acceptodds
Under review as a conference paper at ICLR 2027

Why Learn What You Can Compute? Closed-Form Operators for Periodic Time-Series Forecasting

Abstract

Periodic patterns are a defining characteristic of many real-world time series, and exploiting them has become central to modern long-horizon forecasting. Yet most existing learned forecasters that leverage such patterns still rely on iterative gradient-based optimization. We show that period folding transforms a period- wide-sense cyclostationary process into phase-indexed sequences with cycle-invariant first- and second-order statistics, and that multi-horizon forecasting on the resulting finite delay-coordinate dictionary is a Tikhonov-regularized Galerkin projection with a unique closed-form solution. Under ergodic sampling, the empirical estimator converges to its population projected counterpart, while exact Koopman closure arises as a special case when the folded process has an affine cycle-conditional mean. This theoretical result directly yields **CAPE** (**C**losed-form **A**pproximation of **P**eriodic **E**mbeddings), a gradient-free algorithm that estimates a compact cycle-level predictor, supports direct transfer when source and target datasets have compatible projected cycle structure, and adaptively reduces to a direct multi-output ridge autoregressor when . To evaluate CAPE, we conduct extensive experiments across zero-shot, few-shot, in-distribution, and cross-domain tasks. CAPE ranks first in 20 of 28 in-distribution MSE settings, wins seven of eight within-family and all three cross-domain zero-shot transfers, and achieves the best few-shot result, without gradient-based or GPU training and with under one second of CPU fitting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_YQ4pOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces CAPE, a no-gradient forecasting method for time series forecasting and zero-shot. They fold a time series based on a detected period and align observations by phase across cycles, and then estimate a forecasting operator through a single Tikhonov-regularized closed-form solution. They connect this construction to a finite-dimensional EDMD approximation of cycle-level Koopman evolution. The method is lightweight and empirically evaluated on standard forecasting and transfer benchmarks. The paper also includes useful ablations against ridge regression.

Critical strengths

Computational simplicity and efficiency that reduces forecasting to a small closed-form linear solution and avoids iterative gradient-based training.

Critical weaknesses

1 Figure 1 and Figure 2 are low quality and hard to understand. Especially for figure 2, there are no clear differences except color between the upper panels with the lower panels. Given that these figures introduce the main idea, I believe they should be substantially redesigned.

2 The theoretical contributions appear to be limited. The four parts of theorem 1 seem to be very standard. I don’t feel proposition 1 explains the “stronger zero-short transfer” more than comparing two different spectral filters. Besides, these two spectral filters are both smooth spectral filters.

3 The main Table 1 does not use a matched lookback length across methods. Besides, the baseline comparison is not sufficiently comprehensive to support a strong SOTA claim. Several relevant recent forecasting methods using periodic information are absent. For example, WaveletMixer and WPMixer use multi-resolution wavelet decomposition and mixing; D-PAD explicitly models multi-frequency decomposition and fusion. Without comparison against such methods, especially those targeting closely related multi-scale or frequency-structured forecasting, it is difficult to assess the claimed state-of-the-art advantage.

[1] Zhang, Zichi, et al. "WaveletMixer: a multi-resolution wavelets based MLP-mixer for multivariate long-term time series forecasting." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 39. No. 21. 2025.

[2] Murad, Md Mahmuddun Nabi, Mehmet Aktukmak, and Yasin Yilmaz. "Wpmixer: Efficient multi-resolution mixing for long-term time series forecasting." Proceedings of the AAAI conference on artificial intelligence. Vol. 39. No. 18. 2025.

[3] Yuan, Xiaobing, and Ling Chen. "D-pad: Deep-shallow multi-frequency patterns disentangling for time series forecasting." arXiv preprint arXiv:2403.17814 (2024).

4 The paper does not report the number of seeds. The reported number seems to be very close in Table 14, the reported improvement over SparseTSF is only 0.1% at H=96 (0.3615 vs. 0.3620), and all less than 3% difference, yet this result is used to support the claimed advantage over gradient-based methods.

5 Appendix Table 12 exists the resolution mismatched issues labeled ETTm1→ETTh1 transfers with the same numerical 24 to 24 periods as “matched,” despite the same resolution mismatch. The paper also states that a daily cycle in ETTm corresponds to 96 steps.

6 The evaluation does not sufficiently test the claimed robustness beyond the standard benchmark setting.

Questions for authors

Questions:

1 Can authors guide me through the logic in figure 1 and figure 2. Especially the figure 2 to understand what it actually means.

