A neural-network extension of the method of analogues for iterated time series prediction

We describe an algorithm for nonlinear iterated prediction of time series based on a neural network extension of the method of analogues proposed by Lorenz (1969). The present method is investigated in the context of iterated time series forecasting using embeddings of a nonlinear dynamical system. The approach yields significantly better results than published work on some of the Santa Fe competition data sets. The proposed technique is demonstrated by an application to a real world time series data of electricity load demand