acceptodds
Under review as a conference paper at ICLR 2027

Patterns Persist, Expressions Change: Long-Horizon Time-Series Forecasting via Temporal Pattern Deformation

Abstract

Long-horizon forecasting implicitly relies on temporal patterns observed in history remaining informative beyond the observed window. Under structured non-stationarity, recurring patterns may remain informative even as their future expressions change in amplitude, phase, and cross-channel temporal alignment. We term this variation temporal pattern deformation. Existing methods model related phenomena through statistical shifts, spectral variation, latent dynamics, or future phase evolution. We study how recurring temporal structure can be represented together with time-varying amplitude and phase over the forecast horizon. We introduce NSTime, which applies a recursive moving-average decomposition to form a smoothed background and a hierarchy of detail residuals. Each detail residual is represented by an amplitude–phase component and a residual component. Continuous Circular Temporal Generators (CCTG) map historical and forecast time points onto shared differentiable cyclic coordinates, with learned channel offsets representing persistent channel-level temporal shifts. Along these coordinates, the AM–PM Non-Stationary Forecasting (NSF) module evaluates learned amplitude and phase fields over the historical and forecast intervals, while a residual path captures dynamics beyond the amplitude–phase formulation. Across 11 benchmarks, NSTime ranks first in 16 of 22 dataset–metric comparisons within the evaluated comparison pool and among the top two in 19.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Latent State, Closed Loop

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KcTcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces NSTime for long-horizon time-series forecasting. The method applies a recursive moving-average decomposition to obtain a smoothed background and a hierarchy of detail residuals. The authors evaluate NSTime on 11 benchmark datasets and demonstrate its effectiveness.

Critical strengths

1.NSTime offers a new perspective on addressing non-stationarity through amplitude and phase modeling, an approach overlooked by previous work.

2.NSTime achieves strong or competitive forecasting performance across the 11 benchmark datasets.

Critical weaknesses

1.Although the paper aims to address non-stationarity through amplitude and phase modeling, the theoretical analysis and experimental evidence do not yet sufficiently establish why this formulation is more effective than existing approaches. The connection between the proposed mechanism and the observed improvements could be demonstrated more convincingly.

2.NSTime appears closely related to CycleNet, particularly in its use of cyclic structure and residual forecasting. Although NSTime introduces explicit amplitude–phase modeling and additional architectural components, its distinctive contribution relative to CycleNet could be explained more clearly. Moreover, Table 6 shows comparable performance between the two methods in several settings. While the controlled comparison with CycleNet is helpful, the authors should further clarify the methodological differences and identify the conditions under which NSTime provides meaningful advantages.

Questions for authors

1.Table 1 would be more appropriately placed in Section 4, alongside the experimental results, rather than in Section 3.

2.Table 1 contains two consecutive rows of MSE/MAE labels. Please remove the duplicated header.

3.The input length is fixed at S=96S=96 in the main comparisons. However, some methods, such as DLinear, can benefit from longer input sequences. Although Appendix G includes an input-length study on Electricity, extending this study to other datasets and additional baselines, such as DLinear and PatchTST, would help establish whether NSTime’s advantages persist with longer inputs.

4.The experiments do not include the commonly used Traffic dataset. I suggest adding comparative results on this dataset to complement the existing PEMS experiments and facilitate comparison with previous studies.

5.The paper proposes NSTime to address non-stationarity. Could the authors clarify how non-stationarity is defined in this work and how NSTime’s ability to address it is evaluated? What evidence supports this claim beyond improvements in forecasting accuracy?

14 comments
AuthorsOpenReview ↗

Response to Reviewer KcTc

W1.

Re. Thanks. We focus on structured second-order non-stationarity within recurring components. Representative existing methods address non-stationarity through statistical calibration or latent dynamics. For example, Non-stationary Transformers handle statistical changes through series stationarization and de-stationary attention, while Koopa distinguishes time-variant from time-invariant dynamics. From their formulations, these methods do not jointly express amplitude variation, temporal alignment variation, and their second-order statistical effects through an explicit waveform generation structure.