2 What specifically in Proposition 1 justifies the stronger zero-shot transfer claim?

3 What are the lookback lengths across methods in the main Table 1? Are they under fair comparisons? Can the author provide statistical significance when compared with these different methods with such close performances?

4 Can the authors compare against more recent frequency-/multi-scale methods such as WaveletMixer, WPMixer, and D-PAD?

5 What exactly does “matched” mean in Appendix Table 12, given the 24-step vs. 96-step daily-cycle difference between ETTh and ETTm?

6 Can the authors evaluate robustness under missing data, irregular sampling, weaker periodicity, and broader zero-shot benchmarks such as GIFT-Eval?

Aksu, Taha, et al. "Gift-eval: A benchmark for general time series forecasting model evaluation." arXiv preprint arXiv:2410.10393 (2024).

5 comments

We thank the reviewer for the constructive questions. We ran every requested experiment: matched-protocol comparisons with WPMixer and WaveletMixer, a 5-seed significance test, and a missing-data study. We answer each weakness together with its related question.

W1 + Q1. Figures 1 and 2

Figure 1. The series is folded into an S×pS\times p matrix, with cycles as rows and phases as columns. Each column is one phase sequence, which Theorem 1(1) makes cycle-stationary. A single W∈RS×MW\in\mathbb{R}^{S\times M} maps the last SS cycles of every column to its next MM cycles.

Figure 2. The solid path estimates WW in one closed-form solve on the training data. The dashed path is zero-shot: the source WW is applied to the folded target windows with no fitting. The two panels differ in exactly this respect: estimate WW, or reuse it. We added this walkthrough to the captions.

W2 + Q2. Theoretical contribution and Proposition 1

Theorem 1 provides what no prior period-sparse forecaster states: why one operator can be shared across phases and channels (Part 1), what it estimates, namely the Galerkin projection of the cycle Koopman operator (Part 2), and when it is exact (Part 4). Part 4 yields a transfer result: under shared cycle dynamics the source operator is optimal on the target (see our response to Reviewer uZZE, W1).

Proposition 1 isolates the single difference between CAPE and an SGD-trained forecaster on the same design: the spectral filter. Both filters are smooth, as you note. The gradient-flow filter depends on the stopping time and learning rate, which are fixed by source-specific optimization. The Tikhonov filter is explicit and computed in one solve. The direct test is SparseTSF, the same folded model class trained by SGD: CAPE achieves lower zero-shot MSE on all 8 ETT transfer pairs (table in our response to Reviewer uZZE, W1).

W3 + Q3. Look-back lengths, fairness, and significance

Table 18 compares all methods at L=96 in one pipeline. We added TimeMixer++. Traffic is compared under the matched L=720 protocol in the table below. CAPE has the best MSE on ETTh1 (0.438), ETTh2 (0.373) and ETTm1 (0.386). At L=96 the dictionary holds only S=L/p=4S=L/p=4 daily cycles on hourly data, and none on Weather, where one day spans 144 steps. CAPE leads wherever the window contains several cycles, which is its intended regime.

Significance and strength of the claim. CAPE is deterministic, so variance comes only from the baselines. SparseTSF's seed std is about 0.004 (1%). Under the matched L=720 protocol below, CAPE's MSE margins over the better of WPMixer and WaveletMixer are 2.4% to 9.3% on all seven datasets, well beyond that variance. We now state each claim with its protocol.

W3 + Q4. WPMixer, WaveletMixer, D-PAD

Same protocol as CAPE (L=720, same splits and pipeline, official code and hyperparameters). MSE / MAE averaged over H:

DatasetWPMixerWaveletMixerCAPE
ETTh10.429 / 0.4290.432 / 0.4400.399 / 0.414
ETTh20.361 / 0.3920.365 / 0.3980.332 / 0.380
ETTm10.382 / 0.3930.378 / 0.3920.355 / 0.376
ETTm20.276 / 0.3220.270 / 0.3270.249 / 0.312
Weather0.247 / 0.2730.246 / 0.2750.240 / 0.279
Electricity0.181 / 0.2710.174 / 0.2690.165 / 0.262
Traffic0.497 / 0.3030.451 / 0.3020.409 / 0.291

CAPE has the best MSE on all 7 datasets and the best MAE on 6 of 7, with a single closed-form solve. D-PAD has no public implementation that we could find, so it cannot be run under the matched protocol. We cite and discuss it in related work.

W4. Seeds and the comparison with SparseTSF (Table 14)

SparseTSF at the same period (p=24) over 5 seeds, against CAPE, which is deterministic:

HSparseTSF (mean ± std)CAPE
960.3640 ± 0.00450.3615
1920.4025 ± 0.00300.4002
3360.4341 ± 0.00400.4217
7200.4258 ± 0.00500.4169

CAPE is better at every horizon, and significantly so at H=336 and H=720 (one-sample t-test, p<0.05). At short horizons it matches SparseTSF within seed variance, with no gradient training. The advantage over gradient-based training is therefore equal or better accuracy from a single solve.

