FITS: Lightweight Time-Series Forecasting by Frequency Interpolation
Summary
The document explains FITS, a compact neural-network approach to time-series forecasting. It transforms an input window with the fast Fourier transform, interpolates its complex frequency representation using a complex linear layer, then applies the inverse transform to produce forecasts. Complex values encode both component amplitude and phase, allowing the model to learn changes in each during interpolation.
The method uses reversible instance normalization to handle the large zero-frequency component associated with a nonzero mean, and a low-pass filter to reduce model size by removing higher-frequency components. It is trained in the time domain with standard losses and can be used for forecasting or reconstruction. The article describes results from the cited research, including minimal distortion when retaining only a portion of the frequency representation, but gives no detailed benchmark figures here. Its own MQL5 implementation, tested on historical data, did not produce the desired results; the author cautions that this outcome may not reflect the original algorithm.
Key ideas
- FITS uses Fourier transforms to move time-series windows into a complex frequency representation for interpolation.
- A complex linear layer can learn amplitude scaling and phase shifts across the interpolated frequencies.
- Reversible normalization addresses the zero-frequency component created by a nonzero segment mean.
- A low-pass filter reduces model size by discarding higher-frequency components that may contain noise.
- The article’s MQL5 implementation did not achieve the desired results, so its findings should not be generalized to the original method.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.