126

Informatica Economic, nr. 2 (42)/2007

More General Credibility Models

Virginia ATANASIU Department of Mathematics, Academy of Economic Studies e-mail: virginia_atanasiu@yahoo.com This communication gives some extensions of the original Bhlmann model. The paper is devoted to semi-linear credibility, where one examines functions of the random variables representing claim amounts, rather than the claim amounts themselves. The main purpose of semi-linear credibility theory is the estimation of 0 ( ) = E [ f 0 ( X t +1 ) ] (the net premium for a contract with risk parameter: ) by a linear combination of given functions of the observ' able variables: X = ( X 1 , X 2 ,..., X t ) . So the estimators mainly considered here are linear functions of several functions f1 , f 2 ,..., f n of the observable random variables. The approximation to 0 ( ) based on prescribed approximating functions f1 , f 2 ,..., f n leads to the optimal non-homogeneous linearized estimator for the semi-linear credibility model. Also we discuss the case when taking f p = f for all: p , try to find the optimal function f . It should be

noted that the approximation to 0 ( ) based on a unique optimal approximating function f is always better than the one furnished in the semi-linear credibility model based on prescribed approximating functions: f1 , f 2 ,..., f n . The usefulness of the latter approximation is that it is easy to apply, since it is sufficient to know estimates for the structural parameters appearing in the credibility factors. From this reason we give some unbiased estimators for the structure parameters. For this purpose we embed the contract in a collective of contracts, all providing independent information on the structure distribution. We close this paper by giving the semi-linear hierarchical model used in the applications chapter. Mathematics Subject Classification: 62P05. Keywords: contracts, unbiased estimators, structure parameters, several approximating functions, semi-linear credibility theory, unique optimal function, parameter estimation, hierarchical semi-linear credibility theory. ntroduction In this article we first give the semi-linear credibility model (see Section 1), which involves only one isolated contract. Our problem (from Section1) is the estimation of 0 ( ) = E[ f 0 ( X t +1 ) ] (the net premium for a

contract with risk parameter: ) by a linear combination of given functions f1 , f 2 ,..., f n solution of this problem: of the observable variables: 2 n t Min E 0 ( ) 0 pr f p ( X r ) , where: = ( pr ) p ,r , 0 , p =1 r =1 is the optimal non-homogeneous linearized the unique optimal function f . It should be estimator (namely the semi-linear credibility noted that the approximation to ( ) based 0 result). In Section 2 we discuss the case on a unique optimal approximating function when taking f p = f for all: p , try to find f is always better than the one furnished in

0 ( ) = E [ f 0 ( X t +1 ) ] in the least squares sense, where is the structure variable. The

X = ( X 1 , X 2 ,..., X t ) . So our problem (from Section 1) is the determination of the linear combination of 1 and the random variables: closest to p = 1, n , r = 1, t f p (X r ) ,
'

Informatica Economic, nr. 2 (42)/2007

127

the semi-linear credibility model based on prescribed approximating functions: f1 , f 2 ,..., f n . The usefulness of the latter approximation is that it is easy to apply, since it is sufficient to know estimates for the structural parameters: a pq , b pq (with p ,
q = 0, n ) appearing in the credibility factors

z p (where p = 1, n ). To obtain estimates for

these structure parameters from the semilinear credibility model, in Section 3 we embed the contract in a collective of contracts, all providing independent information on the structure distribution. We close this paper by giving the semi-linear hierarchical model used in the applications chapter (see Section 4). Section 1 (The approximation to 0 ( ) based on prescribed approximating functions: f1 , f 2 ,..., f n ) We use the notation: p ( ) = E f p ( X r ) |

In this section, we consider one contract with unknown and fixed risk parameter: , during a period of t years. The yearly claim amounts are denoted by: X 1 ,..., X t . The risk parameter is supposed to be drawn from some structure distribution function: U () . It is assumed that, for given: , the claims are conditionally independent and identically distributed (conditionally i.i.d.) with known common distribution function FX ( x, ) . The random variables X 1 ,..., X t are observable, and the random variable X t +1 is considered as being not (yet) observable. We assume that: f p ( X r ) , p = 0, n , r = 1, t + 1 have finite variance. For: f 0 , we take the function of X t +1 we want to forecast.