W5 + Q5. "Matched" in Table 12

"Matched" referred to sampling resolution, not to the fold. We relabel each pair by fold type (true period, divisor, misaligned). We also tested the remedy on the cross-resolution pairs: re-detecting the period on the target (p=96 on ETTm2) improves both pairs at both look-backs (ETTh1→ETTm2: 0.297 → 0.291 and 0.296 → 0.289. ETTh2→ETTm2: 0.275 → 0.271 and 0.273 → 0.269).

W6 + Q6. Robustness beyond the standard setting

Models are trained on clean data. At test time 5/10/20% of look-back values are hidden at random (3 masks per setting). Change in MSE relative to clean input:

DatasetModelCleanZero-fill 5%Zero-fill 10%Zero-fill 20%Interp. 20%
ETTh1CAPE0.399+0.7%+2.2%+7.9%+0.2%
ETTh1SparseTSF0.406+0.9%+2.6%+8.5%+0.4%
ETTh1DLinear0.437+1.8%+4.8%+12.7%+0.8%
ETTh2CAPE0.332+4.6%+13.1%+41.5%+0.1%
ETTh2SparseTSF0.361+5.0%+14.0%+43.2%+0.3%
ETTh2DLinear0.479+11.6%+25.8%+57.4%+0.9%

CAPE degrades least in every setting and remains the most accurate model. With linear interpolation before folding (Appendix F), 20% missing data changes its MSE by at most +0.2%. Weak periodicity is covered by Exchange-Rate (Table 6), where the detector selects p=1 and CAPE reduces to a ridge autoregressor. Robustness to the folding period is tested on four datasets (see our response to Reviewer uZZE, W3). A full GIFT-Eval evaluation is left for future work.

We hope these results address your concerns and would appreciate you reconsidering your score.

Dear Reviewer YQ4p, thank you again for your review. As the discussion period is ending, we wanted to check whether our response and the new results (WPMixer and WaveletMixer under a matched protocol, the 5-seed SparseTSF comparison, and the missing-data study) address your concerns. We are happy to clarify anything further.

Reviewer_uZZEOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CAPE, a closed-form forecasting method that combines period folding, a finite-dimensional delay-coordinate representation, and Tikhonov-regularized least-squares estimation. The authors provide theoretical results connecting period folding to phase-wise second-order stationarity and interpret the resulting estimator as a Galerkin projection of cycle-level dynamics. They further establish consistency and a generalization bound under specific assumptions, and evaluate CAPE on in-distribution, zero-shot transfer, and few-shot forecasting tasks.

Critical strengths

s1: CAPE replaces iterative gradient-based optimization with a closed-form solution, which is appealing for resource-constrained forecasting applications.

s2: Period folding provides an interpretable way to incorporate recurring temporal patterns, and the paper makes a useful attempt to connect this design choice to classical cyclostationary processes and Koopman operator theory.

s3: The paper considers several forecasting regimes, including in-distribution, zero-shot, and few-shot settings, and includes ablations examining period selection, regularization, and preprocessing.

Critical weaknesses

w1. The theoretical analysis establishes the consistency of CAPE's regularized projection, but does not sufficiently explain why it should generalize across domains. The exact Koopman closure result relies on restrictive assumptions, and the conditions under which source and target dynamics are compatible remain unclear.

w2. CAPE combines period detection, folding, smoothing, and regularized estimation, making it difficult to determine which components drive the improvements. The ablations suggest that smoothing is important and that gains narrow when periods are matched with SparseTSF.

w3. CAPE depends on identifying appropriate periodic structure, yet performance can degrade when the selected period is mismatched. The ETTh1-to-ETTm2 example illustrates the difficulty of transferring across different sampling frequencies. The paper should more systematically evaluate period-detection errors and robustness to changes in temporal resolution.

Questions for authors

See the Weaknesses.

3 comments

We thank the reviewer for the feedback. In response, we added (i) a transfer result with new zero-shot controls on all 8 ETT pairs, (ii) a step-by-step component analysis, and (iii) a period sweep on four datasets with a cross-resolution test.

