claims were properly supported, the architectural idea of pairing a linear Transformer with a bidirectional Mamba in separate channels matched to different tokenization strategies would be a plausible and potentially useful contribution to the multivariate time series forecasting literature. The paper provides a clear description of the architecture, algorithmic pseudocode for the forecasting procedure, and ablation experiments that probe the contribution of each channel and tokenization strategy. However, the central empirical claim is not established: the paper's own Table II contains multiple entries where DC-Mamber is worse than the S-Mamba baseline that it is supposed to outperform, and the comparison protocol for other baselines is not controlled. The claimed average improvements are therefore unsupported by the evidence presented.

major comments (3)
  1. [Table II, ETTm1 block and Electricity/Weather 192 rows] The paper's own results directly contradict the abstract's claim of 'superior accuracy over existing models' and the introduction's claim of 'consistent gains.' In Table II, DC-Mamber reports worse MSE/MAE than S-Mamba on the entire ETTm1 long-horizon block (Avg 0.408/0.415 vs. 0.398/0.407), on Electricity at horizon 192 (0.163/0.259 vs. 0.161/0.258), and on Weather at horizon 192 (0.218/0.260 vs. 0.215/0.255). The text reports improvements only on selected datasets and omits ETTm1, where DC-Mamber is consistently worse, from the enumerated gains. Since the central claim of the paper is empirical superiority, this internal contradiction is load-bearing.
  2. [Section IV-C, baseline comparison protocol] The baseline comparison is not fair. Section IV-C states that all baseline numbers except S-Mamba* were 'directly sourced from the published results in the original iTransformer paper,' while only S-Mamba was reproduced with official code. However, the paper fixes the input look-back length at L=96 for all datasets, whereas iTransformer's published results are typically reported as the best over multiple look-back lengths (e.g., 96, 192, 336, 720). There is no evidence that the copied baselines were evaluated under the same normalization, train/validation/test splits, or evaluation protocol. If the copied baselines used different look-back lengths or tuning procedures, the reported 4.2% MSE and 4.9% MAE average reductions could be artifacts of inconsistent evaluation rather than genuine model improvements. This concern is load-bearing because the entire superiority claim rests on this comparison.
  3. [Table II, '1st Count' row] The '1st Count' row is internally inconsistent. It reports that S-Mamba achieves only 4 first-place MSE results, yet the table contains at least five entries where S-Mamba has strictly lower MSE than DC-Mamber: ETTm1-192 (0.378 vs. 0.388), ETTm1-336 (0.410 vs. 0.423), ETTm1-720 (0.474 vs. 0.490), Electricity-192 (0.161 vs. 0.163), and Weather-192 (0.215 vs. 0.218). If '1st Count' is meant to count the number of horizons where a model has the best result, these entries should be counted for S-Mamba. If '1st Count' has a different definition, it is not stated. This inconsistency further undermines the reliability of the reported results.
minor comments (4)
  1. [Section IV-B] The text states 'implemented in PyTorch 3.10.9'; PyTorch is a deep learning library, while 3.10.9 is a Python version. This should likely read 'Python 3.10.9 with PyTorch.'
  2. [Throughout] The model name is inconsistently written: 'DC-Mamber' in the title and most of the text, but 'DC-Mamba' in the abstract's first sentence. The intended name should be used consistently.
  3. [Table II, PEMS08, horizon 48] The SCINet column entry contains a typographical error: '0.189x0.270' should presumably be '0.189 0.270'.
  4. [Introduction, contribution 3] The claimed 'average reduction of 4.2% in MSE and 4.9% in MAE' is not accompanied by a clear definition of how the average is computed across datasets and horizons; specifying the aggregation would help reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the performance claim rests on external benchmark tables, the architecture equations are explicit compositions of standard components, and the manuscript's self-citations (FMamba, PCDformer, TCLN) are background positioning rather than load-bearing evidence; the copied-baseline protocol and Table II internal inconsistencies are evaluation-fairness and correctness…

