Smooth-Transition Regression Models for Non-Stationary Extremes

Abstract We introduce a smooth-transition generalized Pareto (GP) regression model to study the time-varying dependence structure between extreme losses and set of economic factors. In this model, distribution loss size is approximated by GP distribution, its parameters are related explanatory variables through functions, which themselves depend on predictor structural changes. use approach dynamics in monthly severity operational at major European bank. Using VIX as transition variable, our analysis reveals that when uncertainty high, high number recent past indicative less future, consistent with self-inhibition hypothesis. On contrary, times low uncertainty, only growth rate economy seems be relevant likelihood losses.