lopment are recalled here for the reader who is moderately familiar with
the theory on an intermediate level (without the use of measure theory).
Chapter 2 treats the risk process and at the same time tlie tools of the
theory of stochastic processes are elucidated. Chapter 3 explains the
concept of the collective and develops the related risk quantities. Conse-
quences of the mathematical model form the content of the second part.
Chapter 4 deals with premium calculation and Chapter 5 with the
retention problern. Finally, the real operational problems are taken up
in Chapter 6, the subject of which are the risk carrier's stability criteria.
In addition to the probability of ruin criterion, the dividend policy and
utility criteria are also discussed. The general tendency toward forming
a bridge between economic and actuarial theory is particularly visible
in this last chapter.
It should also be mentioned explicitly what the book does not treat
(at least in so far as what the reader may have expected). No statistical
estimatioiz methods are brought up in connection with the mathematical
models developed here. I believe that the Separation of the trains of
thought into
a) construction of models and
b) measurement of the Parameters which appear in the models
should be preserved -as in the classical actuarial mathematics of life in-
surance. The present book therefore deals intentionnally only with point a).
The publication of this book would not have been possible without
the vigorous Support of my colleagues. Dr. H.-U. Gerber assisted with
a preliminary version of the manuscript and also made a real contribution
to the content of section 6.4 with his dissertatioii. My assistants, Messrs.
F. Pfenninger and W. Maurer undertook a careful examination of the
manuscript and the corrections. I should particularly like to thank the
translator, Mr. C. E. Brooks, F. S.A. for his cooperation. Through him
it has been possible to publish the book in the language which will make
it accessible to the largest possible circle of readers. My appreciation is
also due to Mrs. E. Minzloff, secretary at the Mathematical Research
Institute of the Federal Institute of Technology for typing the manuscript
in final form.
1.1.1. Definition
A random variable X assumes certain real numerical values in accord-
ance with well-defined probabilities and we therefore speak of the law
of prohability governing a given random variable. This law is described
by means of the variable's (cumulative) distribution function F*(X). The
distribution function is defined as follows:
Fig. 1. Empirical claim amount distribution (motor accident bodily injury claims over
50,000 francs)
4 Chapter 1. Probability Aspects of Risk
etc. We shall usually exclude the case S =O. This is not absolutely neces-
sary, but within the framework of our theory-as from the practical
viewpoint-it is useful to make this exclusion.
Fig. 1 shows us an empirical distribution function related to bodily
injuries in motor accidents producing losses in excess of 50,000 francs.
We can read from the graph for example P [ S > 100,000 francs]=
1- Fs(lOO,OOO)= 0.20. Corresponding to empirical distribution functions
we can develop theoretical distributionfunctions by "smoothing out'' the
former. Deriving theoretical distribution functions from empirical ones
is a statistical problem which will not be dealt within this book. In what
follows it is assumed that the distribution functions employed have
already been "smoothed out". Fig. 2 gives us a theoretical distribution
function for motor accident bodily injury claims in excess of 50,OOO francs.
Fig. 2. Theoretical claim amount distribution (motor accident bodily injury claims over
50,000 francs)
and
b) step distribution functions of type
Distribution function :
The form of the frequency curve depends basically upon whether the
Parameter y is greater than, less than or equal t o 1. In the case y > 1 the
gamma distribution has a shape similar to the log-normal distribution.
When y = 1 the curve is referred to as an exponential distribution. y < 1
produces a density function which is no longer bounded at the origin.
exponential d~stribution
1.1. Random variables explaiiied by the example of claim amount 7
f'(4= 6B(%)
1
( (1 - for 0 < x < 6
where
W .
B (a, ) = .
r ( a + )
Graph of the density function:
V)Cauchy distribution
(Parameterspandd, - m < M < + o o , O < O < c o )
Density function :
0
for - C G < X < + C G ,
z[02+(x- M)2]
bilateral
Cauchy
distribution
20
for M ~ x < c G .
Z[O~+(X-M)~]
A
unilateral
Cauchy
distribution
units of money
o Jump points. Increases of the function at the jump points are called jump amounts
where nk is the number of claims among n cases of similar type for which
the amount of loss S = k. Thus
E Cg - {C (4
g
total claim payments
n J l n = total number of claims
in terms of the claim amount distributions ii), iii), V) and vi). (In the last
case distinguish between a S 1 and cr > 1.)
b) Prove that
random variable):
i) g (X) = xk:E [Xk] = pk: k-th moment about the origin of F, (X)(also,
more simply: the k-th moment).
Special case : p1 = p expected value of F, (X).
ii) g(x)= (X - E [(X-,U)~]= ak:k-th central moment of F,(x).
Special case: a, = E [(X -p)'] = 0' :variance
o: standard deviation.
In addition to these measures
Intuitively we can say that the expected value measures the Center
of gravity of the distribution, if we think of the probabilities pi as weights
at the corresponding points. Analogously, the variance can be regarded
as the moment of inertia about the expected value. It is perhaps even
clearer intuitively if we think of the variance simply as a measure of
dispersion or mean quadratic deviation from the expected value. Finally,
the skewness measures the extent to which the mass of the probability
distribution extends to the left (negative skewness) or to the right (posi-
tive skewness) of the expected value.
For the sake of completeness we also define the median of a distribu-
tion. This is the number M for which F ( M ) ~ and & at the Same time
Chapter 1. Probability Aspects of Risk
i) Normal
distribution
ii) Log-normal
distribution
iii) Gamma
distribution
iv) Beta
distribution
U . C=
vi) Pareto --
xl+m
distribution
ik
viii) Poisson e-A-
distribution k!
a+k-1 I -P
ix) Negative binomial 0, l,2, ... Otuca U . ---
distribution O<p< l P
X) Logarithmic
distribution
F(M - c ) <+for all E > 0. In the example below is M = 20, since F ( M )= 0.6
and F ( M - E ) <0.3.
Fix) t
This function is said to exist if the defining integral exists for at least one
strictly positive value t , >O. If M(t,) < co also for a strictly negative value
t , <O, then all mornents of Fx(x) exist and the following relations hold:
M'(0) =E [X]
M1'(0) = E [ X 2 ]
M'k'(0) = E [ X k ]
ii) Characteristic function
cp(u)=E [eiuX]=J eiuxdFx(x), U a real number,
i = 1/-i(imaginary unit)
The characteristic function always exists. For the moments of F,(x)-if
these exist-similar relations hold as in the case of the moment generating
function,
(0)= (i)k- E [ X k ].
It is important to note that the relation between the characteristic
function (resp. the moment generating function, ifit exists) und the distribu-
1.1. Random variables explained by the example of claim amount 21
It is clear even from these few values that the estimates given by
Chebyshev's inequality are very crude. They can be improved, however,
if certain information about the type of the distribution function is
available. The reader may consult C373 or [48] for generalizations of
Chebyshev's inequality.
Chapter 1. Probability Aspects of Risk
Exercises
1) Calculate several moments and auxiliary functions from the table
under 1.1.4.
2) a) Show that the expected value V has the following property
for all C:
E [(X - S E [(X - c ) ~ ] .
b) Show that the median M of a distribution function has the fol-
lowing property for all C :
In terms of our intuitive interpretation, the left side represents the claim
amount distribution for the k-th claim event and the right side is the
amount distribution for the first n claim events, but with a condition
placed on the k-th claim amount only. A relation which we shall use
often in the sequel follows at once from the foregoing:
Distribution function:
where
where Bayes' Theorem merely states that the band defined in the entire
probability space (plane) by the condition X , = S now bears the total
probability mass 1 ; this is achieved by multiplying the previous (uncon-
ditional) probabilities in the band by a constant (depending only on S,
not on X,).
and designate the expression on the left as the conditional density ,function
of X, given X,. In the case of the two-dimensional log-normal distribu-
26 Chapter 1. Probability Aspects of Risk
tion this is
1.2.3. Independence
Lemma :
= exp { [i
-
i= 1
(log xi - 2 n}'l[(2nr" . sn x, . x, .. .X,,]
i :Y
and the characteristic function cpxk(u)= 1 - -- .
1.2. Sequences of random variables
= I / v ~ ~ ( x )Var(Y)=o(X).o(Y).
.
The proof is simple : For any t we have E [(t X + Y)'] 2 0, or
=CE[(Xi-ECxil). (xi-ECxjl>].
i, j
Remarks.
a) The convergence behavior of Var(Sn)/n2 measures the velocity
with which stability is attained.
b) Our estimate for Var(Sn)makes no distinction between positively
and negatively correlated quantities. Quite generally
Exercises
1) A and B are two events such that P[A] = 114, P[B/A] = 112 and
P [A/B] = 114. We define random variables X and Y such that X = 1 if
A occurs and X=O if A does not occur; Y= 1 if B occurs and Y=O if B
does not occur. Calculate E(X), E(Y), Var(X), Var(Y), p(X, Y). Are X
and Y independent?
2) Let Sn be the sum of n independent, identically distributed X,:
1 with probability p
Sn=xk!, X,, where X, =
0 with probability q= 1- p .
a) Show that Sn is binomially distributed.
S -np
b) Develop the characteristic function cpn(u)of -E---and show that
fi
cpn(u)+ e-U2i2 when n+co
(Central limit theorem for (0, 1)-random variables).
c) Determine $,(U), the characteristic function of S and show that
( U )-+ e when n +co,
provided n p = A is constant (Poisson convergence theorem).
