A R e v i e w o f the Collective T h e o r y of R i s k

Part I. Comments on the deve|opment of the theory

By Carl Philipson, Stockholm

1. Introduction and definitions

Let the random variable ~: be distributed with distribution function G(~) of the con-
tinuous, discontinuous or mixed type, and the random variables x~ be distributed
with the distribution functions V¢(x) with the corresponding characteristic functions
~(~) = ~_+~ e~lx dV£(x), ~j being a real variable, and i the imaginary unit, and with the
generating functions ~ ( z ) defined by W~(-ilog z). ]f the range of x is restricted to
the right semi-plane the lower limit of the integral in tile definition of ~£(~j) can be
replaced by zero.
A random variable x having the characteristic function ~1[~p2(7~)] is said to be
(equivalent to) a variable x~ generalized by the generalizing variable x2 (Fr.: variable
x~ gdndralisde par la variable gdn#ralisante x2 [348, 180It).
If the distribution function V~(x) for different values of~ are mutually independent,
the distribution function corresponding to exp [~ log ~07)dG(~)], the integral being a
Stieltjes integral taken over the range off, shall be denoted 1I(~) V~(x). If~ is allowed to
take also non-integral values, this function shall be called the convolution in the
extended sense of V£(x) over the range of £. For the opposite case, see the next para-
graph.
If ~ assumes only integer values, the asterisk product of V£(x) over the range of
is the convolution of V~(x) for ~ = 1, 2..... n and defines the distribution function of
z.~-~x£. The convolution of V~(x) and Vt(x) can be written in the following well-
known form, where, if, particularly, V2(0)=0, the limits of the integral may be re-
placed by zero and x.

N<~ ,,~Ydx) = V~(x)*V&)= V~(x-~)av~(~). (I a)

- oo

If, in addition, x~ are identically distributed with the distribution function V(x)
indepcndcntly of ~, the convolution takes the form Vn*(x), where V°*(x) shall be
taken equal to unity, and V~*(x)= V(x).
If n is a random variable assuming only integer values, and distributed with the
probability distribution Qn(r), ~ being a parameter or parameter vector, and, if the
variables ~r,, xr ...... ..~r, are mutually independent, and identically distributed with
the distribution V(x) independently of r~, then the distribution function of the sum
of these variables for all z~ <~ is for each given vahle o f r defined by F(x, r) with the

Numbers in square brackets refer to the list of literature in Part I1, this journal 1968, 3-4.
The sign § followed by a number refers to a section of this Part I.
1 - 693822 Skand. AktuarTidskr. 1968
2 Carl Philipson

corresponding characteristic function ~(~, ~), given by the following relations (1 b),
(1 c), where v'(q) corresponds to V(x).

F(x, ~) = ~ Qn(r) Vn'(x), (1 b)

n=o

¢(~, ~) = r i_~
co
O,(O ~," (,7). (1 c)

