ATFNet Adaptive Time and Frequency Forecasting for Trading Series

Summary

This article presents an implementation of ATFNet, a multivariate time-series forecasting approach that combines two forecasting paths. The T-Block models patterns in the time domain, while the F-Block analyzes periodic and global structure in the frequency domain using Fourier decomposition and complex-valued attention. The model adaptively combines the two forecasts. The article describes implementation choices for a neural-network layer, including normalization, patching, positional encoding, and padding the Fourier input to a power-of-two length.

The authors’ approach is implemented in MQL5 and applied to market data, with the conclusion reporting that tests suggest potential for trading strategies. The excerpt provides no detailed performance metrics, benchmark comparison, or enough information to assess robustness, so that claim should be treated as preliminary. Forecast quality does not by itself show that a trading strategy will remain profitable after costs or across market regimes.

Key ideas

• ATFNet combines a time-domain path for local temporal patterns with a frequency-domain path for periodic structure.
• The frequency path uses Fourier representation and complex-valued attention to model dependencies in the spectrum.
• The model adaptively merges the forecasts from its two paths.
• Its implementation processes multivariate series and uses normalization and patching as part of the time-domain path.
• The reported trading potential is not accompanied here by detailed performance or robustness evidence.

Tags

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