3) Let X X X ... be independent claim amount reproductions
and suppose that X , has the density function
1
(i.e.y, varies with k and c is constant). It is assumed that y = lim -
n+a> n
y,C;=,
exists. Determine then for S,=C;=, X,.
a) Var(S,), E(Sn);
b) the frequency function of xYn(x).
C) Show that Sn/n converges in probability toward y/c (law of large
numbers for gamma-distributed random variables).
d) Show that for cpn(u),the characteristic function of
2.1. Fundamentals
P* ( t )= E [ P ( t ) l ,
+
S* ( t )= S ( t )- P(t) E [P(t)],
Pdm u n t
~'~arnount,
lump
jurnp
jurnp points
incrernent
i.e. if all increments of intervals of the same length have the same
distribution.
The following two theorems will be particularly important to us.
7beorem I : Suppose that { X , ; t >= 0 ) is a counting process with independ-
ent increments, i.e. the sampling functions are step functions with constant
positive amount jump 1. l f E(X,)=m(t) is a continuous, monotonically
increasing function of t , then m ( t ) can always he introduced as the new
time variable in accordance with the relation ( f o r simplicity's sake we
assume m (0)= 0 )
so that
a) {XI .r 2 0 ) has stationary increments,
b) { X ,J T 2 0 ) has independent increments,
- 7k
C) P[Xr=k]= e r - k=O, 1,2, ... (i.e. is Poisson distrib-
k! uted).
2.1. Fundamentals 39
P[X,Ix]= C
k= 0
ecAf-
k!
~ * k ( ~ )
depends thereby only upon the last known value of the process (in the
present or the past). This property is known as the Markov condition M:
P C X r m + , I x m +...,X,"
l, s~~/x~,=xi,...,X,~=x,l
(M)
... > X * , ~ X ~ / X , ~ = ~ , I ,
=PCXtm+,5xm+Lr
where the sequence
the claim event in such a way that only one such event can occur at a
given point of time. Cumulative (simultaneous) claims are counted in
this sense as one single claim event.
In the classical collective risk theory it is assumed at this point that
the claim number process has independent increments. If we now apply
theorems I or I1 of section 2.1.2, we discover (after changing over to
operational time) that N, is a Poisson process, i.e.
For simplicity's sake we shall also assume the existence of the following
limits for all t and n. (For a discussion of the existence of these limits
See Chung [20].)
Relations b) and C) mean intuitively that for small time intervals the
probability of a claim occurring is proportional in first approximation
to the length of these intervals and that the sampling functions of the
claim number process have only positive jumps of amount 1.
2.2. The clairn number process
The function A,(t) is called the intensity oj'j'requency of the n-th claim
(transition from n - 1 to n claims). Observe that in the case of stationary
transition probabilities the intensities of frequency are no longer depend-
ent upon time. Knowing these intensities enables us to calculate all of
the transition probabilities by use of Kolmogorov's (forward) dijj"erentia1
equation system (see e.g. [32]):
which can be easily verified by insertion into the system. From this it
follows in particular that the pnm(t,S ) are always non-negatizie for any
given Set of non-negative intensities of frequency A,(s) m= I , 2, . . . and
also that they are differentiable in t and .Y.
Chapter 2. The Risk Process
Iheorem: rlhe "normal case" Sn(t,s)= 1 for all n, t, s applies if und only
" 1
if E --
m = 1 Am
diverges (Am,m = l,2, ... are the non-negative time-independent
claim frequencies).
Proof: From the foregoing differential equation (plus boundary con-
ditions) for S(t, s) we get
S
and
m
If 1
m=n+l
l/Am diverges, ~ ( s must
) be Zero, which implies that Sn(s)=1.
Conversely
remains bounded if 1
l/Am converges; hence S(T) cannot converge
to 1 (which by the theorem of dominated convergence (cf. Appendix)
X
would imply the erroneous Statement s 5 (l/A,) is bounded for all s).
In the case of time-dependent intensities of frequency no simple
characterization is known, but the following theorem is useful:
rlheorem: Let {Nt; t 2 0) be a number of claims process with intensities
of frequency Ai(z) (i= 1,2, ...) und {&*; t 2 0) another number of claims
process with intensities of jkequency AXT)(i= 1,2, ...).
7hen
Ai(z)rA;(z) for all i Z N , t s z s s
implies
S , ( t , z ) ~ S ~ ( t , z )for all n r N , t j z s s .
Ak(s)=O for s = t .
Now An(s)= p(t, s) - p,i,(t, s) 5 0 from the explicit formula for pnn(t,S)
and the assumption A ,(T)2 An+ ,(T).
If Ak-,(s)-0 then
W,({&; t 2 0)) = point of time at which the first claim event in the realiza-
tion {&; t 2 0 ) occurs;
W,({N,; t 2 0))= point of time at which the n-th claim event in the realiza-
tion {Nt;t 2 0) occurs.
Graphically, this has the following appearance:
Proof :
This relation tells us that the claim interoccurrence time has no memory.
This means that if no further claim is incurred up to time W ,+ s then the
future distribution will be the Same as if the last (n-th) claim had occurred
at time W,+s. This property, which we have derived by analytic devel-
opment, can also be deduced as a direct consequence of the Markou
condition.
The following two theorems are important:
Theorem I : If the claim .frequency intensities are independent qf time,
then the claim interoccurrence times are independent of the points oj'time
at which claims occur.
independent of W,.
This is the proof for T, and T,.In the case of k claim interoccurrence
times the procedure is analogous. We See from these two theorems that
time-independent claim intensities of frequency are very much to be
desired. The next section deals with the question of where and how this
time independence may be attained.
2.2. The claim number process 49
1 -P&, t+h) -
lim -Cn,
h-0 h
from which
We have thus far shown that the equation (*) holds for t of the form
t = p - ' ( T ) . That it is valid for any value o f t can be seen, however, from
the fact that, if p - ' ( z + O ) + p - ' ( T ) , we conclude from
must hold.
Finally, it can be seen from (*) by differentiation that the condition
of the theorem is satisfied for y ( t )= p f ( t ) ,An+, =C,. (For the reader who
is familiar with measure theory we may add "up to changes of Lebesgue
measure Zero in the /Z,(t)".)
Remark on the "classical case" in collective risk theory: In this case
it is postulated that all the claim frequency intensities are identical, i.e.
A,(t)=A(t) forall n .
Thus an operational time always exists.
The claim number process becomes the homogeneous Poisson process and
we therefore have the case of the classical collective risk theory.
I,+k+ih. An+l...An+k
~ n , n + k ( ~ ) = ~ -
be(2b)...(k b ) - ( e b h - l ) k , b = ~ , - A
2.2. The claim nurnber process 53
where the fraction on the right is set equal to 1 for k=O. The formula is
valid for k = 0 and follows by induction for any other value of k from the
recursion formula at the beginning of 2.2.4. It is easy to see that the
relation holds for both positive and negative contagion-for the latter,
however, only as long as A, >O. From here on we shall discuss the
two cases separately.
and therefore
These are the probabilies given by the negative binomial distribution. The
reasoning used here is similar to that of Polya-Eggenberger [30] in terms
of which contagion is explained by an urn model with contagion. It
should be noted that this distribution is satisfied not only by the total
number of claims but also by the conditional distribution given the
number of claims already incurred.
11,. Negative contagion (b<0)
An= a + b(n- 1) if the right side is positive. There is a final index m
such that A,r0; we then set A,=O for all k r 1. Let p(h)=ebh and
An+ 1
CI,=- . Then p(h)=O for s < n, s> m whereas for n s s 5 m- 1 we
-b
have the general formula
Exercises
1) Since the Poisson distribution represents the no-contagion case,
can it be derived as a limiting case of the
a) binomial distribution (negative contagion),
b) negative binomial distribution (positive contagion)?
i) How must the Parameters be varied in cases a) and b)?
ii) Take the limit with the help of the characteristic functions.
2) Let us adapt the claim number process to the case of a single
death risk (i.e. A n ( t ) = O for all n r 2 ) .
i) Show that there is always an operational time.
ii) Calculate pol(t, t + h) for the usual time scale,
, p, (h) for the operational time scale.
iii) Interpret the meaning of the operational time scale.
3) Give an interpretation of linear contagion in the homogeneous
case with regard to the effect on the claim interoccurrence times:
i) distribution of claim interoccurrence times,
ii) expected claim interoccurrence times,
iii) sum of the first n claim interoccurrence times (characteristic
function).
4) Prove that if the claim number process has independent increments
then all intensities of frequency are identical.
waiting times (the sums of claim interoccurrence times) Aln) can also be
defined as [ W , = t W,=t ..., Wn=t Wn+,>t].
% j n ' is the class of all the sets Al").
'
I R n = {[.X . ..,X,- ,I) is the set of all ( n- 1)-tuples[ X x 2 , ...,X,- ,I.
The integration is over 'LI)")X IRn-', i.e. over the class of all sets A r )
and all (n - 1)-tuples [ X x 2 , ...,x ~ - ~ ] .
We might add the following constructiue explanation: G,(x/A~"')
t h e distribution function of the sum of the Y,,, (tiAYi)-is obtained
for given AIni from a generalized Chapman-Kolmogorou equation:
G,(X) = 2
n= 0 %("I
j G, (x/AY1)dP [ A r i ]
with
or, by integration over 91, (= the union of all the 2l Y'), where Al represents
any element of <II, :
G,(4= 1G,(x/A,)dP(A,).
%,
Insertion of this relation into the first formula of the proof yields the
desired result.
is dependent only upon the number of points in Aj"', and we therefore have
and hence
We shall consider first the case b $. O and revert to the case b = O later.
