Beard Risk Theory - The Stochastic Basis

MONOGRAPHS
ON APPLIED PROBABILITY AND STATISTICS
General Editor8: M. S. BARTLETT, F.R.S.

and D. R. COX, F.R.S.
RISK THEORY
The Stochastic Basis
of Insurance
Risk Theory
THE STOCHASTIC BASIS
OF INSURANCE
R. E. BEARD, O.B.E., F.I.A., F.I.M.A.

London, England
T. PENTIKAINEN, Phil. Dr.
Helsinki, Finland
E. PESONEN, Phil. Dr.
Helsinki, Finland
SECOND EDITION
LONDON
CHAPMAN AND HALL

A Halsted Press book
John Wiley & Sons, New York
First published 1969
by Methuen & 00. Ltd
Second edition 1977
published by Ohapman and Hall Ltd
11 New Fetter Lane,
London E04P 4EE
© 1969, 1977 R. E. Beard,
T. Pentikiiinen, E. Pesonen
ISBN-13: 978-94-009-5783-1
All rights reserved. No part of
this book may be reprinted, or reproduced
or utilized in any form or by any electronic,
mechanical or other means, now known or hereafter
invented, including photocopying and recording,
or in any information storage or retrieval
system, without permission in writing
from the publisher
Distributed in the U.S.A. by Halsted Press,
a Division of John Wiley & Sons, Inc., New York
Library of Congress Cataloging in Publication Data
Beard, Robert Eric.
Risk theory.
(Monographs on applied probability a.nd statistics)
Bibliography: p.
Includes indexes.
1. Insurance - Mathematics. 2. Risk (Insurance)
I. Pentikainen, Teivo, joint author. II. Pesonen,
Erkki, joint author. III. Title.
HG8781.B34 1977 368'.001 '519 77-1637
ISBN-13: 978-94-009-5783-1 e-ISBN-13: 978-94-009-5781-7
DOl: 10.1007/978-94-009-5781-7
Softcover reprint of the hardcover 1st edition 1977
Contents
Preface to Second Edition page xi

Preface to First Edition. xiii
I. Definitions and Notations
1.1. The Purpose of the Theory of Risk I
1.2. Random Processes in General 3
1.3. Positive and Negative Risk Sums 3
1.4. Main Problems 5
2. Process with Constant Size of One Claim

2.1. Introduction 7
2.2. The Poisson Process 8
2.3. Discussion of Assumptions 8
2.4. Numerical Calculations 10
2.5. Application I 12
2.6. Application 2 15
3. Generalized Poisson Distribution

3.1. The Distribution Function of the Size of a Claim 18
3.2. Generalized Poisson Function 21
3.3. The Mean and Standard Deviation of F(x) 22
3.4. Characteristic Function 24
3.5. Estimation of S(z)
3.5.1. Individual Method 25
3.5.2. Statistical Method 27
3.5.3. Problems Arising from Large Olaims 29
3.5.4. The Dependence of the S-Function on Reinsurance 30
3.5.5. Analytical Methods 33
3.5.6. Exponential Function 34
3.5.7. A Generalization of the Exponential Type 36
3.5.8. Other Types of Distribution 37
3.6. Decomposition of S(z) 37
vii
CONTENTS
4. Normal Approximation and Edgeworth Series for F(x)

4.1. The Normal Approximation 41
4.2. Edgeworth Series 42
4.3. Normal Power Expansion 43
4.4. The Accuracy of the Normal Approximation 47
5. Applications of the Normal Approximation

5.1. The Basic Equation 52
5.2. Net Retention 54
5.3. Reserve Funds 58
5.4. Statutory Basis of Reserve Funds 63
5.5. The Rule of Greatest Retention 64
5.6. The Case of Several M's 65
5.7. An Application to Insurance Statistics 68
5.8. Experience Rating, Credibility Theory 69
6. The Esscher Approximation

6.1. Introduction 76
6.2. The Accuracy of the Esscher Formula 79
6.3. Some Hints for Numerical Computations 82
6.4. Examples of Numerical Applications 84
7. Monte Carlo Method

7.1. Random Numbers 91
7.2. Simulation of Generalized Poisson Function 94
7.3. Discussion on the Accuracy and a Modification 95
8. Other Methods of Calculating the Generalized Poisson

Function
8.1. Inversion of the Characteristic Function 98
8.2. A Modification of the Esscher Method 99
8.3. Step Function Approximation of S(z) 99
8.4. Exponent Polynomials 100
8.5. Mixed Methods 100
8.6. Statistical Method 101
9. Variance as a Measure of Stability
9.1. Optimum Form of Reinsurance 103
9.2. Reciprocity of Two Companies 106
viii
CONTENTS
10. Varying Basic Probabilities

10.2. Compound Poisson Process 114
10.3. Direct Numerical Computation of the Compound
Poisson Function 121
10.4. The Polya Process 124
10.5. Application to Stop Loss Reinsurance 129
n. The Ruin Probability During a Finite Time Period
11.1. The Ruin Function in Finite Time Periods 132
11.2. Calculation of lJ'N(U) by a Monte Carlo Method 134
12. The Ruin Probability During an Infinite Time Period
12.2. Ruin Probability 138
12.3. Applications 147
! 2.4. Some Approximation Formulae 151
12.5. Discussion on the Different Methods 156
13. Application of Risk Theory to Business Planning 160
Appendix A. Derivation of the Poisson Process and Compound
Poisson Processes 167
Appendix B. The Edgeworth Expansion 173
Solutions to the Exercises 175
Bibliography 182
Author Index 187
Subject Index 188
IX
Preface to Second Edition
Since the publication of Risk Theory in 1969, there has been a

continued growth of interest in the subject. ASTIN, the section of
the International Actuarial Association concerned with the subject,
now has well over a thousand members and there are few actuarial
societies which do not include some aspects of risk theory in their
education and training. A number of Universities and technical
institutions now have courses of study and, on the application side,
the growth in the concept of risk management, namely the technique
of total financial management planning, has emphasized the im-
portant part played by the theory of risk.
We have taken this opportunity to correct a number of misprints
which have come to light. Fortunately developments have not
invalidated the text as an elementary introduction to the subject,
but the opportunity has also been taken to rewrite Chapter 13 to
reflect the current direction of development in applications. Were
a new book being written, some changes of emphasis would be
appropriate and we have indicated these in the Preface, together
with additional references to avoid modification of the list and
alterations to the basic text.
In Chapter 5.4, dealing with the statutory requirements for
excess reserves above the usual provisions for unexpired risks and
outstanding claims, reasons were advanced that as regards fluctu-
ations in risk this margin should be fixed in proportion to the square
root of the size of the business. The European Economic Community
have now issued a non-life directive setting out the rules for calcu-
lation of the minimum solvency margin so that member countries
will be modifying their own legislation to conform to the provisions
of the directive.
The discussion on Credibility Theory in Chapter 5.8 was largely
concerned with the basic principles originally developed in the
U.S.A. for premium rating purposes. Over the past few years there
has been a considerable research in this subject but the extent to
Xl
PREFACE
whioh the developments are appropriate in an elementary text book

is open to doubt. Fortunately the proceedings of the conference
arranged by the Society of Actuaries Research Committee in
September 1974 provide an effective review of the ourrent position
(Credibility, Theory and Applications, Ed. P. M. Kahn, Academic
Press, 1975).
It is doubtful if any practical use is now made of the Esscher
approximation and the N-P method is much more convenient and
of adequate accuracy in most practical work. Thus the first half of
Chapter 6 is now largely of historical interest.
Chapter 11 dealing with ruin probability during a finite time
interval does not give an adequate view of the current importanoe
of this topic but the position is fluid because of the considerable
effort being expended in the search for practical methods of calcu-
lation. Formulae are, in general, complicated and involve extensive
computer based quadratures or simulation techniques. The paper
by Seal in the Scandinavian Actuarial Journal (The Numerical
Calculation of U(w,t) the Probability of Non-ruin in an Interval
(O,t) 1974) gives a recent treatment and a fairly complete list of
relevant references.
In many countries studies are currently in progress in the develop-
ment of models for business planning where the basic operations
involve a stochastic process. Not only are insurance companies
interested but in many commercial and industrial firms the needs
are significant so that a very large field exists for applications.
Chapter 13 has been recast to provide a natural starting point for
developments. This is largely based on recent work by Pentikiiinen,
and further references are included in that chapter.
As regards other publications, reference should be made to the
important text books by Biihlmann (Mathematical Methods iIi Risk
Theory, Springer, 1970) and by Seal (Stochastic Theory of a Risk
Business, Wiley, 1969) which provide more advanced treatment of
some of the topics in Risk Theory. Numerous papers have been
presented at the various colloquia organized by ASTIN and pub-
lished in the ASTIN Bulletin. Further papers will also be found in
SAJ (Scandinavian Actuarial Journal) and in the Transactions of
the Casualty Actuarial Society.
Robert Eric Beard
London and Helsinki Teivo Pentikiiinen
August 1976 Erkki Pesonen
xii
Preface to First Edition
The theory of risk already has its traditions. A review of its classical
results is contained in Bohlmann's paper published in the transactions
of the International Congress of Actuaries, Vienna, 1909. This classical
theory was associated with life assurance mathematics and dealt
mainly with deviations, which were expected to be produced by
random fluctuations in individual policies. According to this theory,
these deviations are discounted to some initial instant; the square
root of the sum of the squares of the capital values calculated in this
way then gives a measure for the stability of the portfolio. A theory
constituted in this manner is not, however, very appropriate for
practical purposes. The fact is that it does not give an answer to such
questions as, for example, within what limits a company's probable
gain or loss will lie during different periods. Further, non-life
assurance, to which risk theory has, in fact, its most rewarding
applications, was mainly outside the risk theorists' interest. Thus
it is quite understandable that this theory did not receive very
much attention and that its applications to practical problems of
insurance activity remained rather unimportant.
A new phase of development began following the studies of Filip
Lundberg, which, thanks to H. Cramer, C. O. Segerdahl, and other
Swedish authors, has become generally known as the 'collective
theory of risk'. As regards questions of insurance the problem was
essentially the study of the progress of the business from a pro-
babilistic point of view. In this form the theory has its applications
to non-life insurance as well as to life assurance. This new way of
expressing the problem has proved fruitful and the development of
the theory has since been continued by several other authors. In
recent years the fundamental assumptions of the theory, and thus
the range of its applications, have been significantly enlarged by the
use of more general probability models, which allow, for example,
for certain types of fluctuation in the basic probabilities.
xiii
PREFACE
Today the theory of risk generates an interesting and far· reaching
field for research and the development of the theory is still far from
complete, as is demonstrated by the many papers which continue
to be published on the subject. Studies concerning both its basic
foundations and its applications have been numerous and there is
no reason to suppose that the development will cease during forth·
coming years. A defect, much the same as in so many other new and
rapidly developing branches of human knowledge, is that the theory
has become difficult for practising actuaries to follow. This is
regrettable, because a knowledge of this theory deepens actuarial
intuition and helps the understanding of insurance business as
a process characterized by varying progress and fluctuations from
year to year. The modern theory of risk can also give an actuary
concrete assistance in the form of practical applications. It is true
that many problems in this field, for example, problems of a com·
pany's solvency, reinsurance requirements, safety loadings in the
premiums, and many others, are such that risk theory alone is
incapable of providing a definite solution, because in practical work
it is often necessary to take into consideration many aspects with
which risk theory is not competent to deal. In reinsurance arrange·
ments, for example, attention has to be given to many insurance.
political aspects such as reserves, reciprocity, liquidity, and others.
In spite of this, when choosing a form of reinsurance and calculating
suitable net retentions and safety loadings, risk theory provides
effective tools to estimate the fluctuations in the business retained
by a company; such fluctuations should obviously always be kept
within the limits of the company's resources. Thus the theory of
risk can facilitate important considerations of financial interest and
be useful in making final decisions.
To disseminate knowledge of the theory of risk it seemed essential
to provide an introduction to the theory based upon the elements
of probability theory which form part of actuarial study and which
provide some of the basic ideas concerning risk theory. Furthermore
there is a need for a summary of the results of the present theory,
easily available for practical application.
For this reason, one of the authors, Pentikiiinen, published an
elementary textbook of the risk theory in the Finnish language in
1955, primarily designed for the use of Finnish actuaries, as an
introduction to the theory. Risk theory is included in the syllabus
xiv
PREFACE
for the actuarial examinations in Finland, and candidates are set
a practical exercise on the application of the theory.
Many participants at meetings of ASTIN* have expressed a wish
for a concise book of this kind in English and directed primarily
to practical applications. The authors have attempted to produce
such a book. The basis of compilation is that the Finnish authors
have rewritten and brought up-to-date the above-mentioned Finnish
textbook, and have passed it to the English author, who, for his
part, has worked it into shape, taking into consideration British
circumstances and paying special attention to the general actuarial
education in English-speaking countries.
To prevent the book becoming too large and developing beyond
the limits of a primary textbook it has been necessary to limit the
subject matter. This has been a very difficult task, having regard
to the very abundant field which the theory and its applications
comprise today, and from necessity many interesting aspects of the
theory have been omitted. Furthermore various alternatives, lines
and methods of presentation are possible. Our aim has been for
simplicity. The more so because the main purpose of this book is to
serve as a first introduction to the theory of risk since there are
several publications dealing with advanced aspects of parts of the
theory. On the other hand the authors have been quite conscious of
the risk of oversimplification, which could reduce the theory to
'pseudoscience' and ignorance of the basic assumptions of the theory
could lead to serious mistakes when applying the theory to various
actuarial problems. For this reason the basic foundations of the
theory have not been omitted. The main lines, the practical one and
the theoretical one, are, unfortunately, not easy to fit together and
in the present state of the theory a firm bridge between the practical
problems and exact theory is often not fully developed. Having
regard to the fact that our main purpose is practical, we have also
been obliged to present formulae which are based on approximations
without well-mapped confidence limits. We have also attempted a
compromise between accuracy and simplicity, transferring some
cumbersome considerations to the appendices, which can be omitted
at a first reading.
We sincerely hope that this book will prove to be only a first
* ASTIN = Actuarial Studies in Non-Life Insurance - a section of
the International Actuarial Association.
xv
PREFACE
step to each reader in his introduction to the theory, and that
sufficient interest will be stimulated to provoke a more extensive
and profound investigation. An extensive bibliography has been
included to assist the reader in this direction.
The book has been written on the assumption that the reader
has a knowledge of elementary probability theory, as presented for
example in H. Cramer's 'The Elements of Probability Theory and
Some of Its Applications' (Uppsala, 1954). For some derivations a
knowledge of complex integrals is implied but readers who are not
familiar with this theory should not lose the general trend by
omitting the details concerned. Some of the proofs also use the
techniques of advanced probability calculus and, although not
requiring special knowledge of advanced mathematics, are com-
plicated and cumbersome. This applies, for example, to the proof
of Esscher's formula (Section 6.1) and especially to the introduction
of the ruin probability during an infinite time period (Section 12.1).
As indicated in the relevant paragraphs, the proofs can be omitted
on a first reading and the reader can proceed directly to the final
formulae.
Our particular thanks are due to ASTIN for its initiative and
interest in this publication and especially to Messrs G. Benktander,
P. R. Cox, J. Jung, Carl Philipson, and J. Sousselier, who kindly
read our manuscript and made many valuable suggestions for im-
provement. The typing of the manuscript for printing from numerous
texts and formulas was undertaken by Brita Aalto, to whom we
owe our special thanks for accurate work requiring great patience.
London and Helsinki Robert Eric Beard
February 1968 Teivo Pentikiiinen
Erkki Pesonen
XVI
CHAPTER 1
Definitions and Notations
1.1. The Purpose of the Theory of Risk

Conventional actuarial techniques are largely based on frequencies
and the average amounts of claims. For example, if an insurer has
a portfolio of N policies at risk and if the expected mean value of
the claim frequency for these policies during a specified period is q
and the expected average size of the claim is m, then the expected
total amount of claims is N qm. However the actual amounts arising
from several successive periods will differ from this expected figure
and will fluctuate around it. In probabilistic terms, the actual
amount of claims is a random variable. Conventional actuarial
techniques are in fact based on a simplified model of an insurance
portfolio in which random variables are replaced by this mean value,
i.e. the fluctuation phenomenon is disregarded. Whilst for many
purposes this simplified model is sufficient in the hands of experts,
it is undeniably an over-simplification of the facts and it is both
useful and interesting to develop the principles of insurance mathe-
matics on a more general basis, in which both the number and size
of claims are considered as random variables. Studies ofthe different
kinds of fluctuation appearing in an insurance portfolio which start
from this point of view constitute the branch of actuarial mathe-
matics termed the theory of risk.
Of course, the financial structure of an insurance company depends
on management costs and investment of capital in addition to the
claims aspects, but these two factors are not subject to random
fluctuation in the same way as claims, and the theory of risk is not
therefore an appropriate technique for use in their study. Unless
otherwise mentioned the following analysis is therefore restricted
to the study of claims and to that part of the premiums which
remains when loadings for expenses of management have been
deducted, i.e. risk (net) premiums increased by a safety loading. In
I
RISK THEORY
particular interest earnings are disregarded. The fact that this book
has been confined, for the sake of simplicity, to studies relating to
pure risk business only, should not be construed as implying that
these are the only significant aspects; there are many occasions
when other considerations are of greater significance than the purely
risk aspects.
The claim process can be described graphically as in Fig. 1.1
Every occurrence, from which a claim arises, is represented by a
vertical step, the height of the step showing the amount of the claim.
Time is measured to the right along the x-axis and the difference
in the altitude ~ of the stepped line at points tl and t2 shows the
total amount of claims during this time interval. The process is, in
fact, a compound random process in the sense 'that the time of
occurrence and the number of occurrences is a random phenomenon
and the amount of each claim is also a random variable.
If the whole risk business of an insurance portfolio is considered,
this can be illustrated graphically as shown in Fig. 1.2. The net
premium P together with a safety loading is continuously flowing
in; this is accumulated in a risk reserve of an initial amount U0' so
that the income is represented by a line sloping upwards towards
the right. The claims, which can be regarded as negative income,
Fig. 1.1.
2
DEFINITIONS AND NOTATIONS
Risk rese rye
o t Time
Fig. 1.2.
are paid out from this reserve and are represented by downward
steps. The difference U - U 0 gives the profit ( +) or loss ( - ) arising
during time t.
1.2. Random Processes in General

The theory of risk is essentially a special case of the theory of random
or stochastic processes which has grown rapidly in recent years and
now constitutes a large branch of probability theory. Other examples
of such processes are the number of calls in telephone systems, the
emission of radiation from radioactive substances, the movement
of equity prices on stock exchanges, or the different kinds of
'random walks'. These and other processes have similarities with
the risk process of an insurance portfolio and a number of textbooks
are now available for those who wish to study the subject more
deeply (e.g. Doob, 1953, or Cox and Miller, 1965).
1.3. Positive and Negative Risk Sums

Fig. 1.3 represents a realization or a sample function of a general
type of risk process in which any occurrence giving rise to a down·
ward step represents a loss, this being the case in classes such as
fire, marine, and life (death risk only). A different situation arises
with immediate annuity business since the initial fund is continuously
depleted until an annuitant dies, when the reserve released gives rise
3
RISK THEORY
Positive risk sums
Fig. 1.3(a).
Nega.tive risk sums
Fig. 1.3(b).
to an upward step. This latter type is called a risk process with

negative risk sums. In this book, positive risk sums only are dealt
with as being the case of greater interest having regard to a pplications
but the cases of negative or mixed risk sums have been considered by
some authors whose works can be consulted (Cramer, 1955).
4
DEFINITIONS AND NOTATIONS
1.4. Main Problems

The fundamental objects of the theory of risk to be presented can,
in the first place, be classified into three main groups, which can be
eonveniently described as the provision of meaningful answers to the
questions:
(i) What is the result of the business at the end of a certain
period T (e.g. 1 year)~
In terms of Fig. 1.4(a) this is equivalent to finding the probability
for the different values that UT can assume, particularly the
probability that the loss during this period will equal or exceed the
initial reserve U 0' hence making U T negative (the so-called ruin
probability).
(ii) What is the result if the observations are extended to each
point of time period T; i.e. what is the probability that ruin
will occur at some point of time during this period (Fig. 1.4(a))~
A modification of this problem arises when observations are taken
only at certain set time points T l' T 2' ••• T 11 of the interval T (e.g.
at the end of each fiscal year). The question of ruin arises if U is
negative at one or more of these special points of time (Fig. 1.4(b)).
(iii) What is the result if the time T under (ii) tends to infinity,
i.e. what is the probability that the business will never be
ruined~
Fig. I.4(a). Checking only at end of the observation period T

(problem (i)) or at each point of the period T (problem (ti».
S
RISK THEORY
Uo Uv
o T, TV=T
Fig. 1.4(b). Checking at several points T l' T 2 ••• Tv. during the
observation period.
These questions and others which arise naturally from them are
treated in this book. In Chapter 2 the simplified special case is
considered in which it is assumed that all claims (e.g. the sums
insured in non-life insurance or the risk sums in life assurance) are
of equal amount; in this case, fluctuations arise solely from the
random variation in the number of claims.
The general case, where the individual amount of a claim may
vary, forms the subject matter of Chapter 3 and later parts of the
book.
6
CHAPTER 2
Process with Constant Size

of One Claim
2.1. Introduction
As already mentioned the simplest case is considered first, namely,
where the claims arising from the insurance portfolio under con-
sideration are all for the same amount. H this constant amount is
taken as the monetary unit, the total outgo will be equal to the
number of claims. The problem is to find the probability function of
the number of claims, i.e. a function Pk(t) which gives the probability
that the number of claims in time t is equal to k.
The following analysis is independent of whether the portfolio
concerned represents the aggregate of all risks insured by a company
or only a special part thereof.
The problem can be solved in a number of different ways. One
method is to start by regarding the portfolio in question as made up
of a number of individual policies, each of which has a certain
probability of claim (e.g. in life assurance it is assumed that the
probability that a life aged x dies within a year is qx). Then the total
number of claims is the sum of the contributions from the individual
policies and the Pk-function can be derived by means of the addition
theorem of probability calculus from the primary probabilities.
Basically the probabilities are mainly binomial in character but to
carry out this 'addition' in a rigorous way leads to rather intricate
calculations and involves some restrictive assumptions.
An alternative approach, which has led to fruitful development,
is to follow the collective method adopted by Lundberg. In this
method the individual policy structure is disregarded and instead
the portfolio is considered as a whole, i.e. a 'process' is considered
in which only time points and the number of events (i.e. claims) are
recorded and in which no attention is paid to the particular policies
7
RISK THEORY
from which the claims have arisen. By starting with some general
conditions which the random process has to obey, it can be deduced
that the Pk-function takes the well-known form of a Poisson process.
The Poisson process is often referred to in probability calculus as
the theory of rare phenomena and is well known, for example, in the
theory of disintegration of radioactive atoms. However, as it is
necessary in practical problems to know in which cases the Poisson
function is applicable and in which cases it is not, some discussion of
the assumptions underlying this probability distribution is essential.
2.2. The Poisson Process

By considering the portfolio as a whole and restricting attention to
the claims arising, the sequence of events (i.e. claims) is a random
process. The following three conditions are assumed:
(i) Events occurring in two disjointed time intervals are inde-
pendent (independence of increments).
(ii) The number of events in a time interval (tl' t 2 ) is dependent
only on the length of the interval t = t 2 -t 1 and not on the
initial value tl (stationariness of increments).
(iii) The probability that more than one event will occur at the
same time and the probability that an infinite number of
events will occur in some finite time interval are both zero
(exclusion of multiple events).
In Appendix A it is shown that with these conditions the process
is represented by the well-known Poisson function
(2.1)
where Pk(t} is the probability that exactly k events occur in the

half-closed time interval of length t and q is a parameter indicating
the average number of claims in a unit of time.
2.3. Discussion of Assumptions

Condition (i) means, in fact, that an event (e.g. a fire) cannot give
rise to any other events (exclusion of 'chain reactions'). In practice,
however, a fire can often spread from one risk to another in contra-
diction to this condition.
8
PROCESS WITH CONSTANT SIZE OF ONE CLAIM
Condition (i) can, however, often be met by defining, as is customary

in reinsurance practice, a risk unit as a combination of all those
risks lying near to each other, between which contamination is
possible (e.g. all property in a building irrespective of whether it is
formally insured by one policy or by several or being under single or
multiple ownership). In the same way a ship and its cargo are
considered as one risk unit, and so on. However, it is not always
possible to build up risk units in such a way that outside contamina-
tion would not occur. Such is the case with contagious diseases
in sickness insurance or epidemics in life assurance. The Poisson
function is not then applicable, at least not without suitable
modifications.
Condition (ii) concerning the stationariness of increments means
that the collective flow of the events is stationary, i.e. neither steadily
increasing or decreasing nor oscillating more than can be explained
by normal random fluctuations. This is the usual case in insurance,
particularly during short periods, when the numbers of policies or
other circumstances are not subject to marked changes. This con-
dition implies that the portfolio is so large that the exit of individual
policies by reason of claims or from other causes and the entry of
new cases cannot affect the collective flow of the events to any
significant degree.
However, there are often circumstances arising from practical
conditions, where conditions (i) and (ii) cannot be met. For example,
fire insurance can be greatly affected by weather conditions and a long
dry sunny period can give rise to numerous abnormal fires; in some
countries, hurricanes or other natural catastrophes can give rise to
an enormous increase in events. It is also well known that economic
conditions have considerable influence in many branches of non-life
insurance. Times of economic depression sometimes give rise to a
considerable increase in the number of fires, as well as influencing
credit insurance business. Circumstances like these are so general
that the application of the Poisson function is greatly limited, and
so there is a need for a development of the theory omitting the
conditions concerning independence and stationarity. The essential
features will be developed in Chapter 10. However, in spite of these
limitations, the Poisson function often gives at least a good first
approximation, particularly for short time intervals. Furthermore,
the risk of changes and of variations disturbing stationarity can often
9
RISK THEORY
be dealt with by simply adding a cautionary amount to the constant q.

As will be shown in Appendix A, the condition of stationarity can
sometimes be met by the use of an appropriate non-linear transforma-
tion of the time axis. In some problems even this transformation is
unnecessary. Changes in the stationary flow of events often do not
change the functions, except for the constant q. This arises from the
fact that the sum of two or more Poisson variables is again a Poisson
variable. Thus, even if the process has recurrent seasonal or other
oscillations or changes, which are predictable in advance, the time
axis can be divided into parts (e.g. into months or seasons) and the
Poisson law holds in a satisfactory way for each of the sub-processes.
Then the sum of the whole period can again be Poisson-distributed,
in spite of the fact that the stationarity condition does not strictly
hold. The question is reduced then to the choice of an appropriate
value for the constant q.
At a first glance it would appear that condition (iii), the exclusion
of multiple events, does not always hold. For example, in motor
insurance, two vehicles may collide giving rise to a double event.
Similar incidents can occur in marine and in some other branches.
This difficulty can, however, be circumvented by a suitable choice
of definition, for example by regarding the case of collisions between
two cars as a single claim. This means, however, that the sum of the
claims of both parties is used when building up the statistics of the
distribution of the size of one claim, which is regarded separately as
another random variable, considered in Chapter 3 and later. The
exclusion of an infinite number of claims is no restriction from the
point of view of applications.
2.4. Numerical Calculations

For practical applications it is necessary to compute the Poisson
distribution function. Putting qt (in 2.1) = n = expected number of
claims and writing Pk for Pk(t):
L L~
h h k
F(x) = Pk = e- n (2.2)
k=o k-o k!
where F(x) is the probability that the number of claims g ~ x and

h = [x] is the integer satisfying h ~ x<h+l.
10
The mean and standard deviations ofthis function are given by:
E{s} = .2 k n-k! e-n = n

~
k=O
k
and
respectively, i.e.:
E{S} = n
a= In
If the constant claim amount is S these become:
E{s} = nS
a =SJn
Fairly extensive tables of the function F are available (Tables of
Poisson Distribution, Molina, 1942), but, if n is small, values may be
easily calculated directly from (2.2). For larger values of n, use can
be made of Stirling's approximation:
1 1
k! = J(21Tk) (kje)k (1 + - + - -2- Rk )
12k 288k
where the remainder Rk is of the order Ijk 3 • If n is an integer (which
is no practical restriction if n is large), it may be noted that if Rn
can be neglected:
288n J(nj21T)
Pn = 288n2+24n+1
Having found this term, which is the largest in the series, the terms
on either side can be calculated from the recurrence formula
Pk+1 = npk/(k+ 1).
It may also be noted that F can be expressed in terms of the
incomplete gamma function:
Je-tthdt
00
F(x) =~ (k = [x])
k!n
as can easily be verified.
11
RISK THEORY
TABLE 2.1. Comparison between Poisson distribution (F)

and normal distribution (4)).
n z I-F 1-4>
1 2 0·080 0·159
3 0·019 0·023
4: 0·004 0·001
5 8 0·068 0·090
10 0·014 0·013
12 0·002 0·001
10 15 0·0487 0·0569
18 0·0072 0·0057
21 0·0007 0·0003
50 60 0·0722 0·0787
65 0·0173 0·0170
70 0·0030 0·0023
75 0·0004 0·0002
100 115 0·0632 0·0668
120 0·0227 0·0228
125 0·0068 0·0062
130 0·0017 0·0013
If n is not small, use may be made of the central limit theorem

of probability theory, according to which F tends asymptotically
to the norma.l distribution function when n ~ 00, i.e.:
F(z) ~ ~ ( z;nn ) (2.3)
where ~ denotes a normal distribution function with zero mean and

unit standard deviation. The values given in Table 2.1 above
8how the closeness of this approximation which improves rapidly
with increase in n. In particular, for values of n ~ 100 and
O·05~ I-F~O·OI the relative error is <5%.
2.5. Application 1
Consider an insurance portfolio in which each person is insured for
the same sum, S. An example would be a friendly society operating
on a pure risk premium basis in which the benefits are uniform sums
12
payable on death for funeral expenses, or, if level annual premiums

are charged, a society in its early stages so that reserves are small
and the sums at risk are approximately the same as the sum!!
assured. The question may now be asked as to how large the reserve
U should be so that the probability that the claims exhaust U is
equal to e during a certain time interval, which may be taken as
1 year. To answer this question the following information is needed:
n = expected number of claims

P = nS(1 +,\.) = premium income, where
.\ = safety loading
The conditions of the problem imply:
U = x.-p
where the number x. satisfies the equation:
F(x) = l-e (2.4)
and F(x) is the Poisson distribution function (2.2). For illustrative

purposes the approximation (2.3) will be used in the form:
F(x) ~ (/J (x-ns)

SJn (2.5)
since the size of one claim is S.

Let y. be the solution of the equation (/J( - y) = e. Then it follows
that:
U = nS+y.SJn-P
or
U = y.SJn-.\nS (2.6)
Solutions ofthis equation are shown in graphical form in Fig. 2.1.
Consideration of this figure shows that the reserve requirements
are essentially dependent on the safety loading .\. If A = 0, U is
proportional to the square root of the expected number of claims
n and thus increases with increases in the value of n. If .\ >0 and,
in addition, if:
(2.7)
or
.\ ~ y.!Jn (2.8)
13
RISK THEORY
n
100 200 300 400 500 600
Fig. 2.1. U/S = Ye v'n-,\n. (e = 0·01)
no reserve U is required since the safety loading is more than

enough to cover the fluctuation of claims with probability 1- e.
Even though the assumptions made above are very special, it
will be seen later that the results presented in Fig. 2.1 are also
characteristic of more general cases where the claims are not
necessarily equal. The essential difference being that in the case
of variable claim amounts either a larger reserve or a greater safety
loading is needed.
Exercise 2.5.1. A friendly society has 1000 members. In the event

of death a fixed sum of £500 is paid. The mean value of the rate of
mortality is 0·01 and the safety loading ,\ = 0·1. The actuarial
status of the society is examined every 5 years. How large a security
reserve U should the society have to be sure, at a 99% probability
level, that after a 5 year period the balance does not show any
deficit 1
Exercise 2.5.2. How many members should the society, in the above
exercise, have for no security reserve to be necessary under the
conditions mentioned 1
14
Exercise 2.5.3. The same as Exercise 2.5.1 except that the status
is examined every year instead of every 5 years. Apply both the
Poisson distribution and the normal approximation and compare
the results.
Exercise 2.5.4. A friendly society grants funeral expense benefits

on the death of a member, the benefit being fixed at £100. The
expected number of claims n = 1. The society has a stop loss
reinsurance in accordance with which, if the number of deaths
exceeds two, the reinsurer pays the third and subsequent benefits.
What is the net premium for the reinsurance?
2.6. Application 2
Consider an excess of loss reinsurance arrangement for a life
assurance portfolio or other class of insurance where the amount of
claim under a particular policy can be for a fixed sum only, equality
of risk sums under different policies not being a necessary condition.
It is desired to find a criterion to show whether the reinsurance
arrangement is 'overeffective' (the application of (2.6) to this
problem was first suggested by Hafer!, 1940).
Arrange the policies in increasing order of magnitude of the risk
sum S. The expected number of claims under the ith policy is
denoted by qi and the net retention by M. The reinsurer is assumed to
receive a safety loading A. If now the net retention is increased
by a small quantity -dM then the interest of the reinsurer is reduced
by an amount
(2.9)
where the first term represents the change in the reinsurance
premium, the second the corresponding change in the claims paid
by the reinsurer and
(2.10)
i.e. the expected number of claims of amount at least M. tM is

the actual number of claims of amount at least M in respect of
which the reinsurer has to make a payment under the terms of the
treaty.
Now an occurrence of a claim ~ M is again apparently a
phenomenon where the number of the observed claims tM is a
15
RISK THEORY
~========k~M
Number of policies
Fig. 2.2.
Poisson variable, and therefore what was said in Sections 2.3 and
2.4 can be applied. Hence by (2.6) with probability 1- e,
gM ~ nM+y•.jnM (2.11)
Using (2.9) it follows with the same probability that:
'/.1M ;:;; >..nM-y•.jnM
For a rational reinsurance arrangement the quantity , should

show positive and negative values in different years. If , were
always positive (with probability 1- e), it would imply that the
cedant would be suffering a constant loss without any benefit
in return.
This is the case if:
or
(2.12)
Hence the inequality (2.12) is a criterion as to whether the net

retention M is too low. It will apparently be to the cedant's benefit
to increase the net retention M at least until the mean value of
the number of claims nM reaches the limit (2.12).
16
As an illustration let the mean probability of a claim q be 0'01,

N M be the number of policies with Si ~ M, ,\ = 0·2, and e = 0'01,
then the critical value is:
(2·326)%
nM = 0·01 N M = = 135
0.2 2
or
NM = 13500
If the period of observation is taken as 5 years instead of 1 year,
then N M = 2700.
This analysis shows that a limit exists for the maximum number
of policies which it is rational to reinsure by means of an excess of
loss reinsurance arrangement. This method is not, however, of any
value in finding a useful rule for fixing a maximum retention M.
The equation of criticality (2.12) appears to give unreasonably low
values of M in practice, but the example has been given because
of the theoretical interest in this method of analysing the re-
insurance problem.
17
CHAPTER 3
Generalized Poisson Distribution
3.1. The Distribution Function of the Size of a Claim

To obtain the distribution of the general risk process it is necessary
to have as an auxiliary function the distribution function S(z) of
the size of a claim. The function S(z) gives the probability that, when
a claim occurs in an insurance portfolio, it is of amount ~ z.
The existence of the distribution function S(z) is an axiom in the
theory of risk. In fact, there are two variables in the risk process,
the number of claims v and the size, of the claim, which are
assumed to be mutually independent.
The existence of a function S(z) is in conformity with general
experience, at least as regards time periods of moderate length, and
provided the effect of changes in monetary values is eliminated. The
actual claims can be recorded and numerical estimates obtained for
S(z). At the outset it will only be assumed that the function exists
and that it is known. Subsequently the details of its practical com·
putation will be considered.
As mentioned in Section 1.3 only distributions with positive risk
sums, i.e. z ~ 0, are dealt with in this book.
Three different types of S·functions occur in the applications and
are shown in Fig. 3.1 with the corresponding densities s(z).
In Fig. 3.1(a) S(z) is continuous; this form of S(z) is very common,
because a large portfolio of insurance policies will consist of a wide
variety of different insured amounts, and consequently the claims
will be of all amounts from zero to very large. In non·life business
the continuity becomes more apparent because of partial damages;
this also has the effect of increasing substantially the relative inci·
dence of the smaller claims.
The function of Fig. 3.1(b) is discontinuous and could arise from
a friendly society granting fixed funeral expense benefits or from a
company which has standardized the benefits under its policies.
18
GENERALIZED POISSON DISTRIBUTION
A mixed type of function is shown in Fig. 3.l(c); this can arise

from a basic distribution of type (a), subject to an excess of loss
reinsurance arrangement. In general the step in S(z), corresponding
to the block in s(z), arises from the reinsurance arrangements, which
has the effect of cutting off the top layer of the basic risks. Similar
steps can arise from legal or contractual upper limits of indemnity.
If the arrangement involves different net retentions or limits for
different branches or classes of risks, several steps may be shown
in the S-function.
s(Z) s(Z) s(Z)
$(Z)
(0) (c)
Fig. 3.1(a) A continuous function. (b) A discontinuous function.
(c) Mixed type.
The above treatment has been formal and idealized. In case (a),
the continuity of S and the existence of the derivative s = dSJdz is
assumed. In reality, S is always a step function, since the smallest
unit of money (zo) provides a quantum which, beginning with the
point z = 0, creates a set of intervals (0, zo), (zo, 2zo) ... in each
of which S is a constant. Hence 8 can never exist as a frequency
function in the sense of probability theory. The non-existence arising
from the concept of discrete monetary units is always removed by
smoothing, i.e. by introducing an idealized concept of continuous
money. This same reasoning does not hold where there is an actual
step in the distribution function caused, for example, by a re-
insurance arrangement or by a legal upper limit of indemnity; in
19
RISK THEORY
these circumstances the non-existence of an s-function has essentially

to be accepted.
For practical application it is sufficient to assume that the S-
function is one of the types mentioned above. The theory of risk
can be formally developed however on the basis of concepts of
the distribution theory of Laurent-Schwartz, in which density is
not a function but a measure.
In the following it is assumed that the frequency function 8 =
dS/dz exists except at a finite number of points, where the distribu-
tion function S has steps. This means that the S-function is of mixed
type (c), the continuous and discontinuous types then becoming
special cases.
In the subsequent development, integrals of S(z) will often be
needed. In the general (i.e. mixed) case, moments of S(z) are found
by the formula:
f
00
fXk = zk S'(z)dz+ L Zi,kPi (3.1)

o
where the sum is extended over all the discontinuous step-points

and Pi> P2' ••• Pi ••• are the heights of the steps.
Zl' Z2' ••• , Zi •..
To reach a short and convenient notation this equation will be
generally denoted as follows:
f
00
fXk = zk dS(z) (3.1')

o
Readers who are familiar with the concept of Stieltjes integrals

will realize that the notation of this integration theory is being used
and they can regard the integral as a Stieltjes integral. Readers
who are not familiar with this topic can regard this as an abbrevia-
tion and can always replace any integral by the sum as follows:
f f f
b b
f (z) dS(z) = f (z) : dz + f (Zi) Pi (3.2)

a a
It is easy to see that this extended integral has the same general
20
features as the conventional integral of a continuous function. In

b
most cases the fact that Sf dB is a notation for (3.2) can be ignored
G
and correct general results obtained by taking S(z) as a continuous

function, developing the final formulae in accordance with (3.2).
3.2. Generalized Poisson Function

