FreDF: Frequency-Domain Loss for Autocorrelated Time-Series Forecasts
Summary
The document explains FreDF, a training approach for direct multi-step time-series forecasting. Direct forecast models predict several future values at once and commonly calculate their loss step by step in the time domain. The article argues that this can overlook autocorrelation among the target values, creating a mismatch between the loss and the data structure.
FreDF transforms forecasts and labels into the frequency domain and adds a frequency-based error to the usual time-domain loss. A weighting parameter controls the balance between the two; the article reports experiments favoring a mixed objective and says Fourier transforms performed better than the other transformations considered. It also describes integrating the loss into an existing forecasting model without changing its architecture.
The evidence is presented as experimental results reported by the method’s authors and an implementation using a FEDformer model. The document gives no detailed datasets, metrics, or independent replication in the supplied text. It cautions that relying only on frequency-domain loss reduced accuracy in the cited experiments, so the approach should not be read as universally superior.
Key ideas
- Direct multi-step forecasting can overlook autocorrelation among predicted target values when each step is treated independently.
- FreDF compares forecasts and labels in the frequency domain as well as the time domain.
- The combined loss uses a weighting parameter to balance time-domain and frequency-domain errors.
- The method is a training objective and does not require changing the forecasting model architecture.
- The article reports favorable experiments but provides limited detail in the supplied text.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.