Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Wavelet-based Disentangled Adaptive Normalization for Non-stationary Times Series Forecasting

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Wavelet split lifts forecast accuracy across four backbone models by normalizing trends and fluctuations separately.

desk verdict WDAN is a solid, useful normalization wrapper with real gains and a few overstated claims; worth refereeing with requests for error bars and a boundary artifact check. read the letter →

arxiv 2506.05857 v1 pith:DVCE2DDP submitted 2025-06-06 cs.LG

Junpeng Lin , Tian Lan , Bo Zhang , Ke Lin , Dandan Miao , Huiru He , Jiantao Ye , Chen Zhang
show 1 more author
Yan-fu Li
This is my paper
classification cs.LG MSC 62M1042C4068T07
keywords timeseriesforecastingnon-stationaritywavelettransforminstancenormalizationadaptivetrenddecompositiondenormalizationdeeplearning
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

WDAN is a model-agnostic normalization wrapper for multivariate time series forecasting. It claims that existing instance normalization underperforms on non-stationary data because it treats the whole series as one block, blending slow-moving trends with high-frequency fluctuations. WDAN uses a discrete wavelet transform to separate each input into a low-frequency trend and a high-frequency residual, then normalizes with the trend as a time-varying mean and a sliding-window variance from the residual. A lightweight MLP predicts future normalization statistics using first-order differencing of the trend, and the backbone output is denormalized with those predicted statistics. The paper reports consistent MSE and MAE improvements across four forecasting backbones on several benchmarks, with the largest gains on the most non-stationary datasets and long horizon.

What carries the argument

The central object is the wavelet-based disentangled normalization step paired with a statistics prediction module. A level-$K$ discrete wavelet transform (DWT) with basis functions $\phi_{l,h}$ decomposes each input $\mathbf{x}_i$ into a low-frequency trend $\mathbf{x}_l$ and a high-frequency residual $\mathbf{x}_h$ via inverse DWT; the trend supplies the time-varying mean $\mu_x[t] = x_l[t]$, and a sliding window of size $2w+1$ over $x_h$ supplies the point-level variance $\sigma_x^2[t]$. A first-order difference of the centered trend $\Delta\tilde{\mu}_x$ is concatenated with low- and high-frequency embeddings and passed through MLPs to predict the future $\hat{\mu}_y$ and $\hat{\sigma}_y$ used for denormalization. A three-stage training procedure pretrains the statistics predictor, trains the backbone on the normalized series, then jointly fine-tunes both modules.

What would settle it

Run WDAN on a synthetic series with a trend that the fixed coiflet3 wavelet at the chosen level cannot represent compactly, such as a sharp ramp or step trend, and check whether the high-frequency residual $\mathbf{x}_h$ still has a non-stationary mean; if the residual's ADF statistic remains close to that of the original series, the normalization fails to remove the shift and WDAN should not beat RevIN.

Watch

Extended reading notes

Core claim

The central claim is that disentangling the sources of non-stationarity before instance normalization yields reliable, model-agnostic gains. Concretely, WDAN replaces the single global mean and variance used by RevIN with point-level statistics computed from DWT components: the low-frequency approximation $\mathbf{x}_l$ acts as a time-varying mean $\mu_x[t] = x_l[t]$, and a sliding-window standard deviation over the high-frequency residual $\mathbf{x}_h$ acts as $\sigma_x[t]$. The normalized series is fed to a backbone, and a statistics prediction module predicts future $\hat{\mu}_y$ and $\hat{\sigma}_y$ from low- and high-frequency features, with the trend first-differenced to stabilize it. The backbone output is then denormalized using these predicted statistics. The paper demonstrates on Exchange, ETT, Weather, and Electricity that WDAN improves iTransformer, PatchTST, Crossformer, and FEDformer in most configurations, and outperforms existing normalization frameworks such as SAN and DDN on the strongly non-stationary datasets.

Load-bearing premise

The method assumes the DWT low-frequency component is the true time-varying mean and the high-frequency residual carries the variance, so that normalizing with these quantities makes the input stationary; no automatic rule for choosing the wavelet basis or decomposition level is provided.

Editorial extensions

If this is right

  • On long horizons such as $H=720$, WDAN's gains are consistently larger than on short horizons, making it especially valuable for long-term forecasting of slowly drifting series.
  • WDAN can replace RevIN inside models that already include it (iTransformer and PatchTST) and still improve accuracy, indicating it is a stronger instance-normalization treatment rather than just an additive fix.
  • Because WDAN is model-agnostic, the same normalization layer can be dropped into other backbones beyond the four tested without changing their architecture.
  • On datasets with weak non-stationarity, WDAN performs comparably to much simpler variants, so the added complexity only pays off when there is a real trend to disentangle.
  • The three-stage training strategy gives a stronger lower bound on performance than two-stage or single-stage alternatives, meaning the decoupling of statistics prediction from backbone learning is itself load-bearing.

