Learning to Forecast Dynamical Systems from Streaming Data

Kernel analog forecasting (KAF) is a methodology for data-driven, nonparametric of dynamically generated time series data. This approach has rigorous foundation in Koopman operator theory and it produces good forecasts practice, but suffers from the heavy computational costs common to kernel methods. paper proposes streaming algorithm KAF that only requires single pass over training dramatically reduces prediction without sacrificing skill. Computational experiments demonstrate method can successfully forecast several classes dynamical systems (periodic, quasi-periodic, chaotic) both data-scarce data-rich regimes. The overall may have wider interest as new template regression.