acceptodds
Under review as a conference paper at ICLR 2027

FAPA: A Dual-Branch Framework for Time Series Forecasting with Frequency Amplitude Shaping and Harmonic Period Alignment

Abstract

Time series forecasting supports applications in energy, transportation, and meteorology. Spectral analysis reveals that time series often consist of high-amplitude periodic components and low-amplitude irregular residuals. Fitting structured periodic patterns and irregular residuals jointly with one set of shared weights is harder than modeling them separately, since the two components have very different statistical properties. Predicting the periodic component in a separate branch removes this coupling but raises two problems. First, per-window normalization, the standard remedy for distribution shift, maps windows that differ only in scale to the same input, so the amplitude of the periodic component is inherited from the input window instead of being predicted. Second, the branch has no directly available training target: re-estimating its periods on the ground truth gives periods that are unavailable at inference, while the input-window periods can fall between the frequency bins of the target window, where nearest-bin reconstruction can lose more than half of the component's energy. We propose FAPA, a plug-and-play dual-branch framework that improves existing forecasters. Its periodic branch predicts the periodic component at its original scale and refines the predicted amplitudes with a lightweight frequency Amplitude Shaping Module (AmpShape). This branch is supervised by a dedicated periodic loss whose target is built by Harmonic Period Alignment (HPA), which projects the ground truth onto sinusoids at the input-window periods and is exact for any horizon at no inference cost. A residual branch passes the residual to the backbone. Experiments on eight backbones (e.g., PatchTST, TimesNet) and five datasets show that FAPA lowers the MSE on the five-dataset average for every backbone: it falls by 10.7% overall, and by 19.6% and 18.9% on ECL and Traffic, respectively. A controlled synthetic study further shows that the gain grows monotonically with the amplitude imbalance between the periodic component and the residual.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Time Series Decomposition, Frequency-domain Analysis, Harmonic Projection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_aRj4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FAPA, a dual-branch framework for time-series forecasting that separates periodic components from irregular residuals. The periodic branch predicts the periodic component without window normalization and introduces an amplitude shaping module (AmpShape). Harmonic Period Alignment (HPA) is further used to construct training targets at the input-window frequencies. Experiments across eight forecasting backbones and five datasets report an overall MSE reduction of 10.7%. The results are encouraging, although the methodological novelty and generality of the improvements need further justification.

Critical strengths
  1. The paper addresses a meaningful problem in time-series forecasting. Separating periodic components from irregular residuals is reasonable, especially when their amplitude scales differ significantly.

  2. I find the HPA design interesting. Using input-window periods to construct the future periodic target avoids the frequency-grid mismatch caused by different input and prediction lengths.

  3. The empirical results are fairly convincing. FAPA improves the five-dataset average across all eight backbones, and the synthetic experiments provide useful evidence about the role of amplitude imbalance.

Critical weaknesses
  1. Novelty is somewhat limited. Frequency decomposition and separate periodic prediction have already been explored, particularly in FAN. HPA mainly builds on harmonic projection, and I am not fully convinced that the overall framework represents a substantial methodological advance.

  2. The improvements vary considerably across datasets. Most gains come from ECL and Traffic, while ETTm2 shows almost no improvement.

  3. The ablation study is somewhat narrow. It only uses PatchTST on ETTm1 and ETTm2. I would like to see whether the same conclusions hold on datasets with larger improvements.

  4. The comparison with FITS may be affected by the fixed input length of 96. Stronger comparisons with FAN and simpler decomposition-based alternatives would also be helpful.

Questions for authors
  1. What is the main methodological difference between FAPA and FAN? Could the authors provide a direct comparison under the same forecasting backbones and experimental settings?

  2. HPA uses least-squares projection onto sinusoidal bases. What makes this formulation particularly advantageous over simpler harmonic regression or frequency-alignment approaches?

  3. Could the authors extend the ablation study to ECL and Traffic? I would like to understand whether AmpShape and HPA contribute equally on these datasets.

  4. How sensitive is FAPA to the input length and the number of selected frequencies? Would the improvements remain stable when the dominant periodic patterns change over time?

7 comments
AuthorsOpenReview ↗

Response 1

Thank you for this suggestion. We added a direct comparison using the same backbones and experimental protocol, with FAN’s K set to the dataset-specific values reported in its paper. Across five datasets and four horizons, FAPA achieves lower reported MSE in all 20 settings for each of PatchTST and iTransformer, with relative reductions in average MSE of approximately 11.31% and 10.19%, respectively.

