Escolar Documentos
Profissional Documentos
Cultura Documentos
Mathematics
I
' to second edition
s~ sPreface
\443
For the summer semester 1983, I gave a lecture course on non-life insurance
mathematics at the Faculty of Mathematics and Informatics at the University of
Mannheim, and I was asked by the Dean of the Faculty, Prof. K.-J. Ramspott, to
write out notes based on my lectures. When I gave a copy of these notes to Prof. E.
Helten, he asked me whether I would like the notes to be published in the Mannr
heimer Reihe, which I happily accepted. However, I said that I would rather wait
till I had tried them out on a lecture course at the University of Oslo in the spring
Die Deutsche Bibliothek- CIP-Einheitsaufnahme of 1984.
One problem that became clear was that writing internal lecture notes is
Sundt, Bjorn: something different from writing a book for publication; in internal notes you can
more freely reproduce from other sources. My notes were not very original, and
An Introduction to Non-Life Insurance Mathematics I von Bjern
Sundt.- 3. ed.- Karlsruhe: WW, 1993 they were not intended to be original; they were just intended to give a documenta-
tion of the lectures that I gave in Mannheim, and these lectures were to a great
(Veroffentlichungen des lnstituts fi.ir Versicherungswissenschaft der extent based on existing texts. The text on which I leant most heavily, was the set
Universitat Mannheim; Bd. 28)
ISBN 3-88487-398-9 of unpublished lecture notes by Prof. J.M. Hoem (1973). Chapters 2, 3, and 12 and
Sections 9.1-3 have to different extents adopted both disposition and results from
NE: lnstitut fi.ir Versicherungswissenschaft Mannheim: those notes, and I am grateful to Prof. Hoem for letting me publish these parts.
Veroffentlichungen des lnstituts ...
Since the first edition of the book was published in 1984, I have lectured on it
on several occasions, and I soon got ideas for amendments that I wanted to include
in a future edition of the book. These amendments started as hand-written notes. It
Later .I began up-dating them on the PC-based scientific word processor T3. The
amendments included recent research that fitted into the frame-work. I have also
profited from comments from readers of the book, in particular my students.
A special thank, and apology, goes to the students of a lecture course I gave at
the University of Oslo in the autumn semester 1990. Many improvements to the
manuscript originated from experiences with that course, and the students had to
endure a steadily changing text.
The most extensive changes from the first edition of the book, are the follow-
ing: A new Section 6.8 on hierarchical credibility and a new Chapter 8 on multipli-
Verlag Versicherungswirtschaft e.V. Karlsruhe 1993 cative rating models are included. In Chapter 7 on bonus systems, the asymptotic
Druck Prazis-Druck GmbH Karlsruhe optimality criterion has been replaced with a non-asymptotic one. In Chapter 9 a
ISSN 0170-2254 different proof of Lundberg's Inequality is given. To give a better flow in the pre-
ISBN 3-88487-398-9 3., neubearbeitete Auflage sentation, the material on moment-generating functions, Laplace transforms, and
(ISBN 3-88487-255-9 2., neubearbeitete Auflage) convex functions has been transferred to appendices. A simpler proof of Ohlin's
--------~T------------------------ ..............................111111111111111111111111. ~- ~
The most extensive change from the second edition to the third edition, is the
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
inclusion of exercises. Many of these exercises are based on exercises that have been
2. Classification of insurance forms . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Principles for calculating the compensation . . . . . . . . . . . . . . . . . 6
used at exercise courses in non-life insurance mathematics at University of Oslo in
connection with lecture courses on earlier editions of the present book and, before Exercises ........................................ 9
4. Premium principles . . . . . . . . . . . . . . . . . . . . . . . . . 10
that, on Hoem (1972). For these courses, Ragnar Norberg compiled a collection of
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
exercises in 1972 (Norberg (1972)); I am grateful to Norberg for letting me use ma-
5. Reinsurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
terial from that collection. Furthermore, I am grateful to Ole Hesselager for letting
Exercises ................................................ . 22
me use exercises that he has complied for similar courses at the University of Copen-
6. Credibility theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
hagen.
In the exercises, I have to a large extent included questions where one should 6.1. Introduction ...................................... 23
6.2. A simple credibility model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
comment on assumptions or results. In a practically oriented subject like insurance
6.3. The general set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
mathematics, it is important that the material becomes not only mathematics, but
that one also continuously considers questions like, what does this imply, are these 6.4. Bayes vs. empirical Bayes .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
assumptions realistic, does this result seem reasonable, etc. At an exercise course, 6.5. Recursive credibility estimation ............................ 42
such questions could be discussed between the teacher and the students. 6.6. Credibility estimators incorporating risk volumes . . . . . . . . . . . . . . . 47
The text of the book has been much less changed than was the case with the 6. 7. A credibility regression model . . . .. . . . .. . . . . . . . . . . . . . . . . . . . . . . 48
second edition. Most of the changes have been aimed at simplifying and clarifying, 6.8. Hierarchical credibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
and correcting errors in the second edition. The most important changes are the
Exercises .................................... 59
7. Bonns systems ..................................
following: New material includes subsections 5F, 6.4D-E, 7B, 8.3B, 10.2B,
10.6C-D, and 12D, and Appendix C. Section 8.2 has been extended. The proofs of 8. Multiplicative rating models .....................
Theorems 6.5, 10.2, and 10.6, and Lemma 9.2 have been changed. The material in 8.1. Introduction ............................................ .
the old section on ruin theory in Chapter 9 has been reorganised and divided into 8.2. The method of marginal totals ............................. 79
8.3. Classification of passenger cars . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
three sections (Sections 9.4-6).
Once more I am grateful to Howard Waters checking the language in the new
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
9. The risk process . . . . . . . . . . . . . . . . . . . . . . 88
material.
9.1. Introduction ..................................... 88
9.2. The claim number process ................................... 88
9.3. The claim amount process .. ......................... . 100
Oslo, September 1993.
9.4. Ruin theory ........................................... . 104
Bj0rn Sundt 9.5. Choice of compensation function 114
9.6. Choice of reinsurance form 117
Exercises ....................................... 120
10. The accumulated claim distribution ................ 128
10.1. Introduction .............................................
128
-1-
lB. The above remarks should make it clear that stochastic models and me-
thods are of great importance for the insurance industry. The following questions
are vital: It
less reliable estimates of the model parameters. And in particular in an introducto- max(x,O).
ry text like this book one should not present unnecessarily complex models; one
vi) For a function /we denote by f(x+) and /(x-) the limits lim f(x+t)
should rather try to give the readers some general ideas and stimulate them to fur- tl 0 and
ther work on the matter.
~1~ f(x~t) if the.se limits exist. Furthermore we denote by ~(x) and ~/(x) the
left and nght denvatives of 1at x.
lD. You cannot insure yourself against all sorts of economic risks. Typically
you can only insure yourself against damage incurred suddenly and unforeseen by a vii) we make the convention that E .n =0 and rr i=m-
n -1 f
z=m 1 n<m.
cause that you could not easily prevent. For instance, an ordinary personal accident
insurance usually does not cover accidents during mountain climbing as by climbing
mountains you deliberately expose yourself to risk. And you cannot get an insur-
ance against the need to paint your house. That need arises gradually, not sudden-
ly.
It is illuminating to notice the difference between burglary and ordinary theft.
Ordinary theft is covered only to a small extent. It is often difficult to say whether
the victim could have prevented a theft or not. On the other hand, a house-<>wner
should not be blamed if somebody breaks into his properly locked house.
It should be noted that the insurance company usually reduces the compensa-
tion if the policyholder has shown great neglect. If you have dried clothes on a
stove, and your flat catches fire, you will not get full compensation, perhaps nothing
at all.
fire
2A. The main division line in insurance lies between life and non-life. In this
windstorm
book we shall concentrate on non-life. We propose the following classification crite-
hail
ria:
earthquake
shipwreck
character of what is insured
theft
cause of the claim event
burglary
type of insured object in property insurance
water damage
conditions under which the claim event occurs.
sickness
accident
2B.Classification by the character of what is insured. We have the three
fraud.
main groups
2C. Classification by the cause of the claim event. Many insurance forms give
-6- -7-
By this principle the loss is fully covered as long as it does not exceed the sum in-
3A. In many classes of property insurance, a sum insured is specified. In most sured; if it execeeds the sum insured, then the sum insured is covered. The first risk
cases this is an upper limit for the compensation the insurance company will pay in principle is often used when it is difficult to determine an insurable value, e.g. insur-
the event of a claim. Usually the sum insured is equal to the estimated value of the ance of files and old buildings.
insured object, but it may be lower. In the latter case we have under-insurance. In fire insurance, water-pipe insurance, and comprehensive insurance of build-
By the insurable value we mean the value of the insured object. Even if the ings one can get full-value insurance. Then the policyholder will get full compensati-
insurable value and the sum insured were equal when the policy was issued, they are on even if the insurable value exceeds the sum insured. The premium is based on
not necessarily equal later; the insurable value may have changed because of inflati- the sum insured and is adjusted for inflation. If the policyholder makes alterations
to the building without informing the insurance company, we have under-insurance,
on, wear and tear, etc.
When an insurance claim occurs, one typically calculates both the costs incur- and the pro-rata principle will be applied.
red by the claim and the insurable value immediately before the claim event. There In vehicle damage insurance for ordinary passenger cars in Norway there is no
are various principles for the calculation of the insurable value. For most insurance sum insured, and normally the compensation is equal to the loss (reduced by a de-
forms one applies the day value, that is, the price of a new similar object with ductible). However, here also there is a possibility of under-insurance; in Norway
reduction for age, wear and tear, reduced applicability, or other conditions. By the mileage is applied as rating factor, and if the distance driven exceeds the upper limit
replacement value we mean the price of a new similar object with reduction for re- of the rating class, then the compensation is reduced by the ratio between the paid
duced applicability and other conditions, but not age and wear and tear. The re- premium and the premium for the class corresponding to the mileage when the loss
placement value is sometimes used when the difference between it and the day value occurred.
Finally, we introduce the extreme cases
is small.
3D. In accident insurance one usually has sums insured. On death, the death
sum insured is paid; on full disability the disability sum insured, and on partial
if the insurable value exceeds the sum insured, then the compensation is reduced
disability a proportion of the latter sum, determined by rules specified in the policy.
proportionally by the ratio between the sum insured and the insurable value.
Another principle is the first-risk principle. For a pure first-risk insurance the
3E. Now let y be the compensation as described in subsections 3A-C , that is ,
compensation is the loss, possibly adjusted for under-insurance. In most cases the compensation is
reduced by a deductible. There are several reasons for introducing deductibles:
We shall call r the compensation function, and its complement s defined by G(y) = 1 - [~] a. (y>c; a,c>O)
s(x) = x-1\x) This distribution is often used for claim amounts. Find the distribution of bY ( b>O)
and the conditional distribution of Y given that Y>d~c. Comment on the Pareto
the selfinsurance function. We call g(x)-1\x) the retention of the insured. We have distribution as a claim amount distribution.
of course
-10- -11-
4. Premium principles would say that a risk with a large dispersion is more dangerous than a risk with the
same mean and a small dispersion and hence should have a higher premium. Thus
it seems that some measure of dispersion should be included in the premium formu-
4A. Let X be the total (random) monetary amount of some economic risk; we la. This idea is taken care of by the two other principles. In particular we see that
shall briefly call X a risk. (In this book the word "risk" will be used in several dif- in the degenerate case Var X= 0, the risk loading is equal to zero. This seems rea-
ferent meanings, but we hope that this will not be too much confusing.) By such a sonable; when there is no uncertainty, there is no need for a risk loading.
risk we shall always mean a non-negative random variable; risks taking negative
values do not seem realistic in non-life insurance. In this chapter we also assume 4B. In this subsection we shall look at an undesirable property of the variance
that all risks have finite mean. principle. Let n be an integer greater than one, and suppose that 0 < Var X< w.
By the pure premium of the risk X we shall mean the mean of X. In practice, Then we have
one would add a risk loading to the pure premium, and then one gets the net premi-
um. When administration costs are added to the net premium, we get the gross
premium. In this book we shall for simplicity make the (very unrealistic) assumpti-
on that there are no administration costs. For the rest of this chapter "premium"
will mean "net premium". and thus
By a premium principle we mean a rule H that to any risk X assigns a non-
negative net premium H(X). We shall assume that the premium principle is a func-
tion of the distribution of X. Thus the premium H(X) is .non-random, and if the
risks X and Yare identically distributed, then H(X)=H( Y). This means that it will be profitable for the risk-holder to split the risk into n equal
We introduce the three most common premium principles: proportional policies instead of insuring it in one policy. As
(b>O) 4C. The previous subsection motivates the question: What properties do we
want a premium principle H to satisfy? In this subsection we shall discuss some
properties that might seem desirable.
iii) the variance principle
Pr(X<H(X)) < 1.
Property 2. For any risks X and Y one should have
H(X) ~ H(X+ Y). For the risk-holder this property is obvious; if the risk is almost surely less
than the premium, he will not insure. Thus, if the insurance company really wants
This property means that when one insurance alternative gives a more exten- to cover the risk, it has to make the premium at least so low that Property 4 is
satisfied.
sive cover than another alternative, then the latter alternative should not have a
higher premium than the former one. Property 4 is often replaced with the weaker condition:
Property 3. For any risk X one should have Property 4'. For any risk X satisfying Pr(X~m)=1 for some constant m, one
should have H(X)~m.
H(X) ~EX.
4D. In this subsection we are going to see which of the four properties are
This property says that the net premium should not be lower than the pure satisfied by the three premium principles presented in subsection 4A.
premium, that is, the risk loading should be non-negative. To show that this is a
i) The expected value principle.
reasonable demand, let us assume that we have a portfolio of n independent and
identically distributed risks x 1,... ,Xn, and that H(X)<EX. Then, by the law of . It is clearly seen that Properties 1, 2, and 3 are satisfied. Property 4' is not
large numbers, satisfied; for a constant risk m>O we have H(m)=(l+a)m>m and Pr(m<H(m))=l.
Thus, when the size of the portfolio goes to infinity, the probability of deficit goes to JV'ar(X+ Y) ~ .fV&IX +.;vary,
one. On the other hand, if H(X)>EX, then
which is satisfied as
lim Pr(E .~ 1 X.> nH(X)) = 0,
nT., ,_ '
v'Var(X+ Y) = v'Var X + Var Y + 2 Cov(X, Y) ~
that is, the probability of deficit goes to zero.
Property 3 will not always be satisfied in practice. Sometimes market conditi- ./ Var X + Var Y + 2 .fV&IX .fVa.IY = .jVii7{ + ..;vary,
ons make it desirable or necessary to let an insurance class make a deficit for a while
and subsidise it with surplus from other classes of insurance or other sources of pro- Consider a risk X with distribution given by
fit. However in some countries, e.g. Norway, this would in principle be illegal; the
Pr(X=m) = p = 1- Pr(X=O). (0<p<1; m>O)
Norwegian Insurance Company Act says that the premium should be reasonable
compared to the risk taken over by the insurance company.
We have
-14-
-15-
Differentiation gives 4E. This really looks bad. None of the three most common premium
principles satisfy all the desirable properties of subsection 4C, and none of them
f'(p)=m[ 1 +b(1-2p) ] satisfy Property 4'! However, in practice it is not necessarily so bad after all.
2v'P(l-PJ Firstly, to show that a property is not satisfied, we constructed some counter-
example that was not always too realistic. Secondly, insurance people are human
As beings, some of them possessing a good portion of common sense. When they see a
risk X satisfying Pr(X<H(X}}=1, they probably exclaim, "Hey, this is unfair; we
lim f'(p) = -m,
cannot charge that premium." And then they offer insurance cover for a premium
PP P<H(X} satisfying Pr(X<P)<L
there must exist a p<1 for which H2(X}>H2(m}=m. Thus, as Pr(X$m}=1, Proper- Furthermore, the properties of subsection 4C were stated under the tacit
ty 4' is not satisfied. Furthermore, as m-~0, Property 2 is not satisfied either. assumption that the insurance company is willing to cover any risk offered to it for
Property 3 is trivially satisfied. the premium assigned by a given premium principle. However, one aspect that we
have not touched, is the possibility that the insurance company can refuse an
iii) The variance principle. arrangement. This aspect may be important when discussing Property 1. One
In subsection 4B we have already shown that Property 1 is not satisfied. could assume that an insurance company using a premium principle H not satisfying
To show that Properties 2 and 4' are not satisfied, we consider the same class Property 1, would refuse to cover both of the two risks X and Y with separate pr
of risk distributions as when discussing these properties for the standard deviation rniums if H(X}+H( Y)<H(X+ Y). With such a strategy a premium principle would
principle. We have not need to satisfy Property 1.
H3(X} = mp(l+cm(1-p}], 4F. Other aspects of premium calculation will be discussed later. An exten-
sive treatment of premium principles is given in Goovaerts, De Vylder, & Haezen-
and we get that H (X}>m for m>(cp)-1. Thus Property 4' is not satisfied, and as donck (1984}; in particular they describe the Orlicz principle, which satisfies Proper-
3 ties 1,2,3, and 4'.
H3(m)=m, we see that Property 2 is not satisfied either.
Property 3 is trivially satisfied.
Exercise 4.1
Check whether Properties 1, 2, 3, and 4' of subsection 4C are satisfied by the
following premium principles:
-16- -17-
Under normal conditions the claim ratio will be between zero and one, and it is
a) The ezponential principle
often assumed to be Beta distributed with density
1 aX
H(X) = a;ln Ee . (a>O) a+fjl
g(z) =I'(r(a) I' za--1( 1-z)~1 . (O<z<1; a,{J>O)
b) The Esscher principle
This assumption is also made in the present exercise.
a) Describe the distribution of X.
aX
H(X) _ EXe (a>O) Find expressions for the net premium of X when it is calculated by:
-~
b) The expected value principle.
c) The variance principle.
c) The quantile principle
d) The standard deviation principle.
H(X) =min {h: Pr(X>h)~a}. (a>O)
Exercise 4.4
Let X be a risk. We assume that there can occur at most one claim during the
Goovaerts, De Vylder, & Haezendonck (1984) give an extensive discussion on
insurance period. Let Y be the claim amount if a claim occurs, and let p be the
the properties of various premium principles, i.a. those studied in this exercise.
probability that a claim occurs. We assume that Y is Pareto distributed with
cumulative distribution
Exercise 4.2
Find the net premium for a risk X which is Gamma distributed with density
G(y) = 1- (~)a. (y>c; a,c>O)
____t:__ a--1 -{Jx
f(x) = 1\(ij x e , (x>O; a,{J>O)
Find expressions for the net premium of X when it is calculated by:
a) The expected value principle.
when using the following premium principles:
b) The variance principle.
a) The variance principle.
c) The standard deviation principle.
b) The exponential principle.
d) The quantile principle.
c) The Esscher principle.
Exercise 4.5
Exercise 4.3
Let X be a risk with sum insured S. We assume that the risk is not under- Assume that two insurers A and B offer to insure the risk X fully or partially
by proportional insurance. The two insurers apply the same premium principle, but
insured, and that there can occur at most one claim during the insurance period.
A applies a higher value for the safety loading parameter than B. How should the
Let y be the claim amount if the claim occurs. Thus X= Y if a claim occurs, and
risk holder allocate the risk between the two insurers when the premium is calcu-
X=O otherwise. Let p be the probability that a claim occurs. We introduce the
lated according to:
claim ratio
a) The expected value principle?
y b) The standard deviation principle?
Z= -s c) The variance principle?
-18- -19-
5D. In this subsection we present four different types of treaties; two proporti-
onal ones (quota share and surplus) and two non-proportional ones (excess of loss
-
and stop loss).
In quota share reinsurance a fixed percentage is determined for each treaty,
and the reinsurer covers that percentage of each claim from the risks covered by the
treaty.
In surplus reinsurance the cedant cedes a part of the sum insured of the indivi-
dual risk, and the recoveries are calculated by the pro-rata principle. The cedant
keeps at most a specified amount, called one line, of the sum insured. Amounts in
excess of one line are ceded to the reinsurer, though these are usually bounded by a
specified multiple of the retention, e.g. ten lines. If the whole sum insured is still
not covered, the cedant has to cover the rest himself or cede to another reinsurer,
e.g. by a second surplus treaty covering. a .further number of lines above the lines
covered by the first surplus treaty.
-20-
-21-
6. Credibility theory
Execrises
(i=l, ... ,n) It is natural to compare Xn with p,; by defini~ion tt should be equal to the mean of
Xn. But if now X n is very much greater than p,, what do we do then? Should we go
b) Show that Var Zi is minimised when 1\x)=xfn. Comment.
on using p, as a pure premium? Or should we charge a higher premium? The two
extreme choices of pure premiums seem to be p, and Xn; a compromise would be a
weighted mean znXn+(1-zn)J.t. It seems reasonable to let the weight zn be increas-
ing in n; when the number of years increases, one should attach more credibility ~o
the observed average claim amount from the policy.
Now, why is actually Xn so different from J.t? Has the policyholder just been
unlucky, or is it a bad risk? But what do we mean by a bad risk? It seems that we
need some model assumptions. We start with a simple example.
6.1B. Let us consider a motor insurance policy. We assume that the annual
claim amounts x 1,x2,... are independent and identically distributed, and for simpli-
city we assume that Xj can take only the values 1 and 0. The probability that X:F1
is different for different car models; if the car is of model i (say, VW Golf), then this
probability is equal to Oi, and the pure premium for year n+1 is EXn+ =0i.
