acceptodds
Under review as a conference paper at ICLR 2027

Spectrum-DualNet: Spectrum-Guided Routing of Time-Frequency Dual-Network via Adaptive Decomposition for Time Series Forecasting

Abstract

Multivariate time series often contain temporal patterns evolving at different characteristic timescales, resulting in entangled long-term trends and short-term oscillations. Disentangling these heterogeneous dynamics using spectrum information can facilitate more effective forecasting by enabling specialized modeling of different temporal scales. However, widely used trend-seasonal decomposition methods typically rely on predefined decomposition structures and are not designed to adaptively identify multiple spectrum modes with distinct central frequencies across different datasets; existing decomposition-based forecasting methods often apply the same modeling mechanism to different resulting components, without explicitly accounting for their distinct spectrum characteristics. Yet, how to separate band-limited modes according to their spectra and assign them to spectrum-specialized branches remains underexplored in general multivariate time series forecasting. To address this, we propose Spectrum-DualNet, a dual-branch forecasting network that adaptively separates multivariate signals from various datasets into band-limited modes with distinct characteristic spectra using Multivariate Variational Mode Decomposition (MVMD). The lowest-frequency mode is routed to a trend branch in the time domain to capture long-term dependencies, while the remaining high-frequency modes are routed to a fluctuation branch that models non-stationary dynamics in the frequency domain. The two branches are fused through a gating network in which a multi-layer perceptron computes mixing coefficients that weight the trend and fluctuation components. The decomposition residual is re-injected through a second gate to restore information lost during decomposition. A prediction head maps the fused representation to the forecast. Extensive experiments on real-world multivariate time series datasets demonstrate that Spectrum-DualNet achieves superior forecasting performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time series forecasting, Spectrum-conditioned mode routing, Time–frequency dual-branch network, MVMD, Adaptive gating fusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3vvVOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents Spectrum-DualNet, a dual-branch forecasting network that adaptively separates multivariate signals according to the distributions of their datasets into band-limited modes that belong to distinct frequency ranges as characteristic spectra using Multivariate Variational Mode Decomposition (MVMD). Experiments use seven long-term benchmarks and compare the performance of Spectrum-DualNet at look-back 512 against eight baselines.

Critical strengths

This paper aims to solve an important problem.

Critical weaknesses

This paper carries limited novelty. The proposed architecture combines MVMD (2019), a PatchTST encoder, APRNet's amplitude–phase attention, KAN layers and gates. The large body of VMD-based hybrid forecasting work is not discussed.

This paper claims that MVMD-derived modes are a better basis than predefined decompositions, and routing modes to domain-specific branches helps. The experiments presented in this paper cannot support such claims.

  • There is no measurable gain over APRNet, whose mechanism is reused. Average MSE versus APRNet is 0.347 vs 0.348 (ETTm1), 0.253 vs 0.253 (ETTm2), 0.388 vs 0.388 (Traffic), 0.220 vs 0.223 (Weather) and 0.403 vs 0.411 (ETTh1); it is worse on ETTh2 (0.352 vs 0.336, and 0.331 for PatchTST) and Electricity (0.159 vs 0.155). All are produced through single runs without reporting variance.
  • The model is trained with a weighted L1 + L2 loss whose weight is a tuned hyperparameter; most baselines are trained with MSE. On these benchmarks an L1 term alone typically lowers MAE, and the MAE margins are the larger ones in Table 1 (e.g., Weather 0.253 vs 0.262).
  • Some numbers are taken from papers "with the same settings" and others reproduced, but the DLinear numbers are those of its original paper at look-back 336, not 512. Several baseline entries look erroneous: TimeKAN's Electricity MAE is 0.340 / 0.384 / 0.406 at MSE 0.137–0.165, iTransformer's Weather-96 MSE is 0.253, and DLinear's ETTh2-192 MSE of 0.283 is inconsistent with its reported average.
  • The ablation contradicts the conclusion drawn from it. Table 2 (which matches the horizon-96 rows only) shows that removing the Trend branch improves Electricity (0.124 / 0.216 vs 0.127 / 0.217) and changes Weather by 0.001, yet the text states that "all the three components are essential". Removing the decomposition changes MSE by 0.003 on ETTm1, ETTm2 and Electricity. The Traffic entry for "w/o Fluctuation" (MSE 0.369, MAE 0.357) looks like an error.
  • The ablations that would test the claims are absent. There is no comparison with a moving-average or fixed band-split decomposition in the same network, no swap of the two branches, no variant that gives all modes to both branches, and no ablation of the residual gate, the cross-attention, KAN versus MLP, K or α.

The proposed method does not match its description as adaptive, spectrum-guided and physically meaningful.

