acceptodds
Under review as a conference paper at ICLR 2027

HarMoTC: Harmonic Temporal Conditioning with Phase–Amplitude Spectral Modeling

Abstract

Existing time-series forecasting methods primarily model pointwise dependencies in the time domain, making it difficult to preserve both periodic strength and temporal alignment. Moreover, periodic temporal cues are commonly treated only as auxiliary input features. To address these limitations, we propose HarMoTC, a framework that performs temporally conditioned phase–amplitude spectral modeling for multivariate time-series forecasting. HarMoTC explicitly decouples the amplitude and phase of the complex spectrum and employs learnable amplitude rescaling and phase calibration to preserve periodic strength while maintaining temporal alignment. Furthermore, we introduce a conditional spectral modulation mechanism driven by multi-harmonic Fourier encodings of temporal indices, enabling the model to directly modulate spectral transformations according to temporal context and thereby adapt its frequency responses. The resulting architecture jointly captures periodic structure, context-dependent spectral variations, and cross-variable dependencies. Experiments on multiple real-world datasets demonstrate that HarMoTC achieves competitive forecasting performance.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate time-series forecasting, Frequency-domain modeling, Phase–amplitude decoupling, Harmonic temporal conditioning, Conditional spectral modulation

Reviews

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

HarMoTC is a frequency-domain framework for multivariate long-term time-series forecasting. On top of a shared spectral backbone (instance normalization → RFFT → adaptive spectral gating → linear cross-variable mixing Q∈R^{C×C} → IRFFT → temporal projection), it adds two components: (1) calendar-conditioned spectral modulation, which encodes the forecast-origin hour-of-day/day-of-week indices into K Fourier harmonics and maps them through an MLP (a FiLM-style generator, per Table 8) to channel- and frequency-specific parameters (γ,β) applied in asymmetric pre- and post-modulation steps (Eqs. 8–9); and (2) phase–amplitude decoupled spectral modeling, which polar-decomposes the conditioned spectrum and applies learnable frequency-wise amplitude gains g and phase shifts δ generated from two temporal kernels (Eqs. 11–12). Training uses a pure complex-spectral MAE objective (Eq. 14). Validation covers 7 benchmarks × 4 horizons against 9 baselines, with core-design ablations, calendar-stratified spectral fidelity, a counterfactual calendar-alignment intervention with bootstrap CIs, efficiency, and a three-seed robustness study. Headline claims: best or tied-best on 11/14 dataset–metric pairs; macro-averaged MSE/MAE reductions of 3.18%/1.64% over the strongest baseline (SSformer).

Critical strengths

A causal-style evidence chain for the central mechanism, rare in this genre. The counterfactual intervention (Tables 4/9) demonstrates — per horizon, with CIs — that the model functionally relies on correctly aligned calendar semantics, and the calendar-stratified spectral analysis (Table 3, 96/96 phase-group wins) independently corroborates the channel through which conditioning acts. This is consequential because it elevates the paper from "ablation shows a drop" to genuine mechanism-level evidence, and it is the strongest reason the work deserves serious consideration.

Critical weaknesses

The second headline contribution — phase–amplitude decoupled modeling — is weakly evidenced and arbitrarily designed. The ablation shows the two spectral branches jointly contribute markedly less than calendar conditioning (w/o Amp&Phase degrades 0.2785→0.2908 MSE, +4.4%; w/o TC degrades to 0.2968, +6.55%), and the conditioning path itself barely touches phase: pre-modulation (Eq. 8) rescales only the real part and divides by (1+|γ|+ε), post-modulation (Eq. 9) multiplies by (1+γ) — neither symmetric nor inverse — and Appendix A.3 concedes the mechanism "is not a fully general complex-valued transformation." This is consequential because the title, abstract, and contribution list place phase modeling on equal footing with temporal conditioning; the evidence instead indicates the effective contribution is calendar FiLM, with the phase–amplitude machinery adding a modest increment whose design is never justified.

Questions for authors

1: What justifies the asymmetry between Eq. 8 (real-part only, divide by 1+|γ|+ε) and Eq. 9 (multiply by 1+γ, add β)? Did you try full complex-valued modulation of real/imaginary (or amplitude/phase) parts, and what were the results?

2: The ablation omits Traffic. Does removing temporal conditioning improve Traffic (i.e., is calendar conditioning harmful for heterogeneous sensor networks)? This materially affects the interpretation of the Traffic failure.

