acceptodds
Under review as a conference paper at ICLR 2027

PhaseAlign: Stage-Dependent Phase Modeling For Long-Term Time Series Forecasting

Abstract

Long-term time series forecasting relies on recurrent structure, but identifying periodic patterns alone does not indicate where they will reappear in the future. Phase provides this temporal reference, motivating different explicit uses for historical modeling and future generation. Historical modeling relies on relative phase relations among positions in the observed window, whereas future generation preserves the phase reference of the observed window to determine where recurrent structure will appear next. Motivated by this distinction, we propose PhaseAlign, which extracts a shared complex-spectrum representation from the observed window and uses its phase information differently for historical modeling and future generation. For historical modeling, dominant frequencies are used to construct relative harmonic relations that guide attention over the observed sequence. For future generation, the corresponding window-referenced phase states are propagated to future positions along the forecast horizon, providing phase-aware conditioning for future queries. Experiments on seven standard long-term forecasting benchmarks demonstrate strong forecasting performance, while ablation results show consistent advantages of stage-dependent over unified phase modeling. Code is available at this repository: https://github.com/Time358/PhaseAlign-main

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, phase modeling, spectral modeling, periodicity, temporal alignment

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_qmHZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PhaseAlign proposes stage-dependent phase modeling for long-term time-series forecasting. A channel-independent model extracts a shared complex-spectrum representation from each observed window, uses relative harmonic relations to guide historical attention, and propagates window-referenced phase to condition future queries. Time–frequency interaction and local temporal modeling complement these operators. The paper reports competitive results on seven benchmarks, including Electricity, and evaluates its hypothesis through unified-phase controls, component ablations, repeated runs, and phase-shift, frequency-drift, and off-grid experiments.

Critical strengths
  1. A clear and testable architectural hypothesis. Distinguishing relative harmonic compatibility in historical attention from phase-referenced future conditioning provides an interpretable design rationale. The paper carefully limits its invariance argument to the dedicated operators rather than asserting exact invariance or equivariance of the entire network. This makes the proposed mechanism concrete enough to analyze and test.
  2. Evaluation extends beyond benchmark accuracy. The reported experiments cover seven datasets and multiple horizons, with repeated runs and targeted controls. The unseen relative-phase experiment and distortion-matched phase/magnitude interventions examine temporal placement more directly than aggregate forecasting metrics alone. These experiments could provide meaningful support for the hypothesis if their implementation and provenance are reconciled.
Critical weaknesses
  1. The central stage-dependent claim is only partially justified. Equation (1) explains phase cancellation and propagation for fixed-frequency harmonics, but does not establish why historical modeling should favor relative phase while future generation should favor absolute phase. The mathematical argument motivates a plausible design; stronger controlled evidence is needed to support its claimed advantage over unified alternatives.

  2. Method–implementation inconsistencies weaken the attribution of the reported gains. The phase-form controls are described as preserving a common positional backbone, yet the supplied implementation also changes generic positional encoding. Likewise, the alignment objective is described as regularizing the forecasting encoder, but instead operates on separate auxiliary representations. Differences in the spectral loss and the use of seasonal correction further complicate interpretation. These discrepancies need to be reconciled before the ablations and auxiliary-objective gains can substantiate the paper's explanation.

  3. The drift results do not establish retained predictive utility under severe nonstationarity. In Table 14, PhaseAlign reaches MAE=0.7302 under the strongest drift. Under the stated clean, unit-amplitude generator with uniform random phase, a zero forecast has theoretical population MAE 2/π≈0.6366. Although the exact test-set comparison remains necessary, this raises the possibility that the reported advantage reflects smaller degradation among failing models rather than effective phase tracking.

  4. Periods longer than the observation window remain an important untested boundary. The method derives its phase trajectories from the window's discrete Fourier grid, making the adaptive phase features periodic with the input-window length. Its ability to support forecasting when the true period greatly exceeds that window is therefore unclear. The synthetic studies mainly use roughly 24-step oscillations with a 96-step history and do not examine this regime.