  • The decomposition is a fixed linear filter bank. After Stage I, the model applies K fixed filters determined by K scalars per dataset and one global α. Nothing adapts to the input window and nothing in it is learned, so "differentiable" has no consequence. This is a preset band split, which is the kind of predefined structure the introduction criticizes, and it is not compared with simpler ways of choosing K bands (FFT peak picking, learnable band-pass filters).
  • The lowest mode cannot be a monthly component inside the window. The reported Mode-1 periods are 10 to 35 days (and 1,000–2,500 samples on ETTm), while the window is 512 samples (about 21 days hourly, 5.3 days at 15 minutes). Within a window this mode is a low-pass of a mean-removed signal, at most one or two bins from DC, not a periodicity.
  • The "physical meaning" claim is not supported by Table 3. ETTm1 and ETTm2 have the same sampling and domain but yield 21.0 h and 6.1 h versus a single mode at 3.4 h. Mode-1 periods range from 10.4 to 34.7 days and are described as "monthly". Only ETTh1, ETTh2 and Electricity show a daily mode. Weather and Traffic are not reported.
Questions for authors

Please report, per dataset: K, α, D, patch size and stride, whether the projection is enabled, the loss weight, the search space and whether selection used the validation set; and mean ± std over at least 3 seeds for Table 1 and Table 2.

What are the results with a pure MSE loss, or with APRNet, PatchTST and TimeKAN retrained with the same L1 + L2 loss and look-back 512?

Replace the MVMD filter bank with (a) a moving-average trend/seasonal split and (b) K fixed bands at the top FFT peaks of the training set, keeping everything else the same. Does MVMD matter?

6 comments
AuthorsOpenReview ↗

Responses to Critical Weaknesses for Reviewer 3vvV - Weakness 1

Our contribution concerns how spectral decomposition organizes the forecasting architecture, rather than introducing these individual components. Specifically, Spectrum-DualNet couples dataset-level spectral estimation, consistent window-level extraction, and frequency-based branch assignment. Stage I estimates shared center frequencies from the training sequence. Stage II reuses these centers in a differentiable sequential filtering pass, providing a consistent spectral basis across input windows. The resulting frequency ordering determines the modeling mechanism: the lowest-frequency component enters a time-domain branch, while the higher-frequency components enter a frequency-domain branch. A separate residual gate makes content outside the extracted components available for prediction. This coupling is the intended methodological contribution.

The separation between dataset-level estimation and window-level processing also addresses a practical issue: some characteristic timescales exceed the input-window length, making their frequencies difficult to estimate independently within each window (As you observed, the reported lowest-frequency periods can exceed the 512-sample context, particularly on ETTm, where a window spans approximately 5.3 days). Estimating centers over the longer training sequence supplies a shared spectral prior, while the window-level outputs represent local slow and oscillatory variation. This design supports consistent branch specialization without repeatedly optimizing MVMD for every window.

We added extra discussion about VMD-based hybrid forecasting methods in the Related Work section, please check the revised paper for the corresponding content. However, to the extent of our knowledge, current VMD-based hybrid forecasting methods primarily focusing on a specific domain rather working as a general-purpures time series forecaster (like Spectrum-DualNet). For example, NA-MVMD-iTransformer-DDQN model proposed by paper "Power Load Forecasting Using Deep Learning and Reinforcement Learning" focusing on power load forecasting; CEDV-TCN-GTAF proposed by "A Frequency-domain Decomposition and TCN-GTAF Fusion Framework for GNSS Sequence Forecasting" is for Global Navigation Satellite System time series.

AuthorsOpenReview ↗

Responses to Critical Questions for Reviewer 3vvV - Weakness 2 Point 1

  • Weakness 2 Point 1.

    Thank you for highlighting the close comparison with APRNet . These margins are comparable in scale to improvements reported by recent forecasting methods. For example, TimeKAN (ICLR 2025) improves average Weather MSE from 0.245 to 0.242 relative to TimeMixer, approximately 1.22%, while tying it on ETTm2. FilterNet (NeurIPS 2024) likewise reports some improvements below 1% over PatchTST.

    Modest numerical margins on these established benchmarks are also reported by recent methods. For example, TimeKAN (ICLR 2025) reduces average Weather MSE from 0.245 to 0.242 relative to TimeMixer, while tying it on ETTm2; FilterNet (NeurIPS 2024) likewise reports improvements below 1% over PatchTST in some forecasting settings. These comparisons place the magnitude of our observed gains in context, although they do not establish their statistical significance. As emphasized by Hewamalage et al. (2023) in paper “Forecast evaluation for data scientists: common pitfalls and best practices”, reliable interpretation requires appropriate evaluation and statistical testing.

    We also clarify that Spectrum-DualNet adapts APRNet’s bidirectional amplitude–phase cross-attention mechanism within its Fluctuation branch; it does not incorporate APRNet as a complete forecasting backbone (or use any code from APRNet in SepctrumDualNet). APRNet additionally includes temporal–channel dual-view Fourier modeling, a discrete–spline D-KAN calibration operator, and temporal–channel feature reconstruction and fusion. Spectrum-DualNet does not adopt those specific architectural mechanisms.

