ATFNet: Combining Time and Frequency Domains for Forecasting

Summary

This article explains ATFNet, a neural forecasting architecture that combines time-domain and frequency-domain models. Its time block uses segmented input to capture local patterns, while a complex-valued frequency block models global spectral structure. Their forecasts are combined with weights derived from the share of spectral energy in the dominant harmonic series, intended to adapt the balance to each series’ periodicity.

The method also uses an extended discrete Fourier transform to align the observed spectrum with the frequency grid of the complete input-plus-forecast series, and complex spectral attention to learn across frequency responses. The article describes channel-independent processing and normalization in both domains. It reports that experiments on eight real datasets found promising results, with ATFNet outperforming other contemporary forecasting methods on many of them. These are time-series forecasting results rather than evidence of trading profitability; the article’s practical implementation is incomplete, with the complex attention layer discussed as one component and further development deferred.

Key ideas

• ATFNet combines a time-domain block for local patterns with a frequency-domain block for global spectral dependencies.
• The model adapts the contribution of each block using the dominant harmonic series’ share of total spectral energy.
• An extended Fourier transform aligns observed frequencies with the grid for the full input and forecast horizon.
• Complex spectral attention learns relationships among frequency responses, while channels are processed independently.
• The article reports favorable benchmark comparisons on many of eight real datasets, but does not establish trading performance.

Tags

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