Reading between the lines

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

  • A direct extension the paper leaves implicit is adaptive wavelet basis selection; since the default coiflet3 wavelet and level $K$ are fixed, learning the basis or choosing $K$ per series could extend the gains to trends the fixed basis represents poorly.
  • The mechanism suggests a testable prediction: on purely stationary or noise-like series with no distinguishable trend, WDAN's predicted statistics should collapse to constants and the method should degrade to ordinary instance normalization.
  • The paper only tests channel-independent backbones, so how wavelet disentanglement interacts with cross-variable attention remains open; combining WDAN with a multivariate-attention model could reveal whether the benefit is independent of variable-level dependencies.
  • The ablation replacing DWT with a moving average hints that the frequency localization of wavelets, not just any decomposition, is what helps; a direct comparison against other multiscale decompositions such as empirical mode decomposition would clarify this.
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

4 major / 4 minor

Summary. The paper proposes WDAN (Wavelet-based Disentangled Adaptive Normalization), a model-agnostic normalization wrapper for multivariate time series forecasting. The method applies a discrete wavelet transform to each input channel, reconstructs a low-frequency trend component and a high-frequency residual component, uses the trend as a pointwise mean and a sliding-window statistic of the residual as a pointwise variance, normalizes the input, predicts future normalization statistics with MLPs (including first-order differenced trend features), and denormalizes the backbone output. The authors evaluate WDAN with four backbone models (iTransformer, PatchTST, Crossformer, FEDformer) on seven benchmark datasets, compare it with SAN and DDN, and provide ablations of the wavelet decomposition and the differencing feature, plus a training-strategy ablation.

Significance. If the central claim were fully supported, WDAN would be a useful and practical contribution: it is a simple, model-agnostic normalization layer that could be dropped into existing forecasting models, it is released with code, and the three-stage training strategy is a reasonable engineering choice. The experiments cover several backbones and datasets, and the ablation of the differencing and DWT components is informative. However, the evidence is weaker than the abstract's 'consistently improves' and 'significantly improves' wording: Table 3 contains non-positive improvements for two iTransformer combinations, the horizon-level tables show cases where WDAN is worse than a normalization baseline, no error bars or significance tests are reported despite three runs, and the DWT-based decomposition that underpins the method is not validated for boundary effects or for the choice of wavelet and decomposition level. These are fixable issues rather than fatal flaws, but they need to be addressed before the paper's main claim can be accepted at face value.