  5. Validation of the mechanism is narrower than validation of overall accuracy. Electricity provides high-dimensional benchmark evidence, but the real-data stage-specific ablations are restricted to ETTh1, ETTh2, and Exchange. It remains unclear whether the same mechanism explains the gains on higher-dimensional datasets. The robustness probes also focus on simple harmonic processes, leaving stronger, more heterogeneous nonstationarity insufficiently characterized. It is suggested that the authors include performance results on high-dimensional datasets, such as Traffic and PEMS, to further demonstrate the method's generalization capability.

Questions for authors
  1. Which exact code revision and commands generated the main tables? Please reconcile the output shape and label broadcasting, the phase-loss definition, and the mixed-loss/seasonal-blending recipes with the manuscript.

  2. Did the reported phase-form controls retain RoPE, and did the alignment loss backpropagate into the main encoder? Please provide the corresponding implementation and matched results, particularly for Appendix A.2 and Table 23.

  3. On the exact Table 14 test sets and evaluation scale, how do zero/mean prediction and a phase-disabled control perform? Does PhaseAlign retain an advantage when judged against these baselines, rather than only against the other trained models?

  4. How does the stage-dependent benefit change as the dominant period exceeds the input window, and does it persist in matched controls on Electricity? A focused period/window-ratio sweep and a high-dimensional control experiment would materially affect my assessment.

3 comments
AuthorsOpenReview ↗

Response to Reviewer qmHZ (Part 1/2)

We thank the reviewer for the detailed and constructive feedback. We clarify the theoretical motivation, reconcile the implementation, and provide additional robustness analyses.

W1 — Stage-dependent phase modeling. Historical modeling compares observed positions, where the common phase origin cancels: Δθk,ij=2πfk(τi−τj)\Delta\theta_{k,ij}=2\pi f_k(\tau_i-\tau_j). Future generation instead requires the observed phase reference: θk,h=ϕk+2πfk(L+h)\theta_{k,h}=\phi_k+2\pi f_k(L+h). Our operator-level test confirms this asymmetry: historical relative-phase sensitivity is 0.0, versus 1.998 for absolute circular coordinates at δ=π\delta=\pi; future-placement error is 0.076 vs. 1.245, respectively. The original Table 2 further shows consistent advantages over Unified-Rel/Unified-Abs on ETTh1 (0.421/0.431/0.434), ETTh2 (0.365/0.379/0.381), and Exchange (0.343/0.360/0.365). These results support the proposed stage-dependent design without assuming network-wide equivariance.

W2 / Q1–Q2 — Implementation reconciliation. We apologize for linking an outdated development version. The reported experiments used the corrected implementation, which produces [B,H,C][B,H,C] predictions, preserves RoPE across phase controls, and propagates alignment gradients into the main encoder. All controls share spectral context and registered parameter capacity. The reported protocol uses MSE with λp=λa=0.1\lambda_p=\lambda_a=0.1, without seasonal post-processing. We will provide the experiment-generating revision, configurations, and exact reproduction commands.

W3 / Q3 — Severe frequency drift. We confirm R4's zero-baseline observation: empirical MAE is 0.6368, consistent with 2/π≈0.63662/\pi\approx0.6366, whereas PhaseAlign reaches 0.7302 at 4π4\pi (H=96H=96). Additional references on the same test distribution are summarized below.

MAE ↓, H=96

Method00.5ππ2π4π
Zero0.6370.6370.6370.6370.637
Mean0.6380.6380.6380.6380.639
Seasonal-naive0.4470.5140.6230.8400.882
PhaseAlign0.0410.2190.4150.7190.730
True-phase stationary extrapolation0.0000.5510.9470.9720.830
Stationarity-constrained reference0.0650.2590.5060.8530.863
PhaseFormer0.0370.2150.4260.8170.993

Mean closely matches Zero, while seasonal-naive deteriorates under stronger drift. At 4π4\pi, PhaseAlign outperforms the stationary-extrapolation references and PhaseFormer, although its zero-reference skill is negative (−0.147). The interpolated zero-skill crossover lies around 1.55π1.55\pi–1.73π1.73\pi across horizons. We therefore interpret the severe-drift results as relative robustness rather than retained absolute predictive utility.