(p = 0, n; r = 1, t + 1)
a pq

(1.1)

This expression does not depend on r. We define the following structure parameters: m p = E p ( ) = E {E f p ( X r ) | } = E f p ( X r )
b pq c pq d pq

[ ] [ ] = E {Cov [ f ( X ), f ( X ) | ]} = Cov [ ( ), ( )] = Cov[ f ( X ), f ( X )] = Cov [ f ( X ), ( )]

p r q r
p q

(1.2), (1.3), (1.4), (1.5), (1.6),

for: p , q = 0, n r = 1, t + 1 . These expressions do not depend on: r = 1, t + 1 . The structure parameters are connected by the following relations: c pq = a pq + b pq
d pq = b pq

for: p, q = 0, n . This follows from the covariance relations obtained in the probability theory where they are very well-known. Just as 0 ( ) = E [ f 0 ( X t +1 ) | ] and to f 0 ( X t +1 ) in in the case of considering linear combina- the least squares sense equals: tions of the observable variables themselves, n t n 1 M = z p f p ( X r ) + m0 z p m p (1.9), p =1 r =1 t p =1

we can also obtain non-homogeneous credibility estimates, taking as estimators the class of linear combinations of given functions of the observable variables, as shown in the following theorem: (1.7), linTheorem 1.1 (Optimal non-homogeneous earized estimators) (1.8), The linear combination of 1 and the random variables f p ( X r ), p = 1, n; r = 1, t closest to

128

Informatica Economic, nr. 2 (42)/2007

where z1 , z 2 ,..., z n is a solution to the linear system of equations:

[c
n p =1 n

pq

+ (t 1)d pq z p = td 0 q ( q = 1, n ) + tb pq )z p = tb0 q ( q = 1, n )

(1.10)

or to the equivalent linear system of equations:

(a
p =1

pq

(1.11)

Proof: we have to examine the solution of the problem: 2 n t Min E 0 ( ) 0 pr f p ( X r ) 0 , p =1 r =1 Taking the derivative with respect to 0 gives:
n t

(1.12)

E [ 0 ( )] pr E f p ( X r ) = 0 , or: 0 = m0 pr m p . Inserting this expression for

p =1 r =1 p =1 r =1

0 into (1.12) leads to the following problem:

2 n t (1.13) Min E 0 ( ) m0 pr ( f p ( X r ) m p ) 1 1 = = p r On putting the derivatives with respect to qr ' equal to zero, we get the following system of

equations ( q = 1, n ; r ' = 1, t ):

Cov 0 ( ), f q ( X r ' ) = pr Cov f p ( X r ), f q ( X r ' )

p =1 r =1

(1.14)

Because of the symmetry in time clearly: p1 = p 2 = ... = pt = p , so using the covariance results, for q = 1, n this system of equations can be written as:

Let us forget now about this structure of f and look for any function f such that (2.1) is closest to: 0 ( ) . If are considered only functions f such that f ( X 1 ) has finite variance, then the optimal approximating func(1.15) tion f results from the following theorem: Theorem 2.1 (Optimal approximating function) f ( X 1 ) + ... + f ( X t ) is closest to 0 ( ) and to f 0 ( X t +1 ) in the least squares sense, if and only if f is a solution of the equation:

b0 q = p c pq + (t 1)d pq
p =1

Now (1.15) and (1.13) lead to (1.9) with: zp p = , p = 1, n . t Section 2 (The approximation to 0 ( ) based on a unique optimal approximating function: f ) The estimator M for 0 ( ) of Theorem 1.1 can be displayed as: M = f ( X 1 ) + ... + f ( X t ) (2.1), where: 1 n 1 1 n f ( x ) = z p f p ( x ) + m0 z p m p . t p =1 t t p =1

f ( X 1 ) + (t 1)E [ f ( X 2 ) X 1 ] E [ f 0 ( X 2 ) X 1 ] 0 (2.2) Proof: we have to solve the following minimization problem: 2 Min E [ f 0 ( X t +1 ) g ( X 1 ) ... g ( X t )] (2.3)
g

Suppose that f denotes the solution to this problem, then we consider: g ( X ) = f ( X ) + h( X ) , with h() arbitrary, like in variational calculus. Let:

Informatica Economic, nr. 2 (42)/2007

129

( ) = E [ f 0 ( X t +1 ) f ( X 1 ) ... f ( X t ) h( X 1 ) ... h( X t )]2 Clearly for f to be optimal, ' (0) = 0 , so for every choice of h :
E{[ f 0 ( X t +1 ) f ( X 1 ) ... f ( X t )][h( X 1 ) + ... + h( X t )]} = 0 must hold. This can be rewritten as: E [tf 0 ( X 2 )h( X 1 ) tf ( X 1 )h( X 1 ) t (t 1) f ( X 2 )h( X 1 )] = 0 or: E [h( X 1 ){ f ( X 1 ) (t 1)E[ f ( X 2 ) X 1 ] + E[ f 0 ( X 2 ) X 1 ]}] = 0 Because this equation has to be satisfied for every choice of the function h one obtains, the expression in brackets in (2.7) must be identical to zero, which proves (2.2). An application of Theorem 2.1: If X 1 ,..., X t +1 can only take the values 0,1,..., n and p qr = P[X 1 = q, X 2 = r ] for: q ,
f (q ) p qr + (t 1) f (r ) p qr = f 0 (r ) p qr
r =0 r =0 r =0 n n n

