acceptodds
Withdrawn from ICLR 2027

Time-Series Forecasting Models Care Less About What Humans See

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series forecasting predicts future values from historical observations to support energy scheduling, weather prediction, and traffic management. In recent years, Fourier operations have been widely used in time-series forecasting models. Their reported gains are widely attributed to the explicit representation of periodic structure with a prescribed Fourier basis. Fourier operations commonly introduce multiple filtered views of the same history, but existing studies have not analyzed whether the reported gains come from the Fourier basis or the accompanying view expansion. We examine this attribution with matched Random controls. We replace Fourier filters or bases with norm- and parameter-matched Random counterparts. We conduct 96 matched comparisons across generic backbones, backbone scales, and eight published models, covering input histories with varying degrees of periodic repetition. Fourier shows no aggregate forecasting advantage over Random, with 47 Fourier wins, 47 Random wins, and two ties. Both representations often improve over Raw. Further analysis shows that their shared gain primarily reflects expanding each history into multiple distinct filtered views. Increasing the response count usually improves forecasting, whereas longer filters provide no corresponding benefit. Prescribed frequency structure can provide additional gains when it matches repeated patterns within the input window.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series forecasting, Fourier transform, Filter banks, Random features.

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_455qOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates whether the benefits of Fourier-based time-series forecasting methods actually originate from their prescribed frequency structures. The authors replace Fourier transformations with fixed random filters and find no consistent advantage of Fourier over Random, while both often outperform the raw-input baseline, suggesting that the improvements primarily come from providing multiple distinct filtered views rather than the Fourier basis itself.

Critical strengths

This paper raises a question about the benefits brought by a widely used design (Fourier-based spectrum) in time series forecasting, questioning whether it comes from the inductive biases from Fourier basis. This is a meaningful question, and experiments (Fourier vs. Random, especially for the eight published models) do show that the Fourier basis won't consistently help.

Critical weaknesses

It supports that Fourier basis is not consistently helpful, but does not validly proves what the benefits comes from. This paper tries to do this by comparing Fourier and Random with Raw, and polar decomposition. The key problem is, all the methods with expanded filter views repeat model forwarding 9 times (1 for original xx and 8 for each bixb_ix), and use a small linear layer (the aia_i and β\beta in Appendix A.1.3) to ensemble the 9 repeats. Though model parameters are barely increased, but computation budgets are multiplied by 9. Polar decomposition experiments do not resolve this; in the contrast, it supports my concern. Table 9 shows that Repeat and H brings marginal improvement but Q and QH show significant gains. Notably, there are only 2 times authentic computation budget for Repeat, as all the bib_i are the same for all ii. Experiments that controls computation budgets are needed to analyze whether the benefits just come from more computation.

This repeating setting itself is also a problem. First, sharing model parameters brings irrelavant variables. The model has to learn to deal with 9 different input space, and we have no idae whether this will affect performance. Also, generalization of existing conclusion is limited, as no other input paradigms are examined. I look forward to experiments for more input paradigms.

Questions for authors
  • Conduct experiments that controls computation budgets.
  • Apart from Q, H analysis, provides results with a random matrix for B.
  • Conduct experiments for more paradigms of providing the filtering views as inputs.
  • Provides grounded analysis about what contributes to the gains of Fourier spectrum. See 'Critical Weaknesses' for more information.

If these four questions are resolved (especially the last one), I will be happy to raise my score to 3.

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

This paper investigates whether the gains of Fourier-based time-series forecasting models come from the prescribed Fourier basis itself or from the accompanying expansion into multiple filtered views. The authors construct norm- and parameter-matched Random controls, evaluate 96 paired settings across generic backbones, model scales, and published Fourier-based models, and further use a polar decomposition B=QH together with Repeat controls to attribute the shared gain. The main conclusion is that Fourier shows no aggregate advantage over Random, while multiple distinct filtered responses appear to explain much of the gain over Raw.

Critical strengths
  1. The paper asks an important and well-motivated attribution question: when a forecasting model improves after receiving Fourier-transformed information, does the benefit come from the specific sinusoidal basis, or simply from receiving several differently filtered versions of the same history?

  2. The Fourier–Random comparison is carefully controlled within the generic interface. Fourier and Random share the same filter length, response count, corresponding row norms, Raw branch, predictor, and training protocol, while differing primarily in the filter directions.

  3. The polar-decomposition analysis and Repeat controls are conceptually strong and provide useful mechanistic insight. The decomposition B=QH provides a principled way to separate the metric modification induced by the filter bank from the construction of multiple distinct responses.

  4. The empirical coverage across different backbone families, backbone scales, and published Fourier-based models is broad.

  5. The paper is clearly written and generally provides a detailed and nuanced discussion of its findings

Critical weaknesses
  1. The paper reports 47 Fourier wins, 47 Rnadom wins, and two ties, but the 96 settings are neither independent nor dataset-balanced. In particular, ETTm1 contributes a very large fraction of the settings (48/96) because it appears in the generic experiment, the scaling experiments, and the published-model evaluation. Moreover, a win count ignores effect magnitude: a marginal improvement and a several-percent degradation each count as one win. Therefore, the claim of no aggregate advantage would be stronger with a paired effect-size analysis or a dataset-balanced summary.

