Ferreira and Canto e Castro [6] introduces a power max-autoregressive process, in short pARMAX, as an alternative
to heavy tailed ARMA. An extension of pARMAX was considered
in Ferreira and Canto e Castro [7], by including a random
component, and hence called pRARMAX, which makes the model more flexible to applications. It was then developed a
methodology settled on minimizing the Bayes risk in classification theory, but only considering standard uniform random components. We now extend this procedure to the more general Beta distribution. We illustrate the method with an application to a financial data series. In order to improve estimates of the exceedance probabilities of levels of interest, we use Bortot and Tawn [2] approach and derive a threshold-dependent extremal index which relates with the coefficient of tail dependence of Ledford
and Tawn [8] and with the pRARMAX parameter.

CEMAT - Center for Computational and Stochastic Mathematics