We now define a probability density function f (T)= b eb'/(ebt- 1) in the
interval CO, t ] . The function G,(x) can then be written as
The probabilities pn(t)are identical with those in 2.2.4 under the discus-
sions of positive and negative contagion. If we now make use of the
definition
60 Chapter 2. The Risk Process
we have
Gi(x)= C,"=, p,(t) G,(x/n), the reduced basic form.
In the case of identical intensities ( b= 0) the same expression holds.
We choose
1 1 '
f E ( x / Y ) =l~
&(x/Y)
0
and find that the p,(t) are the Poisson probabilities e-"- ("' . (This
special case was discovered by Jung [40].) n!
To summarize, we have the following situations in the case of time-
independent intensities of jrequency which follow a linear Pattern:
2.3.5. An example
We shall now illustrate the mechanism just discussed by means of an
example. Suppose that
-pn(t) is the probability as given by the negative binomial distribution
(positive linear contagion b > 0 ) ;
-&(X)= 1-e-"" is the claim amount distribution; the density func-
tion f,(x) is therefore C t . e-"";
-the claim amounts are independent of each other.
We then find that
2.3. The accumulated claim process
and thus
cb
/,<X>= [[(b - c X) t - 11 e ' b c x ) ' +l]/(b - cx)*.
Exercises
1) Consider the example in 2.3.5 and
a) prove without using the transformation for f,(x) that
C) Show that the following relation must therefore hold for a risk
process with independent and stationary increments:
For the risk taken at random from the collective O with structure func-
tion U(9) it holds then that
8
P [N(s) = m/N(t,) = n,] =
1P"' [N(t2)= n,] dU(9)
8
3.2. The weighted risk process as description of the risk 69
(Cf. the case without contagion in 2.2.4.) Then we have the following
7heorem: If the structure function U(9) corresponds to an infinitely
divisible distribution, then
70 Chapter 3. The Risk in the Collective
Remark: One should take care not to draw conclusions about the
probability law of the weighted process {Nt; t 1 0 ) from this theorem; the
latter only refers to the random variable Nt for a fixed time t .
Proof: Let $(U) be the characteristic function of U(9) and <pt(u)that
of the random variables 4.
Thus
1
*(U)= eiU9d U($)
and
) = S l1 dU(9)
< P ~ ( ~ est[p'U-
i.e.
and
where
1 -
C
p(t)=---.
t c+t
1+-
C
(Cf. the positive contagion case in 2.2.4.) Note that we allow only the
Parameter 9 to vary, and not p(t) which we assume to be independent of 9.
With analogous means to those used with the Poisson basic proba-
bilities we can show that in the present case too
72 Chapter 3. The Risk in the Collective
where
Exercises
1) Prove the theorem stated in 3.2.4 for negative binomial prob-
abilities.
2) Let
9+k-1
P'" [Nt = k ] =
and
variance
the quotient 2 1 for any structure function U(A).
expected value
where
74 Chapter 3. The Risk in the Collective
We now set p, ( ln
:(U))
- and
=)i(~ ) wish to show that ),(U) is a charac-
teristic function; this follows from the relation
The proof given there is based however on the assumptions that the
amount jump distribution is independent of time and that it has a finite
first moment.
Remurk 2 : The following conditions are equivalent:
A) U ( 9 )is infinitely divisible.
B) po(t)=J ep9' d U ( 9 ) is of the form ew"', where w(O)=O; w ( t ) can
be differentiated as many times as we please with ( - 1)"w'")(t)rOfor all
n r 1.
X
m
dn log e-" d U ( S )
and the equivalence of B) and C) follows from this relation. The two
conditions B) and C ) are found in Thyrion [66] and Pesonen [58].
and thus for the risk in the collective O with structure function U ( 9 )
m
P [S,5X ] =G I ( x=
) 1 1 j G ~ S ' ( ~ /d ~~ ~" '[A:"']
n ' ) dU(9).
n = 0 <U:") 8
P [S, I
X] = G, (X)= 1 rnpj G, (XIA;~))d~ [A;")],
n= 0
where
P [Al"'] = j P'" [Al"'] d U (9).
0
Exercises
1) Carry through the complete proof of the converse of the represen-
tation theorem in 3.2.5.a.
2) Let <p~s'(u)=e3'LX'""1be the characteristic function of the ac-
cumulated claim of the risk 9 and let the structure function be
and
to denote the number of claims process and the accumulated claim process
respectively of the i-th risk in the portfolio. The number of claims process
and the accumulated claim process of the eniire portfolio are then denoted
3.3. Portfolios in the collective
and
For such a portfolio what can we now say about the stability of the
portfolio processes { N t ;t 2 O ) and { S , ;t r O } ? By stability we understand
the convergence in probability of the average toward a constant; in
principle we can expect two kinds of possible stabilizations:
a) Stabilizing in time: convergence of SJt and Ni/t if t +cc and n
is fixed.
b) Stabilizing in size: convergence of S,/n and NJn if n +n; and t
is fixed.
We shall consider in what follows only the quantity S,, since the same
argumentation applies to 4. We have
1
1-m
t( U ) - I] =m U -
- i uE (taking the derivative of
+ W
u/t the characteristic
function).
Thus by interchanging the order of the limit and the integration
(cf. Appendix, Section 5),
lim ip,
1-g>
(4)
=[ ciuE9dU(9),
i.e. the distribution function of S,/t tends toward the distribution function
of 9 . E. With the exception of the case in which U($) concentrates the
total mass in one point (homogeneous collective) therefore, we have no
stability in time. This connection is known in the literature as the theorem
of Ove Lundberg [47]. It can be generalized as follows:
7heorem: 1j the following stability property holds for every risk with
parameter S ,from the collective O :
Sj9)/t-+ m(9) in probability as t -+CO, m(9) being a constant depending
only on 9 (Luw of large numbers), then
3.3. Portfolios in the collective 79
and thus the limit function can be given any value we like (e.g. 1) on this
Set. We thus obtain by interchanging the order of the limit and the inte-
gration (cf. Appendix, section 5)
The proof is thereby completed for all "continuity points" and the
theorem follows at once from this fact for all remaining points.
Interpretation: The structure function U(9) defines a distribution
function Hm((x)on the "mean values" m(9). The average over time SJt
converges then toward a random variable having a distribution identical
with that of these mean values m(9).
The connection between this general theorem and the theorem of
Ove Lundberg can be established by using the relation
2) Suppose that the accumulated claim process SIi) has the charac-
teristic function m
Premium Calculation
4.1.1. General
!fj[Gs(x)]=EIS]+a.o[S] (a>O)
where
a [SI = i / j (X - E [SI)' d ~ ~ ( x )
for a " utility function " u(x) with ul(x)2 0 and u"(x) 5 0. (These hypotheses
tell us that utility increases monotonically and that marginal utility
decreases monotonically-consistent with the usual assumptions of
economic theory.)
This principle requires that the utility u(0) before assuming respon-
sibility for the claims experience be equal to the expected utility
E[u(P-S)] after taking over such responsibility in exchange for a
premium P
1
Special case of the quadratic utility ,function: u(x)= X -- X' (X 5 C);
2c
p=E[S]=JxdG,(x)
and
o2 =Var [SI =J (X -P)' dGs(x).
The quantities p and a2 of the premium in the collective can either be
derived on the basis of relation A or from the elements p(9) and a2(9)
4.2. The risk premium and the collective premium
and
In Summary:
The symbols "E" and "Var" in these two formulae are to be understood
as representing the operations expected value and variance with respect
to the structure function, whereas p(9) and a2(9) are the quantities
explained in the preceding section which characterize the risk 9.
Finally, we find for the collective premium P:
a) P = p(1+ 1)in the case of the expected value principle,
b) P = p + M . o in the case of the standard deviation principle,
C) P = p + - a2 in the case of the variance principle.
4.2.4. The dilemma and the connection between risk and collective premium
The most typical problem in calculating a premium for a given risk
can be formulated as follows: The risk premium must be determined for a
risk for which the collective premium is known.
As we have Seen, the collective premium is essentially derived from
statistical observation data. It is natural, however, that we wish to know
the "true" risk premium for a given risk. The essential point here is the
fact that this risk premium will simply remain unknown in most cases.
There are only two exceptions to the rule that the risk premium remains
unknown :
a) if the statistical collective is homogeneous: In this case-as can
easily be verified-the risk premium is identical with the collective
premium;
b) if the risk can be observed over a very long Span of time (ideally
over an infinitely long duration) and if the claims experience during this
observation period is stationary. In less abstract terms, this means that
the risk conditions remain unchanged over the Course of time.
It is moreover easy to See on the basis of our definitions that in life
insurance the risk premium can be estimated within groups of reasonable
size because (case a) the working hypothesis ofthe homogeneous collectiue
is more or less acceptable in many cases. This working hypothesis is only
rarely appropriate in non-life insurance, however. Case b) is even more
seldom encountered in practice than (at least approximately) case a).
Although the risk premium may not be obtained from the data
a n d this fact is of primary importance-it plays an important role as
a concept (as idealization). It represents as it were the "theoretically
correct premium" for the individual risk. As we shall See later, this
theoretically correct premium can at least be approximated step-wise
(cf. 4.3). The conceptuul difference between collective and risk premium
is however of fundamental significance. It seems to us that this difference
has not been sufficiently underlined in many earlier theoretical investiga-
tions; indeed, the failure to distinguish conceptually between the two
quantities has been precisely the cause of erroneous conclusions having
serious consequences. Having demonstrated this "dilemma" between
risk and collective premium, we shall now turn to the formal connection
between the two.
