acceptodds
Under review as a conference paper at ICLR 2027

WiSPformer: A Wiener-informed Signal-Preserving Spectral Transformer for Multivariate Time Series Forecasting

Abstract

Frequency-domain time-series forecasting has emerged as an effective alternative to time-domain modelling for multivariate time-series forecasting, because spectral representations compactly capture global trends and periodic patterns. However, the difficulty of frequency-based forecasting is the global spectral shrinkage which suppresses not only nuisance noise but also predictive local transients. In this paper, we propose WiSPformer, a Wiener-informed Signal-Preserving Transformer that combines a learnable Wiener Fourier filter, an adaptive signal-preserving mechanism, and a dual-channel complex spectral Transformer. Our learnable Wiener filter infers a frequency-wise spectral gain from the estimated signal and noise spectra. To overcome over-smoothing while retaining the benefits of spectral denoising, our signal-preserving method imposes a learnable protection on attenuation for channels with informative local dynamics. Lastly, we incorporate a dual-branch complex spectral Transformer to leverage synergistic dependencies between channels and frequency components. Extensive experiments on nine benchmark datasets show that WiSPformer consistently outperforms Wiener-only spectral filtering and achieves competitive performance compared with strong state-of-the-art baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Time Series Forecasting, Wiener Filtering, Fourier Transform, Transient Preservation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jBaSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents WiSPformer, a frequency-domain forecaster with three parts: (1) A learnable Wiener filter: a free parameter per channel and frequency bin is passed through a sigmoid to give a spectral gain H_c(k) in [0, 1], motivated by writing the Wiener gain as a sigmoid of the log-SNR; (2) A signal-preserving protection: two window statistics per channel (maximum deviation from the mean and mean absolute first difference, both divided by the standard deviation) are combined into a score P_c, which sets a per-channel coefficient π_c that interpolates the gain towards 1, W_c(k) = π_c + (1 − π_c) H_c(k); and (3) A dual-branch Transformer over channel tokens that processes the real and imaginary parts of the concatenated raw and filtered spectra separately, followed by a per-bin complex multiplication ("calibration") and an inverse FFT. Experiments use nine long-term benchmarks at look-back 96 against 12 baselines, a three-channel synthetic dataset, a component ablation at horizon 96, a plug-in study on three backbones, look-back sensitivity and efficiency plots.

Critical strengths

This paper aims to solve an important problem and presents a plug-in study. Whether a spectral denoising gain should be relaxed for channels with strong local dynamics is an important research question. The proposed mechanism is simple enough to attach to other models.

This paper compares the proposed solution against baselines over nine datasets. The experiments show some improvements over baselines.

Critical weaknesses

Some designs and claims are problematic. First, the Wiener filter is a static learnable mask. H_c(k) = σ(z_{c,k}) with z is a free parameter. Neither signal nor noise spectrum has been estimated, and the gain does not depend on the input window. This is the same object as a learnable frequency filter with a sigmoid (e.g., FilterNet's plain shaping filter). Appendix E.2 argues that z converges to the log-SNR, but it assumes the loss is the reconstruction error E|HZ − S|² against a clean signal S. In the actual model, the loss is the forecasting MSE against noisy future values, taken after a Transformer that can rescale any bin. The derivation therefore does not apply to the trained model. The only empirical support (Pearson 0.80) comes from a synthetic task whose targets are the clean signals, which is a denoising setting that the real benchmarks do not have. Second, the protection is a uniform shrinkage towards the identity. Eq. 59 gives W − H = π_c (1 − H_c(k)). The gain lift is therefore largest on exactly the bins the filter attenuates most, and it is the same for every bin with the same H. Nothing in it identifies transient-related frequencies. Appendix F.2 shows only that the ordering of gains is preserved, which does not support the claim that "signal components are preferentially recovered". With P_c → 1 the filter is simply switched off for that channel. Third, the raw spectrum appears to be fed to the Transformer. In Section 3.2.1, Z_cat ∈ C^{N×2F×d} concatenates the raw and the filtered spectra. If so, nothing is removed from the model's input, and the over-smoothing problem that motivates the protection does not arise in this architecture. Fourth, no evidence that protected transients are predictive on real data. The synthetic steps, pulses and jagged segments occur at random times, so they are unpredictable from the history by construction; Step and Jagged MSEs are about 0.67–0.77 for all variants, and Jagged differs by 0.005–0.007. On real data there is no analysis stratified by P_c (error on high- versus low-protection channels or windows, with and without protection). Fifth, the reported statistics do not match Eq. 6. For any non-constant window, max_t |x − μ| / σ is at least about 1. Yet the text reports ρ^max below 0.2 on Traffic and near 0 on Electricity (Figures 4 and 12). The plotted quantity is therefore rescaled or computed differently from what is defined.

