Pith. sign in

REVIEW 4 major objections 5 minor 36 references

ST-MTM: Masked Time Series Modeling with Seasonal-Trend Decomposition for Time Series Forecasting

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Decomposing a time series into seasonal and trend parts before masking gives a better pre-training task for forecasting than masking the raw series.

desk verdict A genuinely new decomposition-aware masking recipe for time series forecasting, but the 'consistently superior' claim is undercut by test-set kernel selection and missing error bars. read the letter →

arxiv 2507.00013 v1 pith:NE6ODQAY submitted 2025-06-13 cs.LG cs.AIstat.ML

classification cs.LGcs.AIstat.ML
keywords maskedtime-seriesmodelingseasonal-trenddecompositiontimeseriesforecastingself-supervisedlearningcontrastiveperiodmaskingautocorrelation
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 argues that masked time-series modeling learns spurious temporal patterns when it masks the raw series, because raw data entangles at least two kinds of semantic structure. Its proposed fix, the ST-MTM framework, first decomposes each series into a smooth trend and a seasonal residual, then masks each component with a task tuned to that component: period-based masking for the season, sub-series masking for the trend. These masked views are encoded separately, merged by a learned gating layer, and forced to agree through a contrastive loss, after which the original series is reconstructed. On nine forecasting benchmarks the paper reports the best or second-best forecast error in most scenarios, including cross-domain transfer and a low-seasonality dataset. If the claim holds, the practical message is that the masking scheme, not just the encoder, should respect the semantic structure of the input.

What carries the argument

The load-bearing mechanism is seasonal-trend decomposition used as a masking prior. A moving average extracts the trend, the remainder is the seasonal part, and the model's autocorrelation (computed by FFT) on the seasonal part supplies the top-$K$ lags it treats as periods; for each lag it samples an anchor sub-series and masks every sub-series an integer number of periods away, producing $K$ masked seasonal views. The trend is divided into non-overlapping sub-series and a fraction are masked. Separate encoders handle the two component types, an autocorrelation-weighted sum merges the seasonal views, a contextual contrastive loss pulls those views' instance representations together, and a component-wise gating layer decides the seasonal-versus-trend blend before the decoder reconstructs the original series.

What would settle it

Build a synthetic series with a known trend and a seasonality whose period changes halfway through, pre-train ST-MTM and a raw-masking baseline with the same encoder, and compare forecast errors: if ST-MTM does not beat raw masking, or if its period mask routinely hides misaligned positions, the decomposition assumption fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the temporal semantics of a time series live in its components, so a masked pre-training task should be defined per component. It claims that seasonal structure is multi-periodic, and that masking all sub-series aligned with the top autocorrelation lags makes the model learn periodicity rather than raw-value shortcuts; on the trend side, it claims that masking contiguous sub-series of a smooth trend teaches local coherence. The same decomposition is kept through representation learning, with a frequency-domain MLP for the seasonal views, a Transformer for the trend, and a gating layer that blends the two per timestamp. The paper then claims that this component-wise masking plus a contrastive alignment of the masked seasonal views transfers better across domains and forecasts better than raw masked modeling, contrastive learning, and decomposition-based supervised methods on the nine benchmarks tested.

Load-bearing premise

The method assumes that a moving-average trend plus a short list of autocorrelation-detected periods cleanly separates a series into seasonal and trend components, so that the positions hidden by period masking really are the semantically related ones.

Editorial extensions

If this is right

  • Pre-training a forecasting model by masking raw values is a weaker objective than masking each decomposed component according to what that component means.
  • Autocorrelation-based period detection can replace hand-specified frequency lists in masked time-series pre-training, because the top lags are found from the data itself.
  • Adding more masked views only helps if those views are aligned by a contrastive objective; unaligned extra views in the paper's ablations hurt performance.
  • The learned gating layer gives a mechanism for handling series with weak or missing seasonality, because the model can down-weight the seasonal representation when periodicity is absent.

