acceptodds
Withdrawn from ICLR 2027

PaCo: Evidence-Supported Phase Coordinates for Alignment-Aware Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-term time series forecasting is challenging because recurring temporal patterns may reappear at different temporal positions across observation windows. Such alignment variation changes local observations while preserving the underlying pattern, making forecasting representations prone to mixing pattern content with alignment state. Existing temporal and spectral forecasting models have improved pattern extraction, and recent phase-aware methods further introduce spectral phase into temporal modeling. However, using phase is not equivalent to representing it as an alignment state. Spectral phase is a circular variable associated with temporal shifts rather than an unconstrained scalar, while the stability of an observed phase direction is related to the amplitude of the corresponding Fourier coefficient. In this work, we propose PaCo, an evidence-supported phase coordinate conditioning framework for time series forecasting. PaCo constructs spectral phase coordinates that preserve circular phase direction and incorporate amplitude-derived evidence for the observed direction. These coordinates serve as local alignment states to condition temporal representations and guide variable-level context modeling. Experiments on public long-term forecasting benchmarks and an industrial Wastewater Treatment Aeration (WWTA) dataset show that PaCo achieves competitive forecasting performance, while ablation and phase-response analyses support the importance of the proposed evidence-supported phase coordinate design. The code is available at https://anonymous.4open.science/r/Anonymization-68A4/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Frequency-Domain Learning, Spectral Representation Learning, Phase Information, Representation Learning

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

This paper proposes PaCo, a phase-aware framework for long-term time series forecasting. It combines circular phase features with an amplitude-based indicator to construct spectral phase coordinates. These coordinates modulate temporal patch representations and guide interactions between variables. Experiments on public benchmarks and an industrial wastewater dataset show strong forecasting performance with a moderate parameter count and relatively low reported computational cost.

Critical strengths

S1. PaCo achieves strong forecasting results on multiple datasets, with particularly good results on Exchange and the industrial WWTA dataset.

S2. The model obtains these results with a moderate number of parameters and relatively low reported computational cost.

S3. The experiments cover recent forecasting baselines, component ablations, parameter sensitivity, and efficiency comparisons.

Critical weaknesses

W1. The presentation is difficult to follow. The introduction discusses shifts across observation windows, but the transition to patch-level modulation and cross-variable interaction does not make clear which alignment variation each module addresses. Figure 1 illustrates sensitivity to coefficient perturbations rather than the temporal misalignment problem itself. A concrete example would make the motivation and structure easier to understand.

W2. The FFT is computed on patches of only 16 time steps. When the relevant periods are longer than a patch, spectral leakage and boundary effects may make the extracted phase sensitive to the window position. The paper acknowledges boundary effects, but does not examine phase reliability or forecasting performance across patch lengths. This leaves the suitability of such short patches insufficiently justified.

W3. Affine modulation changes feature scales and offsets, but this alone does not establish that the learned coordinates track temporal alignment. Figure 3 probes constructed descriptors rather than actual shifted input sequences. More direct evidence is needed to distinguish alignment-aware representations from general spectral conditioning.

W4. The amplitude-based indicator is motivated by equal-magnitude coefficient perturbations, but its relation to phase reliability under realistic noise remains unclear. Since amplitude is already provided to the encoder, comparisons with alternative amplitude normalizations would help justify this particular indicator.

W5. The ablations remove evidence, phase direction, and phase-guided context, but do not isolate the contribution of SPRL itself. A version without the affine modulation, together with a capacity-matched comparison using conventional Fourier features, would better identify the source of the improvement.

Questions for authors

1.Could the authors illustrate the intended alignment variation on real input windows and show how the learned coordinates and forecasting errors respond to controlled temporal shifts?

2.How do phase stability and forecasting accuracy change with patch length, particularly when a patch covers less than one period of the underlying pattern?

3.Could the authors separately remove SPRL and compare the proposed coordinates with conventional Fourier features under the same model capacity? An alternative amplitude normalization would also help assess the evidence indicator.

Reviewer_5uG2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of phase shifts in periodic patterns across different observation windows in long-term time series forecasting and proposes the PaCo framework. Based on the Fourier shift theorem, the authors treat spectral phase as a circular variable describing the state of local alignment and point out that components with smaller amplitudes exhibit more unstable phases. Therefore, the sine and cosine of the phase are used to preserve the circular structure, and an evidence metric derived from amplitude is added to indicate the reliability of the phase direction. This information is encoded into phase coordinates, which are used to perform time-gated residual affine modulation and to compute cross-variable attention. Experiments were conducted on seven public benchmarks and an industrial wastewater treatment dataset, where PaCo achieved optimal or second-best results in most settings.