full rationale

The paper's central claim is empirical — DC-Mamber achieves lower MSE/MAE than 11 baselines on 8 public datasets — and its evidence is a results table, not a derivation. The architecture (Eqs. 1-11) is an explicit composition of standard building blocks (ReLU linear embeddings, Linformer-style linear attention, Bi-Mamba blocks, concatenation-plus-MLP fusion); no predicted quantity is defined in terms of the quantity it claims to forecast, and no fitted parameter is relabeled as a prediction. The only self-referential elements are citations to the authors' own prior work (PCDformer [6], FMamba [24], TCLN [26]) in the related-work and novelty positioning; these do not carry the empirical claim, which is measured against numbers sourced from the iTransformer paper and from an official-code reproduction of S-Mamba. No uniqueness theorem is imported, and no architectural ansatz is laundered through self-citation: the Bi-Mamba and linear-attention choices are explicitly attributed to external works (S-Mamba [14], Linformer [33]). The manuscript itself flags the strongest concern in Section IV-C: 'other baseline comparison data were directly sourced from the published results in the original iTransformer paper,' while only S-Mamba was reproduced; if those published numbers were produced under different look-back lengths (the paper fixes L=96), normalization, or hyperparameters, the comparison is not apples-to-apples. Table II also contradicts parts of the text (e.g., S-Mamba beats DC-Mamber on the full ETTm1 block and on several individual horizons, and the '1st Count' row undercounts S-Mamba's first-place results). Those are correctness and fairness risks, not circularity: the reported gains are not equivalent by construction to any input, and no result reduces to its own premises. Hence the low circularity score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The empirical claim rests on the chosen hyperparameters and on the fairness of external baseline numbers. No new physical or mathematical entities are introduced.

free parameters (6)
  • d_model (hidden dimension) = 128, 512, or 1024 depending on dataset/horizon
    Tuned per dataset (Tables V-VI); controls representational capacity.
  • d_state (Mamba state dimension) = 8, 32, 128, or 256
    Tuned per dataset/horizon; controls SSM memory and influences accuracy.
  • learning_rate = 1e-4 to 1e-3 depending on dataset/horizon
    Tuned per dataset/horizon; affects convergence and final error.
  • encoder_layers (el) = 2, 3, or 4
    Tuned per dataset; depth of the stacked dual-channel encoder.
  • batch_size (bs) = 16 or 32
    Tuned per dataset/horizon; affects training dynamics.
  • linear_attention_projection k = not specified
    The Linformer-style projection dimension in Eq. (3) is a free parameter that is never reported.
assumptions (4)
  • standard math Zero-order hold discretization of the SSM in Eq. (9) produces the valid discrete recurrence in Eq. (10).
    Standard result from state space model literature, cited via references [13], [34], [35].
  • domain assumption The datasets and train/validation/test splits follow the standard protocols of the iTransformer paper.
    Section IV-A says 'All datasets were split... following standard protocols,' and baseline numbers are taken from the iTransformer paper.
  • ad hoc to paper The learned linear projection from temporal length L to variable count V before fusion (Algorithm 2, line 8) does not discard forecasting-critical temporal information.
    This projection is a design choice with no theoretical or empirical justification in the paper; if it discards information, the claimed gains may not hold.
  • ad hoc to paper The linear-attention projection with k << L in Eq. (3) retains the global dependency information needed for forecasting.
    Linformer-style projection is an approximation; the paper does not analyze its approximation error for time series.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DC-Mamber: A Dual Channel Prediction Model based on Mamba and Linear Transformer for Multivariate Time Series Forecasting." pith.science (2026). https://pith.science/paper/MIMY5WO4

@misc{pith2026250704381,
  author       = {Pith},
  title        = {Pith review of: DC-Mamber: A Dual Channel Prediction Model based on Mamba and Linear Transformer for Multivariate Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MIMY5WO4}},
  note         = {Machine review of arXiv:2507.04381}
}
read the original abstract

