A functional data based method for time series classification

We propose using the integrated periodogram to classify time series. The method assigns a new element to the group minimizing the distance from the integrated periodogram of the element to the group mean of integrated periodograms. Local computation of these periodograms allows the application of the approach to nonstationary time series. Since the integrated periodograms are functional data, we apply depth-based techniques to make the classification robust. The method provides small error rates with both simulated and real data, and shows good computational behaviour.