Pith. sign in

REVIEW 4 major objections 5 minor 49 references

CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables

T0 review · 4 major / 5 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 CrossLinear claims that a single 1D convolution over endogenous and exogenous series can capture the variable dependencies that matter for many-to-one forecasting, outperforming attention-based TimeXer on most benchmark datasets at O(T)…

desk verdict Honest, useful empirical paper whose headline module is a reparameterized convolution; the optimization story is unresolved but the plug-and-play results are worth citing. read the letter →

arxiv 2505.23116 v1 pith:LXX3PJ5K submitted 2025-05-29 cs.LG cs.AI

Pengfei Zhou , Yunlong Liu , Junli Liang , Qi Song , Xiangyang Li This is my paper
classification cs.LGcs.AI
keywords timeseriesforecastingexogenousvariablesmany-to-onecross-correlationembedding1Dconvolutionchannel-independentmodelspatchlinearmodel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that forecasting a target series with external drivers does not need expensive attention: a single 1D convolution over the stacked target and exogenous series can serve as a cross-correlation embedding that captures the dependencies that matter. Built into a linear backbone with patch embedding and a global linear head, this module gives CrossLinear the best MSE in 30 of the many-to-one benchmark settings it reports, beating the Transformer-based TimeXer on most datasets. The same module can be inserted into existing channel-independent models and improves them, with the largest gains when many exogenous variables are present. If the claims hold, forecasting with exogenous variables becomes an $O(T)$ problem rather than an $O((T/p)^2)$ attention problem, with a cheap plug-in for models that previously ignored cross-variable structure.

What carries the argument

The load-bearing object is the cross-correlation embedding module: a 1D convolution with kernel size 3 and stride 1 over the $N \times T$ stack of normalized endogenous and exogenous variables, whose output $X^{cross} = \operatorname{Conv1D}(\operatorname{Stack}(X^{exo}, X^{endo}))$ is combined with the endogenous variable as $X^{emb} = \alpha X^{endo*} + (1-\alpha) X^{cross}$, with $\alpha$ learnable. It is what lets a channel-independent backbone see exogenous information without an attention mechanism. The rest of the architecture is standard: patch embedding with positional embedding for short-range temporal structure, a global linear head for long-range structure, and reversible instance normalization to handle non-stationarity. The paper also uses a weight-sharing extension to make the same module serve multivariate forecasting.

What would settle it

Run CrossLinear on a synthetic many-to-one dataset where one exogenous variable's effect on the target changes sign within the lookback window or lags the target by more than three steps, and compare with a model whose dependency kernel is long enough to see that structure; if the small-kernel model still wins, the time-invariant/direct assumption is not what drives the reported gains.

Watch

Extended reading notes

Core claim

The central claim, on the paper's own terms, is that the right way to model exogenous variables is to be deliberately selective: extract only time-invariant, direct dependencies between each exogenous variable and the target, and leave time-varying and indirect dependencies alone. CrossLinear realizes this with a single-layer 1D convolution of kernel size 3 applied to the matrix formed by stacking the normalized endogenous and exogenous series, producing a cross-correlation embedding that is mixed with the original endogenous series through a learnable weight $\alpha$. The paper argues that this crude dependency model avoids the overfitting that plagues elaborate channel-dependent models, and reports that on 12 real-world datasets the resulting linear model outperforms the Transformer-based TimeXer in most configurations, with top rankings in 30 MSE and 29 MAE comparisons for many-to-one forecasting. For multivariate forecasting, CrossLinear reinterprets the task as many one-to-one problems with shared weights and reports first place in 31 MSE and 28 MAE comparisons. The paper also shows the embedding improves five existing models when plugged in, and proves that its weighted-sum design is mathematically equivalent to using the convolution output alone, so the practical benefit comes from parameterization rather than representational capacity.