The improvements over baselines are marginal and the results are not well analyzed. First, in Table 4 (horizon 96 only), removing protection changes MSE by 0.001–0.004 on five of nine datasets (ETTm2, ETTh1, ETTh2, Weather, Exchange), and on Weather the no-protection filter is worse than no filter at all (0.153 vs 0.152). No seeds or standard deviations are given anywhere. Second, the "full model" in the ablation study is not the main-table model. At horizon 96 the full model scores 0.166 on ETTm2 in Table 4 but 0.160 in Table 6; ECL is 0.132 vs 0.128, Traffic 0.390 vs 0.387, ETTm1 0.305 vs 0.303, Weather 0.150 vs 0.148. The ETTm2 gap (0.006) is larger than most ablation effects and larger than the lead over xPatch (0.164). Run-to-run or configuration variance is thus comparable to the effects being claimed. Third, it is unclear where the headline gains come from. The "18.1% on Exchange" figure is driven by one cell (horizon 720: 0.691 vs 0.75–0.89 for the others) on a small, high-variance dataset; at horizon 96 the gap is 0.079 vs 0.082. The 6.7% average gain is quoted against xPatch, a channel-independent model that is weak on Solar, Traffic and ECL; against EMAFormer the gaps on ETTm1, ETTh1, ECL and ETTm2 are 0.002–0.006. Fourth, the dual real/imaginary channel-token Transformer follows FreEformer. Table 4's "Raw" variant (no filter) is at 0.150 on ECL-96, worse than iTransformer (0.148), so the paper's result rests on the filter; given the points in Weakness 1, what the filter contributes beyond per-channel input reweighting is not explained. Fifth, there are many errors in tables. For example, in Table 6 the Traffic rows for EMAFormer average to MAE 0.255, but 0.282 is reported (and used in Table 1); Alio's rows average to 0.277 but 0.255 is reported. With the corrected value, the Traffic MAE lead over EMAFormer falls from 0.033 to 0.006. Amplifier's Solar MAE average (0.298) also does not match its rows (0.286). In Table 7 the four TimeKAN Traffic rows are identical to the FreTS rows, with and without WiSP, while the reported TimeKAN averages (0.431 → 0.423) are different numbers; the iTransformer ETTh1 rows differ from those in Table 6.

Questions for authors
  1. Is the raw spectrum concatenated with the filtered one as input to the Transformer (Z_cat with 2F bins)? If yes, what does the "over-suppression" mean in this model, and what happens if only Z_flt is used?

  2. Please report mean ± std over at least 3 seeds for Table 1 and Table 4, with the ablation run on all four horizons and with the same configuration as the main table. Why do the full-model numbers at horizon 96 differ between Table 4 and Table 6?

  3. Controls for the protection: (a) π_c as a free per-channel parameter with no input statistics; (b) P_c shuffled across channels or windows; (c) a single global scalar. Does the input-dependent score matter?

  4. How does the learned gain differ from FilterNet's plain shaping filter, and what is the result of replacing it with that filter in the same backbone?

  5. How are ρ^max and ρ̄ computed in Figures 4 and 12, given that Eq. 6 is bounded below by about 1? Are they per window or per dataset? Is the filter applied before or after RevIN?

  6. On real data, is the gain from protection concentrated on high-P_c channels or windows? A per-channel error breakdown against P_c would test the central claim.

  7. Were baseline numbers re-run (Appendix L.1 says so) or taken from the original papers? Several match the published values exactly.

3 comments

Part 1/2

1. Is the raw spectrum concatenated with the filtered one as input to the Transformer (Z_cat with 2F bins)? If yes, what does the "over-suppression" mean in this model, and what happens if only Z_flt is used?


Yes. WiSPformer concatenates ZrawZ_{\mathrm{raw}} and ZfltZ_{\mathrm{flt}}, giving ZcatZ_{\mathrm{cat}} with (2F2F) spectral entries. In our comparison, using only (ZfltZ_{\mathrm{flt}}) gives slightly worse performance than concatenation.

Over-suppression refers to excessive attenuation in the filtered branch, which can make useful structures less accessible to the forecasting backbone. In Table 4, we only use ZfltZ_{\mathrm{flt}} as an input to WiSPformer, Wiener without Z_raw because we need to compare with RAW configuration. It provides the evidence for the effectiveness of the WiSPformer filtered-only input (ZfltZ_{\mathrm{flt}}) which limits over-suppression.

