Pith. sign in

REVIEW 5 major objections 5 minor 27 references

Accurate Parameter-Efficient Test-Time Adaptation for Time Series Forecasting

T0 review · 5 major / 5 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 PETSA is a parameter-efficient test-time adaptation method for time series forecasting that updates only small gated low-rank calibration modules on the frozen model's input and output, and on six benchmarks it matches or improves the MSE…

desk verdict PETSA's efficiency story is plausible, but Table 1 is visibly corrupted (Weather rows duplicate Exchange rows; ETTm1 averages don't match), so the headline 127-vs-88 win count is unsupported until the data is fixed. read the letter →

arxiv 2506.23424 v1 pith:HZR37Z7U submitted 2025-06-29 cs.LG cs.AI

Heitor R. Medeiros , Hossein Sharifi-Noghabi , Gabriel L. Oliveira , Saghar Irandoust This is my paper
classification cs.LGcs.AI
keywords test-timeadaptationtimeseriesforecastingparameter-efficientfine-tuninglow-rankdynamicgatingfrequency-domainlosspatch-wisestructuralnon-stationary
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 proposes PETSA, a way to adapt a pre-trained time-series forecaster to distribution shift at inference time without touching the frozen model's weights. Instead, PETSA inserts two light calibration modules, one on the input and one on the output, built from low-rank adapters and input-dependent gating, and updates only those during test time. The claim is that this parameter-efficient scheme is not merely cheaper but actually competitive or better: across six datasets and six forecasting backbones, PETSA wins 127 best-MSE cells versus 88 for the full-model baseline TAFAS, with up to 33.6x fewer trainable parameters. The accuracy is supported by a three-part loss that combines Huber, frequency-domain, and patch-wise structural terms to compensate for the reduced adaptation capacity. If right, this means accurate test-time adaptation for forecasting does not require updating full models.

What carries the argument

The load-bearing mechanism is the dynamic gated low-rank calibration module applied at both the input and output of a frozen forecaster. A learnable per-variable gate $\alpha$ modulates the signal through a tanh before a low-rank projection $W = A B$ of rank $r$, plus a bias, giving the module input-conditioned behavior with very few trainable parameters. The second pillar is the composite PETSA loss: a Huber term for outlier robustness, an $\ell^1$ loss on the FFT spectra of predictions versus ground truth to preserve periodicity, and a patch-wise structural loss on local correlation, mean, and variance for structural alignment. Together these allow the small calibration modules to compensate for the limited adaptation capacity of a low-rank update.

What would settle it

Run PETSA on a benchmark where the partial-ground-truth window is withheld, so labels arrive only after the full forecast horizon; then the $L_{\text{pt}}$ term vanishes and PETSA can only update on delayed full labels. If its reported MSE advantage over no adaptation disappears in that setting, the improvement depends on the label assumption rather than on the gated low-rank mechanism, and the experiment is directly runnable on the released code by setting the partial-label window size to zero.

Watch

Extended reading notes

Core claim

PETSA's central claim is that test-time adaptation of time-series forecasters can be made parameter-efficient without sacrificing accuracy. It calibrates the input $X_{t^*}$ and the output $\hat{Y}_{t^*}$ of a frozen forecaster with gated low-rank transformations $\hat{X}^{\text{cali}}_{t^*} = X_{t^*} + (\tanh(\alpha \odot X_{t^*}) W + b)$ and similarly for the output, where $W = A B$ is a low-rank weight matrix and $\alpha$ is a learnable per-variable gate. Only these calibration parameters are updated during inference, using a loss $L_{\text{PETSA}} = L_T + L_{\text{pt}}$ that combines a Huber term, an FFT frequency-domain term, and a patch-wise structural term, computed against delayed full and partial ground truth. In experiments, PETSA achieves the best MSE in 127 of the compared cells versus 88 for TAFAS, and it uses far fewer parameters, for example 33.6x fewer on ETTh1 with iTransformer at horizon 720.

Load-bearing premise

PETSA's loss requires ground-truth labels: partial ground truth must arrive shortly after each prediction and full ground truth after the forecast window, so the method is supervised test-time fine-tuning rather than adaptation to an unlabeled stream.