(2.4) (2.5), (2.6), (2.7)

sense, if and only if for q = 0, n , f (q ) is a solution of the linear system:

to 0 ( ) and to f 0 ( X t +1 ) in the least squares

r = 0, n , then f ( X 1 ) + ... + f ( X t ) is closest

(2.8)

f (q ) n f (q ) n = Indeed: f ( X 1 ) : ; [ ] , q = 0 , n ( ) E f X X = f (r )P ( X 2 = r X 1 = 2 1 P( X = q ) p qr r =0 1 r =0 n n n p qr p qr = q ) = f (r ) n ; E [ f 0 ( X 2 ) X 1 ] = f 0 (r )P( X 2 = r X 1 = q ) = f 0 (r ) n . r =0 r =0 r =0 pqr pqr

r =0

E [ f ( X 2 ) X 1 ] and E [ f 0 ( X 2 ) X 1 ] into (2.2) leads to (2.8).

Section 3 (Parameter estimation) It should be noted that the approximation to 0 ( ) based on a unique optimal approximating function f is always better than the one furnished in Section 1 based on prescribed approximating functions: f1 , f 2 ,..., f n . The usefulness of the latter approximation is that it is easy to apply, since it is sufficient to know estimates for the structural parameters a pq , b pq (with p , q = 0, n )

Inserting these expressions for:

f (X 1 ),

r =0

ters. For this purpose we consider k contracts, j = 1, k , and k ( 2 ) independent and identically distributed vectors ' j , X j = ( j , X j1 ,..., X jt ), for j = 1, k . The contract indexed j is a random vector consisting of a random structure parameter j and

observations: X j1 ,..., X jt , where j = 1, k . For every contract j = 1, k and for j fixed, the variables: X j1 ,..., X jt are conditionally independent and identically distributed. Theorem 3.1 (Unbiased estimators for the structure parameters) Let: ^ 1 1 k t m p = X ..p = f p ( X jr ) (3.1) kt kt j =1 r =1

appearing in the credibility factors z p (where

p = 1, n ). From this reason we give some unbiased estimators for the structure paramek t ^ 1 p 1 p q 1 q a pq = X jr X j . X jr X j . k (t 1) j =1 r =1 t t

(3.2)

130

Informatica Economic, nr. 2 (42)/2007

1 k 1 p 1 p 1 q 1 q a pq (3.3), b pq = X j . X .. X j. kt X .. k 1 j =1 t kt t t t t ^ ^ ^ p , then: E m p = m p , E a pq = a pq , E b pq = b pq , where: X jp. = X jr , Xq = Xq j. jr , r =1 r =1

^

p X ..p = X jr , j =1 r =1

X ..q = X q jr ,
j =1 r =1

with

p X jr = f p (X jr ),

( j = 1, k

and

r = 1, t ),

Xq jr = f q (X jr ) , ( j = 1, k and r = 1, t ), for p , q = 0, n , such that p < q .

Proof: note that the usual definitions of the structure parameters apply, with j replacing

kt 1 ^ 1 and X jr replacing X r so: E m p = E f p (X jr ) = m p = m p = m p ; kt j ,r kt kt j ,r 1 1 q p 1 q ^ p p q p E a pq = Cov X jr ,Xq jr + E X jr E X jr Cov X jr , X j . E X jr E X j . t k (t 1) j ,r t

[ (

) ( )( )

( )

1 1 p 1 q 1 p 1 q 1 p 1 q Cov X jp. , X q jr E X j . E ( X jr ) + Cov X j . , X j . + E X j . E X j . = t t t t k (t 1) t t 1 1 1 (a pq + b pq ) + m p mq a pq + b pq m p mq a pq + b pq m p mq + a pq + b pq + m p mq t t t j ,r (t 1) a = a ; E b^ = 1 1 1 1 1 kt Cov = a pq + b pq a pq b pq = pq pq pq k (t 1) j ,r t t k 1 j t k (t 1)