6 comments
AuthorsOpenReview ↗

Clarifying HarMoTC's Technical Contributions, Modulation Design, and Traffic Ablation

Response to Reviewer CSiB (1/2): Contribution positioning and design rationale

Thank you for recognizing the calendar-alignment intervention and spectral analysis. We clarify the contribution hierarchy and design below; Part 2 provides the modulation comparisons and Traffic ablation. Table numbers refer to the revised manuscript.

Contribution and supporting evidence

Calendar-conditioned modulation is the core mechanism for adapting spectral transformations to temporal context. It generates variable- and frequency-specific parameters from calendar indices. Within this framework, phase-amplitude decoupled modeling provides explicit amplitude and phase controls. Both designs can affect amplitude and phase, but their parameter sources and roles differ.

For Z(q)=A(q)exp⁡[iΦ(q)]Z(q)=A(q)\exp[i\Phi(q)], Eq. (12) gives

U(q)=[A(q)⊙g]⊙exp⁡{i[Φ(q)+δ]}.U(q)=[A(q)\odot g]\odot\exp\{i[\Phi(q)+\delta]\}.

Separate temporal kernels generate the frequency-dependent gain gg and phase shift δ\delta, shared across samples and variables. Calendar dependence is supplied by the modulation.

Table 2 measures the module's additional benefit with calendar conditioning retained. Removing both branches increases macro-average MSE from 0.2785 to 0.2908 (+4.4%) across four datasets and four horizons. Removing amplitude rescaling alone gives 0.2870; removing phase calibration gives 0.2843 (+2.08%). The latter retains the input phase, calendar modulation, and complex spectral loss. The phase kernel has only 96 parameters at L=96L=96.

Revised abstract, introduction, and contributions

The revised contribution list first presents the integrated framework, then details its core calendar-modulation mechanism. Amplitude-phase modeling remains part of the framework contribution, with its additional benefit evaluated separately. The title describes the complete architecture; the model and equations are unchanged.

Abstract, method description:

To address these limitations, we propose HarMoTC, a spectral framework for multivariate long-term time-series forecasting that combines calendar-conditioned modulation with phase-amplitude decoupled modeling. Calendar-conditioned modulation forms the core of the framework, using Fourier calendar encodings to adapt spectral transformations to temporal context. Separate temporal kernels provide amplitude rescaling and phase calibration within the conditioned framework, adjusting periodic strength and relative phase.

Introduction overview (Section 1):

HarMoTC integrates calendar-conditioned modulation and phase-amplitude decoupled modeling within a shared spectral forecasting backbone. As the core mechanism, calendar conditioning uses Fourier encodings of hour and/or weekday indices to generate variable- and frequency-specific modulation parameters, adapting spectral transformations to temporal context. Within this framework, separate temporal kernels generate frequency-dependent amplitude gains and phase shifts for spectral reconstruction. These responses are shared across samples and variables, with calendar dependence supplied by the modulation.

Contributions 1 and 2:

We propose HarMoTC, a multivariate long-term spectral forecasting framework centered on calendar conditioning and incorporating phase-amplitude decoupled modeling. It adapts to temporal context and uses separate temporal kernels to adjust periodic strength and relative phase.

We design a spectral modulation mechanism based on Fourier calendar encodings, generating variable- and frequency-specific parameters to directly condition spectral transformations on temporal context.

Q1. Why use asymmetric modulation?

Eq. (8) subtracts β\beta and divides by 1+∣γ∣+ϵ1+|\gamma|+\epsilon. The divisor is bounded below by 1+ϵ1+\epsilon, avoiding division by a near-zero value. Eq. (9) applies (1+γ)⊙Rv+β(1+\gamma)\odot R_v+\beta after spectral transformation and cross-variable interaction, allowing signed affine adjustment. These operations serve input scaling and output adjustment, respectively, and are not intended as inverse transformations.

Keeping the imaginary part unchanged does not fix the phase: scaling or shifting the real part can change the argument of the complex coefficient. The calendar-stratified analysis (Table 4) reports 5.18% lower weighted phase error, with improvements in all 96 hour-horizon groups. This evaluates calendar conditioning, while Table 2 separately measures the benefit of learnable phase calibration. Part 2 compares the requested alternatives.

Response to Reviewer CSiB (2/2): Alternative modulations and Traffic ablation

Table numbers refer to the revised manuscript.