Moreoever, in Table 2, WiSP with removed dual-branch transformer provides the inputs ZfltZ_{\mathrm{flt}} to FreTS, TimeKAN, and iTransformer. On Electricity, for example, average MSE across the four prediction horizons decreases from 0.200 to 0.183 for FreTS, 0.195 to 0.184 for TimeKAN, and 0.178 to 0.171 for iTransformer. These results show that WiSP improves forecasting across different backbones even without an additional raw input. We will make this setup clearer in the revision.


2. Please report mean ± std over at least 3 seeds for Table 1 and Table 4, with the ablation run on all four horizons and with the same configuration as the main table. Why do the full-model numbers at horizon 96 differ between Table 4 and Table 6?


Thank you for pointing this out. Tables 1 and 6 use Zcat=[Zflt;Zraw]Z_{\mathrm{cat}} = [Z_{\mathrm{flt}}; Z_{\mathrm{raw}}], which performed slightly better than the filtered-only input ZfltZ_{\mathrm{flt}}. Table 4 compares Raw, Wiener, and WiSP without concatenation, avoiding an additional raw-spectrum input into Wiener and WiSP to ensure a fair comparision.

Our WiSPformer results in Table 1 and Table 6 are averaged across five seeds. We will report the standard deviations and clearly explain the input differences between Tables 4 and 6 in the revision.


3. Controls for the protection: (a) π_c as a free per-channel parameter with no input statistics; (b) P_c shuffled across channels or windows; (c) a single global scalar. Does the input-dependent score matter?


Eq. (9) combines PcP_c with a learned channel-specific baseline to obtain this coefficient. Replacing only PcP_c would leave that baseline active, which is particularly problematic for the global-scalar control case. Therefore, we apply the controls to λcprot\lambda_c^{\mathrm{prot}} because it directly determines the protection added to the learned Wiener gain in Eq. (10): Wc(k)−Hc(k)=λcprot(1−Hc(k)).W_c(k) - H_c(k) = \lambda_c^{\mathrm{prot}} \bigl(1 - H_c(k)\bigr).

As requested, we evaluate three control configurations:

  1. Learn per-channel πc\pi_c: Replace λcprot\lambda_c^{\mathrm{prot}} with a learnable πc\pi_c, shared across all windows of channel cc.
  2. Channel shuffle: Compute λcprot\lambda_c^{\mathrm{prot}} normally, then shuffle it across channels.
  3. Learn a global scalar π\pi: Replace λcprot\lambda_c^{\mathrm{prot}} with a single learnable scalar π\pi, shared across all channels.

Results (96 →\rightarrow 96, formatted as MSE / MAE)

SettingElectricityTraffic
Wiener-only0.142 / 0.2300.397 / 0.239
(1) Learned per-channel (πc\pi_c)0.138 / 0.2260.396 / 0.237
(2) Channel-shuffled0.146 / 0.2350.401 / 0.243
(3) Learned global scalar (π\pi)0.143 / 0.2320.399 / 0.241
WiSPformer0.132 / 0.2220.390 / 0.232

Per-channel constants retain some benefit on both datasets, but offer little improvement compared to Wiener-only. Channel shuffling degrades performance, and the global scalar also hurts results on both datasets.


4. How does the learned gain differ from FilterNet's plain shaping filter, and what is the result of replacing it with that filter in the same backbone?


Thank you for this suggestion. We could not complete this experiment within the rebuttal period, but will include it in the revised manuscript.

Part 2/2


5. How are ρ^max and ρ̄ computed in Figures 4 and 12, given that Eq. 6 is bounded below by about 1? Are they per window or per dataset? Is the filter applied before or after RevIN?


Thank you for pointing out the missing plotting detail. The statistics (Maximum Channel Deviation Δmax⁡c\Delta_{\max}^c and Average Change Intensity Δˉc\bar{\Delta}_c) in Eqs. (6)–(7) are computed separately for each input window and channel.

Before plotting, we independently normalised Δmax⁡c\Delta_{\max}^c and Δˉc\bar{\Delta}_c across the plotted values by min–max scaling: z~=z−min⁡(z)max⁡(z)−min⁡(z)+10−8.\tilde{z} = \frac{z - \min(z)}{\max(z) - \min(z) + 10^{-8}}.

Thus, the 0–1 axes in Figures 4 and 12 show rescaled statistics, not the raw value in Eq. (6). The protection scores retain the model outputs.

RevIN is applied on the inputs before our models, and inverse RevIN restores the predictions to the original scale. We will clarify this RevIN order and label the rescaled axes explicitly in the revision.


6. On real data, is the gain from protection concentrated on high-P_c channels or windows? A per-channel error breakdown against P_c would test the central claim.


Figure 12 shows the distribution of protection score PcP_c. Most channels have low protection scores PcP_c.

To break down how our WiSP model benefits from the protection, we illustrate the data fraction and MSEs in the table below.