Editorial extensions

If this is right

  • PETSA can be dropped into existing forecasting pipelines without retraining or unfreezing the backbone: only the two calibration modules are optimized at test time.
  • The parameter savings grow with the forecast horizon: at window 720 on ETTh1 with iTransformer, PETSA uses 33.6x fewer trainable parameters than TAFAS while keeping MSE lower.
  • The method is architecture-agnostic in the tested range: it improves transformer-, linear-, and MLP-based forecasters including iTransformer, PatchTST, DLinear, OLS, FreTS, and MICN.
  • Because the calibration modules are small, the per-step memory and compute overhead of adaptation is reduced, making online adaptation more feasible on resource-constrained devices.
  • Ablations on ETTh1 with OLS indicate that the combined three-part loss is needed: MSE loss alone or Huber loss alone underperform the full PETSA loss.

Reading between the lines

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

  • The input- and output-side gating suggests the method could be applied layer-wise inside a deep forecaster rather than only at the boundaries, but the paper does not test this variant.
  • Because the loss requires labels, a natural extension is to pair PETSA's gated calibration with an unsupervised signal such as entropy or consistency on the forecast window for settings where partial ground truth is unavailable; the paper does not address that case.
  • The gating initialization $\alpha$ is a sensitive hyperparameter in the reported ablations, so a practical deployment would likely need per-dataset tuning or an automatic schedule, which the paper recommends but does not provide.
  • The frequency-domain loss is best with $\beta = 0$ on some datasets, suggesting the three-term loss is not universally beneficial and that the weighting should be tuned per model and dataset, as the paper itself notes.
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

5 major / 5 minor

Summary. The paper proposes PETSA, a parameter-efficient test-time adaptation method for time-series forecasting. It wraps a frozen pre-trained forecaster with input and output calibration modules that use low-rank adapters and dynamic gating, and updates only these modules during test time using a loss that combines Huber, frequency-domain, and patch-wise structural terms. The paper reports MSE results across six datasets and six forecasting backbones, claiming that PETSA achieves competitive or better accuracy than the full-model TTA baseline TAFAS while updating far fewer parameters.

Significance. If the empirical claims held, the paper would make a useful contribution: parameter counts in Figures 4-10 indicate large reductions (up to 33.6x in one setting), and the calibration-module design is simple and applicable to arbitrary forecasters. The paper is also transparent about building on TAFAS and provides a code link. However, the central quantitative claim rests on a single table (Table 1) that shows signs of corruption, and the comparison lacks any uncertainty quantification. As presented, the headline 127-versus-88 win count and the 'across six datasets' statement are not currently supported.