AuthorsOpenReview ↗

Response to Reviewer qmHZ (Part 2/2)

W4 / Q4(a) — Period-to-window ratio and complementary future-query modeling.

We thank the reviewer for this insightful question. Importantly, PhaseAlign does not rely solely on FFT-derived phase extrapolation. As defined in Eqs. (8)–(10), future queries combine adaptive phase features with learned horizon queries and deterministic Fourier features, while retrieving historical context through cross-attention. Thus, the LL-periodicity of FFT-grid phase features does not impose the same constraint on the complete decoder.

We further evaluate frequency estimation across different period-to-window ratios:

Estimator diagnosticT/L=0.5T/L=1T/L=2T/L=4
Relative frequency error5.1%5.1%100%301%
Grid MAE/MAE₀0.610.321.511.01
4× zero-padding + interpolation0.220.491.061.24

The results demonstrate accurate frequency estimation when sufficient periodic evidence is available, with 4× interpolation reducing normalized error by 2.77× at T/L=0.5T/L=0.5. More importantly, the adaptive phase trajectory operates jointly with learned horizon queries and historical context, enabling multi-source future-query construction rather than relying on direct spectral extrapolation alone.

W5 / Q4(b) — High-dimensional generalization and heterogeneous nonstationarity. We additionally examine phase-specific controls on Electricity (321 variables) and Traffic (862 variables).

ΔMSE vs FullETTh1ElectricityTraffic
w/o RelPhase+4.0%+2.9%+3.3%
w/o FutPhase+4.8%+4.4%+4.0%
Unified-Rel+2.4%+4.8%+4.5%
Unified-Abs+3.1%+3.1%+3.6%

The positive differences support the stage-dependent mechanism across different dimensionalities. Additional tests cover close frequencies, amplitude modulation, intermittent cycles, abrupt changes, non-periodic mixtures, and noise:

ConditionGain vs best baseline
Close frequencies (1/5/10%)+14.2 / +11.7 / +6.9%
Amplitude modulation (0.2/0.5)+9.4 / +5.1%
Intermittent cycles (0.3/0.5/0.7)+8.8 / +12.3 / +10.6%
Abrupt changes (±π/2, ±π)+7.2 / +3.4%
Non-periodic mixture (α=0/0.5/0.75/1)+15.8 / +6.4 / +2.1 / −1.3%
Noise (σ=0/0.3/0.5/1.0)+16.1 / +9.2 / +4.6 / +1.2%

The advantage is positive in 17 of 18 conditions, gradually diminishing as periodicity weakens or noise increases. The fully non-periodic setting is the only negative case, indicating a natural limitation of the periodic inductive bias.

Thank you for the detailed rebuttal and additional experiments. While my concerns have been partially addressed, several key weaknesses remain unresolved. Therefore, I am maintaining my original score.

Reviewer_uibqOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PhaseAlign, a phase-aware framework for long-term time series forecasting. Its central idea is to use phase information differently in two stages: relative harmonic relations derived from dominant frequencies are incorporated into historical attention, while window-referenced phase states are propagated to future positions to condition the decoder. The model also includes time–frequency interaction, a local temporal pathway, and auxiliary phase-consistency and representation-alignment objectives. Experiments on seven standard forecasting benchmarks demonstrate its effectiveness.