It is here anticipated, that the sum of ~cr,, all r~ ~<~, is a random function of~,
,Y(r) say, which fulfils the conditions for the probabilities being well-defined (see
§ 2, below). (1 b) defines, then, a stochastic process constituted by the discontinuous
random function )7(r) with a discontinuous or continuous parameter or parameter
vector r. In this context r shall always be used for the original parameter measured
on the absolute scale (or absolute scales), which is (are) independent of any properties
of Qn(z); a function of ~ shall always be denoted by a bar.
In the particular case, where 0n(T) denotes the probability distribution o f the number
n o f claims occurring in a group of insurances, when the parameter passes the domain
from zero to • in the parametric space, and, where V(x) is taken to mean the condi-
tional distribution function of the size of one claim at any parameter point, relative
to the hypothesis that one claim has occurred at the parameter point, here called the
claim distribution, the random function ,V(r) represents the total amount of the claims
paid for in the group, when the parameter passes the domain considered. Then, the
process constituted by X(T) is called the risk process, and if, particularly, V(x)=
E(x-k), where k is a given constant and e(y), here and in the following context, the
unity distribution, being equal to zero for negative and to unity for non-negative
arguments, the process is said to be elementary and, in the opposite case, non-ele-
mentary [230]. In a non-elementary process )((r) is, always, a variable n generalized
by the generalizing variable x, wherefore some authors have used the term a generalized
process (Ge.: ein verallgemeinerter Prozess) for the non-elementary process. In French
only the term un processus gdn#ralisd is used. The term generalized is, however, a
wider concept, as in some cases, referred to in the following context, the charac-
teristic functions of an elementary process can be written in the form ~1[~P~(~7)],
so that the random function of an elementary process also may be a generalized
variable, therefore the present author prefers the use of the terms in [230]. Some
other authors have for a non-elementary process used the term a compound process
(Ge.: ein zusammengesetzter Prozess), which here shall be kept for another purpose
(see the next paragraph), cf. for the terminology [82, 152, 185, 268, 348].
Let fin(r) be a probability distribution defined by the following Stieltjes integral.
The distribution defined by fin(r) is, in this context, called a compoundPoisson distribu-
tion (Ge.: eine zusammengesetzte Poisson-Verteilung; Fr.: une distribution de Poisson
composde). It might be remarked here, that in French the term compos6e is used
also for a composed Poisson distribution to be defined in § 4 here below.

.P, (~) = e - ~ c~ [v~(r)]'~d~ O(v, ~) / n !, (1 d)

Skand. AktuarTidskr. 1968

 Review of the collective theory of risk. I 3

where i(T) will be defined below, and, where U(v, T) for every fixed value of 3 is the
distribution function of the non-negative variable v, fulfilling certain conditions for
v = 0. It is here called the risk distribution (by certain authors also called the structure
function). If in (1 b) fi~(3) is substituted for 0~(3), the process defined by the expres-
sions obtained is called a c o m p o u n d Poisson process, cPp, (Ge.: ein zusammenge-
setzter Poisson-Prozess [185]; Fr.:processus de Poisson compos6 [348]). Other authors
use the terms mixed Poisson process (Ge.: ein gemischter Poisson Prozess) or weighted
Poisson process (Ge.: ehl gewichteter Poisson-Prozess [82]). All quotations given
previously in this paragraph concern a particular case of (i d), where O(v, 3) is equal
to U(v) independently of 3. To differentiate between this case, and the general case,
where O(v, 3) may or may not depend on 3, the processes will be denoted with addi-
tion of the words in the narrow sense (i.n.s.), and in the wide sense (i.w.s.) for the
particular and the general case respectively (Fr.: au sens restreint, au sens large,
respectively). The Poisson process is a cPp i.n.s, where, particularly, U(v)=e(v-y~),
ya a given positive constant, and a Polya process a cPp i.n.s., where, particularly,
dU(v) is represented by a Pearson Type I l l frequency curve beginning at origo.
It is often advantageous to transform the parameter 3 (see the second paragraph
of § 4 here below), for a cPp i.n.s., by the relation t = t(T), and for a cPp i.w.s, by this
transformation in the Poisson expression, and by the substitution of U(v, s) for U(v, 3),
where the relation between s and r shall be determined with regard to the form of
O(v, ~). After the transformation the functions appearing in (1 d), and in (1 b), (I c)
after the insertion of (1 d) will be denoted without a bar as functions of t or t, s. The
transformation of z leads to simple expressions even if the assumptions are extended
by the assumption, that the claim distribution depends on the parameter point for the
occurrence of the claim, denoted, after the transformation of 3, by V(x, t) with the
corresponding characteristic functions 'P01, t). If, particularly, t is one-dimensional,
and V(x, t) is continuous in t, this assumption leads to the following expressions
(1 e), (1 f), which define a cPp i.w.s, with the claim distribution V(x, t).