Load-bearing premise

The load-bearing premise is that a single short 1D convolution over the stacked series can faithfully represent how exogenous variables move the target; dependencies that are time-varying, nonlinear, or longer than the kernel window will be invisible to the embedding.

Editorial extensions

If this is right

  • Many-to-one forecasting with exogenous variables can be done at $O(T)$ complexity with a linear backbone, making the paradigm practical on long series and resource-constrained settings.
  • Channel-independent models such as SparseTSF, RLinear, PatchTST, DLinear, and Autoformer can be upgraded to use exogenous information by adding the module; reported error reductions include a 27.8% MSE drop for RLinear on Traffic.
  • Datasets with more exogenous variables benefit more from the embedding, so the module is most valuable when rich external data sources are available.
  • Because the convolution has kernel size 3, datasets with longer lead-lag effects need larger kernels; the paper observes ECL improves with larger kernel size, suggesting the module's effective receptive field can be tuned per dataset.
  • The weight-sharing extension means the model covers multivariate forecasting with one shared temporal backbone, not separate per-channel networks.

Reading between the lines

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

  • Beyond the paper's benchmarks, the mathematical equivalence between the weighted-sum and convolution-only designs predicts that with abundant training data the ablation gap should shrink; the reported advantage is an optimization effect, not a difference in what the architectures can represent.
  • The time-invariance assumption is testable: on data where an exogenous variable's effect flips sign within the lookback window, the module should underperform attention-based models, and the paper's own Weather outlier discussion suggests such regimes exist in practice.
  • The plug-in nature suggests a direct extension to anomaly detection or classification with exogenous features, which the paper flags as future work but does not evaluate.
  • The kernel-size sensitivity implies a simple way to extend the method to longer lead-lag dependencies without abandoning linearity: replace the single convolution with a dilated or multi-scale convolution while keeping the same residual mixing; this is not tested in the paper.
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 / 5 minor

Summary. The paper proposes CrossLinear, a linear-architecture model for time series forecasting with exogenous variables. Its main component is a cross-correlation embedding module: a single 1D convolution with kernel size 3 applied to the stacked endogenous and exogenous series, followed by a learned weighted sum with the endogenous series (Eq. 8). The model also uses patching, a positional embedding, a global linear head, and RevIN. The authors evaluate on 12 datasets (long-term and short-term) against 10 baselines, reporting state-of-the-art MSE/MAE in most cases, and an ablation study that they interpret as showing the cross-correlation module is effective. The appendix (Appendix B) contains a proof that the weighted-sum formulation is mathematically equivalent to a reparameterized Cross Only convolution (Eq. 17).

Significance. If the results are taken at face value, CrossLinear is an attractive lightweight (O(T)) alternative to attention-based models for forecasting with exogenous variables, with a plug-and-play module that can be added to CI models. The paper makes several concrete contributions: a clearly specified architecture, a broad empirical evaluation, a formal equivalence proof in Appendix B, and a public code release. These are genuine strengths. However, the central claim that the cross-correlation embedding drives the gains by capturing variable dependencies is weakened by an internal tension: the paper's own proof shows that the Summation variant is equivalent to a Cross Only convolution, yet the ablation shows large performance gaps between them. This means the observed improvement cannot be attributed to the model's capacity to represent dependencies, and the paper's explanation (training dynamics, Section 4.2) is not tested. The lack of error bars and significance tests further limits the empirical claim, as many reported margins are at the third decimal.