In multivariate time series forecasting (MTSF), existing strategies for processing sequences are typically categorized as channel-independent and channel-mixing. The former treats all temporal information of each variable as a token, focusing on capturing local temporal features of individual variables, while the latter constructs a token from the multivariate information at each time step, emphasizing the modeling of global temporal dependencies. Current mainstream models are mostly based on Transformer and the emerging Mamba. Transformers excel at modeling global dependencies through self-attention mechanisms but exhibit limited sensitivity to local temporal patterns and suffer from quadratic computational complexity, restricting their efficiency in long-sequence processing. In contrast, Mamba, based on state space models (SSMs), achieves linear complexity and efficient long-range modeling but struggles to aggregate global contextual information in parallel. To overcome the limitations of both models, we propose DC-Mamber, a dual-channel forecasting model based on Mamba and linear Transformer for time series forecasting. Specifically, the Mamba-based channel employs a channel-independent strategy to extract intra-variable features, while the Transformer-based channel adopts a channel-mixing strategy to model cross-timestep global dependencies. DC-Mamber first maps the raw input into two distinct feature representations via separate embedding layers. These representations are then processed by a variable encoder (built on Mamba) and a temporal encoder (built on linear Transformer), respectively. Finally, a fusion layer integrates the dual-channel features for prediction. Extensive experiments on eight public datasets confirm DC-Mamber's superior accuracy over existing models.

Figures

Figures reproduced from arXiv: 2507.04381 by the authors.

Figure 1
Figure 1. Model structure temporal information into the data, thereby preserving the se￾quential continuity. After applying PE, we project the variable dimension V to a high-dimensional feature space D, generat￾ing the temporal embedding token: Xtem ∈ R B×L×D, where B denotes the batch size dimension. This process preserves the instantaneous correlations among variables within the time series and provides the necessary featur… view at source ↗
Figure 2
Figure 2. Comparison of Prediction Performance between DC-Mamber and S-Mamba [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Sensitivity analysis of hyperparameters on ETTm1, Weather, Solar-Energy, and ECL datasets under a fixed look-back [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 26 canonical work pages

  1. [1]

    Probabilistic Regularized Extreme Learning for Robust Modeling of Traffic Flow Forecasting,

    J. Lou, Y . Jiang, Q. Shen, R. Wang, and Z. Li, “Probabilistic Regularized Extreme Learning for Robust Modeling of Traffic Flow Forecasting,” IEEE Transactions on Neural Networks and Learning Systems , vol. 34, no. 4, pp. 1732–1741, 2023

    work page 2023

  2. [2]

    Probabilistic spatio-temporal graph con- volutional network for traffic forecasting,

    A. A. Karim and N. Nower, “Probabilistic spatio-temporal graph con- volutional network for traffic forecasting,” Applied Intelligence, vol. 54, no. 11, pp. 7070–7085, 2024

    work page 2024

  3. [3]

    Routeformer: Transformer utilizing routing mecha- nism for traffic flow forecasting,

    J. Qi and H. Fan, “Routeformer: Transformer utilizing routing mecha- nism for traffic flow forecasting,” Neurocomputing, p. 129753, 2025

    work page 2025

  4. [4]

    A flexible and lightweight deep learning weather forecasting model,

    G. Zenkner and S. Navarro-Martinez, “A flexible and lightweight deep learning weather forecasting model,” Applied Intelligence , vol. 53, no. 21, pp. 24 991–25 002, 2023

    work page 2023

  5. [5]

    Solar power time series forecasting utilising wavelet coefficients,

    S. Almaghrabi, M. Rana, M. Hamilton, and M. S. Rahaman, “Solar power time series forecasting utilising wavelet coefficients,” Neurocom- puting, vol. 508, pp. 182–207, 2022

    work page 2022

  6. [6]

    Multivariate time series modeling and forecasting with parallelized convolution and decomposed sparse-transformer,

    S. Ma, Y .-B. Zhao, Y . Kang, and P. Bai, “Multivariate time series modeling and forecasting with parallelized convolution and decomposed sparse-transformer,” IEEE Transactions on Artificial Intelligence , 2024

    work page 2024

  7. [7]