F(x; t, s) = ~ P~ (t, s) Wn'(x, t), (1 e)

rl~O

~(,?; t, s) =P0{t[l - Z(r/, t)], s), (1 f)

where W(x, t) =l/t f~V(x, u)du, and Z(•, t) = lit ~ ( ~ , u)du, which under mild condi-
tions of regularity are consistent expressions with Z(~, t) being the characteristic func-
tion corresponding to W(x, t). The modifications of(! e), (1 f) in cases, where t is a vec-
tor, and where V(x, t) is discontinuous in t, are self-evident.
The conditional probability of the occurrence of v claims, if t is one-dimensional,
in the interval (t~, t2) relative to the hypothesis, that n claims have occurred in the
interval (0, t~), t~ <t2, is denoted Pn.n+v(t~, t,). If T, x + d 3 and 1 are substituted for
t,, t, and v, the conditional probability in this case reduces for a cPp i.n.s, and for a
wide set of cPp i.w.s, to pn(3)d~ +o(dr). Here ~ ( 3 ) is called the intensity function of
the process.
Skand. AktuarTidskr. 1968
4 Carl Philipson

If t is one-dimensional and p~(z)/n is uniformly bounded for all n, tbe mean of

P,(t) is given by the following expression, y~(s) being the vth semi-invariant of U(v, s).

y~(s)t=~,t(s)~(r)=ff [~o~,(u)P~(.)]d.. (lg)

The conditional mean of v, i.e. with respect to Pn.n+v(tx, ta) can be written in the
following form for a cPp i.n.s.

(t~ - h)p~(h). (I h

By the normalization of U(v, s), so that, in a cPp i.n.s., the mean becomes equal
to unity and, in a cPp i.w.s., the mean becomes equal to s, (1 g) and (i h) will be simpli-
fied.

2. The risk process

In [I 11] Cram6r has treated the risk theory from the point of view of the theory of
stochastic processes. He has inter alia proved for the risk process h~ its classical form,
to be defined below, that the sum of the claim amounts in the interval (0, t) of the
one-dimensional, transformed parameter, X(t), is--if confined to a restricted space
to be defined below--a random function associated with well-defined probabilities,
induced by the measures of the variables o~ = {t,, x,; ~, = 1, 2, ...}, where t,, xv for
the vth claim are the parameter point of occurrence, and the amount of the claim
respectively; co is said to belong to a reference space of an enumerable number of
dimensions. Thus, each point co represents the actual development of the claims in
one particular case, which corresponds to one and only one individual function X(t)
called a sample fimction or a realization of the process. Thus, the reference space is
mapped on the restricted functional space according to X(t, ~ ) ~ X ( t ) , where X(t, co)
for any fixed value of ~ is a sample function, and for any fixed value of t, ~t0 say,
represents, for different values of co, different values of the random variable X(t°)
distributed with the distribution function F(x, to). The restriction of the fimctional
space is defined by allowing only for such sample functions, which are relevant in the
risk theory, i.e. step-functions with the discontinuity points t~, defined here above,
and in the intervals between consecutive such points of constant value. The classical
form of the risk process is taken to mean a Poisson process with a claim distribution
equal to V(x) independently of the parameter. A great part of [111] concerns this
form, Cram6r has, however, indicated an extension to the Polya process with reference
to [230], and another extension to a Poisson process with a claim distribution depend-
ent on t with reference to [145].
Later in this review (§ 5) it shall be referred to further extensions of the classical
form. The discussion reviewed in the previous paragraph has been extended [230]
to include all processes, which fulfil the conditions for the validity of Markov's
differential equations (I.e., p. 33), by Ove Lundberg and to include a wide set of
cPp i.w.s, by the present author [303].
Skand. AktuarTidskr. 1968
 Review o f the collective theory o f risk. I 5