FAPA and FAN differ in the purpose of the periodic branch and how its predictions are supervised and refined. FAN uses frequency decomposition for reversible normalization: it removes dominant nonstationary components to improve residual stationarity, predicts their future values, and adds them back to the backbone forecast. Across the compared datasets, FAN uses the K values reported in its paper (K = 2–30), with DC included among the selection candidates. FAPA instead dedicates a lightweight branch to forecasting a compact set of dominant periodic components, using K = 1–2 non-DC frequencies while retaining the mean in the residual. This design separates regular periodic structure from the remaining dynamics, allowing the periodic predictor and backbone to model components with different structures and amplitude scales. FAPA further provides consistent supervision for this periodic forecast through Harmonic Period Alignment (HPA). Whereas FAN independently selects the future window’s Top-K frequencies and reconstructs its target from FFT bins, HPA constructs the future periodic target at the input-selected frequencies through harmonic projection. The input and target therefore share the same frequency basis, even when these frequencies fall between the target window’s FFT bins. FAPA applies auxiliary supervision specifically to this aligned periodic target, while the residual branch learns through the final forecasting loss; FAN’s official implementation supervises both components separately. Finally, AmpShape provides explicit frequency-wise, channel-wise amplitude refinement of the MLP prediction, a mechanism absent from FAN. FAPA thus combines focused periodic modeling, aligned supervision, and explicit spectral refinement. We will clarify these differences and include the direct comparison in the revised manuscript.

Table 1. FAN comparison.

DatasetPatchTST +FANPatchTST +FAPAiTransformer +FANiTransformer +FAPA
ETTm10.4110.3790.4140.404
ETTm20.3360.2780.3380.292
ECL0.2100.1990.1940.186
Traffic0.6430.5440.5760.465
Weather0.2550.2450.2520.247
Avg.0.3710.3290.3550.319
AuthorsOpenReview ↗

Response 2

Thank you for this question. HPA is fundamentally harmonic regression; its advantage lies in how the supervision target is constructed. It fits the future periodic component directly at the input-selected frequencies, preserving periodic consistency between the input and target while avoiding approximation errors from mapping those frequencies to the nearest FFT bins. The projection is used only during training and adds no inference overhead. We will clarify how HPA uses harmonic regression to construct consistent supervision targets.

We held the input-selected periods fixed and changed only the target reconstruction method:

Table 2. Harmonic projection vs. nearest-bin reconstruction.

HPatchTST Nearest binPatchTST FullDLinear Nearest binDLinear FulliTransformer Nearest biniTransformer Full
960.3180.3180.3270.3270.3420.342
1920.3590.3590.3650.3650.3810.381
3360.3960.3880.4290.3970.4250.413
7200.4480.4500.4870.4600.4970.480

The table reports MSE averaged over three seeds on ETTm1 with input length 96; lower is better.

At H = 96/192, the input-selected frequencies exactly match FFT-bin frequencies of the target window, so the two reconstruction methods are theoretically equivalent without regularization and yield identical reported results. At H = 336/720, some input-selected frequencies fall between target bins, and HPA avoids the errors introduced by nearest-bin substitution. The forecasting benefit still depends on the energy and predictability of the corresponding periodic components; improvements do not occur in every setting.

AuthorsOpenReview ↗

Response 3

Thank you for this suggestion. We extended the PatchTST ablation study to ECL and Traffic. Given the time and computational constraints during rebuttal, we focused on representative datasets and forecasting horizons.

Table 3. ECL and Traffic ablations.

DatasetHorizonFull−HPA−AmpShape
ECL960.1820.1820.188
ECL7200.2340.2350.240
Traffic960.5410.5410.547
Traffic7200.5750.5750.580

HPA has a small effect in these settings. This is consistent with the paper’s periodicity analysis: the position consistency values are 0.90 for ECL and 0.98 for Traffic, indicating stable period selection. Their dominant daily period is also 24 time steps, and all forecasting horizons are multiples of 24, so the corresponding frequency aligns with the target window’s frequency grid. AmpShape further improves forecasting in these settings. We will include these additional ablations and discuss how each component contributes across datasets.