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

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  8. [8]

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “iTransformer: Inverted Transformers Are Effective for Time Series Forecasting,” in The Twelfth International Conference on Learning Representations(ICLR), 2024

    work page 2024

Show all 37 references
  1. [9]

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers,” in The Eleventh International Conference on Learning Representations(ICLR) , 2023

    2023

  2. [10]

    Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,

    Y . Zhang and J. Yan, “Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting,” in The Eleventh International Conference on Learning Representations(ICLR) , 2023

    2023

  3. [11]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  4. [12]

    Attention is All you Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is All you Need,” in Advances in Neural Information Processing Systems (NeurIPS) , vol. 30, 2017

    2017

  5. [13]

    Efficiently Modeling Long Sequences with Structured State Spaces,

    A. Gu, K. Goel, and C. R’e, “Efficiently Modeling Long Sequences with Structured State Spaces,” ArXiv, vol. abs/2111.00396, 2021

    2021 arXiv

  6. [14]

    Is mamba effective for time series forecasting?

    Z. Wang, F. Kong, S. Feng, M. Wang, X. Yang, H. Zhao, D. Wang, and Y . Zhang, “Is mamba effective for time series forecasting?” Neurocom- puting, vol. 619, p. 129178, 2025

    2025

  7. [15]

    Timemachine: A time series is worth 4 mambas for long-term forecasting,

    M. A. Ahamed and Q. Cheng, “Timemachine: A time series is worth 4 mambas for long-term forecasting,” in 27th European Conference on Artificial Intelligence, vol. 392, 2024, pp. 1688–1695

    2024

  8. [16]

    Enhancing transformer- based models for long sequence time series forecasting via structured matrix,

    Z. Zhang, Y . Wang, S. Tan, B. Xia, and Y . Luo, “Enhancing transformer- based models for long sequence time series forecasting via structured matrix,” Neurocomputing, vol. 625, p. 129429, 2025

    2025

  9. [17]

    Bert: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of deep bidirectional transformers for language understanding,” in Pro- ceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo...

    2019

  10. [18]

    Deformable DETR: Deformable Transformers for End-to-End Object Detection,

    X. Zhu, W. Su, L. Lu, B. Li, X. Wang, and J. Dai, “Deformable DETR: Deformable Transformers for End-to-End Object Detection,” in The Ninth International Conference on Learning Representations(ICLR), 2021

    2021

  11. [19]

    Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting,

    Y . Liu, H. Wu, J. Wang, and M. Long, “Non-stationary Transformers: Exploring the Stationarity in Time Series Forecasting,” in Advances in Neural Information Processing Systems (NeurIPS) , vol. 35, 2022, pp. 9881–9893

    2022

  12. [20]

    FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting,

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting,” in International Conference on Machine Learning(ICML) , 2022, pp. 27 268–27 286

    2022

  13. [21]

    TimesNet: Tem- poral 2D-Variation Modeling for General Time Series Analysis,

    H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: Tem- poral 2D-Variation Modeling for General Time Series Analysis,” in The Eleventh International Conference on Learning Representations(ICLR) , 2023

    2023

  14. [22]

    Bi-mamba+: Bidirectional mamba for time series forecasting,

    A. Liang, X. Jiang, Y . Sun, X. Shi, and K. Li, “Bi-mamba+: Bidirectional mamba for time series forecasting,” arXiv preprint arXiv:2404.15772 , 2024

    2024 arXiv

  15. [23]

    Mixmamba: Time series modeling with adaptive expertise,

    K. Alkilane, Y . He, and D.-H. Lee, “Mixmamba: Time series modeling with adaptive expertise,” Information Fusion, vol. 112, p. 102589, 2024

    2024

  16. [24]

    Fmamba: Mamba based on fast-attention for multivariate time-series forecasting,

    S. Ma, Y . Kang, P. Bai, and Y .-B. Zhao, “Fmamba: Mamba based on fast-attention for multivariate time-series forecasting,” arXiv preprint arXiv:2407.14814, 2024

