Evolving nearest neighbor time series forecasters

This article proposes a nearest neighbors—differential evolution (NNDE) short-term forecasting technique. The values for the parameters time delay \(\tau \), embedding dimension m, and neighborhood size \(\epsilon \), for nearest neighbors forecasting, are optimized using differential evolution. The advantages of nearest neighbors with respect to popular approaches such as ARIMA and artificial neural networks are the capability of dealing properly with nonlinear and chaotic time series. We propose an optimization scheme based on differential evolution for finding a good approximation to the optimal parameter values. Our optimized nearest neighbors method is compared with its deterministic version, demonstrating superior performance with respect to it and the classical algorithms; this comparison is performed using a set of four synthetic chaotic time series and four market stocks time series. We also tested NNDE in noisy scenarios, where deterministic methods are not capable to produce well-approximated models. NNDE outperforms the other approaches.