W1. Why CAPE generalizes across domains, and when source and target are compatible

Theorem 1 yields an explicit compatibility condition. If source and target share the cycle dynamics E[Zs+1∣hs]=a⊤hs+b\mathbb{E}[Z_{s+1}\mid h_s]=a^\top h_s+b, then A=GW⋆A=GW^\star in both domains, so the source operator is optimal on the target even when Gsrc≠GtgtG_{\mathrm{src}}\neq G_{\mathrm{tgt}}. In general, the excess target risk is

tr[ΔW⊤(Gtgt+λI) ΔW],ΔW=(Gsrc+λI)−1(ΔA−ΔGWtgt).\mathrm{tr}\big[\Delta W^\top (G_{\mathrm{tgt}}+\lambda I)\,\Delta W\big],\qquad \Delta W=(G_{\mathrm{src}}+\lambda I)^{-1}(\Delta_A-\Delta_G W_{\mathrm{tgt}}).

Source and target are therefore compatible when their cycle moments (G,A)(G, A) are close. Exact closure is a sufficient condition, not a requirement: without it, CAPE transfers the Galerkin projection (Theorem 1, Part 4).

New controls confirm this. On all 8 ETT pairs, both period-folded methods (CAPE, SparseTSF) outperform both unfolded methods (ridge, DLinear), and CAPE is best on every pair:

Source → TargetCAPERidgeSparseTSFDLinear
ETTh1 → ETTh20.3490.3820.3530.481
ETTh1 → ETTm20.2950.3260.3010.406
ETTh2 → ETTh10.4160.5090.4250.465
ETTh2 → ETTm20.2720.3070.2850.441
ETTm1 → ETTh20.3540.3790.3630.380
ETTm1 → ETTm20.2560.2810.2640.279
ETTm2 → ETTh20.3540.3860.3680.378
ETTm2 → ETTm10.4040.4380.4130.420

(MSE averaged over H. SparseTSF and DLinear averaged over 5 seeds.)

W2. Which component drives the improvement, and the comparison with SparseTSF

We build CAPE step by step on ETTh1 (MSE averaged over H, zero-shot from ETTh2):

StepIn-distZero-shotEffect
Raw ridge, full window0.42410.5092baseline
+ period folding0.42390.4801576x fewer coefficients at H=96, zero-shot -5.7%
+ residual smoothing (CAPE)0.40070.4164in-dist -5.5%, zero-shot -13.3%
SparseTSF (same class, SGD, 5 seeds)0.40660.425CAPE is better with no training
  • Folding and smoothing act together. Smoothing without folding gives 0.4117 / 0.4424 at every kernel width, because the filter is invertible and an unfolded map absorbs it. Together they form the folded dictionary of Theorem 1.
  • Period detection selects the right fold. The detected period is the best or near-best fold in the sweep below.
  • The narrow gap to SparseTSF is the point. SparseTSF is the same folded model class trained by SGD. Matching it would already show that training can be removed. CAPE exceeds it at every horizon (significantly at H=336 and H=720 over 5 seeds) and on all 8 zero-shot pairs, while being deterministic.

W3. Period-detection errors and temporal resolution

Period errors. The sweep of Table 7 now covers four datasets (MSE):

DatasetHp=2p=4p=8p=12p=20p=22p=24p=26p=48
ETTh1960.3720.3760.3850.3780.4380.4920.3620.5180.374
ETTh17200.4230.4240.4280.4250.4640.4970.4170.5420.425
ETTh2960.2730.2740.2790.2770.2990.3110.2700.3180.282
ETTh27200.3790.3790.3820.3800.3940.4010.3770.4060.384
ETTm1960.3090.3100.3070.3060.3330.3470.3040.3520.301
ETTm17200.4160.4170.4150.4160.4310.4380.4140.4420.411
Electricity960.1400.1410.1450.1430.1740.2110.1380.2230.149
Electricity7200.2070.2080.2120.2100.2340.2650.2060.2780.216

On every dataset and horizon, the near-miss periods (20, 22, 26) are the three worst (+4% to +62%), while every divisor or multiple of the true period stays within 8% of the best. On ETTm1 the best fold is p=48, a divisor of the 96-step daily period. This is what Theorem 1(1) predicts.

Temporal resolution. We applied the remedy of Appendix D.3 to the ETTh→ETTm2 example: re-detect the period on the target and fold the 15-minute series at its daily period (p=96).

Source → TargetLNo remedyRemedy
ETTh1 → ETTm26720.2970.291
ETTh1 → ETTm27680.2960.289
ETTh2 → ETTm26720.2750.271
ETTh2 → ETTm27680.2730.269

The remedy improves every pair, confirming that the cross-resolution loss comes from fold misalignment. One unlabelled target window is enough.

We hope these results address your concerns and would appreciate you reconsidering your score.

Given the clarifications and the additional results, I’m raising my score.

