Pith. sign in

REVIEW 4 major objections 6 minor 22 references

KARMA: A Multilevel Decomposition Hybrid Mamba Framework for Multivariate Long-Term Time Series Forecasting

T0 review · 4 major / 6 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 KARMA beats seven baselines in long-term forecasting.

desk verdict KARMA is a new and plausible architecture, but the 'significantly outperforms' claim is not supported by the reported evidence. read the letter →

arxiv 2506.08939 v1 pith:FVLYPYU5 submitted 2025-06-10 cs.LG

Hang Ye , Gaoxiang Duan , Haoran Zeng , Yangxin Zhu , Lingxue Meng , Xiaoying Zheng , Yongxin Zhu This is my paper
classification cs.LG
keywords multivariatetimeseriesforecastinglong-termMambastatespacemodelsdecompositionwavelettransformmulti-headattentionfrequency-domainloss
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 sets out to establish that long multivariate time series are best forecast by repeatedly decomposing them into interpretable components rather than feeding raw sequences to a single network. It argues that an attention-based module can adaptively split each series into trend and seasonal parts, and that a wavelet-based module can then split the seasonal part into frequency and time views. Each component is processed by a small Mamba state-space block, keeping the whole pipeline linear in sequence length. On eight real-world datasets across electricity, transport, weather, and finance, KARMA reports lower average error than seven Transformer-, Mamba-, MLP-, and CNN-based baselines on most benchmarks, along with faster training on long lookback windows. If correct, the result suggests that learned multilevel decomposition is not a preprocessing detail but a central lever for accurate and efficient long-horizon forecasting.

What carries the argument

The load-bearing mechanism is the pairing of ATCD with HFTD inside KarmaBlocks. ATCD learns a trend as $\mathrm{SiLU}(\mathrm{MHA}(\tilde{X}_{in}))$ and a seasonal component as the residual, which is the identity $\tilde{X}_s = \tilde{X}_{in} - \tilde{X}_t$. HFTD then decomposes the seasonal stream with a one-level discrete wavelet transform, producing high-frequency and low-frequency coefficient sequences plus a normalized temporal stream and its channel-flipped augmentation; these are processed by separate Mamba state-space layers and recombined by inverse wavelet synthesis. The framework also mixes a time-domain MSE loss with a frequency-domain $\ell^1$ loss, weighted by $\alpha = 0.2$.

What would settle it

Build a synthetic series with a known additive trend and seasonality, run ATCD on it, and compare the extracted trend to the true trend; if the correlation is low, or if replacing ATCD with a fixed STL split matches KARMA's accuracy, the claimed adaptive decomposition is not doing the causal work.

Watch

Extended reading notes

Core claim

On its own terms, KARMA's central claim is that fixed decomposition rules are a bottleneck and that both the trend-season split and the frequency-time split should be learned and adapted to each dataset. The trend branch is taken to be $\mathrm{SiLU}(\mathrm{MHA}(\tilde{X}_{in}))$, and seasonality is defined as the residual $\tilde{X}_s = \tilde{X}_{in} - \tilde{X}_t$. The seasonal embedding then passes through HFTD, which applies a one-level discrete wavelet transform to produce low-frequency and high-frequency coefficient sequences, alongside an RMS-normed temporal stream and a channel-flipped copy of it. Three Mamba variants process these streams inside KarmaBlocks, the inverse wavelet transform recombines them, and the trend stream is handled by a Global Mamba. The paper reports that the resulting system beats all seven baselines on most of the eight datasets, with the largest gains on strongly periodic series, and that training time grows roughly linearly with sequence length.

Load-bearing premise

The entire framework rests on the premise that the output of multi-head attention followed by SiLU actually is the trend of the series, so that subtracting it leaves a meaningful seasonal component; the paper offers no direct check of that decomposition, only final accuracy.

Editorial extensions

If this is right

  • Multivariate long-term forecasting can be done with linear-time Mamba blocks if decomposition is adaptive rather than fixed.
  • Datasets with strong periodic structure should see the largest gains, because the ablation shows HFTD is the decisive module on the highly periodic ECL dataset.
  • Replacing fixed STL decomposition or a Transformer-based temporal block with learned decomposition and Mamba blocks improves accuracy, as the replacement ablations report.
  • Training time should scale roughly linearly with sequence length, in contrast to quadratic Transformer attention, making very long lookback windows practical.
  • Adding a frequency-domain loss term alongside MSE can improve overall forecasting fit on top of the architectural changes.