DatasetProtectData fractionWiener MSEWiSP MSE
ElectricityLow: <0.4<0.488.8%0.1340.128
ElectricityMid: 0.4–0.70.4–0.75.9%0.1730.129
ElectricityHigh: >0.7>0.75.3%0.2410.187
ElectricityAll (Table 4)100%0.1420.132
TrafficLow: <0.4<0.497.1%0.3940.391
TrafficMid: 0.4–0.70.4–0.72.4%0.4970.349
TrafficHigh: >0.7>0.70.5%0.4980.396
TrafficAll (Table 4)100%0.3970.390

The above table indicates that WiSP would show larger improvements in the medium and high protection groups, while the low protection group would remain closer to the Wiener model. We can find that the medium group benefits more than the high group on Traffic, and on Electricity both groups would benefit from the protection. We will report this part in the revision.


7. Were baseline numbers re-run (Appendix L.1 says so) or taken from the original papers? Several match the published values exactly.


We reran most baselines using their official implementations. However, some papers did not provide the hyperparameter configurations for every dataset. For those situations, we used the corresponding results reported in the original papers. We will clarify this in Appendix L.1 and label the source of each baseline result, distinguishing our reruns from cited values.

We hope these clarifications and additional results address your concerns and encourage you to consider raising your score.

Thanks for the detailed response and additional experiments. However, some of my key concerns remain unsolved. After reviewing the updated results, I am not convinced the performance improvements provided by WiSPformer are significant compared to existing solutions. Thus, while I acknowledge the improvements made during the rebuttal period, I retain my score.

Reviewer_1QvQOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes WiSPformer, a Wiener-informed signal-preserving Transformer for multivariate time-series forecasting. The authors observe that Wiener-style spectral denoising, while effective at suppressing background noise, over-attenuates informative local transients (steps, pulses, jagged fluctuations) because these transients manifest as coordinated energy spread across many frequency bins, each of which is individually shrunk according to its global spectral SNR. WiSPformer addresses this via a learnable Wiener filter, a signal-preserving channel-wise protection mechanism, and a dual-branch complex spectral Transformer. Experiments cover 9 real-world benchmarks against 12 baselines, a synthetic Step/Pulse/Jagged dataset with ground-truth SNR analysis, and a plug-in transferability study.

Critical strengths
  1. Well-motivated and empirically demonstrated problem\textbf{Well-motivated and empirically demonstrated problem}. The over-smoothing of predictive transients by spectral denoising is a real failure mode; Figure 1 and the synthetic study make the case convincingly.

  2. Principled Wiener-consistent parameterization\textbf{Principled Wiener-consistent parameterization}. The sigmoid-of-log-SNR gain is derived as the exact optimum of the reconstruction objective, with a gradient analysis explaining end-to-end learnability (Appendix E.2). This is more principled than the ad hoc learnable masks in several baselines.

  3. Broad and consistent empirical evaluation\textbf{Broad and consistent empirical evaluation}. 9 datasets, 12 baselines, best in 14/18 main cases, plus a transferability study demonstrating architecture-agnosticism.

Critical weaknesses
  1. Moderate novelty relative to the frequency-filtering literature.

Learnable Wiener-style spectral filtering is well established (FilterNet, FiLM, FreTS, Amplifier); the dual-branch real/imaginary Transformer follows FreEformer/FreTS closely; the calibration layer is a complex linear map. The genuinely new element — channel-wise protection — is a relatively small, heuristic addition.

  1. The protection score is untested under adversarial noise.

No experiment varies the noise level to check whether high −Δˉc-\bar{\Delta}_c noisy channels are incorrectly protected. This directly threatens the method's central claim of "suppressing noise while preserving transients." A plot of Pc\mathcal{P}_c and of MSE(WiSP) vs. MSE(Wiener) as a function of injected SNR is essential.

  1. No component-wise ablation.

Lack of component-wise ablation makes it impossible to attribute gains among the three contributions or to determine whether the dual-branch Transformer is necessary (versus a single-branch complex MLP).

  1. Unexplained per-dataset behavior.

WiSPformer does not beat xPatch on ETTh1/ETTh2, and the average gains are driven by Exchange/Traffic/Solar/ECL. This paper does not analyze whether this correlates with transient richness (e.g., average Pc\mathcal{P}_c per dataset). Such an analysis would either strengthen the story or reveal a limitation.

  1. Incomplete reproducibility details.

λcfilter\lambda^\text{filter}_c parameterization, τc\tau_c initialization, layer/head counts, and whether Pc\mathcal{P}_c is computed per-window or pooled are unspecified in the main text.

  1. The synthetic experiments are self-confirming by construction and cannot support the generalization claims drawn from them.

