FEDformer for Long-Term Time-Series Forecasting in Trading

Summary

This article explains FEDformer, a Transformer architecture for long-horizon time-series forecasting that combines seasonal-trend decomposition with frequency-domain processing. Its Fourier version randomly selects a subset of frequency components, while its wavelet version decomposes signals into localized time and frequency information. These blocks replace standard attention components to capture broader temporal structure while reducing computational and memory costs; the article describes the target complexity as linear rather than quadratic.

The article also presents a trading-oriented implementation and reports that the model showed potential on historical market data. However, the excerpt describes uneven performance, including losing periods and a profit factor of 1.02 with a win share below 46 percent during the cited test. The author concludes that losses need further work. These results are specific to the implementation and test interval, and do not establish that the forecasting architecture will perform reliably across instruments or market regimes.

Key ideas

• FEDformer combines trend-seasonal decomposition with frequency-based processing for long-term forecasts.
• Fourier blocks sample frequency components, while wavelet blocks retain localized time and frequency structure.
• Frequency-enhanced blocks are intended to lower the computational cost of long-sequence Transformer models.
• The trading implementation showed potential but had uneven results and a low reported profit factor.
• Performance in the described test does not establish robustness across markets or regimes.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.