4.2. The risk premium and the collective premium 91
Exercises
1) Let the accumulated claim distribution of the risk be compound
Poisson, i. e. m
G&"(x)= p?)F*"(x),
n= 0
where =
9"
-- and F(x) is an arbitrary distribution function
n!
with mean value p and variance a 2 (independent of 9). In addition let 9
be distributed with the r-distribution as structure function, i.e. with the
density function
Determine
a) the risk premium according to the expected value, standard
deviation, variance and Zero utility principles (in the last case with a
quadratic utility function),
b) the collective premium according to the same four principles of
premium calculation,
C) the iterative premium P* (with approximations) for the first three
premium calculation principles.
2) Show that for the variance principle the following relations
between the collective premium P and the iterative premium P* hold
generally :
PZP* ifVar[a(9)]+2.Cov[p(S),o(9)]20
and
PSP* if Var[a(9)]+2.Cov[p(9),o(9)]50
4.3. The credibility premium 93
that of 2.3; the symbol as used here would represent S i - S , , in the latter
section.)
-We shall assume that the structure function U ( $ )of the collective
and the collective premium P (but not the risk premium P($))are known.
We now wish to find the
-credibility premium P, (Sb S2, ..., Sn) which approximates P ( $ )
in a way yet to be specified. We can think of this as the premium to be
applied for year n + 1 if S S2, ... ,Sn for the years 1,2, . . . , n are known.
We shall attempt to solve this problem in what follows. The simplest
credibility theory is based upon the expected value principle of premium
calculation and it is therefore not surprising that all the works cited at the
beginning of this section ure bused on this principle. We shall now extend
the theory of credibility to the variance principle; naturally the theory
becomes somewhat more complicated thereby. It also becomes far more
useful, however, since it then allows us to approximate not only the
mean claims cost proper to the risk (as in the case of the expected value
principle) but also the related loading for fluctuations.
5j [S/9] = P(9)= E [S/,9] + . Var [S/9] (E [-I91 and Var [./9] for the
distribution for given 9)
und
$[SI = P = E [SI + D -Var [SI (E [-I and Var[-] for the distribution
in the collective as explained below.)
Definition D,) does not contradict D,); the latter is rather a special
case since
P(S)=$j[S/S] = E [S/S] +. Var[S/S] = S + . O=S
according to D,) and thus by DZ)
For the case in which the claim amounts in the past S,, S,, ... , Sn are
independent of the claim amount S in the coming time interval for a
given risk we have the following result which is the cornerstone in the
construction of the credibility premium:
Theorem: Let S und (SI, S, , ...,Sn) be stochastically independent ,for a
giuen parameter value 9. Then
But R,(S ..., Sn)plays the role of a constant in the inner integration
and we thus know that setting this equal to the conditional expected
value will produce a minimum. Thus
100 Chapter 4. Premium Calculation
Concerning b), C) and d): The proof of these statements is beyond the
scope of this book. Statement b) has been proved by the author [15].
Statements C)and d) follow from the fact that for increasing n the a poste-
riori distribution d W(9/Sl, ...,Sn)degenerates more and more toward a
point distribution. Doob C281 has shown that this is the case with prob-
ability 1 (with respect to the structure function) and that the point so
obtained then coincides with the true risk Parameter. (Schwartz [62]
and Freedmann C341 have proved generalizations of this theorem.)
Concerning e):
I f no claims data SI, .. .,Sn are available, we have
can be broken down into the three components (by an additional break-
down of the approximation Part as defined in 4.3.5)
a) expected value part,
(a)+ b) = approximation part)
b) variance part, and
C) fluctuation part.
Now it is very important for practical applications-where we often
have no (or no exact) knowledge of the parametric distributions for the
accumulated claim distributions and the structure function-to obtain
manageable formulae for the three components. Such formulae, which
replace the exact form of the credibility premium, are called credibility
formulae in the terminology chosen here. Credibility formulae thus take
4.3. The credibility premium 101
b= v a r [ P ($11 -
-
v a r [P @)I - n
var[~] 1 n+k
+
- . E [02(9)] Var [p(9)]
n
and therefore for the fluctuation Part with linearized expected value Part
This last formula does not bring us very much further unless we
average thisfiuctuation part once more over the entire collective. Carrying
this through we have then
n
+
E [02(,!J)] Var [ P (G)]
4.3. The credibility premium
The designation "excess variance" is based on the fact that in the pure
Poisson case Var [SI = E [S]/n.
b) In the variance part
Here we have found for the credibility
Var [L2/9]
where E,($) is the k-th central moment of S for a given risk parameter 9
.4(9) - 3,
or, by use of the excess y2 (9)=----
.t
(9)
(It is well-known that every normal distribution has the excess 0.)
4.3. The credibility premium 105
we obtain
( n-1
(1 - b ) . ~ a r [ ~ ( 9 ) ] =
r (n . ~ a[SI
(Var [SI - ~ a[SI)
Var [SI
r - Var [SI)
The following data for the year 1961 give the number of automobile
liability policies in Switzerland for private cars which were struck by
k claims.
0 103,704 0 0
1 14,075 14,075 14,075
2 1,766 3,532 7.064
3' 255 765 2,295
4 45 180 720
5 6 30 150
6 2 12 72
7 and over 0 0 0
P
-- ~~P~~~
This is valid under the assumption that k claims in respect of the given
risk have been observed on average over n years.
C,) For the uariance part the same approximation holds as for the
expected value part. This is so because in the Poisson cciw
-k provides a "better " estimate of 02(9)than does the dispersion Z2
determined from the risk,
o 2(9)= ,u(9) and thus
4.4. A practical exarnple 109
( 0 otherwise.
b) S,, S, , .. ., Snare independent and identically distributed. Si= S . k i ,
where the ki are binomially distributed with parameter 9. The structure
function has the density function
1 0 otherwise.
C) S I , S2, ...,Sn are independent and identically distributed. Si is
normally distributed with mean value 9 and fixed variance 0;. The
structure function has the density function
P,[Ssx]= 1P'"[N=k]
k= 0
.GAk(x).
where C") denotes the price for taking over the risk completely. This
form of reinsurance is also known as "surplus reinsurance".
ii) The non-proportional case:
gi [Sii)] = Sii)- 1
jumps in
max [amount jump - M i O]
,
[Os11
('I
The connection between Sii) and Rii) can be seen most clearly in the
above diagram (cf. also section 2.3). Each function gi is uniquely charac-
terized once the bound M , is given.
-Mi is called the first risk or (maximum) retention.
T h e price for this type of reinsurance is given in the form of the
function #')(Mi)= where e"'(0) = 8")again denotes the price for
completely taking over Sii).
This form of reinsurance is also called "excess of loss".
In both cases-proportional and non-proportional-the retention
problem amounts to determining for each risk in a given risk mass a
corresponding number ( U , or M,). In this connection, these numbers are
usually broken down as follows:
ai=C.lli ( i = 1 , 2 , ..., N )
and
M,=K.bi ( i = 1 , 2 ,..., N ) .
E [S"'] = pi ( 9 )
and
Var [s")]
=a' (9) (cf. 4.2)
E [S")]= E [pi(,Y)]
and
Var [S")]= E [o' ( 9 ) ]+ Var [pi(,Y)] (cf. 4.2).
or more briefly
Ai
gi = 1 $(,'(Mi)
j= 0
with T(i)(Mi)=min [M,, 1);!]',
where the Yji)(Mi)are independent and identically distributed for any Mi.
We have then the general formulae (cf. exercise 4 in 2.3 and [13])
5.2. The relative retention problem
+
Var [gi]= E [ A i ] .Var [q(i'(Mi)] Var [Ai] . E2 [ q i ) ( M i ) ]
for the variance.
It remains to determine E [q("'(Mi)]and Var [?("'(Mi)]. Writing
simply F for F"), we find
denotes the "profit remaining within the retention", where P") stands
for the premium which the direct insurer receives for the risk with
accumulated claim variable S").Since the g i are independent, we have
and
N
Var [ Z ] = 1Var [gi]
i= 1
@=Var[Z] + A .E [ Z ]
N N
= C Var [gi]+ 2 . C {P(')(1 + U , )
- + cci . E Csi])
. E [Si']
i= 1 i= 1
and therefore
C) Tt remains now to introduce the expressions for E [g,] and Var [g,]
found under a) in the relation (G). Now (if F continuous at M,)
and
a
--- Var [g,] = 2 M , E CAj] . Cl- F(M,)]
aMj
M,
+2 [ I - F ( M i ) ] .[Var [Ai] - E [ A i ] ] 1 [ I - F ( x ) ]d x .
0
5.2. The rclative rctention prohlem
For this reason and also because of the fact that j [1 -F'j)(x)] dx
0
(the mean claim amount for the retained portion) usually lies considerably
below the bound Mj in practical applications, the second term in formula
(M) is of relatively little significance.
Finally, it should be pointed out that in the determination of the
bounds Mj, it remains to be verified that our solutions do in fact solve
the minimum variance problem. The proof is very laborious and we shall
not give it here, especially since the gradation of the Mj which we have
found appears to be optimal on intuitive grounds.
5.2.3. The risk with given risk parameter and the risk in the collective
under non-proportional reinsurance
In our treatment of the relative retention problem in non-propor-
tional reinsurance we postponed the discussion of the changes which are
necessary in our considerations according as the latter refer to a risk
with known parameter or to a risk in the collective. We shall go into
this question now.