Q1. Results of alternative modulation designs

For z=R+iIz=R+iI and v=Rv+iIvv=R_v+iI_v, we evaluated:

  • Symmetric real: s=exp⁡(γ)>0s=\exp(\gamma)>0; pre-modulation (R−β)/s+iI(R-\beta)/s+iI, post-modulation sRv+β+iIvsR_v+\beta+iI_v.
  • Full complex affine: a=1+γ+iγia=1+\gamma+i\gamma_i, b=β+iβib=\beta+i\beta_i; pre-modulation (z−b)a‾/(∣a∣2+ϵ)(z-b)\overline a/(|a|^2+\epsilon), post-modulation av+bav+b. Two additional heads generate the imaginary gain and bias.
  • Direct calendar amplitude-phase: c(q)=exp⁡[γ(q)+iβ(q)]c(q)=\exp[\gamma(q)+i\beta(q)]; pre-modulation z/c(q)z/c(q), post-modulation c(q)vc(q)v, with log-gain γ\gamma and phase rotation β\beta.

Both complex variants preserve the real-valued constraints at DC and Nyquist frequencies. All three retain the original amplitude and phase branches, so this comparison tests calendar injection. These results are included in revised Table 3.

Results use L=T=96L=T=96 and seeds {2024, 2025, 2026}. Entries are mean ± sample SD; means and parameters are truncated to three decimals, SDs to four. Parameters are in millions (M).

DatasetMetricOriginalSymmetric realComplex affineCalendar amp-phase
ETTh1MSE0.353±0.00110.360±0.00770.364±0.00470.371±0.0034
MAE0.385±0.00130.387±0.00430.391±0.00150.396±0.0020
Params (M)0.9060.9061.0830.906
WeatherMSE0.147±0.00080.149±0.00070.150±0.00060.158±0.0005
MAE0.189±0.00110.190±0.00100.192±0.00090.202±0.0006
Params (M)0.8010.8010.9340.801
ElectricityMSE0.140±0.00010.142±0.00040.151±0.00030.161±0.0004
MAE0.239±0.00020.241±0.00040.251±0.00080.261±0.0005
Params (M)2.3952.3953.4332.395
TrafficMSE0.435±0.00090.438±0.00190.434±0.00090.468±0.0025
MAE0.287±0.00110.291±0.00290.288±0.00020.314±0.0009
Params (M)4.7844.7847.5724.784

The original has the lowest mean MSE and MAE on ETTh1, Weather, and Electricity. On Electricity, complex affine and direct calendar amplitude-phase modulation increase mean MSE by 7.39% and 15.00%.

On Traffic, complex affine has 0.19% lower mean MSE, but 0.32% higher mean MAE and 58.26% more parameters; the original has the lowest mean MAE. Percentages use untruncated values. These results support the original modulation's accuracy/parameter trade-off in the reported settings; additional complex modulation freedom does not consistently improve both metrics.

Q2. Does removing temporal conditioning improve Traffic?

No. We evaluated w/o TC at four horizons with L=96L=96 and seeds {2024, 2025, 2026}, removing both calendar-conditioned pre- and post-modulation while retaining the other forecasting components.

Entries are MSE / MAE. The full-model column is the single-run reference from Table 13; the remaining columns are additional w/o TC runs.

HorizonHarMoTC (Table 13)w/o TC, 2024w/o TC, 2025w/o TC, 2026
960.435 / 0.2870.468 / 0.3140.466 / 0.3110.466 / 0.311
1920.452 / 0.2980.482 / 0.3190.479 / 0.3180.481 / 0.319
3360.473 / 0.3140.497 / 0.3260.494 / 0.3240.497 / 0.326
7200.505 / 0.3360.555 / 0.3590.551 / 0.3570.548 / 0.354

Averaging horizons within each seed, then computing mean ± sample SD across seeds from the displayed values, gives approximately MSE 0.4986 ± 0.0016 and MAE 0.3281 ± 0.0011 (truncated to four decimals).

For the same four-horizon summary, Table 6 reports the full model's three-seed results as MSE 0.466 ± 0.003 and MAE 0.309 ± 0.003. Removing conditioning therefore increases average MSE/MAE by approximately 7.0%/6.2%. Every w/o TC entry also exceeds the corresponding Table 13 full-model reference.

Calendar conditioning thus remains beneficial on Traffic. Removing it does not resolve the gap to strong baselines, so that gap cannot simply be attributed to harmful calendar conditioning in heterogeneous sensor networks.

Thank you again for your constructive feedback. We would be happy to provide further clarification on any remaining concerns.

Dear Reviewer CSiB,

On Weather (L = T = 96), we further examine phase calibration through forecasting results and an output-phase diagnostic from repeated runs. Both variants retain calendar conditioning and amplitude rescaling.