1
This procedure is OK if the model is known, as it probably would be; in most
rating systems car model is used as a rating factor. But let us assume that the
model is unknown. Then Oi is also unknown. However, we know something about
the distribution of cars. When we randomly draw a policy from the portfolio or a
-24-
-25-
new policy enters the portfolio, we could say that with probability p it is a Mazda
626, with probability p2 a Ford Granada, ... , and with probability pi 1it is a model i. we obtain
Thus we can consider car model as the outcome of a random variable I. However,
we are not as interested in the random variable I as in the random variable El=Op it
is e that contains the risk characteristics of the policy. We shall call e the risk a+ nX +1-1 fJ+n(1-Xn)-1
::: /~ 0 n (1-0) . dO
parameter of the policy and its cumulative distribution U the structural distribution
Xn+1 = a+nX -1 fJ+n(1-Xn)-1
of the portfolio since U describes the risk structure. In the present set-up, with 1~ 0 n (1-0) dO
unknown car model, we have that x 1,x2,... are conditionally independent and iden-
tically distributed given the value of e.
r(a+nXn+1) r( fJ+n(1-X))
n r(a+fJ+n) = a+fi+n'
a+nXn
It is no longer clear what is meant by the pure premium. We have to distin- I' ( a+/i+n+I) r( a+nXn)r ( fJ+n(1-Xn))
guish between two types of pure premiums. By the collective premium for year n+ 1
we mean EXn+ 1=EE>=Jt. This is the premium that we would apply if we did not that is,
know anything about the risk parameter. By the individual premium we mean
E[Xn+ 1 19}=9. This is the premium that we would apply if the risk parameter
were known.
(6.1)
After n years we still do not know 9, but we have observed the claim amounts
X1,... ,Xn, from which we can estimate e. A natural estimator is with
:::
r n . a
Xn+l = E[E>InXJ. ~n = n+a+/i' Jt = EEl = a+fi'
= f.Q, 1)
0(
i=1 J
0 n (1-0) n dU(O)
n+1 n X (1 X ) . With this interpretation we have no possi'bTt
I I y t o say anything about a new . policy
. .
J (_ll
(0,1) J=1 f(XjjO))dU(O) j On n(1-0)n - n dU(O)
(0,1) t . the portfolio. It would be better if we could say that the probability IS
en
U( 0)enng . does not excee d 0. Thus we shall say
that the 9-value of this new policy . that
Under the additional assumption that 9 is continuously distributed with density U is an underlying mathematical distribution of e, and that the El's of di~fere~t
r . al eady within or about to enter the portfolio, are independent and Identi-
po ICies,
cally r
distributed random variables With. distnbutwn
. . . U.. Thi s is by no means
. -
(0<01; a,(J>O) strange at the beginning of the previous subsection we also assumed that the .diffe~
ent claim
.' amounts of a poI'Icy WI'th known car model were independent and Identi-
-26- -27-
cally distributed. criterion we use minimisation of expected quadratic loss. This means that if we
want to estimate a real-valued quantity Y (random variable or constant}, we would
prefer the estimator Yminimising the ~xpected loss E( Y- ii within some class of
6.2. A simple credibility model estimators.
ture. Then each rating class has a structural distribution Ubeing the distribution of
m(nX) = E[m(e) I nX] a.s.
the risk characteristics among all policies satisfying the rating criteria of that parti-
Hence
cular class.
I'll
6.2B. We introduce the collective premium p,=EX of a rating class and the
Xn+l = E[m(0}1nX]
individual premium m(0}=E[XI 0].
As in the previous example, we want to estimate m(e), and as optimality
-28- -29-
is the best estimator (the Bayes estimator) of m(e) based on nX. Em( e)- a- b EX = 0
n (6.3}
We see that
E[(m(e)-a-bX )X ) =
n n o. (6.4)
Rl
Cov(m(e),X ) - b Var X =
n n o,
Thus, by replacing m(e) with Xn+l in (6.2), we see that Xn+l is the best estima-
that is,
tor of Xn+l based on nX. We shall call Xn+l the Bayes premium of Xn+l based
on nx.
b- Cov(m(H) ,In) - Var m(H) n
- Var 1n - k EVar[IIH] + Var m(H) = n+~t
6.2C. Under the assumptions of subsection 6.1B it was easy to find Xn+l'
and this estimator had a very satisfactory form. It was easy to compute and easy to with
explain to customers and insurance agents. It was increasing in Xn (the higher the
average claim amount, the higher the premium), and the weight given to Xn was ~t= EVar[X~e)
Var m( ) '
increasing in n (the more experience, the more weight to the experience). Unfortu-
nately it is not that easy in general. For many choices of U and F, the calculation and by insertion in (6.3} we obtain
Rl
of Xn+l is very messy, and the resulting formula does not necessarily satisfy the
above mentioned properties. Another problem is that U and F are usually un-
known. Therefore we somehow have to estimate these distributions or at least some
Thus,
Rl
parameters from them when estimating m(e). For the Bayes premium Xn+ this
1
can be rather complicated.
(6.5)
The linear form (6.1} had some nice properties, so why not restrict ourselves
to linear estimators? By the credibility premium l"n+l based on nX we shall mean
We see that the weight given to Xn is decreasing in EVar[XI e); the greater
the best linear estimator of m(e); that is, the best estimator in the form
dispersion in the claim amounts from one policy, the less information they give
g0 + EJ-! 1 gjXj For reasons of symmetry we must have g1=g2= ... =gn in l"n+ 1, about the risk level of the policy. On the other hand, the weight is increasing in
and we therefore have to minimise Var E(XI e), that is, the variance of the individual premium; the greater differences
between policies, the more weight we should assign to individual experiences.
Q = E(m(e)-a-bXi. We developed l"n+ 1 as the best estimator of m(e) based on nX, but as
::s
is an unbiased estimator ofVar[Xi1 16i], and thus
EXn+lXn = EXn+IXn = Em(e)Xn,
we see that the equations (6.3) and (6.4) remain unchanged if we replace m(6) with 1 N r 2
rp* = nrr;;;::::'l'l E E (x.~x.)
::s "\1'-~1 i=l j=l ZJ Z
Xn+l or Xn+l; that is, in our present model the credibility premium is the best
linear estimator (the credibility estimator) of the individual premium, the claim is an unbiased estimator of rp. The empirical variance
amount of the next year, and the Bayes premium.
We note that 1 EN (X -X )2
1T-'f i=l i ..
that is, the individual premium, the Bayes premium, and the credibility premium Var X. = !P.+ .-\,
have mean values equal to the collective premium. z r
We see that l'n+ 1 satisfies the desirable properties discussed in the first para- we see that
graph of this subsection. Furthermore, if U and F are unknown, we do not need to
::s
.A =
1
"1<T"T E Z=
N - 2
. 1 (X.Z -X ) - !...
*
make such an extensive estimation as with Xn+ ; it is sufficient to estimate the IV-L r
1
three moments
is an unbiased estimator of .A. However, it can take negative values whereas .A is
p, =EX; rp = EVar[XI 6]; .-\ = Var E[XI 6], non-negative, and we therefore replace it with the estimator ,\*=.A+. This esti-
mator is not unbiased, but it has lower expected loss than .A. The estimators p,*, rp*,
which we call structural parameters. For this estimation, assume that we have ob- and .A* are consistent as Nfro.
served a portfolio of N identical and independent policies for r (~2) years, and let
We estimate K by K*=rp*f.-\*. Insertion of p,* and K* in the credibility premi-
Xij denote the claim amount of policy i in year j. We introduce
um ~n+ 1 gives the empirical credibility premium
X . =r-1 E. r
J= 1
x,~ X.. = Prl E.Nl X .
Z ZJ '!= Z
- - - - - - - - -
-32- -33-
6.3A. Let us now consider credibility estimators under more general conditi- for some scalar d and vector d. We get
0
ons. We assume that the random vector X=(X ,... ,Xn)' is known, and that we
1
want to estimate the unknown random variable M. For this estimation we restrict
to the class of linear estimators based on X, that is, estimators in the form g +g'X
E(M-M)(M-M} = Cov(M-M,M-M} = Cov(M-M,X') d = o
0
where the scalar g0 and the nx1 vector g are non-random. By the credibility esti-
mator X! of M based on X, we shall mean the best linear estimator of M based on X, as Mwas assumed to satisfy the normal equations. Thus
using minimisation of expected quadratic loss as optimality criterion. We assume
that Cov X is finite.
E(M-M) = 0
(6.6)
Cov(M-M,X') =D.
(6.7) for all linear estimators Mbased on X; that is, Mis a credibility estimator. If M is
also a credibility estimator, then
..
and thus M=M a.s. But then M must also satisfy the normal equations, and i) is For all estimators M of M based on X we have
proved.
ii) The first equality in (6.8) follows from (6.7) as Kfis a linear combination 2 2
E(M-M} = EE[{(M-m(e))+(m(e)-M}} 1 e)= EVar[MI e)+ E(m(e)-M} 2,
of X1'" .. ,Xn. Furthermore,
Var(M-Xt} = Var M- 2 Cov(M,Xt} + Var Kf = Var M- Var Kf, and we see that minimising E(M-M} 2 is equivalent to minimising E(m(e)-M} 2. In
particular, this means that the credibility estimator of M is also the credibility esti-
completing the proof of ii). mator of m(e). By replacing e with X, we see that the credibility estimator of Mis
Theorem 6.1 is now proved. Q.E.D. the best linear approximation to the Bayes estimator E[ Ml X).
6.3B. In Section 6.2 we first considered only one insurance policy and deve-
Remarks. i) As EKf=EM, the expected loss E(M-Xt} 2 is equal to Var(M-Xt}.
loped a credibility estimator for this policy based on its observed claims. Then we
ii) Let 'Yo and 'Y be coefficients of the credibility estimator Kf, that is, used claims from the other policies in the same portfolio to estimate structural para-
meters. A natural question is then, could we improve the credibility estimator by
Kf = 10 + 1 'x. (6.9) basing it on claims data from the whole portfolio, and not only data from the indivi-
dual policy? Intuitively one would say no; data from the other policies are indepen-
Then the normal equations (6.6) and (6.7) can be written as dent of the policy considered and thus irrelevant. In the following theorem we prove
this more formally.
'Yo= EM- 1'EX (6.10)
1'Cov X= Cov(M,X'). (6.11) Theorem 6.2. Let X=(Xi,:JS)', where~ is independent oJX and M. Then
1
Kf depends on X only through x .
1
If Cov X is invertible, then (6.11) gives
Proof. Let
1' = Cov(M,X') Cov-1X, (6.12)
Kf = 'Yo + 1i x1 + 12~
and insertion of (6.10) and (6.12) in (6.9) gives
M= E[KfiX1) ='Yo+ 11x1 + 12Ex2.
E(M-M) = 0 (6.6')
rf -0
Pr(XZ=xJ 9 Z=0) = ;;:r
X.
e . (~0,1,2, ... )
f(x)
ox e- od U( 0)
= Pr(X =x) = f ;;:r r*(x) = (x+l) ~(~Jl
(~0,1,2, ... )
Z (O,m) X.
with
t).z = (X.+
z (1-()Ee
with
- f Oxe OdU(O).
(O,m) and, using that
we see that would contact an engineer and get his opinion on the risk, and perhaps also seek
other sources of information. After this phase of collecting information, he would
A= r* -~-t* make an estimate of the pure premium based on his present knowledge. That is, a
pure Bayesian method.
is an unbiased estimator of A = Var E[XI 0]. As A should be non-negative, we esti-
A
6.4D. In Sections 6.1-2 we discussed situations where we wanted to determine
mate it by A*=A+ and propose
the insurance premium for a risk, utilising the claim experience of that particular
risk. In these set-ups, a risk with a bad claim experience would get a higher premi-
"*-A*
~--:r um than a similar risk with a much better claim experience. The practice of making
the premium depend on the claim experience of the individual policy is called experi-
as estimator of (. We now get the empiricallinearised Bayes estimator ence rating. Credibility theory was originally developed within the context of expe-
rience rating, but has later found other applications both within and outside the
tl't: = (*X.+ (1-(*)~-t* insurance area. Another approach to experience rating is the bonus systems in
~ ~
motor insurance, which we shall discuss in Chapter 7.
It should be emphasised that in the approach we take to experience rating in
this book, the goal of experience rating is to fit the premium to the risk level of the
We note that (*~0, and thus t.li is non-decreasing in Xi, which is intuitively policy. We shall not consider experience rating as a way of encouraging the policy-
~
reasonable. On the other hand, 07 would typically have a ragged appearance. Fur- holder to be more careful, although this might be a relevant aspect in practical in-
~ surance. Furthermore, experience rating should not be considered as a punishment
thermore, it would be difficult to generalise e i to situations where the Xi's do not to a policyholder for having claims. It is important that the insurance company
have identical design. These are some arguments for preferring empiricallinearised explains this to the policyholder. Otherwise the policyholder may consider it unfair
Bayes estimators to empirical Bayes estimators. when the premium increases after a claim; he has already paid a premium for cover-
From the present example one clearly sees the two-stage nature of empirical ing such claims, and now he has to continue to pay for it in the future!
Bayes methods. At the first stage one constructs a Bayes method; at the second As mentioned above, in practice experience rating could encourage the policy-
stage one estimates structural parameters and inserts these estimates in the Bayes holder to be more careful. A related aspect is that when the policyholder suffers a
method. small loss caused by an event covered by the policy, then he will sometimes omit to
make a claim as he considers the increase in future premiums too high compared to
6.4C. Should one give the distribution U a subjectivist (pure Bayesian) or a the claim amount. This effect of an experience rating scheme, is called bonus hun-
frequentist (empirical Bayesian) interpretation in insurance applications? I would ger. It is very relevant in practice, but unfortunately it is complicated to model,
say both. The answer should depend on the nature of the available data. In an , and in my personal opinion there does not yet exist a satisfactory treatment in the
application where we have several different samples with different risk parameters, actuarial literature. An interesting contribution is Norberg (1975). Sundt (1989)
the actuary should clearly use the empirical Bayes set-up. However, the actuary discusses some of the problems related to modeling bonus hunger. We shall not
often has to make decisions with few or no objective data. To quote Norberg discuss bonus hunger further in this book.
(1979), the actuary cannot say to a customer, "Your gas-tanker is, of course, a most
interesting object of insurance, and we look forward to negotiate the terms as soon 6.4E. When should one apply experience rating? When answering this questi-
as the hazard can be assessed from objective facts, say in 10 years or so." In such a on, we have to distinguish between the empirical Bayes approach and the pure
situation the actuary would ask for more details about the ship and its use. Then he Bayes approach.
-42- -43-
6.4F. For a general treatment of empirical Bayes theory we refer to Maritz &
Lwin (1989), and for pure Bayes theory e.g. DeGroot (1970) or de Finetti (1975).
of l"n+ 1 considered as an estimator of m(8n+l).
The credibility estimator given by (6.5) has been criticised for giving the claim (6.15)
amounts from all previous years the same weight; intuitively one should believe that ~
new claims should have more weight than old claims. However, nothing in the (6.16)
model assumptions indicates that new claims should be more important than old
ones. The following model is an attempt to overcome this intuitive deficiency.
We assume that Xl'X2,... are conditionally independent given an unknown
-44- -45-
(6.17)
and (6.16) is thereby proved.
It remains to show that '1/Jn satisfies the recursion (6.15). By (6.8) we have
where a 110 ,an1, and an2 are constants. That we actually can write ~n+ in the
1
form (6.17), is proved if we can determine a ,an , and an such that the normal
110 1 2
equations are satisfied.
From (6.7) we obtain
'Y-=5.-y-i..J...1
nJ J n,J,
-46- -47-
5.=
J In the simple model of Section 6.2 we assumed that the risk volume was the
same for all years. In particular in reinsurance one wants to allow for varying risk
volumes. For that purpose Biihlmann & Straub (1970) introduced the following
and insertion of (6.15) gives model.
We consider a ceded insurance portfolio. Let Sj denote the total claim amount
and Pj some measure of risk volume in year j, e.g. the direct insurance premium; in
(6.20)
subsection 10.6A we shall discuss some other measures of risk volume. By the loss
ratio of year j we shall mean XJ=Sj Pi We assume that x ,X ,... are conditionally
1 2
From (6.15) we see that independent given an unknown random risk parameter e that characterises the
portfolio, that
and hence we can divide numerator and denominator in (6.20) by '1/Jj to obtain
independent of j, and that
Var[X I e) =
J ~J
. (6.21)
The reasoning behind the assumption (6.21) is perhaps most clearly seen if Pj
is the number of risks in the portfolio in year j. If we assume that the claim
amounts Y a_, . , Y. p of the P. risks in year j are conditionally independent and
J J, j J
On the other hand it is clear that 5?0, and thus 0<5{1. From (6.16) we see that
identically distributed given e, then
0<1'nn <1, and we get (6.19). Q.E.D.
(1975). He had portfolios from different states of the USA. Let us consider one of
(6.22)
these portfolios, and let X. be the average amount of one claim in year j from that
J
portfolio. Because of inflation etc. we are not willing to assume that EXj is inde-
with
pendent of j. Instead we make the regression assumption EX]YjfJ. where the
design vector y j is known; one could e.g. have EX;/31+j{32 .
P= E.n P; 1 n
X =pE PX; When estimating the vector f3 for different states, Hachemeister found conside-
J= 1 J n J= 1 J J
rable differences. He therefore assumed that to each state there was related an un-
Proof Let Xn+ 1 denote the right-hand side of (6.22). We want to show that known random risk parameter e representing the risk characteristics of that state,
and that es from different states were independent and identically distributed.
~n+l=Xn+1' Again considering one particular state, we assume that E[Xjl 8]=yjb(8) with
In the special case described above, it follows from the theory of Section 6.2
Eb(G)=/1.
that the credibility estimator of m(e) based on the Yj/s (i=1, ... Pi
j=1, ... ,n) is
It is clear that this assumption can also be used in other situations, both with-
equal to X +1. As this estimator depends on the Y ..s only through the X !s, it in and outside the insurance area. In Hachemeister's case it was natural to interpret
n Jt J
must also be the credibility estimator of m(e) based on nX. j as calendar time; we considered the inflation at some calendar time. However,
From this result it is natural to believe that ln+ 1=Xn+ 1 also in the general there are also cases where it is more natural to let j be the age of an insurance poli-
cy. Then the development in E[X I e] can be interpreted as a learning effect. That
case. To prove this, it is sufficient to show that .Xn+ satisfies the normal equati- J
1 interpretation seems apt in motor insurance; the driving abilities of the policyholder
ons (6.6) and (6.7). We obviously have that E(Xn+l-m(8))=0. Furthermore, for develop over time.
j=1, ... ,n we have
6.7B. After these motivating remarks we shall now state our model assump-
tions in more detail.
Let X=(X1,... ,Xn)' be an observed random nx1 vector and e an unknown
random risk parameter. We assume that
E[XI e] = Yb(e),
This completes the proof of Theorem 6.5. Q.E.D.
where the nxq design matrix Y is known and has full rank q~n. We introduce
The model of Section 6.2 appears as a special case of the present model by
letting the risk volume P}1 for all j.
A= Cov b(e); += ECov[XIGJ; P= Eb(e),
For estimation of structural parameters in the present model, we refer to
Biihlmann & Straub (1970) and Sundt (1983b).
assuming that the matrices A and + are positive definite.
with
b= (Y'9-1Y)-1Y'9-1X
(6.24) and insertion in (6.27)
Z = AY'+-1Y(l+AY'9-1Y)-1.
(6.25)
Proof Let
where 'Yo is a non-random qxl vector and r a non-random qxn matrix. From (6. 7')
we obtain
with b given by (6.24), and we get
r Cov X= Cov(b(0),X').
(6.26)
Inserting
that is,
Cov X= ECov(Xj0] + Cov E(Xj0] = 9 + Cov(Yb(0)) = 9 +YAY'
Cov(b(0),X') = Cov(b(0),E(X'j0]) = (Cov b(0))Y' = AY' rX=Zb (6.29)
r(++YAY') = AY', Let M be an unknown random variable, and assume that M is conditionally ~
independent of X given 9, and that
that is,
E(MI 0] = a'b(S).
r+ = (I-rY)AY'.
Then Remark iii) after Theorem 6.1 shows that the credibility estimator of M is
Multiplication from the right by first 9-1 and then y gives equal to the credibility estimator of a'b(0), and by Corollary 6.1 this estimator is
equal to a' I). If X represents the observed average claim amounts from n past years
r = (I-rY)AY'9-1 as in Hachemeister's case, M could be interpreted as the average claim amount from
(6.27)
rY = (I-rY)AY'9-1Y. a future year.
(6.28)
Solving (6.28) for ry gives 6. 7C. In this subsection we shall give an interpretation of b.
From Theorems 6.6 and 6.7 we see that the credibility estimator of b(9) is a 6.8A. In subsection 6.2A we briefly discussed the subdivision of an insurance
weighted mean of the best linear 9-unbiased estimator of b(9) and the mean of portfolio into rating classes. We said that the unknown random risk parameter of
b(9). an insurance policy represented the risk characteristics of the policy that were not
captured by the rating criteria that were applied, and we discussed why such charac-
6.7D. It is often assumed that teristics were not incorporated in the rating structure. One reason was that a too
detailed rating structure could complicate estimation of structural parameters. Let
-1 us look a little more closely at this aspect.
~ = t,oP ' (6.30)
To illustrate the ideas, we look at the simple credibility model of Section 6.2.
Let X. be the claim amount of an insurance policy from the jth year it is in force. It
where 10 is an unknown scalar parameter and P is a known matrix. Under this as-
sumption we get is ass~med that the X !s are conditionally independent and identically distributed
given an unknown random risk parameter e. The risk parameters of the individual
policies in the rating class considered are independent, and their risk parameters are
Z = AY'PY(t,oi+AY'PY)-1. identically distributed with common distribution U. The structural distribution U
will of course depend on how the rating class is defined.
In particular we note that the expression for b does not contain unknown parame- In subsection 6.2C we showed that the credibility estimator ~n+ 1 of Xn+l is
ters. given by
Hachemeister (1975) assumed that X1'" .. ,Xn were conditionally independent
given e with conditional variances
(6.31)
p,=EX; rp = EVar(XI 6]; >.. = Var E(XI 6] for a good empirical credibility estimator.
If,= rpf >..; X =n-1 E. n x. One way to utilise more rating criteria, but at the same time have more data
n J= 1 J available for the estimation of structural parameters, is to introduce a hierarchical
model. To apply such a model, we should be able to assume that the different
The expected loss of l is
n+l rating classes in a way have something in common.
Let us look at a simple example. We consider the class of motor insurance
policies of policyholders living in Mannheim and assume that the structural distribu-
tion U belongs to a parametric class 11 = { U( ;71): 71EII} with some parameter set II.