Critical strengths
  1. The paper provides an intuitive motivation for distinguishing between identifying recurring patterns in historical observations and locating those patterns in the forecast horizon. This distinction offers a rationale for modeling phase differently across the two stages.
  2. The method delivers substantial performance improvements across multiple benchmark datasets.
Critical weaknesses
  1. The overall innovation of this paper is limited, as phase alignment and frequency domain analysis have been extensively studied in existing literature. The architecture appears to combine multiple established components, and the experiments do not yet isolate the value of the central design.
  2. The technical contributions appear largely incremental. Beyond the two phase operators, the architecture combines spectral-context extraction, time–frequency cross-attention, local convolutional processing, and multi-range historical aggregation. These components are individually familiar, and their integration is not a compelling technical contribution.
  3. The paper would benefit from clearer writing. In particular, the key distinctions from existing approaches are not sufficiently clear.
Questions for authors
  1. Time series foundation models (TSFMs) have increasingly become the dominant paradigm for time series forecasting. Can the proposed method be applied to foundation model pre-/post-training?
  2. The future phase propagation mechanism extrapolates phase information estimated from historical observations. How robust is this assumption when dominant frequencies change over time or when the series contains substantial non-periodic components?
  3. Phase estimates can be sensitive to noise, finite observation windows, and spectral leakage. How does PhaseAlign perform under these conditions, and how does estimation error affect its future predictions?
  4. What are the additional computational and parameter costs introduced by PhaseAlign compared with simpler phase-aware baselines?
3 comments
AuthorsOpenReview ↗

Response to Reviewer uibq (Part 1/2)

We thank the reviewer for the direct assessment. We would like to clarify that PhaseAlign's central contribution is a novel stage-dependent phase modeling principle, grounded in the distinct requirements of historical interaction modeling and future temporal placement. This principle goes beyond architectural integration by specifying how phase information should be used at each forecasting stage. We substantiate its novelty and effectiveness through theoretical analysis, operator-level validation, and controlled ablation studies, as detailed below.

W1–W3 — Novelty, contribution, and clarity.

Prior phase-aware methods use phase in one of four roles: as an alignment signal between representations, as part of a frequency-domain representation for self-supervision, as a modulation of features, or as a residual correction term. In each case a single phase convention is applied uniformly across the pipeline, and that convention is treated as a design detail rather than a requirement — nothing in these formulations specifies what the phase operator must be invariant or equivariant to. PhaseFormer, the closest phase-aware baseline, likewise applies one convention to a single attention mechanism.

PhaseAlign's contribution is to make the convention itself stage-dependent and to state it as a testable requirement: the historical operator must be invariant to the phase origin, because it scores compatibility between two observed positions; the future operator must be equivariant to that origin, because it places components at known offsets from it. The roles differ, so the operators must differ. This is a modelling principle rather than an additional module, and it predicts a specific failure for any single-convention design.

That prediction could have failed. If the two stages were interchangeable, both unified conventions would match PhaseAlign; they do not (ETTh1 0.421 vs 0.431 and 0.434, with the same ordering on ETTh2 and Exchange). At the operator level the requirement is satisfied exactly rather than approximately: historical sensitivity to a phase-origin shift is 0.0 against 1.998 for absolute coordinates at δ=π\delta=\pi, and future placement error is 0.076 against 1.245 without the origin anchor.

The principle is also not a free parameter choice. Conditioned on a component's magnitude and frequency, the only remaining degree of freedom is its phase, and the MMSE predictor given (Ak,fk)(A_k,f_k) but not ϕk\phi_k has error E[y2]=0.6250\mathbb{E}[y^2]=0.6250 versus 0.0000 when ϕk\phi_k is known. Magnitude priors alone confine any predictor to shrinkage, so fkf_k cannot be exploited except through ϕk\phi_k — which is precisely why the two stages cannot share one convention.

To distinguish the contribution of the phase operators from that of the surrounding architecture, Table R1 provides additional ablations. The phase-operator controls retain identical parameter counts (11.07M), enabling capacity-matched comparisons, while auxiliary module removals separately quantify their contributions.

Table R1. Contribution of phase operators and architectural components.

VariantParams (M)avg MSEΔ
Full11.070.421—
Phase operators → capacity-matched phase-blind11.070.468+11.2%
w/o both phase operators11.070.457+8.5%
w/o TF-interaction, fixed basis, local conv6.800.434+3.2%
− additionally − shared spectral context6.180.462+9.7%

Three results answer the concern. The phase operators are the largest single contributor (+8.5%); the time–frequency interaction, fixed basis and local convolution together contribute only +3.2%. A parameter-matched but phase-blind replacement (free-form learned attention bias and horizon embedding of identical shape) is worse than removing the operators entirely (+11.2% vs +8.5%) — extra capacity in the same slot overfits, while the phase structure generalises. The value is therefore in the operator structure, not in the number of modules.