    DatasetHorizonMSEMAE
    ETTh1960.361 ± 0.00320.393 ± 0.0031
    ETTh11920.401 ± 0.00250.416 ± 0.0015
    ETTh13360.419 ± 0.00610.428 ± 0.0035
    ETTh17200.433 ± 0.00100.452 ± 0.0020
    ETTh1Avg.0.403 ± 0.00070.422 ± 0.0006
    ETTh2960.276 ± 0.00610.336 ± 0.0052
    ETTh21920.367 ± 0.01170.391 ± 0.0080
    ETTh23360.382 ± 0.00450.408 ± 0.0035
    ETTh27200.405 ± 0.00150.436 ± 0.0006
    ETTh2Avg.0.357 ± 0.00530.393 ± 0.0038
    ETTm1960.285 ± 0.00150.336 ± 0.0017
    ETTm11920.328 ± 0.00260.359 ± 0.0015
    ETTm13360.360 ± 0.00150.384 ± 0.0023
    ETTm17200.418 ± 0.00200.411 ± 0.0006
    ETTm1Avg.0.348 ± 0.00180.373 ± 0.0002
    ETTm2960.163 ± 0.00150.251 ± 0.0021
    ETTm21920.221 ± 0.00210.291 ± 0.0021
    ETTm23360.273 ± 0.00210.326 ± 0.0012
    ETTm27200.358 ± 0.00060.381 ± 0.0010
    ETTm2Avg.0.254 ± 0.00130.312 ± 0.0006
    Weather960.142 ± 0.00150.187 ± 0.0020
    Weather1920.186 ± 0.00060.229 ± 0.0006
    Weather3360.237 ± 0.00100.270 ± 0.0000
    Weather7200.315 ± 0.00210.325 ± 0.0012
    WeatherAvg.0.220 ± 0.00080.252 ± 0.0005
    Electricity960.126 ± 0.00060.217 ± 0.0006
    Electricity1920.148 ± 0.00120.237 ± 0.0006
    Electricity3360.162 ± 0.00060.253 ± 0.0006
    Electricity7200.198 ± 0.00100.285 ± 0.0006
    ElectricityAvg.0.158 ± 0.00040.248 ± 0.0001
    Traffic960.354 ± 0.00060.239 ± 0.0006
    Traffic1920.376 ± 0.00150.251 ± 0.0029
    Traffic3360.389 ± 0.00230.259 ± 0.0038
    Traffic7200.427 ± 0.00120.281 ± 0.0023
    TrafficAvg.0.387 ± 0.00070.258 ± 0.0021

    As the table indicated, the improvements remain statistically and practically reliable under multiple random seeds. The improvements in Table 1 are compared between best results of different models.

AuthorsOpenReview ↗

Responses to Critical Weaknesses for Reviewer 3vvV - Weakness 2 Point 2 - 4

  • Weakness 2 Point 2. | Dataset | Original (full model) MSE | Original (full model) MAE | Moving average MSE | Moving average MAE | Top-K MSE | Top-K MAE | Random learnable centers MSE | Random learnable centers MAE | Reversed branches MSE | Reversed branches MAE | w/o Residual MSE | w/o Residual MAE | w/o L1 loss MSE | w/o L1 loss MAE | | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | ETTh1 | 0.360 | 0.391 | 0.367 | 0.397 | 0.380 | 0.415 | 0.416 | 0.434 | 0.371 | 0.398 | 0.374 | 0.405 | 0.365 | 0.395 | | ETTh2 | 0.269 | 0.330 | 0.278 | 0.338 | 0.298 | 0.357 | 0.296 | 0.354 | 0.305 | 0.369 | 0.285 | 0.347 | 0.277 | 0.342 | | ETTm1 | 0.284 | 0.338 | 0.287 | 0.340 | 0.293 | 0.340 | 0.297 | 0.344 | 0.292 | 0.337 | 0.295 | 0.341 | 0.289 | 0.345 | | ETTm2 | 0.162 | 0.251 | 0.164 | 0.253 | 0.166 | 0.257 | 0.169 | 0.254 | 0.167 | 0.254 | 0.167 | 0.254 | 0.165 | 0.255 | | Electricity | 0.127 | 0.217 | 0.129 | 0.220 | 0.128 | 0.219 | 0.130 | 0.223 | 0.130 | 0.220 | 0.137 | 0.234 | 0.128 | 0.221 | | Weather | 0.143 | 0.187 | 0.145 | 0.190 | 0.146 | 0.193 | 0.144 | 0.188 | 0.152 | 0.201 | 0.144 | 0.188 | 0.144 | 0.239 | | Traffic | 0.354 | 0.239 | 0.356 | 0.242 | 0.377 | 0.267 | 0.362 | 0.246 | 0.359 | 0.241 | 0.364 | 0.249 | 0.358 | 0.247 |

    We agree that the mixed objective contributes to performance. However, the additional ablations indicate that the gains are not attributable to the loss choice alone. For example, on ETTh1, removing L1 increases MSE/MAE by 0.005/0.004, whereas replacing the MVMD centers with randomly initialized learnable centers increases them by 0.056/0.043, replacing decomposition with Top-K increases them by 0.020/0.024, and removing the residual increases them by 0.014/0.014. Under the same mixed objective, these architectural changes therefore cause larger degradations than removing L1 from the full model. This supports the contribution of the spectral design and residual pathway alongside the training objective.

  • Weakness 2 Point 3. As we mentioned in Experiments section Baseline paragraph “We reference the results of the models that have the same experimental settings from the papers. For models that have different experimental settings, we reproduce the model under our experimental settings”, here are the details:

    APRNet – References the best results from the paper because the input and prediction configurations are the same with SpectrumDualNet.

