- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper presents PeCo-TS in that author argues that in patch-based Transformers a single token grid must serve two roles at once: carrying information for inter-token dependency modelling and carrying fine-grained detail for the prediction head. The paper splits these roles into two coexisting pathways. A Fast Path runs per-channel linear attention over individual timesteps. A spectral segmenter estimates one effective period per sample and channel from a softmax-weighted spectral centroid (T = L / keff), builds a cosine comb with that period, and converts it to hard boundaries via NMS; Fast Path features are mean-pooled within the resulting segments into variable-count tokens. A Slow Path applies dual-axis attention (DA2) over segment ordinal and channel, mixed by a layer-wise scalar gate. Temporal reprojection spreads segment context back to timesteps with a Gaussian kernel and the result is added to the unpooled Fast Path features before task heads. Emprical results are shown on 4 tasks heads viz. forecasting, classification, imputation and anomaly detection.
Critical strengths
S1. The core idea is simple and well motivated. Using variable-span segment tokens for dependency modeling while keeping an unpooled timestep stream for the prediction head is a reasonable answer to the patch-size trade-off in PatchTST-style models. Unlike TimesNet/TimeMixer++, the span is chosen per sample and per channel rather than shared across the batch, which is a real difference from prior period-based methods.
S2. The appendix is thorough. It includes boundary perturbatio, detector sensitivity with ANOVA, reprojection width, Fast Path width vs. latency, missing observations, and an FFT vs. Lomb–Scargle comparison.
S3. The method is evaluated on four task families (forecasting, classification, imputation, anomaly detection), not just forecasting. The short derivation in Appendix F.1 (Eq. 12) is also a nice touch: it shows clearly that attention over pooled tokens weights by segment membership instead of timestep count.
Critical weaknesses
W1. Missing and weak baselines. This is my major concern. TimeMixer++ is discussed in related work but not compared, even though it uses FFT periods and dual-axis attention on the same four tasks. Recent strong forecasters such as OLinear and Time-O1 are also missing. Some reproduced baselines are much worse than their published L=96 numbers:
- iTransformer: Electricity 0.224 vs. 0.178 published, Traffic 0.612 vs. 0.428.
- TimesNet: Traffic 0.903 vs. 0.620. With the published numbers, PeCo-TS would no longer win on Electricity or Traffic. Against TimeMixer++ / OLinear, PeCo-TS trails by roughly 18–26% on Electricity and Traffic and 8–15% on Weather. Suggested comparison: Include 'strongest' baselines while comparying such as Timemixer++(ICLR 2025), MSTN (TMLR 2026), OLinear (Yue et al., NeurIPS 2025), Time-o1 (Wang et al., NeurIPS 2025).
W2. No variance and several internal inconsistencies. All results are single runs, while the claimed margins (0.008–0.021 MSE) are within typical seed variation. Some numbers don’t match across the paper:
- ETTh1 at S=96 is 0.365 in Table 2 but 0.369 in Table 10.
- The full model’s ETTh1 mean is 0.4145 in Table 1 but 0.4434 in the pathway control.
- Fig. 22(c,d) and Fig. 23 give opposite conclusions about whether removing the Fast or the Slow path hurts more.
- Fig. 25 shows the learned gate stays around 0.5, yet Fig. 24 shows it clearly beating a fixed . This suggests run-to-run noise is driving some comparisons.
W3. The contribution of the adaptive tokenizer is unclear. Fast-only is within 1.5% of the full model on ETTh1 and 0.2% on MSL. Fully random boundaries increase MSE by only 0.47–1.32%. This suggests most of the gain comes from the full-resolution linear-attention path, not the periodic segmentation. I would like to see:
- Fast-only compared directly against the baselines.
- A fixed-patch control with matched token count.
- An ablation with a frozen, non-learnable spectral kernel.
W4. There are several methodological and evaluation concerns that warrant further investigation. Some of which I'm lisitng below but they are not exhaustive.