AuthorsOpenReview ↗

Response to Reviewer uibq (Part 2/2)

Q1 — Applicability to foundation models.

The operators are an attention bias and a query offset, not a backbone: architecture-agnostic, and initialised so that a frozen model is unperturbed (the phase gate starts near zero). Both objectives (Eq. 14, Eq. 15) are self-supervised and usable in pre-training. We verify post-hoc adaptation directly:

SettingTrainable paramsavg MSE
Frozen backbone, head only0.04M0.441
+ PhaseAlign operators0.19M0.418
PhaseAlign-minimal, from scratch6.80M0.434

Q2 — Changing dominant frequencies and non-periodic content.

The spectrum is re-estimated per window, so across-window changes are handled by re-estimation rather than by an inherited stale estimate; one important failure mode is change within a window, which is what the drift study measures. Another limitation arises when the dominant period exceeds the observation window, leaving insufficient periodic evidence for reliable frequency estimation. PhaseAlign degrades more slowly than every neural baseline up to a zero-reference skill boundary of approximately 1.55π1.55\pi–1.73π1.73\pi, beyond which it is worse than predicting zero — a boundary we now state explicitly. For non-periodic mixtures the gain decreases monotonically from +15.8% (α=0\alpha=0) through +2.1% (α=0.75\alpha=0.75) to −1.3% (α=1\alpha=1); we report the negative case rather than omitting it.

Q3 — Noise, finite windows, and spectral leakage.

Estimation error propagates deterministically: the state is θk(h)=ϕ^k+2πf^k(L+h)\theta_k(h)=\hat{\phi}_k+2\pi\hat{f}_k(L+h), so a frequency error Δf\Delta f produces an additional horizon-dependent phase deviation Δθ(H)=2πΔfH\Delta\theta(H)=2\pi\Delta fH. With f^\hat{f} from top-K rFFT bins, the nearest-bin quantization error is at most Δf=1/(2L)\Delta f=1/(2L). At L=96L=96, a quarter-bin offset accumulates 0.25–1.88 cycles over H=96H=96–720720, while the worst-case half-bin offset accumulates 0.50–3.75 cycles. For on-grid components, quantization error is zero; the 24-step daily period corresponds to FFT bin m=L/24=4m=L/24=4 at L=96L=96, helping explain why the gain does not decay with horizon on the hourly benchmarks. Three mechanisms bound the damage: learned horizon queries, the fixed Fourier basis carrying no estimated frequency, and a learned gate that discounts the extrapolated phase. The gate shrinks from 0.71 to 0.28 as grid misalignment grows, i.e. the model hedges exactly where the analysis predicts. At the hardest half-bin offset, PhaseAlign reaches 0.0711 MSE versus 0.1067 for the next-best model; under additive noise, gains remain positive through σ=1.0\sigma=1.0.

Q4 — Computational and parameter cost.

Against simpler phase-aware baselines introduces no additional trainable parameters: all gating-only variants register identical parameters. PhaseAlign uses 11.07M parameters with 16.3 ms latency and 350 MiB peak memory under the matched setting, approximately 2.1× faster than FEDformer. Disabling both operators reduces latency by only 13.5%, so the phase machinery is not the bottleneck; the marginal cost is asymmetric — the harmonic bias adds 0.06 ms (a fixed cosine matrix) while the future phase state adds 2.1 ms. Accuracy is not purchased by budget either: a 4.19M-parameter configuration reaches 0.424 versus the default 0.421.

Thank you for your response. My concerns have been partially addressed. I still believe that the paper is not ready for publication and requires polishing in writing and figure presentation. However, based on the authors' clarification regarding the technical novelty, I increase my score to 2.

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

Like previous time series forecasting methods, PhaseAlign also leverages periodic patterns in time series data for prediction. However, PhaseAlign primarily focuses on frequency priors, phase alignment, and future phase prediction. Motivated by the observation that historical frequency information alone is insufficient to characterize future phase variations, the method incorporates FFT and frequency priors into feature construction to improve long-term time series forecasting. The experimental results demonstrate that the proposed method achieves certain improvements in forecasting performance.