Owing to the strong connections between the risk theory and the stochastic process
theory, which besidcs in [111] have been elucidated in [230, 40], references to some
studies into the general stochastic process theory have been included in the list of
literature in Part 1I of this review, even if these studies do not particularly deal with
the risk process [39, 97, 112-113, 115, 117-120, 135-138, 152, 174, 203-204, 234,
364]. Some investigations into pure mathematics, the results of which have been used
by authors dealing with the risk theory, have also been included in the reference list
[4, 36, 45, 63, 149, 259, 326]. Of all the items in the list of literature, in total 365
items, thus, 26 items do not directly concern contributions to the risk theory.
The collective theory of risk was, originally, created by Filip Lundberg. A great
part of his contributions were published before 1930, thereinafter, he published
two papers a few years later, his first paper was published in 1903 [223-229]. Cram6r
reviewed and developed his theory in 1919, 1926 and 1946 [98-99, 107]. According to
Cram6r, Lundberg anticipated ideas, which later were propounded in the general
theory of stochastic processes; the modern development of the general theory started
in the early thirties with two important papers by Kolmogoroff [203-204], and was
developed by Bartlett, Cram6r, Doob, Feller, Gnedenko, Khintchine and many
others. As Cram6r stated in [I07], Filip Lundberg's theory is to be considered an
important particular case of the general theory of stochastic processes, the early
contributions to the risk theory can, therefore, be considered an auspicious pioneer
work for the knowledge of stochastic processes, accomplished a long time before the
general principles of the theory of such processes had been established. On the other
hand, the modern development of the general theory has deepened our understanding
of the problems involved in the risk theory, and facilitated the rigorous deduction of
the results in this theory, by giving more satisfactory tools for the solution of such
problems (cf. [40, 111, 107]).
In [82] Biihlmann acknowledges Filip Lundberg's contributions, by using modern
terms, saying that he investigated stochastic processes with independent increments,
and with sample functions--see the first paragraph of this section--in the form of
step-functions, a long time before such processes had been rigorously deducted.
Biihlmann places in this sense Bachelier beside Lundberg; Bachelier had in 1906
introduced a mathematical theory for the Brownian movement of molecules (Th~orie
des probabilit6s continues, Journ. Math. Pures et Appliqudes). At the Astin Colloquium
in Arnhem, 1966, Borch pronounced in an oral contribution, that Bachelier had
stimulated a continued study of the ideas propounded by him, and that this stimula-
tion had led to numerous new important contributions to the field of these ideas.
Borch added, that this should only to a limitcd extent apply to Filip Lundberg.
It is true, that Lundberg based most of his developments on assumptions, which
lead to the classical form of the risk theory (see § 4 here below). This approximation
of the reality has been used by several other authors, and has, in fact, led to very
remarkable results. A great part of these results, particularly with regard to the ruin
theory (see ~j 9 and 10 below), should have been very difficult to reach, if more
realistic models had been introduced from the beginning. A certain criticism of
this simplification of the distribution functions defining the risk process has been

Skand. AktuarTidskr. 1968

6 Carl Philipson

given by Almer [3-7]. Among other critiques the works by Giovanni and Giuseppe
Ottaviani, Campagne, Tedeschi and de Finetti may be mentioned in the first hand;
their criticism was mainly directed against the criterion for the decisions by an in-
surance company especially with regard to the reinsurance policy. Also Borch criticised
this criterion, and suggested new methods for the formulation of the decision problems
(see § 10 here below). With respect to the distribution functions of the risk process,
numerous papers have been based on more realistic assumptions than those leading
to the classical form (see § 5 here below). In a few cases such assumptions have been
applied to the ruin theory and to the decision criterion of this theory. As is seen above,
339 items in the reference list deal with the risk theory, and all these items are, more
or less, based on the fundamental ideas introduced by Filip Lundberg, and later
developed by him and by his followers. This statement holds even for the critiques of
his theory, also for the papers by Borch, as in these papers the fundamental ideas have
been accepted, though some parts of the theory have been modified. It might here
be remarked that Borch has in some of his papers used distribution functions of the
total claim cost, which have been based on an even less realistic model, than the
classical form of the risk theory. In the opinion of the reviewer the discussion of this
paragraph affords a strong argumentation for the statement, that Filip Lundberg's
contributions have to a very wide extent stimulated the continued study of the risk
theory, and that this has led to numerous valuable contributions to the problems
within the scope of this theory.

