acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Phase Alignment: Correcting Phase Drift in Multivariate Time Series Forecasting

Abstract

Phase drift is pervasive in real-world periodic time series, causing historical periodic structures to become misaligned with future observations and degrading multivariate forecasting performance. Existing methods for modeling periodic patterns often adopt fixed phase assignments within each cycle or establish correspondence using a shared set of phase indices. Such representations do not characterize frequency-specific phase variations or adapt periodic correspondence to their changes across observation windows, and may consequently preserve unreliable temporal information corrupted by phase drift and noise. To address these limitations, we characterize phase drift through the window-dependent variation of phase offsets across individual frequency components, providing a basis for adapting periodic representations to changing phase structures. Building on this formulation, we propose Adaptive Phase Alignment (APA), a lightweight and plug-and-play module for correcting phase drift. APA decomposes each variate into frequency components and applies a compact complex-valued affine mapping whose relative phase response depends on the current spectral state, allowing shared parameters to induce component- and window-dependent phase variations. It then converts these variations into representation-level phase alignment through temporal gating, which adaptively reweights periodic information across temporal positions. The resulting representation enhances reliable periodic cues while attenuating information corrupted by phase drift and noise. Extensive experiments on nine real-world benchmarks demonstrate state-of-the-art forecasting performance and consistent improvements across diverse forecasting backbones. Notably, APA achieves strong performance even with a shallow MLP backbone while introducing low computational overhead.

Primary area:
learning on time series and dynamical systems
Keywords:
multivariate time series forecasting, non-stationary time series, periodicity, phase drift, phase alignment

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_m5UPOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes APA, a lightweight module for addressing phase drift in multivariate time series forecasting. It applies a complex-valued affine transformation to Fourier coefficients to induce adaptive phase responses, followed by an inverse Fourier transform and temporal gating mechanism that reweights a learnable periodic template.

Critical strengths

1.This formulation is computationally economical and provides a clear mathematical basis for the proposed mechanism.

2.The experiment is thorough.

Critical weaknesses

1.The central claim concerns correcting phase drift, but Equation (5) characterizes the phase response induced by an affine transformation rather than explicitly estimating the discrepancy between historical and target phases.

2.Figures 3–5 illustrate improved representations and forecasting behavior, but they do not directly quantify how accurately the method recovers or corrects known phase offsets.

3.The gated periodic representation rather than the original historical observations raises a concrete concern about the model's ability to distinguish series with identical periodic structure but different absolute levels.

Questions for authors

See critical weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer m5UP

We thank the reviewer for the constructive review and for recognizing the economical formulation, its clear mathematical basis, and the thoroughness of our experiments. We address the three weaknesses below; all revisions are in blue.

W1. Eq. (5) describes a phase response rather than an explicit phase estimate

We agree, and this is by design: APA does not explicitly estimate the discrepancy between the historical and target phases (Sec. 3.3). Instead, two references are coupled so that their product compensates it. The template P(t)\mathbf{P}^{(t)}, read from the cyclic index s(t)s^{(t)} (Eq. 8), gives the nominal phase, i.e., where the window should lie on the cycle. The gate gives the observed phase: in the time domain, the multiplicative part of Eq. (4) equals Re(wn)xn−Im(wn)H(xn)\mathrm{Re}(w_n)\mathbf{x}_n-\mathrm{Im}(w_n)\mathcal{H}(\mathbf{x}_n), with H\mathcal{H} the Hilbert transform, so for a dominant component acos⁡(ωτ+φ(t))a\cos(\omega\tau+\varphi^{(t)}) it is approximately ∣wn∣acos⁡(ωτ+φ(t)+arg⁡wn)|w_n|a\cos(\omega\tau+\varphi^{(t)}+\arg w_n), and the gate peaks follow φ(t)\varphi^{(t)} (Sec. 3.3, A.14). Hence, H=G⊙P\mathbf{H}=\mathbf{G}\odot\mathbf{P} reads the template at positions selected by the observed phase, which corrects the offset between the nominal and the observed phase without estimating it. Eq. (5) explains why the response depends on the window, whereas this coupling explains how the drift is corrected. Avoiding an explicit estimate keeps the gate at four real parameters per variate and avoids committing to a single offset when several periodic components are present. W2 verifies that this implicit correction recovers known phase offsets, measured on the forecasts against the phase of the target.

