Time Series Forecasting in the Frequency Domain

This direction is split into other directions: frequency-domain feature extraction. Its last version stays readable and citable here.

Version 1 of 1271 papers328 conceptsfrozen 2026-10-04

Traditional time-series models treat data as a sequence of discrete points in time. Whether using Recurrent Neural Networks (RNNs) or standard Transformers, these models attempt to learn dependencies by looking at how one moment relates to the next. However, for long-term forecasting, this "point-wise" perspective is often overwhelmed by high-frequency noise and the sheer computational cost of modeling long-range dependencies.

A growing body of research suggests a more effective alternative: shifting the entire modeling process into the frequency domain. By using the Fast Fourier Transform (FFT) to represent time series as a collection of spectral components, researchers are finding they can capture global periodic patterns more accurately, more efficiently, and with greater robustness to noise.

The Limits of Point-wise Attention

In a standard Transformer, the self-attention mechanism calculates the relationship between every pair of time steps. As the look-back window grows, this cost increases quadratically. More importantly, as Haixu Wu and Mingsheng Long argue in their work on Autoformer, point-wise attention struggles to distinguish the underlying "signal" from the "noise" in long-term sequences. A single outlier at time $t$ can disproportionately affect the attention weights for the entire sequence.

By contrast, the frequency domain offers a global perspective. When a signal is transformed into its spectral representation, periodic patterns—like the daily cycle of traffic or the seasonal ebb of retail sales—emerge as distinct, stationary peaks. This allows models to focus on the fundamental "rhythms" of the data rather than volatile step-by-step transitions.

Sparsity and Spectral Efficiency

The primary intuition behind frequency-domain modeling is sparsity. While a time-series signal might have thousands of time steps, its essential structure is often defined by just a few dominant frequencies.

In FEDformer, Tian Zhou and Qingsong Wen exploit this by performing attention in the Fourier space. Instead of attending to all time points, the model selects a random subset of frequency components. Because most of the signal's energy is concentrated in these low-frequency modes, the model can reconstruct the global trend with linear computational complexity ($O(L)$) without losing predictive power.

Similarly, Jonathan Crabbé demonstrates in Time Series Diffusion in the Frequency Domain that generative models are more efficient when they operate on these localized spectral components. Learning to generate a few important frequencies is significantly easier than trying to resolve the dependencies of every individual time step.

Decomposition as a First-Class Citizen

Real-world time series are rarely pure signals; they are mixtures of long-term trends and cyclical "seasonal" patterns. Modern architectures now embed series decomposition directly into the model layers.

  1. Trend Extraction: Models like Autoformer use moving averages to isolate the smooth trend.
  2. Seasonal Modeling: The remaining "seasonal" part is then passed to a frequency-domain module. Gerald Woo and Steven Hoi in CoST use a learnable Fourier layer to apply linear transformations to these spectral components. This allows the model to identify periodicities without the user having to manually specify the period length (e.g., "24 hours").

This disentanglement is crucial for stability. As Duane S. Boning and Fan-Keng Sun show with FreDo, modeling these components in the frequency domain prevents the error accumulation that typically plagues long-term time-domain forecasts.

Capturing Multivariate Correlations

When dealing with multiple related time series (e.g., sensors in a factory or traffic at different intersections), the challenge is capturing both temporal and inter-series correlations.

Defu Cao and Yujing Wang addressed this in StemGNN by combining Graph Fourier Transforms (GFT) with Discrete Fourier Transforms (DFT). The GFT captures how different variables relate to one another in a "spectral graph" space, while the DFT handles the temporal rhythms. By operating entirely in the spectral domain, the model filters out high-frequency noise that would otherwise confuse a standard graph neural network.

Robustness to Temporal Shifts

One of the most practical advantages of the frequency domain is its invariance to temporal shifts. In mobile sensing, for instance, a user's gait might be captured by an accelerometer, but the exact starting point of a step is arbitrary.

In DeepSense, Shuochao Yao and colleagues demonstrate that converting raw sensor data into frequency magnitudes and phases makes convolutional networks much more robust. Because the "shape" of the frequency spectrum remains the same even if the signal is shifted in time, the model can identify rhythmic patterns (like walking or running) regardless of when the sensor started recording.

This same principle applies to domain adaptation. Romain Mussard and Paul Honeine show in their work on UniJDOT that spectral features act as "robust invariants." When moving from one dataset to another—where sensors or environments might change—the underlying spectral signature of the process often remains the most reliable feature for the model to track.

Papers in this version

Showing 60 of 271 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_time_series_forecasting_in_the_frequency_c29f10b3,
  title = {Time Series Forecasting in the Frequency Domain},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-series-forecasting-in-the-frequency-domain/rdr_c29f10b3266b4e0f9f3fafdf9289c4e0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_time_series_forecasting_in_the_frequency_c29f10b3_v1,
  title = {Time Series Forecasting in the Frequency Domain},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/time-series-forecasting-in-the-frequency-domain/rdr_c29f10b3266b4e0f9f3fafdf9289c4e0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}