Berlin 2015 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 24: Extreme Events (joint session DY / SOE)
SOE 24.5: Talk
Thursday, March 19, 2015, 16:15–16:30, BH-N 243
A data-adaptive definition of extreme events in time series exhibiting seasonality — Eva K. Hauber1,2,3 and •Reik V. Donner1 — 1Potsdam Intsitute for Climate Impact Research, Potsdam, Germany — 2University of Copenhagen, Denmark — 3University of Natural Resources and Life Sciences, Vienna, Austria
Environmental time series are often characterized by strong seasonal variations. In such a case, extreme events are traditionally defined by removing the underlying seasonal component in the mean and applying a threshold-based definition of an extreme to the residuals. However, this approach is only valid if the probability distribution function (PDF) shows a seasonal modulation exclusively of its mean. In turn, real-world climatological records exhibit heteroskedasticity, implying that their variance (but often also the shape of the PDF) changes over the year as well. Here, we present a data-adaptive method that allows defining extreme events under such conditions. Our approach is based on kernel estimates of the conditional PDF of the data taking the phase during the year as a covariate, which allow estimating any given quantile of the PDF as a function of this phase. We demonstrate the capabilities of this new approach for artificial time series as well as real-world observational data. Our results indicate that even for short time series covering only a few periods, the data-adaptive method leads to a systematic reduction of false identifications of extremes in comparison to standard techniques.