Outlier Detection in Multivariate Time Series by Projection Pursuit

Abstract

In this article we use projection pursuit methods to develop a procedure for detecting outliers in a multivariate time series. We show that testing for outliers in some projection directions can be more powerful than testing the multivariate series directly. The optimal directions for detecting outliers are found by numerical optimization of the kurtosis coefficient of the projected series. We propose an iterative procedure to detect and handle multiple outliers based on a univariate search in these optimal directions. In contrast with the existing methods, the proposed procedure can identify outliers without prespecifying a vector ARMA model for the data. The good performance of the proposed method is illustrated in a Monte Carlo study and in a real data analysis.