major comments (4)
  1. [Methodology, Eqs. (1)-(6)] The DWT reconstruction that defines x_l and x_h in Eq. (3) is applied to finite-length windows (T=720), but the paper does not specify the wavelet padding mode, the number of boundary coefficients discarded, or any check that the reconstructed components are free of edge artifacts. This matters because μ_x[t] is set directly to x_l[t] and σ_x[t] is computed from a sliding window over x_h (Eqs. (4)-(5)); artifacts at the right edge of the input window contaminate exactly the normalization statistics used at the forecasting boundary and the ground-truth statistics used to train the statistics-prediction module. Please state the DWT mode (e.g., periodization, symmetric, or replication), report how many coefficients are affected, and provide a numerical or visual check that x_h near t=T is not distorted by boundary effects.
  2. [Main Results and Table 3] The abstract and the Main Results section claim that WDAN 'consistently improves forecasting accuracy' and 'significantly improves the predictive performance of all backbone models in most cases,' but Table 3 itself shows non-positive improvements: iTransformer+WDAN has IMP -0.12% on ETTm2 and IMP -0.02% on Electricity. The horizon-level tables also contain cases where WDAN is worse than a normalization baseline; for example, Table 6 shows FEDformer+WDAN on ETTm2 at H=720 with MSE 0.395 versus SAN's 0.386. Since the paper reports only three-run averages without standard deviations or significance tests, the word 'significantly' is not supported. Please report variance information and either add significance tests or temper the claim to 'improves accuracy in the majority of evaluated settings.'
  3. [Hyperparameter Sensitivity and Experiment Setting] The method requires choosing a wavelet basis and a decomposition level K, but no data-driven selection criterion is provided; the experiment setting fixes 'coiflet 3' as the default, and Fig. 4 shows that on Exchange the results are sensitive to K and to the MLP hidden dimension and number of layers. Since the paper's load-bearing premise is that DWT cleanly separates a non-stationary trend from a stationary residual, the sensitivity of the performance to K on the most non-stationary dataset is a concern. Please provide a principled rule for choosing K and the wavelet, or demonstrate robustness to these choices across all datasets. The conclusion's mention of 'adaptive wavelet basis selection' as future work confirms that this issue is not yet addressed.
  4. [Ablation Study, Table 4] The ablation in Table 4 covers only two backbones (iTransformer and Crossformer) and three datasets (Exchange, ETTh2, Weather), and it does not test the full pipeline on the datasets where WDAN shows the smallest gains (ETTm2, Electricity). In particular, the 'No Differencing' variant is not evaluated on those datasets, so the contribution of the differencing module is only partially established. Please extend the ablation to include at least the datasets and backbones where the main result is weakest, or explicitly state the computational constraints that prevent this.
minor comments (4)
  1. [Methodology, Eq. (15)] The text contains several typos that should be corrected: 'We proposes' at the start of the Methodology section, 'statstics' in Eq. (15), 'lf cif' after Eq. (16), and 'metrcis' in the Table 4 caption.
  2. [Table 3 caption] The caption of Table 3 says 'IMP indicates the percentage improvement in MSE of WDAN,' but the baseline for the improvement is not specified. Please clarify whether IMP is computed against the best of SAN and DDN, against a specific baseline, or against the average of the two.
  3. [Related Work] The related work cites Dish-TS and DAIN as normalization methods, but neither is included in the experimental comparison. A sentence explaining why these methods are omitted (e.g., lack of code or incompatibility with the tested backbones) would strengthen the comparison.
  4. [Fig. 2] Figure 2 is referenced in the Methodology section but does not appear in the submitted text; if this is a rendering issue, please ensure the figure is included in the final version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: WDAN's normalization and statistics-prediction pipeline is self-contained and evaluated on held-out horizons.

full rationale

The paper's derivation chain is not circular. Eqs. (1)-(6) define wavelet-based normalization statistics directly from the input series: the low-frequency reconstruction x_l is used as the pointwise mean and a sliding-window variance is computed from the high-frequency residual x_h. Eqs. (7)-(14) define a learned statistics-prediction module that maps the input-derived statistics series to output-domain statistics; Eqs. (15)-(17) train this module with ground-truth statistics from the training split in the first stage, then train and fine-tune the backbone jointly. Although the statistics predictor is supervised on target statistics, at inference time the denormalization statistics are produced from the input series only and the final forecast is evaluated on held-out test horizons, so the prediction does not reduce to the target by construction. No load-bearing premise is imported by self-citation: the comparisons are against SAN, DDN, RevIN and other methods from other groups, and the wavelet decomposition is justified by standard references plus an illustrative example, not by an author-uniqueness theorem. The paper's central claim of consistent improvement is empirically questionable in isolated cells (e.g., PatchTST+WDAN is worse than PatchTST on ETTm2 at H=720), but that is a correctness or robustness concern, not a circularity concern.

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

The central claim depends on the wavelet decomposition assumptions and on a learned statistics predictor; the listed free parameters are all chosen by grid search and not fully reported, so the method is not parameter-free.

free parameters (5)
  • DWT decomposition level K = not reported in main text
    Controls the separation between trend and residual; Fig. 4 shows strong sensitivity on Exchange.
  • sliding window size w = not reported
    Defines the local window for the variance estimate in Eq (5).
  • hidden dimension D = not reported
    Dimension of the single-layer MLP feature maps in the statistics prediction module.
  • number of MLP layers for MLP_mu and MLP_sigma = hyperparameter
    Depth of the output projection MLPs; treated as a hyperparameter in the appendix.
  • wavelet basis = coiflet 3
    Chosen as default without justification; different bases change the decomposition.
assumptions (5)
  • standard math DWT provides perfect reconstruction: x = x_l + x_h
    Used in Eqs (1)-(3) to define the additive decomposition.
  • domain assumption The low-frequency trend x_l approximates the time-varying mean of the series
    Central modeling choice in Eq (4); motivated only by Fig. 1.
  • domain assumption The high-frequency residual x_h is approximately zero-mean and its local sliding-window variance is the series scale
    Used to define sigma_x in Eq (5).
  • domain assumption First-order differencing of the trend statistics produces more stationary features that aid prediction
    Integrated into Eq (10); motivated by Fig. 1(d).
  • domain assumption MLP-predicted statistics match the true future statistics at test time
    Required for denormalization in Eq (14) to be valid; trained on ground-truth statistics in stage 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Wavelet-based Disentangled Adaptive Normalization for Non-stationary Times Series Forecasting." pith.science (2026). https://pith.science/paper/DVCE2DDP