major comments (5)
  1. [Table 1] The Weather block of Table 1 reproduces the Exchange block with each row shifted by one forecast horizon. For example, the Weather-96 row (0.366/0.345/0.341/0.372/0.353/0.367/0.372/0.359/0.357/0.353/0.294/0.286/0.354/0.335/0.327/0.558/0.307/0.357) is identical to the Exchange-192 row, Weather-192 matches Exchange-336, Weather-336 matches Exchange-720, and Weather-720 matches Exchange-Avg. The ETTm1 'Avg' row (0.257/0.255/...) is also not the mean of the four horizon rows for that dataset, which average approximately 0.51. Since Table 1 is the sole source of the headline '127 wins versus 88' claim, the win counts and the 'six datasets' statement must be recomputed after correcting this table; if the Weather rows are not genuine, the comparison effectively covers five datasets.
  2. [Section 4.2, Table 1] Table 1 reports a single MSE value per cell with no standard deviations, number of seeds, or significance tests. Many of the differences between PETSA and TAFAS are below 0.005 (e.g., ETTh1 window 96, iTransformer: 0.432 vs 0.435; PatchTST: 0.426 vs 0.426). At this granularity, the 127-versus-88 win tally could be dominated by noise. The authors should provide per-seed results, error bars, and paired statistical tests before the superiority claim is accepted.
  3. [Section 3.1, Eqs. (5)-(6)] The method is not unsupervised test-time adaptation. Section 3.1 states that PETSA 'adopts this setup, using both partial and full labels,' and the loss terms in Eqs. (5)-(6) are computed directly against ground-truth Y. This is supervised test-time fine-tuning with delayed labels, and it is inapplicable in settings where future values are never observable. The abstract and introduction should state this assumption prominently, since the current framing overstates the applicability of the method.
  4. [Appendix, 'Ablation on Loss Components'] The reported PETSA results appear to be selected after tuning the frequency-loss weight beta (with beta=0.0 best for ETTh1/OLS, beta=0.1 best for FreTS, and other values recommended for other datasets), as well as the low-rank rank and gating initialization via ablations. The appendix concludes that 'we recommend hyperparameter tuning for optimal performance.' Without a fixed protocol or a sensitivity analysis reported in the main text, the advantage over TAFAS may be a tuning artifact rather than an intrinsic property of the method. The chosen hyperparameters for every dataset/backbone combination should be reported.
  5. [Section 4.2 and Figures 4-10] The 'fewer parameters than baselines' claim is supported only by Figures 4-10, which mostly cover the OLS backbone (Figure 4 covers iTransformer on ETTh1 only). Table 1 reports no parameter counts for the 36 dataset-by-backbone combinations, and no aggregated parameter-efficiency table is given. To substantiate the claim that PETSA uses fewer parameters 'across all models,' the authors should provide parameter counts for all experimental conditions, not just for a subset.
minor comments (5)
  1. [Eq. (1)] The shapes in Eq. (1) are dimensionally inconsistent as written: X_t* is in R^{B x L x V}, alpha is in R^V, and W = A.B with A in R^{L x r} and B in R^{r x L x V}; the expression tanh(alpha * X_t*) . W + b needs a precise contraction rule over variables and sequence positions.
  2. [Figure 11 caption] The caption says 'windows 96, 196, 336, and 720'; the second window should be 192.
  3. [Appendix, first paragraph] The text says 'we provide additional ablations in the next session'; this should read 'next section.'
  4. [Table 1 caption] The caption uses 'PETSA - PT' and marks columns with 'PT', which is confusing because 'PT' is also used for 'partial ground truth' in Figure 1 and Section 3.1. Please use a distinct notation for the PETSA columns.
  5. [Section 4.1] The sentence 'we used PyTorch for PETSA implementation, and training/adapt the models using one NVIDIA A100' has a grammar error; also, exact optimizer settings, learning rates, and batch sizes are not reported in the main text or appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PETSA's reported gains are empirical and not derived from its inputs by construction.

full rationale

The paper's central claim is that PETSA improves or maintains forecasting MSE relative to TAFAS while updating fewer parameters. This is an empirical claim supported by Table 1 and Figure 4. The derivation chain is self-contained: PETSA defines a composite adaptation loss (Huber, frequency-domain, and patch-wise structural terms) and updates only lightweight calibration modules on test-time labels. The reported evaluation metric is MSE, which is not identical to the adaptation loss by construction—Huber loss differs for large errors, and the frequency and structural terms are distinct objectives. Moreover, evaluation covers the full forecast window, including portions not used for the online adaptation update, so the reported MSE is not the optimized objective itself. No equation in the paper makes a claimed prediction equal to an input or fitted parameter. There are no self-citations: all cited prior works (TAFAS, LoRA, DyT, FreDF, patch-wise structural loss) are external, and none is invoked as a load-bearing uniqueness theorem. The manuscript does contain a legitimate correctness concern—Table 1 appears to duplicate Weather rows and has an implausible ETTm1 average, and hyperparameters are tuned via ablations on the test sets—but data corruption and test-set tuning are not circularity. They affect the reliability of the empirical comparison, not the logical structure of the derivation. Therefore, no circular step is present, and the paper deserves a low circularity score.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

PETSA introduces no new physical entities or latent variables. Its central claim depends on two domain assumptions inherited from TAFAS: availability of partial and delayed labels at test time, and the use of an FFT-derived dominant period to define adaptation windows. The method also has several tunable hyperparameters (rank, beta, gating init, delta) whose default values are not fully specified in the main text.

