New Method Enhances Time-Series Forecasting with Spectral Retrieval

This research introduces SpecReTF, an innovative approach to time-series forecasting that significantly improves upon traditional retrieval-augmented methods. Existing techniques often struggle with complex, non-stationary data patterns and tend to overlook crucial frequency-domain information, which is vital for capturing underlying periodic structures. Additionally, they frequently treat all historical data equally, neglecting the higher relevance of more recent patterns. SpecReTF tackles these challenges by transforming time series into windowed frequency representations. It then measures similarity using a comprehensive metric that considers both amplitude and phase information from the spectral domain. To ensure that recent, more pertinent data is prioritized, the method applies an exponential moving average weighting scheme. Extensive evaluations on various benchmark datasets demonstrate that SpecReTF achieves superior forecasting accuracy compared to conventional time-domain retrieval methods, particularly for diverse and non-stationary time series.