Reviewer_4zp9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes FAPA, a time series forecasting framework that decomposes the input into a dominant periodic component and a residual. A dedicated periodic branch and the existing backbone predict these components separately, and their outputs are added to form the final forecast. The periodic branch preserves per-window scale information and uses AmpShape to adjust the predicted amplitudes in the frequency domain. HPA constructs the periodic supervision target by projecting the future ground-truth series onto the sinusoidal subspace defined by frequencies selected from the historical input window.

Critical strengths
  1. The research problem is specific, and the method has practical value. The framework focuses on periodic component prediction and supervision target construction. It can be applied to different types of backbones without changing their internal architectures; HPA is used only during training.
  2. Experiments cover multiple backbones and datasets. The overall results support the method's effectiveness and applicability across different backbones.
  3. The analyses support some specific design choices. AmpShape and preserving per-window scale information both improve performance in the reported ablation settings. HPA provides a well-defined way to construct supervision targets for a given set of frequencies selected from the historical input window. Controlled synthetic experiments and periodic subspace error analysis also help clarify the signal structures under which the full framework achieves gains.
Critical weaknesses
  1. The incremental value over FAN needs more direct support. The paper cites FAN and acknowledges that it predicts dominant frequency components separately. However, FAN can also be applied to different backbones, and the two methods have substantial structural similarities in dominant frequency decomposition, separate branch prediction, and auxiliary supervision. The ablations provide some support for AmpShape and HPA, but there is no direct comparison with FAN using the same backbone and a common experimental protocol. This makes it difficult to assess the incremental value of these changes over the existing framework. The paper should explain more specifically which limitations of FAN the new designs address and provide corresponding comparative evidence.
  2. The limitations of normalization for amplitude prediction are overstated. Per-window normalization removes absolute scale information from the predictor's input, but with denormalization, the model can still predict relative changes in amplitude. Thus, the claim that amplitude can only be inherited rather than predicted is inaccurate. The ablations support preserving scale information in the periodic branch, but do not establish that models using normalization cannot predict amplitude changes.
Questions for authors
  1. Relative to FAN, what is the central additional predictive capability or modeling constraint introduced by this work? Please clarify which specific limitations of the existing framework AmpShape and HPA each address, and how the evidence distinguishes their benefits from those of the dual-branch framework itself. A direct comparison with FAN under the same experimental settings would help assess this incremental value.
  2. Could the conclusion about normalization be more precisely limited to the predictor's restricted access to absolute scale information? With denormalization, the model can still predict relative changes in amplitude. The statement that amplitude is inherited rather than predicted should be narrowed accordingly.
3 comments
AuthorsOpenReview ↗

Response 1

Thank you for this question. We added a direct comparison using the same backbones and experimental protocol, with FAN’s K set to the dataset-specific values reported in its paper. Across five datasets and four horizons, FAPA achieves lower reported MSE in all 20 settings for each of PatchTST and iTransformer, with relative reductions in average MSE of approximately 11.31% and 10.19%, respectively.

FAPA and FAN differ in the purpose of the periodic branch and how its predictions are supervised and refined. FAN uses frequency decomposition for reversible normalization: it removes dominant nonstationary components to improve residual stationarity, predicts their future values, and adds them back to the backbone forecast. Across the compared datasets, FAN uses the K values reported in its paper (K = 2–30), with DC included among the selection candidates. FAPA instead dedicates a lightweight branch to forecasting a compact set of dominant periodic components, using K = 1–2 non-DC frequencies while retaining the mean in the residual. This design separates regular periodic structure from the remaining dynamics, allowing the periodic predictor and backbone to model components with different structures and amplitude scales. FAPA further provides consistent supervision for this periodic forecast through Harmonic Period Alignment (HPA). Whereas FAN independently selects the future window’s Top-K frequencies and reconstructs its target from FFT bins, HPA constructs the future periodic target at the input-selected frequencies through harmonic projection. The input and target therefore share the same frequency basis, even when these frequencies fall between the target window’s FFT bins. FAPA applies auxiliary supervision specifically to this aligned periodic target, while the residual branch learns through the final forecasting loss; FAN’s official implementation supervises both components separately. Finally, AmpShape provides explicit frequency-wise, channel-wise amplitude refinement of the MLP prediction, a mechanism absent from FAN. FAPA thus combines focused periodic modeling, aligned supervision, and explicit spectral refinement. We will clarify these differences and include the direct comparison in the revised manuscript.

Table 1. FAN comparison.