    TimeKAN – Retrained and tested using the official Github code and training scripts, with input sequence set to be 512.

    Amplifier - Retrained and tested using the official Github code and training scripts, with input sequence set to be 512.

    iTransformer – Referenced from the APRNet paper.

    FilterNet - Referenced from the APRNet paper.

    PatchTST - References the best results from the paper because the input and prediction configurations are the same with SpectrumDualNet.

    DLinear - Referenced from the APRNet paper.

  • Weakness 2 Point 4. The electricity dataset (also known as ECL) includes hourly electricity consumption data from 370 clients, providing insights into consumer-level load patterns. Compared with other datasets, the electricity provides a microcosmic perspective as it records data from individual clients. Thus, the trend of the electricity dataset could be vulnerable but the fluctuations could be robust, and that makes the trend branch defect the predictions.

    We double checked our logs and updated the Traffic entry in Table 2, thank you for pointing this error out. The table below provides the full ablation study results of the Fluctuation branch using the traffic dataset.

    HorizonMSEMAE
    960.3570.242
    1920.3780.252
    3360.3970.268
    7200.4360.284
Reviewer_PWiXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Spectrum-DualNet, a dual-branch architecture for multivariate time-series forecasting that uses spectral decomposition to guide the assignment of temporal components to different modeling mechanisms. The method consists of three main components. First, it uses multivariate variational mode decomposition (MVMD) on the training data to estimate dataset-specific center frequencies, then reuses these frequencies in a differentiable, window-level filtering procedure. Second, it routes the lowest-frequency mode to a time-domain Trend branch based on a patch Transformer, while routing the remaining modes to a frequency-domain Fluctuation branch that models amplitude–phase interactions using cross-attention and KAN-based spectral corrections. Third, it employs two input-dependent gates to combine the branch representations and incorporate the decomposition residual. The authors evaluate Spectrum-DualNet on seven commonly used time-series forecasting benchmarks and compare it with eight representative baselines.

Critical strengths
  1. A coherent motivation. The paper identifies a meaningful modeling challenge: temporal components with different spectral characteristics may benefit from different forecasting mechanisms. The proposed architecture follows this motivation consistently.

  2. Residual information is not discarded. The method explicitly retains the decomposition residual and incorporates it through a learned gate. This is preferable to assuming that a finite set of extracted modes captures all useful input information.

  3. Broad baseline coverage for the chosen benchmarks. The paper compares against eight forecasting methods spanning linear, Transformer-based, decomposition-based, and frequency-oriented approaches.

Critical weaknesses
  1. Unexplained inconsistency between reported results.

The ETTh1 and ETTh2 MSE values in Table 1 differ substantially from those of the full model in Table 2, despite the statement that the ablations use the same settings. This raises questions about whether the tables reflect comparable experimental configurations and weakens confidence in the ablation conclusions.

  1. The central routing hypothesis is insufficiently tested.

The paper argues that low-frequency modes should be assigned to the time-domain branch and higher-frequency modes to the frequency-domain branch. However, it does not directly compare this assignment with reversed routing, alternative mode groupings, or parameter-matched architectures that do not use spectrum-guided routing.

  1. Novelty relative to prior work needs sharper delineation.

The manuscript discusses decomposition, frequency-domain forecasting, amplitude–phase modeling, and multi-branch prediction, but the methodological distinction from the closest prior approaches is not established strongly enough. The contribution appears to be primarily an integration of existing ideas, and the paper should explain which component or interaction is genuinely new and why it is necessary.

  1. Limited evidence of statistical reliability.

The paper does not clearly report repeated-seed results or uncertainty estimates. Consequently, it is difficult to determine whether small differences from strong baselines reflect robust improvements.

  1. The approximation's forecasting relevance is not adequately established.

The spectral plots in Section 4.3 offer qualitative evidence that the window-level filtering preserves dominant modes, but they do not quantitatively establish how much decomposition error is introduced or whether this error affects forecasting performance.

  1. Efficiency and practical cost are underreported.

The method includes MVMD frequency estimation, sequential filtering, a Transformer, cross-attention, and KAN-based correction modules. The manuscript does not provide enough quantitative analysis of training time, inference latency, parameter count, or memory use to establish the practical cost of these components relative to simpler alternatives.

  1. Generalization evidence is limited.

The evaluation uses seven established benchmarks. More diverse datasets and explicit distribution-shift tests would help establish whether dataset-level frequency estimation is broadly useful rather than particularly effective on the selected benchmarks.

Questions for authors

See the weaknesses.

6 comments
AuthorsOpenReview ↗

Responses to Critical Weaknesses and Questions for Reviewer PWiX - 1 to 5

  1. The ablation study results presented in Table 2 is conducted with prediction length 96. Table 1 only presents the averaged results of the 4 prediction lengths due to the page limitation of ICLR main text, full results of Table 1 is presented in Appendix A2, which includes the results with prediction length 96.