If we examine the assumptions made under b) in 5.2.2 somewhat
AL
more closely, we find that the representation gi= ?"'(Mi)
with in-
j= 0
dependent and identically distributed ?")(M~)is only possible if the risk
120 Chapter 5. Retentions and Reserves
parameter for ?")(Mi) and thus the distrihution function F")(x) are known.
If we do not assume these to be known then some of the considerations
made in 5.5.4 are no longer valid. (We shall not look into the resulting
case here.)
For our present discussion then let us make the following assumption:
Ai
The random variable S") = 1 Y,") is described
j= 0
where
and
T?($)= 2 [ k - v j ( ~ ) l 2p!j)(,9);
k= 0
U ) Proportional reinsurance
We have found the following proportionality factors for the risk in
the collective (cf. 5.2.1):
Here also the a posteriori expected values and variances which appear
in the formula can be replaced by the approximations of 4.3.6 and 4.3.7.
and independent and have the Same distribution function F(x), which is
independent of p;
b) the Same assumptions as under a) hold for the risk in the collective
described by a structure function U(p) over the probabilities of a claim
occurring.
W(Mj, S(j')=K. a j ,
where
-W=functional which associates a number with the probability
distribution of the accumulated claim process S'j) and the bound Mj;
-~ -
At first sight it would seem more natural to describe a gradation of retentions as
follows: u j = C . U(yj, S"') with an appropriate functional U . Our formulation V ( u P)=
C . y, produces this functional U very simply. however e.g. under the assumption of a
monotonically property with respect to thc U, on the basis of the defining relation
C . U(y,, S'J1)=u, with the property that V(u,, S(ji)=C . y j
The monotonicity guarantees the uniqueness of U , .
The advantage of the "less natural" formulation given here-~over and above the Fact
that the formulation of the proportional case becomes symmetric with that of the non-
proportional case~-will become more apparent in the sequel.
5.3. The absolute retention problem 125
Moreover, for this choice of V and W for a risk mass which is reinsured
partly on a proportional and partly on a non-proportional basis, the
two constants C and K are identical. For it is easy to See that the two
forms of reinsurance coincide for a risk which always produces the
maximum (full) claim amount, if a claim occurs.
Aj
Let S(j)be equal then to 1
i= 0
Y,(j), where Y,(j) is a fixed number (no longer
a random variable!) independent of i, and choose a j and Mj such that
UJ . .Y(j)=Mj.
t Then
V*(aj, P ) ) =
Var
[ 1a . . X(" ]
J
-
Var [A,] . M)
' = Mj . ( E ~1).
+
E CAj] . Mj
In order to get the risk reinsured in exactly the Same way under
proportional and non-proportional reinsurance (which are the Same
here), K must be equal to C.
We come finally to the following formulation of the absolute retention
problem:
-Let V and W be the functionals as above for proportional and
non-proportional reinsurance respectively.
126 Chapter 5. Retentions and Reserves
(By suitably normalizing the functionals as e.g. in the case of the optimal
V* and W* we can always ensure that the Same proportionality constant
appears in both instances.)
-The absolute retention problem consists in determining the con-
stant C.
5.3.2. The random walk of the risk carrier's free reserves generated by
the risk mass
It is characteristic of the absolute retention problem that no solution
is possible without some formulation of the risk carrier's business goals.
This is in contrast to our remarks on gradations of retentions (relative
retention problem), where we merely employed a minimum variance
principle which would probably be acceptable within any-resonable-
framework of business goals.
Before we can formulate these goals, however, we must describe the
consequences of various retentions levels, such consequences arising
from the fact that the risk carrier's free reserves fluctuate in accordance
with these retention levels. We speak of these fluctuations as the "ran-
dom walk" of the risk carrier's free reserves. For a mass of N risks the
latter has the following appearance:
-Let SIi)( i = l , 2, ... , N) be the accumulated claim process of risk i.
For simplycity's sake we shall assume (cf. 2.3.3) that for known risk
parameter this process is compound Poisson, thus that
and
other, then the accumulated claim process of the entire risk mass is also
weighted compound Poisson. Formulae ( A )giue the (conditional) intensity
und the (conditional) claim amount distribution for this process.
Over against the accumulated claim process S, we have the accu-
mulated premium income
N
C* = 1p,
i= 1
where C"' is the risk carrier's accumulated premium for risk i in the time
interval [0, t ] . The differente
represents then the random walk of the profit realized on the risks in
question. If we also take into account the risk carrier's available free
reserves Q at time t = O , then
is the random walk of the risk carrier's free reserues. The following
diagram will make these definitions clearer:
5.4. Reserves
Exercises
1 ) Let Zt be the random walk of free reserves,
S, the accumulated claim process, subject to the compound
Poisson law of probability : ( U ) = ep'[X'u'-ll,
Show that
2, has independent increments and determine the
distribution function FZt(x),and the
characteristic function cpZt(u).
2) From the process Z, of exercise 1) form the differences
and
where z(n)= time at which the n-th claim occurs in the process S,.
What are the distribution functions of X, and Y,,?
Are ( X , , X,, ...,X,, ...) and (Y,, Y,, ..., Y,, ...) sequences of independ-
ent random variables?
3) Examine the random walk 2, defined above and the sequences
{X,; n = l , 2 , ...) and {Y,,; n= l , 2 , ...) for the case of the risk in the
collective with structure function U ( p ) .The accumulated claim process S,
then has the characteristic function
Chapter 6
o discrete time
points of interest
134 Chapter 6. The Insurance Carrier's Stability Criteria
For each of the four cases listed there is thus a set of unaccep-
table realizations of the random walk. We shall designate these Sets as
follows:
case 1a) : A T , , -planning horizon 7:
-time points of interest are multiples of h ;
case 1 b): AT,, -planning horizon 7:
-all time points are of interest;
case 2a): A,, , -planning horizon co,
-time points of interest are multiples of h ;
case 2 b): Am,, -planning horizon co,
-all time points are of interest.
It is easy to verify that
and
Thus for every random walk of free reserves case l a ) produces the
smallest and case 2b) the largest probability of ruin of the four possible
cases.
It is also easy to see on intuitive grounds that
These relations express nothing other than the continuity of the probabi-
lity measure.
For the exact proof (especially for the second relation) the reader is
referred to [23].
6.2. The probability of ruin as stability criterion 135
and
-,
Since the increments Sk - Sk are independent for a given risk para-
mater and since
G('' (tk)= P"' [Sk- Sk 1 5t k ] ,
P
we therefore have
t= -a>
with
P'" [Afi] = P") [W ; Afi] = P(') [Q + E* ; A i _
and the moment generating function is identical in form (cf. 1.1.4) except
that here we must set
+ a>
X*( 4 3 )= J euxd F (XIS).
- a>
Now let B, denote the event in which ruin occurs precisely at time n,
that is
B n = [ S l s Q + c ; S 2 S Q + 2 c ; ... ; S n - l i Q + ( n - l ) c , S n > Q + n c ] .
Let
P'") [Bn/Q= U] = <P!,") (U)
and n
P'") [ AJQ = U] = $2)(U)= q$)(u).
k= 1
(N.B. In this context <P?) is used with a new meaning as a probability and
not as a characteristic function.)
These functions are meaningful only for non-negative arguments U.
We have, with similar recursion formulae as under U),
(P\s) (U) = 1- G") (U + C) for u 2 0
and
and thus
where o
~=e-'"-l+G")(u+c)- 1 er"dG"'(v+u+c).
V= -m
140 Chapter 6. The Insurance Carrier's Stability Criteria
where R"'= smullest upper bound for r which sutisfies corlditions 1) und 2).
It is important to point out at this juncture that R'"' (as suggested by
the notation) varies with the risk Parameter 9.
It can further be shown that x*(r1/9)exists and is finite for all rfe[O, r]
and that the function
W(t)< 0 for all t >0 for which the function exists. Therefore R") = 0.
is valid for all T and h. Thus we have from the limit relations mentioned
in 6.2.1 also
$~)<U)=P'"[A,,,/Q = U] - ~ e - ~ ' " ' "
for infinite planning horizon and discrete points of interest, and
$?)(U) = $@)(U)
=1 for aii u if R")=O ("certain" ruin)
This can be justified for example by taking the limit in the solutions of
the integral equation below.
The method described here yields an exact calculation of the proba-
bilities of ruin and more particularly the asymptotic formulae for large u
and
from which the equations for the continuous case may be obtained by
taking the further limit h + 0. The latter limiting process is rather labo-
rious, however, and since we shall discuss another method for deriving
the same integral equation, we shall not reprodnce the proof here. (The
latter can be found for example in slightly different terminology in C221
or [23].) The resulting equation is
The integral equations (R,) and (R,) have been solved by Cramkr C221
and Tcklind C631 according to the Wiener-Hopf method (i.e. using
Fourier transforms and function-theoretical means in the complex plane).
We shall only summarize here the most important results in this con-
nection.
i) F(x/9) has measure only the positive half-axis
(only positive claims).
For large U we have
in the continuous case
and
in the discrete case (and where the claim distribution is absolutely
continuous)
This method does not give us any new results for the case of the
infinite planning horizon beyond those that could be obtained by follow-
ing the Cramir approach. It is appealing, however, because of its elegance
and clarity.