The actual parameter value 71 is unknown. Analogously we can consider policy-
We introduce the following normed versions of 1n ).. and' . 1
"' ' '~'n+l" holders living in Ludwigshafen and assume that the structural distribution for this
class also belongs to 11 with parameter value 712, which is also unknown. We are not
IPo = rpf J.t2; >..o = >../ J.2 willing to assume that 71 and 71 are equal. On the other hand, we believe that the
1 2
2 1 two towns have so much in common that we could use data from one of them for
7/JO,n+l = 1/Jn+1/J.t = IPo + n 1 ;
rate-making in the other one. Quite analogously to the introduction of random risk
IPo + A()
parameters 6 for individual policies, we therefore assume that 711 and 712 are the
and can now express as values of random variables H and H that are independent and identically distri-
K 1 2
buted. Thus we shall assume that to each community there is related an unknown
random risk parameter H characterising the community, and that risk parameters
from different communities are independent and identically distributed. Given the
value of the risk parameter H of a community, the policies from that community are
We see that K and 7/Jo n+ 1 are independent of the "average" risk level of the rating
class. T"ll ' that when the rating structure is made more detailed th conditionally independent and their risk parameters 6 are conditionally identically
yp1ca Y we have
>..0 will decrease as the sub-portfolio of the rating class becomes more homoge~eo:s~ distributed. We see that we now have random risk parameters on two levels; 6's on
On the other hand, it is often realistic to assume that rp, is approximately constant. the policy level and H's on the community level. We shall call credibility mooels
Under these conditions, 7/Jo n+1 will decrease; we get Oa better estimate of X with random risk parameters on more than one level hierarchical credibility models.
when the sub-portfolio bec~mes more homogeneous. Furthermore, K will incr::S!
6.8B. Let us now define more formally our hierarchical generalisation of the
~d n/(~+~t) decrease; when the sub-portfolio becomes more homogeneous, the indi-
credibility model of Section 6.2. We consider a motor insurance portfolio consisting
VJ~ual d1fferences between the policies will decrease such that we can give more
of policies from different communities. Claims data from different communities are
weight to the mean J.t. We can interpret this to mean that we should have a detail-
ed rating structure. However, in practice one does not apply the credibility premi- assumed to be independent. Each community is characterised by an unknown ran-
dom risk parameter, and we assume that risk parameters of different communities
um ln+ 1 but an empirical credibility premium
are independent and identically distributed. We concentrate on the sub-portfolio
consisting of policies from one particular community with risk parameter H. The
sub-portfolio consists of N policies that are conditionally independent given H. Poli-
cy i is characterised by an unknown random risk parameter 6 i and has been ob-
wh.ere ~t* and J.t* are estimators of K and J.t. These estimators will often become less served for ni years. We assume that the 6 /S are conditionally independent and
reliable _the more detailed we make the rating structure, as we get less data from identically distributed given H. Let \j denote the claim amount of policy i in year
each ratmg class; the desire for a good credibility estimator conflicts with the desire j. We assume that Xil'Xi , ... are conditionally independent and identically distri-
2
-56- -57-
buted with common distribution F( I O,TJ) given the values ()and 1J of 0 i and H. based on X.
Let
Proof Let M denote the right-hand side of (6.32). From Theorem 6.1' we see
m{0i,H) = E[Xijl 0i,H); J.t(H) = E[m(ei,H)IHJ; J.t=EX
rp = EVar[Xi)0i,H); A= EVar[m(ei,H)IHJ; e = Var J.t(H).
that M is equal to&: if it satisfies the normal equations (6.6') and (6.7'). We im-
mediately see that (6.6') is fulfilled. We want to show that (6. 7') also holds. We
Compared to the non-hierarchical model we see that we now have an additional have
e
structural parameter representing differences in risk level between communities; J.t
represents the over-all risk level for the whole portfolio, rp the fluctuations between Cov(M,X') = r Cov X + Cov(70,X') =
claim amounts from an individual policy, and A the differences between policies r(ECov[XIHJ + Cov E[XIH)) + Cov(70 (H),X') =
within the same community. ECov[M,X'IHJ + r Cov E[XIHJ + Cov(E[M-fXIH),E[X'IH)) =
Let ~k n +1 be the credibility estimator of Xk based on the observed ECov[M,X'JH) + Cov(E[MJH),E[X'JH)) = Cov(M,X'),
. 'k ,nk+ 1
clatm amounts. In the non-hierarchical model, this estimator depended only on data
which proves Theorem 6.8. Q.E.D.
from policy k. However, as this policy now depends on the other policies from the
same community, data from these policies also appear in ~ In the hierarchical model presented in subsection 6.8B, the expected condition-
k,nk+1"
al first and second order moments of (X,M) given H had the same structure as the
corresponding unconditional moments in the corresponding non-hierarchical model
6.8C. Before developing an expression for ~k , we want to present a
. ,nk+1 of Section 6.2. This is often the case with hierarchical credibility models. They are
vanant of Theorem 6.1' for hierarchical credibility models. Let X be an observable developed as generalisations of non-hierarchical models, and the structure of the
random nx1 vector, Man unknown random rx1 vector, and Han unknown random expected conditional first and second order moments given the hyper-parameter H is
variable. the same as the structure of the corresponding moments in the corresponding non-
hierarchical model. By comparing Theorems 6.1' and 6.8, we see that in such clses
Theorem 6.8. The credibility estimator ti of M based on X can be written in the coefficients of X in the credibility estimator ti also have the same structure in
the form
both models. Furthermore, the constant term in the credibility estimator in the
non-hierarchical model considered as an expectation is in the hierarchical model
replaced by the credibility estimator of the corresponding conditional expectation
(6.32)
given the hyper-parameter H. From these considerations we conclude that we can
where the non-random rx n matrix r satisfies the normal equation utilise results from the non-hierarchical model when developing credibility estima-
tors in the hierarchical generalisation. This technique can be useful even if we have
r ECov[XI H) = ECov[M,X' IH) not previously developed credibility estimators in the non-hierarchical model as it is
often easier to handle this model than the hierarchical generalisation.
risk parameter characterising the county. Such multi-level hierarchical models have
been treated by Sundt (1980) and Norberg ( 1986); the latter paper also discusses the
with estimation of structural parameters.
Exercise 6.2
We assume that the life-time T of a light-bulb is exponentially distributed
with
with parameter 0, i.e. the life-time distribution is continuous with density
a) Find the conditional expectation of T-t given that T>t. Comment on this
It is interesting to observe that in the estimation of p(H), the credibility weights
result.
(1, ... ,(N take the position of measures ofrisk volume. ,
The parameter 0 depends on the quality of the glow thread and does not have
A similar hierarchical extension of the credibility regression model of Section
6. 7 is described by Sundt (1979b). the same value for all bulbs. We assume that in a production of light-bulbs, the 0-
value of a randomly chosen bulb is a realisation of a random variable e, which is
Gamma distributed with density
6.8E. From the discussion of Theorem 6.8 it should be obvious how to treat
hierarchies with more levels. We could e.g. assume that data from different commu-
nities in the same county are conditionally independent given an unknown random ( 0>0; a,,B>O)
-60- -61-
b) Find under this assumption the conditional expectation of T-t given that
(y>c; a>O, O<c<Q
T>t. How would you explain to a non-statistician why this expectation increases
with t?
and Nis Poisson distributed with parameter 0.
c) How would you estimate the parameters a and {J by using the observed life-
times T1,... , Tn with independent realisations of 9?
Exercise 6.4
Assume now that we have a sample of N light-bulbs that have the same value
of e. These bulbs were lit at the same time. At time t, n of them have burnt out We consider the situation in subsection 6.1B. Show that if 1-n+l=~n+ a.s.1
and this happened at times t , ... ,tn. for all n, then the structure distribution U has to be either continuous with density
1
d) How would you estimate the value of e? of the form
Exercise 6.3
u( ())- r( a+f1 oa--1(1-0)~1 (0<0<1; a,{J>O)
Let N be the number of claims incurred in an insurance portfolio during a -r(a)l' )
given year, and let Yi be the amount of the ith of these claims. We assume that the
or concentrated at one point.
Y/s are mutually independent and identically distributed and independent of N.
We introduce the accumulated claim amount
Exercise 6.5
We consider an insurance portfolio with N independent policies. Policy i has
risk parameter 0-EIR, and its claim amount in year j is Xij We assume that
Xil'xi ,... are inlependent and identically distributed with mean m(Oi) and vari-
a) Show that 2
ance 'If... Oi).
We randomly draw a policy from the portfolio. Let I denote the number of
EX=ENEY; Var X= ENVar Y + E2 YVar N. this policy. Then
b) Find expressions for EX and Var X when N is Poisson distributed with Pr(l=i) = ~- (i=1,2, ... ,N)
parameter 0, that is,
be _the aggregate claim amount of year j. We assume that N ,N ,... are conditional- This exercise is based on Sundt (1991a).
1 2
ly mdependent and identically distributed given a random variable e, and that the
indi~idual claim amounts Yij are mutually independent and identically distributed Exercise 6.9
and mdependent of e and the claim numbers. In subsection 6.2C we concluded somewhat informally that because of symmet-
a) Discuss the model assumptions. ry, the credibility estimator depended on the observations only through their ave-
b) Find the credibility estimators x( 1 ) x( 2 ) and x( 3) of X based on rage. In the present exercise we shall look more closely at this sort of symmetry.
. n+ 1' n+1' n+1 n+1
respectively nX, N, and ( X, N). We shall say that n random vectors X1'" .. ,Xn of the same dimension are ex-
n n n
c) Show that changeable relative to a random vector Y if Xi ,... ,Xi ,Y have the same joint distri-
1 n
bution for all permutations (i ,... ,in) of (1, ... ,n).
1
E(X -l"( 2 ))2 < E(X -A'( 1 ))2
n+l n+l - n+1 n+1 Show that if X1' ... ,Xn are exchangeable relative to (M,X ), then (X ,X) with
0 0
Is this result intuitively reasonable? X= n-
1
1::;
1 Xjis linearly sufficient for Mbased on (X ,xl'"
0
.. ,Xn).
The concept of relative exchangeability was introduced in Sundt (1979a),
Exercise 6. 7 where the result of this exercise was also proved. A different proof was given in
Sundt (1991a).
Let Y and M be random sx1 vecors and X a random nx1 vecor, and let ~
denote the credibility estimator of M based on X. We assume that y is independent Exercise 6.10
of M and X. Show that the credibility estimator of Y' M is equal to (EY)' ~- We assume that we have observed an insurance portfolio with n policies for N
years. Let Xij denote the claim amount of policy i in year j. Each policy i is
Exercise 6.8 characterised by an unknown random risk parameter e i and each year j by an un-
Let X be an observable random nx1 vector, M a~ unknown random variable, known random risk parameter Hj We assume that the ei's are mutually independ-
and T a non-random rxn matrix. We say that TX is linearly sufficient for M based ent and identically distributed, that the H !s are mutually independent and identic-
on ~ if the credibility estimator of M based on X depends on X only through TX; ally distributed, and that the 9 i's are inde~endent of the H.'s. The X. !s are condi-
eqmvalently we can say that TX is linearly sufficient for M based on X if the credi- tionally independent given the 9 's and the H 's, and the c~nditional dfstribution of
r s
bility estimator of M based on X is equal to the credibility estimator of M based on Xijgiven the Sr's and the H8's is in the form F( I ei,H}.
TX. a) Comment on the model assumptions.
a) Discuss the concept of linear sufficiency in credibility estimation and give b) Find an expression for the credibility estimator of Xk based on the
,n+ 1
examples of linear sufficent statistics in the theory of Chapter 6. observed claim amounts. (Hint: The deduction can be simplified by application of
b) Let 9 be an unknown random variable and m a function. Show that TX is the result of Exercise 6.9.) Consider in particular the cases:
linearly sufficient for m(S) based on X if E[XI TX,Sj cam be written in the form i) Var E[X11 jH1] = 0.
ii) N= 1.
E[XjTX,ej =a+ ATX, iii) N _, 111.
Comment on these results.
where the vector a and the matrix A are non-random. This exercise is based on Sundt (1979c).
c) Use the result in b) to prove Theorem 6.2.
tl) Use the result in b) to prove the result in Question c) in Exercise 6.6. Exercise 6.11
Sometimes when one has tried to apply credibility theory for experience rating
-64- -65- l
I
in practice, it has been argued that the credibility premium gives individual experi-
ences too much weight; the policyholders will not understand that the premium
should fluctuate that much from year to year. Thus the insurance companies not
only want the premium to give a good fit to the risk profile of the policy, but also
e) Show that
that it should not fluctuate too much from year to year. In this exercise we shall
look at an optimality criterion that takes this criticism into account.
We consider an insurance policy. Let X. be a random vector that represents
J
the observations that we make concerning claims that occur for this policy during (Hint: Use the result in Exercise 6.7.)
the jth year that the policy is in force, and let M be an unknown random variable This exercise is based on Sundt (1992a).
that we want to estimate each year. Let ll!tn+l be the credibility estimator of M
Exercise 6.12
based on X 1,... ,Xn, and let 1/ln+l denote its expected quadratic loss
Let X and e be non-negative random variables; X integer-valued. Conditional
on 8= (}, X is Poisson distributed with parameter 8, that is,
r1 -{}
a) Show that J(xl 0) = Pr(X=xl 8=0) = Xf e . (x=0,1,2, ... )
We assume that E[e IX] is linear in X, that is, there exist constants c and d such
(n=O,l, ... )
that
and discuss this result.
E[8IX=x) =ex+ d. (x=0,1,2, ... )
b) Deduce upper bounds for Pr(l M-ll!tn+ll >E) and Pr( I Mn+l-Mnl >E) for
some E>O, and discuss these probabilities. Let U denote the cumulative distribution of e.
Let vll!tn+l be the linear estimator Mn+l based on x 1,... ,Xn that minimises Show that U must be either continuous with density of the form
( 0>0; a,{J>O)
for a given weight vE[O,l); we let vll!tl =EM. or concentrated at one point.
c) Discuss this optimality criterion.
Let Y be a random variable independent of the X.'s and M with distribution Exercise 6.13
z Let x ,x ,... be conditionally independent and identically normally distri-
given by 1 2
buted random variables with mean e and variance ifJ given a random variable e,
which is normally distributed with mean p, and variance >.. Show that the conditi-
Pr( Y=l) = 1- Pr( Y=O) = v.
onal distribution of e given nX is normal with mean _+n
nK-nX + nK-
+K- p, and variance
d) Show that
_!LwithX =!E.n x.andK-=Ifi/A. Comment.
n+K- n n J= 1 J
-66- -67-
Exercise 6.14 In Section 6.6 we showed that the credibility estimator of m(e k) based on the obser-
The log-normal distribution is often used as a distribution for claim amounts. ved loss ratios is given by
We say that a random variable X is log-normally distributed with parameters p, and
a if In X is normally distributed with mean p, and standard deviation a.
a) Find expressions for the density, mean, and variance of the log-normal
distribution with parameters p, and a.
with
Let X1,X2,... be conditionally independent and identically log-normally dis-
tributed with parameters e and a given a random variable e which is normally
distributed with mean 1J and standard deviation r. Let m(e)=E[XI e) and Z:F
Xjm(e).
b) Show that the Z !s are mutually independent and independent of e.
J
Xk = .J....
E .nl Pk.Xk'
. rk J= l ::J
Comment on this result.
~ a) Show that the best linear unbiased estimator of p, based on the observed
c) Find the credibility estimator ~n+l and the Bayes estimator Xn+l of
loss ratios is given by
m(e) based on nX. Comment.
N -1 N
Exercise 6.15 p, = (E ,_
._ (.) ._ (.X
E ,_ ..
1 ' 1 t t
In this exercise we use the assumptions and notation of Section 6.5. Show that
1/Jn converges to a limit when n goes to infinity and find an expression for this limit. By the homogeneous credibility estimator ..:tk,n+l of m(ek) based on the ob-
Discuss this result.
served loss ratios we mean the estimator in the form E .N E .n g .. X .. that mini-
This exercise is based on subsection 4.2.2.4 in Sundt (1981 ). t= 1 J= 1 'J 'J
rnises the expected quadratic loss E(m(ek)- E .'!_ E .n g. X ..)2 under the con-
,_ 1 J= 1 t) tJ
Exercise 6.16 straint.
In Section 6.6 we proved Theorem 6.5 by first motivating an expression for the
credibility estimator and then checking that that expression satisfied the normal N n
E. IE. 1g .. =l. (3.33)
t= J= 'J
equations. Prove Theorem 6.5 by solving the normal equations.
b) Motivate the constraint (3.33).
Exercise 6.17 c) Show that
We consider the model studied in Section 6.6. We have N independent ceded
portfolios that have been observed for n years. Let Xij be the loss ratio and Pij the
risk volume of portfolio i in year j. We assume that Xil'Xi ,... are conditionally
2
independent given a random risk parameter e i that characterises the risk of portfo-
d) Let
lio i, and that
v( e.)
E[Xijlei) = m(ei); Var[X .. Ie)=~
'J ' ij
Em(ei) = p,; Var m(ei) = ..\; Ev(ei) = cp. Show that
-68- -69-
* =~E.
1 N E. n P. (X . ~X . )2 v(e .,H)
1n
.,.. 1
"\n-~, 1
z= J= ~J ~J ~ E(Xi)E>i,H) = m(8i,H); Var(X .. je.,H) = ~
~J ~ ij
P N -2 .
).. = }:2
- E.
N ?,. (E._ 1 P. (X. -X) -(N-1)tp)
~- ~ ~
E[m(E>i,H)jH) = v(H); Ev(H) = J.l
~=1 ~- EVar(m(ei,H)IH) =>-.; Ev(8i,H) = tp; Var v(H) =e.
are unbiased estimators of 1fJ and>-.. Would you estimate).. by>-.? a) Find an expression for the credibility estimator of m(e k,H) based on the
This exercise is based on Biihlmann & Straub {1970). Sundt {1983) generalises observed loss ratios. Consider in particular the limiting cases where egoes to zero or
the estimators of the structural parameters to the case where the ceded portfolios do infinity. Comment.
not need to be observed for the same number of years. b) Make an analogous hierarchical extension of the regression model and cre-
dibility estimator studied in Section 6.7.
Exercise 6.18
a) Generalise Theorem 6.3 to the situation where
E(X.je) = m(e)
J J
EVar(Xjl 9) = tpi
).,.+1 = p).. ..
~ ,J J ~J
Exercise 6.19
Show that Theorem 6.5 is a special case of Theorem 6.6.
Exercise 6.20
In this exercise we look at a hierarchical generalisation of the model studied in
Section 6.6. We consider N ceded insurance portfolios within the same geographical
area. We assume that the portfolios are conditionally independent given a random
risk parameter H that characterises the geographical area. The portfolios have been
observed for n years. Let Xij be the loss ratio and Pij the risk volume of portfolio i
in year j. We assume that Xi ,xi , ... are conditionally independent given H and a
1 2
random variable 9 i that characterises the risk of portfolio i, and that
-70- -71-
7B. In bonus systems applied in practice, the bonus rules sometimes say that
one has to stay for a certain number, say n, of claim-free years in the second highest where
bonus class before ~ne can proceed to the highest class. At first glance it may seem
that such bonus rules do not fit into the frame-work that we have introduced above. PT,r/.i,J) = Pr(ZR,n+l=jj(8=1J)n(zR,n=i)) = Pr(T(i,M)=jj8=1J).
However, such a bonus system is easily redefined to fit into our frame-work by split-
ting the second highest bonus class into n bonus classes with the same premium. We introduce the conditional distribution of ZR
,n
When a policyholder enters this set of n bonus classes, he starts in the lowest of
them. Then after each claim-free year he advances to the next higher class, and
after n claim-free years he is in the highest bonus class.
( n)
PR o= (PR( n)1'11' 1) ,... ,pk( n)(K))'
IJ
' ' '
with
7C. We consider one particular policy. Let Mn denote the number of claims
and Xn the total claim amount of the policy during insurance year n. It is assumed
that the pairs (M1,X1),(M2,x2), ... are conditionally independent and identically
distributed given the value of an unknown random risk parameter e characterising
the policy. The cumulative distribution of e is denoted by U. The individual pre- The phn~s can be calculated recursively by
mium for yearn is '
-72- -73-
Em w p(n)(j)l(n~
n=1 n l l,J
7E. When using minimisation of expected quadratic loss as optimality criteri- E m w p(n)(j) .
on, we clearly want for each n
n=1 n l
to be as small as possible. For fixed R this gives the optimal bonus scale for year n
Intuitively one would expect that ~R 4 >~R 3 as class 4 includes bad drivers
from classes 1 and 2 having had one claim-free ye~r whereas class 3 consists of dri-
vers from the better classes 5, 7,9,11, and 13 having had bad luck in one year. This
is also confirmed by numerical examples in Horgan, Haem, & Norberg (1981). But We call iR the credibility scale and
~R, 4 >~R, 3 contradicts the desirable property 11'j+ 1S11'/or all j. One possible way to
overcome this problem is to change the bonus rules; Norberg (1976} has made some
suggestions under a different optimality criterion. Another solution is to leave the
Bayes scale and consider more restricted classes of bonus scales. One natural class
seems to be the linear scales satisfying the credibility risk of the bonus rules R.