An insurance portfolio is again considered and it is desired to find the
probability distribution of the total amount of claims (= g) which
occurs during a time interval (e.g. 1 year). The number of claims
is assumed to follow the Poisson distribution (2.2) and the distribu-
tion function S(z) of one claim is assumed to be known.
The required distribution F(x) gives the probability of the event
g ~ x. This event can occur in the following alternative ways:
(i) In the time interval no claim occurs.
(ii) The number of claims = 1 and the amount of the claim is ~ x.
(iii) The number of claims = 2 and the sum of the amounts of
these is ~ x.
(iv) The number of claims = 3 and the sum of the amounts of
these is ~ x.
etc.
The probability that the number of claims equals k is again
denoted by Pk and the conditional probability that, if the number
of claims is exactly k, the sum of these k claims is ~ x, is denoted
by Sk(X). Using the combined addition and multiplication rules of
probability, it follows that:
L Pk Sk(X)
00
F(x) = (3.3)
k=o
If it is assumed that the amounts of the claims are mutually

independent, the function Sk(X) is well known from probability
calculus as the ktk convolution of the distribution function Sex),
which can be calculated from the recurrence formula:
I Sk_l (x-z) dS(z)

x
Sk(X) = = S<k-1l* * S = Sk*(x) (3.4)
o
21
RISK THEORY
Replacing Pk in (3.3) by the Poisson function (cf. Section 2.4), the

following important formula is obtained:
00
e-nnk
F(x) = ~ _-Sk*(x)
k!
(3.5)
k=O
This is one of the basic functions of risk theory. The name of this
function is not, as yet, well established. In this book it will be called
the generalized Poisson function in contrast to the (ordinary) Poisson
function considered in Chapter 2 where all claims have the same
unit value. Some authors refer to the latter as the elementary
Poisson function and the function (3.5) as the non-elementary
Poisson function, using the term "generalized" for processes where
the assumptions of Section 2.2 are replaced by some weaker ones,
thus introducing more general categories of processes.
The generalized Poisson function (3.5) is subject to assumptions
(i) to (iii)* in Section 2.2 as well as to the assumption that the
amounts of the claims are mutually independent.
(3.5) is, unfortunately, directly useful for numerical computation
only by making very special assumptions concerning the function
S(z) or if n is a very small number. The main difficulty of the theory
is to find approximations or methods of computation for this
formula suitable for practical applications. Several such methods
are given later, but first some general features of F(x) are con-
sidered.
3.3. The Mean and Standard Deviation ofF(x)

The calculation of the mean of the distribution of g defined in (3.5)
follows easily by noting that the mean of the distribution Sk*(x) is
km, where m denotes the mean of Sex):
f zdS(z)
00
m= (3.6)
o
* As mentioned earlier, the first part of the postulate (iii) is not

necessary if a multiple claim is considered as a single claim, the amount
of which is defined as the sum of the amounts of the component claims.
22
and Sk*(x) is the~istribution of the sum + + ... +'k, each

member of this sum being a random variable with mean m. Then
'1 '2
the mean of F(x) is = nm, which can be verified as follows:
I xdF(x) = .2 ~ I xdSk*(x) = 2 ~ km = nm
OCJ OCJ e-nnk OCJ OCJ e-nnk
EW =
o k=O 0 k=O
Let <X2 be the second moment of S(x), i.e.:
I
OCJ
<X2 = Z2 dS(z) (3.7)

o
Then the variance of S(x) is <X2-m2; since it is assumed that the

sizes of claims are independent, the formula for the second moment
of Sk*(x) follows as a consequence of the well-known properties of
the sum of k independent random variables.
I x 2dSk*(x) = k<X2-km2+k2m2 = k<X2+k(k-l)m2

OCJ
Hence
Using the above value of E{g}, the value of the standard deviation
of F(x), denoted by u, follows from the equation:
I (x-nm)2 dF(x)
OCJ
u2 = = nOC2
o
Thus the values of the mean and standard deviations of the
generalized Poisson function F(x) are given by:
E{g} = nm (3.8)
u = .j(n<X2) (3.9)
23
RISK THEORY
Exercise 3.3.1. A friendly society grants funeral expense benefits

which may be £100 or £200 according to the choice of each member.
The sum £100 is chosen by two-thirds of the members, the re-
mainder choosing £200. The number of members is 100 and it is
assumed that the mean death rate for each member is 0·01. Compute
the distribution function F(x) of the annual amount of the claims.
Observe the step character of F.
Exercise 3.3.2. Compute E{~} and a for the society mentioned in the
previous exercise.
3.4. Characteristic Function

In order to make this book appeal to as wide a group of readers as
possible, mathematical aspects involving integrals of complex
variables have been avoided, although their inclusion would have
considerably simplified the development. However, for readers who
are familiar with this topic the characteristic function of the
generalized Poisson distribution is now given. This frequently used
function of probability calculus is needed as an auxiliary function
for many further developments of the theory referred to in the
bibliography at the end of this book, and use will also be made of
it later for calculations which otherwise would be too cumbersome.
The characteristic function of the distribution function F(x) is
defined by the complex integral:
f
+00
I{>(s) = eiSXdF(x) (3.10)

_00
where s is an auxiliary real variable.
The characteristic function of the distribution function S(z) of the
Je
+00
size of one claim is denoted by '" = t3XdS(x). From a theorem,

-00
which states that the characteristic function of two independent

variables is the product of the corresponding characteristic functions,
it follows that the characteristic function of the kth convolution
24
G1ilNERALIZED POISSON DISTRIBUTION
Sk*\x) of S(x) is if;k. The characteristic function of the generalized

Poisson function (3.5) is now found to be:
cp(s) =
00
' " e-nn

~ k!
k
I
+00 00 k
eUXdSk*(x) = ' " e-nn if;k
~ k!
k=O _00 k=o
= e- n ~ (nif;)k = eno/J-n
~ k!
k=O
and thus
1>(8) = eno/J-n (3.11)
In passing, it may also be noted that a similar formula to (3.11)

holds if characteristic functions are replaced by the Laplace trans-
forms:
00
Ie-Ix dF(x)
o
since S(x) vanishes identically (cf. Section 1.3) for x < O. Since
extensive tables are available in a number of standard mathematical
texts, the Laplace transforms may be useful in practice and can
be used without fear of complications arising from considerations
of convergency.
3.5. Estimation of S(z)

3.5.1. Individual Method
In most applIcations of the theol'y ofrisk it is necessary to know the
distribution function S(z) more or less accurately. It should be so
fitted that the representation corresponds as closely as possible to
the true distribution of the amount of one claim in the portfolio.
The fit should align itself to the data which is, in general, empirical.
Some methods of estimating the S-function from the data available
are now presented.
First a method is given for computing the S-function starting from
the individual policies of an insurance portfolio. The risk units
(policies, cf. Section 2.3) are numbered 1, 2, 3 ... , N and the
25
RISK THEORY
corresponding frequency rates* are denoted by Ql' Q2' ••• QN. It is

further assumed that each risk unit has a distribution function
St{z) giving the distribution of the amount of the claim if it occurs
for this risk unit. For life assurance and other classes of insurance
where no partial claims occur this function is simple to obtain. In
this case the function can be conveniently represented as a formula
making use of the degenerate distribution function e:
e{x) = {I when x~O (3.12)

o " x<O
If the sum payable at death (or other event producing a claim)
is Zt, then the function in question is:
For insurance classes where partial claims are possible, the

function St is not easily obtained and the individual method is not
very suitable in these cases. It can, however, be used even for these
classes when no other methods are applicable. This will be discussed
in Section 3.5.3. Here it is supposed that the distribution functions
St are known for each policy.
The distribution function of a claim arising from the whole
portfolio can be found if the risk system is interpreted as an urn
experiment. The different risks are visualized as different urns and
a selection is made of an urn. The probability that the one selected
is the ith is:
Since Bt(z) is the conditional probability that the claim is ~ z the

addition and multiplication rules of probability give immediately:
(3.13)
* The term probability of claim is often used in this connection, but

ql must, in fact, be regarded as a frequency or, what is the same, the
expected number of events (which might even be ~ 1). If the number
of claims is distributed in a Poisson form during a certain interval, and
the parameter q is very small, the probability of occurrence of at least
one event is clearly p = 1 -e-'l ""'q. In this sense, reference is sornetiInes
made in a rather loose way to the probability of an event, when, in
fact, the expected number of events during this interval is meant.
26
As an illustration consider the distribution function of a life

assurance portfolio. For this purpose all policies are divided according
to the size of the risk sum into classes 1, 2, ... i, ... The rate of
death for the policies belonging to ith class are added giving L qvi;
v
this sum, and the direct application of (3.13) gives:
(3.14)
The index v runs over policies of class i; the first sum in the numerator
runs over classes where the risk sum Zt is less or equal to Z and in
the denominator over all classes. In other words: total the expected
number of deaths for all policies having a risk sum ~ Z and compare
this with the expected number of deaths for the whole portfolio.
Exercise 3.5.1.1. A company grants insurance for accidental death
the sums payable at death being standardized at £lOO, £250, or
£500. The number of policies in these classes are 5000, lOOO, and
2000, respectively. It is known that the rate of death in the two
lower classes can be expected to be equal, but that, owing to anti-
selection, the rate in the £500 class is estimated to be double that
in the other classes. What is the S(z)-function for this business?
3.5.2. Statistical Method
In this method the actual claims of the portfolio in question are
collected in a table according to the amounts of the claims, as in
Table 3.1, which sets out claims arising from a combined ex-
perience of Finnish insurance portfolios comprising industrial
fire risks. The reason that the n-column includes decimals is that
claims exceeding £lO 000 are taken from several years experience
and then reduced to correspond to one year's frequency (see
Section 3.5.3).
If the claims are collected over a period during which monetary
values change, it is advisable to eliminate the influences of these
changes by multiplying the earlier claims by the ratio of the
corresponding values of a suitably chosen price index.
Because the structure of the portfolio and many other circum-
stances are always changing, even if slowly, the observation period
should not be very long. On the other hand very short observation
27
TAB L E 3.1. Compilation of the claims statistics.
n,
£ z,
Number of
claims in class
JS = ni/n
x 10'
S = ~JS,
x 10'
Zi_l<Z ~ Z/
1·0 10 3514 3514

1·6 10 3514 7028
2·5 16 5622 12650
4·0 71 24947 37597
6·3 87 30569 68166
10 122 42867 111033

16 161 56571 167604
25 220 77 301 244905
40 283 99438 344343
63 304 106816 451159
100 291 102248 553407

159 243 85383 638790
251 214 75193 713983
398 188 66058 780041
631 166 58327 838368
1000 103 36191 874559
1585 80 28110 902669
2512 70 24596 927265
3981 54 18974 946239
6310 43 15109 961348
10000 31·4 11 033 972381

15850 27·6 9698 982079
25120 15·0 5271 987350
39810 13·2 4638 991988
63100 8·3 2917 994905
100000 10·6 3725 998630
158500 2·6 914 999544
251200 0'7 246 999790
398100 0·3 105 999895
631000 0·0 0 999895
1000000 0·1 35 999930

1413000 0·2 70 1000000
n=2846·0 1000000
periods do not include enough large claims to be representative.

The task of the actuary is to weigh these different aspects and to
try and find for each case the most appropriate method of proceeding.
The problem arising from the large claims can often be dealt with
as described in the next section.
A problem associated with the statistical method is the selec-
tion of the class interval. In table 3.1 a geometrical interval
is used for the claim amounts, a method found convenient in many
cases. A few trials with different class intervals on numerical data
will soon show that a rough partitioning is sufficient for many
purposes. In fact as will be shown later the risk process is not very
sensitive for changes of the S-function except as regards the tail
arising from large claims.
3.5.3. Problem.s Arising from Large Claims

In many practical problems, where the top risks are not cut away
by reinsurance, the larger values of z in S(z) are of critical importance
and the derivation of the distribution function in the region of these
larger values is difficult to determine with confidence from observed
data. Paradoxically the least-known part of S(z) has in these cases
the greatest effect on the numerical results. One possibility is to
determine values of S(z) for small values of z (thus based on the
greatest number of claims) from an experience extending over a
short period only. Since such an experience will, in general, not
include many large claims, a further study of large claims over
a longer period is needed. Thus one year's statistics might suffice
for claims ~ £10 000 whilst for claims > £10 000 data for perhaps
20 years may be necessary. In this case the higher claims should
be adjusted by weights consistent with the relative amount of
business. How far this method can be used depends on how much
time elapses before the data is so changed in structure that it
cannot be regarded as reliable. The truncated distribution of small
claims may be useful as a control in testing the significance of
such a structural alteration.
Sometimes statistics relating to the larger claims are unsatisfactory
because of relatively rapid changes in the risk structure of the
portfolio or simply because they are not available. In such cases
the following rough method may be of use, making use of (3.13).
The largest risks of the portfolio are dealt with individually and
29
RISK THEORY
the net premium for each is determined; next the expected average
extent of damage is estimated for each policy and the results
tabulated as follows:
Net Damage class in £ 1 000

Policy Sum premiums
no. insured (£) 0100 10-50 50-100 100-200
XXX 40000 1 0·2

YYY 100000 2 0·4 0·3
ZZZ 80000 1·5 0·3 0·2
etc.
The placing of the estimated frequencies of partial damages in

the different damage classes would be done by an appropriately
experienced claims specialist. The method is clearly quite subjective,
but in the absence of other methods it does provide some basis for
further calculation. For life assurance the method is easier to use,
because of the absence of partial damages. The method involves a
rough idealization, since for example the risk premium is used as
a measure of individual risk, whereas in practice the basis of a risk
premium involves an equalization over some groups of policies.
If the portfolio is large, so that there are many cases over the
limit (£10000), suitably selected samples for the various risk sums
may be taken and only the largest cases treated individually.
If some information is available relating to large claims it is
sometimes possible to introduce a 'shadow claim', which, having
regard to the actual portfolio, can be considered realistic, although
very seldom occurring. The frequency of this shadow claim can be
assumed, for example, to be one claim in lO, 20, 30, or 40 years in
the whole portfolio (cf. (3.26) concerning the technique).
3.5.4. The Dependence of the S-function on Reinsurance

It should be noted that if the problem concerns the net retained
liability, the amount of claim z is, of course, only that part of the
total claim Ztot which is retained, i.e. the total claim reduced by the
share Zre taken by the reinsurer, i.e.;
(3.15)
30
In this formula the reinsurance arrangement is such that each

claim is separately divided between cedant and reinsurer, i.e. in
quota share, surplus or excess of loss reinsurance but not in stop
loss reinsurance.
As regards an excess of loss treaty, if the global distribution
function of the amount of one claim is S(z) and the retained amount
by the cedant is M, then the distribution function of the amount
of one claim so far as the cedant is concerned is:
S (z) = {S(Z) for z < M (3.16)

M I for z ~ M
From (3.1) the moments of SM are given by:
f zkdS(z)+Mk(I-S(M))
M
rt.k = (3.17)
o
Where a surplus reinsurance arrangement is in use, the allocation

of the sum insured (Q) between the cedant (M) and the reinsurer
(Q - M) is fixed in advance and in the event of a claim:
M
z= Q Ztot (for Q>M) (3.18)
= Ztot (for Q~M)
so that the cedant's own liability S M(Z) can be plotted on a double

logarithmic graph as in Fig. 3.2.
From this graph the value of S M(Zo) for any Zo is the quotient
of the number of points in the ruled area and the total number of
points.
In practice in an insurance company, different net retentions are
often used for different branches and for different types of risks.
A distribution function S M(Z) can then be constructed for each of
these branches and groups, and a joint distribution function - perhaps
also having several steps corresponding to different retentions - is
obtained as will be shown later in Section 3.6.
Another elementary, but often laborious, method is to record the
claims statistics as in the following table. The company's share of
the actual claims is computed for different reinsurance arrangements
and set out in separate columns. The arrangements can vary as
31
RISK THEORY
regards maximum net retention as well as varying in other features.

The functions S,(zo) are obtained for each reinsurance arrangement
by summing up from each column the appropriate cases, i.e. z ~ Zoo
Sum Claim on the company's own retention

Claim insured Total
number £ claim Reins. I Reins. II
1 1000 624 624 511

2 2500 1810 1420 908
3 1500 1500 1200 820
Exercise 3.5.4.1. A company having the distribution function S(z)

for the total size of claims, has two reinsurance treaties in force:
(i) a quota share treaty, under which the reinsurer pays a proportion
p of each claim and (ii) an excess ofloss treaty covering the retained
Z
1000000
100000
10000
100
Fig. 3.2.
Derivation of SM(Z) from a portfolio reinsured by a surplus arrange-
ment. Each point represents one claim from the total claims recorded.
Each point is plotted from the coordinates Q and Ztot (owing to the
partial claims z is often < Q). Double logarithmic scale.
32
business with a maximum net retention M. What is the distribution

function of the size of one claim for the company's own retention?
Exercise 3.5.4.2. Readers having access to some actual claims
statistics are recommended to make up a function S(z) based on
them. Special attention should be paid to the large claims. The
function S M(Z) should then be constructed following some of the
reinsurance principles mentioned above or, if possible, the actual
reinsurance practice of the company. The function SM(Z) can be
used subsequently in different exercises.
Readers not having access to any actual statistics can make use
of the data of Table 3.1. To facilitate the numerical computations
the interval z can be longer than that in the table, for example by
taking every third z.
3.5.5. Analytical Methods

It is often very desirable to try and find an explicit analytical
representation for a claim curve. If a well-fitting curve compared
with the observed data can be found, which at the same time is of
such a form that the convolutions Sk* can be carried out in a closed
form then an explicit expression for the function F(x) (3.5) can
sometimes be found and thus the approximate calculation of F(x)
avoided. Particularly when the portfolio concerned is not very
large, i.e. when the variable n is small, approximations cannot
always be found that give a good fit for F(x) and the need for
exact computational methods becomes apparent. An analytical
expression for S(z) can, of course, be of considerable value to the
actuary in other connections, e.g. tariff calculations, statistical
analysis, etc., and if use can be made of some well-known elementary
functions, such as an exponential function for S(z), the known
properties of the function can be used to enable some insight to
be gained into the characteristic features of the claim distribution.
On the other hand it must be accepted that replacing the actual
data by an analytical expression always implies smoothing. The
goodness of fit of S(z) can easily be estimated by various well-known
methods, but it is often of much greater importance to study the
error introduced in the function F(x). This is a drawback of the
analytical method. However, the fact that the amount of the claim
, is a random variable and, consequently, functions derived from
33
RISK THEORY
the claim statistics are also subject to random fluctuations must

also be borne in mind. If the model found for S(z) is such that the
deviations from the observed data are of the same or a lower order
of magnitude than the random fluctuations, then it may be assumed
that the distortion due to smoothing is imbedded in the inaccuracy
due to the random fluctuation, which is always present when
deducing the function S(z). To ascertain the magnitude of the error
caused by this phenomenon, different functions S(z) can be experi-
mented with so that they approximate the available data. In practice
these fluctuations and the influence of the smoothing are, however,
often ignored and the answer as to how good the results are remains
open to some extent. Fortunately, experience shows that F(x) is
not very sensitive to changes in S(z) in those cases where the tail
is truncated by means of reinsurance, as it normally is in practice.
In the following sections, consideration is given to some frequently
used analytical models, some adopted because of their convenience
in the calculation of (3.5) and some for other reasons.
3.5.6. Exponential Function

In general, claims distributions show the highest frequency for the
small claims, the frequency declining with increasing claim size.
Thus the exponential function will provide at least a first approxi-
mation for dSJdz. Let:
S(z) = 1- e- az (for z ~ 0)
The constant a can be fixed for each application to obtain the
best possible fit. This expression has the advantage that an explicit
expression for the generalized Poisson function F(x) can be found
by direct calculation using (3.4) and (3.5). The calculations are
left to the reader and only the result is given:
0 for x<O
F(x) = {
f
az (3.19)
e- n + g(y)dy for x ~ 0
where
co k
g(y) = n e-(n+y) "'" (ny) (3.20)
~
o
k!(k+l)!
34
If .j(ny) is very small the series converges rapidly and direct

calculation is easily performed. If .j(nax) is also small, even the
double series obtained by integration is very convenient for direct
calculation. But as soon as .j(ny) becomes large, direct calculation
rapidly becomes hopelessly laborious. Fortunately it is then possible
to use the asymptotic properties of Bessel functions, to the modified
type of which g(y) is closely related. Even for relatively small
values of .j(ny) the following asymptotic expansion, which is given
without proof, gives very accurate values for g(y):
g(y) = Jy
n e-(vy-vn)2
2.j1T~(ny)
[3 15
1- 16(.j(ny ))-1- 512 (.j(ny))-2
105 4725 72765 ]

- SI92(.j(ny ))-3- 524288(.j(ny ))-C 8388608(.j(ny ))-S
+ Remainder (3.21)
If the semiconvergent series, the first terms of which are given

co
in brackets, is denoted by L Ak (.j(ny))-k, then the coefficients
o
Ak are obtained from
From this expression, g(y) is obtained correct to eight digits for

.j(ny) = 10 and more accurately for larger values. Even for a
value of .j(ny) as small as 3, the expression gives five correct first
digits if two more terms are included. For smaller values this
formula is of no use, owing to the semiconvergent character of the
series. However, as mentioned earlier, for small values of .j(ny)
it is possible to calculate g(y) directly from the series (3.20).
(It should be noted that the remainder contains terms of types
other than those shown in the brackets.)
The use of an exponential function as a model for S(z) can clearly
only be occasionally useful since this simple function can only be
a crude approximation to the truth and is hardly ever applicable
if reinsurance cuts off the top risks. However, it has some practical
importance in connection with a more general method discussed
in Chapter S.
35
RISK THEORY
3.5.7. A Generalization of the Exponential Type

It is easy to verify that a frequency function of the form:
dS(z) a b +1
- - = _ _-zbe-az for z ~ 0 (b > -1) (3.22)
dz r(b+1)
can still be handled by forming convolutions, giving rise to functions
of the form:
g(y) - nyb e-(n+lI)

-
2: -=r-(
00
0
(n b+l)k
y
(k~+-1-)(-b-+-1)-)(~k+-1-)!
corresponding to (3.20).
Using the lemma of Section 3.6 a more general polynomial can
be constructed where each term is of the type (3.22). There are
good reasons to expect that by using these mixed polynomials,
the number of terms required will, in general, be fewer than when
using pure exponentials. However, numerical calculations become
more complicated.
a(z)
3 b=50
3 4 Z
ab +1
Fig. 3.3. 8(Z) = F(b + 1) zbe-a,. In the figure a = b + 1 (and conse·
quently the mean = 1) and b has the values 0, 1, 4, 9 and 50.
36
3.5.8. Other Types of Distribution

Many of the actual distributions of claims that arise in insurance
applications can be reasonably well approximated by the Pareto
distribution :
dS(z) lX-I
--=-- l<z<oo
dz zl1.
Unfortunately the mathematical properties of this distribution do

not lead to simple expressions for convolutions, particularly if the
curve is truncated at an upper point in order that the higher
moments exist (if lX>3 the first and second moments always exist).
Benktander and Segerdahl (1960) have considered some special
claim distributions in their study of excess of loss reinsurance.
The logarithmic normal distribution:
dS(z)
-
dz
= J1
Zu (21T)
exp {I - - (log Z-/L)2
2u 2
}
has also been shown to represent many actual distributions and,
although the moments exist, it is not possible to express the con-
volutions in a simple form so that normal analytical development
is precluded.
Looked at in terms of the Pearson system, these distributions fall
between types III and V into the region of the type VI distribution
Le.:
dS(z)
- - = k(z-a) lloz- Ql
dz
where
k = aQl-Q.-l r(ql)
r(ql-q2-1)F(q2+ 1)
but here again analytical development is precluded.
3.6. Decomposition of S(z)

From earlier computations the distribution functions Fk of different
parts of a portfolio may be known or it may be easy to compute
them separately for each of these parts, whereas the joint distribu-
tion can be troublesome to handle directly. In these circumstances
the distribution function of the whole portfolio can be obtained
via these components as follows. If the parts in question are mutually
37
RISK THEORY
independent, the joint distribution function F(x) (3.5) of the

portfolio can simply - by definition - be obtained from the functions
Fk(x) by convolution:
(3.23)
since the total amount of claims x is the sum of the corresponding
partial amounts:
x = X1+X 2+ ... +x"
The parameter n of (3.5) obeys the corresponding rule:
n = n 1+n 2+ ... +n"
It may be noted that (3.23) holds irrespective as to whether the
F-functions are generalized Poisson functions or any other functions.
(3.23) can also be derived in another way by means of the
characteristic functions (3.10). Consider a linear combination:
S(z) = "
L OkSk(Z) (0 1+02+ ... +0" = 1) (3.24)
k=l
of II, distribution functions Sk(Z) and suppose S(z) is again a distri-
bution function. Let nk = nCk. Denote by tP, tP1' tP2' ... , tP" the
characteristic functions of the distribution functions S, Sl ... , S".
Then according to (3.11) the characteristic function 4> of the corres-
ponding generalized Poisson distribution function F(x) is:
4>(8) = exp (ntP-n) = exp (nLok,h-n)
= n" exp (ncktPk-nk) = k=ln" 4>k(S)

k=l
(3.25)
where 4>k(8) is the characteristic function corresponding to the kth

part of the portfolio. A well-known theorem of probability calculus
states that if the characteristic function of a distribution is a
product of characteristic functions of other distributions then the
former distribution is a convolution of the latter; this proves (3.23).
The result arrived at by means of characteristic functions is,
however, in a certain sense more general. Whilst S(z) must, of course,
be a distribution function of the size of one claim in insurance
applications, there is no reason to restrict consideration to distribu-
tion functions in so far as component functions Sk are concerned.
38
On the contrary - in order to facilitate computations - it is often

advisable to consider S(z) as made up of components which cannot
be interpreted as distribution functions of one claim of any actual
part of the portfolio. In this wider case the functions Fk are not,
in general, distribution functions, but this feature does not have
any essential influence on the calculations.
For example, suppose there are difficulties in calculating F(x)
directly from (3.5), but that it is possible to select a set of h functions
Sk(Z), for which the corresponding Fk can be calculated in practice
and these functions satisfy (3.23). Then, to get F(x), it is only
necessary to calculate h convolutions, an essentially easier course
than the calculation of a nearly unlimited number of convolutions
which is the case when (3.5) is used directly. Thus the expression
(3.23) could be taken as the first h terms of the expansion of S(z)
in a series of some auxiliary functions Sk(Z), this series having
convergence properties so that the remainder may be neglected.
As an illustration of the direct approach, a case where F(x) is
computed for an S-function may be considered. Suppose that, when
deducing the latter, it was difficult to test the role of very large
claims. To estimate the influence of these, an imagined 'shadow'
claim of catastrophic size, say X, of frequency q (e.g. for a
catastrophe once in 10 years, q = 1/10) is added to the claim
statistics. Then the modified distribution of the total amount of
claims is derived by use of (3.23) as follows (where q = expected
number of 'shadow' claims):
(i) If only one 'shadow' claim is possible (binomial distribution):
F{x) = (1-q)F(x)+qF(x-X),
(ii) If 'shadow' claims are Poisson-distributed:

00
qk
F(x) = e-q
Lo
k! F(x-kX) = e-qF(x)+qe-qF(x-X)
It is essential to notice that, in case (i), the proof given above

using characteristic functions does not hold, the result being a
direct consequence of the probabilistic definition of convolutions.
39
RISK THEORY
The change in the ruin probability e = 1- F(x) in case (i) is:
Ae = q[F(x)-F(x-X)] (3.26)
If this equation is applied to the example used later in Fig. 4.2,

it is found for q = 0·1, n = 10000, M = £83700, X ~ M, and
e = 0·001 that the value of Ae ~ -q(dF/dx). Ax is 0·00005*.
It is thus seen that, when the reinsurance arrangements cut away
all large claims, the 'tail' of the S-function (i.e. the probabilities
of very large claims) is apparently not of very great importance, a
very natural result. But if the maximum net retention is very
high, then the tail of the S-function can be more significant. For
example, in Fig. 4.2 for n = 500, M = 00, q = 0'1, X = £1 000000,
and e = 0'001, the change Ae would be 0·0014. To be on the safe
side in these circumstances it is advisable to add to the claims
statistics some large 'shadow' claims if it is known that such are
possible, even though the statistics from which S(z) is deduced do
not happen to include any.
It will be noticed that the moments of the total business are
obtained from those of the components as follows:
(3.27)
(3.28)
Exercise 3.6.1. Suppose that owing to the conflagration risk it can

be considered possible that events exist such that a claim can occur
which exceeds the estimated maximum net retention M, e.g. of
size 3M (assuming such an event occurs, say, once in 10 years).
Estimate the influence of this risk on e, if the distribution of Fig.
4.1 applies and M = £34500, n = 500, and e = 0·001. The mean
value m corresponding to this value of M is £1 203. Other necessary
data are to be read approximately from Fig. 4.2, Me-curve.
*The value of dFJdx has been estimated from Fig. 4.2. For this purpose
the mean values of claims m(M), is required and in the actual case con-
sidered this was £1 450 for M = £83 700 and £1 732 for M = 00.
40
CHAPTER 4
Normal Approximation and

Edgeworth Series for F (x)
4.1. The Normal Approximation

It will be apparent from the foregoing chapters that the generalized
Poisson function F(x) which gives the distribution of the annual
result of an insurance operation is, unfortunately, complicated as
regards computation particularly in practical applications. Direct
methods of attack on the numerical treatment of F(x) often lead
to very cumbersome expressions so that it is, in general, not easy
to deal with problems concerned with, for example, different
methods of reinsurance, net retentions, and safety loadings. Further-
more, it is extremely difficult to obtain a broad survey of the
problems. Even if the nature of the problems justifies the more
detailed computations, simple working approximations are necessary
and it follows that one of the problems of applied risk theory is the
finding of proper approximations.
The classical approximation to F(x), and one which has most
frequently been applied, is found by applying the central limit
theorem of probability calculus. This theorem says that F(x) tends
to the normal distribution when n tends to infinity, i.e.:
F(x) ~ q, (x-P)
fY..jn
(4.1)
where
fo zdS(z)
<Xl
P = mn = n
(4.2)
f z dS(z)
<Xl
fY.2 = fY.2 = 2
o
the integrals being defined as in Sections 3.1 and 3.5.4.
41
RISK THEORY
In Chapter 5 and later it will be shown that the normal approxi-

mation considerably simplifies the calculations and makes it possible
to provide a broad survey of problems involving many variables
and basic functions in a way which is not otherwise possible or can
only be done with considerable difficulties. Unfortunately, however,
the mapping of the errors of the approximation is still to a large
extent an open question. This question will be discussed in Section
4.4 together with cases where the normal approximation does not
give any acceptable accuracy. If however the reinsurance arrange-
ments cut the top risk away from the portfolio being considered
and if the number of risks is not very small, the normal approxi-
mation seems to be applicable for many practical purposes.
4.2. Edgeworth Series

The normal approximation (4.1) is in fact only a special case of a
more general formula, which is known as an Edgeworth expansion:
F(x) = cp (x-P)
--
oc.jn
- -OCs- cp(S)
6n t oc s
(x-P) + __ (x-P)
__
oc.jn
oc,
24noc'
cp(4) __
oc.jn
+ ~CP(6)
72noc 6
(X-P)
rx.jn
+ O(n- 3/ 2) (4.3)
Here cp(k) denote the derivatives of cP and OCk the moments of S(z)
(3.1).
The Edgeworth expansion is most simply obtained by means of
the characteristic function of F, expanding the exponential in a
MacLaurin series and reverting back to the distribution functions
after integration, making use of the correspondence of the charac-
teristic function and the distribution function. Details of the
derivation of the formula are given in Appendix B.
Reference to (4.1) shows that the normal approximation is merely
the form given by the Edgeworth expansion when the first term
only is retained, i.e. by ignoring terms of O(I/.jn).
(4.3), like (4.1), is an asymptotically correct expression for F(x)
and thus in certain conditions it is possible to take more terms
than in (4.3). It is necessary, however, to pay attention to the
semi convergent nature of (4.3) in that it is not possible merely by
increasing the number of terms to obtain an arbitrarily close
approximation for a fixed n.
42
NORMAL APPROXIMATION AND EDGEWORTH SERIES
4.3. Normal Power Expansion

If the explicit expressions of higher derivatives of the normal
function cP are introduced into (4.3), it can be shown that the
relative error of the Edgeworth expansion tends to infinity as
x _ 00. The Edgeworth expansion is not a convergent but a
divergent series. However by taking a suitable number of terms it
gives acceptable results in the neighbourhood of the mean value.
It can be generally expected that the result is good up to a distance
of twice the standard deviation from the mean, but for points
outside this interval the result soon deteriorates. From the point
of view of risk theory this is unfortunate since in most problems
the main interest arises from points at a distance of two to three
times the standard deviation to the right of the mean. For this
reason some improvement on this series is needed. The best - in the
sense of giving the closest results - is the so-called Esscher method
which will be presented in Chapter 6. There is, however, another
method which leads to essentially simpler calculations than the
Esscher method and which gives an error only slightly greater than
the Esscher method.
The normal approximation operated, as represented in Section
4.1, by means of a normalized random variable 'YJ having zero
mean and unit standard deviation. The total amount of claims ~ is
replaced by the random variable mn+'YJoc.Jn, which leads to the
basic equation (4.1). The idea is now to amend 7J by means of an
expansion:
(4.4)
in which the coefficients a o' aI' a 2 , ••• will be determined by means

of the Edgeworth expansion as follows.
Whereas the normal approximation was established on the basis
of the equation:
l-e = CP(y) (4.5)
giving the probability 'YJ ~ y, the amended variable' will obey the
equation (cf. (4.3)):
l-e = CP(y+.dy)- lYI cp(3) (y+.dy)+fir2 cp(4) (y+.dy)

(4.6)
43
RISK THEORY
giving the probability that' ~ y + Ay. Here:

Yl = (X:~ (skewness of g)
(X:3yn
{ (4.7)
(X: ,
Y2= -
a:'n
The coefficients in (4.4) are now determined by substituting in
(4.6):
(4.8)
and then equating the right-hand sides of (4.5) and (4.6). As the
Edgeworth expansion is semiconvergent, a new expansion (4.8) is
now obtained which can also be expected to be semiconvergent,
because the operation in question means in fact a reversion of the
former expansion. The terms having higher powers than the second
in the quantity l/.Jn are omitted in the following.
To find the coefficients av, Ay is to be determined from:
f (Ay) = <P(y) - (the four first terms on right-hand side of (4.6)) = O.
This can be done by using Newton's method in accordance with
the expansion:
_ f(x) 1f"(x) [f(X)]2
X = X - 1'(x) - "2 1'(x) l'(x) - ....
which gives the solution of the equation f (x) = 0, where x is some

approximate solution. Let x = Ay and x = O. Mter some straight-
forward calculations the method gives:
A irl(y2_1)+-hr2(y3_3Y)+71.2YI2(y5_lOy3+ 15y)
4.Jy = I +
[t
1 +-6YI(y3_3y)+ O(n-I)
y+O(n- 1/2 ) YI (y2-1) + O(n-I) J2 0 n- 3/2
11 + O(n-1/2) 1 + O(n- 1/2 ) + ( )
= trl(y2_1)+fiY2(y3_3y) -l6Y12(2y 3_5y)+ O(n- 3/2 )
or, writing the terms involving different powers of l/.Jn in successive

rows:
x-p
cx.Jn = y+Ay = y
+lrl(y2_1)
+ "fiY2(Y s - 3y) -..ftr12(2y S-5y)
+O(n- 3/2 ) (4.9)
44
(4.9) is not much more difficult to handle numerically than the

normal approximation, which is again a special case, i.e. the first
row, of this expansion. If for example (4.9) is used in the form:
(4.10)
and 1- F(x) = e is. given, the point y is found from the normal
function tables using the equation 1- e = (/l(y), and then x is
simply calculated from (4.10). If, on the other hand, x is given,
then y is found from (4.10) and the corresponding e is picked up
from tables:
F(x) = (/l [J( 9 6(x - P) )

-2+1+.j
'Yl
3 ]
--
'Yl oc n 'Yl
(4.11)
This method will briefly be called the NP-method (= normal

power method) or NP-approximation. The formula is, of course,
useful only when the quantity under the root is non-negative.
Since the proof is based on the assumption that 1- e can be
represented by a finite number of the leading terms of the Edgeworth
expansion, the result would not be expected to be better than that
obtained by the direct use of the expansion (4.3). This argument,
however, proves, quite surprisingly, to be wrong. Experience of
the application of (4.10) by several Finnish writers (Kauppi and
Ojantakanen, 1966) shows that it nearly always (more precisely,
if 1'1 is not very great, say more than 2) gives fairly good results,
whereas the Edgeworth expansion is usually unsatisfactory for
deviations some distance from the mean. Examples of the com-
parisons are given in Fig. 4.1. At first sight this seems to be
unsound, because it would not be expected that an inaccurate
formula. would be improved by making further approximations
when inverting it. But it is not necessarily so. It is incorrect to
say that the Edgeworth expansion is not very good, even if it is
only applicable in certain conditions. It is in fact an exact expansion
which, however, is not a convergent but a divergent series. The
same argument holds true for expansion (4.9), and it is no miracle
if the leading terms of one expansion give more accurate approx-
imations than those of another expansion. This is in fact a well-
known feature also for convergent series.
45
E
(&)
·002
·001
2 3 4 5
(b)
:~
:008
·007
·0
·00
·00
·003
·002
Y
·001
2 4 5
Fig. 4.1. See foot of next page

In Table 4.1 further examples are presented for very different

parameter values so as to provide a comparison of this technique
with other methods.
4.4 The Accuracy of the Normal Approximation

As mentioned above, the mapping of the accuracy of the normal
approximation and of the normal power expansion are still open
questions. There exists, however, a number of numerical investiga-
tions where the generalized Poisson function is computed by more
complicated methods, whereby the accuracy can be controlled, and
on which these approximations can be used. Comparison of the
values obtained by both methods can give some illustration of the
suitability of these functions for practical purposes. Some results
are given here, although a detailed consideration of these other
methods will follow in Chapters 6 to 8. At this stage the reader
can accept the numbers and curves obtained by the methods
referred to as 'exact' values and their deviation from the normal
and NP-values can be regarded as representing the inaccuracy of
the latter methods.
One of these investigations was made by a Swedish group
(Bohman and Esscher, 1963) by means of the inverse method referred
to in Section 8.1. It showed that the normal approximation does not
give satisfactory results for the function 1- F, particularly if the
magnitude of 1- F is less than 0'001, the goodness of fit being
worse for small values of the expected number of claims n. When
the NP-method is applied to the same distributions, the comparisons
are much improved when the skewness Yl does not exceed 2
(Kauppi and Ojantakanen, 1966). In the Swedish study some
Fig. 4.1. Comparison of different approximations: (a) n = 100,

Yl = 1.86; (b) n = 500, Yl = 0.84; (c) n = 2,000, Yl = 0.42.
N = normal approximation, Edg. = Edgeworth expansion, N P
= normal power series, MO = Monte Carlo method (cf. Chapter 7, the
slight irregularities in these latter curves arise from the simulation
errors of this method).
Risk distributions were derived from traffic insurance of Finnish
companies. To illustrate the applicability of the NP-approximation
the examples were chosen from unreinsured portfolios (M = O()),
which are so heterogeneous that the normal approximation and
Edgeworth expansion do not give reasonable representation.
47
RISK THEORY
distribution functions S(z) of the size of one claim derived from

statistics of actual claims were used as a basis. Reinsurance was
ignored. In this respect the results obtained do not illustrate the
types of distribution often appearing in practice, where· the top
risks are cut out by means of reinsurance. The NP-approximation
seems to give more satisfactory results even in these extreme cases.
Another investigation has recently been made by a group of
Finnish actuaries (Kauppi and Ojantakanen, 1966) by means of
the Monte Carlo method (Chapter 7). They started from some
distribution functions S(z) corresponding to gross claims deduced
from the statistics of actual claims in the field of non-life insurance.
A number of functions S M(Z) were then derived involving reinsurance
with a maximum net retention M, the latter quantity being a new
variable in the investigations. It was shown that the approximations
were much better behaved, if M is of a moderate size, as is usual
in practice. Fig. 4.2 gives typical examples of the results. The
values in the figures were obtained for a distribution function
derived from joint statistics of some Finnish industrial fire com-
panies. It had a long 'dangerous' tail, the statistics in question
containing claims from £1 up to £1 400000. The reinsurance was
of excess of loss type.
In the same figure values computed by the Esscher formula
(Chapter 6) and the NP-approximation are also shown.
The small irregular deviations in the Monte Carlo curves, par-
ticularly for small 1- Fvalues are due to the random errors of this
method, which makes use of simulation. The larger deviations,
for M = 00, n = 500, arise from the very large risk units of the
S(z) distribution. The values of the maximum net retention Mare
chosen as multiples (1/2, 2, 5, (0) of the standard deviation (J of
the distribution of the size of one claim.
In Fig. 4.2 the ruin probability e = 1- F was represented as a
function of x/P, but in practical applications e is often fixed in
advance. As is seen from Fig. 4.2 (n.b. probability scale) the
relative change of e as a function of x/Pis very rapid in the critical
area e = 0'01-0·001, or expressing the same fact the other way
round: a change of e involves only a slow change in x/Po Hence
the relative accuracy, when computed in the direction from e to
x/ P, can be considerably better. To illustrate this fact, which can
greatly help the usefulness of the normal and NP-approximations,
48
1-F
- - Monte Ca.rlo
'05 - - - Norma..1
_._ .- E$$cher
·03
n., 10 000
·01
·005
·003
'001
1'0 1'1 1·2 1'3
1-F \
'05 \
\
\
'03 \
\
\
\
\
-01
·005
'003
,001
2·50 3'00 3'50 ~
Fig. 4.2. Comparison of 1- F values computed by different methods.

Industrial fire distribution, excess of loss reinsurance having maxi·
mum net retention M. Probability scale.
comparisons associated with the same distribution as Fig. 4.2 are

given in Table 4.1.
The values in the table confirm the earlier known fact that the
reliability of the normal function can be questionable when the
distribution is 'dangerous', i.e. when the portfolio contains large
risks without effective reinsurance. This is particularly the case
when the expected number of claims n is also small.
The circumstances seem to be more satisfactory when the
maximum net retention M is of moderate size, as it mostly is in
49
RISK THEORY
practice. Then the normal approximation may still remain accept-

able, at least if the problem is to solve for x starting from some
fixed ll. In any case it is advisable to apply more accurate methods,
if possible, and the role of the normal formula is to be regarded
as giving a first approximation when the order of magnitude is
needed quickly or when no greater accuracy is necessary.
The NP-method seems to give an essentially better fit than the
normal approximation and its use can thus be recommended as a
rule if the skewness Yl is available and if the calculations in question
do not become too complicated when introducing this correction
to the normal approximation.
TABLE 4.1. Comparison of x-values Ilomputed by Monte

Carlo method (xc), normal approximation (x N ), Esscher method
(x.) and the NP-method (XNP)' Industrial fire distribution.
x __
_N -p 100 x -p 100
_e_ _ xNP-p 100
Xc - p xc- P xc- P
M n
£ 8=0·01 e=O·OOI 8=0·01 8 =0·001 8=0·01 8=0·001
9900 10000 100 100 101 102 101 102

2000 98 95 100 99 100 99
500 96 89 101 96 101 96
100 89 86 99 100 100 101
20 83 72 103 97 105 100
34500 10000 100 102 102 105 102 104

2000 96 92 100 98 100 97
500 92 85 100 95 100 95
100 85 78 100 98 101 100
20 72 66 99 101 103 107
83700 10000 99 98 102 102 102 102

2000 92 88 98 95 98 96
500 89 80 100 94 100 95
100 82 75 103 103 104 106
20 67 59 102 108 113
ex:> 10000 86 81 99 101 100 101

2000 73 63 99 97 101 99
500 60 48 106 102
50
A large collection of comparisons on the basis of numbers com-

puted by Swedish actuaries (Bohman and Esscher, 1964) has been
given by Pesonen in connection with the Risk Theory Symposium,
Stockholm, 1968. This collection illustrates the behaviour of the
NP-method and the Esscher method for portfolios which are not
reinsured, being thus exceptionally heterogeneous. Here also the
NP-method and the Esscher method seem to give about the same
accuracy, acceptable for most practical purposes unless n is small.
In considering the applicability of the normal approximation in
relation to its error it must not be overlooked that several other
sources of error exist in the application of risk theory to practical
problems; these may constitute a greater source of error than
arises from the numerical approximation. Thus the basic assump-
tions and the way of posing the problem contain considerable
arbitrariness, for example in selecting the ruin probability at 0·01,
0·001, or 0·0001 and in estimating the safety loading A. Thus it is
important not to develop mathematical tools of disproportionate
accuracy (and complication) without regard to the context in the
problem being solved.
The domain of the applicability of the normal distribution can
often be enlarged by dividing the distribution S(z) into two parts.
Claims which amount to less than a certain limit Zo can be safely
approximated by a normal or NP method, whereas the tail
z ~ Zo can be calculated separately having regard to the fact that
the number of expected claims n is so small that it is possible to
compute without great difficulty. The required distribution function
F can then be found by convoluting the two partial distributions
as shown in Section 3.6.
51
CHAPTER 5
Applications of the Normal

Approximation
5.1. The Basic Equation

The great advantage of the normal approximation is that it leads
to formulae in which most of the variables concerned are in an
explicit form. The basic expression for this purpose is obtained as
follows.
Let y, (the shorter form y will sometimes be used in place of Ye)
be the solution of the equation:
e = ~ (-y) (5.1)
e.g. YO'Ol = 2·326 and YO'OOI = 3·090. When possible, y-values
should be corrected before further use by means of the NP-method.
Assuming (4.1) is correct the corresponding value of x is given by:
x-P = y.a.Jn (5.2)
In practical applications, x is shown as representing the insurer's
resources, i.e. the maximum amount he wishes to regard as covering
his liabilities. The difference x- P represents the amount he will
have to use in addition to the risk premium P received. The number
B gives the probability of the event that the total amount of claims
will be >x. All this applies to a certain time period, e.g. 1 year.
If a safety loading "AP exists in the premiums then:
x = U +(l+"A)P (5.3)
where U represents the insurer's reserves referred to above. U may
include the company's so-called hidden reserves, i.e. margins in
technical reserves, in valuations and other balancing technical items
in addition to the specific reserves. A conservative approach is not
to include any visible reserves in U, but merely include some or all
52
APPLICATIONS OF THE NORMAL APPROXIMATION
of the hidden items mentioned above. Since the term Adefines the
safety loading in the premiums, the amount AP is available to cover
possible losses.
From (5.2) and (5.3) the following fundamental equation of
normal approximation techniques can be derived.
U = Ylx..jn->.mn (5.4)
or if the NP-approximation is applied
(5.4')
These equations contain the quantities

e,>',n,M, U (5.5)
and depend on the function S(z). M is here the maximum net
retention of one claim on which S(z) depends (cf. Section 3.5.4) as
do m and ex also. If the company applies different maxima for differ-
ent branches and risk groups, then, instead of one variable M,
several variables M l' M 2' ••• may appear. This case is considered
in Section 5.6.
The main problems of risk theory are of the type where the
function S(z) is known, together with four of the quantities (5.5),
the fifth being the object of study. A frequent question is to deter-
mine how the net retention M should be chosen. Also the size of the
required safety loading>. or ofthe reserves U can be sought. Problems
requiring solution for e or n arise less frequently than the other
cases mentioned. Some examples of the various types of problems
are discussed below.
A more general family of problems (in the mathematical sense) is
introduced by treating more than one of the quantities (5.5) as
unknown. In this case the fundamental equation does not give a
definite solution, but auxiliary conditions can be added, for example,
maximizing the expected value of profit. To do this it is necessary to
obtain the expression for the mean value of the profit, according
to the actual circumstances, and to derive a solution of the bound
extremal problem which satisfies (5.4) or (5.4') and at the same time
gives a maximum to the profit function. Another case arises where
the company wishes to use different net retentions M for different
branches of business, in which event there are several unknown
53
RISK THEORY
retention limits M instead of a single value. This again leads to a

bound extremal problem in several unknown variables; an example
of this type will be considered in Section 5.6.
A novel approach to this type of problem has been proposed
recently by K. Borch who introduced concepts developed in the
theory of games. He considers a so-called utility function of the
reinsurance and general operating philosophy of the company and
considers the various quantities (5.5) by giving a maximum value
to the utility. Full details will be found in the original papers by
Borch (1963).
5.2. Net Retention

One of the main problems of the risk theoretical approach is the
determination of the appropriate level of the net retention.
It is assumed that the reinsurance arrangements are such that
each separate risk is reinsured individually; this is, of course, the
usual arrangement for surplus reinsurance, but the model is ob-
viously applicable to excess of loss covers. Stop loss or the various
forms of aggregate covers are considered separately in Section 10.5.
It is assumed that the distribution SM(Z) is known for each value
of M. In the case of the normal or NP-approximation the solution
is almost trivial, since all that is necessary is to solve (5.4) or (5.4')
with respect to M. The simplest method is graphic. The basic
moments are first computed (cf. Section 3.5.4) as in the following
example.
M 9900 34500 84700 00
800 1200 1450 1730

4,84.10 6 2'08.10 7 4'76. 10 7 2'71 . 108
4·15 . 1010 5·78 . lOll 2·90. 1012 2·49. IOu
U is then obtained immediately for the corresponding M-values

by means of (5.4'), and a graph as in Fig. 5.1 can be drawn. From
this the maxima corresponding to different values of U can be
directly read off. Thus, for example, the retention for U = £400 000
is £32.000.
The quantities nand e should be selected, of course, to comply as
54
U/£1ooo
2000
M/£1000
1oo r-----~----~~-+------~----------------~
20 30 40 50 100 200 300
Fig. 5.1. U as a function of M. Industrial fire, e = 0·001, n = 1000,

,\ = 0·05. The dotted line represents the approximation (5.12b).
closely as possible to the real circumstances, although the value of

e, i.e. the accepted level of the ruin probability, is necessarily
arbitrary. If the number n of expected events refers to I year only,
a value of e equal to 0·01 is probably not sufficiently cautious. On
the other hand U may be conservatively estimated by bringing in
part only of the actual reserve, for example, one-half or one-third
of the visible and hidden amounts which the company regards as
an acceptable loss in the year. Should a loss of this size eventuate
then the amount of U must be re-estimated to conform to the
changed circumstances and the maximum retention M modified
accordingly. There is necessarily some vagueness about such con-
siderations because the estimation of U is very much a matter of
taste, as is the fixing of e and the estimation of '\. However, in spite
of these uncertainties some useful guidance regarding suitable
values of the maximum net retention can be obtained by means of
the calculation. The situation arising from a sequence of several
unfavourable years is considered later in this book.
Note. There is a problem attaching to the consideration outlined
above which at first sight appears troublesome. To make the position
clear consider, for example, a hypothetical company insuring only
one risk, the size of which is very large but the probability of a claim
55
RISK THEORY
under which is less than e during the period under consideration.

If the company's criterion is whether the ruin probability is more
or less than e, then the situation is quite acceptable from the
company's point of view. The ruin probability is certainly less than
e, and what is perhaps more surprising, the company needs no
capital U and it can compete effectively in premiums with other
companies, since every positive premium is good. From the super-
vising authority's point of view, the situation can be, formally at
least, satisfactory too, since, if e is small enough, the objects insured
seem to be properly safeguarded. The position of the policy-holder
is, however, problematic. If a claim really occurs, there is absolutely
no possibility that it can be met, but it is probable that the policy-
holder is completely ignorant of the position. This case therefore
has to be considered on the basis of business morals. It is sometimes
eliminated by legislative or supervisory measures.
Another more realistic example would be a company, owing to
reinsurance arrangements or otherwise, having a great number of
risks of limited size with a ruin probability Ie and in addition a
very large risk also having the probability of claim Ie. (3.26) shows
that the probability of ruin is still less than e, irrespective of the
size of the latter claim. This is again, of course, an unreasonable
situation and no responsible company would expose itself in this
way. In practice the condition M < U is always satisfied or it must
be arranged to hold by reinsuring all large risks or excluding such
risks as cannot be reinsured (war, riots, atomic energy, etc.) from
the insurance contracts by special clauses.
More complicated questions arise when dealing with those branches
of insurance where the risk under the policy cannot be defined, such
as liability insurance written with no theoretical upper limit for
claims. Such cases may be subject to reinsurance, and the cedant
thus limits his exposure; the lack of upper limit then passes to the
reinsurer, who in turn may pass it on, and so on. Since however
the number of companies is finite, there must be at least one company
which has an unlimited exposure and who cannot therefore set
M < U. Fortunately this latter condition need not be imposed
provided the probability of a super catastrophe is very small. It
may also be remarked that to talk about an unlimited size of claim
is unrealistic since there is always a certain upper bound to the
total wealth in the world and thus of any smaller part of the total.
56
Exercise 5.2.1. From the claims statistics of a company the quantities

m and IX have been computed for different values of the maximum
net retention M as follows.
M m <X2 <Xa <X,
600 37 9,1.10 3 4,4.10 8 2.4.109

1500 40 24'6.10 3 28 . 10 6 37 . 109
3000 44 42,3.10 3 87 . 10 8 215 . 109
6000 48 66,7.10 3 249.10 6 1215 . 109
The monetary unit in the table is £1. The company is prepared

to lose at most £10 000 in 1 year and requires the probability e of such
a loss to be not greater than 0·01. What is the maximum net reten-
tion, if ,\ = 0·05 and n = 1 ;0001 The normal approximation may
be assumed to give a satisfactory approximation and the function
U(M) should be interpolated graphically (logarithmic scale) from
the values corresponding to the given values of M.
Exercise 5.2.2. The same as Exercise 5.2.1 but using the NP-method.
Exercise 5.2.3. From the claims statistics of a company the square
root of the second moment of the size of one claim is found to be
.J0(2 = £2 240 and the expected number of claims n = 1000. The
reserve fund of the company, together with the annual premiums, is

able to cover the claims with a probability 0·99. The company has
decided to accept a new policy for which, if a claim occurs, the total
size of claim is the fixed sum £10 7 • The company's own net retention
on this risk will be M. It is assumed that under the new policy a
claim will occur once in 10 years and that Afor the original portfolio
as well as for the new policy is 0·05. How large must M be, if the
ruin probability e = 0·01 is to be unchanged after acceptance of this
new risk1 Assume that both the old and the new portfolio obey the
generalized Poisson distribution and that the normal approximation
can be used.
Exercise 5.2.4. Compute the maximum net retention for the distribu-
tion mentioned in Exercise 3.5.4.2 making suitable assumptions
regarding U, A, and e.
57
RISK THEORY
5.S. Reserve Funds

Another important application of risk theory is to determine the
company's reserve funds U as a function of the other basic quantities
(5.5). This problem can be solved directly from (5.4'), as in the pre-
ceding section. From Fig. 5.1 it will be seen, for example, that
corresponding to the value of M = £20000, U = £300000.
Although the above process is not very troublesome in practice,
it does not provide a very general appreciation ofthe interdependence
of the quantities e, A, n, M and U. To obtain a comprehensive
survey it is more useful to derive explicit expressions which involve
the retention M as well, even though the error introduced may be
larger than that already present. Furthermore, situations are not
rare in which general ideas are sufficient, for example, a knowledge
of the magnitude of M or U even when the function S(z) is not
known or when for other reasons it is not possible to make use of
it in calculations involving (5.4'). In these circumstances (5.4') is
written in the form:
(5.6)
An auxiliary parameter is now introduced putting:

«2
K=- (5.7)
mM
for which an upper limit
K ~ 1 (5.8)
can be obtained from the inequality:
J J
M M
~ = zi dBM(Z) ~ M zi-1 dSM(Z) = M CXi_1 (5.9)
o 0
the equality being valid only if the size of one claim is constant and
equal to M. Experience has shown that the parameter K is only
rather weakly dependent on the quantity M and the function S(z)
if M is not very large. If lOmro ~ M ~ lOOmro (mro being used for
m when M = (0), K generally lies in the interval 0·2 to 0·6; it is a
decreasing function of M which --? 0 when M --? 00.
58
From (5.9) another approximation can be found:
oc3(y2-1) < M(y2-1)

60cVn = 6oc.jn
Putting this limit and (5.8) in (5.6) and using the convention P = mn
a distribution free upper limit for U is obtained:
U ~y.j(PM)- >'P+i(y2_I)M (5.10)
If the magnitude of K is known approximately, then the following
formula is available for day-to-day rapid use:
U ~y.j(KPM)->.p+l(y2-I)M (5.11)
If K is not known, the value K = 0·6 can be used with relative
safety, because for the distributions appearing in practice K does
not exceed this value, especially for large values of M. Thus:,
U ~ 1·9.j(PM)->.P (e = 0·01) (5.12a)
~ 2·5.j(PM)->.P (e = 0'001) (5.12b)
The last term in (5.11) being omitted because it is significant only
for large values of M, and, should M be large, the first term on the
right hand side is too large and thus compensates for the omission.
Since these expressions are approximations independent of the
distribution S(z) it is not to be expected that the accuracy will be
good for values of M very much greater than the mean claim moo.
The expressions give values, which are, however, on the safe side,
i.e. the U values are too large.
In Fig. 5.1 the approximation (5.12b) is indicated by a dotted
line. The closeness of the approximation is due to the fact that, for
the distribution used in the figure, K does not deviate very much
from its assumed value of 0·6.
Some examples of how the general theoretical properties of a
portfolio can be studied using the above approximations are now
considered. First considering the magnitude of reserves, the number
of variables is reduced by using M as the unit of measure of U and
P and transforming (5.12b) into:
U = 2'5JP -AP (e = 0·001) (5.13)

M M M
59
RISK THEORY
u
M
40
30
20
50 100 150 200 250
Fig. 5.2.
When M is constant, or more generally, when the quantities U

and P are measured in terms of M, this expression gives a view as
to how the reserves depend on the size of the premium income and
of the safety loading. Specimen values are shown in graphical form
in Fig. 5.2.
As noted in Section 2.5, the magnitude of the reserves required
depends essentially on the value of the safety loading A. If Ais non-
positive then U increases at least proportionately to the square root
of the premium income P. If however Ais positive, the curve of U
has a maximum given by:
U 1·6 P
- =- for (e = 0'001) (5.14)
M A M
Even if (5.13) is only approximate the general shape of the results
is, however, correct, the true values of U for large values of M being,
in fact, smaller. Exact values of the maximum and of the critical
point U = 0 are derived from (5.4') in the same way as the results
in Section 2.5.
That the need for reserve funds is dependent on the size of the
company can also be illustrated by means of the following example.
Let 0 1 and O2 be two insurance companies. The respective minimum
60
reserve funds Uland U 2 are computed by means of (5.4). The

question can now be asked: how is the need for reserve funds
changed, if these two companies are incorporated by merger into
one company O.
By means of (5.4), (3.27), and (3.28) and assuming independence
of the portfolios the following expression for the minimum reserve
U of the merged company 0 is obtained:
U= Y J{ nlCXI2+n2CX22
n 1+n 2
(nl +n 2)
} A1m 1n 1+A 2m 2n 2
mInI +m 2n 2
(mInI +m 2n 2)
(5.15)
The lower indices, 1 and 2 respectively, indicate the companies 0 1
and O2 , If, on the other hand, the same formula (5.4) is applied
separately to the companies Oland O2 the following formula is
found for the difference of the theoretical reserves:
U1+ U2- U = y.j(nlCXI2)+y.j(n2CX22)-y.j(nlCXI2+n2CX22) >0

(5.16)
The inequality follows on the inequality .ja +.jb >.j(a + b) which
is valid for every a >0 and b >0.
The inequality (5.16) proves that the reserves needed by the merged
company are always less than that of the component companies together,
if the security level 1 - e is unchanged.
The rule can be extended to deal with incorporation of more than
two companies (cf. Exercise 5.3.2).
The result is of general interest; it proves that relatively the
larger the company, the smaller the reserves needed. The same fact
can also be directly seen by transforming (5.4) into the form:
U cx 1
P = Y .jm .jp-A (5.17)
which shows that the relative minimum reserve U/ p is a decreasing

function of the premium income P.
It is thus seen that the merger of insurance collectives into one
large collective gives an advantage from the point of view of
stability, i.e. in this way the existing reserves are used in the most
effective way, or what amounts to the same thing, a merger releases
"idle reserves" if the security level is not changed. In practice a
61
RISK THEORY
fiscal merger of the companies is, of course, not necessary. The same
advantages can also be reached by exchange of reinsurance on a
reciprocal basis cf. Exercise 5.3.6. This problem will be further
considered in Section 9.2.
In the following exercises it may be assumed that the normal
approximation can be applied.
Exercise 5.3.1. The following characteristics are computed from the
statistics of an insurance company: m 1 = £100, OCl = £300, n 1 =
1 000. The company has a reserve fund U 1 = £20 000 and security
loading Al = 0·1.
Another insurance company with the following characteristics,
m 2 = £50, OC2 = £400, n 2 = 200, A2 = 0'05, is merged with this
company.
If the ruin probability e of the former company is not to increase
following the merger, how large should the reserve fund U be for
the merged company1
Exercise 5.3.2. Prove that if k companies are amalgamated the joint
company needs a smaller reserve fund U than the sum of the funds
of the original companies provided that the security level measured
by e is the same in all cases.
Exercise 5.3.3. It is assumed that the risk properties of an insurance
portfolio are so improved that the frequency of claim decreases
equally for each risk unit by 10%. How much could the reserve
fund U be decreased if the ruin probability e is maintained at the
original level ? It is assumed that the company has the same values
for n, m, oc, A, and U as company No.1 in Exercise 5.3.1. and that
the premiums remain unchanged.
Exerci8e 5.3.4. A friendly society grants funeral expense benefits and
each member of the society can choose a benefit of either £100 or
£200. It is assumed that ,\ = 0·5, n = 20, and e = 0·01. How large
must the reserve fund U be if it is not known in advance how many
members will choose the option £100 and how many the option
£200 and consequently the exercise of these options is to be assumed
to be such as to maximize the danger1
Exerci8e 5.3.5. The distribution function of one claim of a company
can be represented (taking the average size of claims as the monetary
unit) by the exponential function dS/dz = e- Z (z~O) and the
62
expected number of claims n = 1 000. How large must the security

loading ,\ be, if the company has no reinsurance and no reserve fund,
if the total amount of claims can be approximated by the normal
distribution and if e is fixed at 0'001?
Exercise 5.3.6. There are r insurance companies with premium
incomes PI' P 2' ••• , P r. The companies exchange reinsurance on
a reciprocal basis. The contract provides that company i takes
from each risk of all companies a share PdP, where P = P 1 +P 2
+ ... + P r. Prove that the total reserve funds required in these
companies is the same as if the companies were incorporated into
one single company.
5.4. Statutory Basis of Reserve Funds

In come countries there is a statutory defined minimum basis for
"a solvency margin", i.e. for the reserves to be held by insurance
companies. The foregoing results indicate that a theoretical basis
exists for these reserves. It will be appreciated that the fact that a
company requires reserve funds for other reasons than to cover risk
fluctuations, such as investment losses, is not considered here.
Whilst the normal situation of the insurance business is that ,\ has
a positive value, it is not unreasonable that the object of the legisla-
tive requirements is to ensure that the values of reserve funds are
adequate when applied to the weakest cases. This means that a
positive value for ,\ cannot be regarded as axiomatic; furthermore
when fluctuations in basic probabilities are considered later in
Chapter 10 the effect may be equivalent to negative values for ,\
for a number of years. Taking the value ,\ = 0 (or ,\ very small) as
a starting point, it can be said that the reserve funds have to be
jW;ed in proportion to the square root of the size of the business.
A rule of this type is incorporated in the legislation of some
countries, for example in Great Britain the excess of assets over
liabilities must be not less than the following:
Premium Income Minimum Excess
Not exceeding £250000 £50000

> £250 000 but not > £2 500 000 20 % of income
>£2500000 £500000 plus 1/10th of income
in excess of £2 500 000
63
RISK THEORY
In Finland the requirement is met by three linear segments. Both

of these cases can be considered as an approximation to the parabola
in question.
To impose a constant size for the reserve funds by legislation is
open to the criticism that if, for example, it is determined by (5.14)
it would be unreasonably onerous for small companies and, having
regard to the earlier comments on the variability of A, perhaps
insufficient for large companies.
5.5. The Rule of Greatest Retention

In Section 5.2 the magnitude of the net retention was considered
and this is now approached by the application of the rule of thumb
(5.12b), which is rewritten in the form:
M = 0.16 (I+APjU)2 (e = 0·001) (5.18)
U PjU
and is shown graphically in Fig. 5.3.
These curves show that the variation of MjU with the premium
is remarkably limited but that a far greater variation is shown
according to the safety loading. For a given U the curves each have
a minimum at the point P = UjA of:
Mmln = 0·64 AU (5.19)
M
'20 U
~ ,.·20~
~~;:-:::
~~='OS
A=O :=
5 10 p
U
Fig. 5.3.
64
The above formula (5.19) leads to a rule of thumb which is often

used in practice by reinsurers, i.e. that the net retention should be
a certain percentage, say 5%, of the reserves U which the company
is willing to lose by way of covering losses during a year. If A is
put equal to 10%, the rule gives M ~0·06 U or, if A = 5%, the
retention M should be 3% of the reserves U. These figures are
based on e = 0·00l; if e = 0'0l the rule of thumb becomes:
M min = 1·1 AU (5.20)
A rule of thumb which is not infrequently met in practice is one

according to which M is fixed as a certain percentage of the premiums
P or of the total sums insured. It will be apparent from the fore·
going discussion that, in general, this type of rule is not satisfactory
since the critical quantities entering into this decision should be
,\ and U.
5.6. The Case of Several M's

The fundamental equation for the normal approximation (5.4) will
now be utilized to deal with the case when the portfolio is sub·
divided into sections, say 1, 2, ... , i, ... , r, each of which has its
own claim distribution St(z), safety loading At> 0, expected number
of claims nt, and net retention Mi.
The problem is so to determine the Mt's that:
(i) the expected amount of surplus as a function of the Mt's:
j(Ml' M 2 , · · · , Mr) = LAtntmt (5.21)
is maximized; and
(ii) the fundamental equation (5.4) is satisfied for the whole
portfolio:
Q(M 1 , M 2 , · · · , Mr) = -yrx..Jn+ L Atntmt
U = 0 (5.22)
where ac is derived from the corresponding coefficients of the sections
by the formula nac 2 = Ini<Xi2 (cf. (3.28)).
This bound extremal problem is solved by the use of the so·
called Lagrange method by introducing a function:
F =j-pQ,
where p is an auxiliary variable.
For a real solution some assumptions concerning the portfolio
65
RISK THEORY
and reinsurance are needed. As an example the problem will be

solved assuming excess of loss reinsurance. To facilitate the calcula-
tions it is further assumed that Ai is independent of the retention Mi
and U>O
To avoid trivialities it is necessary to ensure that reinsurance is
required, or in other words, that with sufficient increase in all the
Mt's, Q becomes negative. The existence of continuous derivatives
Si' can also be assumed without any essential restriction, since
the possible step points of Si can be closely approximated by a
continuous derivative. To circumvent troubles on boundaries the
simplest way is to define S1' = 0 for Mi ~ O. This ensures that the
derivatives introduced in the following analysis vanish for these
values which then need not be specially considered.
Using these assumptions:
It follows after some manipulations that:
8F = n t (l-Si(Mt)){(I- P) "i+
8Mt
P~
!X."n
Mi}
The extremal values can be found among the joint zero-points

of these derivatives and Q. Putting the latter factors equal to zero
and denoting:
!X.~n p-l
JL=--
Y P
it follows that:
(5.23)
Substituting these values into (5.22) J.1, can be determined (cf.
Exercise 5.6.1).
In actual cases where the St,'s are known it may be directly possible
66
to see that (5.23) gives an absolute maximum for the profit function
f on the surface Q = O. This can also be proved by the general
conditions given above, analysing the behaviour of the derivatives
of the functions f and Q directly, but details of the proof are left
to the reader.
The following theorem is now obtained on the assumptions men·
tioned at the beginning of this section.
The liml~ts of retention Mt have to be chosen proportional to the
corresponding safety loadings Ai.
This theorem has been proved on the assumption that the normal
approximation is justified, but it is important to note that the
normal approximation has not been applied separately to each
component i. It thus follows that the theorem is certainly valid so
long as the normal approximation is applicable to the whole business:
in view of the central limit theorem this is a much less restrictive
condition than if the corresponding assumptions were made for
the various subgroups.
The foregoing result is of general interest when consideration is
being given to the overall reinsurance policy of a company. It is not
an unexpected result in the sense that it indicates that the greater
Ai is (or, what is equivalent, the greater the expected profitability
of a branch of insurance or group of policies) the greater the amount
that should be retained for the net account. In this context it is
not necessary to have regard to such items as the cost of reinsurance
in relation to the arrangements made, since it can be assumed that
allowance has already been made for these in arriving at the estimates
of At. As a special case it follows that the proper course, if all the
At's are equal, is to make the retentions M t equal. This rule contradicts
the oft· quoted practice which requires the retentions on "dangerous"
risks to be reduced. If the safety loading is actually the same for
such risks as for simple risks then this has no theoretical background.
If, of course, the dangerous risks are actually written at inadequate
premiums, then the margin Ai is lower and the retention should be
reduced. However the analysis shows that intelligent comments
about the retention can only be made by reference to the adequacy
of the premiums.
Exercise 5.6.1. Prove that, with the conditions given in Section 5.6.,
there is only one JL for which Q vanishes.
67
RISK THEORY
5.7. An Application to Insurance Statistics

It is by no means necessary to limit the application of the theory of
risk to the whole portfolio of an insurance company. On the contrary
it is applicable to any insurance collective. A common example is
in the collection of insurance statistics, where the collective in
question can be a group of policies for similar risks, and for which
a premium tariff is being calculated separately taking claim fre-
quencies as net premiums. In this connection the problem of the
accuracy of the claims frequency derived from the statistics often
arises; or, in other words, how large must the group in question be
so as to give an adequate statistical basis for rate making.
As an example consider a certain group of similar policies which
have been observed during a certain period - in general, several
years. Suppose that the total amount of the claims during this
period has been £80 000. The total amount of the insurance in force
in this group is £70000 000, this amount being the sum of the
average sums of each year. The so-called burning cost is now
obtained as q = 80000/70000000 = 1·14%0 and the problem is
to estimate the accuracy of this quantity.
The solution is a direct application of (5.2). The total amount of
claims (£80000) can be interpreted as a net premium P, and
x - P as the error due to random fluctuation. Then the relative
error is Jq/q = (x-P)/P. Hence:
IJq I 1
IX IX 1
q ~ Y• .Jm .J P = Y. ;, .In (5.24)
with probability I-e. If e = 0·05 is taken as the confidence level,

then Ye = 1·96. The values of m and IX can easily be computed from
the statistics in question, or (if they are not known and an advance
estimation of the error is needed) they may be obtained from the
general experience of the insurance branch in question. In this
case, let n = 1000 and IX/m = 5. Then:
I~ql ~ 1'96.5·J1~ = 0·31
Because the risk groups in statistical considerations are often

68
rather limited, the applicability of the normal approximation may

be uncertain. However, this approximation gives fairly reliable
values particularly if the frequency is not very small. e need not
be as small as it customarily is in risk theory which also adds to the
usefulness of the formulae. Often it is necessary to know only the
order of magnitude of the relative error.
It may be further noted that the same estimation problem can
be easily solved, and often in a much more reliable way, by dividing
the observation period into sub-periods, preferably individual years,
and by calculating the quotient qv separately for each year v. Then
the estimation:
IAqq I ~- Ye
ij
J( L (qV-ij)2)
N(N-l)
(5.25)
holds, where ij = L qv/N and N is the number of sub-periods.

The details of this method will not be developed, because they lie
outside the scope of the theory of risk. In practice the risk theoretical
formula (5.24) is perhaps most useful in cases where the claim
statistics are being forecast in advance and some estimation of the
error is needed at this stage when a statistical basis for (5.25)
does not exist. (5.24) can be useful in some cases in defining
an appropriate classification of risks or in deciding how many
years of statistics are needed to obtain a satisfactory reliability.
Exercise 5.7.1. It is known that for certain fire risks q :;::j 1%0 and
the number of claims per annum is about 1000. How many years
statistics are needed to estimate q with an accuracy of 20% on a
10% confidence leveH a./m is estimated, from experience, to be 10.
5.S. Experience Rating, Credibility Theory

In connection with some insurance branches, different kinds of
bonus systems are applied, for example the no-claim bonus of
motor-car insurance is well known in many countries. Another
bonus system is sometimes attached to reinsurance treaties and also
to direct insurance contracts. If the original premium is (1 + >..)P
and the actual amount of claims in a year is ~, then a bonus or
"profit return" can be agreed, e.g.:
G = k[(I+>..)P-g] G~O (5.26)

69
RISK THEORY
where k is some constant < I and P = E{,} is the mean value of

the annual amount of claims and .\ is the safety loading. This type
of bonus system is mostly used in cases where the statistical basis
of the premium rating is narrow or does not exist at all and conse-
quently the true premium can be very uncertain. To secure an
equitable rating the policy-holder (or cedent) can ask for a bonus,
or in more general terms, for an "experience rating" which adjusts
his final cost of insurance (= (I + .\) P - G) to a level corresponding
to the actual circumstances of the risk.
In practice the most appropriate circumstances in which to apply
experience rating are where the number of claims per annum is
not very small, for example, the different kinds of collective
accident and pension insurance covers for groups of persons such
as the staff of companies. The principle can, of course, also be
applied to other policies involving a large group of risks and it is
often used in connection with reinsurance treaties, where the
cedent is often entitled to a profit return in accordance with a
formula similar in principle to (5.26).
The theory of experience rating is built on the same basic function
F(x), the distribution function of the total amount of claims, as the
theory of risk. Accordingly some elements of this theory will be
discussed to provide a further illustration of the actuarial use of
this function even if, in fact, the topic is to a large extent outside
the scope of the theory of risk.
To present the matter in a more general way the bonus formula
(5.26) can be written as follows:
G = [a(l +.\)P-b'] G;;;O (5.27)
a and b are constants defined in the bonus contract.

The basic inequality of experience rating is found by the condition
of unbiased rating:
(5.28)
which defines an appropriate absolute upper lhnit for the bonus.

To give a more explicit form to (5.28) use is made of the distri-
bution function F(x) of the variable " i.e. the same generalized
Poisson function that has been considered in this chapter, but now
70
applied, of course, only to that limited collective which concerns

this particular contract. Then, denoting Xo = a(1 + A)E{g}/b
Xo
E{G} = a(1 +A)Eg}F(xo)-b J xdF(x) ~ AE{t}

o
Partial integration gives:
Xo
b J F(x) dx ~ AE{t} (5.29)

o
The limits of the integral correspond to the interval in which gmust
lie for a bonus to be given in accordance with (5.27). This equation
defines one of the constants a and b, the other being free. This free
variable can then be fixed by some subsidiary condition, selected so
as to give the bonus contract some further desired features. Am-
meter (1963) has studied systems where the variance of G is mini-
mized or where the net premium P-G is as correct as possible even
if the original P rating is wrong.
As a numerical application the particular case (5.26) where
a = b = k is considered; it is assumed that F is normal (4.1). Hence:
~7; (x:;;) dx ~ Anm

which gives, after some simplification:
k < Ayo
= AYO
J CP(y) dy
where Yo = mJn/lX. If, for example, ,\ = 0·1, IX/m
= 4 and = 100, n
thenk~ 0·48. This means that the insurer can return at the most
48% of the profit, the remainder of the profit, i.e. 52%, being
needed to cover the claims including the risk of loss x > P.
The structure of the above-mentioned contract between the
policy-holder and the insurer means that it is the same as the con-
ventional stop loss reinsurance. The policy-holder bears the small
fluctuations of risk on his own retention whereas the risk of large
total losses is insured. Stop loss theory is dealt with later on. Because
71
RISK THEORY
the collective considered in experience rating is often small, the

applicability of the normal distribution, even as a first approxima-
tion, may be uncertain. In particular, fluctuations in the basic
probabilities (Chapter 10) may be important. It is therefore ad-
visable to use for F(x) in (5.29) the compound functions studied in
Chapter 10 and to experiment also with approximations other than
the normal.
It should be noted that it is unusual in practice for E{g} and
hence also ,\ to be known in advance. The formula can then be
helpful only if assumptions are made regarding these quantities
(e.g. on the basis of general experience based on similar contracts)
but the formula is useful in understanding the structure of the
premiums for experience rating.
The whole philosophy of experience rating has not yet been
discussed. In general, the main reason for the practical application
of experience rating is to try to reach reasonable premiums by
starting from a hypothetical value Po and subsequently correcting
it by using the actual claims experience by means of some agreed
rule, for example similar to that given below in (5.30). When
defining what are considered to be "reasonable premiums", attention
is to be paid partly to the requirement that, at least over several
years, the mean premium should not be too far from the actual
expected value of claims and partly to the requirement that the
premium should not show too much random fluctuation. An example
of this kind of arrangement is set out in the following system of
"sliding premiums" which has been extensively studied in America
under the name "credibility theory".
Consider a risk or a group of risks which have the same initial
premium Po. This group can be, as mentioned at the beginning of
this section, a collective of persons or objects subject to some
group contract. The same method can, however, also be applied
for the adjustment of the general tariffs of a company, in which
case it is applied separately to the different tariff groups, e.g. brick
houses in some defined area, etc.
Suppose further that it is agreed that the premium for the next
year is calculated according to the formula:
(5.30)
where go is the total amount of claims in this collective in the year O.

72
The "braking constant" Z, called "credibility", is chosen from the

interval:
(5.31)
and it will be fixed small enough to eliminate excessively large

random fluctuations. More precisely it is subject to the condition
that pure random fluctuations will not, with probability 1- e,
result in a change in the premium P in excess of 100 p%. Expressed
in symbols this is the case if the constant Z satisfies the condition:
(5.32)
where Llx is obtained from:
F(E{g}+Llx)-F(E{g}-Llx) = l-e (5.33)
which assumes that F is known or preassumed. Then the absolute

value of the deviation Llg = g-E{g} can be larger than Llx only
with probability e.
If it be assumed that the normal approximation gives a satis-
factory approximation for F, then (5.24) of the preceding section
is immediately applicable because LlqJq is in fact = LlxJE{g} and
thus:
1
m-In = P
Q(
ZY' (5.34)
from which is obtained the formula for Z:
p m
Z = - -
Y. IX
-In (5.35)
The expected number of claims n which makes Z = 1 is of special

interest, i.e.:
n = y.2 (~)2 (5.36)
o p2 m
Following the terminology of the American credibility theory it

is said that if Z = 1 there is "full credibility". In the special case
where the risk sums are all equal or, what is equivalent, if only
number of claims is recorded for calculating the frequency of the
73
RISK THEORY
claims, then IX/m = 1 and the values of no which are large enough
for full credibility are immediately obtained by means of tables of
the normal distribution as follows:
Values of no for full credibility:
p 10% 5% 1%
0'01 27057 38416 66347

0·05 1082 1537 2654
0·1 271 384 663
0·2 68 96 166
In most practical cases the risk sums are not equal and hence
IX/m is not 1. The variation in the value of this quantity depends
significantly on the degree of heterogeneity of the risk sums and
consequently the limit of full credibility can be considerably larger
than is given in the table above. The values of IX/m may often be of
the order 3 to 5, but in cases where very large risk sums can occur
the values can be much larger.
If the expected number of claims n is smaller than the value
obtained from (5.36) then the constant Z has values smaller than 1
and the term "partial credibility" is used. From (5.35) and (5.36),
one ofthe well-known formulae of credibility theory can be obtained
immediately by eliminating the joint coefficient (pm)/(IXY), i.e.:
Z = .jn/.jno (5.37)
In the foregoing it was assumed that the normal approximation
could be used because this assumption or some other assumption
concerning the distribution function F is needed for the computation
of no. However, owing to the small size of the risk collective which
often arises in cases subject to experience rating or to credibility
theory, the applicability of the normal distribution can be doubtful,
even if very small values of e are not needed for which the accuracy
of the normal distribution is most unsatisfactory (cf. Section 4.4).
The normal distribution can of course be avoided by calculating the
quantity Y. in (5.36) using some other method of computation of the
generalized Poisson function or by applying the compound functions
of Chapter lO. A drawback is, however, that Ye may depend on n.
74
The experience of American actuaries, however, suggests that the

normal approximation gives values which are satisfactory in practical
work.
Without going into further details it should be noted that an
experience rating scheme is often conditioned by special rules for
excluding catastrophic claims, for which a special loading is included
in the premiums. Formula (5.30) can also be generalized taking
ei
into account the claims of several consecutive years 1, 2, ... by
means of some appropriate formula, Le.:
For further studies the reader is referred to Ammeter (1963).

Biihlmann (1964), Philipson (1967), and Lundberg (1967).
Exercise 5.8.1. A large commercial firm having 1000 lorries has
insured them under a collective treaty, the premium (without
loadings) being £100 per vehicle. It is agreed that the premium
will be adjusted in accordance with credibility theory. What should
the premium be for the next year if the total sum of claims in the
first year is £600001 The number of claims was 200, ex = £1500,
P = 0·1 and Ii = 0·1.
75
CHAPTER 6
The Esscher Approximation
6.1. Introduction
Since it is known that the normal distribution does not give a good
approximation if the basic distribution S(z) is very heterogeneous,
particularly if at the same time the expected number of claims n
is small, another approximation found by Esscher (1932) has been
widely used. Unfortunately the error of Esscher's formula also
remains a feature which is difficult to handle mathematically, but
even so it has been found in practice to have a broad domain of
applicability. The new methods recently developed, i.e. NP-approxi-
mation, Section 4.3, Monte Carlo method, Chapter 7, an inversion
method, Section 8.1, may provide arguments for re-evaluation of the
expediency of the Esscher formula, but nevertheless this method is
one of the important tools available for computing the numerical
values of the generalized Poisson function and it will now be intro-
duced. A reader who is not familiar with the manipulation of
integral transformations and series can at the first reading jump
directly to the result (6.7) on page 79.
In fact the Esscher formula makes use of the normal distribution
or of the first few terms of the Edgeworth series, but the idea is
first to transform the generalized Poisson function in such a way
that the required value of x is moved to the area where the fit of
the normal approximation is closest. The transform in question is:
.!. f eMI dS(y)

:l:
Sex) = (6.1)
Po 0
where the constant Po is fixed by the condition S(oo) = 1, which

makes Sex) again a distribution function. h is an auxiliary constant,
76
THE ESSCHER APPROXIMATION
the value of which is left open at first and will be fixed later. For
the following a set of constants is defined:
f y" e
00
13" = hll dS(y) (6.2)

o
It is seen immediately that f30 above is also obtained from this
equation as a special case when k = O. The generalized Poisson
function F(x) with claim size distribution B and mean claim number
n will now be calculated in terms of the original distribution F(x).
In order to obtain the k'th convolution ofB it suffices to use only
the theorem (Section 3.4) stating that the distribution function
corresponding to a product of characteristic functions of several
distribution functions is the convolution of these distribution func-
tions. Calculation gives, by induction from k-l to k:
dS"*(x) = (f3o)" e- hz dS"*(x)
hence:
= e-n+nPo-hx dF(x)
dF(x) = e-n+nPo-hx dF(x) (6.3)
Since the mean of B is clearly iii = f31/f30' the mean of F is nm
= nf31' The standard deviation of F is equal to .j(iif32/f30) = .j(nf32)'
The Edgeworth expansion (4.3) is now applied to the distribution
function. Since iici" = nf3", (6.3) gives:
(6.4)
where z = (x-n (31)/.j(n (32); R is the remainder and the coefficients

e" are obtained from those of (4.3) by replacing (Xk by f3". A closer
study indicates that the absolute value of the quotient of the latter
and former term in parenthesis tends to infinity with x. The fit
dF ~ d(F - R) is best in the neighbourhood of the mean z = 0
or x = nf31' There is always some value in the interval
(nf31-.j(nf32)' nf31 +.j(nf32))' where the fit is exact, but it is not
possible to give a general rule as to where in this interval the fit is
77
RISK THEORY
best. It is therefore advisable to so arrange that the point x for

which the value of F(x) is required, is chosen to be the mean value
of F. This is possible, since the constant h has not yet been fixed.
f
00
x = nfJl = n ye llU dS(y) (6.5)

o
Then evidently if x<nm (~nm), h<O (~O). The intention is
to omit the remainder term dR in (6.4) and integrate. Having
regard to the sign rule of the constant h, it is advisable to integrate
(6.4) from - 00 to x when x<nm and from x to + 00 when x~nm.
Then for values of the variable of integration y near to x the fit
(6.4), excluding dR, can be expected to be good. For values remote
from x the exponential factor e-hy decreases very quickly thus
making the poorness of the fit of no significance. Suppose for example
that x~nm. Then integration gives:
f e- hZv(nfJ
00
1- F(x) = e-n+nJJo-nhfJl 2) dzLck <J>(k)(z)

o
For numel"ical calculations the so-called Esscher functions:
f e-uzd<J>(k)(z)
00
Ek(U) =
o
are introduced. These functions are easily derived from the recursion
formula:
f
00 00
Ek(U) = [ e-Uz<J>(k)(z) ] +U e-uzd<J>(k-l)(z)

o 0
starting from:
E (u) =
o
f
00
0
e-uzd<J>(z) = l-<J>(u) _ A(u) sa
.J(21T)<J>'(U) - .J(21T) , y.
The lowest order Esscher functions are as follows:

.J(21T)E 1 (u) = uA(u)-l
.J(21T)E 2(u) = u 2A(u)-u (6.6)
.J (21T)E s(u) = u 3A(u) - u 2 + 1
78
Using the Esscher functions the result is:
1-F(x) = e-nw{Eo(u)+csEs(u)+ .,. +Remainder} (6.7)
where:
U = hJ(nP2) (6.8)
w = 1-Po+hPl (6.9)
The explicit values of the coefficients Ck are derived from the

expansion (4.3). In practice it is seldom found necessary to go
beyond cs' which can be expressed as:
(6.10)
Summarizing the above the calculation scheme is as follows. The

value of F(x) is required for a given value x, which is supposed to
be ;:;; nm. The number n and the distribution Sex) being given, the
value h is first found from (6.5). In the general case, h can only be
found by successive approximation, i.e. calculate the right. hand side
for a few values of h until it is possible to interpolate (cf. Section
6.3). If Sex) is expressed in suitable mathematical terms, an
analytical solution may be possible. The values of Pk for k # 1 are
then found from (6.2) and u and w from (6.8) and (6.9). The constant
ca is found from (6.10) and the values of Ek(U) from the relationships
(6.6) as soon as Eo(u) has been calculated. The expression (6.7) then
gives the result.
Short tables of values Eo, .JEo, uE o, E a, .JE a, and u2ES follow.
The columns uEo and u2Ea are more convenient for interpolation
for larger values of u.
6.2. The Accuracy of the Esscher Formula

Some methods are available for estimating the error in the Esscher
approximation (Esscher, 1932; Pentikiiinen, 1952). They have shown
that in many cases where the primary distribution S(z) is not very
heterogeneous the formula gives a very good fit. These estimation
methods are, however, very laborious to apply. In fact the present
knowledge of the accuracy of the Esscher approximation is based
mainly on comparisons which are obtained by parallel computations
79
TABLE 6·1. The Esscher functions Eo and Ea
u Eo iJEo Ea iJEa u Eo iJEo Es iJEa

0·0 0·5000 375 0·3989 35 2·1 0·1620 56 0·1402 73
0·1 0·4625 333 0·3954 90 2·2 0·1564 54 0·1329 67
0·2 0·4292 295 0·3864 126 2·3 0·1510 50 0·1262 65
0·3 0·3997 264 0·3738 148 2·4 0·1460 47 0·1197 58
0·4 0·3733 237 0·3590 161 2·5 0·1413 44 0·1139 56
0·5 0·3496 213 0·3429 167 2·6 0·1369 42 0·1083 52
0·6 0·3283 192 0·3262 167 2·7 0·1327 39 0·1031 50
0·7 0·3091 173 0·3095 165 2·8 0·1288 38 0·0981 45
0·8 0·2918 158 0·2930 160 2·9 0·1250 35 0·0936 42
0·9 0·2760 144 0·2770 154 3·0 0·1215 152 0·0894 186
1·0 0·2616 132 0·2616 147 3·5 0·1063 119 0·0708 114
1·1 0·2484 120 0·2469 139 4·0 0·0944 96 0·0594 III
1·2 0·2364 III 0·2330 132 4'5 0·0848 79 0·0483 79
1·3 0·2253 101 0·2198 124 5·0 0·0769 121 0·0404 110
1·4 0·2152 94 0·2074 116 6·0 0·0648 89 0·0294 72
1·5 0·2058 87 0·1958 109 7·0 0·0559 68 0·0222 48
1·6 0·1971 81 0·1849 101 8·0 0·0491 53 0·0174 35
1·7 0·1890 74 0·1748 96 9·0 0·0438 43 0·0139 25
1·8 0·1816 70 0·1652 89 10·0 0·0395 130 0·0114 63
1·9 0·1746 65 0·1563 83 15·0 0·0265 66 0·0051 22
2·0 0·1681 61 0·1480 78 20·0 0·0199 0·0029
TABLE 6.2. Esscher functions uEo and u2Ea for large values
of the argument.
u uEo u2Ea
5·0 0·38460 1·00746

6·0 0·38868 1·05668
7·0 0·39124 1·08952
8·0 0·39298 1·11232
9·0 0·39420 1-12871
10·0 0·39507 1·14083
15·0 0·39719 1·17103
20·0 0·39795 1·18212
25·0 0·39830 1·18736
30·0 0·39849 1-19023
35·0 0·39862 1'19197
40·0 0·39869 1·19310
45·0 0·39875 1·19388
50-0 0-39878 1-19450
of values for the generalized Poisson function using methods under

which the errors can be controlled.
The largest collection of comparisons mentioned above is published
by Bohman and Esscher (1964). In Table 6.3 some typical examples
from this work are set out.
Attention must be paid to the fact that the examples given in
the table are extremities in that the distribution functions S(z) have
a long dangerous tail and no reassurance is assumed. The accuracy
seems to be very satisfactory and in fact considerable deviations only
appear as a rule in cases where the correct curve of 1- F has
irregularities like those shown in Fig. 4.2. The Esscher approxima-
tion cannot follow these irregularities, which can appear if the basic
distribution is very heterogeneous and if at the same time the
expected number of claims n is small.
TABLE 6.3 Comparison of values computed by means of inversion

method and Esscher formula.
Esscher
Distribution n x-p 100· (I-F) values % of
cxy'n I-F
Life 100 3 2·00 177

4 1·84 92
6 0·36 100
500 3 1·59 110
4 0·50 106
6 0·03 117
1000 3 1·26 99
4 0·28 102
6 0·01 104
Non- 100 3 1·71 223

industrial 4 0·89 229
fire 6 0·38 149
500 3 1·92 101
4 0·76 88
6 0·06 119
1000 3 1·40 98
4 0·35 105
6 0·02 100
81
RISK THEORY
Examples are given in Fig. 4.2 of another comparison obtained by

means of the Monte Carlo method (Kauppi and Ojantakanen, 1966).
In these cases also the accuracy is very good and deviations appear
only in extremal cases, where no reinsurance is assumed and n is
relatively small. The Esscher line on a probability scale straightens
the irregularities in these cases.
As mentioned already in Section 4.4 and as is seen in Fig. 4.2,
the NP-approximation and Esscher approximation seem to give
results which are fairly near to each other even in cases where the
normal approximation completely fails to give even the right order
of magnitude. Further analysis seems to show that in fact the
Esscher formula gives somewhat better accuracy, at least in critical
cases, but the differences can be diminished by taking one further
term into account in (4.9) in the NP-method. Present experience is,
however, too narrow to enable any firm conclusion concerning
comparison of these two approximations to be made and the
ultimate limits of the applicability of both formulae are still open.
Consequently, if possible, it is advisable, at least for critical cases,
to carry out parallel computations by different methods to try to
obtain an understanding of the actual circumstances.
6.3. Some Hints for Numerical Computations

If electronic computors are available, the Esscher formula can be
programmed. For example if S(z) is defined as a step function by
means of the statistical method (Section 3.5.2):
S(z) = L Pi (i = 1, 2, ... ) (6.11)

Zi~Z
the programme can be arranged so that the numbers Zi and Pi are

put in first. Then the different values of the auxiliary function h
can be put in and the values x and 1- F obtained as output. If the
risk units are reinsured by means of some individual method the
risks are limited at Zi ~ M. If this is not the case then "the tail"
can be expressed by means of some analytical function and the
mixed S(z) again be programmed.
If programmed computors are not available the computations
can be done by means of conventional calculating machines in
82
accordance with the scheme mentioned above. Some work is saved

by making use of the formula:
f zkehZdS(z) = f zk(ehZ-l)dS(z)+ f zkdS(z)

00 00 00
13k =
0 0 0
f zk(ehZ-l)dS(z)+OCk
00
= (6.12)
o
since ehz-l is more convenient to integrate numerically than e hz

for small values of hz.
If h is small, the expansions:
= f zk(l +hz+ (hZ)2

2! + ... )dS(z)
= f zkdS(z)+h f zk+1dS(z)+2"!h f zk+2dS(z)+ •••

2
h2
= OCk+ hOCk+1 + 2! OCk+2+'"
are convergent enough for numerical calculations. Putting these

expansions into the Esscher formulae the calculation scheme:
_ 1 h 3 OC3
T3 - - -
3 m
f zkdS(z).
00
is obtained where m = OC1 and OCk is as defined above, i.e.

o
83
RISK THEORY
Then:
P = mn (= net premium income)
U = J[P(2 ;+T 3 +2T 4 ) ]
c' = [::3/2 -h ::312(1'5 ::-::)] J6m

e = I-F = e-Pw1m[ Eo(u)- J~ Ea(U)] (6.13)
x 1
)C = p = 1+h" (2·00 T2 +1·50 T a+l'33 T,) (6.14)
(6.13) and (6.14) are valid only if h is small enough to give a rapid
convergence to the expansions in question. If the risks are limited
(due to reinsurance or otherwise), i.e. S(z) = 1 for some M = z, the
quantities w, U and )C can be found with an inaccuracy of 1% at
most, provided:
h < O'4jM (6.15)
If I-F>O·OI the maximum relative error is 4% and for I-F
>0·001 the corresponding error is 10%.
An appropriate "trial value" to start with is:
h ~ J(5jn1X2) (6.16)
giving I-F in the neighbourhood of 0·01.
6.4. Examples of Numerical Applications

In applications there appear four variables:
P, M, e, and )C (6.17)
or, instead of P, the expected number of claims n. The calculations
are arranged giving M suitable values and the parameter h some
"trial values". For each h, M pair a set of values of P, e, and )C is
then obtained. It is convenient to represent the result graphically.
A special problem arises in obtaining a general view regarding the
interdependence of the variables (6.17) which is, in fact, a surface
in four-dimensional space. In the following an example is given of
84
manipulation by means of two-dimensional figures. The numbers

are derived from a non-industrial fire distribution applying surplus
line reinsurance.
First the variable P is fixed and the results corresponding to
different h-values are plotted on a double logarithmic scale as used
in Fig. 6.1. For different fixed P-values similar graphs are obtained.
From Fig. 6.1, x-I can be read as a function of M and e. If e
is also fixed (the dotted vertical line in the figure), x-I is obtained
as a function of M only. These values can again be arranged as in
Fig. 6.2, where the values corresponding to different P-values are
on the same vertical line.
This figure gives a survey of the interdependence of the three
variables x, M and P. If now P, )" and U are known, then x-I is:
U+)'P
x-I = - - (6.18)
P
and the corresponding maximum net retention M can be read from
the figure. In fact, however, this method is rather inconvenient
because the initial value P (and x-I too) also depends on M. To
find a set of values U, P and M corresponding to each other it
would therefore be necessary to guess a value for M, then compute
P and x-I, and finally to see from the figure whether the value
~-1
1'0
·5
·2
·1
=
r-----t--- '---~-r I ~
'001 '10 E
Fig. 6.1. P = 1000, the unit being the mean value of the gross claims
= moo·
85
RISK THEORY
x-1
1'00
·50
·40
·30
·20
'10
'05
1000 5000 10000 p
Fig. 6.2. e = 0'01, unit for P is moo.
of M obtained from these values is the same as the initial guessed

value or not. The result is soon obtained by iteration, but this
method is not the best possible for the computation of M, in fact
Fig. 6.2 has been constructed for other purposes which will be
explained later.
A more convenient way to find M is to construct a figure like
Fig. 6.1 fixing the variable n instead of the variable P. Then U can
be computed directly on the vertical scale in the figure instead of
,,-lor, if the latter is used, U can be obtained by means of (6.18).
Then U as a function of M for a fixed e (from a dotted vertical line
like that in Fig. 6.1) can be plotted in a figure like Fig. 6.3. From
it the value of M can be read immediately if U is known.
Returning now to Fig. 6.2 a striking feature of this graph is the
near linearity of the M-curves, the slope of the lines being -1. The
same feature is also found in graphs for e - values of 10-3 and
10-' - and for numerous other distribution functions for S(z) of
different types. Expressing this analytically in terms of the double-
logarithmic scale, it follows that:
log (,,-1) = -llog P+log g(e, M)

or
(,,-l).jP = gee, M) (6.19)
86
Here an auxiliary function gee, M) is introduced which shows the
position of the straight lines in question. The function g depends
on the function S(z). It can be calculated indirectly from Esscher's
formula by constructing a graph similar to that in Fig. 6.2. This
gives g = (x -1)J P where x -1 and P are values read from the
straight line in question. If the relation (6.19) were exact, the values
of g obtained would always be the same irrespective of which point
of the line is used to obtain the pair of values x-I, P. In reality a
small divergence is observed if different P-values are used, corre-
sponding to the inaccuracy of the approximation, and owing to the
lines not being exactly straight. In particular if P is low (less than
100, when moo is the unit) and M large the results are less
satisfactory (Pentikiiinen, 1952).
It should be noted that the results outlined above have been
obtained by means of numerous examples computed from actual
claim distributions S(z). Whether they are valid or the extent to
which they are generally true has not yet been established. The
method itself is however of interest and an actuary studying the
risk distributions of his own company may experiment to see
whether similar observations can be obtained from his own actual
1000 U
10~----------~----r----+--~~~rr~r-~
5 10 20 304050 100
Fig. 6.3. U = U(M) for n = 1000, e = 0'01, and'\ = 0·05.
Unit £1 000.
87
RISK THEORY
statistics. In any case graphs as shown above are very convenient for
giving a broad survey of the interdependence of the variables.
It is a very interesting fact that (6.19) is of exactly the same
form as (5.4) obtained by means of the normal approximation.
Proceeding by (6.18) and using the variables U, A, and P instead
of K, (6.19) leads to:
U= g(e, M)JP-AP (6.19')
whilst on the other hand from (5.4) it follows that:
U= y Jm JP-AP (6.19")
Comparing the coefficients and using (5.7)
g(e, M) = y Jm = yJ(KM) (6.20)
Both g and YIX./Jm for a fire insurance distribution S(z) are plotted
in Fig. 6.4. The slight divergence of the lines indicates the difference
between the results of the normal and the Esscher's approximations.
This difference could be reduced if (5.4') were taken as a basis for
comparison, but then the variable n would complicate the picture.
20
10
2 3 45 10 2030 SO 100 200 500 M
Fig. 6.4. Fire I, g = g (10- 1 , M). The dotted line indica.tes the va.lues
computed by (6.20).
88
Fig. 6.5. g = g (10-', M).
The function g represents in a concise way the risk theoretic

characteristics of the insurance portfolio and it may therefore be
called a characteristic curve. The higher this curve is, so the more
dangerous, i.e. less stable, is the risk business. If the g-curve is
known, then by means of (6.19) or (6.19') different values of U, M,
P, and ,\ can be approximately computed.
The characteristic curves make it possible to compare different
risk distributions. For this purpose several functions are drawn in
Fig. 6.5.
The nature of the portfolios is indicated by the labels of the curves.
Fire I is a mixture of large and small risks. Fire II contains mostly
small risks and fire III mostly very big risks. Fire IV was obtained
by combining fire II and fire III in the proportions 4:1. Accident I
contains industrial accident risks of large enterprises. The portfolio
accident II chiefly contains rural employees divided into very small
units, mostly consisting of only one employee. The number of claims
in this type of insurance is large and the risk sums very homogeneous.
The life insurance distribution was extraordinarily heterogeneous,
since inflation had caused the addition to the old portfolio of new
policies whose average sum assured was nearly 10 times larger than
that of the old policies. Moreover the statistics contained a number
of claims for which the inflation of the First World War had so
reduced the real value that it was only about 1% of its original
figure. The slight irregular bend in this curve might be due to the
fact that the statistics contain only a few large risk sums and these
89
RISK THEORY
lie in an isolated position far above the ordinary mass. Marine I is a

portfolio restricted to cargo risks. Marine II contains only hull risks.
In the latter case the statistics did not contain any very high
claims, which might be a reason for the unexpectedly flat form of
the corresponding curve.
In Fig 6.5 the curves have a joint tangent This straight line
obtained "empirically" from the computed curves is:
g = 1·9-/M (e = 0·01 and unit for M is moo) (6.21)
This straight line can again be used as an approximation independent

of the distribution 8(z). For values 1 <M <50 most of the g-curves
seem to fit fairly well with this line. It is striking that the same
approximation (5.12) found by means of the normal approximation
has been obtained again, even including the numerical value of the
constant.
Because the approximation (6.21) is independent of the distribu-
tion function 8(z), it is reasonable to expect that it is equally inde-
pendent of the type of reinsurance (surplus, quota share or excess of
loss or some combination of these) provided, however, that the reinsur-
ance is arranged in some proportional way (each risk unit is reinsured
individually). Changing the reinsurance system to another only
changes the function 8M(Z). It should be remembered that 8M
tends to S as M -+ 00 for any reinsurance system. Consequently, as
M tends to infinity, the g-curves corresponding to different methods
of reinsurance have a joint asymptote, the line g = g (e, 00).
It must again be emphasized that approximations like (6.21) are
often very rough and in many special cases appear to give high
values of U, for example. Therefore, it is always desirable to carry
out all calculations by means of more accurate formulae if this can
be done. If rough approximations are needed or when the distribution
function 8(z) is not known or, for example, when some "safe"
universal formula for all companies jointly is needed, then these
formulae can be used. An example of the use in the latter case has
been briefly discussed in connection with the question oflegal reserve
funds in Section 5.4.
Exercise 6.4.1. Compute the curves presented in Figs. 6.3 and 6.4
for the distribution mentioned in Exercise 3.5.4.2.
90
CHAPTER 7
Monte Carlo Method
7.1. Random Numbers

The Monte Carlo technique makes use of so-called random numbers
which can be defined as follows. Consider a random variable p,
which is equally distributed over the interval (0, 1). Its distribution
function is:
0 for r:$;O
R(r) = ( rfor O<r~ 1
1 for r > 1
and the distribution is called rectangular owing to the shape of its
frequency function. A sample of 8 observed values of p is assumed
to be taken in some way or another, for example by means of a
lottery, and so a sequence of numbers (rl' r 2 , • • • , rs) is obtained.
These numbers are called random numbers. They are like the list
of winning numbers in a lottery: each number in the interval
(rounded to a certain number of decimals) has exactly the same
probability of appearing in the sequence.
As a rule the Monte Carlo method needs such a large number
of random numbers that these have to be generated by electronic
computers. Ready-made programmes already exist for the purpose.
Roughly speaking, a computer takes about the same time to
generate a random number as it takes for a multiplication.
From the rectangularly distributed random numbers rf, sequencies
of new random numbers which are samples of a random variable
obeying any other given distribution function F(x) can easily be
obtained. Suppose a distribution function F(x) is given. Each
random number rf is reflected by means of this distribution function
as is seen in Fig. 7.l. The result x( is the root of the equation
rt = F(xt). In other words the procedure transforms a random
sample of the distribution R(r) into a random sample of the
distribution F(x).
91
RISK THEORY
F(X)11-_ _ _ _ _ _ _ _ _ -=:;:;:;:o_--
x
Fig. 7.1.
To illustrate the simulation technique, which is also called the

Monte Carlo method, the calculation of the convolution of two
distribution functions F I{X) and F 2{X) is considered, i.e. to find
the distribution function of the sum g = g1+ g2 which is denoted
by F = F I * F 2. To do this s pairs of random numbers (rll' r 2l ;
'12' r 22 ; r 13 , '23 ... , riS' r2S) are produced. Each number rlt is now
transformed as mentioned above to another number Xu being a
random value of the variable gl' In the same way the numbers
'2i are transformed by means of the function F 2 to numbers X 2i.
Putting x, = Xli+X2i a sequence (Xl' X 2 , ••• , xs) is obtained. This
is a random sample distributed in accordance with the function
F = F I * F 2' To obtain the numerical values of this distribution
function, or more precisely, an estimate thereof, the numbers Xi
are rearranged according to their magnitudes and, denoting by
kz the number of all xc;;;x, the estimate F{x) = kz/s immediately
follows.
It will thus be seen that the simulation method is, in fact, exactly
equivalent to a method of observing the actual values appearing
in some experiment and building up the statistical estimate of the
distribution function. No physical experiment is in reality carried
out but instead it is 'played' or 'simulated' by means of random
numbers. This method has been increasingly used in connection
with various research projects during recent years, particularly in
cases where the direct calculation of the distribution functions is
92
MONTE CARLO METHOD
too complicated. It can also be usefully applied in the field of the

theory of risk as will be shown in Section 7.2.
The error depends on the number s. Suppose that the real value
of F(x) is p, then the distribution function of kz/s is a binomial
distribution of which the standard deviation is:
a = J(P(I ;P) ) ~ ~ J(~ )

It then follows that it can be said with 99% confidence that the
observed value will lie within ± 2·576 a from p. Table 7.1 has been
calculated at this level to show the maximum error.
TABLE 7.1
Maximum
8 error
100 0·129
1000 0·0408
10 000 0·0129
100000 0·00408
1000000 0·00129
Thus to get one more decimal place the work must be multiplied
a hundred fold. The table gives a kind of maximum error near the
mean point but in practice one may expect the error to be less
than one-half the maximum. For practical applications some
10 000 to 100000 sample values Xi are needed. If x is such that
F(x) = 0·0l, say, then the 'maximum' error is only 20% of the
values in the above table, and for F(x) = 0·001, about 6%.
In many applications of risk theory, values of I-F of the order
0·01 to 0·001 are of most interest. Thus it is also useful to estimate
the relative maximum error:
2·576. ~= 2·576 J(l;sP)

Examples for two values of p are given in Table 7.2. If a high
relative accuracy is required, then the necessary number of random
numbers grows large and the method soon becomes inapplicable.
93
RISK THEORY
TABLE 7.2. Relative maximum error 100.2'576 alp
8 P = 10- 2 P = 10- 3
100 256 814

1000 81 257
10000 26 81
100000 8 26
1000000 3 8
On the other hand, if only the order of magnitude is needed, which

is often the case, and especially if 1- F is about 0·01, then the
method can still be convenient.
7.2. Simulation of Generalized Poisson Function

For simulation of the generalized Poisson function a random
number corresponding to the number of claims has to be generated
for each trial, and another for the corresponding claim amounts.
In principle it is possible to proceed as follows: Let v be the random
number of claims, , the random size of one claim. The distribution
functions of these two random variables are:
2: k!
N
nk
peN) = e- n - and S(z)
o
A random number, say riO, is generated. Let Nl be the solution

of the equation rIO = peN), or in other words, denoting by p-l
the inverse function of peN), Nl = P-I(r IO )' Then Nl random
numbers rw r 12 , ••• , r lN ! are generated, and numbers zli=S-I(r li )
are derived. Evidently:
can be considered as the first member of a sample of a random

variable the distribution function of which is F(x), i.e. the first total
annual amount of claims of the simulated company. The procedure
is repeated say 10 000 times. Unfortunately this procedure requires
on the average 10000 (n+l) random numbers or in many cases
94
MONTE CARLO METHOD
a total of more than 10 8, which is beyond reasonable computing

costs.
This procedure can be rationalized if some convolutions are
calculated before the simulations. If an integer N is written as a
sum of integers hi, >0, then evidently:
In particular, if N is written as a binary number
the representation is:

EN * = TI* E2k *
ak=l
Therefore, as 2k = 2k - l +2 k - l , convolutions E2k* are calculated

beforehand by means of the recursion formula:
Eak * = Eak - I ** E2k - l *
up to a sufficiently large k. The subsequent procedure is as follows.
A random number rIO is generated and the claim number Nl
derived from Nl = p-l(rlO)' Then Nl is expressed as a binary
number L ak 2k , and random numbers r 11 , r 12 , . . . , rn::a k are
generated. Finally the numbers:
Zlk = (E2 k*)-1(r li )
are derived for those values of k for which ak = 1, starting from the
smallest k. Then evidently:
gives the first sample point of the distribution F(x). If 10 000

sample points are required, then only 10 000(llog2n+ 1) random
numbers on average are now needed. The total of 10 8 in the first
method is thus reduced to 80 000, so reducing the time by a factor
of about 10- 3 •
7.3. Discussion on the Accuracy and a Modification

The error estimation described in Section 7.1 remains the same if
convolutionsE2k *are computed correctly. This is, however, extremely
95
RISK THEORY
laborious if k is relatively large and 8 has a long tail with small but
positive 1-8(z). In order to clarify the ideas, consider as an example
the case where n = 10 000, and:
8(x) ~ 1-10- 5 for x < 10- 1M

= 1-10- 5 for 1O-1M~x<M
= 1 for M~x
This corresponds to a case where there is one catastrophic claim

M once in 10 years on the average. If in integration the function
8 can only be taken as a step function with steps of 10- 4, say, as
is the case in most rapidly convergent convolution calculation
schemes, the catastrophic claim does not enter into the convolutions.
The error in F(x) can thus be of magnitude 10- 1, which ruins the
result.
This new error factor motivates another modification of the
method. This can be done with the help of the procedure described
in Section 3.6. If 8 is represented in the form 8 = (1-p)8 1 +p82 ,
where:
8(x)
- - for x<xo
{ I-p
1 for x~xo
I-p
then, by choosing p small enough, the tail of 8 2 becomes manageable

and the tail of 8 1 can, for practical purposes, be taken as non-
existent. According to (3.23) the function F(x) can be partitioned
into the form F 1 * F 2' where F 1 has mean claim number (1- p)n
and claim distribution 8 1 , whereas F 2 has mean claim number pn
and claim distribution 8 2• The distributions F 1 and F 2 can both
be calculated separately by means of the Monte Carlo method
96
MONTE CARLO METHOD
described above. Since the convolution F 1 * F 2 can also be calculated

by the Monte Carlo method, it is advisable to combine this with
the other simulations, as follows:
F 1-sample F 2-sample F-sample
xn x 12 xn +X 12
x 21 X 22 x 21 +X 22
......
X S1 X S2 Xu +XS2
so that the final convolution F 1 * F 2 is reduced merely to 8 additions.

If it so happens that the tails still cause difficulties, it is of course
possible to extend the partition of S into more than two components.
When all convolutions have been calculated, the error estimation
still remains as in Section 7.1, so that in practical calculations a
sample of 10 000 to 100000 simulation years gives an average
accuracy of 0·003 to 0·001 near the mean value and rather better
in the tails.
The Monte Carlo method is at its best if n is small. If very great
accuracy is required, e.g. if the error is not to exceed 0·0001, the
method becomes expensive. Fortunately, from a practical point of
view, such great accuracy is seldom required.
It may be noted that the Monte Carlo method can also be applied
if P(N) is not a Poisson function, provided that it is obtained in
some way or other in a form which is suitable for computer input.
More details can be found in Hovinen (1964 and 1967).
97
CHAPTER 8
Other Methods of Calculating the

Generalized Poisson Function
S.l. Inversion of the Characteristic Function

According to a general theorem of Fourier transforms the transform
(3.10) F -+ tP has an inverse transform, which gives F uniquely,
provided there is a definite convention as to how F(x-), F(x),
F(x+) are defined at points of discontinuity as is the case for
probability distributions. The result is given without proof. At
continuity points:
1 +fT l-e- i8x 1

F(x) = -lim . tP(s)ds +- F(O) (8.1)
2rrT~00 u 2
-T
Thus, if nand S are known, it is sufficient to calculate
00
"'(s) = J e iBXdS(x), and thereafter tP(s)
according to (3.11). Finally
-00
F(x) is calculated from (8.1). Though apparently simple, this method
is extremely laborious, as is seen by writing (8.1) as an integral in
real terms. Further it is not easy to estimate how large T should be
chosen so as to get sufficiently accurate results. Bohman (1964)
has successfully attacked the difficulties arising from the choice of
T by substituting for (8.1) two other inverse Fourier transforms
for which the integrands vanish outside the interval (- T, + T),
and which give an upper and a lower limit to F(x), thus allowing
estimation of the error.
Though laborious, this method has been valuable as a means by
which approximate methods can be tested for some sample values
of the parameter n and for different kinds of S-functions. If S is
approximated as a sum of exponent functions and step functions
(which type of sum with a convenient choice of parameters repre-
98
CALCULATING THE GENERALIZED POISSON FUNCTION
sents a wide range of different kinds of S-functions) the calculation

of cp(s) and the above integrand is not hopelessly laborious, and
test calculations have been successfully carried out.
8.2. A Modification of the Esscher Method

The Esscher method emerges from (6.4) by integration from x to
+ 00 (or if x<nm, from - 00 to x). The constant h was chosen
such that x = nf31 so as to obtain a good approximation for dF(x)
at the point x for which F(x) is to be calculated. This approach
was motivated by the fact that the integrand becomes very small
as soon as the integration variable deviates considerably from x.
A natural modification is obtained if the constant h is calculated
separately for each value of the integration variable before integra.
tion. Then h becomes dependent on the integrand variable as well
as on all the f3k's andck's. If, further, his chosen as the new integration
variable, the result becomes (including the terms expressed explicitly
in the expansion (4.3)):
J(2:)
hi
F(Xl)-F(X2) = e- n [1 exp[ -hnf3l(h)

+nf3o(h)) J(f32(h)){1 +A(h) - B(h)}dk (8.2)
where
f zkehZdS(z); A(h)=f3,(h)/Snf322(h); B(h)=5f3s2(h)/24nf32S(h);

00
f3k(h)=
o
Xl = nf3l(h l ) and X2 = nf3l(h 2)· The detailed proof can be found
in Pesonen (1965).
8.3. Step Function Approximation of S(z)

As stated in Section 3.6 the decomposition (3.24) of S gives the
function F(x) as a convolution of several generalized Poisson
functions (3.23), where each function Fk has the corresponding
component function Sk(Z) as its claim size distribution. This result
can be used with advantage if the number of components is not
great and if each component function Fk(Z) is easily computed.
The simplest case is where each component Sk(Z) consists of a
single step, i.e. Sk(Z) = e(z-zk). Then the distribution Fk(x) is an
ordinary Poisson distribution, which is easily calculated.
99
RISK THEORY
The given distribution S(z) is therefore first approximated by

using a step function and the convolution between corresponding
Poisson functions is calculated. Intuitively it may be expected that
if the step points are appropriately chosen, some ten or fifteen
components should suffice for most applications.
The efficiency of this method depends heavily on the effectiveness
of the device for calculating the convolutions.
8.4. Exponent Polynomials

Another example of the decomposition technique is to choose each
component functionSk as an exponential distribution. The component
functions Fk(X) are given in (3.19) and, as stated in Section 3.5.6,
J
for small values of (nax), the calculation of Fk(X) is easily done
directly from the double series obtained by integration. For greater
values of J(nax), the use of approximation (3.21) is possible.
This technique requires that the original distribution S(z) can
be approximated sufficiently closely by an exponent polynomial.
It can be proved that this is always possible to any degree of
accuracy, but the number of components required can become
quite large particularly if there are steps in the original distribution.
A very good method is a combination of the above and the step
function approximation method. The actual steps of the original
distribution are taken as steps of the step function in Section 8.3,
and the remaining part of the distribution is approximated by an
exponent polynomial. Generally two or three exponent components
are sufficient to give a very good approximation. Finally F is
written as a convolution of two kinds of Fk-functions, i.e. ordinary
Poisson functions and modified Bessel functions.
8.5. Mixed Methods

The decomposition of S can be done in various ways and a convenient
method may be used for each component. One possibility is to
divide S into components according to the size class of claim.
If the largest class consisting of ordinary claims (~Zl) is taken
as one component, it can be expected that in many cases the normal
approximation will suffice, since the failure of the normal approxi-
mation in practice largely arises from the long tail of the S-distri-
bution so soon as the number of claims is sufficiently large, say a
100
CALCULATING THE GE~ERALIZED POISSON FUNCTION
Fig. 8.1.
few thousands. The N P-method is probably more appropriate as is

the Esscher method but both require more calculation.
The tail which is left outside of the large class of ordinary claims
can then be divided into two or more components. Since each class
derived in this way generally includes only a small number of
claims, one possible method is to use a Monte Carlo method for
each component, or, perhaps, a combination of step function
approximation and exponential polynomial approximation.
8.6. Statistical Method

Equations (4.2) and (4.7) give the moments of F(x) about the
mean in terms of the moments of Sex) about zero, i.e.:
Mean = nm
a2 = noc 2
1'1 = ocs/ocS.Jn
1'2 = oc4/noc 4
Since for the normal curve 1'1 = 1'2 = 0, it is readily seen that the
approach of F(x) to the normal distribution depends on the expected
number of claims n and on the magnitude of the factors OC3/OCS
and OC4/ oc 4•
Instead of expanding F(x) in an Edgeworth expansion, an
alternative approach is to assume an analytical form for F(x)
101
RISK THEORY
and to estimate its parameters from the moments calculated from

S(x) and n. If the Pearson system of curves is adopted then the
use of two moments gives the normal approximation described
earlier. The use of three moments implies that the assumed distri-
bution of F(x) is the Pearson type III (gamma distribution) and
the use of four moments implies a distribution of type I (beta
distribution). The available tabulations of the various Pearson
distributions can be used when the parameters have been found.
For some applications the table of values of standardized
deviates at various probability levels for given values of 1'1 and 1'2
(Biometrika, Vol. 50) may be sufficient for practical purposes,
and may afford a rapid method of approximation since the effect
of the factors 1'1 and 1'2 can be easily assessed. The utility of this
method would be materially extended if the tables referred to were
extended for values of P below 0·0025, the lowest figure for which
values are tabulated.
102
CHAPTER 9
Variance as a Measure of Stability
9.1. Optimum Form of Reinsurance

In Chapter 5 criteria were developed based on the assumption that
the normal approximation is a suitable approximation in real
circumstances. A closely related method is to measure the stability
of the yearly results by the variance of the total amount of claims:
I (x_E{g})2 dF(x)
<1J
V{g} = E{(g-E{g})2} = (9.1)

o
This has been explicitly calculated in Section 3.3 in the case of
the generalized Poisson distribution, i.e.:
(9.2)
However, in the present approach it is not necessary to make
the assumptions that the Poisson or normal properties apply,
provided (9.2) is not used, and it opens the way to some general
risk theoretical questions of a rather broad nature. If (9.2) is used,
then the generalized Poisson properties are assumed but not
necessarily the normal ones.
An example of the philosophy mentioned above is to regard an
optimum reinsurance arrangement as one which makes V{g} a
minimum when other conditions remain constant. This is equivalent
to minimizing the ruin probability at least in cases where F is
normal. The following example, the proof of which is due to P. M.
Kahn, deals with the best form of reinsurance in the above sense.
Consider the following problem: a company wishes to arrange a
reinsurance at a prefixed net premium P which is of such special
type that the reinsurer pays the sum g- RW if the total amount
of claims during the period in question is g, subject only to the
condition:
o ~ RW ~ g (9.3)
103
RISK THEORY
where R(x) is uniquely defined for each x. The question is how to

choose the function R(x) so that the variance V R = V{RW} is
minimal, assuming the reinsurance premium to be a fair net premium
without any safety margin and subject to the condition:
P = E{g-R(O} = some fixed constant
It should first be noted that stop loss reinsurance satisfies the
requirements of the special type of reinsurance arrangements
under consideration and that among all possible stop loss re-
insurance arrangements is a uniquely defined one which satisfies
the conditions prescribed above and where:
R*(x) = { x for x ~ M (9.4)

Mforx ~ M
where M is the well-defined unique solution of the equation:
f (x-M)dF(x)
00
P =
M
It is asserted that the reinsurance R*(x) is a solution of the
problem. To prove this it is merely necessary to calculate V R*
and V R and verify that V R* ~ V R for all R. Let P l = E{g}-P.
Then:
f R2(X) dF(x)
00
V R+P1 2 =
o
f [R(x)-M]2 dF(x)+2MP -M2

00
1
o
f [R(x)-M]2 dF{x)+2MP -M2

M
~ 1
o
f (X-M)2 dF(x)+2MP -M2

M
~ l
o
f (R*(x)-M)2 dF(x)+2MP -M2

00
l
o
VARIANCE AS A MEASURE OF STABILITY
which proves the assertion. This assertion has been proved under
the assumption that reinsurance is postulated to be of function
type R(x). The theorem is general however, because it is possible
to prove that for an arbitrary reinsurance there is always another
reinsurance of the special function type, which has the same net
premium but gives smaller variances for both cedant and reinsurer
(Pesonen, 1967).
This result shows that the 8tOp l088 reinsurance (9.4) i8 the optimal
80lution from all reinsurance method8 in the 8ense that it give8 for a
fixed reinsurance net premium P the 8mallest variance V R(X) of the
companY'8 net retention. In other words: if a level of V R(x) = Vt is
fixed corresponding to some (estimated) ruin probability e, the stop
loss reinsurance leaves the maximum premium income to the
ceding company and thus minimizes the reinsurance net premium.
In practice, however, there can be heavy safety loadings in stop
loss premiums, because this form of reinsurance gives a maximum
relative variance to the reinsurer. Another problem is therefore
considered. Again let g be the total amount of claims, 7] the retained
part thereof, and g-7] = , the reinsurer's share. Suppose that the
reinsurance premium is equal to:
(9.5)
where f is supposed to be a non-decreasing function. The net

reinsurance cost is evidently f(Vg}) on the average. Suppose
further that the cedant wishes to retain a variance of fixed size,
say Vb} = V. The problem is to determine how the form of
reinsurance should be chosen if reinsurance costs are to be as low
as possible. In other words, which reinsurance arrangement
minimizes Vg} if V{7]} is fixed. Now:
Vg} = V{g} + VbJ-2(E{g7]}-E{g} E{7]})

From this expression it can be concluded, since V{g} and V{7]}
are constants, that the reinsurance cost attains its minimum on the
average if the correlation coefficient of the random variables g
and 7], namely,
E{g7]}-E{g} Eb}
~(V{g}V)
reaches its maximum value + 1. As is well known from probability

105
RISK THEORY
calculus this is the case if 1] = a" where the positive constant a

is defined from V = V{7}}, giving:
(9.6)
Thus a reinsurance of proportional form gives the desired result.

This theorem is also quite general. Hence, if the reinsurance premium
increases with the reinsurer's variance, and is thus of the form (9.5),
the most inexpensive way to reach a given variance is to use the
reinsurance form (9.6).
9.2. Reciprocity of Two Companies
'1 '2
The problem to be considered arises when two companies 0 1 and
O2 , whose total amounts of claims and are supposed to be
independent, wish to exchange reinsurance on a reciprocal basis
and desire to find an optimum method which satisfies the two
conditions:
(i) The expected profit on the exchange must be zero.
(ii) The variance of the net retained business after the exchange
must be minimized as far as possible for the two companies.
In order to fulfil the requirement (i) it is simply assumed that
the reinsurance premiums are net premiums, i.e. of form (9.5)
with f = O. If then condition (ii) is the only remaining decision
criterion, it is easily seen that the exchange should be of pro-
portional type, i.e. the final total amounts should be:
for 0 1 a1'1 +(I-a 2)'2

for O2 (I-a1)'1 +a 2'2
To prove this suppose that the final arrangement has given the
variance VIR to company 0 1 , This can be written V 11 + V 21> where
'1 '2
V11 originates from 01'S remaining original business and V 21 from
accepted reinsurance, since the amounts and were assumed
to be independent of each other. Analogiously the final variance
of O2 can be written into the form V 12 + V 22' Suppose now that the
reinsurance 0 1 ~ O2 was not of proportional form. Then without
changing the other variances the variance V 12 of O2 could be
reduced by means of reinsurance (9.6). Similarly without changing
106
Fig. 9.1.
other variances the optimal variance V 21 is reached by using a

reinsurance of form (9.6).
The remaining question is to determine how the constants a 1 and
a 2 must be chosen. Evidently a necessary condition for agreement
is that. VIR ~ VI and V 2R ~ V 2, where VI and V 2 are original
variances, that is to say:
aI2Vl+(I-a2)2V2 ~ VI
a22V2+(I-al)2Vl ~ V2
Geometrically this implies that the point (aI' a 2) should be
situated in the common area of two ellipses, one of them having
the midpoint (0, 1) and the principal axes 1 and .j(V II V 2)' the
other the midpoint (1, 0) and the principal axes .j(V2IVl) and 1.
The relatively best agreement can, however, be specified slightly
more closely. Suppose that the final variance VIR is found. This
can be realised by means of different values of a 1 and a 2• Mter
some calculation it will be found that for fixed VIR the minimum
of V 2R is reached if a 1 +a 2 = 1. This condition gives a straight
line which goes through the midpoints of the ellipses. There is
thus a conflicting situation. Company 0 1 prefers to go as near the
107
RISK THEORY
midpoint (0, 1) as possible, the other company prefers points near

the midpoint (I, 0). A compromise has to be found somewhere in
that part of this straight line which lies inside both ellipses.
Suppose now that the companies have agreed to meet in addition
to the conditions (i) and (ii) yet a third one:
(iii) The volume of exchange must be balanced.
That is to say, the reinsurance premium from 0 1 ---+ O2 must be
equal to the reinsurance premium from O2 ---+ al' This new agreement
gives rise to a conflicting situation of a new kind. Evidently both
companies will cede more reinsurance of stop loss type and accept
more business with small variance. If it is assumed that as a
compromise a proportional arrangement has emerged, then the
constants a 1 and a 2 must satisfy in addition to the above inequalities
the condition:
(l-a 1 )P 1 = (l-a 2)P 2
where PI = E{gl} and P 2 = E{g2}' The graph of this equation is a
straight line, which goes through the point (I, I) and which has
common points with the area inside both ellipses. The optimal
point for 0 1 is given by:
() V2P 12
a1 1 =
V 1 P 22+ V 2P 1 2
and the optimal point O2 by:
a (2) _ P I V 2 a (0
1 - P V 1
2 1
Generally there is still a conflicting situation, and a compromise

must be found somewhere between these two values, still remaining,
of course, inside both ellipses. Only in one case are the interests
of both parties completely parallel (after the compromise which
led to the proportional form), namely in the case where:
P 1 /P 2 = V 1/V 2
In this case al = P 1/(Pl +P 2) and a2 = P 2/(Pl +P 2).
Exercise 9.2.1. There are r insurance companies having the same
distribution function of the total amount of claims ~i, the claims
of each company being independent of those of the others. The
companies want to find a reciprocal exchange of business so that
108
each company i pays the amount R(gj) of the claims of each other
company j, and the standard deviation of the claims on each
company's own retention (including the amounts of the other
companies' claims received) must be minimized. Prove that the
function R(x) = xlr is the required solution.
Exercise 9.2.2. A main concept in the classical theory of risk was
the so-called relative mean risk p, which was defined as follows:
V and P being the variance and net premium income respectively.

P was used as a measure of the stability of the insurance collective
concerned.
(i) What is p = Po for a collective having equal risk sums?
(ii) Prove that Po is the minimum for all collectives having the
same P and n, if the generalized Poisson distribution is
assumed.
(iii) Rewrite the basic equations (5.4) making use of p assuming
the generalized Poisson distribution and its normal approxi-
mation.
109
CHAPTER 10
Varying Basic Probabilities
10.1. Introduction
In the preceding chapters it has been assumed, with minor exceptions,
that the basic probability remains constant so that the number of
claims each year has the same expected value n. Furthermore, it has
been assumed that the underlying processes obeyed the laws (i) to
(iii) of Chapter 2, i.e. the process is assumed to be a Poisson process,
elementary or generalized, depending on whether the amount of
one claim is a constant or a random variable. In addition it has been
assumed that the distribution function S(z) of the amount of one
claim is stable or at least independent of the number of claims.
These assumptions are obviously an idealization of the practical
situation.
In fact, the assumption of constancy of the expected number of
claims n (implying that the basic probabilities of claims are constant)
and the assumption of constancy in the safety loading ,\ (which is
of marked importance in long term processes) are in contradiction
with practical experience in many classes of insurance. Owing to
variation in the basic probability and in the value of ,\ some of the
formulae derived on the simpler assumptions generally lead to unduly
optimistic values for the insolvency criterion and similar character-
istics. This may be readily seen intuitively since if some parameter
is assumed to be a constant whereas in fact it is a variable, then
values obtained for the variance of risk are too small and thus values
for derived functions, e.g. insolvency characteristics, are also too
small. Furthermore, in the e-R U technique, which will be considered
in Chapter 12, although some attention is given to these possibilities
by the assumption of an upper limit to the insolvency function which
allows for some changes in the basic probabilities, unsatisfactory
results will be obtained, if, for example, the safety loading occasion-
ally becomes negative.
110
VAR YING BASIC PROBABILITIES
However, in spite of these difficulties the preceding results are far

from useless and, provided attention is given to possible changes of
risk during the period of observation, they can be applied to practical
problems. Some problems, particularly those concerning short time
periods such as one year, may be adequately dealt with by the
foregoing techniques, although even here a strong element of con·
tagion can unduly influence the calculation. On the other hand there
'are some problems, particularly those involving the calculation of
reserve requirements, where the simpler theory is inadequate and the
important role played by the changes in basic probability makes it
necessary to adopt a more complex model. It is therefore necessary
to extend the calculations beyond those based on the idealized
assumptions.
To elaborate the theoretical model it is important to analyse the
different types of variation of the basic probabilities. This can be
conveniently done in the following three categories:
(i) Trend8 - These emerge as a slow moving change of the claim

probabilities which must be properly defined as the ratio of the
numbers of claims in a specified time interval to an appropriate
index of the number of risks exposed to the chance of a claim.
Examples are the improvement in mortality rates or the changes
in frequency of fire due to changes in methods of building
construction or changes in materials used. Trends in the overall
experience from an insurance portfolio can also arise from
changes in its constitution, for example, the proportion of
newer type houses may increase in relationship to older houses
so that the overall frequency will reflect any differences between
the two groups.
(ii) Long period o8cillations - An example of this type of variation
is provided by the association of fire claims with general
economic conditions. The results of consecutive years are not
mutually independent and an oscillation can thus have a length
of several years. If, for some reason, such as the structure of
the insurance market, a time lag exists between the emergence
of an increased frequency and the correction of the premiums,
the safety loading ,\ can rapidly become negative with very
important repercussions on the business. (The USA market
illustrates these points very well.)
111
RISK THEORY
(iii) Short period o8cillations - These can be caused for example

by meteorological changes or from epidemic diseases. Thus a
long dry summer is almost certain to give rise to a marked
increase in the frequency of fires and the incidence of severe
wind storms in certain parts of the world has a pronounced
effect on the results of non-life business. Epidemic diseases, in
addition to affecting the results of sickness insurance, are also
of significance in life insurance, giving rise to fluctuation in the'
mortality experience.
In addition to the three main types mentioned above there are, of

course, seasonal fluctuations of which the variation in claims fre-
quencies between summer and winter months in motor insurance
may be instanced. These can, however, generally be disregarded when
consideration is being given to results based on calendar years.
The simplest extention of the theory represented in Chapters 1 to
9 is developed to deal with variations of type (iii), where it can be
assumed that the variations between years are independent. How-
ever the effects of types (i) and (ii) are also of significance for long
term consideration and some indications of methods of making
allowance for them also are given in the following sections. In some
cases, e.g. in connection with stop loss reinsurance, the theory for
type (iii) can in certain conditions be applied to the fluctuating basic
probabilities underlying type (ii).
Figs. 10.1 and 10.2 are typical curves taken from a portfolio of
motor cycle insurances and illustrate the various types of fluctuation.
The observed data of Fig. 10.1 is the monthly claim frequency
over the years 1960 to 1962, during which period the exposures
increased from about 19000 to about 27 000 policies. The smoothed
curve is derived by a system of moving weighted averages, the
weights being (1, 2, 3, 2, 1), i.e. a double summation in 3's. The
maxima occur in the autumn months and the minima in late winter,
a reflection of the fact that when conditions are unpleasant the
motor cyclist reduces his exposure. The trend line suggests a declin-
ing tendency over the 3 years.
Fig. 10.2 gives the annual claim frequency for the period 1951 to
1963, the exposure increasing from about 7000 to 27000 over the
period and shows a long term cyclical effect, with a probable trend
upwards over the period. This longer period shows that the declining
112
90
80
8070
s;1
•~60
"-
u
c
II
i-so
J:
40
1960 1961 1962
Fig. 10.1. Motor cycles. 4.weekly claim frequency 1960-1962.
SOOI+-~--~~--~~--~~--~-,--~-,--,------
'~ 1 9~ ' %31~mS~6m7m8W~WW1~1H2~3
Fig. 10.2. Motor cycles. Annual claim frequency 1951-1963.

RISK THEORY
tendency in Fig. 10.1 was a downward phase of one of the longer

term cycles and shows the need to look at fairly long series of values.
The data exhibit: (i) a long term trend upwards; (ii) a long term
cycle; (iii) an annual cycle; (iv) random variation.
10.2. Compound Poisson Process

If the expected number of claims n during a year is constant and the
assumptions (i) to (iii) of Section 2.2 apply, then it has been shown
that the probability that exactly k claims arise during a year is:
nk
Pk = e-n - (10.1)
k!
independently of the year in question. This is, in general, not true
in practical applications, but it may happen that the parameter n
in (10.1) is an ordinary function of time (say, for example, dependent
on the absolute value of t for the 1 year interval (t, t + 1»). In this case
the process is a so-called deterministic, i.e. "predicted" process,
provided the other assumptions hold true. This is so if n = n(t)
exhibits changes solely generated by the following effects, all
assumed to be defined functionally: (a) changes in the number of
policies; (b) secular trend (type (i) p. Ill) of the claim frequency
(claims/policies); (c) oscillation of the claim frequency (types (ii)
and (iii), pp. 111-2, but estimated one way or another in advance).
Suppose that the components (a) to (c) are given, or at least can
be estimated for a reasonable number of future years. Then the
theory represented above is applicable without any modifications.
n has to be given suitable estimated values which take into account
the changes in question. The theory concerning long time periods
will be developed on this basis in Chapters 11 and 12.
The position becomes much more complicated if the "process"
n(t) is not deterministic but is instead a stochastic process. This can
be caused by random effects of some external factors, such as
weather conditions, epidemics, etc. In order to describe the collective
of these external factors the constant n is now replaced by a new
variable nx indicating the "expected" number of claims in a year.
Here X is a random variable, which represents the disturbances of
these external effects. n is still a constant. Without loss of generality
X can be normed to have a mean value equal to 1. In this way, so
114
VARYING BASIC PROBABILITIES
to speak, the expected number of claims = nX is a random variable

with mean n.
Now let U(q) = P{x;::;;q} be the distribution function of x. Then
the probability of k claims will be
:h(n) = fo00 e-nq (nq)k!k dU(q) (10.2)
The problem can also be solved in a more general way by supposing

that the disturbances depend on time t. Then X is no longer a random
variable but a random function of time, or in other words the
disturbances can be represented as a stochastic process X(t). Then
(10.2) can be replaced by a more general equation:
'hen, t) = [ e-ntq (n:~)k dU(q, t) (10.2')
as obtained under some general assumptions in Appendix A.

Now consider the distribution function G(x) of the total amount
of claims during 1 year. Since the amounts of individual claims
are still assumed to be mutually independent, by proceeding as in
the case of the generalized Poisson process, it follows from (3.5) that:
G(x) = 2:00
k=O
:h(n) Sk*(x) = 2:00(00f
k-O 0
e- nq
(k
n~) )
dU(q) Sk*(x)
= f 2:
00(00
e- nq (n~) k) Sk*(x) dU(q) = f00 Fnq(x) dU(q)
o k=O k. 0 (10.3)
where F nq(x) denotes a generalized Poisson function with mean value
nq of the number of claims. It is thus seen that G(x) is, so to speak,
the weighted mean of all possible generalized Poisson functions with
the same claim curve S(z), the weights for each value of nq being
the probability of occurrences of this expected number of claims.
If the total amounts of claims in various years are independent,
and if only the annual results are of interest, this process can be
illustrated by means of an urn or lottery model as follows. The level
of the expected number of claims is first fixed for a year by drawing
115
RISK THEORY
a lot. The lottery is arranged so that the probability of getting a

value q<X~q+dq is dU(q). Then X is fixed to be equal to q for
the whole year. The expected number of claims is nq for this year
and the actual number varies at random in the neighbourhood of
this mean value as has been seen in preceding chapters. For the
next period a new value is drawn for X and so on. This urn model
is in fact able to take into account only oscillations of type (iii)
mentioned in Section 10.1 (p. 112), but not types (i) or (ii).
The processes introduced above are called compound Poisson
processes (elementary or generalized depending on whether the
amount of one claim is constant or a random variable). If the
function U is independent of time, the process is called a compound
Poisson process in the narrow sense, otherwise the process is described
as in the wide sense. Some writers refer to compound processes as
mixed Poisson processes, but in this book the more specific term is
preferred (Haight, 1967).
It may be noted that the assumption regarding the random
fluctuations of the basic probabilities is true, for example, in cases
where all the probabilities of claims of each risk unit are changed
simultaneously owing to meteorological conditions, changes in
economic conditions, etc. Simultaneity is not, however, necessary
and it is equally proper to regard some groups only of the risk units
as exposed to changes providing, however, that S(z) is not changed.
How the physical phenomena behind the probabilities are brought
about and what kinds of phenomena exist is, of course, quite
immaterial from the point of view of risk theory. It is only
necessary to assume the existence of some function U(q) which
concerns the total amount of the number of claims. In fact the urn
model mentioned previously is only a simple way to illustrate how
(10.2) can be obtained. There are, however, other and more general
ways to introduce the same formula. An example is the Polya.
process considered later in Section 10.4, where the process allows
for changes of X as a function of time during the period considered
as due to contamination.
Further generalizations are obtained if the function S(z) is no
longer assumed to be independent of time or of other claims. Such
considerations, however, would stand beyond the scope of an
elementary book. For further studies interested readers are referred
to the papers of Philipson, Thyrion, and others. One type of time
116
dependence of the function S(z) will, however, be especially mention-

ed and is often observed. When the value of money is deteriorating
owing to inflation or other reasons from year to year, it has an
influence on the function S(z). This can be dealt with by transforming
the monetary variables to correspond to the present value of money,
as recommended earlier in Section 3.5. If longer future periods are
considered, then the eventual change in the value of money can
often be allowed for by simple methods. It emerges as a proportional
change of both premiums and the size of claims.
A difficult problem in practice is to find the distribution function
U(q). Cases may well exist where empirical data describing the
function U(q) are available. If, for example, statistics covering many
years exist for a branch of insurance showing little variation, it is
possible to estimate how far the data deviates from pure Poisson
variation and so get some idea of the character of the function U(q).
It is of course impossible to separate completely the two components,
Poisson variation and fluctuation of the basic number n, but in
every case it is possible to construct models for the purpose. Two
different methods are then available for computing the compound
function G(x). The function U(q) can be obtained from the statistics
as an empirical step function and these values used without any
other modifications or equalizations; this is dealt with in the following
Section 10.3. Another method is to seek assumptions concerning
the analytical form of U(q). An example of this method will be
given in Section 10.4.
The general formulae for the mean value P and the variance
Vof the compound Poisson function G are now derived. The moment
formula is as follows:
f xk dG(x)
<Xl
Pk = (10.4)
o
f fh(nq)dqU(q)
<Xl
=
o
117
RISK THEORY
where (hen) is the moment of the original generalized Poisson

function Fn(x). Hence (cf. Section 3.3):
J
00
PI = P = mnqdU(q) = mn = P (10.5)
o
and
J(X_P)2 dG(x)
00
V= = ~2_P2
o
JnqIX2 dU(q)+ Jn 2m 2q2 dU(q) - p2

00 00
=
o 0
=nIX2+n2m2 (1
o
q2 dU(q)-1 ) (10.6)
= nlX2+n2m2Vq
where Vq is the variance of x. The first term is the same as the
variance of the generalized Poisson function (cf. 3.9). The second
term indicates the influence of the random fluctuation of the basic
probabilities and it is always positive unless U(q) degenerates to the
one step function e(x-1). In this latter case, however, the problem
reduces to the generalized Poisson function with constant basic
probabilities. This formula proves that variation of basic proba-
bilities always increases the variance of the process.
TABLE 10.1 Relative increase r of the variance
When U(q) is
n Polya
Empirical
Sect. 10.3 k = 100 k=1000
10 0·002 0·001 0·000

100 0·024 0·011 0·001
1000 0·24 0·11 0·011
10000 2·36 1·11 0·11
U8
As an illustration Table 10.1 has been computed and sets out some
examples of the ratio of terms in (10.6):
n 2m 2 Vq m2
r= - - - = n - Vq
nat 2 at2
for an industrial fire distribution (M = 00) and assuming U(q) to

obey the values presented later in Sections 10.3 and 10.4.
It can be seen how the ordinary random fluctuation of the number
and the size of claims is predominant for small companies whereas
for large companies the position is reversed.
It is of interest to transform (10.6) into the form:
v = nm2+n(at2-m2)+n2m2Vq
= m2Vk+n Vz +n 2m 2Vq (10.7)
where Vz is the variance of the variable' indicating the size of one

claim. The first term shows the influence on the total variance of
the variance Vk of the number of claims (elementary Poisson
function), the second term gives the influence of the size of the
claims and the third term the influence of the variation of the basic
probabilities.
In the following table an example is shown of the magnitude of
these three components which control the stability of an insurance
collective.
It can be concluded from (10.7) that the relative standard
deviation .jV!P does not tend to 0 for P -+ 00 as it does in the
case of the generalized Poisson function but tends instead to the
TABLE 10.2. The components of the variance
Fire I, M = 00, U(q) as in Section 10.3. Unit [£1000)1.
n m2Vk nVz n 2 m 2 Vq
10 30 2680 6
100 300 26800 638
1000 3000 268000 63800
10000 30000 2680000 6380000
119
RISK THEORY
limit .JV q• This means that the relative deviation of the total
amount of claims (x-P)/P does not vanish when P -+ 00. In fact
the central limit theorem does not hold but G tends to a function
having the same shape as U(q). (G(x) ~ U(xlnm) for large n.) This
is proved by O. Lundberg (1964).
It is worth noting that the passage -+ U(q) for P -+ 00 is valid
by conditions stated at the beginning of this section, i.e. if the value
nq is fixed for each year and U(q) remains unchanged when P -+ 00.
In practical application "passage" P -+ 00 often means that the
considerations concern a comparison of the behaviour of the risk
theoretical quantities of small and large companies or perhaps an
individual company versus the joint business of several companies.
A passage from a small company to a large company can, in fact,
mean that the latter has several branches and sub-branches of
business and it may quite well happen that these are mutually
independent even if each of them can obey the model introduced
in this section. Then the sum, the total amount of claims, does not
tend to any U(q) but tends instead rather to the normal distribution
in accordance with the central limit theorem. In other words: if
P -+ 00 so that new independent groups are incorporated into the
100
50
20
10
5
2
1'3 1·4 1·5 1'6 1'7 1'8 1-9 2·0 2·1 .!.
P
Fig. 10.3. The generalized and compound Poisson function. Distribu-
tion non-industrial fire, M = £5 000, n = 200, U(q) as in Table 10.3.
Probability Beale for 1- F.
120
V AR YING BASIC PROBABILITIES
collective, then G -+ 1/>. In practice different mixed cases can occur,

e.g. the changes in general economic conditions may have simulta.
neous parallel effects on many branches whereas some other fluctua·
tions may have effects limited to one branch only. Hence care is
needed in thinking out whether or not different kinds of assumptions
are applicable to each actual case. Philipson (1964 and 1968) has
treated these kinds of passages.
Finally the characteristic function of the compound Poisson
function is obtained from the characteristic function of the general.
ized Poisson function (3.11) replacing n by nq and integrating
(cf. (10.3))
f
00
4>(8) = enq<jl-nq dU(q) (10.8)

o
10.3. Direct Numerical Computation of the Compound Poisson

Function
If an estimate for U(q) is obtained from certain data, it may well be
in the form of a table like 10.3, which gives an approximate variation
of the basic probabilities obtained in data from fire insurance.
TABLE 10.3. Function U(q).
100 (l-qj) +25 +20 +15 +10 +5 0 -5 -10 -15 -20 -25
100· pj""
100· .:1U 10 5 5 10 12 16 12 10 5 5 10
The compound function can now be obtained from (10.3) quite

roughly as:
G(X) = L Fnqi(x) ·Pi (10.9)
"
The advantage of this method is that a table like 10.3 can often fit
the empirical data in a better way than any analytical estimate of
U(q) and no further idealizations are needed.
Application of (10.9) requires knowledge of the generalized
Poisson function F for a number of neighbouring nqi-values. These
121
RISK THEORY
can be directly calculated at least for certain pairs of values of x, nq

and then by means of graphical interpolation the values of F can
be obtained for the required values of x and nq.
Besides the direct computation formula (10.9) the different
approximation and computation methods derived for the generalized
Poisson function can also be extended to the compound Poisson
function. So, for example, the normal approximation is available in
this case if the standard deviation of the total amount of claims is
replaced by the corresponding quantity for the compound function
obtained from (10.6), i.e. if in (4.1) the replacement:
(10.10)
is made. As to the fit of the normal approximation it is to be

noticed that the compound function (10.3) does not even approxi-
mately approach the normal function, as did the generalized Poisson
function, when n -+ 00, but instead tends to a function which has
the same shape as U(q) (cf. p. 120). For this reason and because the
fluctuation of the basic probabilities for small values of n also
increases the heterogeneity, the domain of acceptable accuracy of
the normal approximation can be expected to be narrower than in
the case of generalized Poisson function.
The NP-method is applicable to the compound Poisson function,
if in addition to the replacement (10.10) in (4.10) the skewness 'Yl is
computed from the formula:
(10.11)
The coefficients of the formula are obtained by means of similar

calculations to those applied in Section 10.2 (cf. Exercise 10.3.1).
As yet little experience is available on the accuracy of the NP-
approximation, but examples (Kauppi and Ojantakanen, 1966)
computed for the Polya case (Section 10.4) seem to show that it is
satisfactory at least if 'Yl does not exceed 2.
The approximation formula (4.9) was constructed with the aid of
the Edgeworth expansion (4.6), where the explicit terms were chosen
according to their magnitude (measured by powers of 1/Jn). In all
other compound Poisson cases, i.e. if U(q) is not the degenerate
122
distribution e(q-1), the same magnitude criterion gives another order

to the chosen terms in the Edgeworth expansion. One consequence
is, for example, that (4.10) is no longer asymptotically correct
(for n ~ 00). But since the approximation is needed for some given
finite value of n, it is not so important to know whether the approxi-
mation is asymptotically correct but rather to know if it gives a
satisfactory approximation for the actual value of n.
The Esscher formula can also be generalized to the compound
cases. The procedure is not difficult in principle, but leads to rather
more algebra than in Chapter 6. The details are not given here, only
the final result is given for the expansion corresponding to (6.7)
using the two first terms. Let ~k be defined by (6.12) and let:
f e-nq(l-It,) dU(q)
00
C=
o
Nt =C
1
f e-nq(l-Ito)niqi dU(q)
00
o
(6.7) is then modified as follows.
Define u and c 3 , instead of (6.8) and (6.10), by:
u = h.jP,2
1 !L3
c3 = -6 (j.tz)3/Z
where
P,2 = Nlf12+N2~12_NI2~12
li3 = N 1f13 +3N zPlf12 + N zP1 3- 3N IN 2f11 3- 3N12f11f12 +2N 13f113

Finally replace e- nw in (6.7) by Ce- hx . Then:
(10.12)
where h is found from the equation x = N 1~1> U and C3 being now
defined as above.
The original Esscher method requires one numerical integration
for each ~k' Now, if the integrals defining C and Nt cannot be solved
in a closed form, each of them requires one more numerical integra-
tion. Roughly speaking, the calculation work is nearly doubled.
The same argument, represented in connection with the generaliza-
123
RISK THEORY
tion of the NP-method to the compound case, holds true, of course,

also for the above Esscher expansion: it is no longer asymptotically
correct for n ~ 00. Even if this observation does not undermine the
result, it gives rise to the question as to whether it is possible to
generalize the Esscher method by replacing the Edgeworth expansion
of G by some asymptotically correct entity. This has been done in
fact by Bohman and Esscher successfully for a special case, i.e. a
Polya process (cf. Bohman and Esscher, 1964).
For the application of the Monte Carlo method (Chapter 7) the
expected number of claims n is first simulated by means of the
function U(q) and then the computation can be continued in
principle in the same way as described in Chapter 7. There will be
some complications because the convolution theorem (3.23) is no
longer applicable.
The inversion method (Section 8.1) also can be applied to the
present case.
Exercise 10.3.1. Prove (10.11).
Exercise 10.3.2. Compute the compound function G for the distribu-
tion mentioned in Exercise 3.5.4.2 and in Table 10.3 (or some other
similar distribution which is available) by means of the normal
approximation, NP-method, and Esscher formula. The results are
to be represented graphically as in Fig. 10.3.
10.4. The Polya Process

A very interesting method of treating compound Poisson processes
on the basis of an analytical distribution function U(q) arises from
the use of the incomplete r-function, which has been extensively
studied by Ammeter as a theory of fluctuating basic probabilities.
It has become rather popular because of its good behaviour in the
sense that it provides a reasonably practical extension of almost
all the results of ordinary Poisson processes.
In this method the distribution function of the relative deviation
of the basic probabilities X is assumed to be a r-distribution:
f e-
kq
U(q) = _1_ Z Zk-l dz (10.13)
r(k) 0
<Xl
where k is the so-called fluctuation constant and r(k) = Je-Z zk- 1 dz

o
124
VARYING RASIC PROBABILITIES
denotes the well-known r-function satisfying r(i + 1) = i! for

each non-negative integer i. The mean, the variance, and the third
moment are easily obtained by integration and by observing
that r(x) = (x-l)r(x-l):
E{X} = 1 (10.14)
V{X} = 11k
co
f (q_l)3 dU(q) = 21k 2
o
In Fig. 10.4 the shape of the r-distribution is shown for various
values of the parameter k. Values can be found in Pearson (1954).
The probability function (10.2) is now transformed as follows:
_
Pi(n) = f e-nq -(nq)ii! r(k)
co
o
-
1
e-kq (kq)k-l kdq
ikk co
= -n--f e-(n+k)q qk+i-l dq
r(k)i! 0
nikk r(k+i)
r(k)i! (n+k)k+i
= r(k+i) (~)k(n+k)-(k+i)
r{k)r{i+l) n n
If the meaning of the binomial coefficient (J) is extended by use
dU
dq
6 k=200
5
4
3
:2
q
1·5
Fig. 10.4. Polys.frequency functions.

125
RISK THEORY
of the r-function to cover the cases where the quantities i, j may

be non-integral, this may be written:
_
Pi(n) -_ (k+i-l)
.
~
(-
k n
-
k+n
)k(k+n )i (10.15)
As is easily verified, this reduces to the Poisson formula when

k -+ 00.
It is very interesting to note that (10.15) is exactly the same as
the well-known Polya-Eggenberger probability function representing
the number of claims in contamination processes. These are random
processes, where the probability of a further claim increases when-
ever a claim has occurred. It can be interpreted as though every
claim gives rise to other claims in the form of an increased probability.
For example, in fire insurance this can be looked upon as a spread
of the fire to neighbouring risk units or in case of sickness insurance
or life assurance as the appearance of a contaminating disease. The
constant 11k indicates the intensity of the contamination. If k= 00,
then no contamination exists.
The distribution function (10.13) can also be obtained thus,
assuming the process to have contamination properties. For this
reason, a process is generally referred to as a Polya process, if the
probability of i claims in every fixed time interval (0, t) is equal to
Pi(nt), defined by (10.15). It should be noted that in some investiga-
tions of Ammeter, for example, processes are dealt with where
the number of claims in one year is independent of the number in
another year. Even if the number of claims each year satisfies (10.15)
the process as a whole is not a Polya process but a compound
Poisson process in the wide sense of a special type. This difference
is immediately seen by considering the probability of no claims
during a period of 2 years. For Polya, this is:
po(n, 2) = po(2n) = (_~_)k

k+2n
and for the process mentioned above:
The distribution function of the total amount of claims is now

obtained from (10.3) by replacing pt(n) by (10.15).
126
From (10.5), (10.6), (10.11), and (10.14), the mean, variance, and
skewness are obtained immediately as:
p = P = nm (10.16)
the moments tXt being given by (3.1).

As might well have been expected, the mean is identical with the
Poisson case, but, as anticipated, the standard deviation is definitely
greater, thus making the chance of insolvency more likely.
Many of the preceding considerations can be applied to the Polya
process. For the normal and NP-approximations the values (10.16)
are to be substituted into (10.10) and (10.11).
For the Esscher formula (10.12) it is observed that the integrations
needed for 0 and Nt can be performed explicitly. After some calcula-
tions the following expressions are obtained:
II, is found from the equation:
knPl
x = --....:....::-..
k+n(l-po)
and Pt from (6.12).
As the compound function G(x) tends asymptotically to a function
which has the same shape as U(q) (cf. p. 120); Bohman and Esscher
have proposed to approximate G by a function which has the same
127
RISK THEORY
form. The approximating function is constructed so that it and the

function G have the three lowest moments identical. The approxima-
tion is as follows:
(10.17)
Here the constant a is:

(10.18)
The values of the integral can be obtained from calculated tables

of the incomplete r-function.
Bohman and Esscher have computed a number of examples by
means of (10.17) and by the inversion method and the fit has been
good.
The considerations for calculation of the maximum net retention
or reserve funds can also be extended to the compound Poisson
function in general or to the Polya process in particular. For example,
a figure corresponding to Fig. 5.1 can be computed by means of the
G-function. Depending on the assumptions concerning the fluctuation
function U(q) the results obtained deviate more or less from those
computed by means of the generalized Poisson function. Because
the fluctuation of the basic probabilities increases the variance, the
deviations are always "on the safe side" compared with the results
obtained in Chapter 5. For example, the values of the maximum
net retention M are somewhat smaller.
As has been seen, the Polya process is fairly convenient for
numerical applications leading to many important formulae, which
are easy or at least possible for computational use. On the other hand
the function (10.13) often requires a smoothing procedure and
some idealization of the process and it does not seem easy to estimate
the error due to this smoothing.
It will be a matter of consideration for the actuary to weigh
whether and in what way the fluctuations of the basic probabilities
are to be taken into account in actual problems. The discussion
given later in Section 12.5 should be referred to. The need to
introduce the function U(q) into the calculations can at least in
certain circumstances be reduced by making the assumptions con-
cerning the safety loading .>., the funds, etc. sufficiently cautious.
128
On the other hand if experience suggests the existence of significant

fluctuations of the basic probabilities, a stronger case exists for the
use of the techniques presented in this chapter, because neglect of
the basic fluctuations is liable to make the results too optimistic
from the point of the security of the insurance company.
The characteristic function of the Polya function is obtained from:
cp(s) = ! eisx dG(x) = [ 1-~ (~(S)-I)Tk (10.19)
where
fe
00
~(s) = i8Z dS(z) (10.20)

o
If k ---?- 00, (3.11) is obtained making use of the well-known
formula lim (1 +ajn)n = ea.
n-+oo
10.5. Application to Stop Loss Reinsurance

In stop loss reinsurance the reinsurer pays that part of the total
amount of claims which exceeds a certain amount, say A. A modifica-
tion often introduced is for the amount of the reinsurance liability
to be restricted to an amount B - A so that the payment is limited to
this if the claims exceed B. Furthermore it is common for the
reinsurer's liability to be limited to a certain percentage, the balance
being met by the cedent. The reinsurer's interests may be summarized
as follows, where x is the total claim.
Ifx<A reinsurer pays nil
(1-0) (x-A)
"
(1-0) (B-A)
"
The net reinsurance premium P s is easily expressed as follows:
P
_ s. =
1-0
f
B
(x-A)dG(x)+(B-A) f
00
dG(x)
.A B
where G(x) denotes the distribution function of the total amount

129
RISK THEORY
of claims. By integrating by parts the above expression can be

written:
B
(B-A)G(B)- JG(x)dx+(B-A)-(B-A)G(B)
A
so that the net premium becomes:

B
Ps= (I-C) J (I-G(x))dx (10.21)

A
The use of the generalized Poisson distribution F(x) instead of

G(x) would certainly give too low premiums in stop loss reinsurance.
The random fluctuation of the basic probability gives rise to a
greater variation than in the Poisson case. Furthermore there is in
general a cyclical variation with a period of several years asso-
ciated with economic conditions. This last variation cannot be
combined in ruin problems with the random fluctuation of a
stationary basic probability but in stop loss premium calculations,
provided they and the reinsurance contract extend over the period
of oscillation, it can be treated as a random effect. It is then only
smaller trends which need special attention. Thus, if there is no
secular trend, it is sufficient to assume the process is a stationary
compound Poisson process where only the function U(q) need be
calculated. When this is done, the function G(x) is available directly
from (10.9) or by means of the other methods explained above.
TABLE 10.4
100 P,IP
8
Basic fluctuation
Assumed Disregarded
1·00 7·7 4·9

1·05 5·5 2·8
1·10 3·8 1·5
1-15 2·5 0·7
1·20 1·6 0·33
1·25 1·0 0·13
130
VARYING BASIC PROBABILITIE::
To illustrate this point the U(q) function as shown in Table 10.3

has been taken and the net premium Ps has been calculated for a
stop loss cover of an industrial fire portfolio with n = 500, A = sP
(P = net premiums of the total business) and B = 00. The results
are set out in Table 10.4.
These figures show the significant effect that fluctuations in basic
probability have on the premium calculations and confirm that a
stop loss theory which does not allow for this feature is unsatis-
factory for practical use.
Ammeter (1953) and Bohman and Esscher (1964) have developed
extensions in the case of a Polya process for the stop loss premium
(with B = (0). A direct formula corresponding to the Esscher
formula can also be obtained.
131
CHAPTER 11
The Ruin Probability

During a Finite Time Period
11.1. The Ruin Function in Finite Time Periods

In Section 1.4 a brief reference was made to some important problems
arising from considerations of the financial operations of an insurance
company. In the preceding chapters an insurance business has been
considered at the end-points of certain time intervals, mainly at the
end of the year. Attention will next be given to problems where the
random fluctuations are considered during some longer time period.
The advent of electronic computers has led to considerable advances
in the finite time case and this will be considered first.
The approach to this problem will be numerical rather than
analytical. The function of interest is lJ'N( U), the probability that
at the end of at least one of the years I, 2, ... N the claims will
exceed the company's available resources, i.e. that the company
will be ruined, and a recursion formula will be developed for its
numerical evaluation.
Assume that the company reserves at the time points 0 (initially),
I, 2, ... N are 7)0' 7)1> ••• 7)N' respectively, each of these being a
random variable except for 7)0 = U which is known. The probability
sought is then:
lJ'N(U) = P{7)k;;;;O for some l;;;;k;;;;N 17)0 = U}
Let HI (x, U) be the probability that the reserves at point I are

>x given 7)0 = U and generally let Hk(x, U) be the probability of
the event:
{1]1>0, 1]2>0, .. . 1]k-l>O, 1]k>X 11]0 = U}

It is assumed that the function F(x), the distribution of the total
132
RUIN PROBABILITY DURING A FINITE TIME PERIOD
amount of claims during 1 year, remains stable and independent

of the events of preceding years. The functions Hk can then be
calculated as follows:
For an arbitrary integer s the definition gives for k> 1:
Hk(X, U) = ~P{7Jk >x\~ <7Jk-1 ~ i: }pH ~ i:

1 <7Jk-1 1, 7J1 > 0, ...
7Jk-2>0 I7JO = u}
The first term under the above summation sign is clearly H l(X, i/s)
+ O(I/s) whereas the second term is equal to Hk_1 (i/s, U)-
Hk_1((i+l)/s, U). Hence by letting s --')- 00:
f H 1(x, t)
00
Hk(X, U) = - d tHk_1(t, U) (ILl)

o
In order to use this recursion formula the starting function H 1
is needed. This is obtained directly from the definition and is
found to be:
H 1(x, U) = F{U+(I+,\)P-x) (11.2)
Finally:
(11.3)
To evaluate the functions Hk, the function F(x) is first calculated
numerically by the methods discussed in the preceding chapters
and the function Hk can be obtained by some method of numerical
integration. When k is small, this can be done manually on con-
ventional calculating machines, although with increasing k the
labour soon reaches unreasonable dimensions. However, with an
electronic computer the calculation of a limited number of "trun-
cated convolutions" such as those defined by (11.1) as well as the
numerical calculation of F(x) are feasible. Provided N is not too
large, and in practice there is no reason why it should be, the task
is quite acceptable.
A slight modification of the above approach enables the restriction
that F(x) remains' unchanged in successive years to be dispensed
with, since, if the law of change is known or can be estimated, it is
only necessary to make corresponding modifications to the first
factor of each integrand in (11.1). Of course, this procedure requires
133
RISK THEORY
the calculation of the corresponding functions Fk(X), each of them

requiring more work, but, provided N is not unreasonably large,
the task is far from hopeless.
In particular this modification enables the different variations of
the basic probabilities discussed in Chapter 10 to be taken into
account. The simplest case is where F in the above formulae is
replaced by the compound function G. But even long term oscillations
and secular trends can be allowed for, as well as changes in the S(z)
function due to future inflation, and also increases in the number n
due to the probable growth of the company.
It may be expected with the growth of automation that this
method, subject to modifications, will become of increasing im-
portance. However, it must not be forgotten that analytical methods
are, and will probably remain, more important from the theoretical
point of view, and frequently lead to very good tools for everyday
purposes. Furthermore, the accuracy required for a particular
application may not justify relatively heavy computing programmes.
For example, the e-RU method discussed in the following chapter
can provide a broader understanding of the inter-relationships of
the quantities U, E,'\ M, n, etc., over a wider range than is obtained
from numerical calculation, if the latter is limited to particular
values of the selected quantities.
11.2. Calculation of 'l'N(U) by a Monte Carlo Method

The Monte Carlo method is a very practical tool for the calculation
of the integral (11.1). In order to demonstrate the flexible possibilities
of the simulation technique, a case is considered where the distribu-
tion function of the total amount of claims each year is a compound
Poisson distribution and where trend and long term oscillations
are also operative.
Suppose that each year G is of the form (lO.9) or, more precisely,
that during:
the first year Gl(X) = LFmPAi
the second year G2(X) = LFniP~i
the Nth year

It is also supposed that the generalized Poisson distributions F n 1
134
RUIN PROBABILITY DURING A FINITE TIME PERIOD
F n z, ... are given beforehand, as also are each of the distributions

{p~} = Uk(z). Further, it is supposed that the numbers (I +Ak)Pk,
the net premium + safety loading, are given for each year I;;; k;;; N.
The ruin probability lJ'N(U) is calculated by simulation in such
a way that a sample of all possible realizations of the risk process
is constructed. If some realization causes ruin before the Nth year,
it is unnecessary to continue until the end of the whole period.
The first sample realization is reached as follows.
Generate a random number, rIP' say. Let ns = (Ul)-l(rlP)' i.e. na
is the (sample value of the) expected value of the total number of
claims in the first year. This random experiment fixes the distribution
function of the total amount of claims during the first year to be
Fns. Now generate another random number r lF , say, and put:
Xl = F1i!(r 1F)
this can be considered as the total amount of claims during the
first year in the first sample realization. The amount left to the
company is then:
U I = U+(I+AI)PI-X I
making an obvious modification if the rate of interest is taken into
consideration. Now if U 1 is < 0, ruin has occurred and the first
sample realization is complete. If not, then analogously the amount
U 2 = UI+(I+A2)P2-X2 is calculated, and so on. If all numbers
U l' U 2' ••. UN ~ 0, the sample realization has given the result
"non ruin", otherwise a ruin has occurred.
The procedure is now repeated a sufficient number of times, and
the estimate for lJ'N(U) is given by the quotient of the number of
e.
ruins to the sample size. Suppose that this quotient is Then the
binomial rule gives the standard deviation:
J(e(l;e))
where h is the size of the sample.
The above risk process model presupposes that the trend and the
long term oscillation are both deterministic. This assumption was
made for the sake of clarity. It is however not difficult to include
random components in these fluctuation elements.
If the assumption that S(z) remains the same over the whole
135
RISK THEORY
period is dropped, the process appears to become more complicated.

It should be noted however, that it has not been assumed that each
function Fn. has the same S-function as its claim size distribution,
so that the above procedure is directly applicable if the S-function
varies.
In practical applications N should not be too large and the same
requirement applies to the number of terms in the sums 'LFn.Pn. k ,
otherwise the simulation will become unduly expensive. The practical
effect of this limitation may remain quite academic however, since
it is never possible to estimate the fluctuation functions Uk(z) or
any parameters over long periods so precisely that the above
limitation would lead to an error of larger magnitude than the
estimation error. If the N P-method is combined in simulation,
instead of the functions F n. their fIrst few moments only are
required.
136
CHAPTER 12
The Ruin Probability

During an Infinite Time Period
12.1. Introduction
In the previous chapter a brief survey was given of methods of
determining numerical values of the ruin probability for moderate
time periods. For a further appreciation of the nature of this
probability a study of the limiting position for very long term
periods is essential and this chapter will be concerned with this
problem.
It will be assumed that the distribution function F(x) of the
total amount of claims during 1 year remains constant, or at least
that its changes are generated merely by changes in the size of
the portfolio giving rise to changes in the expected number of
claims each year. However, it will always be assumed that the
distribution of the size of one claim S(z) will be independent of
time. In this type of problem, i.e. involving an infinite time period,
this can hardly be avoided since a more general theory is difficult
to apply in practice because of lack of knowledge of changes in
future risk distributions. A more serious restriction is the assumption
which is going to be made about the stability of the safety loading ,\.
This hypothesis implies that changes in the expected number of
claims which are not sufficiently rapidly corrected by premium
adjustments are ignored. Such correction is, of course, most
important if the question of insolvency arises, since insolvency is
obviously most evident if premiums are insufficient to meet the
risk. However, the problem can be divided into two parts by dealing
with the premium insufficiency as a separate aspect from the effect
of random fluctuation which is the subject of the present chapter.
Suppose the company has a risk reserve U at the beginning of
the time period T under consideration and let g be the total amount
137
RISK THEORY
of claims during the period, regarded as a random variable. T can

be 1 year, as it was in the preceding chapters, or any other finite
time. Then the quantity studied is the surplus (or loss if negative)
of the company, i.e.:
(12.1)
where A is the (constant) positive safety loading and P = E{g} is
the net (risk) premium income in the period T, also assumed to be
constant. The expected amount of the surplus is then E(1)} = "AP> O.
If the surpluses of N consecutive periods are considered, i.e.:
(12.2)
then W N is a random function of time attached to a discrete random

process. The probability of the possibility:
Wj ~ - U for some 1 ~ j ~ N (12.3)
which, by definition, indicates insolvency of the company in
question, is to be studied. This probability-:
lJfN(U) = I-P{Wt>-U for i ~ N} (12.4)
is the same as was considered in the preceding chapter, but in this

case N will be allowed to tend to infinity. It is assumed in what
follows that claims amounts g, during several periods are mutually
independent.
12.2. Ruin Probability

The approach in this section is a modification of that adopted
by Wald in his sequential analysis (Wald, 1947). His approach can
be regarded as a model consisting of two gamblers A and B playing
a series of games, the result of each game being a random variable
7J the distribution function of which is assumed to be known.
Applied to an insurance model, A is the insurer and B the portfolio,
i.e. all the policy-holders as a joint group. A has a finite initial
stake U and the gamble is terminated if A loses his whole stake,
the result being the classical case of gambler's ruin. No further
reference will be made to the parallels between the theory of risk
and the theory of games_ Instead the mathematical apparatus
will be directly applied to this special case. Readers who are not
138
RUIN PROBABILITY DURING AN INFINITE TIME PERIOD
familiar with the techniques of random functions may at first reading

omit this paragraph and find the solution of the problem in
Section 12.3.
It is assumed that the random variable "'It> indicating the surplus
of the ith period, has the distribution function G(y), which is
independent of i. No special assumption concerning G is needed;
for example, it is not necessary to assume it to be the generalized
Poisson function. It can equally well be the Polya function or any
other compound Poisson function provided, however, that the
integrals needed later are convergent. Also the practical assumption
is made that 7J can take both positive and negative values with
positive probabilities.
-ur-------------~·'~~-------------
'.I
Fig. 12.1.
The smallest integer for which ruin occurs is denoted by v.

That is to say, v is the number for which w.~ - U, Wj> - U (j < v),
v being a random variable. Fig. 12.1 illustrates a realization of a
process with v = 7 (dotted line), the company then being ruined,
and another case where no ruin occurs during the period of
observations.
The technique makes use of the so-called moment generating
function as an auxiliary function, i.e.:
f
+CX)
M(s) = e-8YdG(y) = E{e-B'li} (12.5)
Another auxiliary function, which is closely related to the above,

is E{e-''''N}, N being a constant integer. This mean value function
139
RISK THEORY
is clearly the moment generating function of the Nth convolution

of the function G(y). The principle is to derive two different
expansions for this function, which then directly give the result
sought. The first expansion is the following:
E{e-SWN} = E{e-8171 • e-817 2 • • • e-817.v} = [M(s)]N (12.6)
Another expansion for the same function is obtained by separating
the realizations into two groups. In the first group all the realizations
of the proce~s that lead to ruin during the first N periods are taken.
The probability of getting a realization belonging to this group is
by (12.4) 'l'N. All other realizations are left in the second group,
that is to say, realizations which lead to ruin subsequently
(v>N) or do not lead to ruin at all (referred to as v = 00). Now
the expectation in question can be partitioned into two conditional
expectations:
E{e-sw.v} = 'l'N. E{e- 8W ,," Iv ~ N}+ (1-'l'N) . E{e-SWNIN < v ~ oo}
(12.7)
For each fixed value of the process parameter the mean value
functions here as well as in the preceding equations are calculated
according to the rules for calculating the means and conditional
means of a random variable, or in other words: by multiplying
the value of e-IWN by the probabilities of all the realizations of the
process subject to the condition specified and then summing
(integrating) over the group of all these realizations and WN and
dividing the result by the sum of these probabilities. The first
term of the right-hand side can be further developed:
'l'N. E{e-8WNlv~N} = 'l'N. E{e- BWv . e-B(Wrwv)lv~N}
N
= 'l'N L P{v = ilv~N}E{e-BWv[M(s)]N-ilv = i}
i-I
= 'l'NE{e- swv . [M(s)]N-Vlv~N}
The latter form of the expansion is obtained having regard to
the definitions of WN and M(s) and the independence of WN-Wv
and W v •
(12.7) can now be written as follows:
E{e-, wlI} = 'I'N. E{e-'wV[M(s)]N-Vlv;;£N}
+ (1-'l'N) . E{e-SWN!N < v ~ oo}
140
By multiplying the partition above by the quantity [M(s)]-N

the following equation is obtained:
1 = lJ'N. E{e-SWv[M(s)]-Vlv~N}
+ (l-lJ'N) . E{e-SWN[M(s}]-NIN < v ~ oo} (12.8)
So far the choice of the auxiliary variable s has been open. To

obtain convenient values it is given the value which makes:
M(s) = 1 (12.9)
To prove that there really exists a positive root s = R satisfying this

equation observe the following facts: M(O) = 1, M'(O) = -E{1J}<O,
and M"(s) = E{1J2e-S'1} is positive for every s. It can now be
°
concluded that when s increases from the function M decreases
from the value 1 and the curve is convex for all positive values
of s. On the other hand, because by definition 1J can take negative
values, by giving a large enough value Sl to S, M(s) will reach a
value larger than 1. Hence it follows that M(s) decreases to a
minimum after s = 0, which is obtained at a point s = So say,
and the minimum is < 1. From this point the function is monotonely
increasing and hence there is between So and slone and only one
value s = R which satisfies (12.9). Replacing s in (12.8) by R:
1 = lJ'NE{e-Rwvlv ~N}+ (1-lJ'N)E{e- RwN IN < v ~ oo} (12.10)
Two approximations are now made. First the second term is

omitted, as it is known to be positive or zero. For the second
approximation note that from the definition of v the quantity
Wy in the first term indicates the ruining value of the (negative)
profit, hence it is always ~ - U. If then Wy is replaced by - U

the first term of the right-hand side of the equation is decreased.
Hence:
(12.11)
or
(12.12)
This formula is one of the main results of the theory of risk. In it

N can be any positive integer and it also holds when N tends to
infinity.
Before developing the theory further it is of great interest to
141
RISK THEORY
get some idea as to how much the function lJ' = lJ'<1J differs in
reality from the upper limit e- RU, i.e. would it be feasible to use
the approximation:
lJ'(U) = O(U)e- RU ~ e- RU (12.13)
where O( U) is an unknown function being ~ 1.
The evaluation of the error in this approximation has been the
subject of intensive studies by, for example, Segerdahl, Arfwedson,
Cramer, Saxer, and Ammeter, and has been solved separately for
different definitions of the function G, for finite and infinite time
intervals and for different configurations of the ruin test points.
These details are not considered here because it would be beyond
the scope of an elementary textbook and especially because it has
not yet been possible to derive simple general formulae which
would be useful in practical applications. An extensive survey of
the results has been given by Segerdahl (1959), but for a full
account reference must be made to the original papers. The following
shows however that as a rule the approximation is not too rough,
at least not when T ~ O.
Let N tend to infinity in (12.10). To find the limit obtained by
the second term on the right-hand side another partition will be
made:
(l-lJ'N) E{e-RcoNjN < v ~ oo} = lJ'IE{e-RCONlwN < B,N < v ~ oo}
+ lJ'2E{e-Rco,vIB ~ wN,N < v ~ oo}
where lJ'1 = P{wN<B,N<v~oo}, lJ'2 = P{B~wN,N<v~oo}and
B is a number which will shortly be fixed. The object is to prove
that both terms vanish when N ~ 00. To show this it is sufficient
to prove that when Band N are chosen large enough, both terms
are ~ e/2 where e can be fixed in advance to be arbitrarily small.
This is so as regards the latter term if B = -log(e/2)/R because
then e- Rco N~ e- RB = e/2 in this term. As regards the first term
the well-known Chebycheff's inequality P{IWN-E{WN}I f; IklaCON}~
l/k2 is applied for k = N1/4 and gives P{IWN-E{WN}I f;a"N3/4}~
1/.JN where E{WN} = NE{TJ} and a" is the joint standard deviation
of the variables "It. Now taking N large enough to satisfy both
inequalities Nf; (4/e 2) e 2RU and NE{TJ}-N3/4a"f; B itfollows that:
lJ'1 ~ P{WN~ B} ~ P{IWN-NE{TJ}I f;N3/4a,,}~ l/.JN ~ (e/2)e- RU
142
Noting that by the definition of v, WN in the first term of the above

partition is > - V and hence e-RCON<eRU, it follows that this
term < e/2 and so the proof is completed.
The second approximation when obtaining (12.11) was the
replacement of e- Rwv by e RU . This means' in fact that ruin is
always assumed to occur only by exactly reaching the ruin barrier
w. = - V. In reality this is, of course, not true as a rule and the
barrier is overpassed more or less. But in the case when N = 00
the effect of this approximation is generally not very great.
Since the second term of (12.10) vanishes, it follows:
1
1= E{e-Rwvlv< oo}. 1JI(V) = O(V) e RU '1'(V)
from which:
1
- - = E{e-R(wv+U)lv< oo} (12.14)
O(V)
From this formula it can be directly seen that O(V) is close to 1
if the absolute value of the ruining loss w. is in general only
relatively slightly in excess of V. This is the case, for example,
when the test period T is very short or equals 0 and reinsurance
cuts off the large claims. When T -+ 0, only one claim can with
probability 1, cause the 'final ruin'. If due to the reinsurance, the
size of one claim is limited and ~ M, then also Iw.+ Vi <M and a
limit formula is obtained as follows:
C(V»e- RM (valid for T = 0) (12.14')
It should further be noted that even if O(V) were as low as 0·5
and e- RU is of the order of magnitude of 0·01, then in replacing
the true '1' by the approximation (12.13), the original V would
be replaced by 1·15 V, the adjusting figure compensating for the
uncertainty of O(V). If 0 were greater than 0·5 the error arising
from the use of the approximate formula would be still smaller.
Cases where this type of margin in V is significant will seldom arise
in practical actuarial work since the choice of the permitted
magnitude of the ruin probability is itself arbitrary. From numerical
examples presented later in this book it will be seen that the safety
loading and, in many cases, the net retention, are factors which
have considerable influence on the insolvency function, far more,
143
RISK THEORY
in fact, than the uncertainty of O(U). Of course, in certain critical

cases, the study of the exact theory as developed in the papers
referred to above may be desirable.
As has been shown, (12.12) holds for every length of the test
period T (if the 'I'}'S remain independent and stationary) and also
when T --* 0, i.e. when testing is carried out continually at every
time point. At the same time the ruin probability 1]1 (U) increases.
This is easily seen since the ruin probability is less when the solvency
is tested once a year, for example, as compared with continuous
testing. The former case allows temporary insolvency during the
observation years provided solvency is regained at the end of
each year.
(12.9) (with 8 = R) is inconvenient for computations and it is
useful to develop it in a more suitable form.
According to the definition (12.1):
1 = M(R) = E{e- Rq} = e-(l+).)PRE{eR~}
f eRXdF(x)
00
i.e. e(l +).)PR = (12.15)

o
In the particular case when ~ has a generalized Poisson distri-
bution, (3.11) (being also true for moment generating functions)
gives:
E{eR~ = exp[nE{e~-n]
where , is the amount of one claim. By taking logarithms and by

using (12.15) this gives:
f eRZdS(z)
00
n+(l+'\)PR = n
o
Noting that as in Section 3.3, P = E{g} = nm, where m is the
mean size of one claim, it follows, after dividing by n, that:
f eRZdS(z)
00
1 + (1 +'\)mR = (12.16)
o
From this it is seen that 'the insolvency constant' R, the non-
trivial root of (12.9), is not dependent on the choice of the time
unit but solely on the distribution function S(z) of the amount
144
of one claim. Thus it will again be appreciated that the inequality

(12.12) is valid independently of how frequently the insolvency
measurements are made and thus it will hold good even if the
possibility of insolvency is measured at every time point.
A formula corresponding to (12.16) can also be obtained when
the fluctuating basic probabilities are assumed to be in accordance
with the distribution (10.15). (10.19) and (10.20) also hold true
for the real transforms obtained if i is omitted, and thus the
equation:
JeRZdS(z))]-k
00
e(1+).)mnR = [1 + (n/k)(1-
o
may be written down. This equation may be expressed as:
oof 1_e-{1+).)nmRlk
eRZdS(z) = 1+ - - - - - (12.17)
n/k
o
from which the insolvency constant R may be found.
In this formula the reservation made on page 126 must be kept
in mind: It is not valid for an ordinary Polya process but only for
the process which is cut into independent annual periods, and the
accumulated profit or loss, being the sum of those for each pre-
ceding period, is observed at the end of each period.
It can be noted that for k =F 00 there is an essential difference
between (12.16) and (12.17). In the latter the constant R depends
on the size of the company.
Consider the case of a generalized Poisson distribution with test
period T --+ O. In order to estimate O(U) in (12.14) suppose that
ruin occurs and write IX = Wv-l + U. Clearly in this case T --+ 0
gives 7Jv --+ - , where , has the conditional distribution of the
amount of one claim on condition that this claim is ~ IX, or
Pg~xl'G:IX}. The amount IX is a random variable which can have
only positive values, and:
The term on the right-hand side is a weighted average of

the numbers e-RxE{eR{I'G:x}, with probabilities P{x<lX~x+dx} as
weights. Intuitively it can be expected that these probabilities are
145
RISK THEORY
at least approximately proportional to the probabilities Pg ~ x}dx.

If this approximation is used, then:
f p{n~x}dx = const. f [l-S(x)]dx

00 00
1 = P{oc< <Xl} ~ const.

o 0
f xdS(x)
00
= const.
o
and hence:
1 x
P{oc;:£x} ~ -
m 0
f
Pg~x}dx
This gives:
1
E{e-R(rov+U>jv< <Xl} ~ ;;;, f
00
e-RXE{eR'j'~x}P{'~x}dx
o
1
f e-Rxdx(Pg;:;;x}E{eR'I';:;;x})]
00 .
= mR [E{eR'I';:;;O}Pg;:;;O}+
o
Now P{'~O} = 1, and, as is easily seen,
dx(P{'~x}E{eR'j'~x}) = -eRXdS(x).
Hence by (12.16):
RU
E{e-Rrovjv<<Xl} ~ e_ [l+(l+'\)mR-
mR
f dS(x)]
0
00
= e RU (l+,\)
This gives an approximation:

1 1
O(U) ~ - . lJf (U) ~ - e- RU (12.18)
. 1+,\' 00 1+,\
Since 1/(1 +,\) ~ 1, it is seen that the approximation (12.13) is
well motivated in the case of a generalized Poisson distribution with
a short test period T.
It is possible to prove that the result (12.18) is exact for
S(x) = l-e- ax .
Exercise 12.2.1. Consider the case of a generalized Poisson dis-
tribution with test period T -+ O. Prove that: (i) for S(x) = I -
e-ax,O(U) = 1/(1+,\), and (ii) for S(x) = e(x-M), 1/(1+2,\+2,\2)
<O(U) < 1.
146
12.3. Applications
Summarizing the main results of the preceding section it can be
stated that the probability that the company will be ruined is:
P{U) = O(U) . e-RU (12.19)
where O(U) ~ 1 and in the Poisson case R is the positive root of:
f
00
1+(I+A)mR = eRZdS(z) (12.20)

o
and in the "Polya" case the positive root of
k
1+ -
n
(l_e-(l+'\)mnRlk) = f
00
0
eRZdS(z) (12.21)
(12.20) and (12.21) can be solved by means of well-known numerical

or graphical methods. Using the exponential expansion of e Rz and
integrating term by term, (12.20) is transformed into the form
(cf.3.1):
(12.20')
from which R can be found by a few trials, the first approximation
being:
In Fig. 12.2 the ruin probabilities for a non-industrial fire insurance
1.oo-=--------~A=--"'0-----------
'10
'01
U
.001+---£-::-1-0+000-----:£:-2--'0'rOOO-~£-30-"f--OOO--£-::-4-0--.,OOOr-----t---...
Fig. 12.2. Non-industrial fire. M = £20000.
147
RISK. THEORY
1'00 .~::::::::::::--_ _
50 100 150 200 250 300
Fig. 12.3. Non·industria'! fire, A = 0.05, unit £1 000.
with M = £20 000 are presented on a semi-logarithmic scale. It is

seen how intimately the value of 1pIO is linked to the safety loading
A and the initial reserve U. If A = 0, it can be proved that 1p = I,
i.e. ruin is certain.
Computing 1p for different values of the maximum net retention M
and fixing A, the inter-dependence between M and U can be found
as in Fig. 12.3 which is calculated for a fire insurance distribution
M
50
//
ES5cher /'
I
20 I
I /
I
I /Eq.(12.36a)
10 I //
I
I
I
I ////
5 /
I /
I
2 I ./
/
,//' /
./"
u
10 20 50 100 200 500
Fig. 12.4. M as function of U. Non·industrial fire, e = 0'01, A = 0'05.
Unit £1000. As a comparison a curve corresponding to Fig. 6.3 for
one year ruin probability, n = 1000, is computed by the Esscher
method.
148
with ,\ = 0·05. From the straight line e = 0·01 the function M =

M(U) can be read. In this way Fig. 12.4 is obtained. In the same
figure the curve obtained by means of the Esscher formula is
presented on the basis of one year's ruin probability. The e- RU
method gives, of course, for a given U the smallest, i.e. most
cautiou8, values.
A strange feature of the e- RU approximation, mentioned but not
emphasized in the preceding section, is that (12.20) depends only on
the distribution function of one claim. This statement means,
paradoxically, that the size of the company plays no part in calculat-
ing, for example, the reserves U when the Poisson model is used.
This is difficult to appreciate, especially as in Chapter 5 it was
specifically stated that the reserves should be proportional to the
square root of the size of the company. However, the contradiction
is only apparent. To illustrate the phenomenon consider again the
profit we which is accumulated during time t. Using the normal
approximation (cf. (5.2)) the limits can be written down as:
(12.22)
the probability that we lies between these limits being 1-2e. In a
slightly heuristic way it can be said that there is a large probability
that the process curves Wt (t = I, 2, ... ), indicating the running of
the risk business, fall in the hatched area in Fig. 12.5 defined by
the limits (12.22).
-u - ---- - - - - - - - - - - - - - - - - - - -
Fig. 12.5.
149
RISK THEORY
The solvency of the company is tested at the end of each calendar

year (the lower time scale on the horizontal axis). The lower boundary
of the hatched area, which can be called "the critical curve", decreases
to the point:
(12.23)
and is from then on increasing because the linear growth of >..tP

overtakes the downwards effect of the square root term in (12.22).
The point (12.23) is critical in the sense that the area of probable
passages is there nearest the ruin limit - U.
To show the difference between a small company and a large
one the variable on the horizontal axis is changed from time into
the accumulated premium income tP (upper scale in the figure).
Then another company is considered to have the same U, >.., and
S(z) as the first one but the size, measured by P, is 10 times as great.
When the curves (Fig. 12.5) of this larger company are placed into
the figure on the basis of P-scale (upper scale), they.prove to be
exactly the same as the curves of the smaller company. The only
difference is that, if again the time scale is introduced (lower scale),
the unit is 10 times that of the original, i.e. the first test point is
placed where originally the end point of the tenth year was (new
time scale corresponding to this larger company in brackets). Hence
in the first year the business reaches directly in one step the most
critical part of the curve. If a third example were taken of a company
having again the same U but with the size of its business, say,
25 times that of the first mentioned company, then the first testing
point of this company is already beyond the most critical point.
Thus it can be clearly seen why the size of the company has an
influence on the probability of ruin attached to one year testing
and why it has no influence for ruin probability with a continuously
tested infinite time period. If the company is very large, the danger
point might be at one week from the start, but for a very small
company this point may fall perhaps in the tenth year and for a
medium size company perhaps just at the end of the first year.
Because regard is not had to possible insolvency during the fiscal
year, the one year's ruin probability depends on that part of the
curve at which the end of the first year falls and this depends on the
size ofthe company. From the same figure it is also easy to appreciate
150
why e- RU is, and must be, independent of the size of the company,
or, what amounts to the same thing, the expected number of claims
during one year. This is due to the fact that the testing is assumed
to be performed continually and so no difference exists between a
small and large company. The underlying concept is that both for
small as well as for large companies the business runs under the
same probability limits if both have the same U, ,\ and S(z) and
testing is continuous at each time point. The only difference is that
the process proceeds much more rapidly for the larger company,
but both have the same insolvency probability when the observation
time is infinite.
If, however, testing for lJI(U) is only at the end of each fiscal year
even if the observations are continued without limit, then a large
company has the smaller probability of insolvency, i.e. the function
C(U) (cf. (12.14)) deviates more from 1. Since companies are,
in general, mainly interested in the measurement of insolvency at
the end of each fiscal year, it can thus be expected that the e- RU
method leads to larger reserves than necessary, especially for large
companies.
Exact expressions for the lJI-function for a few special cases of
S(z) can be found in the literature. One of them is given here without
proof. For the exponential distribution function (cf. Section 3.5.6):
S(z) = 1-e-a~
it will be found that for a continuous ruin test:

1 ,\
lJI(U) =- e-aul+'\ (12.24)
1+'\
.As already noticed in the "Polya" case - contrary to the Poisson
case - e- RU depends on the size of the company. -
Exercise 12.3.1. Calculate e- RU for the distribution mentioned in
Exercise 3.5.4.2. What is M for U = £lOO 000, e = lO-3 and
,\ = 0·051
12.4. Some Approximation Formulae

The ruin probability "l'depends on the distribution S(z) of the size
of one claim. It is however possible to find approximate formulae
which are valid for all possible distl'ibution functions S if the size
151
RISK THEORY
of the claim z is maximized and ~ M, as is normally secured by

reinsurance. The method directly parallels that applied in Chapter 5
for the one year ruin probability.
Consideration is limited to the Poisson case and (12.20) is developed
as follows:
M
1 + (1 +A)mR = f eRZdS(z)
o
=f M ( 1 R2Z 2+ -R3
1+Rz+ __ 1
2!
Z 3+ ... ) dS(z)
3!
o
= 1+Rm+ 1X2 R2+ 1X3 R3+ ... (12.25)
2! 3!
Use will be made of the inequalities:
m(KM)i-1 ~ lXi ~ 1X 2Mi- 2 (j> 1) (12.26)
where (cf. 5.7) again:
K=~ $;1 (12.27)
mM -
The right-hand side of (12.26) follows from:
M M
lXi = f zidS;;:; f Z2 Mi- 2dS
o 0
and the left-hand side from:

lXi 1X1+1
;;:;- (12.28)
1X1-1 lXi
(cf. Exercise 12.4.1). Substituting (12.26) into (12.25) the inequalities:
(_. MR
~. KMR ~ eK~IR-1-KMR
$; eMR -1-MR. (12.29)
K -
are obtained after some calculations.
In order to obtain the desired approximation consider the function
l(x;A) = (1+A)x+1-e x. The inequalities (12.29) may then be
written I(KMR;A)~O and f(MR; ,\fK);;:;O. For a fixed A the
function 1 (x; A) is equal to 0 for x = O. Since f' (x; A) = (1 +,\) -ex,
152
thenf'(O;.\) = .\>0. Sincef' is a decreasing function, the function

f (x; .\) increases from 0 to a certain positive value when x increases
from 0, and then decreases monotonously tending to the limit - 00
for x ~ 00 . Thus the equation:
(1+.\)x+1-e x = 0 (12.30)
has one and only one positive root, say X(A) > O.
The conditions (12.29) then become:
KMR ~ X(A)
M R ~ x(AjK),
or
x(AjK) ~MR ~x(A)jK (12.31)
The function X(A) is an increasing function of A, since
dx x x
dA eX-1 = --:(-1---1-'--) > 0 for x > 0
L xt
-00--.
eX- - - - - --
x 0 i! (i+1)!
The error by this approximation for normal values of .\ and K is
very small as seen from graphs in Fig. 12.6.
·1
K
'5 1'0 .
Fig. 12.6. The quotient Q = K.: A~/K) for some va.luelil of .\.
153
RISK THEORY
It can be verified (cf. Exercise 12.4.2) that the quotient Q of the

lower and upper limits of (12.31) satisfies:
K(I+A) ~ Q ~ 1
K+A - -
For example, if A = 0·05 and if K = 0·5, then the above inequality

gives 0·955 ~ Q ~ 1. The true value of Q is 0·970. Hence in
practical problems where as a rule ,\ is small compared to K, the
ruin probability does not depend to any significant extent on
moments of the distribution S(z) higher than the second.
The lower limit can be made independent of K, i.e. of the first and
second moments of the distribution S(z), by observing that X(A)
~x(AfK). Hence, by eliminating R from e = e- RU, if follows that:
X(A) x(AfK) M X(A)

--<--<-< (12.32)
log(lje) = log(lje) = U = Klog(lje)
X(A) can be solved from (12.30) and some calculated values are
given in Table (12.1).
TABLE 12.1. Function X(A).
A 0·05 0·10 0'15 0·20 0·25 0'30 0'40 0'50 0'75 1'00
X(A) 0'10 0·19 0·27 0·35 0·43 0'50 0'64 0'76 1'03 1'26
As is seen from the table (or by means of a more exact analysis of

(12.30) ):
1·75A < X(A) < 2A (12.33)
the lower limit being valid for A~0·2. Hence
1·75 AU < x(AjK)U ~ M < 2 AU
(12.34)
log(lje) log(lje) - Klog(lje)
and
Klog(lje) M < U ~ log(lje) . M < log(lje) M (12 •
2A - x(AfK) 1.75A .35)
which, summarizing the conditions, are valid for O<A;;;;O·2, if

S(z) = 1 for z = M and the generalized Poisson is assumed.
154
It will be recalled that the other lower limit in (12.32) and the
other upper limit in (12.35) are independent of the distribution S(z),
i.e. the formula is universal and holds good for every portfolio from
which the top risks are cut away, for example by reinsurance. If
the size ofthe claim is fixed equal to M, then all the limits in (12.32)
coincide and the formula is exact. This proves that the first of the
lower limits in (12.32) is the best possible as a distribution-free
approximation. The other limits of the formulae depend on K and
consequently on the function S(z). If the maximum net retention is
relatively low, the limits are not very far from each other in fact.
E.g. in (12.34) the upper limit exceeds the lower one by 20% only if
A~ 0.1 and K"?: 1/3, which again proves that the function S(z) often
has only a minor influence on the solvency. This, however, holds
good only if M is not very large and if the risk business on the
company's own retention is not too heterogeneous, the latter being
indicated by a value of K which exceeds 1/3.
If a limit is taken which is on the safe side, i.e. which gives the
better security, and e is taken as 10- 2 and 10- 3 respectively, the
following approximations are obtained:
M = 0·43 AU for e = 10- 2 (12.36a)
M = 0·29 AU for e = 10- 3 (12.36b)
These formulae are exact parallels to (5.19) and (5.20). The coeffi-
cients are smaller in this instance because infinite time periods are
being considered.
Formulae like the above are convienient for very rapid estimation
of the order of magnitude of M, A, etc. particularly if the S(z)-
function is not known or not available; for exact calculations,
however, (12.19) to (12.21) must be used.
In Fig. 12.4 values of (12.36a) are compared with actual values of
e- RU .
Using inequalities (12.14'), (12.31) and (12.33) it follows that:
Exercise 12.4.1. Prove the inequality (12.28) for every j supposing

that S(z) = 0 for z < 0 and that the moments in question exist.
155
RISK THEORY
Exercise 12.4.2. Prove the inequality K(l +'A)/(K +'A) ~ Q.

Exercise 12.4.3. Estimate M, when U = £100000, e = 10- 3,
and
'A = 0·05. What is the upper limit of the error of the estimate,
if K = 0·61
Exercise 12.4.4. Compare results obtained by (12.35) and (12.36)
with the results obtained in Exercise 12.3.1.
12.5. Discussion on the Different Methods

In this section some further comments will be made concerning
the different ways of approach to the main problems, which were
summarized at the beginning of the book in Section 1.4 and in
accordance with which the problems have been dealt with in the
preceding chapters.
The great advantage of considerations limited to one year only
(or any other finite time period with a single test point) is that the
formulae can be obtained in a form which lends itself to easy
computation and which provides a good guide to the interdependence
of the main variables of the theory of risk n, 'A, e, M and U. The
different trends and variations of basic probabilities have no
conclusive influence during short periods and they can therefore
be estimated in a rather crude way. The great disadvantage of the
short period method is that it does not directly deal with the long
term continuation of the company's life. From the point of view
of the company's management it is not sufficient to arrange
reinsurance, different reserves, etc. so that the company will still
have a minimum degree of solvency after one year. On the contrary,
it is necessary to plan all the security measures affecting the
company's existence for an unlimited number of years in the future.
Regard was given to this aspect in the applications of Chapter 5
recommending the cautious choice of the variables, particularly
by taking only a part of the actual free reserves as the initial
capital U. Often it may be intuitively possible to estimate what
part of the reserve the company's management is prepared to lose
in one year taking the worst possible view and to take for this
purpose, for example, one-third of the actual reserves. Should this
amount really be lost, then reinsurance and other security measures
must be re-evaluated and necessary amendments made im-
mediately.
156
The foregoing approach suffers from arbitrariness in regard to

the choice of the variable U and for this reason long term con-
siderations are developed. The extreme alternative for the future
safety of the company is to take an infinite time period. For this
reason the e- RU technique has been given considerable attention
in actuarial literature. However, this method also has some obscure
points. In particular, the results, as is seen in Fig. 12.2, are all
very sensitive to the size of ,\ and in other respects the method
also involves assumptions concerning all future time. Even if it
were possible to estimate ,\ for some few future years, it is very
difficult to say what it will be subsequently. In fact experience
shows that different kind of trends, amendments of premium tariff
and also other circumstances are always changing ,\. On the other
hand it is sometimes observed that only the very beginning of the
process influences the value of e- RU (cf. Fig. 12.5). When the
dangerous part of the process is passed then the longer future has
no significant influence. This is, however, only true as a consequence
of the condition that ,\ is continuously greater than some fixed
positive value and that the reserves grow to infinity. In practice this
is clearly an unrealistic situation. If a company is accumulating
very large profits, then a mutual company probably will introduce
different bonus systems and return surplus to the policy-holders
and joint stock companies will probably give increased profits to
the stockholders. In practice, in addition, amendments to tariffs,
etc. are expected, and deterioration in the value of money erodes
reserves. It would not seem an unrealistic assumption to provide
that if the accumulated profit is growing very large the company
will take steps to reduce the safety loading '\, the more so because
the interest yield on the accumulated profit is also available. This
then leads to a paradox. If it is assumed that there exists some
upper limit for the accumulated profit, however large it may be,
over which the company will not increase its reserves, then it is
easily seen that 1Jf = 1. In other words, all the companies subject
to this assumption will sooner or later be ruined!
It is true that some of the trends have no fatal influence on the
insolvency probabilities, perhaps no influence at all. This is so as
regards the growth in volume, which is indicated by the increase
in the annual number of claims n. This only means that the test
points (cf. Fig. 12.5, P-scale) are placed less and less dense in future
157
RISK THEORY
if testing is performed at the end of each calendar year or at the

end of some other time period T. If the testing period T -+ 0 the
only result is that the process runs more rapidly, but the insolvency
coefficient R is not changed. Unfortunately, however, other trends
and long term oscillations seem to have an influence which is difficult
to estimate, and, as noted above, some of the future changes can
have such serious consequences that simple methods become of
no value.
One weak point in the e- RU method was the assumption T = 0,
which can give excessively large values for large companies. Another
arbitrariness exists in the selection of a suitable insolvency
probability E = e- RU . The same selection is also certainly necessary
for short period systems, but as regards infinite or long periods
the problem has another dimension. This is a rather nice question
which may be seen by considering the alternatives whether to
allow an insolvency in 1 year with probability E = 10- 4 or during
10 years with probability E = 10- 3 or during an infinite time period
with E = 10- 2 • The reasoning is closely connected with the approaches
recommended by Borch (1963) whose idea is to approach the
problem by means of a quantity called the utility of different
decisions. Many operational problems can then be solved by
optimizing this quantity. Reference is again made to this topic
in Chapter 13.
An optimum choice may be available between two extremes, i.e.
between the one year system and the infinite system, taking into
consideration some finite number of fiscal years. As shown in broad
outline in Section 11.2. it would be feasible, at least if electronic
computors are available, to study the insolvency problem attached
to a suitable series of consecutive years in a very flexible way and
to experiment with different assumptions concerning the behaviour
of the basic variables and the function S(z). In this way it may be
possible to utilize methods which are known from operational
research studies. Different kinds of assumptions and operational
rules can be experimented with; for example it may be possible to
programme in advance rules of procedure which are to apply if
the profit is rapidly accumulating, or vice versa if it is diminishing.
These rules can concern bonus systems, amendments of tariffs,
interest yield on the accumulated reserves, changes of the maximum
net retention as a function of the accumulated security margins,
158
etc. It also seems necessary to include conditions concerning special

circumstances like competition and taxation.
The question of which of the different methods is suitable and
which is superior to the others is, of course, very much dependent
on the special circumstances and the way in which the question
is asked. It may often be advisable to apply the different methods
and ways of proceeding in parallel and it will probably prove both
interesting and illustrative to compare the results reached.
159
CHAPTER 13
Application of Risk Theory to

Business Planning
The many applications of the theory of risk described in the previous

chapters, such as the estimation of a suitable level for the maximum
net retention, the evaluation of stability, the safety loading and
the magnitude of the funds, have been treated as isolated aspects
of an insurance business. In this chapter an endeavour will be
made to build up a picture of the management process in its
entirety and to place the risk theoretical aspects in the context
of other management aspects most of which are not of an actuarial
nature. In this way many of the applications previously dealt with
can be integrated with the concepts of modern business planning,
in particular the techniques of long-range planning and dynamic
programming.
The object is to describe the actual management process of the
insurance business in as comprehensive a way as possible by means
of a mathematical model. To do this it is necessary to define,
a priori the goals of the company.
(a) Solvency. In most of the problems dealt with in risk theory the
concept of solvency has been a predominant assumption. More
precisely only those business policies are adopted for which the ruin
probabilities are less than some predetermined E, selected to be
small enough in the relevant context. In many of the problems
considered in this book ruin probabilities for a period of one year
only have been dealt with for clarity but there is no essential
difficulty in developing the methods for periods of more than one
year.
160
APPLICATION OF RISK THEORY 'fO BUSINESS PLANNING
(b) Maximization of Profit. This is another classical business goal

and although the theory of risk emphasis the solvency criterion the
maximization of profit is generally involved, either explicitly or
implicitly. For example the problem of finding the maximum net
retention on the condition that some minimum stability is guaran-
teed, implicitly involves a maximization of the retained premium
income and thus involves the profit margin. De Finetti, Borch and
other authors have studied models in which the maximization ~ or,
more generally, optimization ~ of dividends is a criterion. This is
clearly closely akin to the maximization of profit.
For the definition of business goals it is further necessary to
distinguish the different ways in which profit may be used. These
are mainly (1) transfer to reserve funds U in order to improve the
solvency and wealth of the company (cf. (a)), (2) distribution to
shareholders and/or policyholders and (3) allocation for expansion
and other development (see (c)).
When the profit is of a stochastic character, as is always the case
in insurance, direct maximization is not possible. Instead, perhaps,
the mean (or expected) value of the profit may be maximized.
However, even this may not correspond satisfactorily to the needs
or wishes of management. For example a very large profit can give
rise to suspicion among clients or perhaps control authorities
regarding the fairness of the premium rates. On the other hand a
very small or negative profit can be so inconvenient that it should
be given increased weight in the decision process. Probably a
medium size surplus is considered appropriate. Decision Theory
attempts to express these aspects in the form of the mathematical
model by introducing the concept of 'utility', which is a function
constructed in an intuitive or other way to describe and measure
numerically the degree to which the result of some economic oper-
ation is desirable. In the present case the capital z available at the
end of the year is the result under consideration. The so-called
utility function (von Neumann and Morgenstern) is seldom linear
and generally fulfils the condition:
(13.1)
That is to say the company prefers to have a large rather than a
small capital.
It may be expected that a rational company, which has a clear
161
RISK THEORY
opinion regarding the size of the utility, aims to maXImIze the

expected value of the utility of z instead of the expected amount of
the surplus. In other words the company chooses that alternative
which maximizes the quantity:
J U(z)dG(z)
+00
U(G) = (13.2)
-00
which is called the utility of the distribution G(z). G depends on

reinsurance, premium rates and other facts which are to be chosen
to maximize (13.2). Olearly this criterion is the same as the preceding
one if the company's utility function is of the form U(z) = Az+B.
In its general form this new criterion is, however, much more useful
because, for example, the company can use a very much smaller
utility for critical negative z values than is gi ven by the linear formula.
(c) Expansion of the Oompany. In recent studies on the economic
and business behaviour of management some authors, e.g. Galbraith,
describe the development of modern society and the changing
philosophies of business and management. Originally owner-
entrepreneurs played a predominant role in business activity.
Nowadays they are very often replaced by professional managers
who are not owners of the enterprises concerned. At the same time,
the maximization of profit as the final goal of the business has
been replaced by other goals which are a guarantee for the con-
tinuation of the business and the expansion of the enterprise. The
size of the enterprise, measured e.g. by means of its turnover or
maybe as a share of the market and the increase of this share, are
features which may be called 'status symbols'. Hence the expansion
of the company may be the most important business goal and sol-
vency and profits subsidiary, though of course necessary, conditions.
Business philosophies will not be further discussed here. Model
building must have regard to all aspects mentioned above and it is
the task and duty of the company's management to choose from
among the various goals. The best strategy is determined by this
choice.
The first step in building a model is to find a suitable set, a
'vector' e (t) of parameters which describe the state of the company
at any time point t. Such parameters may be the premiums
p = p 1 + p 2 + ... + p k where the suffixes indicate the branches
or other suitable sections of the portfolio, technical reserves
162
APPLICATION OF RISK THEORY TO BUSINESS PLANNING
v = V I +V 2 + ... +Vk ' the net retentions M I , M 2, .•• , M k , the

reserve funds U, etc.
The next step is to construct a set of rules which translate the
state () (t) into () (t+ 1). For this purpose a number of functions are
to be found which describe the development of the business as
adequately as possible. These functions contain a number of par-
ameters. Some are fixed and beyond the control of the management.
Some others can be more or less freely chosen, such as the allocation
of the profit and the level of the net retention. The set of these
parameters is called strategy Z. The transition rules also depend on
a random effect r (t) subject to the total amount of claims in the
interval (t, t+ 1). Hence the transition can be written symbolically
() (t+ 1) = F (()(t); Z, r(t)). This formula is in a suitable form for
computer programming; the Monte Carlo method (Section 7.2) is
particularly convenient for numerical management of the stochastic
feature r(t).
For the application of the method outlined above a strategy Z is
first chosen. Then the states () (t) (t = 1, 2, ... , N) can be calculated
when () (0) is given. In accordance with the idea of the Monte Carlo
method this calculation is repeated numerous times and a set of
final state vectors () (N), the 'realization' of the process, is obtained.
It is then possible to get estimates for the mean values, standard
deviations, confidence limits, etc. of the state parameters P(N),
U(N), . ... The procedure is much akin to the theory of dynamic
programming or industrial dynamics developed by Bellman,
:Forrester, Nemhauser and others. A further development is to
define a utility function describing the goals of the management
when the problem becomes the maximization of this function.
As an illustration of the application of the foregoing model some
comments on the main features follow. Further details can be
found in the papers by Pentikiiinen (1975, 1976).
The basic equation of the transition rules is
Y(t)+J(t)+H(t) = D(t+I)+O(t+I)+ilU(t) (13.3)
Y (t) is the surplus of the risk business:

Y (t) = (,\ - K(t) - w(t))P -1X(t).Z(t). v' {n(t) (1 + K(t)+ w(t)}
_ 1X3(t) . Z2(t) -1 (13.4)

1X2(t) 6
163
RISK THEORY
where K(t) indicates the short period oscillations (p. 112) of the
basic probabilities and w(t) long period oscillations (p. Ill). Here
z(t) is a normally (0,1) distributed random number and Athe safety
loading in the premium. The formula is a direct application of
(4.10) and the Monte Carlo method. The value of K is obtained by
simulation, different independent values being used in each simu-
lation step. The variable w is estimated in a deterministic way,
possibly making use of different alternative values.
It may be questioned whether it is realistic to adopt a model
which provides for A to be varied in accordance with company
policy. Certainly because of competition and other market circum-
stances the possibilities of so doing are often rather restricted but
in certain situations it may, however, be possible. In this latter
case A is one of the decision parameters; in the former case A is a
parameter which is fixed or varying, in accordance with circum-
stances, beyond the control of the management. It may also be
possible to extend the model assuming some rules which change the
premium rates in each simulation step in accordance with the flow
of Y(t), e.g. making use of the theory of experience rating or of the
adaptive control systems of the theory of dynamic programming
(or engineering control theory).
The use of formula (4.10) is justified instead of the formulae of
Chapter 10 because K and ware each held constant during each
simulation year although, of course, their values vary from year
to year.
J(t) is the interest surplus, i.e. the excess of actual interest over
the amount needed for the maintenance of technical reserves. It
may be noted that assumptions regarding investment policy can be
introduced into dynamic programming via this quantity. It is also
useful to note that a convenient way to take inflation into account
is to make all calculations without allowance for inflation and to
adjust the assumed rate of interest to allow for the effect of inflation.
In this way the monetary quantities are obtained in terms of the
current value of money and, if necessary, can be converted into
amounts based on inflated values.
H (t) is the surplus of management expenses (i.e. the premium
loading less management expenses). A prognosis for H can be
constructed by means of conventional long-range planning tech-
niques and can be incorporated into dynamic programming as a
164
sub-analysis. An alternative wa? "'ould be to operate by means of

gross premiums in formulae (13.3) and (13...!) and to use the actual
expenses in the same equations. In this case it would not be neces1:iary
to separate ,\ and H.
D(t + 1) represents bonus and lor dividends paid out. This term
can be defined in different ways corresponding to the business goals.
One 'm~- is to consolidate the bonus on some prefixed level e.g.
D(t+l) = d.P(t). where d is a time independent parameter. How-
eyer thp implied non-variation from year to year would be incon-
wnient from the point of view of competition and public relations.
Another way would be to consider D(t+ 1) as some kind of 'output
of the businpss and to take its maximization as a management goal
(cf. (b)).
C(t + 1) is the allocation for salps promotion. discussed later.
~C(t) represents the increase in reserve funds which is a balancing
item in (13.3).
The terms on the left hand side of the basic equation (13.3)
indicate the resources available and the terms on the right hand side
indicate the different purposes to which these resources are allocated
in accordance with the strategy Z approved. However, before the
model can be made to work some more definitions and explanations
are needed.
The state of the company will be defined by means of U(t) and
by fixing the soh-ency levels as follows: U l(t) = limit of safe state;
U 2(t) = alarm limit; i.e. if U1> U> U 2 the state is no longer very
satisfactory and, therefore, increasing U is to be given priority in
the allocation of profit; U 3(t) = the winding up barrier, which may
be prescribed in legislation. If U 2> U> U 3 the company is already
in an unsatisfactory state and urgent emergency measures are
necessary.
Uland U 2 are strategy parameters or can be defined by means
of other parameters e.g. U 2 can be of type (5.4') where Y. and ,\ are
decision parameters.
The Sales Promotion item C(t) is obviously the most problematic
term offormula (13.3). It is assumed that the company has a normal
acquisition organization which is paid for by the normal loading
for management expenses. C is understood to mean re-enforced
acquisition. The idea is to find a formula which allocates a fairly
large part of the resources to this re-enforced acquisition if the
165
RISK THEORY
solvency is good (U > U 1)' and less or none if the solvency is weak.
It may be of the form 0 = 0 (U(t), Y (t), (J, )', 8 ... ), (J, )', 8 ...
being decision parameters involved in the function O.
It may be advisable to construct 0 to be negative in the event
that U < U 2' i.e. when the company is alerted to an unsatisfactory
state and, therefore, emergency measures are necessary. Thus
normal management expenses must be reduced, the normal cost of
acquisition decreased and other costs reduced or postponed if
possible. By such means it is hoped to achieve a 'negative allocation'
and thus to save the company.
The total amount 0 is distributed to different sections through,
say
(13.5)
The coefficients Iv are again strategy parameters indicating the
weights which are given to different sections of the business when
the expansion of business is planned, e.g. greater weight can be
given to the more profitable branches.
The purpose of the re-enforced acquisition 0 is, of course, to get
further expansion for the company. A 'sales response' function Gis
needed to provide this effect; it may be of the form
This function is to be found in some empirical way and indicates

the dependence of sales on the cost of sales. A linear combination
of the arguments may be an appropriate first approximation.
Normal growth, independent of the enforced sales 0, has also to be
taken into account.
Where dynamic control is involved, e.g. for net retentions M v'
some further formulae may still be needed but these are not dis-
cussed here and the original papers should be consulted. The model
will then be ready for working and can be programmed for the
computer.
The 80lvency criterion is a by-product of the calculation. If
U(t) < U 3(t) occurs for any t = 1,2, ... ,N, it indicates the ruin of
the company. The simulation program records the number of
occasions when ruin occurs among the realizations and thus provides
an indicator of the solvency of the original state and the particular
strategy applied.
166
The idea of the model is, of course, to experiment with different

strategies and to find the 'best possible'. This provides the defi·
nition of the goals in an exact form. Because the state parameters
are numerous, the problem of determining which state is the best
can be complicated. For example, is it better to secure a high rate
of expansion for the company (measured perhaps by premium
income P) and a lower level of solvency (measured by the free
reserves U) and lower dividends D or vice versa? If the problem is
dealt with mathematically a utility function is needed. In this case
it will be a function of several variables (U, P, D, .. . ).
To operate with the multi·dimensional utility function in an
elementary and illustrative way which circumvents the difficulties
involved with its construction is simply to calculate numerous
examples employing different proposed alternative strategies. The
results are then compared graphically, in other words a 'trial and
error' method is applied. To illustrate the ideas Fig. 13.1 and 13.2
are taken from numerical applications by I>entikii.inen (1976).
In these figures the main state parameters P and free reserves U
are shown. For the sake of simplicity only the mean values for the
end point t = N = 5 years are shown. Thus each strategy always
gives only one point in the figure. Each point is the result of simu·
lation and the solvency aspect is indicated by marking each case
separately where the simulation gave a number of ruins which
exceeded a certain fixed limit, 0'5%.
® No ruin s
o \ I
A-RUins
,
105
100
Fig. 13.1. Safety loading ,\ and sales allocation parameter f (cf. 13.5)
are varying. All other decision parameters are temporarily fixed.
The 'danger area' where ruins occur is shaded.
167
RISK THEORY
® No ruins
->~-RUins
94
92
90
~rl----~I------~I----~I------+---.
170 175 190 195
Fig. 13.2. Sales intensity parameter f3 (included in the function 0)

and the sales allocation parameter f are varying, other decision
parameters are fixed.
When figures like 13.1 and 13.2 have been obtained, it is relatively
easy to see the character of different strategies and the inter-
dependence of the variables involved. This helps in the selection of
the one which most closely corresponds to the goals and wishes of
the management. E.g. if the expansion of the company is a pre-
dominant goal, then the strategy giving as large a premium income
P as possible is 'the best' on condition that the risk of ruin is less
than some prefixed 'tolerable' limit and intended dividends are
achieved. If the maximisation of the reserves U or the dividend D
are of main interest then, of course, corresponding figures are con-
structed showing the quantities of interest. In practice it is likely
that management will aim to achieve a suitable mixture of U, D
and expansion of P.
Dynamic programming is carried out with the aid of computers.
Pentikiiinen (1976) found that simplified models such as those
illustrated in the figures and using two sections (v = 2) required
8 minutes of computer time for each 1000 realisations of each
strategy. Because the number of potential strategies is very large,
owing to the great number of strategy parameters, it is highly
168
desirable to find short cut methods to identify, at least approximately,

the region in the multi-dimensional space where the optimal values
of decision parameters are likely to be found.
A suitable way is to omit temporarily the stochastic element of
the program and to let the computer always select the mean value
only for the random variables involved (amount of claims, seasonal
and other variations). In this way only one outgo, a 'deterministic
flow', is obtained for each strategy. The final values U(t), P(t) . ..
are then close to the mean values of the respective variables - not
quite exactly, however, because of the biassed character of the
probability distributions.
In this way it was found possible to compute the prognosis in only
2 seconds for each strategy. The position of the values, e.g. in the P,
U-plane, gives a good view of the nature ofthe strategies involved.
A serious deficiency of the 'mean value' method outlined above is
that it does not give any information on the solvency aspect. Some
help can, however, be given by making use of the zone of 'probable
flow' illustrated in Fig. 13.3. If the zone intersects the ruin barrier
U 3(t) it is an indication of danger. This method is further developed
by Pentikiiinen (1976) and Hirvonen (1976).
When an approximate optimal strategy has been found, the
results can be tested and adjusted by s~mulation. Thus possible
inaccuracies in the short cut method can be corrected and a con-
siderable saving in computer time effected. By using this approach
it is probable that dynamic programming can be extended to
embrace models which would otherwise be too complicated to handle.
Fig. 13.3. U = Ut ± ka,,(t) are the borders of the zone of probable

flow.
169
APPENDIX A
Derivation of the Poisson Process

and Compound Poisson Processes
1. Poisson Process
Let v(t) be the number of claims in the half-closed interval (0, t]
(t>O). Define v(O) = 0, and let s>O and O~tl <t2~t3<t4' The
conditions below are supposed to be valid:
(i) The variables v(t 2) - V(tl) and v(t 4) - v(t 3) are independent.
(ii) The variables v(s+t)- v(t) and v(s) have the same distribution.
(iii) The probability that more than one claim occurs at the same
time or that an infinite number of claims occur in a finite
interval are both zero.
A somewhat weaker condition can be substituted for condition (ii):
(ii)' P{v(s+t)- v(t) = O} = P{v(s) = O}.
Let Pk(t) = P{ v(t) = k}. For s, t> °according to (i) and (ii):
Po(s+t) = P{v(s+t) = O} = P{v(s) = 0, v(s+t)- v(s) = O}
= P{v(s) = O} P {vet) = O} = Po(s)Po(t)
or
(A.I)
Hence Po(t) is everywhere a non-increasing function. Let sand t
be positive rational numbers. Then according to (A.I)
°
giving the result [Po(t)Jl/t = const. ~ l. This non-negative constant
cannot be 0, since, if Po(t) = for every rational t, then according to
(ii) any interval should contain at least one claim with the probability
I; thus the interval (0, t] would contain an infinite number of claims
170
APPENDIX A
with the probability 1, contrary to (iii). Consequently the constant

can be denoted e- q (q;;;O) giving:
Po(t) = e- qt (q;;; 0) (A.2)
This is true for any positive rational number t. Further, since Po(t)
is everywhere non-increasing, Po(t) cannot have steps and is ac-
°
cordingly continuous. Thus (A.2) is also true for irrational t's.
In order to calculate Pk(t) for k> let h be an integer such that
n = 2h> k and write down the disjoint partition of the interval (0, t]
(i = 1, 2, ... , n)
This partition has the property that when h increases then the
pre-existent division points remain division points. From the set of
all possible realizations of the process (both the number of claims
v and their placement in (0, t] vary) two sub-sets are now selected
for a fixed h as follows:
Ah = the set of all realizations (sample functions) such that at
least one claim occurs in exactly k intervals Ii; that is
to say, realizations for which exactly n - k intervals remain
without claims (now v;;;k)
Bh = the set of all realizations such that at least in one interval
Ii at least two claims occur (hence v;;;2)
Evidently:
since, in order that exactly k claims occur, it is necessary that either

k different intervals must include claims or some interval must
have at least two claims. On the other hand:
since the left-hand side includes only realizations where claims

occur in exactly k intervals and nowhere more than 1. Thus:
or, a fortiori:
P{Ah}- P{B h} ~ Pk(t) ~ P{Ah} + P{B h} (A.3)
171
RISK THEORY
To prove that P{B lI,} ~ 0 let h ~ 00 and let X be an arbitrary

realization. There are two possibilities:
(1) X is such that from a certain value of h every interval contains
at most one claim.
(2) X is such that independently of h there is always at least one
interval where at least two claims occur.
Clearly lim P{B h } can be > 0 only if the realizations ofthe type
(2) have positive probability. There are two ways of belonging to
category (2). Either some division point must contain two (or more)
claims, or otherwise the claim points have to have an accumulation
point. These two possibilities both have probability 0 according to
condition (iii).
Returning to (A.3) it remains to calculate Pk(t) = lim P{Ah}. The
probability that in a fixed It at least one claim occurs is, according
to (A.2):
1-po(tjn) = 1-e- qt1n
Thus the probability that claims occur in exactly k fixed intervals,
is:
(1- e-qtln)k(e-qtln)n-k
These k intervals can be chosen in (~) different ways, so that:

P{Ah} = (~) (1- e-qt1n)k (e-qt1n)n-k
= 1 ( k-1) ( 1)
k! e-qt(1-e-qtln)k n k (eqtk)lln 1--;- ... 1-;;:
~ e- qt (qt)k
k! .
Hence:
(qt)k
Pk(t) = e-qt - (q~O) (A.4)
k!
172
APPENDIX A
According to (A.2) this is also true for k = O. The process v(t) is

thus a Poisson process.
2. Extensions
The theory of Poisson processes can be extended by replacing one
or more of the assumptions (i) to (iii) (page 167) by weaker ones.
The new processes obtained in this way playa central role in the
advanced theory of risk.
As was seen above, the conditions (i), (ii) - or (ii)' - lead to (A.2),
which gives Po(t) as an exponential function of t. Suppose now that
the condition (ii) is substituted by a weaker condition assuming
only that:
(ii)* Po(t) is a continuous function of t tending to 0 for t --* 00.
Obviously Po(t) - which certainly cannot be anywhere increasing

- can also be denoted by e-qT(tl, where T(t) = - (l/q) log Po(t)
is another continuous function, tending to + 00 for t --* 00, and
being nowhere decreasing. The constant q can be chosen so that 1'(1)
= 1. For every positive l' there exists at least one t so that l' = T(t).
The original process v(t) can be considered as a random process of
a new variable 1', denoted V'(T), defining simply v'(r) = v(t), where
t is any such number, that l' = r(t)t
Now let 1'1' 1'2> 0, and let t 1, t 2, and t be such that 1'1 = T(t 1),
1'2 = T(t 2), and 1'1 + 1'2 = T(t). Then according to the condition (i):
P{V'(Tl+T2) = O} = P{v'(rl+r2)-v'(Tl) = O} P{v'h) = O}

hence, because P{v'(r) = O} = e- q7":
P{v'h + 7"2) - V'(Tl) = O} = e- qT 2 = P{v'h) = O}
which proves that the condition (ii)' is verified for the process v'{r).
Accordingly
P{v(t) = k} = e-qT(t) [qr(t)]k (A.5)
k!
t If Po{t) is absolutely decreasing in point t, the correspondence

between T and t is a one-to·one correspondence. Now let (to, t 1 ) be an
interval, where Po{t) is constant. Then for any t within this interval
Po{to) = Po{t) = P {v{t) - v{t o) = O} P {v{to) = O} = P {v{t) - v{to) = O}.
Po{t o)' Thus P{v{t) -v{t o) = O} = 1, and accordingly, with probability
one, v{t) = v{t o), so that V'{T) is also uniquely defined in this case.
173
RISK THEORY
It can be said that the process vet) is a Poisson process in the

transformed new time scale T, in so called operational time.
The conditions (i) to (iii) lead to a process, where only the constant
q remains to be estimated in applications. The weakened condition
(ii)* instead of (ii) leads to a process where the function T(t) remains
as a "parameter" to be estimated or assumed in applications. The
product qT(t) gives, in this case, the expected number of claims in
the interval (0, t]. The derivative T'(t) can be called the intensity
of the process. In applications the intensity can be assumed to be,
for example, increasing in accordance with some prognosis con-
cerning the future volume of the insurance collective in question or,
perhaps, due to the anticipated changes in the frequencies of claims.
The process (A.5) gives an example of proce88e8 with non-8tationary
increment8, also called heterogeneou8 in time whereas the Poisson
process (A.4) is a proce88 with 8tationary increment8, also called
homogeneou8 in time.
A further extension of risk processes is obtained if the constant
q is thought of as a random variable X' which varies owing to some
outer factors, e.g. due to random effects of weather conditions, etc.
Suppose that the claim number vet) satisfies the conditions (i), (ii)*,
and (iii) on condition that X has a given value q. Then the con-
ditional distribution of vet) is again a Poisson distribution.
A more general case is obtained, which is also more realistic,
if X is dependent on time, hence being a general stochastic process
X(t). In order to give a short survey, let q(t) be a realization, i.e. a
sample function of this process, and suppose that for the fixed q(t),
the conditions (i) (ii)*, and (iii) are satisfied. Then again, for any
value of t, P{v(t) = k} (on condition that this sample function q(t)
of the process "occurs") is evidently dependent only on the expected
value of the number of claims in the interval (0, t], i.e. of the
product qT, where q is the value which the sample function takes
for t, but it is not dependent on the values that the sample function
takes for other values of time. Generally, since the operational time
T is calculated separately for different realizations q(t), it is dependent
on the value q as well, thus the notation T = T(t, q) is used in the
following. Hence for the unconditioned process:
P{v(t) = k} = 1o
e-qT(t,q) [qT(~! q)]~ d q V(q, t)
174
APPENDIX A
where the structure function P{x(t) ~q} = V(q, t) IS m general

dependent on time.
Finally the integration variable is changed introducing r =
q'T(t, q)!t and solving q from this equation. Then the solution is placed
in V(q, t) and a new structure function U(r, t) = V(q(r, t),t)* is
constructed. Writing q instead of r, the above probability can be
written:
(qt)k
k!
<XlI
P{v(t) = k} = e- qt dqU(q, t) (A.6)
o
A process which satisfies condition (A.6) is called a compound
Poisson process in the wide sense, or, especially if the structure
function U is independent of time t, a compound Poisson process in the
narrow sense. It might happen that a compound Poisson process in
the wide sense in time t is a compound Poisson process in the narrow
sense in a suitably chosen operational time 'T. t
A process with property (A.6) no longer obeys the conditions (i)
and (ii) in general. It should be remarked that (A.6) does not yet
define all properties of the process v(t), since, for example, it displays
only the distribution of random variables X(t) for various values of
t, but not the way that these variables are dependent on each other.
So it might happen that two processes have the same distribution
(A.6) for every t, but that the probabilities that k claims occur in
the time interval (0, t], on the condition that i claims have occurred
in the time interval (0, t/2], are different.
'" More rigorously, U(r, t) = J dg V(g, t) where the integration is taken

over all such values of g, for which rt ~ g T (t, g).
t An example of this is the process satisfying condition (A.5), which
can be modified into the form (A.6) and which is thus a compound
Poisson process, leading to a structure function dependent on time. In
fact if, for clarity, the integration variable is denoted by x, U(x, t) =
e (x -g T (t)/t), where e(x) denotes the degenerated distribution function.
175
APPENDIX B
The Edgeworth Expansion
If all derivatives of the function G(x) exist and if GCTd ( ± (0) =

all k ~ 1, then:
° for
Ie Ie
+00 +00
ch(8) = dG(k)(x) = [ e
i8X i8X G(k)(x) ]~: - i8 i8X dG(k-l)(x)
-00 -00
= (-i8)ch_l(8) = ... = (-i8)k~o(8)
Since the normal distribution:
<P(Xj m, a) = ~
V (21T)a
-00
f exp[_!(z-m)2]
x
2 a
dz
= <P (x-m )= (x-m)

-a- j 0, 1 </> -a-
clearly satisfies these conditions, the characteristic function of its

kth derivative is:
~k(8; m, a) = (-i8)k ~o(8; m, a) (B.1)
where the characteristic function of </>(x; m, a) is:
~O(8; m, a) = exp (i8m-8 2a 2/2) (B.2)
Evidently:
</>(k)(x; m, a) = (x-m)
a- k </>(k) -a- (B.3)
The characteristic function of the generalized Poisson function

176
APPENDIX B
F(x), given in (3.11), has the following expansion, according to

(B.2):
cP(s) = ent/ls)-n = exp{ n iif!(k)(O)~~+nO(S5)-n}

o
1 { S3 S4
= exp (isnm- 2s 2n"'2) exp nif!(3)(O) 3!+n if!(4)(O) 4!
+n O(S5)}
n"'3 n"'4 n 2"'3 2

= cPo (s; nm, J(n"'2)) { 1 +6 (is)3+ 24 (is)4+ 72 (is)6
+nO(s5)+n 20(s')+n 30(s9)+ ... }+ Remainder
Hence by inverting the characteristic function, according to (B.l):
F(x) = cP (x; nm, J(n"'2))- ~3 n cp(3) (x; nm, J(n"'2)) +

2
;~ n cp(4)(X; nm, J(n", 2))+ "';2 n 2CP(6)(x; nm, J(n"'2))
+ ... + Remainder
where the terms omitted are of the form Onicp(j}(x; nm, J(n"'2)) with
jj2-i'?,3j2, 0 being independent of n. Hence according to (B.3),
z = (x-nm)jJ(n"'2):
1 1
F(x) = <1>(z)-6" "'an (J(n"'2))-a <1>(a)(z)+ 24 "'4n (J(n"'2))-4<1>(4)(z)
+-1 "'a2n2(J(n"'2))-6 <1>(6)(Z) + ... + Remamder

.
72
where the terms omitted are of the form:
0' n i (J(n"'2))-J CP(j}(z) = 0"nH/2 cp(j}(z) = O"n- k CP(j}(z)
with k'?,3j2. Hence:
1 1
F(x) = CP(z)-6 "'a "'2-3/2n-1/2 cp(a)(z) +24 "'4"'2-2n-1 CP(4)(Z)
1
+72 "'a 2"'2- 3n- 1 <1>(6)(Z) +O(n- 3/2 )
177
Solutions to the Exercises
2.5.1. n = 1000,.0,01 .5 = 50, U = £5740
2.5.2. N = 10 900
2.5.3. n = 10, U = £3 180

or by means of the Poisson function, noting that k = 18 is the
smallest integer for which I-FIO(k) ~0'01,
U = (18-10-0·1.10) £500 = £3500
'2 '2
00 00 1
2.5.4. P = 100 (k-2)Pk(l) = 100 (k-2) e~!
k=3 k=3
= 100e- 1
00
1 2 ""
oolJ
~ -k'
[
~---
" "
k=3 (k-I)! k=3'
= 100e- 1 [
00
""
1
--2-2 "" - +5
001 J
~k! ~k!
k=O k=O
= 100e- 1 (e-2-2e+5) ~ £10·4
3.3.1. n = 100 .0·01 = 1

It is convenient for the computations to take £100 as the monetary
unit. S(z) is a step function having a step 2/3 at 1 and 1/3 at 2.
The total amount of the claims can be only a non-negative integer
x = 0, 1,2, ... , N, ... Constructing all the possible combinations
of the sums 1 and 2 which can lead to N, the following expansions
are obtained making use of the abbreviation Pk(I) = Pk (the
178
SOLUTIONS TO THE EXERCISES
difference compared with (3.5) is that N, not k, is taken as the

variable):
F(x) = Po = e- 1 = 0·37 when 0 ~x< 1
2
F(x) = PO+Pl' -3 = 0·61 when 1 ~x<2
F(x) = F(I) +P2 (2)2 1

3" +Pl' 3" = 0·82 when 2~x<3
1
F(x) = F(2)+P2 . 2 . 3"' 3"+P3
2 (2)3
3" = 0·92 when 3~x<4
F(x) = F(3)+P2' GY +P3' 3. GYG)

+P4GY = 0·97 when 4~x<5
The following steps are constructed in a similar way:

F(5) = 0'99, F(6) = 1·00
3.3.2. E{t} = 1 . ( 1 . ~ +2.~) = 1·33 or £133
u = J[ 1 . ( 12 . ~ + 22 . ~ ) ] = 1·41 or £141
3.5.1.1. Denote the unknown rate of death by q. Then the following

values for the step function 8 follow from (3.14).
5000q
8(100)= =0·5; 8(250)=0·6 and 8(500)= 1
5 OOOq+ 1 000q+2. 2000q
3.5.4.1. 8 M (z) =
r8 (-I-p- z
1 ) for z<M
l 1 for z"?,M
3.6.1. When X is not very large compared to P, (3.26) can be
changed approximately into the form:
LJF LJF X
LJe = q - X = q .-- . -
LJx LJx/P P
179
RISK THEORY
It follows from the given data and values read from the figure (a
change of F from 0·001 to 0·002 changes x/p by 0·063):
P = nm = 500 . 1 203 = £601 500
and
0·001 .3.34500
LIe = 0·1 = 0·0003
0·063 . 601 500
5.2.1. By means of a graph like Fig. 5.1 it will be found that

M ~ £1700.
5.2.2. M ~ £1 400.
5.2.3. The basic equations (5.4) for the original and the new
portfolio (cf. 3.28) are respectively:
U = y,j(tJ.2n)-Amn
U = y,j(tJ.2n+0·IM2)-A(mn+0·IM)
, Hence:
5.3.1. The quantity y, is solved from U 1 = ytJ.l,jn1-A1m1n 1 and

put in (5.15). U ~ £24000.
5.3.2. U = yJ('i.PitJ.i 2)- ~)imini

Ut = y,j(nitJ.i 2)-Aim in i
LUi - U = yL,j(nitJ.i2)-y,j(LnitJ.i2)
The result follows from the general inequality ,ja 1+,ja 2 +.Ja 3 + ...
>,j(a 1+a 2+a 3 + ... ) when each ai>O.
5.3.3. U = YtJ.,j(0·9n)-Amn-0·lmn ~ £8500

LlU = 20000-8500 = £11500
The exercise shows how sensitive the risk status is to changes in
(or assumptions regarding) the safety loading included in the
premiums.
180
5.3.4. If £lOO is taken as the monetary unit and the unknown

frequencies of the sums 1 and 2 are denoted by p and I-p then:
m = p+2(I-p) = 2-p
1X2 = p+4(I-p) = 4-3p
U = y""[(4-3p)n]-A(2-p)n
U as a function of p reaches its maximum for
4 3y2
P = 3- 4nA2 ~ 0·52
and the maximum U = 1·5 = £150.
f ze-zdz = I; f z e- zdz = 2. From (5.4) A can

00 00
5.3.5. m = 1X2 = 2
o 0
be solved: A = 0·14.
5.3.6. Because each company participates in each risk, n is the

same for all companies and Si(Z) = S(Pz/Pi ) where S(z) is the
distribution function of the size of the claims of the 'united'
company. The moments appropriate to the ith company are:
IXk W = [ 00
zkdSi(z)
.
= (P)k
pi [ukdS(U) = (P)k
; IXk
00
where IXk is the kth moment attached to the united company. In

accordance with (5.4) the theoretical minimum reserve of the ith
company is:
Pi
Ui = - (YIX""n-Amn)
P
Hence U1 + U2 + . . . + Uris, as can be seen, equal to an expression
that is obtained by applying (5.4) directly to the united company.
5.6.1. dQ(fL)
dfL
= I
i=l
r
aQ dMi =
aMi dfL
(1- y~ I IX" n
) niAi2(I-Si(Mi))
Let fL be a zero of the equation Q(fL) = O. Then 1- (YfL/IX""n) must

181
RISK THEORY
be <0, since, if this expression were non-negative, i.e. y~(a.JnJp.),

the following inequalities would hold:
U =ya.Jn- .2 Atn t1ni ~ a.:n - .2 Atntmt = .2; nta.t 2- .2Atnt1ni
and, observing that p. = MdA, and a.,2~Mt1ni (cf. 5.9):
But this is in contradiction with the assumption in Section 5.6.

that U>O. Hence dQjdp.<O for all zeros p. satisfying p.<p.'( ~ 00),
where p.' is the smallest value for which all the factors I-St(M,)
are = O. On the other hand, by the assumed necessity of reinsurance,
Q(p.) <0 for all p.~p.', so that all zeros p. satisfy the stated property.
A function Q(p.) has thus been found, which has the following
properties: it is continuous, it has everywhere a continuous derivative,
Q(O) = U > 0, Q( 00) < 0, and dQJdp. < 0 for all zeros. Consequently
it has exactly one zero.
5.7.1. 6·7 ~ 7 years
£60000 a. 1500
5.8.1. n = 200 m = - - = £300 - = - - = 5
, 200 'm 300
~200
no = 271.25 = 6775, Z =j-- = 0·172
'\f 6 775
Pl = 0,172.60+0,828.100 = £93
9.2.1. The variance of company i is, denoting g = gt and R = R(gt):

V = E{[g-(r-l)R]2}-E2{g-(r-l)R}
+ (r-l)E{R2}- (r-I)E2{R}
= E{g2}_E2{g} + (r-l)[rE{R2}-rE2{R}-2E{gR}+2E{g}E{R}]
and for the special case R(gt) = gdr
r-l
Vo = E{g2}_E2{g}_ - [E{g2}-E2{m
r
182
Hence:
_r_ (V _ Vol = E{(rR)2} - E2{rR} - 2(E{grR} - E{g}E{rR})

r-1
+ E{g2} - E2{g}
= E{[(g-rR)-E{g-rR}]2} ~ 0; or V ~ Vo
1
9.2.2. (i) Po = .In
I z2dS = n I (z-m)2dS+nm2
00 00
(ii) V = ex2n = n
o 0
The proof follows from the fact that the latter term is the variance
Vo of the distribution having all claims equal to m and the former
term is non-negative. P ~ Po follows from V> V 0 by the given
conditions.
(iii) U = ypP->"P
12.2.1. (i) For a fixed ex = x:
e-RxE{eR'I' ~ x} = e- Rx - -
1-S(x)
1
I eRZdS(z)
00
=
a-R
a
x
Since this is independent of the choice of x, the weighted average
is also independent of the weighting distribution. Hence the
use of an approximate ex-distribution in (12.18) does not cause any
error in this case.
(ii) For a fixed ex = x:
e-RxE{eR'I' ~ x} = e RM- Rx (O<x ~ M)

Clearly ex obtains with a probability > 0 any nondegenerated
subinterval of (O,M), and hence:
1 < 1/G(U) <e RM
From (12.16) it can be verified that:
1+(1+>")RM = e RM >1+RM+R2M2/2
or RM<2>", and hence e RM = 1+(1+>")RM< 1 +2>"+2)..2.
183
RISK THEORY
12.4.1. Let gl and g2 be two mutually independent random variables

having a common distribution function and consequently the same
moments lXi' The required inequality is then equivalent to the
following:
2(lXi-1IXi+1-lXl) = E{U-lg2i+l} +E{gli+1g 2i -l} - 2E{glig2i}
= E{(glg2)i- 1(gl- g2)2} ;;;; 0
12.4.2. The inequality:
-3+v'(9+24A) < A + ~
(A > 0)
2 I+A
can be verified by some calculations. From (12.30) it follows that:
(I+A)x(A)+1 > l+x(A)+ [X(A)J2 + [X(A)]3

2 6
which gives X(A) < t(-3+v'(9+24A)) < A+Aj(I+A) for all A> 0
Hence x(AjK) < AjK +Aj(K +A), which gives:
A(K +A) x(AjK)
o > x(AjK) - --xz . (1 +A/K)X(AjK)-AjK = g(K), say,
since (I+A)x(A)-A = X(A)jx'(A) >0 for A>O
Consider the function f(K) = (K+A)X(AJK) for A>O, O<K ~ 1.
It follows that f'(K) = g(K) < 0, so that f(K) is a decreasing
function for 0 < K ~ 1. Hence f(K) ;;;; f(l) = (1 +A)X(A). This
immediately gives the required result.
12.4.3. From (12.32) the limits £1450<M <£2410 are obtained.

If M is selected from the safe end of the interval, M = £1 450,
then the error of the approximation is at most = £2410-£1 450
= £960.
184
Bibliography
SA = Skandinavisk AktuarietidBkrift
MS = Mitteilungen der Vereinigung SchweizeriBcher VerBicherungB-
mathematiker
AB = The ABtin Bulletin
CA = TranBactionB of the International CongreBB of ActuarieB
BARAB = Bulletin de l'ABBociation RoyaleB deB Actuaire8 BelgeB
AMMETER, R., (1946), 'Das Maximum des Selbstbehaltes in der Lebens-

versicherung unter Beriicksichtigung der Riickversicherungskosten',
MS.
AMMETER, R., (1948), 'A Generalization of the Collective Theory of
Risk in Regard to Fluctuating Basic Probabilities', SA.
AMMETER, R., (1949), 'Die Elemente der kollektiven Risikotheorie von
festen und zufallsartig schwankenden Grundwahrscheinlichkeiten,'
MS.
AMMETER, R., (1953), 'The Calculation of Premium Rates for Excess
of Loss and Stop Loss Reinsurance Treaties', CA.
AMMETER, R., (1954) I, 'Risikotheoretische Methoden in der Riickversi-
cherung', CA.
AMMETER, R., (1954) II, 'La theorie collective du risque et l'assurance
de choses', MS., p. 185.
AMMETER, R., (1955), 'The Calculation of Premium Rates for Excess
of Loss and Stop Loss Reassurance Treaties', Arithbel, SA, Brussels.
AMMETER, R., (1963), 'Experience Rating', AB.
AMMETER, R., DEPOID, P., and DE FINETTI, B., (1959), 'L'Etude
Mathematique des Assurances Non Viagers dans l'Europe', AB.
BEARD, R. E., (1954), 'Some Statistical Aspects of Non-Life Insurance',
J. InBt. ActB. StudentB' Soc., 13, Pt. 3.
BEARD, R. E., (1959), 'Three R's of Insurance - Risk, Retention, and
Reinsurance', J. InBt. ACtB. StudentB' Soc., 15, Pt. 6.
BEARD, R. E., 'Technical Reserves in Non·Life Insurance with Particu-
lar Reference to Motor Insurance'. To be published in AB.
BENKTANDER, G., and SEGERDAHL, C·O., (1960), 'On the Analytical
Representation of Claim Distribution with Special Reference to
Excess of Loss Reinsurance', CA.
BJERRESKOV, s., (1954) I, 'On the Principles for the Choice of Re-
insurance Method and for the Fixing of Net Retention for an Insur-
ance Company', CA.
185
RISK THEORY
BJERRESKOV, S., (1954) II, 'Om kollektiv risikoteorie', DanishSuper-

vising Office Jubilee Volume.
BOHLMANN, G., (1909), 'Die Theorie des mittleren Risikos', CA.
BOHMAN, H., (1964), 'To Compute the Distribution Function when thc
Characteristic Function is Known', SA.
BOHMAN, H. and ESSCHER, F., (1964), 'Studies in Risk Theory with
Numerical Illustrations Concerning Distribution Functions and Stop
Loss Premiums', SA.
BORCH, K., (1960), 'Reciprocal Reinsurance Treaties', AB.
BORCH, K., (1961), 'The Utility Concept Applied to the Theory of
Insurance', AB.
BORCH, K., (1963), 'Economic Theory and their Application to Insur-
ance', AB.
BORCH, K., (1966), 'Control of a Portfolio of Insurance Contracts',
AB.
BORCH, K., (1967), 'The Economic Theory of Insurance', AB.
BUHLMANN, H. (1964), A Distribution Free Method for General Risk
Problems', AB.
BUHLMANN, H., (1967) I, 'Experience Rating and Credibility', AB.
BUHLMANN,H., (1967) II, 'Kollektive Risikotheorien', MS.
CAMPAGNE, C., (1952), 'La theorie mathematique de lareserve technique
dans l'assurance incendie', Institut des Sciences Economiques
Appliquees, Louvain.
COX, D. R., and MILLER, H. D., (1965), The Theory of Stochastic Processes
(Methuen, London).
CRAMER, H., (1926), 'Review of F. Lundberg' (45), SA.
CRAMER, H., (1930), 'On the Mathematical Theory of Risk', Skandia
Jubilee Volume, Stockholm.
CRAMER, H., (1946) I, Mathematical Methods of Statistics (Princeton
Univ. Press).
CRAMER, H., (1946) II, 'Den Lundbergska. riskteorin och teorin fOr
stokastiska processer', F. Lundberg Jubilee Volume, Stockholm.
CRAMER, H., (1954) I, The Elements of Probability Theory and some of
its Applications (Geber, Stockholm) and (Wiley, New York).
CRAM ER, H., (1954) II, On some Questions Connected with Mathematical
Risk, Univ. of California Publ. in Statistics, 2.
CRAMER, B., (1955), 'Collective Risk Theory, a Survey of the Theory
from the Point of View of the Theory of Stochastic Processes',
Skandia Jubilee Volume, Stockholm.
DEPOID, P., (1967), Applications de La statistique aux assurances
accidenl$ et dommages, Editions Berger-Levrault, (Paris).
DOOB, J. L., (1953)), Stochastic Processes, (Wiley, New York).
DUBOURDIEU, J., (1952), Theone mathematique des assurances, Premier
livre: Theorie mathematique du risque dans les assurances de
repartition (Gauthier-Villars, Paris).
ESSCHER, F., (1932), 'On the Probability Function in the Collective
Theory of Risk', SA.
186
BIBLIOGRAPHY
HAFERL, E., (1940), 'Die Bestimmung der Selbstbenalte in der

Lebensversicherung', CA.
HAIGHT, F. A., (1967), Handbook of the Poisson Distribution (Wiley,
New York).
HOVINEN, E., (1964), 'A Method to Compute Convolutions', CA.
HOVINEN, E., (1967), 'Procedure to Compute the Values of the
Generalized Poisson Function', AB.
HULTMAN, K., (1942), 'Einige numerische Untersuchungen auf Grund
der kollektiven Risikotheorie', SA.
JOHANSEN, P., (1964), 'Mathematical Models in General Insurance',
Nye Danske Centenary Volume.
JOHNSON, N. L., NIXON, E., AMOS, D. E., and PEARSON, E. S., (1963),
'Tables of Percentage Points of Pearson Curves for Given V /31' and
/3 2 Expressed in Standard Measure', Biometrika, 50, pp. 459-498.
KAHN, P. M., (1961), 'Some Remarks on a Recent Paper by Borch', AB.
KAUPPI, L. and OJANTAKANEN, P., 'Approximations of the General-
ized Poisson Function', To be published in AB.
KUPPER, J., (1962), Wahrscheinlichkeitstheoretische Modelle in der
Schadenversicherung, Wiirtzburg bei Konrad Triltsch.
LA URIN, I., (1930), 'An Introduction into Lundberg's Theory of Risk',
SA.
LUNDBERG, F., (1909), 'Zur Theorie der Riickversicherung', CA.
LUNDBERG, F., (1919), Teori for riskmassor, Stockholm (Forsakrings-
inspektionen) .
LUNDBERG, F., (1926-1928), FOrsakringsteknisk riskutjamning, Stock-
holm (F. Englund).
LUNDBERG, F., (1930), Approximerad framstallning av sannolikhets-
funktionen. Aterjorsakring av kollektivrisker, Akad, Afhandling,
Uppsala (Almqvist o. Wiksell)_
LUNDBERG, F., (1930), 'Uber die Wahrscheinlichkeitsfunktion einer
Risikenmasse', SA.
LUNDBERG, F., (1932), 'Some Supplementary Researches on the
Collective Risk Theory', SA.
LUNDBERG, F., (1934), 'On the Numerical Application of the Collective
Risk Theory', De Forenade Jubilee Volume, Stockholm.
LUNDBERG, 0., (1954), 'Actuarial Problems Appertaining to Rein-
surance of Disability Risk', CA.
LUNDBERG, 0., (1964), 'On Random Processes and their Application
to Sickness and Accident Statistics', Reprint of thesis 1940, Uppsala,
Almqvist & Wiksells Boktryckeri.
LUNDBERG, 0., (1967), 'Une note sur des Systemes de Tarification
bases sur des Modeles du type Poisson compose'.
MOLINA, E. C., (1942), Tables oj the Individual and Cumulative Terms
oj Poisson Distribution, D. van Nostrand Co., Princeton, N.J.
PEARSON, K., (1954), 'Tables of Incomplete r-Function', Biometrika
Office.
187
RISK THEORY
PENTIKAINEN, T., (1947), 'Einige numerische Untersuchungen tiber

das risikotheoretische Verhalten von Sterbekassen, SA.
PENTIKAINEN, T., (1952), 'On the Net Retention and Solvency of
Insurance Companies', SA.
PENTIKAINEN, T., (1954), 'On the Reinsurance of an Insurance
Company, GA
PENTIKAINEN, T., (1962), 'Reserves of Motor Vehicle Insurance in
Finland', AB.
PENTIKAINEN, T., (1967), "On the Solvency of Insurance Companies',
AB.
PESONEN, E., (1964), 'Solvency Measurement', GA.
PESONEN, E., (1965), 'A Modification of the Esscher Method', SA.
PESONEN, E., (1967) I, 'On the Calculation of the Generalized Poisson
Function', AB.
PESONEN, E., (1967) II, 'On Optimal Properties of the Stop Loss
Reinsurance', AB.
PESONEN, E., (1967) III, 'Magnitude Control of Technical Reserves
in Finland', AB.
PHILIPSON, c., (1961), 'Note on the Background to the Subject: Theory
of Risk, Fundamental Mathematics and Applications', AB.
PHILIPSON, c., (1964) I, 'On the Risk Theory of Motor Insurance',
SA.
PHILIPSON, c., (1964) II, 'Sur la transformation d'un processus de
Poisson (selon les methodes de P. Thyrion et F. Esscher)" BARAB.
PHILIPSON, c., (1965), 'A Generalized Model for the Risk Process
and its Application to a Tentative Evaluation of Outstanding
Liabilities', AB.
PHILIPSON, c., (1967) I, 'Note on the Relation between Compound
and Composed Poisson Processes', AB.
PHILIPSON, c., (1967) II, 'Comments on Different Deduction of
Conditional Expectation', AB.
PHILIPSON, c., (1968), 'Some Compound Poisson Processes Applicable
in the Theory of Risk', GA.
RIEBESELL, P., (1940), 'Die mathematischen Grundlagen der Sach-
versicherung', GA.
SAXEN, T., (1948), 'On the Probability of Ruin in the Collective Risk
Theory for Insurance Enterprises with only Negative Risk Sums',
SA.
SAXEN, T., (1951), 'Sur les mouvements aIeatoires et Ie probleme de
ruine de Is. theorie de risque collective', Soc. Sc. Fennica, Gomm.
PhY8.-Math., 16.
SAXER, w., (1958), Ver8icherungsmathematik II, (Springer Verlag,
Berlin), (Ch. III Mathematik AUgemeiner Risikover8wherungen).
SEGERDAHL, C.-O., (1939), 'On Homogeneous Random Processes and
Collective Risk Theory', Thesis, Stockholm.
SEGERDAHL, C.-O., (1942), 'Ober einige risikotheoretische Fragestel-
lungen', SA.
188
BIBLIOGRAPHY
SEGERDAHL, c.·o., (1948), 'Some Properties of the Ruin Function in

the Collective Theory of Risk', SA.
SEGERDAHL, c.·o., (1959) 'A survey of results in the collective theory
of risk', Probability and Statistics, The Harald Cramer Volume,
Uppsala.
SIMONSEN, w., (1946), 'Om grundlaget for den kollektive risikoteori',
F. Lu7!dberg Jubilee Volume, Stockholm.
THEPAUT, A., (1954), Essai de determination pratique du plein de
conservation (Dulac, Paris).
THYRION, P., (1959), 'Sur une proprieM des processus Poisson
generalises', BARAB.
THYRION, P., (1960), 'Note sur les distributions "par grappes" "
BARAB.
THYRION, P., (1963), 'Note sur une transformation des distributions
generalisees', BARAB.
THYRION, P., (1964), 'Les lois exponentielles composes', BARAB.
THYRION, P., (1967), 'Regards sur Ie Developpement recent de Is
TMorie du Risque', AB.
WALD, A., (1947), Sequential Analysis, New York.
WILHELMSEN, L., (1954), 'On the Stipulation of Maximum Net
Retentions in Insurance Companies', CA.
WYSS, H., (1953), 'Die Risikotheorie und ihre Bedeutung fiir die
Versicherungsmathematik', MS.
WYSS, H., (1954), 'Die Risikotheorie als Grundlage fiirdieVersicherungs.
mathematische Behandlung von Riickversicherungsproblemen', CA.
At the time of printing of this book a comprehensive bibliography
of the theory of risk is being prepared by Philipson.
A comprehensive review of the modern theory of risk compiled by
several authors will be published by the Filip Lundberg Foundation in
the near future as well as the proceedings of a symposium on this topic
held in Stockholm in March 1968.
Additional Bibliography
BELLMAN, R., and KALABA, R., (1965), Dynamic Programming a7!d
Modern Control Theory (Academic Press, New York and London).
FORRESTER JAY, w., (1961), 17!dustrial Dynamics (Massachusetts
Institute of Technology).
HIRVONEN, T., (1976), Applications of the dynamic prognosis method
C.A.
HOVINEN, E., (1969), 'Basic Statistics to be used in control of equal.
isation reserves'. A.B.
NEMHAUSER, G. L., (1966), 'Introduction to Dynamic Programming'.
(John Wiley & Sons).
PENTIKAINEN, T., (1975), 'Stochastic·Dynamic Prognosis'. S.A.J.
1975.
PENTIKAINEN, T., (1976), 'Stochastic.Dynamic Prognosis'. C.A.
189
Author Index
(See also alphabetical list of authors in Bibliography)

Ammeter, 71, 75, 124, 126, 131, Kahn, 103
142 Kauppi, 45, 47, 48, 82, 122
Arfwedson,142
Lundberg, F., xi, 7
Benktander,37 Lundberg, 0., 75, 120
Bohlmann, xi
Bohman, 47, 51, 81, 98, 124, 127, Miller, 3
128, 131 Molina, 11
Borch, 54,158,162
Biihlmann,75 Ojantakanen, 45, 47, 48, 82, 122
Cramer, xi, xiv, 4, 142 Pearson, 125

Cox, 3 Pentikiiinen, xii, 79, 87
Pesonen, 51, 99, 105
Doob,3 Philipson, 75, 116, 121
Esscher, 47, 51, 76, 79, 81, 124, Saxer, 142

127,128,131 Segerdahl,xi,37,142
Haferl,15 Thyrion,116
Haight, 116
Hovinen,97 Wald,138
191
Subject Index
Addition theorem, 7 Distribution free limit, 59

ASTIN, xiii Distribution function of one claim,
Axioms of von Neuman and Mor- 18-21,25,110,144
genstern, 161 Divergent series, 45
Dynamic programming, 163
Beta distribution, 102
Bessel function, 35, 100 Edgeworth series, 41-45, 76
Binary number, 95 expansion, 42,101,122-4, 176-7
Binomial distribution, 39, 93, 135 Epidemics, 9
Braking constant, 73 Esscher functions, 78-80
Burning cost, 68 Esscher method (formula), 43, 48,
Business planning, 160-9 50-51,76-90,99,101,123,127,
131, 149
Central limit theorem, 12, 41, 67, Excess of loss reinsurance, 15, 17,
120 19, 31,37, 49,54,66,90
Characteristic curve, 89 Exclusion of multiple events, 8-10
Characteristic function, 24, 25, 38, Experience rating, 69-75
39,42,77,98,121,173 Exponent polynomials, 100-10 I
Chebycheff's inequality, 142 Exponential functions, 33-35, 98,
Claim distribution, 18, 27, 28, 34, 151
48,54-7,62,87,89
Classical theory, xi Fluctuating basic probabilities,
Collective method, 7 112,124,145
Compound Poisson process, 114, Fluctuation constant, 124
116,117,121-2,124,126,128, Fourier transforms, 98
130, 134, 139, 170, 175
Convergent series, 45 Games, theory of, 54, 138
Convolution, 21, 24, 33, 36, 38-9, Gamma distribution, 102
77,92,95-97,99,133,140 Generalized Poisson function (see
Correlation coefficient, 105 Poisson function)
Credibility theory, 69, 72, 73
partial,74 Hidden reserves, 52, 55
Decision theory, 161 Incomplete r function, II, 102,

Decomposition of distribution 124,128
function, 37-40, 99,100 Independence of increments, 8
Degenerate distribution function, Inflation, 27, 89, 117, 164
26,122, 172 Insolvency constant, 144-5
Deterministic process, 114 function, 143
193
SUBJECT INDEX
Insurance collectives, 61, 119 Operational time, 174

International Actuarial Associa· Oscillations, 111·2, 114, 130, 134·5
tion, xiii
International Congress of Actu· Pareto distribution, 37
aries, xi Partial credibility, 74
Inverse transforms, 98 Pearson system, 37, 102
Inversion (of characteristic func. Poisson distribution, 10·3, 15, 21,
tion), 76, 81, 98, 124, 128, 177 39,99
Poisson function, 8, 9, 22, 97, 100
Lagrange method, 65 generalized, 18, 21·25, 34, 38,
Laplace transforms, 25 41,47, 57,70,74,76·7,94,99,
Large claims, 29·30 103, 109, 115, 118, 121.2, 128,
Logarithmic normal distribution, 130, 134, 139, 144·6, 152, 154,
37 176
elementary, 22, 110, 116, 119
MacLaurin series, 42 non·elementary,22
Maximum error, 93 Poisson process, 8, 110, 124, 167
Merging of companies, 61·3 compound, 114, 116·7, 121·2,
Mixed Poisson processes, 116 124,126, 128, 130, 134, 139,
Mixed exponential polynomials, 36 167
Mixed risk sums, 4 narrow sense, 116, 175
Moment generating function, 139, wide sense, 116, 126, 175
144 heterogeneous, 174
Monte Carlo method, 48, 50, 76, homogeneous, 174
82,91·7,101,124,134, 163 mixed,116
Multiple events, 8, 10 Poisson variable, 16
Polya process, 116, 122, 124, 126·8,
Negative risk sums, 4 131, 139, 145, 147, 151
Net retention, xii, 15·7, 31, 40, Polya - Eggenberger probability
48·9, 54, 55, 64·5, 67, 85, 105, function, 126
128, 148, 161, 163, 166 Positive risk sums, 4
Newton's method, 44 Premium, risk, 1
No claim bonus, 69 Process - stochastic, 3, 114
Non·linear transformation 10 Processes with non·stationary in·
Normal approximation, 15, 41·2, crements, 174
47, 50, 52·3, 65, 69, 72, 73, 88, Process with stationary incre·
90,100,103,122,127 ments, 174
Normal distribution function, 12,
41, 120, 176 Quota share reinsurance, 31, 90
Normal Power expansion, 43, 45,
47·8, 50·4, 76, 82, 101, 122, Random numbers, 91.4
124,127,136 processes, 2·3, 8
N.P method (see Normal Power sample, 91·2
expansion) walks, 3
Reciprocity, xii, 106
One.step function, 118 Rectangular distribution, 91
194
SUBJECT IND EX
Reinsurance, xii, 9, 62, 67, 103, Sequential analysis, 138

105, 143, 155, 156 Shadow claim, 30, 39-40
Excess ofIoss, 15, 17, 19,31,37, Simulation (techniques) 92, 94
49,54,66,90 Sliding premiums, 72
Quota share, 31, 90 Solvency margin, xii, 63
Stop loss, 15, 31, 54, 71, 104-5, Stability, 103
108, 112, 129-30 Stationariness of increments, 8, 9
Surplus, 31,54,85,90 Statistical method, 101-2
Relative Mean risk, 109 Statutory basis of reserves, 63
Reserves, insurance, 13, 58-64, Step function approximation, 99-
128,156 101, 117
Risk premium, 1 Stieltjes integral, 20
Risk reserve, 2 Stochastic process, 3, 114
Risk Theory symposium, 1968, 51 Stop loss reinsurance, 15, 31, 54,
Ruin function-finite time, 132-6 71, 104-5, 108, 112, 129-130
infinite time, 137·159 Surplus reinsurance, 31, 54, 85, 90
Ruin probability, 5, 40, 48, 55, 62
Theory classical, xi
103-4, 132, 135, 137-8, 143,
Theory of games, 54, 138
147, 150·1, 154, 160
Trends, Ill, 114, 134-5, 156, 158
Rule of greatest retention, 64
Utility function, 54, 158, 161-3,
Safety loading, xii, 1-2, 13-4, 51-3,
167
60, 64, 67, 70, 105, llO, 128,
135, 137, 143, 148, 160, 164 Variance, 23,103·9,119
Sample function, 3 Varying basic probabilities,110-131
Security reserve, 14
]95

Beard Risk Theory - The Stochastic Basis

Uploaded by

Copyright:

Available Formats

Beard Risk Theory - The Stochastic Basis

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Beard Risk Theory - The Stochastic Basis

Uploaded by

Copyright:

Available Formats

What is the topic of the book?

What is the topic of the book?

What are some of the main concepts discussed in the book?

What are some of the main concepts discussed in the book?

MONOGRAPHS

ON APPLIED PROBABILITY AND STATISTICS

General Editor8: M. S. BARTLETT, F.R.S.

R. E. BEARD, O.B.E., F.I.A., F.I.M.A.

CHAPMAN AND HALL

Preface to Second Edition page xi

2. Process with Constant Size of One Claim

3. Generalized Poisson Distribution

4. Normal Approximation and Edgeworth Series for F(x)

5. Applications of the Normal Approximation

6. The Esscher Approximation

7. Monte Carlo Method

8. Other Methods of Calculating the Generalized Poisson

10. Varying Basic Probabilities

Since the publication of Risk Theory in 1969, there has been a

whioh the developments are appropriate in an elementary text book

Definitions and Notations

1.1. The Purpose of the Theory of Risk

Risk rese rye

1.2. Random Processes in General

1.3. Positive and Negative Risk Sums

Positive risk sums

Nega.tive risk sums

to an upward step. This latter type is called a risk process with

1.4. Main Problems

Fig. I.4(a). Checking only at end of the observation period T

Process with Constant Size

2.2. The Poisson Process

where Pk(t} is the probability that exactly k events occur in the

2.3. Discussion of Assumptions

Condition (i) can, however, often be met by defining, as is customary

be dealt with by simply adding a cautionary amount to the constant q.

2.4. Numerical Calculations

where F(x) is the probability that the number of claims g ~ x and

E{s} = .2 k n-k! e-n = n

TABLE 2.1. Comparison between Poisson distribution (F)

If n is not small, use may be made of the central limit theorem

F(z) ~ ~ ( z;nn ) (2.3)

where ~ denotes a normal distribution function with zero mean and

payable on death for funeral expenses, or, if level annual premiums

n = expected number of claims

The conditions of the problem imply:

F(x) = l-e (2.4)

and F(x) is the Poisson distribution function (2.2). For illustrative

F(x) ~ (/J (x-ns)

since the size of one claim is S.

no reserve U is required since the safety loading is more than

Exercise 2.5.1. A friendly society has 1000 members. In the event

Exercise 2.5.4. A friendly society grants funeral expense benefits

i.e. the expected number of claims of amount at least M. tM is

'/.1M ;:;; >..nM-y•.jnM

For a rational reinsurance arrangement the quantity , should

Hence the inequality (2.12) is a criterion as to whether the net

As an illustration let the mean probability of a claim q be 0'01,

Generalized Poisson Distribution