We set ourselves the task of calculating
i.e. the probability of ruin in the continuous case with infinite planning
horizon, and where the initial reserves are equal to U .In order to simplify
the notation, we shall omit the characterizing risk parameter 9 from our
symbols in what follows, although the arguments used are only valid for
a given, known risk parameter 9. According to our hypothesis about the
accumulated claim process St (in 6.2.3) the latter is then (for a given risk
parameter) compound Poisson with distribution function
It can be Seen with the aid of the drawing that ruin can only occur
at the random time points W i.e. when a claim occurs. Thus the set
A m . , can be characterized as follows:
The Y, are independent (also from the T,), identically distributed claim
amounts (with distribution function F(x)).
The event A m , , can then also be defined as follows:
It is noteworthy that the X' are not only identically distributed and
independent; they also have a (common) density function:
The point (n, Sn)is called a record point (ladder point) if Si<Sn for
all i < n. In this case we refer to Sn as a record height (ladder height) and
designate it by M, if k further record points lie to the left of the point in
question. (We always Count (0, So) as a record point.) The differences
M , - M , - , = X are known as record jumps (ladder j ~ m p s ) .The ~
(k = 1,2,3, . ..) are identically distributed random variables with
distribution function
P[.gSx]=H(x).
The record jumps are dependent on each other to the extent that for
a non-existent 2,all later record jumps are also non-existent. Otherwise
they are independent of each other. It is obvious that H(O)=O; on the
other hand it is possible that H(co)< 1. (1- H(m) is the probability that
no further record point will be found. In this case we say that the distribu-
tion function H(x) has the defect 1 - H(co). Defective distribution func-
tions are also acceptable within the framework of the method of interest
to us here.)
The terminology chosen here differs slightly from that usually employed in renewal
theory; See Feller 1331 for example.
6.2. The probability of ruin as stability criterion
U(x)= E
H*k(~).
k= 0
for initial free reserves X no ruin occurs. The following connection with
the renewal measure then holds:
Theorem: If H ( m ) < 1 ( i . e . if the distribution of record jumps is
defectiue) then
a) every random path (realization of random walk) has only a finite
number of record heights (set of exceptions has probability measure 0)
und
b) 6(x)=[1 -H(Go)] U(x).
Proof of a): U(m) represents intuitively the expected number of
record points. If H(m) < 1, however,
But then the number of record points is a fortiori finite on almost all
random paths (i.e. the Set consisting of all random paths with infinitely
many record points has probability measure zero). In other terms the
associated renewal process breaks down (terminates).
Proof of b):
For the details of the derivation of this formula the reader is referred
to pages 363 and 364 of Feller's book [33].
The main problem arising from this elegant method consists in
determining H(x) (distribution of record jumps) from the distribution
F(x) (distribution of claim amounts). In the following we shall again
solve this problem for non-negative claim amounts.
A comparison with the definition of the probability of ruin *(U)
given at the beginning of 6.2.5 shows us immediately the interconnection
with the above function. We have simply
Setting s = u + C z gives us
150 Chapter 6. The Insurance Carrier's Stability Criteria
where W W
-Pj e K Y [ l - F ( y ) ] d y = l and
C 0
This formula is identical with that found in 6.2.5, for we have (through
integration by parts)
P
1=-J eKY[l-F(y)] dy= ---+PSP P m
eKY
dF(y),
C o C.K C.K 0
from which - W
and thus
and @I')
and
e*=c. t
of ruin satisfies
$(U)5 Po
for given u (usually large) and P, (usually small, e.g. 1/).
Note: We have denoted the constant to be determined by K here,
whereas in 5.3.1 this constant was called C. We make this change of
notation on purpose since the letter C already has a specific meaning
in the asymptotic expression for the probability of ruin.
It is intuitively obvious (and will also be demonstrated) that-pro-
vided the loadings in the retained portion are unchanged-the smaller K
is, the smaller is the probability of ruin $(U). Since there is a secondary
interest to retain as much risk (and therewith expected profit) as possible,
the case $ (U)=P, is of particular significance in practice.
First of all we shall retain the assumptions of 5.3.2 with regard to
the accumulated claim process S, of the total risk mass. In addition let
us assume that the claim amounts of the individual risks may have
different distributions, but that the latter do not depend on the Parameters
to be drawn from the collectives; this assumption will simplify our work
considerably, but it is not essential in principle to the procedure which
we shall follow.
In terms of the characteristic function, therefore, we have then
(N.B.: ~ * ( udepends
) upon 9, although the xi(u) do not.) For the accu-
mulated retained claim process gK(Si)we then have (The index K is
intended to recall the constant used in the gradation of retentions.):
where
154 Chapter 6 . The Insurance Carrier's Stability Criteria
E [Ajl
In accordance with the discussion of 6.2 the constant R"' is then deter-
mined from the equation
and thus
= P, we obtain
and from the relation eP2"lpK
in formula (C) and the Same definition for hi as in the preceding section
for the percentage loading on the expected value in the retention, we
find that
We can now split up the risks into the two parts described and
proceed as in the preceding sections. Let
E,=insurer's expected profit in the retention from risks which are
reinsured proportionally and
E,=insurer's expected profit in the retention from risks which are
reinsured non-proportionally.
Using the same notation as before, we have then
and
where our notation expresses clearly the fact that the values P,($) are
dependent upon the risk Parameters.
b) For the case of proportional reinsurance
For the optimal gradation derived from formula a) in 6.3.2 we have
Using these symbols in the relation for R@)found at the end of the
last sub-section, we have
The following relation holds here for the quantities 6,(9), where the
values on the right refer to the collective:
[1 +&($)I pi(9)=(l + 4 ) . pi for any 9 .
This simply expresses the fact that the premium income for any 9 is
always the same. The relation actually represents the definitin of the
Si (9).
Further, R'" must be set equal to Zero ($'")(M)=1 for all M) as soon
as the relation
1pi(9) di(9) Ei(K)>0
1
(sum of all loadings positive) (I)
is not satisfied. This follows from our considerations about the proba-
bility of ruin in 6.2.4d) (case I) and in 6.2.5a). Let us simplify further as
(Note that -
7, y, in practice.)
We set
1 - C ~i (9)6i (9)Ei ( K ) - expectedexpected retained profits s
p (9) C pi(9) Yi Ei(K) reinsurance profits if retention
is taken over by reinsurer
and we thus have
R'" = 2/[K .p ($)I
and
$(')(U)jexp { - 2u/[K. p(9)I).
Finally,
where (I) refers to the above-mentioned requirement that the sum of all
loadings be positive. The probability of ruin can be approximated
Interpretation holds if we assume ii- yi and that the true Parameter is known to the
reinsurer.
162 Chapter 6. The Insurance Carrier's Stability Criteria
represents a lower bound beneath which $(U) can never lie regardless of
how large u is chosen (except in the trivial case in which the entire risk
muss is ceded to the reinsurer)!
where
6.3. The absolute retention
6.4.2. Hypotheses about the model variables when the dividend policy
is used as stability criterion
-, ,
L, = Z , - Z , = C + Si- -Si represents then the annual increase in
the free reserves Z, .
Under the assumption made the L, are likewise integer valued,
independent and identically distributed (except for L,=Q,). Let the
(common) probability distribution be described by the numbers
g,: P[L,=j] = g j (j=O, f 1, +2, ...) for all positive integers 2.
Let the dividend to be paid out at the end of the year be denoted
by D,; as a result of this payment the original random walk of free
reserves
Z,=Qo+L1+...+L,
is replaced by a random walk modified by the dividend payments (last
dividend payment not yet counted)
Using the discount factor 0 < V < 1 we then evaluate the policy Y (for
initial free reserves Q,) as follows:
The small circles are to suggest that f (X)is of interest only at the integral
lattice points. The intuitive interpretation is quite clear: No dividend is
payable for free reserves Q in the interval (b,, a,]; otherwise Q is reduced
by dividend payment to the next lower a-value.
Of particular simplicity are the barrier strategies, which are a special
case of band strategies with n=O; in this case
In his original work de Finetti took only such strategies into account
(and made very special assumptions about the probabilities gj). Miya-
sawa C493 has shown that an optimum barrier strategy exists for the
case g, =g, = ... =O. This optimum property does not hold, however, in
the general case. (See examples in Morill [50].)We shall nevertheless
pursue our study of the barrier strategy in the next section because of
its practical significance. The chief disadvantage of the general band
strategy is the fact that the dividend payment does not depend mono-
tonically upon the available capital.
Chapter 6. The Insurance Carrier's Stability Criteria
ifOIQoIao PI)
Qo-ao + v ( u o , ao) if Qo
The middle equation is obtained by considering the situation after the
+
expiration of one time unit (year). If L, =j, then Q, = Qo j (since D, = 0
according to the strategy); the sum of the discounted dividends from the
second year is thus V(Qo+j,U,). This is then discounted for an addi-
tional year and the expected value is taken over all possible values j.
The other two equations are self-evident. In his original work C261
de Finetti considers the case where
gl =P, g-, = 1 -p=q and all other gj=O.
The middle equation then reads
and we find by using the boundary conditions for Qo= O and Qo=ao
that .Y-(Qo,U,)= C,. rpO+ C, . r)o with suitable constants C C,, where
r, and r2 (r, > r,) represent the solutions of the characteristic equation
r=upr2+uq.
The reader is referred to the work of de Finetti for the details. (See
also exercise 2 in this section.) Additional material on the general barrier
strategy in the discrete case can be found in Morill [50].
Q, has fallen below U, because of claims. 0 Q, has reached the value U,.