Critical strengths

This paper presents a relatively clear research direction by leveraging frequency priors and phase structures to improve long-term time series forecasting. Its strengths lie in the well-motivated research problem grounded in physical principles and signal processing theory, a comprehensive model architecture, and the provision of a basic framework for experimental reproducibility.

Critical weaknesses
  1. It is unclear whether the performance improvements primarily stem from the core idea of phase modeling or simply from the increased representational capacity introduced by multiple architectural components.
  2. The method incorporates both frequency priors and phase alignment. However, the necessity of combining these two components is not sufficiently justified. Why must frequency priors be exploited through phase alignment to achieve effective forecasting?
  3. Does the FuturePhaseDecoder suffer from accumulated phase errors as the forecasting horizon increases?
  4. In Figure 4, although the authors use the same perturbation strength ρ\rho to control input corruption, an identical L2 distance does not necessarily imply an equivalent degree of semantic distortion. Phase perturbations may alter peak locations, periodic structures, and local temporal alignment, whereas magnitude perturbations primarily affect the strength of frequency components. Therefore, the observation that solid curves lie above dashed curves is insufficient to conclude that phase information is inherently more important than magnitude information.
Questions for authors
  1. It is unclear whether the performance improvements primarily stem from the core idea of phase modeling or simply from the increased representational capacity introduced by multiple architectural components.
  2. The method incorporates both frequency priors and phase alignment. However, the necessity of combining these two components is not sufficiently justified. Why must frequency priors be exploited through phase alignment to achieve effective forecasting?
  3. Does the FuturePhaseDecoder suffer from accumulated phase errors as the forecasting horizon increases?
  4. In Figure 4, although the authors use the same perturbation strength ρ\rho to control input corruption, an identical L2 distance does not necessarily imply an equivalent degree of semantic distortion. Phase perturbations may alter peak locations, periodic structures, and local temporal alignment, whereas magnitude perturbations primarily affect the strength of frequency components. Therefore, the observation that solid curves lie above dashed curves is insufficient to conclude that phase information is inherently more important than magnitude information.
2 comments
AuthorsOpenReview ↗

Response to Reviewer WdCB

We thank the reviewer for the constructive review. We address the four concerns through capacity-controlled ablations, theoretical clarification, phase-error analysis, and improved corruption controls.

W1 / Q1 — Capacity or mechanism? All original ablations retain identical parameter counts (11.07M, Appendix A.2), ruling out parameter count alone as an explanation for their differences. Table 2 already confirms the importance of both phase operators. More importantly, our additional capacity-matched phase-blind control yields 0.468 MSE, compared with 0.421 for PhaseAlign, supporting the contribution of structured phase modeling. Furthermore, a 4.19M-parameter variant achieves 0.424 MSE, suggesting that the improvement is not simply driven by model size.

W2 / Q2 — Why combine frequency and phase?

Frequency determines the rate of periodic evolution, while phase determines temporal placement. For yt=Asin⁡(2πft+ϕ)y_t=A\sin(2\pi ft+\phi), knowing AA and ff cannot determine the future waveform without its phase reference; under a uniform phase prior, E[yt∣A,f]=0\mathbb{E}[y_t\mid A,f]=0. Conversely, phase alone does not determine future phase advance without frequency.

This distinction motivates the stage-dependent design. During historical modeling, relative phase differences satisfy Δθk,ij=2πfk(τi−τj)\Delta\theta_{k,ij}=2\pi f_k(\tau_i-\tau_j), where the common phase origin cancels, enabling phase-origin-invariant harmonic interactions. Future prediction instead requires the observed reference ϕk\phi_k together with frequency fkf_k to determine temporal placement through θk,h=ϕk+2πfk(L+h)\theta_{k,h}=\phi_k+2\pi f_k(L+h).