Reading between the lines

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

  • The ATCD trend-as-attention assumption could be tested directly: on synthetic series with known additive trend and seasonality, one could measure whether $\mathrm{SiLU}(\mathrm{MHA}(\cdot))$ actually recovers the injected trend; if it does not, the benchmark gains may come from model capacity rather than from the decomposition semantics.
  • If the seasonal residual is misspecified, the wavelet branch may inject spurious high-frequency content; the paper itself notes an accuracy drop on the high-dimensional Traffic dataset, which is consistent with that risk.
  • The channel-flip augmentation used in the temporal branch is a cheap idea that could transfer to other Mamba and state-space forecasters as a regularizer.
  • Because the loss mixes time and frequency objectives with a fixed weight, a natural testable extension is to vary $\alpha$ per dataset and check whether the frequency term alone explains most of the gain on periodic benchmark series.
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 / 6 minor

Summary. The paper proposes KARMA, a Mamba-based multivariate long-term forecasting architecture that combines an Adaptive Time Channel Decomposition module (ATCD), a Hybrid Frequency-Time Decomposition module (HFTD), and stacked multi-scale KarmaBlocks. ATCD uses multi-head attention followed by SiLU to split the normalized input into trend and seasonal components (Eqs. 4-6), while HFTD applies a one-level discrete wavelet transform plus temporal residual branches to the seasonal component (Eqs. 7-10). KarmaBlocks process the decomposed frequency and time components with three Mamba variants (Eqs. 11-14), and the loss in Eq. (2) combines a time-domain MSE term with a frequency-domain L1 term. Experiments on eight datasets claim state-of-the-art accuracy and efficiency against seven baselines, with ablations on Weather, ECL, and ETTm1.

Significance. If the empirical claims held, KARMA would be a useful contribution: a linear-complexity architecture with dynamic decomposition, offering a credible alternative to Transformer, Mamba, and MLP baselines for long-term forecasting. The modular design is clearly described, and the manuscript appropriately points to a code repository with full results, which supports reproducibility. However, the significance as presented is undercut by the lack of uncertainty quantification, by ablation outcomes that do not confirm the contribution of ATCD, and by results where KARMA is not the best model. The architecture is not shown to be state of the art on all datasets, and the reported margins over the strongest baseline are often small enough to be within run-to-run variation.