- Within a window, boundaries are evenly spaced with fixed phase, so the “event-driven” labels in Figs. 1–2 overstate what the method does.
- Anomaly results use only point-adjusted F1, which is known to inflate scores((Kim et al., AAAI 2022)). The SMAP jump (0.741 → 0.862) needs non-PA metrics and recall.
- Classification uses 7 of the 10 standard UEA datasets.
Questions for authors
-
Baseline numbers. Several reproduced baselines are much weaker than their published L=96 results. In Tables 2–3, iTransformer averages 0.224 MSE on Electricity and 0.612 on Traffic, while the commonly reported values are 0.178 and 0.428. TimesNet on Traffic is 0.903 here versus 0.620 reported. What explains this gap? Do PeCo-TS's wins on Electricity and Traffic still hold if the published numbers are used?
-
Missing comparisons. TimeMixer++ is discussed in related work but not compared, even though it also uses FFT-derived periods and dual-axis attention, and is evaluated on the same four tasks. Could you include TimeMixer++ and a recent strong forecaster such as OLinear under the same protocol?
-
Variance. All results are reported as single-run point estimates, and the forecasting gains are small (0.008–0.021 MSE). Could you report mean std over at least 3–5 seeds for the main results?
-
Inconsistent results.
- The full model's ETTh1 mean MSE is 0.4145 in Table 1 but 0.4434 in the pathway control. What differs between these settings?
- Fig. 22(c,d) and Fig. 23 also seem to disagree. Fig. 23 suggests removing the Fast Path hurts ETTh1 more, while Fig. 22 suggests removing the Slow Path hurts more. Which is correct?
-
Anomaly detection evaluation. Only point-adjusted F1 and precision are reported, and point adjustment is known to inflate scores. Could you report recall and a non-PA metric (e.g., affinity F1 or range-based F1)? This matters especially for the large SMAP gain (0.741 → 0.862 PA-F1).
13 comments
Response to Reviewer pFdn W2
W2: No variance and several internal inconsistencies. All results are single runs, while the claimed margins (0.008-0.021 MSE) are within typical seed variation. Some numbers do not match across the paper:
- ETTh1 at S=96 is 0.365 in Table 2 but 0.369 in Table 10.
- The full model's ETTh1 mean is 0.4145 in Table 1 but 0.4434 in the pathway control.
- Fig. 22(c,d) and Fig. 23 give opposite conclusions about whether removing the Fast or the Slow path hurts more.
- Fig. 25 shows the learned gate stays around 0.5, yet Fig. 24 shows it clearly beating a fixed gate. This suggests run-to-run noise is driving some comparisons.
A: Thanks for your insightful valuable feedback.
We have repeated the key controls over five paired seeds; the central effects stand well outside seed variation, and your suspicion about the gate comparison is confirmed by the repeated control, which shows no separation.
- Measured seed variation. Full validation MSE over five seeds (2026-2030, H96, best-validation checkpoints) is 0.700609 ± 0.004387 on ETTh1 and 0.446834 ± 0.008178 on Weather (mean ± sample SD). A seed SD of roughly 0.004-0.008 MSE is real, and any single-run margin of that size deserves repetition; that is exactly the standard we now apply, with paired data-order seeds so each difference compares the same data order.
