Pith. sign in

REVIEW 3 major objections 4 minor 44 references

CAMP: A Cycle-Aware Multi-Scale Patch Mixer for Time Series Forecasting

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

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

Pith's one-line read CAMP claims state-of-the-art long-term forecasting by learning each window's own cycles and modeling the leftover at multiple scales.

desk verdict Solid empirical architecture paper with a genuinely new combination and plausible results, but missing code and ambiguous look-back selection keep the SOTA claim conditional. read the letter →

arxiv 2608.04051 v1 pith:Y6ACJH7D submitted 2026-08-04 cs.LG

Jung Min Choi , Vijaya Krishna yalavarthi , Lars Schmidt-Thieme This is my paper · ORCID
classification cs.LG
keywords timeseriesforecastingcyclelearningfastFouriertransformstationarywaveletpatchmixingmulti-scalerepresentationmultivariateadaptiveperiodicity
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

This paper introduces CAMP, a forecasting architecture that separates each input window into a cyclic part and a residual part, learns the dominant periods per window rather than per dataset, and models the leftover dynamics at several time scales. The aim is to show that cycle-aware forecasting can work without a predefined cycle length and still beat strong baselines. On seven long-term benchmarks the paper reports the best average MSE on six datasets and the best or tied-best MAE on six, and on four PEMS traffic benchmarks it reports the highest MSE win count across sixteen settings. Ablations attribute the gains to the combination of adaptive cycle learning, multi-resolution residual modeling, and horizon-guided patch mixing.

What carries the argument

The carrying object is the Adaptive Cycle Learning module, which takes the channel-averaged FFT of each input window, suppresses DC, picks the K largest bins, builds sine/cosine phase features over the lookback and horizon, and maps them through a shared MLP with channel-wise gates to produce historical and future cycle components. Around it, the Horizon-Guided Patch Mixer applies progressively fewer cross-patch updates to patches as they approach the forecast boundary, while the stationary wavelet transform splits the de-cycled residual into length-preserving approximation and detail streams that are patched and encoded with independent MLP mixers. The cycle components are subtracted before the residual backbone and reinjected after it, so the model's inductive bias is that periodic structure can be extrapolated by a frequency-conditioned function while residual structure is captured at multiple resolutions.

What would settle it

Construct a synthetic multivariate series whose dominant period is constant in training and lookback but shifts by a fixed factor within every forecast horizon (e.g., daily seasonality that becomes 12-hour after the boundary). If CAMP's future cycle is extrapolated from the lookback's top-K bins, the reinjected cycle should be misaligned and its MSE should degrade relative to a model that re-estimates phase inside the horizon; measuring that gap on such a controlled dataset would settle whether the stability assumption holds.

Watch

Extended reading notes

Core claim

The central claim is that CAMP establishes a new state of the art for long-term multivariate forecasting on the ETT, Weather, Electricity, and Traffic benchmarks, reaching 0.219 MSE on Weather, 0.154 MSE on Electricity, and 0.309 MSE on ETTh2, while also winning 10 of 16 PEMS settings on MSE. The mechanism is that instead of fixing one cycle per dataset, CAMP selects the top-K FFT frequencies separately for each input window, synthesizes a historical and future cycle with a shared MLP, subtracts the historical cycle, and models the residual through stationary-wavelet streams that are patched and mixed independently. The paper also claims that replacing a fixed recurrent cycle module with this adaptive cycle learning improves an existing cycle model, CycleNet, on three datasets.

Load-bearing premise

The load-bearing premise is that the dominant frequencies identified in each lookback window remain approximately representative of the future horizon, so the same sinusoidal basis evaluated at future times produces a useful cycle.

Editorial extensions

If this is right

  • Forecasting models no longer need a dataset-level cycle length: CAMP learns the dominant periods from each input window, which should help when periodicity changes across lookback lengths or regimes.
  • Combining explicit cycle removal with multi-resolution residual modeling appears complementary: removing ACL, SWT, HGPM, or intra-patch mixing each lowers accuracy on the reported benchmarks.
  • Adaptive cycle learning transfers: plugging ACL into CycleNet improves MSE and MAE on the three datasets tested, suggesting it is a reusable component rather than only a CAMP-specific trick.
  • Short-horizon traffic forecasting benefits strongly: CAMP reports the best MSE and MAE at horizon 12 on all four PEMS datasets and the best MSE across all horizons on PEMS08.

