Detecting outliers in time series

In parametric time series analysis there is the implicit assumption of no aberrant observations, so-called outliers. Outliers are observations that seem to be inconsistent with the assumed model. When these observations are included to estimate the model parameters, the resulting estimates are biased. The fact that markets have been affected by shocks (i.e. East Asian crisis, Dot-com bubble, sub-prime mortgage crisis) make the assumption that no outlier is present questionable. This paper addresses the problem of detecting outlying observations in time series. Outliers can be understood as a short transient change of the underlying parameters. Unfortunately tests designed to detect structural breaks cannot be used to find outlying observations. To overcome this problem a test normally used to detect structural breaks is modified. This test is based on the cumulative sum (CUSUM) of the squared observations. In comparison to a likelihood-ratio test neither the underlying model nor the functional form of the outliers have to be specified. In a simulation study the finite sample behaviour of the proposed test is analysed. The simulation study shows that the test has reasonable power against a variety of alternatives. Moreover, to illustrate the behaviour of the proposed test we analyse the returns of the Volkswagen stock.