MetricFullw/o PhaseReduction
MSE0.1460.1492.01%
MAE0.1870.1901.58%
Weighted phase error0.6022640.6255553.72%

MSE/MAE are truncated to three decimals, and their reductions use the displayed values. The phase diagnostic (seed 2024; all 10,444 test windows) shows a 3.72% reduction, with a 95% paired block-bootstrap interval of [0.008178, 0.036860] for w/o Phase − Full. This complements our ablations by linking phase calibration to improved output phase alignment within the calendar-conditioned framework.

We are writing to ask whether our rebuttal has addressed your concerns. We would be happy to discuss any remaining questions.

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

This paper presents HarMoTC, a spectral framework for multivariate long-term time series forecasting. It combines calendar-conditioned spectral modulation with separate amplitude and phase modeling to capture context-dependent periodic patterns. Calendar indices are represented through multi-harmonic Fourier encodings, while learnable spectral responses adjust periodic strength and temporal alignment. Experiments on seven benchmark datasets show competitive forecasting performance and support the usefulness of the proposed components.

Critical strengths
  1. A focused design with direct ablation support. Using calendar information to control spectral transformations is the most valuable aspect of the paper. It allows the model to adapt periodic patterns to the context at the forecast origin, while separate amplitude and phase responses provide a clear modeling interface. On the four ablation datasets, the full model consistently outperforms the component-removal variants; removing temporal conditioning increases macro-average MSE by approximately 6.55%. These results support the practical value of the combination and are a principal reason for my positive recommendation.

  2. Validation examines meaningful use of calendar information. On ETTm1, calendar conditioning reduces the reported phase error across all 96 forecast steps. Changing only the calendar index while fixing the input and model substantially worsens prediction quality, supporting the importance of correct calendar alignment. The moving-block bootstrap addresses dependence between overlapping windows. Together with the seven-dataset comparison and three-seed analysis, these results provide a persuasive empirical case. The explicit reporting of weaker Traffic performance also keeps the overall claims appropriately bounded.

  3. A well-motivated focus on context-dependent periodicity. One aspect I find valuable is the connection between the motivating observations and the modeling objective. The calendar-stratified amplitude analysis in Figure 1 illustrates how periodic strength varies across calendar contexts, while the controlled amplitude and phase perturbations highlight their effects on reconstruction quality. These observations provide a concrete rationale for considering calendar context, periodic strength, and spectral alignment together. This is a meaningful forecasting objective, and the paper makes a clear connection between the observed data characteristics and its architectural choices.

Critical weaknesses
  1. The asymmetric calendar modulation needs stronger empirical justification. The conditional transformations in Equations (8) and (9) operate only on the real component, leaving the imaginary component unchanged. Although the paper explains this asymmetric design, it does not compare it with alternatives, such as modulating both the real and imaginary components. Consequently, the specific advantages of this design remain unclear.

  2. The detailed validation of calendar conditioning is largely limited to ETTm1. The calendar-stratified spectral error analysis and calendar-alignment interventions provide convincing evidence, but both are conducted only on ETTm1. Although ablation studies on other datasets support the effectiveness of temporal conditioning, they do not directly verify whether similar spectral improvements and alignment behavior hold across datasets. Repeating one key diagnostic on another representative dataset would strengthen the generality of this interpretation.

  3. The efficiency evaluation does not fully characterize the model’s resource requirements. Figure 4 reports parameter counts and training time per epoch, but does not include inference latency or peak GPU memory usage. Moreover, the comparison is limited to ETTm1 under the 96→96 setting. Given that HarMoTC has more parameters than most of the compared baselines, additional resource measurements on a dataset with more variables would provide a more complete assessment of the trade-off between forecasting performance and computational cost.

Questions for authors
  1. What empirical evidence supports the asymmetric modulation in Equations (8) and (9), particularly compared with a variant that modulates both real and imaginary components?

  2. To what extent do the spectral improvements and calendar-alignment effects observed on ETTm1 generalize to other datasets? Can the authors provide a representative diagnostic beyond ETTm1?

  3. Given HarMoTC’s relatively high parameter count, how does it compare with representative baselines in inference latency and peak GPU memory usage on datasets with more variables, such as Electricity or Traffic?

4 comments
AuthorsOpenReview ↗

Clarifying HarMoTC's Modulation Design, Cross-Dataset Validation, and Efficiency

Response to Reviewer UGkz (1/2): Modulation Design

Thank you for the constructive questions. We provide additional controlled comparisons, cross-dataset mechanism diagnostics, and inference resource measurements below.