Reading between the lines

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

  • If the mechanism is correct, time-series foundation models could adopt the same decomposition-aware masking as their pre-training task instead of masking raw patches; the paper gestures at this direction but does not test it.
  • A testable engineering extension is to make the decomposition adaptive: on series with drifting periods, re-estimating the autocorrelation lags or moving-average kernel over time would prevent the period mask from hiding stale positions.
  • One indirect consequence is that datasets with strong known periodicities (electricity, traffic) should show the largest gains from period masking, while aperiodic financial series should rely mostly on the trend branch; the paper's Exchange experiments are consistent with that split.
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 ST-MTM, a self-supervised masked time-series modeling framework that first decomposes a series into seasonal and trend components via moving average (Eq. 1), then applies a period-masking strategy to the seasonal component based on top-K autocorrelation lags (Eqs. 2-3) and a sub-series masking strategy to the trend component. The seasonal component is encoded by a frequency-domain MLP (SFM), the trend component by a Transformer, and the two are merged through a component-wise gating layer (Eq. 8). Pre-training combines a reconstruction loss (Eq. 9) with a contextual contrastive loss (Eq. 10). The authors evaluate the method on nine forecasting benchmarks against thirteen baselines, report in-domain, cross-domain, ablation, sensitivity, and runtime results, and claim that ST-MTM achieves consistently superior forecasting performance compared to masked modeling, contrastive learning, and supervised forecasting methods.

Significance. If the empirical claims are sound, the paper makes a useful contribution: it identifies a limitation of masking raw time series, proposes a decomposition-aware masking pretext task with a plausible semantic motivation, and validates it across a broad benchmark suite. The authors state that code is available, which supports reproducibility. The paper is, however, primarily an empirical contribution, and its significance is bounded by the strength of the experimental evidence; the current evidence is weakened by at least one test-set selection issue and by several fairness concerns in the baseline comparisons.

major comments (4)
  1. [Appendix C.2 and Figure 9] The moving-average kernel size is selected using the same test MSE metric reported in Tables 1 and 2; the text states that "these optimal kernel sizes are the hyper-parameters with which we report performance." On ETTh1 and ETTh2 the sensitivity curves in Figure 9 show MSE variation of roughly 0.02-0.04 across kernel sizes, which is the same order as the margins over SimMTM, SCNN, and DLinear in Tables 1 and 2. The reported results for these two datasets are therefore selected on the test set rather than evaluated on it, which can inflate the apparent gains. This is a load-bearing issue for the central claim of consistent superiority, and it must be addressed by fixing the kernel size through a validation split or by reporting performance for all kernel sizes and showing that the ranking is unchanged.
  2. [Section 4.2.1, Tables 1 and 2] The abstract's claim of "consistently superior forecasting performance" is contradicted by the paper's own tables: PatchTST outperforms ST-MTM on ETTm1 (MSE 0.343 vs. 0.350) and on ILI (MSE 2.264 vs. 2.757), DLinear outperforms ST-MTM on Electricity (MSE 0.169 vs. 0.170 with lower MAE), and the additional TimeSiam comparison in Appendix Table 10 shows TimeSiam better on Weather, Electricity, PEMS08, and ILI. The claim should be qualified to reflect the many scenarios where ST-MTM is not the best method.
  3. [Section 4.1.3 and Section 4.2] The paper states that all experiments are repeated five times, but no standard deviations or significance tests are reported anywhere. Several decisive margins are very small (e.g., Electricity in Table 2: 0.170 vs. 0.169), so without variance information the reader cannot determine whether the reported ranking is meaningful. Please report standard deviations for at least the main tables, or a paired significance test across the five repeats.
  4. [Appendix A.2] The baseline configuration is not equal across methods: SimMTM is run with reduced pre-training epochs and reduced model size on Electricity and PEMS08, and Ti-MAE is not run at several long horizons on those datasets because of resource limits. The paper justifies the Ti-MAE omission by saying its performance was "generally inadequate," but this is not a substitute for a fair comparison. The unequal compute budget can inflate ST-MTM's wins on exactly the datasets where it shows the largest margins, so the comparative claim is weaker than presented.
minor comments (5)
  1. [Table 3] The caption lists prediction lengths as {96, 196, 336, 720}; 196 should presumably be 192.
  2. [Reference [11]] The author name "Minsheong Long" appears to be a typo for "Mingsheng Long."
  3. [Section 3.1.1, Eq. (3)] The notation argTopk is used without definition; please define it as returning the indices of the K largest autocorrelation values.
  4. [Section 3.1.1] The phrase "mask all sub-series at positions that are n multiples of the period away" should say "integer multiples" for clarity.
  5. [Section 4.4.3] The missing-data experiment removes points from both train and test sets and then predicts the original values; please clarify how the corrupted test inputs are handled by the forecasting head and how this differs from a standard forecasting evaluation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ST-MTM is an empirical pre-training recipe evaluated on external benchmarks, with a test-set kernel-selection concern that is a correctness risk rather than a circularity.