major comments (4)
  1. [§4.2, Appendix B, Eq. (17), Table 4] The proof in Appendix B shows that the Summation method (Eq. 8) is equivalent to Cross Only with kernel K' = (1-alpha)K + alpha*S. Since the two parameterizations define the same function class, the large performance gaps in Table 4 (e.g., ECL: 0.323 vs 0.396 MSE; Traffic: 0.152 vs 0.182) cannot come from the model's representational power; they must come from optimization, initialization, or regularization effects. The paper's claim that the cross-correlation embedding 'captures variable dependencies' is therefore not supported by the ablation. The attribution should be revised: either the contribution is a parameterization/regularization choice, or the authors should provide evidence (e.g., training Cross Only with the reparameterized kernel, or showing that the learned kernel differs meaningfully from a random kernel of the same capacity) that the convolution itself is responsible for the gains.
  2. [Tables 2 and 3, §4.1] The empirical claim of 'superior performance' and 'top rankings in 30 cases for MSE and 29 for MAE' is not backed by repeated runs or significance tests. Many margins are at the third decimal (e.g., ECL-96: 0.251 vs 0.261 MSE; Weather: all models around 0.001/0.002 MSE). The Weather dataset is explicitly compared at four decimals, but several models achieve identical rounded values, making the superiority claim meaningless on that dataset. The authors should report mean and standard deviation over multiple seeds, or at least perform pairwise significance tests, to establish that the reported differences are not noise.
  3. [§4.3.2, Table 6] The missing-values experiment shows that masking exogenous variables degrades performance, which is evidence that exogenous information matters. However, it does not isolate the cross-correlation embedding module. Given the equivalence result in Appendix B, the performance gains of the Summation variant could arise from the residual weighting (alpha close to 1) rather than from the convolution extracting exogenous structure. To support the mechanistic claim, the authors should compare against a version with a fixed random convolution kernel (with the same capacity) or a model that uses exogenous variables through a simple linear projection; if the learned kernel does not provide additional improvement, the module's contribution is not 'cross-correlation' but rather a regularized residual connection.
  4. [§3.4 and Appendix C] The model assumes that a single 1D convolution with kernel size 3 captures time-invariant direct dependencies and lead-lag effects. A kernel size of 3 only permits lead-lag of one time step, and the paper does not validate that the learned kernels correspond to actual variable dependencies. The variate-wise correlation analysis in §4.4.2 is qualitative and offers no quantitative check against known true dependencies or against chance. The hyperparameter sensitivity in Appendix C shows that a larger kernel helps on ECL but the default is 3. To support the claim that the module captures dependencies rather than acting as a generic linear filter, the authors should provide an analysis of the learned kernel, such as comparing its predictive value against a version with shuffled exogenous channels or with a fixed kernel.
minor comments (5)
  1. [Table 5 caption] The caption contains a typo: 'Performace promotion' should be 'Performance promotion'.
  2. [Section 4.2] The sentence 'Although mathematically, our method could be seen as equivalent to Cross Only under certain conditions' is vague; the appendix proves this equivalence unconditionally, so the sentence should be rephrased to state the precise relationship.
  3. [Figure 3] The three subplots in Figure 3 are not labeled (a), (b), (c), making it hard to map the caption to the panels; please add panel labels.
  4. [Appendix B, Eq. (16)] The notation 'S * Stack(...)' is ambiguous: it would be clearer to write 'S · Stack(...)' or explicitly 'S times the stacked matrix' to indicate matrix multiplication that selects the endogenous row.
  5. [Section 3.4] The 'Note:' paragraph after Eq. (8) is an important caveat and should be moved to the main text or at least referenced again in the ablation discussion, since it directly bears on the interpretation of Table 4.

Circularity Check

1 steps flagged · score 6.0 of 10

The paper's own appendix reduces the cross-correlation embedding to a single reparameterized 1D convolution, undermining the ablation's attribution; external benchmark comparisons remain independent.

