|
Abstract:
|
We propose a new flexible approach, based on finite mixture models, for modeling right censored loss insurance data. The most general case of partial right censoring is considered in presence of multiple censoring points in the right tail of the distribution representing the situation when losses are aggregated from different coverage corresponding to different limits. A special case of right censoring is also considered for a line of business in presence of a single policy limit. We will consider non-Gaussian parametric families of distributions that are suitable for modeling left and right skewed data on a positive domain. The expectation-maximization algorithm is employed for the estimation of parameters as well as the number of components, K, for each considered mixture model. Model selection is performed based on Akaike and Bayesian Information Criterions. A simulation study is designed to validate the proposed approach. The proposed modeling approach is illustrated on two real data sets previously considered in the literature. The methodology considered in this paper brings valuable contributions in the area of predictive insurance loss modeling.
|