Q1. Empirical justification for modulating only the real component

Response. To directly test the suggested alternative, we compared the original real-component modulation with full complex affine modulation, which acts on both real and imaginary components. For complex spectra z and v before pre- and post-modulation, respectively, the alternative uses

a=1+γ+iγi,b=β+iβi,a=1+\gamma+i\gamma_i,\qquad b=\beta+i\beta_i,

z′=(z−b)a‾∣a∣2+ϵ,v′=av+b.z'=\frac{(z-b)\overline a}{|a|^2+\epsilon},\qquad v'=av+b.

Two additional calendar-conditioned heads generate the imaginary gain and bias. The amplitude-rescaling and phase-calibration branches and the remaining backbone are retained, with real-valued constraints at the DC and Nyquist bins. Thus, this comparison directly evaluates whether the additional complex modulation freedom improves prediction.

We evaluated both schemes on four datasets with L=T=96 and seeds {2024, 2025, 2026}. Entries below report MSE / MAE, each as mean ± sample standard deviation; parameter counts are trainable parameters in millions.

DatasetOriginal: MSE / MAEComplex affine: MSE / MAEParameters: original / complex affine (M)
ETTh10.353±0.0011 / 0.385±0.00130.364±0.0047 / 0.391±0.00150.906 / 1.083
Weather0.147±0.0008 / 0.189±0.00110.150±0.0006 / 0.192±0.00090.801 / 0.934
Electricity0.140±0.0001 / 0.239±0.00020.151±0.0003 / 0.251±0.00082.395 / 3.433
Traffic0.435±0.0009 / 0.287±0.00110.434±0.0009 / 0.288±0.00024.784 / 7.572

The original modulation achieves lower mean MSE and MAE on ETTh1, Weather, and Electricity. On Traffic, complex affine modulation gives a slightly lower mean MSE (0.19%), while increasing MAE by 0.32% and parameter count by 58.26%. These results support the original design as an effective accuracy–parameter trade-off across the tested settings.

The design also allows calendar-dependent phase adjustment while leaving the imaginary component unchanged. After the original pre-modulation,

ϕ′=atan2⁡ ⁣(Ig,Rg−β1+∣γ∣+ϵ),\phi'=\operatorname{atan2}\!\left(I_g,\frac{R_g-\beta}{1+|\gamma|+\epsilon}\right),

so changing the calendar-dependent γ\gamma or β\beta generally changes the phase. Furthermore, Equation (8) uses a denominator lower-bounded by 1+ϵ1+\epsilon, avoiding amplification caused by a near-zero divisor, while Equation (9) permits a signed affine adjustment after spectral transformation and cross-variable interaction. The empirical comparison above provides direct support for this constrained parameterization relative to the tested alternative that modulates both components.

Response to Reviewer UGkz (2/2): Cross-Dataset Diagnostics and Efficiency

Q2. Generalization of spectral improvements and calendar alignment beyond ETTm1

Response. We repeated both mechanism diagnostics on Electricity (321 variables), using L=96 and prediction lengths {96, 192, 336, 720} (seed 2024). This configuration conditions on the forecast-origin weekday, extending the hour-conditioned analysis on ETTm1 to a different dataset and calendar cycle.

First, grouping test windows by weekday at each horizon yields 28 weekday–horizon groups. Over non-DC frequencies, we measure normalized magnitude error and target-amplitude-weighted circular phase error. HarMoTC improves both metrics in all 28 groups relative to w/o TC:

MetricHarMoTCw/o TCReductionImproved groups
Magnitude error0.33230.35486.34%28/28
Phase error (rad)0.79150.83044.67%28/28

Values are macro-averaged over the 28 groups.

Second, we fix the best-validation checkpoint, input series, targets, normalization, and sample order, and change only the weekday index. Shifts of ±1, ±2, and ±3 days increase MSE by 28.01%, 60.08%, and 72.08%, respectively; shuffled weekday indices increase it by 45.75%. These relative increases are averaged over the four horizons, with opposite directions and five full-test-set permutations averaged within their respective conditions. Magnitude and phase errors also increase under every intervention at every horizon. Paired circular moving-block bootstrap with 2,000 replicates and block length equal to the prediction length yields 95% confidence intervals above zero for every intervention's MSE increase at every horizon.

These diagnostics establish spectral improvements and dependence on correct calendar alignment beyond aggregate MSE/MAE ablations. Their consistent results on Electricity extend the evidence beyond ETTm1 to a higher-dimensional dataset with a different calendar condition.