In the synthetic dataset, Step/Pulse/Jagged transients are explicitly defined as part of the clean signal, while noise is independent additive white Gaussian noise. Under this construction, "protecting transients = protecting signal" holds by definition, and classical Wiener filtering is guaranteed to fail, so the experiment cannot in principle produce a negative result. The same issue affects the log-SNR correlation analysis (Figure 9): since the gain is parameterized as Hc(k)=σ(log⁡ρ^c(k))H_c(k)=\sigma(\log\hat{ρ}_c(k)), a positive correlation with the ground-truth log-SNR is largely a structural consequence of the parameterization rather than an independent empirical finding. Consequently, neither the synthetic results nor the main-table results on real data can distinguish whether gains arise from preserving predictive transients or merely from attenuating less overall.

Questions for authors

See the weaknesses.

4 comments

Part 1/3

1. Moderate novelty relative to the frequency-filtering literature. Learnable Wiener-style spectral filtering is well established (FilterNet, FiLM, FreTS, Amplifier); the dual-branch real/imaginary Transformer follows Freeformer/FreTS closely; the calibration layer is a complex linear map. The genuinely new element — channel-wise protection — is a relatively small, heuristic addition.


Our main contribution is the WiSP filtering module: learned Wiener-style filter combined with input-dependent attenuation protection. The comparison below clarifies how this relates to existing frequency-domain methods.

MethodFull-spectrum learningNoise suppressionInput adaptationTransient preservationBackbone flexibility
FiLM—✓\checkmark——✓\checkmark
FreTS✓\checkmark✓\checkmark———
FilterNet✓\checkmark✓\checkmark——✓\checkmark
Amplifier✓\checkmark————
Freeformer✓\checkmark————
Learnable Wiener (ours)✓\checkmark✓\checkmark——✓\checkmark
WiSP (ours)✓\checkmark✓\checkmark✓\checkmark✓\checkmark✓\checkmark

Column definitions

  • Full-spectrum learning: Learns from the full frequency range without restricting processing to a predefined subset of frequency modes.
  • Noise suppression: Includes processing intended to reduce the influence of noisy components on forecasting, through explicit filtering or implicit suppression.
  • Input adaptation: Adjusts filtering weights according to the current input window.
  • Transient preservation: Includes a dedicated mechanism to limit excessive attenuation of local changes, such as peaks, pulses, and steps.
  • Backbone flexibility: Provides a component that can be combined with different forecasting backbones, with interface adaptation where needed.

Notes ✓\checkmark indicates a described design purpose or mechanism, not guaranteed forecasting performance.


2. The protection score is untested under adversarial noise. No experiment varies the noise level to check whether high noisy channels are incorrectly protected. This directly threatens the method's central claim of "suppressing noise while preserving transients." A plot of and of MSE(WiSP) vs. MSE(Wiener) as a function of injected SNR is essential.


We added a controlled forecasting experiment with randomly positioned pulse noise at 20 dB and 10 dB, together with a no-noise control. Raw, learnable Wiener-only, and WiSP use the same dual-branch Transformer backbone.

Their MSE/MAE and Protection are in the table below:

InputWiSP MSE/MAEWiener MSE/MAERaw MSE/MAEProtection score
No noise0.058/0.1850.064/0.1970.056/0.1820.687
Random pulse noise, 20 dB0.062/0.1880.068/0.1990.073/0.2080.571
Random pulse noise, 10 dB0.066/0.1960.074/0.2120.101/0.2310.558

WiSP achieves the lowest MSE among the variants at both noise levels, improving over learnable Wiener-only by 8.8% at 20 dB and 10.8% at 10 dB.

Figures 4 and 12 show that protection depends on both Maximum Channel Deviation and Average Change Intensity. In our experiment, adding pulse noise decrease the Maximum Channel and can increase Average Change Intensity. The measured average protection decreases as pulse-noise power increases.

Part 2/3


3. No component-wise ablation. Lack of component-wise ablation makes it impossible to attribute gains among the three contributions or to determine whether the dual-branch Transformer is necessary (versus a single-branch complex MLP).


Our main contributions are the learnable Wiener filter in Eqs. (3)–(4), the protection mechanism in Eqs. (8)–(10), and the complex-consistent calibration layer in Eq. (18). The dual-branch Transformer builds on earlier spectral forecasting designs, including FreEformer, as discussed in Section 3.2. Table 2 addresses whether WiSP needs our Transformer backbone. We also remove the dual-branch Transformer and feed the WiSP-filtered input directly to FreTS, TimeKAN, or iTransformer.