Equation (3) in the revised Sec. 3.1 characterizes this structure. Consider the harmonic process

\qquad \Theta\sim\mathrm{Unif}[0,2\pi).$$ Here, $A(t)\geq 0$ and $\phi(t)$ are deterministic functions, and a single random phase $\Theta$ is shared across time. Its second-order statistics are $$\operatorname{Var}[h(t)]=\frac{1}{2}A(t)^2,$$ $$\operatorname{Cov}[h(t),h(t+\tau)] \\ = \\ \frac{1}{2}A(t)A(t+\tau) \\ \cos\left(\omega\tau+\phi(t+\tau)-\phi(t)\right).$$ Even when the carrier frequency is fixed, amplitude variation can make the variance depend on temporal position, while changes in amplitude products or phase increments can make the covariance at a fixed lag depend on temporal position. In particular, even with constant amplitude, mean, and variance, changing phase increments can still make the covariance time-dependent. Thus, this form of non-stationarity has an identifiable amplitude–phase origin, and calibrating the mean and variance alone is insufficient to fully characterize its variation. **NSTime explicitly represents this structure and its generation mechanism.** When a recurring waveform is learned directly, changes in magnitude and temporal position are jointly encoded in waveform values, leaving their separation and interaction implicit. NSTime separately parameterizes amplitude and phase fields for each detail component on the continuous cyclic coordinates provided by CCTG, then generates the waveform through the AM–PM synthesis in Eq. (16). This gives oscillation magnitude and temporal alignment distinct modeling variables and explicitly specifies how they jointly change the waveform. The cyclic fields represent amplitude–phase structure that recurs by cycle position, while the residual predictor uses historical observations to compensate for window-specific deviations. This modeling structure is directly related to forecasting error. Equation (4) shows that, for predicted and target harmonics sharing the same carrier and the same random phase, phase mismatch contributes $$A\widehat A\left[1-\cos(\Delta\phi)\right]$$ to the expected squared error. Thus, with other factors held fixed, the error caused by a given phase mismatch increases with the product of the predicted and target amplitudes. AM–PM synthesis incorporates this interaction between magnitude and temporal alignment directly into waveform generation, providing a targeted representation of the structured non-stationary component described above. The matched synthesis comparison below evaluates the forecasting benefit of this modeling choice. **The NSF module ablation directly tests the forecasting contribution of this structure.** The *w/o NSF* variant retains the residual forecasting path while removing NSF’s structured AM–PM path. Table 10 in Appendix D.1 (p. 23) shows that removing NSF increases MSE in all 12 dataset–horizon settings across ETTh1, ETTm1, and Weather, with a mean relative increase of **6.50%** over Full. Fixing the amplitude to one (*w/o Amp*) or setting phase modulation to zero (*w/o Phase*) increases MSE by **2.77%** and **2.67%** on average, respectively. These ablations show that NSF’s explicit amplitude–phase structure provides additional forecasting gains beyond the residual forecasting path. **A matched comparison further examines the contribution of the amplitude–phase synthesis rule.** Full NSTime and **Direct Template** use the same two CCTGs, the same number of trainable parameters, the same cyclic lookup operations, and the same initial waveform values at the sampled cycle positions. Direct Template adds the two generator outputs directly. Full instead applies Softplus to one output to determine the amplitude, uses the other to modulate the carrier phase, and generates the waveform through AM–PM synthesis.
AuthorsOpenReview ↗

Response to Reviewer KcTc

All runs use a 96-step input and a 720-step forecast horizon. The table reports the mean ± sample standard deviation across three runs per configuration.

DatasetDirect TemplateNSTimeMeanDirect TemplateNSTimeMean
MSEMSEReductionMAEMAEReduction
ETTm20.4007 ± 0.00850.3860 ± 0.00263.7%0.3967 ± 0.00650.3910 ± 0.00101.44%
Weather0.3457 ± 0.00210.3383 ± 0.00422.1%0.3460 ± 0.00260.3423 ± 0.00151.07%

Full achieves lower MSE in all six paired runs. These results show that explicit amplitude–phase synthesis improves forecasting in the tested settings, and that the gains cannot be explained solely by increasing the number of generators or trainable parameters.

Forecast diagnostics further connect the amplitude–phase structure to the observed improvements. In the harmonic analysis in Appendix E (pp. 23–25, Fig. 6 on p. 25), we fit harmonics to the final forecast and the observed future at the same frequency selected from the input history and on the same time grid. This diagnostic uses Direct Waveform, a different control from Direct Template above, and reports results for random seed 2024. On ETTm2, Full reduces amplitude MAE from 0.0874 to 0.0774 and circular phase MAE from 28.4° to 22.0°.