 

  1. In the supplementary ablation studies, we inversed the two branches, and the results provide a support to this assumption. Results shown in the table below. | Dataset | Original (full model) MSE | Original (full model) MAE | Moving average MSE | Moving average MAE | Top-K MSE | Top-K MAE | Random learnable centers MSE | Random learnable centers MAE | Reversed branches MSE | Reversed branches MAE | w/o Residual MSE | w/o Residual MAE | | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | ETTh1 | 0.360 | 0.391 | 0.367 | 0.397 | 0.380 | 0.415 | 0.416 | 0.434 | 0.371 | 0.398 | 0.374 | 0.405 | | ETTh2 | 0.269 | 0.330 | 0.278 | 0.338 | 0.298 | 0.357 | 0.296 | 0.354 | 0.305 | 0.369 | 0.285 | 0.347 | | ETTm1 | 0.284 | 0.338 | 0.287 | 0.340 | 0.293 | 0.340 | 0.297 | 0.344 | 0.292 | 0.337 | 0.295 | 0.341 | | ETTm2 | 0.162 | 0.251 | 0.164 | 0.253 | 0.166 | 0.257 | 0.169 | 0.254 | 0.167 | 0.254 | 0.167 | 0.254 | | Electricity | 0.127 | 0.217 | 0.129 | 0.220 | 0.128 | 0.219 | 0.130 | 0.223 | 0.130 | 0.220 | 0.137 | 0.234 | | Weather | 0.143 | 0.187 | 0.145 | 0.190 | 0.146 | 0.193 | 0.144 | 0.188 | 0.152 | 0.201 | 0.144 | 0.188 | | Traffic | 0.354 | 0.239 | 0.356 | 0.242 | 0.377 | 0.267 | 0.362 | 0.246 | 0.359 | 0.241 | 0.364 | 0.249 |

 

  1. Please refer to the revised paper, as this question is directly related to the content of the paper.

 

  1. As the table below indicated, the improvements remain statistically and practically reliable under multiple random seeds. The improvements in Table 1 are compared between best results of different models. | Dataset | Horizon | MSE | MAE | | :--- | :---: | :---: | :---: | | ETTh1 | 96 | 0.361 ± 0.0032 | 0.393 ± 0.0031 | | ETTh1 | 192 | 0.401 ± 0.0025 | 0.416 ± 0.0015 | | ETTh1 | 336 | 0.419 ± 0.0061 | 0.428 ± 0.0035 | | ETTh1 | 720 | 0.433 ± 0.0010 | 0.452 ± 0.0020 | | ETTh1 | Avg. | 0.403 ± 0.0007 | 0.422 ± 0.0006 | | ETTh2 | 96 | 0.276 ± 0.0061 | 0.336 ± 0.0052 | | ETTh2 | 192 | 0.367 ± 0.0117 | 0.391 ± 0.0080 | | ETTh2 | 336 | 0.382 ± 0.0045 | 0.408 ± 0.0035 | | ETTh2 | 720 | 0.405 ± 0.0015 | 0.436 ± 0.0006 | | ETTh2 | Avg. | 0.357 ± 0.0053 | 0.393 ± 0.0038 | | ETTm1 | 96 | 0.285 ± 0.0015 | 0.336 ± 0.0017 | | ETTm1 | 192 | 0.328 ± 0.0026 | 0.359 ± 0.0015 | | ETTm1 | 336 | 0.360 ± 0.0015 | 0.384 ± 0.0023 | | ETTm1 | 720 | 0.418 ± 0.0020 | 0.411 ± 0.0006 | | ETTm1 | Avg. | 0.348 ± 0.0018 | 0.373 ± 0.0002 | | ETTm2 | 96 | 0.163 ± 0.0015 | 0.251 ± 0.0021 | | ETTm2 | 192 | 0.221 ± 0.0021 | 0.291 ± 0.0021 | | ETTm2 | 336 | 0.273 ± 0.0021 | 0.326 ± 0.0012 | | ETTm2 | 720 | 0.358 ± 0.0006 | 0.381 ± 0.0010 | | ETTm2 | Avg. | 0.254 ± 0.0013 | 0.312 ± 0.0006 | | Weather | 96 | 0.142 ± 0.0015 | 0.187 ± 0.0020 | | Weather | 192 | 0.186 ± 0.0006 | 0.229 ± 0.0006 | | Weather | 336 | 0.237 ± 0.0010 | 0.270 ± 0.0000 | | Weather | 720 | 0.315 ± 0.0021 | 0.325 ± 0.0012 | | Weather | Avg. | 0.220 ± 0.0008 | 0.252 ± 0.0005 | | Electricity | 96 | 0.126 ± 0.0006 | 0.217 ± 0.0006 | | Electricity | 192 | 0.148 ± 0.0012 | 0.237 ± 0.0006 | | Electricity | 336 | 0.162 ± 0.0006 | 0.253 ± 0.0006 | | Electricity | 720 | 0.198 ± 0.0010 | 0.285 ± 0.0006 | | Electricity | Avg. | 0.158 ± 0.0004 | 0.248 ± 0.0001 | | Traffic | 96 | 0.354 ± 0.0006 | 0.239 ± 0.0006 | | Traffic | 192 | 0.376 ± 0.0015 | 0.251 ± 0.0029 | | Traffic | 336 | 0.389 ± 0.0023 | 0.259 ± 0.0038 | | Traffic | 720 | 0.427 ± 0.0012 | 0.281 ± 0.0023 | | Traffic | Avg. | 0.387 ± 0.0007 | 0.258 ± 0.0021 |