1 1 1 1 1 p 1 q 1 p 1 q X .. E X jp. E X ..q Cov X ..p , X jp. , X q j . + E X j . E X j . Cov X j . , t kt kt t kt t t t 1 1 1 1 p 1 q 1 1 1 a pq X .. E X j . + Cov X ..p , X ..q + E X ..p E X ..q , Xq = j. E t t kt k 1 kt t kt kt kt 1 1 1 1 1 1 a pq + b pq + m p mq a pq + b pq m p mq a pq + b pq m p mq + a pq + k k kt kt kt j t a pq 1 1 1 1 1 k 1 1 a pq + b pq + m p mq = = k b pq + a pq + b pq a pq b pq t t k 1 j t kt k k k 1 k a pq a pq a pq 1 k 1 + k a pq = b pq + = b pq . t t t k 1 kt

Section 4 (Applications of semi-linear credibility theory) We close this paper by giving the semilinear hierarchical model used in the applications chapter. Like in Jewells hierarchical model we consider a portfolio of contracts, which can be broken up into P sectors each sector p consisting of k p groups of con-

( p ) = E X p , j ,t +1 p (the pure net risk premium of the sector p ), we now estimate: f 0 (X p , j ,t +1 ) ,

0 p , p j = E f 0 (X p , j ,t +1 ) p , p j (the pure

) [
[

0 ( p ) = E f 0 (X p , j ,t +1 ) p (the pure net risk

premium of the sector p ), where p = 1, P

net risk premium of the contract

( p, j ) ),

tracts.

p , p j = E X p , j ,t +1 p , p j (the pure net

risk premium of the contract

Instead

) [

of

estimating:

X p , j ,t +1 ,

( p, j ) ),

and j = 1, k p . In semi-linear credibility theory the following class of estimators is con-

Informatica Economic, nr. 2 (42)/2007

131

sidered: 0 + pqir f p (X qir ),

n P t

kq

where f 1 (),..., f n () are functions given in advance. Let us consider the case of one given function f1 in order to approximate f 0 (X p , j ,t +1 ) or 0 ( p ) and 0 p , p j . We

p =1 q =1 i =1 r =1

( ) D = Cov(X , X ) , j j ' , C = Cov(X , X ) = Var (X ).

0 D01 = Cov X pjw ,X1 pj ' w , 11 1 pjw 1 pj ' w 1 pjw 1 pjw 1 pjw

11

Remark 4.1: the linear combination of 1 and the random variables X 1 ( p = 1, P , pjr

formulate the following theorem: Theorem 4.1 (Hierarchical semi-linear credibility) Using the same notations as introduced for the hierarchical model of Jewell and denoting 0 X pjs = f 0 (X pjs ) and X 1 pjs = f 1 ( X pjs ) one obtains the following least squares estimates for the pure net risk premiums:

to 0 ( p ) in the least squares sense equals

j = 1, k p , r = 1, t ) closest to f 0 (X p , j ,t +1 ) and
^

0 ( p ) , and the linear combination of 1 and

the random variables

X1 pjr
^

j = 1, k p , r = 1, t ) closest to 0 p , p j

( p = 1, P ,

) in

0 ( p ) = (m0 z p m1 ) + z p X 1 pzw ,
^

the least squares sense equals 0 ( p , pj ) .

References [1] Goovaerts, M.J., Kaas, R., Van Herwaarden, A.E., Bauwelinckx,T.: Insurance Series, volume 3, Efective Actuarial Methods, University of Amsterdam, The Netherlands, 1991. [2] Pentikinen, T., Daykin, C.D., and Pesonen, M.: Practical Risk Theory for Actuaries, Universit Pierr et Marie Curie, 1990. [3] Sundt, B.: An Introduction to Non-Life Insurance Mathematics, Veroffentlichungen des Instituts fr Versicherungswissenschaft der Universitt Mannheim Band 28 , VVW Karlsruhe, 1984.

0 ( p , pj ) = (m0 z pj m1 ) + z pj X 1 (3.1) pjw

^

where: X 1 pjw =
r =1

w pjr w pj .

X1 pjr ,

X1 pzw =

kp

z pj z p.

z pj = w pj . d 01 / c11 + (w pj . 1)d 11

j =1

X1 pjw ,

]
1 pjr

(the credibility factor on contract level), with: 0 1 1 d 01 = Cov X pjr ,X1 pjr ' , d 11 = Cov X pjr , X pjr ' ,

r r' ,

c11

and: z p = z p. 01 11 p. 11 credibility factor at sector level), with:

) ( ) = Cov(X , X ) = Var (X ) , D / [C + (z 1)D ] (the

1 pjr 1 pjr