Our operator analysis supports this asymmetry, with historical phase-origin sensitivity of 0.0 versus 1.998 for absolute coordinates, and future-placement error of 0.076 versus 1.245 without the origin anchor. Table 2 further shows that PhaseAlign outperforms Unified-Rel and Unified-Abs on ETTh1 (0.421/0.431/0.434), with consistent results on ETTh2 and Exchange. These findings support the complementary roles of frequency and phase and the stage-dependent operator design.

W3 / Q3 — Horizon-dependent phase error. Frequency estimation errors introduce additional phase deviation 2πΔfH2\pi\Delta fH over a forecast horizon HH. For L=96L=96, a quarter-bin offset corresponds to 0.25, 0.50, 0.88, and 1.88 cycles at H=96,192,336,720H=96,192,336,720, respectively. PhaseAlign mitigates reliance on extrapolated phase through learned horizon queries, fixed Fourier features, and gated phase conditioning (Eqs. 8–9). Thus, future prediction is not solely determined by accumulated phase estimates.

W4 / Q4 — Semantic calibration of Figure 4. We agree that equal L2L_2 distortion does not imply equal remaining predictability. The original Figure 4 reports 52.7% versus 36.5% MSE increases under phase and magnitude corruption at ρ≈0.31\rho\approx0.31. We therefore supplement it with corruption levels matched by a 10% increase in reference-predictor MSE.

Table R3. Predictability-matched corruption.

CorruptionRelative L₂Reference MSE ↑PhaseAlign MSE ↑
Phase shift0.15+10.0%+41.2%
Magnitude0.22+10.0%+18.6%
Additive noise0.18+10.0%+21.3%

Phase corruption remains 22.6 percentage points more damaging than magnitude corruption under this calibration. With consistent history–future transformations, PhaseAlign exhibits a 7.3% MSE increase, versus 31.8% for Unified-Rel and 44.5% for PatchTST. These results strengthen the phase-sensitivity finding without assuming network-wide equivariance.

My concerns have been well addressed.

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

This paper studies long-term multivariate time series forecasting. It proposes PhaseAlign, which shares a complex-spectrum representation but uses relative harmonic attention for historical modeling and window-referenced phase propagation for future queries. Experiments on seven benchmarks, matched phase controls, and synthetic and real-data interventions report competitive forecasting accuracy and support using different phase operators at the two stages.

Critical strengths
  • The distinction between historical comparison and future temporal placement is a useful way to formulate phase-aware forecasting.

  • PhaseAlign implements that distinction through two explicit, complementary operators: sample-dependent harmonic attention biases and phase-propagated future queries.

  • Experimental results reports the best or tied-best average MAE on seven datasets.

Critical weaknesses
  • The paper correctly restricts invariance to dedicated operators rather than the entire network. However, Appendix A.2 implements Unified-Abs by replacing a pairwise harmonic bias with token-wise conditioning and sharing the future-phase projection.

  • The practical cost of the complete architecture has not been established. Appendix E gives useful asymptotic complexity, but neither it nor the main results reports absolute parameter counts, latency, peak memory, or training cost against comparison methods.

  • The corruption study changes history while keeping future targets fixed. Equal relative L2 distortion does not equalize timing mismatch or remaining predictability. As shown in Section 4.4.2, the 52.7% versus 36.5% MSE increases therefore show different sensitivities, but do not uniquely identify the claimed phase mechanism.

  • Robustness tests use single-tone or two-component harmonic signals in Appendix. They do not cover close-frequency mixtures, changing amplitudes, intermittent cycles, or abrupt regime changes.

Questions for authors
  1. Does the proposed pairing remain superior when access to the shared phase-bearing descriptors is controlled identically?

  2. Please provide more practical cost analysis of PhaseAlign.

  3. Do the phase-versus-magnitude conclusions persist when the full history-future trajectory is transformed consistently, or when an oracle using the corrupted history quantifies the remaining predictability?

  4. Do the phase-robustness conclusions extend beyond the harmonic and single-tone settings?

3 comments
AuthorsOpenReview ↗

Response to Reviewer NCkF (Part 1/2)

We thank the reviewer for the constructive feedback. We clarify the stage-dependent controls and provide additional analyses of computational efficiency, corruption sensitivity, and robustness.

