Bulletin Français d'Actuariat
Bulletin n°36 / vol. 19 / Janvier 2019 - Juin 2019 Le BFA sur internet
Page précédente - Envoyer ce document


Quantile Mixing and Model Uncertainty Measures

COHIGNAC T.; KAZI-TANI N.


Consulter l'article complet




Abstract

In this paper, we introduce a new simple methodology for combining two models, which are given in the form of two probability distributions. We use convex combination of quantile functions, with weights depending on the quantile level. We choose the weights by comparing, for each quantile level, a given measure of model uncertainty calculated for the two probability distributions that we want to combine. This methodology is particularly useful in insurance and reinsurance of natural disasters, for which there are various physical models available, along with historical data. We apply our procedure to a real portfolio of insurance losses, and show that the model uncertainty measures have a similar behaviour on the set of various insurance losses that we consider. This article serves also as an introduction to the use of model uncertainty measures in actuarial practice.