 

  1. The residual passes the decomposition error to a fusion gate of the model, so the potential information is also preserved. As for the quantitatively establish of the decomposition error, we calculated the overall energy ratios (%) passes through via the residual, as the table below shows. For ablation of the residual, please refer to the ablation table. Residual energy ratios (%) on the training, validation, and test splits of seven datasets. | Dataset | Train | Validation | Test | | --- | ---: | ---: | ---: | | ETTh1 | 14.66 | 9.19 | 14.76 | | ETTh2 | 21.85 | 20.84 | 23.63 | | ETTm1 | 8.44 | 7.15 | 8.93 | | ETTm2 | 21.91 | 22.66 | 24.29 | | Electricity | 9.39 | 10.40 | 9.96 | | Weather | 9.42 | 7.85 | 9.42 | | Traffic | 16.08 | 16.04 | 16.41 |
AuthorsOpenReview ↗

Responses to Critical Weaknesses and Questions for Reviewer PWiX - 6 to 7

  1. Please refer to the table below. | Dataset | Training time per epoch | Training epochs (early stopping) | Inference time | GPU memory (MB) | Parameters (M) | |:---|:---:|:---:|:---:|:---:|:---:| | ETTh1 | 20 s | 26 | 1 s | 3383 | 17.9 | | ETTh2 | 20 s | 34 | 1 s | 3383 | 17.9 | | ETTm1 | 1 min 20 s | 24 | 4 s | 3383 | 17.9 | | ETTm2 | 1 min 30 s | 24 | 4 s | 3203 | 17.9 | | Electricity | 1 min 40 s | 111 | 5 s | 20399 | 1 | | Weather | 15 s | 29 | 1 s | 1259 | 2.1 | | Traffic | 2 min 5 s | 114 | 7 s | 19081 | 2.1 |

  2. Our benchmark selection follows established practice in long-term time-series forecasting: SparseTSF (ICML 2024) evaluates the same seven datasets across its main experiments and appendix, while TimeKAN (ICLR 2025) evaluates six of these datasets. This shared protocol enables direct comparison with existing methods. Nevertheless, we are aware that recent benchmarking research, like fev-bench (NeurIPS 2026), identifies limited domain coverage and overrepresentation of related datasets as limitations of the conventional LTSF suite. In future work, we plan to broaden the dataset coverage and examine changes in dominant frequencies to better establish the method’s range of applicability.

Thank you for the detailed response and additional experiments. The clarification regarding Tables 1 and 2 resolves my concern about the reported results, while the supplementary ablations and repeated-seed experiments strengthen the empirical evidence.

However, my primary concern regarding limited methodological novelty remains unresolved. Spectrum-DualNet primarily integrates existing techniques, including MVMD, Transformer-based modeling, amplitude–phase interactions, and gated fusion. Although the new experiments support the proposed design, they do not sufficiently establish a distinctive methodological contribution beyond existing decomposition-based forecasting approaches. Moreover, the benefits of spectrum-guided routing over parameter-matched alternatives and the trade-off between architectural complexity and forecasting improvements remain insufficiently demonstrated.

Therefore, while I acknowledge the improvements made in the rebuttal, I maintain my original score.

Reviewer_WncSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Spectrum-DualNet, a multivariate long-term forecasting model that estimates dataset-level center frequencies with multivariate variational mode decomposition (MVMD), then reuses those fixed frequencies in a one-pass, window-level filter bank. The lowest-frequency component is routed to a patch-based time-domain Transformer branch, while the remaining components are routed to a frequency-domain branch that models amplitude-phase interactions with cross-attention and KAN-based corrections. A first gate fuses the trend and fluctuation representations, and a second gate mixes the fused representation with the residual left by the approximate decomposition.

Critical strengths

The paper makes a clear architectural hypothesis: slowly varying components should be handled in the time domain and faster oscillations in the frequency domain. Estimating centers only from the training split avoids direct test-set leakage, and reusing those centers makes the front end cheaper than running iterative MVMD on every window.

Critical weaknesses
  1. The proposed Stage-II operation is an ordered single sweep of overlapping Lorentzian filters rather than window-level MVMD, and the remaining signal is explicitly available through a residual gate. The ablations therefore need to separate at least four factors: data-adaptive centers, ordered mode extraction, branch specialization/routing, and residual bypass. The present “without decomposition/trend/fluctuation” variants alter multiple factors simultaneously and do not include random/fixed/learned centers, swapped routes, shared branch architectures, or removal of the residual path. On Electricity, the trend-branch ablation actually obtains lower MSE than the full model (0.124 versus 0.127), contradicting the statement that all components are essential. Consequently, the performance table supports the composite architecture but not the claimed benefit of spectrum-guided routing.

