Thursday 23 Mar 2017: Statistical Science Seminar: Extreme Value Threshold Estimation and Mixture Modelling
Carl Scarrott - University of Canterbury, NZ
Extreme value theory is used to develop asymptotically motivated models for describing the likelihood of rare event occurrence. Such models are typically used to approximate the behaviour of the tail(s) of the population distribution. An important challenge in the application of such extreme value models is the choice of a threshold, beyond which point the extreme value tail models can provide reliable extrapolation.
Various approaches have been developed to aid the selection of this threshold, each with their own advantages and disadvantages. Extreme value mixture models are one such approach, combining a suitable model for the bulk of the distribution below the threshold along with an extreme value tail model above. The threshold is typically treated as a parameter to be inferred, thus permitting both estimation and uncertainty quantification. These models potentially provide a more objective approach to threshold choice.
We have created an R package called "evmix", available on CRAN, which implements most of the existing mixture models in the literature. This talk will place most of these models in a generalised framework, discuss their features, demonstrate usage of the package and provide advice for developers and users of such models based on recent simulation studies evaluating their performance.