To get a linear bonus scale 11'.F a-bj decreasing in j, one should have b>O, and
11'= a- bj to get positive premiums one should in addition have a>bK. If these two conditions
J
are not satisfied by the credibility scale of some given set of bonus rules, then pro-
for some constants a and b. Another reason for restriction to linear scales is that bably something is unreasonable with the bonus rules, and one should search for
they are easily understood by agents and policyholders; the premium is reduced by b other rules.
for each class one moves upwards, and increased by b for each class one moves An interesting question is of course, how much do we lose by applying the
downwards. The scale that was actually used in this bonus system, was linear, credibility scale instead of the Bayes scale for given bonus rules? To answer this
satisfying 1r.F(1.6-0.lJ)11'(6), that is, a=l.61r and b=0.11!' . question we introduce the credibility efficiency
6 6
under the bonus rules R. Numerical examples in Gilde & Sundt (1989) give rela-
tively high values of the credibility efficiency, indicating that we do not lose much
with respect to a and b. This is a well-known situation from Chapter 6, and Theo- by applying the credibility scale instead of the Bayes scale.
rem 6.1 i) shows that the optimal choice of a and b is
7H. When discussing the simple credibility model in subsection 6.2C, we em-
phasised that one of the advantages of working with credibility premiums instead of
Bayes premiums was that with credibility premiums we did not need a full specifica-
tion of the structural distribution and the conditional distribution of the claims
Thus for given R the optimal bonus scale is given the risk parameter; we only needed moments of first and second order, and
these moments could be estimated from empirical data. Unfortunately it is not that
simple with credibility scales of bonus systems. For calculating the credibility scale
and the credibility risk of the bonus rules R we still need only moments of first and
with second order; that is, EZR N' Em(0)=EX, Var ZR N' and Cov(m(e),ZR N)=
Cov(XN+l'ZR,N), and these ~aments can be estimated from portfolio data if k are
the bonus rules that have actually been applied in the portfolio. However, if we
-76- -77-
want to calculate the credibility scale and the credibility risk of another set of bonus much wider class of experience rating schemes than the area of bonus systems, and
rules R', it is a problem that these moments depend on R'. therefore there do not necessarily exist bonus systems with efficiency close to 1. But
In this case, one possibility is to make a complete specification of the joint if all the bonus systems that one tries out, have low efficiency, one should perhaps
distribution of El, N, and the (Mi,Xi)'s, from which we calculate the necessary consider leaving the area of bonus systems and use e.g. the credibility scheme of
moments. If these calculations are complicated, stochastic simulation could be an Section 6.2 instead.
interesting alternative.
If we have observed not only the ZR 's, but also the age of each policy in the 7J. In subsection 7C we assumed that the (Mi,Xi)'s were identically distribut-
,n
portfolio and the number of claims for each year the policy has been in force, then ed, but the set-up can easily be generalised to situations with e.g. learning effects.
we can calculate the ZR, ,n's by (7.1), and from these values we can estimate the For further details we refer to Borgan, Hoem, & Norberg (1981) and Gilde & Sundt
moments like in the case of the bonus rules R. (1989).
An extensive study of the theory of bonus systems is presented by Boos
7!. As has already been mentioned, when looking for good bonus rules, one (1991).
has to rely on the trial-and-error method. By computing the risk (i.e. the expected
loss) of different bonus systems one gets a measure of how much better one system
is than another. However, this comparison does not show how much there is to gain
by evaluating further systems. If we break out of the restrictions of bonus systems
as defined in subsection 7A, we know that the best we can do for year n+ 1 is to
apply the Bayes premium
with risk
2
Q0 ,n = E(m(El)-fhn+1) = EVar[m(El)i nX].
N
% = E(m(El)-fhN+l) 2 = EVar[m(El)l,yXJ,
N
which we shall call the absolute Bayes risk. It is clear that Q(S)>% for any bonus
system S, and thus we have obtained a. lower bound for the risk of bonus systems.
We introduce the efficiency
e(S) = %/ Q(S)
of the bonus system S. One should keep in mind that % has been deduced in a
-78- -79-
B. Multiplicative rating models can therefore often give a better approximation to reality than an additive rating
model. However, multiplying parameters is slightly more complicated than adding
them. We therefore use the transformation
8.1. Introduction
In Chapter 6 we discussed briefly the choice of rating factors; that is, obser-
vable characteristics of the risk that we use to determine the premium. In motor
which can be written in the form
insurance we could have rating factors such as district, mileage, car model, age of
the driver, age of the car, and individual experience rating (bonus system); in home
K
insurance, district, size, value, age of the house, and building material. Q.1 . = v + Ek=l/1k 1. '
1 1K ' k
Let us assume that we apply K rating factors, and that factor k has a finite
number Ik of levels (k=l, ... ,K). For a risk with level ik of factor k (k=l, ... ,K), the and obtain
premium is P; ; . It is clear that even with a reasonable number of rating factors
'tK
and levels, the number of possible values of P. . could become high, and we Q. .
1 11 1K
1""" 1K P.1 . =a b .
could get a voluminous rating book. The rating model should preferably be easy to t'K
apply and interpret. Thus we need some sort of structure so that we do not need a
separate entry for each combination of rating factors. Instead of displaying tt and the aki's in the rating book, one often displays v and the
The simplest solution would presumably be an additive rating model; that is, {1k!s. In addition, a table of abx is given. The numbers a and bare chosen so as to
the premium is determined by gi;e a reasonable magnitude to v and the /1ki's. Each level of each rating factor gets
a number of points. The insurance agent finds the premium of a risk by looking up
the premium corresponding to the total number of points of the risk. The advan-
tage of including the base number tt is that by adjusting this number, one can adjust
the overall premium level, e.g. for inflation, without adjusting the parameters for
Then the insurance agent could add up the parameters for the appropriate levels of
any of the rating factors. We see that the premiums remain unchanged if we multi-
the rating factors, and the rating structure could simply consist of the base number
tt and the parameter ak,ik for each level ik of each factor k. We see that the pretni- ply 'each parameter ak . by a constant ck and divide tt by 11~1 ck.
'k
ums remain unchanged if we add a constant ck to each parameter a . and subtract The search for simple rating structures will presumably become less important
k,zk in the future when rating books are gradually replaced with pre-programmed calcu-
K
Ek=l ckfrom J-t. lators.
Although an additive rating model seems attractive from a practical point of
view, it is not necessarily that attractive as an approximation to the risk structure
of the portfolio. Intuitively one would believe that factors like mileage in motor 8.2. The method of marginal totals
insurance and size of the dwelling in home insurance would rather have a multiplica-
tive than an additive effect. A multiplicative rating model, The method of marginal totals is a simple, intuitively reasonable method for
simultaneous determination of parameters in a multiplicative rating model approxi-
mating pure premiums.
-80- -81-
Let ni ... i denote the total number of observed risk years and x. . the Often one does not want to adjust the parameters for all the rating factors in a
. 1 K z1... zK
'"'"1 ,... ,K) . A reasonabl e
total clrum amount of all policies with level ik of factor k ("= revision of rates. If we only want to adjust the parameters for factors l, ... ,L, we can
restric.tion when fitting a multiplicative structure to observed data, is that, when consider the parameters for the remaining factors as given and solve the equations in
sumrmng over all levels of all the other factors, the total fitted premium for all (8.2) for k=l, ... ,L.
policies with level ik of factor k (k=1, ... ,K) should equal the total observed claims We see that the method of marginal totals does not impose any structure on
for these policies, that is, the parameters to be determined. Thus, the method can produce results that could
seem unreasonable to insurance agents and policyholders. In motor insurance, one
would for instance want a parameter for mileage to be increasing, but this property
is not guaranteed by the method of marginal totals. In practice the parameters
would often be adjusted by judgement. One way of determining the final values of
the marginal totals should be exact. We rewrite (8.1) as the parameters could be to apply first the method of marginal totals to all the rat-
ing factors. Then one could adjust the parameters for one or more of the factors by
judgement and consider these parameters as given when applying the method of
marginal totals for a revised determination of the parameters for the remaining
factors, and so on, until one has constructed a reasonable set of rates. Not very
scientific, but it should be emphasised that constructing a set of rates is not quite
the same as statistical estimation. One does not necessarily want to apply the rat-
where we have put tt=1 without loss of generality. Unfortunately these equations ing structure that gives the best description of the risk structure of the portfolio. In
cannot in general be solved explicitly. However, solutions can be found by iteration. addition to good description of the risk structure, there are other considerations that
When a solution is found, it can be rescaled as described in the previous section if should be taken into account:
desired i) Consistency. The motor insurance premium should for instance be increas-
It is perhaps easier to see how the method works, if we look at the special case ing in mileage. Consistency could be taken care of by imposing more structure on
K=2. With I=I1, J=I2 , ai=a1i, and fl:Ja 2i we can rewrite (8.1) as the parameter values, for instance by saying that the parameter should be an increa- \o
sing function of the level. However, the method of marginal totals does not allow
J J for structure on the parameters so that one would then have to apply other
EJ==. 1 n.P .. = E. X .. ( i=1, ... ,1)
- ZJ ZJ J== 1 ZJ methods, e.g. least squares estimation.
I I
E. n.P .. = E. X ii) Simplicity. The rating structure should be easy to apply and interpret.
Z= 1 ZJ ZJ Z=1 ij (j=1, ... ,J)
That was the reason for concentrating on a simple structure like a multiplicative
and (8.2) as rating model; the risk structure of the portfolio is presumably not multiplicative.
iii) Fluctuations. The difference from the previous rates should not be unrea-
sonable.
( i=1, ... ,1) iv) Market conditions. The insurance company has to consider how much the
market is willing to pay. In particular, one should consider the rates of competitors.
We conclude that when determining insurance premiums in practice, a statis-
(j=1, ... ,J) tical estimation procedure will often give the decision-maker a point of reference
rather than the final answer.
It should also be emphasised that when fitting a multiplicative rating model
-82- -83-
by a method such as the method of marginal totals, we determine the premiums for
insurance cover in the future by fitting these premiums to observed claims from the
past. Hence, although we have not made a proper definition of a probabilistic It is assumed that X .1'" .. ,X. are conditionally independent given e i' and that
z z,ni
model, there seems to be an implicit assumption that the observed claims from the
past are comparable to the unknown claims of the future. In particular we have not
E[X--10-]
ZJ Z
= m~e.);
Z Z
Var[X ..I0] = v(0.)/P
ZJ Z Z ZJ
taken future inflation and other expected changes in the claim pattern into account,
Var m~e .) =A; J.z = Em~e.)
z z = y'.p.
z
and the final premiums will normally have to be adjusted for such effects. z z
The method of marginal totals can also be applied to other rating structures
In the last expression, yi is a qxl design vector based on observable technical data of
than multiplicative ones, in particular for additive rating models.
the car model; in Uni Storebrand they apply (1, engine power, price/weight)' for
The method of marginal totals and some other methods for determination of
vehicle damage and (1, engine power, weight)' for third party liability. We intro-
the parameters in multiplicative and additive rating models have been discussed by
duce
van Eeghen, Greup, & Nijssen (1983). Sundt (1992a) discusses the relation between
insurance rating and statistical estimation.
rp = Ev(e); K = rpf A
n. -1 n i -1 n i
p.=E. lp.;
z X.=p. E. p.X .. =p. E. s.,
Z J= ZJ Z Z J= 1 ZJ ZJ Z J= 1 ZJ
8.3. Classification of passenger cars
We see that when considering one car model separately, our model assumpti-
8.3A. In our presentation of the method of marginal totals we did not make ons are analogous to the assumptions of Section 6.6. However, when considering the
any model assumptions about the mechanisms which generate the claims. We often data from all the car models simultaneously, the present model is more general since
want to make such assumptions, thus including relevant insight we may have. Pre- ' in the model of Section 6.6 we assumed that Ernie) had the same value for all i.
ferably we should consider the rating factors simultaneously as we did in the previ- Thus parameter estimation will be different in the present model; for details we
ous section. However, when we make more detailed model assumptions, this soon refer to subsection 3.3 in Sundt (1987).
becomes messy. Therefore we often consider only one rating factor at the time and From Theorem 6.5 we see that the credibility estimator of mi( e i) is
regard the parameters for the other factors as given. To illustrate this, we shall look
at an approach for classification of passenger cars in motor insurance which is appli-
ed by the Norwegian insurance company Uni Storebrand. The approach can be used
m..=
z (X+
z z (1-(.)J.
z z
both for classification of new car models and for adjusting the parameters for car
with
models that have been classified earlier.
We consider a group of Ncar models. It is assumed that risk data from differ-
ent car models are independent. Furthermore we assume that car model i is charac-
terised by an unknown random parameter ei, and that 01' ... ,0N are independent
and identically distributed.
The estimation error
We apply a multiplicative rating model, and car model is one of the rating
factors. For car model i we have applied the car model factor r. if this car model
z
z E(m~e.)-m.)
2
has been rated previously. We have observed ni risk units (policies) from car model '1/J= z z z = A(I-(;)
i. Let Sij denote the total claim amount from the jth of these risk units during the
exposure period and P .. the premium. We introduce gives the decision maker an impression of the reliability of the estimate.
ZJ
-84- -85- -~
mi is not immediately comparable to the old car
The credibility estimator 8.3B. Car model seems to be a natural rating factor for motor insurance. In
vehicle damage insurance, one obvious reason is repair costs. That motivates the
model factor ri; the scale is wrong. Therefore we want to multiply mi by a scaling
1
use of price in the design vector; an expensive car would normally have high repair
factor "f If we had applied the factor costs. 1
However, repair costs is not the only reason for using car model as a rating
factor. We also mentioned engine power as a technical variable; a more powerful
car could be driven at higher speed and could therefore be more exposed to acci-
in our observed portfolio, then the total premium income would have been dents. From this one should not conclude that a more powerful car in itself is more
'E .N p .pN A reasonable way to determine "f is to say that this premium income dangerous than a less powerful car. A person who considers buying a new car,
~= 1 ~ ~
should not necessarily conclude that it is safer for him to buy a less powerful car as
should be equal to the actual premium income
a more powerful car would be more dangerous to him; perhaps it could actually be
the opposite as with the more powerful car he would have the power available when
he needs it. However, the enigine power of the car could indicate something about
the driver. It is normally rather expensive to insure GTI variants of small car
This gives models; such cars are usually bought by persons who like to drive fast.
From this discussion it should be clear that premium differentiation between
car models is also to a large extent classification of drivers. From this we realise
that the class of possible relevant technical variables is much wider than it would
have been if it had only been a question of classifying the car in itself. For instance,
In Uni Storebrand there are 70 classes for car models in motor vehicle damage the number of doors could be an interesting variable; a four doors variant of a car
insurance. These classes are numbered from 30 to 99, and the factor for class c is model would to a larger extent be a family car than a corresponding two doors
equal to 1.04 c-30 . Thus, when we have found the factor pi, we transform it to a variant. The colour of the car could be relevant; could the driver of a black car be
different from the driver of a bright red car?
ln p.
class 'Ci, letting 'Ci be the integer between 30 and 94 closest to 30 + InTlk As the buyers of the cars are so important for the risk level, one should not
rely blindly on the technical formula when performing the initial classification of a
As argued in Section 8.2, the class ci
should not be the final answer, but a
new car model. By studying photos of the car, an experienced person could get an
point of reference for the decision maker who determines the final class ci' In par-
impression of what sort of people would buy such a car. Therefore, the initial statis-
ticular, one does not want the new class to deviate too much from the old one.
tical classification should be adjusted by subjective judgement.
When the final classification of the car models is completed, the number
From the above, it should be clear that car model is a relevant rating factor
not only for vehicle damage insurance, but also for third party liability insurance.
However, as repair costs are less relevant for third party liability, it is natural that
the price of the car is not used as a technical variable for this insurance form.
is calculated. This is the factor by which we have to multiply the base number of 8.3C. A natural extension of the present model would be to introduce a
the rating structure to obtain the same premium income as before in our observed hierarchical model, assuming that data from different car models of the same make
portfolio. are conditionally independent given an unknown random risk parameter character-
ising the make; the drivers of a Mercedes Benz could be different from the drivers of
---------------------- ----
-86- -87-
an Opel. This extension is discussed in Section 4 in Sundt (1987). P .. =a-+fl~ (i=l, ... ,I; j=l, ... ,J)
'J ~ J
We shall look at some different methods for determination of the factors a. and {J
' is Pois-
a) Show that if the accumulated claim amount of a policy in cell (i,J) J
son distributed with mean afJj then the method of marginal totals produces maxi-
mum likelihood estimators of the a.'s and {J !s.
' minimises
By the least squares method one J
Q1 =E. I E. J 2
r- 1
n.(Y.~a.{J.).
~= 1 'J 'J ' r
b) Show how the als and fljs can be determined by the least squares method,
and show that this method produces maximum likelihood estimators when the accu-
mulated claim amount of a policy in cell (i,J) is normally distributed with mean afJj
and variance u2.
By the Bailey-Simon method one minimises
c) Show how we can determine the als and fljs by the Bailey-Simon method.
d) Consider the three methods applied to an additive model, that is,
-89-
-88-
that is, 1-pnn(s,) is the conditional cumulative distribution of Tn+l given that
9.2A. Let N*(t) be the number of claims incurred in the time interval (O,t).
We clearly have that N*(t) is right continuous and non-decreasing in t, and that it N(s)=n.
-90-
-91-
lim Pr(N(t+h)>n+IIN(t)=n) _
hlO h - 0 (9.2)
-p (s t) [Pm,m+l(t,t+h) + Pr(N(t+h)>m-fi!N(t)=m)] From the following lemma and assumptions i) and iii), it follows that the right-hand
mm ' h h '
side of this expression is continuous in t, and thus
and from assumptions ii) and iv) we obtain
(9.6)
Since from assumptions i) and iii) the right-hand side of this expression is continu-
ous in t, we must have Proof The right continuity follows as Pmn(s,t) is right differentiable in t.
For O<Mt-s we have
(9.3)
-92- -93-
9.2D. We now return to the general assumptions of subsection 9.2B and intro-
duce T =0. The interoccurrence time (waiting time) between claim number n-1
0
and claim number n is
which from {9.5) goes to zero when h goes to zero. Thus Pmn(s,t) is left continuous,
and the lemma is proved. Q.E.D.
For n>m we have Pmn(s,s)=O, and from this and {9.6) we obtain
if Tn- <ro; if Tn-1=ro, we could e.g. let Wn=O. We have
1
giving a recursive determination of Pmn(s,t). Insertion of {9.5) gives by the Markov property, and we have
{9.8)
{9.10)
which could also have been seen by direct reasoning.
In the_ special case where >.n( t)=>.( t) inde~endent of n, the process { N{ t)} t>O is The claim number process is called homogeneous if >.n(t)=>.n independent oft
called a Pozsson process. We introduce - for all n. For a homogeneous claim number process, {9.10) reduces to
A(s,t) = J! >.(r)dr,
and from {9.5) and {9.7) we obtain by induction and, using the Markov property of the claim number process, we see that if >.n+ 1>0
for all n, then the interoccurrence times are independent and exponentially distr~
n-m( buted; if in addition the claim number process is a Poisson process, the interoccur-
p (s t) = A s t) e-A(s,t)
mn ' (n-m).1 (O~s~t; O~m~n)
rence times are also identically distributed.
We see that
Theorem 9.1. The claim number process {N{ t)} t>O is homogeneous if and only
if for all nand O~s~t, Pnn(s,t) depends on sand t only through t-s.
k
Pr(N(t)-N(s)=kl N(s)=m) = Pm,m+k(s,t) = ~ e-A(s,t). {9.9)
Proof Let us first assume that
is independent of m. This implies that N{ t)-N{ s) is independent of N{ s); that is, a
Poisson process has independent increments. Furthermore, the distribution given
by {9.9) is the Poisson distribution with parameter A(s,t), and thus we see that the
independent increments of a Poisson process are Poisson distributed. A great part depends on sand t only through t-s. Then, using {9.5), we get
of classical risk theory is based on the assumption that the claim number process is
a Poisson process.
-94-
-95-
Theorem 9.2. There exists an operational time for {N( t)} t~O if and only if
there exists a non-negative function 'Y such that
which depends on sand t only through t-s, and Theorem 9.1 is proved.
Q.E.D.
J'Q 'Y( t)dt = m (9.12)
Theorem 9.1 shows that if the claim number process is homogeneous, then the
transition probabilities pnn(s,t) depend on sand t only through the difference t-s. and for all n and t
By using induction on (9.8) we see that this property holds more generally for all
the transition probabilities Pmn(s,t). (9.13)
9.2E. Often it is easier to work with a homogeneous claim number process with ,. n independent oft.
than with an inhomogeneous one. In this subsection we shall prove a result intro-
duced by Biihlmann (1970) showing how one can in some cases transform an inho- Proof. Let us first assume that there exists a function 'Y satisfying (9.12) and
mogeneous claim number process into a homogeneous one. (9.13). We want to show that
Let p be a continuous, non-decreasing function satisfying
1
p- ( r) = inf{ t: p( t)=r}.
1
We see that p- is non-decreasing, and that
Let
which depends on r 1 and r 2 only t hrough. T 2-r~ Hence by Theorem 9.1
{ N( . )} is homogeneous. and p is an operatiOnal time.
r,~o;~~et p be an operational time with corresponding claim intensities "n
-97- I
-96- ! I
I I!
Thus (9.13) holds with "(=p', and from (9.11) this 'Y satisfies (9.12).
This completes the proof of Theorem 9.2. Q.E.D.
The practical impact of operational time is perhaps most clear for Poisson
that is,
processes so let us assume for a moment that >.n( t)=>.( t) independent of n. Then
the probability that a claim should occur in the time interval ( t, t+dt], is >.( t)dt, and
this event is independent of the claims history up to time t. Thus >.( t) could be
interpreted as a measure of the risk exposure in the portfolio at time t. We could
for instance have >.( t)=k( t)>. with k( t) denoting the number of policies in the portfo-
We want to show that the second integral in (9.14) is equal to zero. As
6
lio at time t. When using p( t) = f k( s)ds as operational time, the increase in ope-
rational time in the time interval ( t,t+dt] is k( t)dt, which is proportional to the
number of policies in the portfolio. Thus the operational time moves more quickly
when there are more policies in the portfolio.
(9.15)
Intuitively we would believe that P11lm(s,t)=1, but there are processes for which
P11lm(s,t)<1 for some s,t, and m. In such cases we interpret 1-P11lm(s,t) as the pro-
bability Pr(N( t)=wl N(s)=m), that is, the probability that the process "explodes" in
the time interval (s,t]. In risk theory we usually assume that P11lm(s,t)=1 for all m,
Thus (9.15) holds, and (9.14) reduces to s, and t.
Theorem 9.3. If
(9.16}
Proof By Theorem 9.2 there exists a function 'Y and constants K. such that
n and if (9.19) holds, then for all m,s, and t we have Pmn(s,t)=l.