full rationale

The paper's central claim is an empirical performance comparison, not a derivation. ST-MTM's components—moving-average seasonal-trend decomposition (Eq. 1), autocorrelation-based period selection (Eqs. 2-3), period and sub-series masking, SFM/Transformer encoders, and the reconstruction plus contrastive losses (Eqs. 9-11)—are defined independently of the forecasting target and evaluated on external benchmarks (ETT, Weather, Electricity, PEMS08, ILI, Solar, Exchange). No parameter is defined in terms of the reported MSE, no equation reduces to another by construction, and no load-bearing result is imported from the authors' own prior work (the reference list contains no self-citations). The only notable concern is Appendix C.2, where the moving-average kernel size for ETTh1 and ETTh2 is chosen as the value with the best test-set MSE and then used to report performance; this is a test-set selection/leakage issue that could inflate those two datasets' apparent gains and is a correctness or reproducibility risk, but it is not circularity because the reported MSE is a measurement of forecasting error, not the selected hyperparameter renamed as a prediction. The derivation chain therefore does not reduce to its inputs.

Assumptions & free parameters 9 free parameters · 6 assumptions · 0 invented entities

ST-MTM does not derive its method from first principles; it builds on a chain of modeling assumptions. The free parameters are the hyperparameters chosen for masking, loss weighting, and architecture; these are tuned per dataset and some appear to be selected using the same datasets on which performance is reported. The axioms are the decomposition assumption, the autocorrelation-as-period assumption, channel independence, the usefulness of masked reconstruction, and the assumption that masked seasonal views share context. No new entities are introduced.

free parameters (9)
  • number_of_masked_seasonal_series_K = 3
    Hyperparameter in Sections 3.1.1 and 4.1.3; sensitivity in Appendix C.3 shows larger K can increase MSE.
  • trend_masking_ratio_p = 0.2
    Set in Section 4.1.3; sensitivity in Appendix C.3 shows performance is fairly robust to p.
  • segment_length_l = 25 (3 for ILI)
    Set in Section 4.1.3 to control the sub-series granularity for both trend and period masking.
  • moving_average_kernel_size = dataset-specific, e.g., 200 for ETTh1, 50 for ETTh2
    Used in Eq. 1 for trend extraction; Appendix C.2 reports best kernel sizes were chosen and used for reported performance.
  • contrastive_loss_weight_alpha = 0.5
    Set in Section 4.1.3; sensitivity shown in Appendix C.1.
  • contrastive_temperature_tau = 0.1
    Set in Section 4.1.3; sensitivity shown in Appendix C.1.
  • input_length_L = 336 (36 for ILI)
    Look-back window fixed in Section 4.1.3; chosen per standard benchmark protocol.
  • pre_training_epochs = 50 (10 for Electricity and PEMS08)
    Section 4.1.3; reduced due to time constraints.
  • encoder_hyperparameters = layers {1,2}, d_model {16,32,64}, heads {4,8,16}, d_ff {32,64,128} per dataset
    Appendix A.3 Table 7 lists candidate sets; chosen per dataset without a described held-out protocol.
assumptions (6)
  • domain assumption Moving average additively separates a time series into trend and seasonal components (Eq. 1).
    The whole masking strategy assumes this decomposition captures semantically distinct temporal patterns; misspecified for non-additive or irregular series.
  • domain assumption Top-K autocorrelation lags identify the true inherent periods (Eqs. 2-3).
    Assumes periodic structure is visible as autocorrelation peaks on finite, non-stationary, multi-variable series.
  • standard math Wiener-Khinchin theorem relates autocorrelation to the power spectrum via FFT.
    Used in Section 3.1.1 for efficient period computation.
  • domain assumption Each channel can be modeled independently (channel independence).
    Section 4.1.3; discards cross-variable dependencies, a strong assumption for electricity and traffic datasets with correlated sensors.
  • domain assumption Masked reconstruction is an effective pretext task for forecasting.
    Adopted from prior MTM literature (SimMTM, PatchTST); not derived here.
  • ad hoc to paper Different masked seasonal views of the same series share identical contextual information.
    Section 3.3.2; this assumption motivates the contrastive loss and is not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ST-MTM: Masked Time Series Modeling with Seasonal-Trend Decomposition for Time Series Forecasting." pith.science (2026). https://pith.science/paper/NE6ODQAY

@misc{pith2026250700013,
  author       = {Pith},
  title        = {Pith review of: ST-MTM: Masked Time Series Modeling with Seasonal-Trend Decomposition for Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NE6ODQAY}},
  note         = {Machine review of arXiv:2507.00013}
}
read the original abstract