    2024 arXiv

  17. [25]

    A distance correlation-based approach to characterize the effectiveness of recurrent neural networks for time series forecasting,

    C. Salazar and A. G. Banerjee, “A distance correlation-based approach to characterize the effectiveness of recurrent neural networks for time series forecasting,” Neurocomputing, p. 129641, 2025

    2025

  18. [26]

    TCLN: A Transformer-based Conv-LSTM network for multivariate time series forecasting,

    S. Ma, T. Zhang, Y .-B. Zhao, Y . Kang, and P. Bai, “TCLN: A Transformer-based Conv-LSTM network for multivariate time series forecasting,” Applied Intelligence , vol. 53, no. 23, pp. 28 401–28 417, 2023

    2023

  19. [27]

    Addressing prediction delays in time series forecasting: A continuous gru approach with derivative regularization,

    S. Y . Jhin, S. Kim, and N. Park, “Addressing prediction delays in time series forecasting: A continuous gru approach with derivative regularization,” in Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2024, pp. 1234–1245

    2024

  20. [28]

    Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks,

    G. Lai, W.-C. Chang, Y . Yang, and H. Liu, “Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks,” in The 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, 2018, pp. 95–104

    2018

  21. [29]

    Going deeper with convolutions,

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. E. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 1–9

    2015

  22. [30]

    Are transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transformers effective for time series forecasting?” in Proceedings of the AAAI Conference on Artificial Intelligence, 2023, pp. 11 121–11 128

    2023

  23. [31]

    Revisiting long-term time se- ries forecasting: An investigation on linear mapping,

    Z. Li, S. Qi, Y . Li, and Z. Xu, “Revisiting long-term time se- ries forecasting: An investigation on linear mapping,” arXiv preprint arXiv:2305.10721, 2023

    2023 arXiv

  24. [32]

    Tsmixer: An all-mlp architecture for time series forecasting,

    S.-A. Chen, C.-L. Li, N. Yoder, S. O. Arik, and T. Pfister, “Tsmixer: An all-mlp architecture for time series forecasting,” arXiv preprint arXiv:2303.06053, 2023

    2023 arXiv

  25. [33]

    Linformer: Self-attention with linear complexity,

    S. Wang, B. Z. Li, M. Khabsa, H. Fang, and H. Ma, “Linformer: Self-attention with linear complexity,” arXiv preprint arXiv:2006.04768, 2020

    2006 arXiv

  26. [34]

    Combining Recurrent, Convolutional, and Continuous-time Models with Linear State-Space Layers,

    A. Gu, I. Johnson, K. Goel, K. K. Saab, T. Dao, A. Rudra, and C. R’e, “Combining Recurrent, Convolutional, and Continuous-time Models with Linear State-Space Layers,” in Advances in Neural Information Processing Systems (NeurIPS) , 2021

    2021

  27. [35]

    Hippo: Recurrent memory with optimal polynomial projections,

    A. Gu, T. Dao, S. Ermon, A. Rudra, and C. R ´e, “Hippo: Recurrent memory with optimal polynomial projections,” Advances in neural information processing systems , vol. 33, pp. 1474–1487, 2020

    2020

  28. [36]

    Long- term forecasting with tide: Time-series dense encoder,

    A. Das, W. Kong, A. Leach, S. Mathur, R. Sen, and R. Yu, “Long- term forecasting with tide: Time-series dense encoder,” arXiv preprint arXiv:2304.08424, 2023. 12

    2023 arXiv

  29. [37]

    SCINet: time series modeling and forecasting with sample convolution and interaction,

    M. Liu, A. Zeng, M. Chen, Z. Xu, Q. Lai, L. Ma, and Q. Xu, “SCINet: time series modeling and forecasting with sample convolution and interaction,” in Advances in Neural Information Processing Systems (NeurIPS), 2022, pp. 5816–5828. A. Appendix 13 TABLE V: The hyperparameters o...

    2022

Pith tools

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