# Frequency-domain MLPs are More Effective Learners in Time Series Forecasting

**Authors:** Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Defu Lian, Ning An, Longbing Cao, Zhendong Niu  
**arXiv:** [2311.06184](https://arxiv.org/abs/2311.06184)  
**Date:** 2023-11-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.06184v1/

## Key points

- Time series forecasting is crucial in industries like finance, traffic, energy, and healthcare.
- Traditional MLP-based methods face challenges such as point-wise mappings and information bottlenecks.
- A novel approach involving the application of MLPs in the frequency domain has been proposed for time series forecasting.
- Frequency-domain MLPs offer a global view and energy compaction for more effective forecasting.
- The FreTS architecture utilizes Frequency-domain MLPs in two main stages: Domain Conversion and Frequency Learning.
- FreTS demonstrates consistent superiority over state-of-the-art methods for short-term and long-term forecasting tasks in real-world benchmarks.
- Ablation studies confirm the effectiveness of frequency channel and temporal learners in improving short-term forecasting performance.

## Summary

Time series forecasting is a crucial aspect in industries such as finance, traffic, energy, and healthcare. While existing literature has developed sophisticated architectures based on RNNs, GNNs, or Transformers for this purpose, there is a growing interest in exploring the use of multi-layer perceptrons (MLPs) due to their simple structure, low complexity, and superior performance. However, traditional MLP-based methods face challenges such as point-wise mappings and information bottlenecks that limit their capabilities. To address these limitations, a novel approach involving the application of MLPs in the frequency domain has been proposed for time series forecasting. By analyzing the learned patterns of frequency-domain MLPs, two key characteristics have been identified that benefit forecasting: global view and energy compaction. The frequency spectrum allows MLPs to have a comprehensive view of signals and learn global dependencies more effectively. Additionally, focusing on smaller key parts of frequency components with compact signal energy enhances forecasting accuracy. In response to these findings, the FreTS architecture has been introduced as a simple yet effective solution built upon Frequency-domain MLPs for time series forecasting. It consists of two main stages: Domain Conversion and Frequency Learning. The former transforms time-domain signals into complex numbers in the frequency domain while the latter utilizes redesigned MLPs to learn the real and imaginary parts of frequency components. These stages operate at both inter-series and intra-series scales to facilitate channel-wise and time-wise dependency learning. Extensive experiments conducted on 13 real-world benchmarks demonstrate the consistent superiority of FreTS over state-of-the-art methods for short-term and long-term forecasting tasks. Ablation studies further validate the effectiveness of frequency channel and temporal learners in improving short-term forecasting performance. Overall, utilizing Frequency-domain MLPs shows promise as an innovative approach to enhancing time series forecasting accuracy across various domains.