Q3. Inference latency and peak GPU memory on a dataset with more variables

Response. We additionally evaluated HarMoTC, DLinear, and PatchTST on Electricity under the 96→96 setting. All measurements use the same NVIDIA GeForce RTX 3050 Ti Laptop GPU, FP32 precision, and matched batch sizes. Each model/batch configuration is measured in three independent processes, each with 50 warm-up iterations and 200 timed forward passes. Latency includes the forward call and GPU synchronization, with inputs already on the GPU. Peak memory is total PyTorch-allocated GPU memory during inference, including parameters, resident inputs, and temporary tensors. The test MSE values below are from the seed-2024 runs.

ModelTest MSEB=1 latency (ms)B=1 peak memory (MiB)B=16 latency (ms/batch)B=16 peak memory (MiB)
HarMoTC0.1402.362 ± 0.06846.2075.904 ± 0.163107.925
DLinear0.1970.395 ± 0.00932.8960.575 ± 0.05245.720
PatchTST0.1815.191 ± 0.01261.05465.179 ± 0.149447.939

Latency is mean ± sample standard deviation across the three measurement processes. Resource reductions use unrounded measurements; MSE reductions are calculated from the values shown in the table and truncated to two decimal places.

Despite having more parameters than PatchTST in this configuration (2,395,224 versus 548,704), HarMoTC reduces mean latency and peak allocated memory by 54.50% and 24.31% at B=1, and 90.94% and 75.90% at B=16. Its test MSE is also 22.65% lower. DLinear requires fewer resources, while HarMoTC achieves 28.93% lower MSE in this experiment. These measurements quantify the accuracy–resource trade-off on a dataset with substantially more variables and show that the larger parameter count does not translate into higher inference latency or memory than PatchTST under the reported configurations.

Thank you for the thorough response; the additional complex-affine comparison, cross-dataset diagnostics on Electricity, and inference latency/memory benchmarks directly address all of my previous concerns. These new results convincingly support HarMoTC's design choices and practical efficiency, so please incorporate them into the revised manuscript as I maintain my positive recommendation for acceptance.

Dear Reviewer UGkz,

Thank you for your constructive feedback and continued positive recommendation. We are pleased that the additional experiments and analyses have addressed your concerns.

Some results are already included in the current revised manuscript. We will submit a final revised manuscript incorporating the remaining results and clarifications.

Thank you again for your time.

Reviewer_53JyOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose HarMoTC, a framework that performs temporally conditioned phase–amplitude spectral modeling for multivariate time-series forecasting. Specifically, they introduce calendar-conditioned spectral modulation as well as phase–amplitude decoupled spectral modeling.

The paper targets an important problem, and the motivation is clearly explained. The related work section is generally well explained. The authors conduct experiments on seven real time-series datasets and describe the effectiveness. However, the technical contribution of this paper is somewhat unclear, since many existing methods also seem to focus on calendar information and phase–amplitude modeling. Also, in the related work section, the authors discuss many existing approaches; I would like to know why the authors selected these 8 competitors for Table 1.

Critical strengths

(S) The paper targets an important problem, and the motivation is clearly explained

(S) The authors conduct experiments on seven real time-series datasets

Critical weaknesses

(W) The technical contribution of this paper is somewhat unclear

Questions for authors

(Q) In the related work section, the authors discuss many existing approaches; I would like to know why the authors selected these 8 competitors for Table 1.

5 comments
AuthorsOpenReview ↗

Clarifying HarMoTC's Technical Contribution and Baseline Selection

Response to Reviewer 53Jy (1/2): Technical contribution and experimental evidence

We thank the reviewer for the positive assessment of our motivation and experimental coverage, and for raising questions about technical contribution and baseline selection.

1 Technical contribution

HarMoTC is designed to preserve periodic strength and temporal alignment as spectral patterns change with calendar context. Its main contribution is to make calendar information directly control spectral transformations. Fourier encodings of the hour and/or weekday at the forecast origin generate a gain and bias for each variable–frequency pair (Eqs. (7)–(9)). These parameters modulate the spectrum’s real component before amplitude–phase modeling and again after cross-variable interaction. The same historical spectrum can therefore receive different adjustments at different calendar states.

Within this conditioned representation, two learnable kernels generate frequency-wise amplitude gains and phase displacements (Eqs. (10)–(12)). The spectrum is then reconstructed from the adjusted amplitude and phase. The kernels are shared across variables, while calendar dependence comes from the surrounding modulation.

2 Evidence for the proposed design

The ablations use the same backbone, training objective, and protocol. Removing calendar-conditioned modulation increases macro-average MSE and MAE by 6.55% and 4.00%, respectively, across ETTm1, ETTh2, Weather, and Electricity. Removing either the amplitude or phase branch also raises the error; removing both produces larger errors than either individual removal.