The drift-stratified experiments further examine the forecasting contribution of each modulation field (Sec. 4.2, Table 3 and Figs. 3–4). For 720-step forecasting on ETTm2, Full reduces MSE by 5.50% relative to w/o Amp in the highest amplitude-drift group and by 4.36% relative to w/o Phase in the highest phase-drift group. The corresponding reductions on Electricity are 0.21% and 0.38%, indicating that the magnitude of the benefit varies across datasets.

Together, the module ablations, matched synthesis comparison, harmonic diagnostics, and drift-stratified experiments support the forecasting value of explicitly modeling amplitude–phase-induced, structured second-order variation within recurring components and specifying how amplitude and phase jointly generate the waveform.

AuthorsOpenReview ↗

Response to Reviewer KcTc

W2.

Thank you for pointing out that we should explain the distinction from CycleNet more clearly. Circular linear interpolation makes CCTG's lookup piecewise differentiable with respect to real-valued channel offsets. This allows the model to learn fractional shifts relative to the cyclic template through backpropagation. For example, at inference, if the current cycle index is 12 and the learned offset is +0.35+0.35 sampling steps, CCTG queries 12+0.35=12.3512+0.35=12.35 and interpolates between entries 12 and 13 with weights of 0.65 and 0.35, respectively. RCF reads entry 12 directly. This sub-sample alignment helps track local fluctuations in the recurring waveform, particularly where it rises or falls sharply. Figure 9 visualizes the structured components produced by NSF and illustrates changes in amplitude envelope and timing within recurring patterns.

In the controlled comparison within CycleNet (Table 4), replacing integer RCF lookup with CCTG lowers MSE in 14 of 16 settings and ties in the other two. This comparison tests the combined effect of circular interpolation and learned channel offsets.

We also compare the two lookup rules within NSTime on Weather, Electricity, and Traffic. In the RCF variant below, integer lookup generates the amplitude and phase fields while the decomposition, AM–PM synthesis, and residual predictors are retained. The Full NSTime column is the matched reference for each row. Replacing CCTG with RCF raises MSE in all 12 settings.

DatasetHorizonRCFCCTGMSE reductionRCFCCTGMAE reduction
MSEMSEMAEMAE
Weather960.1590.1552.52%0.2040.2011.47%
Weather1920.2080.2032.40%0.2490.2480.402%
Weather3360.2640.2591.89%0.2900.2871.03%
Weather7200.3450.3430.580%0.3430.344-0.292%
WeatherAvg0.2440.2401.64%0.2720.2700.552%
Electricity960.1370.1351.46%0.2330.2301.29%
Electricity1920.1560.1503.85%0.2490.2442.01%
Electricity3360.1730.1682.89%0.2660.2660.00%
Electricity7200.2110.2081.42%0.3040.3030.329%
ElectricityAvg0.1690.1652.36%0.2630.2610.856%
Traffic960.3910.3851.53%0.2740.2711.09%
Traffic1920.4040.3971.73%0.2750.2730.727%
Traffic3360.4140.4071.69%0.2810.2800.356%
Traffic7200.4520.4441.77%0.3000.2980.667%
TrafficAvg0.4150.4081.69%0.2830.2810.708%

These results demonstrate the effectiveness of CCTG. A plausible explanation for the comparable performance in several settings of Table 6 is that integer cyclic templates already capture the recurring waveform well, leaving limited additional benefit from finer alignment and explicit amplitude–phase modulation. CCTG is expected to provide greater benefits when predictable amplitude variations and temporal alignment differences are inadequately represented by the existing cyclic template.

Q1.

Thank you for pointing this out. LaTeX’s float placement put Table 1 in Section 3 of the original submission. We have moved it to Section 4 of the revised manuscript, alongside the experimental results.

Q2.

Thank you for catching this. We have removed the duplicate MSE/MAE header row from Table 1 in the revised manuscript.

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

In this paper, NSTime is proposed for long-term time series forecasting via temporal pattern deformation. Recursive moving-average decomposition separates trend and detail components, while continuous cyclic temporal generators model amplitude and phase variations along shared periodic coordinates. Residual forecasters capture remaining dynamics. Experiments across eleven datasets demonstrate competitive forecasting performance, with ablation studies supporting the proposed components.

Critical strengths
  1. The paper is well-written and generally easy to follow.