Let us now assume that P mn(s,t)=1 for all t~s. Then, by monotone conver-
I'Q 'Y( t)dt = m
(9.18) gence we obtain from the second inequality in (9.20)
,... m 1
L.Jn=1 -;;-= m.
n (9.19) As this inequality holds for all t, we must have
Thus we see that for a homogeneous claim number process there exists a t such that EX(t) = EE[X(t)jN(t)] = E[N(t) EY],
p
mw'(s t)<l if and only if lim E[Tn IN(s)=m] < w. The following example shows
nTw
that this result does not hold in general. that is,
Furthermore,
and thus ETn="' for all n. However, as for all r 9.3B. Let us now assume that {N( t)} t>O is a Poisson process with claim inten-
sity A(t). Then we call {X(t)}t>O a compound Poisson process. As {N(t)}t~O has
., ..
""'n=1
1 (
) ., .. -2
DrJ = r+1 ""'n=1 n < "' independent increments, {X( t)} t~O has independent increments too .
n With
9.3A. Let Yi denote the amount of the ith claim. We assume that Y , Y ,...
1 2
are non-negative, mutually independent, identically distributed with common distri-
bution F, and independent of the claim number process. These assumptions are
from which we obtain
very strong. For instance they do not allow for inflation. However, they simplify
the mathematics. For a more general treatment we refer to Biihlmann (1970).
EN(t) = Var N(t) = A(t).
Hogg & Klugman (1984) discuss various classes of claim amount distributions.
Let X(t) be the accumulated claim amount in the time interval (O,t], that is,
Insertion in (9.21) and (9.22) gives
9.3C. Let us assume that the claims can be of m different types, and let K.
denote the type of the ith claim. It is assumed that the (Ki, Yi)'s are mutually ind~
pendent and identically distributed and independent of the claim number process.
We denote by Fk the cumulative conditional distribution of Y given that K=k and with
introduce
From Theorem 9.5 it follows that the processes {X1(t)}t>O, ... ,{Xm(t)}t>O are
independent, and that {Xk( t)} t>O is a compound Poisson process with claim intensi-
ty qk>.(t) and cumulative claim -amount distribution Fk. By combining these results
We obviously have
with (9.25) and (9.26) we obtain the following theorem. It
N(t) = Ek:-1 Nk(t) Theorem 9.6. Let {X 1( t)} t>O'" .. ,{ Xm( t)} t>O be independent compound Poisson
X(t) = Ek:-l Xk(t). processes. The process {Xk(t)}t;O has claim intensity 1Jk'Y(t) and cumulative claim
(9.26)
amount distribution Fk" Let -
Furthermore, we have the following result.
Theorem 9.5. The processes {N1(t)}t>O, ... ,{Nm(t)}t>O are independent, and
{Nk(t)} ~O is a Poisson process with claim intensity qk>.(t). - Then {X( t)} t~O is a compound Poisson process with claim intensity 1J'Y( t) with
Proof. We easily see that the m-dimensional process {(N (t), ... ,Nm(t))}t>O
1
has independent increments. Furthermore, with n = Ek~\ nk we have -
and claim amount distribution
-104- -105-
1/J(O) = 1. (9.28)
The assumption about the claim intensities means that the m processes have For u>O we have
the same operational time. We see that the claim amount distribution of {X(t)}t>O
is a weighted average of the claim amount distributions of the m processes with (9.29)
weights proportional to the claim intensities. This result holds in general for sums
of compound Poisson processes. However, in the general case the weights, and If the player wins the first game ( v =1), then after the game the reserve of the
1
hence also the weighted claim amount distributions, could depend on time. casino becomes u-1 and the ruin probability 1/J( u-1). On the other hand, if ihe
player loses ( v =-1), then the reserve becomes u-1 and the ruin probability
1
1/J( u-1). Insertion in (9.29) gives
9.4. Ruin theory
1/J(u) = p'I/J(u-1) + (1-p)'I/J(u+l). (9.30)
9.4A. In this section we shall discuss ruin theory, a topic about which it has
been said that never have so many people written so much about such a small pro- Thus we have a second order difference equation (9.30) in u and one constraint
bability. (9.28). In subsection 9.4B we shall prove Lundberg's Inequality, which gives an
To clarify ideas, we start with a simple example. We consider a roulette upper bound for 1/J( u). As this bound goes to zero when u goes to infinity, we have
game. The casino has an initial reserve of u DEM (u being a positive integer), and the second constraint
there is only one player. In each game this player puts 1 DEM on red. If red comes
out, he gains 1 DEM; otherwise he loses 1 DEM. We want to find the probability lim 1/J(u) = 0. (9.31)
1/J( u) that the casino at some time is ruined, that is, the probability that its reserves
uT.,
at some time are less than or equal to zero. By solving (9.30) and using the constraints (9.28) and (9.31) we obtain
Let Vi be the net loss of the casino in game i. We assume that v1, v2,... are
independent and identically distributed with distribution given by
With e=0.01 this gives u~86; that is, if the casino does not want to have its ruin
probability greater than 1 %, then it needs an initial reserve of at least 86 DEM.
and the probability that this reserve becomes less than or equal to zero, is
9.4B. We shall now discuss ruin probabilities in a more general setting. Let
v1, v2,... be independent and identically distributed random variables with cumula-
tive distribution Hand moment generating function p Vl and let
We clearly have
-106- -107-
(u~O)
Analogous to (9.33} we have
We see that 1/J is the tail of the distribution of max {Sn: n=0,1,2, ... }.
By conditioning on v1, we get the following generalisation of (9.30} .,p (u) = If(u-) + f .,Pn_1(u-v)dH(v); (u>O; n=1,2, ... ) (9.36)
n (--m,u)
1
(9.35) and this equation has the solutions R=O and R=ln(p- -1). As p<t, the second
solution is positive. Insertion in (9.35) gives
Proof We introduce
-108- -109-
and by comparison with (9.32), we see that in this case Lundberg's Inequality holds Under the present assumptions the claim number process {N( t)} t>O is called a re-
with equality. newal process. It is a Markov process if and only if the interoccurrence times are
exponentially distributed; in that case {N( t)} t>O is a homogeneous Poisson process.
9.4C. In our roulette example it was fairly simple to evaluate 1/J( u) exactly. We assume that the premium is paid continuously with constant rate P, that is, in
Unfortunately it is not that simple in general. However, under the assumptions of each time interval of length t, the premium income is Pt. Let u denote the initial
Lemma 9.2 we showed that reserve of the insurance company at time 0 and U(t) the reserve at time t~O, that is,
U(t) = u+ Pt-X(t).
V.=
~
Y.-PW..
~ ~
9.4D. We shall now apply Lundberg's Inequality to an insurance portfolio in
some different situations.
As the V/s are independent and identically distributed and
i) Let Yi denote the amount of claim i and w. the interoccurrence time be-
~
tween
.
claim i-1 and claim i. It is assumed that the pairs ( Y1' W1) ,( Y.2' W.2) , ... are
mdependent and identically distributed, and that the Y/s are non-negative. Corre-
sponding to our previous notation we introduce we can apply the theory of subsection 9.4B.
-110- -111-
This inequality says that the expected premium income between two claims should
P>>.EY. (9.42)
exceed the expected amount of one claim. We also assume that there exists an R>O
satisfying
From (9.23) we see that the expected accumulated claim in any time interval of
length tis equal to >.t E Y. As the premium income in such a time interval is equal
py(R) = 1. (9.39) to Pt, (9.42) implies that in any time interval the premium income should exceed
the expected accumulated claim, that is, to obtain a ruin probability less than one,
Then Lundberg's Inequality gives
we should have a positive risk loading. This is consistent with Property 3 of subsec-
tion 4C.
1/J(u) S e-Ru.
Insertion of (9.41) in (9.40) and some reorganisation give
As
1+R~-py(R)=O. (9.43)
Pr(WSw) = 1- e->.w, ii) In the previous case we looked at the probability of ruin on the set of all
positive time points. We shall now study the probability of ruin on a countable set
and we easily obtain of time points.
Let X. denote the total claim amount in year i. The premium income is P
each year. ;tis assumed that x ,x ,... are independent and identically distributed.
1 2
-112- -113-
Let We now leave the assumption about normality. Let Ni denote the number of
claims incurred in year i and Yij the amount of the jth of these claims. Then
p, =EX; tp = Var X.
We assume that P>p,. The initial reserve at time 0 is equal to u, and the reserve
after n years is
It is assumed that N is Poisson distributed with parameter A, and that the Yijs are
mutually independent and identically distributed and independent of Ni' Let r N
denote the probability generating function of Nand p y the moment generating func-
tion of Y. We have
Let
A(py(r)-1)
px(r) = e .
with R being the positive solution of Insertion in (9.44) and some reorganisation give
p
1 + R X- p y( R) = 0.
(9.44)
where PX is the moment generating function of X; we assume that such a solution But this equation looks exactly like (9.43). It is therefore interesting to compare '
exists. Like in the continuous case we have that R is non-decreasing in P. the situations under which these two equations were found. Let vh
u) and 1/Jd( u)
For a large portfolio of relatively similar policies it could be .reasonable to denote the ruin probabilities studied in the continuous and the discrete case respec-
approximate the distribution of X with a normal distribution. Then tively. We see that in the continuous case X(1),X(2)-X(1),X(3)-X(2), ... are inde-
pendent and identically distributed with the same distribution as Xi in the discrete
2 case. Furthermore, in the continuous case the premium income in one year is P.
p,r+!Ef-
px(r) = e . Thus we can move from the continuous to the discrete case by putting
We see that in this case the adjustment coefficient is decreasing in tp. This is intui-
tively reasonable; the larger variance, the more risky is the portfolio.
-114- -115-
9.4E. We have defined the ruin probability 1/J( u) as the probability that the Er(Z)/E w. The initial reserve of the insurance company at time zero is u, and the
reserve of the insurance company at some time becomes less than or equal to zero. reserve at time t is
It may seem more logical to define it as the probability that it becomes strictly less
than zero. That definition is used in most of the literature on ruin theory. In most U(t,r) = u+ Prt-X(t;r).
cases of practical interest we consider continuously distributed variables, and then
the two definitions give identical results (except for u=O). However, in our roulette We are now within the frame-work of case i) of subsection 9.4D, and the ad-
example the second definition would have given the lower ruin probability justment coefficient R( r) for the ruin probability
(p/{1-p))u+l.
1/J(u;r) = Pr{Ut~o( U(t;r)~O))
9.4F. In our presentation of ruin theory we have not taken interest into
is the positive root of the equation
account. This may be a rather serious deficiency when making such long-term con-
siderations. Ruin theory incorporating interest has been studied by Delbaen &
pr(s)cp(P rs) = 1 (9.45)
Haezendonck {1987).
The insurance company applies a compensation function r. Thus the accumulated Proof By (9.45)
claim amount of claims incurred in the time interval {O,t] is
Theorems 9.7 and 9.8 both show that if pr ~Pr on [O,ro), then the compensati- Pr(r*(Z)~y) = 1 ~ Pr(r(Z)~y).
1 2
on function r1 is at least as attractive for the insurance company as r . Theorem
2
9.8 says that as r1 is at least as attractive as r , the insurance company could
2
Hence Theorem 9.9 follows from Lemma 9.4. Q.E.D.
charge at least as low a premium for r1 as for r without increasing the ruin proba-
2
bility. On the other hand, Theorem 9.7 says in particular that if the company were Theorem 9.10. Let
to charge the same premium for both compensation functions, then it would have at
least as low a ruin probability with r as with r . r**(z) = I(z~d)z.
1 2
As ex is strictly convex, the following two lemmas are trivial consequences of
Ohlin's Lemma and Corollary A.2. Then p~pr** for all non-decreasing compensation functions r satisfying
Er(Z)=Er**(Z).
Lemma 9.4. Let r1 and r2 be two compensation functions satisfying
Erl (Z)=Er2(Z). We assume that there exists a y such that Proof We clearly have
0
From Theorems 9.9 and 9.10 we see that among non-decreasing compensation
Lemma 9.5. Let r 1 and r 2 be two non-decreasing compensation functions functions, the insurance company would prefer a first-risk cover, whereas a franchise
satisfying Er1(Z)=Er2(Z), and assume that there exists a z such that
0 deductible is least favourable. It is interesting to note that a first-risk cover cuts off
the tail of the loss distribution whereas with a franchise deductible it is only the tail
r 1 (z)~r (z)
2 that is covered. Thus the corollaries indicate that the tail is what the insurance ~
r 1(z)~r (z).
2 company should fear. In Chapter 12 we shall show that the insurance company has
analogous preferences when using maximisation of expected utility as its optimality
criterion.
Theorem 9.9. Let r*(z)=min(z,S). Then p~pr* for all compensation functions
r satisfying Er(Z)=Er*(Z). 9.6. Choice of reinsurance form
Proof For y<S we have 9.6A. Let us now reinterpret the model of the Section 9.5 so that Zi is the
amount of the ith claim, of which the company retains r(Zi) after reinsurance; we
Pr(r*(Z)~y) = Pr(Z~y) ~ Pr(r(Z)~z), shall call r the retention function. Under this interpretation we can apply the re-
sults of the previous subsection to the choice of reinsurance form. In particular,
and for y~S
Theorem 9.9 shows that a pure excess of loss cover is at least as attractive to the
cedant as any combination of excess of loss and quota share with the same pure
reinsurance premium.
-119-
9.6B. From the result about excess of loss and quota share one should expect
We shall now, keeping the
9~:~ection an~
a similar result for stop loss and quota share. For the former result we worked with- . among global reinsurance forms. . ranee form, there corre-
in the frame-work of case i) of subsection 9.4D; for the latter we enter case ii). in subsection 9.6B, show that to local re;nc::ant as the local form.
Let Zi denote the total claim amount in year i. We assume that z ,z , ... are frame-work of s h'ch is at least as attractive to th . d C the amount of
sponds a global form w I b f claims incurred in year z an ij
independent and identically distributed, and that the moment generating function
1 2 of Let N. denote the num er o
z io finite on [O,.J. Let ~Z;) be the "'""'' doim '"'oomt of'"' i .., P, "'' the jth of t~ese claims. Then
retained premium, that is, the total premium less reinsurance premium. We assume
the Pr>Er(Z).
thatreserve after nThe is reserve of the insurance company at time zero is u, and
initial
years
H.
"' z
Zi = .uj=l c.,
ZJ
Theorem 9 11
To any local reinsurance there corresponds a global form
r~:; t form
t an adjustment coef.J.czen .
with at least as grea f and intro-
is the positive root of the equation
. of a local reinsurance orm
Proo.f Let r be the retention function
t'
-P s duce a global form WI.th retention func Ion
pr( s)e r == 1.
E[E~l r{Cl)IZ1=z].
J~l
with pr denoting the moment generating function of r(Z). -
We see that we are within the mathematical frame-work of Section 9.5 with >"(') - ' b th '""'"'"" fo>mo,
ntis the same tor 0 . ._
W:l, and thus all the results of that subsection are easily adapted to our present As the expected total retained clru.m am:e introduce the moment generatmg fui:ti
tained premmm.
situation. In particular Theorem 9.9 shows that a pure stop loss cover is at least as they have the same re fficients R of the local
attractive to the cedant as any combination of stop loss and quota share with the N Z) d the adjustment coe
same pure premium. 1 ...t clj) and
ons p of Ej=l'\ . p* of r*(From
an Jensens's Inequ art
I y we obtain
and R* of the global remsurance.
Z1 ~ Ee sE[r:~!11(Cl)lz1]
9.6C. Unfortunately we cannot immediately conclude from Theorem 9.9 that
the cedant would prefer stop loss reinsurance to excess of loss. We can only say that E .1
p(s) = Ee s J=l
N r{Cl} = [ Hl
EE esEj=l'\ 1
...tc}ll
stop loss is the optimal reinsurance form among all forms for which the total retain-
ed claim amount for one year is uniquely determined by the total claim amount of sr*(Z1) *( )
that year. We shall call such reinsurance forms global. For excess of loss reinsur- E ~p
' ' Q.E.D.
ance a retained amount is calculated separately for each claim. Such reinsurance 9 7 shows that R*~R.
forms are called local. A quota share treaty can be considered as local as well as and Theorem f is at
globol. In '""'"'Uon 9.6A we oo..,;d.,OO optimality """ng i<><ai "'"'"""'" bn.,;
By Theorem 9.8 we can al so prove that is
th cedant the global
as willing reinsurance orm
to pay at least as much for
least as attractive as the local form, e
-120- -121-
(n=1,2, ... )
Use this result to interpret the result in Question b).
with a~O. This model is called the contagion model, and we distinguish between d) Generalise Questions a)-c) to the case where { N( t)} t>O is an inhomogene-
positive contagion (/1>0), no contagion (/1=0), and negative contagion (/1<0). We ous Poisson process with intensity >.( t). -
introduce
Exercise 9.4
_ >. n+1. p(h) = e-l.Bih_ We consider a simple mortality risk. Let >.(t) be the rate of mortality at age t.
'Yn- Tlfl'
Show that there exists an operational time for the claim number process, and find
a) Discuss the contagion model. the transition probabilities p ( t,t+h) and ]101 ( r, r+h) relative to respectively true
01
b) Show that with positive contagion we have and operational time. Give an interpretation of the operational time.
accidents, the workers are replaced with substitutes, who are also covered by the
tities in Question b) in this special case. . . mma
system. At timet there are a( t) workers workingJor the company. d) Show that if { N( t)} t~O has an operational time, then U ts etther a Ga
a) Comment on the model assumptions.
distribution or concentrated at one point. .
b) Find an operational time for the claim number process. Consider in parti- e) Show that a mixed Poisson proces~ ts homogeneous if and only if the
cular the case where >.( t) is independent of t.
mixing distribution is concentrated at one pomt. Give an intuitive explanation of
Let us now assume that the system includes only those workers who worked
for the company when the system was introduced, and that they leave the system this result. f L db (1964) In
Mixed Poisson processes are discussed in Chapter V o un erg .
after their first accident so that each worker receives compensation for at most one particular the result in Question d) of this exercise is proved as Theorem lOa.
claim. We assume that the group has no other causes of decrement than accidents.
c) Does there exist an operational time for the claim number process under
these assumptions? Exercise 9.7that the claim amount process {X( t)} t~O defin ed in subsection 9.3A is
Show
not necessarily a Markov process.
Exercise 9.6
(r>O)
a) Discuss the Ammeter process and compare it with a mixed Poisson process.
X'(t) = E~(t)
t=1
ytt)
,.... I
Let Y. be the amount of the ith claim. We assume that the claim amounts are
t
non-negative, mutually independent and identically distributed with moment gene-
where Y1( t), Y2( t), ... are mutually independent and identically distributed and inde-
rating function p, and independent of the claim number process. Let {X( t)} t~O be
pendent of N(t), and find the distribution of Y(t).
the corresponding claim amount process. The initial reserve of the insurance com-
c) Find expressions for EX(t) and Var X(t).
pany is u, and the premium income is P each year.
We now introduce a fixed amount deductible b. Let 1V(t) be the number of b) Is this risk process a Sparre Andersen process?
claims with amount greater than b. We assume that thee .'s are Gamma distributed with density
t
d) Show that {1V( t)} t>O is an inhomogeneous Poisson process, and find the
claim intensity of this process. Consider in particular the case -y=O, that is, no in-
(0>0; a,{l>O)
flation.
c) Show that the adjustment coefficient R for the probability that the insur-
Exercise 9.9
ance company is ruined by the end of a calendar year, is the positive solution of the
Prove Theorem 9.6 by using Laplace transforms.
equation
Exercise 9.10
We consider m independent insurance portfolios. For portfolio k (k=l, ... ,m),
the claim amount process {Xk(t)}t>O is a compou~d Poisson process with intensity
>.k and claim amount distribution Fk. The initial reserve at time 0 is uk, and the
d) Show that the adjustment coefficient is decreasing in Var e when we keep
premium is paid continuously with constant rate
Ee fixed, and give an intuitive explanation of this result.
Exercise 9.12
For this exercise we shall need the following result from renewal theory:
a) Find an upper bound for the probability that ruin occurs for all the m pro-
cesses. Theorem 9.13. Let L be a continuous defective probability distribution {L(w) =
We now assume that Fk=F and Pk=p>.k for k=l, ... ,m and merge the k portfo- lim L(t) < I} with density l. We assume that there exists a K- satisfying
lios into one portfolio. tl w
b) Find an upper bound for the ruin probability of the merged portfolio.
c) Show that the upper bounds in Questions a) and b) are equal. How does 1~ eK-tl( t)dt = 1.
this result appeal to you?
Assume that the function Z satisfies the renewal equation
Exercise 9.11
n:
Let { N( t)} t>O be a claim number process and 0={ en} 1 a sequence of inde- Z(t) = c + 15 Z(t-y)~y)dy.
pendent and identically distributed non-negative random variables. We call the
claim number process an Ammeter process if, given e, it is an inhomogeneous Then
Poisson process with claim intensity en during yearn.
-126- -127-
the t~me interval (O,m) when the initial reserve is u, and let ~ u)=1-1/J( u). (In this (Hint: Solve the integral equation in Question b).)
exercise we mean by ruin that the reserve becomess strictly negative.) Theorem 9.13 is proved in Section XI.6 of Feller (1971). Question a) of this
a) Explain the relation exercise is solved in Section V.5 of that book and Questions b~d) in subsection
XI.7(a). The case with exponentially distributed claim amounts is proved in Secti-
Vi(u) = I'Q Ae-At I ~u+Pt-y)dF(y) dt on 2.5 of Straub (1988).
(O,u+Pt]
Exercise 9.13
and utilise it to show that
We make the same assumptions as in Exercise 9.12 and use ordinary notation,
but we now assume that the reserve receives an interest continuously at a non-nega-
d ;;r1 A A
au'~'\ u) = p Vie u) - p I Vie u-y)dF( y). tive rate 6, which is added to the reserve. This will of course reduce the ruin proba-
(0, u]
bility.
b) Show that a) Write down an expression for the reserve at time t when the initial reserve
is u.
b) Explain the relation
Vi(u) = ~0) + ~~~~u-y)(1-F(y))dy.
c) Show that
1/1(0) =9ft.
and utilise it to show that
d) Show that
d A A
au -:TI
Vie u) = "P+"DU Vie u) - "P+"DU ( I I 1t-y)dF( y).
