B2001
Title: Clustering of multivariate nonparametric time trends
Authors: Marina Khismatullina - Erasmus University Rotterdam (Netherlands) [presenting]
Abstract: In this paper, we consider multivariate nonparametric time series and develop a clustering algorithm that allows us to categorize the observed time series into groups that exhibit the same trends. This algorithm is an extension of a multiscale testing approach developed for the comparison of univariate time trends. With the help of this approach, it is possible to formally test the hypothesis that all of the time trends across multiple univariate time series are the same; moreover, it allows us to pinpoint the time regions where the trends are different from each other. The clustering algorithm may be very useful for the practitioners in case of the null hypothesis being rejected: even though some of the trends are different, part of the time series may still exhibit the same trend. The clustering algorithm helps uncover this hidden group structure from the observed time series. We extend the univariate multiscale method to the multivariate nonparametric time trends. With our method, it is possible to find groups of time series which have the same time trend. We show some asymptotic properties of our clustering algorithm, and we illustrate it with an application to the Spanish weather dataset.
