A Nearest Trajectory Strategy for Time Series Prediction
James McNames · 2000
Abstract-A method of local modeling for predicting time series generated by nonlinear dynamic systems is proposed that incorporates a weighted Euclidean metric and a novel ρ-steps ahead crossvalidation error to assess model accuracy. The tradeoff between the cost of computation and model accuracy is discussed in the context of optimizing model parameters. A fast nearest neighbor algorithm and a novel modification to find neighboring trajectory segments are described. I.