  2. All forecasting experiments use horizon 96, while the generic, scaling, and mechanism studies also use history length 96. However, Section 5.2 argues that Fourier structure becomes more useful when repeated cycles occur within the observed window, and this point is emphasized throughout the paper. A stronger test would vary the history length on the same dataset, e.g., 48, 96, and 192, so that the number of repeated cycles changes while the underlying data source remains fixed.

  3. The paper observes that Electricity contains several repeated daily cycles in the input window and that Fourier consistently outperforms Random on the generic backbones. However, Electricity differs from ETTm1 and Weather in many properties beyond periodic repetition. Therefore, the current evidence is more consistent with an association than with an isolated causal effect of periodicity or frequency alignment. Controlled synthetic signals with known periods and noise levels would provide stronger support for this conclusion.

  4. Changing K appears to alter branch count, filter dictionary, spectral coverage, brandwise power allocation, computation, and the resulting Q. The authors are encouraged to clarify whether the observed improvement is primarily due to a larger K, a larger number of distinct views, or the accompanying representation changes.

  5. Based on the stated attribution formula and the values reported in Table 1, the average Mamba gain is approximately 5.5736%, whereas Tables 10 and 12 report 6.0051. The authors should reconcile these values or clarify whether the mechanism results correspond to a different run or experimental configuration.

  6. Several important references are missed:

  • FreEformer[1] studies alternative bases such as wavelet, polynomial, and identity transformations. Although its controls are not equivalent to the present Random/polar framework, it is highly relevant to the paper's basis-selection argument.
  • FAN[2] and DeRiTS[3] are relevant for distinguishing prescribed Fourier structure from adaptive frequency modeling and multi-branch mechanisms for handling non-stationarity.

[1] Yue, Wenzhen, et al. "Freeformer: Frequency enhanced transformer for multivariate time series forecasting." arXiv preprint arXiv:2501.13989 (2025).

[2] Ye, Weiwei, et al. "Frequency adaptive normalization for non-stationary time series forecasting." Advances in neural information processing systems 37 (2024): 31350-31379.

[3] Fan, Wei, et al. "Deep frequency derivative learning for non-stationary time series forecasting." arXiv preprint arXiv:2407.00502 (2024).

Questions for authors
  1. Could the authors provide a paired Fourier-versus-Random effect-size analysis, preferably with dataset-balanced aggregation, rather than relying mainly on win counts?
  2. Could the authors evaluate multiple history lengths on the same dataset to directly test whether the relative advantage of Fourier changes as more repeated cycles become visible in the input window?
  3. Could the authors provide a controlled experiment, such as synthetic signals with known periodicities to more directly verify the proposed relationship between frequency alignment and forecasting benefit?
  4. Since changing K also changes other aspects of the representation, could the authors design an experiment with a fixed response dictionary or frequency grid so that the effect of response count can be isolated more cleanly?
  5. Could the authors explain why the Mamba gain recomputed from Table 1 is approximately 5.5736%, while Tables 10 and 12 report 6.0051%? Were these results obtained from different runs or configurations?
Reviewer_QhGWOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper examines whether the gains from Fourier operations in time-series forecasting arise from their explicit representation of periodic structure or from view expansion. By comparing Fourier filters with norm- and parameter-matched random filters, the authors find that both outperform Raw, with little difference between Fourier and Random. The results suggest that the gains primarily come from expanding each input history into multiple distinct filtered views, although Fourier's frequency structure can provide additional benefits when aligned with repeated patterns.

Critical strengths

S1. The research question is interesting. Fourier operations are widely used in recent time-series forecasting works, and understanding where their performance gains actually come from is an important question in the field.

S2. The experimental design is interesting. Using different input views to anlayze performance gains across generic backbones is a useful approach.

S3. The finding that Fourier operations do not always improve performance by capturing periodic patterns is valuable.

Critical weaknesses

W1. I'm not fully convinced that Fourier and Random provide similar performance gains.

W1-1. Figures 2 and 3 seem somewhat misleading. Since MSE values are largely dependent on the dataset and backbone, Fourier and Random results for the same setting may naturally fall within a similar range and appear close to the diagonal. I think if we 3-dimensional plot Random vs Fourier vs Raw, the points could also like near the diagonal. So, current figures make it hard to tell whether they actually provide similar improvements over Raw. I think it would be more informative to plot the relative MSE improvements over Raw, with Random on the x-axis and Fourier on the y-axis.

W1-2. The experiments seem quite limited. Most settings use a history length and forecasting horizon of 96, with only four datasets.(ETTm1, Weather, Electricity, Exchange) Given the broad claim of the paper, this feels more like a preliminary study. Longer input and prediction windows, along with more standard datasets (e.g., ETTh1, ETTh2, ETTm2, Traffic), especially those with clear periodic patterns, might lead to different conclusions. I would like to see more experiments, or at least an explanation of why these datasets and settings were chosen.

W2. I also have some concerns about hyperparameter tuning and baseline performance. The experiments seem to use the same model sizes and learning rates for Random and Fourier. While this is reasonable for a controlled comparison, the best hyperparameters might differ between the two methods. Also, increasing model size often leads to worse performance in the scaling experiments, rather than improving it. This makes me wonder whether the baselines were sufficiently tuned. Could the authors provide more details on how the hyperparameters were chosen and whether method-specific tuning would change the conclusions?

Questions for authors

Please see weakness above.

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

Related papers

Discussion (0)

Sign in to comment.