  1. self definitional [Section 3.4 (Eqs. 7-8, Note); Appendix B (Eq. 17)]
    "In theory, X_emb and X_cross are mathematically equivalent (see Appendix B for proof). ... K′ =(1−𝛼)K+𝛼𝑆"

    The module's final embedding (Eq. 8) is defined as alpha times the endogenous series plus (1-alpha) times Conv1D(Stack(...)) (Eq. 7). Appendix B constructs K'=(1-alpha)K+alpha S and proves the same output is obtained by a single Conv1D with kernel K' on the same stack, i.e., exactly the 'Cross Only' embedding. Thus the Summation variant is, by construction, a reparameterization of Cross Only; the ablation in Table 4 cannot test two different dependency-capturing mechanisms. The paper's own equations reduce the claimed 'cross-correlation embedding' mechanism to a standard 1D convolution, so the large performance gaps (e.g., ECL 0.323 vs 0.396 MSE) must be optimization/regularization effects, not evidence for the module's ability to capture exogenous structure.

full rationale

The empirical forecasting comparisons in Tables 2-3 are self-contained against external baselines and are not derived from the embedding equations, so the main predictive claims are not circular. The circularity burden falls on the design/attribution claim: Section 3.4 defines the final embedding as a weighted sum (Eq. 8) of the endogenous series and the Conv1D output (Eq. 7), and Appendix B proves that this weighted sum is exactly a single Conv1D with kernel K'=(1-alpha)K+alpha S. Consequently, the 'Summation' variant is, by the paper's own equations, the same function class as 'Cross Only'; the large MSE differences in Table 4 cannot be evidence for the cross-correlation embedding mechanism, only for parameterization or optimization effects. The paper acknowledges the equivalence in its Note and Appendix B but does not resolve the tension; it attributes the difference to training dynamics without testing that explanation. No self-citations are load-bearing, and no prediction is a fitted parameter renamed as a result. Score 6 reflects partial, not total, circularity: the central novelty claim reduces by construction to a standard convolution, while the benchmark results retain independent empirical content.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The model's empirical performance depends on learned weights and scalar gates, and its design rests on the unproven premise that variable dependencies are time-invariant and direct. The Appendix B equivalence shows the fusion is a reparameterization of a convolution, not a new operation.

free parameters (5)
  • alpha = learned, not reported; sensitive to initialization near 1
    Balances endogenous vs cross-correlation embedding in Eq. 8; the paper shows initialization near 1 improves performance, so it is effectively tuned.
  • beta = learned, not reported
    Balances patch projection vs positional embedding in Eq. 10.
  • patch size p = chosen per dataset from {2,4,8,16,32} (Figure 5)
    Hyperparameter controlling short-term dependency capture; sensitivity shown.
  • kernel size = 3
    Fixed at 3 by hand; sensitivity shown in Figure 5.
  • hidden dimension d = chosen from 128-1024
    Capacity hyperparameter.
assumptions (4)
  • domain assumption Variable dependencies between endogenous and exogenous variables are time-invariant and direct.
    The core premise of the cross-correlation embedding (Sections 1 and 3.4); if false, the module's fixed convolution cannot capture relevant relationships.
  • ad hoc to paper A single 1D convolution layer with kernel size 3 can represent the relevant lead-lag effects.
    The paper chooses kernel 3 to capture dependencies, without deriving why this suffices.
  • domain assumption Instance normalization (RevIN) improves non-stationary forecasting.
    Borrowed from RevIN [13] and used without re-validation in this setting.
  • standard math L2 loss is the appropriate objective.
    Standard for regression, but not justified for the specific data characteristics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables." pith.science (2026). https://pith.science/paper/LXX3PJ5K

@misc{pith2026250523116,
  author       = {Pith},
  title        = {Pith review of: CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LXX3PJ5K}},
  note         = {Machine review of arXiv:2505.23116}
}
read the original abstract