W2. Quantifying the correction of known phase offsets

Beyond Figs. 3–5, the revision adds quantitative evaluations against known phases (new Sec. 4.5, A.8–A.11):

  • Residual offset relative to a known target phase (A.11). Under the smooth drift of Sec. 4.4, with maximum displacements of 1.2 to 6 h, the residual offset of the forecast relative to the known drifted phase of the target stays within 0.066–0.102 h for APA, versus 0.292–0.454 h for shared phase indices (PhaseFormer).
  • Known injected shifts (A.10). We resample Electricity and PEMS03 with piecewise-constant shifts of up to 3 h while keeping the cyclic index unshifted, so that the template and the data differ by the known phase 2πδ(t)/Pday2\pi\delta(t)/P_{\mathrm{day}}. The template alone cannot follow the shift: on Electricity, its MSE rises from 0.179 to 0.405 and its lag jitter (the spread of the forecast lag among windows with the same known shift) from 0.037 to 0.128 h, whereas the MSE of Full APA only rises from 0.133 to 0.165. Full APA attains lower MSE and MAE than all controlled variants at every strength and seed, including windows whose shift changes within the horizon, where the target phase departs from the historical one.
  • Phase error on real data (A.9). For window–variate pairs with a pronounced periodic component, the median phase error with respect to the ground truth drops from 1.29 h for the fixed template to 0.16 h for APA on Electricity, and from 0.41–0.67 h to 0.08–0.11 h on PEMS.
  • Timing vs. shape (A.8). Relative to Magnitude-Only, which differs from Full APA only in phase adaptation, Full APA reduces the temporal distortion index (TDI) by 18.8% on average over the nine datasets (32.5–44.0% on PEMS), more than twice its 7.8% reduction in the DTW shape error.

W3. Distinguishing series with the same periodic structure but different levels

The level is not discarded but handled outside the gate. Following common practice, APA uses reversible instance normalization (RevIN; Kim et al., 2022), as now stated in App. A.2: each input window is normalized by its own mean and standard deviation before APA, and the forecast is de-normalized with the same statistics. APA thus models the normalized periodic structure, while the level of each window is restored exactly at the output, so two series with identical periodic structure but different levels receive forecasts that differ exactly by their levels. This division is consistent with the design of the gate: a level only changes the DC term of the spectrum, which adds the same constant to every position of R(t)\mathbf{R}^{(t)}, and the softmax is invariant to such a constant. The prediction therefore combines three complementary sources: the level and scale of the window through the normalization statistics, its phase through G(t)\mathbf{G}^{(t)}, and the periodic shape learned over the dataset through the template.

We hope these clarifications address the reviewer's concerns, and we are happy to discuss any further questions.

Thank you for the detailed response and additional experiments. The RevIN clarification resolves my concern about absolute levels, and the quantitative phase evaluations substantially strengthen the empirical evidence.

My remaining reservation concerns the connection between the proposed mechanism and the central phase-correction interpretation. The derivation explains how the multiplicative spectral term responds to the observed phase, but the explanation of how the full affine transformation and subsequent template gating correct the phase mismatch remains largely heuristic. The improved forecast alignment supports the method’s effectiveness, but does not fully resolve this mechanistic concern.

I recognize the lightweight design, the extensive evaluation, and the substantial effort made in the rebuttal. However, because the phase-correction interpretation is central to the paper’s contribution, the remaining gap between the mechanism and its justification keeps my assessment slightly below the acceptance threshold. I therefore maintain my weak rejection recommendation

Reviewer_h2fbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Adaptive Phase Alignment (APA), a lightweight module for periodic multivariate time series forecasting. APA applies a complex affine transformation to the spectrum of each variable, converts the transformed representation into a temporal gate, and uses that gate to modulate a learned periodic template. The transformation can induce input-dependent relative phase changes despite sharing parameters across frequencies. The resulting representation is evaluated both with a shallow forecasting network and as a component of multiple existing backbones. Experiments cover nine datasets and multiple forecasting horizons, with additional ablations, synthetic drift experiments, efficiency measurements, and five-seed results in the appendix.

Critical strengths
  1. clear motivation. The shared complex affine map and periodic-template gate form an understandable, lightweight architecture.
  2. Broad empirical coverage and useful reported gains.
Critical weaknesses

My recommendation is primarily driven by the limited establishment of a distinct methodological contribution and the evidence for its claimed mechanism, detailed in points 1–3. Points 4–5 are supporting limitations.

  1. The methodological advance over existing spectral and periodic alignment methods is insufficiently established. The ingredients have close precedents: FITS learns amplitude and phase transformations with a complex-valued linear layer; CycleNet learns recurrent cycle templates and combines them with a simple forecasting backbone; and AOSNet already combines a learnable oscillatory prior with a state-conditioned gate for adaptive alignment. APA's specific contribution appears to be the combination of a frequency-shared complex affine map, temporal softmax gating, and a periodically indexed template. The paper acknowledges related methods, but does not sufficiently establish what new capability or insight this particular combination provides. This is a concern about incremental contribution, not a claim that the full architecture is identical to a prior method.