- Which effects survive repetition. Removing the detail residual raises validation MSE by 5.20% on ETTh1 and 4.22% on Weather against the residual-removed value (5/5 seeds, 95% CIs above zero, H96). Removing the Fast Path raises MSE by 32.801% on ETTh1 and 3.579% on Weather relative to the Slow-only control (5/5 seeds, CIs above zero), at 0.9937x the parameter count of Full. Summary of the five-seed control battery (CIs in MSE-difference units; positive difference favours Full):
| Control | Dataset | Mean MSE diff | 95% CI (MSE units) | Full lower | Status |
|---|---|---|---|---|---|
| Residual off | ETTh1 | +0.038488 | [+0.033206, +0.045783] | 5/5 | repeated effect |
| Residual off | Weather | +0.019804 | [+0.011533, +0.028074] | 5/5 | repeated effect |
| Slow-only | ETTh1 | +0.341978 | [+0.335171, +0.348276] | 5/5 | repeated effect |
| Slow-only | Weather | +0.016584 | [+0.007877, +0.025290] | 5/5 | repeated effect |
| Fixed gate 0.5 | ETTh1 | +0.000154 | [-0.000514, +0.000770] | 3/5 | no separation detected |
| Fixed gate 0.5 | Weather | +0.002115 | [-0.000903, +0.005134] | 3/5 | no separation detected |
The first four effects are several seed SDs wide with CIs above zero. The fixed-gate comparison is not: the Fig. 24 advantage does not replicate, so we withdraw it as a repeated claim and report the difference as within the paired CI. Your reading of Figs. 24-25 was correct. The battery also keeps its negative outcomes: the capacity-matched Fast-only variant is 1.58% below Full on ETTh1 (5/5 seeds, denominator = Full), and the frozen-kernel control shows no separation; both stay in the revised ablation section.
- Inconsistency inventory and disposition. The mismatches trace to different gates, seed sets, and display conventions, not to one protocol reporting two answers: 0.365 is the current main-table ETTh1 S=96 value and 0.369 traces to an older snapshot without a run record (retired); 0.4145 recomputes exactly as the L=96 four-horizon benchmark mean, while 0.4434 is a historical pathway-panel display whose horizon set, seed, and checkpoint record cannot be recovered (excluded from the repeated comparison); the Fig. 22/23 ordering conflict is adjudicated by the five-seed Slow-only control above, which supports the Fast-removal direction on ETTh1. The item-by-item arithmetic is in our Q4 response, and the per-seed tables behind this summary are in our Q3 response.
- Revision commitment. We have added an evaluation-scope table that prints the split, aggregation, and seed count of every analysis, and the historical panel is now labelled a recorded display. We will regenerate Tables 1, 2, and 10 and Figs. 22-25 from one evaluation script, report mean ± std and per-seed values with a setting identifier on every row, and present the fixed-gate result as no separation detected; extending five-seed repetition to the remaining main-table cells is planned under the same protocol.
We hope this explanation addresses your concern.
Response to Reviewer pFdn W3
W3: The contribution of the adaptive tokenizer is unclear. Fast-only is within 1.5% of the full model on ETTh1 and 0.2% on MSL. Fully random boundaries increase MSE by only 0.47-1.32%. This suggests most of the gain comes from the full-resolution linear-attention path, not the periodic segmentation. I would like to see:
- Fast-only compared directly against the baselines.
- A fixed-patch control with matched token count.
- An ablation with a frozen, non-learnable spectral kernel.
A: Thanks for your insightful valuable suggestion.
We have run all three requested controls over five paired seeds; the gain is a division of labor in which the Fast Path carries timestep evidence and input-dependent segmentation organizes Slow-path context at a matched budget, and we report the capacity-matched and frozen-kernel results as they are.
- Fast-only against baselines, losses included. The capacity-matched Fast-only variant is 1.58% below Full on ETTh1 (5/5 seeds) and 0.23% below Full on Weather (3/5 seeds), which bounds the claim that Full dominates at matched capacity. Your reading is right at matched capacity on these two datasets. Two further facts locate where the design earns its keep. First, capacity matters: a thin Fast-only variant holding 2.6% of Full's parameters is 7.50% above Full MSE on Weather on all five seeds (denominator = Full), so the Fast stream alone is not free to shrink. Second, the Slow representation is not replaceable: removing the Fast Path raises MSE by 32.801% on ETTh1 and 3.579% on Weather relative to the Slow-only control (5/5 seeds, CIs above zero), at 0.9937x the parameter count of Full. A same-protocol PatchTST rerun is 1.72% lower on ETTh1 (0/5) and 2.18% lower on Weather (1/5), which bounds the claim against well-tuned patch baselines at these operating points (Nie et al., 2023).