Forecasting complex time series is an important yet challenging problem that involves various industrial applications. Recently, masked time-series modeling has been proposed to effectively model temporal dependencies for forecasting by reconstructing masked segments from unmasked ones. However, since the semantic information in time series is involved in intricate temporal variations generated by multiple time series components, simply masking a raw time series ignores the inherent semantic structure, which may cause MTM to learn spurious temporal patterns present in the raw data. To capture distinct temporal semantics, we show that masked modeling techniques should address entangled patterns through a decomposition approach. Specifically, we propose ST-MTM, a masked time-series modeling framework with seasonal-trend decomposition, which includes a novel masking method for the seasonal-trend components that incorporates different temporal variations from each component. ST-MTM uses a period masking strategy for seasonal components to produce multiple masked seasonal series based on inherent multi-periodicity and a sub-series masking strategy for trend components to mask temporal regions that share similar variations. The proposed masking method presents an effective pre-training task for learning intricate temporal variations and dependencies. Additionally, ST-MTM introduces a contrastive learning task to support masked modeling by enhancing contextual consistency among multiple masked seasonal representations. Experimental results show that our proposed ST-MTM achieves consistently superior forecasting performance compared to existing masked modeling, contrastive learning, and supervised forecasting methods.

Figures

Figures reproduced from arXiv: 2507.00013 by the authors.

Figure 1
Figure 1. The attention score distributions of the MTM Trans [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall architecture of ST-MTM. The ST-MTM architecture includes self-supervised pre-training and fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Seasonal frequency MLP 3.2.1 Encoding seasonal series. To obtain a seasonal time series representation from a set of multiple masked seasonal series, ST￾MTM first encodes each masked seasonal series into point-wise representations. The seasonal component of a time series exhibits multiple periodic properties, generated from its constituent fre￾quencies [26]. To effectively capture the periodic information in masked … view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Ablation of ST-MTM on the reconstruction task [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Visualization of the outputs from the component [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: The left part shows MSE performance with vary [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Sensitivity analysis of alpha (upper left), batch size [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: The MSE performance of ST-MTM on the ETTh2 [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 12
Figure 12. Figure 12: Prediction cases on the Weather dataset for pre [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 20 canonical work pages

  1. [1]

    Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. 2022. Data2vec: A general framework for self-supervised learning in speech, vision and language. InInternational Conference on Machine Learning. PMLR, 1298–1312

  2. [2]

    Hubert Banville, Isabela Albuquerque, Aapo Hyvärinen, Graeme Moffat, Denis- Alexander Engemann, and Alexandre Gramfort. 2019. Self-supervised representa- tion learning from electroencephalography signals. In2019 IEEE 29th International Workshop on Machine Learning for Signal Processing (MLSP). IEEE, 1–6

  3. [3]

    Ling Cai, Krzysztof Janowicz, Gengchen Mai, Bo Yan, and Rui Zhu. 2020. Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting.Transactions in GIS24, 3 (2020), 736–755

  4. [4]

    Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. 2018. Deep clustering for unsupervised learning of visual features. InProceedings of the European conference on computer vision (ECCV). 132–149

  5. [5]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional conference on machine learning. PMLR, 1597–1607

  6. [6]

    Ranak Roy Chowdhury, Xiyuan Zhang, Jingbo Shang, Rajesh K Gupta, and Dezhi Hong. 2022. Tarnet: Task-aware reconstruction for time-series transformer. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 212–220

  7. [7]

    Robert B Cleveland, William S Cleveland, Jean E McRae, and Irma Terpenning

  8. [8]

    Jinliang Deng, Xiusi Chen, Renhe Jiang, Du Yin, Yi Yang, Xuan Song, and Ivor W Tsang. 2024. Disentangling Structured Components: Towards Adaptive, Inter- pretable and Scalable Time Series Forecasting.IEEE Transactions on Knowledge and Data Engineering(2024)

Show all 36 references
  1. [9]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding.arXiv preprint arXiv:1810.04805(2018)

  2. [10]

    Jiaxiang Dong, Haixu Wu, Yuxuan Wang, Yunzhong Qiu, Li Zhang, Jianmin Wang, and Mingsheng Long. 2024. TimeSiam: A Pre-Training Framework for Siamese Time-Series Modeling. InForty-first International Conference on Machine Learning

  3. [11]

    Jiaxiang Dong, Haixu Wu, Haoran Zhang, Li Zhang, Jianmin Wang, and Min- sheong Long. 2023. SimMTM : A Simple Pre-Training Framework for Masked Time-Series Modeling.Advances in Neural Information Processing Systems(2023)

  4. [12]

    Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Chee Keong, Xiaoli Li Kwoh, and Cuntai Guan. 2021. Time-Series Representation Learning via Temporal and Contextual Contrasting. (2021)

  5. [13]

    Pradeep Hewage, Ardhendu Behera, Marcello Trovati, Ella Pereira, Morteza Ghahremani, Francesco Palmieri, and Yonghuai Liu. 2020. Temporal convolutional neural (TCN) network for an effective weather forecasting using time-series data from the local weather station.Soft Computin...