Solving for h7
ch
because of the continuity of V ( Q ,a ) in Q
0
Using this we form the expected value of J C ec' dt, which leads to
0
m
p* J ( { c e - t d t ) ecN*'ds=- P* J (1 - e-') e-p*s ds
0 0 0
and thus
C 1
C(U)= a+O -
h'(4
( ~ * + ) h ( a ) - P *J h(a-y).dF(y)
0
because of (D,).
We therefore have
ch'(~)=(p*+)h(x)-p*.~Jh(x-~)e-~~dy.
0
By differentiation we obtain
X
If we replace the last member by the relation from the original integro-
differential relation, it follows that
As can easily be verified, the relation r, + y >0 always holds and thus C,
and C, have different signs. (For h ( x )to be positive we must have C, >0,
C, < O and C, + C, >O.) The function h'(x) then takes on its minimum
174 Chapter 6. The Insurance Carrier's Stability Criteria
value at the point a, within [0, CO) and we find from hV(a,)=O that
The optimal bounds are tabulated in the following table for =0.05 and
several values of P*, y and C.
Intuitively C stands for the annual premium income, P* for the annual
total number of claims, and l l y as the expected amount of a single claim.
=0.05 corresponds then approximately to an annual rate of interest
of 5 %. There is a certain satisfaction in noting that the optimal free
reserves ao (in respect of other criteria also) vary within quite reasonable
orders of magnitude.
By setting expression (A) into (D,) we obtain for the new function 6(x)
the integro-differential equation
6.4. Dividend policy as criterion of stability
where
W
1
A= J e-"dF(y) and dF(y)=- epSyd ~ ( y ) .
0 A.
But this is precisely equation (R,) from 6.2.6~)and thus 6(x) can be
represented as the solution to the renewal equation (cf. (R))
> d
V(0) V(0, Y)2 [I - $(0)] -
(-d = present value of perpetual dividend
of intensity d)
Remark: Intuitively speaking, the theorem states that for large risk
masses the optimal barrier a, also becomes large.
d) Let us now determine the asymptotic behavior of 6(x) and thus
of h(x) also. It is practical to norm 6(x) in such a way that 6(00)= 1. It
then follows from (D,) that 6(x) is identical with the probability of
survival which appeared in 6.2.6. If
p*A.
--- 1[ l - p(x)] dx < 1 (and this always holds in the case which is of
C o interest to us: C > p* J X dF(x)),
we can take over directly formula (Ru)from 6.2.6c), i.e.
d(x)-l-Ce-"" (forlargex),
with K such that
and
where
C - Co, where C, =
"0 Pt
and , ~ ~ P*
= ~ ~ y e ~ ~ ~ [ l - F ( y ) ] d ~ .
0
where we recall once again the intuitive meaning of the symbols, viz.,
Q = initial free reserves,
E* = accumulated premium income in the time interval [0, t],and
S, = accumulated claim process variable (accumulated claim amounts)
in the interval CO, t].
2, thus represents the remaining amount of free reserves at time t. As
we have explained already in 5.3.2, the stability problems of the insurance
enterprise cannot be handled in a meaningful fashion without formula-
6.5. Utility as criterion of stability 179
q z ; ;t r o ] = @ [ z ; ' ;t 2 0 ]
and we will choose the second in preference to the first if
180 Chapter 6. The lnsurance Carrier's Stability Criteria
we mean the random walk which arises by carrying out the random
walk (2;; t z O } with probability a and the random walk {Z:';t r O }
with probability 1-U. (The probability laws of 2; and 2;' are assumed
here to be independent.) Obviously a must be a number between Zero
and one. Then the following linearity condition (in respect of the mixture
operation) must hold for the evaluation 9 in order that 9 be called
linear:
9 [ Z , ; t r o ] = a 9 [ Z : ;t 2 0 ] + ( 1 - a ) 9 [ z ; ' ; t 2 0 1 .
%?[Z,;tzO]=sup
Y { r: I r: 11
a E jep'dY,' + ( l - @ ) E J e - ' d ~ " ,
where Y,' and Y," are assigned by the dividend policy 9'to the random
walks 2; and ZJ' respectively. By carrying out the supremum operation
separately for each member, it follows that
6.5. Utility as criterion of stability 181
1
-e-# (Y;)
182 Chapter 6. The lnsurance Carrier's Stability Criteria
or using our abbreviations, that A > B > C > D (which is correct for
example if V = B = In (714)).
Because of the linearity of the evaluation we have then for the mixture
{Z,; t20}={aZ;,(l-a)Z;'; t Z 0 )
discounted sum of dividends under ,Y, : a C + (1 - a) B,
discounted sum of dividends under : a A + (1 - E ) D.
evaluation with %
Assuming that the axioms Al) to A,) (including the strongcr forms)
hold, we have the
Auxiliary theorem 1 : Let
@[Z1] =@[U Wl, (1 -a) W,] und
@CZ"I = % [ P w (1 - P ) W,].
7hen
%[AZ',(l-A.)Z1']=%[y Wi,(l -y) W,], where y=A.a+(l-A).
Proof: It follows from A,) that
--
probability of
choosing mixture
I and therein W,
probability of
choosing mixture
2 and therein W,
and from axiom A,) that the right hand side must be greater than or
equal to the expression
K [Zr']
K [Z] + I [Zu] - K [ Z ] - -- .
- K [Z'] - K [Z"]
A B
Principal theorem 2
Let 9 ' he an evaluation which is likewise linear und equivalent to Uli.
Then 9 ' [ Z ]= A 9 [ Z ] + B ( A>0, B real).
Proof: From auxiliary theorem 5.
These two principal theorems complete the reasoning of von Neu-
mann and Morgenstern. In the following sections we shall treat some of
the consequences of the considerations made here.
Chapter 6. The Insurance Carrier's Stability Criteria
(K(t, X) and V(t) are assumed to be such that the defining integral yields
a finite value.)
This evaluation is obviously linear-and thus a utility-since the
distribution function A H: (X)+ (1 - A) H:'(x) corresponds to the mixture
{nz:,(i-n)z;; t z o } .
Intuitively, K(t, X) measures the evaluation of the free reserves of
amount X at time t , whereas V(t) (resp. its increments) assigns to each
time interval the weight corresponding to our estimation of its impor-
tance.
For example : K (t, X)= K (X)e - ' ( = force of interest), where K (X)
expressing the idea that all equally long time intervals in the planning
period [0, T] are judged to be of equal importance.
6.5. Utility as criterion of stability 189
Note that we are using the symbol Y in a different sense here from
that of 2.1. Y'j) is a random variable which tells us what insurance
carrier j's free reserves are after the exchange. The exchange takes place
here as the result of ugreements (reinsurance treaties) among the insurance
carriers; we can describe these agreements by real junctions
with the intuitive meaning that they should produce the share of insurance
carrierj after the exchange in dependence upon the quantities X,, X,, . . .,X ,
which relate to the situation before the exchange.
From the nature of the agreements as exchange we must have the
following admissibility condition for the functions f j ( j = 1,2, ...,N):
N N
C f,(x
j= 1
X,, . . . , X,)= C xi
i= l
identically in X X ... ,X,.
In what follows Y"', . .. , Y',' resp. Y"', ... , Y',)will always denote random
variables which arise from the Zu', ... , Z',) through exchange, i.e. by
applying udmissihle functions &(X,, . .. ,X,) ( j= 1,2, ... , N ) , as just defined.
Let us now suppose that insurance carrier j evaluates its random
walk at time t , with a given utility kerne1 K j ( x ) having the properties
postulated in 6.5.5. We then have the following evaluations of the
situation before and after the exchange:
Theorem: Under ihe usual assumplions (cj: 6.5.5) aboui lhe utiliiy
kernels Kl(x),K,(x), ...,K N ( x ) i.e.
, where
i) K i ( x ) is strictly monotonically increasing und differentiuble every-
where for all i, und
ii) K:(x)is strictly monotonically decreasing for all i, the
f j ( ~ 1 > ~...>
2 ?xN), ( J = 2,
are Pareto-optimal among all admissible exchange agreements if' und only
if they satisfy the condition
K ; ( & ( x ~..., , x ~ ) ) = c , K \ ( f ~ (...,
x ~X,))
>
identicully in
x1,x2,..., X , (j=1,2,..., N). (z)
The c,, ... , C , (C, = 1) are here arbitrary positive constants.
Remark: Since the evaluation of random walks of free reserves is
made by using expected values, it suffices of Course to require condi-
tion (Z) for any X ...,X , in a Set M of N-tuples ( X ... , X,) with proba-
bility 1. What occurs outside M has no influence at all upon the expected
value.
Proof: Here we follow essentially Du Mouchel (cf. [29]).
U )( Z )is sufficient
We assume that (2) holds for the N-tuple of the functions
f1(x,, ..., X,), ..., fN(xl,... , X,) for constants C,, C,, ..., C , which are all
positive. Let el(x ..., X,), ... ,eN(xl,... , X,) be a further N-tuple of func-
tions with which we form the admissible functions fj=fj+ej. N
From the admissibility condition (N) it follows at once that ej
j= 1
must be equal to zero. By applying the- f;. to the Z"), . .. ,Z'N' we obtain
It follows from the existence of the derivative K;(x) that the expression
in brackets tends itself toward Zero as E approaches zero; the convergence
is monotonic since &(X) is monotonic and thus the expected value on
the right hand side also tends toward 0. It is therefore possible to find
a small positive C such that @ l ( Y ( l ) ) - % l ( ~ " ) ) > ~ .