Reading between the lines

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

  • Because ACL extrapolates the same sinusoidal basis into the horizon, the method should be most reliable when the dominant period is stable across the forecast boundary; in series with phase drift or period shifts inside the horizon, the reinjected cycle could actively misalign the forecast.
  • The channel-averaged FFT selection shares one harmonic basis across all variables, so strongly variable-specific cycles (e.g., weekly electricity use for one sensor and daily for another) may be diluted; a per-channel or grouped selection could be a natural extension.
  • The fixed, non-learnable SWT with no inverse transform keeps all streams aligned, which makes the architecture simple, but also means the transform choice matters; testing other undecimated decompositions could reveal whether the multi-resolution alignment or the specific transform drives the gains.
  • One could probe the false-alarm regime: on data with no real periodicity, ACL's top-K selection will still pick the largest noise frequencies, and the zero-initialized gates plus gradient training should learn to suppress them; measuring how quickly the gates shrink on white-noise benchmarks would quantify the adaptive gate's robustness.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes CAMP, a multivariate time-series forecaster that decomposes each input window into a sample-adaptive cyclic component and a residual. The cyclic component is generated by the Adaptive Cycle Learning module, which selects top-K FFT bins per window, builds sinusoidal features over the look-back and forecast horizon, and maps them to channel-wise cycles through an MLP with separate gates. The residual is decomposed by a stationary wavelet transform into aligned multi-scale streams, each patched and encoded by MLP mixers; a Horizon-Guided Patch Mixer preserves recent patches while progressively mixing earlier prefixes. The model is evaluated on seven long-term forecasting benchmarks and four PEMS traffic benchmarks, reporting best average MSE on six long-term datasets and the highest PEMS MSE win count in 10 of 16 settings. Ablation studies show that each module contributes, and an insertion experiment shows that ACL also improves CycleNet.

Significance. If the empirical results are taken at face value, CAMP is a strong new architecture that combines adaptive frequency selection, wavelet residual streams, and position-dependent patch mixing, and its reported gains over strong baselines would be of interest to the forecasting community. The strengths of the evaluation are that results are averaged over three seeds, splits are standard, each module is ablated, and the ACL transfer experiment to CycleNet tests generalizability. I also found no internal circularity: the model is trained on training windows and evaluated on held-out test horizons, and the reported tables match the abstract's summary. The main risk is that ACL's future-cycle extrapolation is built on quantized FFT bins, and the paper does not quantify the resulting phase drift over long horizons or isolate that cost in the ablation, which is directly relevant to the claimed improvements.

major comments (3)
  1. [Adaptive Cycle Learning (equations defining bin selection, phi_k(t), and Cpred)] The selected frequency is quantized to i_k/L, so for any true period that is not an exact integer number of samples, the phase error in Cpred grows linearly with t rather than remaining a constant offset. For example, with L=96 and H=720, a true period of 24.3 samples maps to bin i=4 (period 24), giving roughly 0.37 cycles (about 2.3 radians) of drift over the forecast horizon. Since Cpred is added to the residual forecast in Eq. (12), this drift is directly load-bearing for the claimed gains. The current ablation removes ACL entirely and therefore cannot separate frequency-resolution error from the benefit of cycle modeling. Please add a quantitative evaluation of phase drift, such as a horizon-by-horizon breakdown, and test a variant with continuous-frequency estimation or per-window phase update; if no fix is adopted, the limitation should be stated with evidence in the main text.
  2. [Experiments, Evaluation Protocol] The text states that the look-back length is selected from L in {96,336,512} and that 'the best result is reported' for each model, dataset, and prediction horizon. If 'best result' means selection by test performance rather than validation performance, the reported SOTA numbers in Tables 1 and 2 are not a fair comparison. Please state explicitly that L is selected on the validation split and report the chosen L values for each setting; this is essential for reproducing the average results and for the claim that CAMP often works with shorter look-backs than SRSNet.
  3. [Ablation Study, Table 3] The ablation variants are described as 'independently hyperparameter-tuned,' but the selected hyperparameters are deferred to Supplementary Material A, which is not included in the manuscript. Without knowing the search ranges and selected configurations for each variant, the claim that each module contributes could reflect tuning budget rather than architectural merit. Please make the supplement available or summarize the selected configurations in the paper so the ablation results in Table 3 are verifiable.
minor comments (4)
  1. [Reproducibility] The paper does not provide a code link or release plan; since the central claim is empirical SOTA, making code available would substantially improve reproducibility.
  2. [Supplementary Material] Several tables in Supplementary Material A (Tables A.1-A.7) are referenced but the supplement is not included with the manuscript; please ensure it is accessible to reviewers and readers.
  3. [Table 2 / Related Work] Table 2 lists both 'CycleNet/MLP' and 'CycleNet/Linear' as baselines, but the baselines described in the Experiments section mention only CycleNet; the relationship between these variants should be clarified.
  4. [Conclusion] The conclusion notes higher computational cost, but no runtime, parameter-count, or FLOPs comparison is given; a small complexity table would help readers judge the practical trade-off.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: all central claims are benchmark-tested, and the ACL future cycle is an extrapolation from input-window frequencies rather than a refit of the target.

