01.06.2015 Views

Actuarial Modelling of Claim Counts Risk Classification, Credibility ...

Actuarial Modelling of Claim Counts Risk Classification, Credibility ...

Actuarial Modelling of Claim Counts Risk Classification, Credibility ...

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Efficiency and Bonus Hunger 229<br />

Application to the <strong>Claim</strong> Costs Recorded in Portfolio C<br />

The initial values obtained with the the moment method applied to large claims <strong>of</strong> Portfolio<br />

C are ̂ 0 = 0319 and ̂ 0 = 180 1767. The maximum likelihood estimates <strong>of</strong> the Generalized<br />

Pareto parameters are ̂ = 04152 and ̂ = 156 7374. Different starting values have been<br />

used and the convergence always occurred.<br />

Note that the limited sample size for the large losses does not allow us to draw reliable<br />

conclusions about major claims (in particular, large sample properties <strong>of</strong> the maximum<br />

likelihood estimators cannot be invoked with a sample size as small as 17). In practice, the<br />

insurance company must gather the large losses together in a data base to perform detailed<br />

analysis. The amounts <strong>of</strong> these large losses have to be corrected for different sources <strong>of</strong><br />

inflation. The assistance <strong>of</strong> a reinsurance company is useful in this respect, especially for<br />

insurers with small to moderate portfolios.<br />

5.2.4 <strong>Modelling</strong> the Number <strong>of</strong> Large <strong>Claim</strong>s<br />

The number <strong>of</strong> large claims N large<br />

i for policyholder i is modelled using the oi large<br />

i <br />

distribution. This is in line with the fact that large claims occur purely at random. Poisson<br />

regression is then used to incorporate the information available about policyholder i,<br />

summarized in a vector xi<br />

T = x i1 x ip consisting <strong>of</strong> explanatory variables (assumed<br />

here to be categorical and coded by means <strong>of</strong> binary variables), in the expected frequency<br />

<strong>of</strong> large claims through a linear predictor by means <strong>of</strong> an exponential link function.<br />

The regression coefficients are estimated with Poisson regression. All the explanatory<br />

variables introduced in Section 5.1.6 have been excluded from the model. Only the intercept<br />

remained in the Poisson regression model. This is not surprising with such a small number <strong>of</strong><br />

policies producing a large claim. We obtained ̂ 0 =−90555, with a standard deviation equal<br />

to 0.2425. The resulting frequency <strong>of</strong> large claims is<br />

̂<br />

large<br />

i<br />

= 00117 % for all policyholders.<br />

Remark 5.1 (Logistic regression) In most cases, policyholders report 0 or just 1 large<br />

claim. Therefore, the number <strong>of</strong> large claims could also be modelled with a binary variable<br />

instead <strong>of</strong> a Poisson count. The probability that policyholder i reports (at least) a large claim<br />

can also be modelled with the help <strong>of</strong> logistic regression. Specifically, let us define the<br />

binary random variable J i<br />

{ 0 if policyholder i does not report any large claim during the observation period<br />

J i =<br />

1 if policyholder i reports at least one large claim during the observation period<br />

The aim is to model the probability PrJ i = 0 = q i x i that policyholder i does not report<br />

any large claim during the coverage period.<br />

Since q i x i ∈ 0 1, we resort to some distribution function F to link q i x i to the linear<br />

predictor 0 + ∑ p<br />

j=1 jx ij , that is<br />

(<br />

)<br />

p∑<br />

p∑<br />

q i x i = F 0 + j x ij ⇔ 0 + j x ij = F −1 q i x i <br />

j=1<br />

j=1

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!