O,u
'1'\
Comment.
-128- -129-
10. The accumulated claim distribution We assume that V , V ,... , VRare independent, and that Vi has cumulative distribu-
1 2
tion Hi" Then
10.1. Introduction
. Let X be the accumulated claim of an insurance portfolio in a specified year, ii) Individual Model B. We make the same assumptions as in Individual Model
that 1s, the sum of the amounts of the claims occurring in the period. In this chap- A, but add the assumption that HFHindependent of i. Then we get
ter we shall discuss the calculation of the distribution G of X. There are several
reasons for studying G. We shall mention some of them:
R*
i} Calculation of one-year ruin probabilities. Let p be the total premium G=H .
income in the period and u the reserve at the beginning of the period. The one-year
In Individual Model A we explicitly assumed that the different policies could
ruin probability, the probability that the premiums and the reserve are insufficient
to cover the claims, is 1-G(u+P). have different risk characteristics; in Individual Model B we consider these differen-
ces as random. One therefore gets a natural transition from Individual Model A to
ii} Calculation of stop loss premiums (that is, preluiums for stop loss reinsur-
ance) for the portfolio. The pure stop loss preluium with retention Mis equal to Individual Model B by putting
E(X-M)+ = f M(x-M)dG(x).
iii} Collective Model A. This model resembles the way we worked in Chapter
iii} To see how an excess of loss cover would influence the total result of the
9. Let N be the number of claims in the period and Yi the size of the ith of these
portfolio. For such investigations we could compare the accumulated claim distri-
claims. It is assumed that the Yls are mutually independent and identically distri-
bution for the case with no reinsurance to the accumulated claim distribution with
excess of loss reinsurance with different retentions. buted and independent of N. As usual we assume that negative claims cannot
occur. Let F be the cumulative distribution of Yand
iv) To see how different deductibles would influence the total result of the
portfolio.
(n=0,1,2, ... )
An extensive pres~ntation of the theory of accumulated claim distributions is Pn = Pr(N=n).
given by Panjer & Willmot (1992}.
We have
10.2. Four different models G(x) = Pr(X~x) = 1:~ 0 Pr((X~x)n(N=n)) = En,:O Pr(N=n) Pr(X~xl N=n) =
1:~ 0 Pr(N=n) Pr(Ei: 1 Yi~x),
10.2A. There are different ways of modeling the mechanism generating x. In
this subsection we shall discuss four cases.
that is,
i) Individual Model A. Let R be the number of policies in the portfolio and v.
the total claim amount of policy i in the period. Then '
(10.1}
- - - --------- - - -
-130- -131-
10.2B. A distribution that can be expressed in the form {10.1) is called a com- z' = 1- q.t + q.F{v).
H'v) z z
pound distribution, that is, a compound distribution is a distribution that can be
Let
expressed as the distribution of a random number of mutually independent and
identically distributed random variables whose common distribution is independent
of the number of terms in the sum. As compound distributions also appear in other
contexts than the present one, we introduce the more neutral terms counting distri-
bution and severity distribution for the claim number distribution and the claim be the number of policies with positive claims. We have
amount distribution. A compound distribution with Poisson counting distribution
is called a compound Poisson distribution. Compound distributions with counting
distributions belonging to other specified classes are named analogously.
The following'-theorem follows trivially from Theorem 9.6. Furthermore, let
and severity distribution We want to approximate the present model by Collective Model B, and the
question is now how to choose the distributions of N and Y. Let H be the total
claim distribution of a policy randomly drawn from the portfolio, that is,
10.2C. In this subsection we shall discuss how to get from Individual Model A 1 R
H= R Ei=l Hi.
to Collective Model B.
We use the assumptions and notation of Individual Model A, but as we are It seems natural to choose
going to compare that model to collective models, we introduce the notation
ri'l ) _ Hfv1-H~O)
"\ v - - (0 ,
d
'nd
= *t=R 1 Ht..
that is F is the conditional distribution of the total claim amount of a policy ran-
Let
domly 'drawn from the portfolio, given that the total claim amount of this policy is
-132- -133-
positive. We have We see that we would have obtained the same compound Poisson approximati-
on if all the policies had the total claim distribution H. Thus the compound Poisson
1 R q.F:
F=,-E.
1\ I= 1 t t 1 m = E Y = ,..1 E .R1 q .m. approximation can be considered as a two-stage procedure. At the first stage, Indi-
EX= ENm. " I= ' ' vidual Model A is approximated by Individual Model B; at the second, Individual
(10.3)
Model B is approximated by Collective Model B. Intuitively one would believe that
Our next question is how to choose the distribution of N. A natural require- we would obtain a better approximation by performing only the first stage. This is
ment seems to be that X should have the same mean value in the individual model supported by a numerical example in Sundt (1985). However, this approximation
and the collective approximation. From (10.2) and (10.3) this requirement gives does not give an upper bound for the stop loss premiums like the compound Poisson
approximation.
EN= A,
~'
which is also the mean value of N in the individual model. A common choice of the 10.3. Stop loss inequalities
distribution of Nis the Poisson distribution with discrete density
In this section we shall prove some inequalities for stop loss premiums. For
simplicity, we shall assume that all distributions considered in this section are de-
(n=0,1, ... ) fined on the non-negative numbers, but most of the results are easily extended to
distributions on all reals.
As we shall see in Section 10.4, evaluation of
The stop loss transform V of a cumulative distribution G on the non-negative
numbers is defined by
metic, that is, concentrated at the points h,2h,3h, ... for some h. Furthermore, let
.
In particular V(o) is equal to the mean of G. We see that if G is the cumulative
distribution of the accumulated claim from an insurance portfolio during a specified
dnd(t) =If, (x-t)ddnd(x); cf0 ll( t) = If, (:c-t)d cf011 ( x) period, then V( t) is the pure stop loss premium with retention t for that period.
We write
be the stop loss premiums with retention t for the two distributions. From the fol-
lowing theorem we see that the collective approximation with Poisson counting
distribution gives an upper bound for the stop loss premium of the individual model. V(t) = J't f~ ds dG(x).
Theorem 10.2. For all t~O we have By changing the order of integration we obtain
F(O)$V(O). (10.8)
"?'H(t) = E(X+ Y-t)+ = EE({X-(t-Y))+il1 = EF(t-Y),
We assume that there exists an r such that
letting
F(x) $ G(x) (x<r)
F(x) ~ G(x). (~r) F(s) = F(O)- s; V(s) = V(O)- s
Proof For t~r the inequality F(t)$V(t) follows from {10.4); for t<r from EF(t-Y) ~ EV(t-Y),
{10.5) and (10.8). Q.E.D.
which proves the lemma. Q.E.D.
In the light of Theorem 10.3 it is interesting to look back at Lemma 9.4; we
see that if the compensation functions r and r satisfy the conditions of Lemma Lemma 10.4. Let F1'" .. ,Fn'Gl' ... ,Gn be cumulative distributions satisfying
1 2
9.4, then by Theorem 10.3 the stop loss transform of r (Z) is less than or equal to
1
the stop loss transform of r 2(Z), and hence stop loss ordering is consistent with the {i=1, ... ,n)
ordering implied by Lemma 9.4.
---- --
Then
-136-
r
I
I
I
the general case by induction.
-137-
Q.E.D.
Lemma 10.6. Let {pn} be the discrete density of a distribution on the non-
negative integers, and let F be a cumulative distribution. Then
7
Proof. By Lemma 10.5, nF~F for all n, and hence
The practical impact of Theorem 10.4 should be ,clear: If we assume that the
individual claim amounts are mutually independent and identically distributed and
independent of the number of claims, then we get an upper bound for the stop loss This proves Theorem 10.2 in the special case R=l. For general R, application of
premium by replacing the distribution of the individual claim amounts with a distri- Lemma 10.4 gives
bution for which the stop loss transform is greater than or equal to the stop loss
transform of the original claim amount distribution.
dnd < .R
-
*'1=1 Jf.Oll.
t
Lemma 10.5. For any cumulative distributions F ,... ,Fn we have
1 By Theorem 10.1
Proof The case n=2 follows from the inequality and thus Theorem 10.2 is proved.
Many of the results given in Section 10.3 are taken from Biihlmann et al.
(x,y,t~O) (1977). There exists an extensive literature on inequalities for stop loss premiums;
for a more extensive treatment we refer to van Heerwaarden (1991). and solving for hR* (x) concludes the proof of Lemma 10.6. Q.E.D.
h(v) = Pr(V=v). Proof. We introduce Vi= Vr-m and X' =Ei~ Vi=X-mR. The distribution
1
(V= ... ,-1,0,1, ... )
of X' can be calculated by Lemma 10.6, and Theorem 10.5 follows by a trivial
It is assumed that transformation. Q.E.D.
(=0) 10.4C. We shall now look at recursive methods in the collective models.
Let N be the number of claims and Y. the amount of the ith claim. It is
(=1,2, ... ) '
assumed that the Y.>s are independent of N, and that they are mutually independ-
'
ent and identically distributed on the positive integers with discrete density f. The
Proof. We trivially have hR* (O)=h(O)R. discrete density g of X is given by
Let us introduce an auxiliary random variable v0, independent of X, with
discrete density h. For x>O we have (10.9)
that is, x
g(x) = En=O Pnf n* (x). (10.10)
rem 10.4 or 10.40. In this subsection we shall look at three classes of counting distributi-
ons satisfying the recursion {10.12). Sundt & Jewell {1981) show that these classes
(n=O) are the only ones satisfying that recursion.
{10.11)
(n=1,2, ... )
i) Poisson distribution.
and then use (10.10). However, such procedur~s are very time-eonsuming, and we
therefore want some simpler algorithm. (n=0,1,2, ... )
The following theorem (Panjer {1981)) gives a recursive algorithm that can be
used when the counting distribution belongs to a particular class. We have
Theorem 10.6. Assume that there exist constants a and b such that
(n=1,2,3, ... ) ( 10.12) that is, a=O, b=>-., and Theorem 10.6 gives
Then (=0)
->-.
g(x) = e
We have
g(x) = ~n:l Pr/n*(x) = ~n: 1 [a+*]Pn_/n*(x) =
~n:1 ~y!1 ( a+~]Pn-1/(y)f(n-1 )\:D-y) = that is, a=p, b=p(a--1), and Theorem 10.6 gives
We see that {10.13) reduces to (10.12) in the special case where f(x)=ox , that iii) Binomial distribution.
1
is, all claim amounts are equal to one.
(n=O,l, ... ,M)
- - -- --- - --- - -- --
~~
I i.l
I
-143-
-142-
(10.15)
and as Pn=O for all n>M, the recursion (10.14) holds for all n>O. Thus (10.12)
holds with
10.4F. We have
(=0)
(z=-1)
(=1,2, ... ) cf(x) = ~ 1
l cf(x-1)- g(x) (z=O, 1,2, ... )
We obviously have that x(h)~x, and thus the cumulative distribution a{h) of x(h)
-144- -145-
satisfies the inequality a( h)<_ G. The ~ro11owing theorem gives upper and lower that within the model assumptions they give exact values of the accumulated claim
bo)lnds for V based on the approxima~ion a( h)_ distribution. There also exist a number of approximations for the accumulated
claim distribution using only some moments of the counting distribution and the
Theorem 10.7. We have severity distribution.
One could of course ask why one should use such approximations when there
exist exact methods. There are several reasons. For large portfolios, the exact
methods are usually very time-consuming, even in our age with powerful computers.
Furthermore, in particular for large portfolios, we can get numerical problems when
Proof From (10.7) we obtain applying them. The approximations have the advantage that they are fast, and for
many applications they are sufficiently accurate.
For the choice between exact methods and approximations it is also important
to know what information is available. In practice the information about the severi-
ty distribution often amounts to only estimates of a few moments. In such cases we
Furthermore
cannot apply exact methods; we have to rely on approximations.
In Section 10.5 we shall concentrate on the NP approximation (Normal
CPiJ(o)- V(O) = E_x(h)- EX= ENE( y(h)_ Y) ~ hEN, _ower). In practical applications this approximation seems to perform well. How-
ever, unfortunately we do not have any bounds for the approximation error.
which completes the proof of Theorem 10.7. Let u1_E denote the 1-E fractile of G (that is, u1_E is the smallest number u
Q.E.D. such that 1-G( u)~t), and let z1_E be the 1-E fractile of the standard normal distri-
bution N(0,1). The NP approximation says that we have approximately
We see that we can make the difference h EN between the upper and lower
bound for V as small as we want by making h small.
To get an upper bound for G we can round the Y/s down wards instead of up- (10.16) ~
wards. Upper and lower bounds for V based on this approximation can be found b
For a derivation of this approximation we refer to Beard, Pentikli.inen, & Pesonen
applying (10.6) instead of (10. 7) in the proof of Theorem 10. 1. Y
(1984).
. ~rithmetic approximations to F can be useful even ifF is already arithmetic,
smce 1f the monetary unit is very small, say 0.01 DEM, the computations can be
10.5B. The most common application of the NP approximation is the case
v_ery ti~e-consuming, and the accuracy is sometimes sufficient with an approxima-
tiOn usmg, say, 100 DEM or 1000 DEMas unit. where G is a compound Poisson distribution, that is,
where N is Poisson distributed, and the Y/s are mutually independent and identi-
10.5. The normal power approximation cally distributed and independent of N. Let A be the Poisson parameter. Then from
(9.23) and (9.24) we have
10.5A. The recursive methods described in Section 10.4 are exact in the sense
-146- -147-
EX=>. EY; pany were to wind up its business at present, then the probability that the reserves
(10.17)
would be insufficient to cover the future liabilities, would not exceed E.
We can also easily find The insurance technical liability umin can be split into the expectation liability
~=EX
(10.18)
and the fluctuation liability
and insertion in (10.16) gives
The expectation liability represents the expected claims. However, as the claims are
likely to exceed their expectation, we find it insufficient to claim that the reserves
This approximation seems to work well if the skewness of X , that is , should exceed ~- Therefore we add the fluctuation liability as a safety loading for
random fluctuations.
By using the NP approximation, we obtain
10.5C. A more extensive treatment of the NP approximation is given by with c being the 1-E fractile of the standard normal distribution N(O,l), and
1
Beard, Pentikli.inen, & Pesonen (1984), who also discuss other approximations to the c =( c~-1 )/6. With t=0.01 we have c1=2.32635 and c2=0. 73532.
accumulated claim distribution. 2
Let us assume that the company works in H different insurance classes, and let
Xh be total amount of the future claims in class h, for which the company is respon~
sible at present. Then we have
10.6. Solvency control
for some specified E (e.g. E=0.01), that is, the reserves should be greater than or
For each class h, let Nh denote the number of claims included in Xh, and let
equal to umin' the smallest number satisfying (10.19). We call u . the insurance
technical liability of the company. The requirement u~umin mean~~hat if the com- Yhi be the amount of the ith of these claims. Then
(10.26)
N . It is. assume.d t~at given the value 0 of an unknown random risk parameter e with
h1 IS Pmsson
w distnbuted with parameter p h0' whe re p h IS a measure of riskh'
vo ume. e further assume that the claim amounts y y .
pendent and identically distributed and independent of~' a h~~ are mutually mde-
01 hk = EV::.1
b~ r:ndo~
The reason for assuming the risk parameter eh to is that th
{3h1 = EE>h;
extern~, unpredictable phenomena that could affect the whole 'portfolio :re a:: We assume that the parameters ahk and flhk are the same for all companies
mo:r msu~ance
a cold winter with icy roads could lead to more accide~ts ~~an and estimated by the supervisory authority from statistics provided by the compani-
usu . ~t. IS perhaps unrealistic to assume that eh does not affect the claim es. We shall not go further into this estimation. The companies obtain the esti-
amounts, m the above example it is plausible that the icy roads will not only af~ t mates of the ahk's and {Jhks from the supervisory authority and insert these toge-
the number of claims ' but also the indiVI'dual cI aJm
. amounts. ec
ther with their risk volumes in formulae (10.24)-(10.26) for each class h. Then the
. For choice of measure ph there are various possibilities. In motor insurance moments of X are found by insertion of the moments of Xh in (10.21)-(10.23), and
mileage could be .a natural measure; in other classes one could use years at risk oli~ finally umin is found by insertion of the moments of X in {10.20).
:ylyears, or su~
msured years. Earned premium is often applied as measure ~f~isk Up to now we have considered X as the total amount of future claims for
o ume: but this measure is somewhat awkward; one does not reduce the risk by which the company is responsible at present, without mentioning anything about
decreasmg the premium rates.
reinsurance. It is clear that reinsurance is of great importance for the risk of the
From the above assumptions and formulae (10.17) and {10.18) we obtain company and ought to be taken into account when making solvency requirements.
We therefore assume that the Yhls are claim amounts for own account, that is, the
EXh = EE(XhjE>h) = E(PhE>hEYh) = PhEE>hEYh total payments on the claims less reinsurance recoveries. This assumption excludes
reinsurance forms like stop loss. It is assumed that the reinsurance arrangement
Var Xh = EVar(XhjE>h) + Var E(XhjE>h) = PhEE>hE~ + (PhEYh)2 Var E>h does not affect the Nh's. The model assumptions could be realistic for reinsuravce
forms like quota share, excess of loss, and surplus.
It is clear that the ahk's will depend on the the reinsurance cover, and for
E(Xh-EXi = EE((Xh-E(XhiE>h)+E(XhiE>h]-EXh)31E>h] =
reinsurance forms like surplus and excess of loss, the supervisory authorities would
EE((Xh-E(XhjE>h])31E>h] + 3E((E(XhiE>h]-EXh) Var(XhjE>h]) + have to specify tables of these parameters for different retentions. For quota share
E(E(XhiE>h]-EXh)3 = E(PhE>hE~) + the situation is much simpler. Let q be the fraction paid by the cedant, and let
-- -- ---
- -- - - - - -
- ---------
-150- -151-
described in the previous subsection, one should require that for each class h, the giving
company should have a reserve uh satisfying
for some specified t', presumably greater than the f of the previous subsection. To and u ,... ,uHcan be found by insertion in (10.27).
1
the present author, such a requirement does not make sense. When a company is
insolvent, it is the company as a whole that is insolvent; a company is not insolvent 10.6C. Let us consider the ratio between the expectation liability ~ and the
in travel insurance while fire insurance is doing excellently. However, both for in- fluctuation liability Uy
ternal purposes and for supervision, it could be desirable to distribute the insurance
technical liability umin on the different classes. An interesting question both for the r= Uyl~
company and for the supervisory authority is: Are the present rates of class h
reasonable compared to the risk of that class, or is this class subsidising, or being As this ratio could be considered as a measure of riskiness of the portfolio, it is in-
subsidised by, other classes? teresting to discuss how it depends on changes in the portfolio. For simplicity we
One approach for distributing the insurance technical liability, is the follow- first assume that H=1, that is, the company writes only one class of insurance.
ing: Let uh be the share of umin to be allocated to class h. We want the same ruin If the risk volume of the company increases, then r will normally decrease;
probability for all classes, that is, Pr(Xh>uh) should be the same for all h. The NP within a larger portfolio the uncertainty will be spread over more risks, and the
approximation shows that there exists a constant c independent of h such that relative fluctuations will be lower.
Quota share reinsurance will have no effect on r. Although the scale will de-
crease, the relative fluctuations will be the same. On the other hand, an unlimited
(h=1, ... ,H) (10.27) excess of loss reinsurance will truncate the large claims, and such a reinsurance will
therefore normally reduce r. Surplus reinsurance will usually also have this effect.
As If the company writes more than one class (H>1), the picture becomes ~less
clear. It is not always true that increased risk volume in one of .e classes will re-
H duce r. To take an extreme case, if the company writes only home insurance and oil
Eh= 1 uh = umm
.
rig insurance, then an increase in the latter class could increase r as oil rigs are more
risky than homes.
we have
10.6D. In the solvency control system outlined above, the minimum require-
1 2
E X+cs+,.(c-1)r=u.
o Irun ment for technical reserves is based on a fixed ruin probability. If different insur-
ance companies have a relatively similar composition of their portfolios, then nor-
with mally a large company would have a lower r than a small company. This seems
intuitively reasonable as the relative uncertainty would be lower in the large compa-
H E(Xh-Elh)3
ny.
r- E On the other hand, is it reasonable to require the same minimum value of the
- h=1 Var 1 '
h ruin probability for a large insurance company as for a small company? Although
bankruptcy of a small company could be disastrous for its policyholders, it would
-152- -153-
have much more severe consequences for society as a whole if a large company went
broke. Given these considerations, it would seem reasonable to require a lower ruin
probability for a large company. We introduce
We could also ask whether it would be reasonable to have the same solvency
requirements for a mutual company as for a stock company since a mutual compny
1 h(n)Fn* i m-1 h( nl'-
Vm=En=O )-::ii"'.
would normally have the possibility to claim extra premiums from its policyholders Gm = Em-
n=O
retrospectively if the reserves were to be insufficient.
Show that
10.6E. The NP approximation is used in the Finnish solvency control system
m*
- Gm + If(m-1) F
described by Pentikainen & Rantala (1982) and in the Norwegian solvency control G
m-< G <
system, of which a preliminary draft is given in Norberg & Sundt (1985), on which
subsection 10.6A has been based to a great extent. 7r + If(m-1) Fm* ~ V ~
m
v;;; + If(m-1) Fm* + 'H(m) E Y.
The principle of allocating the insurance technical liability described in subsec-
tion 10.6B has been used by Straub (1988, Chapter 7) in connection with infinite This exercise is based on Sundt (1991b).
time ruin probabilities and Lundberg's Inequality.
