Survey_MY_2008

Abstract

In doing this, we used the multivariate normal latent-variable model as benchmark model to which we calibrated the other models. Throughout the simulations, we assumed n = 1 000 loss processes, to match with the studies in Frey et al. (2001) and Frey and McNeil (2001). For the mixture models, we simulated a factor realization ? and calculated conditional occurrence probabilities p(?) which were then used to conduct n Bernoulli or Poisson trials. After summing up the number of event occurrences, we repeated 100 000 times and calculated VaR and ES of the resulting empirical distribution. For the common Poisson-shock model, we assumed nc = 1 and calculated ?,?? and ?cj from ? and ?Y . These quantities were then used to conduct Poisson trials and proceed further as in the mixture model setup.