Thank you for reading our response and for raising your score. If any concern remains, we would be glad to clarify it before the discussion closes.

Reviewer_ti7NOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CAPE, a gradient-free forecaster. A dominant period p is detected with an FFT/ACF score restricted to divisors of the look-back L. Each instance-mean-removed window has a centered moving average of width ≈ p added back to itself ("residual smoothing"), and is then folded so that each phase of each channel yields a row of S = L/p past values at lag multiples of p; the target is the M = H/p future values at the same phase. One S×M matrix, shared across phases and channels, is obtained by ridge regression in closed form. With p = 1 the method is ridge autoregression. The paper frames this as a Tikhonov-regularized EDMD / Galerkin projection of a "cycle Koopman operator", proves that folding a wide-sense cyclostationary process yields cycle-stationary phase sequences, gives a consistency statement, an O(N^{-1/2}) generalization bound under a dependence assumption, and the spectral-filter forms of gradient flow and ridge. Experiments report in-distribution results on seven benchmarks (L = 720; a matched L = 96 table in the appendix), zero-shot transfer within ETT and to Weather/Electricity by reusing the source matrix and period, a 10% few-shot result, Exchange-Rate, and ablations on period, ridge variants, channel independence and cost. Anonymous code and a NumPy implementation are provided.

Critical strengths
  1. A useful and reproducible empirical data point. A 120–900-coefficient ridge solve that fits in under a second and is competitive with heavily engineered deep models on the standard long-term benchmarks is worth knowing about, and the full implementation is printed in the paper. This is the paper's real contribution and the reason I consider it close to the threshold.

  2. Ablations that expose where the performance comes from. Table 13 (seasonal naive, full ridge, size-matched ridge, folded ridge without smoothing), Table 14 (SparseTSF at the same period), the period sweep, the matched-look-back table and the limitations section let a reader evaluate the method independently of the framing.

Critical weaknesses
  1. Stripped of the framing, CAPE is the closed-form solution of SparseTSF's linear model plus a heuristic input filter, and the paper's own ablations show that the theorised ingredient is not what produces the gain.

    • SparseTSF already regresses phase-aligned, period-subsampled histories onto period-subsampled futures with a matrix shared across phases, with the same 120 parameters. CAPE solves the same least-squares problem by the normal equations. Table 14 confirms this: at matched period CAPE differs from SparseTSF by 0.1% at H = 96. That least-squares linear forecasters admit closed-form solutions that match their SGD-trained versions is established (e.g. Toner & Darlow, ICML 2024, which is not cited). The Koopman/EDMD language adds no content: Theorem 1(2) is the ridge normal equation, (1) restates the definition of cyclostationarity, (3) is the ergodic theorem, and Proposition 1 is a textbook identity.
    • Table 13: folded ridge without smoothing scores 0.4239 in-distribution, the same as unfolded full ridge (0.4241). The whole improvement to 0.4007 comes from "residual smoothing". That step is justified as a low-pass filter that "attenuates high-frequency components", but the operation is X + MA_k(X), i.e. a filter 1 + H(f): it attenuates nothing and doubles the low-frequency gain. What it actually does is add a local (≈ one-cycle) mean to every phase-specific regressor, which lets cross-phase level information into a regression that folding had removed. That may be a good idea, but it is outside the theory and is the opposite of the stated explanation.
    • The candidate periods must divide L = 720, so the daily period of 15-minute data (96) can never be selected; ETTm1/ETTm2 are folded at p = 24 (six hours), which is not a period of the process. The cyclostationary theory therefore does not apply to two of the datasets where CAPE does best. The same paper attributes the ETTh1→ETTm2 transfer loss to exactly this mismatch, while ETTh1→Weather (10-minute data folded at 24 steps) is presented as a success of "shared operator structure".
    • Proposition 1 is said to "explain" better zero-shot transfer and Theorem 2 to explain data efficiency. Neither follows. The baselines are not gradient flow from zero on this design matrix; Theorem 2 assumes a summable-covariance condition over rows built from heavily overlapping windows (N_total = 588,000 counts each observation hundreds of times), obtains its rate from Markov's inequality with an unspecified constant, and the measured slope is −0.35, not −0.5.
  2. The empirical claims (20/28 wins, 7/8 zero-shot wins, best few-shot) rest on non-matched protocols and on tables that are internally inconsistent.