  2. The paper mixes results copied from prior publications with locally reproduced results without marking their provenance or giving matched hyperparameter budgets. It reports point estimates from apparently single runs and no variance, despite many differences being only a few thousandths. The manuscript omits learning rate, batch size, optimizer, epochs/early stopping, branch dimensions, patch settings, KAN settings, the chosen L1/L2 weight, and dataset-specific configurations.

Questions for authors
  1. Please provide controlled ablations for (a) MVMD-estimated versus random, uniformly spaced, and learned centers; (b) spectrum-guided versus swapped or shared-branch routing; (c) one-pass filtering versus full window-level MVMD or a capacity-matched learned filter bank; and (d) the model with the residual gate disabled. How much forecast signal and energy pass through the residual gate on each dataset?

  2. Please identify the exact source and configuration of every baseline result, rerun the strongest baselines and Spectrum-DualNet with the same preprocessing and tuning budget over multiple seeds, and report mean, standard deviation, and paired differences. Do the small improvements remain statistically and practically reliable?

  3. Please provide the complete training configuration, including optimizer, learning rate, batch size, epoch/early-stopping policy, all branch dimensions, patch size/stride, KAN settings, L1/L2 mixture weight, and dataset-specific choices.

6 comments
AuthorsOpenReview ↗

Responses to Critical Weaknesses for Reviewer WnCS - Weakness 1

  1. Filtering is also intrinsic to original MVMD. In the original MVMD paper, each mode update applies a Wiener-type filter with a Lorentzian-shaped response to the input spectrum after subtracting the other modes and incorporating the dual variable (Eq. 29). Thus, describing Stage II as filtering does not itself distinguish it from MVMD. The important distinction is how the filters and modes are estimated: original MVMD repeatedly updates the modes, shared centers, and dual variables, whereas our Stage II reuses the centers estimated by Stage I and performs one sequential sweep. It therefore retains the frequency-selective structure of the MVMD update.

     

    Window-level MVMD would introduce repeated optimization and potentially inconsistent spectral assignments. As reviewer 3vvV observes, the reported lowest-frequency periods can exceed the 512-sample context, particularly on ETTm, where a window spans approximately 5.3 days. Such windows contain less than one cycle of these slow components, making independent estimation of their characteristic frequencies poorly constrained. Our design instead estimates dataset-specific centers using only the training sequence and reuses them across windows, providing a consistent spectral basis for routing components to the trend and fluctuation branches. ...

    As for the anomaly of the ablation of the electricity dataset, the electricity dataset (also known as ECL) includes hourly electricity consumption data from 370 clients, providing insights into consumer-level load patterns. Compared with other datasets, the electricity provides a microcosmic perspective as it records data from individual clients. Thus, the trend of the electricity dataset could be vulnerable but the fluctuations could be robust, and that makes the trend branch defect the predictions.

     

    For extra ablation studies, please refer to the table below. More details are described in the revised paper. | Dataset | Original (full model) MSE | Original (full model) MAE | Moving average MSE | Moving average MAE | Top-K MSE | Top-K MAE | Random learnable centers MSE | Random learnable centers MAE | Reversed branches MSE | Reversed branches MAE | w/o Residual MSE | w/o Residual MAE | | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | | ETTh1 | 0.360 | 0.391 | 0.367 | 0.397 | 0.380 | 0.415 | 0.416 | 0.434 | 0.371 | 0.398 | 0.374 | 0.405 | | ETTh2 | 0.269 | 0.330 | 0.278 | 0.338 | 0.298 | 0.357 | 0.296 | 0.354 | 0.305 | 0.369 | 0.285 | 0.347 | | ETTm1 | 0.284 | 0.338 | 0.287 | 0.340 | 0.293 | 0.340 | 0.297 | 0.344 | 0.292 | 0.337 | 0.295 | 0.341 | | ETTm2 | 0.162 | 0.251 | 0.164 | 0.253 | 0.166 | 0.257 | 0.169 | 0.254 | 0.167 | 0.254 | 0.167 | 0.254 | | Electricity | 0.127 | 0.217 | 0.129 | 0.220 | 0.128 | 0.219 | 0.130 | 0.223 | 0.130 | 0.220 | 0.137 | 0.234 | | Weather | 0.143 | 0.187 | 0.145 | 0.190 | 0.146 | 0.193 | 0.144 | 0.188 | 0.152 | 0.201 | 0.144 | 0.188 | | Traffic | 0.354 | 0.239 | 0.356 | 0.242 | 0.377 | 0.267 | 0.362 | 0.246 | 0.359 | 0.241 | 0.364 | 0.249 |

AuthorsOpenReview ↗

Responses to Questions for Reviewer WnCS - Question 1

Please refer to the supplementary ablation study table((a) - exp 4; (b) - exp5; (c) - exp 2 & 4; (d) - exp 6).