DatasetPatchTST +FANPatchTST +FAPAiTransformer +FANiTransformer +FAPA
ETTm10.4110.3790.4140.404
ETTm20.3360.2780.3380.292
ECL0.2100.1990.1940.186
Traffic0.6430.5440.5760.465
Weather0.2550.2450.2520.247
Avg.0.3710.3290.3550.319

To assess the contributions of HPA and AmpShape beyond the dual-branch structure, we retain both branches and remove each component separately, as reported in the paper’s ablation table. These ablations examine the respective benefits of aligned periodic supervision and explicit amplitude refinement. A comparison between harmonic projection and nearest-FFT-bin target reconstruction is provided in our response to Reviewer YhAo’s second question.

AuthorsOpenReview ↗

Response 2

Thank you for this correction. We agree that a model with denormalization can still predict relative amplitude changes. The original statement that amplitude is “inherited rather than predicted” was too absolute, and we will revise it.

Our concern is that per-window normalization restricts the predictor’s direct access to absolute scale information. We compared two normalized periodic-branch settings: providing the historical mean and standard deviation to the predictor, or filling the corresponding input positions with zeros.

Table 2. Normalization controls.

BackboneFull UnnormalizedNormalized Without statisticsNormalized With statistics
PatchTST0.28180.28650.2820
iTransformer0.28850.29670.2891

MSE is averaged over three seeds on ETTm1, ETTm2, ECL, and Weather at H = 96/720; lower is better. Full values are aggregated from the paper’s results. The two normalized variants have identical parameter counts, each adding 512 parameters relative to Full.

Providing the statistics brings the normalized branch close to Full, supporting the design choice to preserve scale information. Leaving FAPA’s periodic branch unnormalized is a direct way to retain this information.

Reviewer_4zp9OpenReview ↗

Follow-up Comment

Thank you for the detailed responses and additional experiments. The direct comparison with FAN strengthens the evidence for FAPA’s incremental value, and the normalization controls clarify the role of absolute scale information. I also appreciate the correction and clarification of the −HPA ablation results. These responses substantially address my main concerns, and I will maintain my rating of 3.

Reviewer_YhAoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FAPA, FAPA is a forecasting wrapper that separates each input window into its largest non-DC Fourier components and a residual. An MLP predicts the periodic component using both the extracted component and the original input, without instance normalization in that branch. A learned, channel-specific spectral filter, AmpShape, refines the prediction. The original forecasting backbone processes the residual, and the two outputs are added. During training, an auxiliary periodic loss uses a target obtained by least-squares projection of the future observations onto sinusoidal bases at the input-selected frequencies. This Harmonic Period Alignment (HPA) avoids rounding those frequencies to the future window's Fourier grid. Experiments evaluate eight backbones on five datasets and four horizons, with component ablations and synthetic analysis of periodic-to-residual amplitude imbalance.

Critical strengths
  1. The construction is clear. The method projects the future sequence onto the input-selected sinusoidal basis instead of rounding those frequencies onto the target grid. For signals in that selected span, ordinary least squares reconstructs the component even when the horizon contains a non-integer number of cycles. This is an understandable way to define consistent auxiliary supervision, and its training-only computation avoids adding projection cost at inference. The claim should be understood as exactness within the selected subspace, rather than recovery of every true periodic component in arbitrary data.

  2. The empirical results are strong. The gains on ECL and Traffic are substantial for several architectures, rather than confined to one model. The component ablations, synthetic amplitude sweep, and periodic-versus-residual error analysis provide additional evidence about when the framework is useful.

Critical weaknesses
  1. The distinction from FAN is not established at the level of the actual computation, and the closest direct comparison is missing. FAN[1] already extracts instance-wise Top-K Fourier components, sends the residual to a backbone, predicts the removed component with an MLP using both that component and the raw input, adds the two predictions, and uses an auxiliary component-prediction loss.

  2. In Section 4.2, removing HPA reselects Top-K frequencies from the future window; it does not hold the input frequencies fixed and replace harmonic projection with nearest-bin reconstruction. The experiment therefore changes the identity of the supervised components as well as their representation. This distinction is visible in Table 2: on ETTm2, the largest degradations occur at H=96 and H=192, where MSE rises from 0.176 to 0.313 and from 0.238 to 0.351. Since T=96, every input bin maps to an integer target bin at both horizons, so those differences cannot be explained by target-grid mismatch.