3. Earlier reviews of the risk theory and the list of literature in

Part H of this review
Besides Cram~r's survey in [111], which has been referred to in the previous section,
a reference list was published by Ammeter in 1956 [15], and a survey of the risk
theory and other problems within non-life insurance in 1959 by Ammeter, Depoid &
de Finetti [19]. General surveys of the risk theory were given by Wilhelmsen in 1955
[358], by Segerdahl in 1959 [323], by Philipson in 1961 [281], by Thyrion in 1963,
1965, 1967 [349, 351-352], and, in 1967, by Bi.ihlmann [82], and by Kupper [210]. Fur-
ther, a book on the risk theory, not included in the list of literature, is under prepa-
ration by Beard, Pentik/iinen & Pesonen.
In addition to these reviews the Astin Bulletin for the whole time of its existence,
and Skandinavisk Aktuarietidskrift from the year 1961 inclusive have been consulted
for all contributions to the risk theory published in these journals during the periods
mentioned. Further, the reference lists with items from other journals published in the
Scandinavian journal in the same period have been read through, as well as reference
lists published by different authors. The reviewer has endeavoured to render a list of
literature as complete as possible; he regrets, if, notwithstanding his endeavours,
imoortant papers should have been neglected.

Skand. AktuarTidskr. 1968

 Review o f the collective theory o f risk. 1 7

4. Fundamental assumptions for the risk process

The deduction of a Poisson process is, generally, based on three assumptions, namely
(i) homogeneity with respect to the parameter, (ii) homogeneity in space, and (iii)
rarity of multiple events. (i) is often formulated stationarity of the increments, (ii)
independency of the increments, and (iii) the probability of the occurrence of more
than one event in an infinitesimal domain of the parametric space, dr, is of lower
order than the order of dr J i l l , 114]. Other formulations of (iii) are found in [2,
191, 307, 310], published by R6nyi et alia.
It shall be remarked here, that there exist cases, for which (ii) and (iii) hold, while
(i) does not hold with respect to r. This implies that pn(r) is equal to p(r) independently
r m

of n, in this case (I g) reduces to 7'~ ~oP(U)du, so that the probability of one claim
in an interval of length dt is equal to yldt, and the process with the transformed para-
meter t fulfils (i) with respect to t. Then, the process before the transformation may be
called a Poisson process, heterogeneous with respect to z. Such processes have been
included in the definition of the classical form of the risk theory [l I 1].
Biihlmann [82], seems to have anticipated a theorem (according to a letter from
Buhlmann to the reviewer, it should be proved in [83]), which should imply, that (iii)
is a consequence of (i), (ii) and of the properties of the sample functions in the re-
stricted space; (iii) should, thus, according to Btihlmann, not be necessarily included
in the conditions for the process being a Poisson process. In [ll4] Cram6r gives an
example of a case, where (iii) is not fulfilled which leads to a cPp i.n.s, with a risk
distribution of the discontinuous type. In this case, however, (ii) is not strictly ful-
filled. Cram6r refers in [114] to a case, where the probability distribution of the length
of the mutually independent time intervals between consecutive discontinuity points
is given in a general form, which for the Poisson process is exponential. This process
has been called a "process of limited after effects" and has been introduced by
C. Palm. According to Goldmann (Ann. Math. Soc. 38, 3, 1967), there exist processes
with Poisson-distributed number of events, for which (ii) does not hold, so that they
are not Poisson processes (cf. also [89, 91,201,313, 329, 341]).
According to Bi.ihlmann, who has published his thesis [77] on exchangeable vari-
ables, a theorem is given by de Finetti for such variables (by de Finetti called humeri
aleatori equivalenti [154], cf. also [172]), which should lead to the following fundamen-
tal assumption for general cPp i.n.s. (instead of (i) and (ii)). For an arbitrary number
of non-overlapping parameter intervals of the same length the amount of the claims
occurring in each of the intervals can be arbitrarily exchanged without change in the
probability distributions of the process. It seems likely, that it should be possible to
find a similar condition leading to a cPp i.w.s. Cram6r remarks in [114] that, if (i)
is given up, the theory of so-called harmonizable processes defined by a spectrum
distribution with correlated increments may lead to better understanding of the risk
process, in this case with non-stationary increments. As the Polya process can be
deducted both from the Polya-Eggenberger urn scheme [230], and from the Lexis
urn scheme [9, 230] such a process may be the consequence of heterogeneity either
in space or in time or in both space and time. An analysis of the effects of these
different types of heterogeneity has been given in [230].
Skand. AktuarTidskr. 1968
8 Carl Philipson