  2. NSTime achieves the best average cyclic Mean Squared Error (MSE) on ten out of the eleven datasets evaluated.

  3. Controlled ablation experiments and comparisons against CycleNet validate the benefits of the proposed components.

Critical weaknesses
  1. It is unclear how the fixed cyclic fields adapt to changes in amplitude and phase that do not follow the learned periodic patterns.

  2. Details regarding cycle lengths, higher-order carrier harmonics, and training hyperparameters are not sufficiently described to allow for the reproduction of the results.

  3. The primary comparison relies on a single random seed and omits models such as Chronos [1] and SCALER [2]; furthermore, TimeLLM [3] is evaluated only within a few-shot learning setting.

[1] Ansari, A. F., et al. Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research, 2024.

[2] Le, X.-M., et al. Efficient Test-Time Scaling for LLM-based Time Series Forecasting (SCALER). KDD, 2026.

[3] Jin, M., et al. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. ICLR, 2024.

Questions for authors

See the weakness section!

4 comments
AuthorsOpenReview ↗

Response to Reviewer ccSE

W1.

When amplitude and phase variations do not follow the learned periodic patterns, such changes are difficult to characterize directly using the fixed cyclic prior. We therefore model them using the nonlinear representational capacity of the residual predictor. Specifically, it takes the difference between the historical detail component and the gated cyclic prior as input and maps these historical changes to future corrections. The resulting corrections are added to the cyclic forecast to compensate for temporal changes that do not follow the learned periodic patterns.

W2.

Re. Thanks. we choose a calendar-based reference period TrefT_{\mathrm{ref}} for each dataset based on its expected daily or weekly recurrence. We then divide that period by the sampling interval Δt\Delta t to obtain the cycle length L=Tref/ΔtL=T_{\mathrm{ref}}/\Delta t in sampling steps.

DatasetReference periodSampling intervalCycle length LL
ETTh1, ETTh224 hours1 hour24/1=2424/1=24
ETTm1, ETTm224 hours15 minutes1,440/15=961{,}440/15=96
Weather, Solar24 hours10 minutes1,440/10=1441{,}440/10=144
Electricity7 days1 hour(7×24)/1=168(7\times24)/1=168
PEMS03, PEMS04, PEMS07, PEMS0824 hours5 minutes1,440/5=2881{,}440/5=288

The daily reference aligns observations taken at the same time of day across days. For Electricity, we use a weekly reference so that the same hour on a weekday and a weekend occupies different positions in the cycle.

Carrier harmonics (mjm_j). We start from the fundamental of the reference cycle and double the carrier frequency at each successive detail branch. This gives the harmonic rule mj=2j−1m_j=2^{j-1} and equally spaced carrier frequencies on a logarithmic scale. For dataset dd, branch jj therefore has angular frequency ωj,d=2πmj/Ld\omega_{j,d}=2\pi m_j/L_d and carrier period Ld/mjL_d/m_j. Over one reference cycle, the carrier phase advances by 2πmj2\pi m_j, returning to the same phase. The amplitude and phase fields also repeat every LdL_d steps, so the modulated components share LdL_d as a common period.

We have updated the main text of the revised manuscript to explain this selection rule and report the cycle length and carrier harmonics used for each dataset. Appendix C.4 provides the full parameter configuration.

Table 7 give the learning-rate, batch-size, and dropout configurations, including the detail-branch dropout of 0.1. Table 14 reports the residual-MLP widths by dataset and horizon. We will also make the NSTime implementation publicly available to facilitate reproduction of the reported results.

AuthorsOpenReview ↗

Response to Reviewer ccSE

W3.

Re. Thanks.

  1. In the main comparative experiments, we used the same random seed for all methods we reproduced. Table 16 reports NSTime’s results for three random seeds (2024, 2025, and 2026), along with the corresponding means and sample standard deviations. The results show relatively small variations in forecasting errors across seeds in most evaluated settings, providing additional evidence of NSTime’s stability across runs in those settings.

  2. In the original few-shot experiments using only 5% of the training data, we included LLM-based methods such as Time-LLM to compare NSTime’s forecasting performance with that of methods that draw on pretrained knowledge for generalization under limited-data conditions, and to examine whether NSTime can effectively capture temporal patterns with a small number of parameters.