Time series forecasting with exogenous variables is a critical emerging paradigm that presents unique challenges in modeling dependencies between variables. Traditional models often struggle to differentiate between endogenous and exogenous variables, leading to inefficiencies and overfitting. In this paper, we introduce CrossLinear, a novel Linear-based forecasting model that addresses these challenges by incorporating a plug-and-play cross-correlation embedding module. This lightweight module captures the dependencies between variables with minimal computational cost and seamlessly integrates into existing neural networks. Specifically, it captures time-invariant and direct variable dependencies while disregarding time-varying or indirect dependencies, thereby mitigating the risk of overfitting in dependency modeling and contributing to consistent performance improvements. Furthermore, CrossLinear employs patch-wise processing and a global linear head to effectively capture both short-term and long-term temporal dependencies, further improving its forecasting precision. Extensive experiments on 12 real-world datasets demonstrate that CrossLinear achieves superior performance in both short-term and long-term forecasting tasks. The ablation study underscores the effectiveness of the cross-correlation embedding module. Additionally, the generalizability of this module makes it a valuable plug-in for various forecasting tasks across different domains. Codes are available at https://github.com/mumiao2000/CrossLinear.

Figures

Figures reproduced from arXiv: 2505.23116 by the authors.

Figure 1
Figure 1. Different forecasting paradigms. Here, X𝑖 represents the 𝑖 th variable, X 𝑒𝑛𝑑𝑜 1 is the endogenous variable, and X 𝑒𝑥𝑜 𝑖 denotes the 𝑖 th exogenous variable. (a) Univariate forecasting. (b) Multivariate forecasting. (c) Forecasting with exogenous variables. Compared with traditional paradigms, the many-to-one time series forecasting paradigm introduces new challenges in model￾ing the dependencies between endogenous … view at source ↗
Figure 2
Figure 2. Architecture of CrossLinear. (a) The cross-correlation embedding module captures variable dependencies. (b) The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Hyperparameter sensitivity with respect to [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Model anlysis. (a) Variate-wise correlation analysis. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Hyperparameter sensitivity with respect to patch [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 35 canonical work pages

  1. [1]

    Mohammed Alnour, Ashar Awan, and Md Emran Hossain. 2024. Towards a green transportation system in Mexico: The role of renewable energy and transport public-private partnership to curb emissions.Journal of Cleaner Production442 (2024), 140984

    work page 2024

  2. [2]

    Christoph Bergmeir. 2024. Fundamental limitations of foundational forecasting models: The need for multimodality and rigorous evaluation. https://cbergmeir. com/talks/neurips2024/

    2024

  3. [3]

    George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung. 2015. Time series analysis: forecasting and control. John Wiley & Sons

    2015

  4. [4]

    Wanlin Cai, Yuxuan Liang, Xianggen Liu, Jianshuai Feng, and Yuankai Wu. 2024. Msgnet: Learning multi-scale inter-series correlations for multivariate time series forecasting. InAAAI, Vol. 38. 11141–11149

    work page 2024

  5. [5]

    Taehee Chang, Kyung-Duk Min, Sung-il Cho, and Yoonhee Kim. 2024. Asso- ciations of meteorological factors and dynamics of scrub typhus incidence in South Korea: A nationwide time-series study.Environmental Research245 (2024), 117994

    work page 2024

  6. [6]

    Alice Chuang. 1991. Time series analysis: univariate and multivariate methods

    work page 1991

  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)

    arXiv 2023

  8. [8]

    Alexey Dosovitskiy. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    arXiv 2020

Show all 49 references
  1. [9]

    Seyyed Ahmad Edalatpanah, Farnaz Sheikh Hassani, Florentin Smarandache, Ali Sorourkhah, Dragan Pamucar, and Bing Cui. 2024. A hybrid time series forecasting method based on neutrosophic logic with applications in financial issues.Engineering applications of artificial intellig...

    2024

  2. [10]

    Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2023. Tsmixer: Lightweight mlp-mixer model for multivariate time series forecasting. InKDD. 459–469

    2023

  3. [11]

    Xusen Guo, Qiming Zhang, Mingxing Peng, Meixin Zhua, et al . 2024. Ex- plainable Traffic Flow Prediction with Large Language Models.arXiv preprint arXiv:2404.02937(2024)

    2024

  4. [12]

    HyunGi Kim, Siwon Kim, Jisoo Mok, and Sungroh Yoon. 2025. Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation.AAAI (2025)

    2025

  5. [13]

