ASSESSMENT OF PEAKS OVER THRESHOLD METHODS FOR ESTIMATING EXTREME-VALUE DISTRIBUTION TAILS

In the past twenty years a vast new body of extreme value theory was d eveloped, referred to as 'peaks over threshold modeling.' This theory allows the use in the analysis of all data exceeding a sufficiently hi gh threshold, a feature that may result in improved extreme value esti mates. The application of the theory depends upon the performance of m ethods for estimating the distribution parameters Corresponding to any given set of extreme data. We present a comparative assessment of the performance of three such methods. The assessment is based on Monte C arlo simulations from populations with four distributions: Gumbel, Wei bull, generalized Pareto, and normal. The simulation results showed th at the de Haan and the Conditional Mean Exceedance (CME) methods perfo rmed consistently better than the Pickands method (NIST implementation ). For the distributions, parameter values, and mean recurrence interv als assumed in this work, the CME method outperformed the de Haan meth od only when the percent estimation errors were about one percent or s maller, a case unlikely to be encountered in wind engineering practice .