    • Table 1 runs CAPE at L = 720 against literature numbers at other look-backs. In the matched L = 96 table CAPE loses on ETTm2, Weather and Electricity; Traffic and TimeMixer++ are omitted there. Many "wins" in Table 19 are by 0.001–0.002 on values "truncated to three decimals", from a deterministic single run against copied numbers.
    • The same baseline has different numbers in different tables: TimeMixer++ on ETTh1 is 0.380/0.415/0.439/0.433 in Table 19 but 0.386/0.416/0.430/0.467 in Table 3; TimeMixer++ "supervised" on Weather is 0.239 MSE in Table 1 and 0.205 in Table 4; the Table 4 "supervised upper bound" for ETTh2 (0.367) is the zero-shot value from Table 2. CAPE's own ETTh1 and ETTm1 numbers differ between Table 15 and Table 19. Table 19 has entries that cannot be right (ETTm1: TimeMixer++ MSE 0.369 / MAE 0.313 and PatchTST MAE 0.314 at H = 96 but 0.366 at H = 720; Electricity: iTransformer MAE 0.155 at H = 192), which suggests shifted columns. Section 4.1 says Traffic "remains a limitation" while Table 1 shows CAPE with the best Traffic MSE.
    • Zero-shot baselines are copied from the TimeMixer++ and Time-LLM papers (different look-backs and pipelines). The controls that matter are missing: zero-shot SparseTSF, DLinear and plain ridge on all eight ETT pairs. The one such control reported (Table 13, ETTh2→ETTh1) shows the smoothing heuristic, not folding or the closed-form solve, accounts for most of the transfer gain (0.480 → 0.416). With instance normalization, transfer between ETT subsets is a weak test, and comparing MSE across different target datasets ("ETTh1→Weather outperforms ETTh1→ETTh2 by 20%") is not meaningful. The protocol also standardizes with target-training statistics, which should be stated as a caveat on "zero-shot".
    • The 13.6× speed-up compares 0.5 s with 7 s on ETTh1 and has no practical significance.
Questions for authors

Clarifications that could change my assessment:

  1. What exactly differs between CAPE and SparseTSF trained to convergence with the same smoothing pre-filter and weight decay? Please add SparseTSF + smoothing and full ridge + smoothing to Table 13, in-distribution and zero-shot.
  2. Please justify residual smoothing correctly: X + MA_k(X) is not a low-pass filter. What happens with MA only, with other kernel widths, and with the cycle mean appended as an extra regressor?
  3. On ETTm1/ETTm2 the selected p = 24 is not the daily period. Why does the method work there, and what do L = 672 or 768 with p = 96 give?
  4. Please reconcile the inconsistent numbers listed above (TimeMixer++ across Tables 1/3/4/19, CAPE across Tables 15/19, the anomalous cells in Table 19) and state the look-back and source of every baseline number.
  5. Please provide a complete matched-protocol comparison (same L for all methods, including Traffic and TimeMixer++), and zero-shot results for SparseTSF, DLinear and ridge on all transfer pairs.
  6. Is λ ever selected using the test split or the target dataset? What does "one-pass cross-validation" mean operationally?

Suggestion for future work (does not affect my rating): multiple-period dictionaries and online updates, as the authors already note.

5 comments

We thank the reviewer for the most detailed reading of the paper. We ran every requested control. In short: CAPE beats SparseTSF, DLinear and ridge zero-shot on all 8 ETT pairs, the smoothing controls isolate its role, and p=96 behaves as the theory predicts. We answer each weakness together with its related question.

W1(a) + Q1. CAPE vs. SparseTSF, Toner & Darlow, and the content of the theory

CAPE and SparseTSF share a model class. SparseTSF learns its aggregation kernel and its linear map by SGD. CAPE fixes the kernel to a one-cycle mean and solves the map in closed form. With the kernel fixed and weight decay λ′\lambda', the unique minimizer of the SparseTSF objective is CAPE's solution with λ=Nλ′\lambda=N\lambda', so "SparseTSF with this pre-filter and weight decay, trained to convergence" is CAPE, reached in one solve.

Against SparseTSF with its learned kernel over 5 seeds (p=24, ETTh1):

HSparseTSF (mean ± std)CAPE
960.3640 ± 0.00450.3615
1920.4025 ± 0.00300.4002
3360.4341 ± 0.00400.4217
7200.4258 ± 0.00500.4169

CAPE is better at every horizon, significantly at H=336 and H=720 (one-sample t-test, p<0.05), and better zero-shot on all 8 pairs (see W2(c) below). Full ridge with smoothing is reported in the next section.

We now cite Toner & Darlow (2024). Their result covers unconstrained linear maps. Ours covers the period-folded dictionary, its cyclostationary justification, and cross-dataset transfer.