  3. The original papers do not specify complete configurations for the 96-step input setting used in our study. We provide a broader comparison with SCALER and Chronos using the input lengths reported in their respective papers. We report comparative results under the 96-step input setting, where NSTime demonstrates an overall performance advantage.

DatasetHorizon HTime-LLMTime-LLMSCALERSCALERChronosChronosNSTimeNSTime
MSEMAEMSEMAEMSEMAEMSEMAE
Weather960.1980.2350.1660.2260.1860.2080.1560.202
Weather1920.2400.2690.2120.2760.2380.2580.2040.246
Weather3360.2950.3080.2630.3090.3130.3530.2610.289
Weather7200.3680.3530.3410.3760.4160.4150.3430.344
WeatherAvg0.2750.2910.2460.2970.2880.3090.2410.270
Electricity960.1720.2650.1370.2410.1410.2430.1350.230
Electricity1920.1820.2790.1580.2570.1740.2600.1520.245
Electricity3360.1950.2880.1770.2720.1810.2740.1690.265
Electricity7200.2330.3200.2210.3210.2340.3270.2100.301
ElectricityAvg0.1960.2880.1730.2730.1830.2760.1670.260

The reported results for Time-LLM and Chronos use an input length of 512. We therefore evaluate NSTime under the same input-length setting for comparison.

DatasetHorizon HTimeLLMTimeLLMSCALERSCALERChronos(46M)Chronos(46M)Chronos(200M)Chronos(200M)Chronos(710M)Chronos(710M)NSTimeNSTime
MSEMAEMSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
Weather960.1470.2010.1490.1970.2120.2380.1980.2310.2070.2320.1500.205
Weather1920.1890.2340.1920.2440.2660.2830.2480.2760.2600.2770.1860.240
Weather3360.2620.2790.2410.2820.3210.3210.3010.3140.3100.3130.2390.280
Weather7200.3040.3160.3140.3250.3930.3670.3860.3610.3900.3650.3180.334
WeatherAvg0.2260.2580.2240.2620.2980.3020.2830.2960.2920.2970.2230.265
Electricity960.1310.2240.1300.2310.1990.2670.1960.2730.1980.2780.1260.221
Electricity1920.1520.2410.1510.2380.2180.2890.2070.2850.2140.2940.1450.238
Electricity3360.1600.2480.1670.2570.2440.3210.2380.3140.2390.3140.1640.259
Electricity7200.1920.2980.2010.2900.3240.3710.3100.3550.3160.3620.1970.289
ElectricityAvg0.1590.2530.1620.2540.2460.3120.2380.3070.2420.3120.1580.252
AuthorsOpenReview ↗

Response to Reviewer ccSE

The SCALER comparison uses a 672-step input window, providing seven times as many historical observations as our main experiments and covering more recurring cycles and longer-term variations. Accordingly, we compare NSTime with SCALER, Chronos, and Time-LLM using a 672-step input window.

DatasetHorizon HSCALERSCALERChronosChronosTime-LLMTime-LLMNSTimeNSTime
MSEMAEMSEMAEMSEMAEMSEMAE
Weather960.1410.1930.1450.1980.1490.2000.1510.206
Weather1920.1860.2380.1880.2420.1950.2430.1850.237
Weather3360.2340.2770.2380.2810.2450.2820.2360.278
Weather7200.3060.3200.3090.3310.3180.3380.3150.337
WeatherAvg0.2170.2570.2200.2630.2270.2660.2220.265
Electricity960.1200.2220.1310.2330.1370.2440.1270.221
Electricity1920.1430.2310.1540.2440.1620.2710.1430.237
Electricity3360.1580.2460.1670.2620.1750.2790.1620.259
Electricity7200.1940.2810.2040.2930.2070.3060.1950.288
ElectricityAvg0.1540.2450.1640.2580.1700.2750.1570.251

NSTime has a higher average MSE than SCALER on both Electricity and Weather. Taken together, the results across different input lengths indicate an overall forecasting advantage for NSTime.

Current LLM-based methods effectively exploit transferable knowledge acquired through pretraining, while NSTime uses fewer parameters.

ModelSCALERChronos familyTime-LLMNSTime
Parameters (M)1380.699 to 7101380.560.181

We hope the reviewer will also consider differences in model scale and pretraining when interpreting these results. For LLM-based time series forecasting methods, comparisons with other LLM-based approaches of comparable scale and similar pretraining provide a more informative basis for assessment.