major comments (4)
  1. [Abstract and Section 4.2, Table 2] The central claim that KARMA significantly outperforms mainstream baselines is not supported by the evidence in Table 2. On Traffic, SMamba achieves a lower MSE than KARMA (0.414 vs. 0.453), and on Exchange, DLinear achieves a lower MSE (0.354 vs. 0.360). On the remaining reported datasets, the KARMA advantage over the strongest baseline is tiny: 0.168 vs. 0.170 on ECL and 0.250 vs. 0.252 on Weather. These margins are smaller than the effect of removing ATCD in Table 3, where ECL MSE improves from 0.168 to 0.163. Because no standard deviations, repeated-seed runs, or significance tests are reported, the claimed ranking cannot be distinguished from run-to-run variation. The word "significantly" should either be backed by statistical evidence or replaced with a more measured claim.
  2. [Table 2 and footnote 1] The paper claims experiments on eight real-world datasets, but Table 2 reports only an averaged row labeled "ETT1" for the four ETT datasets and a footnote says "we achieved SOTA performance across all ETT datasets in actuality." The per-dataset results for ETTh1, ETTh2, ETTm1, and ETTm2 are not shown anywhere in the manuscript; they are only promised in a repository. This makes the eight-dataset claim unverifiable from the paper itself. The authors should include the full per-dataset results, at least in an appendix or in the supplementary material, and the footnote's unsupported assertion should be replaced with a reference to the actual table.
  3. [Section 3.3, Eq. (5), and Table 3] The load-bearing modeling premise of ATCD is that the output of multi-head attention followed by SiLU is a usable trend component and that the residual is the seasonal component. The manuscript offers no theoretical justification, synthetic data validation, or visualization demonstrating that this decomposition actually separates trend from seasonality. The only indirect evidence is the ablation in Table 3, and that evidence is mixed: removing ATCD improves ECL MSE from 0.168 to 0.163, and the text in Section 4.2 concedes that ATCD may introduce noise on the Traffic dataset. The paper should either provide direct validation of the decomposition (for example, by inspecting extracted components on series with known trend/seasonality) or substantially soften the claims about ATCD's contribution.
  4. [Section 4.2, Fig. 3] The efficiency claim is not quantitatively supported. Fig. 3 plots "Hybird Time (ms/iter)" and training time versus lookback length, but no exact numerical values are given, no error bars or repeated-measurement variation are reported, and the setup is limited to the ECL dataset. The abstract claims superiority "in computational efficiency," but the reader cannot assess the magnitude of the efficiency gain or its statistical reliability. The authors should provide a table with measured training and inference times, parameter counts, and hardware conditions, and ideally a complexity analysis.
minor comments (6)
  1. [Section 3.4, Eq. (8)] The second basis function in Eq. (8) uses "2^{-i/2}" in the exponent; this should be "2^{-j/2}" to match the index j.
  2. [Section 3.2, Eq. (2)] The frequency transform function F is introduced in the text but not explicitly defined, and the frequency-domain L1 term is written for possibly complex coefficients. Please define the transform and clarify how the L1 norm is computed on complex values.
  3. [Section 4.1] The statement that alpha is set to 0.2 because it is "optimal in comprehensive situations" is not accompanied by any sensitivity analysis. Please report the effect of alpha on at least one dataset or state that it is a fixed constant.
  4. [Figure 3 and Table 1] There are several presentation issues: the left axis of Fig. 3 reads "Hybird Time" rather than "Hybrid Time"; Table 1 has spacing problems such as "T ime stepsis"; and Table 2's header shows "KARMASMamba" without a space. These should be corrected.
  5. [Table 2] The row label "ETT1" is ambiguous because it denotes the average over ETTh1, ETTh2, ETTm1, and ETTm2. The caption or row label should state this explicitly rather than relying on the footnote.
  6. [Section 3.4, Eqs. (13)-(14)] The matrix A is used both as the state transition matrix in Eq. (13) and as the discretized transition in Eq. (14), creating notational confusion. Rename one of the two quantities.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: KARMA's components are defined constructively and validated against external benchmarks.

full rationale

The paper's derivation chain is constructive rather than circular. ATCD defines the trend as SiLU(MHA(X_in)) and the seasonality as the residual X_in - X_t (Eqs. 4-5); this is an architectural ansatz, not a derived equivalence whose conclusion is assumed in its input. HFTD uses DWT to obtain frequency components and a residual temporal branch; Eq. 15 inverts the DWT representation in Eq. 7, which is a deliberate reconstruction design rather than a circular prediction. All performance claims are checked against external baselines (SMamba, iTransformer, PatchTST, etc.) on eight public datasets, so the central empirical claim is not defined in terms of KARMA's own outputs. The loss in Eq. 2 combines time-domain MSE with a frequency-domain term, and the weight alpha is tuned on the benchmark suite; this is standard empirical practice, not a fitted parameter renamed as a prediction. There are no self-citations: references [8] and [19] are external prior works, and no load-bearing argument reduces to an author-cited theorem. The acknowledged weakness that 'significantly outperforms' is not backed by repeated-seed variance or significance tests concerns statistical evidence and correctness risk, not circularity. Overall, the paper is self-contained against external benchmarks and exhibits no equation-level or citation-level circularity.

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

No physical entities are postulated; the proposed modules are architectural designs whose validity is measured only on the reported benchmarks. The load-bearing ingredients are the decomposition assumptions in ATCD and HFTD plus the tuned hyperparameters listed above.