Our ETTm1 diagnostics also show lower amplitude and phase errors with calendar conditioning (1.65% and 5.18%). Changing only the hour indices raises MSE from 0.3597 to 0.4706 under ±3-hour shifts and 0.6141 under shuffled indices.

Transfer to iTransformer on Weather

We further evaluated the transferability of the core design to iTransformer. We added calendar-conditioned spectral modulation and the phase–amplitude transformation as a residual adapter between instance normalization and inverted embedding, using near-identity initialization. The original encoder, prediction head, and timestamp covariates are retained.

Table R1 Transfer of the core design to iTransformer on Weather

HorizoniTransformerWith HarMoTC coreMSE gain (%)MAE gain (%)
960.174 / 0.2140.1591 / 0.20368.5524.848
1920.221 / 0.2540.2079 / 0.24975.9021.656
3360.278 / 0.2960.2669 / 0.29203.9591.320
7200.358 / 0.3470.3469 / 0.34383.0950.909
Avg0.2577 / 0.27770.2452 / 0.27234.8511.948

Entries show MSE / MAE. Gains are approximate; displayed adapter values and averages are truncated.

The adapter improves both metrics at all four horizons, reducing the four-horizon average MSE by approximately 4.851% and MAE by 1.948%. This extends the evidence for the core spectral design to iTransformer on Weather.

Calendar conditions and amplitude–phase errors

At H=96, we analyzed amplitude and phase errors on all 10,444 test windows and tested calendar alignment with all trained weights fixed (Table R2).

Table R2 Spectral and calendar diagnostics on Weather at H=96

Model / calendar conditionMSEMAEAmplitude NRMSEPhase error
iTransformer0.1740.2140.82300.6970
With core, correct calendar0.15910.20360.81440.6680
With core, wrong calendar0.17810.22360.82910.7005

Lower is better. Baseline and correct-calendar MSE/MAE match Table R1 at H=96. Wrong calendar pools errors across all 23 nonzero hour shifts. Amplitude NRMSE uses positive Fourier frequencies; phase error is target-energy-weighted 1−cos(phase difference).

The adapter reduces amplitude NRMSE by 1.044% and weighted circular phase error by 4.152% relative to iTransformer, linking its accuracy gains to better spectral magnitude and phase alignment. Keeping all weights, historical inputs, and backbone time-feature tokens fixed, we changed only the adapter’s hour condition. Pooled MSE increases by 11.943% under wrong hour indices. Correct indices give lower errors in all four metrics for each of the 23 nonzero shifts. These results support improved spectral fidelity and effective calendar use within iTransformer.

Response to Reviewer 53Jy (2/2): Baseline selection and additional comparisons

3 Baseline selection and relation to closely related methods

We chose the eight baselines to cover established forecasting architectures alongside recent methods under the fixed look-back setting. SSformer, iTransformer, and PatchTST represent Transformer-based models. DLinear provides a linear reference, and TimeMixer represents multiscale MLP modeling. Times2D, RAFT, and FITS add two-dimensional temporal modeling, retrieval-augmented prediction, and frequency-domain forecasting, respectively.

We supplement these comparisons with TimeLinear for explicit timestamp modeling and with APDNet and FIRE for spectral modeling. These methods allow a closer comparison with the two components discussed above.

TimeLinear combines TimeSter’s timestamp-based forecast with an observation-based linear forecast through a weighted sum. HarMoTC uses calendar context inside the spectral computation, where it modulates the operator’s input and output.

APDNet adjusts spectral features through mappings of the real and imaginary parts. HarMoTC explicitly parameterizes amplitude gain and phase displacement and applies these responses within the calendar-modulated representation.

FIRE also separates amplitude and phase. It uses frequency patches and adaptive weighting of frequency bases to model concept drift and basis evolution. HarMoTC uses explicit calendar indices to determine how the spectrum is modulated at each forecast origin.

Table R3 uses L=96. Each cell shows MSE / MAE averaged over horizons 96, 192, 336, and 720. Bold and underline mark the best and second-best distinct displayed values for each metric; ties share the same formatting.