We will make this distinction clearer in the contribution statement. Our paper already includes component-wise ablations:

Component testedComparisonSolar MSE (96→9696 \rightarrow 96)
Learnable Wiener filteringRaw →\rightarrow Wiener0.180→0.1690.180 \rightarrow 0.169
ProtectionWiener →\rightarrow WiSPformer0.169→0.1610.169 \rightarrow 0.161
CalibrationUncalibrated →\rightarrow WiSPformer0.166→0.1610.166 \rightarrow 0.161
Branch structureShare →\rightarrow WiSPformer0.172→0.1610.172 \rightarrow 0.161

The table above already evaluates these components: Raw vs. Wiener isolates learned filtering, Wiener vs. WiSPformer isolates protection, and Uncalibrated vs. WiSPformer isolates calibration. Additionally, the Share variant evaluates a single-branch Transformer.


4. Unexplained per-dataset behavior. WiSPformer does not beat xPatch on ETTh1/ETTh2, and the average gains are driven by Exchange/Traffic/Solar/ECL. This paper does not analyze whether this correlates with transient richness (e.g., average per dataset). Such an analysis would either strengthen the story or reveal a limitation.


xPatch predicts each channel independently. WiSPformer uses dual-channel transformer backend to obtain information across channels. Dramatic changes in the cross-channel relationships may therefore decrease our model’s performance.

We examined whether cross-channel relationships remain stable between training and testing in ETTh1 and ETTh2. Some relationships weaken substantially, while others reverse:

DatasetChannel pairTraining correlationTest correlation
ETTh1HULL–OT0.6010.601−0.004-0.004
ETTh1MULL–OT0.5240.524−0.043-0.043
ETTh2MULL–OT0.3490.3490.0530.053
ETTh2HUFL–MUFL−0.102-0.1020.9340.934

The Pearson correlations show large changes in the data relationships, particularly OT channel. To test whether including the OT channel affects the remaining channels and model performance, we retrained both models (xPatch and WiSPformer) without OT and compared the retrained six-channel models with the original seven-channel models.

Results (Each entry is MSE / MAE):

DatasetTraining channelsEvaluation channelsWiSPformerxPatch
ETTh1All sevenSix load channels0.410/0.4120.410 / 0.4120.414/0.4090.414 / 0.409
ETTh1Six, excluding OTSame six load channels0.402/0.4060.402 / 0.4060.413/0.4070.413 / 0.407
ETTh2All sevenSix load channels0.340/0.3710.340 / 0.3710.312/0.3560.312 / 0.356
ETTh2Six, excluding OTSame six load channels0.320/0.3620.320 / 0.3620.309/0.3550.309 / 0.355

Removing OT would reduce WiSPformer's six-channel MSE by approximately 2.0% on ETTh1 and 5.9% on ETTh2, compared with 0.2% and 0.3% for xPatch. On ETTh2, the MSE gap would narrow from 0.028 to 0.011.

Traffic, Solar, Exchange and Electricity show stronger and more stable cross-channel correlations than ETTh1 and ETTh2 .

DatasetChannelsTrain mean correlationTest mean correlation
ETTh170.3110.259
ETTh270.3490.391
Traffic8620.5790.566
Solar1370.9140.917
Electricity3210.4970.499
Exchange80.4840.705

5. Incomplete reproducibility details. parameterization, initialization, layer/head counts, and whether is computed per-window or pooled are unspecified in the main text.


Attention heads: 8 per layer. Model dimension: {256,512}. Embedding dimension:16. Maximum Channel Deviation Δmax⁡c\Delta_{\max}^c and Average Change Intensity Δˉc\bar{\Delta}_c: computed per input window.

We will add the parameter settings and initialization values to the revised paper.

Part 3/3

6. The synthetic experiments are self-confirming by construction and cannot support the generalization claims drawn from them. In the synthetic dataset, Step/Pulse/Jagged transients are explicitly defined as part of the clean signal, while noise is independent additive white Gaussian noise. Under this construction, "protecting transients = protecting signal" holds by definition, and classical Wiener filtering is guaranteed to fail, so the experiment cannot in principle produce a negative result. The same issue affects the log-SNR correlation analysis (Figure 9): since the gain is parameterized as , a positive correlation with the ground-truth log-SNR is largely a structural consequence of the parameterization rather than an independent empirical finding. Consequently, neither the synthetic results nor the main-table results on real data can distinguish whether gains arise from preserving predictive transients or merely from attenuating less overall.

In the additional pulse-noise experiments below, Raw, Wiener, and WiSP all share the same dual-branch Transformer architecture and differ only in their spectral inputs:

  • Raw: uses the unfiltered input without Wiener filtering or protection.
  • Wiener: applies our learnable Wiener filter from Eqs. (3)–(4) without protection.
  • WiSP: combines the learnable Wiener filter with the protection mechanism in Eqs. (8)–(10).

