REFERENCES:y ´Risk Classification In Non Life Insuranceµ Paper

presented by Katrien Antonio and Jan Beirlant on November 28,2006

KEYWORDS:y Non Life Insurance y Risk Classification y Generalized Linear Models y Longitudinal Data y Mixed Model y Predicted y Bayesian Statistics y Heavy Tailed Regression

Main Aim:To know the construction of fair Tariff Structure means Premium charge from policyholder

Main Idea behind Risk Classification is to split an insurance portfolio into classes that consists:Risks with a similar profile. To design fair tariff for each of them.


Priori Rating System

Posteriori Rating System

Regression Model For A Priori Risk Classification
y To build a tariff that reflects the various risk

profiles in a portfolio in a reasonable way, we rely on regression techniques. y There are two Typical Response variables:y Number of claims y Its corresponding severity

For Priori Risk Classification we use two approach:yGeneralized Linear Models. yFlexible, Parametric Families of

Distribution and Regression.

Generalized Linear Models(GLM)
y GLMs extend the framework of general(normal) linear models to the class of distribution from exponential family. y A whole variety of possible outcome measures can be modelled within this framework. y We uses canonical form specification of densities from the exponential family

f(y)=exp({[y - ( )]\ }+c(y, ))
where y (:) and c(:) are known functions, y is the natural and y the scale parameter

Parametric Families of Distribution and Regression.
y In this we model severity of claims as a

function of their risk characteristics might require statistical distributions outside the exponential family. y We use distribution with heavy tail. y We will use BURR XII and GB2(Generalized Beta of second kind) Distribution.



y Cumulative Distribution Function for the BURR

XII and GB2 distribution are given by

FBurr,Y(y)= 1 [ \( +y^ )]^ y>0 , , , >0 FGB2,Y(y)= B {[(y/b)^a]/[1+(y/b)^a]; 1, 2} y>0, a 0,b, 1 , 2>0
Where B(«) is the incomplete Beta Function

ILLUSTRATION 2:Fire Insurance Portfolio
y Consider a fire insurance portfolio which consists

of 1823 observations. y We want to assess how the loss distribution changes with sum insured and the type of building. y Claims expressed as a fraction of the sum insured are used as the response. y Explanatory variables are the type of building and the sum insured. y Parameters are estimated with maximum likelihood.

Residual QQplots can be used to judge the goodnessof-fit of the proposed regression models

Figure 1: Fire Insurance Portfolio: residual QQplots for Burr and GB2 regression.

A Posteriori Ratemaking
y To update the priori tariff or to predict

future claims when historical claims are available y We uses statistical modes for longitudinal or panel data(i.e., data observed on a group of subjects) y These statistical models are known as credibility models (mixed or randomeffects models)

Credibility Ratemaking problem is concerned determination of a risk premium that combines y Observed risks y Individual claims experience of a risk and y The experience regarding related risks

ILLUSTRATION:Workers· Compensation Insurance
y 133 occupation or risk classes are followed over a

period of 7 years.
y Frequency counts in workers· compensation

insurance are observed on yearly basis.
y Explanatory variable are Year and Payroll.

The following models are considered:Yij|bi ~ Poisson( ij)

log( ij) = log( Payrollij ) +



Yearij + bi,0

versus log(

= log( Payrollij ) + + bi,1 Yearij


Yearij + bi,0

Hereby y Yij represents the jth measurement on the ith subject of the response Count. and 1 are fixed effects and y bi,0, versus bi,1, is a risk class specific intercept, versus slope. y It is assumed that bi = (bi,0; bi,1) ~N(0;D)

y The results of both a maximum likelihood and a Bayesian analysis are given in Table
PQL Est. Model (7) 0 1 0 Model (8) 0 1 0 1 0;1 SE adaptive G-H Est. SE Bayesian Mean 90%Cred. Int.

-3.529 0.083 -3.557 0.084 0.01 0.005 0.01 0.005 0.790 0.110 0.807 0.114 -3.532 0.009 0.790 0.006 / 0.083 -3.565 0.084 0.011 0.009 0.011 0.111 0.810 0.115 0.002 0.006 0.002 / 0.001 0.01

-3.565 0.01 0.825 -3.585 0.008 0.834 0.024 0.006

(-3.704, -3.428) (0.001, 0.018) (0.648, 1.034) (-3.726, -3.445) (-0.02, 0.04) (0.658, 1.047) (0.018, 0.032) (-0.021, 0.034)

Table 1: Workers' compensation data (frequencies): results of maximum likelihood and Bayesian analysis. 0 = Var(bi;0), 1 = Var(bi;1) and 0;1 = 1;0 = Cov(bi;0; bi;1)

Given Table compares the predictions for some selected risk classes with the observed values.
Actual Values PQL Class 11 20 70 89 112 Payrolli7 Counti7 230 1,315 54.81 79.63 18,810 8 22 0 40 45 Mean s.e. Expected number of claims adaptive G-H Bayesian Mean 11.296 33.396 0.361 47.628 33.191 s.e. Mean 90%Cred.Int.

11.294 2.726 33.386 4.109 0.373 0.23 47.558 5.903 33.278 4.842

2.728 12.18 (5,21) 4.12 32.63 (22,45) 0.23 0.416 (0,2) 6.023 50.18 (35,67) 4.931 32.66 (21,46)

Table 2: Workers' compensation data (frequencies): predictions for selected risk classes.

y Predictive distributions obtained with a Bayesian

analysis are illustrated in Figure .

Figure 2: Workers Compensation Insurance (Counts): predictive distributions for selection of risk classes.