AuthorsOpenReview ↗

Response to Reviewer ccSE

Dear Reviewer ccSE,

Thank you again for your time and constructive feedback. As the discussion period is drawing to a close, we would greatly appreciate your feedback on our responses and the revised manuscript.

Could you please let us know whether our responses have addressed your main concerns, or whether any points still require clarification? We would be happy to provide further explanations before the discussion closes.

Thank you for your time and consideration.

Best regards, The Authors

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

This paper studies long-horizon time series forecasting when recurring patterns remain informative but change in amplitude, phase or cross-channel temporal alignment. The author stated that forecasting should account for how persistent patterns are re-expressed beyond the observed window. Based on this insight, the paper proposed NSTime, which decomposes the input into a smoothed background and hierarchical detail residuals, then models each detail through an amplitude–phase component and a residual predictor. Continuous Circular Temporal Generators provide shared cyclic coordinates for evaluating historical and future amplitude–phase fields. Experiments on diverse benchmarks demonstrate competitive performance of proposed method.

Critical strengths
  1. The distinction between recurring temporal structure and its changing expression provides an intuitive perspective on forecasting. Also, the modeling changes in amplitude and timing is relevant to applications such as traffic flow and energy consumption, thus the research question is quite important.
  2. The paper includes substantial component analysis. Matched ablations, drift-based stratification, and the controlled CCTG replacement in CycleNet provide useful evidence about the contributions of the proposed components.
Critical weaknesses
  1. The background could explain more clearly how temporal pattern deformation differs from the changing periodic patterns and non-stationarity addressed by previous methods. The examples are intuitive, but the specific limitation of existing approaches that motivates a new framework is less clear.
  2. Important configuration choices need clearer specification, particularly the cycle length (L), carrier harmonics (m_j), smoothing window, and low-rank settings. The selection on these parameters is central to reproducing the method and understanding which recurring structures it assumes.
Questions for authors
  1. How are the recurring cycle lengths selected for each dataset, and what information is used in that selection? I noticed that the benchmarks have different sampling rates and recurring patterns, so this choice seems important for understanding the method’s applicability and reproducing the results.
  2. I’m curious about why does NSTime model amplitude and phase separately rather than directly learn a recurring waveform? The ablations show that both components contribute, but it remains unclear for me what advantage this separation provides for modeling temporal pattern deformation compared with a similarly expressive cyclic template.
  3. Can the author explain why the input length is fixed to 96 in the main experiments? Also, the input length analyses covers only Electricity, is there evidence that the advantage of NSTime remain consistent with different amounts of historical information on other datasets?
5 comments
AuthorsOpenReview ↗

Response to Reviewer SZJt

W1.

Re. Thank you for pointing out this issue. We have revised the introduction (Sec. 1) to clarify the relationships between these concepts and the modeling motivation behind NSTime, and moved the main second-order analysis from the original Appendix B to Sec. 3.1 of the main text.

Previous work addresses non-stationarity through statistical shifts, latent dynamics, spectral changes, or dependency structure. Within this broader context, we use temporal pattern deformation to describe variations in the magnitude and temporal alignment of a recurring component relative to a cyclic reference. It captures a specific aspect of changing periodic patterns and identifies a mechanism that can give rise to statistical non-stationarity. Our motivation is to explicitly represent this waveform-level structure in the model to better capture difficult-to-model non-stationary components and improve forecasting performance.

The specific modeling issue concerns how variations in magnitude and timing are represented. When a recurring waveform is learned directly, both types of variation are jointly encoded in waveform values. A sufficiently expressive template can represent these changes, but their separation and interaction remain implicit. NSTime separately parameterizes amplitude and phase fields on the continuous cyclic coordinates provided by CCTG, allowing it to characterize amplitude variations and temporal shifts in recurring patterns within a common cyclic reference. It then combines these fields through AM–PM synthesis in NSF, introducing an explicit modeling structure.

The second-order analysis in Sec. 3.1 further clarifies the connection between our approach and statistical non-stationarity. Even with a fixed carrier frequency, amplitude variation can make the variance depend on temporal position, while variations in amplitude products or phase increments can make the covariance at a fixed lag depend on temporal position. This analysis explains the statistical implications of the pattern deformation targeted by NSTime and provides a rationale for explicitly modeling its amplitude and phase structure.

AuthorsOpenReview ↗

Response to Reviewer SZJt

W2.