  2. A changing relative phase response is not sufficient evidence of phase-drift correction. In Eqs. (4)–(5), Z~k,n=wnZk,n+bn\widetilde Z_{k,n}=w_nZ_{k,n}+b_n induces Δϕk,n=arg⁡(wn+bn/Zk,n)\Delta\phi_{k,n}=\arg(w_n+b_n/Z_{k,n}). This describes the phase change produced by the affine operation; it does not establish that the change compensates for the underlying drift. Before gating, the operation is a fixed affine map in the time domain. The nonlinear gate can still be useful, and the residual-phase illustration in Figure 1 and representation plots in Figures 3–4 provide suggestive evidence. However, they do not adequately isolate phase correction from amplitude filtering and temporal reweighting across controlled cases. Since phase correction is the organizing contribution, this uncertainty limits the paper's scientific insight.

  3. The main comparisons do not adequately distinguish APA from methods already targeting adaptive periodic alignment. Tables 1–2 include useful comparisons with CycleNet, TQNet, and PhaseFormer. However, Section 2.2 also discusses Correctformer, which corrects attention-induced periodic drift, and AOSNet, which adaptively aligns oscillatory states; neither appears in the main benchmark. Their mechanisms and problem formulations differ, so their relevance should be examined explicitly rather than assumed identical. A focused comparison with a close applicable method, or a precise account of the capability not covered by it, would be more informative about APA's contribution than additional comparisons against general forecasting architectures. The existing generic-gate control is helpful but does not substitute for this distinction.

Questions for authors
  1. What capability or insight does the proposed combination add beyond existing complex spectral transformations, learned periodic templates, and adaptive state-conditioned gating? Please make the distinction from FITS, CycleNet, and especially AOSNet explicit at the level of operations and assumptions.

  2. Beyond the illustrative residual-phase and representation plots, can you quantify alignment against known component phases across controlled cases? How does the result distinguish phase correction from amplitude filtering or generic temporal reweighting?

  3. Why are the adaptive drift/alignment methods discussed in Section 2.2 absent from the main comparison? Is a matched comparison with one applicable close method feasible, using the same data and forecasting setup? If their settings are incompatible, please explain the relevant distinction.

  4. What advantage is expected when component phase changes cannot be represented by the shared displacement used in Section 4.4? A small controlled example, or a narrower statement of scope, would help assess the claimed departure from shared phase correspondence.

  5. Can you report paired-run differences for one near-tie and one substantial gain over a strong periodic backbone in Table 1? This would clarify the settings in which the additional mechanism provides a reproducible practical advantage.

5 comments
AuthorsOpenReview ↗

Response to Reviewer h2fb (1/2)

We thank the reviewer for the detailed review and for finding the motivation clear, the architecture lightweight, and the empirical coverage broad. W3/Q3, Q4, and Q5 are addressed in Part 2. All revisions are in blue.

W1 / Q1. Capability beyond FITS, CycleNet, and AOSNet

New App. A.14 (Table 14) compares the methods by operations and assumptions. The combination rests on a division of labor: where a window should lie on the cycle is a property of the dataset and can be stored in a cycle-indexed template, whereas where it actually lies is a property of the window and can be read from its spectrum. CycleNet uses only the former and AOSNet only the latter. APA couples the two: it assumes a fixed period LL but lets the phase drift across windows, and follows this drift without estimating the phase or modifying the input.

  • vs. FITS. FITS applies the same complex linear frequency response to every window and outputs the forecast. In APA, the affine term bn/Zk,n(t)b_n/Z_{k,n}^{(t)} makes the response depend on the window (Eq. 5), and the output serves only as gate logits. Setting bn=0b_n=0 raises the average MSE on Electricity, Weather, and PEMS03 from 0.164 to 0.173 (A.6).
  • vs. CycleNet. CycleNet locks the phase to the time index: it subtracts the cycle from the input and adds it back to the output, leaving phase deviations to the backbone. APA multiplies the template by a gate that follows the observed phase. Adding APA to CycleNet reduces its MSE/MAE by 12.4%/9.0% (Table 1). Under injected shifts of 0–3 h on Electricity (A.10), the template alone degrades from 0.179 to 0.405 MSE, whereas Full APA only goes from 0.133 to 0.165.
  • vs. AOSNet, the closest method, with three differences:
    1. Cycle-indexed reference. The prior of AOSNet has the window length and is applied at the same positions in every window, whereas APA reads the template from the cyclic start s(t)s^{(t)} (Eq. 8), which locates the window on the cycle even when L>TL>T.
    2. Observation-only, phase-locked gate. AOSNet computes a per-position sigmoid gate with a convolutional network from descriptors of both X\mathbf{X} and P\mathbf{P}; APA computes a softmax from X\mathbf{X} alone, with four real parameters per variate. Replacing only the gate computation with a parameter-matched time-domain module raises the average MSE from 0.197 to 0.242 (A.7).
    3. Multiplicative combination. AOSNet forms (1−G)⊙X+G⊙P(\mathbf{1}-\mathbf{G})\odot\mathbf{X}+\mathbf{G}\odot\mathbf{P} and replaces mismatched observations by the prior; APA forms G⊙P\mathbf{G}\odot\mathbf{P}, so the representation takes its shape from the template and its timing from the window (evidence in W2 below).