free parameters (6)
  • Loss balance weight alpha = 0.2
    Balances time-domain MSE and frequency-domain L1 loss in Eq. 2; described as optimal on benchmarks with no sensitivity analysis.
  • DWT decomposition level J = 1
    Number of wavelet levels in Eqs. 7-9 is chosen by hand; deeper decomposition is not tested.
  • Wavelet family = not reported
    Choice of wavelet basis in Eq. 8 affects the high/low frequency split; the specific wavelet is not stated in the manuscript.
  • Number of KarmaBlocks N = not reported
    Depth of stacked Mamba blocks in Section 3.4 is a central capacity hyperparameter and is not specified.
  • ATCD attention heads and embedding dimension = not reported
    Multi-head attention configuration in Eq. 4 is not stated, though it determines the decomposition capacity.
  • Dropout probability p = not reported
    Dropout used in Eq. 3 is mentioned but its value is not given.
assumptions (4)
  • ad hoc to paper Multi-head attention output after SiLU is a usable trend component and the residual is seasonal (Eq. 5).
    No proof or independent test that this learned decomposition separates true trend and seasonality; the ATCD branch depends entirely on this assumption.
  • ad hoc to paper One-level DWT with an unspecified wavelet adequately separates high- and low-frequency seasonal information (Eq. 9).
    J=1 and the wavelet choice are asserted without comparison; if the wavelet mismatches the signal structure, the frequency branches introduce noise.
  • domain assumption Frequency-domain L1 loss in Eq. 2 improves forecasting (following reference [19]).
    No ablation isolates this loss, and alpha is fixed, so its actual contribution is not measured.
  • domain assumption The benchmark datasets and baseline implementations in Table 1 are representative and fairly comparable.
    Comparability requires identical preprocessing, hyperparameters, and seeds, which are not documented for the baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KARMA: A Multilevel Decomposition Hybrid Mamba Framework for Multivariate Long-Term Time Series Forecasting." pith.science (2026). https://pith.science/paper/FVLYPYU5

@misc{pith2026250608939,
  author       = {Pith},
  title        = {Pith review of: KARMA: A Multilevel Decomposition Hybrid Mamba Framework for Multivariate Long-Term Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FVLYPYU5}},
  note         = {Machine review of arXiv:2506.08939}
}
read the original abstract

Multivariate long-term and efficient time series forecasting is a key requirement for a variety of practical applications, and there are complex interleaving time dynamics in time series data that require decomposition modeling. Traditional time series decomposition methods are single and rely on fixed rules, which are insufficient for mining the potential information of the series and adapting to the dynamic characteristics of complex series. On the other hand, the Transformer-based models for time series forecasting struggle to effectively model long sequences and intricate dynamic relationships due to their high computational complexity. To overcome these limitations, we introduce KARMA, with an Adaptive Time Channel Decomposition module (ATCD) to dynamically extract trend and seasonal components. It further integrates a Hybrid Frequency-Time Decomposition module (HFTD) to further decompose Series into frequency-domain and time-domain. These components are coupled with multi-scale Mamba-based KarmaBlock to efficiently process global and local information in a coordinated manner. Experiments on eight real-world datasets from diverse domains well demonstrated that KARMA significantly outperforms mainstream baseline methods in both predictive accuracy and computational efficiency. Code and full results are available at this repository: https://github.com/yedadasd/KARMA

Figures

Figures reproduced from arXiv: 2506.08939 by the authors.

Figure 1
Figure 1. Overall structure of KARMA. (a) ATCD module, which operates along the channel dimension. (b) Details of Mamba, the core is the selective SSM, Composed of High Frequency Mamba, Low Frequency Mamba and Temporal Mamba to form Karmablock, and Global Mamba. (c) HFTD module, integrated with KarmaBlocks, leveraging Mambas across multiple scales, which functions along the temporal dimension. Additionally, the loss function … view at source ↗
Figure 2
Figure 2. Visualization results of input-96-predict-96 results on ECL, Traffic, ETTm1, and Weather datasets [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Model efficiency comparison under input-720-predict-720 of ECL (left) . Com￾parison of training efficiency corresponding to different input and output lengths (right) Efficiency. We analyzed the efficiency of KARMA on the ECL dataset from two perspectives [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 canonical work pages

  1. [1]

    In: NeurIPS (2021)

    Wu, H., Xu, J., Wang, J., et al.: Autoformer: Decomposition transformers with Auto-Correlation for long-term series forecasting. In: NeurIPS (2021)

    work page 2021

  2. [2]

    In: ICASSP (2024)

    Das, S., NaderiAlizadeh, N., Ribeiro, A.: State-augmented information routing in communication systems with graph neural networks. In: ICASSP (2024)

    work page 2024

  3. [3]