On the theory: each part of Theorem 1 uses standard tools. Together they state why one operator can be shared across phases and channels (Part 1), what it estimates (Part 2), and when it is exact (Part 4, equivalent to a well-specified cycle-level AR(S)). Part 4 yields a transfer result: under shared cycle dynamics, the source operator is optimal on the target (see our response to Reviewer uZZE, W1). The predictions hold: near-miss folds are the three worst on all four datasets of the extended period sweep, and folded methods beat unfolded ones on all 8 transfer pairs.

W1(b) + Q2. Residual smoothing

Requested controls (ETTh1 in-distribution and ETTh2→ETTh1 zero-shot, MSE averaged over H):

MethodkIn-distZero-shot
CAPE250.40070.4164
Full ridge on X+MAk(X)X+\mathrm{MA}_k(X)13 / 25 / 490.41170.4424
Folded ridge on MAk(X)\mathrm{MA}_k(X) only130.53270.5484
250.64570.6674
490.40370.4394
Folded ridge + cycle means130.42470.4414
250.40770.4254
490.40670.4244
CAPE130.42170.4344
CAPE490.40500.4230

You are right that X+MAk(X)X+\mathrm{MA}_k(X) is not a low-pass filter. It is a low-shelf filter with gain 1+Dk(ω)∈[0.78, 2]1+D_k(\omega)\in[0.78,\,2], and we corrected Appendix C.1. The controls show its role:

  • It brings nothing on the unfolded model (identical for every k, since I+MkI+M_k is invertible and the map absorbs it).
  • MA alone is worse than CAPE at every kernel width.
  • Appending the cycle mean as an extra regressor is worse, even with twice the parameters.
  • The period-matched kernel (k=25) is best in both settings.

Smoothing is inside the theory. The filter is linear time-invariant, so it preserves wide-sense cyclostationarity and Theorem 1 applies unchanged. After folding, it adds a phase-aligned local aggregate to each row, which enlarges the observable dictionary.

W1(c) + Q3. Candidate periods, p=24 on ETTm, and ETTh1→Weather

CAPE at the daily period (MSE averaged over H):

DatasetLp=96p=24
ETTm16720.3510.358
ETTm17680.3490.354
ETTm26720.2740.252
ETTm27680.2710.248

On ETTm1 the daily fold improves on p=24, and at L=768 it beats our main result at L=720 (0.355). On ETTm2, p=24 stays better. This is the trade-off between phase alignment and dictionary size S=L/pS=L/p (7 to 8 cycles at p=96 vs. 28 to 32 at p=24), and it is why CAPE selects a forecast-effective period (Appendix G.1).

The theory covers this case. Since 24 divides 96, the 24-step fold keeps phase moments periodic across cycles, and the consistency result of Theorem 1 applies. The same holds for Weather (24 divides the 144-step day), so ETTh1→ETTm2 and ETTh1→Weather are both divisor folds and we now describe them consistently. Re-detecting the period on ETTm2 quantifies the remaining alignment cost at 1.5% to 2.4% (see our response to Reviewer uZZE, W3).

W1(d). Proposition 1 and Theorem 2

Proposition 1 states that CAPE and SGD-trained forecasters on the same design differ only in their spectral filter. SparseTSF is exactly that comparison (same folded design, SGD), and CAPE wins in-distribution at every horizon and zero-shot on all 8 pairs.

Theorem 2. Assumption (22) explicitly allows overlapping and correlated rows: overlap enters VV, which does not depend on NN. The constant is explicit in Appendix A.2: the bound is R∗4+4R∗2+1 V/(Nδ)\sqrt{R_*^4+4R_*^2+1}\,\sqrt{V/(N\delta)} with R∗=2(∥A∥F+1)/γR_*=2(\|A\|_F+1)/\gamma, and we now state it in the theorem. The measured gap uses a later test period, so it contains a temporal-shift term that does not shrink with NN, which accounts for a slope of -0.35.

W2(a) + Q5. Matched protocol and small margins

Table 18 is the matched comparison: every method at L=96 in one pipeline, now with TimeMixer++. CAPE has the best MSE on ETTh1, ETTh2 and ETTm1. At L=96 the dictionary holds only S=4S=4 daily cycles on hourly data and none on Weather (one day = 144 steps), which is where the other methods lead. For Traffic, the matched comparison is at L=720, where CAPE's dictionary holds 30 daily cycles: CAPE reaches 0.409 against 0.451 for WaveletMixer and 0.497 for WPMixer under the same pipeline (see our response to Reviewer YQ4p).

In Table 19, baselines are reproduced in our pipeline at their published look-backs, which the caption now states.

