Quantification of Uncertainty in Return Values for Extreme Precipitation Events in the Western US

When observations of extreme precipitation events are not plentiful computer simulations are used to model these extremes. When high computing costs limit the simulations performed, extreme value theory is sometimes implemented to extrapolate the distribution of simulated data to its extremes. However, it remains unclear how limited a sample size can be extrapolated to estimate extreme precipitation events with a robust degree of certainty. We tackle this using several thousand years worth of daily data simulated by a seasonal- forecast-resolution climate model capable of capturing typical synoptic conditions associated with extreme precipitation events, and focus on differing climatic regions of the western U.S. We first empirically estimate the value of the extreme events, and associated uncertainty bounds, from the simulations before comparing to results from applying extreme value theory to a range of limited sub-samples of these simulations. We find that the more extreme an event is, the larger the sample size required to robustly estimate the associated uncertainty bounds, but optimal sample sizes ranged from 50 to 170 simulations.