DatasetOriginal (full model) MSEOriginal (full model) MAEMoving average MSEMoving average MAETop-K MSETop-K MAERandom learnable centers MSERandom learnable centers MAEReversed branches MSEReversed branches MAEw/o Residual MSEw/o Residual MAE
ETTh10.3600.3910.3670.3970.3800.4150.4160.4340.3710.3980.3740.405
ETTh20.2690.3300.2780.3380.2980.3570.2960.3540.3050.3690.2850.347
ETTm10.2840.3380.2870.3400.2930.3400.2970.3440.2920.3370.2950.341
ETTm20.1620.2510.1640.2530.1660.2570.1690.2540.1670.2540.1670.254
Electricity0.1270.2170.1290.2200.1280.2190.1300.2230.1300.2200.1370.234
Weather0.1430.1870.1450.1900.1460.1930.1440.1880.1520.2010.1440.188
Traffic0.3540.2390.3560.2420.3770.2670.3620.2460.3590.2410.3640.249

For energy passes through the residual, we calculated and summed energy of all windows, the calculated the energy rations (%) that pass through the residual gate. The table below shows the results from different datasets:

DatasetTrainValidationTest
ETTh114.669.1914.76
ETTh221.8520.8423.63
ETTm18.447.158.93
ETTm221.9122.6624.29
Electricity9.3910.409.96
Weather9.427.859.42
Traffic16.0816.0416.41
AuthorsOpenReview ↗

Responses to Questions for Reviewer WnCS - Question 2

Freq Cycle – Retrained and tested using the official Github code and training scripts, with input sequence set to be 512.

APRNet – References the best results from the paper because the input and prediction configurations are the same with SpectrumDualNet.

TimeKAN – Retrained and tested using the official Github code and training scripts, with input sequence set to be 512.

Amplifier - Retrained and tested using the official Github code and training scripts, with input sequence set to be 512.

iTransformer – Referenced from the APRNet paper.

FilterNet - Referenced from the APRNet paper.

PatchTST - References the best results from the paper because the input and prediction configurations are the same with SpectrumDualNet.

DLinear - Referenced from the APRNet paper.

We reference the best results of APRNet and PatchTST presented in the original papers, as these models share the same input-prediction configuration(Input seq length 512, predict 96, 192, 336, and 720, with window stride 1). Results of iTransformer, FilterNet, and DLinear are referenced from the APRNet paper, as it has the same input-prediction configurations as SpectrumDualNet. For TimeKAN and Amplifier, we downloaded code and configured the environments followed their official Github page instructions, and reproduced the results using the official scripts with input length set to be 512. Reproduced models and SpectrumDualNet running with a same machine.

DatasetHorizonMSEMAE
ETTh1960.361 ± 0.00320.393 ± 0.0031
ETTh11920.401 ± 0.00250.416 ± 0.0015
ETTh13360.419 ± 0.00610.428 ± 0.0035
ETTh17200.433 ± 0.00100.452 ± 0.0020
ETTh1Avg.0.403 ± 0.00070.422 ± 0.0006
ETTh2960.276 ± 0.00610.336 ± 0.0052
ETTh21920.367 ± 0.01170.391 ± 0.0080
ETTh23360.382 ± 0.00450.408 ± 0.0035
ETTh27200.405 ± 0.00150.436 ± 0.0006
ETTh2Avg.0.357 ± 0.00530.393 ± 0.0038
ETTm1960.285 ± 0.00150.336 ± 0.0017
ETTm11920.328 ± 0.00260.359 ± 0.0015
ETTm13360.360 ± 0.00150.384 ± 0.0023
ETTm17200.418 ± 0.00200.411 ± 0.0006
ETTm1Avg.0.348 ± 0.00180.373 ± 0.0002
ETTm2960.163 ± 0.00150.251 ± 0.0021
ETTm21920.221 ± 0.00210.291 ± 0.0021
ETTm23360.273 ± 0.00210.326 ± 0.0012
ETTm27200.358 ± 0.00060.381 ± 0.0010
ETTm2Avg.0.254 ± 0.00130.312 ± 0.0006
Weather960.142 ± 0.00150.187 ± 0.0020
Weather1920.186 ± 0.00060.229 ± 0.0006
Weather3360.237 ± 0.00100.270 ± 0.0000
Weather7200.315 ± 0.00210.325 ± 0.0012
WeatherAvg.0.220 ± 0.00080.252 ± 0.0005
Electricity960.126 ± 0.00060.217 ± 0.0006
Electricity1920.148 ± 0.00120.237 ± 0.0006
Electricity3360.162 ± 0.00060.253 ± 0.0006
Electricity7200.198 ± 0.00100.285 ± 0.0006
ElectricityAvg.0.158 ± 0.00040.248 ± 0.0001
Traffic960.354 ± 0.00060.239 ± 0.0006
Traffic1920.376 ± 0.00150.251 ± 0.0029
Traffic3360.389 ± 0.00230.259 ± 0.0038
Traffic7200.427 ± 0.00120.281 ± 0.0023
TrafficAvg.0.387 ± 0.00070.258 ± 0.0021

As the table indicated, the improvements remain statistically and practically reliable under multiple random seeds. The improvements in Table 1 are compared between best results of different models.

Test code of the model alongside the hyper parameter configurations are available via official comment "Anonymous Project Repo".

Comments on the paper

2 comments

The rebuttal revision of this paper has been uploaded, the modified contents are highlighted with red color.

Area_Chair_A4yCOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewers who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions. Thank you all for your time and contributions to a constructive discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.