full rationale

CAMP's central claim is an empirical state-of-the-art result evaluated against external baselines on held-out test splits. The model is trained with MSE loss on training windows, and hyperparameters such as look-back length, patch size, and initial cycle scales are selected using validation performance, so no target-derived constant is renamed as a prediction. The Adaptive Cycle Learning module selects top-K FFT bins from the input window X_in in Eq. (3), defines sinusoidal phases in Eq. (4), evaluates the shared synthesis MLP over both historical and future times in Eq. (6), and reinjects the future cycle in Eq. (12). This is an extrapolation under the paper's explicitly stated assumption that the dominant frequency support remains approximately stable from lookback to horizon; it is not an equivalence to a fitted future quantity, because the future cyclic component is not fitted to the target Y but generated from input-window spectra and a shared MLP trained end-to-end on forecasting loss. The paper's one self-citation (Abdelmalak et al. 2026, which includes three of the present authors) supports the claim that look-back length can affect performance, but the paper independently searches L in {96, 336, 512} for all models and reports results under the same protocol, so the citation is not load-bearing. Acknowledged limitations, such as FFT-based cycle estimation being less reliable for short or weakly periodic inputs, are stated modeling weaknesses rather than circular steps. No equation in the paper reduces to its own input by construction, and no fitted parameter is presented as an independent prediction.

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

CAMP does not introduce new physical or mathematical entities; the components are algorithmic. The central claim depends on a small set of hyperparameters whose values are deferred to a missing supplementary and on the domain assumption that periodicity observed in the past persists into the horizon. The FFT/SWT tools are standard. The most fragile input is the frequency-stability assumption, which the paper states explicitly but does not empirically quantify.

free parameters (6)
  • K (number of dominant frequencies per window)
    Controls the number of sinusoidal components used by Adaptive Cycle Learning; selected via validation (Supplementary A).
  • J (SWT decomposition level)
    Determines the number of aligned residual streams (A_J, D_J, ..., D_1); selected via validation.
  • Patch length P and stride S
    Define temporal granularity of patching; values are deferred to Supplementary A and affect all encoders.
  • Embedding dimension D
    Width of patch projections and MLP mixers; set via validation.
  • Initial global cycle scale gamma and gate logits
    Treated as hyperparameters selected using validation performance (Methodology, Adaptive Cycle Learning).
  • Look-back length L per model/dataset/horizon
    Selected from {96, 336, 512} to report the best result for each configuration; a model-selection choice that affects all comparisons.
assumptions (4)
  • domain assumption The dominant frequency support of an input window remains approximately stable over the prediction horizon.
    ACL extrapolates the cycle from frequencies estimated on the look-back window (Methodology, Adaptive Cycle Learning: 'ACL assumes that the dominant frequency support remains approximately stable from the lookback window to the prediction horizon').
  • domain assumption Circular padding in the stationary wavelet transform preserves informative boundary coefficients.
    SWT is applied with circular padding and no inverse transform; the model relies on these stream coefficients as features (Methodology, Multi-Resolution Residual Decomposition).
  • standard math FFT and SWT are valid signal decompositions for these time series.
    The paper uses FFT for frequency selection and SWT for multi-scale decomposition as standard tools, citing Cochran et al. (1967) and Nason and Silverman (1995).
  • domain assumption Baseline results are comparable because all models search the same look-back lengths and use the same splits.
    The comparison mixes CAMP's three-seed runs with published baselines; without shared code, parity cannot be independently verified (Experiments, Experimental Setup).

how reviews work

0 comments
Cite this review

Pith. "Pith review of CAMP: A Cycle-Aware Multi-Scale Patch Mixer for Time Series Forecasting." pith.science (2026). https://pith.science/paper/Y6ACJH7D

@misc{pith2026260804051,
  author       = {Pith},
  title        = {Pith review of: CAMP: A Cycle-Aware Multi-Scale Patch Mixer for Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y6ACJH7D}},
  note         = {Machine review of arXiv:2608.04051}
}
read the original abstract

Real-world time series are often governed by recurring patterns, but their dominant periods may vary across datasets, forecasting settings, and individual input windows. Existing cycle-aware forecasters commonly rely on a single period selected at the dataset level, which can be restrictive when periodic behavior changes over time or when multiple cycles coexist. Moreover, patch-based models typically process all patch positions uni- formly, although patches farther from the forecast boundary may require broader contextual refinement, while recent patches contain information that should be preserved more directly. Af- ter cyclic behavior is removed, the remaining dynamics may also span multiple temporal resolutions and cannot be adequately de- scribed at a single scale. We introduce CAMP, a Cycle-Aware Multi-Scale Patch Mixer designed to address these challenges. The Adaptive Cycle Learning module identifies dominant fre- quencies separately for each input window and generates both historical and future cyclic components without requiring a pre- defined cycle length. The Horizon-Guided Patch Mixer intro- duces position-dependent refinement, allowing earlier patches to incorporate broader temporal context while preserving infor- mation close to the forecast boundary. CAMP further models the de-cycled residual through temporally aligned multi-resolution representations, enabling complementary dynamics at different scales to be captured within one forecasting framework. Across seven long-term forecasting benchmarks, CAMP achieves the best average MSE on six datasets and the best or tied-best MAE on six. It also obtains the highest MSE win count across sixteen settings on four PEMS traffic benchmarks.

Figures

Figures reproduced from arXiv: 2608.04051 by the authors.

Figure 1
Figure 1. Median dominant period across test input windows [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of CAMP. The RevIN-normalized input is decomposed into a learned cyclic component and a residual. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Architecture of Adaptive Cycle Learning. Per-sample [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Architecture of the Horizon-Guided Patch Mixer. Ear [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 18 canonical work pages

  1. [1]

    Jena Climate Dataset , year =

  2. [2]

    2015 , howpublished =

    Trindade, Artur , title =. 2015 , howpublished =

    2015

  3. [3]

    2024 , howpublished =

    Performance Measurement System (. 2024 , howpublished =

    2024

  4. [4]

    International Conference on Learning Representations , year=

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers , author=. International Conference on Learning Representations , year=

  5. [5]

    International Conference on Learning Representations , year=

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting , author=. International Conference on Learning Representations , year=

  6. [6]

    Hui Chen and Viet Luong and Lopamudra Mukherjee and Vikas Singh , booktitle=. Simple

  7. [7]

    International Conference on Learning Representations , year=

    TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting , author=. International Conference on Learning Representations , year=

  8. [8]

    Si-An Chen and Chun-Liang Li and Sercan O Arik and Nathanael Christian Yoder and Tomas Pfister , journal=

Show all 44 references
  1. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    WaveForM: Graph Enhanced Wavelet Learning for Long Sequence Forecasting of Multivariate Time Series , abstractNote=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2023 , month=

    2023

  2. [10]

    2024 , eprint=

    CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns , author=. 2024 , eprint=

    2024

  3. [11]

    Nason, G. P. and Silverman, B. W. The Stationary Wavelet Transform and some Statistical Applications. Wavelets and Statistics. 1995

    1995

  4. [12]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Are transformers effective for time series forecasting? , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

  5. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    HDMixer: Hierarchical Dependency with Extendable Patch for Multivariate Time Series Forecasting , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2024 , month=. doi:10.1609/aaai.v38i11.29155 , abstractNote=

    2024 doi

  6. [14]

    2024 , eprint=

    WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting , author=. 2024 , eprint=

    2024

  7. [15]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

    Tsmixer: Lightweight mlp-mixer model for multivariate time series forecasting , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

  8. [16]

    International Conference on Learning Representations , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations , year=

  9. [17]

    Proceedings of the IEEE , volume=

    What is the fast Fourier transform? , author=. Proceedings of the IEEE , volume=. 1967 , publisher=

    1967

  10. [18]

    Advances in Neural Information Processing Systems , year=

    Frequency-domain MLPs are more effective learners in time series forecasting , author=. Advances in Neural Information Processing Systems , year=

  11. [19]

    SIAM review , volume=

    Continuous and discrete wavelet transforms , author=. SIAM review , volume=. 1989 , publisher=

    1989

  12. [20]

    International Conference on Learning Representations , year=

    Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting , author=. International Conference on Learning Representations , year=

  13. [21]

