Thresholding Using Extreme Value Theory Threshold Models

Automated techniques of anomaly and outlier detection usually rely on thresholding. The quality of the method performance often depends heavily on the threshold selection. In this paper, assuming that the signal values are bounded, we propose a simple, practical method of threshold determination based on signal maxima sampling. It can be shown that only modeling block maxima is a wasteful approach to extreme value determination if other data on extremes are available. In our contribution, we reformulate the method of threshold selection based on blocks into a method using so-called threshold models. Threshold models rely on extracting the peak values from signal values which exceed a certain threshold. The experimental results confirm the performance of the method.