free parameters (4)
  • low-rank rank r = ablated 8 to 128; default not stated in text
    Controls the number of trainable parameters in the input and output calibration modules; the paper ablates it but does not state a single default value used in Table 1.
  • frequency loss weight beta = best per dataset/model; e.g., 0.1 for FreTS, 0.0 for ETTh1 OLS
    Weights the frequency-domain loss; the ablation shows the best value varies by dataset and model, so Table 1 likely uses tuned betas.
  • gating initialization alpha_init = ablated 0.0 to 1.0; default not stated
    Initial value of the per-variable gating parameter alpha, which affects the sign and magnitude of the low-rank update.
  • Huber delta = 0.5
    Fixed at 0.5, controls robustness to outliers.
assumptions (3)
  • domain assumption Partial ground truth (PT) and delayed full ground truth (T) become available at test time for each look-back window.
    Adopted from TAFAS (Kim et al. 2025), this label availability is required to compute L_pt and L_T; without it the adaptation loss cannot be evaluated.
  • domain assumption The dominant period estimated by FFT defines the adaptation window.
    In Section 3.1, PETSA 'adopts this setup' from TAFAS, so the adaptation interval depends on a frequency-domain estimate of periodicity.
  • standard math The tensor contraction in Eq. 1 is well-defined for the stated shapes (B,L,V) and (L,L,V).
    The paper does not specify how tanh(alpha * X) (B,L,V) is contracted with W (L,L,V); it assumes a standard Einstein-style contraction over the look-back dimension.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accurate Parameter-Efficient Test-Time Adaptation for Time Series Forecasting." pith.science (2026). https://pith.science/paper/HZR37Z7U

@misc{pith2026250623424,
  author       = {Pith},
  title        = {Pith review of: Accurate Parameter-Efficient Test-Time Adaptation for Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZR37Z7U}},
  note         = {Machine review of arXiv:2506.23424}
}
read the original abstract

Real-world time series often exhibit a non-stationary nature, degrading the performance of pre-trained forecasting models. Test-Time Adaptation (TTA) addresses this by adjusting models during inference, but existing methods typically update the full model, increasing memory and compute costs. We propose PETSA, a parameter-efficient method that adapts forecasters at test time by only updating small calibration modules on the input and output. PETSA uses low-rank adapters and dynamic gating to adjust representations without retraining. To maintain accuracy despite limited adaptation capacity, we introduce a specialized loss combining three components: (1) a robust term, (2) a frequency-domain term to preserve periodicity, and (3) a patch-wise structural term for structural alignment. PETSA improves the adaptability of various forecasting backbones while requiring fewer parameters than baselines. Experimental results on benchmark datasets show that PETSA achieves competitive or better performance across all horizons. Our code is available at: https://github.com/BorealisAI/PETSA

Figures

Figures reproduced from arXiv: 2506.23424 by the authors.

