About the Justification of Experience Rating: Bonus Malus System and a new Poisson Mixture Model

TL;DR

This paper explores alternative mixing distributions in Poisson models for insurance claim data, introducing an inverse-gamma distribution to better capture tail behavior and risk, and proposes a new concept called resolution for risk classification.

Contribution

It introduces the inverse-gamma mixing distribution for Poisson models in insurance, providing closed-form solutions and analyzing tail behavior, enhancing risk assessment methods.

Findings

The inverse-gamma mixture distribution has a closed form involving Bessel functions.

The tail behavior of the inverse-gamma mixture differs significantly from the gamma mixture.

The concept of resolution helps evaluate the adequacy of risk group classifications.

Abstract

The claim experience of the past is a very important information to calculate the fair price of an insurance contract. In a lot of European countries for instance the prices for motor car insurance depend on the number of claims the driver has reported to the insurance company during the last years. Classically these prices are calculated on the basis of a mixed Poisson model with a gamma mixing distribution. The mixing distribution models the car drivers' qualities across the insured portfolio. This is just one example for experience rating. In the classical context the price is equal to the expectation of the Bayesian posterior distribution. In some lines of business (especially third party liability and lines with exposure to extreme weather events) we that the real world data cannot be described well enough by the classical Poisson - gamma model. Therefore we investigate the influence of the mixing distribution on the posterior distribution conditional on the experienced number of claims. This enables the application of other - more risk adequate premium principles than the expectation principle. We introduce the inverse - gamma distribution as a new mixing distribution to model claim numbers and compare it to the classical gamma distribution. In both cases a closed analytic representation of the mixed distribution can be found: In the classic case the well known negative binomial distribution, in our new one a representation using the Bessel functions. Additionally we present numerical results about the tail behaviour of the mixed Poisson - inverse - gamma distribution. Finally we introduce the concept of resolution. It enables us to decide if the classification of risk groups via the number of experienced claims is a risk adequate procedure.