@misc{pith2026250605857,
  author       = {Pith},
  title        = {Pith review of: Wavelet-based Disentangled Adaptive Normalization for Non-stationary Times Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVCE2DDP}},
  note         = {Machine review of arXiv:2506.05857}
}
read the original abstract

Forecasting non-stationary time series is a challenging task because their statistical properties often change over time, making it hard for deep models to generalize well. Instance-level normalization techniques can help address shifts in temporal distribution. However, most existing methods overlook the multi-component nature of time series, where different components exhibit distinct non-stationary behaviors. In this paper, we propose Wavelet-based Disentangled Adaptive Normalization (WDAN), a model-agnostic framework designed to address non-stationarity in time series forecasting. WDAN uses discrete wavelet transforms to break down the input into low-frequency trends and high-frequency fluctuations. It then applies tailored normalization strategies to each part. For trend components that exhibit strong non-stationarity, we apply first-order differencing to extract stable features used for predicting normalization parameters. Extensive experiments on multiple benchmarks demonstrate that WDAN consistently improves forecasting accuracy across various backbone model. Code is available at this repository: https://github.com/MonBG/WDAN.

Figures

Figures reproduced from arXiv: 2506.05857 by the authors.

Figure 1
Figure 1. An illustration of a forecast sample indicating the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The illustration of the proposed WDAN framework [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The long-term forecasting MSE evaluations of different backbone models under various input lengths. Large values [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity with respect to the number of MLP layers of statistics prediction module, the hidden [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 24 canonical work pages

  1. [1]

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

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

  2. [2]

    write newline

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

  3. [3]

    Ahmadi, M.; Jafarzadeh-Ghoushchi, S.; Taghizadeh, R.; and Sharifi, A. 2019. Presentation of a new hybrid approach for forecasting economic growth using artificial intelligence approaches. Neural Computing and Applications, 31: 8661--8680

    work page 2019

  4. [4]

    Bai, L.; Yao, L.; Li, C.; Wang, X.; and Wang, C. 2020. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in neural information processing systems, 33: 17804--17815

    2020

  5. [5]

    Z.; and Koltun, V

    Bai, S.; Kolter, J. Z.; and Koltun, V. 2018. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271

    arXiv 2018

  6. [6]

    E.; Jenkins, G

    Box, G. E.; Jenkins, G. M.; Reinsel, G. C.; and Ljung, G. M. 2015. Time series analysis: forecasting and control. John Wiley & Sons

    work page 2015

  7. [7]

    Chaovalit, P.; Gangopadhyay, A.; Karabatis, G.; and Chen, Z. 2011. Discrete wavelet transform-based time series analysis and mining. ACM Computing Surveys (CSUR), 43(2): 1--37

    work page 2011

  8. [8]

    Cho, K.; Van Merri \"e nboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; and Bengio, Y. 2014. Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078

    arXiv 2014

Show all 39 references
  1. [9]

    C.; and Bengio, Y

    Chung, J.; Kastner, K.; Dinh, L.; Goel, K.; Courville, A. C.; and Bengio, Y. 2015. A recurrent latent variable model for sequential data. Advances in neural information processing systems, 28

    2015

  2. [10]

    Dai, T.; Wu, B.; Liu, P.; Li, N.; Yuerong, X.; Xia, S.-T.; and Zhu, Z. 2024. DDN: Dual-domain dynamic normalization for non-stationary time series forecasting. Advances in Neural Information Processing Systems, 37: 108490--108517

    2024

  3. [11]

    Fan, W.; Wang, P.; Wang, D.; Wang, D.; Zhou, Y.; and Fu, Y. 2023. Dish-TS: a general paradigm for alleviating distribution shift in time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, 7522--7529

    2023

  4. [12]

    Han, L.; Ye, H.-J.; and Zhan, D.-C. 2024. The capacity and robustness trade-off: Revisiting the channel independent strategy for multivariate time series forecasting. IEEE Transactions on Knowledge and Data Engineering

    2024

  5. [13]

    Hochreiter, S.; and Schmidhuber, J. 1997. Long short-term memory. Neural computation, 9(8): 1735--1780

    1997

  6. [14]

    Huang, S.-J.; and Hsieh, C.-T. 2002. Coiflet wavelet transform applied to inspect power system disturbance-generated signals. IEEE Transactions on Aerospace and Electronic Systems, 38(1): 204--210

    2002

  7. [15]

    K.; Dasgupta, N.; Natarajan, S.; Pickett, L

    Kaushik, S.; Choudhury, A.; Sheron, P. K.; Dasgupta, N.; Natarajan, S.; Pickett, L. A.; and Dutt, V. 2020. AI in healthcare: time-series forecasting using statistical, neural, and ensemble architectures. Frontiers in big data, 3: 4

    2020

  8. [16]

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

    2021

  9. [17]

    P.; and Ba, J

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

    2014 arXiv

  10. [18]

    Kitaev, N.; Kaiser, L.; and Levskaya, A. 2020. Reformer: The Efficient Transformer. In International Conference on Learning Representations

    2020

  11. [19]

    Lai, G.; Chang, W.-C.; Yang, Y.; and Liu, H. 2018. Modeling long-and short-term temporal patterns with deep neural networks. In The 41st international ACM SIGIR conference on research & development in information retrieval, 95--104

    2018

  12. [20]

    Li, W.; Yang, X.; Liu, W.; Xia, Y.; and Bian, J. 2022. Ddg-da: Data distribution generation for predictable concept drift adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 4092--4100

    2022

  13. [21]

    Liu, Y.; Hu, T.; Zhang, H.; Wu, H.; Wang, S.; Ma, L.; and Long, M. 2023 a . iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. arXiv preprint arXiv:2310.06625

    2023 arXiv

  14. [22]

    Liu, Y.; Li, C.; Wang, J.; and Long, M. 2023 b . Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors. arXiv preprint arXiv:2305.18803

    2023 arXiv

  15. [23]

    Liu, Z.; Cheng, M.; Li, Z.; Huang, Z.; Liu, Q.; Xie, Y.; and Chen, E. 2023 c . Adaptive Normalization for Non-stationary Time Series Forecasting: A Temporal Slice Perspective. In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  16. [24]

    Lu, J.; Liu, A.; Dong, F.; Gu, F.; Gama, J.; and Zhang, G. 2018. Learning under concept drift: A review. IEEE transactions on knowledge and data engineering, 31(12): 2346--2363

    2018

  17. [25]

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

    Nie, Y.; H. Nguyen, N.; Sinthong, P.; and Kalagnanam, J. 2023. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. In International Conference on Learning Representations

    2023

  18. [26]

    Passalis, N.; Tefas, A.; Kanniainen, J.; Gabbouj, M.; and Iosifidis, A. 2019. Deep adaptive input normalization for time series forecasting. IEEE transactions on neural networks and learning systems, 31(9): 3760--3765

    2019

  19. [27]

    RB, C. 1990. STL: A seasonal-trend decomposition procedure based on loess. J Off Stat, 6: 3--73

    1990

  20. [28]

    Sen, R.; Yu, H.-F.; and Dhillon, I. S. 2019. Think globally, act locally: A deep neural network approach to high-dimensional time series forecasting. Advances in neural information processing systems, 32

    2019

  21. [29]

    K.; Ibraheem, S

    Singh, A. K.; Ibraheem, S. K.; Muazzam, M.; and Chaturvedi, D. 2013. An overview of electricity demand forecasting techniques. Network and complex systems, 3(3): 38--48

    2013

  22. [30]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  23. [31]

    \"O .; and Yu, R

    Wang, R.; Dong, Y.; Arik, S. \"O .; and Yu, R. 2022. Koopman neural forecaster for time series with temporal distribution shifts. arXiv preprint arXiv:2210.03675

    2022 arXiv

  24. [32]

    Wen, Q.; Chen, W.; Sun, L.; Zhang, Z.; Wang, L.; Jin, R.; Tan, T.; et al. 2023. Onenet: Enhancing time series forecasting models under concept drift by online ensembling. Advances in Neural Information Processing Systems, 36: 69949--69980

    2023

  25. [33]

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

    2021

  26. [34]

    Wu, Z.; Pan, S.; Long, G.; Jiang, J.; Chang, X.; and Zhang, C. 2020. Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 753--763

    2020

  27. [35]

    Yu, B.; Yin, H.; and Zhu, Z. 2018. Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, 3634--3640

    2018

  28. [36]

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

    2023

  29. [37]

    Zhang, Y.; and Yan, J. 2022. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. In The Eleventh International Conference on Learning Representations

    2022

  30. [38]

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

    2021

  31. [39]

    Zhou, T.; Ma, Z.; Wen, Q.; Wang, X.; Sun, L.; and Jin, R. 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In International Conference on Machine Learning, 27268--27286. PMLR

    2022

Pith tools

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