Figure 1
Figure 1. Illustration of the test-time adaptation setup in PETSA. The model observes a look-back window and makes predictions over the forecast window. A partial portion of the ground truth (PT) becomes available shortly after prediction (light yellow), which is used to adapt the model online. Full ground truth (T) may also be observed after the forecast window completes (shaded yellow). PETSA uses both partial and delayed T… view at source ↗
Figure 2
Figure 2. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Comparison of PETSA and TAFAS on ETTh1 for iTransformer. Top: MSE across different window sizes with no adaptation, TAFAS, and PETSA. Bottom: Number of trainable parameters used for adaptation. PETSA achieves similar or better accuracy while using up to 33.6× fewer parameters at window size 720. Memory usage is annotated in MB. 4.2. Results In [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (10 more)
Figure 3
Figure 3. Figure 3: Total number of best-value wins grouped by window length for TAFAS and PETSA approaches. PETSA consistently outperforms TAFAS across all horizons. 4. Experiments 4.1. Experimental Protocol (a) Datasets: We demonstrate the effectiveness of our method, PETSA, using widel…
Figure 5
Figure 5. Figure 5: Comparison of PETSA and TAFAS on ETTh1 for OLS. Left: MSE across different window sizes with TAFAS, and PETSA. Right: Number of trainable parameters used for adaptation. Memory usage is annotated in MB. MSE vs. Window Size (ETTm1 - OLS) Parameters vs. Window Size (ETTm…
Figure 6
Figure 6. Figure 6: Comparison of PETSA and TAFAS on ETTm1 for OLS. Left: MSE across different window sizes with TAFAS, and PETSA. Right: Number of trainable parameters used for adaptation. Memory usage is annotated in MB. Ablation on Dynamic Gating parameter In this ablation, [PITH_FULL…
Figure 7
Figure 7. Figure 7: Comparison of PETSA and TAFAS on ETTh2 for OLS. Left: MSE across different window sizes with TAFAS, and PETSA. Right: Number of trainable parameters used for adaptation. Memory usage is annotated in MB. MSE vs. Window Size (ETTm2 - OLS) Parameters vs. Window Size (ETTm…
Figure 8
Figure 8. Figure 8: Comparison of PETSA and TAFAS on ETTm2 for OLS. Left: MSE across different window sizes with TAFAS, and PETSA. Right: Number of trainable parameters used for adaptation. Memory usage is annotated in MB. MSE vs. Window Size (exchange_rate - OLS) Parameters vs. Window Si…
Figure 9
Figure 9. Figure 9: Comparison of PETSA and TAFAS on Exchange Rate for OLS. Left: MSE across different window sizes with TAFAS, and PETSA. Right: Number of trainable parameters used for adaptation. Memory usage is annotated in MB. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Comparison of PETSA and TAFAS on Weather for OLS. Left: MSE across different window sizes with TAFAS, and PETSA. Right: Number of trainable parameters used for adaptation. Memory usage is annotated in MB. 8 16 32 64 128 Rank 0.440 0.442 0.444 0.446 0.448 0.450 Test MS…
Figure 11
Figure 11. Figure 11: Comparison of the original model, TAFAS, and PETSA on ETTh1 for OLS. MSE across different ranks for windows 96, 196, 336, and 720. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_11.png]
Figure 12
Figure 12. Figure 12: Comparison of the original model, TAFAS, and PETSA on ETTh1 for OLS. MSE across different gating initial values for windows 96, 196, 336, and 720. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: Comparison of different loss terms in PETSA on ETTh1 for OLS. MSE across different beta values for windows 96, 196, 336, and 720. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Parameter-free online test-time adaptation

    Boudiaf, M., Mueller, R., Ben Ayed, I., and Bertinetto, L. Parameter-free online test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 8344--8353, 2022

    work page 2022

  3. [3]

    Adaptformer: Adapting vision transformers for scalable visual recognition

    Chen, S., Ge, C., Tong, Z., Wang, J., Song, Y., Wang, J., and Luo, P. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35: 0 16664--16678, 2022

    work page 2022

  4. [4]

    Low-rank adaptation of time series foundational models for out-of-domain modality forecasting

    Gupta, D., Bhatti, A., Parmar, S., Dan, C., Liu, Y., Shen, B., and Lee, S. Low-rank adaptation of time series foundational models for out-of-domain modality forecasting. In Proceedings of the 26th International Conference on Multimodal Interaction, pp.\ 382--386, 2024

    work page 2024

  5. [5]

    J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al

    Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

    2022

  6. [6]

    Huber, P. J. Robust estimation of a location parameter. In Breakthroughs in statistics: Methodology and distribution, pp.\ 492--518. Springer, 1992

    work page 1992

  7. [7]

    Visual prompt tuning

    Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., and Lim, S.-N. Visual prompt tuning. In European conference on computer vision, pp.\ 709--727. Springer, 2022

    work page 2022

  8. [8]

    When model meets new normals: test-time adaptation for unsupervised time-series anomaly detection

    Kim, D., Park, S., and Choo, J. When model meets new normals: test-time adaptation for unsupervised time-series anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pp.\ 13113--13121, 2024

    work page 2024

Show all 27 references
  1. [9]

    Battling the non-stationarity in time series forecasting via test-time adaptation

