Likelihood-based inference for extreme value model (Q1979087)

The problem of generalized extreme value family parameters estimation is considered. The distributions of this family are defined by \[ \Pr\{\zeta<z\}=F(z)=G(z;\mu,\sigma,\xi) =\exp\left( -\left[ 1+\xi \left((z-\mu)/\sigma \right)^{-1/\xi} \right] \right) \] for \(\{z:\;1+\xi(z-\mu)/\sigma>0\}\), the case of \(\xi=0\) is interpreted as \(\xi\to 0\). The parameters \(\mu\), \(\sigma\), \(\xi\) are estimated by an i.i.d. sample \(z_1\),\dots \(z_n\). Three estimation methods are considered: maximum likelihood (ML), penalized ML (PML) and probability weighted moments (PWM). In PWM the theoretical \(r\)-th probability weighted moments of a variable \(\zeta\) are considered: \[ \beta_r=\beta_r(\mu,\sigma,\xi)=E\zeta(F(\zeta))^r, \;r=0,1,2,\dots \] and estimators of the parameters are obtained by equating \(\beta_r\) with their empirical counterparts \(\tilde\beta_r=n^{-1}\sum_{j=1}^n z_j (\tilde F_n(z_j))^r,\) where \(\tilde F_n(z_j)\) is an empirical d.f. These methods are compared via simulation studies. Results of analysis of annual maximum sea-level data from 58 UK coast sites are described. In these studies a model of the linear trend is considered: \(z_t\sim G(\cdot ;a+bt,\sigma,\xi)\).