Their spectral weights follow Eqs. (3)–(4) and are trained through the MSE loss. Following Eq. (4), Figure 9 compares the learned log⁡ρ^c(k)\log \hat{\rho}_c(k) with the ground-truth log⁡ρc(k)\log \rho_c(k), computed separately from the clean signal and injected noise. While Eq. (4) guarantees that the gain increases with its learned log-SNR parameter, it does not guarantee that this parameter correlates with the ground-truth log-SNR. The observed correlation is therefore an empirical finding.

To examine whether pulse-shaped noise triggers protection and how this affects prediction, we replaced the Gaussian noise with two randomly positioned noise pulses per input window and channel, while keeping the clean signals and forecasting targets unchanged. We evaluated SNRs of 20 dB and 10 dB alongside a no-noise control.

The results are shown in the table above (Q2). Raw performs slightly better on clean (no noise) inputs, while WiSP stays close to RAW and outperforms Wiener-only. Under random pulse noise, WiSP achieves the lowest MSE and MAE at both noise levels, reducing MSE relative to Wiener-only by 8.8% at 20 dB and 10.8% at 10 dB. These results support the usefulness of protection for forecasting even when the interference itself is pulse-shaped.

We hope these clarifications and additional results address your concerns and encourage you to consider raising your score.

I thank the authors for the clarity of their responses and for the additional information provided. This has allowed me to better appreciate certain points; however, it does not lead me to change my rating, which I consider fair given the overall criticisms raised regarding the submitted manuscript.

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

The paper proposes WiSPformer, a frequency-domain model for multivariate time series forecasting. The main idea is to reduce noise while preserving useful local changes that might otherwise be lost during spectral filtering. To achieve this, the authors combine a learnable Wiener filter, an adaptive signal-preserving mechanism, and a dual-branch spectral Transformer for modeling cross-channel dependencies. Experiments on nine benchmark datasets show that WiSPformer achieves competitive forecasting performance, and the ablation studies further demonstrate the benefits of its main components.

Critical strengths
  1. The paper addresses an interesting problem in frequency-domain time series forecasting, where spectral denoising may also suppress useful local patterns.
  2. The paper evaluates WiSPformer on nine benchmark datasets and shows competitive results against several baselines. The authors also provide ablation studies, synthetic experiments, and tests on different forecasting models to further support the effectiveness of their method.
Critical weaknesses
  1. Although the authors compare their method with 12 baselines, the comparisons could be further strengthened by including more recent models from ICLR 2026 and ICML 2026, such as frequency-domain models like DMANet[1] and other recent approaches like GTR[2] and Phaseformer[3].
  2. Given that the main motivation is to balance noise suppression and informative transient preservation, it would be helpful to further evaluate WiSPformer under different noise levels and noise types.

References:

[1] Anti-Aliasing Matters: A Dynamic Network for Time Series Forecasting (ICML 2026)

[2] ENHANCING MULTIVARIATE TIME SERIES FORECASTING WITH GLOBAL TEMPORAL RETRIEVAL (ICLR 2026)

[3] PHASEFORMER: FROM PATCHES TO PHASES FOR EFFICIENT AND EFFECTIVE TIME SERIES FORECASTING (ICLR 2026)

Questions for authors
  1. As the authors acknowledge in Lines 054–055, “Wiener filtering still relies on global spectral statistics, and informative local transients may be over-suppressed when their global SNR is low.” Given this limitation, why do the authors still choose Wiener filtering as the basis for their model design? Could the authors further justify this choice and explain why other filtering methods were not considered?
  2. According to the visualization results in Figure 1, the authors claim that “WiSPformer retains more energy in the low- and mid-frequency range (WiSP).” Could a simple frequency truncation strategy that retains only low- and mid-frequency components achieve similar results? I suspect that it might. If so, what additional benefits does WiSPformer provide over such a straightforward approach?
  3. How is “X Clean” in Figure 1 obtained? Is it generated by retaining only the low- and mid-frequency components of “X Noisy”?
  4. How do the authors distinguish between background noise and informative local transitions?
2 comments

Part 1/2

1. As the authors acknowledge in Lines 054–055, “Wiener filtering still relies on global spectral statistics, and informative local transients may be over-suppressed when their global SNR is low.” Given this limitation, why do the authors still choose Wiener filtering as the basis for their model design? Could the authors further justify this choice and explain why other filtering methods were not considered?


Classical Wiener filtering calculates weights from estimated signal and noise spectra. In contrast, our learnable Wiener filter learns weights directly from the forecasting MSE loss without estimating the two spectra separately. The learnable components therefore determine attenuation at each frequency bin based on the forecasting objective rather than a separate signal-reconstruction criterion. This enables the model to reduce noise while retaining components that contribute to forecasting.