    Kim, H., Kim, S., Mok, J., and Yoon, S. Battling the non-stationarity in time series forecasting via test-time adaptation. arXiv preprint arXiv:2501.04970, 2025

    2025 arXiv

  2. [10]

    Patch-wise structural loss for time series forecasting

    Kudrat, D., Xie, Z., Sun, Y., Jia, T., and Hu, Q. Patch-wise structural loss for time series forecasting. arXiv preprint arXiv:2503.00877, 2025

    2025 arXiv

  3. [11]

    A comprehensive survey on test-time adaptation under distribution shifts

    Liang, J., He, R., and Tan, T. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, 133 0 (1): 0 31--64, 2025

    2025

  4. [12]

    F., Cheng, K.-T., and Chen, M.-H

    Liu, S.-Y., Wang, C.-Y., Yin, H., Molchanov, P., Wang, Y.-C. F., Cheng, K.-T., and Chen, M.-H. Dora: Weight-decomposed low-rank adaptation. In Forty-first International Conference on Machine Learning, 2024 a

    2024

  5. [13]

    itransformer: Inverted transformers are effective for time series forecasting

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., and Long, M. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024 b

    2024

  6. [14]

    Channel-aware low-rank adaptation in time series forecasting

    Nie, T., Mei, Y., Qin, G., Sun, J., and Ma, W. Channel-aware low-rank adaptation in time series forecasting. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pp.\ 3959--3963, 2024

    2024

  7. [15]

    H., Sinthong, P., and Kalagnanam, J

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

    2023

  8. [16]

    Low-rank adaptation for spatio-temporal forecasting

    Ruan, W., Chen, W., Dang, X., Zhou, J., Li, W., Liu, X., and Liang, Y. Low-rank adaptation for spatio-temporal forecasting. arXiv preprint arXiv:2404.07919, 2024

    2024 arXiv

  9. [17]

    Scaleformer: Iterative multi-scale refining transformers for time series forecasting

    Shabani, A., Abdi, A., Meng, L., and Sylvain, T. Scaleformer: Iterative multi-scale refining transformers for time series forecasting. arXiv preprint arXiv:2206.04038, 2022

    2022 arXiv

  10. [18]

    and Darlow, L

    Toner, W. and Darlow, L. N. An analysis of linear time series forecasting models. In International Conference on Machine Learning, pp.\ 48404--48427. PMLR, 2024

    2024

  11. [19]

    Tent: Fully test-time adaptation by entropy minimization

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., and Darrell, T. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020

    2006 arXiv

  12. [20]

    Micn: Multi-scale local and global context modeling for long-term series forecasting

    Wang, H., Peng, J., Huang, F., Wang, J., Chen, J., and Xiao, Y. Micn: Multi-scale local and global context modeling for long-term series forecasting. In The eleventh international conference on learning representations, 2023

    2023

  13. [21]

    Fredf: Learning to forecast in the frequency domain

    Wang, H., Pan, L., Chen, Z., Yang, D., Zhang, S., Yang, Y., Liu, X., Li, H., and Tao, D. Fredf: Learning to forecast in the frequency domain. In ICLR, 2025

    2025

  14. [22]

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

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

    2021

  15. [23]

    Frequency-domain mlps are more effective learners in time series forecasting

    Yi, K., Zhang, Q., Fan, W., Wang, S., Wang, P., He, H., An, N., Lian, D., Cao, L., and Niu, Z. Frequency-domain mlps are more effective learners in time series forecasting. Advances in Neural Information Processing Systems, 36: 0 76656--76679, 2023

    2023

  16. [24]

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

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

    2023

  17. [25]

    On pitfalls of test-time adaptation

    Zhao, H., Liu, Y., Alahi, A., and Lin, T. On pitfalls of test-time adaptation. In Proceedings of the 40th International Conference on Machine Learning, pp.\ 42058--42080, 2023

    2023

  18. [26]

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

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

    2021

  19. [27]

    Transformers without normalization

    Zhu, J., Chen, X., He, K., LeCun, Y., and Liu, Z. Transformers without normalization. arXiv preprint arXiv:2503.10622, 2025

    2025 arXiv

Pith tools

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