A survey of solvency theory is given by Kastilijn & Remmerswaal (1986). Exercise 10.3 . 1
Pentikainen et al. (1989) give an extensive analysis of the solvency problem. We assume that all probability distributions studied in thts prob em, are con-
centrated on the non-negative numbers.
a) Show that if
Exercises
(t~O)
a~ V(t)- 'F(tH b,
Exercise 10.1
Let X be a non-negative random variable with cumulative distribution G and then
finite mean. Show that
Exercise 10.2
then
Let Y1, Y2,... be independent and identically distributed non-negative random
variables with common cumulative distribution F, and let N be a non-negative in-
teger-valued random variable with cumulative distribution Hand discrete density h. (~0)
We assume that N is independent of the Y/s. Let G denote the cumulative distri-
bution of
c) Show that
-154-
-155-
--;:;.-
n F(t) 5 Fn (t) 5 (n-1) F(o) + F(t). c) Show that
(n=1,2, ... ; t~O)
This exercise is based on De Pril & Dhaene (1992). This exercise is based on Kaas, van Heerwaarden, & Goovaerts (1988).
Comment.
Exercise 10.4
05
Exercise 1 distribution on the integers 0,1, ... ,m with discrete
Theorem 10.2 says that we obtain an upper bound for the stop loss transform Let F be the U niform
a) Show that x
E y=:~J-m ) - n x [H"* (x)-Hn*{:ll-m-1)].
yhn*( Y-n+l
Exercise
Let 10.6 . mteger
N be a non-negative . -valued random variable with discrete density
As an global measure of the quality of the approximation GA we introduce the
efficiency
(n=1,2, ... )
b) Discuss this efficiency measure,
then N must b e ei'ther Poisson distributed, binmnially distributed, negative binomi-
-156- -157-
ally distributed, or almost surely equal to zero. Deduce which values of a and b give Exercise 10.8 . .
Let N be the number of claims incurred in an insurance portfolio dunng a
distributions from these classes, and indicate these values in an (a,b) diagram.
The solution of this exercise is proved as Theorem 1 in Sundt & Jewell ( 1981). fixed year, yi the amount of the ith of these claims, and
Exercise 10.7 N
X= L:i=1 yi
Let N be the number of claims occurring in an insurance portfolio during a
specified period, and let Yi be the amount of the ith of these claims. We assume the accumulated claim. We assume that the distribution of N is geometric with
that the claim amounts are mutually independent and identically distributed and discrete density
independent of the number of claims. Let F denote the claim amount distribution
and (n=O,l,2, ... ; O<p<1))
p = Pr(N=n) = (1-p)pn.
n
Pn = Pr(N=n). (n=0,1, ... ) The claim amounts yl' y , ... are assumed to be positive, ~utually independent and
2
identically distributed with common distribution F, and mdependent of N. Let G
Let fiQ be the number of claims exceeding a fixed limit I. denote the cumulative distribution of X.
a) Find an expression for We assume that there exists a constant K, satisfying
(10.28)
(n=0,1, ... ) ~ = I (O,w) eK,YdF(y),
c) Show that if the distribution of N belongs to one of the three classes dis-
b) Show that in the continuous case
cussed in subsection 10.4D, then the distribution of fiQ belongs to the same class.
and find similar asymptotic expressions for the density g and the stop-loss transform g) Show that the ruin probability 1/J in Exercise 9.12 is the tail of a compound
geometric distribution, and discuss the relation between the severity distribution of
V. (Hint: Cf. Exercise 9.12).
the compound distribution and the claim amount distribution of the risk process in
c) Show that the asymptotic expressions in Question b) are exact when the
terms of stop loss transforms. Explain how one could calculate upper and lower
distribution F is exponential with density
bounds for 1/J recursively.
h) Discuss how the result in Question /) can be used to generalise Lundberg's
_JI. Inequality for compound Poisson risk processes to the case where the moment gene-
f(y) = 1- e J.. (y>O; J,>0) rating function of the claim amount distribution does not exist.
Questions a)-c) are based on Sundt (1982), Question e) on Willmot (1993),
d) Suggest recursive methods for evaluating g(x) and G(x) in the discrete case.
and Questions /)-h) on Dickson (1993). Embrechts, Maejima, & Teugels (1985)
e) Show that in the discrete case
generalise the asymptotic expression (10.29) to the case with negative binomial
claim number distribution.
(X=0,1,2, .. ) (10.30)
Exercise 10.9
and deduce a similar exponential bound for the stop loss transform V. (Hint: Re- Let N be a non-negative integer-valued random variable with discrete density
write the equation in Question a) as
Pn = Pr(N=n) ( n=0,1,2, ... )
Var N 1
~=r-a,
/) Show that
and discuss this expression in the three main cases of subsection 10.4D.
d) Show that
(x=:0,1,2, .. ,M)
and discuss this result and its relation to the result in Question e).
-161-
-160-
(n=k+l,k+2, ... )
probability gener~ting function rY' and moment generating function p We assume
that the Yls are mdependent of N. Let y
for some non-negative integer k. Consider in particular the case where pn=O for n=
o, ... ,k--1.
This exercise is based on Section 5 of Sundt & Jewell (1981).
the accumulated claim amount. The claim amounts Y1, Y2,... are positive, integer-
and use this result to prove Theorem 10.6. valued, independent, and identically distributed with discrete density
/) Show that
(y=1,2, ... )
f(y) = Pr( Y=y)
(1-apy(s))p_x(s) = (a+b)py(s)p_x(s), I
and independent of N, which is logarithmically distributed with discrete density
and use this result to show that
1 n (n=1,2, ... ; O<p<1)
Pn = Pr(N=n) = lln(l-p)l ;-.
(~1,2, ... )
Find a recursive method for calculating the distribution of X.
Deduce a similar recursion for This exercise is based on Example 1 in Sundt & Jewell (1981).
discrete density
( IJ>O; a,/3>0)
and assume that the random variable N is conditionally Poisson distributed with Pn = Pr(N=n) = [ M] n
njP {1-p
)M-n
, {n=0,1, ... ,M)
parameter 0 given 6= 0. Show that N is unconditionally negative binomially distri-
buted. and let
b) Show that a compound Poisson distribution with logarithmic claim amount
distribution is a negative binomial distribution.
c) Discuss the results in Questions a) and b).
We introduce the discrete densities
Exercise 10.13
We consider a wind-storm insurance portfolio. Let N be the number of wind- f(y) = Pr( Y=y); g(x) = Pr(X=x)
storms during a given period, Mi the number of claims incurred by the ith wind-
storm, and Zij the amount of the jth claim incurred by the ith wind-storm. We and assume that
want to calculate the distribution of the accumulated claim
K = -inf{y: /(y)>O} < m.
Discuss how to calculate g(x) when: For simplicity we assume that the moment generation function of the Yls exists.
Let
m
a) q = ~e-J-. (m=0,1, ... )
m m.
- 1 ~
0. =
~
(Y.)+;
~
(i=1,2, ... )
b) qm- lln(l-p) I m (m=1,2, ... )
x+ = E.Nl 0.; N
~= ~
X=Ei=lyi"
Exercise 10.14
Let Y1, Y2,... be independent and identically distributed integer-valued ran- Show that x+ and x- are stochastically independent, and explain how the
dom variables, independent of the binomially distributed random variable N with distribution of X can be calculated when the distribution of the Yls is known.
-164- -165-
(n=1,2, ... )
with Pn=O for n<O. We shall denote this distribution by RJa,b] with a=( a1' ... ,ak)
and b=(bl' ... ,b,t).
a) Show that
(=1,2, ... )
'<' k
d ...., -1 ( ~a.+
. b .) si-1
a(s)=aslnr(s)= F k ~ ~,
1 - Ei= 1 ais
-166- -167-
11.1B. We now see that the solvency control problem is more complicated calendar year i. Nir of these claims are reported in development year r, that is,
than indicated in Section 10.6 as the random variable X should in principle also calendar year i+r, we have
include future payments on IBNS claims in additiion to CBNI claims (.Qovered Jiut
Not Incurred). As future payments on unsettled claims could depend on earlier 00
E N.=N ..
claims data, the unconditional distribution of X in (10.19) should be replaced with r= 0 ~r ~
the conditional distribution given the claims data available at the time of evaluati-
The amount of the jth of these claims is Yiri and we denote by Yirjd the cumula-
on. Furthermore, the expectation liability should be equal to the conditional expec-
tive payments on this claim up to the end of development year d. We have YirjCO
tation of X given these data. The expectation liability can be split into a CBNI
for d< r. Furthermore,
liability and an IBNS liability; the latter can be further split into an IBNR liability
and an RBNS liability. For more details we refer to Norberg & Sundt (1985).
limY. d= Y. '
In the following we consider the reserves for unsettled claims as predictions of dj., trJ zrJ
the payments. Thus we assume that random fluctuations are provided for by sepa-
rate fluctuation reserves. The total payments up to the end of development year d on claims incurred in
calendar year i and reported in development year r is
11.1C. A pragmatic method for RBNS reserves is that the claims evaluator
makes for each single claim a subjective estimate of the final amount based on his N.~r
X.trd =E.
J= 1 Y.ZTJ'd
present knowledge. When he receives new information, he revises his estimate.
This procedure is not very mathematical, but it is often the best method. An expe-
We also introduce
rienced claims evaluator possesses an extensive knowledge and intuition that can
never be replaced even by an extremely elaborate mathematical model. However, in
N.
particular for portfolios with a great number of similar claims, the method could be X. = 1i m X. d = E .~lr Yir'
~r dj., zr 3- J
unreasonably time-consuming.
In practice one often applies a compromise between subjective individual case
being the total amount of claims incurred in calendar year i and reported in develop-'
estimation and statistical methods: If the claims evaluator believes that the final
ment year r.
amount of a claim will not exceed a specified limit, then the amount is estimated by
At the end of calendar year C the total amount of not yet reported claims
statistical methods; otherwise he makes a subjective assessment.
incurred in calendar year i (~C) is
Statistical methods can also be applied as a control of the subjective estimates.
If the statistical methods give higher estimates, the reserves based on subjective
judgement may be insufficient.
Subjective judgement can be a valuable tool for the insurance company for
rate-making and reserving, but for insurance supervision one would presumably need and our IBNR reserve should be a prediction of
more objective methods for prediction of unsettled claims; should the Supervisory
Service be satisfied with the argument, "By subjective judgement our claims depart-
:ment finds our reserves sufficient."?
The aggregate future payments on RBNS claims incurred in calendar year i is
11.1D. We shall now describe the prediction problem in more mathematical
terms. Let Ni denote the number of claims incurred in an insurance portfolio in
-171-
-170-
On the other hand, it would be more consistent with the concept of IBNR claims to
split the predictions by the year when the claims are reported, that is, to predict
W'. = E d (X X )
'd r=O ira ir,d-1
11.2. A probabilistic model
separately for iS C, d> G-i. Analogously one should make separate predictions of
IBNR and RBNS claims for each claims year and preferably also split these predicti- 11.2A. We now introduce a probabilistic model which is consistent witi the
ons by development year. For RBNS claims one should thus predict assumptions made in Section 10.6. Within this model, we want to determine opti-
mal reserves according to quadratic loss.
V. = EC-i (X ) It is assumed that data related with claims incurred in different calendar years
'dC r=O iraXir,d-1
are independent, and that all claims will be settled before the end of development
year D. Thus NFE s~O Nis' xir=XirD and yirJ yirjD for all ( i, r,J).
For IBNR claims there are two possibilities. One of them is to estimate
For each claim year i we assume that Ni is conditionally Poisson distributed
U.
,de = E r=C-H1
d. (X X )
ira ir,d-1
with parameter P/J, given the value (}of an unknown random variable e i that cha-
racterises the year; Pi is an observable measure of risk volume.
that is, the payments in development year d on claims incurred in calendar year i It is assumed that a claim incurred in calendar year i is reported in develop-
~ut not yet reported by the end of calendar year C. This approach would fit nicely ment year r with probability J(r) independent of Ni and ei" From Theorem 9.5 we
mto the splitting of the predictions of IBNS and RBNS claims by development year see that Ni0, ... ,NiD are conditionally independent and Poisson distributed with
as we have parameters P/(O)fJ, ... ,P /(D)fJ given the value (}of 0 i"
The (YirjJ'"""'yirjD)'s are assumed to be mutually independent for all (i,r,J)
-173-
-172-
and independent of the claim numbers. Furthermore, they are identically distribu-
ted for fixed r. A claim amount is allowed to depend on the time when it is report-
ed; perhaps a particular type of claim is reported late. We assume that
with
EY . . = JL ). = Var 0;
~rJr r (r<d~D)
JL = E0;
E[ Y.~rJdi Y~r,. . . d- 1] = 'Yr dy~rJ,
. ...., ... , Y~rJ, . . d- 1;
For i~ C and G-i< d~ D we have
in practice we would normally have 'Yrd~l.
At the end of calendar year C, the available data about claims incurred in
calendar year i (~C) are
The parameters of the present model are easily estimated from observed
claims data. Biihlmann, Schnieper, & Straub (1980) discuss parameter estimation
in the special case where Ni is Poisson distributed with parameter P/)- They also
present a simulation study.
VidC= E[VidCI1Ji0 = E~~ E[XirdXir,d-111JiO =
11.2B. Let us now deduce Bayes estimators for the unsettled claims based on
. N. J
the available data at the end of calendar year C. As data about claims incurred in E~~ Ej!~ E[Yirjdyirj,d-11 yirj,G-i =
different calendar years were assumed to be independent, we need only data from c- N. d-1
claims incurred in calendar year i to predict unsettled claims incurred in that year E r=~ Ej!~ Yirj,G-i ( 'Yrti1) Ilh=G-i+1 'Yrh =
G-i d-1
(except for parameter estimation). Er=O Xir,G-~{'Yrti1 ) Ilh=G-i+l 'Yrh
For i~ C, the Bayes estimator of 0 i is
As a Bayes estimator of a sum is equal to the sum of the Ba:es estimators of the
elements of the sum to be estimated, we obtain the Bayes estimators of the a~gre
gate quantities in subsections 11.1D-E by summing the appropriate Bayes estima-
tors.
with
Exercises
Exercice 11.1
. Deduce the credibility estimator tJ iC in section 11.2B and suggest estimators
We approximate 0 iC by the credibility estimator based on 0 iC' that is,
h t ~ -tJ in the special case where
for unknown structure parameters. Show t a iCi iC
-174- -175-
~i.
E. 0 a.f3.= x.D-. (i=O, ... ,D)
J= t J ,, z (11.2)
~. ~-
E; 0J a.f3.= E. J Y .. (j=O, ... ,D)
= t J '= 0 ZJ (11.3)
Po=L (11.4)
b) Show that
c) Express
(j=1, .. ,D)
- - - - - -- -=-=-- - - - - - - - - -
---------
-176- -177-
12. Utility theory utility function u such that the person will prefer the fortune X to the fortune Y if
and only if Eu(X)>Eu( Y) (see e.g. DeGroot (1971)). This utility function is unique-
ly determined up to linear transformations V=au+b with a>O; it is clear that v gives
12A. We assume that an insurance company would charge a premium p to the same preference structure as u. However, in my opinion, some of these axioms
cover a risk X. Usually P>EX. In that case the expected gain EX-P of the risk- seem somewhat artificial so we shall not go further into the axiomatic development
holder by insuring the risk is negative, that is, the expected loss is positive. How- of utility functions; we assume only that the utility function exists and satisfies the
ever, despite of this, people insure their risks, and insurance companies flourish. following two properties:
Are all policyholders crazy? Presumably not.
As a simple example, let us consider the risk distribution given i) u is strictly increasing.
ii) u( w+t)-u( w) is strictly decreasing in w for all t>O.
Pr(X = 100 000 DEM) = 0.001 = 1- Pr(X = 0).
We have already discussed property i).
Then EX = 100 DEM, but the risk-holder would in most cases prefer to insure the Property ii) seems to be satisfied in most cases; you would presumably appre-
risk for a premium P > 100 DEM rather than risking to lose 100 000 DEM. He ciate an increase of 1 DEM in your fortune more if you have a small fortune than if
finds the event of losing 100 000 DEM so unpleasant that, although the probability you have a large fortune. However, in some situations the assumption seems more
that this event should occur, is very low, he is willing to pay more than 100 DEM, questionable; if you want to buy a video recorder that costs 1 200 DEM, then an
more than the expected monetary amount of the risk, to avoid the risk. This im- addition of 500 DEM to your fortune is more useful to you if your fortune is 1 000
plies that the utility of money is not linear. DEM than if it is 500 DEM.
In utility theory we assume that the risk-holder has a utility function u, and his We have assumed that there exists a utility function u satisfying properties i)
utility of having a fortune xis u(x). It is clear that u is strictly increasing; if x>y, and ii). But what does u look like? Is it exponential? Logarithmic? Quadratic? It
then the utility of the fortune x is clearly greater than the utility of the fortune y is clear that to specify the utility function completely, we have to make very strong
(although one sometimes feels a bit confused about this when filling in a tax declara- assumptions. In great parts of utility theory it is assumed that the utility function
tion!). is completely specified. However, as we shall see in the following, one can prove ~
The utility function becomes more interesting when comparing random for- several interesting results by only assuming that properties i) and ii) are satisfied.
tunes. Let X and Y be two random fortunes. We assume that the person would
prefer the fortune X to the fortune Yif Eu(X)>Eu( Y). 12B. The main result of this subsection is that the utility function u is strictly
To make this more concrete, let us return to the question whether or not to concave. To prove this we shall need the following lemma.
insure a risk X for a premium P. We assume that the risk-holder's fortune at the
beginning of the insurance year is w, and that the development of his fortune during Lemma 12.1. The utility function u is continuous.
the year is not_ influenced by any other factors than the risk X and the premium P.
Then his fortune at the end of the year is w-P if he insures, and w-X if he does not, Proof From property i) follows that for all w the limits u( w+) and u( w-)
and he insures if and only if exist and are finite, and that u(w+)~u(w-). We have to show that u(w+)=u(w-).
Let v<w. By property ii), we have for all positive integers n
u( w-P) ~ Eu( w-X).
u( w)- u( v) = E~ 1 [ u[ v + k(~v)] - u[ v + (k- 1 )~w-v)]] >
In utility theory, axioms for rational behaviour of a person have been stated,
and it has been shown that if these axioms are satisfied, then there must exist a n[u[w+Pn] -u[w-Fn]],
-178- -179-
(i=1,2) (12.3)
which gives
){ = un:,1 Jln.
u( t' v' +(1-t' )w) ~ t' u( v') + (1-t' )u( w),
We shall prove by induction that (12.1) is satisfied for all tEJl, that is, for each
n=1,2, ... , (12.1) is satisfied for all tEJln. that is,
For tEJl1 , that is t=L (12.1) can be rewritten as
u(tv+(1-t)w) ~ t'u(sv+(1-s)w) + (1-t')u(w). (12.4)
which is satisfied by property ii). Now, assume that (12.1) is satisfied for all tEJln, (12.5)
u(sv+(1-s)w) > su(v) + (1-s)u(w),
and let tEJln+lN){n. Then t=(t +f:!)/2 with t =t-2-(n+l) and t =t+2-(n+l), and
1 1 2
property ii) gives and from (12.4) and (12,5) we obtain
which shows that (12.1) holds for all tE(0,1)NJl. Hence, (12.1) holds for all tE(0,1),
and the first part of the theorem is proved.
-180- -181-
Now, assume that u is strictly concave, and let v<w and t>O. With the nota- Inequality gives
tion of the proof of Theorem A.1 we have
u(w-II(w)) = Eu(w-X) < u(w-EX),
S(v,v+t) > S(v,w+t) > S(w,w+t)
and as u is strictly increasing, we must have II( w)>EX. This completes the proof of
which gives Theorem 12.2. Q.E.D.
u( v+t)- u( v) > u( w+t)- u( w), Normally we would have II( w)<m. The case II( w)=m means that the risk X is
so terrible that there is no limit to what the risk-holder is willing to pay to get rid of
implying that property ii) is satisfied. This completes the proof of Theorem 12.1. it.
Q.E.D.
12D. Let us now briefly look at something that seems like almost the opposite
12C. We assume that the risk X is a non-negative and non-degenerate random of buying insurance, namely participating in a hazard game. Whereas a person who
variable with finite mean. Let w denote the fortune of the risk-holder at the begin- buys insurance, is willing to pay for getting protection against uncertainty, a person
ning of the insurance year. We assume that this fortune will not be influenced by who participates in a hazard game, is willing to pay for obtaining uncertainty.
other factors than the risk X during the year. The risk-holder has the possibility to Let us consider the same person that we considered in the previous subsection.
insure the risk for the premium P and will do so if and only if However, we now discard the risk X. Instead, we assume that the person partici-
pates in a hazard game. Let V denote his net monetary gain from the game, that is,
u( w-P) ~ Eu( w-X). (12.6) what he receives, less what he has paid to participate. As there is some hazard, the
random variable V is non-degenerate. Furthermore, as the organiser of the game
The inequality (12.6) is obviously satisfied for P=O. Let II( w) be the supremum of would want to make a profit, we assume that E V<O.
the P's that satisfy (12.6). If II( w)<m, then The expected utility of the player is Eu( w+ V). By Jensen's Inequality we
have
u(w-II(w)) = Eu(w-X)
Eu(w+V) < u(w+EV) < u(w);
by the continuity and strict monotonicity of u. We call II( w) the risk-holder's zero
utility premium for the risk X. It is clear that (12.6) is satisfied if and only if the last inequality follows from the monotonicity of u and the inequality E V<O. We
P~II( w), that is, the risk-holder insures if and only if P~II( w). see that the player reduces his expected utility by participating in the game. Fur-
In subsection 12A we wondered whether the risk-holder was crazy if he insured thermore, often the same person participates in hazard games and has some sort of
X for a premium greater than EX. The following theorem implies that this is not insurance.
necessarily the case. The explanation of this apparent paradox seems to be that the expected utility
of money does not tell the whole story. By participation in a hazard game, there is
Theorem 12.2. We have also the excitement involved which is not taken into account in our utility argu-
ments. However, the player would normally have home insurance as he does not
II(w) >EX. feel the same positive excitement with regard to the uncertainty of his home burn-
ing down as with regard to the uncertainty of winning in a lottery.
Proof. The theorem obviously holds if II( w)=m. If II( w)<m, then Jensen's
-183-
-182-
12F. In this subsection we shall study whether it can be profitable for the
12E. We now return to the insurance context of subsection 12C. As we have
discussed in Chapter 3, the risk-holder often insures only a part r(X) of the risk and risk-holder to under-insure the risk X.
