Additive Outlier Detection via Extreme-Value Theory.

This article is concerned with detecting additive outliers using extreme value methods. The test recently proposed for use with possibly non-stationary time series by Perron and Rodriguez [Journal of Time Series Analysis (2003) vol. 24, pp. 193–220], is, as they point out, extremely sensitive to departures from their assumption of Gaussianity, even asymptotically. As an alternative, we investigate the robustness to distributional form of a test based on weighted spacings of the sample order statistics. Difficulties arising from uncertainty about the number of potential outliers are discussed, and a simple algorithm requiring minimal distributional assumptions is proposed and its performance evaluated. The new algorithm has dramatically lower level-inflation in face of departures from Gaussianity than the Perron–Rodriguez test, yet retains good power in the presence of outliers.