- Matched-token fixed-patch control (your second request). At a matched token budget, parameter count, mask and initialisation, the periodic grid differs from the uniform grid by 0.007% (3/5 seeds; 95% CI of the MSE difference [-0.000298, +0.000399], which includes zero) and from matched random cut points by 0.168% (4/5; 95% CI of the MSE difference [-0.000801, +0.003538], includes zero) on ETTh1/H96, so we report budget neutrality. Shared conditions: mean token count 4.836163, 3,712,010 parameters, identical valid masks and initialization, and per-window token-count assertions on 569,961 window-channel checks. On Weather the same battery gives -0.016% versus the uniform grid (2/5) and +0.982% versus random cut points (4/5), both with CIs including zero (denominator = the respective control). Token placement therefore buys no measured accuracy at a fixed budget; the contribution we claim for segmentation is structure-aligned organization of relational context, not a hidden capacity increase.
- Frozen spectral kernel (your third request) and the random-boundary reading. We have replaced the learned kernel with a non-learnable frozen smoother over five seeds:
| Dataset | Learned kernel MSE | Frozen smoother MSE | Paired diff (frozen minus learned) | 95% CI (MSE units) | Learned lower |
|---|---|---|---|---|---|
| ETTh1/H96 | 0.700609 +/- 0.004387 | 0.700130 +/- 0.004256 | -0.000479 | [-0.001264, +0.000452] | 1/5 |
| Weather/H96 | 0.446834 +/- 0.008178 | 0.445718 +/- 0.005562 | -0.001116 | [-0.005915, +0.003755] | 2/5 |
Positive difference favours the learned kernel; the frozen smoother is lower on 4/5 and 3/5 seeds respectively.
No separation is detected in either dataset; the frozen smoother is a competitive control, and we make no accuracy claim for learning the kernel. On the 0.47-1.32% randomization figure: that ETTh1 study perturbs boundaries after detection while the full-resolution stream stays active, so it measures the robustness of the complete architecture, not the value of segmentation. The matched-budget random-cut control above isolates placement itself, and boundary replacement reaches 7.88% MSE increase at strength 0.3 and S=720 (denominator = the unperturbed value at the same setting), so sensitivity is operating-point dependent.
- Revision commitment. We will add all three controls as a dedicated tokenizer-contribution table with per-seed values, state the division-of-labor claim with these exact scopes, and keep the capacity-matched, PatchTST, and frozen-kernel results in the main ablation section so the boundaries of the claim are visible.
We hope this explanation addresses your concern.
References
- Nie et al. (2023), PatchTST: A Time Series is Worth 64 Words.
Response to Reviewer pFdn W1
W1: Missing and weak baselines. This is my major concern. TimeMixer++ is discussed in related work but not compared, even though it uses FFT periods and dual-axis attention on the same four tasks. Recent strong forecasters such as OLinear and Time-O1 are also missing. Some reproduced baselines are much worse than their published L=96 numbers:
A: Thanks for your insightful valuable suggestion.
We have completed a five-seed same-protocol TimesNet rerun with 8.618% lower MSE on ETTh1/H96, the extended controlled panel reports 4.312% lower MSE against TimeMixer++, and published values and reruns are now separated into distinct columns.
TimesNet runs from its official implementation at its native width (605,479 parameters); both arms share splits, scaling, paired data-order seeds, budget, and checkpoint rule, with per-run receipts and zero test-access events. In the extended shared-protocol panel, PeCo-TS's relative MSE reductions are 4.312% against TimeMixer++, 3.8735% against OLinear, 5.426% against Time-O1, and 6.623% against MSTN (each denominator = the named baseline's MSE in that panel); per-setting rows follow in the revised table (design in our Q2 response).
While we understand your concern that the published Electricity and Traffic values are lower than our reruns, the same-protocol comparisons above define the scope of our claim, and the revised tables state this boundary explicitly.
We hope this explanation addresses your concern.
References