Critical strengths
  1. The paper examines phase from the perspective of alignment-aware representations, pointing out that using phase does not equate to representing it as an alignment state; its motivation is clear and it offers some novel insights.
  2. The gate-controlled residual affine modulation has a simple design; the same set of phase coordinates is reused for cross-variable attention, eliminating the need for additional Fourier transforms.
  3. The experimental evaluation is reasonably broad. The paper includes recent strong baselines and ablations of its core modules. The results suggest that the method is effective to some extent and offers a cost advantage over certain heavier models.
Critical weaknesses
  1. Equation (24) in Appendix C keeps the amplitude at all frequency bins equal and fixed, yet sets ρ to 0.2, 0.5, and 0.8. Under Equation (3), if α is unchanged, ρ should also remain unchanged. Figures 3 and 5 therefore show the network's response to manually specified features. The authors do not directly demonstrate that the model can identify local misalignment in real time series and use it to improve forecasts. The paper also lacks forecasting experiments with actual temporal misalignment.
  2. In terms of MSE, PaCo has five wins and two losses across the seven public datasets. The gains on public datasets are modest, and the paper does not report variability across runs or confidence intervals, or provide sufficient comparisons with closely related methods. Gains are larger on the industrial dataset, but the dataset is not public, making independent replication difficult. Moreover, Table 6 reports 0.36M parameters for PaCo without specifying the corresponding input and prediction lengths, making it hard to determine which experimental configurations the efficiency figures apply to.
  3. The paper lacks experiments to verify key hypotheses, such as testing on synthetic data or artificially shifted windows, to determine whether PaCo is truly more robust to alignment changes and whether the benefits stem from alignment modeling.
Questions for authors
  1. If α is held constant in Equation (24), how can Equation (3) yield different values of ρ? How should Figures 3 and 5 be interpreted in relation to features derived from real inputs?
  2. Could the authors introduce controlled local temporal misalignment into public datasets and report forecasting results and mechanism responses at different degrees of misalignment? Could they provide variability across multiple runs or confidence intervals, as well as comparisons with more closely related methods?
  3. Which input and prediction lengths correspond to the 0.36M parameters in Table 6? Could the authors report computational cost, inference latency, and peak GPU memory usage for each configuration?
Reviewer_Ft6nOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PaCo, a phase-coordinate conditioning framework for long-term time-series forecasting. For each normalized temporal patch, it computes selected non-DC Fourier coefficients and encodes log amplitude, circular phase direction (sine and cosine), and a bounded relative-amplitude indicator into a learned phase coordinate. The coordinate generates gated affine modulation of temporal tokens and also contributes to queries and keys for dense variable-level contextual attention. The authors evaluate on eight datasets at four forecast horizons, compare against eight baselines, and report component ablations, parameter sensitivity, phase-response probes, parameter counts, and FLOPs.

Critical strengths

The paper gives a coherent representation-level motivation for treating spectral phase as a circular alignment cue rather than an unconstrained scalar. The patch-aligned encoding preserves phase direction with sine/cosine features and pairs it with amplitude information, then reuses the learned coordinate for both local temporal modulation and variable-level context.

Critical weaknesses

The paper motivates PaCo as a way to handle local alignment variation, but it does not test forecasting under controlled shifts of the actual input. Figure 3 and Appendix C perturb phase descriptors while leaving the input unchanged, so they show how the conditioning pathway responds rather than whether forecasts become more robust to real input shifts. The ablation suggests that the relative-amplitude indicator can help prediction, but does not establish that it tracks phase reliability on real noisy patches. The theoretical bound depends on perturbation magnitude, whereas the implemented indicator is relative amplitude within a patch. The reported results leave uncertainty about robustness and deployment efficiency. PaCo is run with three seeds, but the tables provide means without dispersion, and the baseline seed counts are not clearly stated; this makes close comparisons and ablation differences difficult to assess. WWTA adds an industrial case, but it covers one month at one plant and is not publicly released, limiting independent verification. The efficiency analysis reports parameter counts and FLOPs rather than latency or memory, while the variable-level attention in Equation 9 is dense over variables and therefore has quadratic attention-score cost in the number of variables. The largest evaluated public dataset has 321 variables, so the reported results do not establish scaling to substantially larger variable counts.

Questions for authors
  • Do the authors have an experiment applying controlled temporal shifts to the actual input patches or windows? If not, such a comparison against the phase/evidence ablations would directly test whether PaCo improves forecasting under alignment changes, beyond showing that its conditioning pathway responds to synthetic phase descriptors.
  • Does the implemented relative-amplitude indicator ρ rank the reliability of observed phase directions on real patches under controlled noise or perturbations? Please clarify how this quantity relates empirically to realized phase error, given that the theoretical bound depends on perturbation magnitude and ρ is not calibrated as a probability.
  • Please report the number of seeds used for each baseline and provide variability (for example, standard deviations or confidence intervals) for PaCo, baselines, and the principal ablations, especially for close results.
  • Could the authors report actual inference latency and memory as the number of variables grows, and discuss the quadratic variable-attention cost in Equation 9? The current FLOPs summary and evaluation up to 321 public variables do not establish deployment behavior at larger scales.

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

Related papers

Discussion (0)

Sign in to comment.