Keywords
generalized extreme value distribution, variable selection, extreme rainfall
Start Date
1-7-2010 12:00 AM
Abstract
In this paper, we outline and explore the use of RaVE, a sparse variable selectionmethod that can be used for selecting variables when modelling extreme values. We illustrate itsuse by modelling the location parameter of a series of length 49 of annual rainfall maxima at twostations in North-West Western Australia. The ensemble of potential predictors consists of 1980atmospheric variables. Preliminary results show that RaVE can produce parsimonious yet sensiblemodels. Future work will focus on devising criteria to select the best choice of hyperparametersand on methods to choose groups of spatially contiguous variables so as to aid interpretation.
Fast Variable Selection for Extreme Values
In this paper, we outline and explore the use of RaVE, a sparse variable selectionmethod that can be used for selecting variables when modelling extreme values. We illustrate itsuse by modelling the location parameter of a series of length 49 of annual rainfall maxima at twostations in North-West Western Australia. The ensemble of potential predictors consists of 1980atmospheric variables. Preliminary results show that RaVE can produce parsimonious yet sensiblemodels. Future work will focus on devising criteria to select the best choice of hyperparametersand on methods to choose groups of spatially contiguous variables so as to aid interpretation.
