-
Couret & Venter hazard group method
- want to assign different credibility by loss type within a class within a hazard group
- within a class, relative freq of injury type is correlated
-
Couret & Venter
- μ(θ)=E(Xj|Θ =θ) = ZX* + (1-Z)μ
- v(θ) = Var(Xj|Θ =θ)
- μ(θ) = aj + sum(ajxj); a0 = (1-z)μ, ai = zn
- z=n / (n+k), k = v/a
- optimal z = VHM / (EVPV/n + VHM) = n / (n + EVPV/VHM)
-
Testing Couret & Venter's model
- first glance: disappointing compared to just using HG prediction
- reason 1: influenced by data too much to match held back data
- reason 2: class data by year is very volatile. some covariances where not accounted for
- quintiles test: apply model to all classes, sort and split into 5 class grp of roughly equal size. There was a correlation btwn odd & even years suggesting that the model adds info
- exception: hazard group A is already very homogeneous so the credibility procedure does not provide any add'l info
|
|