    In: WASA

    Yan, L., Hu, H., Yang, K., et al.: Variational autoencoder based automatic clustering for multivariate time series anomaly detection. In: WASA. pp. 151–162 (2024)

    work page 2024

  4. [4]

    ICLR (2023)

    Nie, Y., Nguyen, N.H., Sinthong, P., et al.: A time series is worth 64 words: Long-term forecasting with transformers. ICLR (2023)

    work page 2023

  5. [5]

    In: ICLR (2024)

    Liu, Y., Hu, T., Zhang, H., et al.: iTransformer: Inverted transformers are effective for time series forecasting. In: ICLR (2024)

    work page 2024

  6. [6]

    In: NeurIPS (2022)

    Liu, Y., Wu, H., Wang, J., et al.: Non-stationary Transformers: Exploring the stationarity in time series forecasting. In: NeurIPS (2022)

    work page 2022

  7. [7]

    In: CoLM (2024)

    Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. In: CoLM (2024)

    2024

  8. [8]

    Wang, Z., Kong, F., Feng, S., et al.: Is mamba effective for time series forecasting? ArXivabs/2403.11144(2024)

    arXiv 2024

Show all 22 references
  1. [9]

    Journal of Official Statistics6(1), 3–73 (1990)

    Cleveland, R.B., Cleveland, W.S., et al.: STL: A seasonal-trend decomposition procedure based on loess. Journal of Official Statistics6(1), 3–73 (1990)

    1990

  2. [10]

    Mathematics of Computation19(90), 297–301 (1965)

    Cooley, J.W., Tukey, J.W.: An algorithm for the machine calculation of complex fourier series. Mathematics of Computation19(90), 297–301 (1965)

    1965

  3. [11]

    ICML (2022)

    Zhou, T., Ma, Z., Wen, Q., et al.: FEDformer: Frequency enhanced decomposed transformer for long-term series forecasting. ICML (2022)

    2022

  4. [12]

    International Journal of Forecasting (2020)

    Salinas, D., Flunkert, V., Gasthaus, J., et al.: DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International Journal of Forecasting (2020)

    2020

  5. [13]

    ICLR (2023)

    Wu, H., Hu, T., Liu, Y., et al.: TimesNet: Temporal 2d-variation modeling for general time series analysis. ICLR (2023)

    2023

  6. [14]

    Chen, S., Li, C.L., Yoder, N., et al.: TSMixer: An all-mlp architecture for time series forecasting. Trans. Mach. Learn. Res. (2023)

    2023

  7. [15]

    NeurIPS (2017)

    Vaswani, A., Shazeer, N., Parmar, N., et al.: Attention is all you need. NeurIPS (2017)

    2017

  8. [16]

    ICLR (2021)

    Liu, S., Yu, H., Liao, C., et al.: Pyraformer: Low-complexity pyramidal attention for long-range time series modeling and forecasting. ICLR (2021)

    2021

  9. [17]

    In: ICLR (2020)

    Oreshkin, B.N., Carpov, D., Chapados, N., Bengio, Y.: N-BEATS: neural basis expansion analysis for interpretable time series forecasting. In: ICLR (2020)

    2020

  10. [18]

    Zeng, A., Chen, M., Zhang, L., et al.: Are transformers effective for time series forecasting? AAAI (2023)

    2023

  11. [19]

    ArXivabs/2402.02399(2024)

    Wang, H., Pan, L., Chen, Z., et al.: Fredf: Learning to forecast in frequency domain. ArXivabs/2402.02399(2024)

    2024 arXiv

  12. [20]

    IEEE Transactions on Information Theory36(5), 961–1005 (1990)

    Daubechies, I.: The wavelet transform, time-frequency localization and signal analysis. IEEE Transactions on Information Theory36(5), 961–1005 (1990)

    1990

  13. [21]

    In: AAAI (2021)

    Zhou, H., Zhang, S., Peng, J., et al.: Informer: Beyond efficient transformer for long sequence time-series forecasting. In: AAAI (2021)

    2021

  14. [22]

    SIGIR (2018)

    Lai, G., Chang, W.C., Yang, Y., et al.: Modeling long- and short-term temporal patterns with deep neural networks. SIGIR (2018)

    2018

Pith tools

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