R6nyi et alia, quoted here above, have introduced the concept composed Poisson
processes, which fulfil the condition (ii), and are, therefore, principally different from
the compound Poisson processes, which, with exception of the Poisson process, have
dependent increments. The composed Poisson processes have been discussed in [289,
301] by the reviewer. The characteristic functions defining a cPp can be transformed
into a form similar to the form of these functions for a composed Poisson process.

5. Extensions of the classical form of the risk process

The claim distribution of the classical form is, by definition, independent of t. Very
often, however, the claim distribution actually varies with t, as proved by extensive
statistics. Esscher extended the Poisson process with regard to a claim distribution,
dependent on the parameter point for the occurrence of the claim, in 1932 [145].
This was extended to a Polya process in 1957 [272], and to a general cPp in 1965
[299]. The last-mentioned result was reached independently of a theorem published
by Jung in 1963, [195], according to which a symmetric function of the increments of a
random function attached to a cPp i.n.s, with t-dependent claim distribution is
distributed with a distribution function of similar form, as that obtained in [299],
and given in (le) here above.
According to [19], Dubois was the first author, who accounted for a dependency
between the events in non-overlapping time intervals, for his calculation of the
variance of F(x, t), in a paper published in 1936 [139] (cf. also [142]). In [19] the
compound Poisson distribution with t-independent risk distribution was said to have
been used by Ammeter with reference to his paper in 1949 [10]. No reference is made
in [19] to [230], which in the first edition was published in 1940 by Ove Lundberg.
[230] contained a systematic study of the cPp i.n.s, with particular consideration of
the Polya process with references to [305, 144, 176, 250, 203]. Similar models were
introduced by Ammeter in 1948 [9], independently of Ove Lundberg, and generalized
by him in 1949 [10]. In 1954 [13] Ammeter gave Lundberg credit for his priority.
Both Ore Lundberg and Ammeter [230, 9] deducted the limiting distribution, as
t tends to infinity, for the variable x/t, i f x is distributed in a compound Poisson distri-
bution with t-independent risk distribution U(v), and found this limiting distribution
to be in the form of U(v), provided that cl • 0. Lundberg has also for this case given
an asymptotic expansion of F(x, t) for a Polya process in terms of U(v), and for cl = 0
in an Edgeworth series for F(x, t) of a cPp i.n.s. Ammeter proved that, if, for a
Polya model, ty~/~'~ remains finite, when t tends to infinity, the limiting distribution
is in the form of the normal distribution function, and he has also expanded F(x, t)
in this case in terms of the normal distribution function after the transformation
according to Esscher (see § 7 here below). The reviewer [293, 303] proved that the
intensity function of a cPp i.w.s, is in principle proportional to the volume of the
population, as the intensity function of a main heterogeneous group is equal to the
weighted sum of the intensity functions of homogeneous sub-groups. Consequently,
an increase in t may by (1 g) be due either to an increase in the volume of the popula-
Skand. AktuarTidskr. 1968
 Review of the collective theory o f risk. I 9

tion or to an increase in • (or both). The limiting distribution, when t tends to infinity
only due to the increase in volume, is in the form of the normal distribution, and, when
the increase of t is only due to an increase in r, in the form of U(v, s). The condition
in the former case can be replaced by a condition of boundedness for the functions
t~-ay~(s)/~(s), which for the Polya process reduces to Ammeter's condition for this
limit passage. An asymptotic expansion of V0~; t, s) in an Edgeworth series has been
given for this case [293].
Arfwedson [35] extended the Poisson process by the omission of (iii) in § 4, and
found, that the extension rendered the same result at the end of time-intervals of
finite length as Ammeter's model in the case, where tydy~ is bounded even for infinite
values of t. It has been proved [271], that this model can be interpreted as a transform
of either a sequence of Polya processes or of Poisson processes defined only for
discrete parameter points.
Hofmann [185] introduced a wide subset of cPp i.n.s, by defining Po(t) as the
solution of the following differential equation, where k, q and a are constants k >0,
q > 0 , a>~0.