W1 / Q1 — Operator form and descriptor access.

We thank R2 for this observation. Our controls compare stage-specific phase operators: historical relative phase models pairwise harmonic relations, while future absolute phase preserves the observed temporal reference. Their different injection forms reflect these distinct roles.

Crucially, all variants share the same spectral descriptors, positional backbone, training objective, and parameter count. Unified-Abs reuses the existing projection without adding parameters. As shown in Table 2, PhaseAlign consistently outperforms both unified-phase controls across all three datasets, supporting the stage-matched operator design. These controls compare complete phase operators rather than isolating phase convention independently of injection.

W2 / Q2 — Practical computational cost.

We benchmark all models under identical hardware and input settings (L=H=96L=H=96, C=7C=7, batch size 32).

Table R2. Efficiency comparison.

ModelParams (M)Latency (ms)Throughput (win/s)Memory (MiB)Train step (ms)
DLinear0.142.115,238386.2
PatchTST1.218.73,67817627.4
iTransformer2.6310.92,93620833.1
Crossformer5.4222.31,43538268.9
TimesNet1.1427.61,15942481.5
FEDformer14.5134.2936618104.3
PhaseAlign11.0716.31,96935053.0

PhaseAlign is 2.1× faster than FEDformer, with approximately 24% fewer parameters and 43% less peak memory, although it is slower than PatchTST. Disabling both phase operators reduces latency by 13.5%, indicating that these operators are not the dominant computational cost. All original gating-only controls retain identical registered parameter counts. A supplementary width sweep also examines smaller configurations to characterize the accuracy–efficiency trade-off.

AuthorsOpenReview ↗

Response to Reviewer NCkF (Part 2/2)

W3 / Q3 — Corruption sensitivity.

We agree that equal input-space L2L_2 distortion does not necessarily imply equal remaining predictability. In the original manuscript (Fig. 4), at matched relative L2L_2 distortion (ρ≈0.31\rho\approx0.31), phase and magnitude corruption increase MSE by 52.7% and 36.5%, respectively. We additionally evaluate corruption calibrated to matched reference-predictor degradation.

Table R5. Corruption at matched reference degradation.

CorruptionRelative L₂Reference MSE ↑PhaseAlign MSE ↑
Phase shift0.15+10.0%+41.2%
Magnitude0.22+10.0%+18.6%
Additive noise0.18+10.0%+21.3%

At matched reference degradation, phase corruption causes a 22.6-percentage-point larger MSE increase than magnitude corruption, supporting the distinct importance of temporal placement. Under consistent history–future transformations, PhaseAlign exhibits a 7.3% MSE increase, compared with 31.8% for Unified-Rel and 44.5% for PatchTST. These results support the operator-level design without assuming exact network-wide equivariance.

W4 / Q4 — Robustness beyond simple harmonics.

We extend the evaluation to heterogeneous temporal conditions, including close frequencies, amplitude modulation, intermittency, regime changes, non-periodic components, and noise.

Table R6. Robustness across heterogeneous conditions.

ConditionGain vs best baseline
Close-frequency mixtures (1/5/10%)+14.2 / +11.7 / +6.9%
Amplitude modulation (0.2/0.5)+9.4 / +5.1%
Intermittent cycles (0.3/0.5/0.7)+8.8 / +12.3 / +10.6%
Abrupt changes (±π/2, ±π)+7.2 / +3.4%
Non-periodic mixture (α=0/0.5/0.75/1)+15.8 / +6.4 / +2.1 / −1.3%
Noise (σ=0/0.3/0.5/1.0)+16.1 / +9.2 / +4.6 / +1.2%

PhaseAlign achieves positive gains in 17 of 18 settings, with generally decreasing benefits as periodic structure weakens or noise increases. The fully non-periodic case is the only negative setting, revealing a limitation of the periodic inductive bias.

Thank you for the clarifications and additional experiments. Most of my concerns have been addressed. I'll keep my original rating.

Comments on the paper

1 comment

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.