The claim does not rest on 0.001 to 0.002 differences. Under the matched L=720 protocol, CAPE improves on the better of WPMixer and WaveletMixer by 2.4% to 9.3% in MSE on all seven datasets, and on SparseTSF significantly at H=336 and H=720 over 5 seeds. In Table 19 we now mark differences within 0.002 as ties.

W2(c) + Q5. Zero-shot controls

All 8 ETT pairs (MSE averaged over H, SparseTSF and DLinear averaged over 5 seeds):

Source → TargetCAPERidgeSparseTSFDLinear
ETTh1 → ETTh20.3490.3820.3530.481
ETTh1 → ETTm20.2950.3260.3010.406
ETTh2 → ETTh10.4160.5090.4250.465
ETTh2 → ETTm20.2720.3070.2850.441
ETTm1 → ETTh20.3540.3790.3630.380
ETTm1 → ETTm20.2560.2810.2640.279
ETTm2 → ETTh20.3540.3860.3680.378
ETTm2 → ETTm10.4040.4380.4130.420
  • CAPE is best on all 8 pairs, against controls run in our own pipeline.
  • Instance normalization does not explain transfer. All four methods use it, yet both folded methods beat both unfolded methods on every pair. Against ridge (same closed form, same normalization, no folding) CAPE improves by 6.6% to 18.3%.
  • Smoothing and folding both contribute. On ETTh2→ETTh1, folding takes ridge from 0.509 to 0.480, and smoothing takes it to 0.416. Smoothing on the unfolded model gives no such gain (0.4424).
  • Cross-target comparison. We removed the sentence comparing ETTh1→Weather with ETTh1→ETTh2. Each transfer is now compared only against baselines on the same target.
  • Target statistics. Standardization with target training statistics (Appendix D.1) is now stated in the main text.

W2(b) + Q4. Table consistency

Thank you for catching these.

  • Table 15: the CAPE rows contained a typo. The correct values are those of Table 19.
  • Tables 3 and 4 vs. Table 19: Tables 3 and 4 used published baseline values, while Table 19 uses our reproduced runs. Each table now uses one source, stated in its caption with the look-back.
  • Table 4: the "TM++ sup." column mixed sources. The ETTh2 entry was the zero-shot value and is corrected to the supervised one (0.334 / 0.382, as in Table 1). The Weather entry now uses the same source as Table 1.
  • Table 19: the anomalous cells (ETTm1 for TimeMixer++ and PatchTST, Electricity MAE for iTransformer) were transcription errors and are corrected.
  • Section 4.1: the sentence on Traffic is corrected to match Table 1.

W2(d). Speed-up

Table 8 is kept as cost information: no GPU, no epochs, and under one second of CPU fitting. We no longer list 13.6x as a contribution.

Q6. Selection of λ

λ is never selected on the test split or on the target dataset. For each source trainig dataset, λ is chosen from {0.01, 0.1, 1, 5, 10, 100} using source trainig data only: the Gram matrices X⊤XX^\top X and X⊤YX^\top Y are accumulated in a single pass over the source training split, the six candidate operators are solved from these same matrices, and the candidate with the lowest MSE on the source validation split is kept. "One-pass" refers to this single accumulation. In zero-shot transfer, the selected λ is carried over with the operator (Appendix D.1). Selection is per dataset, which is why Exchange-Rate uses λ = 100. MSE is flat over λ ∈ [0.1, 10] (Table 10).

We hope these results address your concerns and would appreciate you reconsidering your score.

Dear Reviewer ti7N, thank you again for your detailed review. As the discussion period is ending, we wanted to check whether our response and the new results (zero-shot SparseTSF, DLinear and ridge on all 8 ETT pairs, the smoothing controls, the p=96 runs on ETTm, and the corrected tables) address your concerns. We are happy to clarify anything further.

Comments on the paper

1 comment

Dear Area Chair,

We thank you and the reviewers for the discussion. A brief summary for the AC discussion:

  • Every requested experiment was run. Zero-shot SparseTSF, DLinear and ridge on all 8 ETT pairs (CAPE best on all 8), the smoothing and kernel-width ablations, p=96 on ETTm, matched-protocol comparisons with TimeMixer++, WPMixer and WaveletMixer (CAPE best MSE on all 7 datasets), 5-seed significance tests, and a missing-data study.
  • Two reviewers raised their scores after these results: uZZE (1 → 2) and YQ4p (2 → 3).
  • Reviewer ti7N's two remaining points are addressed. Table 2 is now complete (CAPE outperforms TimeMixer++ and iTransformer on all 8 pairs), and the step-by-step ablation isolates the contribution of folding, smoothing and the closed-form solve.

All corrected tables and new results will be in the final version. Thank you for your time.

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.