W2 / Q2. Phase correction vs. amplitude filtering and generic reweighting

We agree that, before gating, the operation is a fixed affine map in the time domain, and Sec. 3.3 now makes this explicit: its multiplicative part equals Re(wn)xn−Im(wn)H(xn)\mathrm{Re}(w_n)\mathbf{x}_n-\mathrm{Im}(w_n)\mathcal{H}(\mathbf{x}_n), with H\mathcal{H} the Hilbert transform, i.e., it rotates every component by the same angle arg⁡wn\arg w_n. Window dependence therefore does not require the map to change across windows: the map carries the window's own phase into the gate. For a dominant component acos⁡(ωτ+φ(t))a\cos(\omega\tau+\varphi^{(t)}), the gate peaks follow φ(t)\varphi^{(t)}, while P\mathbf{P} stays on the cycle grid, so G⊙P\mathbf{G}\odot\mathbf{P} reads the template at positions selected by the observed phase.

A.7 adds two controls on all nine datasets: Generic Temporal Gate (a parameter-matched time-domain gate, i.e., generic reweighting) and Magnitude-Only (magnitude adapted, phase preserved, i.e., amplitude filtering). Full APA is best on all nine, and alignment metrics locate the gain:

  • Timing vs. shape (A.8). Relative to Magnitude-Only, Full APA reduces TDI by 18.8% (32.5–44.0% on PEMS), more than twice its 7.8% DTW reduction.
  • Known injected shifts (A.10). Full APA attains lower MSE and MAE than both controls at every strength and seed. Its lag jitter is 25.9–33.8% lower than that of Generic Temporal Gate on Electricity; on PEMS03, it grows from 0.592 to 0.855 h for the generic gate but stays within 0.546–0.585 h for Full APA.
  • Amplitude vs. timing (A.10). On Electricity, Full APA and Magnitude-Only differ by only 2.7–3.7% in the phase-insensitive amplitude error, but by up to 25.7% in lag jitter, so the gain over amplitude filtering lies in timing.
  • Known component phase (A.9, A.11). On real data, the median phase error of pronounced periodic components drops from 1.29 h (template) to 0.16 h (APA) on Electricity and from 0.41–0.67 h to 0.08–0.11 h on PEMS. Under the known smooth drift, the residual offset of APA is 0.066–0.102 h, versus 0.292–0.454 h for shared phase indices.
AuthorsOpenReview ↗

Response to Reviewer h2fb (2/2)

W3 / Q3. Comparison with Correctformer, AOSNet, and other drift-oriented methods

New App. A.13 compares APA with the four methods discussed in Sec. 2.2: POEM, AOSNet, Correctformer, and AlignTime. Each reports only part of our nine datasets, so we use the three shared by all of them (ETTh1, ETTm1, Weather) under the same protocol (T=96T=96, S∈S\in {96, 192, 336, 720}), with results from the original papers (for Correctformer, its PatchTST-based variant). MSE/MAE averaged over horizons:

APAPOEMAOSNetCorrectformerAlignTime
ETTh10.422/0.4260.422/0.4310.427/0.4290.435/0.4300.423/0.427
ETTm10.367/0.3910.371/0.3930.377/0.3890.388/0.3980.396/0.399
Weather0.239/0.2670.240/0.2710.235/0.2630.258/0.2780.254/0.280

APA ranks first in three and second in three of the six cases, the best average rank. As the reviewer notes, the problem formulations differ: Correctformer corrects periodic drift induced by attention inside a Transformer backbone, whereas APA addresses phase drift in the data and is backbone-agnostic. AOSNet, the closest method, has no public implementation, so a matched rerun was not feasible; instead, A.14 gives a precise account of the capabilities it does not cover (W1 in Part 1).

Q4. Component phase changes not representable by a shared displacement

We first restate the claim more precisely. APA departs from shared phase correspondence in where the correspondence comes from: shared phase indices assign the same correspondence to every window, whereas APA reads it from the spectrum of each window and each variate. Frequency specificity refers to the first stage, drift modeling: every component keeps its observed phase in R(t)\mathbf{R}^{(t)} up to a common rotation, and its relative response depends on its own coefficient Zk,n(t)Z_{k,n}^{(t)} (Eq. 5). The second stage, alignment, converts the resulting phase structure into one temporal weighting per variate and window. APA therefore does not estimate or re-align the phase of each component separately.

