Frequency Domain MLP: A More Effective Learner for Time Series Forecasting: Applying MLP to the Frequency Domain to Overcome Point Mapping and Information Bottleneck in Time Domain MLP, Improving Prediction Performance
The paper proposes the FreTS architecture, which transforms time series to the frequency domain via discrete Fourier transform and uses frequency domain MLP to learn real and imaginary parts, outperforming existing methods on 13 real-world benchmarks, including 7 short-term and 6 long-term forecasting tasks.
This work finds that frequency domain MLP has global perspective and energy compaction properties, enabling easier learning of global dependencies and focusing on key frequency components, thus overcoming the point mapping and information bottleneck issues of traditional time domain MLP.
FreTS consists of two stages: domain transformation and frequency learning. Domain transformation uses FFT to convert time domain signals to frequency domain complex numbers. Frequency learning uses redesigned MLP to process real and imaginary parts separately, operating at inter-sequence and intra-sequence scales to learn dependencies along channel and time dimensions. Experiments cover 13 benchmarks including ETT, Exchange, Weather, etc., consistently outperforming Transformer, RNN, and other baselines on MSE and MAE metrics. Limitations: mainly for univariate and multivariate time series, not involving high-dimensional spatiotemporal data.
In industries relying on time series forecasting such as finance, transportation, energy, and healthcare, FreTS offers a low-complexity, high-performance alternative that can replace existing Transformer or RNN models, reducing deployment costs.
It is recommended to integrate the FreTS module into time series forecasting platforms (e.g., as a Prophet alternative), and conduct A/B testing in scenarios such as power load forecasting or stock price prediction to evaluate accuracy improvement and computational cost savings.
Future directions for verification include: scalability on larger datasets (e.g., 100+ variables), integration with pre-trained models, and transferability of frequency domain MLP to anomaly detection or classification tasks.