    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. InICLR

    2021

  6. [14]

    Diederik P Kingma. 2014. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980(2014)

    2014 arXiv

  7. [15]

    Jesus Lago, Grzegorz Marcjasz, Bart De Schutter, and Rafał Weron. 2021. Fore- casting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark.Applied Energy293 (2021), 116983

    2021

  8. [16]

    Xiang Lan, Hanshu Yan, Shenda Hong, and Mengling Feng. 2024. Towards En- hancing Time Series Contrastive Learning: A Dynamic Bad Pair Mining Approach. InICLR

    2024

  9. [17]

    Zhe Li, Shiyi Qi, Yiduo Li, and Zenglin Xu. 2023. Revisiting long-term time series forecasting: An investigation on linear mapping.arXiv preprint arXiv:2305.10721 (2023)

    2023 arXiv

  10. [18]

    Daojun Liang, Haixia Zhang, Dongfeng Yuan, Bingzheng Zhang, and Minggao Zhang. 2024. Minusformer: Improving Time Series Forecasting by Progressively Learning Residuals.arXiv preprint arXiv:2402.02332(2024)

    2024 arXiv

  11. [19]

    ChungYi Lin, Shen-Lung Tung, Hung-Ting Su, and Winston H Hsu. 2024. Tel- Trans: Applying Multi-Type Telecom Data to Transportation Evaluation and Prediction via Multifaceted Graph Modeling. InAAAI, Vol. 38. 22927–22933

    2024

  12. [20]

    Changgui Lin, Kun Yang, Jianping Huang, Wenjun Tang, Jun Qin, Xiaolei Niu, Yingying Chen, Deliang Chen, Ning Lu, and Rong Fu. 2015. Impacts of wind stilling on solar radiation variability in China.Scientific reports5, 1 (2015), 15135

    2015

  13. [21]

    Shengsheng Lin, Weiwei Lin, Wentai Wu, Haojun Chen, and Junjie Yang. 2024. SparseTSF: Modeling Long-term Time Series Forecasting with 1k Parameters. arXiv preprint arXiv:2405.00946(2024)

    2024 arXiv

  14. [22]

    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. InICLR

    2024

  15. [23]

    Zhiding Liu, Mingyue Cheng, Zhi Li, Zhenya Huang, Qi Liu, Yanhu Xie, and Enhong Chen. 2024. Adaptive normalization for non-stationary time series forecasting: A temporal slice perspective.NeurIPS36 (2024)

    2024

  16. [24]

    Minrong Lu and Xuerong Xu. 2024. TRNN: An efficient time-series recurrent neural network for stock price prediction.Information Sciences657 (2024), 119951

    2024

  17. [25]

    Donghao Luo and Xue Wang. 2024. Moderntcn: A modern pure convolution structure for general time series analysis. InICLR

    2024

  18. [26]

    Shahenaz Mulla, Chaitanya B Pande, and Sudhir K Singh. 2024. Times Series Forecasting of Monthly Rainfall using Seasonal Auto Regressive Integrated Mov- ing Average with EXogenous Variables (SARIMAX) Model.Water Resources Management38, 6 (2024), 1825–1846

    2024

  19. [27]

    Paul Newbold and Clive WJ Granger. 1974. Experience with forecasting univariate time series and the combination of forecasts.Journal of the Royal Statistical Society: Series A (General)137, 2 (1974), 131–146

    1974

  20. [28]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A time series is worth 64 words: Long-term forecasting with transformers.ICLR (2022)

    2022

  21. [29]

    Kin G Olivares, Cristian Challu, Grzegorz Marcjasz, Rafał Weron, and Artur Dubrawski. 2023. Neural basis expansion analysis with exogenous variables: Forecasting electricity prices with NBEATSx.International Journal of Forecasting 39, 2 (2023), 884–900

    2023

  22. [30]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library.NeurIPS 32 (2019)

    2019

  23. [31]

    Montaser NA Ramadan, Mohammed AH Ali, Shin Yee Khoo, Mohammad Alkhed- her, and Mohammad Alherbawi. 2024. Real-time IoT-powered AI system for monitoring and forecasting of air pollution in industrial environment.Ecotoxi- cology and environmental safety283 (2024), 116856

    2024

  24. [32]

    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 Data...

