Statistical downscaling of extreme precipitation events using extreme value theory (Q549636): Difference between revisions

The problem of local probabilistic calibration for daily precipitation forecasts made by global weather forecast models is considered. An approach based on conditional quantiles is discussed. Three methods of conditional quantiles estimation are considered: the semiparametric censored quantile regression (QR), fitting of a non-homogeneous Poisson model (PP), and the peaks-over-non-stationary thresholds model (POT) with the generalized Pareto distribution for tails modeling. The performance of these approaches is compared on the data from German weather stations from 1957 by 2002. The author's conclusion is that PP and POT are appropriate parametric alternatives to the semi-parametric QR approach, especially when very high quantiles are to be forecasted.