Re. Thanks Appendix C.4 now specifies the cycle lengths, carrier harmonics, smoothing settings, and factorization widths. Table 7 in the same appendix summarizes the training configurations.

Cycle length (LL). We specify LL at the dataset level before fitting. We choose a reference period TrefT_{\mathrm{ref}} based on the expected physical or operational recurrence of each dataset and convert it to sampling steps as L=Tref/ΔtL=T_{\mathrm{ref}}/\Delta t, where Δt\Delta t is the sampling interval.

DatasetSampling intervalReference periodLL
ETTh1, ETTh21 hour1 day24
ETTm1, ETTm215 minutes1 day96
Weather10 minutes1 day144
Solar10 minutes1 day144
PEMS03, PEMS04, PEMS07, PEMS085 minutes1 day288
Electricity1 hour1 week168

The daily reference aligns intraday patterns such as variations in transformer load and oil temperature, diurnal weather changes, daylight-driven solar generation, and recurring traffic peaks. For Electricity, the weekly reference lets the same hour on weekdays and weekends occupy different cycle positions. The background, amplitude, and phase tables share the configured LL within each dataset.

Carrier harmonics (mjm_j). We start from the fundamental of the reference cycle and double the carrier frequency at each successive detail branch. This gives the harmonic rule mj=2j−1m_j=2^{j-1} and equally spaced carrier frequencies on a logarithmic scale. For dataset dd, branch jj therefore has angular frequency ωj,d=2πmj/Ld\omega_{j,d}=2\pi m_j/L_d and carrier period Ld/mjL_d/m_j. Over one reference cycle, the carrier phase advances by 2πmj2\pi m_j, returning to the same phase. The amplitude and phase fields also repeat every LdL_d steps, so the modulated components share LdL_d as a common period.

Smoothing. The reference configuration applies a width-five circular moving average twice to the background cycle table. The two passes produce a center-weighted triangular kernel and smooth across the cycle boundary. Circular smoothing is disabled for the amplitude and phase tables to preserve finer within-cycle variation. In Table 10, removing circular smoothing raises MSE in 11 of 12 matched settings, with a mean relative increase of 0.97%.

Factorization width. Each cycle table is independently parameterized as M=UVM=UV, where U∈RL×rU\in\mathbb{R}^{L\times r} contains temporal basis profiles and V∈Rr×CV\in\mathbb{R}^{r\times C} contains channel-specific coefficients. Appendix C.4 specifies the reference width r=64r=64 for the background, amplitude, and phase tables. The effective rank is at most min⁡(L,C,r)\min(L,C,r), so the factorization imposes a strict rank constraint when r<min⁡(L,C)r < \min(L,C). In Table 10, directly parameterizing the cycle tables increases MSE by 2.84% on average across the 12 matched settings.

Q1.

Re. Thanks. We select a reference period TrefT_{\mathrm{ref}} for each dataset based on domain knowledge of its recurring physical or operational patterns. We then use the sampling interval Δt\Delta t to convert the reference period into a number of sampling steps, L=Tref/ΔtL=T_{\mathrm{ref}}/\Delta t.

DatasetSampling intervalReference periodCycle length LL
ETTh1, ETTh21 hour1 day24
ETTm1, ETTm215 minutes1 day96
Weather, Solar10 minutes1 day144
PEMS03, PEMS04, PEMS07, PEMS085 minutes1 day288
Electricity1 hour1 week168

The daily reference aligns the same time of day across days to capture recurring patterns in transformer operation, weather, solar generation, and traffic. For Electricity, the weekly reference lets the same hour on weekdays and weekends occupy different positions in the cycle. Thus, although ETTh and Electricity are both sampled hourly, their different reference periods yield L=24L=24 and L=168L=168, respectively. The selected LL defines the recurrence assumed by the cyclic component. The residual branch models variation that departs from this recurrence.

AuthorsOpenReview ↗

Response to Reviewer SZJt

Q2.

Re. Thanks A sufficiently expressive cyclic template can represent the same waveform at the sampled cycle positions. The advantage of separating amplitude and phase is that NSTime gives changes in magnitude and timing distinct variables and specifies how they combine. A direct template encodes both changes in its waveform values. In NSTime, the amplitude field sets the local oscillation magnitude, and the phase field shifts the oscillation relative to the carrier. As illustrated in Fig. 1 and implemented in Sec. 3.4 (p. 6, Eq. (16)), the resulting component is

hj(t)=Aj(t)cos⁡(ωjt+ϕj(t)).h_j(t)=A_j(t)\cos\bigl(\omega_j t+\phi_j(t)\bigr).