ky" +q(l + t/k)-ay = 0. (5a)

The present author introduced the extended Hofinann processes by defining Po(t)
as a product of the solutions of equations in the form of (5a) with, not necessarily,
different values of k, q, a [280, 282, 290, 297].
The study of the cPp i.n.s. [278-279] led the present author to the introduction
of the cPp i.w.s., as defined in § 1 [284, 290-291, 293, 296]. A wide sub-set of these
processes was introduced in [303, 304] under the name of cPp of the ordcr v,
v = I, 2..... (cPp: ~) which will be defined below. Pesonen and Jung have discusscd
the cPp i.w.s, in recent manuscripts to the Lundberg symposium.
Some of the processes exemplified by Bartlett [40], and the processes studied by
Matern [238] are cPp i.w.s. This can also be said of Ammeter's model with bounded
ty2/yx~. Thyrion introduced [345, 347-348] a very wide class of distributions, the
distributions in bunches (par grappes), and in bunches of bunches (pat" grappes de
grappes) defined by characteristic functions in the following general form.

~,~) = ~ {~b,[~ . . . . (q)~(,1))] } (v I> 2), (5 b)

where ffj(z) are generating functions of integer valued variables, j = 1, 2, ..., v - I.

Originally ~P~(O was the characteristic function of such a variable, but may also be
any characteristic function for v > 2. If particularly, 'P~01) is the characteristic function
of a r-dependent Poisson variable, or of a generalized such variable, (5b) can be
considered the characteristic functions defining a cPp, which, if at least one of the
ffj(z) depends on t, is a general cPp i.w.s. If in (5 b) all ~j(z), j = I, 2..... ~, - 1 are in
the form jP[tj(l -z)], and, if ~(~1) defines a cPp i.n.s., (5b) defines a cPp: v. This
implies, that X(t) is a generalized variable even for an elementary process (cf. the
remark to the terminology in § I here before).
Extensions to processes with parameters of more than one dimension have been
dealt with in [40, 238, 270, 273, 298, 342].
Skand. A ktuarTidskr. 1958
10 Carl Philipson

In [299] the extended risk process was introduced, taken to mean a process, where
the occurrence of the accidents and the extent of the damage caused by them, as well
as the development of the actual payments for a claim during the period, when it is
outstanding, is accounted for. For the deduction the theory of cPp i.w.s, was used
(cf. also [300]).
Almer introduced [3-7] a very general model for the risk process. His fundamental
assumption can be formulated by saying that "behind" the risk process, there exists
another process constituted by a large, but finite number of risk situations, called
risk elements. Each risk element is supposed to be associated with a certain probability
of inducing a claim, and a certain claim distribution. In [275] some of the deductions
were based on this model. In [298] Almer's model was modified by the present author,
by the assumption that the occurrence of a risk element was associated with a change
in a random function of a two-dimensional parameter (time and geographical space).
This random function was, further, assumed to be subject also to changes caused by
changes in environmental conditions, and the occurrence of an accident was supposed
to be correlated with the random function just defined. The extent of the damage of
one accident could be correlated either with the same, or with a similar random
function. Also this theory could be interpreted in terms of the cPp i.w.s.

6. Particular forms of the claim distribution

Cram6r used in [I 11] for the numerical comparison of different approximations for
F(x, t) in one example an exponential distribution, and in another the form
k~ e -~1~ + k2(x + b) -~' for x < 500, and equal to zero for x > 500, as derived by Esscher
from Swedish non-industry fire experience 1948-1951. For the deduction of the ruin
function (see § 9 here below) a form indicated by T~icklind [354], defined in the
following sentence, was used in [111]. The claim distribution was in this case supposed
to be arbitrary in the negative semi-plane, with the restriction of having a finite,
absolute mean over this domain, and given by an exponential polynomial

~ln(l-e -anz) in the positive semi-plane.

