FreTS: Frequency-domain MLPs for Time Series Forecasting
FreTS (Frequency-domain MLPs for Forecasting) · Also known as: Frequency-domain MLPs, FrequencyMLP, FreTS Forecaster, Frekans Alanı MLP
FreTS is a time series forecasting architecture introduced by Yi et al. at NeurIPS 2023. It departs from Transformer-based designs by applying simple Multi-Layer Perceptrons (MLPs) entirely in the frequency domain. The model transforms input sequences with the Discrete Fourier Transform and then learns temporal and channel dependencies through complex-valued MLP layers, achieving competitive or superior long-term forecasting accuracy with substantially lower computational cost.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
FreTS is well-suited for long-horizon multivariate time series forecasting tasks where periodicity and global temporal structure dominate, such as electricity consumption, weather prediction, and traffic flow. It is most appropriate when computational budget is limited, as frequency-domain operations reduce effective sequence length. It assumes the series contains meaningful spectral structure; on highly irregular or non-stationary data without clear periodic components, gains may be smaller. Alternatives include FEDformer for wavelet-based spectral mixing, PatchTST for patch-based attention, and TSMixer for pure channel-mixing MLPs.
Strengths & limitations
- Low computational complexity: operating on T/2+1 frequency bins rather than T time steps reduces memory and FLOPs significantly for long sequences.
- Captures global periodicities and long-range dependencies compactly through spectral coefficients, avoiding the quadratic cost of full self-attention.
- Simple architecture with no positional encoding, tokenization, or complex attention mechanisms, making it easy to implement and tune.
- Demonstrated competitive long-term forecasting accuracy across multiple standard benchmarks against both Transformer-based and MLP-based baselines.
- The Discrete Fourier Transform assumes periodicity; non-stationary or trend-dominated series may require additional preprocessing such as instance normalization or differencing.
- Channel-mixing MLP implicitly assumes a fixed number of variables, limiting zero-shot transfer to datasets with different channel counts.
- Frequency truncation (keeping only the lowest T/2+1 bins) discards high-frequency components, which may matter for irregular or impulsive signals.
- Lacks explicit temporal locality; cannot easily attend to recent local context separately from long-range patterns without architectural modification.
Frequently asked
How does FreTS differ from FEDformer, which also uses frequency-domain operations?
FEDformer integrates frequency-domain operations into a Transformer backbone by replacing the attention mechanism with Fourier or wavelet mixing, retaining positional encodings and feed-forward blocks. FreTS is a simpler, purely MLP-based architecture that discards the Transformer entirely and applies all mixing steps directly in the frequency domain, resulting in lower complexity and fewer components.
Does FreTS require the time series to be strictly stationary?
FreTS does not strictly require stationarity, but it benefits from instance-level normalization such as RevIN applied before the Fourier Transform. Without normalization, distributional shift between training and inference windows can corrupt the spectral representation. For strongly non-stationary series, differencing or trend removal as preprocessing is advisable before applying FreTS.
Can FreTS handle univariate forecasting, or is it only for multivariate settings?
FreTS can handle univariate series; in that case the channel-mixing MLP degenerates to a single-channel operation and the model reduces to a frequency-domain MLP applied solely along the time axis. However, its channel-mixing component provides the most benefit in multivariate settings where correlated variables share spectral structure that the model can exploit across channels.
Sources
- Yi, K., Zhang, Q., Fan, W., Wang, S., Wang, P., He, H., An, N., Lian, D., Cao, L., & Niu, Z. (2023). Frequency-domain MLPs are more effective learners in time series forecasting. NeurIPS. link ↗
How to cite this page
ScholarGate. (2026, June 2). FreTS (Frequency-domain MLPs for Forecasting). ScholarGate. https://scholargate.app/en/deep-learning/frets
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →