MONOGRAPHS ON

STATISTICS AND APPLIED PROBABILITY

General Editors
D.R. Cox, D.V. Hinkley, D. Rubin, B.W. Silverman

Stochastic Population Models in Ecology and Epidemiology

M.S. Bartlett (1960)

2 Queues D.R. Cox and W.L. Smith (1961)

3 Monte Carlo Methods 1.M. Hammersley and D.C. Handscomb (1964)

4 The Statistical Analysis of Series of Events D.R. Cox and

P.A. W. Lewis (1966)

5 Population Genetics W.J. Ewens (1969)

6 Probability, Statistics and Time M.S. Bartlett (1975)

7 Statistical Inference S.D. Silvey (1975)

8 The Analysis of Contingency Tables B.S. Everitt (1977)

9 Multivariate Analysis in Behavioural Research A.E. Maxwell (1977)

10 Stochastic Abundance Models S. Engen (1978)

11 Some Basic Theory for Statistical Inference E.J.G. Pitman (1979)

12 Point Processes D.R. Cox and V. Isham (1980)

13 Identification of Outliers D.M. Hawkins (1980)

14 Optimal Design S.D. Silvey (1980)

15 Finite Mixture Distributions B.S. Everitt and D.J. Hand (1981)

16 Classification A.D. Gordon (1981)

17 Distribution-free Statistical Methods 1.S. Maritz (1981)

18 Residuals and Influence in Regression R.D. Cook and S. Weisberg (1982)

19 Applications of Queueing Theory G.F. Newell (1982)

 20 Risk Theory, 3rd edition R.E. Beard, T. Pentikainen and
E. Pesonen (1984)

21 Analysis of Survival Data D.R. Cox and D. Oakes (1984)

22 An Introduction to Latent Variable Models B.S. Everitt (1984)

23 Bandit Problems D.A. Berry and B. Fristedt (1985)

24 Stochastic Modelling and Control M.B.A. Davis and R. Vinter (1985)

25 The Statistical Analysis of Compositional Data J. Aitchison (1986)

26 Density Estimation for Statistical and Data Analysis

B. W. Silverman (1986)

27 Regression Analysis with Applications G.B. Wetherill (1986)

28 Sequential Methods in Statistics, 3rd edition G.B. Wetherill (1986)

29 Tensor Methods in Statistics P. McCullagh (19~7)

30 Transformation and Weighting in Regression R.J. Carroll and

D. Ruppert (1988)

31 Asymptotic Techniques for Use in Statistics O.E. Barndoff-Nielsen

and D.R. Cox (1989)

32 Analysis of Binary Data, 2nd edition D.R. Cox and E.J. Snell (1989)

33 Analysis of Infectious Disease Data N.G. Becker (1989)

34 Design and Analysis of Cross-Over Trials B. Jones and

M.G. Kenward (1989)

35 Empirical Bayes Methods, 2nd edition J.S. Maritz and T. Lwin (1989)

36 Symmetric Multivariate and Related Distributions K.-T. Fang,

S. Kotz and K. Ng (1989)

37 Generalized Linear Models, 2nd edition P.McCullagh and

J.A. Neider (1989)

38 Cyclic Designs J.A. John (1987)

39 Analog Estimation Methods in Econometrics C.F. Manski (1988)

40 Subset Selection in Regression A.J. Miller (1990)

41 Analysis of Repeated Measures M. Crowder and D.J. Band (1990)

 Risk Theory
THE STOCHASTIC BASIS OF INSURANCE

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

Leicestershire, England

T. PENTIKAINEN
PHIL. Dr, PROFESSOR h.c.

Helsinki, Finland

E. PESONEN
PHIL. Dr
Helsinki, Finland

THIRD EDITION

CHAPMAN AND HALL

LONDON· NEW YORK· TOKYO· MELBOURNE· MADRAS
UK Chapman and Hall, II New Fetter Lane, London EC4P 4EE
USA Chapman and Hall, 29 West 35th Street, New York NYIOOOI
JAPAN Chapman and Hall Japan, Thomson Publishing Japan,
Hirakawacho Nemoto Building, 7F, 1-7-11 Hirakawa-cho,
Chiyoda-ku, Tokyo 102
AUSTRALIA Chapman and Hall Australia, Thomas Nelson Australia,
480 La Trobe Street, PO Box 4725, Melbourne 3000
INDIA Chapman and Hall India, R. Sheshadri, 32 Second Main Road,
CIT East, Madras 600 035

First edition 1969

Second edition 1977
Third edition 1984
Reprinted 1987, 1990
© 1969, 1977, 1984 R.E. Beard, T. Pentikiiinen, E. Pesonen
ISBN 978-94-011-7682-8 ISBN 978-94-011-7680-4 (eBook)
DOI 10.1007/978-94-011-7680-4

All rights reserved. No part of this publication may be

reproduced or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording or otherwise, or
stored in any retrieval system of any nature, without the written
permission of the copyright holder and the publisher, application
for which shall be made to the publisher.
British Library Cataloguing in Publication Data
Beard, R.E.
Risk theory: the stochastic basis of
insurance.--3rd ed.-{Monographs on statistics and
applied probability)
I. Risks (Insurance)
I. Title II. Pentikiiinen, T.
III. Pesonen, E. IV. Series
368 HG8782

Lihrary of Congress Cataloging in Publication Data

Beard, R. E. (Robert Eric)
Risk theory.
(Monographs on statistics and applied probability)
Bibliography: p.
Includes indexes.
I. Insurance-Mathematics. 2. Risk (Insurance)
3. Stochastic processes. I. Pentikiiinen, Teivo.
II. Pesonen, E. (Erkki) III. Title. IV. Series.
HG878J.B34 1984 368 83-25180
 Contents

Preface IX

Nomenclature XIV

1 Definitions and notation 1

1.1 The purpose of the theory of risk I
1.2 Stochastic processes in general 5
1.3 Positive and negative risk sums 6
1.4 Main problems 7
1.5 On the notation 11
1.6 The moment generating function, the character-
istic function, and the Laplace transform 15

2 Claim number process 18

2.1 Introduction 18
2.2 The Poisson process 19
2.3 Discussion of conditions 20
2.4 Some basic formulae 22
2.5 Numerical values of Poisson probabilities 23
2.6 The additivity of Poisson variables 27
2.7 Time-dependent variation of risk exposure 29
2.8 Formulae concerning the mixed Poisson
distribution 35
2.9 The Polya process 38
2.10 Risk exposure variation inside the portfolio 43

3 Compound Poisson process 47

3.1 The distribution of claim size 47
3.2 Compound distribution of the aggregate claim 50
3.3 Basic characteristics of F 52
vi CONTENTS

3.4 The moment generating function 59

3.5 Estimation of S 60
3.5.1 Individual method 60
3.5.2 Statistical method 62
3.5.3 Problems arising from large claims 65
3.5.4 Analytical methods 67
3.5.5 Exponential distribution 68
3.5.6 Gamma distribution 69
3.5.7 Logarithmic-normal distribution 72
3.5.8 The Pareto distribution 74
3.5.9 The two-parametric Pareto and the quasi-
log-normal distributions 76
3.5.10 The family of Benktander distributions 79
3.5.11 Other types of distribution 83
3.6 The dependence of the S function
on reinsurance 84
3.6.1 General aspects 84
3.6.2 Excess of loss reinsurance- 85
3.6.3 Quota share reinsurance 85
3.6.4 Surplus reinsurance 88
3.6.5 Technique using the concept of degree of loss 90
3.7 Decomposition of the portfolio into sections 94
3.8 Recursion formula for F 100
3.9 The normal approximation 104
3.10 Edgeworth series 107
3.11 Normal power approximation 108
3.12 Gamma approximation 121
3.13 Approximations by means offunctions belonging
to the Pearson family 123
3.14 Inversion of the characteristic function 124
3.15 Mixed methods 124

4 Applications related to one-year time-span 126

4.1 The basic equation 126
4.2 Evaluation of the fluctuation range of the annual
underwriting profits and losses 130
4.3 Some approximate formulae 138
4.4 Reserve funds 142
4.5 Rules for the greatest retention 145
CONTENTS vii

4.6 The case of several Ms 152

4.7 Excess of loss reinsurance premium 154
4.8 Application to stop loss reinsurance 156
4.9 An application to insurance statistics 159
4.10 Experience rating, credibility theory 162

5 Variance as a measure of stability 171

5.1 Optimum form of reinsurance 171
5.2 Reciprocity of two companies 175
5.3 Equitability of safety loadings: a link to theory of
multiplayer games 179

6 Risk processes with a time-span of several years 183

6.1 Claims 183
6.2 Premium income P(l, t) 198
6.3 Yield of investments 205
6.4 Portfolio divided in sections 211
6.5 Trading result 214
6.6 Distribution of the solvency ratio u 220
6.7 Ruin probability 'I'T(U), truncated convolution 227
6.8 Monte Carlo method 233
6.8.1 Random numbers 233
6.8.2 Direct simulation of the compound
Poisson function 239
6.8.3 A random number generator for the
cycling mixed compound Poisson variable
X(t) 241
6.8.4 Simulation of the solvency ratio u(t) 245
6.9 Limits for the finite time ruin probability 'P T 250

7 Applications related to finite time-span T 258

7.1 General features of finite time risk processes 258
7.2 The size of the portfolio 263
7.3 Evaluation of net retention M 265
7.4 Effect of cycles 266
7.5 Effect of the time-span T 266
7.6 Effect of inflation 267
7.7 Dynamic control rules 272
7.8 Solvency profile 278
viii CONTENTS

7.9 Evaluation of the variation range of u( t) 281

7.10 Safety loading 284

8 Risk theory analysis of life insurance 288

8.1 Cohort analysis 288
8.2 Link to classic individual risk theory 292
8.3 Extensions of the cohort approach 295
8.4 General system 300

9 Ruin probability during an inrmite time period 308

9.1 Introduction 308
9.2 The infinite time ruin probability 311
9.3 Discussion of the different methods 315

10 Application of risk theory to business planning 319

10.1 General features of the models 319
10.2 An example ofrisk theory models 322
10.3 Stochastic dynamic programming 330
10.4 Business objectives 336
10.5 Competition models 345

Appendixes 349
A Derivation ofthe Poisson and mixed Poisson processes 349
B Edgeworth expansion 355
C Infinite time ruin probability 357
D Computation of the limits for the finite time ruin
probability according to method of Section 6.9 367
E Random numbers 370
F Solutions to the exercises 373

Bibliography 396

Author index 403

Subject index 405

 Preface

The theory of risk already has its traditions. A review of its classical
results is contained in Bohlmann (1909). This classical theory was
associated with life insurance mathematics, and dealt mainly with
deviations which were expected to be produced by random fluctua-
tions 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 insurance, to which
risk theory has, in fact, its most rewarding applications, was mainly
outside the field of interest of the risk theorists. 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 (1909, 1919), which, thanks to H. Cramer (1926), e.O.
Segerdahl and other Swedish authors, became generally known as
the 'collective theory of risk'. As regards questions of insurance, the
problem was essentially the study ofthe progress of the business from
a probabilistic point of view. In this form the theory has its applica-
tions to non-life insurance as well as to life insurance. This new way
of expressing the problem has proved fruitful. In recent years the
fundamental assumptions of the theory, and the range of applica-
tions, have been significantly enlarged. The advancement of the
general theory of stochastic processes and its numerous sub-
branches and applications has been reflected in the development of
risk theory. The explosive development of computers has made it
x PREFACE

feasible to treat problems which previously could not be handled

because of their complicated structure. For example, it is now
possible to create models to describe the insurance business as a
whole, and the interactions between its sectors, instead oflimiting the
consideration to isolated sub-problems like the range of risk
fluctuations, reinsurance, safety loadings, reserve funds, and so on.
Today the theory of risk generates an interesting and far-reaching
field for research. 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. A defect, 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 towards an understanding of the insurance
business as a process characterized by varying progress and fluctua-
tions 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
company's solvency, reinsurance requirements, safety loadings in
premiums and many others, are such that risk theory alone is
incapable of providing a definite solution. This is 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
arrangements, for example, attention has to be given to many
political aspects of insurance such as reserves, reciprocity and liquid-
ity. In spite ofthis, when choosing a form of reinsurance and calculat-
ing suitable net retentions and safety loadings, and in business
planning in general, 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
PREFACE xi

elementary textbook of risk theory in the Finnish language in

1955, primarily designed for the use of Finnish actuaries as an
introduction to the theory. Many participants at meetings of ASTIN
(Actuarial Studies in Non-Life Insurance- a section of the Inter-
national Actuarial Association) expressed a wish for a concise book
of this kind in English and devoted primarily to practical applica-
tions. The authors attempted to produce such a book. The first
edition was completed in 1969. The basis of compilation was that
the Finnish authors rewrote and brought up-to-date the earlier
Finnish textbook and passed it to the English author, who, for his
part, worked it into shape, taking into consideration British circum-
stances and paying special attention to general actuarial education
in English-speaking countries.
Since the publication of the first edition there has been a continued
growth of interest in the subject. ASTIN now has well over a thou-
sand 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 important part played
by the theory of risk.
The present and essentially renewed third edition has been
worked very much along the same lines, taking into account recent
developments of theories and applications. The compilation of the
text was mainly made in connection with the university lectures of
Pentikainen, and many of the new approaches were developed and
tested in a comprehensive research work concerning solvency and
the reserve technique (Pentikainen 1982, Rantala 1982).
Since the first edition was published several other textbooks on
risk theory have become available, for example Biihlmann (1970),
Gerber (1979) and Seal (l966b and 1978). The existence of these
excellent presentations of risk theory has very much facilitated the
compilation of the third edition of this book. Because the other
authors have followed more strictly mathematical lines, it has been
appropriate to assume a more pragmatic approach here. The authors
hope that all these books will complement each other and give the
interested reader a more comprehensive view of the theory than
would be possible in any single monograph.
To prevent the book becoming too large and developing beyond
xii PREFACE

the limits of a primary textbook, it has been necessary to limit the

subject matter. This has been a very difficult task, due to the very
abundant field which the theory and its applications cover 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 danger of over-
simplification, which could reduce the theory to 'pseudoscience';
moreover, 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. Bearing in
mind 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
step for 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 the elementary aspects of probability theory. On
the other hand, familiarity with the axiomatic theory of probability
and stochastic processes is not assumed and the text is compiled
accordingly, not covering, for instance, strictly axiomatic existence
proofs of the processes concerned, as they are irrelevant to the aims
of the book.
In order to render the book more readable, as far as possible,
for those with different interests and different degrees of familiarity
with probability calculus, the parts of the text that utilize special
techniques, or consider details of limited interest only, are marked **
PREFACE xiii

or referred to appendixes. These sections as well as many of the proofs,

can be omitted on the first reading and the reader can proceed
directly to the final formulae.
Martti Pesonen, PhD, participated in the final collation of the
text.
The authors are grateful to the numerous experts in various
countries for the generous donation of very helpful criticisms and
advice.
The typing of the manuscript for printing, taken from numerous
texts and formulae, was undertaken by Brita Aalto, to whom we
owe our special thanks for accurate work requiring great patience.
Last but not least, we wish to thank Chapman and Hall Ltd and
in particular Mr Richard Stileman and Ms Mary Ann Ommanney
for their generous cooperation in the preparation of the book and
for the final linguistic checking of the text.

Leicestershire and Helsinki Robert Eric Beard

April, 1983 Teivo Pentikiiinen
Erkki Pesonen

We are sorry to announce that our co-author and colleague of

long standing Professor Robert E. Beard passed away in November,
1983 without having witnessed the completion of this new edition
of Risk Theory.
In grateful recognition for his life-long work in the development
and advancement of non-life insurance mathematics we dedicate
this book to the memory of Professor Beard.
T.P.
E.P.
 Nomenclature

Equations in parentheses, items without parentheses.

For principles of the notations see Section 1.5.
Many of the variables may be both stochastic or deterministic. For
brevity this feature is not specified, i.e. bold face letters are not used.

A(v) discounted claims (8.2.7). F(X) dJ. of aggregate claims (3.2.3).

aj zero-moments of claim size (3.3.1). F(X; tj' t z ) dJ. of aggregate claims
B gross premiums = P;/(1 - c) for period [tj' tzJ (6.1.23).
(6.2.8). F(x) dJ. of the standardized
B(t!, t z ) gross premiums from period aggregate claim (3.9.4).
[t!, tzJ (6.4.6). f(X) density of X (3.11.18).
bj zero-moment of loss degree f = X/B claims ratio (6.S.4).
(3.6.16). f(1, t) accumulated claim ratio
C loading for expenses of administra- (6.6.S).
tion = cB gross or net 6.S(a), G utility dJ. 1O.4(d).
(8.1.2), 10.2(1). G(u o' u; tl' t z ) (6.6.17a).
c loading factor for expenses 6.2(f), G,(u) dJ. of u(t) (6.6.17b)
(6.S.4). GNP gross national product
c j factor synchronizing inflation G(r) loss degree dJ. (3.6.13).
(7.6.2). G(x) Edgeworth expansion (3.10.1)
cl' C z control caetI. (7.7.1). g = y/6 3.11(h).
D(t) dividends 6.S(a), 1O.2(g); time H(q) structure d.f. (2.7.1).
lag function (6.6.7). h.Polya parameter (2.9) or Yz/24
d = D/B (10.2.8). 3.1I(h).
d(t) number of deaths (8.1.4). I(t) investment income 6.S.(a),
dJ. distribution function 1O.2(d).
E( ) expected value. i growth rate specified by suffices
Eu equilibrium level of u (6.6.9). ig growth rate 6.1 (c).
e base of nat. logarithms = 2.718282. ijgp = ijrgp (6.S.8a)
F(k) = F(k ; n) dJ. ofthe claim number i j interest rate 6.3(a).
(2.4.1). ip rate of premium inflation 6.2(a).
P(k) mixed dJ. of the claim number ix rate of claim inflation 6.1(j).
(2.7.4). J number of sections or classes 3.7(a).
NOMENCLATURE xv

j current index, section index 3.7(a). q claims number intensity (2.6.5),

K error factor (4.3.5). structure variable 2.7(c).
k number of claims 2.1(a). qi discrete claim size frequency
L(s) Laplace transform (1.6.9). (3.8.2).
[(t) cohort size (8.l.l). q(t) mortality (8.1.1).
M net retention 3.6. R insolvency coefficient (9.2.1).
M(s) moment generating function R(r) rectangular dJ. (6.8.1).
(1.6.1). R(t) discounted income (6.6.5).
m mean value of one claim = a 1 R 1 , R 2 , R~, R~ control limits
(3.3.2). (7.7.1), (7.7.2).
m.gJ. moment generating function r growth or accumulation factor
1.6. specified by suffices, degree of
m(t) mean value of claim for year t loss (3.6.12).
6A(b). r2 = azlm 2 risk index (3.3.8).
m(Z) measure of tail (3.5.25). r3 = a3/m3 (3.3.S).
N(x) normal dJ., mean 0, st. dev. I rg = I + ig 6.1(c).
(3.9.3). rgp = rg X rp (6.5.7).
N/x) NP dJ. (3.11.10). rgx = rg x rx (6. I.l S).
n expected number of claims 2A(a), r igp = rJr gp (6.5.Sb).
(6.1.2), 6A(b). rp = 1 + ip (6.2.1), (S.3.6).
n(l, t) expected number of claims r(t 1 , t 2 ) (6.6.la).
for period [I, tJ (6.1.24). /'x = 1 + ix (6.1.10).
Ott) run-off error (10.2.7b). r(Z) extinction rate (3.5.26).
P = E(X) risk premium (3.3.7), S sum insured S.2.
(4.l.l). S(Z) dJ. of claim size (3.l.l).
PSL risk premium of stop loss Sk* k:th convolution of S(Z) (3.2.2).
reinsurance (4.8.1). T time period, planning horizon.
P X!L risk premium of excess of loss Tz length of a cycle (6.IA).
reinsurance (4.7.1). t time
P A = (1 + A)P safety loaded premium tp time lag (6.2.2).
(6.2.5). V risk reserve, solvency margin
P, = P)B (6.5.14). (4.1.5), often the initial minimum
P(t 1 , t 2 ) risk premium from period capital V(O) = V 0 which satisfies
tJ
[t1' (6A.7). the solvency conditions.
p(k; t) claim number probability of V, ruin barrier (4.1.5), 6.7(a).
period (0, tJ (2.l.l). u solvency ratio = V/P or V/B
Pk(n) = prob{k = kin} claim number (6.5A).
probability (2.2.1), (2A.2). u, = V/B
Pk(n) claim number probability of V = var(') variance
mixed dJ. (2.7.2). V(t) premium reserve (8.1.2).
pit) termination frequency (8.2.3). v cycle phase (6.1.3); discounting
Q insured sum 3.6A(a); moving factor I/ri 8.2.
expenses 10.2(f). v(y) = v/y) NP transformation
xvi NOMENCLATURE
(3.11.1), (3.I1.17a). 2.9(a).
W technical reserves 6.5(c). rex, h) incomplete gamma function
Wk waiting time of the k :th event, (2.9.1).
exercise 2.6.2. Y skewness 1.5(g).
W(Q) dJ. of sums insured (3.6.14). y skewness of mixed d.f. (2.8.6).
WT(U) conditioned survival function Y2 kurtosis 1.5(g).
(6.7.2). Y2 kurtosis of mixed dJ. (2.8.6).
w = WIB (6.5.4), maturity age Yx skewness of aggregate claims
(8.1.3). (3.3.7).
wit) exogenous impulse (6.1.6). Y2X kurtosis of aggregate claims
X aggregate claims 3.2(a). (3.3.7).
X(t) aggregate claims assigned to Yq skewness of the structure function
year t (6.1.1). (2.8.5).
X(t l , t 2 ) aggregate claims assigned l1f1u 7.9(c).
to years t 1 , tl + 1, ... , t2 e ruin probability.
(6.1.1). e(t) noise (6.1.5).
x = (X - f1x)l(Jx (3.9.1), also (3.3.13). e(X) unit step function (1.5.3).
x(l, t) (6.6.15). '1 auxiliary variable (3.5.18).
x, = confidence coefficient defined A safety loading (4.1.4), (6.5.12).
by the NP function e = Ny! -xJ Ab gross premium B adjusted safety
Y = P A - X = underwriting profit loading (6.2.10).
Yet), YeO, t) profit (8.1.6), (8.1.7a), Ap risk premium adjusted safety
(8.3.1), (lO.2.7a). loading (6.2.9).
Y standardized normally distributed Atot = A = total safety loading (6.5.12)
variable 3.11 (b). f1r mean value of the claim ratio f
yeO, t) = YeO, t)IB(t) (8.1.7b). (6.6.6).
Yo discounted random profit (8.2.1). l1i = ith central moment (1.5.5b).
y, confidence coefficient, e = N( - yJ l1u mean value of the solvency ratio
Z size of one claim (3.1.1), credibility u (6.6.8).
coeff. (4.10.3). f1x = f1x (1, t) mean value of
Z(t) risk sum (8.2.4). aggregate claims X (3.3.7), (6.1.13).
Z cycle variable (6.1.3). Vj =n/n (3.7.3).
zm cycle amplitude (6.1.3). 1!j = mx/mx (3.7.8).
rJ. Pareto parameter (3.5.20) or p risk intensity parameter (2.2.1),
41Y~ (3.12.1). see also (3.3.11), relative mean
rJ. j moment about zero of the claim risk, exercise 5.2.2.
number (2.4.3), (1.5.5a). pit) lapse ratio (8.3.2).
aj moment about zero of the mixed (J standard deviation.
d.f. (2.8.1). (j standard deviation of mixed dJ.
Pj moment about zero of aggregate (2.8.6).
claims (3.3.3). (J q standard deviation of the
r(h) = complete gamma function structure function (2.8.5).
NOMENCLATURE xvii

O"u standard deviation of the [x] = the integer equal or next

solvency ratio u (6.6.10). smaller than x.
O"x standard deviation of aggregate Au B union of sets A and B = a set
claims X (3.3.7). consisting of all the elements of A
O"x (t l , t 2 ) = standard deviation and B
related to X(t l , t 2 ) (6.1.19). An B intersection of sets = a set
qJe (6.9.5). consisting of the joint elements
qJee (6.9.7). of A and B
qJ(u) characteristic function (1.6.8). A :J B B is a subset of A.
qJ(t) ruin state probability (6.6.19). "Ix1 valid for all x.1
'PiU) ruin probability related to 3x i valid at least for one of the Xi'
period (0, T] (6.7.1), (9.2.1). * e.g. F * G(X) convolution (\.5.8).
w cycle frequency = 2n/Tz (6.1.4). ** mark for technically detailed
section or paragraph which can
Some conventions be passed on first reading.
x+ or ( )+ = X or = the expression f - 1 is the inverse of f.
in the parentheses if x or the - I as a subscript relates to the
expression;:" 0, otherwise = 0. preceding year, e.g. u_ 1 (6.5.2).
 CHAPTER 1

Definitions and notation

1.1 The purpose of the theory of risk

(a) Deterministic versus probabilistic approach Conventional actu-

rial 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 their mean values,
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 oversimplification 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, as well as possibly other key quantities, are considered
as random variables. Studies of the different kinds of fluctuation
appearing in an insurance portfolio which start from this point of
view constitute the branch of actuarial mathematics termed the
theory of risk.

(b) Gradual introduction of stochasticity Of course, the financial

structure of an insurance company depends on management costs
and investment of capital in addition to the claim aspects, but
these two factors are not subject to random fluctuation to the same
extent as are claims. Therefore the analysis is first restricted to
 2 DEFINITIONS AND NOTATIONS

claims and to that part of premiums which remains when the loading
for expenses for management has been deducted, i.e. risk premiums
increased by a safety loading. These restrictions are then relaxed,
leading gradually to the construction of a comprehensive model
(see Chapter 10).
In particular, the high rates of inflation prevalent today cannot
be ignored in practical work. To provide a satisfactory basis for
development it is assumed that, when the horizon under considera-
tion is longer than one year, the size of the claim will be corrected
by a factor depending on the assumed value of money.

(c) Claims as stochastic process The claim process can be described

graphically as in Fig. 1.1.1. 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
horizontal axis and the altitude X of the stepped line shows the
total amount of claims during the time interval (0, t]. The process
is, in fact, a compound stochastic process in the sense that the time
of occurrence and the number of occurrences are random phenomena

1/1
E
·Ci
U

Time t
Figure 1.1.1 A sample path of claim process.
1.1 THE PURPOSE OF THE THEORY OF RISK 3

and the size of each claim is also a random variable. Any particular
realization consisting of an observed flow like that in Fig. 1.1.1
is called a sample path or a realization of the process.
If the observation time t is fixed, then the corresponding outcome,
X(t) in our example, is a random variable having a distribution
function (abbreviated dJ.) F(X; t) = prob {X(t) ~ X}. Random vari-
ables will be denoted by bold-face letters (see Section 1.5). If the
stochastic process is well defined, then F is uniquely determined
for every t of the observation range. On the other hand, however,
mere definition of F, even if it were valid at every t, is not sufficient
to determine a stochastic process. In addition, transition rules are
needed to describe how the X(t) values related to different times t
are correlated. Hence some care is necessary when the terms
'stochastic processes' and 'stochastic variables' or their distributions
are used.

(d) Underwriting process If the whole risk business of an insurance

portfolio is considered, this can be illustrated graphically as shown
in Fig. 1.1.2. For the sake of simplicity the yields of interest and of

...... j
u0 + (1+i
I

I
I X (tl
I

u (tl

Figure 1.1.2 Risk process as difference of incoming premiums and outgoing

claims.
4 DEFINITIONS AND NOTATIONS

many other relevant factors are omitted in this connection (they

will be incorporated into the model in later chapters). The risk
premium P together with a safety loading A is continuously flowing
in; this is accumulated in a risk reserve U of an initial amount
U 0' so that the income is represented by a line sloping upwards to
the right. The claims, which can be regarded as negative income,
are paid out from this reserve and are represented by downward
steps. The difference
U(t) - U0 = P(1 + Je)t - X(t),
gives the net gain (positive or negative) arising during time t.

(e) Risk reserve The concept of risk reserve is closely related to

or possibly identical with the concepts solvency margin or net worth,
frequently used in current general practice. All these terms can be
defined as the difference between the values of assets and liabilities,
even if the definitions diverge in detail, for example whether or not
to regard the assets as book values or as current market values.
Also, in the latter case the underestimation of the book values, as
'hidden reserve', is included in the risk reserve. In the following, the
risk reserve is understood to be a 'reservoir' or 'basin', where the
underwriting gain is flowing in if positive or draining out if negative.
When risk-theoretical models are constructed and operated, the
decision as to whether the risk reserve is composed of the whole
solvency margin (understood in the wide sense), or possibly only
some part of it, can be left open and to be determined in each
application.
In some cases it may be advisable to take U to mean only that
part of freely allocable resources which can be used to cover adverse
fluctuations of the underwriting gains or losses without too great
an inconvenience. In some other contexts it may be necessary to
suppose that U consists of the whole actual solvency margin.

(f) Parameter uncertainty There is still an important feature to

be mentioned. For numerical calculations the risk models always
need the assignment of numerical data for a number of model
parameters and for initial values of the variables to be analysed.
The derivation ofthese from statistics and other available experience
is mostly done by well-known estimation procedures from the
mathematical theory of statistics. This topic is outside the scope of
1.2 STOCHASTIC PROCESSES IN GENERAL 5

this book. Instead it is mostly supposed that the initial values are
readily available. The problems caused by parameter uncertainty
are not only a feature of risk-theoretical considerations. In essence
the same problems are always present and are even more critical
in premium rating and in evaluation of technical reserves. In
fact some of the basic data underlie both risk -theoretical and rating
calculations or can be derived from the same basic files. Premium
rates are ultimately based on past experience and more or less
reliable prognoses of future trends, cycles, inflation and other
relevant factors, and they are understandably subject to inaccuracies
and errors. They affect the trading result, which makes it possible to
evaluate their order of magnitude but not until after a time lag,
which in practice may be two or three years and even longer for
some particular classes of insurance. The rates and reserves can
and should be then corrected (within limitations imposed by
competitive market conditions or statutory regulations, if legally
controlled). This control mechanism, which is inherent just from
the uncertainty in the parameters, is one of the important causes of
the underwriting cycles which will be described in Section 2.7 and
incorporated in the model in Chapter 6. In fact the effect of the
parameter uncertainty will be regarded in this indirect way in the
risk-theoretical considerations discussed in Section 6.2.
The effect of parameter inaccuracy can also be investigated
directly if necessary. The technique of sensitivity analysis, which
will be developed in Section 7.6, can be useful for this purpose.
Simply, variations in the initial data can be fed into the risk-theoreti-
cal models and the sensitivity of the outcomes can be used for
evaluation of the effect of inaccuracies arising from the uncertainty
of the initial data.
Even if the estimation inaccuracy is not considered, it must
always be kept in mind as a relevant background factor. For
example, it may be meaningless to apply very laborious techniques
to get very accurate results if the initial data are uncertain. The
selection of approaches, if alternatives are available, should thus be
consistent with the environment under consideration.

1.2 Stochastic 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
6 DEFINITIONS AND NOTATIONS

and now constitutes a large branch of probability theory. Other

examples of such processes are waiting time in a queue, level of
water in a dam, 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. Chung,
1974; Cox and Miller, 1965).

1.3 Positive and negative risk sums

Figure 1.1.2 represents a realization or a sample path of a general
type of insurance risk process in which any occurrence giving rise to
a downward step represents a loss, this being the case in classes such
as fire, marine and life (death risk only). A different situation arises
in the case of immediate annuity business, since the initial fund is
continuously depleted until an annuitant dies, when the reserve

Time t
Figure 1.3.1 Negative risk sums. A portfolio of current annuities. The death
of an annuitant is reflected as a step upward.
1.4 MAIN PROBLEMS 7

released gives rise to an upward step. This latter type is called a

risk process with negative risk sums (see Fig. 1.3.l). In this book,
positive risk sums only are dealt with as their applications are of
greater interest, and because in this way the considerations could be
kept simpler. However, most of the results are valid both for negative
and mixed sums.

1.4 Main problems

(a) Model building Risk theory can be applied to a wide variety

of situations. Before the procedure can be described the problem
has to be well defined in terms of the necessary variables as well as
the rules and distributions which determine their behaviour. In
conventional terms it is stated that a model will be constructed to
describe the insurance business or some particular function thereof
to be analysed. A simple example is given in Fig. 1.1.2, where the
model consists of incoming premiums and outgoing claims, the
difference of which is accumulated into the risk reserve. In more
advanced models the number of variables involved can be quite
great. Some of them are usually target variables, the behaviour of
which is to be examined under given conditions. Generally it is the
business outcomes, such as the values of the risk reserve U in the
above simple example, which are sought when the initial conditions
and policy alternatives (the so-called business strategies) are given.
The analysis is usually of the 'what .. .if...' type, i.e. meaningful
answers are looked for to such questions as what would happen if
the initial conditions and strategies were as given. Before going
on to more sophisticated developments it is advisable to distinguish
the following types, (i) and (ii), and to pose the relevant questions:
(i) What is the result of the business at the end of a certain time
period (0, T] as illustrated in Fig. 1.4.l? This type of problem
will be considered in Chapters 3 and 4, generally taking T to
equal one year.
(ii) What is the result if the observations are extended to a certain
set of times t l ' t 2 , ... , tv of a (prolonged) interval (0, T] ? This
period can be called a planning horizon. This approach will be
taken in Chapters 6 and 7, where the check points t1' t 2 , ... are
usually the end points of the calendar years included in the
planning horizon, i.e. 1, 2, ... , T. Figure 1.4.2 illustrates this case.
 8 DEFINITIONS AND NOTATIONS

;:)
CII
~
CII
I/)

~
x.
I/)

0:
Uo

u (t)

Time t
T.

Figure 1.4.1 A simple risk process. The state of the process can be checked
either at the end of the period (0, TJ or continuously as in item (b).

(J (t)

3 T
Figure 1.4.2 Checking at points t = 1, 2, ... T during the observation period.

In terms of Fig. 1.4.1 the first question is equivalent to finding the

probability for different values that the risk reserve U(t), or any other
target variable, can assume at time t = T. Of particular interest is
the so called ruin probability, i.e. the probability that the risk reserve
U(t) may become negative or, more generally that it will fall below
some given limit, usually called the ruin barrier and denoted by
Ur(t). In the case of the extended problem setting (ii), the ruin prob-
ability is defined as the probability that U is negative at one or more
of the specified time points t1' t 2 , .... This latter concept can be
called 'finite time ruin probability' as distinct from the simple
'ruin state probability'.
1.4 MAIN PROBLEMS 9

(b) Discrete and continuous procedure A modification of the second

problem arises when observations are taken at each point of
time period (0, T]; i.e. what is the probability that ruin will occur
during this period, while checking is performed continuously at
every time point of the period? The former approach is called
discrete, the latter continuous. Both ofthem are applied in risk theory,
often depending on which of them may be more convenient for the
special technique under consideration. In practice the incoming
and outgoing money flows are followed continuously, but the
valuation of assets and liabilities may generally be carried out
completely only at the end of each accounting year. Consequently,
the financial status is usually checked only at discrete time points,
which suggests the discrete analysis approach. There is, of course,
some difference in the outcomes resulting from the selection of
this problem setting. The discrete approach ignores the possibility
that the status, i.e. the risk reserve U, may be negative between the
checking times t i , t i + 1 but then recovers to a positive level. Con-
sequently it gives a somewhat lower ruin probability. Note that the
continuous testing would consider the example plotted in Fig.
1.4.2 to be a ruin, but discrete testing would not.
In the following, the discrete problem setting is mostly assumed
to take the checking interval equal to one year. However, the same
technique can be used without need of modification for any other
interval length, e.g. if the checking of status is made monthly or at
each quarter of the year.

(c) Finite or infinite time-span Another modification, very central

in the earlier development of collective risk theory, is to ask what
is the result if the time T tends to irifinity. Chapter 9 will be devoted
to this case.

(d) Claim number and compound processes The analysis of risk pro-
cesses begins in Chapter 2 by considering the number of claims
and the process related to it, which is called the claim number process
or often the counting process, the well-known Poisson and negative
binomial processes being treated as special cases.
The general case where the individual amount of a claim, claim
size, may vary forms the subject matter of Chapter 3 and later parts
of the book. This process, where both the claim number and claim
size are random variables, is called the compound process.
10 DEFINITIONS AND NOTATIONS

(e) Break-up and going-concern basis One way to define the solvent
state of an insurer is to require that at the end of each fiscal year
the assets should be at least equal to the total amount of liabilities
(possibly increased by some legally prescribed margin). This
situation may be tested by assuming that the activity of the insurer
would be broken at the test time point and the liabilities, such as
those due to outstanding claims, would be cleared up during a
liquidation process. Then assets should be available in step with
the time of claim and other payments. The risk factors that are
involved and which are to be evaluated are the uncertainties of
the magnitude of the claims, including those claims which have
already occurred before the test time point but which may be
notified later. Furthermore, realization of the assets is affected by
changes in market value and the whole process is subject to inflation.
In other words, this 'break-up basis' is involved with uncertainties
arising when both the liabilities and assets go into hypothetical
liquidation. The problem is to evaluate these inaccuracies and to
find a minimum solvency margin in a way which still gives an
adequate guarantee for the fulfilment of the commitments of the
insurer.
Another possibility is to assume that the business of the insurer
will go on. Then, in addition to the errors and inaccuracies concern-
ing gradual liquidation of the assets and the outstanding claims
and other liabilities that have arisen in the past fiscal years, i.e. the
risks involved just with the break-up situation described above, the
continual flow of new claim-causing events and other business
transactions gives rise to further fluctuations. Because this 'going-
concern' basis, by definition, includes the 'break-up' risks as a
partial element, it generates a larger range of fluctuations and leads
to demand for a greater solvency margin and safety loading than
does the break-up basis alone. This assumes, of course, that
consistent principles are followed in the two bases. It may be re-
marked that if, for example, the break-up basis is defined con-
ventionally, such as by statute, there may be incompatibility with
the practical reality of the going-concern basis.
The going-concern basis was tacitly assumed in the previous items
and it will be followed in this book generally. The problems involved
with the outstanding claims, which constitute the most important
break-up risks, will be discussed in item 3.1 (c). Asset risks will be
considered in Section 6.3.
1.5 ON THE NOTATION 11

(f) Life and general insurance An insurance contract normally

covers an insurer's liability to indemnify losses caused by specified
events like fire, accident, death etc. An insurance contract can, in
addition, include more than these pure risk elements; in particular,
life and pension insurance schemes may also provide savings for
future years. Risk theory is most appropriate to deal with the risk
elements and it will be considered in detail in Chapters 1-7. The
presentation is formulated, as far as possible, to cover generally
all kinds of risk businesses, including the risk elements of life and
pension insurance. Special features oflife insurance will be discussed
in Chapter 8.

1.5 On the notation

The following conventions will be ,applied:

(a) Stochastic variables and processes are denoted by bold-face

letters, e.g. X. If X is a stochastic process, then the state of X at time
t is denoted by X(t) or Xt'
Some ofthe variables, such as the safety loading A., may be stochas-
tic in some applications but deterministic in others. Then ordinary
typeface is generally used, without indicating that stochasticity
may sometimes occur; i.e., A. is used for A.

(b) Monetary variables and rates Variables directly representing

monetary qualities like claims, X, premiums, P, or risk reserves
U are denoted by capital letters, they are of dimension one (or more)
in terms of the monetary unit, e.g. £, $, etc.
If a dimensional monetary variable is transformed into a non-
dimensional 'relative' variable, then the corresponding small letter
is used, e.g. u = U/P or x = (X - E(X))/O"x'

(c) The (cumulative) distribution functions (= d.f.) are denoted by

capital letters and densities by small letters. For example,

F(X) = prob {X ~ X} and f(X) = F'(X),

define the distribution function F and the density f (if it exists) of a

random variable X.
12 DEFINITIONS AND NOTATIONS

(d) Stieltjes integrals A number of different distribution (or other)

functions F will be employed. Some ofthem are assumed continuous,
having a continuous derivative f = F' (possibly with a finite number
of exception points). Some others are discrete, i.e. the variable
X under consideration may take only values Xl' X 2' ... by given
probabilities PI' Pz ' .... Also functions of mixed type are needed,
being continuous except at a set of discrete points Xl' X 2' ... (Fig.
1.5.1). These steps may for example be caused by reinsurance limits.
In the subsequent development, integrals of the type

f g(X)dF(X), (1.5.1)

are often needed, where g is some auxiliary function. For the types
of distribution functions mentioned, this becomes

f g(X)f(X) dX (continuous case)

Lg(Xi)P i (discrete case) (1.5.2)

f g(X)f(X) dX + ~g(X)Pi' (mixed case)

where the sums are to be taken for a ~ Xi ~ b. It is convenient

to make a convention that (1.5.1) will represent all the types (l.5.2).
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 (1.5.1) as a Stieltjes integral. Readers

FIX)
1 - - ______________________________________ .=-_~_~ __

x
Figure 1.5.1 A mixed type df
I.5 ON THE NOTATION 13

who are not familiar with this topic can regard it as an abbreviation
andreplace any such integral by the last form in (1.5.2), since for
our purposes the mixed case, as defined above, is general enough.
This extended integral has the same general features as the conven-
tional integral.

(e) Unit step function It is often convenient to use the unit step
function E(X) defined as follows
E(X )= {o forx<O
(1.5.3)
1 for x ~ o·
It can be interpreted as the d.f. of a 'degenerated' variable, where the
whole probability mass is concentrated at the origin.
An example of the application of the step function in integration

f::
is as follows (see the previous item)

F(x) dE(x - xo) = F(x o)' (1.5.4)

i.e. the step function picks up a selected value of the integrand.

(f) Rates and factors The time-dependent change of numerous

quantities is required in a number of circumstances, for example
reserves and assets are increased by interest, volume variables by
inflation and real growth of the portfolio and risk exposure. The
(mostly annual) growth rates will be denoted by i and distinguished
by a subscript; for example, ii' ig, ix' and ip are the rates of interest,
real growth of the volume variables, claims inflations (x for X,
the notation of claim amount), and premiums, respectively. The
corresponding (accumulation) factors will be denoted by r with the
same subscripts as for the rates, e.g. rx = I + ix • If the change is
caused by two or more factors, then the subscripts are composed
of the corresponding symbols, e.g. rgp is the growth factor represent-
ing the joint effect of real growth and premium inflation (this and
other component combinations will be defined in the appropriate
parts of the text).
To avoid inconvenient notation, the subscripts of the factors
and rates are always printed by lower-case letters and without
indicating whether the quantity concerned is stochastic or not, e.g.
ix for claims ( = X) inflation.
The factors and the rates may be time dependent.
14 DEFINITIONS AND NOTATIONS

(g) Characteristics of distributions The jth moment about zero of

a random variable X, defined as the expected value of the power of
X, will be denoted by rxj or alternatively by aj

rx.J = rx.(X)
J
= E(Xj) = f +OO

-00
xj dF(X). (1.5.5a)

Most random variables concerned in the sequel are non-negative,

in which case the integration in (1.5.5a) can be taken from 0 to

f::
+ 00. The central moments J1. j of X
J1. j = J1.iX ) = E{(X - E(X))j} = (X - E(X))-i dF(X), (1.5.5b)

are obtained for j = 2, 3 and 4 from the moments about zero as

follows
_ _ 2
J1. 2 - rx 2 rx l
J1. 3 = rx3 - 3rx l rx 2 + 2rxi (1.5.6)
J1. 4 = rx 4 - 4rxl rx3 + 6rx;rx2 - 3rx1·
Recall that, for instance, in the case of discrete X, (1.5.5a) becomes
simply (see item 1.5(d))
where Pk = prob{X = XJ. (1.5.7)

The most important characteristics, the mean E(X), variance a 2 =

O"i = J1.2, skewness Y = Yx = J1.3/a3 , and kurtosis Y2 = Y2(X) = J1.4/a4 - 3,
can be expressed using the above moments. It should be noted
that all the characteristics of X depend only on the dJ. F of X. If
necessary X will be given in connection with the characteristic
either as a 'normal' argument, for example as J1. 2(X) or ax, as a
subscript. The latter practice (even though inconsistent with the
former) is used for common symbols like a and Y which have no
subscript related to the concept itself. Furthermore the skewness,
which is one of the parameters frequently used in the following,
is denoted by Y instead of the conventional notation Y1 .

(h) The convolution Let F and G be distribution functions. Their

convolution F * G,

F*G(X)= f +OO
-00 F(X - Y)dG(Y), (1.5.8)
1.6 THE MOMENT GENERATING FUNCTION 15

is also a distribution function. It is well known that if X and Yare

independent random variables with distribution functions F and G,
then the d.f. of their sum X + Y is F * G. Hence, F * G = G *F and the
convolution of several distribution functions does not depend on the
order in which the convolutions are taken.
If, for example, G is continuous with g = G', then (1.5.8) can be
written as

F*G(X)= f +OO
-00 F(X - Y)g(Y)dY, (1.5.9)

which is a continuous function of X. Thus the sum of two indepen-

dent random variables has continuous d.f. if at least one of the
variables has continuous dJ.

(i) A list of symbols and notation has been given after the Preface
for the convenience of readers.

1.6 The moment generating function, the characteristic function,

and the Laplace transform

(a) Three auxiliary functions In this section three well-known

operational functions of probability calculus are introduced. They
transform a given dJ. into a form from which it is, for example, easy
to see the characteristics of the distribution. These transforms are
of particular interest because they facilitate the handling of convolu-
tions and some other operations.
The moment generating function, introduced in item (b), is some-
what simpler than the characteristic function (item (c)) but has the
disadvantage of being not always defined, whereas the characteristic
function, which is defined as a complex integral, always exists. The
characteristic function also has an important continuity property.
In order to avoid the use of complex integrals, only the moment
generating functions will be used in this book, except in Appendix B
where the characteristic function is needed.

(b) The moment generating function (m.g.f.) of a random variable

X (or of its dJ. F) is

M(s) = E(e SX ) = f +OO

_ eSx dF(X).
00 0.6.1 )
16 DEFINITIONS AND NOTATIONS

In the case of continuous dJ., (1.6.1) becomes

M(s) = f+OO
_ 00 eSx f(X) dX, (1.6.2)

and for a discrete dJ.

(1.6.3)
k

where s is an auxiliary real variable; for notation see item 1.5(d).

It will be assumed that for the distributions dealt with in this
section M(s) exists at least in some neighbourhood of the origin,
i.e. for Is I < So for some positive so. This condition must be verified
for each actual application, of course.
The moment generating function has the following important
properties which are derived in standard textbooks and are given
here without proof.

(i) The m.gJ. can be expressed by the moment expansion

00

M(s) = L rt.hSh/h!, (1.6.4)

h=O

where (Xh are the moments (1.5.5a) about zero. If M(s) is known,
they can be obtained in terms of derivatives
(1.6.5)
(ii) The dJ. is uniquely determined by its m.gJ. (if the m.gJ. exists),
i.e. if two distributions have the same m.gJ. they are identical.
(iii) The linear transformation y = ax + b transforms the m.gJ. into
the form
(1.6.6)
(iv) If Xl and x 2 are two independent stochastic variables, then the
m.gJ. of sum Xl + x 2 is obtained by multiplication
M(s) = M 1(s)M 2(S). (1.6.7)
In terms of distribution functions this means that the m.gJ.
transforms convolutions of distributions into products ofm.gJ.s.
1.6 THE MOMENT GENERATING FUNCTION 17

**(c) The characteristic function of X is obtained by replacing the

variable s by an imaginary variable iu in the m.gJ., where i = -1J
q>(u) = E(e iuX ) = f
+oo
_ eiuX dF(X).
00 (1.6.8)

The characteristic function q> has properties similar to (i)-(iv) of the

m.gJ. Owing to the fact that IeiuX I = 1, the convergence in (1.6.8)
does not cause problems. In addition, it has a special benefit in that
the inverse transformation q> --+ F can be expressed explicitly in form
of a Fourier integral, which is sometimes convenient for the calcula-
tion of F or for derivation of some of its properties.
One more important property of the characteristic function is the
continuity property: if a sequence of dJ.s converges to a dJ., then the
corresponding characteristic functions converge to the characteristic
function of the limit function. Conversely, if a sequence of character-
istic functions converges to a limit function which is continuous at
the origin, then the limit function is the characteristic function of the
limit dJ. of the dJ.s corresponding to the sequence of characteristic
functions.

**(d) The Laplace transform belongs to the same family of auxili-

ary functions. The Laplace transform of a non-negative random
variable X is

L(s) = too e- sx dF(X), (1.6.9)

where s now assumes complex values. The Laplace transform also

has basic properties similar to properties (i)-(iv) of the m.gJ.
Furthermore, its inverse can be given in the form of a complex
integral. Extensive tables are available in a number of standard
mathematical texts.
The integral (1.6.9) converges when the real part of s exceeds
some number called the abscissa of convergence. If the real part of
s is zero, the characteristic function (1.6.8) is obtained as a special
case of the Laplace transform of non-negative random variables.
 CHAPTER 2

Claim number process

2.1 Introduction

(a) Definition of the problem As already mentioned, the simplest

problem is considered first. This is to find the probability function of
the number of claims arising in a risk collective, i.e. a function p(k; t)
which gives the probability that the number of claims k t in time t
is equal to k. In terms of conventional symbols

p(k; t) = prob {kt = k}. (2.1.1 )

In the following analysis the collective concerned can be the whole

portfolio of an insurer or some special part thereof.

(b) The individual and collective approaches 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 insurance it is assumed that the probability that a life
aged x dies within a year is qJ Then the total number of claims is
the sum of the contributions from the individual policies and the
probabilities (2.1.1) can be derived by means of the addition theorem
of probability calculus from the primary probabilities. Basically the
probabilities are 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 times and the number of events (claims) are recorded and
2.2 THE POISSON PROCESS 19

in which no attention is paid to the particular policies 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
process takes the well-known Poisson form.
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.
More general number processes are considered in Sections 2.7-2.10.

2.2 The Poisson process

It is now assumed that the claim number process satisfies the follow-
ing three conditions

(i) Events occurring in two disjointed time intervals are indepen-

dent (independence o/increments).
(ii) The number of events in a time interval (t l ' t 2 ) is dependent only
on the length of the interval t = t2 - tl and not on the initial
value tl (stationariness o/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 0/
multiple events).

In Appendix A it is shown that with these conditions the probability

function p(k; t) is represented by the well-known Poisson law

(2.2.1)

°
for every t > 0, where p ~ is a parameter indicating, as will be seen
later, the average number of claims in a time unit. The process k
is called a Poisson process. The occurrence of an event of claim de-
pends on both the number of cases (risk units) exposed to risk and
also the risk intensity, i.e. the chance that a particular case gives
rise to a claim. The Poisson process arises as a product of these
components.
20 CLAIM NUMBER PROCESS

2.3 Discussion of conditions

(a) Independence of increments 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 contradiction to this condition.
Condition (i) can, however, often be met by defining a risk unit, as
is customary in reinsurance practice, 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 insurance. The Poisson
function is not then applicable, at least not without suitable modifica-
tions. These kinds of cases will be treated in Section 2.7 where
condition (i) is replaced by a more general one.

(b) Stationariness of increments This condition means that the

collective flow of the events is stationary, i.e. neither steadily
increasing or decreasing nor osciIIating more than would be caused
by normal random fluctuation. In other words, the intensity of the
claims is constant. This is the usual case in insurance, particularly
during short periods, when the numbers of policies or other circum-
stances are not subject to marked changes. This condition 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.
Quite often, however, there are situations where the stationarity
does not strictly apply; for example, there may be seasonal variations
in claim intensities. Then the time interval concerned can be divided
into subintervals in such a way that the corresponding sub-processes
have (at least approximately) constant intensities and are therefore
Poisson processes. It wiII be shown in Section 2.6 that the sum of
independent Poisson variables is again a Poisson variable. Hence,
the total number of claims during the whole interval is Poisson
distributed. So, if only this total number and its behaviour are of
interest, as is most often the case, then the seasonal variations can
2.3 DISCUSSION OF CONDITIONS 21

be disregarded, providing that the changes in the claim intensities

are deterministic, i.e. they recur in such a way that their prediction
is possible through experience.
The above is also applicable to cases where the risk intensities
are changing, according to some trend, in a way which is deter-
ministically predictable. The total number of claims still remains
Poisson distributed but the parameter pt, which will be later denoted
by n, is to be calculated as a sum of the parameters related to the
subintervals as will be shown in Section 2.6. Another method often
used in risk theory literature to derive the same result is to introduce
the concept of 'operational time'. This approach is presented in
Appendix A.
However, there are often other circumstances arising in practice
where conditions (i) and (ii) cannot be met. For example, fire insu-
rance 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 cause an
enormous increase in claims. It is also well known that economic
conditions have considerable influence in many classes of non-life
insurance. Times of economic booms or recessions give rise to a
considerable increase or decrease in the number of traffic or work
accidents, as well as influencing credit insurance business. Such
circumstances are so general that the application of the elementary
Poisson function is greatly limited, and so there is a need for a
development of the theory omitting the conditions concerning
independence and stationarity. This will be done in Section 2.7.
In spite of these limitations, the Poisson function often gives at
least a good first approximation, particularly for short time intervals.
It is also the basis of more generalized distributions. Furthermore,
the risk of changes and of variations disturbing stationarity can often
be dealt with by simply adding a precautionary amount to the
parameter p.

(c) Exclusion of multiple events At a first glance it would appear

that condition (iii) 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 insurance 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
 CLAIM NUMBER PROCESS
22

statistics of the distribution ofthe 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 Some basic formulae

(a) Distribution function Since the state k t at time t of a Poisson

process obeys the Poisson distribution law with the parameter
pt = n, it is necessary to be familiar with a number of basic character-
istics of the Poisson distribution function.
Let us suppose that k is a Poisson-distributed random variable
with Poisson parameter n. Then the dJ. F of k is
[k)
F(k) = F(k; n) = prob {k ~ k} = L Pi(n), (2.4.1)
i=O
where
(2.4.2)
To facilitate the technical handling of the formulae, non-integer
values for k will also be allowed. By convention, [k] means the
rounded-down integer value of k defined by [k] ~ k < [k] + 1.
In the following, Pk is often used for Pk(n) and F(k) for F(k; n).

(b) Characteristics Making use of the standard formulae given in

item 1.5(g), the moments of the Poisson distribution can be obtained
either by direct summation or by means of the m.gJ. (see exercises
2.4.1 and 2.6.1). The lowest ones are
!Xl n
= (Jl l = 0)
!X 2-- n + n2 Jl 2 = n
(2.4.3)
!X3 = n+ 3n 2 + n3 Jl 3 =n
!X 4 = n + 7n + 6n 3 + n4
2 Jl4 = n + 3n 2
from which the basic characteristics of k are immediately derived
(see item 1.5(g))
mean E(k) = !Xl = n
variance u2 = Jl 2 = n (2.4.4)
skewness 'Y = Jl 3 1u 3 = I/Jn
kurtosis 'Y 2 = Jl4 /u 4 - 3 = lin.
2.5 NUMERICAL VALUES OF POISSON PROBABILITIES 23

(C) Inter-occurrence time For some applications it is useful to note

that the interval between consecutive events (claims) of the Poisson
process, the so-called inter-occurrence time, is exponentially
distributed, i.e. the length tk of the time period between (k - l)th
and kth events satisfies
(2.4.5)
where p is the expected number of claims in the time unit applied in
the relevant application. (For proof see Biihlmann, 1970, for
example.)
A rather more general basis for the development of risk theory
is to assume some dJ. for the inter-occurrence time instead of using
Pk as the basic concept. This is essentially the basis of the so-called
renewal processes in the theory of stochastic processes. The applica-
tion to risk theory was first suggested by Sparre-Andersen (1957)
and the resulting processes when applied in risk theory are common-
ly called Sparre-Andersen processes (see also Thorin, 1971).

Exercise 2.4.1 Derive the expressions for !Xl and !X2 by direct
summation from (2.4.1).

2.5 Numerical values of Poisson probabilities

(a) Exact values If n is not large the value of the probabilities Pk

and F can be calculated directly from (2.4.1). There are also fairly
extensive tables of numerical values, e.g. General Electric (1962), and
values can also be derived from tables of the chi-square distribution.
Programmable calculators can be used to determine Pk and its
cumulative values F, using
Po = e-", (2.5.1 )
as a starting value and applying the recurrence formula
Pk+ 1(n) = Pk(n)n/(k + 1). (2.5.2)
Owing to the wide range of values of n some special technique may
be necessary to avoid problems from overflow (e.g. e- n escapes from
the range of numbers acceptable to the computer). It may help
either to use logarithms or to start with a smaller auxiliary n', putting
Po = e- n' and afterwards correcting the Pk and F(k) by multiplying
them by en'-n. Another computational difficulty may arise from the
24 CLAIM NUMBER PROCESS

accumulation of rounding-off errors when the probability 1 - F(k)

that k is greater than k is calculated for large k. Then a sum of
terms of considerable magnitude is needed and the rounding-off
error of the largest terms is too close to the order of magnitude ofthe
target value 1 - F(k). This can easily be controlled by continuing
the calculation up to the k-values where Pk vanishes from the range
of the desired accuracy. If the calculation is exact, F(k) should equal 1
(because F( CfJ) = 1). The possible deviation of the computed value
from 1 gives the rounding-off correlation for the upper tail of F(k).

(b) The shape of the Poisson distribution for three small values of
n is depicted in Fig. 2.5.1.

(c) Normal approximation If n is large, use may be made of the

central limit theorem of probability theory, according to which F
tends asymptotically to the normal distribution function when
n --- CfJ, I.e.
F(k) ~ N((k - n)/Jn), (2.5.3)
where N denotes a normal distribution function with zero mean and
unit standard deviation. The central limit theorem is applicable
because of the additivity of Poisson variables, i.e. the Poisson

,,,_,,, n= 5

.,
I ,
: !-
,,j- ,
I

I
,
I
15
,,
I

.
I L,
0.1 , ,,
I

,,
~ ..
,,i
,, n
10 20 30 40 50
Figure 2.5.1 Poisson probabilities Pk' (To help visual shaping of the distribu-
tions the discrete probability values are linked as step curves.)
 2.5 NUMERICAL VALUES OF POISSON PROBABILITIES 25

F l-F

.II "1\.
JIf 1!..
"'I
gI ~

r ~
] ;;::J

\n
\
I[ 100 ~~
_1001--
25 ~25-
L..... 10 Normal ' 10
10-3

I
If , II
III \~ ll-,
-4 -3 -2 -1 o 2 3 4 5
z
Figure 2.5.2 F or 1 - F for the normalized Poisson distribution (step lines)
of some values ofn andfor the normal approximation (2.5.3) (logarithmic scale).

variable with parameter n can be expressed, for example, as a sum of

n independent Poisson-distributed variables with the same para-
meter value one (see Section 2.6) (providing that n is an integer).
Figure 2.5.2 demonstrates the accuracy of this approximation for
some values of n. It can be seen that the Poisson dJ. tends closer
and closer towards the normal d.f. as n increases.

(d) Wilson-Hilfertyapproximation A much closer approximation

is provided by the following formula
(2.5.4)
where
K = 3J(l + k) and C = (9n)1/3,
(see Johnson and Kotz, 1969; Section 4.7).
This is again an asymptotic approximation but will be found to
give satisfactory values for n as small as 10, as is shown in Table 2.5.1.
26 CLAIM NUMBER PROCESS

Table 2.5.1 Comparison of the Wilson-Hilferty approximated values with

exact Poisson values. For k ~ n values of F are given and for k > n. values of
1- F.

n= 10 25 50

k Exact Approx. k Exact Approx. k Exact Approx.

1 0.00050 0.00058 10 0.00059 0.00061 29 0.00092 0.00093

4 0.02925 0.02920 15 0.02229 0.02232 36 0.02376 0.02377
7 0.22022 0.21974 20 0.18549 0.18529 43 0.17980 0.17969
10 0.58304 0.58341 25 0.55292 0.55303 50 0.53752 0.53756

13 0.13554 0.13527 30 0.13669 0.13656 57 0.14486 0.14479

16 0.02704 0.02715 35 0.02246 0.02250 64 0.02360 0.02363
19 0.00345 0.00354 40 0.00204 0.00206 71 0.00201 0.00203

(e) Gamma formula The dJ. of the Poisson variable can also be ex-
pressed in terms of the incomplete gamma function as will be
shown in Exercise 2.9.7.

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

death a fixed sum S = £1000 is paid. The mean value of the rate of
mortality is 0.01, the premium p;. = (l + A)S E(k), where A = 0.1 is a
safety loading and k is the number of deaths. The actuarial status of
the society is examined every year. How large a security reserve
U0 should the society have to be sure, at a 99% probability level,
that the balance does not show any deficit?
Make use first of (a) the exact Poisson formulae, then (b) the
normal, and (c) the Wilson-Hilferty approximations. Hint: Poisson
values can be computed starting from values given in Table 2.5.1.

Exercise 2.5.2 How many members should the society of exercise

2.5.1 have for no security reserve to be necessary under the conditions
mentioned? Use the normal approximation.

Exercise 2.5.3 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
2.6 THE ADDITIVITY OF POISSON VARIABLES 27

reinsurer pays the third and subsequent benefits. What is the risk
premium ( = expected amount of claims) for the reinsurance?

2.6 The additivity of Poisson variables

(a) The m.g.f. of the Poisson d.f. can be calculated by substituting

the Poisson probabilities Pk(n) in (1.6.3)
ao
M(s)= L eShe-nnh/h!
h=O
ao
= e- n L (neS)h/h! (2.6.1)
h=O

(b) Additivity of Poisson variables It follows from (2.6.1) that the

m.gJ. for the sum of two independent Poisson variables having
parameters n1 and n2 is
(2.6.2)
This is again of the form of (2.6.1); hence according to property (ii)
of item 1.6(b) the sum variable is also Poisson distributed with the
parameter
(2.6.3)
i.e. the Poisson distribution is additive: the sum of mutually
independent Poisson variables is again a Poisson variable having the
parameter n as the sum of the original parameters.
The additivity is a very important feature. It makes it possible to
divide the risk portfolio into sections, indexedj = 1,2, ... ,according
to the classes and sub classes ofthe insurance concerned, for example.
It is often advisable to evaluate the Poisson parameters nj , the
expected number of claims, separately for each section and then to
determine the parameter for the whole collective by summation

(2.6.4)

It is also possible to divide the time-span into adjacent intervals,

e.g. in months, evaluate nt for each interval, and then to sum the n.
28 CLAIM NUMBER PROCESS

This makes it possible to apply the Poisson law to cases where the
risk exposure, measured by nt' may vary in some defined way, such
as by following a cycle or a trend. This aspect has already been
discussed in item 2.3(b).

(c) As an example consider a life insurance portfolio. The insured

persons have the probability of death during one year qj
(j = 1,2, ... ,J). Then the expected number of deaths in the whole
collective, the Poisson parameter, is
(2.6.5)
This equation is, in fact, a link between the individual risk theory,
which primarily focused attention on the individual risk units, and
the collective risk theory, which takes the collective itself as the
primary issue. A more detailed discussion of the special features of
life insurance will be pursued in Chapter 8.

**(d) The additivity of the Poisson processes The additivity con-

sidered in item (b) concerned the Poisson variables, i.e. in the
present context the number of claims during a fixed period (0, tJ. If
the time t is allowed to vary and the growth of the claim number
k t is followed, a realization of the claim number process is obtained.
The distinction between these concepts was emphasized in item
1.1(c).1t can be proved on the basis of conditions (i)-(iii) of Section 2.2
that additivity also holds for the sums of independent Poisson
processes related to the same time interval (Exercise 2.6.3).

Exercise 2.6.1 Calculate the moments rt.j (j = 1,2,3,4) of the Poisson

dJ. by means of the m.gJ.
** Exercise 2.6.2 Let Wk denote the waiting time of the kth event
of a Poisson process k, i.e. the time of kth event. Then Wk > t implies
that less than k events occurred during the time interval (0, t J. Prove
that

(t> 0).

** Exercise 2.6.3 Prove that the sum of two independent Poisson

processes related to the same time interval is again a Poisson process
2.7 TIME-DEPENDENT VARIATION OF RISK EXPOSURE 29

i.e. also that the sum process satisfies the conditions (i)-(iii) of
Section 2.2.

2.7 Time-dependent variation of risk exposure

(a) Experience of the applicability of the Possion law As mentioned

in Section 2.3, the simple Poisson law frequently fails to provide a
satisfactory representation of the actual claim number distribution.
This feature is demonstrated by Figs 2.7.1 and 2.7.2, which are
typical curves taken from a portfolio of motor cycle insurances and
illustrate the various types offluctuation.
The observed data of Fig. 2.7.1 are the monthly claim frequencies
of motor cyclists over the years 1960 to 1962, during which period
the exposures increased from about 19000 to about 27000 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
3s. 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
declining tendency over the three years.

90

40

1960 1961 1962

Figure 2.7.1 Motor cycles. Four-weekly claim frequency 1960-62.
30 CLAIM NUMBER PROCESS

900

o
8 800
~
...
8.
ij 700
c:
~
~
lJ.. 600

500+-~--~~--~~~~~~-r--~-r--~-r-----
1951 1953 1955 1957 1959 1961 1963

Figure 2.7.2 Motor cycles. Annual claim frequency 1951-63.

0.3 (n-n)/n
30 n/10000
0.2
25
0.1

---
20
0.0
15
-0.1
10
5 -0.2

-0.3
1960 1965 1970 1975
Time

Figure 2.7.3 Number of accidents in workers' compensation insurance in the

period 1958-79 Uoint data of Finnish insurance companies). The right-hand
graph exhibits the relative deviations from the trend-adjusted midline n of the
left-hand figure.

Figure 2.7.2 gives the annual claim frequency for the period 1951
to 1963, the exposure increasing from about 7000 to 27 000 over the
period, and shows a long-term periodic effect, with a probable trend
upwards over the period. This longer period shows that the declining
tendency in Fig. 2.7.1 was a downward phase of one of the longer-
term variations and shows the need to look at fairly long series of
values.
The workers' compensation time series depicted in Fig. 2.7.3 shows
a similar behaviour. Further analysis shows that the cycles are
2.7 TIME-DEPENDENT VARIATION OF RISK EXPOSURE 31

strongly correlated with general economic booms and recessions.

During a boom, industry is at full capacity and overtime working is
frequent. As would be expected, the number of accidents increases
accordingly. On the other hand, during a recession working hours
are reduced, which is immediately reflected in the claim frequencies.
Results like these show that fluctuations observed in the actual
flow of claims number processes may be much greater than would
be expected if the data conformed strictly to the Poisson law. This
observation, which is of the utmost importance for the applicability
of risk theory, has been confirmed by research in various countries,
e.g. McGuinness (1970), Helten (1977), Becker (1981), Bohman
(1979), and James (1981). This phenomenon was widely considered in
the Finnish solvency report (Pentikiiinen, 1982).

(b) Four categories of fluctuation To elaborate the theoretical mod-

el it is necessary to analyse the different types of fluctuation in
the number of claims.
(i) Trends These emerge as a slow moving change of the claim
probabilities which must be properly defined as the ratio of the
number 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. Of course, besides changes in risk intensities, the
growth of the number of risks contained in the portfolio is
another reason for an increase (or decrease) of the expected
number of claims parameter n. This arises from the normal
changes in the number of policies as well as from the internal
growth of risk units under old policies, e.g. in the number of
employees, vehicles, plants and facilities.
(ii) Long-period cycles An example of this type of variation is
provided by the association of accidents with general economic
conditions. The results of consecutive years are not mutually
independent and a cycle period can thus have a length of several
years.
32 CLAIM NUMBER PROCESS

(iii) Short-period oscillations These can be caused for example by

meteorological changes or by 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, may also be of
significance in life insurance, giving rise to fluctuation in the
mortality experience.
(iv) Pure random fluctuations like those considered in previous
sections of this chapter are, of course, always present.
In addition to the four main types mentioned there may be seasonal
fluctuations of which the variation in claims frequencies 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 calender years, as mentioned in
Section 2.3.

(c) Introduction of a structure variable In this chapter, the model

is extended to incorporate short-period oscillations. These are sup-
posed to be so short that oscillations of consecutive years can be
regarded as independent.
A natural way of developing such a model is to assume that n,
the expected number of claims, is itself a random number, n = nq.
Here q (~ 0) is an auxiliary random variable which indicates the
relative deviation ofn from its average value n whereby q is normed
to have mean value 1. Hence this variable represents disturbances
which cause the short-term oscillations beyond the range that can
be explained by means of the Poisson dJ. In the following chapters,
the longer term fluctuations or dependences are incorporated in
other variables; this is unlike the treatment often encountered in the
published literature. These variations are often assumed to be
deterministic; the increments of the risk process during consecutive
years will then be mutually independent. It can be interpreted that
q has a constant value during each year but varies randomly from
year to year, the values related to consecutive years being mutually
independent and equally distributed. Often the behaviour of the
process within the year needs no attention, and thus q can be con-
sidered to be a stochastic process whose state remains unchanged
within each year, as illustrated in Fig. 2.7.4. Equally, one can consider
that each year t has its own variable qt' To keep the notation simple
2.7 TIME-DEPENDENT VARIATION OF RISK EXPOSURE 33

2 3 Time
Figure 2.7.4 A process where the risk parameter q related to consecutive time
periods varies at random.

the subscript t is mostly omitted. Since the states of q related to

different years are equally distributed this does not usually lead to
any misunderstandings.
Now let
H(q) = prob{q:::;; q}, (2.7.1)
be the dJ. of q. Then the conditional probability of k claims in a year
by condition q = q is the Poisson probability Pk(nq). The un-
conditional probability Pk(n) = prob {k = k} is obtained by means of
the addition and multiplication rules of probabilities as 'a weighted
average' (or more exactly as the expected value of Pk(nq)) as follows
oo foo (n )k
Pk(n)= f 0 Pk(nq) dH(q) = 0 e- nq kq! dH(q). (2.7.2)

For discrete H (e.g. when it is approximated in a tabular form)

this reduces using the elementary formula of conditional
probabilities
prob {k = k} = LPk(nqJ prob {q = qJ. (2.7.3)

Now the dJ. ofk can be derived by the summation of (2.7.2)

[k] [k] foo
. F(k) = i~O Pi(n) = i~O 0 Pi (nq) dH(q)

= f oo [ .I pJnq)
[k] ]
dH(q) =
fOO F(k; nq) dH(q), (2.7.4)
o I~O 0
34 CLAIM NUMBER PROCESS

where F(k; nq) denotes the Poisson dJ. with mean value nq of the
number of claims. Note that the formula could also have been
obtained directly using the same reasoning as applied for (2.7.2),
i.e. by considering that P is, so to speak, the weighted mean of all
possible Poisson functions, the weights for each value of nq being
the probability of occurrences of this expected number of claims.
The claim number process k just introduced is called a weighted
or mixed Poisson process as distinguished from the 'simple' Poisson
process considered in previous sections. Naturally, types of claim
number processes other than the Poisson one can be similarly
weighted. The variable q will be designated the structure variable
and its dJ. H the structure distribution.
In the following the notation P is often replaced by F if it is
clear in the context that the weighted Poisson dJ. is under considera-
tion.

(d) Illustration The mixed 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 a lot. The lottery is
arranged so that the probability of getting a value q < q ~ q + dq
is dH(q). Then q is fixed to be equal to q for the whole year. For the
next year a new value is drawn for q, and so on.

(e) Discussion on environment 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 allow for the
sections j of the portfolio to have their own variations and dis-
tributions H/q). Then the total fluctuation is to be calculated by
convolution, as will be shown in Section 3.7. How the physical
phenomena behind the probabilities are brought about and what
kinds of phenomena exist are, of course, quite immaterial from the
point of view of risk theory. It is only necessary to assume the
existence of some function H which relates to the number of claims.
In fact the urn model is only a simple way to illustrate how (2.7.1)
can be obtained. There are, however, other ways to introduce the
same formula. An example is the Polya process considered in
2.8 MIXED POISSON DISTRIBUTION 35

Section 2.9, where the process allows for changes of q to be consider-

ed as due to contamination.

2.8 Formulae concerning the mixed Poisson distribution

(a) The basic characteristics It will be assumed in this chapter

that the portfolio is considered as a whole, not partitioned in
sections, and that the time span to which the number of claims is
related is one calendar year. These restrictions will be relaxed in
Chapter 6.
The moments about zero of the mixed Poisson distribution Fare
obtained by a straightforward application of (1.5.7) and (2.7.2)
ao
ti j = L kjpk(n)
f: f: [k~O
k=O

= k~O k j Pk(nq) dH(q) = kipk(nq) ] dH(q)

= f: CI)nq) dH(q). (2.8.1)

Hence these moments are obtained from those of the simple Poisson
dJ. by a similar weighting as for Pk and F(k) in the previous section.
Substituting the expressions (2.4.3) where n is to be replaced by nq,
it follows that

lil = tao nq dH(q) = n tao q dH(q) = n (2.8.2)

ti2 = {ao (nq + n2q2) dH(q) = n {ao q dH(q) + n2 {ao q2 dH(q)

= n + n2~iq). (2.8.3)
Similarly
ti3 = n + 3n2~2(q) + n3~3(q)
fi. 4 = n + 7n2~2(q) + 6n3~3(q) + n4~4(q), (2.8.4)
where q indicates that these are the moments about zero of the
structure dJ. H in question.
36 CLAIM NUMBER PROCESS

In addition to the moments, the standard deviation, skewness

and kurtosis of the distributions are often needed and will be given
now. First denote
u; = Q(2(q) - 1; Yq = ,u3(q)/U;; y2(q) = ,u4(q)/U: - 3. (2.8.5)
(See item 1.5(g) for notation.) Then, expressing the central moments
by means of the moments about zero (see (1.5.6)) and making use
of the formulae (2.4.3) and (2.4.4) the mean, variance, skewness and
kurtosis of the mixed Poisson distribution F are obtained
m=n
(j2= n + n2 u q2 (2.8.6)
y = (n + 3n 2 u q2 + n 3 u q3 Yq )/ii 3
Y2 = [n + 7n 2 u~ + 6n 3 Yq u; + n4 Y2 (q)u:]/ii 4 .
As might well have been expected, the mean is identical with the
Poisson case, but the standard deviation and skewness are greater,
thus increasing the chance of excessive claim numbers.
Because it is convenient to give the basic characteristics of
the structure distribution in terms of u q and Yq (seldom y2 (q)),
formulae in (2.8.6) include these characteristics instead of moments
,uj(q) which would make the equations slightly simpler.

(b) The moment generating function can be derived by weighting

in the same way as for the moments

M(S) = tXl enq(eS-l) dH(q). (2.8.7)

(c) For computation of the mixed Poisson distribution function

F, three alternative approaches are generally used.
(i) The structure function H will be expressed (approximated) in
an analytic form. This is exemplified in Section 2.9.
(ii) The standard deviation and skewness (or some other character-
istics) of q are first estimated and used in approximation
formulae which will be developed for the computation of P
fitting the moments. Hence no particular assumption of the
strict form of H is needed. This will be the standard method
applied in the following chapters.
(iii) The basic data, from which the estimate for H is to be derived,
2.8 MIXED POISSON DISTRIBUTION 37

may be given or estimated in tabular form such as in Table 2.8.1,

where the relevant range of q is divided in intervals i and the
frequencies hj = nJr.n j are calculated from the numbers nj
of the cases where q falls in the interval i. Then the mixed dJ.
can be computed (see (2.7.2) and (2.7.3)) from
(2.8.8)

which is often convenient for numerical computations.

The advantage of this method is that tabular values like those
in Table 2.8.1 can often fit the empirical data more closely than
any analytical estimate of H, and no further idealizations are
needed.

Table 2.8.2 gives examples of simple and mixed Poisson values.

The mixed function F was computed using the numbers of Table
2.8.1. As seen in the table, the probabilities differ significantly,
even when the H -distribution is symmetric. In many applications
the structure distribution is skew, having a longer tail for large
q-values. In such cases the deviations from the simple Poisson
probabilities are even greater.

**Exercise 2.8.1 Prove (2.8.1) and (2.8.7) directly by applying the

general formula

E(X)=E(E(XIV))= t+ oo
oo
E(Xlv= Y)dH(Y),

where H is the dJ. of the variable V.

Table 2.8.1 Example of the function H given in a tabular form.

(1 - q,) +0.25 +0.20 +0.15 +0.10 +0.05 0 -0.05 -0.10 -0.15 -0.20 -0.25

h, 0.10 0.05 0.05 0.10 0.12 0.16 0.12 0.10 0.05 0.05 0.10

Table 2.8.2 Examples of simple and mixed Poisson probabilities for n = 100
and for the H function are given in Table 2.8.1

k 110 120 130 140

1 - F(k) 0.147 0.023 0.0017 0.0001

1 - F(k) 0.279 0.132 0.0470 0.01 10
38 CLAIM NUMBER PROCESS

2.9 The Polya process

(a) Gamma as structure function For analysis and for practical

purposes it is often advantageous for the structure function to have
a closed analytical form as mentioned in item 2.8 (c)(i) as one of the
alternative approaches. The incomplete gamma function

(x ~ 0, h > 0), (2.9.1)

is often used, where h is a freely available parameter and r(h) =

1 dz denotes the well-known complete r function satisfying
J~ e - Z Zh -
r(i + 1) = i! for each non-negative integer i. Its special benefit is
the possibility of getting many useful formulae calculated in a
closed form, and it can represent a wide variety of distributional
shapes. On the other hand, the existence of only one parameter
limits the free adjustment of the ratio of the breadth and skewness
of the distribution.
The formula suggested is

(2.9.2)

where the upper limit is chosen to give the mean value E(q) = 1,
after which one parameter h is still open to choice.

(b) The moments and other characteristics are easily obtained by

integration and by observing that r(x) = (x - l)r(x - 1) (for
x ~ 1).

E(q) =1
O"~ = I/h O"q = I/Jh
J.l 3 (q) = 21h z Yq = 21Jh
J.l 4 (q) = 61h 3 + 31h z Yz(q) = 61h. (2.9.3)

(c) The shape of the r~distribution is shown in Fig. 2.9.1 for various
values of the parameter h. Values can be found in Pearson (1954)
or they can be computed using expansions which will be given in
Section 3.12.
 2.9 THE POLY A PROCESS 39

H'(q)

9 h=500

0.5 1.0 1.5 2.0

q
Figure 2.9.1 The incomplete gamma junction (2.9.2).

(d) The m.g.f. and convolution For further applications presented

later we give the m.gJ. of the incomplete gamma function r(x; h)
(see exercise 2.9.2)
(for s < 1). (2.9.4)
If M 1 (s) and M 2(S) are the m.gJ.s of r(x; h 1) and r(x; h 2), then
M 1(s)M/s) = (1- S)-hl(1 - S)-h 2 = (1 - S)-(h 1 +h 2 ). (2.9.5)
Hence, according to item 1.6(b) the convolution of two incomplete
gamma functions is again an incomplete gamma function having
h = h 1 + h2 as a parameter
(2.9.6)
where the right-hand side is to be understood as the convolution of
the distributions evaluated at the point x. Because r{cx; h) =
1 - e- cx , it follows that for positive integer values of h the r(cx; h)
is a convolution of h exponential distributions denoted by
r(cx; h) = Gh*(cx) for G(x) = r(x; I) = I - e-X, (2.9.7)
where everywhere x ~ O.
40 CLAIM NUMBER PROCESS

(e) The probability function (2.7.2) is now transformed as follows

- ( ) = foo -nq(nq)k_1_e-hq(h )h-Ihd

Pk n 0 e k! f(h) q q

khh foo e-(n+h)qqh+k-I dq

= _n__ (2.9.8)
f(h)k! 0

nkh h r(h +k)

f(h)k! (n + h)h+k
= r(h + k)(~)h(n + h)-(h+k)
f(h)k! n n

If the meaning of the binomial coefficient C) is extended by use of

the r function to cover the cases where the quantities i,j may be
non-integer, this may be written

i\(n) = (
h+k -
k
1) ( +
n
h
h
)h ( n +n h )k . (2.9.9)

This probability formula is well known as the negative binomial.

It can be easily verified that it reduces to the Poisson formula when
h ~ 00 (exercise 2.9.4).
From (2.8.6) and (2.9.3), the mean, variance, skewness and kurtosis
are obtained (see exercise 2.9.1)
IXI =n
(j2 = n + n 2 /h
(2.9.10)
y = (n + 3n 2 /h + 2n 3 /h 2 )j(j3
y2 = (n + 7n 2/h + 12n 3/h 2 + 6n 4 /h 3 )/(j4.

(I) The moment generating function can be obtained from (2.8.7)

by integration (see exercise 2.9.3):

M(s) = (1 - n(e S - l)/h)-h, (2.9.11)

which is valid for s < In(1 + h/n).

If h ~ 00, (2.6.1) is obtained by making use of the well-known
formula limn~oo (1 + a/n)" = ea.
2.9 THE POL YA PROCESS 41

(g) The numerical calculation of Pk and P can conveniently be

made by applying the recursion technique used in Section 2.5
(see (2.5.1) and (2.5.2)). Thus
Po = (h/(n + h))h, (2.9.12)

and
(2.9.13)
where
a = n/(n + h)
b = n(h - l)j(n + h).
It may be noted that integer-valued distributions having a recursion
rule of the form (2.9.13), where a and b are constants not depending
on k, constitute an important class of distributions. The Poisson
distribution (a = 0, b = n) is a member of this family (see 2.5.2)).
Generalizing a result of Panjer (1981), Jewell and Sundt (1981)
have recently shown that the compound Poisson function is also
computable by a recursion formula in certain conditions. This is
discussed in Section 3.8.
A few numerical values are provided in Table 2.9.1 and give some
idea of the flexibility of (2.9.9) compared with the simple Poisson
dJ.

Table 2.9.1 Examples of I - F per thousand.

h= CIJ 100 20 10 5
n k

10 12 208 217 243 262 281

14 83 93 126 154 188
16 27 34 58 84 121

50 60 72 115 205 250 285

70 3 13 71 122 177
80 0 1 19 53 105
100 110 147 224 312 340 356
120 23 78 194 246 285
130 2 20 112 172 225
140 0 4 60 116 175
42 CLAIM NUMBER PROCESS

(h) Discussion The Polya process is fairly convenient for applica-

tions and leads to many important formulae, which are easy, or
at least possible, for computational use. On the other hand, it
requires a smoothing procedure and some idealization of the process
and it does not seem easy to estimate the error due to this smoothing.
Because there is only one free parameter, h, available, the Polya dJ.
is not always flexible to fit various structure distributions, especialIy
if they are rather skew, as wilI be seen in Section 2.10.

(i) Contamination model It is interesting to note that there is

also another way to derive the negative binomial distribution,
originalIy formulated by Eggenberger and Polya (1923). This
may be done by means of an urn model as folIows. Assume that N I
red balIs and N 2 white balIs (reds representing accidents, diseases,
fires or other casualties) are placed in an urn. A balI is drawn at
random repeatedly s times. After each draw the balI is returned to
the urn together with C balIs having the same colour. Hence before
the sth draw the number of balIs in the urn is N + (s - l)C, which
provides the denominator for the s th probability. Here N = N I + N 2.
By combinatorial reasoning it can be shown that the probability of
getting exactly k red balIs is
k-I s-k-I

S)TI(NI+UC) TI (N 2 +vC)
prob{k = k} = ( k u=o s I v=o
TI (N + wC)
w=O
(2.9.14)

The purpose of this model is to introduce contamination into the

model. Each event gives rise to an increased probability of the same
kind of events in future because balIs of the same colour were put
into the urn.
Defining
I/h = CjN 1 = degree of contamination, (2.9.15)
and then performing the sequence of passages
N --> 00 ; S --> 00 ; sN liN --> n, (2.9.16)
it follows that the contamination, which was originalIy assumed to
occur at discrete times, will occur continuously (i.e. in infinitely
short time intervals) so that the expected number of events during a
2.10 RISK EXPOSURE VARIA nON 43

fixed observation period, say in one year, will be kept constant and
equal to n. It can be proved that the passage leads to the negative
binomial formula (2.9.9). Hence the negative binomial can be
derived by assuming contamination between the risk units, e.g.
epidemic diseases or the spreading of fire. For this reason the
negative binomial distribution is often called a Polya distribution
and the corresponding mixed process is called a Polya process.

Exercise 2.9.1 Prove that the moments about zero of r(x; h) are
rl i = r(h + i)/r(h), (2.9.17)
and calculate the characteristics (2.9.3) and (2.9.10).

Exercise 2.9.2 Prove (2.9.4).

Exercise 2.9.3 Derive the moment generating function (2.9.11).

Exercise 2.9.4 Prove that the negative binomial probability function

(2.9.9) is reduced to the Poisson probability as h --+ CD.

Exercise 2.9.5 Calculate and plot in the same diagram the Poisson
function Pk and the corresponding Polya Pk for n = 5 and h = lO.

Exercise 2.9.6 The aggregate claim numbers of two stochastically

independent portfolios are Polya distributed. Prove that, if the
portfolios are merged, the joint distribution is again of Polya type
providing that the parameters nand h are the same for both the
original portfolios.

Exercise 2.9.7 Prove that the dJ. Fn(k) of the Poisson variable can
be expressed in terms of the gamma dJ.
1 - F(k; n) = r(n, k + 1).
Exercise 2.9.8 For which value of k does Pk(n) as given by (2.9.9)
achieve its maximum?

2.10 Risk exposure variation inside the portfolio

(a) Individual risk proneness The mixed Poisson distributions also

have applications in environments other than those just described.
44 CLAIM NUMBER PROCESS

An example is the situation when the risk proneness ofthe individual

risk units of the insured portfolio is considered, e.g. the problem
may be to find the dJ. of the number of claims arising from a single
motor-car policy. The physical process may justify the assumption of
a Poisson law for the accidents, but the risk parameter n, the expected
number of claims per car, can be expected to vary for different cars
depending on the type, use, exposure time (mileage), etc., of the car
and the skill of the driver. It can be assumed that each risk unit i,
a car in this example, is involved with a proneness parameter ni = nqi
which is the expected number of claims pertinent to this unit. Here
n is an average value and qi a coefficient indicating the deviation
per unit from n. Let H be the dJ. which describes the variation of the
qi values (such a dJ. can be assumed to exist even though it often
may be unknown or only partially estimated in practice). This
function characterizes the distribution of risk inside the portfolio (or
inside some particular part of the portfolio under consideration, e.g.
some class of motor-car).
The distribution of the claim number variable k of the individual
unit which is selected at random from the portfolio can be obtained
by first taking the probability that the risk parameter q is in the
interval qi' qi + dq and then assuming the Poisson law for the
parameter value nqi' The construction of the probability expression
is exactly analogous to that applied in the derivation of formula
(2.7.2). Only the physical environments are different - in (2.7.2) it was
the variation of the Poisson parameter n from one time unit to the
next, in the present case it is the variableness of it from one risk unit
to the next. Hence (2.7.2) is readily applicable

Pk = prob {k = k} = too Pk(nq) dH(q), (2.10.1 )

where Pk is the standard Poisson probability (2.4.2). Also in this

connection H(q) is generally called a structure function (see Ammeter,
1948; Biihlmann, 1970). This is an important concept in many
applications, including credibility theory (Section 4.1 0) in particular
and in rate-making in general.

(b) Example For illustration consider the example given in Table

2.10.1. The statistics are taken from a study by Johnson and Hey
and relate to claims under UK comprehensive motor policies in 1968.
The 421 240 policies were classified according to the number of
2.10 RISK EXPOSURE VARIA nON 45

Table 2.10.1 Comprehensive motor policies according to the number of claims

in 1968.

k Nk Poisson Neg. binomial Two Poissons

0 370412 369246 370460 370460

1 46545 48644 46411 46418
2 3935 3204 4045 4036
3 317 141 301 306
4 28 5 21 20
5 3 1 1

claims in the year 1968, the average number of claims per policy
being 0.131 74 and the variance 0.13852. The column headed
'Poisson' sets out the distribution that would result if the occurrence
of claims had followed the Poisson law with n = 0.131 74, i.e. the
expected number of claims per policy in one year. As will be apparent
the Poisson distribution is theoretically shorter than the data, an
observation confirmed by the chi-squared test. In other words, the
hypothesis that the risk proneness is different for different policies
is confirmed.
The insufficiency of the Poisson law could also be anticipated
from the fact that the variance is greater than the mean, whereas they
should be equal if the Poisson law were valid, as will be seen from
(2.4.3).
The column headed 'Negative binomial' sets out the distribution
according to this law with parameters n = O. 13 174 and h = 2.555,
the latter being found by the method of maximum likelihood. The
value of chi-square is 6.9 which gives a probability of 0.14 for 4
degrees of freedom, so that the representation is acceptable. There
is a slight indication that the negative binomial may be under-
representing the tail and for some applications it might be desirable
to elaborate the model, but for applications which have no signi-
ficantly large skewness the model may be safely used.

(c) Discrete structure function Another approach is to approximate

the structure function H(q) by a discrete dJ. assuming values ql'
q2' ... , q, with probabilities hi' h2' ... , h,. This means, in fact, that
the dJ. is composed of r Poisson terms. The greater the number
of free parameters, the better the possibility of achieving a reason-
able fit even for heterogeneous portfolios.
46 CLAIM NUMBER PROCESS

In the present case a two-term distribution already gives a quite

satisfactory result. The parameter values qt = 0.653 41 and q2 =
2.1293, with probabilities 0.765 19 and 0.23481, can be found from
the equation
hpk(nt) + (1 - h)Pk(n 2 ) = Pk(n),
and equating moments with respect to k. In principle the problem is
related to Gaussian quadrature and a simple treatment with a
number of tables of relevant numerical values will be found in
Beard (1947).
There is a substantial literature about structure functions. The
problem of finding the parameters qi and hi for the discrete approxi-
mation was dealt with by D'Hooge and Goovaerts (1976). Gossiaux
and Lemaire (1981) studied the fit of the above methods and applied
them to motor-car accident statistics. Loimaranta et al. (1980) have
presented a cluster analysis approach as a solution of the same
problem.

(d) Terminology For the purposes of this book the inner variation in
the collective is not relevant. The collective will be treated as a
whole and the heterogeneity taken care of by the expected number
of claims n. Thus in what follows H(q) and the term 'structure
function' will represent only short-term variations in n, i.e. the random
fluctuation from one accounting period to another. Longer-term
variations are dealt with later in Chapter 6.
The reader will appreciate that this terminology deviates from
the practice sometimes assumed in the literature where a structure
function may refer mainly to the internal heterogeneity of collectives.
 CHAPTER 3

Compound Poisson process

3.1 The distribution of claim size

(a) Definitions Consideration is now extended from the claim
number processes to processes which operate the claim amounts,
concerning both the individual claims and their sums, the aggregate
claims. A primary building block is the randomly varying size Z of an
individual claim, i.e. the sum to be paid by the insurer at occurrence
offire, accident or any other event insured against. It is assumed that
the claim sizes Z arising from different claim causing events are
mutually independent and equally distributed, having a d.f.

S(Z) = prob{Z:::;; Z}. (3.1.1 )

Following the collective approach outlined in item 2.l(b) no regard

is paid to the risk unit (policy) from which the claim has arisen. The
dJ. S describes the variability of sizes of the continual flow of claims.
The aspects discussed in Section 2.3 suggest that payments due to
one and the same event should be united as one claim irrespective of
whether or not they formally concern different policies (e.g. different
owners of a property in a damaged building complex or in a vehicle).
Furthermore, if two or more claims are coming from the same risk
units they are considered as different claims if they are caused by
different events.
The approaches, which will be dealt with in Chapters 3-7, will
be formulated so as to be generally applicable, as far as possible, to
all kind of insurance. In the case of life insurance the claim size Z
should be defined as the difference of the sum S paid by the insurer
and the policy reserve V released thereby, i.e. Z = S - V. Special
features concerning life insurance will be discussed in Chapter 8.
The claim size variable introduces to the processes concerned a
new layer of stochasticity in addition to the claim number variation.
48 COMPOUND POISSON PROCESS

The processes constituted by the stochastic variation of both claim

number and claim sizes are called compound processes.
The existence of a function S is in conformity with general experi-
ence, at least as regards periods of moderate length and pro vi oed
the effect of changes in monetary values is eliminated, for example
by methods which will be presented later. The actual claims can be
recorded and numerical estimates obtained for S. At the outset it
will only be assumed that the function exists and that it is known.
Subsequently the details of its practical computation will be con-
sidered and a few more common distributions will be recorded in
Section 3.5.
As mentioned in Section 1.3, only distributions with positive
risk sums, i.e. Z ~ 0, are dealt with in this book.

(b) Three different types of S-function occur in the applications

and are shown in Fig. 3.1.1 with the corresponding densities or
frequencies s.
In Fig. 3.1.1(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 substantially increasing the relative inci-
dence of the smaller claims.
The discrete function of Fig. 3.1.1(b) could arise from a friendly
society granting fixed funeral expense benefits or from a company
which has standardized the benefits under its policies, such as
travel accident insurance (often sold through airport automats)
where the face sums are fixed optional sums.
A mixed type of function is shown in Fig. 3.1.1(c); this can arise
from a basic distribution oftype (a), subject to re-insurance arrange-
ments, 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 classes of risks, several steps may be shown in
the S-funttion.
For practical application it is sufficient to assume that the
S-function is one of the types mentioned above. Furthermore, it is
assumed that the derivative S'(Z) exists for types (a) and (c) and is
continuous except at a finite number of points.
 3.1 THE DISTRIBUTION OF CLAIM SIZE 49

S IZ)

s IZ) s (Z) s (Z)

(0) (b) (c)

Figure 3.1.1(a) A continuous function: (b) a discrete function: (c) mixed type.

(c) Outstanding claims To keep the risk-theoretical models within

reasonable dimensions, it is usually assumed that the claims are paid
out immediately they are incurred. In practice, there is inevitably some
time lag between the occurrence of the event giving rise to a claim
and its settlement, whether from the minor aspect of administrative
procedures or from legal problems such as the determination of
liability or the assessment of amount or from delays in notifying
the insurer of claims which are due the so-called IBNR problem,
(an abbreviation for incurred but not reported). Thus, in addition to
the problems of assessing the probable amount of claims which
have been notified, some allowance has to be made for the expected
late-notified cases.
In principle, any errors in estimates of outstanding claims will
ultimately be corrected when the claims are finally settled, but they
can give rise to a shift of profits or losses between years and in that
way affect also the risk fluctuation; however, they will probably be
of much smaller magnitude than the 'ordinary fluctuations'. It
should be noted that the assumption concerning the immediate
payment of claims does not result in the elimination of the estimation
50 COMPOUND POISSON PROCESS

errors of the outstanding claims from the model. Like any other
inaccuracy of the basic assumption, it gives rise to extra fluctuations
in underwriting results, probably in a periodic manner as assumed
in item 1.1(0 and as will be further discussed in Section 6.2. When
the model parameters are calibrated on the basis of observed actual
fluctuations, the effect of these inaccuracies will be automatically
taken into account. Furthermore, there is no essential obstacle to
the introduction of the outstanding claims or rather their estimation
error as a particular entry to the model. A brief indication of such an
approach will be presented in item 1O.2(e).
On the other hand, systematic under- or overestimation can give
rise to considerable bias in the balance sheet and thus in evaluation
of the actual solvency margins (risk reserves). The consideration of
these as well as many kinds of 'non-stochastic' aspects, e.g. in-
calculable risks jeopardizing the existence of insurers such as major
failures in investments or risk evaluation, misfeasance or malfeas-
ance of management, etc., are essential parts of the general solvency
control of insurance industry, but they fall outside the scope of this
book (see Pentikiiinen, 1982; Section 2.9).
Analysis of the development of claims estimates is a normal part
of business routine and will indicate the need for some extra pro-
visions; it might be thought, for example, that a margin is needed
to deal with inflationary changes with respect to both the average
level of inflation as well as the need for emergency measures in cases
when the rate of inflation may occasionally be soaring. Even though
inflation will be incorporated into the model assumptions, its special
effects on the claims reserve will not be further discussed. Such
adjustments might be needed in assessing the parameters of the
overall model and will have to be dealt with in individual circum-
stances.

3.2 Compound distribution of the aggregate claim

(a) Derivation of the d.f. An insurance portfolio is again considered

and it is desired to find the probability distribution of the total
amount of claims X, or briefly the aggregate claim, which occurs
during a time interval (e.g. I year). The probability Pk that the number
of claims equals k, and the distribution function S of one claim, are
assumed to be known.
3.2 COMPOUND DISTRIBUTION 51

The required distribution F(X) gives the probability of the event

X ~ X. This event can occur in the following alternative ways:
(i) In the time interval no claim occurs.
(ii) The number of claims = I 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 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
00

F(X) = L PkSk(X), (3.2.1)

k=O
If it is assumed that the amounts of the claims are mutually inde-
pendent, the function Sk(X) is well known from probability calculus
as the kth convolution of the distribution function S(X), which can

f:
be calculated from the recurrence formula.

Sk(X) = Sk_t(X -Z)dS(Z) = S(k-ll* *S(X) = Sk*(X) (3.2.2)

and the following important formula is obtained for the dJ. of the
aggregate claims
00

F(X) = L PkSh(X). (3.2.3)

k=O

(b) Terminology The distribution function (3.2.3) is called

compound, referring to the compound process behind it. If the claim
number process related to X is a (mixed) Poisson process then X is
called a (mixed) compound Poisson process and the function (3.2.3) as
(mixed) compound Poisson df. In this case, which is mostly assumed
in the following, Pk is either the simple Poisson probability (2.4.2) or
the mixed version (2.7.2). The prefix 'mixed' is often omitted for
brevity.
The case where Pk is defined as the simple Poisson probability is
often called the Poisson case; where Pk is the negative binomial
probability (2.9.9) this is known as the Polya case.
52 COMPOUND POISSON PROCESS

For the sake of simplicity the notation Pk will be replaced by Pk

provided that the meaning is clear in the particular context.

(c) On applicability The dJ. (3.2.3) is, unfortunately, directly

useful for numerical computations only by making special assump-
tions concerning S or if n is very small. The formulae are mainly
applicable when X is given and F(X) is sought, and not easily in the
reverse direction from F(X) to X. Finding workable approximation
methods is a major problem and several are given later, but first
some general features of F are considered.

3.3 Basic characteristics of F

(a) Basic moments The characteristics of the (mixed) compound

(Poisson) distribution can be expressed in terms of the moments of
the claim number process, which were given in Section 2.8, and of
the moments about zero of the S function

ai = f: Zi dS(Z), (3.3.1 )

of the claim size dJ. S. It is convenient to choose for the lowest

moment, which indicates the mean claim size, the special notation
(3.3.2)

(b) The moments Pi about zero of compound distributions, can be

derived from the following general formula

= too Xi k~/k dSk*(X)

(3.3.3)
k~/k LOO Xi dSk*(X)
00

L Pkajk),
k=O

where a}k) is the jth moment about zero of the sum of k individual
3.3 BASIC CHARACTERISTICS OF F 53

claims
Zl +Z2 + ... +Zk'
The terms of this sum are mutually independent according to
assumptions made in Section 3.1 and have the same dJ. S. Hence
the first moment
(3.3.4)
and the second and third central moments of the sum of claims can
be summed from its components:
f.1.(k)
} }
= f.1.,(Zl) + '" + f.1. ,(Zk)
}
U = 2 or 3).
Then the second moment about zero, needed for (3.3.3), can be
calculated as follows
a~l = f.1.~k) + (a\k l)2
= kf.1. Z(Zl) + k 2 mZ (3.3.5)
= k(a 2 - mZ) + k 2 mZ
= ka 2 + k(k - 1)m 2 .

(c) Characteristics of compound Poisson distribution These results

are valid for compound distributions in general, i.e. for distributions
which are composed of a (claim) number process and of a (claim)
size process. Next the claim number process of a compound Poisson
type is assumed. Substituting these expressions in (3.3.3) and making
use of the earlier results (2.8.2) and (2.8.3), the moments of the aggre-
gate claim are obtained:
00 00

/3 1 = L Pkkm = m L kpk = mn,

k=O k=O
00

/3 2 = L Pk[ka 2 + k(k - l)m Z]

k=O
00 00 (3.3.6)
= (a 2 - m2 ) L kPk + m2 L PPk
k=O k=O
= (a 2 - mZ)n + m2 (n + nZa2 (q))
= na 2 + n2 mZaz(q)·

In a similar way the higher moments can also be obtained. A more

convenient method for their calculation is, however, the use of the
m.gJ., as will be seen in the next section.
54 COMPOUND POISSON PROCESS

The central moments of the aggregate claim distribution are now

readily calculated by means of (1.5.6), and by some further algebra
(see exercise J.3.5) the following important expressions can be
obtained
mean Jlx = E(X) = nm = P,
variance O'i = var(X) = na 2 + n2m20'~.
= (r 2/n + 0'~)P2,
skewness Yx = Jl 3 (X)jO'i = (na 3 + 3n2ma20'~ + n3m3yqO'!)jO'i
= (r 3/n2 + 3r 20'~/n + yqO'!)/(r 2/n + 0'~)3/2
kurtosis yiXl = JliX)/O'~ - 3
= (na 4 + 4n 2ma 3 0'2q + 3n 2a220'2q +
+ 6n 3m2a2yqO'! + n4m4Y2(q)0':)j0'~, (3.3.7)
where q is the structure variable. Because just the basic character-
istic y and Y2 will be used as standard input parameters they are
written into the above expressions, despite the fact that replace-
ment ofyqO'! by Jl 3 (q) and y2(q)0': by Jliq) - 30': would have some-
what simplified the formulae. P = nm is the risk premium income
covering the claim expenditure due to X, as will be defined in
Section 4.1.
The relations a2/m2 and a3/m 3 are needed so frequently that it
is convenient to introduce special notation for them:
r 2 = a 2 /m 2 = risk index
r3 = a3/m3. (3.3.8)
A merit of these indexes is that, as a first approximation, they are not
affected by inflation because the numerator and denominator
increase in the same ratio in case ofa change of the value of money.

(d) In the Poisson case i.e. when the structure variable q is constant
(= 1), (3.3.7}is reduced as follows
Jlx = nm
2_
O'x - na 2
a r
Y = 3 = 3 (3.3.9)
X ai/ 2 In r~/2 In
3.3 BASIC CHARAC. ~l{ISTICS OF F 55

(e) For the Polya case i.e. when the structure function is of gamma
type, the corresponding expressions are derived by using (2.9.10)
I1x= nm
ax = na 2 + n2 m2 /h
2

Yx = (na 3 + 3n 2 ma 2 /h + 2n 3m3/h 2 )jai, (3.3.1 0)

y2 (X) = [na 4 + 3n 2 a;/h + 4n 2 ma 3/h + 12n 3m2 a2 /h 2
+ 6n4m4/k3]/a~.
(t) Analysis of the background effects The above formulae separate
the effects of the components of stochasticity introduced in item
3.1 (a).
The variance ai
is composed of two terms. The first represents
the variance, if the basic parameter n were constant. The second
term is the increment arising from the fluctuation in n according to
the dJ. H (see (2.7.1)).
As an illustration, Table 3.3.1 sets out some examples of the
relative share of the structure component in ax and the 'pure
Poisson' component in
(3.3.11 )
where aD is the standard deviation obtained from equations (3.3.9)
for the pure Poisson case. The model parameters for the examples
where r 2 = 44 and a q = 0.038 which will be chosen as standard
values for examples in item 4.2(b). Furthermore, two Polya cases
are calculated having r 2 = 44 and h = 100 or 1000.
It can be seen how the ordinary random fluctuation ofthe number
and the size of claims is predominant for small companies, whereas
for large companies the position is reversed.

Table 3.3.1 The share p of the standard deviation due to the structure variation.

100o-xiP lOOp
n Standard Standard Polya h = 100 Polya h = 1000

10 209.8 0.0 0.1 0.0

100 66.4 0.2 1.1 0.1
1000 21.3 1.6 9.7 1.1
10000 7.6 13.2 44.7 9.7
100000 4.3 51.7 79.5 44.7
1000000 3.9 82.8 93.4 79.5
56 COMPOUND POISSON PROCESS

Table 3.3.2 The components (3.3.12) as a percentage of the total variance ai.

n lOOax/P var(k) var (Z) var(q)

10 209.8 2.3 97.7 0.0

100 66.4 2.3 97.4 0.3
1000 21.3 2.2 94.6 3.2
10000 7.6 1.7 73.6 24.7
100000 4.3 0.5 22.8 76.6
1000000 3.9 0.1 2.9 97.0

A further decomposition is of interest

var(X) = ai = nm 2 + n(a 2 - m2) + n2m2a~.
= m2 var(k) + n var(Z) + n2m2 var(q), (3.3.12)
where the first term represents the variance if the claim number only
where stochastic, var (k) = n being the Poisson variance of claim
number variable k. The second term arises when the claim size Z is
also made stochastic. The third term represents the increment
when the structure variable q is introduced to the model.
In Table 3.3.2 an example is shown of the magnitude of the three
components which control the stability of an insurance collective,
employing the same standards as above.
It is evident that the component due to the simple Poisson
variation of the claim number is slight. For small collectives the
variation of the claim size is predominant and for the large collectives
the variation caused by the structure variable q is predominant. Of
course, the conclusion concerns only this example and may be
different for other kinds of portfolios.

(g) Limit distributions From (3.3.7) it is also possible to draw some

conclusions regarding the behaviour of the process when the port-
folio grows large, i.e. when n tends to infinity. This can be seen by
considering the relative claim amount variable
x = XI E(X) = Xlnm, (3.3.13)
which has standard deviation
_ (jx _ J(na 2 + n2m2(j~) _ J( I 2)
a - - - - r2 n + a . (3.3.14)
"nm nm q
3.3 BASIC CHARACTERISTICS OF F 57

It is immediately apparent that (J x tends to (J q when n ~ 00. Similarly,

Yx ~ Yq and Yz(X) ~ Yz(q)· These observations suggest that the limit
dJ. of x is H, the dJ. of the structure variable q. This is in fact the
case, as was proved by O. Lundberg (1964) (see exercise 3.3.6).
In the 'Poisson case', i.e. when the structure variable degenerates
to one point 1 ((Jq = 0), the standard deviation (Jx tends to zero (and
only in this case). Then the law of large numbers is applicable, as
can be easily seen by dividing the time interval into equal subintervals
and considering X as a sum of independent and equally distributed
variables, since (J q = O. In this case the central limit theorem gives
the asymptotic relation
F(X) ~ N( (X - nm)/(Jx) for n ~ 00, (J q = O. (3.3.15)
In the general case when (Jq > 0, F is not asymptotically normal but
instead has the shape of the structure dJ. H, which means that
F(X) ~ H(X/nm) for n ~ 00, (Jq > O. (3.3.16)
It obviously depends on the properties of both the claim size dJ. S
and the structure dJ. H how rapidly the approximation (3.3.16)
provides a reasonably accurate approximation for F, i.e. whether
it can be used whilst n is not very large. Table 3.3.2 suggests that
quite large n values are required before the structure component
turns predominant.
In practical application the passage n ~ 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 classes and subclasses of
business and it may 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, may tend to the normal
distribution in accordance with the central limit theorem. In other
words, if n ~ 00 so that new independent groups are incorporated
into the collective, then F can be approximated by the normal
distribution. In practice different mixed cases can occur, e.g. the
changes in general economic conditions may have simultaneous
parallel effects on many classes of the portfolio whereas some other
fluctuations may have effects limited to one class only. Hence care
is needed in deciding which assumptions are applicable to each
58 COMPOUND POISSON PROCESS

actual case. The problems related to the division of the portfolio

into sections will be dealt with in Section 6.4. Philipson (1968)
has treated these kinds of passages.

(h) Direct calculation of F Formulae (2.7.4) and (2.8.8) can easily

be extended to the compound variable X (see exercise 3.3.3)

fleX) = tXl Fnq(X) dH(q), (3.3.17)

of if H is discrete or approximated by a discrete dJ.

(3.3.18)

where Fnq is the simple compound Poisson dJ. (aq = 0) having the
expected number of claims nq. The merit of (3.3.18) is that it is
not necessary to try to find any analytical presentation for H in
cases where it is evaluated, for example from empirical data.

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 ofthe members, the remainder
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(X) and, ax for the society mentioned

in the previous exercise.

Exercise 3.3.3 Prove (3.3.17).

Exercise 3.3.4 Show that if Zl' ... ,Zk are mutually independent
random variables, then the third central moments satisfy the
equation

113 (Jl Jl
Zj) = 113 (Zj)'

provided the moments exist.

Exercise 3.3.5 Derive ax and 'Yx (see (3.3.7)).

3.4 THE MOMENT GENERATING FUNCTION 59

**Exercise 3.3.6 Let Gn be the dJ. of x = X/P (see (3.3.13)). Show

that Gn -+ H as n -+ 00, where H is the structure dJ. It is assumed that
H is discrete.

**3.4 The moment generating function

(a) The m.g.f. of the (mixed) compound Poisson distribution can be

obtained by a straightforward application of definition (1.6.1)
as follows

M(s) = M x(s) = too eXs dF(X)

= f: e Xs dX[k~/k(n)Sk*(X) ]
k~/k(n) too e Xs dSk*(X).
This last integral is the m.gJ. of the variable Zl + ... + Zk' Owing
to the independence of the individual claim sizes and the fact that
they have a joint dJ. S, it can be expressed (see property (iv) of item
1.6(b)) as the kth power of

M z(s) = too eZs dS(Z). (3.4.1)

Furthermore, replacing Pk(n) by (2.7.2) gives

M(s) = k~J too e-nq(:qr dH(q) JMZ(S)k

= foo e- nq I (nqM~(s))k dH(q) (3.4.2)
o k=O k.
= too enq(Mz(s)-l) dH(q).
The moments about zero and the basic characteristics (3.3.7) of
the mixed compound Poisson distribution can be derived making
use of the m.gJ. (see exercise 3.4.1).

(b) In the Poisson case when H(q) = e(q - I) (see (1.5.3)) the follow-
ing formula is obtained as a special case of (3.4.2)

(3.4.3)
60 COMPOUND POISSON PROCESS

(c) In the Polya case the m.gJ. is (see exercise 3.4.3)

M(s) = [ 1- ~(Mz(s) - 1) T
h
• (3.4.4)

M(s) does not necessarily exist if the expression in brackets is

negative. It is, however, positive near the origin if S(Z) is continuous
at the origin. Otherwise its existence must be checked for each
application.

Exercise 3.4.1 Calculate the moments about zero Pi (see (3.3.6))

for i = 1, 2, 3, 4 and the characteristics (3.3.7) making use of the
m.gJ.

Exercise 3.4.2 Check that the m.gJ (2.8.7) is obtained from (3.4.2)
by substituting S(Z) = 8(Z - 1).

Exercise 3.4.3 Prove (3.4.4).

3.5 Estimation of S

3.5.1 INDIVIDUAL METHOD

(a) General aspects In most applications of the theory of risk it is

necessary to know the claim size distribution function S more or
less accurately. It should be so fitted that the representation cor-
responds as closely as possible to the true distribution ofthe amount
of one claim in the portfolio. The fit should align itself to the data
which are, in general, empirical. Insurers always have data files
containing detailed information of both the policies and the claims,
and many kinds of statistics are produced for counting, rate-making
and other purposes. Construction of the claim size distributions
and other data needed for risk-theoretical analyses can be obtained
directly or by some modifications as side products from these data
processes.
Some methods of estimating the S function from the data available
are now presented in this and subsequent sections.

(b) Policy files as basis First a method is given for computing the
S function starting from the individual policies of an insurance
3.5 ESTIMATION OF S 61

portfolio. This approach is convenient for practical calculations

only in special cases. It is presented mainly because it describes in
an illustrative way the connection of the claim size dJ. and the
portfolio structure, which is conventionally recorded as files
containing information on the existing policies.
The risk units (policies) are numbered i = 1, 2, 3, ... , N and the
corresponding frequency rates* are assumed to be known and
are denoted by ql' Q2' ... , QN' It is further assumed that only one
claim size Zi is possible for each unit, i.e. no partial claims can
occur. For insurance classes where partial claims are possible the
method is less convenient, but it can be modified also for that
case when no other method is applicable, as will be shown in
Section 3.5.3.
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
q/n,
where n = I.qi' the sum being extended over the whole portfolio.
Since S(Z) is the conditional probability that the claim is ~ Z
the addition and multiplication rules of probability give immediately
1
S(Z)=- L qi'
(3.5.1)
n Zt~Z

Exercise 3.5.1 A company grants insurance for accidental death,

the sums payable at death being standardized at £100, £250, or
£500. The number of polices in these classes are 5000, 1000, 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

• The term 'probability of claim' is often used in this connection, but qi must, in
fact, be regarded as a frequency or, what is the same, the expected number of events
(which might even be ;;, I). 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 occur-
rence of at least one event is clearly p = I - e -q ::::; q. In this sense, reference is some-
times 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.
62 COMPOUND POISSON PROCESS

anti-selection, the rate in the £500 class is estimated to be double that

in the other classes. What is the S function for this business?

3.5.2 STATISTICAL METHOD

(a) Claim statistics 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 3.5.1 which sets out claims arising from a combined
experience of Finnish insurance portfolios comprising industrial
fire risks.

Table 3.5.1 Compilation of the claims statistics.

2 3 4
n.
Z x 10- 3 £ ni ~S=---.!. S=L~S
n
I 0.010 283 0.033953 0.033953
2 0.016 280 0.037664 0.071617
3 0.025 157 0.045479 0.117096
4 0.040 464 0.055413 0.172 509
5 0.063 710 0.063707 0.236216
6 0.100 781 0.068234 0.304450
7 0.158 530 0.070466 0.374915
8 0.251 446 0.070370 0.445285
9 0.398 491 0.071745 0.517030
10 0.631 673 0.074009 0.591039
II 1.000 779 0.075761 0.666800
12 1.585 741 0.073025 0.739825
13 2.512 520 0.064 899 0.804724
14 3.981 425 0.052757 0.857481
15 6.310 323 0.040 152 0.897633
16 10.000 179 0.029698 0.927331
17 15.849 173 0.021660 0.948990
18 25.119 112 0.Ql5765 0.964755
19 39.811 94 0.011310 0.976065
20 63.096 57 0.008222 0.984287
21 100.000 39 0.005599 0.989886
22 158.489 22 0.003767 0.993653
23 251.189 17 0.002424 0.996077
24 398.107 12 0.001582 0.997659
25 630957 5 0.001022 0.998680
3.5 ESTIMATION OF S 63

Table 3.5.1 (contd.)

2 3 n. 4
Z x 10- 3 £ ni tlS=~ S = L:tlS
n

26 1000.000 5 0.000600 0.999280

27 1584.890 3 0.000330 0.999610
28 2511.890 1 0.000179 0.999789
29 3981.070 0 0.000097 0.999886
30 6309.570 2 0.000052 0.999938

31 10 000.000 0 0.000028 0.999967

32 15848.900 0 0.000015 0.999982
33 25 118.900 0 0.000008 0.999991
34 39810.700 0 0.000005 0.999995
35 63095.700 0 0.000002 0.999997

36 100000.000 0 0.000001 0.999999

37 158489.000 0 0.000001 0.999999
38 251 189.000 0 0.000000 1.000000
39 398108.000 0 0.000000 1.000000
40 630958.000 0 0.000000 1.000000

41 1000000.000 0 0.000000 1.000000

Monetary unit £1000. Zi = 10- 2 .2+.,5. n = Ion i = 8324. ni = number of claims in

class Zi _ I < Z ,;;; Zi' Finnish industrial fire insurance. The tail (i ;;. 27) fitted with
Pareto dJ. 1 - S(Z) = 7.162 88 x Z-1.332585. The data were provided by Ham Lonka
and Jarmo Jacobsson, Statistical Centre of the Finnish Non-Life Insurers.

(b) Tabular or analytic form A claim size table can be used as a

basis for analytical curve fitting, as will be discussed in Section
3.5.4, or it can be used for straightforward numerical calculations.
For example the moment integrals, which are frequently needed
for applications, can be replaced by a discrete sum as follows
ak = fOO Zk dS(Z) ~ IZ~ f!.Si' (3.5.2)
o i
in which Zi and f!.Si are to be taken directly from Table 3.5.l. This
method is often convenient and it avoids rounding-off inaccuracies
which arise when the empirical distribution is replaced by some
analytical curve. In other words, the empirical values as such may
describe better than any analytical function the actual but always
unknown distribution which is behind the observed data. On the
other hand, the statistical data are only a sample from the actual
distribution and care is needed to take into account the sampling
errors.
 64 COMPOUND POISSON PROCESS

(c) Piecewise construction The statistical method provides that

the data base of the statistics is sufficiently large that the inaccuracy
can be expected to be slight. This rarely holds for the upper tail
of the distribution in cases where very large claims are possible.
In such cases it may be advisable to divide the range of the relevant
values of Z into two or more pieces and apply to each of them
the appropriate functions for the environment in question. For
example, claim sizes up to some limit Zo can be employed in a
tabular form and the tail Z ~ Zo approximated by some analytic
function which by experience can be expected to give a good fit.
In fact this was done in Table 3.5.1 for the large Z-values, for which
the shape of the empirical distribution according to Fig. 3.5.1 was
already clearly irregular.

(d) Smoothing The empirical data of Table 3.5.1 and Fig. 3.5.1
were mechanically smoothed by replacing each of them by a moving

AS 10-1
i
~

10-2 : ~

10-3

10-4 iii

10- 5 \

10-6

Figure 3.5.1 Claim size densities of Finnish industrial fire insurance. The data
of Table 3.5.1 are plotted on a double logarithmic graph. The points indicate
observed data. Unit for Z is £1000.
3.5 ESTIMATION OF S 65

average of five values. This kind of smoothing is motivated if the

sample errors are noticeable. On the other hand, some seemingly
irregular bends in the curve may reflect significant special features
in portfolio structure, e.g. clusters of policies of a special type, in
which case their 'smoothing away' is not appropriate. A further
analysis proved that this was the case in Fig. 3.5.l. Therefore the
bends for small Z-values were preserved.

(e) Selection ofthe class interval is a problem associated with the

statistical method. In Table 3.5.1, 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.

(f) Reinsurance A normal practice is to protect the portfolio

against excessive risk fluctuations by ceding the top risks to re-
insurers, as will be dealt with in Section 3.6. If only that part of the
business which is retained on the insurer's own account is analysed
then the estimation and curve fitting problems concerning the tail
become insignificant.

(g) Effect of inflation If the claims are collected over a period

during which monetary values change, it is necessary to rectify
the values by means of a suitably chosen price or other 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
periods do not include sufficient 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 problems arising from large claims can often be dealt with as
described in the next section.

3.5.3 PROBLEMS ARISING FROM LARGE CLAIMS

(a) Prolonged observations period In many practical problems,

where the top risks are not cut away by reinsurance, the larger
66 COMPOUND POISSON PROCESS

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. Para-
doxically the least-known part of S 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 experience extending over a short period
only. Since such experience will, in general, not include many large
claims, a further study oflarge claims over a longer period is needed.
Thus one year's statistics might suffice for claims say :( £100 000,
whilst for claims >£100000 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 are so changed in structure that they 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.

(b) Individual evaluation Sometimes satistics 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 (3.5.1). The largest risks of the portfolio are dealt
with individually and the net premium for each is determined;
next the expected average extent of damage is estimated for each
policy and the results are tabulated as in Table 3.5.2.
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,

Table 3.5.2 IndividuaL evaLuation.

Sum Risk Damage class in £1000

PoLicy insured premiums
no. (£1000) per thousand 100-500 500-1000 1000-2000

001 400 1 0.2

002 1000 2 0.4 0.3
003 800 1.5 0.3 0.2
etc.
3.5 ESTIMATION OF S 67

but in the absence of other methods it does provide some basis for
further calculation. For life insurance 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 (in the above example £100000), suitably selected samples
for the various risk sums may be taken and only the largest cases
treated individually.

(c) Shadow claims If some information is available relating to

large claims it is sometimes possible to introduce one or more
hypothetical shadow claims, which, having regard to the actual
portfolio, can be considered realistic although very seldom occur-
ring. The frequency ofthe shadow claim can be assumed, for example,
to be one claim in 10, 20, 30, or 40 years in the whole portfolio.

(d) Decomposition of the Z-range The approach presented in

item 3.5.2(c) can be followed. It is often effective to use a tabular
form for S(Z) under some limit Z ~ Zo as in Table 3.5.1 and an analy-
tic form for the tail Z > Zo. This method will be further discussed
in Section 3.7, but first the analytic method is treated in the following
sections.

3.5.4 ANALYTICAL METHODS

It is often desirable to try and find an explicit analytical representa-

tion for a claim curve. This is the case especially if the data base
is narrow for the use of the statistical method presented in Section
3.5.2 or if there is good reason to expect that the claim size dJ. is
of some particular form. This approach also has the advantage
that an analytic S function may be convenient to handle in many
calculations; if use can be made of some well-known elementary
functions for S, such as exponential, log-normal, Pareto, etc.,
the known properties of the function can be used to gain some
insight into the characteristic features of the claim distribution.
In some cases S may be of such a form that the convolutions Sk*
can be carried out in a closed form; then an explicit expression for
the dJ. F of the aggregate claim can sometimes be found, and thus
68 COMPOUND POISSON PROCESS

the approximations avoided. Furthermore, an analytical expression

for S can, of course, be of considerable value to the actuary in other
connections, e.g. tariff calculations, statistical analysis, etc.
On the other hand, it must be accepted that replacing the actual
data by an analytical expression always implies smoothing. The
goodness offit of S can be estimated by various well-known methods,
but it is often of much greater importance to study the error introduc-
ed in the function F. This is a drawback of the analytical method.
To ascertain the magnitude of the error caused by this phenomenon,
different functions S can be experimented with so that they approxi-
mate 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 may remain open. Fortunately,
experience shows that F is not very sensitive to changes in S 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 (3.2.3) and some for other reasons.

3.5.5 EXPONENTIAL DISTRIBUTION

(a) Definitions In general, claims distributions show the highest

frequency for the small claims, the frequency declining with increas-
ing claim size. Thus the exponential function may provide at least
a first approximation for claim size distribution
S(Z) = I - e- cz (for Z ~ 0). (3.5.3)
The constant e ( > 0) can be fixed for each application to obtain the
best possible fit. This expression has the advantage that an explicit
expression for the compound Poisson function F can be found by
direct calculation, thanks to the fact that the convolution (3.2.2)
can be obtained in closed form (see exercise 3.5.2) for X > 0

Sk*(X) = I - e-ex[ I + eX + ~(eX)2

2!
+ ... + I (eX)k-1J
(k-I)!
= S(k-\)* (X) - e -eX (eX)k - l/(k - I)! (3.5.4)

Note that according to (2.9.7) this formula could be written also in

the form Sk*(X) = r(eX, k). By programming the claim number
3.5 ESTIMATION OF S 69

probability Pk(n) and Sk* in the form of a recursion formula, F(X)

can be computed at least in the Poisson and Polya cases (see recur-
sion rule (2.5.2) or (2.9.13».

(b) Applicability The use of an exponential function as a model

for S 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. Experience has
proved that often the exponential S(Z) converges too fast for large Z
values.

(c) Exponential polynomials The area of the applicability of the

exponential distribution can be extended if S is constructed as a
sum of exponentials having different parameters c
r

S(Z) = L p (1- e-
i
CiZ ), (3.5.5)
i= 1

where L Pi = 1.

Exercise 3.5.2 Verify (3.5.4) and calculate J.lx and ax assummg

the exponential claim size dJ. (3.5.3). Compute F(2) for c = 1.
n = 1 and H(q) = s(q - 1).

3.5.6 GAMMA DISTRIBUTION

(a) The three-parameter gamma function A way to provide more

flexibility than that provided by the exponential dJ. is to use the
incomplete gamma function (see (2.9.1» in the form

1
r(aZ + b, a) = r(a)
faZ b
0
+
e-uu a - 1 du (Z ~ 0, aZ + b ~ 0),
(3.5.6)
as an estimate for the claim size dJ. S. There are three parameters
available for fitting the curve according to the actual dJ. which
can be determined so that the distribution will have the given mean
(/1), standard deviation (0-) and skewness (y). First it is useful to
standardize the variable Z
z = (Z - J.l)/a, (3.5.7)
70 COMPOUND POISSON PROCESS

to have mean 0 and standard deviation 1. This transformation does

not change the skewness, as may easily be verified. Hence z also
has skewness y. The coefficients a, band rx can then be determined
from the conditions that the function (3.5.6) should have the same
characteristics (exercise 3.5.3)
5(z) = S(Z) = f(rx + zjrx, rx) (z~ -jrx)
1 f~+zv'~
= f(rx) 0 e-uu~-I du, (3.5.8)

where
(3.5.9)

(b) For numerical evaluation of the gamma function the easily

programmable expansion

S(Z) =
w~ 1) [ 1 + ~1
r(
w w w
+ ~1·~2 + ... ] , (3.5.10)
eW rx+ rx+ rx+ rx+
is convenient, where
w=rx+zjrx. (3.5.11)
A good approximation for the complete gamma function is obtained
from the formula

where
bl = - 0.577 191 652 bs = - 0.756 704078
b2 = 0.988205891 b6 = 0.482199394
b3 = - 0.897 056 937 b7 = - 0.193 527818
b4 = 0.918206857 bs = 0.035 868 343.
This formula requires that the parameter rx is 1 ~ rx ~ 2. This can be
achieved by making use of the recursive formula
f(rx) = (rx -l)f(rx - 1). (3.5.13)
The formulae given above are useful if the skewness is not too
small. Troubles arise if this condition is not valid, because rx and w
grow to such an extent that the formulae are no longer easily work-
able and a special technique is needed. For example, the following
3.5 ESTIMATION OF S 71

Wilson-Hilferty formula (Johnson and Kotz, 1970; Section 17.5)

(3.5.14)
where
C1 = y/6 - 6/y; c 2 = 3 X (2/y)2/3; c 3 = 2/y,
is applicable particularly when r:J. is large or (which is the same) when
the skewness y is small, i.e. precisely in the area where the expansion
(3.5.10) becomes impractical. Another approach is to integrate
(3.5.6) numerically (see exercise 3.5.4).
Examples of gamma densities are plotted in Fig. 3.5.2. A handicap
of the gamma dJ. is that when it is used for approximation of rather
skew compound Poisson distributions it is not defined ( or S(z) = 0)
for the values
z < - 2/y. (3.5.15)
This may seriously worsen the fit for the short tail ofthe distribution.
It will be seen in Section 3.12 that the gamma dJ. is useful also
for approximation of the dJ. of the aggregate claims, i.e. instead
of S the 'target' function F itself.

Exercise 3.5.3 Prove (3.5.8) (Hint: note (2.9.17».

s' (z)
1.0

0.5

-3 -2 -1 o 2 3 4
z
Figure 3.5.2 Examples of gamma densities having mean = O,standard devia-
tion = 1 and varying skewness y.
72 COMPOUND POISSON PROCESS

Exercise 3.5.4 Assuming rt. to be large, integrate (3.5.6) by making

use of the Simpson formula. (Hint: remove the factor e - a rt. a - 1
from the integral.)

3.5.7 LOGARITHMIC-NORMAL DISTRIBUTION

(a) Definitions A frequently used claim size distribution is the

logarithmic-normal or briefly log-normal. It is derived by introducing
a variable Z > a ;;:: 0 so that
Y = In(Z - a), (3.5.16)
is normally distributed with parameters Jl and (J. Then the density
of the distribution is (see exercise 3.5.5)

S'(Z) = (J(Z _ ~)J(21!) ex p [ - 2~2 (In (Z - a) - Jl)2 ] (3.5.17)

The parameters a, Jl and (J are determined to fit the lowest moments

with those of the observed or assumed distribution (Cramer 1945,
p. 258). Note that Jl and (J are the mean and variance of Y, not
of Z. The parameter Jl may well be negative. For the solution,
the real root of the following equation is first determined
1]3 + 31] - 'Y = O. (3.5.18)
Then (see exercise 3.5.6)
1
a=a l -ryJm2
(J2 = In(l + 1]2) (3.5.19)
Jl = In(a l - a) - ~(J2,

where a l , m2 and 'Yare respectively the mean, variance and the

skewness of the claim size distribution to be approximated. It is
assumed that they are known, e.g. they are estimated from an actual
claims statistics.

(b) The shapes of the log-normal density curves can be seen in

Fig. 3.5.3. In order to demonstrate the scale selection, the same
functions are plotted both in linear and double logarithmic scales.

(c) Two-parameter version The derivation in item (a) relates to

the three-parameter form of the log-normal dJ. as compared with
 1.0
S' (z)

0.5

-3 -2 -1 o 2 3 4 5 6
z

S' (z)

2 3 5 7 10
z
Figure 3.5.3 A family of log-normal densities having the joint moments a 1 = 0
and mz = 1 but varying skewness y. The whole curve is plotted on a linear scale
(upper figure) and the tail on a double logarithmic scale (lower figure).
74 COMPOUND POISSON PROCESS

the more common two-parameter form in which the parameter a

is equal to zero. The difference between the two forms is solely a
shifting of the curve along the z-axis; this is often useful for represent-
ing data where the proportion of small claims recorded has been
reduced by the operation of policy conditions, such as the imposition
of small excesses or no-claim discount schemes. As compared with
the analytic formulae previously mentioned, the availability of three
parameters provides scope for an improved fit, at least in the early
and middle parts of the distribution. Further information can be
found in Benckert (1962).

Exercise 3.5.5 Prove that the density of the log-normal distribution

is given by (3.5.17).

** Exercise 3.5.6 (i) The m.gJ. of a normally distributed variable

N(J.1, ()2) is

Make use of this function and calculate the moments ak of the log-
normally distributed variable Z in the case a = O.
(ii) Show that if Z is log-normally distributed with parameters
a, J.1,)" then
at = E(Z) = el'etG2 +a
m2 = var(Z) = e2 1' eG2 (e G2 - 1)
)' =)'z = (e G1 + 2)J(e G2 - 1).

(iii) Prove that if S is a log-normal distribution with mean at,

variance m2 and skewness), ( > 0), then the parameters a, J.1 and ()
have the expressions as shown in (3.5.19).

3.5.8 THE PARETO DISTRIBUTION

(a) Definition Many of the actual distributions of claims that arise

in insurance applications can be reasonably well approximated by
the Pareto distribution

S(Z)
Z
= 1 - ( Zo
)-a (Zo~Z;rx>I). (3.5.20)

It should be noted that the moments a j of this dJ. only exist if

3.5 ESTIMATION OF S 75

2 3 5 7 10 20 30 50 70
Z= Z/Zo
Figure 3.5.4 A family of Pareto densities S'(Z) = rxZ~/Z«+ 1 (double logarith-
mic scale).

j< (X (see exercise 3.5.7). In fact the Pareto formula represents

'dangerous' distributions, where very large claims are possible.

Unfortunately, the mathematical properties of the Pareto dis-
tribution do not lead to simple expressions for convolutions, and
therefore numerical integration methods have to be used.
Figure 3.5.4 shows a family of Pareto densities.

(b) Large claims Experience has shown that the Pareto distribution
is often appropriate for representing the tail of distributions where
large claims may occur. As was demonstrated in Section 3.5.2, the
Pareto dJ. can be combined with other types of distributions;
that is, S(Z) can be piecewise composed of several functions, each
of them being valid in disjoint intervals of the Z-axis.

(c) Danger index The parameter (X can be used as an index of the

distribution, or at least of its tail. If (X < 2 it indicates a 'very dangerous'
distribution. Note that only moments of order < (X exist
76 COMPOUND POISSON PROCESS

(exercise 3.5.7). For example, the first and second moments exist
for (X > 2 only.
Seal (1980) has collected empirical (X values.

(d) Modifications If there is evidence that the occurrence of very

large claims is excluded, then the Pareto dJ. is often censored by
letting S(Z) be equal to 1 from some large enough value Z 1 upwards,
and (3.5.20) is replaced by
(3.5.21)
Another modification is truncation, i.e. moving the 'probability mass'
of the interval Z 1 ~ Z into Z 1
forZ~Z1
(3.5.22)
for Z < Z1·
This modification comes up, for instance, when the risk tops are cut
off by an excess of loss reinsurance treaty, as will be considered in
Section 3.6.2.

Exercise 3.5.7 Calculate the moment aj for the Pareto dJ.

3.5.9 THE TWO-PARAMETRIC PARETO AND THE QUASI-

LOG-NORMAL DISTRIBUTIONS

(a) Comparisons Figure 3.5.5 demonstrates the behaviour of the

exponential, log-normal and Pareto densities for large Z values.
The curves are fitted to go through a point P and to have the same
value of the derivatives at this point. The exponential curve con-
verges fastest and the Pareto slowest. In other words, when the
upper tail is considered, the Pareto distribution is the 'most cautious'.
The log-normal distribution underestimates the risk of excessive
claims as compared with the Pareto distribution and the exponential
curve gives practically vanishing probabilities for them. These
features, illustrated by Fig. 3.5.5, are valid in general.

(b) Experience has shown that the behaviour of the tail in practice
is often between that of the Pareto and log-normal types. Therefore
there is an obvious need to find distribution functions which have
greater flexibility. Two such distributions are presented in this
section and others will be considered in subsequent sections.
 5' (z)
p

10-4+-______________~--------~----~--~~--~~--~~~
1 2 3 5 7 10
z
Figure 3.5.5 Comparison of the exponential, log-normal and Pareto densities
(double logarithmic scale).

5' (Z)

0.8
1
1.1

2 3 5 7 10
(Z-Zo) 1Zo

Figure 3.5.6 The two-parameter Pareto densities, 0( = 1.8 and f3 varying. Note
that for f3 = 1 the one-parameter Pareto is obtained (double logarithmic scale).
 78 COMPOUND POISSON PROCESS

(c) The two-parameter Pareto d.C. can be defined as follows

S(Z) = 1- b[l + (Z ;:0 YTa (Z ~ Zo)' (3.5.23)

where IX and P are positive parameters, Zo is the limit for the tail
for which the formula is fitted, and b indicates the weight of the
probability mass, which is situated in the tail area Z ~ Zo' i.e.
b=l-S(Zo)·
Shapes of the distribution are shown in Fig. 3.5.6 for selected
parameter values. As can be seen, the desired flow between the
Pareto case (P = 1) and the log-normal type can be achieved by
varying the parameter P (> 1). Some actuaries, e.g. Gary Patric
(Prudential, New Jersey, unpublished letter) have reported successful
results concerning the fit of (3.5.23) to actual distributions.

(d) The quasi-log-normal d.f. is another approach to obtaining

curves that vary between the extreme cases presented in Fig. 3.5.5.

S'lZl

10-41 + - - - - - - - . . , . . - - - l - - . , . . 1 - - - . . . , . - - J , . . . . - - r - - , - - , - - - . - - - r....
1 2 3 5 7 10
ZIZo

Figure 3.5.7 Quasi-log-normal densities, ex = 1.8 and p varying (double

logarithmic scale).
3.5 ESTIMATION OF S 79

It is defined by formula
Z )-a-Pln(ZIZO)
S(Z) = 1 - b ( Z ' (3.5.24)
o
where the meaning of the parameters is in principle the same as
in (3.5.23).
Examples of the distributions are plotted in Fig. 3.5.7. For
{3 = 0 the Pareto case is obtained. Analysis of this dJ. can be found
in Shpilberg (1977). The name 'quasi-log-normal' reflects the
fact that the curves closely approximate the log-normal ones for
positive {3 values (Dumouchel and Olsten, 1974).

3.5.10 THE FAMILY OF BENKTANDER DISTRIBUTIONS

(a) The idea of finding more flexibility for curve fitting can be
extended. For this purpose Benktander (1970), following the earlier
work of Benktander and Segerdahl (1960), suggested a family of
distributions which contains as special members both the Pareto
and exponential and also approximately the log-normal distribu-
tions. By suitable adjustment of parameters a better fit with actual
data can be obtained; general experience of the type of the portfolio
in question and the crucial choice of the distribution type, as
mentioned above, are not so significant and may be replaced by
parameter estimation.
The analysis is again focused on the tail Z ~ Zo of claim size
distribution above some suitably chosen limit Zo. For values
Z < Zo some other expression or directly observed frequencies in
tabular form can be used.

(b) Extinction rate Suppose that the claim size dJ. S(Z) is known
or assumed for Z ~ Zo and that the necessary integrals and
derivatives exist; an auxiliary function m(Z) will be introduced as
follows
m(Z) = E{Z - Z/Z ~ Z}

=
1
1 - S(Z)
fooZ (V - Z)dS(V) (3.5.25)

=
1
l-S(Z)
foo (l-S(V»dV.
Z
 80 COMPOUND POISSON PROCESS

s' (Z)

Figure 3.5.8 The function m(Z).

The function m(Z) can be interpreted as the mean value ofthe claims
excesses over Z or as the distance of the centre of gravity of the
shaded area in Fig. 3.5.8 from the Z vertical. The latter form is
obtained from the former by partial integration.
If (3.5.25) is differentiated, a differential equation

5'(Z) = 1 + m'(Z) == r(Z) (3.5.26)

1 - S(Z) m(Z) ,

is obtained which determines the interdependence of the auxiliary

function m(Z) and the claim size distribution 5(Z).
The function r(Z) is a straightforward analogy to the rate of
mortality applied in life insurance mathematics. It gives the rate at
which the risk of large claims is decreasing when Z grows.
Benktander calls it the 'mortality of claims' and Shpilberg (1977)
describes it as a 'failure rate' and refers to the terms 'hazard rate',
'force of mortality' and 'intensity function' found in the literature
of reliability theory. This kind of concept has already been dealt
with by Witney (1909). In this connection it could be called the
'extinction rate'. Shpilberg has shown that this function has a
direct connection with the physical progress of fire in fire insurance.
Most fires are stopped at the very beginning and the amount of the
claim remains slight. However, if the early extinction fails, then the
chance of stopping the fire soon decreases, which in the case of
large risk units results in large claims and in long tails ofthe functions
r(Z) and 5'(Z).
If the function S is given, then the extinction rate r is determined
according to (3.5.26). The reverse relation also holds. If r(Z) is
 3.5 ESTIMATION OF S 81

given, the claim size d.f. is obtained by solving this differential

equation

I-S(Z)=[I-S(Zo)]exp [ - f:or(V)dV 1 (3.5.27)

which expresses S in terms ofr (or m).

(c) Examples It is easily verified that for the exponential dJ. (3.5.3)
m(Z) = lie, (3.5.28)
and for the Pareto dJ. (3.5.20)
m(Z) = Z/(IX - 1) (3.5.29)
Benktander and Segerdahl have investigated a number of actual
distributions containing large claims. Their results (see Benktander
and Segerdahl, 1960) have the same general behaviour as that
depicted in Fig. 3.5.9.

(Z)

200

150

100

50

50 100
Z
Figure 3.5.9 The function m calculated for the unsmoothed claims frequencies
given in Fig. 3.5.1. Unit is £1000000.
82 COMPOUND POISSON PROCESS

(d) The Benktander family Observations on m(Z), due to

Benktander and Segerdahl (1960) and Benktander (1970), suggested
an analytic formula for m(Z), either
Z
m(Z) = a + 2b In Z a> 0, b > 0 (Type I). (3.5.30a)

or
Zl-b
m(Z)=- o ~ b ~ 1 (Type II) (3.5.30b)
a
where
(3.5.31 )
and a and b are parameters which can be chosen within the limits
given so that the best possible fit with actual experience can be
achieved. As is seen immediately, the exponential and Pareto
cases are members of these function families. Type I gives a smaller
deviation from the Pareto straight line than type II.
Substituting the functions (3.5.30) into (3.5.27), the following
claim size distributions are obtained
(3.5.32a)
and

(3.5.32b)
The constant c is chosen so that continuous linking with the function
chosen for Z < Zo can be achieved.
Examples of the distribution of type I are given in Fig. 3.5.10.
It is appropriate to use Zo as a unit on the Z-axis.
Benktander (1970) has proved that the log-normal distribution,
which falls (depending, of course, on the parameter choice) between
the exponential and Pareto extreme cases, can be quite closely
approximated by the functions (3.5.32). As is seen from Fig. 3.5.10,
the Pareto distribution (b = 0) is the 'most dangerous' one in that
it gives the greatest probability of occurrence for very large claims.
Benktander has also derived this conclusion analytically.

Exercise 3.5.8 Prove (3.5.28) and (3.5.29).

Exercise 3.5.9 Prove (3.5.32a) and (3.5.32b).

 3.5 ESTIMATION OF S 83

104~____________~________~____- r_ _~r-__~~~~~~~
1 2 3 5 7 10
z
Figure 3.5.10 A bunch of Benktander type I densities; a = 1.8, c = 0.1 and
b = 0, 0.1, 0.2, 0.3, 0.4, 0.5 and 1; Zo = 1.

**3.5.11 OTHER TYPES OF DISTRIBUTION

(a) General aspects In addition to the distributions dealt with in

the previous sections, there are also various other forms of con-
tinuous functions used to describe claim amount distributions.
In general they have proved inconvenient both in calculations
and in analytical studies. However with the general availability of
microcomputers the calculations may no longer be a major consid-
eration. Furthermore the facility of recursive calculation raises the
question of using discrete distributions and is dealt with in Section 3.8.

(b) The Pearson system is one of the major families of distribution

which has proved applicable to a variety of problems. It has been
found that many of the distributions are, in practice, close to type VI
I.e.
(3.5.33)
and lie between type V and type III. This region also includes the
Pareto and log-normal distributions, although these are not
84 COMPOUND POISSON PROCESS

members of the Pearson system. For further details consult, for

example, Kendall and Stuart (1977), Section 6.2, or Johnson and
Kotz (1970), Section 12.4.1.

(c) The Weibull dJ. is also often a suitable alternative and it may
be written (see Johnson and Kotz, Vol. 2, 1970, Chapter 20) as
S(Z) = 1 - exp{ - [(Z - ZoVa]b}. (3.5.34)

(d) The inverse normal for which the density function is (see
Johnson and Kotz, 1970, Chapter 15)
(3.5.35)
is also sometimes suggested for claims size d.f. It is, in fact, a modified
Bessel function.

(e) Distributions related to the extreme values can still be referred

to, type I being (see Johnson & Kotz 1970, Chapter 21)
S(Z) = exp( - IX e- fJZ ). (3.5.36)

3.6 The dependence of the S function on reinsurance

3.6.1 GENERAL ASPECTS

(a) Total and retained claim amounts It should be noted that if

the problem concerns the net retained liability, the amount Z of
the claim 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
(3.6.1 )
In this formula the reinsurance arrangement is such that each claim
is separately divided between cedent and reinsurer, which is the
case in quota share, surplus or excess of loss reinsurance but not
in stop loss reinsurance, these types being treated in subsequent
sections.
If necessary the dJ. of the retained claim size Z is distinguished
from the global dJ. S of Ztol' denoting it, for example, by SM if the
reinsurance treaty is of such a type that it contains a retention limit
3.6 THE DEPENDENCE OF THE S FUNCTION 85

M, which is then a freely disposable parameter or variable of the

system concerned.

(b) Decomposition approach In practice different net retentions

are often used for different classes and for different types of risk.
A distribution function SM can then be constructed for each of these
groups, and a joint distribution function for the whole portfolio-
perhaps also having several steps corresponding to different reten-
tions - is obtained as will be shown in Section 3.7.

3.6.2 EXCESS OF LOSS REINSURANCE

According to the excess of loss treaty the reinsurer pays that part
of each claim Ztot which exceeds an agreed limit M, and hence the
cedent's share is Z = min (Ztot' M). Then the dJ. SM of the amount of
one claim so far as the cedent is concerned can be expressed in terms
of dJ. S of the total claim Ztot as follows
for Z <M
(3.6.2)
for Z~M.

f:
From (3.3.1) the moments of SM are given by

ah = ah(M) = Zh dS(Z) + Mh(l - S(M)). (3.6.3)

It is convenient in practice to calculate the lowest moments (3.6.3)

and the related indexes r2 and r3 (see (3.3.8)) in a tabular form,
because they are frequently needed for various applications. Table
3.6.1 gives an example.

3.6.3 QUOTA SHARE REINSURANCE

In quota share reinsurance any claim, irrespective of its size, is

divided between the cedent and reinsurer in one and same, prefixed
ratio, 'quota', r
Z = rZ tot (O<r<I). (3.6.4)
Hence
Sr(Z) = S(Z/r). (3.6.5)
Table 3.6.1 Moments of the claim size df on the cedent's net retention. Excess of loss treaty. The global S(Z) is the same as in
Table 3.5.1. The moment alM) is according to (3.6.3).

2 3 4 5 6 7 8 9
ZorM AS(Z) S(Z) a 1 (M) a 2 (M) a 3 (M) r2 (M) r 3 (M)

1 1.000E - 02 0.033953 0.033953 1.0ooE -02 l.oo0E -04 1.oo0E - 06 1.oooE + 00 l.oooE+OO
2 1.585E - 02 0.037664 0.071617 1.565E - 02 2.461E - 04 3.880E - 06 1.005E + 00 1.012E+00
3 2.512E - 02 0.045479 0.117096 2.426E - 02 5.986E - 04 1.490E - 05 1.017E + 00 1.044E +00
4 3.981E - 02 0.055413 0.172 509 3.723E -02 1.441E - 03 5.661E - 05 I.040E + 00 1.097E +00
5 6.3IOE - 02 0.063707 0.236216 5.650E - 02 3.424E - 03 2.123E - 04 1.073E + 00 1.177E+00

6 1.000E - 01 0.068234 0.304450 8.468E - 02 8.021E - 03 7.842E-04 1.118E + 00 1.291E+00

7 1.585E - 01 0.070466 0.374915 1.254E - 01 1.854E - 02 2.858E - 03 1.179E + 00 1.450E + 00
8 2.512E - 01 0.070370 0.445285 1.833E - 01 4.228E - 02 1.028E - 02 1.258E + 00 1.668E +00
9 3.981E - 01 0.071745 0.517030 2.648E - 01 9.519E - 02 3.648E - 02 1.357E + 00 1.965E +00
10 6.31OE - 01 0.074009 0.591039 3.773E - 01 2.109E - 01 1.273E - 01 1.482E + 00 2.371E +00

11 1.000E+ 00 0.075761 0.666800 5.282E- 01 4.571E- 01 4.336E - 01 1.638E + 00 2.942E+00

12 1.585E + 00 0.073025 0.739825 7.231E - 01 9.608E - 01 1.427E +00 1.838E + 00 3.774E+00
13 2.512E +00 0.064899 0.804724 9.643E - 01 1.949E + 00 4.515E + 00 2.096E+00 5.035E + 00
14 3.981E+00 0.052757 0.857481 1.251E + 00 3.812E + 00 1.374E + 01 2.435E + 00 7.016E + 00
15 6.31OE + 00 0.040152 0.897633 1.583E + 00 7.227E+00 4.055E + 01 2.884E+00 1.022E + 01

16 1.0ooE + 01 0.029698 0.927331 1.961E + 00 1.339E + 01 1.172E + 02 3.482E + 00 1.555E + 01

17 1.585E + 01 0.021660 0.948990 2.386E + 00 2.437E + 01 3.338E +02 4.282E+00 2.458E + 01
18 2.512E + 01 0.015765 0.964755 2.859E+ 00 4.375E + 01 9.392E + 02 5.353E + 00 4.020E + 01
19 3.981E + 01 0.011 310 0.976065 3.377E + 00 7.737E +01 2.604E+03 6.786E+00 6.766E + 01
20 6.310E + 01 0.008222 0.984287 3.934E+ 00 1.347E +02 7.106E + 03 8.706E + 00 1.167E + 02
21 1.000E + 02 0.005599 0.989886 4.514E + 00 2.293E + 02 1.887E + 04 1.125E + 01 2.052E + 02
22 l.585E + 02 0.003767 0.993653 5.105E + 00 3.822E + 02 4.902E + 04 1.466E + 01 3.684E + 02
23 2.512E + 02 0.002424 0.996077 5.694E + 00 6.232E + 02 1.243E + 05 1.923E + 01 6.737E + 02
24 3.981E + 02 0.001582 0.997659 6.270E + 00 9.975E + 02 3.097E + 05 2.537E + 01 1.256E + 03
25 6.31OE + 02 0.001022 0.998680 6.815E + 00 1.559E + 03 7.501E + 05 3.356E + 01 2.370E + 03

26 1.000E + 03 0.000600 0.999280 7.302E + 00 2.353E + 03 1.738E + 06 4.413E + 01 4.465E + 03

27 l.585E + 03 0.000330 0.999610 7.723E + 00 3.441E + 03 3.884E + 06 5.769E + 01 8.432E + 03
28 2.512E + 03 0.000179 0.999789 8.085E + 00 4.921E + 03 8.509E + 06 7.529E + 01 1.61OE + 04
29 3.981E + 03 0.000097 0.999886 8.394E + 00 6.933E + 03 1.847E + 07 9.838E + 01 3.123E + 04
30 6.31OE + 03 0.000052 0.999938 8.660E + 00 9.666E + 03 3.993E + 07 1.289E + 02 6.148E + 04

31 1.000E + 04 0.000028 0.999967 8.888E + 00 l.338E + 04 8.613E + 07 1.694E + 02 l.227E + 05

32 l.585E + 04 0.000015 0.999982 9.083E + 00 1.842E + 04 1.855E + 08 2.233E + 02 2.475E + 05
33 2.512E + 04 0.000008 0.999991 9.249E + 00 2.523E + 04 3.984E + 08 2.949E +02 5.035E + 05
34 3.981E + 04 0.000005 0.999995 9.391E + 00 3.444E + 04 8.546E + 08 3.906E + 02 1.032E + 06
35 6.31OE + 04 0.000002 0.999997 9.51OE + 00 4.672E + 04 1.819E + 09 5.166E + 02 2.115E + 06

36 1.000E + 05 0.000001 0.999999 9.609E + 00 6.287E + 04 3.827E + 09 6.809E + 02 4.314E + 06

37 1.585E + 05 0.000001 0.999999 9.689E + 00 8.359E + 04 7.914E + 09 8.904E + 02 8.700E + 06
38 2.512E + 05 0.000000 1.000000 9.750E + 00 1.085E + 05 l.570E + 10 1.141E + 03 1.693E + 07
39 3.981E + 05 0.000000 1.000000 9.794E + 00 l.369E + 05 2.978E + 10 1.427E + 03 3.170E + 07
40 6.31OE + 05 0.000000 1.000000 9.808E + 00 l.512E + 05 4.099E + 10 1.572E + 03 4.345E + 07

41 1.000E + 06 0.000000 1.000000 9.808E + 00 l.512E + 05 4.099E + 10 1.572E + 03 4.345E + 07

Monetary unit is £1000. Notations: xE ± n = x x 1O±. r 2 = a,!a~; r3 = a 3 /ai.

88 COMPOUND POISSON PROCESS

Exercise 3.6.1. An insurer having the distribution function S for the

total size of claims, has a combination of two reinsurance treaties
in force: (i) a quota share treaty, under which the reinsurer pays
a proportion r of each claim, and (ii) an excess of loss treaty covering
the retained business with a maximum net retention M. What is
the distribution function of the size of one claim for the insurer's
net retention?

3.6.4 SURPLUS REINSURANCE

(a) Definition The surplus treaty is applicable to the kind of

insurance class where the sum insured, denoted by Q, is defined for
each risk unit (policy or policies related to a property or other in-
sured object as in fire insurance for example). The original idea
behind this type of treaty was to divide not just the claims (as in
excess ofloss treaty) but the whole policy between the cedent and the
reinsurer. More precisely if a claim of size Ztot occurs and if the
insured sum of the corresponding risk unit is Q, then the cedent
retains the whole claim Ztot in the case where Q does not exceed
a fixed limit M. If this is not the case, then the retained amount
is proportional to the ratio M/Q. To summarize, the cedent's share
is
M
Z= QZtot (forQ > M)

(for Q ~M). (3.6.6)

Note that where the total loss Ztot = Q the outcome is the same as
in excess of loss treaty, i.e. Z = M or = Ztot if Q ~ M. In the case
of a partial loss, i.e. when Ztot < Q, the cedent's share is proportional
to the degree of loss Zto/Q. Hence for risk units having Q > M
the reinsurer participates in small claims as well, contrary to the
excess of loss arrangement.

(b) The d.f. SM(Z) Now the cedent's share Z is determined by

Ztot and Q which are both random. In other words, the cedent's
share Z of a claim depends on the distribution of the random
vector (Ztol' Q). An expression for S M is derived on this basis in
Section 3.6.5.
In practice this distribution can be more conveniently constructed
directly from claims statistics. To this end the sum Q must be known
3.6 THE DEPENDENCE OF THE S FUNCTION 89
in addition to the total claim size Ztot for each claim. A number,
N Z' relating to the claims which fulfil conditions (3.6.6) must then
be determined; thus all the claims for which Q:::; M and Ztot :::; Z
must be accounted for, followed by those for which Q > M and
Ztot :::; ZQ/M. If the total number of claims in the statistics in ques-
tion is N, then (see Fig. 3.6.1)
(3.6.7)
The claim size dJ. is a function of two variables M and Z. It can be
tabulated in a two-dimensional table with columns for selected
M values. Each column corresponds to column 4 in Table 3.5.1.
The moments for each M are obtained from the appropriate column

ah(M) = to Zh dS M(Z) ~ i tl Z~· (S M(i) - SM(i - 1)), (3.6.8)

where im is the highest value of the row index taken into the table
(see Table 3.5.1 where im = 41).

(c) Modifications The presentation in item (b) was simplified so

that a single retention limit M was assumed to be applied for the
whole portfolio or for that section of the portfolio under considera-
tion. Practice is more complicated, however, and modifications
to this rule are needed.
In fire insurance it is usual to evaluate a so-called estimated maxi-
mum loss (EML) for each risk unit. This sum QEML may be less than
the insured value Q of the property. This is the case if the risk unit
is large, or if it is of a type such that it is highly improbable that it
wiII be completely lost in one fire. Then QEML is the sum which is
estimated to be the loss in the worst case. The idea of rule (3.6.6) was
to adjust the net retention so that the retained claim Z will be within
the range 0 < Z :::; M. This can now be achieved by applying reten-
tion M' = MQ/QEML" Substituting in (3.6.6) a modified rule is
obtained (for QEML > M)
M' M
Z = -Q Ztot = --Ztot. (3.6.9)
QEML
Hence, the outcome is the same as if the original formula (3.6.6)
were applied replacing Q by QEML. This rule can be taken as guidance
for construction of SM. It ensures that the EML sums QEML are
recorded properly in the policy files.
90 COMPOUND POISSON PROCESS

Furthermore, especially in facultative treaties, the cedent can,

on the basis of a case by case adjustment, choose the retention M
separately for each risk unit. For example, the retentions may be
larger than the average level for risks having the best profitability
and vice versa (see the case of several maxima, Section 4.6). It is
difficult to give a rule for the construction of the S function which
would be applicable in all environments. It will depend on the
practice and working conditions of each insurer. An approach
could be to divide the risk units into subgroups according to the
applied retention M, and then first to derive the S function for each
subgroup and second to apply the composition technique which will
be presented in Section 3.7. The work need not be overwhelmingly
cumbersome if the files, which anyway are necessary for normal
services and reinsurance practice, are planned in a way that also
allows access to the data needed for the present purpose without
renewed manual policy-by-policy handling.

(d) Robustness related to the selection ofrules Some tests have pro-
ved that fortunately the risk-theoretical behaviour of the claims
process is fairly robust for variations in the claim size function S
in so far as the top risks are cut off by reinsurance. It seems even to
be possible to get an idea of the order of magnitude of the fluctua-
tions of the aggregate claims on the insurer's net retention simply
by applying the excess of loss technique presented in Section 3.6.2.
Then for M is to be taken the highest level of M applied in practice
(see Pentikiiinen (1982), Heiskanen (1982)). These results indicate
that the considerations are tolerant of fairly rough approximations
if these are necessary in some special environments.

*3.6.5 TECHNIQUE USING THE CONCEPT OF DEGREE

OF LOSS

(a) Two-dimensional S Another way of handling surplus rein-

surance (and also some problems in rate-making) is to recognize
that the claim share Z on the insurer's net retention is in fact depend-
dent on both the sum insured Q and the amount of the total size
Ztot. The idea is first to build a two-dimensional dJ. of the random
vector (Q, Ztot)
S(Ztot' Q) = prob {Ztot ~ Ztot;Q ~ Q}. (3.6.10)
3.6 THE DEPENDENCE OF THE S FUNCTION 91

Figure 3.6.1 Surplus reinsurance. The shaded area indicates that part of the
range of the two-dimensional total claim size distribution where Z ~ Z.

Then

SM(Z) = prob {Z :s::; Z} = fLz dS(Ztot' Q) = prob {(Ztot' Q)E A z }·

(3.6.11)
The double integral is to be taken over the shaded area A z of Fig.
3.6.1.

(b) Degree of loss denoted by r, is now introduced

r= Ztot/Q. (3.6.12)
It follows from the definition that 0 < r :s::; 1. The conditional dJ.
G of r as well as the (marginal) dJ. W of Q can be obtained from
S(Ztot' Q) (or rather they may be derived directly, e.g. from empiric
data if available) as follows
G(r/Q) = prob{r:S::; r/Q = Q} = prob {Ztot:S::; rQ/Q = Q}

(3.6.13)

and
W(Q) = prob{Q ~ Q} = S(Q, Q). (3.6.14)
92 COMPOUND POISSON PROCESS

Note that W(Q) is not the same d.f. as could be obtained by directly
recording the sums insured of the portfolio files of the policies in
force. The policy sums Q in (3.6.14) are weighted according to the
risk proneness, which causes the differences.

(c) The moments of SM can now be calculated

ab(M) = E(Zh)

= f: E(Zh IQ = Q) d W(Q)

= f: E(Zt~tIQ = Q)dW(Q)
+ f: E(ZhIQ = Q)dW(Q), (3.6.15)

and given that Ztot = rQ and for Q > M Z = rM

ah(M) = f: QhE(r"IQ = Q)dW(Q) + Mh f: E(rhlQ = Q)dW(Q)

= f: Qhbh(Q) d W(Q) + Mh f: bh(Q) d W(Q),

where

bh(Q) = f>h dG(rl Q), (3.6.16)

which are to be interpreted as conditional moments of the loss

degree r. Note the close formal analogy with the excess of loss
formula (3.6.3).
These expressions can be simplified if the distribution of the loss
degree is independent ofthe sum Q, i.e. G(r) is (at least approximately)
the same for small and large insured objects:
(3.6.17)
Then

ah(M)=bh[f: QhdW(Q)+Mh(l- W(M))] , (3.6.18)

(see Straub, 1978 and Venezian and Gaydos, 1980; the last considera-
tion follows Heiskanen, 1982).
3.6 THE DEPENDENCE OF THE S FUNCTION 93

(d) Some experimental data concerning the degree ofloss Unfortu-

nately condition (3.6.17), i.e. independence of the loss degree
of the sum Q, may not in general be well satisfied, as is seen in
Figs 3.6.2 and 3.6.3 which exhibit the moment hl and the density
of r for some Q values.
5 I
b, )(10 5 JI b, )( 105 b, xl0 :
25000-1 Non-industrialfire 25000-1 Industrial fire 25000-1 All fire
I I I
I I 1
I I I
I I 1
20000'1 20000-1 2000()-1
I I I
I I I
I I I
I I 1
15000-1 15000-1 15000-1
I I 1
I I I

--: :~:~:~
=l =i
10000-: 10000-1 10000-1

I
o _ILl..I..L.L.l..lL. ............l ........... L............ Q_IL ............LLL. ..........LLI- ...... LLLLL.L.. 0 _Il .......~"t-L .............. L......... L.LL. ..... L
10 4 105 10 6 101 Q 10 4 105 10 6 10 7 Q 10 4 105 10 6 10 7 Q

Figure 3.6.2 T he expected loss degree b 1 (Q) as a function of the sum Q.

Finnish industrial and non-industrial fire insurance by Heiskanen (1982).
Monetary unit FIM (~£O.1).

6' (r)

-2
10

-3
10

\ >5000
,
1

0.5 1.0
r
Figure 3.6.3 Density of loss degree for some insured sums Q (in units of 1000
FIM as plotted at curves). Industrial fire insurance in Finland 1973-78.
Compiled by Harri Lanka (unpublished materiaO.
94 COMPOUND POISSON PROCESS

The total losses (r = 1) are more frequent for small objects (Q

small) than for large ones. Hence application of (3.6.18) needs
critical examination of condition (3.6.17).
Another feature of interest is the rise of the curves when r in-
creases towards 1. It corresponds to real fire experience. Once a
fire has progressed beyond a certain point in a small risk unit it
can seldom be stopped, and the destruction is total. (This feature
has already been observed by Whitney (1909), who operated the
concept loss ratio and found analogies between claim size distri-
bution and life insurance formulae such as the force of mortality,
similar to that referred to in item 3.5.10(b).)

3.7 Decomposition of the portfolio into sections

(a) Decomposition Basic data and basic distributions are often
conveniently available for different sections of the portfolio, e.g.
for different classes and subclasses. For the present purpose it is
necessary to find techniques for deriving the distributions and
characteristics concerning the whole portfolio if the corresponding
distributions and characteristics are known for the separate sections.
This decomposition into sections is also advantageous for the
reason that the total business may be rather heterogeneous. It is
thus often easier first to construct the distributions for the sections
and then to combine the partial distributions.
Note that the sum process of the independent mixed compound
Poisson section processes is, in general, no longer a compound
Poisson process because its claim size variable depends on the
randomly varying relative values of the structure variables of the
sections. However, if the structure variable q is the same for all sections,
i.e. the variation of the Poisson parameters n., J
j referring to the
section, is fully synchronized, then the sum process is again of mixed
compound Poisson type as will be seen in item (f). The Poisson case,
where q degenerates to the constant 1 (see item 3.2(b)), is an example
(item (e)). However in t,he general case, where this kind of synchro-
nism does not exist, the main characteristics of the sum variable
can be expressed in terms of the section characteristics, which is
often sufficient for applications. This will be carried out such that
the composed formulae are formally similar to those derived for the
undivided portfolio. A benefit of this procedure is that many of the
3.7 DECOMPOSITION OF THE PORTFOLIO 95

further considerations, as presented in later parts of the book, are

the same irrespective of whether or not the portfolio is assumed to
be divided in sections.
Hence, assume that the portfolio is divided in sections j = 1,
2, ... ,J, each section being represented by a (mixed) compound
Poisson process. Let Xj denote the aggregate claim amount of section
j in the time interval of unit length (e.g. one year) considered.
Then the aggregate claim amount of the whole portfolio is the sum
of the section amounts

(3.7.1 )

Let
(3.7.2)

be again the expected number of claims, which is the sum of the

corresponding section numbers. Further, let
n.
v=-.-l. (3.7.3)
j n'

be a set of coefficients. They can be called 'distribution parameters'.

Then the notation
ai = LVPij, (3.7.4)
j

can be introduced where

(3.7.5)

Sj is the dJ. of claim size assumed for the sectionj. As in the foregoing,
fo~ convenience of notation the first moments a l j of the sections
WIll be denoted by mj and a l by m.
The quantity a i can be interpreted as 'a weighted moment about
zero' of the compound portfolio. This concept proves to be useful,
as will be seen in what follows.

(b) Mean value Now the expected amount of the aggregate claims
96 COMPOUND POISSON PROCESS

can be expressed (see (3.3.7)) as follows

Ilx = l>x. = I,njmj = nI,vjmj = nm = P, (3.7.6)
j J j j

where P = mn is (see (3.3.7)) the risk premium income.

(c) Variance In order to avoid covariance terms in the following

formulae it is now assumed that the section variables Xj in (3.7.1)
are mutually independent. The second and third central moments are
then simple sums of the section moments (see exercise 3.3.4)
IlJX) = E{ (X - E(X)n = I,IlJX) for i = 2, 3. (3.7.7)

In addition to the VjS, yet another set of distribution parameters is

needed
E(X.) m.n. P. m.
lr.=~=~=--1. =~V .. (3.7.8)
J E(X) mn P m J
Hence, lr j parameters are weights based on the section premium
volume whereas Vj parameters are based on the claim number
volume.
The variance can now be expressed as a sum of the section variances
by using the notation (3.7.4) as follows (see (3.3.7))
ai = I,ai J = I,(np2j + nJmJa~)
j j

= na 2 + n2m 2aq2 (3.7.9)

= (r 21n + a;)p2,
where (see (2.8.5)) the notation
(3.7.10)

and
r2 _ a 2 I m 2,
- (3.7.11)
are introduced. The former expression is the extension of the
structure variance concept and the latter of the risk index (3.3.8).
Note that q here does not refer to any real structure variable
q; a~ simply conveys the composite effect of the section variables qj
on the portfolio variance. Furthermore, note the complete formal
similarity of the composite expression (3.7.9) with (3.3.7).
3.7 DECOMPOSITION OF THE PORTFOLIO 97

(d) Skewness First, the third central moment of X can be derived

in a similar way as the second (see (3.3.7»
,u3(X) = I,u3(X)
j

= Inp3j+3InJmpzp~j+ In]m],u3(q) (3.7.12)

By convention (see (3.3.8»

(3.7.13)
and by using y 0"3 instead of,u3 as suggested in item 3.3(e) the follow-
ing expression is obtained for the skewness in terms of the given
notation

_,u3(X)_(!i
YX- 3 -
0" 2+
~"zazf~;
L-TC.
n +L-TC.y.O".
nm.) m .
,,3
.)
3)/(rZn
q) q)
+0".2)3 /2
q
x ))) (3.7.14)
This formula as well as the expression in parentheses in the final
formulation of (3.7.9) are of dimension zero in respect of the
monetary unit, which makes them 'immune' to the direct effect of
inflation. This simplifies matters when (as later) periods longer than
one year are studied.

(e) In the Poisson case the S function for the whole portfolio exists
and can be composed of the section functions Sj making use of the
m.gJ.s as follows (see (3.4.3». The m.gJ. of Section j is

Mis) = exp [ nj ( too eSz dS/Z) - 1) J (3.7.15)

and the m.gJ. of the whole portfolio is obtained by multiplication

(see item 1.6(b), property (iv»
M(s) = TI Mis)
j

= ex p [ ~>j( too e SZ
dS/Z) - I) ] (3.7.16)

= exp [ n too e
SZ d( ~ ~ Si Z )) - n 1
98 COMPOUND POISSON PROCESS

But this is again the m.g.f. of a Poisson dJ. for which (see (3.4.3),
(3.4.1) )
(3.7.17)

Hence, according to the unique correspondence of the m.gJ. and the

dJ., it can be concluded that the claims size distribution of the whole
portfolio can be obtained as the weighted sum (3.7.17) of the section
functions where the section distributions are of the simple Poisson
type.
Note that the weight parameters Vj in (3.7.17) vary randomly in
the general case where a time-dependent structure variation is
assumed and is not synchronized within the portfolio, e.g. if the
variation of the sections is mutually independent. Then there exists
no global claim size function that will be the same from year to year.

(f) Synchronized structure variation It follows from (3.4.2) that the

results of the previous item hold also in the special case of varying
basic probabilities when the time variation of the nj parameters is
the synchronized nj = njq, where q is the same for all sections. This
case can be handled by calculating first the moments (3.7.4) and
substituting them as well as the joint (Jq' Yq and y2 (q) into (3.3.7) (see
exercise 3.7.1).

(g) Further applications The result arrived at by means of m.g.f. in

the Poisson case is, however, in one sense more general. Whilst S
must, of course, be a distribution function of the size of one claim
in insurance applications, there is no reason to restrict consideration
to distribution functions in so far as component functions Sj are
concerned. On the contrary, in order to facilitate computations it
may sometimes be advisable to consider S 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 Fj'
formally defined as F in Section 3.2, are not necessarily distribution
functions, but this feature does not have any essential influence on
the calculations, i.e. the distribution function F of the portfolio
can simply be obtained from the functions F j by convolution

F=F 1 *F 2 *···*Fr (3.7.18)

For example, suppose there are difficulties in calculating a mixed
3.7 DECOMPOSITION OF THE PORTFOLIO 99

compound Poisson dJ. F directly from (3.2.1) but, from earlier

computations, the distribution functions F j of different parts of a
portfolio may be known, or it may be easy to compute them separate-
ly for each of these parts, or it is possible to select a set of J functions
Sj for which the corresponding F j can be calculated in practice, and
these functions satisfy (3.7.16). In these circumstances the distribu-
tion function of the whole portfolio can be obtained via these
components according to (3.7.18). Then, to get F, it is necessary only
to calculate J - 1 convolutions, an essentially easier course than the
calculation of a nearly unlimited number of convolutions which is
the case when (3.2.1) is used directly. For example, the expression
(3.7.16) could be taken as the first J terms of the expansion of Sin
a series of some auxiliary functions Sj' this series having convergence
properties so that the remainder may be neglected. The index j in
this connection does not, of course, refer to any actual section of the
portfolio.
As an example, let us suppose that the claim size distribution is
(or can be approximated to be; see Section 3.8) of discrete type, i.e.
only the claim sizes Z l' Z 2' ... ,ZJ are possible, and by probabilities
q l' q 2' ... , qJ and the claims number process is a simple Poisson one.
Then the m.gJ. (3.4.3) is reduced to the form

M(s) = exp [ n( ~ ~eSZj - 1) ]

= fl exp(nj e
szJ - n), (3.7.19)

where nj = nqj' But this is a product of simple Poisson m.gJ.s. That

means that the dJ. ofthe whole portfolio can be obtained by convolu-
tion of J simple Poisson variables, each having only one claim size
Zj' This result can be interpreted so that the portfolio is divided in
hypothetical sectionsj = 1,2, ... ,J. Only one claim size Zj is possible
in each section and the Poisson parameter n, the expected number of
claims, is nj' The sum of the section aggregate claims Xj has the same
dJ. as the original portfolio under consideration.

(h) Link to individual risk theory As briefly referred to in item

2.1(b), risk theory can be built also taking the individual risk units
(policies) as primary building blocks ('atoms') of the risk process.
This was, in fact, the issue in the early history of risk theory. The
100 COMPOUND POISSON PROCESS

foregoing important decomposition rules also lend themselves to

consideration of this approach. The 'sections' j is defined to be just
the risk units, and hence J is the number of policies in the portfolio.
Then X.J in (3.7.1) is the sum of the claims arisen from thejth unit.
Note that Xj may assume also the value 0 and that, of course, any
unit may produce more than one claim. If the distributions, or at
least some of characteristics related to the individual risk units, are
known then the aggregate claim process can be employed using the
methods derived earlier. An example is given in exercise 3.7.1.

Exercise 3.7.1 Consider a block of N = 10 000 risk units j, each

having a distribution of claims of mixed compound Poisson type
with the same expected number nj = 0.1 of claims. The standard
deviation of the structure variation qj of each risk unit j is a qj = 0.2.
Furthermore, the joint claim size dJ. of the units is approximated by
a discrete dJ. as given in terms of some suitable monetary units
(e.g. £1000) in the following table, where Si is the probability that the
size of a claim is Zi.

2 3 4 5
2 4 8 16
0.8 O.l 0.05 0.02 0.03

Calculate J.1 x and ax for the whole block in the following cases
(i) Assume first that structure variables qj = q are the same for all
risk units.
(ii) Assume that the structure variables qj are mutually
independent.

3.8 Recursion formula for F

(a) Derivation of the formula A case where the compound Poisson

dJ. F can be found by direct numerical calculation is that where the
claim size dJ. S is a discrete equidistant (often called 'lattice') distribu-
tion according to which only the values
(i=I,2,3, ... ). (3.8.1)
3.8 RECURSION FORMULA FOR F 101

can occur as the claim size. For brevity Zl will be taken as the
monetary unit: hence Zi = Z/Z 1 = i. Let the corresponding fre-
quencies be
qi = prob{z = i}. (3.8.2)
Of course, any distribution can be approximated by a dJ. of this
kind. In principle it is not necessary to limit the number of Zi values,
but unfortunately the numerical computations very soon become
laborious if the number of the points grows large. Hence the index
is generally limited to some rather small range 1, ... , S so that

for i > S. (3.8.3)

One or more of the probabilities for 1 ~ i < S may be zero.
It is convenient to operate with probability frequencies instead
of the (cumulative) dJ. Then the basic formula (3.2.3) of the com-
pound Poisson distribution is transformed as follows
00

f(x) = I Pkq~* for x = 0, 1, 2, ... , (3.8.4)

k=O

where the total amount of claims is denoted by a positive integer x,

and
qk*
x
+z +
= prob{z 1 2 ... +k
z = x}
' (3.8.5)

the variables Zj being mutually independent and equally distributed

according to (3.8.2). The recursion rule for the convolution is now
x
qxk* =
~
" q.q(k-l)*.
1 X-I
(3.8.6)
i= 1

By convention qg* = 1, q~* = 0 for x ~ l. An auxiliary equation is

needed as an intermediate stage for further development. It is obtain-
ed by analysing the expression
x
Ek = "iq.q(k-.1)*/qk*.
L..J 1 X - I X
(3.8.7)
i= 1

The quotient q8~k~}*/q~* is the conditional probability that Zk = i

for Zl' + ... + Zk = x; hence Ek is the expected value of Zk

Ek = E { Zk I J1 Zi = X}.
 102 COMPOUND POISSON PROCESS

Owing to the symmetry this conditional expected value has the same
value Ek for all Zi (i = 1, 2, ... ,k) and the sum of all these expected
values is x. Hence kEk = x or Ek = x/k. Substituting in (3.8.7), the
equation
k x
qhx = _ "i..J iq.q(k-.I)*
l X-l ,
(3.8.8)
Xi=l

follows. Note that by (3.8.6) this also holds when q~* = O.

It is still assumed that the claims number process belongs to the
family which obeys the recursive formula (see equation (2.9.13))

Pk = (a + b/k)Pk_l . (3.8.9)

This formula was introduced in item 2.9 (g) and it was stated that the
Poisson and negative binomial distributions belong to it. Then the
frequency f(x) for x > 0 can be manipulated into a form where it is
expressed by thefvalues calculated for x-I, x - 2, ....
To obtain this expression it must first be noted that according to
(3.8.9) 00

f(x) = L (a+b/k)pk_lq~*
k=l

_ 00 k* 00 b k*
-Lapk-1qx +L,?k-lqX'
k=l k=l

The convolution can be lowered one step making use of (3.8.6) and
(3.8.8) and by denoting m = min (x, s)

The inner sums are equal to f(x - i) as seen from (3.8.4). So the
recursion formula is obtained
min(x,s)
f(x) = L (a + ib/x)qJ(x - i). (3.8.10)
i= 1

Starting from the value

f(O) = Po, (3.8.11)
f(x) can be calculated step by step for x = 1,2, ....
3.8 RECURSION FORMULA FOR F 103

If the range s of the S distribution is not very long and if n is not

very large, the recursion formula is quite convenient for computa-
tions of f(x) and, of course, the d.f. is immediately obtained by
summation
[xl
F(X) = F(xZ 1 ) = L f(v), (3.8.12)
v=o
where X is the aggregate claim as x multiple of the unit Z 1 .
The remarks given in item 2.5(a) concerning programming are
applicable also for this formula.

(b) References Panjer (1981), Bertram (1981) and Jewell and Sundt
(1981) have proved the recursion rule to be valid for more general
assumptions than above. Jewell and Sundt also present an exhaustive
study of the family satisfying (3.8.9) and extend the consideration to
some more general distributions. The recursion formula (3.8.12) for
the Poisson case was presented by Adelson (1966).

(c) Example The recursion formula (3.8.10) is exact. However, ifas

a first step of the calculation the original claim size dJ., which may
be of continuous type, is first discretized, an inaccuracy will arise due
to rounding off and for other reasons. To test the sensitivity of the
results to this discretization procedure a truncated Pareto dJ.
(Z < 1)
(1 ~ Z < 21) (3.8.13)
(Z ~ 21),
was replaced by a discrete equidistant d.f., the probability mass
being concentrated in r + 1 points
Z = 1, 1 + A, 1 + 2A, ... , 1 + rA = 21
where A = 20jr. In order to test the effect of the length of the interval
the alternative values 20, 5, 2 and 1 were assumed for r. The probabili-
ties assigned to these points were
ql +iA = S(l + (i + t)A) - S(l + (i - t)A) i = 1, 2, '" , r - 1,
and at the end points of the interval [1, 21 ]
ql = S(l + tAl; ql +rA = q21 = 1 - S(1 + (r - t)A).
The examples in Table 3.8.1 demonstrate the sensitivity of the result
104 COMPOUND POISSON PROCESS

Table 3.8.1 Examples of F(x) for x < 0 and 1 - F for x> 0 per thousand.
Pareto claim size d.f. (3.8.13), Poisson claim number dj, n = 50, x = (X - m,J/ux
the normed variable.

Number of intervals r
x 20 10 5 2

3 5.2 5.2 5.8 7.9 0.2

2 34.7 35.1 35.1 38.1 129.5
1 151.5 153.4 155.2 156.5 155.1
-1 152.8 158.9 156.3 160.5 47.5
-2 9.0 8.7 7.4 1.8 10.5

to the density of the interval number r. Obviously the method

tolerates a fairly coarse interval net.
Gerber (1982) has suggested methods according to which the
inaccuracy due to the discretization rounding can be evaluated.

Exercise 3.8.1 Calculate Polya F(X) for the parameter combination

n = 2, h = 10. S(Z) is a two-point discrete function Zl = 1, Z2 = 2,
S(1) = 0.2 and X = 0,1,2,3, ... ,6.

3.9 The normal approximation

(a) Need for approximation methods It will be apparent from the

foregoing chapters that the compound Poisson function F which
gives the distribution of the aggregate claims is, unfortunately,
complicated as regards computation particularly in practical
applications. Direct methods of attack on the numerical treatment of
F 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. The
recursion formula presented in the previous section may be useful
for applications only in small collectives and in problems where X
is given and F is requested. Furthermore, it is extremely difficult to
obtain a broad survey of the problems. Even if the nature of the
problems justifies more detailed computations, simple working
approximations are necessary; it follows that one of the problems of
applied risk theory is the finding of proper approximations.
3.9 THE NORMAL APPROXIMATION 105

(b) The central limit theorem of the probability calculus is a

classical approximation much used in risk theory as well. For this
purpose, first the variable X will be 'standardized' or 'normalized',
transferring it in the form
(3.9.1 )

where (see (3.3.9)), limiting the consideration to the Poisson case,

(3.9.2)

The normed variable x has mean 0 and standard deviation 1. The

central limit theorem says that the dJ. F ofx tends to the normal dJ.

1 IX (3.9.3)
N(x) = J(2n) _ 00 e- tu2 du,

when n tends to infinity, i.e.

F(X) = F(x) ~ N(x). (3.9.4)

(c) Discussion of the area of applicability In Chapter 4 and later it

will be shown that the normal approximation essentially 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 considera-
ble difficulties. Unfortunately, however, the accuracy ofthis approxi-
mation is not satisfactory if the skewness of the distribution is large,
as is demonstrated in Fig. 3.9.1, where normally distributed values
are compared with the so-called N P values calculated by a method
which will be presented in the next item. Provided that the accuracy
of the NP values is satisfactory, the deviations from them indicate
the degree of inaccuracy of the normal approximation.
Figure 3.9.1 suggests that if the skewness y is small- as it is for
portfolios which are protected by normal reinsurance and which are
not very small- the normal approximation gives quite a good fit.
However, if y exceeds 0.1 the deviations grow rapidly, especially at
the tails of the distribution. The normal dJ. then generally under-
estimates the risk of excessive claims and should not be used if safe
evaluations are requested.
 106 COMPOUND POISSON PROCESS

'"1 2.00 i
I I
I I
I
I I :
I
I I I I
I I I
1.00 -------1--
I
--r-------.-------,
I l
I I
I
iI
•

0.50 I I
I
I
0.30

0.10 -I
:
••
I
I
I
0.05
,,I
,,
I

0.03 ,
,
I
I

,••
I

,••
,•
~01+---_+----+_--_+----+_--_+----~--~--~
-3 -2 -1 0 3 4 5
x
Figure 3.9.1 Comparison of the normal df (N) and the NP df (N) as a
function of x and y. The relative deviation 100 (N - N)/Ny for x < 0 and
100[(1- N) - (1 - N y)]/(l - Ny) for x> 0 is computed and then the value
pairs x, yare sought and plotted for which these deviations are equal to - 75,
- 50, ... , 50, 100 so constituting 'a map' to give the altitudes of the deviations.
For example, for x = 2 and skewness y = 0.1 one can read that the relative
deviation is about - 11 %, i.e. the normal approximation gives a value for 1 - F
which is 11 %less than the one obtained by the N P method. The discontinuities at
x = 0 are due to the presentation ofF and 1 - F at either side of this line, due to
the fact that F(O) =1= 1 - F(O) for y > O.
(d) Numerical values of the normal d.f. can be obtained from
standard textbooks or can be programmed making use of the
following expansion (Abramowitz and Stegun, 1970).
First calculate

1
R =J(2n)e -t x2 (b b 2
1t+ 2t + b3t 3 + b4t4 + b 5t 5), (3.9.5a)
3.10 EDGEWORTH SERIES 107

where
t = 1/(1 + 0.2316419Ixl),
and the values of bi' i = 1,2, ... ,5 are respectively
0.319381530, - 0.356 563782, 1.781477937,
- 1.821 255978, 1.330274429.
Then
for x ~ 0
(3.9.5b)
for x > O.
When the value N = N(x) of the function is given and the matching
argument value x is requested, first calculate
J( - 21nN)
{J( for 0 < N~0.5
t= - 21n(1 - N)) for 0.5 < N < 1,
(3.9.6a)

and
(3.9.6b)

where
Co = 2.515 517, c 1 = 0.802 853, c2 = 0.010 328
d 1 = 1.432788, d2 = 0.189 269, d3 = 0.001308.
Then
-y for 0 < N ~ 0.5
x= { (3.9.6c)
y for 0.5 < N < 1.
The absolute amount of the error is estimated to be < 7.5 x 10- 8
for (3.9.5b) and < 4.5 x 10- 4 for (3.9.6c).

3.10 Edgeworth series

The normal approximation (3.9.4) is in fact only a special case of a
more general formula, known as an Edgeworth expansion
F(X) ~ G(x) = N(x) - pN(3)(X) + 2~Y2N(4)(X) + /2 y2N(6)(X) + R(x),
(3.10.1)
where N is again the normal dJ. (3.9.3), x is the normed variable
(3.9.1) and y and Y2 are the skewness and kurtosis (3.3.7) of X. The
108 COMPOUND POISSON PROCESS

remainder term contains lin at least in power 3/2 in the Poisson case.
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 deriva-
tion of the formula are given in Appendix B.
Reference to (3.10.1) 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). If the explicit
expressions of higher derivatives of the normal function N are
introduced into (3.l0.l), it can be shown that the error of the
Edgeworth expansion tends to infinity as the number of terms
increases without limit. The Edgeworth expansion is not a conver-
gent 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 oftwice 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, and this is given in the following
sections.

3.11 Normal power approximation

(a) Background The normal power approximation was originally
found by Kauppi and Ojantakanen (1969). They computed, in a
parallel manner, normal approximated values and actual values of
the compound Poisson F in cases where these are computable
exactly or the accuracy is within well-controlled limits. It was found
that, in general, the differences obeyed a certain simple pattern.
This observation suggested a correction of the standardized
(see (3.9.l)) argument x in a way which transformed it into another,
approximately normally distributed, variable y. Also a theoretical
derivation was soon found for the formula and eventually it proved
to be a special case of a similar transformation originally due to
Cornish and Fisher (1937). A good account of the idea can be found
in Kendall and Stuart (1977, paragraphs 6.25-27).
Generally speaking the situation in the present case, as also in
3.11 NORMAL POWER APPROXIMA nON 109

many other occasions in mathematical statistics, is that a variable x

tends towards normality when the population to which it is related
becomes large. In our application, this means that the volume
variable, the expected number of claims n, should be large, making
the skewness small; according to item 3.9(c), this is an indication
for the acceptability of the normal approximation for the standardiz-
ed compound Poisson variable x. However, if the background
collective is small or a skew structure function is assumed, significant
deviation from normality appears. The idea is to find a
transformation
x = v(y), (3.11.1 )
which converts a normally N(O, 1) distributed variable y into another
variable x, which can be better fitted to actual distributions. It is
convenient to try to find a function v which includes some free
parameters. The latter will be chosen, for example by equating lowest
moments, to give a maximal fit with the compound dJ. employed.
Then we have F(X)=F(x)~N(v-1(X)) where F is the dJ. of the
standardized compound Poisson variable (3.9.1) and v- 1 is the
inverse of the function v.

(b) Derivation of the NP formula It can be proved that a suitable

transformation v in a polynomial form can be obtained by inverting
the Edgeworth expansion G (see (3.10.1)).
First denote y + ily = x. Then the required relationship (i.e.
x = v(y) = y + ily) between x and y can be obtained from the
equation
N(y) = G(y + ily) ~ F(x), (3.11.2)
which determines y so that N(y) gives (approximately) the same value
as the Edgeworth expansion. This equation can be written in the
compound Poisson case (structure variation omitting)
N(y) - N(y + ily) + iyN(3)(y + ily) - 214 y2N(4)(y + Lly)
- 712 y2 N(6)(y + Lly) + O(n - 3/2) = O. (3.11.3)
It is now solved by means of Newton's method, according to which
the solution of an equation f(u) = 0 can be found if some
approximate solution it is known as the expansion
- f(ii) 1 f"(ii) [f(ii)
u=u------- -- J2
!'(ii) 2 !'(ii) !'(ii)
110 COMPOUND POISSON PROCESS

Now f is the l.h.s. of (3.11.3), ii = Y and u = y + L\y. Further, the

derivatives of the normal dJ. N are needed in form
N"(y) = - yN'(y); N(3)(y) = (y2 - 1)N'(y); N(4)(y) = (l- 3y)N'(y); ...
(3.11.4)
which are readily obtained by differentiating (3.9.3). Similarly the
higher derivatives of N(y) can be expressed as polynomials in y
multiplied by the first derivative N'(y) (so-called Chebyshev-
Hermite polynomials, see Kendall and Stuart, 1977, par. 6.14).
Substituting in (3.11.3) and after some straightforward calculations
the method gives
L\ _ b(y2 - 1) + -h-y/l- 3y) + y2(/- h" lOl + l5y)
y- l+b(l-3y)+O(n- 1)

1Y+O(n- 1/2 )[b(l-1)+O(n- 1)]2 -3/2

+ 21 + O(n 1/2) 1 + O(n 1/2) + O(n )
= b(y2 - 1) + 2~y 2(l - 3y) - 316 y2(2y3 - 5y) + O(n - 3/ 2),

from which the sought transformation v(y) is obtained

x = v(y) = y + L\y = y + iy(y2 - 1) + 2~y/y3 - 3y)

- 316 y2 (2y3 - 5y) + O(n- 3/2 ). (3.11.5)
If only the first correcting term is taken into account, then the
following approach to find an approximated value for the compound
Poisson function F(X) is obtained. First solve y from
(3.11.6)
Then
F(X) = F(x) ~ N(y), (3.11.7)
or, substituting y into N(y)

F(X) ~ N[ - ~
Yx
+ J( ~ + +
Yx
1 6X -
y ax
Ilx) J. (3.11.8)
If the next two terms in (3.11.5) are taken into account and Yx and
Y2(X) are again denoted briefly by y and Y2' the extended version is
obtained
(3.11.9)
3.11 NORMAL POWER APPROXIMA nON III

The root y of this equation can be found by direct solution or other

methods and substituted into (3.11. 7).
These formulae are called N P approximations (normal power
approximations). They are properly applicable only for x > 1. For
the remaining range some modifications are given later, but first
some special features and the question of accuracy will be discussed.

(c) Discussion The foregoing results can be interpreted in two

formally different ways. It can be said that the compound Poisson
variable is developed as a semi-convergent series (3.11.5) and a
suitable number of terms will be accepted to give numerical approxi-
mative values for the relationship between x and y.
Another interpretation is obtained when the number of accepted
terms is fixed. The relationships x ---+ y and y ---+ x are then uniquely
defined for all real values of these variables as will be seen in the
sequel, and an explicit transformation v(y) is found as required in
item (a). Ultimately a new dJ., N(v - l(X)), is constituted. If, as will be
the case in what follows, the short version (3.11.6) is assumed, then
v(y) includes one parameter, only, the skewness y, and can be suitably
denoted by vy(y). Furthermore, it is convenient to introduce the
notations
(3.11.10)

The mean, the standard deviation and the skewness of this distribu-
tion are approximately 0, 1 and y (= free parameter) respectively if
the short version (3.11.6) is used. It is convenient to say that x is
NP-distributed, denoted NP(O, 1, y). The variable X = Ilx + xax
(see (3.9.1)) is then also NP distributed having mx ' ax' y = Yx
approximately as mean, standard deviation and skewness. Note that
a linear transformation does not change the skewness, i.e. x and X
have the same skewness. In brief it can be said that X is NP(llx' ax' yJ
distributed having the dJ. N y [ (X - mx)/axJ.
As the compound Poisson distribution was approximated by the
normal dJ. N(llx, ai) in Section 3.9, so it will now be approximated
by the NP(llx' ax, Yx) distribution. The crucial benefit of this
approach is, of course, that now three parameters are available
instead of only two. In fact, the normal dJ. is extended to a family of
functions having an extra parameter y available. For y = 0 the NP
function reduces to the N function.
Since the proof is based on the assumption that F can be represented
112 COMPOUND POISSON PROCESS

by a finite number of the leading terms of the Edgeworth expan-

sion, the result would not be expected to be better than that obtained
by the direct use of (3.10.1). This argument however proves, quite
surprisingly, to be wrong. Experience of the application of (3.11.8)
has shown, as will be presented in subsequent items, that if y is
not great it gives fairly good results, whereas the Edgeworth ex-
pansion is usually unsatisfactory for deviations some distance from
the mean. 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. The Edgeworth expansion is not a convergent but
a divergent series and its fitness depends, among other things, on
how many of its terms are accepted for the approximation. The
same argument holds true for the inverted expansion, and it is no
miracle if the leading terms of one expansion give more accurate
approximations than those of another expansion. This is in fact
a well-known feature for semi-convergent series.

(d) NP function structure It is worth noting that the NP function

(3.11.8) includes as parameters only the mean value fl x ' standard
deviation o"x and skewness Yx of the distribution of the aggregate
claim X. The background distributions of the claim numbers
and claim sizes are not directly needed; they have an effect only via
the moments which determine the above characteristics (see (3.3.9)
or (3.3.7)). Hence for computations only these characteristics are
needed. This is an important observation, because often just the
moments can be directly derived from the empirical data (see
(3.5.2), (3.6.3), (3.6.8) and Section 3.7 and an analytic formulation
of the claim number and claim size distributions is not needed.
On the contrary, an attempt to force these distributions into the
form of some standard analytic function may give rise to rounding-
off errors which are difficult to control and which are, in fact, rather
unnecessary.

(e) Extension of the area of applicability The above formulae were

originally derived only for the simple compound Poisson distribution
for which they are, as the normal approximation too, asymptotically
correct. They can be formally applied to any distribution by substitut-
ing the respective characteristics of the distribution into (3.11.8). For
example, the mixed compound Poisson dJ. can be approximated by
 3.11 NORMAL POWER APPROXIMATION 113

calculating the characteristics using (3.3.7). The tests referred to in

the following item prove that accuracy is still about as good as in the
Poisson case, in so far as the criteria presented in the following are
satisfied. This is understandable since Ny(X/nm) tends to NP(1, O"q' Yq)
as n --+ 00. In other words, it approximates the structure function as a
limit distribution. But this is precisely the function towards which the
mixed compound Poisson dJ. F tends according to equation (3.3.16).

(f) Accuracy The usefulness of the NP method has been tested by

comparing it with other methods, which give exact or at least
controlled values for the compound function F; the works of Kauppi
and Ojantakanen (1969), Pesonen (1967a, 1969), Berger (1972) and
Pentikainen (1977) may be referred to. Typical results are given in
Fig. 3.11.1 and in Table 3.11.1. The accuracy becomes critical when

F 1-F

-1
10

Normal
~\X\
Edgeworth \ \\
NP

\ \

\ \
-2 -1 o 2 3 4 5
x
Figure 3.11.1 Example ofthe Normal, Edgeworth and N P approximated values.
Polya df F with h = 100, n = 25, claim size df Pareto with IX = 2 truncated and
discretized in points Z = 1,2, ... , 21 (see Table 3.11.2). The slight irregular
bends are due to the calculation of this strictly discrete function at equidistant
points which did not coincide with the steps of F, and with irregular rounding off
small errors resulted.
114 COMPOUND POISSON PROCESS

Table 3.11.1 Compound Poisson values computed by the recursive formula

(3.8.12) and NP formula. 1runcated Pareto claim size df. IX = 3, 1 ~ z ~ 11
discretized in II points, y = 0.18, n = 100.

F N y -F

ro
x Ny (Ny - F)/F%

0.0004 0.0004 -3.16 -0.0000

-2.5 0.0039 0.0034 -13.58 -0.0005
F -2.0 0.0184 0.0174 -5.12 -0.0009
-1.5 0.0617 0.0616 -0.19 -0.0001
-1.0 0.1560 0.1594 2.12 0.0033
-0.5 0.3092 0.3178 2.78 0.0086

0.0 0.4762 0.4877 2.41 0.0115

0.5 0.2946 0.3004 1.97 0.0058
1.0 0.1571 0.1587 0.99 0.0016
1.5 0.0721 0.0715 -0.86 -0.0006
I-F 2.0 0.0286 0.0276 -3.55 -0.0010
2.5 0.0099 0.0092 -7.07 -0.0007
3.0 0.0025 0.0026 3.86 0.0001
3.5 0.0007 0.0007 -0.52 -0.0000
4.0 0.0002 0.0001 -5.81 -0.0000

for positive x-values the skewness is increasing up to, say 1.5 or 2, first
at the periphery of the x values and then over the whole range. For
negative values of the argument x the deterioration of the relative
deviations for larger x values begins to appear from skewness values
rather less than 1, even if the absolute differences are still quite small.
Table 3.11.2 is intended to illustrate the critical area. For this
purpose, rather heterogeneous distributions were taken as examples
and in addition the parameter n, the expected number of claims as
an indicator of the size of the portfolio, is very small. Hence another
critical condition is that n should not be very small ( < 25).
Fig. 3.11.1 illustrates typical behaviour of the different formulae.
The normal dJ. is symmetric and therefore incapable of approximat-
ing skewed distributions. The Edgeworth expansion is clearly better,
but nowhere near so effective as the NP formula, which gives, even
for as small n as 25, a quite good approximation over the whole
relevant range.
In practice the size of the portfolio in insurance companies usually
makes the skewness parameter small because the volume parameter
n is in the denominator of its expression (see (3.3.7)). It is mostly of
order of magnitude of 0.1-0.4, and often less. The examples given
3.11 NORMAL POWER APPROXIMATION 115

Table 3.11.2 Polya distributed values, hand n as given in the table. Truncated
Pareto claim size d! (ex = 2, 1 ~ Z ~ 21, discretized in 21 points. In the upper
block Ffor x < 0 or 1 - F for x ~ 0, and in the lower block the relative deviations
ofN P values from the corresponding F or 1 - F values.

h OCJ 100 100 10 10 100 5 100

n 100 100 25 25 10 10 5 5
Y 0.3281 0.3316 0.6453 0.6749 0.7743 1.0280 1.1426 1.4599

-2 F { 0.0131 0.0130 0.0063 0.0031 0.0019 0.0004 0.0456 0.0277

r
-1 0.1605 0.1605 0.1593 0.1518 0.1558 0.1445 0.1644 0.1195

0 0.4789 0.4465 0.4558 0.4479 0.4497 0.4175 0.4076

x I 1- F 0.1534 0.1547 0.1500 0.1550 0.1554 0.1449 0.1530 0.1320
2 793
0.0308 0.0309 0.0361 0.0373 0.0395 0.0468 0.0441 0.0450
3 0.0040 0.0040 0.0067 0.0071 0.0080 0.0112 0.0115 0.0160
4 0.0003 0.0003 0.0010 0.0011 0.0013 0.0024 0.0026 0.0046

[(Ny-F}IF] x 100

-2 1.2 1.3 -9.4 63.8 92.5 167.2 -98.6 -99.7

-I 0.2 0.2 4.9 10.7 9.8 25.2 13.4 71.9
0 -0.2 -0.2 2.4 -0.1 0.2 -3.9 1.7 -0.9
x 3.4 2.6 5.8 2.3 2.1 9.5 3.7 20.2
2 1.2 0.8 6.3 4.5 3.9 -1.7 9.1 19.0
3 -4.2 -3.0 3.6 3.2 5.3 0.9 10.1 2.2
4 -3.4 1.2 -1.4 4.2 6.7 3.6 13.6 0.8

The slight irregularities in relative deviations are partially due to the same rounding
ofT effects as mentioned at Fig. 3.11.1.

here as well as the experience referred to earlier, or otherwise gained

in connection with applications, suggest that the N P approximation
can be safely used in such circumstances. Obviously Table 3.11.1 and
Table 3.11.2 for n ~ 25 and h ~ 100 may give a typical picture of the
accuracy, whereas Table 3.11.2 for small nand h illustrates mainly
borderline conditions.
Note that the different approaches for calculation of the com-
pound Poisson dJ. complement each other. The normal approxima-
tion is preferable for large and only slightly skewed distributions as
Fig. 3.9.1 suggests. The N P formula is applicable also for medium-
size collectives having a considerable skewness. For quite small
collectives, where the N P approximation also often fails, exact
methods like the recursive calculation dealt with in Section 3.8 are
appropriate, as is the straightforward simulation which will be
116 COMPOUND POISSON PROCESS

presented in Section 6.8.2. On the other hand, these accurate methods

are not suitable for numerical calculations if the collective, i.e. n, is
large and the claim size d.f. is not forced into some simplified form.

(g) The long version (3.11.9) often improves the approximation

(Pesonen, 1969; Pentikiiinen, 1977). However, as y increases, i.e. at
the tails of the distributions, the suitability ofthis version may worsen
in an irregular way and its utility is not well mapped as yet. Mainly
for this reason, the short version only will be used in the following.

(h) Extension to negative x values The compound Poisson func-

tion F is a basic building block in the more complicated types of
problems dealt with in the following chapters. Often it is needed for
the whole range of the variable X, i.e. also for negative values of the
standardized variable x in (3.9.1). In some problems the value of the
function F is given and the matching variable value X is to be
calculated. For integrations and simulations fast formulae are
needed for F, its inverse F- 1 and derivative F'. The original form of
the N P approximation is not suitable for all these purposes and it
will now be manipulated accordingly.
For brevity let
9 = y/6 and h = Y2/24.

To derive a workable formula for the 'short tail' (x < 1) in the

extended version (3.11.9), y is expressed as a power series of 9 and h

y = aoo + alOg + a20 g 2 + ao1h + ... ,

where the coefficients aij are unknown and have to be found. Only
terms including second and lower powers of ljjn are taken into the
expansion. Substituting in (3.11.9) and equating to zero the co-
efficients of each power of 9 and h, the expansion
y=x- 9 X (x 2 - 1) + g2 X (4x 3 - 7x) - h X (x 3 - 3x) + ...
(3.11.11)
is obtained after some algebraic reductions.
In the middle of the distribution the first two terms are sufficient,
but for the tail the third term is still needed. To avoid discontinuity
it is appropriate to incorporate the third term at the point
Xo = -j(7/4) (3.11.12)
3.11 NORMAL POWER APPROXIMATION 117

as it is zero at this point. Hence

y=x- g X (x 2 - 1) + g2 X (4x 3 - 7x) x e(xo - x), (3.11.13)
where e is the step function (1.5.3).

(i) Summary For convenience the original formula (3.11.8) and the
modified formulae are summarized as follows
(i) X given, F(X) to be found
x = (X - jJ.,J/ux (for jJ., u and y see (3.3.7))
g = y/6; Xo = -J(7/4)

y= J( ~) - ~
1 + _1_ +
4g 2 g 2g
for x ~ 1

= x - g(x 2 - 1) + g2(4x 3 - 7x)e(xo - x) for x < 1

F(X) ~ N(y). (3.11.14)
By using the notation of (3.11.10) the above relations can be
written briefly
(3.11.14a)

where v; l(X) is the transformation x ~ y defined by the third and

fourth lines of (3.11.14) and depicted in Fig. 3.11.2.
(ii) F(X) given, X to be found
Four auxiliary constants are first calculated
Yo= -JG)-~g
11 1 7
p=----
144g 2 12 (3.11.15)
17 1 5 1 1Y
Q= 1728g 3 + 96g - 8g 2
D =J(P3 + Q2),
where y is the root of F(X) = N(y), i.e. y = N- 1(F(X)). Then
x = y + g(y2 - 1) for y ~ 1

= ~
2g
- )(-1 + 1_~)4g 2 g
for Yo:::; y < 1 (3.11.16)

1
=J(D-Q)-J(D+Q)+12g fory<yo'
118 COMPOUND POISSON PROCESS

y 0·2 0·4
;001/03/0507
. '';: ,./"
10

•
5
IV ~ V V I.--V 1.5
IV: ~f% t/'V V- I.-- ~~ 2.0
4
~Vv V- ~ I- 3.0
V V ~
3
V

.;~I
2

1,,{.,,-
...&i,. I
-r 1-1~
I I I I

-5 -4 1 2 3 4 5 6 7 B 9 10
-3; X

'J
IA' -2
III/m
.aWl 3
!J Q:'r/,
V:'lI flJJrI -4
/,1/ 'I
V '/'/ 1 -5
r = 00·1 0·3 0·5 1 23

Figure 3.11.2 N P transformation x --> y, where y is N(O, 1) distributed and

x = Vy(y) is NP(O, 1, y) distributed.

and finally
X = J.1. x + xO'x' (3.11.17)

The last part of formula (3.11.16) is found from (3.11.14) by means of

Cardan's solution.
By using the notations of (3.11.10) the above relations can be
written briefly
(3.11.17a)

where viY) is now defined by (3.11.16).

Fig. 3.11.2 gives a view of the variable transformation x -+ Y
according to (3.11.14). It can also be used as a nomogram which gives
y approximately when x is given or vice versa.

G> The density function! of F is required for applications where

convolutions and other integrals are to be computed. It can be
3.11 NORMAL POWER APPROXIMATION 119

derived from the equations given by differentiation

-, dx _ 1
f(X) = F (x) dX = f(x) (lx'

- 1 _12dy
f(x) = J(2n) e ,Y dx
(3.11.18)

dy
for x ~ 1
dx 1 + 2gy'
= 1 - 2gx + g2(12x 2 - 7)e(x o - x) for x < 1.

where, as above, g = y/6.

When X is given, the corresponding y is to be calculated from
(3.11.14) and is then to be placed in (3.11.18).
Convenient approximations exist for programming of the normal
function N or its inverse (see (3.9.5) and (3.9.6».

(k) Asymptotic behaviour It was stated in item 3.3(g) that the

relative compound Poisson variable x = X/mn has for n --+ 00 the
limits (l x = (l q and yx = Yq and that its dJ. tends to the structure dJ. H.
Then NP(l, (lx' Yx) tends to NP(l, (lq' yq ), hence approximating the
structure function H(X/mn) (see (3.3.16». How good a fit the N P
function can give for F(x) for large values of the volume parameter
n depends, obviously, on how well the N P function can approximate
the structure dJ. H. This depends on the shape of H. However, in
practical applications this problem often has only academic value,
because the strict form of H is seldom known. Usually only its
standard deviation (lq can be evaluated and, perhaps, also the skew-
ness Yq to some degree. Then obviously the N P function with the
same parameters can represent it as well as any other dJ., unless
perhaps some particular reason justifies other conclusions.
Note that in the Polya case where the structure distribution is
assumed to be a gamma function, the limit distribution is also a
gamma function. Then the gamma approximation presented in
Section 3.12 may seem to be a natural approach. However, as will
be seen, the NP and gamma functions having the same mean value,
standard deviation and skewness give the same numerical values to
a very close degree.
120 COMPOUND POISSON PROCESS

(I) Discussion A drawback of the above approach is that the

derivative (3.11.18) has discontinuities at points 1 and xo. They
can be removed by applying the so-called spline technique for
example, but this would lead to more complicated expressions and
increase the computation time. Moreover, it would not generally
give a much better result than the direct application of (3.11.18),
because these formulae are needed mainly for the calculation of
J
integrals of type A(x)dF(x) where A is some function depending on
the application in question. The errors on either side of the
discontinuity points have opposite signs and hence are likely to offset
each other in integration.
The poor fit for negative values of the standardized variable x
has already been mentioned in item (0. If the parameter n is small,
this can result in a just significant probability that X has a negative
value (whilst X was assumed overall non-negative, see Section 1.3).
This is demonstrated in exercise 3.11.2, giving a simple test for the
applicability of the formula in circumstances where the negative tail
of the distribution is required.
Even if the formulae (3.11.14) to (3.11.18) are not so simple in form
as one could wish - mainly due to the necessity to split the range of
the variable x into sections having different expressions - they
are easily programmable and convenient for fast computation. A
major benefit is that they can be operated in either direction,
X ~ F(X) and F(X) ~ X.

(m) The Wilson-Hilferty formula (3.5.14) is another example of

'normalizing' transformation (3.11.1). Applied now to the direct
calculation of F(x) it gives values very close to those given by the
NP formula when the skewness y is not very large (~ 2 for positive
x and ~ 1 for negative x values).

Exercise 3.11.1 Let the moments about zero of the claim size dJ.
be a 1 = £10 3, a2 = £210 8 and a3 = £3 10 14 and let the standard
deviation and the skewness of the structure distribution be 0.1 and
0.5 and the expected number of claims n = 10000. Calculate the
probability that the total amount of claims exceeds 14 x £10 6
by using (a) the Normal approximation, (b) the NP formula and
(c) the Wilson-Hilferty formula.

Exercise 3.11.2 Show that when a compound Poisson variable is

3.12 GAMMA APPROXIMA nON 121

approximated by the N P formula, a condition for probability of

negative values to be < G is that
n > r 2 x e2 j(1 - Xe2 (J"2)
q , (3.11.19)
where xe is the root of G = Ny (x). Calculate the lower boundary
(3.11.19) of n for G= 10- 4 , r2 = 10, r3 = 200, (J"q = 0.05 and Yq = 0.5.

3.12 Gamma approximation

(a) Three-parameter gamma function The mixed compound

Poisson function F can be approximated also by the incomplete
gamma function r(ax + b, IX), where again x is the normed variable
(3.9.1) and the parameters a, band c are determined from the condi-
tion that the mean, standard deviation and skewness should be
equal to those of F. As was seen in Section 3.5.6, these conditions
result in
X;;::-JIX, (3.12.1 )
where
1X=4/y2.
Now Y is the skewness of the aggregate claim distribution (instead
of claim size dJ. in (3.5.9)).
References for calculation technique were given already in
Section 3.5.6 (see also Bohman and Esscher, 1964; Pentikainen, 1977
and Seal 1977).

(b) Applicability The NP approximation and gamma approxima-

tion are compared in Fig. 3.12.1. In the area which is mostly needed
for applications both approximations give about the same values.
Only when the skewness increases to the level of 2 do the values
diverge for the positive x values and for the negative x values already
before it. But these are variable areas where both approximations
are unreliable. The gamma distribution is not defined for values
less than - 2/y (or is defined to be 0).
Experience ofthe application of these formulae suggests that their
ability to give a satisfactory fit is roughly equal. The choice of the
formula should depend on the convenience for the user in the context
of each application. A handicap is that the gamma approximation
 --
122 COMPOUND POISSON PROCESS

II ~
5
"-..... 10 V ..... / --
'-"
5

/j 11==
4
=~
-i=
3 -40§
-20-- t=
-~( -1LT I f - - - -~
=\=

\
2

I~
1\ --~

1\
---=::::~ r=-
o

1/V ....- V
~ ~l--
~~ ~
-1
-~ -10?20
I / IJ~ F -1.00 , I
-2 --
5

---- ~ II
I
Gamma =0
t:=-- /10,
1.1'--
.........
1== f:ii It'll I !
~ 111111111 '1IIHt"ttt""';~1111
0.1 0.2 0.5 1.0 2.0 Skewness

Figure 3.12.1 Comparison of the gamma and NP functions. The niveau

curves of the relative deviation (r - N y )/N y x 100 for x:>;; 0 and
[(1 - r) - (1 - N y)]/(1 - Ny) x 100 for x> 0 as functions of the normed
variable x and the skewness yare plotted in a semi-logarithmic scale. In the
area bordered by the niveau curves + 5 and - 5 the absolute value of the devia-
tion is less than 5%. In the right hand side lower corner r = o.

is not easy to handle in problems where the function value F(X) is

given and the matching argument X is required.

Exercise 3.12.1 Instead of the three-parameter version just applied,

a two-parameter version r(ax, oc) can also be used. Calculate a and
oc from the condition that the approximating function should have as
mean and standard deviation Ilx and (J x. What is its skewness?

Exercise 3.12.2 Develop the Wilson-Hilferty argument y = the

expression in brackets in (3.5.14) as series and verify that the lowest
terms coincide with the N P expansion for z > 1. (Hint: Manipulate
the expression in a form where the quantity yz/2 can be used as
the argument of the series development.)
3.13 APPROXIMATIONS BY MEANS OF FUNCTIONS 123

**3.13 Approximations by means of functions belonging to the

Pearson family

(a) The Pearson system The idea of both the NP approximation

and the gamma approximation was to find an analytic d.f. which
has a shape known by experience to be generally of close fit to
distributions functions of aggregate claims, and then to equate
the lowest moments. Besides these two functions there are also a
number of others which are suggested for similar use. An approach
to solving the problem is to assume that the approximating function
belongs to the Pearson system of curves (see item 3.5.11(b». Use
of two moments gives the normal approximation described earlier.
The use of three moments implies that the assumed distribution
of F 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 the model
parameters Y1 and Y2 (Johnson et ai., 1963) may be sufficient for
practical purposes, and may afford a rapid method of approximation
since the effect of the factors Y1 and y2 can be easily assessed.

(b) The beta density function

x P - 1 (1 - x)q - 1
f(x; p, q) = B(p, q) for 0 < x < 1

=0 for x ~ 0 or x ~ 1 (3.13.1 )
with

B(p, q) = I uP - 1 (1 - U)q-l du,

was used by Campagne when the rules for the EEC convention for
solvency margins were under consideration. Here x is the claims
ratio defined as the claims paid for the insurer's own account divided
by the premiums including also loading for expenses. p and q are
parameters, the mean of x being
J1x = p/(p + q)
124 COMPOUND POISSON PROCESS

and variance
2 pq
(J'x = -:-(p-+-q-:-;;)2,....:-(p-=-+-q-+-----:-7"I)

A description of the method and some follow-up investigations can

can be found in De Wit and Kastelijn, 1980. The mean rnx for ten
Dutch companies was found to be 0.43, (J'x 0.089 and p = 12.9,
q = 16.9. The probability x> 0.78 = rnx + 3.9(J'x is, according to
these values, 0.0003.

**3.14 Inversion of the characteristic function

According to a general theorem of Fourier transforms the transform
(1.6.8) F --+ cp, which couples any dJ. to its characteristic function,
has an inverse transform, which gives F uniquely at continuity
points

F(X) =
I
2 lim
f+Tl - .e- isX
cp(s) ds + tF(O). (3.14.1)
1tT-+oo -T IS

Thus, in case F is a compound Poisson dJ. and if n, Hand S are

known, the characteristic function can be calculated (see Section
3.4) and thereafter F according to (3.14.1). Though apparently
simple, this method leads to the problem of quadrature of complex
functions, which may oscillate widely, and numerical methods have
to be used with care. Successful applications have been made by
Bohman (1964) and Seal (1971) and more recently by the use offast
Fourier transforms backed up by a computer facility (Brigham,
1974; Bertram, 1981).

3.15 Mixed methods

(a) The decomposition of S can be achieved 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 (see
item 3.7(g)).
If the largest class consisting of claims (~Z 1) is taken as one
component, it can be expected that for the remaining part the normal
or the N P approximation will suffice as soon as the number of
claims is sufficiently large, say a few thousand (Fig. 3.15.1) since the
3.15 MIXED METHODS 125

sr----------:::::=-r--

Z, Z2 Z3 Z
Figure 3.l5.l Decomposition of S function.

failure of these approximations in practice largely arises from the

long tail of the S distribution.
The tail left outside the large class of ordinary claims can then be
divided into one 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 (Section (6.8) for
each component, or the recursion formula (Section 3.8).

(b) An example of the mixed method in handling of catastrophic

risk is given in item 6.8.4(f) and it is dealt with by Rantala (1982), who
also treats the 'multichannel problem' brought about by the fact
that a major catastrophe may affect simultaneously, through quite
numerous treaties, insurers who have accepted risks via the world-
wide reinsurance network.
 CHAPTER 4

Applications related to
one-year time-span

4.1 The basic equation

(a) One year as time-span The various formulae developed in the

previous sections provide the means to solve problems of central
interest in the area of applications of risk theory. In this chapter
consideration is restricted to a one-year time-span only, such as
when evaluating the limits in which the underwriting result will
fluctuate and how it depends on background factors like the size
of the portfolio, the distribution of the claim size, reinsurance, the
level of the safety loadings, etc. Limitation of the study to one year
makes it possible to express many of the interdependences of the
variables involved in a form which is easy to handle when analysing
the structure of the risk process. Of course, a one-year time-span
is not sufficient for consideration of many important problems
concerning e.g. solvency, long-range planning, etc. and this restric-
tion will be relaxed in later chapters. However, results concerning
a short period are of interest in many connections, e.g. it may be
useful to know the range of fluctuation of the annual underwriting
gain or loss. Features unveiled for the one-year case are then easily
extended for longer periods in an analogous way.

(b) Risk premium In previous chapters the claims process only

was examined and now a new variable will be introduced, namely
the risk premium income
P = E(X) = Ilx = mn. (4.1.1)
This, by definition, is the expected value of claims, all the variables
now being related to a one-year period (see (3.3.7)).
4.1 THE BASIC EQUATION 127

(c) Safety loading It is further assumed that the risk premium is

increased by a safety loading. This can be done in a number of
different ways, e.g. by relating it to the risk premium, to the standard
deviation of the assumed risk or to the variance, or by using some
combination of these components as follows

L.l = A'P." + A"ax I. + A"'aX2 I•• (4.1.2)

Here the lambdas are coefficients, the determination of which is one

of the central problems in premium rating. The subscript i refers to
individual policies or group of policies. Of course the lambdas may
be different for different types or groups of insurance. From the
point of view of most considerations in risk theory the question of
how the individual policies are loaded is not relevant. The important
quantity is the total income arising from these loadings, i.e. r.L i •
Then a weighted safety loading coefficient will be defined by
(4.1.3)

It can first be calculated separately for different sections of the

portfolio, after which the coefficient A for the whole business is
obtained from the (average) section loadings applying (4.1.3) to
section data

(4.1.4)

where j refers to section. The weighting factors are the same as

defined already by (3.7.8).
In order to incorporate the safety loading in the model, the loaded
premium (1 + A)P will mostly be used instead of the risk premium
P. Even if the safety loading is formally written in linear form, it
means a weighted average as defined by (4.1.3) and (4.1.4), and the
type of rating formula e.g. (4.1.2) or any other is irrelevant. The
coefficient depends, of course, on the structure and also, for non-
linear loading formulae, on the size of the portfolio, because Ai =
L;/Pi may be different for different risk units i. However, the structure
of the portfolio is normally changing only slowly, so that A is a
fairly stable quantity and can be used as one of the characteristics
ofthe portfolio.

(d) Underwriting process It is now assumed that the premium

 128 APPLICA TlONS RELATED TO ONE-YEAR TIME-SPAN

Ruin barrier Ur

Time

Figure 4.1.1 The risk process as a difference of incoming premiums and out-
going claims.

income is accumulated to a risk reserve U and the claims are paid

out from it as demonstrated in Fig. 4.1.1.
A basic problem to which many other considerations are related
to is to examine the distribution of the risk reserve U at the end of
the accounting period. In fact U - U 0 means underwriting gain
(or loss, if negative). Then the range of variation of this quantity is
sought. This problem can also be formulated to concern the so-called
one-year ruin probability B, the probability that U could drop
under some ruin barrier U, which will be defined for the problem
under consideration. It can be, for example, the statutory minimum
solvency margin. Recall that the ruin probability can be defined in
a continuous or discrete way (see item 1.4(a)), i.e. according to
whether U will be less than U r at any time during the observation
period or only at its end. For reasons discussed in Section 1.4 the
latter possibility, the 'year-end ruin probability', will be used in the
following. The answer to the ruin question is immediately provided
4.1 THE BASIC EQUATION 129

by means of the dJ. F of X as follows

1 - e = prob {U ;?; U,} = prob {X:::::; U0 -Ur + (1 + A)P}
= F(U 0 - Ur + (1 + A)P). (4.1.5)
1 - e is often called the 'survival probability'.

(e) Basic equation To get the equation in a form which gives the
interdependence of the involved variables explicitly, it is assumed
that the N P approximation is applicable.
For brevity U is here and later often scaled by putting Ur = 0,
and U 0 will be denoted by U if the clarity in the context concerned
does not necessitate the subscript. Furthermore, let Y, be the value
of the standardized variable which corresponds to the ruin probabi-
lity e according to the equation
e = N( - y,) = 1 - N(y,). (4.1.6)
For example, YO.Ol = 2.326 and YO.OOl = 3.090. The shorter form
Y will sometimes be used in place of Y, .
Making use of the above notation and the N P formula (3.11.16)
(for Y ;?; 1) and (3.11.17) the relation (4.1.5) can be written in the form
of the following basic equation
(4.1.7a)
Assuming X as a compound Poisson variable and substituting the
expressions of the standard deviation and the skewness from (3.3.7)
this equation can be written
U = y,pJ(r 2 /n + (J~) - AP
+ iP(y; - 1) x (r3/n2 + 3r2(J~/n + Yq (J!)/(r 2/n + (J~). (4.1.7b)
In the particular case where the normal approximation is applicable
the above equations are reduced to the shorter form
U = Y,(Jx - AP = YEP J(r 2/n + (J~) - AP. (4.1.8)
These equations contain either explicitly or implicitly the quantities
e, A, M, U, (Jq' Yq and n or P. (4.1.9)

Furthermore, the coefficients r2 and r3 and the risk premium income

P depend on the claim size dJ. S or rather on its lowest moments a l ,
130 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

a2 and a 3 • The maximum net retention M is included in the list

(4.1.9) because these moments depend on it, M being one of the
'primary control variables' (see Section 3.6). As a primary volume
variable either n or P will be used. When one of them is given as
well as the other variables (4.1.9) and the claim size dJ. S then
the basic equation is determined.
The portfolio under consideration may mostly be divided into
sections and the basic variables (4.1.9) are derived from the section
variables and data as presented in Section 3.7. This phase of the
calculations will be assumed to be ready made and will be no longer
referred to. Then it is sufficient to treat the portfolio as if it were
undivided.
The main problems of concern in this chapter are of the type where
the claim size function S is given together with six of the quantities
(4.1.9), the seventh being the object of study.
A more general family of problems is introduced by treating
more than one of the quantities (4.1.9) as unknown. In this case the
basic equation does not give a definite solution, but auxiliary condi-
tions 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 (4.1.7) or (4.1.8) 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 lines of business,
in which event there are several unknown, 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 con-
sidered in Section 4.6.

4.2 Evaluation of the fluctuation range of the annual underwriting

profits and losses

(a) Problem setting An important application of risk theory is

the determination of the range in which the annual profit or loss
resulting from the underwriting operation may fluctuate. Thus the
lowest point U 1 of the risk reserve may be sought, below which the
risk reserve U will not fall for a given probability 1 - B (Fig. 4.2.1(a)).
Another way of formulating the problem is to ask what must be the
4.2 EVALUATION OF THE FLUCTUATION RANGE 131

u u

---
U(o)
--- ---
U(O)
--- ------ ---
1-oE--:;,,;--------'-I U(O)+~ p

(a) Time (b) Time

Figure 4.2.1 Two ways of formulating the problem of reserve fluctuation:

by seeking (a) the lowest point UI and (b) the initial reserve U = U(O).

initial amount U = U(O) such that the reserve will not be exhausted
at the end of the accounting period (Fig. 4.2.1(b)). The latter ap-
proach will be assumed in this chapter but the former problem
setting will prove important in subsequent chapters.

(b) Standard data A solution can be obtained directly by applying

(4.1.5) or, if the N P approximation is applicable - as it usually
is - the NP formula (4.1.7) or the normal approximation (4.1.8).
Examples showing how U can be examined as a function of the
other six variables (4.1.9) are set out in Figs. 4.2.2-4.2.6. In order to
make the sequence of the applications mutually comparable, and
also comparable with the applications presented in Chapter 7
for longer time spans, the same standard data will be used, unless
otherwise mentioned, for the variables (4.1.7) and for the basic
distributions as follows:
t: = 0.001
A = 0.04
n = 10000 (4.2.1)
M =£10 6

(Jq = 0.038
Yq = 0.25
SM(Z) according to Table 3.6.1.

The value of U which satisfies (4.1.7), when the above data are
substituted in it, is U = 15.72 £10 6 . The risk premium income
corresponding to the data is P = 73.0 £10 6 •
132 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

114=10
U 50

40
4

0.4
0.1

10

50 100 150
p

Figure 4.2.2 U as a function of net premiums P and net retention M. Unit is

£106. Standard data (4.2.1).

(c) Risk reserve V = V (P, M) The initial mlmmum risk reserve

U as a function of the risk premiums P and the retention limit M
is depicted in Fig. 4.2.2. As understandable, U is an increasing func-
tion of both P and M. The structure of both dependencies will be
discussed more fully later. Note that both U and P = n x m(M)
depend on M.

(d) Solvency ratio VIP = f(P, l) In some situations a more ap-

propriate approach is to formulate the problem in terms of the
relative amount for the risk reserve rather than the absolute amount
U. In the following this will be called the solvency ratio and will be
denoted by u. This is done in Fig. 4.2.3 where minimum initial sol-
vency ratio u is calculated as a function of P and A.. It quite obviously
tends asymptotically to some horizontal level, a feature which will
be analytically explained subsequently. If A. is high and the business
volume large, then no initial capital U at all is necessary, i.e. the
safety loading alone is sufficient to cover the adverse fluctuations.
In the special case where the structure variation has degenerated
(aq = 0) u tends to - A. when n ~ 00.
Even though a larger insurer needs more initial capital U than
a smaller one, according to Fig. 4.2.2 the relative need, i.e. the smallest
4.2 EVALUATION OF THE FLUCTUATION RANGE 133

VIP 1.0

0.5

-==-____________________ A=

_===================
_ 0.00

0.10
0.05
0.20
50 100 150 P 200

Figure 4.2.3 The solvency ratio u = VIP as a function of P and A.. Standard
data (4.2.1).

still safe solvency ratio, is a decreasing function of the company

size.

(e) Net retention M = f(n, U) Next the net retention limit M is

examined as a function of U and n in Fig. 4.2.4. In this example M
affects significantly the risk coefficients r2 and r3 (see (3.3.8)) as
seen in Table 3.6.1. In the double logarithmic scale the curves are
approximately linear. Of course, for claim size distributions contain-
ing only small risks the structure of the figure is another type - the
curves have a vertical asymptote when M increases beyond the values
of the maximum claim sizes.
Retention problems are discussed further in Section 4.5.

(f) Ruin probability II =f(n, U) Figure 4.2.5 sets out the ruin pro-
bability t: = 1 - F as a function of nand U.
In order to demonstrate the influence of the structure function
its variance was removed in two cases (a q = 0) dotted in the figure.
As seen, the change for large collectives (n large) is quite crucial.
This means that the structure variation is the main cause of the
fluctuations of large portfolios, whereas the 'pure Poisson' fluctua-
tion is dominant for small portfolios. This very same feature was
already anticipated by Table 3.3.2.
 n/l000 =1 5 10 20 50 100
10.0
M
7.0
5.0

3.0

2.0

1.0
0.7
0.5

0.3

0.2

0.1 ~~~~--~~__~~~~-----4--~~~~~~~
1 2 3 5 7 10 20 30 50 70 100
u
Figure 4.2.4 The net retention M as a function of U and n. Monetary unit is
£10 6 (double logarithmic scale).

U = 2
5
10
0.1000
20

50
0.0100
10

0.0010 -- - - - - - -- -- - -- -- --

2 3 5 7 10 20 30 50 70 100
n/l000

Figure 4.2.5 The ruin probability e as a function of nand U. Two Poisson

cases (uq = O) are plotted by dotted lines.
4.2 EVALUATION OF THE FLUCTUATION RANGE 135

2.0
UIP

ex: 0.9
1.0 1.1
Standard
0.7
1.5
0.5

0.3
=-_-------2.5

0.2

0.2 0.3 0.5 0.7 1.0 2.0 3.0 5.0 7.0 10.0 20.0 ~.O 50.0 100.0
M

Figure 4.2.6 u = UIP as function of M for different Pareto-distributed tails.

Unit for M is £10 6 .

If Ais positive the ruin probability as a function of the size variable

n has, in general, a maximum even though its location may be,
owing to the structure component (O"q > 0, Yq > 0), quite remote.
In the Poisson cases the maximum is clearly seen.

(g) The effect of the claim size d.f. To show the effect of the choice
of claim size function S, calculations based on some Pareto distribu-
tions are set out in Fig. 4.2.6 together with values for our standard
distribution.
It was assumed that S(Z) is in all cases the same and equal to
the standard up to the value Zo = £10 6 , and the tail is upwards
Pareto distributed (see (3.5.20)). Even though only 0.072~~ (see
Table 3.6.1, line 26) of the probability mass of the S(Z) distribution
is located in the area Z> Zo its influence on the behaviour of the
tail of the function u = U IP = u(M) is quite significant. This means,
among other things, that if a portfolio gross of reinsurance is
considered, its solvency properties may depend quite crucially on
the assumptions concerning the upper tail of the claim size distribu-
tion. As already expected (see item 3.5.8(c) and Table 3.5.1) the
standard distribution is to be classified as 'dangerous'.
136 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

(h) Effect of the risk of catastrophic claims The estimation of the

chance of occurrence of very great claims is one of the major practical
difficulties in risk theory, owing to the fact that the experience
basis is nearly always limited and unexpected events may give rise to
losses of unforeseen dimensions. It was suggested in item 3.5.3(c)
that in uncertainty hypothetical 'shadow' claims should be incor-
porated into the claims statistics when the claim size dJ. is derived.
Another way to investigate the sensitivity of the outcomes for large
claims is to assume that, in addition to the 'normal' claims expendi-
ture X, a large single claim Xc will still occur with some estimated
frequency q. Because q is expected to be very small, the occurrence
of two extra large claims in one year can be omitted. Then the dJ.
F(X) of the aggregate claim is changed to the form
F(X) = (1 - q)F(X) + qF(X - XJ, (4.2.2)
which is the straightforward convolution of the original distribution
and the distribution consisting of the single claim Xc' the latter
being binomial with fixed sum Xc.
An example of the application of equation (4.2.2) can be found in
exercise 4.2.5.

(i) Profile figure Because it is not easy to get an adequate view of

the interdependence of the variables involved by merely constructing
figures like Figs 4.2.2-4.2.6, an attempt to give a synthesis is made
in the profile Fig. 4.2.7. The idea is first to calculate U for average
circumstances (basic alternative). Then the different variables are
changed individually so that the sensitivity of U to the background
factors can be examined and a mental picture obtained of the risk
structure.

Exercise 4.2.1 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 and the premiums remain unchanged. Equation
(4.1.8) can be used, in millions of £, with data U = 20, n = 10000,
m = 0.01, r 2 = 30, O"q = 0 and A. = 0.05.

Exercise 4.2.2 A friendly society grants funeral expense benefits

and each member of the society can choose a benefit of either £100
4.2 EVALUATION OF THE FLUCTUATION RANGE 137

~1~1I11~lIIl1l11l11llllll11l1l11l1l11l11llll11m 6.1 Basic alternative

1111111111111111111111111111111111111 4.3 n = 5000

1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 ~1 1 !1 1 1 1 1 1 1 9. 2 n = 20000
111111111111111111111111111113.4 ~ = 0.1
mllllllllllllllllllllllllllllllllllllllllllllllllllill111111111111111111 8.9 A = 0

1111111111111111111111111111111111111111111111111111111111111111111111111111111111 9 . 7 M=1

1111111111111111111111111111111111114.3 !Tq = 0

1111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111112.7 !Tq =0.1

1111111111111111111111111111111111111111111111111111111111111111111111111111119.3 E= 0.001
Figure 4.2.7 Dependence of the risk reserve U according to the basic equation
(4.1.7). U is calculated first to the basic combination n = 10000, M = OA,
A = 0.05, O'q = 0.04, Yq = 0.25, P = 54.6,8 = 0.01, having unit = £10 6 . Then the
variables changed as given in the figure and the value of U is plotted individually
at each bar.

or £200. It is assumed that A. = 0.5, n = 20, O'q = 0 and s = 0.01. How

large should the reserve fund U be according to (4.1.8) if it is not
known in advance how many members will choose the option 100
and how many the option 200 and, consequently, what mix of these
options is to be assumed so as to maximize the risk?

Exercise 4.2.3 Assume that the distribution function of one claim

can be represented by the exponential function S'(z) = e-Z(z ~ 0)
(taking the average size of claims as the monetary unit). The ex-
pected number of claims n is 1000 and 0' q = 0.04. How large should
the safety loading A. be according to (4.1.8) if there is no reinsurance
and no reserve fund, and if s is fixed at 0.01 ?

Exercise 4.2.4 Let the claim size d.f. be of Pareto type, equal to
1 - z-a for z ~ 1 expressed in suitable monetary units; excess of
loss reinsurance is arranged with maximum net retention M. Cal-
culate M by one decimal from (4.1.7) when IY. = 2.5, ~ = 20, n = 100,
A. = 0.1, 0' q = 0 and Ye = 2.33. Hint: Derive an expression for U as
a function of M and then find (by trial and error) the requested
numerical value of M.

Exercise 4.2.5 Let the portfolio be the standard one defined by

the data of (4.2.1). How is the ruin probability s changed if the chance
138 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

of an extra claim Xc according to item 4.2(h) is assumed? Experi-

mental values are Xc = £5, 10 and 20 million respectively. The extra
claim Xc is expected to occur once in 20 years. F can be approximated
by the N P function.

4.3 Some approximate formulae

(a) Asymptotic behaviour Although the procedure presented in

the previous section is not very troublesome in practice, it still does
not provide a clear insight into the interdependence of the basic
quantities (4.1.9). To obtain a comprehensive survey it is more useful
to derive explicit even if approximate expressions.
First manipulate (4.1.7) into the form

VIP = (y + I'x(i- 1)/6)J(r 2 In + 0";) - A. (4.3.1)

When n -+ 00 it follows that

V ( y2 - I )
P -+ Y + -6-l'q O"q - A, (4.3.2)

since according to (3.3.7) the skewness Yx -+ I'q. Hence only the

terms related to the variation of the basic probabilities are left.
This is again the result mentioned in item 3.3(g), i.e. the mixed com-
pound Poisson distribution asymptotically tends to the structure
distribution H.
It is seen from (4.3.2) that in the general case (O"q > 0) the limiting
value of VIP is positive or negative depending on the size of A. The
character of the process is essentially different in the Poisson case
where 0"q = 0, i.e. when variation of the basic probabilities vanishes.
Then VIP -+ - A and the standard deviation 0"x -+ O. As was seen
in item 3.3(g) in fact the variation of the basic pro babilities constitutes
the fluctuation in large collectives. This same feature is seen in
Fig. 4.2.5. The horizontal asymptote, which was anticipated already
from Fig. 4.2.3, can be explained by (4.3.2).

(b) Distribution-free approximations Situations often occur in

which general ideas of the magnitude of M or V are needed, when
the function S is not known or when for other reasons it is not
possible to make use of it in calculations involving (4.1.7). In these
4.3 SOME APPROXIMATE FORMULAE 139

circumstances it is possible to find a distribution-free estimate for

U.
The structure variation affects the expected number of claims
n = nq via the variable q making n larger or smaller than n. While
the premium P is assumed to the fixed P = (l + A) x E(X), the
effect is equivalent with the fact that the safety loading AE(X) =
P - E(Xln(t)) varies from year to year. Furthermore, the
structure variation affects the width of the variation range as seen
by equation (4.1.7b). These observations make it possible to find
upper limits for the various fluctuation ranges replacing the effect
of the structure variations by a conservative deterministic choice of
A and n. This approach is the more appropriate when the long-term
cycles are also to be taken into account (see Section 7.4).
Assuming that the portfolio is reinsured so that the individual

f: f:
claim sizes are limited to ~ M, the inequality

ai = Zi dSM(Z) ~M Zi-l dSM(Z) = Ma i _ 1, (4.3.3)

can be utilized. The equality is valid only if the size of one claim is
constant and equal to M. Putting this limit in (4.3.1) and using the
convention P = mn, a distribution-free upper limit for U is obtained
U ~ yJ(PM) - AP + i(y2 - l)M. (4.3.4)

r- ...
0.5 r-- i'oo.

0.3

0.2

0.1
10 100 1000 10000
M/£1000
Figure 4.3.1 The factor K (see (4.3.5)) as a function of M. Claim size df as
given in Table 3.6.1.
140 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

To get an idea of how rough the approximation (4.3.4) may be,

its main 'error factor', the ratio
(4.3.5)

of the main terms in (4.1.8) and (4.3.4) is plotted in Fig. 4.3.1 for
the standard d.f. (Table 3.6.1). This figure as well as numerous other
examples suggests that for the commonly used values of M the
factor K lies in the interval 0.5 to 0.8 and that the value K ~ 0.6
can be used as an estimate. K is (possibly with some rare local
exceptions) a decreasing function of M and it tends to zero when
M --t 00.
It follows that the approximation (4.3.4) can safely be improved
into the form
u ~ KyJ(PM) -).p +i(/- I)M, (4.3.6)

where K is to be fixed according to the application in question.

Assigning to K the value 0.6 further simplification is obtained

u ~ 1.4J(PM) - )'P (6=0.01) (4.3.7a)

~ 1.9J(PM) - AP (8 = 0.001). (4.3.7b)

The last term in (4.3.6) is 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 an overestimation, as seen from Fig. 4.3.1 and thus
compensates for the omission.
Some curves representing the approximation (4.3.7b) are shown
in Fig. 4.3.2 and, for comparison, a curve calculated directly by
means of the basic equation (4.1.7) omitting the structure variation
(i.e. (J q = 0). 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.

(c) The most dangerous distribution The formula (4.3.4) has a

nice interpretation. It corresponds exactly to a hypothetical port-
folio where all claims are equal to M and where the expected number
of claims n = P/M. For this particular claim size distribution the
moments a i are = Mi and the equality in (4.3.3) and (4.3.4) is valid.
Hence this portfolio is the most dangerous among all having the
4.3 SOME APPROXIMATE FORMULAE 141

40
A:-0.10
U
-0.05

30

°
20

_------0.05

10

Basic eq.
0.15
50 100 150 200
P

Figure 4.3.2 U values, assuming M = £10 6 , computed by approximation

(4.3.7 b) and for .Ie = 0.1 also by the basic equation (4.1. 7).

same P and the claims being limited in the interval 0 ~ Z ~ M,

i.e. it needs the largest initial reserve U.

(d) Some general properties of the solvency structure can be seen

from (4.3.7) and from Fig. 4.3.2. The magnitude of the lowest, still
safe, risk reserve U as a function of the volume of the portfolio,
measured by P, is a parabola. If A is non-positive then U increases
at least proportionately to the square root of the premium income P.
If, however, A is positive, the curve of U has a maximum given by:

(4.3.8)

Even if (4.3.7) is only approximate, the general shape of the results

is correct, however, the true values of U for large values of M being,
in fact, smaller.
The formula (4.3.8) gives a distribution-free upper limit for the
relationship of M to U or to P respectively, which will be further
discussed in Section 4.5. It should be recalled, however, that the
derivation of the above results was made assuming that the omission
142 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

of the structure variation can be compensated for by a conservative

selection of the values to be assigned to the parameters A. and n (or Pl.

4.4 Reserve funds

(a) Two types of objectives Insurance legislation generally contains

provisions which are intended to give guarantees of the solvency
of insurers so that 'insurance consumers' can, as far as possible, be
safeguarded from the serious consequences which could result
if an insurer were incapable of meeting pending claims or fulfil
other commitments. For this purpose, among other measures,
minimum amounts for solvency margins are prescribed.
The dichotomy of the objectives of solvency politics must be
appreciated. The public interest, as presented in legislation, is
restricted to safeguarding the insurance consumers. For this purpose
it is sufficient to prescribe for the solvency margin a legal minimum
which is high enough to make it very unlikely that it could be
exhausted in one accounting year. If the actual solvency margin
drops below the legal minimum, then the insurance company is to
be wound up unless immediate remedial measures are successfully
carried out in order to get the solvency margin back to the adequate
level. This minimum amount of the solvency margin is often called
the wind-up barrier or ruin barrier. The problem of defining it is in
principle the same as that discussed above and described in Fig.
4.2.1(b). The problem was, however, simplified in so far as only the
fluctuations of the claims business are taken into account. In
practice, of course, losses and fluctuations in investments and other
non-insurance risks are, in addition, to be considered (see Benjamin
(1977); Pentikainen (1952, 1982); Pesonen (1964, 1967(b)). On the
other hand, it is a matter of fact that the solvency condition is not
satisfactory from the point of view of the insurers. An objective
must be a safe survival for a period longer than one year. For this
purpose the time span must be extended from one year to longer
periods. This will be done in Chapter 6.

(b) Statutory minimum solvency margins are considered in this

section. The foregoing results offer a ready formula for that part of
them required for underwriting operations. Owing to the general
nature of the problem, just the distribution-free approximation
formulae are appropriate for the purpose. Because the objective of
4.4 RESERVE FUNDS 143

80 Premium 8

Figure 4.4.1 A modelfor statutory solvency margins.

legislation is above all to ensure that the values of reserve funds are
adequate also when applied to the weakest cases, it is reasonable to
assume that the safety loading A. is not positive. Then, according to
(4.3.7) and Fig. 4.3.2, U obviously must be an increasing function of
the volume of the company, i.e. an increasing function of the premium
income P. Hence it is of the type which is plotted by solid line in
Fig. 4.4.1, being a parabola for A. = O. According to practice, instead of
the risk premiums the gross pre~iums E, including also loading for
expenses, are used as a basis for the rule.
The parabolic curve can be approximated by a broken line as
presented in the figure. Expressed as a formula it is
(4.4.1)
Solvency margin rules of just this type are applied, for example,
in the decrees of the EEC (European Economic Community) in
1973. According to the EEC rule, a = 0.18, b = 0.02 and Bo is 10
million monetary units. Alternatively the basis is defined as
aggregate claims instead of the gross premiums and then the
constants are a = 0.26, b = 0.03 and Bo = 7 million units. The
constant U 0 is 0, but instead a certain minimum amount for U is
defined (for details see Kimball and Pfennigstorf, 1981).
In Great Britain up to 1978 a similar rule was applied with
a = 0.2, b = 0.1 and Bo = £2 500000. In Finland, where this type
of formula had already been introduced in 1953, U 0 = 0.2 millions
of FIM, a = 0.2, b = 0.1 and Eo = 4 millions of FIM.

(e) A merger of portfolios That the need for reserve funds is

dependent on the size of the company can also be illustrated by means
144 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

ofthe following example. Let C 1 and C 2 be two insurance companies.

The respective minimum reserve funds V 1 and V 2 are computed by
means of (4.1. 7). The question can now be asked: how is the need for
reserve funds changed, if these two companies are incorporated by
a merger into one company C?
Assuming independence of the portfolios, the following expression
for the minimum reserve V of the merged company C is obtained (see
exercise 3.3.4) by means of (4.1.7)

where the (second) subscripts refer to the original companies.

V1 + V2 - V = yJJo-i + Ja1- jO'i + 0';]
+ y; -
6
1 ('\1 +
0'1 0'2
+ 11~.2) > O.
11322 _ 11 3.;
+ 0'2
0'1
(4.4.2)

The inequality follows on the inequalities j a + jb >j(a + b)

a c a+c
and b + d > b + d which are valid for every positive a, b, c, d.
The inequality (4.4.2) proves that the reserves needed by the
merged company are always less than that of the separate component
companies together, if the security level 1 - c is unchanged.
The rule can be readily extended to deal with the incorporation
of more than two companies and cases where the structure variations
of the companies are not independent.
The result is of general interest; it proves that the larger the com-
pany, the smaller a solvency ratio needed. The same fact can also be
seen directly from (4.3.1), which shows that the relative minimum
reserve VIP is a decreasing function of the volume variable n = Plm.
One can say that a merger helps to use the existing reserves
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 fiscal merger of the companies is, of course, not necessary.
4.5 RULES FOR THE GREATEST RETENTION 145

The same advantages can also be reached by exchange of reinsurance

on a reciprocal basis. This problem will be further considered in
Sections 5.1 and 5.2.

Exercise 4.4.1 The following characteristics are computed from

the statistics of an insurance company C 1 : m1 = £1000, r2 ,l = 40,
r 3 ,l = 400, C1q ,l = 0.05, Yq,l = 0.1 and n 1 = 1000. The company has
a reserve fund U 1 = £500 000 and safety loading A1 = 0.1.
Another insurance company C 2 with the following characteristics:
m2 = £500, r2,2 = 50, r3,2 = 500, C1q ,2 = 0.1, Yq ,2 = 0.5, n2 = 200 and
A2 = 0.05 is merged with company C 1 •
If the ruin probability e according to (4.1.7) of the former company
is not allowed to increase following the merger, how large should the
reserve fund U be for the merged company?

4.5 Rules for the greatest retention

(a) Definition of problem The compound Poisson function or the

basic equation (4.1.7) can also be used for calculating the proper
dimensions for the net retention according to the reinsurance philo-
sophy of the cedent. The problem can be defined as follows. The
cedent has an amount U available to meet adverse fluctuations of the
risk business. In general this amount would not be the same as,
for example, the total solvency margin or the reserve funds, but
would be taken as the amount of resources which could be lost
without too much inconvenience. 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 to include in U, if
possible, merely hidden items. Furthermore, owing to the risk of
several consecutive adverse annual results, it may be advisable to
take as U some part only of the resources. This problem can be treated
more appropriately when the time span is extended from one year to
longer periods (Chapter 6). However, it may often be useful also to
know the range of the annual fluctuation as a function of the chosen
level of net retention. Therefore consideration limited to one year only
is also of interest and will be given in this chapter. By this means the
results can be derived more easily and provide a better qualitative
general view of the structures.
146 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

Hence the problem is formulated as what should be the maximum

net retention M such that, by probability 1 - G, the fluctuating result
of risk business does not consume a given initial reserve V in one
year.
Broader views of the problem are considered by Beard (1959)
and in numerous subsequent text books on reinsurance.

(b) The effect of the level M of the net retention has already been
studied in Figs 4.2.2, 4.2.4 and 4.2.6 for excess of loss treaty applying
the data given in Table 3.6.1. Similar figures can be obtained also for
other types of reinsurance by virtue of the technique considered in
Section 3.6.
Owing to the fact that the solvency properties are fairly robust
for the dJ. of the claim size as long as the net retention limit M is
not very high (see Pentikiiinen, 1982, Section 4.2.3), the values obtain-
ed for excess of loss type of reinsurance may be used as an
approximate guide for surplus treaties as well. This is useful owing to
the fact that the latter type of treaties is rather inconvenient to handle
as discussed in Section 3.6.4. This conclusion was confirmed by
Heiskanen (1982) who calculated various cases by both excess ofloss
and surplus rules.

(c) A straightforward solution can be obtained by plotting V as a

function of M as shown in Fig. 4.5.1, which was calculated applying
the basic equation (4.1. 7).
When the value of V is given, then the matching M can be read
from the graph. For example, if V = 5 then M = 0.087. An alternative
method is, of course, to solve the equation for M directly by numeri-
cal methods; however, this may be somewhat laborious, because
M is buried in the claim size function S (see Section 3.6). By a suitable
computer technique it is, however, fairly tractable. An example can
be found in Fig. 4.2.4 and in Fig. 4.5.2.

(d) The contradiction between profitability and solvency The risk

premium income P = nm(M) calculated net of reinsurance is an
increasing function of M. The increase is fairly rapid for small M
values, slow for large values and is stopped when M is equal to the
largest risks of the portfolio. This is illustrated in Fig. 4.5.1 by a
P curve. Because the ceded reinsurance premiums can be expected to
include safety and expense margins (the interest should be regarded.
4.5 RULES FOR THE GREATEST RETENTION 147

100
----_ ..
U
--- --- -- --
-- --
70

---
p -----
50

-- -- --
--'"
30 ",-

20

10
7
5 ----------------------------
i
3 ii
2

1~--~~~~~--~~~~~~--~~~~~.
!
0.01 0.020.03 0.05 0.1 0.2 0.3 0.5 2 3 5 7 10
lot
Figure 4.5.1 Risk reserve U as a function of the net retention M. Unit is
£10 6 • Standard data according to item 4.2(b).

among other factors; see item 1O.2(b)), it would pay, in expecta-

tion of profit, to have as high a maximum net retention M as
possible. On the other hand the range of fluctuations also increases
with M and the maintenance of solvency puts an upper limit on M
as presented above. This suggests as an optimal policy of having as
high a maximum M as the above basic equation allows within the
limits of the resources U which are available for covering of adverse
fluctuations.
One should appreciate, however, that the real life situation may
be more complicated than that just assumed. For example, the
cedent may have the opportunity of obtaining satisfactory reci-
procity against the ceded business and in this way balance the
profitability. The risk exchange between insurers will be discussed in
Section 5.2.
The problem of the optimal level of net retentions is considered
by Rantala (1982, Chapter 6) and is exemplified in exercise 4.5.5.
The reinsurance cost is further discussed in Section 7.3 and
item IO.2(b), and the problem of balancing contradictory business
objectives such as profitability and solvency in Section 10.4.

(e) The effect of the background factors (see (4.1.9)) can be examined
148 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

2.0
M

0.7

0.5

0.3

0.2

0.1 '-----~-_- ____ ~_---_- ___

2 3 5 7 10 20 30
u
Figure 4.5.2 Mas afunction of U and A.

as in Fig. 4.5.2, where the dependence of M on U and the safety

loading A is investigated.

(1) A distribution-free rule As was presented for the case of reserve

funds in Section 4.3, in considering the net retention M it is some-
times useful to have a rapid rule for evaluation of the order of
magnitude of this quantity. Such a rule is readily obtained by apply-
ing the approximation (4.3.5) for (4.3.2), omitting the correction
term inel uding rx

and solving
(4.5.1)

This equation could be handled like (4.3.6) or (4.3.7) and a figure

like Fig. 4.3.2 could be given. However, a slightly different technique
will be applied, which is sometimes useful when the number of
variables involved is large. Instead of absolute variables P, U and
M their ratios
w=M/U and u=U/P (4.5.2)
4.5 RULES FOR THE GREATEST RETENTION 149

will be employed. Then (4.5.1) can be written in the form

W=~(~+U+2A)'
K Y u
(4.5.3)

where
(4.5.4)
The ratio w is a hyperbolic function of the solvency ratio u as
plotted in Fig. 4.5.3.
If the coefficient f3 is positive, then the curve has a minimum
u1 = J f3
w =-~~-
J
2(A + f3)
(4.5.5)
1 K2/
If f3 is negative, then the curve is increasing for u > 0 and w is negative
for

u < - A + YO"q' (4.5.6)

w=MIU

0.15

0.10

0.05

0.1 0.2 0.3 0.4 0.5

u=UIP
Figure 4.5.3 The relative amount of net retention w = MjU as a function of
the solvency ratio u according to equation (4.5.3).
ISO APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

This means that owing to the structure variation, introduced into

the equations by means of a q' no reinsurance arrangement can
stabilize the business unless the solvency ratio u exceeds the limit
provided in (4.5.6).

(g) Rules of thumb are sometimes presented, according to which

the retention M is supposed to be some particular ('empirical' or
traditional) quota of the premium income P or of the reserves U,
e.g. M = q' P or M = q" U. One rule of this type is readily obtained
from (4.5.5) written as follows

M ~
2()' + J().2 - i( 2)) U.
2 2 q (4.5.7a)
Ky
In the special case, when no variation of the basic probabilities
exists (or it is omitted or included in ).), i.e. a q = 0, we have
4
M=~).U~)'U (for 8 = 0.001) (4.5.7b)
Ky
This formula is, in fact, the same as (4.3.8). It leads to a rule ofthumb
which is often used in practice by insurers: that the net retention
should be a certain percentage of the reserves U which the company
is willing to lose for covering losses during a year. If ). is taken to
be e.g. 5%, M is 0.05U. This estimate is based, however, on such a
weak premise that it may not be very useful except in a few special
cases. Neither does it offer any noticeable simplification compared
with(4.5.1).
Another rule, according to which M is related to the premium
income, could be obtained from the latter part of (4.3.8).

Exercise 4.5.1 Prove that the minimum risk reserve U defined by

(4.1.8) and assuming a q = 0 has a derivative

U'(M) = [JMyJn
(a (M))
2
- ).nJ(1 - S(M)). (4.5.8)

Exercise 4.5.2 Find conditions according to which the derivative

(4.5.8) has a negative value at M = O.

Exercise 4.5.3 Examine the shape of the function U(M) defined

in exercise 4.5.1. When does this function have a minimum for a
4.5 RULES FOR THE GREATEST RETENTION 151

finite M? Note that the type of the solution depends on the sign of the
derivative at the origin (see exercise 4.5.2).

Exercise 4.5.4 Suppose that the standard data (other than M) and
distributions given in item 4.2(b) are valid and U = 10 million.
What should the maximum net retention M be according to the
basic equation (4.1.7)? Calculate also an approximate value by
means of (4.5.1) taking P = £65 million and K = 0.6.

Exercise 4.5.5 The insurer considered in exercise 4.5.4 allocates

from its freely disposable profit a small increment f:..U for its current
risk reserve U and increases the net retention M by an amount f:..M
according to the basic equation (4.1.7). The premium income net
of reinsurance P increases then by f:..P, which includes a safety
loading (corresponding to the profit, solvency and expenses of
the reinsurer) Ar = 0.1. What is the rate of return ir of the amount
f:..U, i.e. the expected increment of the cedent's profit margin divided
by f:..U? The formulae derived for exercise 4.5.1 may be used.

Exercise 4.5.6 Find, applying the idea presented in exercise 4.5.5,

an upper limit for the risk reserve U if it is required that the rate of
return ir must be at least i o ' Evaluate a numerical value for it in the
case treated in exercise 4.5.5 for io = 0.1.

Exercise 4.5.7 A new policy is added to a life insurance portfolio.

Annual rate of death for it is q and the insurer will accept amount
M of it for net retention. What should M be if the ruin probability
e calculated according to (4.1.8) were not changed? The new policy
has the same safety loading ), as for the average of the portfolio.
Derive for M a simple rule, if q is small, the expected number of
claims n large and (Jq = 0 (M is to be expressed as a function of the
variables A, n, U, P and the moments about zero of the claim size
d.f. S).

Exercise 4.5.8 Let the claim size dJ. be, in suitable monetary units,

O.8Z for 0:::;: Z:::;: I

S(Z) = { 1- 0.2Z-2 for Z > 1.

What should the maximum net retention M be, when excess of loss
152 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

treaty is applied and n = 100, U = 20, A = 0.05 and E= 0.01 ? Use

the short version (4.1.8) ofthe basic equation.

4.6 The case of several Ms

(a) The problem The basic equation in its short form (4.1.8) will
now be utilized to deal with the case when the portfolio is subdivided
into independent sections indexed by j = 1, 2, ... , r, each of which
has its own claim size dJ. Sj' safety loading Aj > 0, expected number
of claims n.,
J
net retention M.J and structure variable qJ.. First the
moments ak about zero are defined separately for each section. They
depend on the retentions M j (see (3.6.3) and (3.6.8)) and can be
denoted by akiM/ For brevity, let mj = miM) = a1iM/
The problem is to determine the Mjs so that:
(i) the expected amount of profit as a function of the Mjs
f(Ml' M 2, ... , Mr) = I.Ajnjmj' (4.6.1)
is maximized; and
(ii) the basic equation (4.1.8) is satisfied for the whole portfolio:
(4.6.2)

where (see (3.7.9))

ax = aX(M,···M.l = J( ~(np2iMj) +n;mi M / a ;)).

(b) Solution This bound extremal problem is solved by the use of

the so-called Lagrange method by introducing a function
F=f- pQ,
where p is an auxiliary variable.
For a real solution some assumptions concerning the portfolio
and reinsurance are needed. As an example the problem will be
dealt with by assuming excess of loss reinsurance. To facilitate the
calculations it is further assumed that Aj is independent of the reten-
tion Mj" The existence of continuous derivatives S~ can also be
assumed without any essential restriction, since the possible step
points of Sj can be closely approximated by a continuous differenti-
able segment of curve.
4.6 THE CASE OF SEVERAL Ms 153

By using these assumptions, we have

a'k/Mj) = a~J f: j
ZkS'i Z ) dZ + (1 - SiM))M; ]

= kMr 1(1- Si M)),

The extremal values can be found among the joint zero points of
Q and of these derivatives. Putting the factors in braces equal to
zero it follows that
O"x(p-IL 2
M.} = A.. - n.m.O" .. (4.6.4)
yp } }} q}

Solving the Mjs from these equations and substituting into (4.6.2),
p can be determined.
Equations (4.6.3) give only the necessary conditions for solution.
In actual cases, where the data involved are known, it has to be
investigated whether solutions exist which among other things
depend on the values of U. A further step is finding a numerical
solution when the distributions and data are given. This may
give rise to considerable problems, because the variables M j are
buried in the moment expressions.
The problem concentrates on the search for variables which give
an absolute maximum for the profit function f on the surface
Q = O. Note that also one or more of the factors I - S(M) in (4.6.3)
may be 0, which means that the section in question needs no rein-
surance and the consideration is to be limited to the remaining
part of the portfolio.

(c) The Poisson case It is of interest to observe that the expression

by which Aj is multiplied in (4.6.4) does not depend on the sectionj,
i.e. it is the same for all of them. Therefore in the particular Poisson
case where 0" qj = 0 for every j, the following theorem can be derived
on the assumptions mentioned at the beginning of this section.
The limits of retention M j have to be chosen proportional to the
corresponding safety loadings Aj"
154 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

(d) General discussion It was assumed above (when (4.1.8) was

applied) that the normal approximation is justified, but it is im-
portant to note that the normal approximation has not been
assumed for each component j individually. It thus follows that the
theorem is certainly valid as long as the normal approximation is
applicable to the whole business: in view of the central limit theorem
this is a less restrictive condition than if the corresponding assump-
tions 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 Aj is (or, what is equivalent, the greater the expected pro-
fitability of a class 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 re-
insurance 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 Ai" As a special case it follows that the proper
course is to make the retentions M j equal, if all the Aj s are equal and
the second terms in (4.6.4) due to the variation of the basic pro-
babilities are the same order of magnitude (or can be omitted).

4.7 Excess of loss reinsurance premium

(a) The formula for the excess of loss reinsurance premium is one
of the direct applications of the compound Poisson function. The
treaty is mostly formulated by extending the presentation in
Section 3.6.2, so that the reinsurer pays the share Zre = Ztot - A of
each claim Ztot which exceeds a limit A but no more than B - A,
as seen in Fig. 4.7.1.
Then the reinsurance net premium is, according to the general
formula (4.1.1) for the risk premium,

P XIL(A, B) = nE(Zre) = n[ f: (Z - A) dS(Z) + (B - A)(l - S(B» J

(4.7.1a)

where n is the expected number of all claims in the whole portfolio

concerned, including also claims for which Zre = o.
In the particular case where B = 00, (4.7.l) can be written in the
 4.7 EXCESS OF LOSS REINSURANCE PREMIUM 155

5' rZ)

I I
I I
I
I
I

A z
Figure 4.7.1 Excess of loss treaty. The total amount of a claim Z,o, exceeding
the limit (net retention) A is divided between the cedent and the reinsurer:
Z,o, = Z + Zre'

form (see 3.6.3)

PX/L(A, (0) = n[a1(00) - a1(A)] (4.7.lb)
and hence, regarding Zre(A, B) = Zre(A, (0) - Zre(B, 00)
Px/£(A, B) = n[a1(B) - a1(A)]. (4.7.2)
If the moments are available in a tabular form, such as in Table
3.6.1, then P X/L is readily obtained. For example, let n = 1000,
A = £10 6 and B = £10 7 ; then according to the table P X/L = 1000 x
(8.888 - 7.302) x 10 3 = £1.586 X 10 6 .

(b) Variance It is well known in practice that the application

of the simple formulae just given is rather vulnerable to the in-
accuracy ofthe tail of the claim size distribution S, which is very often
present due to the fact that the experience base is scarce. To com-
pensate for the uncertainty (and the greater volatility), safety
loadings are included in the reinsurance premium which may be
quite substantial and can fundamentally affect the consideration
of the appropriateness of this treaty type. For example, formulae
are often suggested where net premium is loaded by a safety margin
which is proportional to the standard deviation or variance of the
risk involved. The variance of the claims on the reinsurer's liability
can be expressed in terms of the second moments as follows (see
exercise 4.7.1).
a-;e = n[a2 (B) - a2 (A)] - 2AP X/L + PilL (J'~ • (4.7.3)
156 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

The premium is also very sensitive even to slight inflation, as is

demonstrated in exercise 4.7.3.
In this example, and assuming (J q = 0.05, the value
(J;e = 1000 x (1.338 104 - 2.353 X 10 3 ) x 10 6
X

- 2 X 106 x 1.586 X 10 6 + 1.586 2 X 0.05 2

= £10 12 x 7.86;
(Jre = £10 6 x 2.80 is obtained from Table 3.6.1.

Exercise 4.7.1 Prove (4.7.3).

Exercise 4.7.2 The upper tail of the claim size is supposed to follow
the Pareto dJ.
S(Z) = I - b(Zo/ZY
Calculate the excess of loss premium (4.7.2) for Zo < A < B.

Exercise 4.7.3 The claim size distribution S(Z) can be approximated

by 0.9Z for 0 ~ Z ~ I and by the Pareto distribution I - O.1Z- 3
for Z ~ 1 where the unit of the monetary variables is £10 5 • The
limits of the excess of loss treaty are A = 2 and B = 5. Calculate
the net premium P XjL as a percentage of the total premium P.
Suppose that due to inflation the sizes of claims are uniformly
increased by 10%, but the premium is not changed. What is the
expected loss of the reinsurer?

Exercise 4.7.4 The claim size dJ. is exponential S(Z) = I - e- cz

and excess of loss reinsurance is applied with the retention limit
M. What is the dJ. of the claim size on the reinsurer's liability?
What is the variance of the aggregate claims of the reinsurer's share?

4.8 Application to stop loss reinsurance

(a) General formula In stop loss reinsurance the reinsurer pays

that part of the total amount of claims X which exceeds a certain
amount, say A. The reinsurer's liability is often limited to an amount
B - A so that the payment is no more than this if the total claim
X exceeds B. Furthermore it is common for the reinsurer's
liability to be limited to a certain share (1 - c) of the excess X - A,
4.8 APPLICA nON TO STOP LOSS REINSURANCE 157

the remaining share c being met by the cedent. The reinsurer's

share of the claims may be summarized as follows

X ={~l - c)(X - A)
when X ~ A
r• when A <X <B
(1- c)(B - A) when X ~ B,
where X is the total claim.
The reinsurance risk premium PsL(A, B) is easily expressed as
follows

PsL(A, B)/(l - c) = f~ (X - A) dF(X) + (B - A) f:dF(X), (4.8.1a)

where F denotes again the distribution function of the aggregate

claims. By integrating by parts the above expression can be written as

(B - A)F(B) - f~ F(X) dX + (B - A) - (B - A)F(B),

f:
so that the net premium becomes

PsL(A, B) = (1 - c) (1 - F(X» dX. (4.8.1b)

(b) Discussion Like the excess of loss premium, the stop loss pre-
mium is sensitive to uncertainties of the tail of the claim size dis-
tribution as well as to inflation. In addition it is sensitive to in-
accuracies and to the structure variations of the number of claims
n. Table 4.8.1 demonstrates how the stop loss premium depends
on the variation of the basic probabilities indicated by O"q and Yq
(see Section 2.7) as well as on other environmental aspects.
It is important to notice that not only the short-term oscillation
of the basic probabilities (assumed when the mixed compound
Poisson process was defined in item 2.7(c») but also the long-term
cycles and trends mentioned in item 2.7(b) (consideration of which
is left until Section 6.1) very strongly affect the stop loss premium.
It can be assumed that allowance will be made for trends and that
the parameters, especially n, will be adjusted accordingly. Even if
the long-term cycles, e.g. when ruin probabilities are evaluated,
are in general intricate to handle, they can be considered in connec-
tion with the rating problem as a random fluctuation of the same
158 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

character as the short-term oscillation. Obviously a convenient

method is to choose the characteristics aq and Yq so large as to
include the cycles in their range. Clearly this implies, however, that
the reinsurance treaty extends over a reasonably long time. Without
special precautions the premium calculated in this way may be
insufficient during that half of the cycle during which the claim
frequencies are high, but this can be expected to be compensated
during the other half of the wave if the contract is not discontinued
before.

(c) NP formula If the compound function F can be approximated

by the N P formula, the stop loss premium (4.8.l) can be expressed
in terms of the normal d.f. N and its first derivative N'. In the simple
case, when the reinsurer pays the whole excess X - A, the following
expression is obtained

PSL(A) = PsL(A, CX)) = (P - A)(l - N(YA)) + a x (1 + Yxy)6)N'(YA) '

(4.8.2)

where the total risk premium P is defined by (4.1.1), ax and Yx

by (3.3.7) and YA is related to A via the N P transformation, i.e.
(see (3.11.14) and (3.11.l4a))

(see exercise 4.8.1).

Then the premium in the general case is
(4.8.3)
Table 4.8.1 exhibits some examples of the stop loss risk premiums.
The first group in the table demonstrates the effect of the limits
A and B. The other arguments are fixed.
The second group illustrates how the heterogeneity of the in-
dividual risk sums affects the stop loss rate. The risk indexes r 2
and r 3 are chosen from Table 3.6.l, assuming that the portfolio is
first protected by an excess of loss treaty having M = 1, 10 and 100
in £ millions respectively and the retained business is then subject
to the stop loss cover exemplified in the table.
The third group of examples shows how the rates depend on the
size of the portfolio, letting n vary and keeping the other arguments
fixed. As expected, the rates (calculated as percentages) needed to
4.9 AN APPLICATION TO INSURANCE STATISTICS 159

Table 4.8.1 Examples of stop loss risk premiums according to (4.8.2).

m A B P(A) P(B) P(A, B)

n .-!:.L
£i000 '2 1000
(J
q
P P P P P
1000 6.00 20.0 0.7 0.04 0.30 1.00 1.40 0.0587 0.0004 0.0583
1000 6.00 20.0 0.7 0.04 0.30 1.10 1.50 0.0231 0.0001 0.0230
1000 6.00 20.0 0.7 0.04 0.30 1.20 1.60 0.0072 0.0000 0.0072

5000 7.30 44.0 4.5 0.04 0.30 1.25 1.75 0.0005 0.0000 0.0005
5000 8.90 169.0 122.7 0.04 0.30 1.25 1.75 0.0134 0.0001 0.0133
5000 9.60 681.0 4314.0 0.04 0.30 1.25 1.75 0.1014 0.0390 0.0624

100 7.30 44.0 4.5 0.04 0.30 1.25 1.75 0.1868 0.0836 0.1032
300 7.30 44.0 4.5 0.04 0.30 1.25 1.75 0.0721 0.0118 0.0603
1000 7.30 44.0 4.5 0.04 0.30 1.25 1.75 0.0168 0.0002 0.0166
3000 7.30 44.0 4.5 0.04 0.30 1.25 1.75 0.0020 0.0000 0.0020

3000 7.30 44.0 4.5 0.00 0.00 1.25 1.75 0.0015 0.0000 0.0015
3000 7.30 44.0 4.5 0.05 0.50 1.25 1.75 0.0023 0.0000 0.0023
3000 7.30 44.0 4.5 0.10 1.00 1.25 1.75 0.0062 0.0000 0.0062
3000 7.30 44.0 4.5 0.20 1.50 1.25 1.75 0.0275 0.0014 0.0261

cover the risk of a small collective are substantially higher than

those necessary for a large collective.
In the last group of examples the structure parameters uq and Yq
vary. It is seen also that the structure function fundamentally
affects the rates.
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).

Exercise 4.8.1 Prove (4.8.2).

4.9 An application tb insurance statistics

(a) The problem 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 frequencies as risk premiums. In this connection the problem
160 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

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 £800 000. The total amount of the insurance in force in
this group is £700 000 000, this amount being the sum of the average
sums of each year. The so-called burning cost is now obtained as
f = 800000/700000000 = 1.14 per thousand and the problem is to
estimate the accuracy ofthis quantity.

(b) Confidence limits Assuming the mixed Poisson process the

solution is a direct application of the NP formula (4.l.7). The
expected total amount of claims can be interpreted as a risk premium
P, and X - P, the deviation of the observed claim amount X from
its expectation, as the error due to random fluctuation. Then the
relative error is I1f I f = (X - P)I P. Hence with probability 1 - c:
X I - P ::S;_::S;_2
I1f X __ - P
-
P f P
where
F(X 2) = 1 - c:/2 ,
F(X I )=c:/2.
The true value of P is, of course, unknown and therefore estimates
are to be used, derived from past experience or simply by using the
observed amount X as P.
First let c:/2 = N(y l ) and 1 - c:/2 = N(yz)' where N is the normal
dJ. and YI = - yz· Further, let XI and X z be the corresponding
NP corrected variables x=v/y) according to (3.11.16). Then
(i=I,2),
and

(4.9.l)

Here (see (3.3.7))

(4.9.2)
4.9 AN APPLICATION TO INSURANCE STATISTICS 161

and for the formulae (3.11.16) the skewness is the same as given in
(3.3.7) (note that a linear transformation, here X --+ X/P, does not
affect the skewness).
The computation of the indices r 2 and r3 and the evaluation
for (J q and yq can 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 class in question.
As an example let n = 100, r2 = 10, r3 = 200, (Jq = 0.1 and Yq = O.
Then
(Jx/P = j(10/100 + 0.1 2 ) = 0.33,
and
Yx = (200/100 2 + 3 x 10 X 0.1 2 /100 + 0)/(0.3W = 0.64,
and for e = 0.05, y = ± 1.96 according to the nomogram of Fig.
3.11.2 Xl = - 1.7 and x 2 = 2.3. Hence
- 1.7 x 0.33 ~ N If ~ 2.3 x 0.33,
or
- 0.6 ~ N / f ~ 0.8.

(c) The normal approximation can replace the NP formula if only

the order of magnitude is needed, or when the data base is large. Then
(4.9.3)
In this example INIfI ~ 0.7. As usual the normal approximation
has underestimated the risk of large positive deviations. Because
the risk groups in statistical considerations are often rather limited,
the applicability of the normal approximation and even the NP
formula may be uncertain. However, these approximations give
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.
Note that if the statistical material under consideration is derived
from several years' data, then the structure variation of consecutive
years can be expected to offset each other and (J q can be assumed to be
approximately O. On the other hand, serious bias can be expected
162 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

to arise due to trends and cycles (see item 2.7(b)), which should
be estimated in one way or another (see McGuinness, 1970).

(d) A straightforward estimation It may be further noted that,

e.g. for rapid checking ofthe order of magnitude the same estimation
problem can easily be solved by dividing the observation period into
sub-periods, preferably years, and by calculating the quotient Iv
separately for each year v. Then the estimation

Iii ~ j J(~{~~J;)2} (4.9.4)

holds, where J = 'LI)N, N is the number of sub-periods and 8

the confidence level. 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 (4.9.2) 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 (4.9.4) does not exist. Equation (4.9.2) 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 4.9.1 It is known that for certain fire risks I ~ 0.1 %

and the number of claims per annum is about 1000. How many
years of statistics are needed to estimate I with an accuracy of 20%
on 1O~~ confidence level? The risk index r2 is estimated to be 100
and (J q = o. The normal approximation can be used.

4.10 Experience rating, credibility theory

(a) Profit return Bonus systems are applied in connection with

some insurance classes. 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 safety loaded risk premium
is (l + A)P and the actual amount of claims related to this particular
collective (treaty or policy) in a year is X, a bonus or 'profit return'
can be agreed, e.g. according to the formula
G = k[(l + ),)P - X]+. (4.10.1)
4.10 EXPERIENCE RATING, CREDIBILITY THEORY 163

Assuming that on average the safety loading 2P = 2E(X) should cover

the profit return and denoting Xo = (1 + 2)E(X)

E(G) = k(1 + 2)E(X)F(Xo) - k f:o X dF(X) ~ 2E(X).

Partial integration gives

f
k 0xo F(X) dX ~ AE(X). (4.10.2)

This equation defines the constant k. Ammeter (1963) has studied

systems where the variance of G is minimized or where the net
premium P - G is as correct as possible even if the original rating
of P is erroneous.

(b) Example IfF can be approximated by the NP dJ. Ny (3.1l.14a),

the formula (4.10.2) can be written in the form

k Il +A)P Ny (X ax P) dX ~ AP,

which gives, after some simplifications

k ~ Xo Ij:°oo Ny(x) dx,

where x = (X - P)/a x and Xo = AP/ax = A/J(r 2 /n + a;) (see (3.3.7».

If, for example, A = 0.1, r 2 = 25, r3 = 1000, a q = 0.2, Yq = 0.5
and n = 100, then k ~ 0.85. This means that the insurer can return
at most 85% of the profit. The remainder of the profit is needed to
cover the risk of excessive losses X > P.
If instead of the N P function the normal approximation had been
used, the k value would be 92% in the above example.

(c) Discussion The structure of the contract between the policy-

holder and the insurer is, in essence, the same as the conventional
stop loss reinsurance. The policyholder bears the small fluctuations
of claims partially on his own account, whereas the risk of large
losses is insured as has been presented in Section 4.8.
It should be noted that it is unusual in practice for E(X) and hence
also A to be known in advance for each risk collective under
164 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

consideration. In fact the situation is precisely the same as described

in Section 2.10, where the concept 'risk exposure variation inside the
portfolio' was introduced. Hence the standard deviation O"q and the
skewness Yq should now be interpreted as related to the risk variation
inside that group of risks from which the particular unit concerned
is selected. The initial premium P is the average premium derived
for the whole group and the profit return G is designed to compensate
for the deviations of the exposure rates of the individual units from
this joint average level.

(d) 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 reason-
able premiums by starting from a hypothetical value Po and sub-
sequently correcting it by using the actual claims experience by
means of some agreed rule, for example similar to that given in
(4.10.3). When defining what are considered to be 'reasonable pre-
miums', 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 require-
ment that the premium should not show too much random fluctua-
tion. An example of this kind of arrangement is set out in the follow-
ing system of 'sliding premiums', which has been extensively studied
in the USA under the title '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 ofthe general tariffs, 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

(4.10.3)

where Xo is the total amount of claims in this collective in the preced-

ing year. The 'breaking constant' Z, called 'credibility', is chosen from
the interval:
0< Z ~ I,
4.l0 EXPERIENCE RATING, CREDIBILITY THEORY 165

and 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 - 6, result in a
change in the premium P in excess of 100p% calculated from E(X).
Expressed in symbols this is the case if the constant Z satisfies the
condition
Z~X ~pE(X), (4.10.4)
where ~x is obtained from
F(E(X) + ~X) - F(E(X) - ~X) = 1- 6, (4.10.5)
which assumes that the dJ. F of X is known or presumed. Then the
absolute value of the deviation AX = X - E(X) can be larger than
~X only with probability 6.
lf it can be assumed that the N P approximation gives a satisfactory
approximation for F, then item 4.9(b) is applicable because
N I f is in fact AXIE(X) and, confining the analysis to the upwards
jumps of X,
Zx,J(r 21n + O'~) = p,
where according to (3.11.6)
x, = Y, + iYx(Y; - 1),
and Y. is the root of 1 - 6 = N(y.) and r2 is again the risk index (3.3.8).
Hence we have

(4.10.6a)

Because for minor risk collectives, for which the experience rating
is usually applied, the variation of the basic probabilities may be
less significant than the other fluctuation (see Table 3.3.2), the
formula can be simplified by putting 0' q = O. Then

Z=- - pJn (4.1O.6b)

x. rz
The expected number of claims n which makes Z = 1 i.e.

(4.10.7)

is of special interest.
166 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

Table 4.10.1 Values of nofor full credibility (constant claim size).

p 10% 5% 1%

0.01 27057 38416 66347

0.05 1082 1537 2654
0.1 271 384 663
0.2 68 96 166

Following the terminology of the American credibility theory it

is said that if Z = 1, there isfull credibility.
In the special case where the risk sums are all equal or, what is
equivalent, if only the number of claims is recorded for calculating
the frequency of the claims, then r2 = 1 and the values of no which
are large enough for full credibility are immediately obtained by
means of tables of the normal distribution, which is used instead
of N P formula, as in Table 4.10.1.
In most practical cases the risk sums are not equal and hence
r2 is not 1. The variation in the value of this quantity depends signi-
ficantly on the degree of heterogeneity of the risk sums and con-
sequently the limit of full credibility can be considerably larger than
is given in Table 4.10.1. The values of r 2 may often be of the order
of 5 to 10, but in cases where large risk sums can occur the values
can be much larger, as is seen for example in Table 3.6.1.
If the expected number of claims n is smaller than the value
obtained from (4.10.7) then the constant Z has values smaller than
1 and the term partial credibility is used. From (4.10.6b) and (4.10.7),
one of the well-known formulae of credibility theory can be
immediately obtained by eliminating the coefficient ofJn in (4.1 0.6b)
Z = J(n/n o) (4.10.8)
In the foregoing it was assumed that the N P approximation
could be used. However, owing to the small size of the risk collective
which often arises in cases subject to experience rating or to credi-
bility theory, the N P formula can be doubtful, even if very small
values of e are not needed for which the accuracy of the formula is
most unsatisfactory. The uncertainty can of course be avoided by
calculating the quantity x. by using some other method of computa-
tion. A drawback is, however, that x. may depend on n. The
4.10 EXPERIENCE RATING, CREDIBILITY THEOR Y 167

experience of American actuaries suggests nevertheless that even the

normal approximation gives values which are satisfactory in
practical work.

(e) Limit value Applying (4.10.3) for a sequence of t years and

developing the algorithm into a series it follows that
PI=ZXI _ 1 +(1-Z)PI _ 1
t
(4.10.9)
= Z I (1- Z)i-1XI _i + (1- ZYPo'
i= 1

If the expected value /1 = E(X) is assumed to be equal for all i

values, then
I

E(PI ) = Z I (1- Z)i-l/1 + (1- ZYP o

i= 1

= [1 - (1 - ZYJ/1 + (1 - ZYPo' (4.10.10)

which tends to /1 as t ~ 00. Thus, in the long run PI is expected
to tend to the theoretically correct unknown mean value /1. Hence
the formula fulfils the requirement of fairness. The coefficient Z
regulates the fluctuation of the sliding premium rate.
Owing to the fact that PI depends on the claims amounts X t - i
of the preceding years by means of weights having the elapsed time
i in exponents, the algorithm (4.10.9) and the formula (4.10.3) are
sometimes called exponential.

(f) The Bayesian approach Another way to build up experience

rating theory is to make use of the theory of variation of risk inside
the portfolio as discussed in Section 2.10. It is assumed that the
risk variation variable q has dJ. H inside some particular risk
collective, e.g. in a group of a certain type of industrial plants
(see item 2.10(a)). Note that q is now not the time-dependent
structure variable that is employed in most parts of this book,
but instead a variable describing the heterogeneity of the risk
intensities in the collective from which the risk unit under considera-
tion is randomly drawn. It is supposed that the function H is known,
but the value of the risk parameter q for any individual risk unit
is unknown. Then the dJ. F(X ;q) and its density f(X ;q) are both
functions of the parameter q, as illustrated in Fig. 4.1 0.1.
The correct premium for any risk unit depends on unknown
 168 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

((X,qi

Figure 4.10.1 A sample of f(X; q) for different parameter values q and a

cluster of observed X values.

q as follows

P(q) = f: Xf(X; q) dX. (4.10.11)

On the other hand the average premium of the whole collective is

P= t'Xl too Xf(X;q)dXdH(q) = too P(q)dH(q). (4.10.12)

Now let us assume that for the risk unit concerned a sequence of
total claims Xl' ... , X t for t years is observed as illustrated in Fig.
4.10.1. It can be expected that they are clustered in a more or less
narrow area on the X -axis. Heuristically it is easily conceived that
this kind of accrued experience makes it possible to conclude what
is the order of magnitude of the unknown parameter q. Obviously
those values are most probable which correspond to the curves
having their modes in just that area where the X values are clustered.
The well-known Bayesian rule enables us to find an expression for
the probability density of the unknown parameter q by the condition
that X has the observed sequence of values. Then an obvious idea
is to amend the premium formula (4.10.12) by weighting P(q) by
these conditional probabilities as follows
4.10 EXPERIENCE RATING, CREDIBILITY THEOR Y 169

P", ~ E(X", IX" ... X,) ~ f 00

o
p(q)
lJ~ ,
[l!(X p q) dH(q)
t 1.
[l !(X i , q) d H(q)
o i= 1

(4.10.13)
The expression in brackets is just the conditional Bayesian pro-
bability density.
If the density function! is known, as assumed, then in principle
this expression gives an estimate for the premium for the next year.
The expression (4.10.13) has been widely examined and it is
proved inter alia that in the Polya case where H is the incomplete
gamma function and F the mixed Poisson (and, as Jewell (1976)
has shown, also for some other function mixtures) the formula
can be reduced to the very simple form
(4.10.14)
where

(4.10.15)

is the mean value of the observed data and

t
Z =--. (4.10.16)
t t+A
Here A is a constant which is independent of t but dependent on
the size of the collective from which the X values are drawn. This
is formally similar to the exponential formula (4.10.3), but now
the credibility factor Z depends on time t. When t increases, Z
tends to 1, which means that more and more emphasis is given to
the mean value of the observed data X.
There is an important difference between the formula (4.10.3)
and the credibility rule (4.10.14). In the former the effect of each
Xi gradually vanishes, whereas in the latter all X s have an effect with
same weights.
It is of interest to know that a formula of just the type of (4.10.14)
was found, largely by a trial and error method, in the USA some
60 years ago and has been in use since then. The theoretical derivation
was found much later (Bailey, 1945, and Mayerson, 1965). The
presentation given above follows Jewell (1976).
170 APPLICATIONS RELATED TO ONE-YEAR TIME-SPAN

(g) The least squares approach A third way of building credibility

theory has been developed by Biihlmann (1967) and others. The
idea is to construct from the observed data Xi some predicting
function g(X l ' ... , Xt) which, as well as possible, can give a forecast
for the amount of claims X t + 1 ofthe next year. The simplest approach
is to choose g linear
t

gt+ 1 = ao + L aiX i , (4.10.17)

i= 1

where the coefficients ao' at' ... are to be determined by means of

the least-squares principle, i.e. minimizing the expression
(4.10.18)
It can be proved that, subject to some fairly general conditions,
a formula can be found which is of just the same shape as (4.10.14)
and (4.10.16) even if the constant A is different.
A good survey of credibility theory can be found, for example,
in Jewell (1980).

Exercise 4.10.1 Calculate the variance u~ of the sliding premium

P t defined by (4.10.9) assuming that the variables X t are mutually
independent and have an equal variance u 2 • Furthermore, prove
that up is finite when t ~ 00 even in the case where the individual
variances are different but have a finite upper limit.

Exercise 4.10.2 Let the safety loaded risk premium B of a risk

unit be> E(X). A bonus G = k(B - X)+ will be given. What is the
largest value of k which does not render a systematic loss for the
insurer? Express this limit value in terms of Band F(X) and prove
that it is ~ 1.

Exercise 4.10.3 A commercial firm has insured 1000 lorries by

a collective treaty. The premium without loadings is £300 per
vehicle and will be adjusted in accordance with the credibility
formula (4.10.3). The total sum of claims in the first year is £180000
and the number of claims 200. What should the premium be for
the next year if the credibility coefficient Z is determined by using
values rz = 10, uq = 0, p = 0.1 and xe = 2?
 CHAPTER 5

Variance as a measure of stability

5.1 Optimum form of reinsurance

(a) Problem setting In Chapter 4, the ruin probability was the

main criterion used for the solvency of insurers. The implementation
was based on the calculation of the variation range of the aggregate
claims, which mostly lead to use of the normal, the N P or another
approximation. Another way is to use directly the variance

var(X):::::: V(X):::::: ai = t
ro
(X - E(X))2 dF(X) , (5.1.1)

of the aggregate claims as a measure of stability. In the case where

the normal approximation is applicable, the minimum initial risk
reserve is by a given safety level I - G proportional to the standard
deviation ax' as can be seen from (4.1.8). Hence the smaller the
variance is the safer is the position, assuming that the other quanti-
ties involved are not changed, at least not enough to offset the
effect. If several policy options are available, for example different
reinsurance arrangements, the best of them is, from the point of
view of solvency, that which gives the smallest variance (providing,
of course, that it is not realizable too much at the expense of some
other aspects, e.g. weakened safety loading or decreased business
volume). It is natural to expect that this conclusion will also be valid
more generally than merely for the normal approximation. That
is the issue in this chapter. The problem of finding an optimal re-
insurance arrangement is transformed to the problem of minimiza-
tion of the variance VeX).
The variance has been explicitly calculated in Section 3.3 for
the compound Poisson variable. However, in the present approach
it is not necessary to make the assumptions that the compound
172 VARIANCE AS A MEASURE OF STABILITY

Poisson or normal properties apply, and the way is open to some

general risk-theoretical questions of a rather broad nature.

(b) The optimality of stop loss reinsurance Consider the following

problem. A company wants to find a reinsurance policy which
gives the smallest variance for the retained business provided the
reinsurance risk premium P (without any safety loadings) is fixed.
The total amount of claims resulting from the claims process during
a time period (e.g. one year) under consideration is assumed to be
given by the random variable Xtot ' Furthermore, it is assumed that
reinsurance is arranged in a way which defines a uniquely determined
claims amount X
O~X~X tot and E(X) = P, (5.1.2)
on the cedent's net retention for each claim process realization. It
is not necessary to specify further the type or details of the reinsur-
ance treaty; for example any of the standard forms dealt with in
Sections 3.6, 3.7 and 3.8 or their mixes are acceptable. The problem
is now to find that reinsurance arrangement which has the smallest
variance V(X) subject to conditions (5.1.2).
It is suggested that the required solution is the stop loss treaty
(see Section 4.8) defined by
X* = min (Xtot' M), (5.1.3)
where the retention limit M is a constant determined by condition
E(X*) = P.
The problem setting is illustrated in Fig. 5.1.1.
For the proof of the assertion it is first noted that, since 0 ~ X ~
Xtot ' it follows from (5.1.3) that the following sets of realizations are
identical
{X + X* > 2M} = {X > M} = {X - X* > O}. (5.1.4)
All such realizations satisfy
X2 - X*2 = (X + X*)(X - X*) ~ 2M(X - X*). (5.1.5)
But this inequality is valid also for the rest of the realizations, because
by (5.1.4) these satisfy the inequalities 0 ~ X + X* ~ 2M and
X - X* ~ 0, or together (X + X*)(X - X*) ~ 2M(X - X*). By taking
 5.1 OPTIMUM FORM OF REINSURANCE 173

50 100
X tot
Figure 5.1.1 The stop loss treaty (X*) and excess of loss treaty (X). The
realizations (amounts retained on the cedent's own account) of X* are on the
straight lines plotted solidly and the realizations of X, shown by circles,
are distributed in the half axis angle 0 ~ X ~ X'o'. Compound Poisson aggregate
claim df n = 20, claim sizes Pareto-distributed, IX = 2, Zo = 1 (see (3.5.20»,
net retention for the excess of loss arrangement M = 5.

expected values in (5.1.5), and recalling the assumption E(X) =

E(X*), it is obtained conclusively
V(X) - V(X*) = E(X2) - E(X*2)
= E(X2 - X*2) ~ 2M E(X - X*) = o. (5.1.6)
Hence the variance is minimized by the stop loss reinsurance.
Obviously there is equality in (5.1.6) only if X = X* (with probability
1).
This result shows that the stop loss reinsurance (5.1.3) is the
optimal solution among all reinsurance formulae in the sense that it
gives for a fixed reinsurance risk premium the smallest variance for
the company's net retention. In other words, if the level V of the
174 VARIANCE AS A MEASURE OF STABILITY

variance of the retained business is fixed, the stop loss reinsurance

leaves the maximum premium income to the ceding company and
thus minimizes the reinsurance risk premium.
The above proof is due to Martti Pesonen (1983).

(c) Loaded reinsurance premium In practice, however, there can

be heavy safety and expense loadings in stop loss premiums, because
this form of reinsurance gives a large relative variance to the reinsur-
er. The problem is therefore modified accordingly. Again let Xtot be
the total amount of claims, X the retained part thereof, and
Xtot - X = Xre the reinsurer's share. Suppose that the reinsurance
premium is defined as
Pre = E(X re ) + f(V(X re )), (5.1.7)
where f is a given 'loading' function. It is necessary to assume only
that f is non-decreasing. The net reinsurance cost is evidently
f(V(X re )) on average. Suppose further that the cedent wishes to
retain a variance of fixed size, say V(X) = V. The problem is to
determine how the form of reinsurance should be chosen, if re-
insurance costs are to be as low as possible; in other words, which
insurance arrangement minimizes V(Xre ) if V(X) is fixed. Now
V(X re ) = V(Xtot ) + V(X) - 2(E(Xtot X) - E(Xtot)E(X)).
From this expression it can be concluded, since V(XtoJ and V(X) are
constants, that the reinsurance cost attains its minimum on average
if the correlation coefficient
E(XtotX) - E(Xtot)E(X)
J[V(XtO\)V]
of the variables X and Xtot reaches its maximum value + 1. As is
well known from probability calculus, this is the case if X =
aXtot' where the positive constant a is defined from V = V(X)
giving

(5.1.8)

Thus a reinsurance of quota share form (see Section 3.6.3) gives the
desired result.
Hence a quite general theorem is proved that if the reinsurance
premium increases with the reinsurer's variance, and is thus of the
5.2 RECIPROCITY OF TWO COMPANIES 175

form (5.1.7), the most inexpensive way to reach a given variance is to

use the reinsurance form (5.1.8).

bxercise 5.1.1 Let Xtot be the aggregate claim without reinsurance

and X and Xre = Xtot - X the shares ofthe cedent and the reinsurer(s)
when some reinsurance treaty is applied (0 ~ X ~ XtoJ In many
reinsurance arrangements, for example in excess of loss and in
surplus treaty, the value of the aggregate claim Xtot does not deter-
mine the values of X and Xre , these variables depending on how
Xtot is composed as a sum of individual claims. Prove that there is,
however, always a function R such that R = R(X tot ) satisfies the
conditions
(i) 0 ~ R ~ Xtot
(ii) E(R) = E(X)
(iii) V(R) ~ V(X)
(iv) V(Rre) ~ V(X re ),
where Rre = Xtot - R.
Hint: choose R = E(X IXtot )' i.e. for every X let R(X) = E(X IXtot = X)
= mean value of the different outcomes which all result in a joint
total aggregate claim.
Note that by replacing X by R one can improve the variance of
both the cedent and the reinsurer without any extra cost to any of
the parties compared with the original rule (E. Pesonen, I 967a).
In this respect, more important results for general application
have been obtained by Martti Pesonen (1983).

5.2 Reciprocity of two companies

(a) The problem to be considered arises when two companies C l

and C z , whose total amounts of claims Xl and X z 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.

(b) The type of risk exchange In order to fulfil requirement (i) it is

simply assumed that the reinsurance premiums are unloaded
176 VARIANCE AS A MEASURE OF STABILITY

risk premiums, i.e. of form (5.1.7) with f = o. If then condition (ii) is

the only remaining decision criterion, it is easily seen that the
exchange should be of quota share type, i.e. the final total amounts
should be
for C I CIX I + (1 - C1)X 1

for C 1 (1 - CI)X I + C1X 1 · (5.2.l)

To prove this suppose that the final arrangement has given the
variance V~ to company C I . This can be written VII + VlI , where
VII eminates from C I 's remaining original business and VlI from
accepted reinsurance, since the amounts XI and Xl were assumed to
be independent of each other. Analogously the final variance of C l
can be written in the form V I2 + V 22 • Suppose now that the reinsur-
ance C I ---> C2 was not of the quota share form. Then without
changing the other variances the variance V I2 of C2 could be reduced
by means of reinsurance (5.1.8). Similarly without changing other
variances the optimal variance V 21 is reached by using a reinsurance
of form (5.1.8), i.e. of the quota share form (5.2.1) as asserted. Then
V~ = ciVI + (1- C2 )2V2
V:=(l-CI)lV I +C~V2· (5.2.2)

(c) Pareto optimal The remaining question is to determine how

the constants c i and c1 should be chosen.
First observe that V~ is, in the C I' c2 plane, constant on the
periphery of an ellipse having midpoint 0, 1 and principal axes
j(v~fVI) and j(V~/V2). For different values of V~ a family EI
of concentric ellipses is obtained. Similarly the condition for
determines another family E2 of concentric ellipses having midpoint
V:
(1,0) and principal axes j(V:fV I ) and j(v:fVJ
Evidently a necessary condition for agreement is that V~ ~ VI
and V: ~ V2 • Geometrically this means that the point (c I ' cz )
should be located in the common area of the two ellipses E'I and
E~ corresponding to the values V~ = VI and V: = V2 (shaded in
Fig. 5.2.1). Through any point P goes one and only one ellipse E I (V~)
of the family EI and another ellipse of the family E 2. Now let the

V:
point P move along EI(V~). Then V~ is preserved unchanged but
is changing all the time according to which of the ellipses E2 is
intercepted until a point is reached in which El(V~) is tangential to
one of the E2 ellipses. Then V:has reached its minimum. So a point
5.2 RECIPROCITY OF TWO COMPANIES 177

J(

c,
J(~)
Figure 5.2.1 The families of level ellipses. E'1 and E~ plotted by solid lines:
the cases V~ = VIand V~ = V 2.

is reached from which it is no longer possible to move to any direc-

tion without worsening the benefit (variance) of C 1 or C 2 • This
reasoning shows that this is the situation always and only in those
points where one of the ellipses of each family is tangential to an
ellipse of the other family inside the shaded area. It is easily seen
that this set of points is the segment of the straight line c 1 + c2 = 1
which goes through the midpoints of E1 and E2 (see exercise 5.2.3).
This kind of point set is called a Pareto optimal, a concept which in
a similar way is important in the theory of games and economic
behaviour.

(d) Reciprocity condition We have now concluded that it is

reasonable to choose c1 and c2 from
c 1 + c2 = 1. (5.2.3)
But there is still a conflicting situation. Company C 1 prefers to go
as near the midpoint (0, 1) as possible; the other company prefers
points near the midpoint (1, 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), a third one:
(iii) The volume of exchange must be balanced.
178 VARIANCE AS A MEASURE OF STABILITY

That is to say, the reinsurance premium from C, -+ C 2 must be

equal to the reinsurance premium from C 2 -+ C" i.e.
(5.2.4)
Then the constants c, and c2 become uniquely determined from
the equations, the solution being
c, =P,/(P, +P 2 )
c 2 = Pz/(P, + P z)·
(e) Multi-unit risk exchange The above consideration can obvious-
ly be extended to the cases where more than two, say n, insurers
are involved. Instead of level eJlipses, ellipsoids in n-dimensional
space are then operated and the optimal points are sought inside
their joint space segments by methods of the theory of multi player
games. This topic is illustrated in the following section, where
formaJly a different type of problem will be treated, but the idea of
minimization of variances is the same as in the risk exchange
problem.

Exercise 5.2.l There are r insurance companies having the same

distribution function of the total amount of claims Xi' the claims of
each company being independent of those of the others. The com-
panies want to find a reciprocal exchange of business so that each
company i pays the amount R(X} of the claims of each other com-
pany 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. Note that the equality of distributions
implies the equality of business volumes and other parameters.

Exercise 5.2.2 A major concept in the classical theory of risk was

the so-called relative mean risk p, which was defined as
p = axlP

ax and P being the standard deviation and risk 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 and
mixed Poisson claim number process?
5.3 EQUITABILITY OF SAFETY LOADINGS 179

(ii) Prove that Po is the minimum for all collectives having the
same P, n and structure function H, if the mixed compound
Poisson distribution is assumed.
(iii) Rewrite the basic equation (4.l.8) making use of p.

Exercise 5.2.3 Prove that the Pareto optimal derived in item

5.2(c) is on the straight line (5.2.3).

5.3 Equitability of safety loadings: a link to theory

of multiplayer games

(a) Safety loading problem Yet another example may be given of

how variances or equivalently standard deviations may be used as a
yardstick in solvency treatments. Even if the details of premium
rating are beyond the scope of this book, a special aspect related to
safety loading will be dealt with in order to illustrate the variance
approach. As has just been stated, and as will be demonstrated
later, when long-term survival conditions are examined, premiums
must necessarily be loaded by safety margins. It was stated in item
4.1(c) that regarding solvency conditions it is relevant that the total
yield accrued from the whole portfolio is large enough. The solvency
aspects do not determine how the loading is to be distributed among
the individual policies or groups of the insured. This is to be decided
on the basis of rules and principles, which are in the realm rather
of general non-life insurance mathematics or business philosophy
than risk theory, and there are numerous partially contradictory
suggestions for the purpose. We pick up only one of them, which
follows ideas orginally presented by Borch (1962).
The issue is the minimum solvency margin U which is needed to
protect a supposed portfolio. As reasoned above it should be
dimensioned according to the range of fluctuations arising in the
risk business concerned. The first term of the basic equation (4.1.7)
gives an approximation for it, i.e. U must be proportional to the
standard deviation of the aggregate claims amount X
U = yo-x = ypJ(r 2 /n + o-~), (5.3.1)
where y is again a safety factor, say 3. In the case of proprietary
companies it is expected that the portfolio pays interest to this
safety capital. In the case of a mutual company, U will perhaps have
to be created by self-financing and maintained from the safety
180 VARIANCE AS A MEASURE OF STABILITY

loadings, a point which may be crucial especially in inflatory en-

vironments. These aspects suggest safety loading A, which gives a
total yield AP proportional to U, or what is approximately the
same, to the standard deviation a x of the aggregate claims.
Another way to reason the same outcome is to suppose a hypo-
thetical situation where a certain group of policy-holders are going
to establish an insurance company and U is the minimum initial
capital to be got by levying it from them in one way or other.

(b) Multiplayer approach We are now going to discuss what kind

of consequences may result if the safety loading is defined propor-
tional to the minimum initial capital (5.3.1), i.e. AP = kU where k
is a proportionality factor. Let us take a simplified example assum-
ing that the portfolio is composed of three groups of the insureds,
which are each of different types but all internally homogeneous.
Group 1 is comprised of small risks, like motor cars, family property
etc. Group 2 contains large risks like industrial fire, marine, aviation
etc. Group 3 risks are characterized by exceptionally great short-
term variation of the basic probabilities, such as for forests or in other
insurance against natural forces. The basic characteristics are given
in Table 5.3.1.
The quantities U j are the risk reserves according to (5.3.1) if each
group separately built an insurance collective. Furthermore,
applying the composition rules (3.7.9) the risk reverse U l23 = 27.8
of the whole united portfolio is obtained. It is considerably less than
the sum of the components (column 7 in Table 5.3.1). Without loss
of generality the proportionality factor k, relating U to A, can be

Table 5.3.1 Example of an internal composition of a portfolio and minimum

solvency margins (5.3.1) of the groups j = I. 2, 3 and combined groups ij. M one-
tary unit is £10 6 , r2j is the risk index (3.3.8), cr qj is the standard deviation (J{
the structure variation, mJ is the mean claim size and Pj = njmi .

2 3 4 5 6 7 8 9 10 II 12 13
j n.
J r2j cr qj mj PJ Uj ij Uij Gij Gi G~, G,V
1 10000 5 0.05 0.0028 28.0 4.6 12 26.8 4.2 2 1.2 0.3
2 5000 150 0.10 0.0088 44.0 26.4 13 8.1 3.2 5.1 7.1 9.2
3 1000 5 0.80 0.0028 2.8 6.7 23 27.2 5.9 3 1.8 0.6

I 16000 74.8 37.7 10.1 10.1 10.1

5.3 EQUIT ABILITY OF SAFETY LOADINGS 181

taken to be 1. Then the difference can be called 'the gain G' obtained
when the groups are united as one collective:
G I23 = VI + V z + V3 - V I23 = 10.1.
Now the problem is how to divide reasonably this gain among the
groups, i.e. to find amounts Gi satisfying the condition
3
L G; = G123 · (5.3.2)
i= I

A natural aspect is to expect that the shares G; should be at least

equal to what can be gained if any two of the groups establish
together a collective (excluding the third). For example, if groups
1 and 2 join, their reserve would be, by again using (5.3.1), V 12 = 26.8
and the gain G I2 = VI + V 2 - V I2 = 4.2. In a similar way the two
other combinations can be evaluated as given in columns 9 and 10
of Table 5.3.1. An obvious condition for the co-operation of all the
three groups is now
G; + Gj > G;j (i =1= j, i,j = 1,2,3), (5.3.3)
i.e. the gain of each group should be better than what can be achieved
if any two of them joined leaving the third outside. As shown in
column 11 by means of an example there exist shares which satisfy
conditions (5.3.2) and (5.3.3). On the other hand the solution is not
unique; it is still open for further conditions. The situation is
analogous to that found in the previous section in the form of the
Pareto optimal (5.2.9).
The example given above is a typical n players co-operative game
(Borch, 1962). The set of solutions satisfying the above conditions is
called the core of the game and the consideration can be extended to
an arbitrary number of insurers participating in the risk exchange.

(c) Equitabilityaspects An approach sometimes suggested is to

load policies or policy groups by safety loadings proportional to
standard deviation (or variance) of the risk. If the solvency margin
(and safety loading yield proportional to it) were divided according
to the group standard deviations, which by (5.3.1) are proportional
to the risk reserves given in column 7, the gains Gf given in column
12 would result. It is interesting to observe that these shares do not
satisfy conditions (5.3.3). The groups of small risks were overloaded
in benefit of the big risks (group 2).
182 VARIANCE AS A MEASURE OF STABILITY

The situation would be still more blatant if the variance principle

were applied, i.e. the loadings were defined proportional to variances.
This is shown in column 13 of the table.
The above examples were intended, among other things, to warn
of the fact that too straightforward an application of conventional
rating principles can violate the equitable treatment of different
types of policies; in particular, the great mass of small risks may
easily be paying for the risk equalization of big risks.

(d) Company size Another interesting observation is that the

safety loading A = kUjP, according to (5.3.1), is smaller the greater
is the portfolio. This means that in theory a large insurer can apply
a lower safety loading than a small insurer, if the other conditions
are equal.

(e) Utility approach The variance or the standard deviation of the

collective concerned was used as a measure of stability, the
optimization of which was the objective of the reinsurance arrange-
ments. One should appreciate that instead of these quantities other
optimization target functions can also be used. The utilities, which
will be discussed in Section 10.4, are often suggested for the purpose,
and rules for Pareto optimals can be obtained (Borch, 1960, 1961).
Biihlmann and Jewell (1979) have proved that the risk exchange
between insurers can be presented in well-defined forms if it is
assumed that all insurers accept an exponential utility function for
the guidance of their procedure. Martti Pesonen (1983) has extended
the considerations to quite general cases.
 CHAPTER 6

Risk processes with a time-span

of several years

6.1 Claims

(a) Finite time horizon In the preceding chapters the time-span

used was generally limited to one year. This restriction is now
relaxed. In what follows the argument will be extended to an arbit-
rary finite - and in Chapter 9 for infinite - time T.
The financial status of insurers is usually stated at the end of each
calendar (accounting) year, and possibly also at other times, e.g. at
the end of each quarter. Accordingly a discrete treatment of risk and
other business processes is assumed, i.e. the state of the process is
observed and calculated at equidistant time points t = 1, 2, ... , T
(Fig. 6.1.1). For brevity of presentation a year is used as the account-
ing and time unit. This does not restrict the theory because any other
accounting period can be treated as a basic unit (see the discussion
about the continuous and discrete approaches in item 1.4(b)).
The aggregate claims of year t will be denoted X(t). The accumulat-
ed amount X(t1' t 2 ) in years t1' ... , t2 is then
12

X(tl' t 2 ) = L X(t) = X(I, t 2 ) - X(1, tl - 1). (6.1.1 )

1=1,

As was specified in item 3.1 (c), X(t) includes both the paid and
outstanding claims.

(b) Extension ofthe basic assumptions It was stated in Section 2.7,

and shown e.g. in Fig. 2.7.3, that the basic parameters of the claim
process are continually subject to alterations which are partially
revealed as trends and partially as more or less regular, often cyclical,
changes. The effect of these phenomena is so significant that they
184 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

X (1,T)

r
Figure 6.1.1 A realization of a claims process extended over T years.

cannot be omitted in long-term considerations. The assumptions

concerning the mixed compound Poisson process are now extended
accordingly; first, in this section, the parameter n (the expected
number of claims) is discussed, and then, in subsequent sections,
some other bases are also taken up.

(c) The trends in the parameter n are partially due to the gradual
change of the size of portfolio and are partially affected also by
alterations in the risk exposure inside the portfolio. They can be
taken into account by assuming that the model parameter n is
time dependent. This can be done conveniently by means of a
growth factor rit) = n(t)/n(t - I), or equivalently
t
n(t) = n TI ri'r), (6.1.2a)
t= I

where briefly n = n(O). The corresponding rate rg - I is denoted by

ig (see item l.S(f)).
The growth rate need not be the same from year to year. However,
if consideration is limited to relatively short periods, its constancy
can be assumed. The formulae can be considerably simplified in
this way. Hence, in most connections in the following, trends are
introduced into the model by simply assuming them exponential
 6.1 CLAIMS 185

and by putting
n(t) = nr~. (6.1.2b)
Another approach would be to use linear formula
n(t) = n + r~t. (6.1.2c)

(d) The cycles Besides the trends there are the periodic variations
in claims frequencies; these are considered next. Even though the
variations concerned are often in practice rather irregular, they are
by convention called 'cycles'. They are distinguished from the short-
term 'oscillation' already introduced in Section 2.7; these are
composed of only variations appearing as waves which extend over
two or more years, whereas the 'structure variations' of consecutive
years are mutually independent. The simplest way is to find some
suitable deterministic formula to indicate the relative deviations,
denoted by z(t) = Lln(t)/n(t) of n(t) from its trend flow (6.1.2b). A
deterministic sinusoidal form
z(t) = zm sin(wt + v), (6.1.3)
is assumed as an example (Fig. 6.1.2), where Zm is the amplitude,

z (tJ

v=o

, ~~--_
'-, v=-1T/2
,,,
",'

"
,,
,,
\
,
, ,, /
I "
I
,"
I
,, "
,-- " "
;
"

Tz

Figure 6.1.2 Sinusoidal cycles according to (6.1.3).

186 RISK PROCESSES WITH TIME·SPAN OF SEVERAL YEARS

v the phase and w is the frequency coefficient

(6.1.4)
Tz being the length of wave. The observations, which were referred
to in Section 2.7, suggest this shape, at least as a first approximation
in solvency considerations (but not necessarily when the model is
designed for forecasting; see items (g) and (h».

(e) Autocorrelative time series A more sophisticated way to handle

the 'cycles' is to make use of the well-developed theories of time
series.
Fig. 6.1.3 exhibits a couple of typical economic time series. The
'cycle' variable z = /).n/ii periodically soars and then swings down
for another period until another upswing occurs.
A simple approach in modelling z is to assume that its value
z(t) depends on the state of the process in previous times
t - 1, t - 2, .... If z is high at these times, z(t) is also likely to be
high, but several years continued growth can be a portent of stagna-
tion to come. In addition z is subject to more or less strong
irregularities which can be assumed to be purely random
fluctuations. These features provide a basis for building

Figure 6.1.3 Employment accidents n and working man-hours w in industry

1960-81 (Federation of Accident Insurance Institutions in Finland). ii depicts
the trend of n.
6.1 CLAIMS 187

mathematical models to describe the process of the type

z(t) = !(z(t - 1), z(t - 2), ... ,) + 8(t) + b 1 8(t - 1) + b28(t - 2) +"',
(6. 1.5 a)
where the 'descriptor' or 'predictor' function! conveys the feedback
of the past experience to the future. The terms 8(t), the so-called
'noise', represent the random effect, which is often assumed to be
normally (N(O, 0'2) distributed.
Linear functions are often satisfactory approximations for
(6.1.5a)
z(t) = a1z(t - 1) + a2 z(t - 2) + a 3 z(t - 3) + ...
+ 8(t) + b 18(t - 1) + b28(t - 2) ... , (6.1.5b)
which are known as ARMA processes (autoregressive moving
average, see Chatfield, 1978). The cycles can be generated by suitable
choice of coefficients as can be shown analytically (see item 6.2(e)).
In terms of the theory there exists a strong autocorrelation between

°
the consecutive variable values.
If b 1 = b2 = ... = then (6.1.5a) is reduced to a so-called auto-
regressive, briefly AR process. A benefit of the ARMA approach is
that a stationary time series may often be described by a model
involving fewer parameters than a pure AR process (Chatfield 1975,
paragraph 3.4.5).
According to Wold's decomposition theorem (see Chatfield, 1978,
Section 3.5) any discrete stationary process can be expressed as the
sum oftwo uncorrelated processes, one purely deterministic and one
purely indeterministic. The best-known examples of purely deter-
ministic processes are those whose realizations are simply of the
sinusoidal form (6.1.3).

(1) Exogenous impacts Even though, as can be proved (see item

6.2(e)), the time series approach described generates processes which
may be of the same type as those observed in practice, they can often
be improved further by incorporating some exogenous impulses
into the system. A clear indication can be found by comparing the
curves of Fig. 6.1.3. The variation of the number of accidents is
clearly correlated with the working activity of the industry which,
furthermore, depends on the general economic climate of the
national economy, on booms and recessions. This phenomenon has
188 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

already been discussed in item 2.7(b). Such observations suggest an

extension of the time series formulae by introducing exogenous
variables wi(t), which are links with the general economy of the
country; such variables might include the gross national product
(GNP), traffic intensity, volume of investments, total number of
man-hours worked, etc. Then (6.1.5a) is generalized as follows
z(t) = f(z(t - 1), z(t - 2), ... , ; w1(t), w2 (t), ... ) + s(t). (6.1.6)
Suggestions and analyses of this kind of dependence can be found in
Helten (1977) and in the theses of Becker (1981). Similar results are
reported also by e.g. McGuinness (1970), Witt and Miller (1980),
Bohman (1979), Balzar and Benjamin (1980, 1982), Pentikiiinen
(1982) and Rantala (1982).
As an illustration of the method related to (6.1.6), an application
presented by Becker (1981, p. 190) is given in Fig. 6.1.4. The formula
applied is of the following linear form

fI!..f = 0.92 TI!..k + 0.29 N

I!..N
- 0.57u_ 1, (6.1. 7)

where f is motor third-party claims per vehicle, k is total traffic

kilometres in the relevant year, N is the sales of new saloon and
estate cars, and u_ 1 is a residue term related to the preceding year.
The coefficients are determined by the least squares method
(Coherence-Orcutt method, see Becker, 1981, p.73). The autoregres-
sive effect is achieved through the last term having a negative coeffi-
cient. As is seen from Fig. 6.1.4, the suggested formula gave quite a
good fit in the case concerned.

(g) Forecasting If a model is constructed to give a forecast for

the future flow of business, either on the company level or more
generally e.g. for some line of all insurers, obviously a formula of
type (6.1.6) is most appropriate. Exogenous variables Wi have to be
found which are best fitted to the environment in question, as
demonstrated in Fig. 6.1.4. For instance the GNP is often used. This
and other similar variables should first be adjusted for trends by
treating as in item (c). Then deviations of the actual data from the
trend-adjusted flow are often useful indicators, as demonstrated in
Fig. 6.1.5.
Unfortunately the possibilities of finding reliable forecasts are
quite limited as regards general economic indicators. Mostly they
 22

18

14

10

19641
-~9t62

-6

Figure 6.1.4 Relative changes in the amount of losses per vehicle. German
motor third-party insurance. The actual values (solid line) and the forecast
(dashed line) are according to (6.1.7). Reproduced by permission from Becker
(1981).

wit)

t Time

Figure 6.1.5 Formulation of exogenous variables w(t), decomposed into an

exponential trend (dashed line) and a 'cyclic variable' wt(t). The former is
intended to be taken into account via the growth equations (6.1.2).
190 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

have to be limited to one or two years only so that the construction

of predictors for the insurance industry is a matter of some difficulty.

(h) General purpose models Fortunately in many standard risk-

theoretical considerations it is not necessary to know exact times
for the peaks or troughs of future cycles. This is the case for example
when the general structures of the insurer's business are examined,
or when the effect of the cycles and an insurer's capacity to overcome
cyclically bad periods is studied. For example, in the case of solvency
tests it is sufficient to appreciate that such cycles are present and to
have estimates for their amplitudes and wavelengths. To obtain
conservative evaluations it is often advisable to assume that the
adverse half of the cycle begins immediately, which means that the
phase parameter v in (6.1.3) should be put equal to O. In this connec-
tion, in particular, the sinusoidal form of the wave is a natural
assumption.
For technical reasons it is often convenient to keep the long-term
waves purely deterministic, with the possible exception of the phase
v, which can sometimes be assumed stochastic. The random varia-
tion, i.e. the 'noise' of the basic probabilities, will be moved to the
variable q already introduced in Section 2.7, dimensioning its dJ. H
accordingly.

(i) Summarizing the above presentation, the (mixed) compound

Poisson process is now further generalized by replacing the variable
n = nq (see Section 2.8) by
t
n(t) = n TI rirHI + z(t))q(t), (6.1.8a)
t= 1

or for a constant growth rate

n(t) = nx r! x (l + z(t)) x q(t). (6.1.8b)
The cycle variable z(t) can be defined to be either deterministic, e.g.
by applying formula (6.1.3), or it can be generated by means of a time
series technique as drafted in the preceding items. The former
approach will be mainly followed in the subsequent consideration
(hence notation z is used instead of z even though it is not necessary
to exclude the possibility of stochastic approaches in many con-
nections).
The separation between the trend variables and the cycle variables
6.1 CLAIMS 191

can be supposed to be made, so that on the long run the mean

value of the effect of the cycle variable z is zero. This means that
systematic changes in claim number intensities should be introduced
to the model via the trend variable rg(t).
For most subsequent applications the mean value of n(t), denoted
briefly by n(t), is required. Since q(t) was assumed independent of
the growth and cycle variables having E(q) = 1, n(t) is obtained from
t
n(t) = n x Ilr/r) x (1 + z(t)) (6.1.9a)
r; 1

or for a constant growth rate

n(t) = n x r~ x (1 + z(t)), (6.1.9b)
where the value of z(t) is given or generated for each year t separately

G) Inflation is another aspect which must be taken into account

in long-term considerations. Even a modest inflation changes values
sufficiently for the monetary variables calculated at the beginning
and at the end of the period under consideration to deviate so much
that the variation cannot be ignored. The models not only have to
cope with modest inflation but also must be workable in environ-
ments where the inflation may be rather high, or may vary from
year to year.
The effect of inflation is introduced into the model by means of
an inflation rate ix(t). The subindex x refers to claim injlation as
distinct from the premium inflation which will be introduced
in Section 6.2.
In most connections it is convenient to replace the inflation rate
by an 'inflation factor' (for notation see item 1.5.(f).
(6.1.10a)
In principle rx(t) can be stochastic. However, in what follows only
deterministic cases will be considered.
As a special case, e.g. when average flows for long periods are
examined, the rate of inflation can be assumed to be a constant i x .
The corresponding factor will be then denoted by
(6.1.10b)

(k) Distinction of the inflation effects One of the difficulties in

inflation analyses is that variables like claims, rates, etc. of different
192 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

subsections of any portfolio may change by different degrees during

inflation. It is generally expected that personal indemnities mainly
follow salary indexes, whereas property indemnities are more related
to the index of construction or to the index of wholesale prices. In
order to guide the evaluation of the claim index the average sizes
of the individual claims were recorded for some business classes
in Fig. 6.1.6, where some general indexes are also given for compari-
son. A similar analysis can be found in the report of Munich Re-
insurance Company (1971). The numbers show quite considerable
deviations from each other and from the indexes. This is,
of course, due to the fact that, besides inflation, numerous other
factors affect the claim sizes -legislation defining indemnities in
mandatory insurance classes, changes in policy conditions as well
as trends in construction techniques, renewal of buildings, plants
and vehicles, etc. Extension of the insurance contracts to cover many
subsidiary perils and damages and moving to multiline policies
explain the low growth of fire claim sizes in Fig. 6.1.6. It is difficult
in practice to attribute the changes in claim sizes to inflation rather
than other reasons. Therefore only the joint effect will be considered.
The factor rx will be termed the inflation factor despite the fact that

12.4 Worker's compensation

10.5 Index of wages and salaries

9.6 General liability

8.4 Motor third party liability

8.2 All classes total
7.3 Index of construction cost
6.6 Cost of living index

4.2 Fire and allied perils

Figure 6.1.6 Claim inflation 1958-75 of Finnish non-life insurers measured

by the average claim size (Pentikiiinen, 1982).
6.1 CLAIMS 193

it is assumed to represent not merely inflation, but also time depend-

ent changes of the claim sizes as a whole. Note that some of the
background factors, such as changes in policy conditions, may also
affect the shape of the claim size dJ. S, for example the frequencies
of small claims may increase more rapidly than those of large ones.
This may require special attention and necessitate frequent updat-
ing of the calculation bases, the dJ. S of claim sizes included.
The fact that the different variables follow different indexes
obviously continually changes the distributions and parameters
which are needed for the models. Those parts of business which are
strongly affected by inflation have a tendency to be weighted more
than other sections. Obviously the differentiation of the inflation
effects ought to be taken into account in models which are intended
to be used for more detailed analysis, especially on the company
level. Decomposition formulae will be given in Section 6.4. In the
following, however, a joint time-dependent rate of inflation for all
monetary variables is mostly assumed in order to prevent the
presentation from becoming unwieldy. If the purpose of the model is
not to develop accurate forecasting models but instead to evaluate
the range of fluctuations, it is unlikely that serious errors result.

(1) Aspects of the selection of the inflation premise The proper

grounds for calculation of the effect of inflation depend essentially
on the local conditions and are to be fitted accordingly. For example,
in conventional fire and other property insurance it may be a
practice that the face value is adjusted annually and then fixed
for an accounting period, normally one year ahead and the so-called
pro rata rule is applied for partial damages. It may, however, be
difficult for the insurer to make a reduction in partial claims, even
though the face value does not fully correspond to the current value
of the property at the time of loss, if the face value was at least
approximately correct at the beginning of the accounting period.
Hence, despite the normal fixing of the values to be constant for
each policy period, the claims may have a tendency to slide up contin-
ually. This is clearer in lines like liability, where indemnities are
determined on the basis of actual costs or salaries. Note still the
relevance of time lags between the claims inflation and the premium
inflation which will be considered in item 6.2(a).
The practice already may be that the continuation of inflation is
accepted as a normal working hypothesis, and necessary margins in
194 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

premium rates and reserve calculations are provided to meet losses

which cannot be covered from future premiums. Problems arise,
however, when the actual inflation temporarily exceeds the assumed
level, e.g. the technical reserves may need reinforcement, which
causes extra expenditure in accounts. Aspects such as this should be
considered when analysis of the effect of inflation is made and the
bases for calculation, the factor rx(t), are chosen. The problems
caused by inflation shocks will be considered in Section 7.6.
A further complication is the experience that the level of inflation
as a rule is correlated with the growth cycles of the national economy.
For example, a boom may Simultaneously both increase the claim
ratio, as stated in item (f), and push up inflation, and in that way
worsen the claim ratio. This synchronism cannot be ignored; its
consideration will be deferred to item 7.6(d).

(m) Interdependence of inflation and interest rates Another observ-

ation of importance is the interdependence of inflation rate and
interest rate. Both market forces and the government policy, and
the central banks provide an increment in the interest rate to
compensate for the inflation losses of money lenders and savers.
This important relationship will be dealt with in Section 6.3, and it
will be proved that, instead of the inflation rate, a more relevant
factor for the economy and solvency of insurers is the ratio between
the inflation rate and the interest rate.

(n) Inflation immune ratios It will be assumed in the following

that inflation affects the claims amounts of all sizes according to the
same ratio ix(t) (see (6.1.10a)). The moments about the origin of the
claim size distributions, as defined by (3.3.1), are changed as follows
ak(t) = rx(ttak(t - 1). (6.1.11)
It is useful to notice that the ratio
(6.1.12)
is not changed; it is 'immune' to claims inflation. Some analytical
simplifications in many considerations can be achieved by making
use of this property, as will be shown later. The indexes r 2 and r 3
as defined by (3.3.8) are of just this type.
The immune assumption means that the shape of the claim size
function S is not changed by inflation, i.e. only the scale of the
6.1 CLAIMS 195

claim size variables is to be corrected by the inflation factor. In other

words, inflation hits both small and large claims equally. As far
as experience shows, this may be satisfactorily true at least for
short periods and when the inflation rate is not exceptionally high
and does not change much from year to year. It can also be used
as a working hypothesis for calculations without fear of serious bias
because, as is seen, the risk-theoretical behaviour of the processes
is fairly robust against changes in the shape of the S function.
Of course, for actual applications it is desirable to test this assump-
tion; and in any case, for this - and also for many other - reasons
the shape of this function should be recalculated fairly frequently
according to accrued experience. In particular, excessive inflation
shocks may give rise to a need to check the assumptions rapidly
and to take special precautions, e.g. concerning the long-tail busi-
ness.
Note that the assumption about constancy of ratios (6.1.12)
implies that the reinsurance policy is not changed other than that
the monetary limits are changed in proportion to inflation.

(0) Mean values After the definition in item (n) we are now ready to
calculate the main characteristics of the annual claims expenditure
X(t) and its accumulated amount X(l, t). The mean values are obtain-
ed directly from the sum (6.1.1)
t
flx(l, t) = I flx(r), (6.1.13)
r~ 1

where according to (3.3.7), (6.1.8a) and (6.1.11)

flx(r) = nm(l + z(r)) n (rg(u)rx(u)).

u~l
r
(6.1.14)

As before, the notation

n = n(O) and m = a/O), (6.1.15)
is used. The values of z are obtained or generated separately and
substituted in this formula.
In the special case when the growth rate and inflation rate are both
constant, this formula is reduced as follows
flx(r) = nmr;r:(l + z(r)). (6.1.16)
196 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

If the cycles introduced by means of the variable z are omitted,

then the expression (6.1.13) is changed into a geometric series and
can be summed to

(6.1.17)

where the coefficient (see item 1.5(f) for notation)

(6.1.18)
can be interpreted as the factor of the total growth of the nominal
values.

(p) Variance In general cases the variance of the accumulated

aggregate claims (6.1.1) is
ai(tl' t 2 ) = var X(t 1 , t 2 )
12

= L var X(r) + L cov(X(r 1 ), X(r 2 ))· (6.1.19)

r=tl

In most applications in the following the cycles will be taken into

account by means of some deterministic rule, e.g. (6.1.3). The
inflation factor rx(t) will also either be given in a deterministic way
or it is assumed to be subject, if at all, to random fluctuation (noise)
of a short-term nature. Furthermore, given that the structure
variables q(t) of consecutive years are assumed independent, the
covariance terms of the above formula vanish. Hence, we have
12

ai(tl' t 2 ) = L ai(r), (6.1.20)

r=tl

where (see (3.3.7)), if inflation is also deterministically defined,

(6.1.21)

The latter formulation makes use of the immunity aspect, which

was stated in connection with (6.1.12). The effect of inflation enters
into consideration only through the factor Jlx(r) whereas the expres-
sion in the square brackets is not affected. The expected number of
claims n(r) is defined by (6.1.9a).
A more general treatment of (6.1.19) can be found in Rantala
(1982, Section 1.5).
6.1 CLAIMS 197

(q) Skewness For the calculation of the skewness the third central
moment is needed and is obtained in a similar way to the variance
(see (3.3.7))
t

,u3(X(tl' t 2)) = L [r3/n(ti + 3r2a~/n(r) + a!y ],ux(r)3.

Q (6.1.22)
t= 1

Also in this formula the expression in the brackets is immune against

inflation.

(r) Distribution function F In numerous applications the dJ.

F(X; 1, t) = prob {X(l, t) ~ X}, (6.1.23)
of the aggregate claims amount X(l, t) is needed. A rigorous formula
for F could be obtained by t - 1 convolutions of the corresponding
distribution functions related to each year of the period. However,
in practice it can be obtained with satisfactory accuracy by means of
the NP formula or by the gamma approximation. Of course the
parameter n is to be replaced (see (6.1.9a)) by
t
n(1, t) = L n(r), (6.1.24)
t=l

and the mean value and standard deviations are to be taken from
(6.1.13) and (6.1.20). The skewness
Y = Yx(1, t) = ,u3(X(1, t))/a x(1, t)3, (6.1.25)
is obtained by means of (6.1.22).
Then
x = x(l, t) = (X - ,ux(1, t))/ax (1, t), (6.1.26)
is transformed into y according to (3.11.14), after which
F(X; 1, t) = F(x; 1, t) ~ Nix) = N(y). (6.1.27)

(s) Discussion on the assumptions It was assumed that inflation

affects the claim sizes proportionally, the shape of the dJ. S remaining
unchanged. All other changes and variations were directed to affect
only the expected number n of claims. This is one of the model
assumptions which have not been investigated well so far. It could
well be supposed that e.g. the increased claim frequency could cause
198 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

a proportionately larger increase in small claims than in large.

Then the claim size dJ. S would also be deformed in shape. This
would possibly suggest the introduction of a claim size d.f. S, which
is intercorrelated with the parameter n in one way or another. It
would essentially complicate the considerations, even though it
may not be quite intractable in simulation approaches, which will be
dealt with later. Fortunately, as will be seen, the risk-theoretical
behaviour of the solvency structures is fairly robust against changes
in the shape of the function S in so far as the top risks are excluded
by means of reinsurance. So it cannot be expected that the assump-
tions made would result in a serious bias in outcome.

6.2 Premium income P(l, t)

(a) Premium inflation As was stated already in item 6.1(j) it

is necessary to distinguish the inflation effects affecting the claims
and the premiums. If the inflation rate is changed in some particular
year t it usually increases claims immediately, but it may be possible
to amend premium rates after some time lag tp' Some experience
indicates (Revell, 1979; Pentikainen, 1982, Section 3.3.l4) that the
time lag is often of the order of up to 2 years, until the changed rates
can be effective. This is notoriously due to the fact that it takes some
time until the situation is ascertained, decisions made, new rates
calculated and changes implemented. Increased premiums can be
applied for many insurance classes, not before the next period after
their adoption. Market forces and regulatory authorities, if any,
can still render further delays.
Premium inflation is now introduced into the model by means of
rate ip(t) and the corresponding factor (for notation see item 1.5(f)
rp(t) = I + ip(t). (6.2.1)
As a standard assumption, the relation between the claims inflation
and the premium inflation can be expected to be
(6.2.2)
This means that the long-run level of both infation rates is the same;
they differ only by the time lag t p ' which is one of the model para-
meters. The equalization of the levels is not a limitation of the model,
because possible deviations will be dealt with by means of the
safety loading to be defined in item (c) below.
6.2 PREMIUM INCOME P(l, t) 199

For brevity, consideration will be limited to the case where both

the premium inflation and the claim inflation (item 6.l(j» are
deterministic.
The discussion in item 6.1(1) about the importance of considering
the local conditions is equally relevant to time lag. If e.g. the rates
are successfully linked with an index, or in any other way are based
on some quantity such as salary, which automatically follows
inflation, the time lag may be small and possibly relevant only if a
major inflation shock occurs.
Further consideration of problems caused by inflation can be
found in discussion papers of the Casualty Actuarial Society (1980).

(b) Risk premium Now a formula for the pure risk premium for
year t can be given by making use of(4.1.1) and (6.1.14)

(6.2.3)
t= 1

The effects of the cycle variable z and of the structure variable q

were not taken into account, because the mean value of the former
was assumed to be zero for a long period (item 6.l(i», and the mean
value of the latter was equal to I (item 2.7(c)). In other words, it is
appropriate to define the risk premium to correspond to long-run
expected average claim expenditure so that the cycles and the
structure variable can be omitted. So far as the volume increment
controlled by rg is due to the growth of the portfolio, the premium
income provided by (6.2.3) is automatic. Otherwise, of course, the
continuing updating of the rates is supposed.
In the case of steady inflation (see (6.1.16) and (6.1.18» this formula
can be simplified to
P(t) = nm(rlpY = nmr~p' (6.2.4)
where (see item 1.5({) and equation (6.2.2)) rgp = rlp = rgx'
Note that in general P(t) +- EX(t) if cycles are present. However,
when t grows,
P(1, t)/.ux(1, t) ---+ 1,

where P(I, t) = P(l) + ... + P(t).

(c) Safety loading Also the safety loading A, being defined as

weighted average for the portfolio (see item 4.l(c», should be time
200 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

dependent. Hence the safety loaded premium (still without loading

for expenses) is
p./t) = (l + A(t))P(t), (6.2.5)
and the accumulated premium income for period [tl' t 2 ] is
12

PA(tl' t 2 ) = L Pir). (6.2.6)

The safety loading A is one of the crucial control parameters of the

modeL It may have been originally aimed, as its name suggests,
at maintaining the insurer's solvency and, ifnot considered separate-
ly, also to generate necessary underwriting profit. The definition
provided above is, in fact, broader: A indicates the deviation of the
safety loaded risk premium from the level which corresponds
to the long-term mean value of the claims expenditure X.
Consequently A is affected by quite numerous factors, as will be
discussed in the following item and later. It is a general indicator
of the adequacy of the premium rates.

(d) Parameter uncertainty It is a matter of fact that many of the

central data of distributions needed for premium rating, reserve
calculations and risk theory considerations are never known exactly;
there are only estimates subject to statistical inaccuracy, and there
are continual changes in portfolio and risk structures. This un-
certainty eventually affects the safety loading, which is not only
stipulated by the management to be at some target level, say Ao'
but is affected also by various errors in rating, by unexpected claims
development, by deviations of the actual expenses from those
anticipated by expense loadings, and also by the net result of
investments as will be discussed later. This uncertainty is aggravated
because the estimation of the trading result, and hence the actual
safety margin of rates, cannot be made until after some time delay,
which for the so-called long-tail insurance classes may be quite
considerable.
A very instructive approach is to make the safety loading a
control parameter which provides an instrument to correct biases
arising from the various errors and inaccuracies in rating and in
other control measures, as well as providing, as far as possible, the
target safety income Ao P. This mechanism is best demonstrated
by a simplified example in item (e), which makes use of the technique
6.2 PREMIUM INCOME P(l, t) 201

of time series theory. It makes it possible for example to simulate

the behaviour of management when adverse results of underwriting
are faced.
The parameter uncertainty is in no sense a feature only of
risk theory. It concerns nearly all mathematical models which are
intended to describe actual phenomena. Its harmful effects can be
controlled by assigning conservative values to the uncertain para-
meters, or by making sensitivity analyses applying alternative
parameter selections - optimistic, pessimistic and likely.

(e) Profitability control by means of l will be exemplified in this

item. Even though oversimplified, the example will throw light on
some central features of the decision process of insurers. A more
comprehensive development of control systems will be pursued in
Section 7.7. Suppose that an insurer wishes to have for the safety
loading l a target level Ao' Estimates for l are made annually on
basis of the trading results, and the value applied in the rates of the
next following year is corrected for the observed deviations from
target level in the two (or more) most recent years as follows
I(t) = Ao + a 1 (A(t - 1) - Ao) + az(A(t - 2) - Ao) (6.2.7a)
where I(t) is the value of the safety loading to be applied in the
calculation of rates for year t and the AS are estimated values based
on the actual outcomes of the process. The coefficients at and a z
are to be chosen to give suitable control effect for the formula. For
example at = a2 = - 0.5 would mean an attempt to correct the
observed deviations in two years. The safety loading 'forecast'
I is affected by the fluctuation of the claim expenditure X, errors in
the premiums and all the types of parameter uncertainties mention-
ed above, which result in inaccuracies of the estimates of A.
The effects of the control rule provided by (6.2.7a) can be described
by an AR-type process (see item 6.1(e)).
l(t) = Ao + aJl(t - 1) - AoJ + a z [l(t - 2) - AoJ + s(t),
(6.2.7b)
where the s term introduces the stochastic variation to the system.
It is due, among other things, to the uncertainties involved in the
estimation of A(t - 1) and A(t - 2) and furthermore to inflation,
changes in risk exposure etc.
The properties of the solution of(6.2.7) can be suitably investigated
202 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

by the Monte Carlo technique, which is presented in Section 6.8.

An example is given in Fig. 6.2.1 and another in exercise 6.8.2.
Analyses on AR processes can be found in Cox and Miller, 1965,
p. 282, Chatfield, 1975, p. 49 and Rantala, 1982, Section 2.2.1 (see
also exercises 6.2.1 and 6.8.2).
The process is useful for risk theory applications if it is stationary,
which means, broadly speaking, that there is no systematic change
(no trend) in the mean and no systematic change in the variance,
when the strictly periodic variations have been removed (see
Chatfield 1978, paragraph 2.2 and for a strict definition see paragraph
3.2). The necessary and sufficient conditions for the stationarity of
the above AR process are (Chatfield, 1978, paragraph 3.4.4)
a1 + a2 < 1
a1 - a2 > - 1
a2 > - 1.
The above series can be extended in a straightforward way to
include more than two terms.
The process is oscillating if a; + 4a 2 < 0 having the wavelength
T= 2n/cp,
where cp = arctan J( -ai - 4a 2 )/a 1 with 0 < cp < n (see exercise
6.2.1).
The safety loading control will be discussed later on the basis of
more flexible and realistic formulae (Section 7.7). The general
features described prove to be generally valid. The parameter
uncertainty and the necessary subsequent correction mechanism
render oscillation in the safety loading, and through that in trading
results and risk reserve.
Note that (6.2.7a) is, in essence, a modification of the experience
rating (see Section 4.10).

(f) Written premiums In most risk theory considerations attention

is focused on the claims and the risk process related to them. Then
it is natural to use risk premium income P or safety loaded one
P A as one of the basic variables. Often, however, it is necessary
to find a link with the quantities which are employed in accounting
practice and can be found in income statements and balance sheets.
Therefore, in what follows, depending on the context and application
6.2 PREMIUM INCOME P(1, t) 203

5 10 15 20 25
Time t
Figure 6.2.1 A simulated solution of the difference equation (6.2.7b) for
a 1 = - 0.5, a 2 = - 0.5, Ao = 0 and Il a normally distributed variable having
mean 0 and standard deviation (J = 0.05. The variation range is approximately
± 0.1 and the wavelength 3.2 years.

under consideration, risk premiums P and loaded premiums will

be used in a parallel manner. The latter will be denoted by Band
defined mostly as the earned premium income on the insurer's net
retention. For this purpose, loading coefficient for expenses c is
introduced. It combines the premiums as follows
P).=(I-c)B. (6.2.8)
Note that the premium incomes P). and B in the subsequent
considerations may be either stochastic, as follows from control
rules like (6.2.7), or deterministic. However, the notations P). and
B will mostly be used instead of the stochastic symbols P). and B
(see item 1.5(a)).
The safety loading is sometimes calculated on the basis of the
pure risk premium income P and sometimes - which in some
connections is more convenient - on the basis of the written
premiums B. These loading rates will be distinguished by subscripts
p and b. Hence
(6.2.9)
204 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

or

(6.2.10)

Then the risk premium can be expressed in terms of B as follows

(6.2.11)
and the two safety loadings are linked to each other by the equation
(6.2.12)
Note that )'p is the same as was denoted in (6.2.5) and in previous
connections by A. The concept and the notation will be developed
still further in item 6.5(g), taking into account also the effect of
yield of investments.

(g) Variations in expense loading c In practice also the expense

ratio c changes from year to year, which may suggest its definition
as a time-dependent variable. From experience the range of varia-
tions is, however, slight. It is convenient to keep c constant and let
its actual variation be reflected in the safety loading, which is a
conjunction of various variations. Of course, c is also a weighted
average of c values of the portfolio sections and must be as frequently
updated as all other data.

(h) A simplified approach As presented, both the aggregate claims

X and, if a control rule as (4.2.7) is assumed, the premium income
B are stochastic and also subject to long-range 'cycles'. In fact,
what is most relevant is the joint effect on the loss ratio f = X/B.
Unfortunately, handling of this kind of stochastic quotient is often
intricate, even though it may not be an insuperable problem in
simulations. Therefore, in some connections it may be useful to
simplify the model by defining P and B as deterministic and letting X
fluctuate so that the loss ratio f has about the same variation as if
both B and X were stochastic. The effect can be expected to be about
the same in both approaches. This is easily seen by denoting B =
B + AB, P = P + AP and X = X + AX where B, P and X are the
mean values, which can be programmed to be deterministic, and AB,
AP and AX are the stochastically fluctuating parts of these quantities.
Band P and consequently AB and AP (see item (0 above) can be
assumed to deviate only by a constant (or at least approximately
6.3 YIELD OF INVESTMENTS 205

constant) proportionality factor. Hence AB/E ~ AP/P. Then

f X+AX X(I+AX/X) X( AX AP)
l-e-Ab = P+AP = P 1 + AP/P ~ P 1 +g-p

~ ~(1 + AX -= AP) = X+ (A~ - AP) (6.2.13)

p X P

providing that the relative stochastic deviations are not large and
given that X ~ P. Hence, the fluctuation of f is about the same,
whether both AX and AP fluctuate or whether their sum is replaced
by a modified AX' having a range of fluctuation about the same as
the original sum of these variables; that is, if P is taken as determinis-
tic and the standard deviation and skewness of the claim fluctuation
are adjusted to correspond to the fluctuation of the loss ratio.
This approach would be expected to be most useful in cases where the
long-term variations, i.e. trends and 'cycles', are programmed to be
deterministic, and consequently the values of AX (and AP) for
consecutive years are mutually independent.

Exercise 6.2.1 Solve the difference equation (6.2.7b) in the deter-

ministic case s(t) == O. When is the solution an oscillating function
of t? What is the wavelength? (Hint: the solution is of the form
f(t) == A(t) - .1.0 = Abt , A and b being constants; furthermore, if
fl (t) and f 2(t) are solutions, C til (t) + CJ2(t) is as well.) Consider
the special case a l = a2 = - 0.5 with initial conditions f( - 1) = 0.1
and f(O) = 0.05.
The time series (6.2.7b) is further illustrated by exercise 6.8.2.
Readers who are familiar with the Monte Carlo technique should
study it in this connection.

6.3 Yield of investments

(a) Investment income The considerations contained in the fore-

going chapters and sections were limited to premium income and
claims expenditures, the difference in which constitutes the 'under-
writing result' according to common terminology. However,
insurers nearly always have assets in excess of what is needed for
current cash transactions. These excess funds, which may be quite
substantial, are invested in order to earn interest. In fact, the
ordinary underwriting result, calculated without regard to the yield
206 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

of investments, may be inadequate to ensure the continuation of

the business; it can occasionally be even negative. Hence investment
income is so substantial a component of the total business structure
that it is necessary to take it into account when the insurer's
business as a whole is considered, regardless of whether it should
be taken into account when e.g. the premium rates are determined
and the acceptability of the underwriting results are evaluated.
It will now be introduced into the model. The investment income
is, in addition to premiums and claims, a third source of time-
dependent variation.
The interest income will be divided into two parts, one allocated
to the technical reserves and the other to the risk reserve, as will
be considered in detail later. The key variable will be the interest
rate. It refers to the total actual rate at which the insurer earns
from his investments in year t. It will be denoted by iJt) or briefly
i j (subscript i for interest; for notation see item 1.5(f)).
The interest rate depends, of course, on the accounting practice
applied for the calculation of the return gained on investments.
When the actual profitability is analysed, then it is a natural
approach to regard both the cash flow and changes in market
values. The latter is determined by asset valuation principles,
which may vary from country to country, may be different for
different types of investments, and may depend on, among other
things, the capital market properties of the country. The book
values of assets may follow the market values or they may be
different. In the latter case the difference (if positive) between the
market values and the book values constitutes 'a hidden reserve'.
Another method could be described as the 'actuarial' approach
where the net gain from the insurance operation is assumed to be
accumulated in a fund which earns some assumed interest.
It is beyond the scope of this book to discuss valuation principles
in any detail. In what follows it is assumed that the interest rate
i j (t) is defined in one way or another. It will be up to the user of the
model to plan and decide, for example, how the asset valuations are
made, and whether the hidden reserves are to be counted as a
component of the solvency margin and the changes in it as 'interest
income'. The model should be operable for any valuation method.
The important point to be watched is the consistency between the
rates derived from the accounting records and those used for model
building.
6.3 YIELD OF INVESTMENTS 207

Sometimes 'interest rate' refers to the actuarial interest rate,

which is used for premium and reserve calculations. These two
rates need not be equal; normalJy the actuarial rate is less than
the actual one. The actuarial rate is intended to assign some part of
the investment income to support the underwriting business.
For example, for the increment of technical reserves, particularly in
life insurance, the calculation principles preassume certain interest
income attributable for the purpose. Then only the marginal interest
income is freely available for an increase in the solvency margin,
for dividends or other purposes. The technique wiIJ be developed in
item 6.5(h) and later, so that it is not necessary to separately
distinguish the actuarial rate. Therefore, in the subsequent consider-
ations only the actual rates wiIJ be employed.

(b) General aspects An insurer has normalJy several approaches

in planning an investment program, as investment objects may be
e.g. bonds, equities, loans, real estate, etc. On the one hand the return
should be maximized, and on the other hand the risk oflosses should
be minimized, or at least kept within some reasonable limits. A
further aspect is to find, as far as is possible, protection against
inflation. International business is usuaIJy subject to currency risks.
The legislation in some countries may impose restrictions on the
choice of investment objects or it can regulate the return rates.
There is a well developed 'portfolio theory' dealing with the
aspects involved in investment policies, e.g. contradictory aspects of
high return and risk reduction. For examples, textbooks of Christy
and Clendenin (1978), Levine (1975) and Lorie and Hamilton (1973)
may be referred to. The application of the theory to insurance
environment has been discussed by Ferrari (1968) and subsequently
by others such as Butsic (1979). Pioneering work to build a compre-
hensive model whereby the investment transactions are incorporat-
ed into the general framework of the economy of an insurer, in
addition to the traditional treatment of underwriting and other
aspects is attributed to McGuinness (1954).
A further major complication is the fact that the financial effects
of the yield of investments do not depend on the nominal interest
rate only but rather on its relation to the rate of inflation, i.e. the
relevant factor is the real rate of return (which, during high inflation,
may even be occasionally negative). This topic will be briefly
considered in item (e) and in more detail in item 6.6(a).
208 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

Clearly 'the asset risks', i.e. the risks involved with investments,
are much more diversified in different countries than are the 'liability
risks' generated by the claims expenditure and rating. Therefore it
seems to be quite difficult to find any proven detailed technique for
their consideration which could be valid to a satisfactory degree
in the very varying circumstances existing in different countries.
Only some rather general aspects can be discussed in this section, and
then the investment income will be incorporated into the general
model as a special entry in Chapter 10, where business planning is
considered.
While a more sophisticated technique for the handling of asset
risks in connection with risk theoretical considerations is at present
deemed to be beyond the scope of this book, pending future develop-
ment of the theory, this need not occasion any serious gap in the
model. As in the case of many other uncertainties, it is always
possible to make deterministic or semi-deterministic assumptions
about the anticipated future development of the return rates and
asset risks. In fact this will be provided in Chapter 10. If, in turn,
reasonably optimistic, pessimistic and likely hypotheses are used
as input data, useful information about the sensitivity of the model
to asset risks can be obtained. Surely such an approach, even though
seemingly crude, is better than complete ignorance of the asset risks.

(c) Decomposition of the investment portfolio For advanced models

it may be first necessary to evaluate investment yield separately
for different kinds of investments like mortages, equities,
and real estate. The distribution between the investment categories
and asset structure in general, including also cash, deposits, fixed
property, etc., is one of the model bases and is subject to the invest-
ment strategies of the insurer, which will be discussed later in
Chapter 10 in connection with business planning. In the following
sections and chapters it will be mainly assumed that the interest
rate ij(t) is a weighted average of all sorts of investments of the
insurer concerned and the decomposition problem is not dealt with.

(d) Bases for investment income One of the major problems in

model building is to find proper assumptions concerning the future
interest rate. For the purpose it is appropriate first to analyse
experience about the different investment components mentioned,
and then to construct the average expected flow or, if not deter-
6.3 YIELD OF INVESTMENTS 209

15

1975 1980

Figure 6.3.1 Commercial bank lending rates to prime borrowers in West

Germany, in the UK and in the USA. Annual averages for years 1973-81 and
for half of 1982.

mmlstIc, distribution of the rate. Examples of the actual flow of

rates are given for illustration in Figs 6.3.1 and 6.3.2.
Figure 6.3.1 exhibits lending rates in three countries. It is clearly
seen that there are periods of low and high rates having a length
of several years. The short-term random variations have small
amplitude compared with these long waves, which are the predo-
minant source of change. Furthermore, the flow of rates is markedly
parallel in all three countries. The interest rates in major countries
are necessarily highly correlated, because the central banks must
keep them reasonably close in order to prevent violent outflows from
low-interest countries to high-interest countries and the consequent
effects on the relative values of currencies.
The difficulty of forecasting or in finding any useful distribution
is the fact that the general level of interest yields in investment
markets is simultaneously subject both to market forces and also
to the policy of governments and central banks, which in most
countries determine or strongly affect the level of interest rates.
In fact, the major changes are due to political decisions, which may
be related also to international trends and stresses.
A drastic increase of an obviously unforeseeable character in
210 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

~ 300
."
.s
~
1ii"
« 200

100

1972 1975 1980

Figure 6.3.2 Actuaries' all share index, UK 1972-82.

the interest rate is seen in Fig. 6.3.1, in particular in the flow of the
rate in the USA in the years 1979-81.
Still more volatile is the current market value of the shares.
This is illustrated in Fig. 6.3.2, where the actuaries' all share index
in the UK for years 1972-82 is depicted. A temporary plunge down
more than 50 per cent can be seen; however, the yields on shares vary
much less than the share prices. A well-planned investment policy
may very much preclude the possibility of capital losses. Some
shares may decline in value, but if only sound securities are bought,
they will refind the original price level if given enough time.
It is highly doubtful whether modelled forecasting of interest
rates is feasible for more than perhaps a few years ahead. For
analysis of the general behaviour of an insurer's economy and for
some particular applications, simulation of the gain rates may be
useful by employing cycle generating time series (Godolphin and the
Maturity Guarantees working party, 1980). The problem and appli-
cation under consideration will determine which kind of prognoses
and techniques for the calculation of the yield on investments can be
accepted. If the solvency of an insurer is evaluated, then a cautious
assumption as to rate and possibly some kind of hypothetical
depreciation may be advisable. In rate-making and corporate plan-
ning alternative prognoses can be applied in order to get a feel for
the sensitivity of the model outcomes to the various assumptions.
6.4 PORTFOLIO DIVIDED IN SECTIONS 211

In what follows the choice of the prognoses is left open and it

is assumed that the rate ij(t) is given in one way or another either
as a deterministic flow or as a stochastic one or preferably as a mix
of both types.

(e) Real rate of interest One should appreciate that, factually,

the interest rate ij is a less relevant factor for the insurer's financial
position and solvency than is its relationship to the inflation rate
and growth rate, i.e. a modified rate which can be called a 'real rate'. It
will be introduced in item 6.5(g).
Hence, it is not rewarding to make great efforts to find prognoses
for the nominal yield of investments, if the often equally important
rate of inflation cannot be treated in a balanced way.

(f) Investment losses Besides the yield, the depreciation (or possi-
bly appreciation) of the values of investments, especially securities,
needs to be considered. If the market value slides below the book
value, the depreciation of the latter must immediately be reflected
as negative income. The possibility of this event should be taken
into account in one way or another, at least when the factors
affecting the solvency of insurer are evaluated.
A current practice in many countries is not to book the assets
at market values. In particular, due to inflation, considerable
underestimations in assets may arise, which are a cushion against
unexpected losses in investments and can well be taken into account
in solvency evaluations and also when the model assumptions are
deliberated.
More details of the non-stochastic risks including the asset risks
are discussed in Pentikainen 1982, Section 2.9.

6.4 Portfolio divided in sections

(a) Decomposition It is often necessary to divide the portfolio

in sections, as already presented in Section 3.7 in the one year case
as well as in considerations extended over a horizon of several years.
Then by using the following formulae the data related to the total
business can be calculated from the sectional data. In order to
avoid covariance terms it will be assumed that the aggregate claims
Xj of each section are mutually independent (mixed) compound
Poisson variables as defined in Section 6.1.
212 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

The total amount of claims X(t!' t 2) inherited from the years

t1' t1 + 1, ... , t2 is a double sum of the sectional amounts of the
aggregate claims of each period
t2

X(t!' t 2) = L LXIr). (6.4.1 )

r= tl j

(b) The mean value is the straightforward sum of the correspond-

ing sectional mean values

flX(t1' t 2 ) = Lflx/r ) = LLn/r)mlr)

j.r r j
t2

= L n(r)m('r), (6.4.2)
r::::;: tl

where, making use of the decomposition formulae of Section (3.7)

and equations (6.1.9a) and (6.1.14)

n(r) = Lnir) = Lnp

j j
+ zir)) nt

u= 1
rg/u),

and
n .(r)
mer) =
J
n r )m.(r)
L. ~( J

= ~v.(r)m.(r) = ~v.(r)m.
L.
j
J J L.
j
J J
n r .(u).
u= 1
XJ

The coefficients Vj were defined by (3.7.4); they can now be time

dependent. It is often rational first to compute the weighted sectional
values for each year r and then to sum with respect to the time r as
on the second line of (6.4.2).

(c) The variance is derived similarly as a double sum (see (3.7.9)

t2

ai(t!' t 2) = Lai}r) = L L[nir )a 2i r) + flx/r)2a~J (6.4.3)

t,j r=tl j

Here again it is preferable first to determine the sectional weighted

quantities for each year r and then to sum according to r. It is
assumed that the structure function H does not change from year to
year. Hence aq and'}' q are independent of time.

(d) Skewness Finally the third moment, needed for the calculation
6.4 PORTFOLIO DIVIDED IN SECTIONS 213

of the skewness, is obtained from

(6.4.4a)
t,j

After some manipulation the formula can be expressed in the form

12

J.L 3(X(t I' t2)) = I J.LxCr)3[Al (r)jn(rf + A 2(r)/n(r) + A 3(r)]

!::; t1
(6.4.4b)
where (see (3.7.12) and (6.1.22)
a
A I (r) =~
a3
I

A2(r) = 3 I a~ vir)nj(r)2a~j (6.4.5)

j lj
A 3 (r) = Inir)3YqP!r
j

(e) Simplifications The formulae can be simplified if some or

rather all of the variations rates are the same for all sections. For
example, in inflation affects all sections to the same degree so that
rxit) = rx(t) for all j, then the expressions in brackets in (6.4.3) and
in (6.4.4b) are not affected by inflation, because the coefficients
nj are then immune against inflation (see (6.1.12». They may,
however, depend on time as do the factors A in (6.4.5). If the distri-
bution parameters Vj and nj are independent of time, and the cycle
variable z is the same for all sections, the quantities Ai are independ-
ent of time. The writing out of the various alternatives is left to the
reader (see exercise 6.4.1 ).

(f) The premium income of the whole portfolio is always a straight-

forward sum of the section quantities
B(t!' t 2 ) =I Bit!, t 2 ), (6.4.6)
j
and similarly
(6.4.7)

(g) Sectional or global processing When the comprehensive simul-

ations and straightforward calculations are performed, there are
two alternative approaches with regard to the section combination:
214 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(i) The model parameters are first derived and calculated on the
basis of the section parameters as weighted averages as given
above, and then the computation (either simulation or direct
analytical handling of the problem) is performed for the whole
portfolio.
(ii) All the simulations and calculations are done separately for
each section, and finally the section outcomes are summed up.
The first approach is, of course, very much more economical
than the second as far as computation time and other resource
requirements are concerned. It means, however, that the cycles,
real growth and even inflation must be the same for all sections as
mentioned in item (e). If this cannot be assumed, approach (ii)
may be more rational, or a special technique is needed to calculate
the weighted average parameters.

(h) Robustness concerning the claim size d.f. Another feature to be

taken into account when choosing between approaches (i) and (ii)
in item (g) is the effect of the claim size dJ. S. Experience seems to
indicate that the behaviour of the claims process retained after
cutting the risk tops by reinsurance is fairly robust concerning the
choice of this distribution, as long as the net retention M is not
too high (Pentikainen, 1982, Section 4.2.2). This is a feature anti-
cipated already in Section 4.3. Hence for deduction of the aggregate
distribution it may be possible to use a fairly coarse division of the
portfolio into sections, uniting insurance classes which by ex-
perience have fairly similar risk structures.

Exercise 6.4.1 Write (6.4.3) in a simplified form providing that the

parameters Vj and 1tj are independent of time and that the rates
of growth, cycles and inflation are constant and the same for all
sections.

6.5 Trading result

(a) Basic equation The consideration can now be extended from

the one-year basic equation presented in Section 4.1 to a finite time
environment. For this purpose the increment l1V(t) = V(t) -
V(t - 1) of the risk reserve V for any year t will be defined as the
6.5 TRADING RESULT 215

difference of incomes and expenditures as follows

.1U = B + I - X - C - D, (6.5.l )
where B, is the earned premium income included safety loadings
and loading for expenses (see item 6.2(0), I is the net income from
investments (see Section 6.3), X are claims paid and outstanding, C
are expenses, and D are dividends, bonuses, etc. For brevity, the
argument t is omitted.
All amounts are to be calculated on the insurer's net retention,
that is, deducting the shares of reinsurers.
The basic equation (6.5.l) constitutes an algorithm, by means of
which the risk reserve U is obtained for year t when its value for
the year t - 1 is known. The equation provides that all the variables
involved are defined and numerical values (or analytic expressions)
for them can be assigned. Many of the definitions are given already
in preceding sections; the rest will be pursued in the following.

(b) Broad and narrow approaches The basic equation (6.5.1) can
be treated by quite general conditions, taking into account numerous
inside and outside impacts and presses and business strategies. As a
rule the model gets so complicated that often only simulation can
cope with it. This general study will be postponed to Chapter 7.
It is useful before that to deal with the model by making some simpli-
fying assumptions. In this way it is possible to throw light on many
special problems which are of interest and which can serve as build-
ing blocks for more comprehensive approaches.

(c) The yield of interest will first be partitioned into two compo-
nents
(6.5.2)
where W is the technical reserve, i j is the rate of interest (see item
6.3(a)) and the subscript - 1 indicates time shift to the previous year.
By using this formula and replacing .1U by U - U _ l ' the basic
formula (6.5.1) assumes the form
(6.5.3)

(d) Interest period There are different practices for calculation of

216 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

interest and growths. One assumes that transactions like reception

of premiums, payment of claims, etc. are concentrated at the end
of the counting period (year); another concentrates them at the
midpoint. The latter is almost the same as supposing the money
flows to be continuous. The midpoint approach seems to be more
natural and better fitted to real world circumstances. Its drawback
is, however, that many of the variables should be provided with
factors J (1 + i) for the half-year interest yield, which complicates
the formulae and obscures readability. Therefore these square root
factors will be omitted. This renders, of course, a slight bias in
interest and growth amounts. Bias is fortunately offset well if the
rate factors rg , rp etc. are derived from the actual data by applying
the very same calculation method. Then the differences in the final
outcomes are negligible irrespective which of the above methods is
applied, providing of course that the same method is always applied
subsequently.

(e) Dividends D can be included in the expenses C in the considera-

tion in this chapter. This simplification can be justified by the fact
that in practice it is both usual and advisable to create a business
strategy which consolidates dividends at some fixed level rather than
allows them to vary from year to year. Furthermore, the expenses are
assumed to be given deterministically; hence the notation C is
replaced by C.

(f) Ratio form of the basic equation In order to avoid inconveni-

ences caused by inflation and real volume growth in long-run
considerations, it is often useful to use, instead of absolute monetary
quantities, their ratios to the premiums B. These ratios have dimen-
sion zero with respect to the monetary unit, and are thus not directly
affected by inflation. We therefore move on to the relative variables
u = U IB; f = XI B; c=CjB;w=WIB, (6.5.4)

dividing equation (6.5.3) by B

UIB=(1 +iJU_1IB+ I +ijW_IIB-X/B-CjB,
or
B B
u = (I + i.)-----=-!'u
1 B -1 + I + i.w -1 -----=-!.
1 B - f - c. (6.5.5)
6.5 TRADING RESULT 217

The ratio u will be called the solvency ratio. Even though B may be
stochastic in some applications by virtue of the safety loading 1 (see
item (6.2(e», the 'deterministic' symbol B will be mostly used.

(g) Dynamic calculus of the yield of interest Expression (6.5.5) gives

rise to the following definitions and notation. It is convenient
to introduce a new rate and a new factor
and (6.5.6)

to replace the classical interest rate ij and the interest factor rj = 1 + ij •

It combines the interest rate with the growth of the nominal values
of business volume, which is measured by the premium volume B.
The growth ratio BIB _ l' which will be frequently needed later, is
occasioned by (premium) inflation (with rate ip) and by real growth
(with rate ig) (see items 6.2(b) and 6.2(f).

(6.5.7)
Hence
(6.5.8a)
and
rjgp = rJr gp' (6.5.8b)

All the factors and ratios may be time dependent, although the
argument t is not written out for brevity reasons. The subscripts are
chosen to indicate the factors interest, growth and premium inflation
involved (see notation, item 1.5(1)). This rate and factor can be
interpreted as generalizations of the classical ij (interest rate) and
rj (accumulation factor = 1 + i), operated in the 'static' interest

calculus based on constant values of the money. The rate ijgp and
the factor rjgp are their counterparts in 'dynamic calculus' where
both the value of money and the real volume of business are chang-
ing. Note that the rule rj = 1 + ij cannot be extended to r jgp ' i.e.
generally rjgp =1= 1 + ijgp'
Then (6.5.5) can be rewritten
(6.5.9)

Furthermore, (1 - c - .A.b)B is what is left of the premiums when the

218 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

loading for expenses (and dividends) and safety loading Ab are

deducted, i.e. the net premiums P (see (6.2.11))

On the other hand, by definition P = E(X), i.e. the long-term expected

value of the claims. Denoting the mean claims ratio
J= E(X)/B, (6.5.l0a)
it follows that
(6.5.l0b)
and
(6.5.11)

(h) Aggregate safety loading A useful observation made in practice

(Pentikiiinen, 1982, pp. 3. 1-7) is that the ratio w = W/B is fairly
constant from year to year (the average value w = 1.71 was found
for the Finnish non-life companies). This makes it possible to simplify
the above equations by replacing the variable w _ 1 by a constant
parameter w. (However, w may be different for different classes of
insurance and for different insurers). This finding suggests that the
expression
(6.5.l2)

which after the replacement w -1 = W appears in (6.5.l1) should be

defined as an 'aggregate' safety loading. It will be introduced to
this equation in order to put it formally in a more compact shape,
allowing easier use and providing a better general view of the most
essential relationships. This approach will be followed in Chapters 6
and 7 but relaxed in Chapters 8 and 10. The second part of the
investment income I, related to the solvency margin U according to
(6.5.2), remains expressively in the basic equation.
The aggregate safety loading (6.5.l2) will be denoted by Alol if
it needs to be distinguished from Ap and Ab . In the following, A will
mostly be used for it, and it is up to the user of the model whether it
is interpreted as Alol or Ab •
In practice Ab may often be quite low, even occasionally negative,
6.5 TRADING RESULT 219

as recent experience has proved. The introduction of the yield of

interest alone makes the total safety margin sufficiently large.
One must, however, appreciate that the yield of interest is not
necessarily freely disposable; this has already been stated in item
6.3(a) in reference to the actuarial rate of interest. The technical bases
often provide that the premium income is supported by the yield of
interest and the underwriting reserves are credited by some yield
calculated according to a technical rate of interest. This is the
practice in particular in life assurance. From this point of view it
would be perhaps a more natural approach to allocate a correspond-
ing part of the yield of interest directly to enforcing the reserves;
or, which is an equivalent approach, to affiliate it to the premium
income B, i.e. to define B as a sum of 'ordinary' premiums and the
yield of interest, which is technically assigned to finance the under-
writing business in the rating schedules. For the sake of simplicity,
however, this is not done here.

(i) Basic equation The basic equation can now be expressed as

follows
(6.5.13)

which is convenient for the analysis of the underwriting business.

The coefficients may be time dependent and one or more of
them may be stochastic, for example the safety loading, the interest
rate and infla!ion buried in the factor r igp '
The sum f + A represents in a ratio form what remains of the
gross premium B when the loading for expenses is set aside. It is the
safety loaded risk premium related to B and will be denoted by p;..
Regarding (6.5.10b) we have

(6.5.14)

This equation can then be written in an alternative form

(6.5.15)
The interpretation of the basic equation (6.5.13) is obvious. The
relative amount of the risk reserve u (solvency ratio) changes from
year to year for the following reasons:
220 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(i) Yield of interest on the initial amount u_ 1 increases the solvency

ratio, but inflation and real growth according to (6.5.8) reduce
it. All these effects are included, by definition, in the factor r.Igp .
(ii) Safety loading A. increases u owing to the 'ordinary' loading ..1b
and owing to the yield of interest gained by the underwriting
reserve according to (6.5.12).
(iii) The fluctuating deviation of the loss ratio from its mean
value (J - f) is moved to u.

(j) Influence channels of the background factors It must still be

recalled that the important factor rjgp (see (6.5.8b» conveys to the
basic equation the joint effect of yield of investments, inflation and
real growth of the business volume. No one of these three factors
appears in the equation explicitly (implicit influence channels,
however, exist; e.g. ij is included in the formula (6.5.12) of A.. and
rg affects n (according to (6.1.2a -c». This proves the assertion
expressed in item 6.3(e) according to which the relationship of these
three factors is more relevant than each of them separately.
Errors in the expense factor c or (what is the same) in 1, as well
as calculation of estimates for yield of interest, inflation, growth,
cycles etc., affect directly the safety loading A., and this should be
chosen accordingly. When the model is used for simulations, this
effect may provoke cycles in a way which was described in item
6.2(e) and will be discussed in Section 7.7.

6.6 Distribution of the solvency ratio u

(a) Dynamic discounting In what follows, inflation-adjusted (i.e.

moving rather than constant) values of money will be operated, as
was reasoned in item 6.1(j). It is therefore necessary to modify the
conventional concepts of interest calculation accordingly. For this
purpose 'a dynamic discounting factor' will be introduced, synthe-
sizing the rates of interest, growth and inflation (see item 6.5(g» as
follows

r(r, t) = n
.=t+1
rjgp(v) = n t;V)()
v=t+1 r g vrpv
= B((r» x
Bt
n
v=t+1
rj(v). (6.6.1a)
6.6 DISTRIBUTION OF THE SOLVENCY RATIO u 221

In the particular case where all the factors involved are constants
this is reduced to a power expression
r(r, t) = r:;t. (6.6.1b)
If the growth and inflation are omitted, i.e. B is constant, it is further
reduced to the well-known accumulation factor r:-
t = (1 + iJ-t.

The following properties and conventions of the dynamic factor

are needed
r(t l ' t 2 )r(t 2 , t 3 ) = r(t l ' t 3 )
r(t - 1, t) = rjgp(t) (6.6.2)
r(t, t) = 1.

(b) Analysis of the u algorithm Whilst the premises of the trading

process are so general that interest, inflation, growth of the portfolio
and safety loading are stochastic and/or follow some specially
programmed dynamic or other rules, often the simulation which will
be considered in Section 6.8 is the only workable technique for
coping with the process numerically. It is, however, useful also to
investigate the problem with simplified assumptions, as mentioned
already in item 6.5(b). The results so obtained may as such be of
interest, and they may reveal properties which are also characteristic
for more general processes. In fact the conventional risk theory has
been mainly built on these simplified bases. Hence in this section
it is assumed that the future flow of interest, real growth, inflation
and safety loading are all deterministically given but not necessarily
constant from year to year. The stochastic variation is limited to
claims only.
By recursive application of the u algorithm (6.5.15) and by using
the concepts and notation introduced in item (a), the following
expression is obtained
t

u(t) = u(O)r(O, t) + L p;,(r)r(r, t) - L f(r)r(r, t). (6.6.3a)

t= 1 t= 1

In the particular case where the rates of interest, growth and inflation
are not time dependent this equation is modified in form
t t
u(t) = u(O)r: gp + L pir)r:g-;,t - L f(r)r:;pt. (6.6.3b)
t=1 t=1
222 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

It is convenient to rewrite (6.6.3a) as

u(t) = R(t) - f(1, t), (6.6.4)

where
t

R(t) = u(O)r(O, t) + L p..\ (r)r(r, t)

r;1
t
f(1, t) = L f(r)r(r, t). (6.6.5)
r; 1

The variable f(l, t) is the discounted accumulated claim ratio. If

the yields of interest, inflation and growth were disregarded, f(l, t)B(t)
would be the accumulated claim amount X(l, t) considered in
Section 6.1. The first term R(t) in the decomposition (6.6.4) is the
(hypothetical) limit amount which would be achieved if no claims
occurred; i.e. R(t) is the sum of the first two terms in the right-hand
side of (6.6.3a) representing the income issues, yield of interest, and
premiums of the process.
The interrelations between the variables involved can be seen in
Fig. 6.6.1.

u{-t:}

Ruin barrier

Figure 6.6.1 The solvency ratio process u.

6.6 DISTRIBUTION OF THE SOLVENCY RATIO u 223

(C) The mean value of u Some formulae related to the u process

will be now deduced, beginning with mean values. The terms of the
last sum in (6.6.3a) have the mean (see (6.1.14) and (6.2.11))

nm x (1 + z(r)) x nt

rg(v)rx(v)
/lr(r) = t v=1

nm x TI r g(v)r p(v)/(1 - Ab - c)
v=1

= (1 - c -~) x (1 + z(t)) x D(r), (6.6.6)

I
where
D(r) = VD1 rx(v) VD1 rp(v), (6.6.7)

is an auxiliary function showing the effect of the time lag between

claim and premium inflations (see (6.2.2)). If the lag is zero, then
D(r) == 1.
The expectation ofu(t) can now be written (see (6.5.14))
t
/lo(t) = u(O)r(O, t) + (1 + A- Ab - c) L r(r, t)
t
- (1 - c - Ab ) L (1 + z(r) )D(-r)r(-r, t)
t= 1
t
= u(O)r(O, t) + A L r(r, t)
t
+ (1 - C- Ab ) L [1 - (1 + z(r))D(r)]r(r, t). (6.6.8)
t= 1

The first two terms of the last expression give the average flow of
the solvency ratio without the cyclical fluctuation, and the last
term sets forth the cycles.
If rjgp < 1, as often happens in practical applications, the
possible downward or upward trend declines when t grows and
/lo(t) continues oscillation around a horizontal equilibrium level.
This equilibrium level can be obtained from (6.6.8) by dropping
the effect of cycles, i.e. by putting z(r) == 0. Then the formula is
transformed to a geometric series, the sum of which is
_ (O)t
() -u ,1-r:gp (6.6.8a)
/lot rjgp+IL~,
rigp
224 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

ii It I

A
(l-r )
'gp

Figure 6.6.2 The trend of the average solvency ratio and the deviation caused
by the cycles according to (6.6.8) when r igp < 1.

having an asymptotic equilibrium level (see Fig. 6.6.2)

En =
1
l~m Jlu(t) = -_A1
00 r igp
. (6.6.9)

(d) Variance According to the assumptions made in item (b),

the stochasticity is limited only to the last term in (6.6.3), i.e. to the
accumulated claim ratio denoted by f(l, t) in (6.6.5). Hence
er;(t) = er 2(f(1, t)). (6.6.10)
By using equation (6.1.21) the variance of the solvency ratio can be
written
1 1

er;(t) = L er;(r)r(r, t)2 = L [r 2/n(r) + er;]Jlf(r)2r(r, t)2. (6.6.11)

t= 1 t= 1

The asymptotic value corresponding to (6.6.9) is again finite if

r igp < 1. Assuming that rg > I, it is
.
11m 2( ) _ er;(1 - C- Ab)2
ern t - 2 (6.6.12)
1 .... 00 I-rigp

(e) Skewness The third central moment is

1

Jl3(u(r)) = - L [r3/n(r)2 + 3r2er;/n(r) + er!y q] x Jlf(r)3 x r(r, t)3,

t= 1
6.6 DISTRIBUTION OF THE SOLVENCY RATIO u 225

and the skewness

Yu (l, t) = - Yr(1, t) = 1l3(U(t) )/a;(t) (6.6.13)
Its asymptotic value is
( 1 - r2 )3/2
lim Yu (1, t) = - Yq 1 Ig~ , (6.6.14)
t~oo - r igp

provided that r igp < 1 and r g > 1.

The reader should note the minus sign of Yu and 1l3(U).

(f) The distribution of the solvency ratio u A strict calculation of

the dJ. ofu(t) would provide t - 1 convolutions as has already been
stated in item 6.1(r). This approach will be considered in the next
section. It is, however, impracticable for most applications and
again, therefore, approximative methods are needed.
It can be expected that the approximation technique employed
in Chapter 3 for the one year case (normal, N P, gamma or Wilson-
Hilferty formulae) is also applicable to the variables related to
several years' process, the more so because the stochastic element in
(6.6.3), the term f(l, t), is a sum of independent one-year variables
and therefore better smoothed than anyone of them.
Hence, as a straightforward application of the N P approximation
the variable f(l, t) is first standardized
_ (1 )_f(l,t)-Ilr (l,t)
x - x ,t - (1)' (6.6.15)
a r ,t

and the N P transformation x -+ y = v; 1 (x) with skewness Y =

Yx = Yr(l, t) is made (see (3.11.14)). Then we have

prob{x ~ x} ~ Ny(x) = N(y). (6.6.16)

The required dJ. of u(t) can now be obtained with regard to
the connection between the variables f(l, t) and u(t) as defined in
(6.6.4) and shown in Fig. 6.6.1. It will be denoted by

G(uo,u;tl't2)=prob{u(t2)~ulu(tl -1)=u o}. (6.6.17a)

In most applications t1 = 1 and U o is a constant which need not be

given in the notation. Then a short notation is more convenient,
226 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

as follows
Giu) = G(u o , u; 1, t)
= prob{u(t)::;; ulu(O) = uo}
= I-prob{f(I,t)::;;R(t)-ulu(O)=u o }
= 1 - pro b {x ::;; x}
~ 1 - N(y), (6.6.17b)
where, since ,ur(1, t) = R(t) - ,uo(t) and O"r(1, t) = O"u(t) (see Fig.
6.6.1),
(6.6.18)
and, as above, y = v;
l(X) with y = Yr(1, t) = - yu(t).
Here and in the following the case whereby the dJ. F or G is
discontinuous is ignored, which could provide an auxiliary term
prob{u(t) = u} (or similar) when the passage from u(t) to f(l, t)
is made.
In addition, either gamma approximation, presented in Section
3.12, or the Wilson-Hilferty formula (item 3.11(m)) can be used for
the calculation of F(X; 1, t) and Gt(u) in a similar way to the NP
formula based on the same characteristics as above.

(g) The probability of the state of ruin As a special case of the

above problem, one can ask what is the probability cp(t) that a sample
path is in the state of ruin atthe end of the year t, i.e. that u(t) ::;; ur(t) =
Ur(t)/B(t), where Ur(t) is the ruin barrier (see item 4.1(d)) and
B(t) is again the premium income. A straightforward application of
(6.6.17) gives
cp(t) = prob {u(t) < ur(t)}
= Gt(ur(t)). (6.6.19)
This formula will prove a very useful building block for the finite
time ruin probability calculations, as will be shown in Section 6.9.
For readers' convenience, the steps of the calculation of the
'ruin state probability' cp(t) are summarized as follows:
(i) Calculate ,uu(t) from (6.6.8).
(ii) Calculate O"u(t) = O"f(l, t) from (6.6.11).
(iii) Calculate Yr(1, t) = - YuU) from (6.6.13).
(iv) Define the ruin barrier uit)( = Ur(t)jB(t)).
6.7 RUIN PROBABILITY 'PT(u) 227
(v) Put X = (,uu(t) - ur(t))jo)t).
(vi) Perform the NP transformation -+ y = x v;
l(X) according
to (3.11.14) by using the skewness Y = Yf(l, t) = - Yu(t) as the
skewness Y needed for the formula.
(vii) Then cp(t) ~ 1 - N(y).
A common practice is to manipulate the scale of the solvency margin
so that the ruin barrier UJt) is chosen as the zero level. Then ur(t) = 0
and x = ,uu (t)/(Ju (t).

6.7 Ruin probability If'T (u), truncated convolution

(a) Survival probability The ruin problem related to an arbitrary

number T of years can now be considered. First some quite general
formulae are derived. Let the insurer's risk reserve at time points
o(initially), 1, 2, ... , be U(O), U(l), U(2), ... , respectively, each being
a random variable except U(O), which is known and briefly denoted
by U o'
The process consists of yield of interest earned for the risk reserve
U(t), incoming safety loaded premiums and outgoing claims as
depicted in Fig. 6.7.1. Furthermore, a ruin barrier U.(t) is provided

U
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
I
U(O)
U(T)

:ur
I
(t!

t -1 T Time

Figure 6.7.1 A discrete risk process defined by the algorithm U(t) =

riU(t - 1) + p;.(t) - X(t). R(t) is that (hypothetical) risk reserve which would
be obtained if no claims occurred.
228 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

as a function of t. The idea is mainly to demonstrate the convolution

method. The formulae achieved are not very expedient for computa-
tions compared with the approaches dealt with in subsequent
sections.
As reasoned in item 1.4(b), the argument is limited to the discrete
checking of the state of the solvency and the approach will be
numerical rather than analytic. The continuous problem is treated
by Seal (1974, 1978, 1980). Beard (1971) and Taylor (1978) have
presented modifications for its numerical treatment. The trick of
integrating over the tails rather than over the main part of the
distributions, as will be made in item (d), follows Seal's presentation.
The probability of ruin during the first T years is
'l'T = 'l'T(U O ) = prob{U(t) < Ur(t) for at least one t = 1, ... , T}.
(6.7.1)
Furthermore, the 'survival function'
WT(U) = WT(Uo' U) = prob{U(T) ~ U and U(t) ~ Ur(t) for t < T}
(6.7.2)
will be introduced. Then
(6.7.3)
is the 'survival probability', i.e. the probability that a realization
of the process survives at least T years, i.e. U(t) ~ Ur(t) for every
t = 1, ... , T.

(b) Visualization of the process Before proceeding to the derivation

of the survival equations it may be useful to summarize the basic
concepts and to attempt to find a concrete image of them. For
this purpose, each realization can be visualized by a sample path
random walking point by point t = 1, 2, ... and the whole process
as a bundle of these paths. The density of the bundle is high in the
middle and decreases away from it, as illustrated in Fig. 6.7.2,
where crosscuts of the density distribution are plotted for a sequence
of discrete times. In fact, these crosscut densities are just the densities
related to the distribution of U(t). During the process some of the
sample paths fall below the ruin barrier and will be earmarked
as 'ruined' the remainder having 'survived'. As time passes, sample
paths are continually moving from the set of the survived to the set
of the ruined.
6.7 RUIN PROBABILITY 'l'T(U) 229

Figure 6.7.2 A three-dimensional visualization of the random process U.

The function f is the density of the 'crosscut' variable U(t) for different t-values.
U r is the ruin barrier. The 1 per cent confidence limits are plotted on the ground
plane.

Figure 6.7.3 A crosscut of the process, the distribution of the state of the
business measured by means of the risk reserve U(t) at time t. The function gt
is the density of all cases and wt represents the portion of such sample paths which
still survived at t - 1.

The paths may be terminated at the occasion of ruin in some of

the applications and only the set of the sample paths still surviving is
conducted forward. (See the simulated figures, like Fig. 6.8.5 below,
where this visualization is performed).
One of the crosscuts of Fig. 6.7.2 is depicted in Fig. 6.7.3. The
function gt represents the distribution of all sample paths passing
time t. It is (in the continuous case) the derivative of the dJ. Gt of
230 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

U(t) (see 6.6.17)). The function wt plotted in the figure is defined as

the derivative of the survival function
(6.7.4)

f:
or equivalently

Wt(V) = wt(V) dV. (6.7.5)

The function wt can be interpreted as the relative density of those

sample paths not being ruined before time t. In other words wiV) dU
gives the conditional probability that a sample path started at
t = 0 goes through the interval (V, V + dV] at time t and has not
been in a state of ruin before that. This function is not a density
function in the ordinary sense because its integral extended over
the whole range ( - 00, ex)) of U is not equal to I but 1 - 'P t _ 1 .
Now, let A, B, C and D denote the areas ofthe domains determined
by gl' wt and V/t) as shown in Fig. 6.7.3, and, as before, let 1 - cp(t)
be the probability that a sample path is not in state of ruin at t. Then
A + B = 1 - cp(t)
C + D = cp(t)
A = Wt(Vr(t)) = 1 - 'P t (6.7.6)
A +C= 1 - 'P t - 1

C = 1- 'P t - 1 - A = 'P t - 'P t - 1 == ~'Pt.

The area C represents the probability ~ 'P t that the process first
time is ruined at time t.

(c) An algorithm for W, Now let the risk process be constituted

by the recurrence equation
U(t) = rj(t)U(t - 1) + Pit) - X(t), (6.7.7)
for t = 1,2, ... , where rj(t) and P)t) are assumed deterministic,
and the claim amounts X(t) are mutually independent.
If F(X; t) denotes the dJ. of X(t), then W 1 has the expression
W1 (V) = prob{r j (1)Uo + P.«l) -
X(I) ~ V}
=prob{X(1)::::;dl)V o +Pil)- V} (6.7.8)
= F(rj(I)V o + Pi1) - V; 1),
6.7 RUIN PROBABILITY 'l'T(U) 231

in terms of F(X; 1). In order to get a recurrence formula for Wt

let R(t) = rmUo(t -1) + p;.(t) be the upper bound for U(t) (see
Fig. 6.7.1). Then the following sequence of so-called truncated
convolutions is obtained for t = 2, 3, ...

(6.7.9)

The running variable Y is to be interpreted as the value ofU(t - 1)

The integrand gives the probability that a sample path at time t - I
is in state Y < U(t - 1) ~ Y + d Y and at t U(t) ~ U, which provides
that the claims of year t should be less or equal to the argument
assigned for F in the above integrals.
For numerical results the dJ.s F(X; t) of X(t) are first calculated
by the methods discussed in the preceding chapters. Then the func-
tion Wt can be obtained successively for t = 2, 3, ... by some method
of numerical integration. However, even if feasible especially when
computers are used, this method may be laborious. If some approxi-
mation is to be used for F, e.g. the N P or r formula, then the accumu-
lation of inaccuracy as well as the normal rounding-off errors under
the rather long sequence of computations may be difficult to control.
This is so especially because the integration is mostly to be extended
over the modes of distributions, and the order of magnitude of the
required I - 'I'T may thus be too close to the rounding-off errors of
the largest terms involved in the integrations.

(d) Indirect calculation of the ruin probability The calculations can

be greatly facilitated if the convolutions are arranged such that,
instead of almost the whole range of the running variable Y, only a
tail of the distribution is needed for integration. This can be done
as follows.
The set of all sample paths related to the conditional probability
WT(U) can be obtained as a difference of the following two sets

(I) The set of all paths which begin at U(O) = U 0 and satisfy U(T) ~
U, irrespective of whether the process is in a state of ruin before
the year T or not.
232 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(II) The set of all those sample paths which are in a state of ruin,
i.e. satisfy Vet) < Vr(t), at one or more of the intervening points
t = 1,2, ... , T - 1.

Set II can be further divided into T - 1 disjoint subsets according

to the time when the sample path was last in a state of ruin. Then the
survival probability can be obtained as the probability related to the
set I less the sum of the probabilities of the subsets of II, i.e.
WT(V) = prob {VeT) ~ V}
T-l
- I prob {Vet) < Vr(t); U(r) ~ Vr(r)
t= 1

for r = t + 1, ... , T - 1; U(T) ~ V}. (6.7.10)

The calculation of the above expressions is based on the dJ.

G(V 0' V; t1' t 2) ofthe risk reserve Vet) which was derived in Section
6.6 (see (6.6.17a)), where now o(t) = U(t)/B(t).
Also the survival function W/V) is needed for varying periods and
therefore is denoted accordingly
W(Y, V; t1' t 2) = prob{U(r) ~ V.(r) for r = t1' ... ,t2 -1;
U(t2)~ VIU(t 1 -1)= Y}.

Then (6.7.10) can be written

W(V o' V; To, T) = 1 - G(V o' V; To, T)
T-l fUr(tl
- t~ -00 W(Y, V; t, T)dyG(V o, Y; To' t), (6.7.11)
o

which makes the successive calculations possible, taking first To = T

(for this value of To the L term is to be dropped), then To = T - 1
etc., and finally
(6.7.12)
For further calculations, the values of W(Y, V; To, T) are needed
only for the value V = Vr(T) and for the range Y ~ V.(t) (t = T,
T- 1, ... ).
By means of numerical integrations, e.g. by Simpson's formula,
the auxiliary function W can first be calculated for a set of points
Y = Vr(t) - vh with step h and v = 0, 1,2, ... until it vanishes in
the range of the accuracy demanded. Hence a matrix of values can be
6.8 MONTE CARLO METHOD 233

computed
[W(v, t)]v,t = [W(U/t) - vh, U/T);t, T)]v,t
for t = T - 1, T - 2, ,,, , 1; v = 0, 1,2, ... , (6.7.13)
after which 'I'T is obtained from (6.7.12) and (6.7.11) by another
series of numerical integrations.
The benefit of this method is that the intervals over which the
integrals are to be taken are shorter than in the direct truncated
convolution and mainly consist of the tails of the distributions.
Whilst the integration is not extended over the mode of the distri-
bution, the rounding-off errors do not cause problems.
Note that the calculation algorithm can be arranged also by
defining the set II according to the time when the ruin first occurs, as
will be shown in exercise 6.7.1. The benefit of the approach given is,
however, that the matrix (6.7.13) does not depend on the initial
capital U 0 and can be used unchanged when 'I'T is computed for
several values of U o.

Exercise 6.7.1 Rearrange the calculation algorithm (6. 7.l1), distin-

guishing between the subsets of II on the basis of when the ruin
first occurs.

6.8 Monte Carlo method

6.8.1 RANDOM NUMBERS

(a) Uniformly distributed random numbers The Monte Carlo tech-

nique makes use of so-called random numbers, which are sequences
of equally distributed mutually independent random variables or
samples (realizations) of such sequences.
Consider as an example a random variable r which is uniformly
distributed over the interval (0, I), Its distribution function is

~
for r ~ 0
R(r) {; for 0 < r ~ 1 (6.8.l )
for r > 1.
This distribution is also called rectangular owing to the shape of
its density function. A sequence of uniformly distributed random
234 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

numbers is obtained or generated, taking a set of variables r l' r 2' ... ,

which all are uniformly (6.8.1) distributed and mutually independent.
Such sequences are needed as basic building blocks for the Monte
Carlo simulations as will be seen in the sequel. Several approximate
techniques are suggested for their generation. They operate like
a lottery, producing a list of winning numbers: each number in the
interval (0, 1) (rounded to a certain number of decimals) should
have exactly the same probability of appearing in the sequence.
Because the practicable methods can generate sequences which only
approximately satisfy the theoretical assumptions they are often
called pseudorandom numbers.
As a rule the Monte Carlo method needs such a large number of
random numbers that these have to be generated by computer.
Ready-made programmes exist for the purpose and are described in
item (e).

(b) Arbitrarily distributed random numbers Letrbeuniformlydistri-

buted and let F be a given dJ., and transform r into a new
random variable X = F-1(r). F need not be strictly increasing: it can
be for example a discontinuous step function. Therefore F- 1(r) is
defined to mean the smallest X satisfying the inequality F(X) ~ r
where 0 < r < 1. (For technical reasons an open interval is often
used instead of closed 0 :::; r :::; 1, which could lead to inconveniences
e.g. in formulae as (6.8.2) and (6.8.6) where logarithms appear.)
Then
prob {X :::; X} = prob {F - l(r) :::; X}
= prob {r :::; F(X)} = F(X),
which means that F is the dJ. of X. This observation gives (in
principle) a simple procedure for generating random numbers having
an arbitrary given dJ. F. They are obtained as a transformation from
the uniformly distributed numbers. Figure 6.8.1 illustrates the idea.
For example, random numbers
X = -(1/a) In r (r uniformly (0, 1) distributed) (6.8.2)
have the exponential dJ.
F(X) = 1 - e- ax (X ~ 0),
(6.8.3)
as is readily seen (see exercise 6.8.1).
6.8 MONTE CARLO METHOD 235

F F

1---------------------- 1 -----------.-----------------------:~-

~
F(X)
! F~Epi
r f-~~~._j,-----'

x x x x
II

Figure 6.8.1 Generation of random numbers X distributed according to F .. (1)

continuous case .. (11) discrete case.

(c) Monte Carlo method To illustrate the simulation technique,

known as the Monte Carlo method, the convolution of two distri-
bution functions F land F 2 is calculated. The idea is to make use
of the well-known fact that the sum of two independent random
variables having dJ.s F land F 2 is F 1 * F 2 distributed. First two
sequences of random numbers (Xl!' X l2 , .. · ,XlS) and (X 2l , X22 , ... ,
X2s ) are generated, the former being distributed according to F 1
and the latter according to F 2' For example the technique presented
in the preceding item can be applied. Then putting Xi = Xli + X2i
a sequence (Xl' ... ,X.) is obtained. This is a sequence of random
numbers distributed in accordance with the function F = F 1 * F 2.
To obtain numerical values of this distribution function, or more
precisely an estimate thereof, the values Xi of Xi given by the random
number generator are rearranged according to their magnitudes
and, denoting by c(X) 'the counter number' of all Xi ~ X, the
estimate F(X) ~ c(X)/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 actually carried out
but instead is 'played' or 'simulated' by means of random numbers.
This method has been increasingly used in connection with various
research projects, particularly when the direct calculation of the
distribution functions is complicated. It can also be usefully applied
in the field of risk theory, especially in cases where other methods
are intractable owing to the complexity of the model.
236 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(d) Evaluation of accuracy Because the value of p= c(X)/s is

obtained as a random sample it is subject to inaccuracy like any
other sample. In fact it is binomially distributed, having standard
deviation

where p is the true (unknown) value of F(X). Using the normal

approximation, the maximum error at confidence level 1 - 13 is

~ = Y = YJ(P(l - P)) ~ l (6.8.4)

2Js
(j

s • s •

where Y. = N- 1(1 - 13/2). The limit ~s depends on the sample size s

as shown in Table 6.8.1, where both the absolute and relative
(= ~.lp) values of the limit are given for some p values.
Thus to get one more decimal place the sample size must be
multiplied a hundredfold. The table suggests a rather large sample
size when the tail of distribution is to be evaluated, which may make
simulation costly. The textbooks suggest, however, methods to
make the simulation more efficient, e.g. by means of the so-called
variance reduction technique (see Rubinstein 1981, Section 4.3,
and Sections 6.8.2 and 6.8.3 in the following).
Fortunately, if what is needed is mainly an idea of the shape of the
distribution and of its variation range, a reasonably limited sample
size may be sufficient. This is demonstrated in Fig. 6.8.2, where

Table 6.8.l Examples of the absolute and relative limits of the simulation
errors according to (6.8.4). Confidence level 1 - e = 0.95; Y, = 1.96.

0.5 0.1 0.01 0.001

s ~s ~)p(%) ~s ~)p(%) ~s ~)p(%) ~s ~)p(%)

1000.0980 19.6 0.0588 58.8 0.0195 195.0 0.0062 619.5

10000.0310 6.2 0.0186 18.6 0.0062 61.7 0.0020 195.9
10 000 0.0098 2.0 0.0059 5.9 0.0020 19.5 0.0006 61.9
100 000 0.0031 0.6 0.0019 1.9 0.0006 6.2 0.0002 19.6
1 000 000 0.0010 0.2 0.0006 0.6 0.0002 2.0 0.0001 6.2
6.8 MONTE CARLO METHOD 237

0.4
F'(X)

0.3

0.2

0.1

2 3 4 5 6 7 B 9 10
X
Figure 6.8.2 Exact values of r'(X; 2) (solid line) and its simulated estimate
(the heights of the dotted pillars); sample size 2000.

r(X; 2) (see (2.9.1)) is simulated. Already a visual investigation

gives a good insight about the distribution and makes it possible
also to evaluate the position and magnitude of the mode and the
tail.
It is still easier to view the position and the confidence boundaries
of the bundles of sample paths, as will be seen in Section 6.8.4,
where the sample size is often no more than 100. This is due to the
fact that a visual investigation can benefit the entire configuration,
which includes information on a very great number of simulation
outcomes, whereas the registered number of 'ruins', c(X), is derived
from a few samples only. On the other hand, of course, a visual
evaluation is to some extent subjective. This may possibly be helped
by a curve-fitting technique, which supports e.g. the estimation of
the ends of the distribution using all the information included in
the sample and not only the extreme sample outcomes (in fact, this
is the idea in visual investigation, too).
The benefits and shortcomings are discussed further in item
6.8.4(g).
238 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(e) Generation of uniformly distributed random numbers As a pri-

mary input for simulations, uniformly distributed (pseudo)random
numbers are needed. They are often available as ready-made stan-
dard commands in computers or can be readily generated as a
sequence by some standard formula. One in common use is the so-
called mixed congruential formula (see Mihram, 1972, Section 2.8.7)

(d positive integer). (6.8.5)

The kth random number r k is obtained from the preceding number

r k - l taking d last digits from (6.8.5). If b is a positive integer not
divisible by 2 or 5 and a is of the form 1 + 20 i (i any positive integer)
the formula generates a sequence of lOd different numbers until
it begins to repeat the same sequence.
The initial number r 0' the so-called 'seed', can be chosen freely

°
from the open interval (0, 1). For many applications it is necessary
to employ generators which do not give or 1 as output.
Roughly speaking, a computer takes about the same time to
generate a random number as it takes for a couple of multiplications.

(I) Normally distributed random numbers In addition to uniformly

distributed random numbers, normally distributed numbers are
often needed as input for more sophisticated simulators. A suitable
standard method makes use of the following so-called log and trig
formula (see Hammersley and Handscomb, 1964)
Xl = J(-21n(r l )) x cos(2nr 2 )
x2 = J(-2In(r l )) x sin(2nr 2 ), (6.8.6)
where r l' r 2 is a pair of uniformly distributed mutually independent
random numbers on the interval (0, 1) and the generated pair Xl' X 2

°
consists of normally distributed mutually independent random
numbers having mean and standard deviation 1.
A simple way of generating approximately normally distributed
random numbers is first to generate m uniformly distributed mutual-
ly independent random numbers. Their sum X is approximately
normally N(mI2, m/12) distributed. Often m = 12 is chosen. Then
X - 6 is approximately normally N(O, 1) distributed. It gives, how-
ever, a bias in the periphery of the distribution range and is therefore
unsuitable for many risk -theoretical applications.
6.8 MONTE CARLO METHOD 239

Exercise 6.8.1 Construct random number generators which produce

numbers distributed according to the following functions:
(~) Poisson law (2.4.2) distinguishing the cases where the expected
number n is
small
large (use the Wilson-Hilferty formula (3.5.14) and ex. 2.9.7)
and the generator for normally distributed random numbers,
item 6.8.1(f).
(b) Exponential (6.8.3)
(c) Incomplete gamma function qax, h) (2.9.1) where h is a positive
integer (use the convolution rule (2.9.7)).
Note: More general r generators can be found in Rubinstein,
(1982, Section 3.6.2).
(d) Log-normal (3.5.17)
(e) Pareto (3.5.20)

Exercise 6.8.2 Study the behaviour of the time series defined by

the difference equation (6.2.7b) where ..1.( -1) = 0.1, ..1.(0) = 0.05,
a 1 = 0, a z = -1, ..1.0 = 0 and the E terms are normally distributed

°
random variables having mean = 0, (J = 0.03. Plot A{t) for
t = 1,2, ... , 25 first putting (J = and then three (or more) realiza-
tions regarding the given value of (J (readers who have no random
number generator available can use numbers given in Appendix
E). Verify that the wavelength is the same as that obtained by the
analytic calculation presented in exercise 6.2.1.

6.8.2 DIRECT SIMULATION OF THE COMPOUND POISSON

FUNCTION

(a) Straightforward simulation for calculation of the compound

dJ. of the aggregate claim amount (see (3.2.3))
00

F(X) = L PkSk*(X), (6.8.7)

k~O

for a given X, or simultaneously for a given set Xj (j = 1,2, ... )

of X values, can be carried out by repeating the following three
steps s times. The distribution functions of the claim number
process and the claim size are assumed given.
240 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(i) Generate the claim number k applying the technique presented

in the preceding section (see exercise 6.8.1(a)).
(ii) Generate k numbers Z 1 ' Z 2' ... , Z k each being distributed
according to the given claim size dJ., and sum them
k
X sim = I Zi· (6.8.8)
i= 1

(iii) Set the counter: c(X) ..... c(X) + 1 if X sim :::; Xj (j = 1,2, ... ).
Finally
F(X) ~ c(X)/s (j = 1,2, ... ). (6.8.9)

(b) Variance reduction The procedure presented in item (a) can

be made faster and the simulation inaccuracy can be reduced if the
claim number frequency is calculated analytically and only the
second phase, steps (ii) and (iii), generating the claim amount (6.8.8),
are simulated. This is an example of the so-called variance reduction
technique well known in the textbooks (Rubinstein, 1981, Section
4.3).
The procedure described in item (a) is modified as follows.

(i) = sPk (rounded as integer) will be the sample size assigned

Sk
to k (s = LS k ; if S deviates from the 'original's because of the
rounding-offs the 'new's can be adapted).
(ii) Carry out steps (ii) and (iii) of item (a) Sk times for each k for
whichsk > O.
(iii) F(X) is obtained from (6.8.9).

The variance reduction example just given demonstrates a

general feature of the simulation technique. It pays to first calculate
analytically as many steps of the calculations as conveniently
possible and to limit the simulation to those links which are not
otherwise reasonably manageable.

(c) Stratified sampling The efficiency of simulation can still be

improved if the sample is concentrated in those k values for which
the terms of the sum (6.8.7) are the largest, or more exactly in those
which most affect the variance of the simulation outcome.
The procedure described in the preceding item will be modified
so that instead of the 'natural' sample size Sk = Pk S another sample
6.8 MONTE CARLO METHOD 241

of size wk ' the so-called strata size, will be assigned for each k. The
strata sizes are to be determined so that optimal efficiency can be
achieved. For details consult, for example, Rubinstein (1981)
Chapter 4.3.

(d) Discussion The direct Monte Carlo method is at its best if

n is relatively small, i.e. when small collectives are studied. It is
suitable e.g. for treatments of catastrophic claims; an example will
be given in item 6.8.(g). Its applications are given, for example, by
Hovinen (1964), Seal (1969a), Sugars (1974), Roy (1981) and Galitz
(1982).
If n is large, as it often is, computing time and costs soon grow
beyond reasonable limits, especially if the portfolio is to be divided
into sections each needing a separate calculation. Then some other
technique is necessary, for example along the lines which will be
presented in the following section.

Exercise 6.8.3 The claim numbers are assumed to be Poisson

distributed with parameter n = 2 and the claim sizes Pareto distribut-
ed, Z ~ 1, IY. = 3. Simulate F(X) for integer values of X by using
methods described in items (a) and (b). Readers not having a
computer available can use the random numbers given in Appendix
E (and set s small).

6.8.3 A RANDOM NUMBER GENERATOR FOR THE CYCLING

MIXED COMPOUND POISSON VARIABLE X(t)

(a) Recapitulation of definitions A fast method for generating

random numbers which are approximately distributed according
to the premise defined in Section 6.1 for the aggregate claim X(t)
can be obtained by making use of the N P formula in cases where the
collective concerned is not very small. Let us recall the definitions:

(i) The claim number process is of the mixed Poisson type with
a Poisson parameter (the 'conditional' expected number of
claims) n(t) = n(t)q(t) for t = 0, 1,2, ... , T. The structure vari-
ables q(t) are mutually independent and have a joint dJ. H.
(ii) The trends are given by a deterministically defined factor
242 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

rg(t) which affects the average claim numbers n(t) according

to (6.1.2).
(iii) The average claim number n(t) may be also subject to cycles,
which will be given either deterministically or randomly,
as was discussed in items 6.1 (d)-(g).
(iv) The claim sizes obey adJ. S which is changed from year to
year according to a deterministically given 'inflation' factor
rx(t); the shape of S remains, however, unchanged as presented
in item 6.1 (n). This assumption implies that the risk indexes
r 2 and r 3 defined by (3.3.8) are independent of time (see (6.1.12».

In order to get a simulated realization of the cycling mixed com-

pound Poisson variable X(t) (t integer), a value for the cycle variable
z(t) must first be generated if it is stochastic, or calculated if it is
deterministic. Next, a value q of the structure variable q(t) has
to be generated. Which method is appropriate for the purpose
depends on the known or assumed properties of this variable, e.g.
some of those considered in the previous sections. A convenient
approach is to use the NP generator which is presented in item (b)
below. The characteristics (J q and')' q are needed as input parameters.
Then the value of the Poisson parameter n(t) for the year t is, accord-
ing to (6.1.8a)
I

n = n(O) x TI r/r) x (1 + z(t»q(t). (6.8.13)

t=1

The dJ.
Fn(X) = F(X In(t) = n). (6.8.14)

of X(t) under the condition n(t) = n reduces now to a simple com-

pound Poisson dJ. Its mean value, standard deviation and skewness
are, according to (3.3.9) and to assumption (iv) above

Il(t) = nm(t), (6.8.15)

with

t= 1

a(t) = m(t)j(nr2)' (6.8.16)

6.8 MONTE CARLO METHOD 243

and

) r3
y(t = J(nr~) . (6.8.17)

If n is not very small, as was assumed in item 6.8.2(d), the NP

approximation can be applied for F n.

(b) NP generator Suppose that t is fixed and the Poisson parameter

n (see (6.8.13)) has been generated. The simulated value of X(t)
is then obtained by generating a random number distributed
according to the dJ. (6.8.14). This will be done by making use of a
N P generator, which generates random numbers distributed accord-
ing to the standardized NP distribution N y . It is first introduced
in this item. Having the skewness y as an input parameter, the
generator is readily obtained from the N P formulae derived in
Section 3.11. as follows:
(i) Generate first a normally (0, 1) distributed random number y,
as was dealt with in item 6.8.1(f).
(ii) Transform it into NP(O, 1, y) distributed x by means of the
transformation x = Vy(y) defined by (3.11.16) and (3.11.17a).
This transformation is the same one which is depicted in
Fig. 3.11.2 but using the nomogram in the direction y to x.

The final value of X(t) is obtained from

X t = /l(t) + xu(t).
Note that for generation of another random number having the
dJ. ofX(t) a new value for n, too, must be generated.

(c) X generator We shall now summarize the procedure for the

generation of the cycling mixed compound Poisson process X(t)
(t = 1, 2, 3, ... ). This random number generator consists of the
following sequence of operations:

Deterministic initializations, needed only once: Put

g
032
= r /r3/2 ' (6.8.18)
where r2 and r3 are the risk indexes (3.3.8).
244 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

Furthermore, calculate for t = 1, ... , T

ii(t) = rg(t) ii(t - 1); ii(O) = n(O)

m(t) = rx (t) m(t - 1); m(O) given. (6.8.19)

Operation steps to be repeated for each sample path for

t = 1,2, ... , T.

(1) Generate the cycle variable z = z(t). Either a deterministic

formula such as (6.1.3) or a separately constructed time series
generator such as (6.1.6) can be used. (Note that if z is determinis-
tic, this step can be moved to (6.8.19) because it is common
for all realizations.)
(2) Generate the structure variable q distributed according to H.
If only the mean ( = 1), variance and skewness of H are known,
then the N P generator with skewness Yq can be used for genera-
tion of the value q of q(t): if Xl is the value given by the NP
generator, put q = 1 + Xl 11q .
(3) n = ii(t) x (1 + z) x q.
(4) y = go/Jn.
(5) Generate x using the NP generator with skewness y.
(6) X(t) = n x m(t) + x x m(t) x J(nr 2).
A sample path for X(t) is obtained by letting the time variable t
run 1, 2, ... , T. Of course, the variables z, q and x must be generated
separately for each t.

(d) Applicability As the NP formula itself, so also understandably

the generator based on it is satisfactory only if the simulated collec-
tive, measured in terms of the expected number of claims n, is not
very small. For small n values the direct simulation may be workable.
Hence these methods complement each other; the N P generator
is appropriate for large collectives and the direct simulations for
small ones. The critical value of n distinguishing the 'large' and
'small collectives' can be evaluated from (3.11.19), and it seems to
be usually 20-50 depending on the breadth of the claim size distri-
bution, measured by the risk index r2.
6.8 MONTE CARLO METHOD 245

6.8.4 SIMULA nON OF THE SOLVENCY RA no u(t)

(a) u process The whole risk process, i.e. the claims as well as the
flow ofthe solvency ratio u(t), can now be simulated. For this purpose
the sequence X(1), X(2), ... ,X(T) or rather the corresponding
ratios f(t) = X(t)/B(t) are to be generated as presented in Section
6.8.3. Furthermore assumptions and basic data for premium calcula-
tion and for the yield of interest are now needed as described in
Sections 6.2 and 6.3. The approaches may be either deterministic
or stochastic and dynamic control rules can be programmed as
examples in later sections demonstrate. Equation (6.5.15) then
immediately gives u(t) and a sample path can be generated such as
that in Fig. 6.8.3 where, to begin with a simple case, safety loading
and rates for interest, inflation and growth were kept constant.

(b) Stochastic bundle Following the idea of the Monte Carlo

method the simulation is repeated numerous times. A bundle of
sample paths is thus obtained, as shown in Fig. 6.8.4.
The mean line and the confidence boundaries are evaluated by
means of equations (6.6.6) and (6.6.11) and will be discussed in
Section 6.9.
A 'stochastic bundle' like that in Fig. 6.8.4 is an important tool
in analysing risk behaviour of the portfolio. The shape and position
of the bundle make it possible to draw conclusions about the
solvency and other features of the process. If the bundle is safely

1.0
<Xl
"-
;:)

"
"o
:;::
e
~0.5
c
~
oVl

5 10 15 20 25
Time

Figure 6.8.3 A sample path of the business flow process; algorithm (6.5.13)
has been used.
 1.5

Ruin barrier

5 10 15 20 25
Time

Figure 6.8.4 A bundle of realizations of an insurer's business flow process

generated by virtue of the algorithm (6.5.15). Data standards are as given in
item 7.l(d), however, with no cycles and u(O) = 0.9. The confidence boundaries
are determined by the condition that an average 1%of the sample paths at given
time t may be outside them.

1.5

o
:c
~
>-
u
~ 1.0
o
Vl

0.5

Figure 6.8.5 Risk process subject to cyclical variation of the basic probabilities
having the amplitude zm = O.l, and wavelength T z = 12 years. Inflation varying
as will be explained in item 7.6(d). Other data as in Fig. 6.8.4. Ruins 18/100. The
stars indicate the ruin of the sample path.
6.8 MONTE CARLO METHOD 247

over the ruin barrier (e.g. the legal minimum amount of solvency
margin), it indicates a solvent state.

(c) Analysis of background factors The Monte Carlo method can

be used to examine the structure of the risk process and its depend-
ence on the initial conditions and background assumptions. As an
illustration, cyclic variation via (6.1.3) was introduced to the process
of Fig. 6.8.4 and the process changed in a dramatic way.
The example depicted in Fig. 6.8.5 was constructed assuming
that the phase of the cycles is the same all the time, i.e. v = 0 in
(6.1.3). If this assumption is relaxed making v stochastic and uni-
formly distributed in interval [1, 2lt], Fig. 6.8.5 is changed as seen
in Fig. 6.8.6. This example demonstrates the flexibility of the Monte
Carlo method. It allows quite varying conditions and assumptions;
among other things numerous variables may be made stochastic.

(d) Ruin probability The Monte Carlo method also makes it

possible to evaluate the ruin probability. Each sample path which
dropped below the ruin barrier was terminated by an asterisk and
registered as a ruin. Then the registered number of ruins in relation

1.5

"
e
o

>-
u
~ 1.0
~

0.5

5 10 15 20 TIme 25

Figure 6.8.6 The same process as in Fig. 6.8.5, but now the phase of the business
cycle is randomized.
248 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

to the whole number of the sample paths is an estimate for the ruin
probability 'P T (in Fig. 6.8.5, 18/100).

(e) Dynamics The model given is still simplified and was aimed
solely at demonstrating the simulation method. It will be further
developed later, taking into account dynamics which may appear
in actual circumstances. When an insurer approaches a critical
state measures will be probably sought to improve profitability, e.g.
economising in costs, reducing risks such as lowering the net
retentions, acquiring additional solvency margin, etc. On the other
hand if theory and practice are compared, i.e. by asking whether or
not the model outcomes are in acceptable coincidence with actual
experience, it must be kept in mind that insurers, being in a weak
solvency situation, seldom become bankrupt. They are instead
usually merged with other companies. So the actual number of
observed 'ruins' is not comparable with the theoretical ruin number;
rather the latter is comparable with the number of dissolved compan-
ies (mergers included) even though, of course, a merger can also
arise for reasons other than the threat of an imminent ruin.

(f) Catastrophic risks If the expected number of claims n is very

small, then the direct method presented in Section 6.8.2 may be
appropriate to generate the annual claims amount X(t). As an
example, the risk of ' supercatastrophes', i.e. earthquakes, tornados,
etc., is simulated in Fig. 6.8.7.
It is assumed that an insurer has numerous globally distributed
reinsurance contracts and that the maximum estimated loss M
is determined so that the amount of claims originated from one and
the same occurrence, e.g. a wind storm, cannot exceed M per
contract. It is, however, well known that in the case of a major
catastrophe numerous contracts may be affected simultaneously.
Hence, the insurer can be involved in the same event through several
channels of the international reinsurance network, and his liability
may be a major multiple of the per-contract prior estimation M,
i.e. instead of M, say, kM where k is a multiplier to be evaluated.
Landin (1980) has examined this 'channel problem' and estimated
that the multiplier k is roughly proportional to the total amount
X cat of all indemnities caused to all insurers by this one occurrence.
On the other hand, the dJ. of Xc at can be well approximated by the
Pareto dJ. In this way it is possible to construct the necessary
6.8 MONTE CARLO METHOD 249

50

40

20

10

50
Time

Figure 6.8.7 Simulated supercatastrophes. Parameters derived by Rantala

(1982) from data given by D. Landin. Total losses exceeding £15 million counted
as catastrophes (wind storms). Average annual number of events n = 0.5, claim
sizes Z according to Pareto df with index 1.76. The insurer's liability calculated
using the channel multiplier hypothesis.

distributions and perform the straightforward simulation (for details

see Rantala, 1982, Chapter 5). The configurations achieved seem
to fit with the actual experience for example the data published by
McGuinness (1966).
For evaluation of the flow of the total claims expenditure, compris-
ed of ' normal' and catastrophic claims, it may be advisable to perform
the simulation using the two approaches described in parallel.
The normal claims are simulated by means of the random number
generator described in Section 6.8.3, and the supercatastrophes
are simulated simultaneously by the direct method considered in
Section 6.8.2 and modified by the channel risk. Finally the annual
simulated outcomes of both approaches are summed up. Figure 6.8.8
gives an example of how a 'normal' risk process like that given in
Fig. 6.8.4 is changed when the risk of catastrophic claims is incorpo-
rated into the model.

(g) Benefits and disadvantages of simulation The Monte Carlo me-

thod is overall a very powerful tool; it can be applied to rather
complicated models because of the algorithm technique which
proceeds step by step from time t to time t + l. The transition rules
can then be chosen in flexible ways and the assumptions concerning
trends, cycles, inflation, short-term oscillation, claim size dJ.,
reinsurance etc. can be fitted to the actual environment without
250 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

u(ti

1.0

0.5

.... _..... _.....+.+_ ...... _..

0.0 +----~-+-+~---~---~--- ...
5 10 15 20 25

Figure 6.8.8 Simulation of risk business exposed by the risk of supercat-

astrophes. Otherwise the same process as in Fig. 6.8.4.

too urgent a need to simplify, approximate or smooth them in

some analytical or other manner convenient for calculations, which
is a major drawback of the most direct methods.
The main disadvantage of simulation is that it gives only samples
as outcome, and is thus always involved with sample inaccuracy
which may make it necessary to increase the sample size so much
that the method becomes costly or even intractable. However, it
can provide a satisfactory insight into the stochastic bundle and
thus quite a good impression of the structure of the process with
moderate sample sizes, as was discussed in item 6.8.l(d).
Anyway, if direct analytic treatment is applicable it generally
is more expedient, and the role of simulation is above all to deal
with cases where other techniques are impracticable.

6.9 Limits for the finite time ruin probability 'liT

(a) Heuristic observations as guidance The simulation pictures

of the previous section, e.g. Figs 6.8.4 and 6.8.5, assume that the
ruin probability related to the process depends on how far the
stochastic bundle is located from the ruin barrier. To make the
6.9 LIMITS FOR THE FINITE TIME RUIN PROBABILITY 'liT 251

shape of the bundle clearer its confidence boundaries were plotted.

The upper and lower limits were both determined so that on average
at each time only 1/100 of the sample paths fell outside the confined
area. The NP formula was used with the mean values, standard
deviations and skewness derived in Section 6.6. If the bundle
intersects the ruin barrier, it is clearly an indication of significant
risk of ruin and the risk is the greater the longer and deeper the
lower limit runs in the risky area. This heuristic observation can be
used as guidance for finding exactly defined upper and lower limits
for the ruin probability.

(b) Simple limits The ruin state probabilities

qJ(t) = prob {U(t) < U.(t)}, (6.9.1)
considered in item 6.6(g) are now utilized. In this and in the next
item no restrictive assumptions such as those made in Section 6.6
are needed. Hence also elements other than claims can be stochastic
and the values related to the consecutive years need not be
independent.
Now let At be the set of all the sample paths of the process under
consideration which are in state of ruin at t. Then

prob(A t ) ~ probC~l At) ~ Jl prob(A t)· (6.9.2)

The middle term, the union of the sets At' is the set of all the sample
paths which are at one or more times in a state of ruin. Since it is just
the finite time ruin probability required, the limit inequalities
T
max qJ(t) ~ 'P T ~
1 <;;t<;;T
L qJ(t), (6.9.3)
t= 1

are obtained.
The upper limit deduced can be illustrated by means of Fig. 6.7.3,
writing

(6.9.4)
t= 1

The area C in the figure is equal to A'P t = 'Pt - 'Pt - 1 , which is always
a part of the area C + D = qJ(t); from this (6.9.3) follows.
252 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(C) Improved limits The above limits can be still further improved
in a simple way. For this purpose, note that the sample paths
which are in a state of ruin in more than one year are counted several
times in the upper limit. An exact value for ruin probability could
be found if each sample path, being ruined one or more times, could
be counted only once. This is, however, difficult to achieve without
unreasonable complications, but it is fairly easy to remove from
each term cp(t) the counting of those paths which were in state of
ruin already atthe previous time t - 1. To show this we will introduce
an auxiliary quantity, the probability cp(t - 1, t) that a sample path
is in state of ruin at both t - 1 and t. Its evaluation will be dealt
with in item (e). Providing it is obtainable, a corrected probability

(6.9.5)
can be introduced, giving the probability that the sample path is in a
state of ruin at time t but was not at t - 1.
By similar reasoning, the inequalities (6.9.3) can now be replaced
by the following inequalities
T
max [1P(t - 1) + IPc(t)] ~ 'PT ~ L IPc(t). (6.9.6)
2';;t';;T t=l

These inequalities have proved to be very useful owing to the experi-

ence that the upper limit and the lower limit are generally quite
close to each other; this makes it possible to use the upper limit
as a safe approximation for the finite time ruin probability 'P TO
as will be discussed in item (0 and as is illustrated in the next chapter.

**(d) A furtber improvement The limit (6.9.6) is still somewhat

rough since the sample paths which indicate ruin before time
t - 1 but subsequent recovery are unnecessarily counted for
CPc(t). This bias can be somewhat reduced if auxiliary assumptions
are made on the process concerned. They can be of the type which
make sure that prob{U(t) < V r(t)!U(t 1) = V 1}, where t1 < t,
should be a non-increasing function of V 1. This means that the
better the state of a sample path is at time t 1 the better is the chance
that it is not in a state of ruin at any later time t. Intuitively, it seems
natural to assume that risk processes in practical applications are
normally of this type. Instead of attempting to construct a strict
theory, the idea is described heuristically by making use of Fig. 6.7.3.
6.9 LIMITS FOR THE FINITE TIME RUIN PROBABILITY 'fiT 253

The aim is to replace epe(t) by

) l-'I't_1 ()
ep ee(t = 1 _ ep (t _ 1) x ep ct. (6.9.7)

in the upper limit of (6.9.6). As seen from (6.9.4) epe(t) is, in fact,
aimed to approximate L\ 'l't and it should be replaced by another
expression which is more closely related to L\'l't. Now L\ 'l't represents
all those sample paths which are first time ruined at t. Hence they
emanate from the set of paths which are not ruined during the first
t - 1 years. They are presented by the area A t - 1 = I - 'l't-1
(Fig. 6.7.3 applied for the year t - 1). The paths which were counted
above for epc(t) emanate from the whole set of sample paths which
are not in a state of ruin at t - 1, the corresponding probability
being 1 - ep{t - 1) = area A t - 1 + B t - 1 • Evidently L\ 'l't/epe{t) is
smaller than the ratio of the probabilities of the parent sets
(l - 'l't-1)/(1 - ep(t - 1)) = A t - 1 /(A t - 1 + B t - 1 )· This follows from
the fact that those paths which are already ruined before t - 1 are,
according to our auxiliary assumption, on average in a worse
solvency situation than those paths never ruined, and the former
ones are therefore more likely to fall into a state of ruin at t. Hence,
introducing the notation (6.9.7), we have L\ 'l't ~ epcc(t) which justifies
T

'l'T ~ L epec(t)· (6.9.8)

t= 1

For computations 'l't-1 is to be replaced in (6.9.6) by its estimate

:E!: ~ epJr). Without further analysis it cannot be concluded whether
this 'improved upper limit' really is an upper limit any longer because
epee (t - 1) decreased in this replacement. However, it can be expected
that it is an improved estimate for 'l't because the errors in 'l't-1
and in L\'l't have the opposite signs, and so the latter corrects the
accumulating bias in the former.
Clearly the improved formula (6.9.7) can render a marked
improvement in (6.9.6) only when 'l'T is not very small. In extreme
cases where the upper limit of (6.9.6) may exceed 1 the improved
formula may still give meaningful values.
The upper limits of the inequalities given can be used for calcula-
tion of the finite time ruin probability, particularly when a safe
approximation is appropriate:
T
'I' T ~ L epee (t).
t= 1
254 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

(e) An integral expression for q> (t -1, t) can be derived as a straight-

forward application of the dJ. (6.6.17a) of the risk reserve

q>(t - 1, t) = f_
ur <t-l)
00 G(u, ur(t); t - 1, t)d u G(u o' u; 1, t - 1),
(6.9.9)
where ur(t) = U.(t)/B(t) is again the (relative) ruin barrier. The
differential term gives the probability that a sample path at t - 1
is going through the infinitesimal interval (u, u + duJ, and the
integrand term gives the conditional probability that it is still under
the ruin barrier ur(t) at time t.
The numerical calculation can be made, for example, by virtue of
Simpson's rule substituting the G function by the NP (or gamma)
approximations (see items 6.6(f) and (g)). Formulae for the derivative
were given in item 3.11(j).
For readers' convenience the formulae needed for calculation of
the limits derived are recorded in Appendix D in a form which is
readily programmable.

(f) Examples of the limits (6.9.3), (6.9.6) and (6.9.7) are given in
Table 6.9.1 for a process illustrated in Fig. 6.9.1.
As the example illustrates the effectiveness of the limit formula,
i.e. the closeness of the limits, depended on the fact that the test
processes were of a type which dipped down only for a couple of

Table 6.9.1 Probabilities related to the example given in Fig. 6.9.1 'Plower and
'P upper are the limitsfor 'P, as given in theformulae specified in the table.

q>(t) q>(t - I, t) 'P'ower 'Pupper

(6.9.1) (6.9.9) (6.9.6) (6.9.3) (6.9.6) (6.9.8)

1 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000

2 0.000000 0.000000 0.000000 0.000000 0.000000 0.000000
3 0.000007 0.000000 0.000007 0.000007 0.000007 0.000007
4 0.000493 0.000005 0.000495 0.000501 0.000495 0.000495
5 0.003437 0.000390 0.003540 0.003937 0.003542 0.003542
6 0.005709 0.002138 0.007008 0.009646 0.007113 0.007112
7 0.003329 0.002220 0.007008 0.012976 0.008222 0.008220
8 0.000840 0.000700 0.007008 0.013816 0.008363 0.008360
9 0.000124 0.000109 0.007008 0.013940 0.008378 0.008376
10 0.000019 0.000014 0.007008 0.013958 0.008383 0.008380
6.9 LIMITS FOR THE FINITE TIME RUIN PROBABILITY 'liT 255

U
0.9

0.8

0.7 W1r
0.61~
0.5

0.4

0.3

0.2

0.1 i==========~~~:;;::==~R~u:;in~ba;;rr~ler

2 3 4 5 6 7 8 9 10
TIme

Figure 6.9.1 An example of the application of the limit formulae. rp(t) values
calculated by the NP method are given in Table 6.9.1. Standard data given in
item 7.l(d) exceptfor u(O) = 0.7 and the amplitude of sinusoidal cycles zm = 0.05.

years and then recovered. Hence only very few years during the
time span were critical in this example. The same is the case in
examples given in the figures of Section 6.8. If, on the other hand,
the process is of such a type that the lower confidence boundary of
the bundle is directed for a long time parallel with the ruin barrier,
then the limits may be expected to be more distant from each other.
In any case, it is advisable to calculate both the upper and the lower
limit to gauge the accuracy of the approximation when the upper
limit is used as an estimate for 'P T"
A major benefit of the limit formulae is the relative simplicity
of the computations compared with other known techniques, and
the ability to control the accuracy.

(g) Calculation of the initial minimum solvency ratio The 'finite

time' ruin probability derived depends on numerous background
variables, parameters and distributions, which will be specified in
item 7.1(a). The equation
(6.9.10)
is mostly employed in the reverse direction in applications, i.e. the
ruin probability B as well as other variables and parameters is
256 RISK PROCESSES WITH TIME-SPAN OF SEVERAL YEARS

Ij1 (ui
T

u
Figure 6.9.2 The ruin probability'll T(U) as a function of u applying the limit
(6.9.8) for T = 5 or 10. Finding of u when'll T is given. Data according to standards
given in item 7.1(d).

given and the initial minimum solvency ratio u(O) (or briefly u) or
the corresponding initial risk reserve (solvency margin) U = uB is
requested. For the purpose 'P T is first calculated as a function of u
(Fig. 6.9.2). Then the u values corresponding to any given s can be
read from the figure.
It is convenient to program the computer to seek the required
value u = ue directly as a continuation of the calculations producing
'P T so that no graphical or other separate determination of u is
needed. This approach is precisely what is mostly needed for
various applications.
Note that the approximate formulae (6.9.6) or (6.9.8) give still
better accuracy than quoted in item (0 when they are used for
the calculation of u. This is due to the fact that the curve 'P T as a
function u is quite steep for larger u values (note the half-logarithmic
scale in Fig. 6.9.2), hence inaccuracies in 'P T values do not affect u
values considerably.
For comparison of the limit formulae and the Monte Carlo
method some examples are calculated using both (Table 6.9.2).
Generally the direct evaluation dealt with in this section is far
more expedient in cases where it is applicable. The benefit of
6.9 LIMITS FOR THE FINITE TIME RUIN PROBABILITY 'PT 257

the simulation approach is that it can be employed in very complex

models, e.g. when stochastic and dynamic control rules are
assumed.

Table 6.9.2 Examples of the finite time ruin probabilities 'P T calculated using
the limit formulae and by the Monte Carlo method, T = 10 years, ruin barrier =
O.IB (Rantala, 1982, pp. 3.1-5).

Ruins per Corresponding 95%

Limits Limits 1000 simu- confidence interval
u (6.9.3) (6.9.6) lations for simulated 'P T

0.3 0.79-3.6 0.83-0.88 0.858 0.84-0.88

0.4 0.60-2.3 0.65-0.70 0.644 0.61-0.67
0.5 0.39-1.3 0.42-0.47 0.475 0.44-0.51
0.6 0.21-0.60 0.24-0.26 0.230 0.20-0.26
0.7 0.091-0.24 0.11-0.12 0.107 0.087-0.13
0.8 0.032-0.081 0.042-0.045 0.034 0.023-0.045
0.9 0.0093-0.023 0.015-0.018 0.011 0.0044-0.018
1.0 0.0024-0.0055 0.0040-0.0046 0.001 0.000-0.0030

Exercise 6.9.1. Prove (6.9.6).

 CHAPTER 7

Applications related to finite

time-span T

7.1 General features of finite time risk processes

(a) Basic variables Many of the considerations presented in

Chapter 4 for a one-year time-span can be extended to arbitrary
long-time periods, and in addition numerous other applications
can now be treated.
The one-year model was characterized by the seven basic variables
or parameters listed in (4.1.9). For the finite time version a number
of new variables were introduced in Chapter 6; they are as follows
(see the notation list or the list in item (d) below).
U(or u), n,.ie, M, Ciq , Yq , zm' Tz' ip ix, ip ' ig , t p ' e, T. (7.1.1)
Here U is the initial risk reserve and u = U/ B the initial solvency
ratio (which should be denoted by u(O) if any risk of confusion
with the current solvency ratio u(t) exists). In addition the claim
size dJ. S underlies all the calculations and the portfolio may be
divided in sections j = 1, 2, ... , J. If the time unit is taken as other
than one year the parameters must be adjusted accordingly.
In the one year case it was possible to express the interdependence
of the basic variables in the form of the fairly simple 'basic equation'
(4.1.7), this being applicable except for environments involving
very small portfolios. The finite time model is so much more
complicated that it seems to be difficult to find any correspondingly
simple equation, except for some special problems treated in
Section 7.9. Thus numerical computation methods have to be
used either by direct analytic calculation or by simulation. The
former shall be used, mainly replacing the ruin probability by its
upper limit as presented in Section 6.9.
7.1 FINITE TIME RISK PROCESSES 259

qs10 (ul

0.5 1.0 1.5

Figure 7.1.1 The ten year ruin probability as a function of the expected
number of claims n and the initial solvency ratio u. Other variables have the
standard values given in item (d). The dotted line was obtained by shortening
the time-span from 10 to 5 years in the case n = 20000. Explanation for the
slight deviation of '¥ 5 from '¥ 10 can be found in the analysis in Section 7.5.

(b) Types of applications As in Chapter 4, many of the applications

of the model are of such a type that some of the variables are
calculated as functions of the others. Fig. 7.1.1 gives an example, the
ruin probability being calculated by means of the technique pre-
sented in Section 6.9 and being plotted letting nand u vary and
keeping the other variables fixed.
A more useful approach is often to fix the ruin probability e to
some suitable level, say 0.01 or 0.001, and then to study the relations
of the remaining factors either as functions of each other or by
introducing some additional conditions, e.g. optimizing profits,
volumes, utilities or other outcomes.

(c) Selection of the basic data A number of applications of the

finite time technique will be given in this chapter. They are intended
to demonstrate the calculation technique. In order to make the
sequence of the applications, as far as possible, consequent and
mutually comparable the same standard values for the variables
(7.1.1) shall be used. They are listed in the next item and follow
260 APPLICA TlONS RELATED TO FINITE TIME-SPAN T

mainly the material derived in the comprehensive research work

(Pentikiiinen, 1982 and Rantala, 1982) concerning Finnish non-life
insurance. It must be appreciated that the examples given do not
claim to give any universally valid conception of the actual risk
processes, even if the source data were derived from actual insurance
statistics. The environmental conditions and business strategies
maybe so very different in different countries and different companies
that it is completely beyond the scope of this book to try to monitor
the great variety of risk processes and their applications appearing
in real life, even though the general patterns can be expected to be
similar. The purpose of the presentation is essentially to devise
tools which can be used for examination of the processes.

(d) Standard data The great number of the variables and para-
meters involved in the model makes it very difficult to get any
insight into the structure of the system and into the interdependence
of the variables and background assumptions. One way to get
round this difficulty is first to fix average values for the parameters
and variables and then to calculate the required analyses for this
set of basic values. In other words, a special 'standard insurer' will
be first constructed and examined. Then one or more variables
in turn are changed and the resulting reactions calculated. In this
way it is possible to get insight into the multidimensional structure
of the model. This idea is illustrated in the profile in Section 7.8.
The standard data are as follows:
n = 10000, the expected average number of claims; see (6.1.9)
}p = - 0.086, the premium related safety loading which provides
the value 1.71 for the coefficient w, see item 6.5 (h)
Ie = 0.040, the total safety loading; see item 6.5(h)
M = £10 6 , the maximum net retention; see Section 3.6 and Table
3.6.1
c = 0.28, the loading for expenses; see (6.2.8)
(J q = 0.038, the standard deviation of the structure variable q;

see (2.8.5)
Yq = 0.25, the skewness of the structure variable q; see (2.8.5)
T = 10, the time-span of the finite time ruin probability
G = 0.01 (= '1'10)' the ruin probability; see (6.7.1)
ix = 0.09, the constant or the average rate of the claims inflation;
see item 6.1 U)
7.1 FINITE TIME RISK PROCESSES 261

ip = 0.09, the constant or the average rate of the premium

inflation; see item 6.2(a)
tp = 2, the time lag between the claims inflation and the premium
inflation; see (6.2.2)
ig = 0.061, the rate of the real growth of the premium volume;
see item 6.1(c)
i j = 0.085, the rate of the yield of investments; see item 6.3(a)
zm = 0.15, the amplitude of the assumed cycles according to
(6.1.3); if the synchronism with the economic cycles is provid-
ed as will be presented in item 7.6(d) applying c j = 0.5 in
(7.6.2), then Zm is 0.1
Tz = 12, the length of cycles; see (6.1.4).

The claim size dJ. is the one given in Table 3.6.1 providing
excess ofloss reinsurance and the maximum net retention M = £10 6 •
As seen from Table 3.6.1, line 26, the mean claim size is £7302 and
the risk indexes r2 = 44.1 and r3 = 4465. In those applications
which are directly borrowed from the Finnish research works
already mentioned, the dJ. of industrial fire and loss of profit
insurance was used (Rantala, 1982, Appendix 1). For moderate
retentions the results calculated by either of these dJ. do not deviate
notably.
It follows from the above data that:

B = £90.6 X 10 6 , the amount of earned premiums net of re-

insurance; see (6.2.11);
rjgp = 0.938, the dynamic accumulation factor; see (6.5.8).

(e) Four types of processes Before going on to detailed considera-

tion of various applications it is useful to characterize some main
types of processes. A number of examples of risk processes have
already been given in Chapter 6. These showed that the nature of
the process is very much dependent on two basic assumptions,
namely the dynamic 'discounting factor' rigp' giving the joint effect
of interest, growth and inflation (see (6.5.8)), and cyclic behaviour.
If rjgp < 1 then, as stated in item 6.6(c), the process has a finite
equilibrium level (6.6.9). If rjgp > 1, then the stochastic bundle drifts
to infinity (if no special control rules are assumed; see item 7. 7(f)).
An example of the former case is depicted in Fig. 7.1.2(b) and an
262 APPLICATIONS RELATED TO FINITE TIME-SPAN T

U 1.5

1.0 1.0

0.5 0.5

0.0 0.0
--------------
<---~-.~---~-~_

(a)
10 20 t (b)
10 20 t

1.5
U(n 1.5 t
1.0 U(t:.o.~
0.5 0.5 •
_ u _ n _________________ ._. __ :______ ~
0.0 <---~-~-~-~_ 0.0 ~

(e) 10 20 t (d) 10 20 t

Figure 7.1.2 The main types of risk processes generated by means of algorithm
(6.5.13) (a) r igp = 1.027, no cycles; (b) r igp = 0.938, no cycles; (c) r igp = 1.027,
cycles; (d) r igp = 0.938, cycles.

example of the latter type in Fig. 7.1.2(a). The algorithm (6.5.15)

is the basis of these and the subsequent simulations. It is noteworthy
that the breadth of the stochastic bundle of the solvency ratio
u(t) (calculated by some given confidence level) in the former case
is asymptotically finite but in the latter case tends to infinity. This
can be easily verified from (6.6.12).
Illustrations of the effects of the assumption concerning cycles
are provided in diagrams (c) and (d) of Fig. 7.1.2 but not in diagrams
(a) and (b).
Figure 7.1.2 shows how crucial are the assumptions concerning
inflation and cycles. Clearly those processes where both are present
may correspond best to practical circumstances. Figure 7.1.2(d) is
of this type, even though the assumption concerning cycles was
idealized by means of the deterministic sinusoidal formula (6.1.3).
The assumption concerning the strictly deterministic form of the
cycles can, of course, be relaxed either by taking the phase of the
cycle stochastic (Fig. 6.8.6) or by making use of an auto correlated
time series. An example ofthe latter approach will be given in Section
7.7.
In order to prevent the text from growing too much, the case
rjgp < 1 will be mainly considered in what follows. The proposed
7.2 THE SIZE OF THE PORTFOLIO 263

technique, however, is applicable also for processes of type rigp > 1,

as will be shown in items 7.7(f) and 7.10(c).

7.2 The size of the portfolio

(a) Effect of company size As a first example of the use of the

technique derived in Chapter 6 the minimum initial solvency ratio
u = u(O) is examined as a function of the size of the company in
Fig. 7.2.1. The other parameters were kept constant and equal to
the standard values listed in item 7.1(d).
In order to get a link with the insurer's records the earned gross
premium income B on the insurer's own retention will be used as
one of the basic variables instead of the risk premiums P or p;.
(see item 6.2(f)). As an indicator of the financial state of the insurer
the risk reserve (solvency margin) U or mostly its relative value,
the solvency ratio u = U/ B is used and investigated keeping the ruin
probability fixed. The reserve U( = U(O)) means the smallest initial
capital which satisfies the condition
(7.2.1 )

2.0
veo)

1.5

~~
-e. _ e __& _ ::1, cycles

1.0 & - €I- -9 - £1- -G- - e- -Go -T)--G _ e--0

Q.5 \~~ ~
M=~e a a e a a e e

5000 10000 15000 20000

n
Figure 7.2.1 The initial solvency ratio as a function of the expected number
of claims n and the maximum net retention M given in £ million. Other data as
standards in 7.1(d). The dotted curve with circles: cycles assumed, for other
cases not.
264 APPLICATIONS RELATED TO FINITE TIME-SPAN T

Figure 7.2.1 exhibits u as a function of the size variable n and the

maximum net retention M. As expected the required initial solvency
ratio is higher for small than for large companies because the risk
fluctuation is smoothed relatively better in large than in small
collectives. Long time cycles were not assumed except in the case
M = 1 (curves with circles), which was calculated both assuming
(dotted line) and not assuming the cycles, in order to demonstrate
the effect of this particular assumption.

(b) Joint effect of size and retention In practice larger companies

usually apply higher net retentions than smaller ones, because the
internal equalization of the claims fluctuation improves the larger
the risk collective is. This feature, which is commonly experienced,
was clearly seen in the one-year considerations in Section 4.4. It
is again confirmed by Fig. 7.2.1. If the solvency ratio u(O) is fixed,
the allowed net retention M should be considerably smaller for
small companies (n small) than for large ones. Further examples
of the effect of the combinations of the size variable n and retention
M will be given in the form of a profile figure 7.8.2.

(c) Simulation The size of the company can be also examined by

means of Monte Carlo simulation. An example is plotted in Fig.
7.2.2 where all data, the retention M included, were assumed to be
equal to the standards of item 7.l(d) with the exception of the size
variable n. The stochastic bundle of the smaller company is es-
sentially broader than that of the larger one.
Note that the magnitude of the initial solvency ratio u(O) in
these examples as well as in the cases treated later in this chapter
is essentially larger than current statutory demands, many standards
which are often suggested and the amounts derived in Section 4.4.

1.5
U(I)

1.0

0.5

0.0 .I:=::=;=~~~~~~ 0.0 t:.:::=:.:==;.:=~=:;.:.=-~

(0) 10 20 I (b) 10 20 I

Figure 7.2.2 A stochastic bundle of (a) a small company (n = 2000) and of

(b) a large one (n = 20000).
7.3 EVALUATION OF NET RETENTION M 265

This divergence arises from the long time-span (T = 10) and as seen
in Fig. 7.21 from the cycles now assumed.

7.3 Evaluation of net retention M

As a next example on the application of the finite time technique
the net retention M is examined as a function of the initial solvency
ratio u and the expected number of claims n in Fig. 7.3.l.
This figure corresponds to Fig. 4.2.4 which was constructed for
a one-year time-span; however, the initial solvency ratio u = U/B
is now taken as the argument instead of U. A striking feature in
Fig. 7.3.1 is that M curves are located to the right (fairly high values
of u) for small M values. This is due to the fact that the long-term
cycles have a dominant effect on the ruin probabilities when the
time-span is long, and the conventional reinsurance forms, surplus
and excess of loss treaties do not give protection against temporary
increases in the number of claims. Such increases are caused by
the assumed cycles and the short-term structure variation.

n/l000= 20 10 5 2.5 1

/
3.00
M
2.00

1.00
0.70
0.50

0.30
0.20

0.10
0.07
0.05

0.03
0.02

0.5 1.0 1.5 2.0 2.5 3.0

u

Figure 7.3.1 The net retention M in £ million as a function of the initial

solvency ratio u and the expected number of claims n. Standard data of item
7.l(d), excess of loss reinsurance.
266 APPLICATIONS RELATED TO FINITE TIME-SPAN T

2.0
U

1.5
- - - - -_ _ _ _ _ _ _ _=zm;.15

1.0
--------------0.10

--------------0.05
0.5

----------------------0
5000 10000 15000 20000
n
Figure 7.4.l The initial solvency ratio u as a function of the amplitude Zm
of the cycle waves and the volume parameter n. The values of u are calculated
for intervals of length 3000 of n. M = £400 000. Default data according to
standards given in item 7.l(d).

7.4 Effect of cycles

The effect of the cycles which were introduced by the sine
formula (6.1.3) has already been demonstrated in Figs 6.8.5 and 6.8.6.
The same effect can also be examined by means of the analytical
method described in item 6.9(g). Examples are shown in Fig. 7.4.1.
The very significant effect of the cycles is again revealed. The
results depend among other things on the length of the cycle, which
was chosen to be as long as 12 years according to standards in item
7.1 (d). If the cycle length were 6 years, then e.g. u for n = 11 000,
zm = 0.1 would drop from the value 0.99 exhibited in the figure
to 0.60.

7.5 Effect ofthe time-span T

The effect of the planning horizon, i.e. the time-span T, is examined
in Fig. 7.5.1. It was allowed to grow from 1 to 25. The ruin probability
is, of course, an increasing function of T. Here again the effect of the
growth factor r igp and the cycles is seen. If the factor is greater than 1,
then the process is of the type which is seen in Fig. 7.1.2(a) and (c).
7.6 EFFECT OF INFLATION 267

U: 0.5

0.9

-1
10

1.3

10-2 ________ _

-3
10 . . . . . . _.... . ___ A' /____ •

/
/
/
/
/
/---------- 0.9
1O-4-1----'----'-_J,...!-"--_ _~---~---~~-.=.._
5 10 15 20 25
T

Figure 7.5.1 Example of the effect of the time span T. Solid lines: processes
involved with cycles of type (d) in Fig. 7.1.2. Dotted lines: no cycles, types (a)
or (b) in Fig. 7.1.2.

Then clearly the first few years are critical from the point of view of
the solvency effects. When they have passed, the stochastic bundle is
soaring high enough that extension of the period T no longer has
any noticeable influence on the probability ofruin and hence on the
solvency ratio u. The situation is quite different if the growth factor
r igp is less than 1. This is seen e.g. in Fig. 7.1.2(d), where also cycles are
assumed. Then the stochastic bundle dives down periodically to the
critical area. These periods are reflected as upward jumps in the
curves of Fig. 7.5.1, which coincide with the beginning of the adverse
halves of the cycles.

7.6 Effect of inflation

(a) Two kinds of inflation The technique developed in the previous

section can also be used for examination of the effect of inflation
on the risk process. It is appropriate first to study steady inflation,
keeping the rate ix = ip (see item 6.1(j)) constant from year to year
and then varying inflation from year to year.

(b) A steady inflation is continually shrinking the already existing

solvency margin; this effect was conveniently introduced to the
268 APPLICATIONS RELATED TO FINITE TIME-SPAN T

1.5
U III

1.0

0.5

10 20 30 40 50 60 70 80 90 100 years

Figure 7.6.1 Solvency process in the presence of four forces.

model by means of the factor r igp (6.5.8b). If the nominal growth

of the business volume owing to inflation and real growth exceeds
the rate of interest then, as stated in item 6.6(c), the stochastic
bundle asymptotically tends to a horizontal equilibrium level. If
cycles are present, then the bundle fluctuates around this level as
exhibited in Fig. 6.6.1. There are in fact four principal forces affecting
the asymptotic level of the stochastic bundle as well as its breadth.
They are presented in Fig. 7.6.1. The equilibrium level was given
by (6.6.9) and an evaluation for the breadth by (6.6.l2).
In order to show clearly the asymptotic behaviour, the simulation
was extended to 100 years.

(c) Varying inflation has effects which are very different from those
of steady inflation. The claims expenditure can be assumed (see
items 6.l(j) and 6.2(a)) to be affected nearly immediately, but it is
likely that the premium rates are corrected after some time lag tp'
During this time lag a loss occurs which may be compensated later
when the adjustment of rates has become effective. The claim in-
flation and the premium inflation were distinguished in item 6.2(a)
for the study of this effect. Let us recall the connecting equation (6.2.2):
(7.6.1)
7.6 EFFECT OF INFLATION 269

1.5
U(II

1.0

0.5

0.0~------~5------~10--------15--------2~0-------'2·5
t
Figure 7.6.2 Solid line with circles: only a steady inflation rate of 9% per
annum was assumed. The other line: a shock inflation rate of 20% per annum was
effective in years 3 and 4. The shaded area marks the difference between the
original and changed flows.

Furthermore, a special time lag function D(t) was introduced by

(6.6.7). The time lag tp is one of the important control parameters
of the model.
First it will be assumed that the inflation has some steady 'normal'
level; in the example presented in Fig. 7.6.2 this is ix = 0.09. Then
it is allowed to jump up to a higher level (in the example ix = 0.2)
for two years, after which it drops down to the original level. In
Fig. 7.6.2 first one sample path was generated for the steady inflation
and then another for this 'shock inflation'. The figure demonstrates
the use of the simulation model as a sensitivity analyser. The same
random numbers were used for both sample paths. Hence the dif-
ference in the flow of the process was caused solely by the changed
inflation assumption.
The introduction of a shock inflation into the model produces
drastic effects, as can be seen. Owing to the assumed time lag tp = 2
years, the loss accumulates first in two consecutive years until the
change in premium rates becomes effective. Furthermore, the in-
flation shock affects the solvency ratio also via the factor rjgp of
(6.5.13). Owing to the various model assumptions, above all the
270 APPLICATIONS RELATED TO FINITE TlME·SPAN T

factor r igp ' the two curves of Fig. 7.6.2 do not coincide completely
even in the long run. In fact, the correction of rates according to
(7.6.1) overcompensates the losses.
In the example the inflation shock caused an extra loss of some
20% ofthe premium income and is consequently one of the factors to
be taken into account when the dimensioning of solvency margins is
deliberated.

(d) Correlation with economic cycles A question of importance

is whether or not the inflation variations are correlated with the
general economic booms and recessions, which is one of the main
reasons for the cycles of the claim frequencies. Some past ex-
perience has been that a boom also gives rise to increased inflation
and a recession to low inflation. This is quite natural because during
the boom the demand for commodities is high. This increases
the demand of manpower which can cause pressure for increased
wages and salaries, and hence in turn people have more money
available to buy commodities. All these features are apt to provoke
inflation. Tendencies are reversed during a recession. If, as this
experience suggested, booms and recessions are synchronized with
inflation, then these factors reinforce each other in a way which is
significant for solvency considerations. In the figures, which will be
given as examples in the following, this kind of correlation was
assumed. The experimental formula
(7.6.2)
was introduced, where z(t) is the cyclic variable determined as
presented in items 6.1 (d)-(t). The auxiliary condition iJt) ~ i)2
can be justified by the experience that the rate of inflation may not
totally disappear even in a period of recession. ci is another model
parameter, subject to free selection.
It must be appreciated, however, that the correlation between
inflation and the economic climate of the national economy men-
tioned above has been considerably disturbed in recent years
following the well-known 'oil crisis'. In contradiction to the old
rules recession and high inflation have appeared simultaneously in
some countries; this is so-called 'stagflation'. Because the purpose of
this book is only to present risk-theoretical techniques, further
discussion of national economic phenomena is beyond the scope of
this book. The models obviously must be capable of working with
7.6 EFFECT OF INFLATION 271

any assumption according to the choice and judgement of the user.

It is, of course, often advisable to experiment with different assump-
tions in the evaluation of the numerous background features, and
to alter among other things the level and character of inflation.

(e) Effect of the length of the time lagtp was studied assigning to
it values 0, 1 and 2 years respectively; a sample path was driven
for the each of them, keeping the sequence of random numbers
the same (Fig. 7.6.3). The change of inflation was now programmed
according to (7.6.2) with cj = 0.5.
As expected, the time lag gives rise to considerable loss when
inflation is in its increasing phase in this case also. The loss is
compensated during the decreasing phase. The assumed syn-
chronism enlarges the amplitude of the cycles, as concluded easily
from (7.6.2) and as seen in Fig. 7.6.3. In effect, about the same

°
outcome as given by the synchronization (7.6.2) can be achieved by
taking cj = and by making the amplitude zm larger, which
procedure somewhat simplifies the considerations. This was given
as another alternative for zm in the standard list of item 7.1 (d).
More details on this relationship can be found in Pentikainen
(1982, item 4.2.6.4).

1.5
U(tJ

1.0

0.5

0.0 +-----~--~- _ _~_ _ _~---__

5 10 15 20 25
t
Figure 7.6.3 A sample path driven for three inflation time lags. The rates of
inflation were synchronized according to (7.6.1) and (7.6.2).
272 APPLICATIONS RELATED TO FINITE TIME-SPAN T

1.5

"o
;::
~
>-
~ 1.0
>
(5
Ul

0.5

5 10 15 20 25
Time

Figure 7.6.4 For examination of the effect of the time lag tp it is here removed
from the process that was exhibited in Fig. 6.8.5; otherwise the two processes
are the same.

The effect of the time lag is examined also by simulating the

stochastic bundles, first assuming the time lag and then omitting
it. The former case has already been given in Fig. 6.8.5, where tp = 2
years. The latter is depicted in Fig. 7.6.4. As expected, the waves
of the stochastic bundle were clearly smoothed. This is reflected
also in the number of ruins, which was reduced from 18/100 to
0/100.

7.7 Dynamic control rules

(a) Market pressures and self-control It was stated conclusively

in Section 2.7 that claim frequencies are subject to trends, periodic
variations and 'pure' random fluctuation. The trends were in-
corporated into the model in item 6.l(c). The periodical variations
were approximated by sinusoidal waves (6.1.3) in the previous
sections. Furthermore it was stated in Section 6.2 that the biases
and inaccuracies in premium rates are another source of variations
in underwriting results. The third source of variation is the yield of
interest, which was incorporated in the model, amalgamating it to
the joint control coefficient r igp together with inflation and growth
7.7 DYNAMIC CONTROL RULES 273

in item 6.5(g). Ultimately the different effects are concentrated in

the modified (total) safety loading A as defined by (6.5.12).
The somewhat schematic assumptions accepted in the foregoing
can now be improved, taking A as a central control variable. In
principle it is under the control of the insurer in premium rating.
However, in practice this control may be hampered and delayed by
competitive market pressures, by the uncertainty of the actual
level of profitability and by the technique involved in the rate
calculation and their implementation. If the rates are controlled
by supervising authorities, still further delays and restrictions may
arise.
The evidence of market pressures can be revealed when the busi-
ness flow of all insurers operating in the same market is analysed.
A quite clear synchronism of the profitability waves is frequently
observed, resulting in waves also in joint statistics of the total
amounts of claims and underwriting profits or losses (see
Pentikiiinen, 1982, Fig. 2.2.4). This is a general feature of the
insurance market - as the markets of other industries - that tends
to generate cyclic variation. If the profitability is good for a certain
period, intensified competition is provoked, marketing costs increase
and reductions in the rates are used as a competitive tool. As a
consequence profitability is rapidly worsened and safety loadings,
even if possibly still positive, are no longer adequate to provide
increases in the solvency margin in keeping with the growing volume
of business, the growth being caused by inflation and real growth
together. Then solvency ratios u(t) = U(t)/B(t) are generally de-
creasing, and this continues until eventually the presence of too
many insurers becomes critical. Competitive pressures are then
reduced and amendments of rates become inescapable, so giving rise
to an upswing and to a new cycle wave.

(b) Dynamic control This reasoning suggests an extension of the

model by rules which simulate the assumed behaviour of the market
and insurers. This is done by letting the safety loading A depend on
the current level of the solvency ratio u(t).
An example of a rule realizing this idea is given in Fig. 7.7.1 and
formulated as follows
A(t) = {Ao - c2 (u(t - t 1) - R 2 ): (7.7.1 )
Ao + c 1 (R 1 - u(t - t 1» .
 274 APPLICATIONS RELATED TO FINITE TIME-SPAN T

U It)

A ItJ< i\o
~--~~~~-------------------------------~

------------------------------R;

~----------------------~--~-------------~

t
Figure 7.7.1 Dynamics of the safety loading.

If the solvency ratio u(t) exceeds a given limit R2 then the safety
loading is reduced after a time lag t 1 • On the other hand, if u(t)
drops below another limit Rl then the safety loading is enhanced.
Here AO is a target value of the safety loading and c1 and c2 free
parameters. The changed safety loading is valid either until (accord-
ing to the rules (7.7.1)) another change is coming, the absolute value
of which is larger than the previous one, or until the solvency ratio
falls in the normal zone, defined by limits R'l' R~. In this zone the
safety loading has the target value AO'
The rule described is an example of the so-called dynamic or
adaptive processes. The process is made self-correcting by means
of ,autoregressive' rules (7.7.1).
The rule suggested was made fairly simple because the aim is
mainly to illustrate the control technique. The applied simulation
procedure also allows, without serious complications, more sophi-
sticated systems; e.g. the programmed changes in A (and possibly
in other control variables, too) may depend on the profits or losses
of the previous accounting years or on some joint combination of
the profitability and solvency position. For example, the works of
7.7 DYNAMIC CONTROL RULES 275

Frisque (1974), Pentikiiinen (1978), Bohman (1979), Martin-Lof

(1983) and Balzar and Benjamin (1980, 1982) can be referred to.
Of course, the actual possibilities and ways of controlling the rates
and other relevant factors should be considered when models
are planned.

(c) Sensitivity tests Experiments to show the effect of the dynamic

rule are illustrated in Figs 7.7.2 and 7.7.3, where first the lower
control limit and then the upper one were introduced.
The limits of the normal zone were assumed to be
R'l = + 0.3R 2
0.7Rl
(7.7.2)
R; = 0.3R 1 + 0.7R 2 .

(d) Simulation The effect of both rules is illustrated in Fig. 7.7.4.

It exhibits the same process as Fig. 6.8.5, but the dynamic rules are
now incorporated into the model, the limit values being as presented
in the caption to the figure. The number of ruins is expectedly
essentially reduced, being only 1 against 18 of the total number of
100 sample paths in Fig. 6.8.5. The effect of the dynamic control
can also clearly be seen in the form of the stochastic bundle. In

UlfJ

1.0

0.5

0.0 .j.---~---~---~---~--_
5 10 15 20 25
f

Figure 7.7.2 Effect of the lowerlimit RI which is either 0.35 or O. The coefficient
c 1 = 0.3.
 1.5

u(tJ

0.5

o.o~------~--------~-------- __--------~------~
5 10 15 20 25

Figure 7.7.3 Effect of the upper limit R2 = 1.0; c2 = 0.3.

1.5

o
:c
~
>-
u
"~ 1.0
o
If)

0.5

5 10 15 20 25
Time

Figure 7.7.4 A Dynamically controlled process; RI = 0.35, R2 = 1.J, c 1 =

c2 =0.3,t l =1.
7.7 DYNAMIC CONTROL RULES 277

Fig. 7.7.4 the bundle for larger t values is significantly narrower

than the bundle in Fig. 6.8.5. The confidence boundaries were cal-
culated disregarding the dynamics.

(e) Comparison of the deterministic and dynamic control In order

to study the effect of the foregoing dynamics the deterministic
cycles were relaxed (the cycle variable z(t) was put == 0) and the
standard data was changed to make r.'gp > 1. Figure 7.7.5 exhibits a
case where one sample path was generated by providing the dyna-
mics (7.7.1) and by omitting it. Furthermore, 100 samples of a
similarly dynamically controlled process are generated in Fig. 7.7.6
to give an idea of the character of the process concerned. It is seen
that the dynamic control of the safety loading results in a cyclic
variation of the solvency ratio u(t), as has already been anticipated
in item 6.2(e).
It can be expected that processes being mixtures of the above
types, i.e. allowing outside impulses such as those caused by the
cycles of the national economy, the profitability waves of the in-
surance market and the current status of the insurer concerned,
may be best suited to describe the behaviour of the flow of the
2 1. 5
:)

o
.r:
~
>-
u
c:
~
3i 1.0
------- ------------- -------------R2

0.5

5 10 15 20 25
Time

Figure 7.7.5 A sample path generated by providing the dynamic control

(7.7.1) (solid line) and by omitting it (dotted line). R, = 0.3, R2 = 0.9, rj = 1.1,
rx = 1.05, rg = 1, zm = 0, Ap = -0.05; these parameter values result in rjgp =
1.048 and A = 0.11.
278 APPLICATIONS RELATED TO FINITE TIME-SPAN T
1.5

o
;:;
1:'
~
c:
">
01.0
(J)

05

Ruin barrier

5 10 20 . 25
TIme

Figure 7.7.6 A stochastic bundle exhibiting a similar process to Fig. 7.7.5.

trading result and the solvency ratio. The simultaneous applica-

tion of the approaches discussed offers frameworks for this. Figure
7.7.4 illustrates an example where both the (deterministic) cycles
of the basic probabilities and the cyclic control (7.7.1) are present.

(1) The case r > 1 In most of the foregoing examples the important
accumulation factor r igp was < 1 according to the standard bases
given in item 7.1(d). In essence, the same simulation technique and
many of the straightforward calculations as employed earlier are
equally applicable for processes where r igp > 1, i.e. in cases of low
inflation and low real growth of the portfolio. This is demonstrated
in Figs 7.7.5 and 7.7.6. As seen in Fig. 7.7.5, the uncontrolled flow
of u(t) tends to infinity (dotted line). This is, of course, an unrealistic
situation in view of the applications. Hence, additional assumptions
are unavoidable to make the model workable in those cases where
rigp> 1. The rules (7.7.1) suggest an approach for the purpose. More
sophisticated models will be discussed in Chapter 10.

7.8 Solvency profile

(a) Purpose of profiles The structure of the model developed in

Chapter 6 was studied in the previous sections, selecting sets of two,
7.8 SOL VENey PROFILE 279

three or four variables in turn as moving variables and investigating

their interdependences and tie-ins. Owing to the large number of
variables, it is, however, difficult to get an idea of the model's struc-
ture as a whole in that way. Therefore an attempt is now made
to find a general view about the properties of the model. For this
purpose the minimum initial solvency ratio u will be calculated for
numerous combinations of the model variables and then plotted as
horizontal columns in one and the same graph. Thus it is possible to
show in a single picture the sensitivity of the solvency to the different
background factors. The idea is the same as in Fig. 4.2.7 where only
the one-year time-span was considered.

(b) The influence of the basic assumptions of the model is illustrat-

ed in Fig. 7.8.1.
First only the number of claims is assumed to be a random variable;
claim sizes and all other aspects are constant. As expected, the
necessary minimum solvency ratio can be quite small- in this
example, 8.5%.
The next step is to randomize also the claim size. After that the
short-span variations of basic probabilities are introduced and
then the long-term cycles. Finally inflation and all the other assump-
tions are introduced as standards in item 7.1(d). The IO-year ruin
probability of 1% is assumed.
The figure again shows how significant is the assumption concerning

Claim numbers D 8.5

+
Claim size ~ 21.0

+
Structure variation 111111111111111111111111 24.7

+
Cyel es 1111111111111111111111111111111111111111111111111111111111111111111111111111 81.2

+
Inflation 111111111111111111111111111111l1li1111111111111111111111111111111111111111111111111111111111111111111 107.0

Figure 7.8.1 The dependence of the minimum solvency ratio on the different
basic assumptions.
 280 APPLICATIONS RELATED TO FINITE TIME-SPAN T

cycles. The rate of inflation also has a marked effect on the solvency
condition.

(c) A comprehensive profile The sensitivity of the minimum initial

solvency ratio u(O) to num.erous other background factors was
tested by applying the same technique as above in Fig. 7.8.2. A
number of values calculated for the standard insurer (as defined in
item 7.1(d» are exhibited. The text given in the figure details the
entries concerned, so no further explanations are necessary.
Because our purpose is only to illustrate the technique of analysing
and presenting the effects of various backgrounds, more details
are not dealt with. They can be found in Pentikainen (1982, Sections
4.2.15 and 4.2.2).

Standard insurer

1IIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIUllllllllllllllllllllllll107 Inflation
1
Insurer's size 1 ix = 5Of, ..•• 111111111111111""111""'111"'' ' ' "11"'' ' ' ' " 92
n 1 :I~Z: '~ .'. : : : : : : : : : : : : : : : : :~ :~ : : ~: : : ~: : : I ;~ 125
5000 ....... !!!!!!!!!!~!~!~!!!!~!!~~~=!~:::~!l!!:~!:'I:~~122
I
~~~·.·~·.·.··.·.lIllrn'=='III,£ffi~8 TIme lag between claim and premium inflation
o years 1iI111111111111111111111111!1!I!II!l1II1II11II 76 i
Ne~ retention 1 year .. Jllllllllllllllllillllllll!lllllllllll!!!Iilllillllll!! gl
M Cycle amplitude 1
0.1 ........... ~~ 1
0.5............. 1
1.0 ........ . 1
I 107
Size and net retention 1 iiiiiiiiiiiiiiiiiiiiiiiiiiii~_0IID149
1
n M I
5000 0.1 U1i1lu1l1l 11llDJ1i1lmJ1I1II1I1II1II1II1II11I1II1I1II1II 100 Cycle length
10 000 0.5 IIIIIIIlIIlIDJlIIlIIIIlIIlIIIIIIIIIIIUlUlillullllllllllll 100
20000 1.0 111111111111111111111111111111111111111111111111011111111111101
1
Time span T 1
1
1 year ... /OOOIDO 27 1 Safety loadmg Ab
2 years ..~ 45 1 1
5 years . 1111111111111111111111111111111111111111111111111111111111111111104
-0.10 ... "'I"''''""'''IIIIII'''"'''""IiIIllIll""lllllllllllll"i"1ll116
10 years ... 11111111111111111111111111111111111111111111111111111111111111111107 -0.05 ... IIIIIIIIIIIIIIIIUIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII 86 1
1 - 0.00 . UIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIIII 61 I
Structure variation 1 +0.05 .... -.um41 1
1
O"q Yq 1 Real growth 1
o
0 1111111111111111111111111111111111111111111111111111111111111100 1
0.05 0.25 II 1 1 1 II 112 ig =0 . 1111111111111111111111111111111111111111111111111111 85 1

~;.~~ ~:~ :::::::::~:~:~~~:::=l,~i~~"I~ 13 1

1 I II 1 '" 144 : 1~;i. :.l: : : : : : : : : : : : : : : : : : : : : : : : : : : : : : ~I ;I~I~I I 134
1
1

Figure 7.8.2 Values of the minimum initial solvency ratio u(O) for different
parameter sets. T = 10 years. Band M given in £ million.
7.9 EVALUATION OF THE VARIATION RANGE OF u(t) 281

7.9 Evaluation of the variation range of u(t)

(a) The problem setting It is useful to have at least an approximate

idea of the dimensions of the fluctuations of the solvency ratio u(t).
Every insurer must have the capacity and technique to meet fluctu-
ations, i.e. some kind of buffer is needed in the form of hidden or book
reserves or an opportunity to make use of flexibility in the valuation
of the assets and liabilities. In some countries, e.g. in Finland,
Germany and Sweden, this is taken into account in legal regulation
in the form of a special fluctuation (equalization) provision or
margin inside the technical underwriting reserves.
Two ways of setting the problem can be distinguished:
(i) The limits are to be found between which a stochastic bundle
like that shown in Fig. 6.8.5 can be expected to be located.
The range Rv in Fig. 7.9.1 iIIustrates the idea.
(ii) The maximum probable plunge during some period of length
T wiII be taken as a measure for the solvency margin. The
range Rs in Fig. 7.9.1 is an example.
Approach (i) is meaningful only when the stochastic bundle is
limited from above, i.e. is not tending towards infinity. This is
the case ifr jgp < 1 as was seen in item 7.1(e) or if some rule is assumed

::::
"o
:;:
e
>-
u

~
'0
V>

/
(t+s)

t+s Time t

Figure 7.9.1 Two ways of defining the range R of fluctuation.

282 APPLICA TIONS RELATED TO FINITE TIME-SPAN T

according to which the solvency ratio u(t) is kept under some

finite limit, e.g. by suitable application of bonus or dividend policies.

(b) The variation range R. is composed of the variation (= dp)

of the midline flu(t) ofthe bundle and of the breadth ofthe confidence
area of the bundle of the fluctuating u(t). The time t corresponds to
a peak of the bundle and t + s to the next minimum. The midline
can be obtained from (6.6.8) and is mainly due to the cyclical waves.
The breadth of the bundle can be expressed as multiples of the
standard deviations Y10"u(t+s) and YzO"u(t), where Yl and Yz are
the confidence coefficients and the sigmas can be evaluated from
(6.6.11). Owing to the skewness of the process and to the time
difference (s in Fig. 7.9.1), Yl and Y2 may deviate from each other
even though they are both usually approximately equal to Y =
N- 1(1 - 8). Hence
Rv ~ dflu(t, t + s) + 2YO"u(l, t + s). (7.9.1)
The notation introduced in Chapter 6 is made up by writing out,
if necessary, the ends of the periods concerned. It is appropriate
to choose s as half the cycle length (see (6.1.4))
s = T)2. (7.9.2)
The calculations can be made by means of the formulae of Chapter 6
in a way similar to that illustrated in the next item. Note that the
outcome depends on the time t in (7.9.1). It is natural to take t
fairly large so that the breadth of the bundle is already close to
its steady state dimensions (see (6.6.12)). Approximate rapid rules
are presented in Rantala (1982, Chapter 4). They are intended to
define a limit up to which a fluctuation buffer (equalization reserve)
can still be reasonably deemed to be a technical rather than a free
reserve.

(c) Solvency margin can be understood as the capital which an

insurer should have to meet adverse periods. Its minimum
amount, denoted by Rs in Fig. 7.9.1, must obviously be high enough
to cover by a safe probability the loss which may arise during the
unfavourable half of a cycle. This can be expressed in the terms of
the preceding item as follows
(7.9.3)
The limit Rs depends on all the basic variables (7.1.1) of the model.
7.9 EVALUATION OF THE VARIATION RANGE OF u(t) 283

Table 7.9.1 The minimum solvency margin (7.9.3) as function of some basic
parameters (7.1.1), T = Tz •

Zm Tz A ri rx rg tp Ci M Rs

0.10 12 0.040 1.085 1.090 1.061 2 0.50 1.00 0.91

0.00 0.040 1.085 1.090 1.061 2 0.50 1.00 0.43
0.10 .Q 0.040 1.085 1.090 1.061 2 0.50 1.00 0.65
0.10 12 0,176 1.085 1.090 1.061 2 0.50 1.00 0.35
0.10 12 0.210 1200 1.090 1.061 2 0.50 1.00 0.33
0.10 12 0.044 1.085 1.050 1.061 2 0.50 1.00 0.82
0.10 12 0.047 1.085 1.090 LOOO 2 0.50 1.00 0.75
0.10 12 0.040 1.085 1.090 1.061 Q. 0.50 1.00 0.68
0.10 12 0.040 1.085 1.090 1.061 2 0.00 1.00 0.61
0.10 12 0.040 1.085 1.090 1.061 2 0.50 0.40 0.86

It can be investigated by programming it using the formulae of

Chapter 6, which have already been specified in the previous item.
The time horizon T should be at least equal to half of the cycle
length (s in Fig. 7.9.1). If cycles are not assumed or if they are weak,
the breadth of the bundle is decisive and Rs can be determined
as the difference of the initial level u(O) and the lowest point of the
lower boundary of the bundle in time (0, TJ. Depending on the
type of process the outcome may be affected very much by the
choice of T, as demonstrated in Fig. 7.5.1.
Some examples are set out in Table 7.9.1. On the topmost line
is the standard set of values of the variables data as given in item
7.1 (d). Then each variable is changed in turn to ascertain the
sensitivity.

(d) Legal stipulations for the solvency margins were considered al-
ready in Section 4.4, deriving some rules on the basis of a one-year
planning horizon. The background of the short horizon is the view
that public intervention to control the insurance industry can be
limited mainly to safeguarding the interests of those insured.
Therefore only that margin is necessary which is needed to protect
the financial state of a possibly weak insurer from direct bankruptcy
for a 'warning period' during which the weakened solvency is to
be restored on pain of winding up. On the other hand a short
planning period is, of course, not sufficient from the point of view
of the insurer. The objective should be a safe continuation of the
284 APPLICATIONS RELATED TO FINITE TIME-SPAN T

existence of the institution, which provides a long time horizon.

Then understandably also the dimension of the minimum solvency
margin increases substantially, as already seen. Another reason
for the divergence of the outcomes given from conventional
approaches is the fact that the model assumptions applied in this
and in the preceding section are less restrictive than has been the
case in many of the previous studies; in particular, the introduction
of cycles and inflation proved significant.
Further discussion on the solvency margins and equalization
reserves can be found in Karten (1980) and Pentikiiinen (1982,
Chapter 5).

7.10 Safety loading

(a) Minimum condition The technique described can also be

used for evaluation of a minimum requirement for the safety
loading A. The principle is demonstrated in Fig. 7.10.1. An obvious
condition for solvency is that the stochastic bundle does not intersect
the ruin barrier ur ' The position of the bundle depends on the safety
loading A, understood as an aggregate loading including also the
yield of interest earned for the technical reserves as defined in item
6.5 (h). If Ais too small, the bundle falls down with fatal consequences.
The problem is to find the lowest value for A which still prevents
the bundle from hanging below U r • It is necessary to distinguish

Figure 7.10.1 The lowest still solvent positions of the stochastic bundle;
r igp < 1, rg > 1.
7.10 SAFETY LOADING 285

between the cases when the 'dynamic discount factor' rjgp (see
item 6.5(g)) is either < 1 or > 1.

(b) The case Yjgp < 1. The flow of the midline of the bundle was
given by (6.6.8) {see Fig. 6.6.2). Assuming a constant rate for interest,
inflation and growth it tends towards an equilibrium level Eu
(6.6.9). Furthermore, the measure for the breadth of the bundle
is required. It was studied in the previous section and denoted
t
by Rv' Clearly Eu must be at a distance at least Rv from the ruin
barrier, or in terms of the quantities given
Eu > ur + tRy.
Substituting (6.6.9) for Eu and solving an inequality for A, the
following is obtained
(7.1 0.1)
Assuming that ur = O.l, Rv = l.2 (see item 7.9(c)) and (according
to the standards in item 6.l (d)) rjgp = 0.938, a numerical value is
found for the lower limit of A ~ 0.043.
Another way to study minimum conditions for the safety loading
is to use the asymptotic expressions (6.6.9) and (6.6.12). This is
done in Exercise 7.l O.l.
The interpretation of this outcome is as follows. A minimum
amount of safety loading is necessary to counteract the eroding
effect of inflation and to increase the solvency margin in keeping
with the real growth of the business volume, in so far as the yield
of interest is not sufficient for these purposes, as is the case when
Yjgp < l. The situation is more complicated if the insurer can acquire
fresh capital to reinforce the solvency margin V, as may be possible
at least in the case of proprietary companies. Then the demand
of the safety loading which is needed to compensate the eroding
capital is smaller, but on the other hand the stockholders expect a
return for their investment, which increases the need for margins
or loadings in rates.
It will be recalled that the safety loading, as defined above, is
an aggregate quantity composed of the yield of interest earned for
the underwriting reserves and of the 'ordinary' loading accrued
from premiums according to (6.5.12). Furthermore, for the above
considerations the total amount of the safety income AB is relevant,
as was discussed in item 4.l(c), without paying any attention to the
286 APPLICATIONS RELATED TO FINITE TIME-SPAN T

problem of how it is divided between different policies and policy

groups. In essence, the solvency considerations discussed only
provide conditions for just this aggregate amount, not for the
individual rates, which are subject to aspects such as those dealt
with in Section 5.3.

(c) The case rigp > 1 According to (6.6.8) a safety loading of

A > - (r igp - l)u(O)
is sufficient to ensure that the midline of the stochastic bundle
tends to infinity. Note, however, that the presence of cycles may
cause irregular and even unlimited oscillations. If the lower con-
fidence boundary grows toward infinity and has a minimum, the
value of the solvency ratio u(t) at this minimum point depends
among other things on the safety loading, as shown in Fig. 7.10.2.
The required solvency condition is now that A or rather its premium
related component Ap should be larger than the critical value giving
the minimum (Ap = - 0.09 in the figure).

1.5 - 0.01 -0.05

Ap =+0.01 ~-0.031 -0.07 -0.09
u

1.0

0.5

5 10 15 20
Time

Figure 7.10.2 The lower confidence boundary of the stochastic bundle as

a function of safety loading Ap' The standard data, given in item 7.1 (d), was
changed by putting rj = 1.1, rx = 1.05 and rg = 1 resulting in rjgp = 1.048. Note
that despite the assumed negative values of Ap ' the total safety loading according
to (6.5.12) is strongly positive.
7.10 SAFETY LOADING 287

As discussed in item 7.7(0, the uncontrolled flow of the u(t)

when r igp > 1 is an oversimplification of the processes concerned
and the model should necessarily be supported by some control
rules. Then the problem may become more complicated than
before, but the same idea, i.e. finding the lowest safety loading
that still prevents the stochastic bundle from intersecting the ruin
barrier, can be expected to be applicable.
Some necessary conditions for the unlimited growth of the
lower confidence boundary of the stochastic bundle are dealt with
in Exercise 7.10.2.

Exercise 7.10.1 Find a lower limit for A. in the case where rigp < 1
by using the limit expressions (6.6.9) and (6.6.12) and applying a
similar method as in item (b). What is its numerical value calculated
for the standard data of item 7.1(d)?

Exercise 7.10.2 Assume constant rates of interest, inflation and

r;
growth so that r igp > 1, p = ri~: 1 < 1 and omit the cycles and the
structure variation. Find a condition where the lower confidence
boundary of the stochastic bundle, defined as in Section 6.6., grows
asymptotically and without limit when t tends to infinity.
 CHAPTER 8

Risk theory analysis of

life insurance

8.1 Cohort analysis

(a) Special character of life insurance The foregoing analysis

methods were developed to fit, as far as possible, all kinds of
insurance classes, although most examples were chosen from
non-life insurance. There is no doubt that risk theory, as it stands
today, can find its most rewarding applications in just that en-
vironment, but the fluctuation of life insurance business can also
be studied by the same technique. Of course, the individual risk
sums Z are the amounts payable S less the policy reserve V, i.e.
Z = S - V. There are, however, crucially different features in life
insurance, pension insurance included. Conventionally life insurance
contracts are issued to continue decades up to some termination
age or even lifelong without the possibility of adjusting the original
premiums upwards (unless according to some specified rules in the
case of indexation), whereas non-life premiums are normally valid
only for short terms. The special risk involved in life insurance
is the chance of future adverse development in bases relevant for
the solvent flow of the business, such as mortality, invalidity,
interest rate, inflation etc. As a protection against them the premium
rates are calculated with safe margins. If the margins are not needed
to cover losses, they can be returned as bonuses for instance as
premium rebates, additional benefits (to profit participating
policies) or they can be used to compensate inflation in case of
index-linked policies. It can be said that the practice in life insurance
is to have implicit solvency margins in the form of safe calculation
bases, whereas the solvency margins of non-life insurers are mainly
explicit, either shown in balance sheets or possibly also constituted
8.1 COHORT ANALYSIS 289

in underestimations of assets or overestimations of liabilities

('hidden reserves').
In the foregoing considerations the portfolio development was
approximated by the trend factor rg(t) and by the inflation factors
rx(t) and rp(t). In fact, this was justified only by assuming that the
rates are continually, even if after some time lags, adjusted for
changing risk structures and inflation. When the life insurance
business is dealt with then it may be necessary to accept the fact
that the old policies are frozen at their original level or at least
cannot fully follow the changing relevant circumstances. If changes
occur, for instance, due to inflation or other circumstances or
due to the action of the insurer, e.g. revision of premium rates,
they may affect old and new policies in significantly differing
ways. This suggests a more sophisticated modelling of the portfolio
development.
It is beyond the scope of this book to discuss the solvency
problems of life insurers as a whole. Only some hints will be given
as to how the risk theory technique described in the previous
sections may be modified for the analysis of life insurance. It is
also possible to link the presentation to the classical individual
risk theory which was developed specially for life insurance.

(b) Cohort approach A simplified problem is first considered,

which can then be used as a building block for more advanced
considerations. Consider that group of policies which have the
same entry year, entry age x, policy type and face sum. The scaling
will be made taking the entry year equal to 0 and the face
sum equal to 1. Endowment insurance will be considered as an
example where the same sum, now equal to 1, will be paid at death
or at maturity, after w years. Let 1(0) be the number of policies
( = policyholders) of this group at the beginning of the entry year.
Such a group is commonly called a cohort, and it will now be
followed from year to year assuming that the only way of exit from
the cohort is death. Then according to conventional life insurance
mathematics the cohort size is obtained from the algorithm
I(t + 1) = I(t) x (1 - q(t)) t = 0, 1, ... , w - 1, (8.1.1)
where q(t) is the mortality rate, i.e. the expected proportional
number of deaths in year t. A basic idea of life insurance mathe-
matics is to construct a policy (or premium) reserve V(t), where
290 RISK THEORY ANALYSIS OF LIFE INSURANCE

the difference between incomes and expenditures related to the

cohort are accumulated according to the algorithm
I(t + 1) x V(t + 1) = rj x I(t)
x [V(t) + B - C(t)] - I(t) x q(t), (8.1.2)
where V(t) is the policy reserve per living cohort member at the
beginning of year t, B is the premium calculated per cohort member,
C(t) is the expense loading, and rj = 1 + i j is the (actuarial) rate
of interest.
For simplicity of notation it is assumed that the premiums are
received and the expenses paid at the beginning of each year and
the benefits are paid at the end of the year. The sum 1 is paid at
the end of year w - 1, or what is the same, at the beginning of the
year w, to cohort members still alive. Then conditions
V(O) = 0 and V(w) = 1, (8.1.3)
determine the premium B (see exercise 8.2.1) and the reserve V(t)
uniquely.
This conventional deterministic calculation rule can now be
readily made stochastic by, as a first step, letting the annual number
of deaths
d(t) = I(t)q(t), (8.1.4)
be a random variable (with the expectation n = E(d(t)). Then also
the number of the cohort members obtained by algorithm
I(t + 1) = I(t) - d(t) (1(0) == 1(0)), (8.1.5)
will be stochastic. Furthermore, we can define the random profit
(or loss) Y(t) resulting from this cohort in year t and the interest
calculated up to the end ofthe year as follows
Y(t) = rjl(t) [V(t) + B - C(t)] - d(t) -I(t + 1) V(t + 1)
= I(t)[rj V(t) + rjB - rjC(t) - V(t + 1)]
-d(t)[I- V(t+ 1)]. (8.1.6)
As seen from (8.1.2), Y(t) == 0 if d(t) == d(t). Hence Y(t) has arisen
from the random deviations of d(t) from those values (8.1.4) which
were preassumed in the calculation bases when the premiums were
determined.
Furthermore it is useful to follow the accumulated cohort result
8.1 COHORT ANALYSIS 291

2.0
y(O,t)

5 10 15 20 25 35 40
Time t

Figure 8.1.1 The cohort random process (8.1. 7b) calculated using the following
assumptions and data: q(t) = 0.0006 + exp[O, lISt - 8.125],/(0) = 10 000, r i =
LOS, w = 40, d(t) simple Poisson variable with expectation n = I(t)q(t). Expense
loading C(O) = 0.02; C(t) = 0.002 for t ~ 1.

Y(O,t) = I Y(.)r:-<. (S.1.7a)

<=0

The random processes introduced can be handled using the Monte

Carlo method in the same way as considered in the previous sec-
tions. An example is given in Fig. S.1.1, where the ratio
y(O, t) = Y(O, t)/B(t) , (S.1.7b)
i.e. the accumulated profit in relation to the premium income
B(t) = l(t)B is displayed.
The spreading of the bundle in the figure is striking. It is due
to the interest calculated according to (S.l. 7) accumulating to the
randomly arisen profit or loss. Introduction of inflation, real
growth and lapses will fully change the character of the process,
as will be seen later.
This approach can be extended to any other type of policy, and
also to take account of the fact that the face sums are different
inside a cohort and that also bases other than the number of deaths
are varying, as will be shown later. As expected, the variation
range is considerably larger e.g. for cohorts consisting of pure
292 RISK THEORY ANALYSIS OF LIFE INSURANCE

death risk policies (temporary life insurance), and in particular

when sickness or disablement benefits are involved.

8.2 Link to classic individual risk theory

The cohort approach was the basis for the consideration of the
classic, often called individual, risk theory, which is reviewed for
instance by Bohlmann (1909) and Cramer (1930). The practice
was to discount all the expected fluctuations to the entry time of
the cohort
w-l
Yo = Y(O, w - l)v W = L Y(t)V +
t 1, (8.2.1 )
t=O

where v = l/r j is the discounting factor. The uncertainty involved

in the business was measured by the standard deviation of Yo and
the normal approximation was assumed applicable, when the
dJ. of Yo was introduced.
The consideration is restricted in this section only to the net
incomes and outgoings, omitting the expenses and loading (term
C(t) in (8.1.2)). Furthermore, only the number of deaths is assumed
stochastically varying; all the other bases of calculations, such as
interest rate, are assumed to be set exactly according to the deter-
ministic assumptions which were used for premium calculations.
The variance (J~ = var(Y o)' which was a central characteristic
of individual risk theory, is readily deduced from the above issues.
Note that the annual terms Y(t) in (8.2.1) are not mutually in-
dependent because the number of deaths in each year controls
the cohort size of subsequent years. Therefore the required variance
is not directly obtainable as a sum of the variances of the terms.
In order to get a more general result a hypothetical single premium
policy is considered, on the basis of which are paid to the cohort
members still living both an annuity a at the beginning of years
0, 1, ... , w - 1 and, in case of death or of maturity, a sum S at the
end of the relevant year. The single risk premium, i.e. the discounted
sum of the expected future outgoings, is then according to conven-
tionallife insurance mathematics and adopting a discrete calculation
method

(8.2.2)
8.2 LINK TO CLASSIC INDIVIDUAL RISK THEORY 293

The last term is due to the maturity at the end of year w - 1. The
terms related to expenses (see (8.1.6» do not appear, because the
consideration is now limited to risk premiums only.
Note that putting P 1 = 0 and a = - P the conventional endow-
ment treaty with all term risk premium P is obtained as a special
case, and putting S = 0 the capital value of an annuity a.
Now consider one policy entering into the cohort at t = O.
There are w + 1 different ways it can terminate, i.e. either death
in one of the years t = 0, 1, ... , w - 1 or maturity at the beginning
of the year w. These events can be described letting the termination
time t be a discrete random variable with frequencies
d(t) l(t)q(t)
pet) = prob {t = t} = 1(0) = l(O) (t = 0, 1, ... , w - 1)

let)
(t = w).
1(0)
The second line represents the case of maturity. Noting that
lew) = lew - 1) x (1 - q(w - 1», and combining the events related
to the end of the last policy year w - 1 and the beginning of the
next year, these expressions can be rewritten as follows
( ) _ d(t) _ l(t)q(t)
p t - 1(0) -l(O) for t = 0, 1, ... , w - 2

( _ 1) _ dew - 1) lew) _ lew - 1) .

(8.2.3)
pw - 1(0) + 1(0) - 1(0)

Next a profit (or loss) Z(t) arising from termination at year t can
be introduced
1

Z(t) = P 1- a L v' - Sv l + 1 (t = 0,1,2, ... , w - 1)

,=0
1 - vl + 1
= P1 - a - Sv l + 1 (8.2.4)
I-v

=P 1 - a (S -I-v
[--+
I-v
-a)
- v1+1J .

The different profit outcomes can be understood to constitute

another discrete random variable Z, obtaining values (8.2.4) by
probabilities (8.2.3). Its moments about zero are, according to the
standard definitions
w-1
E(Zi) = L p(t)Z(t)i. (8.2.5)
1=0
294 RISK THEORY ANALYSIS OF LIFE INSURANCE

Since PI in (8.2.4) is the pure risk premium, E(Z) = 0, the variance

IS

(8.2.6)
Introducing the notation
w-l
A(v) = I p(t)v l + 1, (8.2.7)
1=0

the single risk premium can be written

PI = L
w-l [I
p(t) a -
vl + 1
+Svl+ 1
]

1=0 I - v

= Wfp(t)[_a
1=0 I - v
+(s __v)v1+1] a
I -

( a )W-l (8.2.8)
a
=--+ s--- I p(t)V + 1 I

I-v I-v 1=0

= _a +
I-v
(s __ I-v
a )A(V),

and the following classic expression for the variance (J2 can be
derived by some algebra (exercise 8.2.2)

(J2 = (s - _a_)2 x (A(v 2) - A(V)2).

I - v
(8.2.9)

Because the discounted cohort profit Yo is the sum of the profits

(Z) of all the 1(0) cohort members, and since these can be assumed
mutually independent, 'the cohort variance' is
(J~ = var(Y o ) = 1(0)(J2. (8.2.1 0)
The following special results are now obtained from this general
formula:

(i) In the case of the endowment policy with whole term premiums,
put a = - P where P is the pure annual risk premium (B
without expense loading C); now.P 1 = O.
(ii) In the case of single premium endowment treaty put a = o.
(iii) In the case of a single premium annuity ( = a) put S = O.

For other types of life policies similar formulae can be derived.

8.3 EXTENSIONS OF THE COHORT APPROACH 295

The standard deviation a and the ratio aI A(v) or preferably the

per cohort calculated values a e and Ae may be used in evaluation
of the need of safety loading in premiums and reserves.
In the example given in Fig. 8.1.1 a = 0.103, A = 0.168 and
a I A = 0.613. If the cohort size is 1(0) = 10000, then the corres-
ponding cohort characteristics are a e = aJ 10 000 = 10.3 and
Ae = 10 OOOA = 1680, and hence the 'relative' standard deviation
ae/Ae = aIAJI(O) = 0.006.

Exercise 8.2.1 Write down the formula for risk premium P (without
loadings) of the endowment policy.

Exercise 8.2.2 Prove (8.2.9).

8.3 Extensions of the cohort approach

(a) Generalized profit algorithm The classic model can be

developed to consider and to test the effects of a great number of
disturbances, basic variations, impulses, trends etc. other than
those arising from the simple fluctuation of the number of deaths
considered in Section 8.2.
The gross premium basis is returned to again, and the profit defi-
nition (8.1.6) will now be written in a more general form
Y(t) = l(t)[rj(t)V(t) + rj(t)B - V(t + 1)]S(t) (8.3.1 )
- d(t)[l - V(t+ l)]S(t) -1(t)C(t)rj(t).

The current insurance sum is now S(t) (hence no longer normed

to unity but still the same for all the cohort members). However,
the reserve V and gross premium B still have the values which
are calculated per unit sum by the primary actuarial bases. The
sum S(t) can change from year to year by virtue of bonus or indexa-
tion. The interest factor rJt) should represent the evaluated or
anticipated actual yield of investment, and it is normally greater
than the primary actuarial rate r j • Also the number of deaths d(t)
can be simulated in a way which corresponds to the anticipated
actual level.
The variable C(t) represents the anticipated actual expenses
per policy in force. It can be expected to change in pace with
inflation.
296 RISK THEORY ANALYSIS OF LIFE INSURANCE

(b) Second-order bases Formula (8.3.1) now offers an instrument

for investigation of the consequences of deviations of the actual
relevant factors like interest rate, mortality and many others from
those values which were primarily assumed when the premiums
were determined at the entry of the cohort.
(i) The actual yield of investment can be taken into account in
(8.3.1) by the rates rJt) which are fitted to the actual average
yield of investments. If, as is usual, rj(t) > rp 'interest profit'
results.
(ii) Mortality deviations from the calculation bases can be in-
corporated into the model, choosing the simulation rule for
the generation of the number of deaths d(t) accordingly. If,
for example, the Poisson law is assumed, then the parameter n,
the expected number of deaths, should be defined by means
of some function q'(t) which is fitted to the actual or anticipated
mortality more accurately than the calculation basis q(t).
If q' < q, then mortality profit arises.
(iii) Cancellations of policies will be introduced next, defining a
'lapse ratio' p(t). It indicates the proportion of the cohort
size I(t) which cancels its policies at year t before the maturity
age. Then the algorithm (8.1.5) is replaced by the following
I(t + I) = I(t) - d(t) -I(t)p(t). (8.3.2)
The lapse ratios should be given as input data. Of course,
the stochastic definition is also feasible. Furthermore 'a lapse
profit' ( ± )
(V(t) - Vs(t))S(t), (8.3.3)
is to be added to the profit equation (8.3.1). Vs(t) is the surrender
value per unit, i.e. the amount Vs (t)S(t) is to be paid, if positive,
when the policyholder breaks the contract before maturity.
Then the reserve V(t)S is released and the difference (8.3.3)
constitutes the profit or loss arising.
(iv) Inflation will be introduced next, denoting its rate by ix(t).
Perhaps the most crucial point is that the expenses increase
in keeping with inflation, but the premiums are either fixed
or - according to the practice in some countries - they may
be increased by virtue of index linkage but possibly with a
time lag. It may be useful to define separately a 'premium
inflation' rate ip (t) for life insurance as was the case for general
8.3 EXTENSIONS OF THE COHORT APPROACH 297

insurance in item 6.2(a). Of course, if no indexation of the

policies is applied, then ip is zero.
Now the sum insured S in (8.3.1) should be made time
dependent according to the algorithm
S(t + 1) = rp(t)S(t), (8.3.4)
where rp (t) = 1 + ip (t). An example of the definition of rate
ip (t) will be given in point (vi).
The expenses are naturally linked with the 'general inflation'
(see item 6.1(k))

C(t) = C(t) nrJr),

t

t= 1
(8.3.5)

where rx(t) = 1 + ix(t). Again prognoses or assumptions about

the inflation rates are needed to make the algorithm (8.3.1)
operable.
(v) Bonuses are commonly given to the policyholders when the
safely determined calculation bases provide excess profit.
Definition for them can be added to the 'profit algorithm'
(8.3.1). A simple example may be a rule according to which
a bonus is given, if the accumulated profit Y(O, t) (see (8.1.7a))
exceeds some barrier. The amount of bonuses should be
deducted from the r.h.s. of (8.3.1).
The bonus barrier controls the allocation of the annual
profits between their reservation for future use, to cover any
adverse period of the business, and the return of a part of
them as bonuses to the policyholders (or as dividends to the
shareholders). What is left of the reserve up to the maturity
as an accumulated not a distributed profit is the final profit
(or loss if negative) arising from the cohort concerned.
(vi) Indexation Another practice is to use investment and other
profits to support index linkage. This is the case in some
so-called equity-linked or similar policies. The idea is to use
that part of the accumulated profit Y(O, t) that exceeds some
limit for enforcement of the policy reserves, and the insurance
sum and the current premiums are simultaneously increased
accordingly. This is conveniently implemented by ruling some
suitable increment for the insurance sum in (8.3.1)
S(t + 1) = rp(t)S(t), (8.3.6)
298 RISK THEORY ANALYSIS OF LIFE INSURANCE

where rp(t) = 1 + ip(t) is a growth factor, which should be

determined so that the financial balance is preserved. In fact
this means that the increment of the policy reserve
l(t + I)V(t + 1)(S(t + I) - S(t)) = l(t + I)V(t + l)ip(t)S(t) of the
cohort should be equal to the amount which is released from
the accumulated profit Y(O, t). The latter amount can be pro-
grammed into the system assuming a bonus barrier the excess
of which will be given away from Y(O,t). It can be introduced
by a constant b which defines the bonus barrier in proportion
to the premium income of the cohort. Hence we have, given
that S(t) and i/t) are now stochastic
l(t + I) x V(t + I) x ip(t) x S(t)
= Y(O, t) - bB x l(t + 1) x S(t), (8.3.7)
from which ip(t) is obtained by solving if the right hand side is
positive, otherwise put ip (t) = 0. Note that ip (t) is a direct
analogy for the 'premium inflation' rate which was introduced
in item 6.2(a).

The above ideas are illustrated in Figs 8.3.1, 8.3.2 and 8.3.3.
In order to get a general idea about the effect of the background
assumptions, first the process was made deterministic in Figs
8.3.1 and 8.3.2. Explanations can be found in the captions of the
figures. Figure 8.3.2, demonstrates how the model can be used to test
what degree of inflation the supposed secondary bases still tolerate.
Then one ofthe processes was selected for Monte Carlo simulation
in Fig. 8.3.3. Both the deterministic average flow without bonus
barrier and the bundle of stochastically varying flows provided
with bonus barrier are exhibited. The hypothetical example set
out in Fig. 8.3.3 illustrates a case where the accumulated profit
was not sufficient to cover the adverse development at the final
phase before maturity. It arose in this example mainly due to the
fact that owing to inflation the 'actual' expenses widely exceeded
the loading calculated for them.
It is seen that the system is sensitive to changes in the primary
bases, whereas the stochastic variation of the number of deaths
has a minor though noticeable effect.
The steepness of curves 1 and 3 confirm the well-known
experience that in particular the interest rate and inflation are of
crucial significance for the long-range balance of that type of life
 y(O,t)
/
1.5 4

1.0

0.5

0.0

-0.5

-1.0

3
-1.5

5 10 15 20 25 30 35 40
Time

Figure 8.3.1 Deterministic flows of accumulated cohort profit when the

basic assumptions given in the caption of Fig. 8.1.1 are changed one at a time
the others being the originals. (1) Interest rate ri(t) = 1.09; (2) lapses are assumed
p(l) = 0.16, p(2) = 0.11, p(3) = 0.06 and p(t) = 0.01 for t> 3, Vs(t) = V(t) if
positive, otherwise 0; (3) iriflation rx = 1.05; (4) mortality rates q are lowered
by 20%; (5) assumption 4 combined with a bonus barrier b = 0.2.

2.0
Y (O,t)
1.5
9 9

1.0

0.5

0.0

-0.5
11

-1.0

-1.5

5 10 15 20 25 30 35 40
Time

Figure 8.3.2 Deterministic flows of the accumulated cohort profit. Interest,

mortality and lapses as in Fig. 8.3.1, and the rate ix of iriflation varying as depicted
in the figure.
300 RISK THEORY ANALYSIS OF LIFE INSURANCE

2.0
y(O,t)

1.5

1.0

0.5

0.0

-0.5

-1.0

-1.5

-2.0 iii ( i i
5 10 15 20 40

Figure 8.3.3 Flows of the accumulated cohort profit. The curve with circles
is deterministic, assuming no bonus barrier. The other curves are simulated
letting deaths be randomly varying according to the Poisson law and adopting
a bonus barrier b = 0.2, rj = 1.09, rx = 1.09, mortality and lapses as in Fig. 8.3.1.

insurance which involves a substantial saving element, such as the

endowment insurance shown in Fig. 8.3.1.

8.4 General system

(a) Portfolio as a sum of cohorts The life insurance portfolio is,

of course, always composed of a number of cohorts of the type
defined in the previous sections. The technique drafted earlier
can be applied separately for each of them and then the data for
the whole business can be obtained summing the cohort outcomes.
The cohorts are specified by
(i) the entry year tc
(ii) the entry age x
(iii) the type ofthe policy including also the maturity age
(iv) the initial insured sums Sc.
In other words, policies having (at least approximately) the
same characteristics constitute a cohort, which will be followed
up for the whole time they are in force.
8.4 GENERAL SYSTEM 301

........
o
.t:.
o
U

f_.., __ _
------- - -- --------- I

L_.-_
I

Current time

Figure 8.4.1 A portfolio consisting of cohorts (horizontal pillars). For clarity

in the figure it is assumed that all the policies entering in each year tc constitute
only one cohort. The total portfolio profit is obtained by summing the profits
emanating from all the cohorts in force that year. Every year a new cohort is
assimilated into the portfolio and the oldest one passes away.

The current time will be denoted by t, putting the order number

of the starting year of the analysis at O. Then the entry years tc of
the cohorts emanating from earlier years have negative values.
For the portfolio analysis it is necessary to have the number
of the members of each cohort as an initial datum for t = o. Pre-
ferably these numbers can be derived from the actual files. In
theoretical considerations they and the initial sums S can also
be derived making some assumptions concerning the original
initial sizes (for t = t c ) of the cohorts, past lapse rates, inflation,
bonuses or the index i p .
When the initial data are obtained, then each cohort can be
operated for t = 0, 1,2, ... as before and the results related to each
year t are summed in order to get the underwriting profit of the
year t. Furthermore, each year a new set of cohorts will be assimilated
into the portfolio and another set of cohorts is removed due to
maturities. For new cohorts some assumption of the acquisition of
new policies is needed, possibly characterized by a growth factor
rg(t). Furthermore, the average values and distributions of the
sums insured for all the entering policies are needed. Possibly
the distributions can be assumed to be unchanged in shape from
year to year, but the average sum and equally the scale of the
variable S will be changed according to some index, perhaps to the
302 RISK THEORY ANALYSIS OF LIFE INSURANCE

same inflation rate ix(t) as already introduced above. This means

that the sums S(t + 1) of the policies entering the portfolio in year
t + 1 are equally distributed as the sums S(t) of entries of year t,
but so that on an average the sums are increased in the
ratio rx(t) = 1 + ix(t).
Practical difficulties may arise because the number of cohorts
to be constructed and followed has the tendency to grow quite
large if the characteristics listed are very strictly applied. It is
likely that, without loss of validity of the model, the number of
cohorts can be kept reasonable by combining into one and the same
cohort, policies having similar characteristics, like entry age, sums S,
etc., i.e. dividing the portfolio into classes according to the character-
istics and applying a fairly coarse class interval. A further rationaliza-
tion may be to combine in one and the same cohort all the policies
of equal type irrespective of the size of the insured sum S and to
introduce the sum dJ., as shown in the next item.

(b) Examples The idea outlined is illustrated in Figs 8.4.2 and

8.4.3. The examples are still simplified in so far as all the new policies
of each year are assumed to be of endowment type having the same
entry and maturity age, but variation of insurance sums inside
the cohort was allowed according to the Pareto law (see Section 3.5.8)
with ex = 3. It was further supposed that the original dJ. of the
sums with original parameter ex will be preserved inside the cohorts
from year to year, but so that the average sum S(te' t) is changing
and the S scale is set accordingly. The parameter Zo in (3.5.20)
was determined by the condition that the average insurance sum
should coincide with the assumed cohort mean value S(te' t), i.e.
Zo = [(oc - 1)/oc]S(te, t), where t is the current year and te the entry
year of the cohort. The mean sums of new cohorts compared with
the next preceding year were assumed to be increasing in pace with
inflation and the sums of 'old cohorts' were increased by the factor
rp = 1 + ip defined by (8.3.7). Hence, by these assumptions we
have

n
t
S(te' t) = S(O, 0) x r!c x rlr), (8.4.1 )
't=tc+ 1

where te = - w + t, - w + t + 1, ... , t, being the termination time

of the cohort concerned (see (8.1.3)).
8.4 GENERAL SYSTEM 303

Furthermore the initial cohort sizes were assumed to grow

by the factor rg from year to year. That means that the policies
entering or having entered the portfolio in year te have the initial
size [(t e, t) = [(0, O)r~c, and after that the cohort size [(t e, t) is changed
according to algorithm (8.3.2), i.e. members are departing from
the cohort before maturity owing to either death or lapse.
These assumptions imply that the example considered is still
rather simplified. All the policies were assumed to be of endowment
type having the same entry age and maturity age. All the new
policies of each year were combined in one single cohort but the
spread of insurance sums S was allowed inside the cohort. The
initial values of cohorts were obtained by making assumptions
about the (past and future) acquisition of new policies, inflation
and indexations.
First again some deterministic flows of the accumulated solvency
ratio are displayed in Fig. 8.4.2. This figure demonstrates how the
model can be used to test how serious an inflation the system still
tolerates. For this purpose different values are assigned for the
inflation rate ix as seen in the figure.
The case ix = 9% and bonus barrier b = 0.2 is simulated in
Fig. 8.4.3.

2.0

0
:i: 1.5
E ix x100= 5
>-
u

""
1.0
>
'0
Vl
0.5 8

0.0
10

-0.5
11
-1.0

-1.5

-2.0
5 10 15 20 25
Time t

Figure 8.4.2 Deterministically calculated flow of solvency ratio. The rate of

inflation ix = rx - 1 is varying, the interest rate is 0.075, other data are as in
Fig. 8.3.1. The bonus barrier b is not adopted.
304 RISK 1HEORY ANALYSIS OF LIFE INSURANCE

~ 0.5
::J
o
e
>-
u
c
~
o
<f) 0.0 j----- ----- ____________ .. ____ ... ___ .. ___________________""

-0.5

-1.0 +-----~---~---~---~--- ...

5 10 15 20 25
Time t

Figure 8.4.3 Simulation of the case inflation rate ix = 0.09 and the bonus
barrier b = 0.2. Other data are as in Fig. 8.4.2.

These examples are still simplified such that the study of the
effect of the differing entry ages and different types of policies
was omitted. However, it can be expected that these examples
already reveal typical features in life insurance structures. As
is well known from earlier experience and already stated in Section
8.3 on the basis of an analysis of one single cohort, the long-range
validity of the basic data like interest rate, mortality, loading and
the effects of inflation are crucial, whereas the 'ordinary' random
fluctuation has minor, even though not negligible, dimensions
compared with the effects of the basis aspects. For pure risk in-
surance (temporary life insurance) and disablement benefits the
latter variation range is expected to be larger than in the present
examples.
The model can be further developed regarding all the aspects
which specify a cohort. Furthermore, time-dependent changes in
relevant factors, for instance inflation, interest rate, mortality,
lapse rates, etc. can be programmed and the sensitivity of the
outcomes of each can be investigated in a similar way to the case
of general insurance in the foregoing chapters. Also dynamic rules
can be incorporated into the model to simulate possible measures
taken by management, e.g. in the case of adverse development.
8.4 GENERAL SYSTEM 305

y(O,tJ

t
Figure 8.4.4 Demonstration of the effect of remedial action when an adverse
flow of the solvency margin is imminent. The solid line shows the flow if no
special measures are taken; the dotted line shows the rectified flow.

Figure 8.4.4 shows an example. Note that the control mechanism

generates cyclic variation as seen also in the analysis of the general
insurance (see Fig. 7.7.5). However, owing to the difficulties in
rendering rapid improvements in relevant factors affecting the
solvency, it can be expected that the wavelengths of the cycles are
considerably longer than in the case of non-life business.
In more detailed analyses, of course, features can arise such as the
necessity to reinforce the reserves of current policies if the primary
calculation bases are weak compared with current experience, and
these can bring about enormous stresses for the continuation of the
business. It can be expected that the stochastic dynamic model
outlined can provide a useful service in the analysis of such kinds of
situation, supplementing the conventional deterministic approaches
which are traditionally in use in life insurance practice and widely
employed in general management information systems as well as in
conventional prognosis and budgeting models.
A further dimension for the use of the model provides the possi-
bility of analysing the behaviour of the components of the portfolio,
e.g. the profitability of the different cohorts. This is shown in Fig.
8.4.5 where the flows of both the whole portfolio and some of
its cohorts are displayed. Even if strongly simplified the example
may be realistic in so far as the profitability owing to inflation and
other circumstances may differ considerably inside the portfolio.

(c) Link with general risk theory If the planning horizon is short
or if the portfolio is already more or less in a state of equilibrium,
306 RISK THEORY ANALYSIS OF LIFE INSURANCE

y(O,t) t
1.5

1.0 ."
~~======~~26
0.5
'-----------~--~~--~~-------~------~~~~
0.0
* * * * * -~<:o:::: l~ -11
--<1<6
-------------36
-0.5

-1.0

-1.5

- 2.0
5 10

Figure 8.4.5 Examples of the profit (or loss) flow of a supposed portfolio
and eight of its cohorts having the duration from the entry 1,6, 11, '" , 36 years
respectively. Interest rate 7%, rg = 1.04 and other data are as in Fig. 8.3.1.
The curve with asterisks represents the joint result of all the cohorts.

the behaviour of the life insurance business can be studied, without

essential modifications, using the same technique as was developed
in the foregoing chapters mainly for general insurance. Now, if
we follow closely the decomposition idea dealt with in Section
3.7, the expected number of claims is readily obtained from (8.3.1),
specifying the cohorts by their entry years tc and denoting the
current year by t.

L I(t, tJq(t - tc )' (8.4.2)

tc= -w+t

The argument t - tc should be adjusted to correspond to the age

of the cohort members in year t, i.e. the current age x = Xc + (t - t c )
where Xc is the age at entry.
Furthermore, the dJ. of the claim sizes, i.e. the risk sums
Z = S(t)(l - V(t - tJ) can be obtained as weighted averages
from the cohort distributions, or more simply they can be found
by statistical methods directly from the records of paid claims.
8.4 GENERAL SYSTEM 307

The basic difference between the 'general approach' and the

'life insurance approach' outlined in this chapter is that the latter
makes it possible to forecast the time-dependent changes in the
basic data for the long run, in particular possible distortions in
the claim size distributions and expense loadings. On the other
hand, very similar models are attainable as for the general system
by using the decomposition of the portfolio into parts, as was
presented in Section 3.7, and by programming suitable growth and
other transition rules for each of them.
 CHAPTER 9

Ruin probability during an

infinite time period

9.1 Introduction

(a) Passage T -+ 00 In the previous chapters a survey was given

of methods of determining numerical values of the ruin probability
for firtite time periods (0, T].1t is of interest to examine the behaviour
of the risk process when T tends to infinity. For example in applica-
tions where the solvency of the insurer is a premise, conservative
estimates can be obtained if the planning horizon is not limited
to any finite time. This approach forms the basis of the traditional
risk theory, resulting in the well-known infinite time formula for ruin
probability, which will be presented in Section 9.2.

(b) Examples Some conception of different outcomes of the infinite

time problem can be found in the figures of Chapters 6 and 7. In
particular, Fig. 7.1.2 is illustrative. In the case exemplified in part
(a) of the figure the stochastic bundle drifts into infinity. When the
process passes the critical point where the lower boundary of the
stochastic bundle (confidence region) has its minimum, the sample
paths with increasing probability are more and more remote from
the ruin barrier and the probability of ruin at large t values rapidly
vanishes. This is seen also from Fig. 7.5.1 where the two lowest
curves represent just this type of process. If the lower boundary
of the stochastic bundle drifts upwards fast enough the infinite
time ruin probability is < 1, as will be shown in item (c). The same
is the case also when the cycles are assumed as in part (c) of Fig.
7.1.2.
The drift to infinity of the solvency ratio u(t) for t -+ 00 is not a
necessary condition for 'P 00 < 1. There are cases where the bundle
9.1 INTRODUCTION 309

,
,, 0.3
I
~ &-&-e-e-e-e.-&-e
,
I
,Ill'
,
I
I

,,
! , -------

0.4

5 10 15 20 Time T 25

Figure 9.1.1 Examples of the asymptotic behaviour of the ruin probability

'¥ T(U(O)) as a function of T calculated for some fixed values of the initial
solvency ratio u(O). Solid lines, no cycles assumed; dotted lines with circles,
deterministic cycles assumed.

gets narrower and narrower when T grows, so that (see (6.6.12))

the standard deviation of the solvency ratio u(t) tends to zero,
making 'P oo(u) still less than one (see exercise 9.1.3).
If the process has a finite equilibrium (see (6.6.9)) and
limau (l, T»O for T-HfJ (see (6.6.12)) as in parts (b) and (d) of
Fig. 7.l.2 then the ruin is usually certain, i.e. 'Pro = 1, as further
analysis can prove (see exercise 9.1.1). This can be anticipated also
from the topmost curve of Fig. 7.5.1.
Figure 9.1.1 exhibits some typical flows of 'PT'

(c) A condition for non-zero survival probability It must be appre-

ciated that these examples do not exhaust all the types of pro-
cesses; nor are the referred conditions always sufficient to keep
'P oo(u) < l. The crucial fact is the speed with which the cross-cut
ruin probability cp(t) (see (6.6.19)) decreases when t ~ 00. This
regulates the increment
(9.1.1 )
310 RUIN PROBABILITY DURING INFINITE TIME PERIOD

and makes the sum

(9.1.2)

equal to one or less than one. It follows directly from (6.9.3) that
the condition
ao
L cp(t) < 1, (9.1.3)
t=to

is sufficient for non-zero survival probability at least for to = 1.

This can be extended to any to by some auxiliary conditions (see
exercise 9.1.2).
Another method of evaluating L\ 'Pt can be obtained by means of
the truncated convolution algorithm (6.7.9) or (6.7.11). It can, for
instance, be seen (exercise 9.1.1) that if any finite limit, however
large, is assumed for the 'wealth' of the insurer, i.e. a finite upper limit
for U(t), then 'P ao is always equal to 1 by quite general conditions.
Such an upper limit is called the reflection barrier. It can be supposed
that any possible excess ofU(t) is divided as bonuses, dividends, etc.
In fact a non-zero survival probability can then be obtained only by
letting U exceed all limits which, indeed, is a rather unrealistic
condition from the point of view of applications. This is a well-known
paradox of classical risk theory. The nice problem of seeking condi-
tions for non-zero survival probability cannot be dealt with in this
book other than by an example in exercise 9.1.3 (see also Ruohonen
(1980) ).

**Exercise 9.1.1 Assume that the underwriting process satisfies

the equation U(t) = rjU(t - 1) + p .. - X(t), U(O) == Uo,for t = 1,2, ... ,
where the variables X(t} are mutually independent. Further, assume
that the variables U(t} have an upper limit of some bound U max
(e.g. possible excesses being given away as bonuses or dividends),
and that the ruin barrier is Ur(t} == O.
(i) Show that if M = rpmax + p .. and if
prob {X(t} > M} ~ e > 0 for every t, (9.1.4)
then 'P ao = 1. p .. and rj are constant (they could be allowed to vary,
however, having some finite upper limits).
9.2 THE INFINITE TIME RUIN PROBABILITY 311

(ii) Show that 'P 00 = 1 still applies when (9.1.4) is replaced by the
weaker condition
00
L prob {X(t) > M} = 00. (9.1.4')
t= 1

Note that (9.1.4) holds, for example, in the case where the variables
X(t) have a d.f. of the compound Poisson type.

** Exercise 9.1.2 Prove that (9.1.3) is a sufficient condition for 'P 00 < 1
provided that the condition ~ 'P t ~ qJ cc(t) holds for t > to (see (6.9.7))
and that 'P to < 1.

** Exercise 9.1.3 Assume that in the u-process dealt with in item

6.6(b) the parameters r i , r g' r p and A are constants and r igp < 1, r g > 1,
(1 q = 0 and uit) = o. Prove that if the condition ~ 'PI ~ qJ cc(t) is

satisfied for t> to' then 'P 00 < 1. (Make use of the Tchebyshev
inequality.)

9.2 The infinite time ruin probability

(a) Special assumptions The famous formula for the infinite time
ruin probability 'P 00' often briefly denoted 'P and called 'ruin pro-
bability', is discussed in this chapter. Consideration is limited to
the very special case, where the risk process is stationary and no
fluctuations in basic probabilities, either short term or long term,
are present. Also the safety loading A, and thus implicitly the rating,
is never changed; furthermore, interest and inflation are omitted
and the continuous checking of the state is adopted (see item 1.4(b)).
The assumption concerning stationarity can be replaced by a more
general assumption allowing a preprogrammed deterministic change
of the parameter n, the expected number of claims. Anyway, the
independence of the stochastically defined increments is assumed.
Hence it will be assumed that the distribution function F of the
total amount of claims during one year remains unchanged, 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 dis-
tribution of the size of a claim is independent of time.
312 RUIN PROBABILITY DURING INFINITE TIME PERIOD

The detailed treatment of the problem is deferred to Appendix C.

(b) Ruin probability Summarizing the main results of Appendix C

it can be stated that the probability that the company will be ruined
is:

(9.2.1 )

where U( = U(O) = U 0) is the initial solvency margin (risk reserve)

and qU) is an auxiliary function having values less than 1. The
coefficient R is a positive constant, termed the 'adjustment coeffi-
cient' or 'insolvency coefficient' or Lundberg's coefficient. In the
Poisson case R is the root of

1 + (1 + Je)mR = too eRZ dS(Z), (9.2.2)

and in the Polya case the root of

1 + -(1
h - e-(l H)mnR1h) = foo e RZ dS(Z). (9.2.3)
n 0

The auxiliary function qU) depends on the special characteristics

of the process concerned, among other things on the claim size
dJ. S. If the claim sizes have an upper limit M, i.e. S(M) = 1 then

e- RM < qU) < 1,

as will be proved in Appendix C, equation (C.l6). Hence we have

(9.2.4)

M may be the retention limit applied in reinsurance. Normally it

is small compared with the initial solvency margin U, which makes
the limits of the above equalities close to each other as will be shown
in exercise 9.2.2. Also more generally the function C is often close
to 1 and consequently 'I'(U) ~ e- RU usually gives a satisfactory
approximation for the infinite time ruin probability.
Equations (9.2.2) and (9.2.3) can be solved by means of well-known
numerical or graphical methods. Using the exponential expansion
9.2 THE INFINITE TIME R urN PROBABILITY 313

of eRZ and integrating term by term, (9.2.2) is transformed into the

form

from which R can be found by a few trials. By using the first or first
two terms and the risk indexes r 2 and r 3 (see (3.3.8)) the following
approximations are obtained:
2A 1
R=-- (9.2.5a)
r2 m

1
R= m
J'[6A~ + 9r~J 3r
4r~ - 2r m'
2 (9.2.5b)
3

(c) Applications Many of the problems considered in the previous

chapters can also be treated by making use of (9.2.1). In fact it fully
corresponds to the basic equation (4.1.7a) and its counterpart
(7.2.1) for the finite time planning horizon. Some examples will be
given in the following items and in the exercises of this section.

(d) qJ as function of U and 1 First the ruin probabilities for the

claim size distribution given in Table 3.6.1 with M = £ million
are presented on a semi-logarithmic scale (Fig. 9.2.1). It is seen how

10 20 30 40 50
U
Figure 9.2.1 The (approximate) ruin probability 'P ~ e- RU as a function of
U and A. The default data are the standards defined in item 7.1(d) (semi-logarith-
mic scale).
314 RUIN PROBABILITY DURING INFINITE TIME PERIOD

-RU
e

10 20 30 40 50
U
Figure 9.2.2 The (approximate) ruin probability as a function of U and M;
default data as standards of item 7.1(d) (semi-logarithmic scale).

closely the value of'llIe is linked to the safety loading Aand the initial
reserve U. If A = 0, it can be proved that'll = 1, i.e. ruin is certain.

(e) 'I' as a function of U and M is computed in Fig. 9.2.2. Fixing the

value of'll the maximum net retention M as a function of U can be
read from the figure (dotted line). This figure is a counterpart of
Fig. 7.3.1. The crucial difference in the numerical values is due to
the fact that cycles of the basic probabilities were assumed in
Section 7.3 but are omitted in the present Chapter 9.

(1) Indifference to the size of the portfolio A strange feature of the

e- RU approximation is that it depends only on the distribution
function of one claim and the safety loading A. This statement means,
paradoxically, that the size of the company plays no part in calculat-
ing, for example, the initial reserves U when the Poisson model is
used. This may be difficult to appreciate, especially as for example
in Fig. 4.2.2 or in Fig. 7.2.1 it was specifically stated that the reserves
should depend on the size of the company. However, the contradic-
tion is only apparent and is due to the assumption that the state
of the process is checked continually at every time t. The process
U(t) is constituted, independently of the size of the company, as the
difference of the incoming premium flow and the outgoing claims
flow. The difference between a small company and a large one is that
in the latter case more premiums and accordingly more claims
9.3 DISCUSSION OF THE DIFFERENT METHODS 315

are coming and going in a time unit. This can be expressed in

another way: it takes perhaps ten years for a small company to
experience all that could happen in one year in a large company.
But because the state is tested continuously and the time-span is
infinite this does not affect the final ruin probability.
If, however, testing for 'P 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 qU)
(see (9.2.1)) 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.

(g) Short-cutformulae can be derived for the rapid use of (9.2.1)

by applying techniques similar to those employed in Section 4.2.

Exercise 9.2.1 Calculate e- RU for U = £10 7 and data given in item

7.1 (d).

Exercise 9.2.2 Evaluate the relation of the limits (9.2.4) assuming

that M ~ 0.02 U and that e - RU is either 0.01 or 0.001.

9.3 Discussion of the different methods

(a) One-year approach In this section some further comments

will be made concerning the main lines of approach to risk theory
problems, as viewed at the beginning of the book in Section 1.4
and implemented in the preceding chapters.
The great advantage of considerations limited to one year only
(or any other finite 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 n, A, 8, M, O"q and U. The 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
316 RUIN PROBABILITY DURING INFINITE TIME PERIOD

contrary, it is necessary to plan all the security measures affecting

the company's existence for an unlimited number of years in the
future. Attention was paid to this aspect in the applications of
Chapter 4, 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 reserves the company's management is prepared to lose
in one year taking the worst possible view, and to take for this pur-
pose, 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.
(b) Finite time approach For elimination of the arbitrariness in
regard to the choice of the variable U, long-term considerations
are developed. This is especially so because the variations in basic
probabilities, like trends and cycles as well as inflation, are of crucial
importance and necessarily demand a long planning horizon. For
this purpose the formula apparatus was extended to an arbitrary
time period (0, T] in Chapters 6 and 7. The rather complicated
structure of the model restricted consideration mainly to numerical
techniques, which are scarely manageable without at least modest
computers. The interconnection of the numerous model variables
could not be given in any simple analytic form. Fortunately the
graphic presentation of the results, which can conveniently be
produced even by small computers, effectively provides a survey over
the structure of the model. A great merit of the finite time technique
is that it can be kept workable without restrictive preconditions
and special assumption, for example concerning the variations-
both short and long - of the basic probabilities, inflation or the
claim size dJ.
(c) Infinite time approach The extreme alternative for the future
safety of the company is to take an infinite period. For this reason
the e - RU technique has been given considerable attention in actuarial
literature. It can be used for evaluation of conservative limits for
finite time probabilities providing, however, assumptions are
made about the specified stationarity of the process. Unfortunately,
elaboration of the calculations and the necessity to keep the ruin
probability from becoming equal to one seems to lead to serious
restrictions in the choice of model assumptions.
9.3 DISCUSSION OF THE DIFFERENT METHODS 317

This method also has some other obscure points. In particular,

the results, as is seen in Fig. 9.2.1, are all very sensitive to the size
of A- and in other respects the method also involves assumptions
concerning all future time. Even if it were possible to estimate A-
for a few future years, it is very difficult to say what it will be sub-
sequently. In fact, experience shows that different kinds of trends,
amendments of premium tariff and also other circumstances are
always changing A-, often cyclically. On the other hand, it is some-
times observed that only the very beginning of the process affects
the value of e- RU (see Figs 7.1.2,7.5.1 and 9.1.1). When the dangerous
part of the process is passed then the longer future has no significant
influence; this would suggest the infinite time ruin probability
as an acceptable upper limit for finite time approaches. This seems,
however, to be true only as a consequence of rather restricted condi-
tions and by letting the reserves grow to infinity. In practice this is
clearly an unrealistic situation, indeed. This is 'the paradox' men-
tioned in item 9.1(c): ifit 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 'I' = 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. In fact, it only means that the
process runs more rapidly as stated in item 9.2(0. 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 point in the e - RU method was the assumption of con-
tinuous checking which can give excessively large values for large
companies. Another arbitrariness exists in the selection of a suitable
insolvency probability t: = 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 of
whether to allow an insolvency in one year with probability B = 10- 4
or over ten years with probability B = 10- 3 or over an infinite
period with B = 10- 2 •
318 RUIN PROBABILITY DURING INFINITE TIME PERIOD

(d) 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
Chapters 6 and 7 it is feasible, at least if computers 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 vari-
ables and the claim size dJ. S. 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 program rules
of procedure which are to apply if the profit is rapidly accumulating
excessively, or vice versa if it is diminishing, as will be referred to in
Chapter 10. 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 solvency
margins, etc. It also seems necessary to include conditions con-
cerning 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.
 CHAPTER 10

Application of risk theory to

business planning

10.1 General features of the models

(a) The purpose of the model 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, many 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 by means of a mathematical model.

(b) Corporate planning The construction of the model depends on

the objectives of the user. If it is aimed at corporate planning it
may be very rich in details simulating all the actions of an insurance
office which may be relevant for the trading result. It is used e.g.
to forecast the future flow of business and to help in evaluation of
different policy alternatives. A corporate planning model can
analyse the business structures and answer questions of the type
'what. .. if... '. To be successful it should be closely incorporated
with the conventional accounting and statistical functions extracting
the necessary data basis from them. Pioneering work in the building
320 APPLICATION OF RISK THEORY

of insurance models is credited to McGuinness (1954). Descriptions

of models have been recently published for instance by Heubeck
(1971), Galitz (1982, so-called ASIR model) Gelder and Schauwers
(1981), Geusau (1981) Reischel (1981) and Roy (1981).

(c) Risk theory models Model building is useful also for the study
of risk-theoretical behaviour of portfolios, e.g. for analysis of
the solvency conditions. It can be developed as a natural extension
to the conventional risk theory, with the purpose of giving a com-
prehensive view over the structures as a whole and of trying to find
tie-ins between the numerous aspects involved. Then it is sufficient
to select from the very numerous variables and details which are
needed in comprehensive corporate models only those which are
most relevant for solvency properties. This line will be followed in
the sequel. A model of this type was experimented with in a Finnish
solvency study (Pentikainen, 1982 and Rantala, 1982).
In order to facilitate the reading only the main lines will be drawn.
For example, the decomposition of the portfolio in sections is
omitted; even though it would not be difficult in principle to accom-
plish using the technique described in Section 6.4, its implementation
can be laborious. Furthermore, problems concerning rate-making,
technical reserves and many other aspects which are relevant for
corporate planning are not dealt with, because it has been necessary
to limit the scope of this book to aspects which are suitably handled
by the technique ofrisk theory. For the same reason 'non-stochastic
risks' like failures in investments, political risks, consequences of
misfeasance or malfeasance by management are not discussed,
notwithstanding the fact that they may be quite important and can
seriously jeopardize the solvency of insurers (see Pentikainen, 1982,
Section 2.9).
The presentation will be limited to approaches which are best
fitted to non-life insurance. However, life insurance can also be
handled in a similar way, in particular if the planning horizon is
fairly short. For the long-range consideration of life insurance,
modifications and attention to the special features along the lines
discussed in Chapter 8 are necessary. It is a matter of course also
that generally, in all kinds of business, the capability of models of
producing reliable analyses rapidly deteriorates when the time-
span grows because most of the basic distributions and data distort
in a way which is barely predictable, and because the environment
10.1 GENERAL FEATURES OF THE MODELS 321

where the future insurers operate is likely to be very different from

the present one. However, prognoses extended to cover rather long
future periods may also be of interest. Despite the fact that they
may be of little value in giving a reliable view of the true total situa-
tion the models can illustrate the structures and properties of in-
sur;nce undertakings and help to understand underlying tie-ins.
Of course, the models can be improved by further developing and
refining the basic assumptions - which were made in the preceding
chapters and will now be adopted as building blocks for the more
comprehensive approaches we are going to deal with in this chapter -
but discussion about them is beyond the scope of the presentation,
which is necessarily to be limited to monitor mainly general features
of the risk theory and its applications.
(d) General structure of the models Revenues and expenditures will
be incorporated in the model as displayed in Fig. 10.1.1. The
difference between the incoming and outgoing money flows is
accumulated into the risk reserve - or as it is nowadays called-
the solvency margin U.
The schedule is essentially the same as is conventionally used in
the accounting of profit and loss statements and balance sheets.
This basic setting also constitutes the framework of various prognos-
es and business analysis schemes according to the current practice
of insurers. The conventional prognoses are mostly deterministic.
The major contribution of risk theory is that the forecast can be
made stochastic. The apparatus developed in previous chapters
can now be incorporated into the schedule. Then a model is achieved
which gives the future state as a distribution of the relevant variables,
i.e. a range in which for example the solvency margin can be expected
to lie after some time interval, whereas a deterministic model gives
only one single value. This approach makes it possible to deal with
problems where solvency is also involved, an aspect which is mostly
not adequately possible in deterministic systems.
Another advantage of the risk-theoretical approach is that plan-
ning schemes can be made dynamic, i.e. the system can be made
self-controlling or, as is often said, adaptive. For example, if the
solvency during the planning horizon T is weakening, special rules
can be programmed which provide restoring measures such as
increases in premium rates, economies in cost, etc. This idea was
exemplified in Section 7.7 and can be benefited from in model build-
ing.
322 APPLICATION OF RISK THEORY

Inflow Risk reserve Ou~flow

I1U Claims
Earned paid
risk ou~s~anding
co premiums p
III
E
:J Expenses
'E
Cl) Safety loading
L.. U Reinsurance
Cl.
Loading for (ne~ balance)
expenses

Ne~ inves~men~
Dividends
earnings

VARIABLES AND ASSUMPTIONS

State variables & parameters Market conditions

Portfolio mix, classes i = 1, 2, ... -real growth of the
Claim size d.f. portfolio ig
Volume indicators n = n 1 + n 2 + ... -sales response etc.
(expected number of Claims) Rate of interest ,;
Reserves
Assets Business strategies
Rates of premiums
Exogenous factors - safety loadings
Short-period fluctuation of Reinsurance
the basic probabilities - net retentions M J
Long-period business cycles Sales efforts, administration
Inflation Dividends and bonuses
-steady rate i x Safety level (ruin probability
-shock impulses and time-span T)

Figure 1O.l.l The basic schedule of the model.

10.2 An example of risk theory models

(a) Synthesis of submodels The idea displayed in the schedule

of Fig. 10.1.1 can be readily written, using the notation shown in
the figure, in the form
B + I = X + Co + C re + C m + D + ~u. (10.2.1)
The variables involved need detailed definition in order to make the
model operative. In fact, each of them represents the outcome of
10.2 AN EXAMPLE OF RISK THEORY MODELS 323

corresponding submodels, as will be seen in the following, and

(10.2.1) represents a synthesis of these submodels.
All the variables may be stochastic (printed in bold letters) even
though it may be necessary often to simplify the model by assuming
that many of the elements involved are deterministic. For example,
it is convenient to concentrate the stochasticity of the premium
income into the safety loading A defining (see (6.2.11))
B = P/(1 - c - Ab),
where the risk premium should be found deterministically by using
the technique developed in Chapter 6. Methods to implement the
the stochasticity of Ab (or A) were discussed in item 6.2(e) and in
Section 7.7.

(b) Three ways of handling ceded reinsurance One way to construct

the basic schedule is to calculate the premium income B,
claims X and other variables gross of reinsurance. It is illustrative
to unite all transactions concerning the ceded reinsurance as one
single entry, the 'reinsurance cost'
(10.2.2)

where B re is the amount of the ceded reinsurance premiums account-

ed as earned amount and net of commissions and bonuses and X re is
the reinsurers' share of claims including the outstanding claims.
Another approach can be used when the trading result on the
insurer's own account, i.e. net of reinsurance, is concerned. Then
the premiums and claims are decomposed in the shares which are
due to the insurer's net retention and due to the reinsurers:

B = Bo + Bre = Bo + P re (1 + Are)
X = Xo + Xre ,
where the subscripts '0' and 're' refer to the insurer's own account
and to the reinsurers. Pre is the ceded part of the risk premiums and
Are the loading including net compensation of the reinsurers'
expenses, safety margin and profit margin. Substituting in (10.2.1),
the reinsurance terms offset each other and the equation is trans-
formed in the following net of reinsurance form
(10.2.3)
324 APPLICATION OF RISK THEORY

In fact this formulation is used conventionally in risk theory, and

that was the case also in the foregoing chapters.
Formula (10.2.3) has, however, the inconvenience that the re-
insurance cost is hidden in the calculation rules and in the rates of
Bo and Xo' A compromise would be to decompose Xre into its
expectation Pre and stochastic variation ~X.( = Xre - Pre) and
then to substitute into (10.2.2) and (10.2.1) X re = Pre + ~Xre'
X = Xo + Pre + ~Xre and Bre = (1 + Are}Pre

(10.2.4)
where
(10.2.5)
(10.2.6)

This approach is based on the fact that ~Xre does not affect the
trading result ~U because it is included, with the + sign in X
and the - sign in C re • Hence, when the outcome on the insurer's net
retention only is concerned it is unnecessary to simulate the stochas-
tic value of Xre because the net result is the same as if it is replaced
by its expected value.
The benefit of the version (10.2.4) is that the reinsurance cost
is expressly shown, which is particularly useful when different
reinsurance policies are investigated and compared (see the example
given in item 4.5(d)). The type of the reinsurance treaty as well as
its parameters like the retention limits M and the target level of the
loading A. re are model parameters.
A more detailed analysis should necessarily concern also the effect
ofthe yield ofinterest earned for the shares of reserves or for the agreed
depositions.
In what follows the choice of the approach is left open: the user
can understand the basic variables either as gross amounts or net
amounts and in the latter case drop the term C re from the equations,
if it is taken into account in the definition of the variable values
calculated on the insurer's net retention. It is also dependent on
the application concerned whether Are and consequently Cre are to be
assumed stochastic or deterministic.

(c) Premiums have already been defined in Sections 4.1 and 6.2,
and the technique to handle stochastic control was discussed in
connection with (10.2.4). The growth of the premium income is
10.2 AN EXAMPLE OF RISK THEORY MODELS 325

controlled by the trend factor rg(t) (6.1.2), 'premium inflation' factor

rp(t) (6.2.1) and safety loading l(t) (6.5.12).
Because the earned premiums, which are provided in the equa-
tions, are calculated by deducting from the written premiums the
increment of the reserve of unearned premiums, the aspect related
to the reserve calculations is also to be considered in advanced
comprehensive models. However, because the relation of this reserve
to the premium volume is fairly constant and not seriously sensitive
to inflation and other disturbances, it was suggested in item 6.S(h)
that the interest earned for the reserve could be amalgamated into
the safety loading, after which this reserve may not be needed any
more as a special entry. The case is essentially another in the long-
term analysis of life insurance, where it may be necessary to follow
the flow of the policy reserves for instance by means of the technique
outlined in Chapter 8.

(d) Investments have already been discussed in Section 6.3. In

advanced models different investment strategies can be assumed and
their outcomes analysed and optimized.
A major difficulty is to find realistic assumptions concerning the
future returns earned for the investment, i.e. for the entry I(t) in
(10.2.1). The interest yield or the concept of rate of return, which
includes also movements in the capital value of the investments,
depends on varying conditions of the general investment markets
of the country and on the investment strategies which the manage-
ment has chosen. Some suggestions have been made that I(t) could
also be taken to be stochastic in order to take into account the
often quite great variations in the rate of interest and of security
values. A major difficulty is, however, to find any meaningful
distributions. The rate of interest notoriously depends very much on
the political decisions which are in the hands of governments or the
national central banks. There is also some correlation between the
variation in different countries. The 'genuine' random fluctuation
in the interest rates seems to be of a smaller order of magnitude than
the long-term political and economic changes. The examples given
in Figs 6.3.1 and 6.3.2 illustrate the situation.
As was seen in previous chapters, a more relevant feature of the
interest rate is its relation to the simultaneous inflation rates and
also to the real growth of the portfolio, all three effects being com-
bined in the important growth factor r igp or, in a more general case,
in r(tt, t 2 ) (see (6.5.8) and (6.6.1)). In advanced models it may be
326 APPLICATION OF RISK THEORY

appropriate to treat the rate of interest ij(t) and the inflation rates
ix (t) and ip (t) as separate variables, trying to find appropriate
assumptions for their combinations. Experience has convincingly
shown that all these quantities are subject to long-term major
variations and, in addition, to short-term oscillations having a
minor amplitude. The abundant econometric literature may suggest
ideas for the construction of proper submodels for these features. At
present there are no well-developed theoretical disciplines available
for insurance applications. As has already been suggested in item
6.3(b), the sensitivity of the model outcomes to the investment
entries may be tested by making varying optimistic and pessimistic
deterministic (or semi-deterministic) assumptions about the anti-
cipated future level of the return rates and their relation to inflation
and growth.
In risk theory models the investment earnings can be handled in
different ways. They can be kept as a separate variable or they can be
divided between the underwriting reserves and the solvency margin
as was suggested in Section 6.5 (see (6.5.2) and (6.5.12)).

(e) Underwriting result and claims As a next step, terms relating

to the underwriting profit will be united
Y(t) = (l + l(t))P(t) - X(t), (lO.2.7a)
where Y means the result of the underwriting business on the
insurer's net retention. The variables P and X have already been
defined above. The safety loading l can conveniently include also the
income gained on the technical reserves following the technique put
forward in item 6.5(h). The definition of this loading depends also
on the handling of reinsurance as discussed in item (b).
In fact, the results derived in Chapter 6 are now directly
applicable.
The claims X in the above formula, as generally in the foregoing,
include both paid and outstanding items. As mentioned already in
item 3.1 (c) this involves an idealization: it is supposed that any claim
is paid immediately when it occurs. It was stated that factually the
inaccuracies in the evaluation of outstanding claims, including also
effects of inflation, do not escape from the model. Their effect is
to move possible profits or losses only from year to year and should
be allowed for in the basic data of the model when these are derived
from actual records. However, for advanced models, which are
10.2 AN EXAMPLE OF RISK THEORY MODELS 327

constructed for corporate planning, it is highly advisable to consider

the outstanding claims as a set of separate variables, in particular
in the case of long-tail business. A shock inflation especially can
give rise to adverse effects which should not be ignored. An extra
reinforcement of the reserve may be necessary, as has already been
discussed in item 6.1(1).
The adequacy of the outstanding claims reserves is commonly
checked a posteriori by means of 'run-off triangles' or by otherwise
comparing the actual claims amounts with the reserve estimates
made earlier. If the deviations appear noticeable compared with the
total business volume and with the 'normal' fluctuation of the under-
writing business a simple way to amend the model is to take the
estimation error of the outstanding claims as a special entry O(t)
into the model replacing (10.2.7a) by
Y(t) = (1 + l)P(t) - X(t) + O(t). (l0.2.7b)
The variable O(t) should be defined and calibrated in a way which
fits with observed experience and with the reasonably anticipated
future development. It can be expressed in terms of a random com-
ponent (so-called pure 'noise') and cyclical and trend components
if necessary. Special attention to inflation can be focused in particular
in case of long-tail business. The standard methods of mathematical
statistics are available as well as the recent rapid development of
the theoretical and practical consideration of the problems about
the outstanding claims. The discussion of further details falls,
however, outside the scope of this book. Reference to the research
work of the Insurance Technical Laboratory of the University
of Copenhagen (a comprehensive report was expected around
1983) as well as the proceedings of the XVI ASTIN Colloquium in
Liege 1982 can be made.

(I) Expenses Usually the insurers also use special accounting for
their expenses. Its outcomes can be linked with the model as a
special entry through the terms Co and Cm. Co represents all the
sales and operating expenses the insurer may have. Cm includes
different kind of miscellaneous entries in the loss and profit account.
These items may be either stochastically or deterministically
defined. Probably it will be convenient to keep them deterministic
and move the uncertainties to the safety loading l (see item 6.2(d)).
This is supposed to have been done in the present considerations,
328 APPLICATION OF RISK THEORY

and these variables are denoted accordingly (no longer using bold
letters).
It is convenient to divide the expenses Co into two parts. On the
one hand every insurer has some normal costs which cannot be
controlled much by the insurer, at least not in any radical way in the
short term. Some of these costs are related to the volume of sales of
new policies, some to the number of policies in force and some to the
number of claims. On the other hand, some other costs, e.g. those
related to the sales, are very much under the control of the manage-
ment, and it is advisable to introduce a special variable Q for them.
Hence the total costs are composed of a normal amount C 1 and a
moving part Q as follows.
Co = C1 + Q.
As will be seen in Section 10.3, Q can be one of the key issues in a
dynamic program, because e.g. sales campaigns and other similar
operation policies can be introduced into the model by using it.
Furthermore, alternative actions concerning personal, office
facilities, etc. can be considered. For example, investments in new
buildings and inventories will obviously first cause a rise in expenses
but later on will reduce them.
The normal costs C 1 should be programmed to depend on the
forecasted business volume. Often a sufficient approximation for a
risk theory model may be to assume them proportional to the
premium income P(t), as was presented in item 6.2(f), i.e. C1(t) =
c'P(t), where c' was adopted as a coefficient related to unloaded risk
premiums P instead of the coefficient c which was used in (6.2.8).
In fact, c' = c/(1 - c - A.b ) as seen from (6.2.11). It is natural in this
connection to base the supposed actual costs on a quantity which is
not stochastically varying. This was the rationale for the change of
the definition.
Tax is one of the relevant items of the model. Its consideration
and programming depends on local legislation and practice. We
assume it to be included in expenses in one way or other. Some
actuaries have suggested it as a separate entry in the basic equation
(10.2.3) and provided rules on its dependence on the book profit
(and dividends).

(g) Dividends is another entry in the basic schedule. It is supposed

that this concept consists ofthe 'ordinary dividends' divided between
10.2 AN EXAMPLE OF RISK THEORY MODELS 329

the stockholders of proprietary companies, as well as bonuses given

possibly to the policy-holders. It depends essentially on the business
objectives. The topic will be discussed in Section 10.4. If the under-
writing business is considered on the basis of the interest of an
outside investor, i.e. the stockholder of the insurance company or
a private underwriter, then a natural idea is to search for strategies
that maximize the dividends. Another approach, natural for bonuses
of mutual companies and also for dividends of many proprietary
companies operating in modern environments, is to consolidate
them to some suitable level preferably in one way or another related
to the volume of the business
D(t + 1) = d x B(t), (10.2.8)
where again B is the premium income and d is a constant which is
one of the free parameters of the model.

(h) Modified basic equation It is often useful to rearrange the

terms of the original schedule (10.2.1) in a way in which the net
results of the different operational sections are entries instead of the
incoming and outgoing money flows
Y x(t) + Yi(t) + Yc(t) - Cre(t) = D(t + 1) + Q(t + 1) + L1U(t), (10.2.9)
where Yx(t) is the underwriting profit (or loss) (1O.2.7b); Yi(t) is the
net income gained on investments; Yc(t) = cB - C 1 - Cm is the
net surplus of the expense budget, which is equal to the loading
minus the actual operating expenses, the term Q being excluded
(item (f)); Cre(t) is the (consolidated) reinsurance cost (10.2.6);
D(t + 1) are the dividends, e.g. according to (10.2.8); Q(t + 1) are the
specified extra sales or other expenses as defined in item (0; and
L1U is the increment ofthe risk reserve, equal to the amount allocated
to enforce the financial status of the insurer, equal to the balancing
term of the equation (10.2.9).
Only four of the variables in (10.2.9) were assumed stochastic
(bold letter), despite the fact that the other variables can also be
modelled stochastic. For example, they can indirectly via the
technique of dynamic programming depend on the outcomes of
'primarily' stochastic variables like the underwriting profit Yx '
Equation (10.2.9) is a direct extension of the basic equation
(6.5.1) which constituted the basis for finite time considerations.
The l.h.s. gives the resources available and the r.h.s. their allocation.
330 APPLICATION OF RISK THEORY

The same simulation technique as applied in the foregoing chapters

can be used also for this generalized model as well as many of the
analytic evaluations and short-cuts derived above. However, the
new features introduced require special attention. They are dealt
with in the subsequent sections.

10.3 Stochastic dynamic programming

(a) Theory and practice of dynamic programming It is useful to note

that the model drafted above is along the same lines as the
approach which was originally called industrial dynamics, due to
Forrester (1972). It is mostly concerned with the flow of work,
decision-making and other kind of operations inside factories and
organizations. An attempt to apply the idea to insurance environ-
ments was reported by Veit (1976). The industrial dynamics method
has been widely used for operational research and planning and its
name has recently changed to 'system dynamics'. A comprehensive
survey of this discipline is edited by Roberts (1978).
Our model also has features similar to so-called dynamic pro-
gramming, which is a tool much used in the operational research
applied to many industries. A well-developed mathematical
discipline exists for these topics; pioneering work was published by
Bellman (1957) (see also Nemhauser (1966) for example).
The idea of dynamic programming is to provide the calculation
algorithm with rules which control the flow of subsequent steps
in a way which depends on the outcome of the preceding steps.
An example of the dynamic control rule has already been given
in Section 7.7, where the premium rates were changed according to
the level of the solvency ratio via the safety loading. The actual
change in safety loading can be obtained either by adjusting the
premium rates or, at least in emergency cases, by economizing
in administration and possibly in other costs as well.
Another way to introduce dynamics into the model is to control
the level of net retention as a function of the solvency ratio. If the
solvency is good, then the company can take more risks in net
retention and in that way gain by a reduction in reinsurance costs.
If the solvency is getting weak, then as one ofthe protective measures
perhaps the net retentions can also be lowered. This approach was
investigated by Pentikainen (1975) and (1982).
10.3 STOCHASTIC DYNAMIC PROGRAMMING 331

Further dynamics can be introduced by linking some relevant

entries with outside impulses like the growth of the national eco-
nomy. An example was given in item 6.1 (t).

(b) Terminology Using the practice of the dynamic programming

theory the variables and rules involved in the model can be
characterized as follows.

State variables define the state of the system (insurer's fiscal status)
at times t = 0,1,2, .... Volume parameters n, assets, liabilities,
solvency margin V, etc. are examples.
Transition equations like (10.2.9) determine the increments of the
state variables from each time t to the next t + 1.
System parameters are needed to define both the state and transitions.
Some of them, such as inflation, depend wholly or at least mainly
on the outside circumstances and environment. Some others,
so-called decision parameters or rather decision variables, can be
controlled at least to some degree by the management of each
insurer, e.g. premium rates, net retentions and allocation of
resources like dividends D and sales efforts Q. Some parameters
are of mixed character, being partially dependent on exogenous
factors and to some degree on the insurer's own policy decisions.
Rate of interest is an example. A set of the numerically assigned
decision variables is called a strategy. Examples of the variables
and factors involved with the model are listed in Fig. 10.1.1.

Aspects of the strategies and business objectives will be discussed

in Section 10.4, but before that the idea of dynamic programming is
illustrated in the remaining items of this section.

(c) Return functions The strategies mentioned can concern various

policy options. In order to get the model operative special rules,
so-called return functions, are needed to make possible the evalua-
tion of the consequences of the policy options. As an example
the situation is considered when the management decides to attempt
to expand the business volume, the market share, launching a sales
campaign and allocating money Q (see item lO.2(f)) for the purpose.
The problem is to find in one way or another an estimate of the
premium increment I1B which can be expected to be received as a
return of the marketing cost - or in other words 'investment' - Q.
332 APPLICATION OF RISK THEORY

There are available a number of textbooks concerning these so-

called sales response junctions, e.g. Kotler (1975).
A sales response function can be, of course, constructed and
calibrated only on the basis of experience. Depending on environ-
ments and special circumstances Kottler suggests two basic types
of functions. The first is linear
I1B = qQ, (10.3.1 )
where q is a so-called elasticity factor. If, as frequently happens,
the return is spread over more than one year, this can be suitably
programmed.
Another alternative function is exponential. Assume that the
extra marketing costs are again Q and in addition that a reduction
of p 100% is made in premium rates as another competitive measure.
Then the normal formula of the annual premium increment (see
(6.5.7)) is changed as follows

B(t + 1) = B(t)rl p ( 1 + ~oJ (1 - p)-fJ. (10.3.2)

The exponential coefficients are to be found empirically. By a

logarithmic differentiation approximate partial effects can be
derived

(l0.3.3a)
and

I1Bp = {Jp (l0.3.3b)

B .
Owing to the form of these equations (J. and {J are called 'elasticities'.
A more accurate treatment of the problem is to provide for the
situation where the effect is spread over several years. Furthermore,
the sales campaign is likely to be concentrated only on some special
type of policies and the return can be expected to concern mainly the
same type of insurance. However, experience shows that if a new
customer is acquired, e.g. owing to lowered rates of motor third-
party insurance, his other insurance cover is often moved to the
same company irrespective of whether or not any special benefit
concerning other classes is offered. This is an extra complication in
finding sales responses.
10.3 STOCHASTIC DYNAMIC PROGRAMMING 333

Furthermore, the return of a sales campaign can quite crucially

depend on the reactions of the competitors in the same market.
For example, the effect of premium reduction will be completely
missed if the competitors follow suit. Aspects like this will be
discussed in Section 10.5.
An example of sales responses and market reactions is displayed
in Fig. 10.5.2.

(d) Algorithm When all the assumptions are settled and values
for the model parameters determined, then the model can be
operated by means of the Monte Carlo method, just as was presented
for the simple underwriting processes in Chapters 6 and 7. The
difference is that the number of model variables may be greater and,
in addition to the claims, other stochastic variables, e.g. the yield
of interest and premiums, via the response functions can also now be
stochastic. For different state variables, and above all for the
solvency margin U, 'a random walk' can again be obtained as is
illustrated in Fig. 10.3.1. In principle it fully corresponds to the
simulation figures of the previous chapters, e.g. Figs 6.8.3 and
6.8.4; however, now only the boundaries of the confidence region
are plotted.

(e) Characterization of business strategies A major difficulty in

the use of models like that just drafted is that the number ofvariables,
distributions and aspects involved can be very great. It is quite
difficult for the user to obtain any insight into the structure of the

I
I
I
I
... I
I
......... _ . I
- - _____________________ -4

Ruin barrier

Time t
Figure 10.3.1 An outcome of a stochastic model.
334 APPLICATION OF RISK THEORY

model which, in fact, is a more or less complicated configuration

in a multidimensional parameter space. It is still more difficult to
explain successfully the properties and outcomes of the model to any
outsider, e.g. to the members of the management team who are not
familiar with the technique of dynamic programming and the risk
theory. A quite useful approach is graphic data processing. Some of
the most relevant variables should be displayed in a way which
reveals the essential features of the properties of the process con-
cerned. By changing different strategy parameters new outcomes are
obtained, which can usefully be plotted in the same figure to get an
idea of the behaviour of the structure. In fact, the graphic display of
the result has been amply used already in the foregoing, and it will
be still further exemplified in the following.
The variable U, the solvency margin, can be selected as one of
the main variables. It represents the 'wealth' of an insurer. Another
important feature to be followed and analysed is the volume. This
can be suitably described by the gross premium income on the
insurer's net retention B, and it will be accepted as the second main
variable. It is useful to give a survey of the total situation by plotting
the variables Band U in the same plane. This is done in Fig. 10.3.2,
where now instead of time (see Fig. 10.3.1) the premium volume B is
displayed on the x-axis.

,
".

v (0) --------~~~ ",

,.",,.,
Ruin barrier
"

8(0) 8
Figure 10.3.2 An outcome of a business strategy in the B, U plane. Stochastic
bundle and its confidence region.
 10.3 STOCHASTIC DYNAMIC PROGRAMMING 335

Each simulation step generates for the variable pair B, U a new

outcome which represents the state of business at time t, and the
next step at time t + 1, etc. A sample path in the B, U plane is a
broken line, a 'random walk' as seen in Fig. 10.3.2. The repetition
of the simulation generates a bundle of sample paths and the final
points represent a sample of the situation at the end of the planning
horizon T. For Fig. 10.3.2 it was assumed that not only the solvency
margin U but also premium income B is stochastic. Then at a given
confidence level the final points are two-dimensionally distributed
over a region which is approximately elliptical in shape.
Graphic configurations like that in Fig. 10.3.2 give an idea of
the assumed flow of business and properties of chosen strategy
options. The more the distribution ellipse is situated to the right,
the more expansive, i.e. volume increasing, is the strategy option.
The vertical position of the ellipse indicates the solvency situation.
If the ellipse or the bundle intersects the ruin barrier, it is an indi-
cation of a risky situation.

(I) Comparison of strategies The purpose of the model is above

all to give a tool for the evaluation of business strategies and for
their comparison. This is illustrated in Fig. 10.3.3, where the out-
comes of two strategy options I and II are plotted.
For clarity only the mid-lines (solid lines) and the lower confidence

U (0) .- .- .-
, .-
_ _ _ _ _ _ _ _ ----

....
,
............ Ruin barrier

.... _------- --
8(0) 8
Figure 10.3.3 A conservative strategy, I, and a risk-taking one, II.
336 APPLICATION OF RISK THEOR Y

boundary (dotted lines) of the stochastic bundles are plotted. The

endpoints PI and PH represent the expected average outcomes ofthe
strategies. The lower boundaries make it possible to evaluate the
risks involved. The strategy option I is intended to exemplify a
conservative policy, not accepting any solvency risks. The strategy
option II is an expansive one. At the beginning of the planning
period money is invested in marketing which is reflected in a faster
growth of the volume B but in a lower growth, even a temporary
reduction, of the solvency margin U until the invested costs are
regained. However, the stochastic bundle B intersects with the ruin
barrier. This can be interpreted as a solvency risk which is involved
with this expansive strategy option.

(g) Ruin probabilities can be found as a by-product of the Monte

Carlo simulation, as presented in Section 6.8.

(h) Straightforward technique It must be appreciated that, instead

of the Monte Carlo simulation, the direct calculation of the final
state and the evaluation of ruin probability may often be feasible
and advantageous following the ideas presented in Section 6.9.
The computation time can thereby be radically shortened. On the
other hand if dynamic rules are provided, their handling in direct
calculations may be out of the question. Then probably no technique
other than simulation can be tractable. Fortunately a fairly small
number of sample paths, when plotted appropriately, may be
sufficient to give quite a satisfactory visual picture of the process
concerned, as was stated in item 6.8.l(d).

10.4 Business objectives

(a) Strategies The above model gives as output the distribution

of the final state of the business at the end of the planning period,
and as a by-product it can also give an appreciation of the ruin
probability during this period. When the strategy assumptions are
changed, a new outcome is obtained. A very interesting problem is
to try to find 'the best strategy' among the infinitely many options.
Is it for instance better to be 'great but poor' or 'smaller but weal-
thier'? To answer these kinds of questions it is necessary to define the
business goals, or the objectives which the management will use as a
guideline for its decision-making.
10.4 BUSINESS OBJECTIVES 337

(b) Strategic triangle In the earlier literature a business goal

often used was the maximization of profit. As recent literature on
business behaviour now indicates, this goal is nowadays too narrow.
Substantial private entrepreneurs have disappeared and most large
enterprises are led by professional managers. Galbraith (1973)
describes how the interest of professional managers may be not so
much the maximization of the dividends to an often large and diffuse
group of stockholders, but rather the safeguarding of the stability
and continuity of the enterprise and often - a common status
symbol- the achievement of a strong expansion in sales volume.
Hence it seems necessary to assume multiple goals. For example,
dividends, growth in sales, and growth in net worth can be alternative
business goals which must be weighted in some way. These three
items correspond to the three r.h.s. terms in (10.2.9). They con-
stitute a 'strategic triangle' as depicted in Fig. 10.4.1. The three goals
are mutually contradictory and the problem is to find in one way or
another a balance between them.

(c) Selection between the goals One reasonable way is first to

provide the solvency constraint ('I'T::::; e) to be fulfilled, then to
fix some acceptable level for dividends or bonuses (so that the
stockholders and/or policyholders are sufficiently satisfied, as

Solvency Expansion
U B

Bes~
stra~egy ?

Dividends
D

Figure 10.4.1 The strategic triangle.

338 APPLICATION OF RISK THEORY

B(O),U(O)
e

B
Figure 10.4.2 Search for an optimal strategy. The labelled points represent
outcomes of different strategies.

Galbraith notes). After that the available resources are to be divided

between expansion in sales and net worth, i.e. between the terms Q
and AU in (10.2.9).
There are at least two approaches here: (i) the 'looking-at-the-
chart' method and (ii) the utility function method. The latter will be
discussed in items (d)-(h). The former operates simply with a figure
like Fig. 10.4.2, where the outcomes of different strategies are points
in the plane. They correspond to the average end points PI and P u
in Fig. 10.3.3, the strategy options being specified by the numbers
1,2, .... For clarity the confidence boundaries of the stochastic
bundles (dotted in Fig. 10.3.3) are not plotted. Instead information
about the solvency aspect is given by marking with an asterisk
the cases where the solvency condition 'P T ~ e is satisfied and with a
cross the cases where the solvency risk is imminent, suggesting
rejection of the strategy concerned.
It is the insurer management's task to define a suitable business
goal from the chart. In Fig. 10.4.2 a straight line is taken as the goal.
It balances Band U and optimization means finding a point on the
target line as far to the right as possible.

(d) The concept of utility The choice of the strategy options in the
approach described was left very much to the subjective weighting
between the contradictory alternative outcomes. Attempts to find
more objective criteria can be made by introducing the so-called
10.4 BUSINESS OBJECTIVES 339

utility functions. This concept derives from the well-developed

theories of games and economic behaviour due to von Neumann and
Morgenstern (1947).
Karl Borch has pioneered the application of the idea of utility
to the insurance environment (1961, 1962, 1963 and 1970).
Utility functions are widely used in the literature to describe
the economic behaviour of individual persons as well as of insti-
tutions. They constitute a useful way of clarifying situations in
which the decision-making is to be done. To illustrate the concept
a simplified example is given which describes a potential policy-
holder who is deliberating whether or not to take insurance cover
for a particular risk object, and if yes, what level of deductible
claim, if available, ought to be chosen (this presentation follows an
idea given by Jewell in his lecture in the Oberwolfach seminar,
1980).
Let a person have an initial capital Uo and let his property be
subject to the incidence of a loss causing event with probability p.
For the sake of simplicity, assume that one risk sum X only is
possible if a claim occurs. Then he has to choose between two
alternatives (Table 10.4.1):

(I) Take an insurance to cover the risk and pay a premium B.

(II) Save the premium B and take the risk of loss of the amount X.

The utility function G(U) is now constructed to describe the

decision-making situation and to give the problem a mathematical
formulation. It attempts to weigh how desirable the different out-
comes are for the decision-maker. Clearly it can be expected that
it is the better for him the larger is his wealth U. It is, however,

Table 10.4.1 Decision options of a potential policyholder

The wealth if
I II
Insurance Insurance
taken not taken

In case of
(i) no incidence of loss
(ii) incidence of loss
340 APPLICATION OF RISK THEORY

reasonable to assume that the 'desirability' is not directly pro-

portional to U. If for some reason or another the decision-maker
loses a substantial part of his wealth, rather great inconveniences can
entail. As a consequence may be e.g. the loss of home or loss of
ability to afford school education for the children. Hence a special
weight should be given to the outcomes where wealth becomes very
low (not to say negative). The desirability of such chances is best
described by negative values. On the other hand, growth of wealth is
generally accepted as a positive event, but the more wealth is
received, probably the less weight can be assumed for the interest
to get still more wealth. These personal aspects, conveniences and
inconveniences will be described by constructing a function G(U)
to give numerical evaluations for the desirability or, as it is called,
'utility' of different outcomes. It can be assumed to be an increasing
function of U and the aspects discussed furthermore suggest that
it should be a concave function as illustrated in Fig. 10.4.3.
Expressed by means of formulae the above conditions are
G'(U) > 0; G"(U) < 0, (10.4.1)
assuming the existence of the derivatives.
The simplified example concerning the decision alternatives of
the potential policyholder related to a hypothetical utility function
is depicted in Fig. 10.4.3. It seems natural to characterize the two
decision alternatives, i.e. to insure or not, by considering on one hand

stU)

Wealth U

Figure 10.4.3 Schematic presentation of the options of a policyholder.

10.4 BUSINESS OBJECTIVES 341

the value of the utility function G and on the other hand the pro-
bability that just the outcome in question may occur. The weighted
utility of the option I is then
GI = (1- p)G(U o - B) + pG(Uo - B) = G(U o - B),
and that of the option II, i.e. not to insure, is
Gn = (1 - p)G(U 0) + pG(U 0 - X).
Insurance should be taken if GI > GIl' otherwise not. This depends
on the selection of the utility function G and also on the initial
state U0 and on the values of p, X and B.
Note that in the particular case in which G is linear, equal to
aU + b (against the axioms (l0.4.1)),
GI - Gn = -aB + apX.

Now pX = E(X) is the risk premium. Hence, if the rating is correct,

i.e. the gross premium B is greater than the risk premium P, this
expression is always negative. This means that strictly statistically
the policyholder is losing. Insurance is, however, meaningful if the
possible loss X is large compared with U 0 and its consequences
harmful, maybe extremely, for the policyholder if not compensated
by insurance. Therefore a linear utility is not suitable to describe
the actual situation and a concave function is rational. Then
GI - Gn may be positive, suggesting that insurance cover should
be taken. However, if X is very small compared with the initial
wealth U 0' then the conclusion may be the opposite, i.e. it is not
useful to cover quite modest risks by insurance. Perhaps a suitable
deduction would then be appropriate.
Another interesting situation arises if the probability p of a
large loss is very small, in which case GI - Gn can also be negative.
This is the classical dilemma of whether or not to take insurance
cover for risks with a remote but positive chance of occurrence.
(The more so because the lack of experience in these cases may result
in a prohibitive safety margin in premium B.) The outcome depends,
in any case, on the subjective choice of the utility function, i.e.
whether or not and to what degree the person in question has a
risk-taking or risk-averting attitude.
The above example was simplified so that only two outcomes
were assumed, namely the loss of fixed amount occurs or does not.
A more general situation is where several, perhaps continuous,
342 APPLICATION OF RISK THEORY

outcomes are present. It may concern the decision situation of

insurance management, e.g. the choice between policy strategies
like those illustrated in Fig. 10.3.3. Then the probability numbers
p and I - p are to be replaced by a distribution function F(U) and
the utilities can be expressed in the form of an integral

G= E( G(U) ) = f_co G( U)dF( U),

+co
(10.4.2)

which is called the utility of a variable U having a distribution F.

The simplified Fig. 10.4.3 is now replaced by Fig. 10.4.4.
The one-variable utility function is often suggested also for
analysis of the insurer's decision-making when strategy options are
considered on the lines illustrated in Fig. 10.4.4. One of the diffi-
culties, when the construction of a concrete utility function is
attempted, is the diffuse problem of the numerical values that ought to
be assigned to the possibility of ruin, i.e. for the values of U which drop
below the ruin barrier. However, this dilemma may be circumvented

--- ----
u

Figure 10.4.4 The idea of utility applied in a decision situation of insurance

management. Two strategy options I (conservative) and II (risk-taking) and
their distributions. The utilities GJ and Gn are obtained weighting U values by
the respective densities according to (10.4.2) and are depicted by vertical bars.
10.4 BUSINESS OBJECTIVES 343

by separating the ruin problem from the utility definition.

For this purpose every strategy which does not satisfy the condition
(10.4.3)
is to be rejected. Then the definition of the utility function is needed
only for the values of U which exceed the ruin barrier (option II in
Fig. 10.4.4 was chosen so that (10.4.3) obviously is not fulfilled, and
consequently it is not acceptable regardless of whether or not
GIl> G,).

(e) Multivariate utilities A serious restriction is the limitation of

the business objectives to one goal only (mostly solvency margin
or equivalent), which is an obvious oversimplification of the real
world situation. This can be helped (see Kotler (1975)) by introducing
multivariable utility functions where at least the three goals of
Fig. 10.4.1 are present. However, a subjective evaluation can never be
avoided. Firstly, the often diffuse business objectives have to be
expressed in the form (or guise) of some mathematical function.
Secondly, values for parameters involved with the formula have
still to be found and this may not be done without a more or less
intuitive selection; on the other hand, subjectivity is always involved
with a decision judgement irrespective of which kind of formal
methods are applied. A benefit of the utility function approach is
that it makes the management procedure more consequent than
decision-making case by case. Similar situations can be handled
in a similar way and different situations can be put in a logical badge
of rank.
Even if the utility function approach is, no doubt, mathematically
appealing and elegant, and can be used for many considerations,
the 'looking-at-the-chart' method may be more practical and more
easily understood for persons who are involved in decision-making
but do not have sufficient training in the mathematics of risks and
utility theory. Mathematically, the difference between the approach-
es need not be significant; it may even be semantic. In any case,
the model should allow a choice and the user should then decide
which of them (or some other method) is accepted.

(I) Optimal strategy Irrespective of which of the approaches is

chosen, the search for an optimal strategy leads to intricate problems
of how to find extreme values of a multivariate function, i.e. a
344 APPLICATION OF RISK THEORY

® No ruins
-
U
\I
,,-Ruins
10

100

185
B
Figure 10.4.5 Example of the search for favourable strategies presented by
Pentikiiinen (1976). A. is the safety loading and f is a parameter that allocates
the sales efforts between the two classes of insurance presented in the example.
The midpoints only for each outcome are plotted. An acceptable or not accept-
able level of the risk of ruin is indicated by circles or asterisks respectively.

maximum for the utility G or an optimal position of the final state

in the B, U plane. A simple procedure is to experiment with different
parameter sets and by trial and error move in the multidimensional
parameter space towards a region where better and better outcomes
can be found. Fig. 10.4.5 illustrates the method. One or two of the
variables in turn are changed whilst the others are held fixed. The
target functions seem to be approximately linear in that fairly
narrow region which is allowed by resources or other limitations.
This observation essentially facilitates the search operation.
More sophisticated techniques can be found in standard text-
books on dynamic programming, e.g. Nemhauser (I 966).

Exercise 10.4.1 Prove that if a utility function G is linearly trans-

formed to another
H=aG+b (a> 0),
then any relation G(F 1) > G(F 2) defined by (10.4.2) implies H(F 1) >
H(F 2 )·
10.5 COMPETITION MODELS 345

10.5 Competition models

(a) Multi-unit models The foregoing consideration was mainly

limited to an isolated insurer. A brief reference to the dynamics
connected with competitive measures was given in item lO.3(c) in the
form of sales response functions. In reality, however, the market
situation can affect the business flow in a rather more complicated
way than was assumed for the isolated case. It is possible, in fact
likely, that an acceleration in sales efforts very soon leads to a
reaction from competitors. It is necessary therefore to extend the
model by bringing in the relevant insurance market, incorporating
consideration of the behaviour of competing insurers. In principle,
this can be done by constructing a model, similar to that drafted
above, separately for each of the companies of the market and then
adjusting their interactions. This is illustrated in Fig. lO.S.1, where,
for brevity, only three companies I, II and III are presented.
It is first supposed that no special competitive activity is being
undertaken by any of the companies. Then the flow of the business
of the three companies may be as plotted by thick curves 1. Each
of the companies has a normal growth both in volume and solvency
margin. Next it is supposed that company I launches a sales cam-
paign making use of the premium reductions, advertising, etc. If
the other two companies do not react, the outcome can be as plotted

B
Figure 10.5.1 Competitive market of three insurers I, II and III. The curves
represent the flow of the average outcome EB(t), EU(t) as functions of time
for respective insurers (see Figs 10.3.2 and 10.3.3).
346 APPLICATION OF RISK THEORY

by curves 2. Company I benefits by a considerable growth in volume.

The solvency margin may first be decreased due to the campaign
expenses, but the increased portfolio later leads to a recovery.
On the other hand, assuming that the market concerned is relatively
small and almost saturated and that each of the insurers has a fairly
sizable share of it, the action of insurer I is reflected in a reduction
of the volume growth of the competitors.
Alternative 3, plotted by the dashed lines, describes the possibility
that after some time lag companies II and III react with a similar
campaign, reducing rates and increasing sales activity. It can be
expected that the sales efforts of all three companies neutralize
each other and equilibrium is again reached after some fluctuations.
As a consequence no extra growth in volume is gained by any of the
companies. Instead, however, the campaign expenses remain thus
giving rise to a negative return. Clearly continuation of the situation
soon leads one or more of the companies to a critical situation and
necessitates stabilizing changes in business strategy. Just this kind
of market mechanism was supposed to be one of the major causes
for the cyclic variation introduced to the model in Section 6.1. The
example could be continued, e.g. by speculating on the consequences
if one or more of the companies first amends its rates and stops the
other competitive efforts.

(b) Simulation of competition To make the multi-unit model work-

able information is needed for each of the market unit and also
formulae for the behaviour of the market shares. It is appropriate to
adjust the response functions as exemplified by (10.3.2) so that first
some kind of average level of premium rates is determined and then
the possible deviation of each insurer from this level. These devia-
tions are just those relevant in competition. In this connection the
premium reduction variable p in (10.3.2) is to be understood as
such a deviation from the average market level.
Although models using general marketing theories are well
developed and frequently applied to other industries, a near complete
lacuna concerning insurance applications surprisingly seems to
exist, as e.g. Jewell in his model survey (I 980) stated. Even if difficult
it may, however, be possible to gain experience, perhaps by trial
and error methods, to get some idea of this very complicated reaction
mechanism. Possibly one reason for the lack of reports is the
tendency not to publish this kind of data. Annual company reports
10.5 COMPETITION MODELS 347

40

_ _ _ _- - 2

20 -1

1970 1971 1972 1973 1974 1975 1976

Figure 10.5.2 Third-party motor insurance. Trend in market shares oj the

five companies as percentages oj the whole market (Pentikiiinen, 1979).

and general insurance statistics may possibly be of use as one of the

sources of informatio~. For illustrative purposes reactions of
Finnish motor third-party liability markets are depicted in Fig.
10.5.2.
In years 1973 and 1974 company 1 applied rate reductions and'
gained in its market share until the rates were equalized.
The market reactions depend among many other things also
on the degree of saturation of the market. Extreme cases are manda-
tory insurance classes such as third-party liability insurance. Be-
cause no market effort can increase the joint volume ofthe insurance,
any growth of market share is possible only at the cost of other
insurers. In the case of most voluntary classes market efforts can be
expected to discover objects which otherwise may have been left
outside insurance cover. In such a situation an accelerated sales
activity partially increases the national volume of insurance and is
likely in addition to move market shares from competitors. Only
experience can give guidan~e for the construction of suitable
response formulae.

(c) The analysis of multi-unit markets belongs to the large realm

of multiplayer games and theories of economic behaviour. For
instance theories of oligopolitic markets can be referred to (Fried-
mann, 1977) where problems related to collusions, equilibriums,
348 APPLICATION OF RISK THEORY

etc. are examined. If the strategy changes are made in short time
intervals the scenario is much akin to systems which are treated by
the theory of differential games, a survey of which theory is edited by
Grote (1975).
Both for staff training and for business planning so-called business
games have been developed. Special teams play the role of the
management of different enterprises and their decisions are simul-
taneously put into a computer programmed to give as output the
market reactions. These are then fed back to the playing teams for
the next step. Obviously the models drafted in this chapter can well
be used for such purposes. If stochasticity is programmed into the
model, it can give a new dimension to conventional game models,
which may mostly be deterministic.
 APPENDIX A

Derivation of the Poisson and

mixed Poisson processes

A.I Poisson process

Let k(t) be the number of claims in the half-closed interval (0, (]
(t > 0). Define k(O) = 0, and let s > 0 and 0 ~ tl < t2 ~ t3 < (4' The
following conditions are assumed to be valid:
(i) The variables k(t 2 ) - k(t l ) and k(t 4 ) - k(t 3) are independent.
(ii) The variables k(s + t) - k(t) and k(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)' prob {k(s + t) - k(t) = O} = prob {k(s) = O}.
Let nk(t) = prob{k(t) = k}. According to (i) and (ii)
no(s + t) = prob {k(s + t) = O} = prob {k(s) = 0, k(s + t) - k(s) = O}
= prob {k(s) = O} prob {k(t) = O} = no(s)no(t)
or
(A.I)
for every s, t > O.
Hence no(t) is everywhere a non-increasing function. Let sand t
be positive rational numbers. Then according to (A.l)
no(st) = [no(s)J = [no(t) ]S,
giving the result [no(t)] l/t = const. ~ 1. This non-negative constant
cannot be 0, since, if no(t) = 0 for every rational t, then according to
350 APPENDIX A

(ii) any interval should contain at least one claim with the probability
1; thus the interval (0, t] would contain an infinite number of claims
with the probability 1, contrary to (iii). Consequently the constant
can be denoted e-P(p ~ 0) giving
(p ~ 0), (A.2)
This is true for any positive rational number t. Further, since
is everywhere non-increasing, 1to(t) cannot have steps and is
1t o(t)
accordingly continuous. Thus (A.2) is also true for irrational ts.
In order to calculate 1tk(t) for k > 0 let h be an integer such that
n = 2h > k and write down the disjoint partition
i- 1 i ]
1.= ( --t,-t (i = 1, 2, ... , n),
, n n
of the interval (0, t].
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
k and their placement in (0, t] vary) two subsets are now selected
for a fixed h as follows:
Ah is the set of all realizations (sample functions) such that at least
one claim occurs in exactly k intervals Ii; that is to say, realiza-
tions for which exactly n - k intervals remain without claims
(nowk ~ k)
Bh is the set of all realizations such that at least in one interval
Ii at least two claims occur (hence k ~ 2)
Evidently
{k(t) = k} c Ah U B h ,
since, in order that exactly k claims occur, it is necessar)1that either
k different intervals must include claims or some interval must have
at least two claims. On the other hand,
Ah - Bh c {k(t) = k},
since the left-hand side includes only realizations where claims
occur in exactly k intervals and nowhere. more than 1. Thus
prob(Ah - B h ) ~ 1t k (t) ~ prob(A h u B h )
or, a fortiori
A.l POISSON PROCESS 351

To prove that prob(B h ) ---t 0 let h ---t 00 and let w be an arbitrary

realization. There are two possibilities:
(1) w is such that from a certain value of h every interval contains
at most one claim.
(2) w is such that independently of h there is always at least one
interval where at least two claims occur.

Clearly lim prob(Bh ) can be > 0 only if realizations of 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 the claim points have to have an accumulation point.
These two possibilities both have probability 0 according to condi-
tion (iii).
Returning to (A.3) it remains to calculate 1tk (t) = lim prob(A h ).
The probability that in a fixed Ii at least one claim occurs is, accord-
ing to (A.2),

Thus the probability that claims occur in exactly k fixed intervals is

These k values can be chosen in (:) different ways, so that

Hence,

(p ~ 0). (A.4)
352 APPENDIX A

According to (A.2) this is also true for k = O. The process k(t) is

called a Poisson process.

A.2 Extensions
The theory of Poisson processes can be extended by replacing one
or more of the assumptions (i) to (iii) by weaker ones. The new
claim number processes obtained in this way playa central role in
the advanced theory of risk.
As has been seen, the conditions (i), (ii) - or (ii)' -lead to (A.2),
which gives 1t o(t) as an exponential function of t. Suppose now that
condition (ii) is substituted by a weaker condition assuming only
that:

(ii)* 1t o(t) is a continuous function of t tending to 0 for t -> 00.

Obviously 1t o(t) - which certainly cannot be anywhere increasing-

can also be denoted by e-pr(t), where ,(t) = -(l/p)log1t o(t) is
another continuous function, tending to + 00 for t -> 00, and being
nowhere decreasing. The constant p can be chosen so that T(l) = 1.
For every positive, there exists at least one t so that, = ,(t). The
original process k(t) can be considered as a random process of a
new variable " denoted k'(,), defining simply k'(,) = k(t), where
t is any such number, that, = ,(t).
(Note: lf1to(t) is absolutely decreasing in point t, the correspon-
dence between , and t is a one-to-one correspondence. Now let
(to' t 1 ) be an interval, where 1t o(t) is constant. Then for any t within
this interval 1t o(to) = 1t o(t) = prob{k(t) - k(to) = O} prob {k(to) = O} =
prob {k(t) - k(to) = O}1t o(to). Thus prob {k(t) - k(to) = O} = 1, and
accordingly, with probability one, k(t) = k(to)' so that k'(r) is also
uniquely defined in this case.)
Now let '1' 'z > 0, and let t 1 , t z ' and t be such that 'I = ,(t 1 ),
'I
'z = ,(t z )' and + 'z = ,(t). Then according to condition (i)

prob{k'('1 + 'z) = O}
= prob{k'('1 + 'z) -k'('I) = O} prob{k'('I) = O},
and hence, because prob{k'(,) = O} = e- pr
prob {k'(, 1 + 'z) - k'(, 1) = O} = e -p r2 = prob {k'(,z) = O}.
A.2 EXTENSIONS 353

This proves that the process k'(r) satisfies the condition (ii)'.
Accordingly

prob {k(t) = k} = e - pt(/) [pr(t) ]k (A.5)

k! .

It can be said that the process k(t) is a Poisson process in the trans-
formed new time scale r, in so-called operational time.
Conditions (i) to (iii) lead to a process where only the constant
p remains to be estimated in applications. The weakened condition
(ii)* instead of (ii) leads to a process where the function r(t) remains
as a 'parameter' to be estimated or assumed in applications. The
product pr(t) gives, in this case, the expected number of claims in
the interval (0, t]. The derivative r'(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 processes with non-stationary
increments, also called heterogeneous in time, whereas the Poisson
process (A.4) is a process with stationary increments, also called
homogeneous in time.
A further extension of risk processes is obtained if the constant
p is thought of as a random variable p, which varies due to some
outer factors, e.g. random effects of weather conditions. Suppose
that the claim number k(t) satisfies the conditions (i), (ii)*, and (iii)
on condition that p has a given value p. Then the conditional dis-
tribution ofk(t) is again a Poisson distribution.
A more general case is obtained, which is also more realistic,
if p is dependent on time, hence being a general stochastic process
p(t). In order to give a short survey, let p(t) be a realization, i.e. a
sample function of this process, and suppose that, for the fixed p(t),
conditions (i) (ii)*, and (iii) are satisfied. Then again, for any value
of t, prob{k(t) = k} (on condition that this sample function p(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
pr, where p 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 r
is calculated separately for different realizations p(t), it is dependent
354 APPENDIX A

on the value p as well; thus the notation t = t(t, p) is used in the

following. Hence for the unconditioned process:

prob{k(t) = k} = too e-pr(I'P)[Pt(~r)]k dpV(p, t),

where the 'structure function' V(p, t) = prob {p(t) ~ p} is in general
dependent on time.
Finally the integration variable is changed by introducing q =
pt(t, p)/t and solving p from this equation. Then the solution is
placed in V(p, t) and a new structure function H(q, t) = V(p(q, t), t)*
is constructed. The probability can be rewritten:

prob{k(t) = k} = 1 o
00 (qt)k
e- ql - , dqH(q, t).
k.
(A.6)

A process which satisfies condition (A.6) is called a mixed Poisson

process in the wide sense or, especially if the structure function H is
independent of time t, a mixed Poisson process (in the narrow sense).
It might happen that a mixed Poisson process in the wide sense
in time t is a mixed Poisson process in the narrow sense in a suitably
chosen operational time r. (An example of this is the process satisfy-
ing condition (A.5), which can be modified into the form (A.6) and
which is thus a mixed Poisson process, leading to a structure function
dependent on time. In fact if, for clarity, the integration variable is
denoted by x, H(x, t) = 6(X - pr(t)/(t), where 6(X) denotes the
degenerated distribution function.)
For further information see for example Philipson (1968) and
Thyrion (1967).
A process with property (A.6) no longer obeys conditions (i)
and (ii) in general. It should be remarked that (A.6) does not yet
define all properties of the process k(t), since, for example, it displays
only the distribution of random variables p(t) for various values of
t,)ut 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, H(q, t) = Jd p V(p, t) where the integration is taken over all such
values of p for which qt > pr (t, pl.
 APPENDIX B

Edgeworth expansion

If all derivatives of the function G exist and if G(k) ( ± (0) = 0 for

all k ~ 1, then:

CPk(S) = f +OO eisxdG(k)(x) =

-00 [eiSxG(k)(x)]~: - is
f+oo
-00 eisxdG(k-l)(x)

= ( - is)CPk-l (s) = ... = ( - is)kcpo (s).

Since the normal distribution

N(X;m,u)=J(2n)u
1 fX [1
-2 (z - m)2 ]
-00 exp -u- dz

=N( X ~ m ; 0, 1) = N( X ~ m ).
clearly satisfies these conditions, the characteristic function of its
kth derivative is
(B.l)
where
(B.2)

is the characteristic function of N(x; m, ( 2 ). Evidently,

N(k)(X; m, u) = u- kN(k{ X ~ m). (B.3)

The characteristic function of the simple compound Poisson

function is
cp(s) = en"'(s)-n,
where '" is the characteristic function of the claim size dJ. It has
356 APPENDIX B

the following expansion, according to (B.2)

cp(S) = enJP(s)-n = exp{nI.I/I(k)(O)l, + nO(s5) - n}

o k.
= exp(isnm - ts Zna z )

Hence by inverting the characteristic function, according to (B.1):

+ a2~nN(4)(X; nm, J(na z )) + a7~nZ N(6)(X; nm, J(na z ))

Z

+ ... + remainder,
where the terms omitted are of the form Cn i NU) (X; n m, j(n az ))
withj/2 - i ~ 3/2, C being independent of n. Hence, if x = (X - mn)/
j(na z ), according to (B.3),
F(X) = N(x) - ia 3n(j(na z ))- 3 N(3)(x)
+ z14a4n(j(naz))-4N(4)(x)
+ /Z a~ nZ( j(na z )) - 6 N(6)(x) + ... + remainder,
where the terms omitted are ofthe form
C'ni(J(na z ))- j N(j)(x) = C" ni - jlZ N(j)(x) = C" n -k N(j)(x),

with k ~ 3/2. Hence

F (X) = N(x) - ia 3a; 3/2 n -liZ N(3)(x) + Z14 a4 a;Z n- 1N(4)(X)
+ /z a~ a; 3 n - 1 N(6)(x) + O(n - 3/Z).
 APPENDIX C

Infinite time ruin probability

C.I Assumptions
We are now going to extend to a special case the consideration
of the risk processes related to a finite time interval as considered
in Section 6.5, letting the length of the time interval grow to infinity.
The assumptions and notation are modified as follows.
Let A denote a time unit. It is assumed that the state of the risk
process is tested at times A, 2A, ... , tA, ....
The aggregate claims amount during the· period (( t - 1)A, tA]
is denoted by X(t) and the underwriting profit (or loss if negative)
will be defined for each period by the equation

Y(t) = (1 + A)P - X(t), (C.1)

where P is the risk premium E(X(t)) and A> 0 is a safety loading.

It is assumed that the variables Y(t) of the consecutive periods
are mutually independent and have the same dJ. G. The only
assumption concerning G is that the integrals needed later are
convergent. Without loss of generality it can also be assumed that
Y(t) assumes negative values with positive probability; otherwise the
probability of ruin would be always zero.
It is assumed that the risk premium income P and the safety
loading A are constant. However, it would be possible to extend
the consideration for cases where the business volume is growing
according to some deterministic rule as presented in Section A.2,
where the concept 'operational time' was introduced, but for
simplicity this is not provided in the following.
Next, the accumulated profit is defined by

W(t) = Y(1) + ... + Y(t). (C.2)

358 APPENDIX C

Wltl

TIme In A Unl~S
o ----..........
'..... .... ........ , ,."
..... --,, I
,----- /'
, I
, I
, I

-uo _.. _-- _. _. _. " I

- - - - - - - - - - ~ _ _ .J._
"II

Figure C.l. Two realizations of the accumulated profit W(t) with L'1 as length
of test period. For the other realization a ruin (asterisk) is observed.

Then the rum probability related to the time period (0, T!1] IS

'PT = 'PT(U O )
= 1 - prob {W(t) ~ - U0 for t = 1,2, ... , T} (C.3)
where U0 is the initial risk reserve (see Fig. C.Ll).
Note that in terms the notation of Chapter 6 W(t) = U(t) - U 0
(if!1 = 1).
The irifinite time ruin probability 'P = 'P 00' which is obtained
when T tends to infinity, depends, as well as on the risk process,
on the choice of the time unit !1: the smaller !1 is the larger 'P be-
comes. When !1 ~ 0 the testing is carried out continually at every
time point.
In general the assumptions given above concerning the variables
Y(t) do not remain valid if !1 is replaced by some other time unit
!1', unless !1' = k!1, where k is some positive integer. However, if
the risk process under consideration is a simple compound Poisson
process, it is obvious that the assumptions remain valid for any
positive !1.

C.2 Estimation of 'P

Estimation is based on a technique which makes use of the joint
C.2. ESTIMATION OF 'I' 359

moment generating function

M(s) == MY(t)(s) = E{exp(sY(t))}, (C.4)
of the variables Y(t).
Suppose first that T < 00 is given. Then the m.g.f. MT of W(T)
satisfies
MT(s) = E{exp[s(Y(I) + ... + Y(T))]} = M(sf. (e.S)
The idea is to derive yet another expansion for this function, which
then directly gives the result sought. It is obtained by introducing the
concept 'the time of (first) ruin' (L\ as time unit), denoted by t 1 •
Then the realizations of the process are separated into two groups.
The first group consists of all the realizations that lead to ruin during
the first T periods. The probability of getting a realization belonging
to this group is, by (C.3), 'PT' All the other realizations are left in
the second group, that is to say, realizations which lead to ruin
subsequently, i.e. after the period (0, T L\] or do not lead to ruin
at all (referred to tl = 00). Now the expansion in question can be
partitioned into two conditional expectations
MT(s) = 'P TE{exp(sW(T))lt 1 ~ T}
+ (1 - 'PT)E{exp(sW(T)) IT< t 1 }. (C.6)
Making use of the definition of M(s) and the independence of the
variables W(T) - W(t) and W(t), t ~ T, one obtains
E{exp(sW(Tnlt1 ~ T}
= E{exp(sW(t 1)) exp [s(W(T) - W(t 1) ] Itl ~ T}
T
= L prob{tl = tltl ~ T}E{exp(sW(t))M(sf- t lt 1 = t}
t= 1

= E{exp(sW(t 1))M(sf- h Itl ~ T}. (e.7)

Hence (C.6) can be rewritten as
M T(S) = 'PTE{exp(sW(t1))M(s)T-h Itl ~ T}
+ (1 - 'P T)E{exp(sW(T)) IT < t 1 }. (e.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. (e.9)
To prove that there really exists a unique negative solution s = - R
360 APPENDIX C

of (C9) observe the following facts

M(O) = I
M'(O) = E(Y(t)) = AP > 0
M"(s) = E(Y(t)2 exp(sY(t))) > 0 for every s.
Hence M is a strictly convex function which is increasing at the
origin where it assumes the value 1. In order to prove the existence
of the desired solution - R it is sufficient to prove that M(s) is greater
than I for some negative s. But by the assumption that Y(t) also
assumes negative values there exists a constant So < 0 such that
p = prob {Y(t) ~ so} > O. Hence, for s < 0
M(s) = E(esY(t») ~ pE(esY(t) /Y(t) ~ so)
as s ---. - 00,

and the existence ofthe solution s = - R of (C9) follows. Combining

(CS) and (CS) and substituting the solution s = - R of (C9) the
following equation is obtained
1= 'I'TE{exp( - RW(t1))/t l ~ T}
+ (1 - 'I' T)E{ exp( - RW(T)) IT < tl}' (CIO)
Two approximations are now made. First the second term is
omitted, as it is known to be positive or zero. For the second approxi-
mation note that from the definition of tl the quantity W(t 1 )
in the first term indicates the ruining value of the (negative) profit;
hence it is always ~ - U o' If then W(t l ) is replaced by - U 0 the
first term of the right-hand side of the equation is decreased. Hence

(CII)
or
(C.l2)
This formula is one of the main results of the theory of risk. In
it T can be any positive integer and therefore it holds also when T
tends to infinity.
Before developing the theory further it is of great interest to get
some idea as to how much 'I' = 'I' 00 differs in reality from the upper
limit e- RUo , i.e. would it be feasible to use the approximation
(C13)
C.2 ESTIMATION OF 'I' 361

where C( U 0) is an unknown function being ~ 1 ; by (C.12) it is known

that necessarily C(U 0):::; 1.
The evaluation of the error in this approximation has been the
subject of intensive studies 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). The following shows however that
as a rule the approximation is not too rough, at least not when L1
is relatively small.
Let T tend to infinity in (C.lO). To find the limit obtained by the
second term on the right-hand side another partition will be made
denoting for brevity W = W(T)
(l - 'I' T)E(e-RWI T < t 1) = tPl E(e-RWI W < B, T < t 1)
+tP2E(e-RWIB:::;W, T<t 1 ), (C.14)

where tP 1 =prob{W<B,T<t 1 }, tP2 =prob{B:::;W, T<t 1 }, and

B is a number which will shortly be fixed. The object is to prove that
both terms vanish when T -+ 00. To show this it is sufficient to prove
that when Band T are chosen large enough, both terms are :::; 6/2
where 6 can be fixed in advance to be arbitrarily small. This is so as
regards the latter term if B = -10g(6/2)/R because then e- RW :::;
e- RB = 6/2 in this term. As regards the first term the well-known
Tchebychev's inequality prob {IW-E(W)I~kaw}:::;1/k2 is ap-
plied for k = Tl/4 and gives
prob{IW - E(W) I ~ aT3/4}:::; T-+,
where E(W) = TAP and a is the joint standard deviation of the
variables Y(t). Now taking T large enough to satisfy both inequalities
T ~ (4/6 2) exp(2RU 0) and TAP - T3/4a ~ B it follows that
tPl :::; prob{W:::; B}
:::; prob{IW - TAPI ~ T3/4a }

Noting that, by the definition of t 1 , W = W(T) in the first term of

362 APPENDIX C

the partition (C.14) is ~ - U o and hence exp( - RW)::::; exp(RUo)'

it follows that this term ::::; e/2 and so the proof is completed.
The second approximation when obtaining (C.12) was the replace-
ment of e - RW by eRUo . This means in fact that ruin is always assumed
to occur only by exactly reaching the ruin barrier W = - U0 .
In reality this is, of course, not true as a rule and the barrier is
exceeded more or less. But in the case when ~ is relatively small the
effect of this approximation is generally not very great.
Since the second term in (C.IO) vanishes when T -+ 00, it follows
that
E(e- RW(t ll lt < oo),¥(U ) = 1 = '¥(U o)
1 0 C(Uo)e RUo'
from which
1
-- = E(e-R(W(ttl+uo)lt < 00) (C.15)
C(Uo) 1·

From this formula it can be directly seen that C(U 0) is close to 1 if

the absolute value of the ruining loss W is in general only relatively
slightly in excess of U o. This is the case, for example, when the test
period A is very short or equals 0 and reinsurance cuts off the large
claims. When A -+ 0, only one claim can with probability 1 cause
the 'final ruin'. If, due to the reinsurance, the size of one claim is
limited by some upper limit M, then also IW(t 1 ) - U 0 I < M and
a limit formula
(C.16)
is obtained, which holds when A -+ 0 provided that the assumptions
introduced in Section C.l remain in force (this is the case if the risk
process is a simple compound Poisson process).
It should further be noted that even if C(U 0) were as low as 0.5 and
e- RUo is of the order of magnitude of 0.01, then in replacing the true
'¥ by the approximation (C.l3), the original U 0 would be replaced
by 1.15 U 0' the adjusting figure compensating for the uncertainty
of C(U 0). If C 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 U0 is significant will seldom arise in practical actuarial
work since the choice of the permitted magnitude of the ruin pro-
bability is itself arbitrary. From numerical examples presented in
this book it is seen that the safety loading and, in many cases, the
C.2 ESTI MA TION OF 'I' 363

net retention, are factors which have considerable influence on the

insolvency function, far more, in fact, than the uncertainty of qUO).
Of course, in certain critical cases the study of the exact theory as
developed in the papers referred to may be desirable.
As has been shown, (C.12) holds for every length of the test period
L\ (if the Ys remain independent and stationary) and also when
L\ ~ 0, i.e. when testing is carried out continually at every time.
At the same time the ruin probability Ij;(U 0) 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 observa-
tion years provided solvency is regained at the end of each year.
Equation (C.9) (with s = - R) is inconvenient for computation and
it is useful to develop it in a more suitable form.
According to the definition (C.4)
1 = M( - R) = E(e-RY(t)) = e-(1 H)PRE(eRX(t)),
I.e.
e(l H)PR = M X(t) (R) • (C.17)
In the particular case when the risk process is a simple compound
Poisson process the m.gJ. (C.l7) can be expressed in terms of the
m.gJ. M Z of the claim size dJ. S
MX(t)(R) = exp[nMiR) - n].
By taking logarithms and by using (C.17) this gives the formula

n + (1 + A)PR = n I'" eRZ dS(Z).

Noting that P = E(X(t)) = nm where m is the mean claim size, it

follows that

1 + (1 + A)mR = f: eRZ dS(Z). (C.18)

From this it is seen that 'the insolvency constant' R, the opposite

of the non-trivial root of (C.9), is not dependent on the choice of the
time unit but solely on the distribution function S of the amount
of one claim. Thus it will again be appreciated that the inequality
(C.12) is valid independently of how frequently the insolvency
364 APPENDIX C

measurements are made, and thus it will hold good even if the
possibility of insolvency is measured at every time.
A formula corresponding to (C.I8) can also be obtained when the
fluctuating basic probabilities are assumed to be in accordance with
the distribution (2.9.9). In this case the m.gJ. in (C.17) is given by
(3.4.4), and thus the equation

e(lH)mnR = [1 + (nih) ( 1- too e RZ dS(Z») l h ,

may be written down. This equation may be expressed as

f oo
o eRZ dS(Z) = 1 +
l_e-(1+J.)nmR/h
nih ' (C.19)

from which the insolvency constant R may be found.

In this formula the reservation made on item C.1 must be kept in
mind. It is valid only when the structure process q and the test
period are synchronized, i.e. when a new value for q is drawn at the
end of each period.
It can be noted that (for h -+ 00) there is an essential difference
between (C.18) and (C.19). In the latter the constant R depends on
the size of the company.
Consider the case of a simple compound Poisson risk process with
test period A -+ O. In order to estimate C( U0) in (C.I5), define
V=W(t 1 -I)+U o '
where the values corresponding to the realizations having t1 = 00
are not under interest, and it is defined V = 00 for those realizations.
Denoting Zl = - Y(t 1) (define Zl = 0 when t1 = 00) it is quite
obvious that the conditional probability prob {Zl ~ Ziti < oo}
approaches to the probability that the size of one claim is greater
than V but not greater than Z, when A -+ 0 (see Fig. C.2.1).
Thus, when testing is carried out continually, i.e. A -+ 0, one gets
the limit formula

E(e RZ , IV = V. t < 00) =

, 1
for all V ;;::: O. Denoting, for brevity, by
1
1 - IS(V) v
00
e RZ dS(Z)
'
(C.20)

E 1(-) = E(·lt1 < 00),

 C.2 ESTIMA TION OF 'I' 365

d--V=W(t~l+U
~----I 1 0

Time
0--------4-----4-----~--------~
<t,-1lt.

,
Wet -1)+U =v-z -------~
, 0 1 ~

Figure C.2.l If 11 is small and tl < 00, then V - [V + Y(t l )] = - Y(t l ) = ZI

is approximately the size of one claim under the condition that the claim is
larger than V.

the expected value operator related to the set of all realizations

having tl < 00, it follows from (C.I5) that
1
-- -=E (e-R(W(f!)+uo»)
C(Uo) 1

= El(e-RV+RZ1) (C.2l)
= El {E 1 (e- RV eRZ1 1 V)}
= El {e-RVE 1 (e RZ1 1 V)}.

This formula together with (C.20) can be used when quo) is

approximated.
Consider, as an example, the case that the claim sizes are exponen-
tially distributed, i.e. S(Z) = 1 - exp( - aZ), a > 0, Z ;;:: O. In this
special case it is easily seen from (C.I5) that (note m = I/a)
A
R=I+A. a, (C.22)
366 APPENDIX C

especially R < a. Using (C.20) it is seen that

e-RVE(eRZ,!V = V, tl < (0)

= e- RV I fro
eRZ dS(Z)
1- S(V) v

=e(a-R)Vf ro ae-(a-R)ZdZ=_a_= I +A,

v a-R
is now independent of the value V. Therefore (C.21) can be calculated
directly:

_I_=E {e-RVE (eRZ'!V)}=E (I+A)=I+A

C(U 0) 1 1 1 .

Hence, for exponential claim size dJ.

where R is given by (C.22).

It should be noted that in this example C(U 0) does not depend on
the parameter a. It is quite clear that C(U 0) does not depend very
strongly on the shape of the claim size dJ. in the general case either.
In fact it can be shown that the value of quo) is also of the same
order 1/(1 + A) for many other claim size distributions.
 APPENDIX D

Computation of the limits for

the finite time ruin probability
according to method of
Section 6.9

0.1 Given quantities

n, M, m = a1 , r2 , r3 , T are given, where the moments are functions
of the net retention M as provided in Section 3.6.
Furthermore, the following deterministically defined time-depen-
dent functions are to be given: ri(t), rg(t), rx(t), rp(t), z(t), A(t), Ur(t).
The process is constituted by the algorithm dealt with in Chapter 6
in form (6.7.7) and the ruin probability as given by (6.9.1). The NP
approximation will be used, and for convenience of notation (see
3.11.14a) Ny(x) will be denoted by N(x; y).

0.2 Calculations

n(t) = n TI rg(r)
<=1

flx(t) = flx(t - l)rg(t)rx(t)(1 + z(t))/(I + z(t - I)),

where flx (0) = nm.
P).(t) = flx(t)(1 + A(t))/(l + z(t))
B(t) = P).(t)/(I - c)
flu(t) = riflu(t -1) + Pit) - flx(t).
For brevity the subscripts U or X are omitted in the following
expressions of the standard deviation, third moments, and skewness.
368 APPENDIX D

Note that by the conditions given the standard deviations are the
same for U and X but the third moments and skewness have opposite
signs.
a 2(t) = (r 2/n(t) + a;)Jlx(t)2
a 2(1, t) = a 2(1, t - I)r; + a 2(t)
Jl3(t) = (r 3/n(tf + 3r 2a~/n(t) + a!yq)Jlx(t)3
Jl 3(1, t) = Jl3(1, t - 1)r; + Jl3(t)
y(t) = Jl 3(t)/a 3(t)
y(1, t) = Jl 3(1, t)/a 3(1, t)

cp(t) = 1 - N[JlU(t) - Ur(t); y(1, t)]

a(1, t)
cp(O) = 0

cp(t - 1, t) = f:~ D(t) dU,

which can be evaluated by Simpson's rule

where
d
D(t)=[I-N(xl'y(t))]dU N(x 2 ;y(l,t-1))

U I = Ur(t - 1)
Xl = (r i U + p;Jt) - Ur(t) - Jlx(t))/a(t)

X2 = (Jlu(t - 1) - U)/a(I, t - 1)).

Furthermore (d/dU)N(x2; y(l, t - 1)) is (3.11.18) where ax should be

substituted for a(1, t - 1) and g for y(I, t - 1)/6.
It is appropriate to take as the Simpson's interval h = a(1, t - I)/k,
where k is a constant to be selected to determine the length of
integration step. From experience, values 2 or 3 are sufficient.
Note that now a(1, t - 1) is reduced from the Simpson's sum,
because it is in the numerator of h and in the denominator of
dN/dU.
Finally, the lower limit of \{' T is
\{'~ = max [cp(t - 1) + cp(t) - cp(t, t - 1)],
t
0.2 CALCULATIONS 369

and the upper limit

T
'I'~ = L [cp(t) - cp(t - I, t)],
1= 1

regarding cp(O, I) = 0.
If (6.9.7) is required, the algorithm should be applied
I _ '1''''
'1'''' = '1''''
I 1-1
+ l_cp(t_I)'t'
t-l (m(t) - m(t t - I)).
't',
 APPENDIX E

Random Numbers

E.1 Uniformly (0,1) distributed (pseudo) random numbers

0.6296 0.2398 0.4581 0.3662 0.4208 0.9293 0.0621 0.0482 0.3030 0.4816
0.8251 0.4668 0.9510 0.7583 0.8647 0.8345 0.7651 0.9910 0.2975 0.7888
0.3556 0.0782 0.0987 0.5638 0.7772 0.6325 0.7109 0.9119 0.5130 0.7772
0.4041 0.7764 0.2097 0.6930 0.3849 0.7126 0.1610 0.9153 0.6557 0.0600
0.5427 0.3808 0.6239 0.8641 0.3968 0.2245 0.4621 0.5486 0.3656 0.3263
0.2164 0.3367 0.1768 0.6085 0.6667 0.4235 0.8771 0.0819 0.9354 0.8975
0.8170 0.6431 0.8471 0.4127 0.3852 0.6215 0.0381 0.6105 0.4928 0.0758
0.8154 0.2912 0.8076 0.8412 0.7663 0.3364 0.7387 0.3261 0.1178 0.3806
0.1709 0.0589 0.8460 0.2330 0.9489 0.9852 0.5352 0.0496 0.1115 0.5144
0.9877 0.2549 0.8766 0.3164 0.5828 0.2l38 0.3750 0.8691 0.3863 0.6536
E.2 Normally (0,1) distributed random numbers

0.6211 -1.1983 -0.5891 1.8947 1.2064 0.9722 -0.7708 0.1443 -0.3138 0.9395
-1.3769 -0.6988 -0.9492 -0.9243 0.2836 -0.6349 -0.3449 -0.0277 0.3379 -0.1070
0.9971 -0.3024 -0.5235 -0.0315 0.2423 0.3936 -0.3664 -1.4520 -0.2174 -2.0894
0.1015 -1.1994 -2.0946 1.1515 2.6083 1.0364 -0.3801 -0.7835 -1.3011 -1.1076
0.3792 -0.1713 -1.2002 -0.6104 0.4016 -0.2125 -0.8872 0.5802 1.2419 0.5373
0.2474 0.7294 0.6982 0.6237 0.6423 -1.3561 -0.1343 0.4261 0.0666 -0.9646
-0.1009 -0.2162 -0.5640 -1.4068 -0.1834 0.0118 -0.3795 -1.9057 0.7346 -1.7216
1.2772 0.3481 0.6472 0.3538 0.4539 -0.0924 -0.3890 1.4981 0.8572 -0.7642
-1.0352 -0.2686 -1.4689 -0.3880 -1.4175 0.3399 1.7439 -1.7855 3.0626 -0.0526
1.9466 -0.6060 -0.2452 0.6367 -0.8552 0.2417 0.3301 0.4039 0.6269 -1.4241

E.3 NP(O, 1, ')I) distributed random numbers

(a) }' = 0.5
1.7366 0.3073 0.5520 -0.7275 -1.0791 0.1868 -1.2969 2.1834 -0.0579 1.1571
-1.7942 0.9956 -0.4829 -1.1483 -0.4519 -0.5708 0.1116 -0.2678 -1.6752 1.8835
0.3308 0.1720 0.0786 1.0786 -0.1472 1.2757 -0.2711 -0.5553 -0.3948 0.0301
-0.1475 2.4209 0.5093 2.8619 0.2556 0.6315 -0.0934 0.1079 0.0994 0.2971
0.0998 -1.9343 1.2952 -1.3509 -0.2523 1.0017 -0.8328 -0.2878 0.2090 1.3646
-0.4914 -0.6692 -1.8089 0.3057 0.2782 0.1107 -1.0372 0.1593 0.1673 -0.4022
-1.5223 -0.5254 -0.8486 -0.0634 -0.8544 -1.5779 -0.8696 1.0558 -0.5158 0.6833
-1.4123 1.8236 -1.6554 0.0246 0.9758 -1.0005 0.0056 -0.1732 -0.7792 0.1885
-1.3262 -0.5699 0.6259 2.2760 0.3705 0.5842 -1.2739 -1.2732 -0.5890 -0.4409
1.3115 0.0324 0.2220 1.0291 -0.9462 -0.1104 0.2320 0.6383 0.1168 1.3788
(b) Y= 1.0

0.6947 -0.4526 -1.4981 2.3178 -0.1857 -0.4655 -1.4804 -1.0227 -1.2948 -0.2342
0.3982 0.1092 1.8420 -1.1016 - 1.4655 -0.1568 -0.4998 -1.0772 -0.4776 0.3044
-1.9760 0.3013 -1.0754 2.2538 0.2810 1.0067 3.8306 -0.1465 -0.2595 -1.1833
1.2749 -1.3010 -0.8147 0.2485 -1.5294 0.2597 -0.3160 -0.7570 3.2355 -0.2540
-l.U~06 -0.1517 -0.6264 -0.3259 -0.2138 0.2344 -1.4570 1.0950 0.2183 0.1515
-1.1553 2.2340 -0.5611 -0.5717 -0.9523 -0.2603 -0.3859 -0.0134 -0.2194 -0.8178
4.2132 -0.6998 0.8881 -0.7762 -0.2138 -0.1504 -0.4196 2.0429 -1.0294 0.2643
0.3140 -0.4378 3.7994 -1.2701 -0.4573 -0.3335 1.0099 -0.6327 0.1963 -0.1462
-0.8154 -0.3648 1.4191 -0.8595 -1.1017 1.7996 -0.0024 0.4220 -0.3136 -1.2075
-0.8118 0.1803 0.3486 2.5939 0.5823 -1.0132 -0.5640 -0.5312 - 1.1353 0.0013
 APPENDIX F

Solutions to the exercises

F.l Chapter 2

2.4.1
co nk co nk- 1
0: 2 = k~O k 2 e- nk! = n e-nk~l (k _ 1)!k (substitute k = k - 1 + 1)

= ne- n[ n k~2 (kn~-;)! + k~l (kn~- ;)J

2.5.1 E(k) = n = 0.01 x 1000 = 10, P A = n(l + ).)S = I1S. The total
claims amount is X = kS. The minimal value of the required security
reserve U 0 is the smallest U which satisfies

prob{U + P A - X ~ O} = prob{X:O:; U + PA}

= prob{k:O:; (U + PA)/S} ~ 0.99.

(a) prob {k:O:; 17} = 0.986 and prob {k:O:; 18} = 0.993. Hence, round-
ed upwards to the next integer value of k

U O=(l8-11)S=7S=7000.

(b) If N((k-n)/Jn)=0.99, then k~17.4 and therefore again

U o = 7000.

(c) U 0 = 7000.
2.5.2 54300
374 APPENDIX F
e-1
L: (k -
00

2.5.3 P/I000 = 2)-

k=3 k!

A shorter way to get the solution is to observe that

Prein,=Ptot-Pcedent= I X lOOO-lOOO[P 1 +2(l-PO-Pl)]
2.6.1
M(s) = exp[n(eS - 1)] = exp[ns + ns 2/2! + ... ]
=1+[ ]+[ ]2/2!+···
= 1 + ns + (n + n2)s2/2! + (n + 3n 2 + n3)s3/3 !
+ (n + 7n 2 + 6n 3 + n4)s4/4! + ...
2.6.2 For t> 0
Fk(t) = prob{Wk ~ t} = 1 - prob{k(t) < k}
k-l
= 1- L: e-Pt(pt)i/i!.
i=O

By differentiation it follows that

k
I'()-F'()- P -ptk-l
Jkt - k t - (k _ I)! e t .

Note that Wk/p is r(t; k) distributed; see (2.9.1).

2.6.3 Conditions (i) and (ii) are obvious. For proof of condition
(iii) define a variable W~ U= 1,2; k = 1,2, ... ) which gives the time
when the kth event of the process j occurs. The variables W: and
W; are independent and their distribution functions continuous
(see exercise 2.6.2). Then W: - W; is also continuously distributed.
Hence, W: i= W; by probability 1, i.e. the probability of a multiple
event is zero.

2.8.1 The proof is straightforward when the relevant expressions

are substituted for E(X IY = Y).
F.l. CHAPTER 2 375
2.9.1

0(. = _1_ foo Zi e- zZh-l dz = _l_foo e- zZi+u-l dz = r(h + i)/r(h)

I r(h) 0 r(h) 0

= h(h + 1) ... (h + i-I).

If the dJ. of q is H(q) = r(hq; h), then the dJ. of hq is r(x; h) and
hence
O(i(q) = O(/h i•
Substituting into the expressions defining the relevant characteristics,
and noting (1.5.6), the requested formulae are obtained.

f
2.9.2
OO 1
M (s) = eSz _ e- z Zh-l d7.
r 0 r(h) -

Substitute u = (1 - s)z

1 1 foo
Mr(z) = (1 _ S)h x r(h) 0 e- uuh - 1 du
=(l-s)-h.

2.9.3 Substitute u = q(h + n - n eS) in

M(s) =- 1 foo e nq (e S - l ) e- hq (hq)h-l h dq

r(h) 0 •

Note that the condition of convergence is h + n - n eS> 0 or

s < In(l + (h/n».
2.9.4

( h+k-l)(_h )h(_n )k
k n+h n+h
(h+k-l)(h+k-2) ... (h+ l)h 1 nk _nn k
-'-------~--,----------'-------'--'---

k!
X
(l+iY X ----+e
(n+h)k
-
k!'

2.9.5
k 0 I 2 3 4 5 6 7 8 9 10 II 12 13 14 15

P.( x 10 000) 67 337 842 1404 1755 1755 1462 1044 653 363 181 82 34 13 5 2
P.( x 10000) 173 578 1060 1413 1531 1429 1191 907643428271 164 96 54 30 14
376 APPENDIX F

0·2

-,L_-,
I
I
L_

0·1 I
L __

5 10 k 15
Figure F.1.l Poisson probabilities (solid line) and Polya probabilities;
n=5,h=lO.

2.9.6 See Fig F.l.l

- 2 [ h J2h [ 2h J2h
M(S) = h + n(l - eS ) = 2h + 2n(1 - e') .

2.9.7 Apply repeatedly integration by parts and note that k is a

positive integer

= F(k, n),
(see exercise 2.6.2).

2.9.8

when k ~ n(l - (l/h)) (providing h > 1).

F.2. CHAPTER 3 377

F.2 Chapter 3
3.3.1
n = 100 x 0.01 =1
It is convenient for the computations to take £100 as the monetary
unit. S is a step function having a step 2/3 at I 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 ofthe abbreviation Pk(l) = Pk (the difference
compared with (3.2.3) is that N, not k, is taken as the variable):
F(X)=p o =e- 1 =0.37 whenO~X<1
F(X) = Po + P1(t) = 0.61 when 1 ~ X < 2
F(X) = F(l) + P2(t)2 + P1(t) = 0.82 when 2 ~ X < 3.
The following steps are constructed in a similar way: F(3) = 0.92,
F(4) = 0.97, F(5) = 0.99, F(6) = 1.00.

3.3.2
E(X) = 1 x (1 x f + 2 x t) = 1.33 or £133
(J = J[1 X (12 x f+ 22 x t)] = 1.41 or £141.

3.3.3

F(X) = k~/k(n)Sh(X) = k~J IV e-nq(nkqr dH(q) )Sk*(X)

= foo ( f e-nq(nkqt Sk*(X)) dH(q) = foo Fnq(X) dH(q).
o k=O' 0

3.3.4

1130':Z) = E(L(Zj - EZ))3

j j

= LE(Zj - EZ/ + 3 LE(Zi - EZ;)E(Zj - EZy

j itj
+6 I E(Zh - EZh)E(Zi - EZ;)E(Zj - EZ).
h<i<j
378 APPENDIX F

The assertion follows from the fact that at least one of the factors
included in the last two sums is always zero.

3.3.5

Substituting az(q) = CT~ + 1, (3.3.7) follows. The skewness is obtained

in a similar way by using the connecting equations between the
central moments and the moments about zero.
3.3.6 Let q l ' ... , qk be the jump points of H. Then H can be written
as

i= 1 i= 1

Let Fn be the simple compound Poisson dJ. with the same claim size
dJ. as X and having n as expected number of claims, and let Gn be
the corresponding standardized dJ., i.e.

Then (see (3.3.l7))

Gn(x) = Fn(Px) = {"J Fnq(Px) dH(q) = Jl Fnq,(Px)h i = Jl Gnqi(x)hi ,

and, by (3.3.l5),
as n -+ 00.

Therefore
k
Gn(x) -+ I hic(x - qJ = H(x) as h -+ 00.
i= 1

3.4.l Expand first (3.4.l) in powers of s according to (1.6.4) and

denote

Then the integrant in (3.4.2) can be expanded as follows

F.2 CHAPTER 3 379

Substituting the expressions of E, arranging according to the powers

of s, substituting in (3.4.2) and integrating, the requested moments
/3.J are obtained,
.
applying again (1.6.4), as the coefficients of the
terms sJ/j!
PI = mn
/3 2 = na 2 + n2m2a2(q)
/3 3 = na 3 + 3n2ma2aiq) + n3 m3 a3(q)
/3 4 = na4 + 4n 2ma 3 a2(q) + 3n 2m2a2a2(q) + 6n 3 m2a 2a3(q) + n4 m4 a4 (q),
where

aiq) = tXl qi dH(q).

The requested formulae (3.3.7) are obtained by passing from the
moments /3j and aiq) about zero to the central moments.

3.4.2 Now M z(s) = eS •

3.4.3 The proof is the same as for exercise 2.9.3 when eS is substi-
tuted by M z(s).

3.5.1 Denote the unknown rate of death by q. Then the following

values for the step function S follow from (3.5.1)
5000q
S( 100) = 5000q + 1000q + 2 x 2000q = 0.5; S(250) = 0.6 and
S(500) = 1.

3.5.2 Assuming the formula valid for k - I the next step is verified

f:
by substituting (3.5.4) and the expression of S into:

Sh(X) = S(k-I)*(X - Y)dS(y),

and integration by parts. Verify still the validity for k = 2.

I1x= nal = n/c; (Jx = Jna 2 = J(2n)/c
F(2) = 0.82.
380 APPENDIX F

3.5.3 From (2.9.17) and the general moment rule

a/az + b) = a\(z) + C )aj-tbaj_t(Z) + ... + bj ,

it follows

at(az + b) = aat(z) + b = a
a,{/!:, + b) = a2 a2 (z) + 2aba t (z) + b2 = a(a + 1)
a3 (az + b) = a3 a3 (z) + 3a 2 ba 2 (z) + 3ab 2 a t (z) + b 3 = a(a + l)(a + 2).

Substituting the moments at (z) = 0, a2 (z) = 1 and a3 (z) = y of the

standardized variable z, the coefficients of (3.5.8) and (3.5.9) are
obtained as solution.

3.5.4 For z > 0 (see (3.5.7))

e-l1.al1.
1 - r(w, a) = r(a + 1)
fro el1.-u (u)l1.-t
w ~ du

h
= 3C(q(w) + 4q(w + h) + 2q(w + 2h) + 4q(w + 3h) + ... ),

where q is the notation for the integrand

11.-1
q(v) = el1.- v ( ~ ) ,

and h is the integration step of the Simpson formula. Furthermore,

applying the Stirling formula to the constant C, preceding the above
integral, we have

1 1 + -~
llC =j(2na) ( 1 + ~- 1 1 + ... ) .
l2a 288a 2

The series should be extended until the terms vanish within the
required limits of accuracy. The integration step h can be taken
as h = (l/k)ja, where k is another coefficient; e.g. k = 10 seems to
give a satisfactory accuracy. For z < 0 the integral can be taken from
w to - 00 giving r(w, a) and h should be negative.
F.2 CHAPTER 3 381

3.5.5
prob{Z:::; Z} = prob {Y:::; In(Z - a)}
1 f1n(z-a)
= -- e-(1/2a 2 )(Y-/l) dy,
afo -00

S'(Z) is the derivative of this function.

3.5.6 (i) For a = 0 we have Z = eY, where Y is a normally N(Il, ( 2)

distributed variable. Its moments about zero are

aJ = E(Zj) = E(e jY) = M(j) = e/l j e~fa2.
(ii) Passage from Z to Z - a transforms the first moment a~ to
a~ + a but does not change the variance and skewness. The
requested formulae are obtained expressing the central moments
by means of the above moments.
(iii) Choose I] = J(e a2 - 1) and use (ii).

3.5.7
rJ. .
a.=--ZJ for j < rJ..
J rx-j 0

3.6.1
for Z < M

for Z ~ M.

3.7.1 Clearly Ilx = nm = Nnjm = 1840 in both cases. In case (i)

ax = J(na 2 + n 2m 2a;) = 383,
and in case (ii) each risk unit forms a section, and therefore

ax = J[ ~(np2 + n;m2a~) ] = 105.

3.8.1
X= 0 1 2 3 4 5 6
F = 0.162 0.215 0.441 0.521 0.695 0.760 0.858
382 APPENDIX F

3.11.1 Y= 0.6364, x = 2.828; (a) 0.0023, (b) 0.0094; (c) 0.0093.

3.11.2
prob{X < - nm 2 2)} = prob{X < O} < s.
J (na 2 +n Z
m a q

This implies
nm/ j(na 2 + n2m2a~) > - xe where xe = N y-l(S).
The condition is obtained from this inequality.
For calculation of xe the skewness y = Yx is needed. However, this
depends, according to (3.3.7), on the unknown n. Hence an iteration
is necessary.
Note that a conservative evaluation is obtained by taking xe =
N- 1(s). In this case x = N- 1(lO-4) = - 3.72 which can be taken as
an initial value for iteration. It follows from (3.11.15) that n = 143.
Furthermore, substituting in (3.3.7) this value and the given value
y = 0.531 are obtained. Then according to (3.11.15) and (3.11.16)
x = - 3.01. Repeating the iteration loop four times more the final
evaluation n = 85 is obtained.
If the initial s were 10- 3 , then n = 60 would result.

3.12.1 According to the conditions of the exercise and (2.9.17),

a 1 = a/a = J1.x
a 2 = a(a + 1}/a 2 = ai + J1.i .
Solving these equations the following is obtained
a= J1.x/ai, a = J1.i/ a i and Yx = 2ax/J1.x·

3.12.2
y = y/6 - 6jy + 3(2jy)2 /3(2/y)1 /3[1 + yz/2]1/ 3.
Develop the term ( )1 /3 as a Taylor series

y = z - !(Z2 - 1) + ~y2 Z3 - •••

6 108 .

F.3 Chapter 4
4.2.1
U = yj(na ) - ).P; U + LlU = yj(0.9na 2) - ).P. Hence LlU =
(U + ).P)(:]0.9 - 1) = (20 + 0.05 x 100)(~O.9 - 1) = -1.28
F.3 CHAPTER 4 383

4.2.2 If £100 is taken as the monetary unit and the unknown

frequencies of the sums 1 and 2 are denoted by p and 1 - p then
m = p + 2(1 - p) = 2 - p
az = p + 4(1 - p) = 4 - 3p
U = yJ[(4 - 3p)n] - A(2 - p)n,

U as a function of p reaches its maximum for

4 3yZ
p = "3 - 4nA Z ~ 0.52,

and the maximum U = 1.5 = £150.

4.2.3
U = yJ(rz/n + (j'~)P - AP = 0
A = yJ(r z/n + (j'~)

a 1 = {OOze- Z dZ=I;a z = f:ZZe-ZdZ=2;rz=az/ai=2

A = 2.33J(2/1000 + 0.04Z) = 0.14.

4.2.4 Substitute

a i M-~+i
ai =--·---· .
a-I a-I

into

and calculate U for suitably chosen M values. After some trials the
requested value is obtained: M = 2.0.

4.2.5 Put into the NP formula X = U + (1 + A)P where U is

15.72£106 according to item 4.2(b). F is 0.0015, 0.0042 and 0.0302
respectively for Xc = 5, 10 and 20£10 6 .

4.4.1 Substitute the values related to the company C 1 into (4.1.7)

and solve y = 2.816. Then modify (4.1.7) for the merged company
applying (3.7.6), (3.7.9) and (3.7.12). U = £512 000.
384 APPENDIX F

4.5.1 Substitute a;(M) = iM i - 1 (1 - S(M)) into

U'(M) = hJ(n/a 2(M)) x a~(M) - Ana'l (M)

4.5.2 Prove first lim M _ oa2(M)/M 2 = 1. It follows that U'(M)

tends to a negative limit as M -+ 0 + if n > (y/A)2.

4.5.3 Write (4.5.8) in the form

U'(M)=An(l- S(M))Ja~M)[AJn - Ja~~)J.

Let Zo be the lowest point for which S(Z) = 1. Zo may be infinite.
Then in the interval [0, Zo) the expression in brackets determines
the sign of U'(M). Note that a2(M)/M2 is in this interval a non-
increasing function of M (proof!). Hence if U'(O) > 0 (see exercise
4.5.2), U(M) is in this same interval an increasing function. If
i
U'(O) < 0 and a2(M)/M 2 < /(A 2n) at Zo then U(M) has a minimum
inside the interval, otherwise it is a decreasing function.

4.5.4 Reading from Fig. 4.5.1 one finds M ~ £0.5 million. The
approximate value is £0.45 million.

4.5.5 ~ U = U'(M)~M and ~P = P'(M)~M and the return R =

A~P. The required rate is
ir(M) = R/~U = ArP'(M)/U'(M).
U'(M) was calculated for exercise 4.5.1 and
P'(M) = na~ (M) = n(1 - S(M)),
(see exercise 4.5.1). Then after some reductions and interpolating
from Table 3.6.1 we obtain

4.5.6 Let M move upwards until the rate of ir(M), calculated in the
previous exercise, reaches i o. It is convenient to solve M 2 /a 2 from
the equation. From the given data the value 1130 is found for it.
The line most nearly corresponding to it in Table 3.6.1 is M = £2.512
million. Then P = £80.8 million and U = £18.4 million are obtained
from (4.1.8).
F.3 CHAPTER 4 385

4.5.7 Consider the new policy as another 'portfolio' to be merged to

the original one. The premium income of it is qM and variance
q(l - q)M 2 ~ qM2. Then
U = yj(na 2 ) -)..p = yj(na 2 + qM2) - )..(P + qM).
Solving this:
M = 2)"(U + )"P)/[(U + )..p)2/na2 - q)..2] ~ 2)..na2/(U + )..P).

4.5.8 For M > 1

4 1 1 7 2
a l =S-SM;a2 =15+S 1nM .

Substituting these expressions and the given data into (4.l.8) the
equation
1
20 = j(253.3 + 217.21nM) - 4+M

follows. By standard methods (or more easily by a few trials) the

solution M = 4.2 is obtained.

4.7.1 Applying (3.3.7) and counting also the cases Zre = 0 into n we
have

f:
where

a 2 (re) = (Z - A)2 dS(Z) + (B - A)2(1 - S(B)).

Decomposing these expressions in terms and substituting the

moments (3.6.3) for M = A and M = B the required formula is
obtained.

4.7.2

nbZ~[_l_ __
IX - 1 A,,-l
1_J
B"- 1 •

4.7.3 Calculate the first moment according to (3.6.3)

al(M) = 0.6 Zo - 0.05Z~/A2.
386 APPENDIX F

Then P = na i (oo) =nO.6Zo andPX/L = n x 0.05 x Z~ x (A -2 - B- 2 ).

Putting Zo = 1, A = 2 and B = 5 we have Px1dP = 1.75%.
For evaluation of inflation replace Zo by 1.1Zo ; then the correct
risk premium percentage is increased in the ratio 1.1 2 = 1.21.

4.7.4
Sre(Zre) = (S(M + Zre) - S(M))/(1 - S(M))
= 1 - e- czre

4.8.1 Substitute into (4.8.1 a) dF(X) = dN(y) and X = Ilx +

O"x(y + Yxy2/6 - Yx/6) which follow from (3.11.14) (lowest line)
and (3.11.16). In manipulating the integrand, note that yN'(y) =
- NI/(y) and y2 N'(y) = NI/,(y) + N'(y) (see (3.11.4)).

4.9.1 7 years.

4.10.1
2 1 - (1 - Z)2t 2 2
0" p = 1 _ (1 _ zf Z 0"

For proof of the second part of the exercise replace the individual
variances by their joint upper limit.

4.10.2

E(G) = k f:(B - X)dF ~B- P

= too BdF - too X dF = too (B - X)dF.

Hence k ~ 1, equality being valid only in the case F(B) = 1.

4.10.3
Z_ 0.1 _
- 2J(10/200) - 0.224
180000
Pi = 0.224 x 1000 + 0.776 x 300 = £273.
 F.4 CHAPl ER 5 387

F.4 Chapter 5

5.1.1 (i) is obvious and (ii) is simply the formula E(E(Y 1 X)) = E(Y)
given in exercise 2.8.1. To prove (iii) note that quite generally
V(Y) = V(E(YIX)) + E(V(YIX)) ~ V(E(YIX)),
where V(Y 1X) = E[ (Y - E(Y 1X) )21 X] is the conditional variance
of Y. In fact, since E(Y 1X) = R(X) is a function of X, it is seen that
V(Y) - V(E(YIX)) = E(y2) - E(R2(X))
= E[E(y2 - R 2(X)IX)]
= E[E(y2 - 2R(X)Y + R 2(X)IX)]
= E{E[(Y - R(X))2IX]} = E[(Y - R(X)f] ~ 0

(iv) is just (iii) with Y replaced by X - Y.

5.2.1 Let Xk be the original aggregate claim and Yk the share of

the company k after the exchange, according to the rule given in
the exercise, and Y~ the share according to the specified rule:

Then

The inequality follows from the general rule

cov(y;, Y k ) ~ Jvar(YJ x Jvar(Y k ),
well known in probability calculus and from the assumption of the
equality of the company distributions which means var(Y;) =
var(Yk )·

5.2.2
388 APPENDIX F

(iii) v= YPP-AP.

5.2.3 The proof can be found either by straightforward calculations

or more easily by observing that a linear co-ordinate transformation,
which stretches x co-ordinates in the ratio J(Vz/V l ) and keeps
Y co-ordinates unchanged transforms the ellipses concerned to
circles. Then the assertion becomes evident, because the transforma-
tion keeps straight lines straight and tangents as tangents.

F.5 Chapter 6

6.2.1 Denote A(t) - Ao = f(t) = Abt • Then

f(t) - ad(t - 1) - azf(t - 2) = Ab t - Z [b Z - al b - az ] = O.
Equating the expression in brackets with zero, two values are
obtained
b= tal ± tJ(ai + 4a 2 ),

and the general solution is

f(t) = Al b~ + A2b~.
The coefficients Al and A z are determined by initial conditions.
The solution is oscillating if b l and b z are complex, i.e. when
ai+4a z <0.
Moving into polar coordinates
b =re±i'l'
1. Z '

with

r = J ( - az ) and cp = arc tan [ J ( - all - 4a z )] (0 < cp < n).

The general solution can be written in form

f(t) = Al rt ei'l't + A z rt e - i'l't,
or, since ei'l't = cos (cpt) + i sin (cpt) and by some straightforward
transformations
f(t) = Art sin (cpt + v),
F.S CHAPTER 6 389

where A and v are other coefficients. "I:he oscillation is damped if

r < 1. The wavelength is directly seen from condition 2n = cpT, i.e.
T = 2n/cp.
In the particular case given in the exercise

f(t) = A(t) - AO = - 0.10690 (~} sin (1.93216t - 0.48669),

with T = 3.25. Verify the result for t = 1,2, ... comparing with the
values which are obtained directly from the difference equation;
the first step being
f(l) = - tf(O) - tf( - 1) = - 0.075.
Compared with Fig. 6.2.1 the oscillation is now rapidly vanishing
as t increases. This shows that the stationary oscillation in Fig. 6.2.1
is maintained by the continued impulses due to the 'noise' e(t).

6.4.1 Now
n(r) = nr;(1 + z(r»
Jlx(r) = n(r)m(r) = nmr~r~(l + z(r»
P(r) = Jlx(r)/(l + z(r»
where nand m are obtained from (3.7.2) and (3.7.4). Then (see (3.7.9»
t2

ai(tl' t 2 ) = L (r2/n(r) + a~)P(r)2,

and the coefficients (6.4.5) become time independent (and the argu-
ment can be dropped from (6.4.5». Hence, for convenience of
calculations, the time dependence is concentrated in the variables
n(r~ Jlx(r) and P(r) only.

6.7.1 The second term of the right hand side of (6.7.11) is replaced
by

6.8.1 As primary entries in all the following cases are either uni-
formly (0,1) generated random numbers r or normally N(O, 1)
distributed numbers y.
390 APPENDIX F

(i)
(I) n small
First make a program for the calculation of
k
F(k) = L Pi(n),
i=O

e.g. by using the algorithm (2.5.2). If n is fixed to be the same

during the simulation algorithm, then it may be rewarding to
calculate F(k) in advance and to save as an array. Then the
simulation algorithm is as follows:

(1) Generate r
(2) k=0
(3) If r :::; F(k) deliver k
(4) k --+ k + 1; go to (3).

Note: If n is not very small, it is useful to start the algorithm

consisting of steps (3) and (4) at k = n and if r > F(n) to proceed
upwards, otherwise downwards (see Rubinstein 1981, Section
3.7).
Furthermore, pay attention to the possibility that owing to
the rounding-off inaccuracies the limit value of F(k) may
occasionally remain less than I when k --+ 00, if F is computed
repeatedly for every sample.

(II) n large
Denote by y the argument of N[ ] in (3.5.14), substitute (see
(2.4.5» Y= llJn and z=(k-n)l)n and solve k:
(a) k = [(y + A)IBr - n - t,
where y is normally (0, I) distributed and

(b) A = 6Jn - 1/(6Jn); B = 3 X 22 / 3 x n 1/ 6 •

t
Note the correction term - in (a). It is due to roundings and
is not obtained directly from (3.5.14).
Then

(1) Generate y
(2) Calculate k from (a).
F.S CHAPTER 6 391

(ii)
1
X = --In(l- F).
a

(1) Generate r
1
(2) X = - -In r.
a
Note that rand 1 - r are equally distributed.
(iii)
(1) Generate rl' r 2 , ••• , rh
h 1 1 h
(2) X = L --Inri = --In Il rio
i=l C C i=l

(iv)
(1) Generate y
(2) X = a + exp(yu -IL).
(v)
(1) Generate r
(2) X = X or- 1 /a..

6.8.2 Examples are plotted in Fig. F.S.l.

). It}

0.4

-0.4

5 10 15 20 25
t
Figure F.S.l.
392 APPENDIX F

6.8.3 Example of an outcome of stratified simulation. Sample size

10000.

x o 2 4

c(X) 1352 0 2394 2013 1494 986 703 439 256 159
F(X) 0.1354 0.1354 0.3749 0.5763 0.7257 0.8244 0.8947 0.9385 0.9641 0.9800

x 10 11 12 13 14 15 16 17 18 19

c(X) 85 40 28 17 15 6 2
F(X) 0.9885 0.9924 0.9952 0.9969 0.9984 0.9990 0.9992 0.9996 0.9997 0.9998

The largest X was 29, making F(29) = l.0000.

F.6 Chapter 7

7.10.1 From

it follows
A
- - - ~ Yt (1 - c - Ab)O'q (1 - r~19p) - t
1 - r.
+ Z m + ur ,
19p

where zm is the amplitude of the cycles. Substitute according to

(6.5.12) Ap = A-i.19p wand solve A
A ~ [(1 - c + iigpw)A + (zm + ur )(1- r igp)]/(1 + A),
where

A = Y,O'qj(1 - rigp).
1 + r igp
Substituting the numerical values from item 7.1(d) (note Ye = 2.33,
ur and w = 1.71) we have A~ 0.028.

7.10.2 Substitute into

L = J1.u (t) - Y,O'u(t),
first the expression (6.6.8a)
1 - rt
J1.u (t) = u(O)rt - -1-A,
-r
F.7 CHAPTER 8 393

where briefly r = r¥p' According to (6.6.11) and (6.6.6) and denoting

p= 1/r2rg = rgr;/rj

Hence

t[A
L=r u(0) +---y
r- l'
J(r2(1- C - Ab)2(1 -
n(l - p)
pt)p)] - -A-
r - l'
which proves that L --+ 00 when t --+ 00 if p < 1 and the expression
in the brackets is positive.

F.7 Chapter 8

8.2.1

P = SA(v) I~t: l(t)v t

•

8.2.2 All the sums following will run from 0 to w - 1

S - 1 : v) LP(t) [A(V)2 - 2A(v)v t + 1 + (V2y+ 1],

2
= (

from which (8.2.8) follows, noting that kp(t) = 1.

F.8 Chapter 9

9.1.1 The introduction of the 'bonus barrier' U max means in fact

394 APPENDIX F

that U(t) is replaced by U*(t) = min (rp*(t - 1) + P A - X(t), Urn • x )'

where U*(O) == Uo. Since
{U*(t) ~ O} c {X(t) ~ M},

and since the variables X(t) are mutually independent it follows that

1 - 'l'ro = prob( t01 {U*(t) ~ O} ) ~ probC01 {X(t) ~ M} )

= II prob{X(t) ~ M} ~ exp ( - J1 prob{X(t) > M} ).

The last estimate is an application of the inequality 1 - x ~ e- x .

9.1.2 By making use of the relations derived in item 6.9(d) we have

for every t > 0

Obviously 'l't is non-increasing. Assuming that (9.1.3) is true, let

t1 ~ to be such that L cp(t)/(l - cp(t - 1)) is < 1. Then
t> tl
_ ~ cpW
'l'ro - 'l't, +L A'I't -.. :: 'l't, + (1 - 'l't) L ) < 1.
t>t, t>t,l - cp(t - 1
9.1.3 Substitute in to the Tchebychev inequality
1
prob {Iu(t) - fluWI ~ kO"u(t)} ~ k2

k = flu(t)/O"u(t).
Then
(a) cp(t) < prob {I u(t) - flu(t) I> flu(t)} < O"u(t)2 / flJt)2.
Denoting briefly r = r igp ' it follows from equation (6.6.8a)
1 - rt
(b) fl (t) = u(O)rt + A - - = A[l + o(rt)],
u 1- r
where A is a constant and the expression 0(') vanishes when its
argument tends to zero. Furthermore, according to (6.6.11) (see
exercise 7.1 0.1)
r 1 _pt
(c) 0"2(t) = --.l.(1 - C- A )2 pr2t _ _ .
U n b I-p
F.9 CHAPTER 10 395

where p = l/r 2 rg • This expression can be written in the form

(J';(t) = [r 21 - (r2p)I]B,

and then
qJ(t) < [r 21 - (r2p)l]C,
where Band C are constant. Because rand r2 p = l/rg are by assump-
tion < 1 the premises of exercise 9.1.2 are satisfied which proves the
assertion.

9.2.1 According to (9.2.5b) R = 0.235 and then e- RU = 0.095.

9.2.2
e -RU/(e -RU e -RM) =e RM =e
(-RU)-M/U .

For e- RU = 0.Q1 and M/U :::; 0.02 this is :::; 1.096 and for e- RU = 0.001
it is :::; 1.148.

F.9 Chapter 10

10.4.1 Substituting in (10.4.2) it follows that

E(H(U)) = aE(G(U)) + b,
from which the assertion follows directly.
 Bibliography

SA Scandinavian Actuarial Journal (earlier Skandinavisk Aktuarietid-

=
skrift)
MS = Mitteilungen der Vereinigung Schweizerischer Versicherungsmathe-
matiker
AB= The Astin Bulletin
CA = Transactions of the International Congress of Actuaries
BARAB = Bulletin de l'Association Royales des Actuaires Belges
GP = Geneva Papers on Risk and Insurance
JIA = Journal of the Institute of Actuaries
JRI = The Journal of Risk and Insurance
PCAS = Proceedings of the Casualty Actuarial Society

Abramowitz, M. and Stegun, I. (1970) Handbook of Mathematical Functions,

(9th print) Doover Publications, New York.
Adelson, R.M. (1966) Compound Poisson distributions, Operations Research
Quarterly, 17.
Ammeter, H. (1948) A generalization of the collective theory of risk in regard
to fluctuating basic probabilities, SA.
Ammeter, H. (1953) The calculation of premium rates for excess of loss
and stop loss reinsurance treaties, CA.
Ammeter, H. (1963) Experience Rating, AB.
Bailey, A.L. (1945) A generalized theory of credibility, PCAS, 32.
Balzar, L.A. and Benjamin, S. (1980) Dynamic response of insurance systems,
JIA.
Balzar, L.A. and Benjamin, S. (1982) Control of insurance systems, JIA.
Beard, R.E. (1947) Some statistical aspects of non-life insurance, JIA.
Beard, R.E. (1959) Three R's of insurance ~ risk, retention, and reinsurance,
JIA.
Beard, R.E. (1971) On the calculation of the ruin probability for a finite
time period, AB.
Becker, F. (1981) Analyse und Prognose von wirtschaftlichen Zeitreihen der
Deutschen Schaden- und Unfallsversicherung, Inaugural- Dissertation,
Universitat Mannheim, Veroffentlichungen des Instituts fUr Versicher-
ungswissenschaft der Universitat Mannheim, 19.
BIBLIOGRAPHY 397

Bellman, R. (1957) Dynamic Programming (reprinted 1965), Princeton.

Benckert, L-G. (1962) The lognormal model for the distribution of one
claim, AB.
Benjamin, B. (1977) General Insurance, Heinemann, London.
Benktander, G. (1970) Schadenverteilung nach Grosse in der Nicht-Leben-
Versicherung, MS.
Benktander, G. and Segerdahl, C-O. (1960) On the analytical representation
of claim distrubution with special reference to excess of loss reinsurance,
CA.
Berger, G. (1972) Integration of the normal power approximation, AB.
Bertram, J. (1981) Numerischen Berechnung von Gesamtschadenverteilung,
Bliitter der Deutschen Gesellschaft fur Versicherungsmathematik.
Bohlmann, G. (1909) Die Theorie des mittleren Risikos, CA.
Bohman, H. (1964) To compute the distribution function when the charac-
teristic function is known, SA.
Bohman, H. (1979) A mathematical model of insurance business and how it
may be used, Geneva Papers on Risk and Insurance no. 11 (January).
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. (1962) Application of game theory to some problems in automobile
insurance, AB.
Borch, K. (1963) Economic theory and their application to insurance, AB.
Borch, K. (1970) The Mathematical Theory of Insurance, D.C. Health & Co.,
Lexington, MA.
Brigham, E.O. (1974) The Fast Fourier Transform, Prentiss Hall, Englewood
Cliffs, New Jersey.
Biihlmann, H. (1967) Experience rating and credibility, AB.
Biihlmann, H. (1970) Mathematical Methods in Risk Theory, Springer,
Heidelberg.
Biihlmann, H. and Jewell, W.S. (1979) Optimal risk exchanges, AB.
Butsic, R.P. (1979) Risk and return for property-casualty insurers, in Total
return due a property-casualty insurance company, Casualty Actuarial
Society.
Casualty Actuarial Society (1980) Inflation implications for property-
casualty insurance, Discussion paper program.
Chatfield, C. (1980) The Analysis of Time Series, 2nd edn, Chapman and Hall,
London.
Christy, G.A. and Clendenin, J.C. (1978) Introduction to Investments,
McGraw-Hill, New York.
Chung, K.L. (1974) A Course in Probability Theory, 2nd edn, Academic Press,
New York.
398 BIBLIOGRAPHY

Cornish, E.A. and Fisher, R.A. (1937) Moments and cumulants in the
specification of distributions, Rev. Int. Statist. Inst., 5, 307.
Cox, D.R. and Miller, H.D. (1965), The Theory of Stochastic Processes,
Methuen. London.
Cramer, H. (1926) Review of F. Lundberg, SA.
Cramer, H. (1930) On the mathematical theory of risk, Skandia Jubilee
Volume, Stockholm.
Cramer, H. (1945) Mathematical Methods of Statistics, Almqvist and Wick-
sells, Uppsala.
Cramer, H. (1955) Collective risk theory, a survey of the theory from the point
of view of the theory of stochastic processes, Skandia Jubilee Volume,
Stockholm.
De Wit, G.W. and Kastelijn, W.M. (1980) The solvency margin in non-life
insurance companies, AB.
D'Hooge, L. and Goovaerts, M.J. (1976) Numerical treatment of the deter-
mination of the structure function of a tariff class, CA.
DuMouchel, W.H. and Olsten, R.A. (1974) On the distribution of claim
costs, Berkeley Actuarial Research Conference on Credibility.
Eggenberger, F. and Polya, G. (1923) Uber die Statistik der vergetteter
Vorgiinge, Zeitschriftfiir angewandte Mathematik und Mechanik, I.
Ferrari, J.R. (1968) The relationship of underwriting, investment, leverage,
and exposure to total return on owners' equity, peAS.
Forrester, J.W. (1972) Industrial Dynamics, The MIT Press, Massachusetts.
Friedman, J.W. (1977) Oligopoly and the Theory of Games, North Holland
Publishing Company, Amsterdam, New York and Oxford.
Frisque, A. (1974) Dynamic model for insurance company's management,
AB.
Galbraith, J.K. (1973) Economics and the Public Purpose.
Galitz, L. (1982) The ASIR model, GP.
Gelder, H.v. and Schauwers, C. (1981) Planning in theory and practice with
reference to insurance, Delta Lloyd Insurance Group, Amsterdam.
General Electric Company (1962) Tables of the Individual and Cumulative
Terms of Poisson Distribution, D. van Nostrand Co., Princeton.
Gerber, H.U. (1979) An Introduction to the Mathematical Risk Theory,
S.S. Huebner foundation monographs, University of Pennsylvania.
Gerber, H.U. (1982) On the numerical evaluation of the distribution of
aggregate claims and its stop loss premiums, Insurance: Mathematics and
Economics, I.
Geusau, A.B.J.J. von (1981) Some applicable actuarial forecasting models,
Nederlandske Reassurantie Groep.
Godolphin, E.J. (1980) Specifying univariate models for the Zoete equity
index, Maturity Guarantees Working Party.
Gossiaux, A-M. and Lemaire, J. (1981) Methodes d'ajustement de distri-
butions de sinistres, MS.
BILIOGRAPHY 399

Grote, J.D. (1975) The Theory and Application of Differential Games, D.

Reidel Publishing Company, Dortrecht, Holland.
Hammersley, J.M. and Handscomb, D.C. (1964) Monte Carlo Methods,
Methuen, London.
Heiskanen, J. (1982) Degree ofloss and surplus reinsurance of fire insurance,
unpublished research work, Helsinki.
Helten, E. (1977) Business cycles and insurance, GP.
Heubeck, K. (1971) Optima Ie Reservepolitik als Markoffscher Entscheid-
ungsprozess, Blatter der Deutschen Gesellschaft fur Versicherungsmathe-
matik, Band X, Heft I.
Hovinen, E. (1964) A method to compute convolutions, CA.
James, K.D. (1981) Underwriting cycles in property-casualty insurance
industry, Discussion papers, Casualty Actuarial Society.
Jewell, W. (1976) A survey of credibility theory, University of California,
Berkeley.
Jewell, W. (1980) Models in insurance, CA.
Jewell, W. and Sundt, B. (1981) Recursive evaluation of compound distri-
butions, AB.
Johnson, N.L. and Kotz, S. (1969) Distributions in Statistics, 1, 'Discrete
Distributions', Wiley, New York.
Johnson, N.L. and Kotz, S. (1970) Distribution in Statistics, I, Continuous
Univariate Distributions, Wiley, New York.
Johnson, N.L., Nixon, E., Amos, D.E. and Pearson, E.S. (1963) Tables of
J
percentage points of Pearson curves for given PI' and P2 expressed in
standard measure, Biometrika, 50.
Karten, W. (1980) The new 'Schwankungsriickstellung' in annual statements
of German insurers. An application of the theory of risks, GP.
Kauppi, L. and Ojantakanen, P. (1969) Approximations of the generalized
Poisson function, AB.
Kendall, M.G. and Stuart, A. (1977) The Advanced Theory of Statistics,
Charles Griffin & Co., London.
Kimball, S.L. and Pfennigstorf W. (1981) The regulation of insurance
companies in the United States and the European communities, a com-
parative study, International Insurance Advisory Council, Chamber of
Commerce of the United States.
Kotler, P. (1975) Marketing Decision Making, a Model Building Approach,
Holt, Rinehart and Winston, New York.
Landin, D. (1980) Risk accumulation, a lecture paper.
Levine, S.H. (ed.) (1975) Financial Analysts' Handbook II: Analysis by
Industry (Vol. 2) Dow Jones-Irwin, Inc. Homewood, II.
Loimaranta, K. Jacobsson, J. and Lonka, H. (1980) On the use of mixture
models in clustering multivariate frequency data, CA.
Lorie, J.H. and Hamilton, M.T. (1973) The Stock Market: Theories and
Evidence, Richard D. Irwin, Inc. Homewood, II.
400 BIBLIOGRAPHY

Lundberg, F. (1909) Zur Theorie der Riickversicherung, CA.

Lundberg, F. (1919) Teori for riskmassor, Stockholm (Forsiikrings-
inspektionen).
Lundberg, O. (1964) On Random Processes and their Application to Sickness
and Accident Statistics (reprint of thesis 1940) Almqvist and Wiksells
Boktryckeri, Uppsala.
Martin-Lo[, A. (1983) Premium control in an insurance system, an approach
using control theory, SA.
Maturity Guarantees Working Party (1980) Report, JIA, 107.
Mayerson, A.L. (1965) A Bayesian view of credibility (discussion) PCAS, 52.
McGuinness, J.S. (1954) Top Management, Organization, and Control of
Insurance Companies, Stanford University Press, Stanford, CA.
McGuinness, l.S. (1966) A Feasibility Study of Flood Insurance, New York.
McGuinness, J.S. (1970) Elements of time-series analysis in liability and
property insurance ratemaking, PCAS.
Mihram, G.A. (1972) Simulation, Academic Press, New York.
Munich Reinsurance Company (1971) The Influence of Inflation on In-
surance, Munich.
Nemhauser, G.L. (1966) Dynamic Programming, Wiley, New York.
Neumann, J. v. and Morgenstern, o. (1964) Theory of Games and Economic
Behaviour, Wiley, New York.
Panjer, H.H. (1981) Recursive evaluation of a family of compound distri-
butions, AB.
Pearson, K. (1954) Tables of incomplete r-function, Biometrika Office.
Pentikiiinen, T. (1952) On the net retention and solvency of insurance com-
panies, SA.
Pentikiiinen, T. (1975) A model of stochastic-dynamic prognosis. An appli-
cation of risk theory to business planning, SA.
Pentikiiinen, T. (1976) Stochastic-dynamic prognosis, CA.
Pentikiiinen, T. (1977) On the approximation of the total amount of claims,
AB.
Pentikiiinen, T. (1978) A solvency testing model building approach for
business planning, SA.
Pentikiiinen, T. (1979) Dynamic programming, an approach for analysing
competition strategies, AB.
Pentikiiinen, T. (1982) Solvency of Insurers and Equalization Reserves (Vol. I)
Insurance Publishing Company, Helsinki.
Pesonen, E. (1964) Solvency measurement, CA.
Pesonen, E. (1967a) On the calculation of the generalized Poisson function,
AB.
Pesonen, E. (1967b) Magnitude control of technical reserves in Finland, AB.
Pesonen, E. (l967c) On optimal properties of the stop loss reinsurance, AB.
Pesonen, E. (1969) NP-approximation, SA.
BIBLIOGRAPHY 401

Pesonen, M. (1983) Optimal reinsurances, to be published in SA.

Philipson, C. (1968) Some compound Poisson processes applicable in the
theory of risk, CA.
Rantala, J. (1982) Solvency of Insurers and Equalization Reserves (Vol. II)
Insurance Publishing Company, Helsinki.
Reischel, M. (1981) Dynamische Riickversicherungs- und Ausschuttungs-
politik, Inaugural Dissertation, Universitat Hamburg.
Revell, J. (1979) Inflation and financial institutions, Financial Times.
Roberts, E. (1978) Managerial Applications of System Dynamics, MIT Press,
Cambridge, Massachusetts.
Roy, Y. (1981) A stochastic simulation model for reinsurance decision-making
by ceding companies, dissertation, University of Pennsylvania.
Rubinstein, R. (1981) Simulation and the Monte Carlo Method, Wiley,
New York.
Ruohonen, M. (1980) On the probability of ruin of risk processes, SA.
Seal, H. (1969a) Simulation of the ruin potential of non-life insurance
companies, Transactions of Society of Actuaries, 21.
Seal, H. (1969b) Stochastic Theory of a Risk Business, Wiley, New York.
Seal, H. (1971) Numerical calculation of the Bohman-Esscher family of
convolution-mixed negative binomial distribution functions, MS.
Seal, H. (1974) The numerical calculation of U(w, t), the probability of
non-ruin in an interval (0, t), SA.
Seal, H. (1977) Approximations to risk theory's F(x, t) by means of the gamma
distribution, AB.
Seal, H. (1978) Survival Probabilities, Wiley, New York.
Seal, H. (1980) Survival probabilities based on Pareto claim distribution, AB.
Segerdahl, c.-a. (1959) A survey of results in the collective theory of risk.
The Harald Cramer Volume, Uppsala.
Shpilberg, D. (1977) The probability distribution of fire loss amount, JRI.
Sparre-Andersen, E. (1957) On the collective theory of risk in the case of
contagion between the claims, CA.
Straub, E. (1978) How to fix retention, MS.
Sugars, E.G. (1974) Selected results from a risk theoretic simulation of an
insurance company, JRI.
Taylor, G.C. (1978) Representation and explicit calculation of finite time
ruin probabilities, SA.
Thorin, O. (1971) An outline of a generalization-started by E. Sparre-
Andersen - of the classical ruin theory, AB.
Thyrion, P. (1967) Regards sur Ie developpement recent de la theorie du
risque, AB.
Veit, K. (1976) The use of system dynamics simulation models for corporate
long range strategic planning, CA.
Venezian, E.C. and Gaydos, E.M. (1980) The effects of variable size of risk
402 BIBLIOGRAPHY

on capital solvency requirements, Research report of Insurance Services

Office, New York.
Whitney, A.W. (1909) The actuarial theory of fire insurance rates, CA.
Witt, R.C. and Miller, H. (1980) A comparative analysis of relative costs
under competitive and non-competitive rate regulatory laws, University
of Texas, Austin.
 Author index

Adelson, 103 Gossiaux, 46

Ammeter, 159 Grote, 348

Bailey, 169 Hamilton, 207

Balzar, 188,275 Heiskanen,90, 146
Becker, 31, 188 Helten, 31, 188
Bellman, 330 Heubeck, 320
Benckert,74
Benjamin, B., 142 James, 31
Benjamin, S., 188,275 Jewell, 103, 169, 170, 182,339,346
Benktander, 79, 81, 82
Bertram, 103, 124 Karten,284
Bohlmann, ix, 292 Kastelijn, 124
Bohman, 31, 124, 159, 188,275 Kauppi, 108
Borch, 179, 181, 182,339 Kimball,143
Biihlmann, xi, 23, 44, 170, 182 Kotler, 332
Butsic,207
Landin, 249
Campagne, 123 Lemaire, 46
Christy, 207 Loimaranta, 46
Clendenin, 207 Lorie,207
Cramer, ix, 292 Lundberg, ix, 18, 57

D'Hooge,46 McGuinness, 31,162, 188,207,249

De Wit, 124 Martin-Lor, 275
Dumouchel, 79 Mayerson, 169
Miller, 188
Eggenberger, 42
Esscher, 159 Nemhauser, 330, 344

Ferrari, 207 Ojantakanen, 108

Forrester, 330 Olsten,79
Frisque, 275
Panjer, 103
Galitz,320 Pentikainen, xi, 31, 50, 90, 142, 146,
Gelder, 320 188, 198,211,214,218,260,
Gerber, xi 271,273,275,284,330,347
Gesau,320 Pesonen, E., 142, 175
Godolphin, 210 Pesonen, M., xiii, 174, 175, 182
Goovaerts, 46 Pfennigstorf, 143
404 AUTHOR INDEX

Philipson, 354 Segerdahl, ix, 79, 81, 82

Polya,42 Shpilberg, 79, 80
Sparre-Andersen, 23
Rantala, xi, 125, 147, 188,202,249, Sundt, 103
257,260,261,282
Reichsel, 320 Thorin,23
Revell, 198
Ro berts, 330 Veit, 330
Roy, 320
Whitney, 80, 94
Seal, xi, 124 Witt, 188
 Subject index

Actuaries' all share index, 210 Compound distribution, 50

Additivity of Poisson variables, 27 possibly mixed, 51
Additivity of the Poisson processes, 28 Compound Poisson process, 47,51
Adjustment coefficient, 312 Compound process, 9
Aggregate claim, 50 Confidence limits, 160
Aggregate safety loading, 218 Contamination model, 42
AR process, 187,201 Continuous procedure, 9
ARMA process, 187 Convolution, 14
Asset risk, 208 Corporate planning, 319
ASTIN, 327 Counting process, 9
Asymptotic behaviour, 138 Credibility, 162, 164
Autocorrelative time series, 186 Cycles, 31, 185,261,266,270

Basic equation, 126, 129, 140,214, Danger index, 75

216,219,329 Decomposition of portfolio, 94, 211
Benktander distributions, 79 Decomposition of the investment
Beta function, 123 portfolio, 208
Bonus, 162,297,304 Degree ofloss, 90
Break-up basis, 10 Differential games, 348
Business games, 348 Discrete equidistant distribution, 100
Business planning, 319 Discrete procedure, 9
Business strategies, 7, 333 Distribution functions, II
Distribution-free approximations,
Cancellations, 296 138,148
Catastrophic claims, 125, 136,248 Dividends, 216, 328
Central limit theorem, 105 Dynamic accumulation factor, 217, 261
Central moments, 14 Dynamic calculus of the yield of
Characteristic function, 17 intereSt, 217
Characteristics of distributions, 14 Dynamic control rules, 272
Claim number process, 9, 191 Dynamic discounting, 220
Claim size 9, 47, 60, 63, 67, 135 Dynamic programming, 248, 330,
Claim statistics, 62
Claims, 9, 47, 183 Edgeworth expansion, 107,355
Claims process, 2, 47 EEC convention, 123, 143
Cohort analysis, 288, 299, 300 Elasticities, 332
Collective approach, 18 EML,89
Collective theory of risk, ix Equitability, 181
Commercial bank lending rates, 209 Equity linked policies, 297
Competition models, 345 Excess of loss reinsurance, 85, 154, 265
406 SUBJECT INDEX

Exclusion of multiple events, 19, 21 Limit distributions, 56

Exogenous variables, 189 Loading for expenses, 203, 260
Expense loading, 204 Logarithmic-normal distribution, 72
Expenses, 327 Lundberg's coefficient, 312
Experience rating, 162, 164
Exponential distribution, 68 Market pressures, 272
Exponential experience rating rule, 167 Mean value
Exponential polynomials, 69 composed, 95
Extinction rate, 79 compound Poisson, 54
Extreme values distribution, 84 definition, 14
finite time, 195,212
Factor K, 139 mixed Poisson, 36
Finite time horizon, 183 Poisson, 22
Forecasting, 188 Polya,40
solvency ratio u, 223
Gamma function, 26, 38, 69, 121 Merger of portfolios, 143
General purpose models, 190 Mixed compound Poisson d.f., 51
GNP, 188 Mixed Poisson process, 34, 35
Going-concern basis, 10 in the narrow sense, 354
in the wide sense, 354
IBNR problem, 49, 326 Model building, 7
Independence of increments, 19,20 Models, 320
Indexation, 297 Moment about zero, 14
Individual approach, 18 Moment generating function, 15,36,
Individual risk theory, 292 40, 59
Industrial dynamics, 330 Monetary variables, II
Inflation, 65, 191, 260, 267, 271, 296 Monte Carlo method, 233, 235, 298
Inflation rate, 304 Mortality, 296
Insolvency coefficient, 312 Multi-unit markets, 347
Insurance statistics, 159 Multi-unit models, 345
Inter-occurrence time, 23 Multi-unit risk exchange, 178
Interest period, 215 Multiplayer games, 179,347
Interest rate, 194, 206 Multivariate utilities, 343
Inverse normal distribution, 84
Inversion of the characteristic function, Negative binomial, 51
124 Negative risk sums, 6
Investment income, 205, 325 Net retention, 133, 145, 152, 260,
264,265, 314, 330
Kurtosis Net worth, 4
compound Poisson, 54 Normal approximation, 24, 104, 115
definition, 14 Normal power approximation, see NP
mixed Poisson, 36 formula
Poisson, 22 NP formula, 108, 109, 117, 121, 158,
Polya,40 161
NP generator, 243
Lag between the claims inflation and NP transformation, 118
the premium inflation, 261
Laplace transform, 17 Objectives 142, 336
Large claims, 75, 136 Optimal reinsurance, 171, 174
Lattice distribution, 100 Optimal strategy, 343
Least squares approach, 170 Outstanding claims, 49, 327
Legal minimum solvency margins, 142
Life insurance, 28, 288 Parameter uncertainty, 4, 200
SUBjf!CT INDEX 407

Pareto distribution, 74, 151,249 Return functions, 331

Pareto distribution, two-parameter, 78 Risk exchange, 175
Pareto optimal, 176, 181, 182 Risk exposure variation inside the
Pearson system, 83, 123 portfolio, 43
Planning horizon, see Time span Risk proneness, individual, 43
Poisson case, 51, 54, 59,97, 153 Risk reserve, 4, 132, 141
Poisson process, 19,349,352 Robustness concerning the claim size
Polya case, 51,55,60, 159, 169 dJ., 214
Polya distribution, 43 Ruin barrier, 8, 142
Polya process, 38, 43 Ruin probability
Positive risk sums, 6 finite time, 8, 199,226,227,247,250,
Predictor function, 187 316,336,343,367
Premium infinite time, 9, 308, 311, 312, 316,
earned, 203 357
life insurance, 29 one year, 8, 128, 131, 133
risk, 126 Ruin state probability, 8, 226, 251
safety loaded, 127, 199 Run-off triangles, 327
t years, 199,213,324
written, 202, 213, 323, 324 Safety loading, 127, 179, 199,260,
Profile figure, 136 274,284,330
Profit algorithm, life insurance, 295 Sales response function, 332
Profit return, 162 Sample path, 3, 245
Profitability control, 201 Second-order bases, 296
Pure random fluctuations, 32 Self-control, 272
Sensitivity tests, 265, 275
Quasi-log-normal d.f., 78, 83 Shadow claims, 67,136
Quota share reinsurance, 85 Short-period oscillations, 32
see also Structure variation
Random numbers Simulation, 233, 239, 249, 264, 275, 346
arbitrary distribution, 234 Size of the portfolio, 263, 314
definition, 233 Skewness
exponential, 239 composed, 97
gamma, 239 compound Poisson, 54
log-normal, 239 definition, 14
mixed compound Poisson, 241, 371 finite time, 197,212
normal, 238, 271 mixed Poisson, 36
Pareto, 239 Poisson, 22
Poisson, 239 Polya,40
pseudorandom, 234 solvency ratio u, 224
uniform, 233, 238, 370 Solvency margin, 4, 282
u process, 245 Solvency profile, 278
Random variable, 3 Solvency ratio, 132, 220, 255, 303
Rate of the real growth, 261 Sparre-Andersen processes, 23
Real rate of return, 207 Standard data, 131, 260
Realization of the process, 3, 6 Standard deviation
Reciprocity of two companies, 175 composed, 96
Rectangular distribution, 233 compound Poisson, 54
Recursion formula for F, 100 definition, 14
Reinsurance, 65, 84, 323 finite time, 196,212
Relative mean risk, 178 mixed Poisson, 36
Renewal processes, 23 Poisson, 22
Reserve funds, 142 Polya,40
Reserve of unearned premiums, 325 solvency ratio u, 224
408 SUBJECT INDEX

Standard insurer, 260 Time-dependent variation of risk

State variables, 331 exposure, 29
Stationariness of increments, 19, 20 Time-span, 7, 9, 126,260, 266
Step function, 13 Trading result, 214
Stieltjes integrals, 12 Transition equations, 331
Stochastic bundle, 245 Trends, 31, 184
Stochastic processes, 5, II
Stochastic variables, II Underwriting process, 3, 127,301,326
Stop loss reinsurance, 156, 172 Utility, 182,338
Strategic triangle, 337
Strategies, 336 Variance as a measure of stability, 171
Stratified sampling, 240 Variance reduction, 240
Structure function, discrete, 45 Variation range of u(t), 281
Structure variation, 32, 46, 260
Supercatastrophes, 248 Waiting time, 28
Surplus reinsurance, 88 Weibull distribution, 84
Survival probability, 227, 309 Weighted process,
Synchronized structure variation, 98 see Mixed Poisson process
System dynamic, 330 Wilson-Hilferty formula, 25, 71, 120
System parameters, 331 Wind-up barrier, 142
Wold's decomposition theorem, 187
The most dangerous distribution, 140
Time lag, 271 Yield of investments, 215, 261, 296