- 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
-
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?
-
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?
-
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?
-
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?
-
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?
-
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.
-
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 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 and Average Change Intensity ) in Eqs. (6)–(7) are computed separately for each input window and channel.
Before plotting, we independently normalised and across the plotted values by min–max scaling:
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 . Most channels have low protection scores .
To break down how our WiSP model benefits from the protection, we illustrate the data fraction and MSEs in the table below.
| Dataset | Protect | Data fraction | Wiener MSE | WiSP MSE |
|---|---|---|---|---|
| Electricity | Low: | 88.8% | 0.134 | 0.128 |
| Electricity | Mid: | 5.9% | 0.173 | 0.129 |
| Electricity | High: | 5.3% | 0.241 | 0.187 |
| Electricity | All (Table 4) | 100% | 0.142 | 0.132 |
| Traffic | Low: | 97.1% | 0.394 | 0.391 |
| Traffic | Mid: | 2.4% | 0.497 | 0.349 |
| Traffic | High: | 0.5% | 0.498 | 0.396 |
| Traffic | All (Table 4) | 100% | 0.397 | 0.390 |
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.
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 Zraw and Zflt, giving Zcat with (2F) spectral entries. In our comparison, using only (Zflt) 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 Zflt 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 (Zflt) which limits over-suppression.
Moreoever, in Table 2, WiSP with removed dual-branch transformer provides the inputs Zflt 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], which performed slightly better than the filtered-only input Zflt. 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 Pc with a learned channel-specific baseline to obtain this coefficient. Replacing only Pc would leave that baseline active, which is particularly problematic for the global-scalar control case. Therefore, we apply the controls to λcprot because it directly determines the protection added to the learned Wiener gain in Eq. (10): Wc(k)−Hc(k)=λcprot(1−Hc(k)).
As requested, we evaluate three control configurations:
Results (96 → 96, formatted as MSE / MAE)
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.