  6. [14]

    Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. 2022. Semmae: Semantic-guided masking for learning masked autoen- coders.Advances in Neural Information Processing Systems35 (2022), 14290–14302

  7. [15]

    Zhe Li, Zhongwen Rao, Lujia Pan, Pengyun Wang, and Zenglin Xu. 2023. Ti-MAE: Self-Supervised Masked Time Series Autoencoders.arXiv preprint arXiv:2301.08871(2023)

  8. [16]

    Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024. Timer: Generative Pre-trained Transformers Are Large Time Series Models. InForty-first International Conference on Machine Learning

  9. [17]

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. 2022. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers. The Eleventh International Conference on Learning Representations

  10. [18]

    1991.Random variables and stochastic processes

    Athanasios Papoulis. 1991.Random variables and stochastic processes. McGraw Hill

  11. [19]

    Zezhi Shao, Fei Wang, Yongjun Xu, Wei Wei, Chengqing Yu, Zhao Zhang, Di Yao, Guangyin Jin, Xin Cao, Gao Cong, et al. 2023. Exploring progress in multi- variate time series forecasting: Comprehensive benchmarking and heterogeneity analysis.arXiv preprint arXiv:2310.06119(2023)

  12. [20]

    Dalwinder Singh and Birmohan Singh. 2020. Investigating the impact of data normalization on classification performance.Applied Soft Computing97 (2020), 105524

  13. [21]

    2018.Probability and random processes for electrical and computer engineers

    Charles Therrien and Murali Tummala. 2018.Probability and random processes for electrical and computer engineers. CRC press

  14. [22]

    Zhiyuan Wang, Xovee Xu, Weifeng Zhang, Goce Trajcevski, Ting Zhong, and Fan Zhou. 2022. Learning latent seasonal-trend representations for time series forecasting.Advances in Neural Information Processing Systems35 (2022), 38775– 38787

  15. [23]

    Norbert Wiener. 1930. Generalized harmonic analysis.Acta mathematica55, 1 (1930), 117–258

  16. [24]

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. 2021. CoST: Contrastive Learning of Disentangled Seasonal-Trend Representations for Time Series Forecasting. InInternational Conference on Learning Representations

  17. [25]

    Gerald Woo, Chenghao Liu, Doyen Sahoo, Akshat Kumar, and Steven Hoi. 2022. Etsformer: Exponential smoothing transformers for time-series forecasting.arXiv preprint arXiv:2202.01381(2022)

  18. [26]

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. 2022. TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis. InThe Eleventh International Conference on Learning Representations

  19. [27]

    Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. 2021. Autoformer: De- composition transformers with auto-correlation for long-term series forecasting. Advances in Neural Information Processing Systems34 (2021), 22419–22430

  20. [28]

    Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, and Zhendong Niu. 2023. Frequency-domain MLPs are More Effective Learners in Time Series Forecasting. InThirty-seventh Conference on Neural Information Processing Systems

  21. [29]

    Zhihan Yue, Yujing Wang, Juanyong Duan, Tianmeng Yang, Congrui Huang, Yunhai Tong, and Bixiong Xu. 2022. Ts2vec: Towards universal representation of time series. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 8980–8987

  22. [30]

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

  23. [31]

    George Zerveas, Srideepika Jayaraman, Dhaval Patel, Anuradha Bhamidipaty, and Carsten Eickhoff. 2021. A transformer-based framework for multivariate time series representation learning. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2114–2124

  24. [32]

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. 2021. Informer: Beyond efficient transformer for long se- quence time-series forecasting. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 11106–11115

  25. [33]

    Tian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin

  26. [36]

    For supervised forecasting baselines, we select SCNN, Times- Net, and ETSformer, as they demonstrate superior performance

    For comparison, we include SimMTM from MTM, which uses a vanilla Transformer encoder similar to our model, and LaST from contrastive learning, which incorporates seasonal-trend decompo- sition. For supervised forecasting baselines, we select SCNN, Times- Net, and ETSformer, as...

  27. [1990]

    STL: A seasonal-trend decomposition.J. Off. Stat6, 1 (1990), 3–73

  28. [2022]

    InInternational Conference on Machine Learning

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. InInternational Conference on Machine Learning. PMLR, 27268– 27286. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Seo and Lim. A Implementation details A.1 Datasets The detailed description...

Pith tools

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