Penalized quasi-likelihood estimation of generalized Pareto regression -- consistent identification of risk factors for extreme losses (Q2138617)

Consider data exceeding a large threshold, whose distribution is assumed to belong to the maximum domain of attraction of an extreme value distribution. Then the exceedances can be modelled by the generalised Pareto distribution \(F_Y(y;\eta^*, \xi^*) = 1-(1+\xi^* y/\eta^*)^{-1/\xi^*}\) if \(\xi^* \ne 0\), and \(F_Y(y;\eta^*, 0) = 1-\exp\{-y/\eta^*\}\) if \(\xi^* = 0\). The main parameter to be estimated is then the shape parameter \(\xi^*\). It is assumed that the scale Parameter \(\eta^*\) depends on auxiliary parameters \(x \in \mathbb{R}^p\), modelled as a regression \(\log(\eta^*) = x^\top \beta^*\) with the unknown regression parameter \(\beta^*\). One then restricts \(\xi^*\) to the interval \((0,1)\) in order to consider the heavy-tailed case and the first moment exists. It is shown that maximum likelihood estimation works and that the truncation does asymptotically not cause problems. Because usually the number of covariates is large, an information criterion is introduced to penalise too many significant covariates. Under a set of conditions, it is shown that the penalised quasi maximum likelihood estimator is consistent. The theory is illustrated by a simulation study and then applied to motor insurance data.