    2024

  25. [33]

    Jingzhe Shi, Qinwei Ma, Huan Ma, and Lei Li. 2024. Scaling Law for Time Series Forecasting.NeurIPS(2024)

    2024

  26. [34]

    2006.Chemistry: The molecular nature of matter and change

    Martin Stuart Silberberg, Patricia Amateis, Rashmi Venkateswaran, and Lydia Chen. 2006.Chemistry: The molecular nature of matter and change. Vol. 4. McGraw- Hill New York

    2006

  27. [35]

    Radek Svoboda, Vojtech Kotik, and Jan Platos. 2024. Data-driven multi-step en- ergy consumption forecasting with complex seasonality patterns and exogenous variables: Model accuracy assessment in change point neighborhoods.Applied Soft Computing150 (2024), 111099

    2024

  28. [36]

    Peiwang Tang and Weitai Zhang. 2024. Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting.arXiv preprint arXiv:2405.13575(2024)

    2024 arXiv

  29. [37]

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. 2021. Mlp-mixer: An all-mlp architecture for vision.NeurIPS34 (2021), 24261–24272

    2021

  30. [38]

    A Vaswani. 2017. Attention is all you need.NeurIPS(2017)

    2017

  31. [39]

    Yihe Wang, Yu Han, Haishuai Wang, and Xiang Zhang. 2024. Contrast everything: A hierarchical contrastive framework for medical time-series.NeurIPS36 (2024)

    2024

  32. [40]

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Yong Liu, Mingsheng Long, and Jianmin Wang. 2024. Deep time series models: A comprehensive survey and benchmark. arXiv preprint arXiv:2407.13278(2024)

    2024 arXiv

  33. [41]

    Yuxuan Wang, Haixu Wu, Jiaxiang Dong, Guo Qin, Haoran Zhang, Yong Liu, Yunzhong Qiu, Jianmin Wang, and Mingsheng Long. 2024. Timexer: Empowering transformers for time series forecasting with exogenous variables.NeurIPS(2024)

    2024

  34. [42]

    Yucheng Wang, Yuecong Xu, Jianfei Yang, Min Wu, Xiaoli Li, Lihua Xie, and Zhenghua Chen. 2024. Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data. InAAAI, Vol. 38. 15715–15724

    2024

  35. [43]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2022. Timesnet: Temporal 2d-variation modeling for general time series analysis.ICLR(2022)

    2022

  36. [44]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. NeurIPS34 (2021), 22419–22430

    2021

  37. [45]

    Kun Yi, Qi Zhang, Wei Fan, Hui He, Liang Hu, Pengyang Wang, Ning An, Long- bing Cao, and Zhendong Niu. 2024. FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective.NeurIPS36 (2024)

    2024

  38. [46]

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting?. InAAAI, Vol. 37. 11121–11128

    2023

  39. [47]

    Yunhao Zhang and Junchi Yan. 2023. Crossformer: Transformer utilizing cross- dimension dependency for multivariate time series forecasting. InICLR

    2023

  40. [48]

    Nanyang Zhu, Ying Wang, Kun Yuan, Jiahao Yan, Yaping Li, and Kaifeng Zhang

  41. [2024]

    CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables KDD ’25, August 3–7, 2025, Toronto, ON, Canada

    GGNet: A novel graph structure for power forecasting in renewable power plants considering temporal lead-lag correlations.Applied Energy364 (2024), 123194. CrossLinear: Plug-and-Play Cross-Correlation Embedding for Time Series Forecasting with Exogenous Variables KDD ’25, Augu...

    2024

Pith tools

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