Under this formulation, phase changes need not follow a shared displacement. (i) Across variates, the gate is computed per variate, so variates may drift by different amounts, whereas the controlled drift of Sec. 4.4 applies one displacement to all variates. (ii) Across components within a variate, Eq. (4) preserves their observed phases, so component-specific changes are retained in R(t)\mathbf{R}^{(t)} and enter H(t)\mathbf{H}^{(t)} through G(t)\mathbf{G}^{(t)}, instead of being overridden by a fixed correspondence. The expected advantage is that the representation follows the actual phase structure of the current window, weighted toward its dominant components.

Real-data evidence is given in A.9, where no drift is injected, so the phase changes are not constrained to a shared displacement. Measured per window and variate against the ground truth, the template lags in some windows and leads in others by one to several hours, whereas APA stays within one hour in 35 of the 36 shown windows; over pairs with a pronounced periodic component, its median phase error is 0.16 h vs. 1.29 h for the template on Electricity, and 0.08–0.11 h vs. 0.41–0.67 h on PEMS.

As a narrower statement of scope, APA targets drift of the dominant periodic structure across windows and variates. When components within a variate drift inconsistently, it follows their combined phase pattern rather than aligning each component separately; a controlled example isolating this case is left to future work.

Q5. Paired runs on a near-tie and a substantial gain

New App. A.12 trains TQNet and TQNet+APA, which differ only in the APA module, with five seeds on the two datasets with the smallest gains in Table 1 (ETTm1, Weather) and the two with the largest (PEMS03, PEMS04). Horizon-averaged results, with the number of seeds on which TQNet+APA wins:

DatasetMSE: TQNet → +APA (wins)MAE: TQNet → +APA (wins)
ETTm10.377 → 0.370 (5/5)0.393 → 0.392 (5/5)
Weather0.242 → 0.240 (5/5)0.269 → 0.267 (2/5)
PEMS030.097 → 0.086 (4/5)0.203 → 0.188 (5/5)
PEMS040.091 → 0.078 (5/5)0.197 → 0.179 (5/5)

On PEMS, the gain is reproducible at nearly every horizon, with at least four wins in 15 of the 16 horizon-specific cases. On the near-tie datasets, the horizon-averaged MSE gain is reproduced on all five seeds, whereas small MAE gains vary across horizons. This pattern is consistent with the mechanism: the PEMS datasets are also where Full APA reduces TDI the most (A.8). APA thus provides the clearest practical advantage when the periodic structure is strong and its phase drifts across windows.

We hope these clarifications address the reviewer's concerns about the contribution and mechanism of APA, and we would be glad to discuss any remaining questions.

Reviewer_h2fbOpenReview ↗

Thank you for your response.

Thank you for the substantial additional analysis. Part of my questions are addresed. However, my remaining reservation is still mainly about the methodological contribution. The response makes the differences from AOSNet more concrete, particularly the cycle-indexed template and the observation-only gate. However, I am still not fully convinced that the evidence isolates the benefit of this particular combination over the closely related alternatives. The added comparison with published results provides useful context, but it is not a matched comparison, and the results are competitive rather than uniformly better. I understand the limitation imposed by the unavailable AOSNet implementation and do not regard that alone as a reason for rejection.

At this stage I lean toward retaining my weak-rejection recommendation because the distinct contribution remains borderline for me.

Reviewer_pa31OpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies phase drift in periodic multivariate time-series forecasting, where temporal misalignment can make historical periodic patterns unreliable for future prediction. It proposes Adaptive Phase Alignment (APA), which applies a frequency-shared complex affine transformation to each variate's spectrum and converts the result into a temporal gate that reweights a cycle-indexed learnable periodic template. The authors evaluate APA on nine benchmarks, compare it with representative forecasting methods and backbone integrations, and provide ablations, controlled temporal-warp experiments, robustness results, and efficiency analyses.

Critical strengths
  1. The frequency-domain motivation is sound: a time-varying temporal displacement naturally induces frequency-dependent phase changes, as expressed by Δϕk(t)=ωkδ(t)\Delta\phi_k(t)=\omega_k\delta(t). The lightweight shared-complex parameterization and cycle-indexed template provide a practical mechanism and a sensible seasonal inductive bias.

  2. The empirical evaluation is broad, covering nine standard benchmarks, multiple forecasting horizons, diverse linear, Transformer, periodic, and phase-aware baselines, and both MSE and MAE. The paper also includes component ablations, period-length sensitivity, seed robustness for APA, and efficiency comparisons.

Critical weaknesses

The paper does not directly verify that APA recovers or corrects frequency-level phase drift, rather than simply improving forecasting through an input-dependent temporal gate. The current evidence demonstrates better forecasting performance and provides qualitative representation analyses, but it does not report a quantitative phase-alignment metric against known ground truth. Consequently, the physical interpretation of APA as phase-drift correction is not yet fully distinguished from the more general fact that its transformation mathematically changes phase. The most relevant controlled variants in Appendix A.7 are evaluated on only three of the nine datasets. These experiments are useful, but they do not yet establish across the full benchmark suite that the gains arise specifically from phase-aware adaptation rather than generic temporal gating, magnitude adaptation, or the periodic template.