This synthesis captures an interaction that matters for forecasting: at a given carrier position, the same phase shift produces a larger change in the waveform value when the local amplitude is higher. We tested whether encoding this interaction explicitly improves forecasting using two controls within NSTime.

Direct Template (matched synthesis control). This comparison most directly addresses the question about a similarly expressive template. Both variants use the same two CCTGs, the same number of trainable parameters, the same cyclic lookup, and the same initial waveform values at the sampled cycle positions. Let t=s+τt=s+\tau be the unshifted timestamp, and let Gj,1(t)G_{j,1}(t) and Gj,2(t)G_{j,2}(t) be the generator outputs after their respective channel offsets are applied. Direct Template forms the waveform as Gj,1(t)+Gj,2(t)G_{j,1}(t)+G_{j,2}(t). Full NSTime instead uses the first output to set the amplitude and the second to shift the carrier phase, yielding softplus⁡(Gj,1(t))cos⁡(ωjt+Gj,2(t))\operatorname{softplus}(G_{j,1}(t))\cos(\omega_j t+G_{j,2}(t)).

NSTime + RCF (waveform control). This variant replaces the AM–PM generator in each detail branch with a learned waveform table accessed by integer cyclic indexing. It provides a complementary comparison with direct waveform lookup. Both the lookup and the waveform construction change in this control.

All runs use a 96-step input and a 720-step forecast horizon. Values are the mean ± sample standard deviation across three runs per configuration.

DatasetVariantMSEMAE
ETTm2Direct Template0.4007 ± 0.00850.3967 ± 0.0065
ETTm2NSTime + RCF0.3897 ± 0.00290.3927 ± 0.0006
ETTm2Full NSTime0.3860 ± 0.00260.3910 ± 0.0010
WeatherDirect Template0.3457 ± 0.00210.3460 ± 0.0026
WeatherNSTime + RCF0.3423 ± 0.00380.3450 ± 0.0020
WeatherFull NSTime0.3383 ± 0.00420.3423 ± 0.0015

With the generators and parameter count matched, Full NSTime reduced mean MSE relative to Direct Template by 3.7% on ETTm2 and 2.1% on Weather, with lower MSE in all six paired runs. Relative to the RCF waveform control, the corresponding reductions were approximately 0.9% and 1.2%.

The separate harmonic-fitting analysis in Fig. 6 (p. 24) shows what improved in the forecasts themselves. On ETTm2, Full reduced amplitude MAE from 0.0874 to 0.0774 and circular phase MAE from 28.4° to 22.0° relative to Direct Waveform. Thus, in the matched comparison, explicit AM–PM synthesis improved forecasting accuracy at the same parameter count. In the harmonic diagnostic, its forecasts also more closely reproduced both the magnitude and the timing of the observed recurring component.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Major Manuscript Revisions

We have highlighted the major revisions in blue.

1. Problem formulation: Clarified temporal pattern deformation and its connection to second-order non-stationarity and forecasting error (Secs. 1–2 and 3.1).

2. Mechanism validation: Added component ablations, forecast amplitude–phase error evaluations, and drift-stratified comparisons (Sec. 4.2; Appendices D–E).

3. Statistical diagnostics: Clarified the relationship between temporal pattern deformation and ADF/KPSS diagnostics of broader statistical properties (Sec. 4.3; Appendix F.2).

4. Experimental coverage: Added Traffic comparisons (Sec. 4, Table 1; Appendix C.5, Table 8) and three-seed stability results (Appendix H.2, Table 16).

5. Reproducibility and presentation: Detailed parameter configurations (Appendix C.4) and corrected Table 1’s placement and duplicated header (Sec. 4).

Area_Chair_isKDOpenReview ↗

Reminder: Reviewer–Author Discussion

Thank you to the reviewer who have already engaged in the discussion and provided valuable feedback. As the discussion deadline is approaching, I would kindly encourage the remaining reviewers to review the authors' responses and share their follow-up comments, particularly on whether their concerns have been adequately addressed.

Authors are also encouraged to respond promptly to any further questions.

Thank you all for your time and contributions to a constructive discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.