Table R3 Additional comparisons with closely related methods

DatasetHarMoTCFIREAPDNetTimeLinear
ETTm10.359 / 0.3800.375‾\underline{0.375} / 0.387‾\underline{0.387}0.380 / 0.3930.385 / 0.395
ETTm20.263 / 0.3080.276 / 0.3210.271‾\underline{0.271} / 0.3170.273 / 0.315‾\underline{0.315}
ETTh10.412 / 0.4250.425‾\underline{0.425} / 0.4250.443 / 0.4320.432 / 0.426‾\underline{0.426}
ETTh20.353 / 0.3860.364 / 0.3920.372 / 0.3980.358‾\underline{0.358} / 0.389‾\underline{0.389}
Weather0.232 / 0.2660.243‾\underline{0.243} / 0.269‾\underline{0.269}0.243‾\underline{0.243} / 0.2720.251 / 0.276
Electricity0.170‾\underline{0.170} / 0.2670.175 / 0.262‾\underline{0.262}0.170‾\underline{0.170} / 0.2660.165 / 0.259
Traffic0.466‾\underline{0.466} / 0.3090.494 / 0.2780.421 / 0.281‾\underline{0.281}0.480 / 0.304

Sources: FIRE and TimeLinear original papers, and APDNet results reported in mLANet.

At the displayed precision, HarMoTC achieves best or tied-best results on 10 of the 14 dataset–metric pairs and second-best or tied-second-best results on 2 more.

We hope these additional experiments and clarifications address your concerns, and we would be happy to discuss any remaining questions.

AuthorsOpenReview ↗

Follow-up: Extremum Magnitude and Timing

Dear Reviewer 53Jy,

We evaluated the existing validation-selected checkpoints on all 10,444 Weather test windows (L = H = 96, seed 2023), without retraining.

MetriciTransformer+ HarMoTC coreReduction (%)95% CI (%)
Extremum magnitude MAE0.29820.29072.519[1.233, 3.706]
Extremum timing MAE (steps)29.667528.19324.969[2.973, 7.033]

Magnitude MAE averages the errors in forecast-horizon maxima and minima using train-standardized values. Timing MAE averages their absolute position errors, excluding target window–variable pairs with a range ≤ 1e-6; 93.49% of pairs are eligible.

The confidence intervals for percentage reductions use 2,000 paired circular block-bootstrap replicates with a block length of 192 windows. Displayed values are truncated; reductions use full precision.

Both errors decrease, with confidence intervals above zero. These results complement our spectral diagnostics by showing more accurate extremal values and temporal positions after transferring the core design to iTransformer.

We hope this revision addresses your concerns. Please let us know if any points require further clarification; we would be happy to discuss any remaining questions.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Manuscript Revisions

Summary of Manuscript Revisions

We thank the reviewers and Area Chair for their feedback. The main revisions are summarized below; all locations refer to the revised PDF.

  1. Contribution clarification. We revise the abstract, introduction, and contribution list to present the overall framework first and its core calendar-modulation mechanism second, with phase-amplitude modeling incorporated into the framework. Related discussions are aligned accordingly.

  2. Modulation comparisons (Section 4.2, Table 3). We add comparisons with symmetric real, full complex affine, and direct calendar amplitude-phase modulation on four datasets at L=T=96L=T=96 over three seeds, reporting forecasting errors and parameter counts.

  3. Transfer experiments and diagnostics (Appendix B, Tables 8-9). We add iTransformer transfer experiments on Weather, together with calendar-shift controls and amplitude/phase error diagnostics.

  4. Presentation and organization. We streamline Section 4.2 and move the former Additional Experimental Details appendix to Appendix C, updating table numbering and cross-references.

The original HarMoTC architecture, training objective, and previously reported benchmark and component-ablation results are retained.

Area_Chair_1mZ1OpenReview ↗

Invitation to Discuss the Authors’ Response

Dear Reviewers,

The authors have now provided their response. I encourage all reviewers to read it and participate in the discussion with the authors to clarify any remaining questions or concerns.

Thank you to both the authors and reviewers for your time and effort throughout the review process.

AuthorsOpenReview ↗

Summary of Further Manuscript Revisions

Summary of Further Manuscript Revisions

We thank the reviewers and Area Chair for their helpful feedback. This revision updates the title and incorporates experimental results already reported in our rebuttal:

  1. Title: Revised to HarMoTC: Harmonic Temporal Conditioning for Spectral Time-Series Forecasting to emphasize calendar-conditioned modulation as the main contribution.
  2. Calendar diagnostics (Appendix C.2, Table 12): Added weekday-stratified spectral errors and calendar-alignment interventions on Electricity.
  3. Inference efficiency (Appendix C.4, Table 14): Added inference latency and GPU memory comparisons with DLinear and PatchTST.

Appendix layout, table numbering, and cross-references have been updated accordingly. The model, training objective, and previously reported main results are unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.