Questions for authors

See Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer pa31

We thank the reviewer for the careful assessment and for recognizing the frequency-domain motivation, the practical shared-complex parameterization with a cycle-indexed template, and the breadth of our evaluation. We agree that higher accuracy, together with the fact that Eq. (4) mathematically changes phase, does not by itself establish phase-drift correction. The revision addresses this directly; changes are in blue, and the new evidence is summarized in the new Sec. 4.5.

W1. Quantitative alignment against a known ground-truth phase

APA does not output an explicit phase estimate; it aligns the periodic representation (Sec. 3.3). We therefore measure the alignment of its forecasts against a known phase, both on real data and under drift of known magnitude. Since a time shift δ\delta changes the phase of every component kk by ωkδ\omega_k\delta, the lag measured below reflects the alignment of all periodic components jointly.

  • Injected shifts of known magnitude (A.10). We resample Electricity and PEMS03 with piecewise-constant shifts δ(t)\delta(t) of up to 3 h while keeping the cyclic index of the template unshifted, so that the template and the data differ by the known phase 2πδ(t)/Pday2\pi\delta(t)/P_{\mathrm{day}}. For each window, the forecast lag τ^\hat{\tau} is estimated by cross-correlation, and the lag jitter is the spread of τ^\hat{\tau} among windows with the same known shift (below 0.001 h for the ground truth). The template alone cannot follow the shift: on Electricity, its MSE rises from 0.179 to 0.405 and its jitter from 0.037 to 0.128 h as the shift grows to 3 h, whereas the MSE of Full APA only rises from 0.133 to 0.165. Full APA attains lower MSE and MAE than every variant at every strength and seed, and its jitter is 25.9–33.8% lower than that of Generic Temporal Gate on Electricity. On PEMS03, the jitter of the generic gate grows from 0.592 to 0.855 h, whereas that of Full APA stays within 0.546–0.585 h.
  • Known smoothly drifting phase (A.11). Under the drift of Sec. 4.4, the residual offset of APA relative to the known drifted phase stays within 0.066–0.102 h for all α>0\alpha>0, versus 0.292–0.454 h for shared phase indices (PhaseFormer), with a standard deviation of 0.083–0.126 h versus 0.388–0.496 h.
  • Real data (A.9). For window–variate pairs with a pronounced periodic component, the median phase error with respect to the ground truth drops from 1.29 h for the fixed template to 0.16 h for APA on Electricity, from 0.41–0.67 h to 0.08–0.11 h on PEMS, and by about half on ETTh1/ETTm1.

W2. Phase correction vs. a transformation that changes phase

Eq. (5) only describes the phase response, so we now state what this response does (Sec. 3.3, A.14). In the time domain, the multiplicative part of Eq. (4) equals Re(wn)xn−Im(wn)H(xn)\mathrm{Re}(w_n)\mathbf{x}_n-\mathrm{Im}(w_n)\mathcal{H}(\mathbf{x}_n), with H\mathcal{H} the Hilbert transform. For a dominant component acos⁡(ωτ+φ(t))a\cos(\omega\tau+\varphi^{(t)}), this is approximately ∣wn∣acos⁡(ωτ+φ(t)+arg⁡wn)|w_n|a\cos(\omega\tau+\varphi^{(t)}+\arg w_n), so the gate peaks follow the observed phase φ(t)\varphi^{(t)} of the current window, while P\mathbf{P} stays on the cycle grid. H=G⊙P\mathbf{H}=\mathbf{G}\odot\mathbf{P} thus reads the template at phase-selected positions, i.e., the phase change is tied to the phase of the window. Empirically, the DTW/TDI decomposition (A.8) separates shape error from timing error: relative to Magnitude-Only, which differs from Full APA only in phase adaptation, Full APA reduces TDI by 18.8% (32.5–44.0% on PEMS), more than twice its 7.8% DTW reduction, and it attains the lowest TDI and Shift on seven of the nine datasets. Phase adaptation therefore mainly reduces timing error, as phase correction should.

W3. Controlled variants on all nine datasets (A.7)

We extended the controlled analysis from three to all nine datasets (Table 8) and evaluated the same variants with the alignment metrics (Table 9). Averages over the nine datasets:

VariantMSEMAEDTWTDIShift
Periodic Template Only0.2940.3420.3443.0559.34
Generic Temporal Gate0.2420.2990.3002.6078.79
Magnitude-Only0.2250.2850.2832.6098.81
Full APA0.1970.2610.2612.1188.11

Full APA is best on every dataset in both MSE and MAE. Relative to Magnitude-Only, the strongest variant on average, its MSE reduction ranges from 3.2% (ETTm1) to 37.3% (PEMS08). Hence, neither the template, generic temporal gating, nor magnitude adaptation alone accounts for the gains.