6.5. Utility as criterion of stability
By the same procedure as before we find that the second term is relatively
insignificant for small E . Considering the first term, we find that
With regard to the practical interpretation this result tells us that Pareto-
optimal risk exchanges of random walks depend only on the sum of these
random walks. The insurance carriers will thus form a pool if they are to
realize a Pareto-optimal solution. The Pareto-optimal criterion says
nothing, however, about how they should divide up the results of this
pool. On the other hand, we would mention again at this point the basic
attitude of cooperation which remains the prerequisite for the insurance
carrier's attaining a Pareto optimum.
6.5. Utility as criterion of stability 197
For the utility kernels to satisfy the usual properties we must have
b
for insurance carrier A : X s--=
2C
H, and
e
for insurance carrier B : X 5--=
2f
(These conditions will always be tacitly assumed to hold in what follows.)
M and thus play the role of a level of saturation (capacity) for insurers
A and B respectively.
Moreover, h, C,e and f must be non-negative constants. We are, by
the way, working with very special quadratic utility kernels here simply
for computational reasons. These quadratic forms can in addition be
justified as approximations to more complicated utility kernels as long
as the values of the random variables X characterizing the risk are small
in comparison to the domains [0, a] and [0, ] of the two utility kernels.
We shall be concerned below with the question: "What is the (minimum)
price (=premium) for which the two insurance carriers are prepared to
take over the risk?"
a) Solution under individual market behavior
In this case each insurer determines the price for the risk X exclusively
on the basis of its own standards. We shall follow through the situation
198 Chapter 6. The Insurance Carrier's Stability Criteria
If we let the equal sign hold in this solution we have the minimum
premium P, (cf. principle of Zero utility in 4.1.2). Because of the quadratic
form we find for this minimum premium
~=E[xI+~-/-,
where a = b/2c and E [X] and a2[X] are the expected value and variance
of X respectively.
Similarly, for insurance carrier B :
e
P,=E[x]+~-~/-, where P=-.
2f
In particular: The higher the level of saturation (capacity) the smaller
the minimum premium.
b ) Solution under cooperatiue market behavior
In this case the two insurance carriers are prepared to work together
and to divide up the risk proportionally, i.e.
insurance carrier A will take over y X and
insurance carrier B will take over (1 - y) X
of every possible claim charge X produced by the risk (Osy 5 1). For
this the minimum premium
4 (YX) + 5 ((1 - y) X) = P(X/y) is to be charged .
The effect of the cooperation is strongest if y is chosen such that P(X/y)
becomes minimal.
Since
~(~/y)=~[~]+a-1/a~-~~a~[~]+-fl-(l-~)~cr
we wish to find that y* which maximizes
Here the two insurance carriers adopt the (very extreme) standpoint
that each seeks to make its utility as large as possible as compared to
that of the other company. More precisely, each wants to maximize the
difference between its own utility and that of its rival.
The following cases are possible for the premiums P, and P, to be
charged by A and B:
i) P, <P, (The risk is taken over by A.)
The utility of insurance carrier A is 4?Ll (P, -X) = E [K, (P, -X)] ;
the utility of insurance carrier B is a2(O)= K, (0);
the difference V(P P,)= E [K, (P,- X)] - K, (0).
ii) &=P, (The risk is taken over by A with probability p and
by B with probability 1 -P.)
+
The utility ofinsurancecarrier A is p E [ K , ( P ,- X)] (1 -p) K,(O);
the utility of insurance carrier B is (1 - p) E [K, (P,- X)] + p K,(O);
the difference V(P,, &)=P. E[K,(P,-X)- K2(0)]-(1 -p)
. ECK,(P,-X)-K,(O)I.
iii) &<P, (The risk is taken over by B.)
The utility of insurance carrier A is a1(O)= K,(O);
the utility of insurance carrier B is @, (P,- X) = E [K, (P,- X)] ;
the difference V(& P,) = K,(O)- E [K, (P,- X)].
In the rivalistic case A seeks through its choice of P, to make V(P P,)
as large as possible and B through an adroit choice of P, seeks to make
the same function as small as possible. This is exactly the situation of a
two-person zero-sum game. (See e.g. von Neumann-Morgenstern [53].)
Because K, (X) and K, (X) are continuous and monotonic, we find that
and
sup inf V(& P,)= V(P, P),
PA PB
Since also
inf sup V(& P,)= V(P, P)
PB PA
produces the same value, the two-person zero-sum game described here
is determined (i.e. it has a pure value).
By setting the special forms of K,(x) and K, (X) into the determining
+
equation E[K,(P- X)] E[K,(P- X)] = K,(O)+ K,(O) we find that
Here too the resultant price is that which a company with level of satura-
tion (capacity)
d ) Comparison
If we assume that a > , we have the following quantitative comparison
of the (minimum) premiums under the different market behaviors discus-
sed above:
Insurer A
Insurer B
Cooperative
(for optimal y*) P a+ _P<$<$
Rivalistic
e ) Discussion
The three different market behaviors represent extreme cases. In
concrete situations it must be assumed that all three behaviors are present
"in a certain mixture" for the individual insurer. An understanding of
the significance of market behavior can be gained, however, by a study
of the extreme cases. The relations between the magnitudes of the resulting
prices are particularly suggestive.
APPENDIX
A.1. Preliminary
Every distribution function F(x) can be broken down uniquely into
a weighted average of two distribution functions of which one is a pure
continuous and the other a pure step distribution function. This finding
can be stated formally as follows:
Lemma: a) F ( x ) = p . F , ( x ) + ( l - p ) F , ( x ) ,
where F(x) = any distribution function,
F, (X) = continuous distribution function,
<(X)= step distribution function und
05~51.
b) ?bis breakdown is unique.
The proof can be found in C441 or C551 for example.
if the limit on the right-hand side exists and is the same for all sequences
of partitions of the interval for which the fineness tends to 0 and for any
intermediate points 5,.
iv) We say that g(x) dF(x) (the indefinite integral over the whole
real axis) exists if
b
J g(x) dF(x) exists for all a, b and
a
b
- lim Ig(x)l dF(x)< cc (absolute convergence of the indefinite
a-+ - m
b++m <I integral)
and we then define:
h
J ~ ( x ) ~ F ( x )lim
= jg(x)d~(x).
a-t-m
b++m
When the above conditions are not fulfilled we say that the (generalized)
Riemann-Stieltjes integral does not exist.
Special case: If F' (X)=f (X)exists, then
We then define:
The integral concept used here is very useful and intuitively easy
to grasp, but is nevertheless seldom encountered in the literature. (The
closest approach is the description in [55].) We shall therefore state
several existence theorems and properties relating to our concept. (For
proofs See for example [52].)
In what follows we shall proceed from the breakdown
tions above are also necessary for the existence of the integral.
b
Important special case: j g(x) dF,(x) exists, if
a
--g(x) is bounded und
g ( x ) has at most a countable number of points ofdiscontinuity.
2. Existence theorem: g(x) dF(x) exists if and only if
b
i) j g(x) dF(x) exists for all a, b und
a
b
ii) lim j lg(x)l dF(x)< cc.
U+-a, ,
b
Remark: S g(x)dF,(x)= C g(ak)p,
a k
likewise exists only if the series
converges absolutely.
rule holds for the limits a, h which are continuity points of F ( x ) and g ( x ) :
52. Nagy, Szkefalvi -, B.: An introduction t o real functions and orthogonal expansion.
New York: Oxford University Press 1965.
53. Neumann,J.von, Morgenstern, 0.: Theory of games and economic behavior. Princeton:
Princeton University Press 1944.
54. Ostrowski, A.: Vorlesung ber Differential- und Integralrechnung, Bd. 11. Basel: Birk-
huser 1951.
55. Parzen, E.: Modern probability theory and its applications. New York: Wiley 1960.
56. - Stochastic processes. San Francisco: Holden & Day 1962.
57. Pesonen, E.: On the calculation of the generalized Poisson function. ASTIN-Bulletin
4 (1967). P. 120-128.
58. - NP-approximation of risk processes. SA 51 (1968), p. 158-164.
59. Philipson, C.: O n the difference between the concepts compound and composed
Poisson processes. ASTIN-Bulletin 2 (1962), p. 445-451.
60. - A review of the collective theory of risk. SA 1968 51 (1968), p.45-68.
61. Rozanov,Y. A.: Introductory probability theory. Englewood Cliffs, New Jersey:
Prentice Hall Inc. 1969.
62. Schwartz, L.: On consistency of Bayes' procedures. Proc. Nat. Acad. Sci. U.S.A. 52
(1964). P.46-48.
63. Seal, H. L.: Stochastic theory of a risk business. New York: Wiley 1969.
64. Tcklind, S.: Sur le risque de ruine dans des jeux inequitables. SA 25 (1942), p. 1-42.
65. Takeuchi. K.: A remark on economic survival games. J. Operations Res. Soc. Japan 4
(1962). p.114-121.
66. Thyrion, P.: Extension de la theorie collective de risque. SA 52 (1969), (to appear).
67. Vajda, S.: Minimum variance reinsurance. ASTIN-Bulletin 2 (1962). p. 257-260.
Index
Absolute retention problem 113 DE FINETTI 113, 114, 115, 132, 164, 168
- - when the risk parameters are drawn De Finetti solution 116
from one or more collectives 159 Density function 5
- - when the risk Parameter is known Discrete random walks 145
156 Distribution function 3
Accumulated claim process 36 Dividend policy 164
Admissible stability policy 135 - in the continuous case 168
AMMETER68 - in the discrete case 165
P