  3. Sections 3.4 and 4.4 add a separate nonlinear predictor and channel-specific weights; for the sub-1K FITS and SparseTSF backbones, the additional 70.7K parameters are particularly substantial. Identical training settings within each pair do not control for that added capacity. Moreover, Eq. (1) makes AmpShape a sample-independent linear spectral filter after the MLP, with Wa+WbW_a+W_b equivalent in the forward computation to one effective weight matrix. Its ablation supports a useful parameterization, but does not yet establish a distinctive amplitude-modeling capability. A comparable auxiliary MLP or simple output-calibration control would help determine how much benefit requires the proposed decomposition, supervision, and spectral refinement together.

  4. Table 1 fixes T=96 throughout. The authors fairly disclose that ETTm2 is essentially unchanged and report both aggregate and per-setting improvements. Nevertheless, using the rounded main-table entries, FITS contributes approximately 49% of the total absolute MSE reduction across the 40 entries; excluding it reduces the aggregate relative improvement from approximately 10.7% to 6.5%. The remaining gain is still useful, but the headline is strongly influenced by that baseline. Since input length determines frequency resolution and the periods used by HPA, evidence at one longer look-back would materially clarify whether the proposed advantage persists when both the baseline forecast and input frequency estimates improve.

[1] Frequency Adaptive Normalization For Non-stationary Time Series Forecasting

Questions for authors
  1. Beyond the terminology of normalization versus decomposition, what operations distinguish FAPA from FAN? Can you compare both wrappers on the same backbone and protocol, then identify the effects of the changed target construction, DC handling, and AmpShape? This comparison would directly affect my novelty and significance assessment.

  2. Can you hold the input-selected frequencies fixed and compare harmonic projection against nearest-bin target reconstruction, alongside the existing future-Top-K target? Please explain the large −HPA degradation at H=96 and H=192, where all input frequencies lie exactly on the target grid. Which gains arise from consistent component selection, and which arise from avoiding off-grid reconstruction?

  3. Is the intended limitation the loss of absolute-scale conditioning, rather than an inability to predict a new amplitude? Can the claim be stated in those terms and checked with a normalized branch that also receives the removed scale statistics? Such a control would distinguish information loss from other effects of changing the branch's input distribution.

  4. How does FAPA compare with a similarly sized auxiliary MLP attached to the backbone under matched training? For AmpShape, does the gain persist against a simple channel-wise output calibration, and what changes when Wa+WbW_a+W_b is replaced by a single weight initialized to one? These controls would clarify whether the benefit is specific to frequency-wise refinement or mainly additional capacity and optimization.

  5. Does the advantage persist at a longer input length, with each Base/+FAPA pair still matched? A focused comparison on FITS and one strong backbone such as iTransformer, including both an ECL/Traffic case and an ETT case, would help determine whether the main result generalizes beyond the fixed T=96 regime.

8 comments
AuthorsOpenReview ↗

Response 1

Thank you for highlighting the structural similarity between FAPA and FAN. We added a direct comparison using the same backbones and training protocol, with FAN’s K set to the dataset-specific values reported in its paper.

FAPA and FAN differ in the purpose of the periodic branch and how its predictions are supervised and refined. FAN uses frequency decomposition for reversible normalization: it removes dominant nonstationary components to improve residual stationarity, predicts their future values, and adds them back to the backbone forecast. Across the compared datasets, FAN uses the K values reported in its paper (K = 2–30), with DC included among the selection candidates. FAPA instead dedicates a lightweight branch to forecasting a compact set of dominant periodic components, using K = 1–2 non-DC frequencies while retaining the mean in the residual. This design separates regular periodic structure from the remaining dynamics, allowing the periodic predictor and backbone to model components with different structures and amplitude scales. FAPA uses Harmonic Period Alignment (HPA) to construct consistent auxiliary supervision for the periodic branch, with benefits that vary across datasets and forecasting horizons. Whereas FAN independently selects the future window’s Top-K frequencies and reconstructs its target from FFT bins, HPA constructs the future periodic target at the input-selected frequencies through harmonic projection. The input and target therefore share the same frequency basis, even when these frequencies fall between the target window’s FFT bins. FAPA applies auxiliary supervision specifically to this aligned periodic target, while the residual branch learns through the final forecasting loss; FAN’s official implementation supervises both components separately. Finally, AmpShape provides explicit frequency-wise, channel-wise amplitude refinement of the MLP prediction, a mechanism absent from FAN. FAN’s Top-K selection can include DC, so its separate branch may also need to predict the window mean. This may partly explain the degradation on ETTm2, consistent with our DC control. We will expand the discussion of DC handling in the revised manuscript. We will clarify these differences and include the direct comparison in the revised manuscript.