We hope the new quantitative evidence addresses the reviewer's concerns, and we are happy to provide further details.

I thank the authors for the revision. Extending the controlled variants to all nine datasets addresses my request on coverage. My central concern—phase-drift correction versus generic input-dependent gating—remains open.


1. Theory: APA is an input-dependent gate, and a positive gate cannot realign the template

  • The gate is a fixed function of the input.
    Since (wn,bn)(w_n, b_n) is shared across bins, Eq. (7) decomposes exactly as: R:,n(t)=Re(wn) xn(t)−Im(wn) H(xn(t))+βnR^{(t)}_{:,n} = \mathrm{Re}(w_n)\,x^{(t)}_n - \mathrm{Im}(w_n)\,\mathcal{H}\big(x^{(t)}_n\big) + \beta_n where βn=F−1(bn1)\beta_n = \mathcal{F}^{-1}(b_n\mathbf{1}) does not depend on the window. The coefficients on the input are identical for every window, and bnb_n is a fixed positional logit bias. The window-dependence of Eq. (5) comes entirely from this fixed bias. APA is thus an input-dependent temporal gate with four real parameters per variate. That gate peaks follow the input (Sec. 3.3) is a statement about GG, not about the template.

  • The product cannot follow arbitrary phase shifts.
    Since G>0G > 0, for any template, input, bnb_n, and TT: ⟨H,P⟩=∑τGτPτ2>0\langle H, P \rangle = \sum_\tau G_\tau P_\tau^2 > 0 The gated template therefore always correlates positively with the unshifted template, so a positive gate cannot produce a shift such as a half-cycle displacement. For a sinusoidal template, the phase of HH at the template frequency stays strictly within ±90∘\pm 90^\circ of the template phase, regardless of the drift. Appendix A.14 states that the product "follows the phase of each window"; that claim is not supported at the operator level and must be shown on H(t)H^{(t)} directly.


2. Experiments: The new evidence does not isolate phase correction

  • Lag jitter is not the metric my review asked for.
    It measures the spread of τ^\hat{\tau} within a shift level, not the offset against the known shift. A forecast that is consistently one hour off has zero jitter.

  • The key comparison is omitted, and its trend runs the wrong way.
    Magnitude-Only differs from Full APA only in phase adaptation. Yet in Table 10, the advantage of Full APA shrinks as the shift grows:

    • PEMS03: The MSE advantage falls from 24.4%24.4\% to 0.5%0.5\%.
    • Electricity: The lag jitter at 3 h3\text{ h} is 0.045 h0.045\text{ h} for Full APA vs. 0.044 h0.044\text{ h} for Magnitude-Only.
  • Appendix A.10 mostly tests static offsets already seen in training.
    81.0%81.0\% of Electricity windows (64.6%64.6\% on PEMS03) have a constant shift. The five shift levels appear in both training and test. All shifts (≤45∘\le 45^\circ on the daily cycle) fall within the range a positive gate can adjust.

  • Appendices A.9 and A.11 lack the controlled variants.
    A.9 compares only against the fixed template, and A.11 compares only against PhaseFormer. Neither attributes the alignment to the phase branch.


What Would Address the Concern

  1. ∣τ^ˉ−δ∣|\bar{\hat{\tau}} - \delta| per shift level for all four variants in A.10, including shifts beyond a quarter cycle of the relevant component.
  2. The A.9/A.11 phase-error metrics for Magnitude-Only and Generic Temporal Gate.
  3. The phase of H(t)H^{(t)} itself against the known shift, since alignment is claimed at the representation level.
  4. An explanation of why the advantage over Magnitude-Only shrinks as the injected shift grows.

These requests do not add new requirements. They follow directly from the paper's own claim: if APA performs representation-level phase-drift correction, then the representation—measured against a known shift and compared with the variant lacking phase adaptation—is where that claim must be verified. Until then, the evidence shows that APA is an effective input-dependent gate on a cycle-indexed template, not that it corrects phase drift. I therefore maintain my assessment.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response: Summary of Revisions and New Evidence on Phase Alignment

We thank all reviewers for their constructive comments, and are glad that they found the frequency-domain motivation sound and the design lightweight and practical (pa31, h2fb), the formulation economical with a clear mathematical basis (m5UP), and the evaluation broad and thorough (all).