However, we observed that the learned weights still exhibit some over-suppression in Figure 1: the learned Wiener-only filter weakens the pulse, causing the forecast to underestimate its amplitude. This observation motivated our protection mechanism, which adjusts the attenuation based on the variation within the current input window . For example, a brief demand increase when equipment starts may carry useful information about the subsequent load and should not be smoothed away. Table 3 supports this adjustment: adding protection reduces Pulse forecasting MSE from 0.075 to 0.065 .

We will revise the text to clearly distinguish classical Wiener filtering, which relies on estimated signal and noise spectra, from our learnable filter, whose weights are trained directly via the forecasting loss.


2. According to the visualization results in Figure 1, the authors claim that “WiSPformer retains more energy in the low- and mid-frequency range (WiSP).” Could a simple frequency truncation strategy that retains only low- and mid-frequency components achieve similar results? I suspect that it might. If so, what additional benefits does WiSPformer provide over such a straightforward approach?


The low- and mid-frequency description in Figure 1 specifically illustrates the Pulse example. Rather than imposing a fixed frequency cutoff, WiSP learns an adaptive weight for each frequency bin and channel, dynamically adjusting the attenuation based on the current input window.

There are also examples of protection at higher frequencies. In Figure 11, the Weather case with (P_c=0.875) retains more spectral magnitude around bin 30 than Wiener-only filtering. The Traffic example in Figure 14(d), with (P_c=0.939), shows the same behavior.

To check whether this protection contributes to prediction, we can remove some frequency bands (30-48) in the high-protected channels while keeping the trained model unchanged. The table below gives the comparisons.

Configuration, 96→9696 \rightarrow 96Weather MSE/MAETraffic MSE/MAE
Full WiSP0.148 / 0.1860.387 / 0.231
Protection removed from bins 25−3025-300.150 / 0.1880.391 / 0.235
Protection removed from bins 30−4830-480.152 / 0.1880.394 / 0.238
Wiener-only0.153 / 0.1890.397 / 0.239

3. How is “X Clean” in Figure 1 obtained? Is it generated by retaining only the low- and mid-frequency components of “X Noisy”?


X Clean is the ground-truth signal from our synthetic generator, not a filtered version of X Noisy. We first generate a background consisting of two sinusoids and a mild linear trend, add Step, Pulse, and Jagged components, respectively, and normalize the resulting signal by its standard deviation. This gives X Clean. We then add independent white Gaussian noise to obtain X Noisy. The generation process is described in Appendix H.1.

Part 2/2


4. How do the authors distinguish between background noise and informative local transitions?


The model jointly learns both the frequency weights and the sensitivity to local changes. As discussed in Appendix F.1, the Wiener filter determines which frequency bins are reliable, whereas the protection mechanism determines whether the current channel contains predictive transient information that should not be excessively suppressed.

Moreover, as shown in Appendix F.2, although channel-wise protection relaxes the overall attenuation strength, it preserves the relative attenuation ordering established by the Wiener filter. Frequency components with higher Wiener gains continue to experience less attenuation, whereas components with lower Wiener gains remain more strongly suppressed.

Furthermore, in Table 2, we integrated WiSP into different backbones: FreTS, TimeKAN, and iTransformer and compared each against its original counterpart. These models take the WiSP-filtered data as input. The resulting performance gains demonstrate that the WiSP remains effective across diverse forecasting backbones.

Lastly, to examine whether pulse-shaped noise triggers protection under WiSP and how this affects prediction, we replaced the Gaussian noise with two randomly positioned noise pulses, while keeping the clean signals and forecasting targets unchanged. We evaluated SNRs of 20 dB and 10 dB alongside a no-noise control.

Their MSE/MAE and Protection are in the table below:

InputWiSP MSE/MAEWiener MSE/MAERaw MSE/MAEProtection score
No noise0.058/0.1850.064/0.1970.056/0.1820.687
Random pulse noise, 20 dB0.062/0.1880.068/0.1990.073/0.2080.571
Random pulse noise, 10 dB0.066/0.1960.074/0.2120.101/0.2310.558

In this experiment, WiSP achieves the lowest MSE among the variants at both noise levels, improving over learnable Wiener-only by 8.8% at 20 dB and 10.8% at 10 dB.

Figures 4 and 12 show that protection depends on both Maximum Channel Deviation and Average Change Intensity. adding pulse noise decrease the Maximum Channel and can increase Average Change Intensity. The measured average protection decreases as pulse-noise power increases.

We hope these clarifications and additional results address your concerns and encourage you to consider raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.