Table 1. FAN comparison.

DatasetPatchTST +FANPatchTST +FAPAiTransformer +FANiTransformer +FAPA
ETTm10.4110.3790.4140.404
ETTm20.3360.2780.3380.292
ECL0.2100.1990.1940.186
Traffic0.6430.5440.5760.465
Weather0.2550.2450.2520.247
Avg.0.3710.3290.3550.319

Table 2. Additional ablations.

SettingETTm1ETTm2
Full0.3790.278
−HPA0.3820.288
−AmpShape0.3840.281
DC moved to the periodic branch0.3930.320
FAPA with FAN’s K0.3830.278
AuthorsOpenReview ↗

Response 2

Thank you for pointing out that these two effects should be examined separately. We added a control that holds the input-selected periods fixed and replaces only the target-side harmonic projection with nearest-FFT-bin reconstruction.

When H/pkH/p_k is an integer (H = 96, 192), the frequencies corresponding to the input-selected periods lie exactly on the target-window frequency grid. The two reconstructions are theoretically equivalent without regularization and yield identical reported results, consistent with the reviewer’s observation. At H = 336/720, some input-selected frequencies fall between target-window FFT bins. For these frequencies, harmonic projection avoids the approximation introduced by nearest-bin reconstruction. For example, on ETTm1, DLinear’s MSE decreases from 0.429 and 0.487 to 0.397 and 0.460, respectively. The forecasting benefit still depends on the energy and predictability of the corresponding periodic components.

Table 3. Harmonic projection vs. nearest-bin reconstruction.

HPatchTST Nearest binPatchTST FullDLinear Nearest binDLinear FulliTransformer Nearest biniTransformer Full
960.3180.3180.3270.3270.3420.342
1920.3590.3590.3650.3650.3810.381
3360.3960.3880.4290.3970.4250.413
7200.4480.4500.4870.4600.4970.480

We investigated the unusually large short-horizon degradation in the original −HPA results and identified an issue in target construction: the DC component was not always excluded for constant-valued channels in future target windows, causing some supervision targets to include the window mean unintentionally. In the submission configurations and execution paths we checked, this issue affected the −HPA variant but not Full or the other main ablations reported in the submission. Whether the issue occurs and how much it affects the results depend on the dataset and window length, with the largest anomalies concentrated in ETTm2 at H = 96 and 192. Table 4 gives the corrected results; all other entries remain unchanged to three decimal places. We will update the corresponding results and analysis in the revised manuscript.

Table 4. Corrected −HPA results.

DatasetHFullOriginal −HPACorrected −HPA
ETTm2960.1760.3130.177
ETTm21920.2380.3510.246

Thus, at H = 96/192, the differences mainly arise from whether the supervision targets reuse the input-selected periods. At H = 336/720, harmonic projection additionally avoids errors from replacing those frequencies with neighboring FFT bins.

AuthorsOpenReview ↗

Response 3

Thank you for this clarification. We agree that normalized models can still predict relative amplitude changes. The periodic predictor can benefit from direct access to scale information, and omitting normalization is a straightforward way to preserve it. We will revise the original claim to focus on the predictor’s access to absolute scale information.

We added two normalized periodic-branch controls, with and without concatenating the mean and standard deviation of the historical input to the predictor’s input. Both use output denormalization and the same architecture; the corresponding input positions are filled with zeros when the statistics are not provided.

Table 5. Normalization controls.

BackboneFull UnnormalizedNormalized Without statisticsNormalized With statistics
PatchTST0.28180.28650.2820
iTransformer0.28850.29670.2891

MSE is averaged over three seeds on ETTm1, ETTm2, ECL, and Weather at H = 96/720; lower is better. Full values are aggregated from the paper’s results. The two normalized variants have identical parameter counts, each adding 512 parameters relative to Full.

Providing the statistics brings the normalized variant’s performance close to that of Full, indicating that preserving scale information is the key factor. FAPA achieves this directly through an unnormalized periodic branch, without concatenating additional statistics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.