The main shared concern is that the submission did not verify against a known reference that APA corrects phase drift, rather than gaining from generic temporal gating, magnitude adaptation, or the periodic template (pa31; h2fb W2/Q2; m5UP W1/W2). Reviewer h2fb also asked how APA differs from FITS, CycleNet, and AOSNet, and why the drift-oriented methods of Sec. 2.2 were not compared. We have revised the paper accordingly. New content is in blue, and blue appendix titles mark new or revised sections.

1. Quantitative evidence of phase alignment (all reviewers; new Sec. 4.5, App. A.7–A.11)

  • Controlled variants on all nine datasets (A.7, previously three). Periodic Template Only removes the gate, Generic Temporal Gate uses a parameter-matched time-domain gate, and Magnitude-Only preserves the spectral phase, differing from Full APA only in phase adaptation. Full APA is best on all nine datasets in MSE and MAE, reducing the average MSE/MAE by 12.4%/8.4% over Magnitude-Only.
  • Shape vs. timing errors (A.8). Following Le Guen & Thome (2019), we measure shape error with DTW and temporal distortion with TDI and Shift. Relative to Magnitude-Only, Full APA reduces TDI by 18.8% (32.5–44.0% on PEMS), more than twice its 7.8% DTW reduction, so phase adaptation mainly reduces temporal distortion.
  • Phase error against the ground truth (A.9). For window–variate pairs with a pronounced periodic component, the median phase error drops from 1.29 h for the fixed template to 0.16 h for APA on Electricity, from 0.41–0.67 h to 0.08–0.11 h on PEMS, and by about half on ETTh1/ETTm1.
  • Injected shifts of known magnitude (A.10). With piecewise-constant shifts of up to 3 h on Electricity and PEMS03, Full APA attains lower MSE and MAE than every variant at every strength and seed, including windows whose shift changes within the horizon. On Electricity, its MSE reduction over Periodic Template Only grows from 25.7% to 59.3% with the shift, and its lag jitter is 25.9–33.8% lower than that of Generic Temporal Gate.
  • Shared phase indices vs. a known drifted phase (A.11). Under the smooth drift of Sec. 4.4, the MSE of APA rises by 18.0% from α=0\alpha=0 to α=1\alpha=1, versus 67.4% for PhaseFormer, and its residual offset relative to the known drifted phase of the target stays at 0.066–0.102 h, versus 0.292–0.454 h. Without drift, PhaseFormer has a slightly smaller offset (0.106 vs. 0.141 h), but the advantage of APA grows with the drift strength: its MSE reduction rises from 22.7% at α=0\alpha=0 to 45.5% at α=1\alpha=1.

2. Distinction from FITS, CycleNet, and AOSNet (h2fb W1/Q1; new App. A.14, Secs. 2.2, 3.3)

We add an operation- and assumption-level comparison. APA couples an absolute, cycle-indexed reference (where the window should lie on the cycle, a property of the dataset) with a window-level spectral phase correction (where the oscillation actually lies, a property of the window). In the time domain, the multiplicative part of Eq. (4) equals Re(wn)xn−Im(wn)H(xn)\mathrm{Re}(w_n)\mathbf{x}_n-\mathrm{Im}(w_n)\mathcal{H}(\mathbf{x}_n), with H\mathcal{H} the Hilbert transform, so the largest gate values follow the observed phase while P\mathbf{P} stays on the cycle grid, and H=G⊙P\mathbf{H}=\mathbf{G}\odot\mathbf{P} reads the template at phase-selected positions. In contrast, FITS applies one frequency response to all windows and outputs the forecast, CycleNet locks the phase to the time index, and the prior of AOSNet carries no cycle position and is mixed convexly with the input.

3. Comparison with drift-oriented methods (h2fb W3/Q3; new App. A.13)

On ETTh1, ETTm1, and Weather, the datasets shared by POEM, AOSNet, Correctformer, and AlignTime (results from the original papers under the same protocol), APA ranks first in three and second in three of the six dataset–metric cases, the best average rank. AOSNet is best on Weather and in ETTm1 MAE; as no public code is available, we could not compare it under identical settings.

4. Reproducibility over a strong periodic backbone (h2fb Q5; new App. A.12)

In five-seed paired runs of TQNet and TQNet+APA on the datasets with the two smallest (ETTm1, Weather) and two largest (PEMS03, PEMS04) gains, TQNet+APA wins on at least four seeds in seven of the eight horizon-averaged cases.

5. Level handling (m5UP W3; App. A.2)

App. A.2 now states that APA uses RevIN (Kim et al., 2022), so the level of each window is restored exactly at the output.

Other reviewer-specific questions are answered in the individual responses. We hope the revision addresses the reviewers' concerns and are happy to clarify further.

Dear reviewers,

The authors have now submitted their responses. Please review them and share your thoughts on whether your concerns have been addressed. I encourage you to engage with one another’s assessments, discuss any remaining disagreements, and update your reviews and scores as appropriate.

Thank you.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.