Adaptive metrics in the nearest neighbours method

We present a modification of the nearest neighbours method in order to improve accuracy of forecast of a time series. This modification is characterized by an adaptive metrics, which is able to adjust its parameters to each segment of a time series. In order to assess the accuracy of the presented method synthetic and real time series were predicted. We show, that this modified metrics is advantageous in prediction in comparison with standard Euclidean metrics or weighted metrics. (c) 2007 Elsevier B.V. All rights reserved.