Let V be the insurable value and S=(1-k) V the sum insured of the object to
covers the rest s(X)=X-r(X) himself. Let Pr be the premium for such an arrange-
which the risk X is related. We assume that the policy gives pro rata cover, and
ment. If he has the choice between several compensation functions r, he will choose
that the premium is proportional to the sum insured; that is, the premium is
the one that maximises his expected utility Eu(w-Pr-s(X)). In particular he will
(1-k)P. The expected utility of the risk-holder is
choose an insurance with compensation function r rather than no insurance if
As u is strictly increasing and s(x)~x, (12.6) is satisfied for p =0. Let n (w) be the We assume that
supremum of the Pr's that satisfy (12.6). It is clear that ([2.6) is sati;fied for all
Pr<llr( w). Furthermore, since + +
hu(k) = ~u(w-(1-k)P-kX).
for all Q>ll( w), we must have O~llr( w)~ll( w). By letting P approach n (w) from
bel~w, it follows by the monotonicity of u and monotone co:vergence tha[ (12.6) is ~U(k) = E((~u(y)\ l(P-x)].
Y y=w-(1-k)P-kXJ
satisfied by llr( w) if llr( w)<ro. Corresponding to the special case r(x):x, we call
llr( w) the zero utility premium of the risk-holder. We have the following analogue
In particular
to Theorem 12.2.
Let us assume that P>EX, that is, the insurance company has a positive risk load-
Eu( w-Er(X)-X+r(X)) ~ Eu( w-X). part r,{ Yi) of Yi" Thus the total cover from the company is Ei~l r,{ Yi), and the
.N1 rt.( Yt.) himself.
risk-holder covers X- E t=
Hence Pr=Er(X) satisfies (12.6), which proves the theorem. Q.E.D. Let us define a compensation function ron the total risk X by
-185-
-184-
of risk j. rinci le
Theorem 12.4. We have We shall look a little more closely at case ii). Let H be the premmm P . p .
. s should the risk-holder buy n policies separately, policy t
used The ques t 10n 1 , r(X)
. ~Y) for the premium H(r.(Y.)), or should he buy the global cover
covermg r~ . i Hl(r(X))? He would ~prefer the global solution if and only if
for the premmm .
N N
Eu(w-P-X+Ei= 1 rt.(Yi)) = EE[u(w-P-X+Ei= 1 ri(Yi))IXJ S By Theorem 12.4 and the monotonicity of u, a sufficient condition for this inequali-
Eu(E[w-P-X+Ei~1 rt{Yi)IXJ) = Eu(w-P-X+r(X)). Q.E.D. ty is
total loss of risk i. Then X= Ei~ 1 Yi, that is, the random variable N is now re- d h (12.8) also holds for the standard deviation principle. Analogously w.e
an ence ) al t' fi d b the van
placed by a constant n. Risk 1 could represent motor insurance, risk 2 home insur- can show that if the n risks are independent, then (12.8 is so sa IS e y -
ance, risk 3 personal accident insurance, etc. ance principle.
-186- -187-
Returning to the general assumptions, we have the following result. Eu( w-P-X+r1(X)) ~ Eu( w-P-X+r(X)). (12.9)
Q.E.D.
and for x~b
To me, this corollary seems even more important than Theorem 12.4. It is by
no means obvious that the risk-holder would have the choice between the local and
the global solution for the same premium. However, in Corollary 12.1 the zero utili-
ty premium can be interpreted as a measure of risk. For the risk-holder, the global
and (12.9) follows from Ohlin's Lemma. Q.E.D.
solution is less risky than the local solution, and thus he is willing to pay more for
the former solution. Similar remarks about the risk-holder's assessment of risk can
Let H be a premium principle. It would of course be nice if one could show
be made in connection to Theorems 12.5 and 12.11, but will not be repeated there.
that
12H. According to Theorem 12.4 the risk-holder would prefer the global cover (12.10)
Eu( w-H(r (X))-X+r1(X)) ~ Eu( w-H(r(X))-X+r(X)).
1
r(X) to the local cover E:1 ri( Yi) if the premium was the same for both solutions.
One argument against using the compensation function r as defined by (12. 7), is This result holds in particular if the premium is uniquely determined by the pure
that it could be complicated to calculate r, and that r could have an impractical premium under the premium principle H, since in that case we have H(~X))=
form. However, the particular form of r given by (12. 7) is not the essential feature H(r (X)) as Er (X)=Er(X), and thus we can use Theorem 12.5. The ~ost I~por
1
of Theorem 12.4. The essential feature is that to any local insurance cover there tant case where the premium is uniquely determined by the pure premmm, IS the
exists a global cover that is at least as attractive for the risk-holder as the local expected value principle.
cover. Thus, when searching for optimal insurance forms, we can restrict ourselves More generally, (12.10) is satisfied if H(r1 (X))~H(r(X)) as the utility function
to global forms. The following theorem says that the optimal global insurance form u is strictly increasing. Unfortunately, this inequality does not always hold.
is a cover with fixed amount deductible.
121. Analogously to the risk-holder, we now assume that also the insurer has a
Theorem 12.5. We assume that the risk X is insured with compensation func-
utility function; let us call it u. We make the same assumptions about u as about u,
tion r and premium P. Then there exists a fixed amount deductible b such that the
corresponding compensation function r1(x)=(x-b)+ satisfies the condition Er (X)= that is, that it is strictly increasing and strictly concave. Let wbe the fortune ~f t_he
1
Er(X). This compensation function satisfies the inequality insurer at the beginning of the insurance year. We make the (extremely unrealistic)
assumption that during the year the fortune of the insurer will not be influenced by
-189-
-188-
any other factors than possibly the risk X. Thus, at the end of the year, the fortune
is w if the insurer does not insure, and iii+ P-r( X) if he covers r( X) for the premi urn
P. He will prefer to cover r( X) for P to no insurance if and only if and a necessary condition for a contract is that II/w)$ll/w).
As u is strictly increasing, {12.11) is not satisfied for any P<O. Let IIr(W) be the 12J. W e now make the same assumptions and use the same notation as in
supremum of the P's that do not satisfy (12.11 ). As ( 12.11) is not satisfied for any subsection 12G.
P<O, we must have IIr(W)~O. Furthermore, as u is strictly increasing, {12.11) is not Theorem 12.7. We have
satisfied for any P<II(W). On the other hand, if IIr(w)<m, then {12.11) is satisfied
for all P>IIr(w). By letting P approach IIr(w) from above, we see by the monotoni-
city of u and monotone convergence that {12.11) is also satisfied for P=II/w). We
Proof Application of Jensen's Inequality gives
call IIr(W) the zero utility premium of the insurer.
Corollary 12.2. If II1(w) and IIg(W) are the zero utility premiums of the insur-
er for the local and global solutions respectively, then II\w)~ IIg(w).
and as u is strictly increasing, {12.12) must hold. This completes the proof of Theo-
rem 12.6. Q.E.D.
Proof From Theorem 12.7
As shown above, the insurer would be willing to insure for all P~IIr(W). Fur-
thermore, as a consequence of Theorem 12.6 he would never cover r(X) for a premi-
um P$Er(X); he would always require a positive loading. This is consistent with
Q.E.D.
Property 3 of Chapter 4. On the other hand, the risk-holder would insure if and
only if P$llr{ w). Hence a contract is made if and only if
In the same way as we used the zero utility premium of the risk-holder as a
-190- -191-
measure of his attitude to risk in connection to Corollary 12.1, we now use the zero ance does not seem to be that attractive for the risk-holder. The fear of large losses
utility premium of the insurer as a measure of his attitude to risk. Like the risk- is perhaps his main reason for buying an insurance policy, but with a first risk insur-
holder, the insurer finds the global solution less risky than the local one, and he is ance the relative compensation r(x)fx is greater for a small loss than for a large loss.
therefore willing to charge less for the former solution. Similar remarks about the For a risk-holder with such a motivation for insurance, it would be desirable if
insurer's attitude to risk can be made in connection with Theorems 12.8-11, but r(x)fx was non-decreasing. A compensation function with this property is called a
will not be repeated there. Vajda function.
Let us look a little more closely at the extreme case ii) of subsection 12G, and
let H be the premium principle used. Then the insurer would prefer the insurance of Theorem 12.9. We assume that the risk X is insured with compensation func-
X with compensation function r to separate policies of Y ,... , Yn if and only if tion r and premium P. Then there exists a unique proportional deductible k such that
1
the corresponding compensation function r3(x)=(1-k)x satisfies the condition
Er (X)=Er(X). If r is a Vajda function, then
Eu(w+ H(r(X))-r(X)) ~ Eu(w+ E.~ H(r{Y.))-E.n r'Y.)).
~- 1
3
~ ~ ~=1 ,, t
From Theorem 12.7 and the monotonicity of u, a sufficient condition for this inequa- Eu(w+P-r (x)) ~ EU(w+P-r(X)).
3
lity is that
Proof With
(12.13)
A comparison of (12.13) and (12.8) indicates that the risk-holder and the insurer
will not necessarily always have the same opinion about the global solution r(X). In r (x)=(1-k)x satisfies Er3(X)=Er(X).
3
particular, the insurer will not necessarily prefer the global solution if His the stan- Let x be such that r(x )sr (x1). Then, if r is a Vajda function, we have for
1 1 3
dard deviation principle. all XE(O,x1)
12K. Unfortunately, the risk-holder and the insurer have very divergent
preferences in regard to compensation functions for global solutions.
Theorem 12.8. We assume that the risk X is insured with compensation func- Thus there must exist an x0 such that
tion r and premium P. Then there exists a sum insured S such that the correspond-
ing first risk insurance with compensation function r: (x)=min(x,S) satisfies the r (x) ~ r(x)
3
condition Er2(X)=Er(X). This compensation function s~isfies the inequality r (x) S r(x),
3
12L. In subsections 12H and 12K we have discussed which compensation func-
The proof of Theorem 12.8 is very similar to the proof of Theorem 12.5 a~d is
therefore omitted. tions are optimal for risk-holder and insurer. In the present subsection we shall look
at a completely opposite question, namely which compensation functions are worst
By Theorem 12.8, first risk insurance is optimal for the insurer if the premi~
for risk-holder and insurer.
is uniquely determined by the pure premium. On the other hand, first risk insur-
-192- -193-
( i=1,2,3)
Proof As r4 and rare non-decreasing and
i) If s(x)~sl(x) for all x, then we would have s(X)=s1(X) almost s~rely, and
r(x) ~ r4 (x) (x<:d)
( 12 _14 ) would hold with equality. Let us therefore assume that there eXIsts an x1
r(x) ~ r4 (x), (x>d)
such that s(x1)>s1(x1). Then this x1 must be greater than b since_fo~ ~b we ha~~
s(x)~=s (x). As 8 is non-decreasing, we have s(x)~s(x1 )~s 1 (x1)-b-sl (x) for
Theorem 12.10 follows from Corollary A.2. Q.E.D. 1
~x1. Hence, there must exist an xo~b such that
It is interesting to see how beautifully Theorems 12.8 and 12.10 match Theo-
s(x) ~ s
rems 9.9 and 9.10 and the remark immediately after Theorem 10.3; the different 1(x)
s(x) ~ s
optimality criteria agree on the conclusion that it is the tail of the claim amount 1(x),
distribution that is dangerous for the insurer.
Analogous to Theorem 12.10 we can show that r (x)=xl(x<t) is the worst that is,
5
possible compensation function for the risk-holder in the sense that for all compensa-
r(x) ~ r1(x)
tion functions rfor which x-r(x) is non-decreasing and Er(X)=Er (X), we have
5 r(x) ~ r1(x),
Eu( w-P-X+r5(X)) ~ Eu( w-P-X+r(X)).
and (12.14) follows from Corollary A.2.
ii) The proof of (12.15) is analogous to the proof of (12.14).
The following theorem says that if we restrict ourselves to the situation where
iii) Let us now assume that r is a Vajda function. Then by the proof of Theo-
both the compensation function and the self-insurance function are non-decreasing,
rem 12.9 there exists an x0 such that
then the insurer and the risk-holder have completely opposite preferences concerning
compensation functions. The intuitive interpretation of this result is that they both
fear the tail of the loss distribution. s(x) ~ s3(x)
s(x) ~ s3(x),
Theorem 12.11. We assume that the risk X is insured with compensation func-
and (12.16) follows from Corollary A.2.
tion r and premium P, and that the compensation function and the selfinsurance Q.E.D.
This completes the proof of Theorem 12.11.
function s(x)=x-r(x) are non-decreasing. Let r 1,r2 , and r be defined as in Theo-
3
rems 12.5, 12.8, and 12.9. Then we have
12M. An insurance policy for the risk X is defined by the pair (r,P), where r
is the compensation function and Pis the premium. Let
Eu(w+P-r1(X)) ~ Eu(w+P-r(X)) (12.14)
Eu( w-P-X+r2(X)) ~ Eu( w-P-X+r(X)), (12.15)
-194- -195-
Exercise 12.1
For a proof of Theorem 12.12 we refer to DuMouchel (1968).
An art collector has a fortune wand a utility function u. His fortune includes
From Theorem 12.12 we see that the Pareto optimal policies will depend very
in particular two valuable paintings of value x and y. He wants to send these paint-
much on the shape of the utility functions and the initial fortunes of the risk-holder
ings to his summer residence by ship. We assume that for this tr.ip, each s~i~ .has a
and the insurer, and thus we shall not go further into the theory of Pareto optimal
probability p of getting lost, and that if it gets lost, then there IS no possibility of
policies here.
rescuing its cargo. Different ships get lost independent of each other.
Show that the expected utility of the art collector after the transporting the
12N. Before leaving utility theory, we mention some aspects that we have not
paintings is higher if he sends them by different ships, than if he sends them by the
treated, and make some critical remarks on what we have done.
same ship. We assume that there are no transport costs, and that it is not possible
i) In subsection 3E we mentioned some aspects concerning the choice of com-
-196- -197-
Exercise 12.2
Exercise 12.5
We consider a risk-holder who wants to insure a risk X. He has a quadratic A risk-holder wants to insure n independent and identically distributed risks
utility function with an insurer who calculates premiums using the expected value principle. The
risk-holder intends to under-insure by insuring only a quota ki ofrisk i ( i=1,2, ... , n).
Exercise 12.8
A risk-holder wants to insure a risk X and is offered insurance with a deductib-
le given by the self-insurance function
can be distributed between the projects in a way under which each of them gets at
least as great expected utility as it would have had when each of them had carried
its own risk. What is this distribution when the risks are independent and X. is
~
s(x) = { ~ + k(~b). g~~~
-198- -199-
The premium is calculated according to the expected value principle. The insurer Appendix A
k~k >0,
requires that
0 and that the expected self-insured amount of the risk-holder
should be equal to a fixed amount 7J>k0EX. Convex and concave functions
a) How should the risk-holder choose k and b to maximise his expected utility
under these conditions?
b) What choice of k and b gives the highest expected utility for the insurer? Let u be a real-valued function. We shall say that u is strictly convex on an
interval I if for all yEI there exists a number k(y) such that for all xEl, #y, we have
Exercise 12.9
(A.1)
We consider a risk-holder who wants to insure a risk X. Both the risk-holder u(x) > u(y) + k(y)(x-y).
and the insurer have an exponential utility function, respectively
We say that u is convex on !if the weaker condition
-ax
u(x)=8_
a (a>O) u(x) ~ u(y) + k(y)(:-y)
(x>O; a.,/1>0) Proof Assume that u is strictly convex, and let x,yEI, #y, and tE(O,l). By
using Jensen's Inequality on a random variable X with distribution given by
Comment on the case where fl is small.
Pr(X=x) = t = 1- Pr(X=y),
we obtain (A.4).
-200-
Now assume that (A.4) is satisfied for all x,yEI, xfy, and tE(O,l). We denote
1 \
and we see that
increasing.
iizu+
-201-
the line through (x,u(x)) and (y,u(y)) by L(x,y) and introduce its slope I
I As a monotone function cannot have more than a countable number of discon-
tinuities, the last part of the theorem follows.
Q.E.D.
S(x,y) = u(yt=~x)_ Theorem A.2 is now proved.
From Theorem A.2 and the definition of strict convexity we obtain the follow-
Let y ,y,xei; y <y<x. From (A.4) it follows that (y,u(y)) lies below L(y0,x),
0 0
and thus S(y ,y)<S(y,x). Analogously we see that S(y,x) decreases strictly when x ing corollary.
0
decreases, and as S(y,x) is bounded from below (by S(y0,y)), the limit
Corollary A.l. A twice differentiable function is strictly convex if and only if its
Theorem A.2. A strictly convex function u is continuous and possesses left and Lemma A.2 (Ohlin's Lemma). Let x1 and x 2 be random variables with the
right derivatives at each point. These derivatives are strictly increasing and same finite mean and cumulative distributions F 1 and F 2 . It is assumed that
X
1
,x2ei a.s. If there exists an x0 such that
F (x)~F (x)
1 2
The derivative of u exists, except possibly at a countable number of points. F 1(x)~F2 (x),
Proof The right continuity and existence of right derivatives follow from the then
proof of Theorem A.l. The left continuity and existence of left derivatives can be
(A.5)
shown by analogous arguments. As for all x and all h>O S(:J>-h,x)<S(x,x+h), we ob-
tain
for all convex functions u on I; we allow infinite values for these expectations.
(A.6) then Eu(g (Y)hEu(g2(Y)) for all convex functions u; we allow infinite values for
1
is non-negative, and t(x0)=0. Furthermore, from Theorem A.2 we see that vis non- these expectations.
increasing on r =In(-w,x0] and non-decreasing on I+ =In[x ,ro). We introduce the -1 ,2, 1et X ~.=g.(
0 Proof F or ~- t Y) and let F.~ be the cumulative distribution of
random variables
Xi' Let xo=g2(y0). For x<xo we have
We have
implying (Xl~x)c(X $x), and we obtain F1(x)$F2(x). On the other hand, for x~x0
2
we have
(yEI)
(gl ( Y)>x) U (g2( Y)>x~gl ( Y)) = (g2( Y)>x),
that is, Y1 is stochastically greater than Y . Furthermore, as Y , Y E1 and vis
2 1 2
non-increasing and non-negative on 1, we have 1 ng (X >x)c(X >x). Hence (X2$x)c(X1$x), and we have F1(x)~F2 (x). Carol-
Imp yi 1 2 QED
lary A.2 now follows from Ohlin's Lemma.
We have now shown some results for convex and strictly convex functi.ons.
Analogous reasoning gives Corresponding results for concave and strictly concave functions u can be obtamed
by application of the present results to -u.
Jl-
Unfortunately p r) is not always finite. The following theorem says that the
area on which it is finite is an interval containing zero. We see that r J s) is finite if IsI< 1.
B3. Let N be a non-negat1ve . ran dom vana bl e, an d let y 1' Y.2' ... be
. independ-
Theorem B.l. i) The set
ent and identically distributed random variables independent of N. We mtroduce
N
X= Ei=l Yi.
is a (possibly degenerate) interval containing zero.
ii) PxisstrictlyconvexonAxif Pr(X=O)<l. The distribution of X is called a compound distribution. We have
Proof If X=O almost surely, then px=l <ro, and i) trivially holds with ).x=IR.
Let us therefore assume that Pr(X=O}<l. As pJ!-O)=l<ro, OE.J.X' Let r,sE.J._x; s/r.
As ex is convex, Theorem A.l shows that for all tE(O,l), that is,
(B.l)
pJ!-tr+(l-t)s) = Ee[tr+( 1-t)s)X < E[terX+(l-t)lX] = pJI-r) = rrJ..p-.j...r)).
tpJ!-r) + (1-t)pJ!-s) < ro.
B4. The Laplace transform lfJx of a non-negative random variable X is defined
It can be shown (cf. e.g. Billingsley (1986)) that if the nth derivative of Px We see that I{J X is finite.
exists at zero, then
B5 For a more extensive treatment of moment generating functions we refer
to Billin~sley (1986) and for more details of Laplace transforms to Feller (1971).
-207-
-206-
In this appendix we shall briefly discuss some properties of this function and some In particular, for a=1 we obtain an exponential distribution.
related distributions.
Theorem C.2. The convolution of two Gamma distributions with parameters
Theorem C.l. r(x)<ro for all x>O . . l ( ~ /3) and (a {3) is a Gamma distribution with parameters ( a 1+a2,/3).
respec t we y ul' 2'
Proof Let x>O. We want to show that r(x)<ro. As Proof The density of the convolution is
(t~M)
This density has the shape of a Gamma density, and as it should integrate to oee,
Q.E.D.
Q.E.D.
which proves the theorem.
By partial integration we obtain
The Beta distribution with parameters ( a.,{J) is a continuous distribution on Author index
the interval (0,1} with density
. fl) -_ r(r(a)a.+fl~
v( x,a., I'(l xa.-1( 1-x){J-1 . (O<x<1; a.,{J>O) Andersen, E.S. 107
Beard, R.E. iii,145,146
In particular, for a.={J=1 we obtain the uniform distribution on (0,1}. Bernoulli, D. 196
Billingsley, P. 204,205
Boos, A. 77
Borch, K. 195
Borgan, 0. 74,77
Bowers, N.L. 120
Biihlmann, H. iii,47,48,68,88,94,100,121,137,172,195
Centeno, L. 120
Chamberlin, G. 167
de Finetti, B. 42
DeGroot, M.H. 42,177
Delbaen, F. 114
De Pril, N. 138,139,154,161
De Vylder, F. 15,16,18
Dhaene, J. 154
Dickson, D.C.M. 159
Dienst, H.R. 19
DuMouchel, W.H. 194,195
Embrechts, P. 111,159
Feller, W. 38,106,127,205
Gerber, H.U. 120
Gilde, V. 75,77
Goovaerts, M. 15,16,18,155
Grandell, J. 88
Graves, L.M. 91
Greup, E.K. 82,87
Hachemeister, C.A. 49,52,53
Haezendonck, J. 15,16,18,114
Heilmann, W.R. iii
Hesselager, 0. 53,120
Hoem, H.M. i,iv,74,77
Hogg, R. 100
-210- -211-