    CoRR , volume =

    Haixu Wu and Jiehui Xu and Jianmin Wang and Mingsheng Long , title =. CoRR , volume =. 2021 , url =. 2106.13008 , timestamp =

    2021 arXiv

  14. [22]

    CoRR , volume =

    Zonghan Wu and Shirui Pan and Guodong Long and Jing Jiang and Xiaojun Chang and Chengqi Zhang , title =. CoRR , volume =. 2020 , url =. 2005.11650 , timestamp =

    2020 arXiv

  15. [23]

    CoRR , volume =

    Haoyi Zhou and Shanghang Zhang and Jieqi Peng and Shuai Zhang and Jianxin Li and Hui Xiong and Wancai Zhang , title =. CoRR , volume =. 2020 , url =. 2012.07436 , timestamp =

    2020 arXiv

  16. [24]

    2025 , eprint=

    Enhancing Time Series Forecasting through Selective Representation Spaces: A Patch Perspective , author=. 2025 , eprint=

    2025

  17. [25]

    2025 , eprint=

    TimeKAN: KAN-based Frequency Decomposition Learning Architecture for Long-term Time Series Forecasting , author=. 2025 , eprint=

    2025

  18. [26]

    2025 , eprint=

    Amplifier: Bringing Attention to Neglected Low-Energy Components in Time Series Forecasting , author=. 2025 , eprint=

    2025

  19. [27]

    2022 , eprint=

    N-HiTS: Neural Hierarchical Interpolation for Time Series Forecasting , author=. 2022 , eprint=

    2022

  20. [28]

    2026 , eprint=

    FreqCycle: A Multi-Scale Time-Frequency Analysis Method for Time Series Forecasting , author=. 2026 , eprint=

    2026

  21. [29]

    arXiv preprint arXiv:2203.07681 , year=

    DEPTS: Deep expansion learning for periodic time series forecasting , author=. arXiv preprint arXiv:2203.07681 , year=

    arXiv

  22. [30]

    arXiv preprint arXiv:2405.00946 , year=

    Sparsetsf: Modeling long-term time series forecasting with 1k parameters , author=. arXiv preprint arXiv:2405.00946 , year=

    arXiv

  23. [31]

    International Conference on Learning Representations , year=

    Reversible instance normalization for accurate time-series forecasting against distribution shift , author=. International Conference on Learning Representations , year=

  24. [32]

    Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=

    Channel Dependence, Limited Lookback Windows, and the Simplicity of Datasets: How Biased is Time Series Forecasting? , author=. Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=. 2026 , organization=

    2026

  25. [33]

    Advances in Neural Information Processing Systems , volume=

    Scinet: Time series modeling and forecasting with sample convolution and interaction , author=. Advances in Neural Information Processing Systems , volume=

  26. [34]

    2023 , eprint=

    TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis , author=. 2023 , eprint=

    2023

  27. [35]

    2023 , eprint=

    Koopa: Learning Non-stationary Time Series Dynamics with Koopman Predictors , author=. 2023 , eprint=

    2023

  28. [36]

    2022 , eprint=

    FEDformer: Frequency Enhanced Decomposed Transformer for Long-term Series Forecasting , author=. 2022 , eprint=

    2022

  29. [37]

    2022 , eprint=

    FiLM: Frequency improved Legendre Memory Model for Long-term Time Series Forecasting , author=. 2022 , eprint=

    2022

  30. [38]

    2026 , howpublished =

    2026

  31. [39]

    Advances in Neural Information Processing Systems , volume=

    Basisformer: Attention-based time series forecasting with learnable and interpretable basis , author=. Advances in Neural Information Processing Systems , volume=

  32. [40]

    2024 , eprint=

    SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion , author=. 2024 , eprint=

    2024

  33. [41]

    2024 , eprint=

    From Similarity to Superiority: Channel Clustering for Time Series Forecasting , author=. 2024 , eprint=

    2024

  34. [42]

    2023 , eprint=

    FourierGNN: Rethinking Multivariate Time Series Forecasting from a Pure Graph Perspective , author=. 2023 , eprint=

    2023

  35. [43]

    2019 , eprint=

    Optuna: A Next-generation Hyperparameter Optimization Framework , author=. 2019 , eprint=

    2019

  36. [44]

    Proceedings of the 9th Machine Learning for Healthcare Conference , series =

    Early Prediction of Causes (not Effects) in Healthcare by Long-Term Clinical Time Series Forecasting , author =. Proceedings of the 9th Machine Learning for Healthcare Conference , series =. 2024 , publisher =

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.