7L, I

Almer, who mainly dealt with non-life insurance, for which V(0)-0, proved for
non-negative values of x, that an upper and lower approximation in the form of
exponential polynomials can, with any desired precision, be found for any distribution
function [3]. He, further graduated extensive statistics for different time periods mainly
from Swedish motor insurance, with exponential polynomials containing three terms
for certain periods, and four terms for other periods of time, giving deviations of at
most one to two per cent. Almer used this form for the deduction of approximation
formulae for F(x, t) (see § 7), this was also done by Hovinen [186-188], and Pesonen
[268] both for such deductions and for their numerical investigations. Bohman and
Esscher [67] modified this form, by replacing, for higher values of x, the exponential
polynomial with actually found frequencies in discrete points spread out over small
Skand. AktuarTidskr. 1968
 Review of the collective theory of risk. I 11

intervals about such points. This modified form was numerically compared with
Swedish experience in life insurance 1957-1961, in third party liability motor insurance
1957, and in fire insurance 1948-1957, differentiated with regard to industry and non-
industry. The agreement between graduated and actual values was very satisfactory.
Benckert [48] studied the application of a log-normal distribution to the claim
distribution. A more systematic treatment of the effect of different forms of V(x)
on the risk process with particular regard to excess of aggregate loss reinsurance was
given by Benktander & Segerdahl [50], and by Benktander [51], cf. also [170, 182].
Finally, Thyrion [350] introduced a general class of functions, the compound exponen-
tial functions, which is a particular case of (7h) of the next section, with I.'~(x) =
1 - e -¢<x-c). A great part of the forms for V(x) used by other authors belong to the
class defined by Thyrion. The exponential polynomials are compound exponentials,
with c =0 and ~ integer valued. If V'g(x)= 1 - e -¢Cx-c)la, and the distribution function
of ~ defined by an incomplete F-function, V(x) used in [236] is obtained, if c =0,
and the Pareto distribution analysed in [50], if c ~ 0 . Thyrion also proved that this
class also contained the functions IV(x)]a, and V[g(x)] for g(0)=0, g ( o o ) = l ;
( - 1)ngC'~(x)-~0, where V(x)is a compound exponential. A particular form is ob-
tained by taking g(x)=k~x k', so that V(x) has the m o m e n t s / ~ = kTr/k'F(r/k~ + 1),
which, eventually, can be used in life insurance technics.

7. Transforms of approximations for F(x, t) and F(x; t, s)

Ove Lundberg proved that an elementary Polya process being heterogeneous with
respect to z and t, is homogeneous with respect to a transformed parameter equal
to - I o g Po(t) (230). Ammeter [9] transformed F(x, t) for a non-elementary Polya
process, by using this transform of the parameter, and further, a transform of V(x).
A similar transform was used by Campagne [91]. These transforms are particular
cases of the transformation of F(x, t) for cPp i.n.s, given by Thyrion [345, 347-348].
The characteristic function corresponding to the transform is in the form of the
characteristic function of a generalized Poisson variable (even for the transform of an
elementary process) by a generalizing variable with the probability distribution
qn(t) = [1/( - 0(t))]. [( - t)n/n!] 0(~)(t), where O(t) ~ log Po(t), in the elementary case. It
is proved that qn(t) can be obtained by a truncation of a distribution defined by (i d)
with O(v, z) - U(v), independently of r, if O(t) is bounded. The characteristic function
corresponding to the transform can be written in the form

and equal to ~(~), in the elementary and non-elementary case respectively. (7 a)

(7a) has been extended [297] to include i.a. non-elementary cPp i.w.s., where U(v, ~)
defines a generalized extended Hofmann-process (see § 5 here above). Thyrion
gave in the papers quoted numerous examples of this and other transforms. One of
Skand. AktuarTidskr. 1968
12 Carl Philipson

these transforms was obtained by the substitution of the distribution function of a

generalized Poisson variable with the mean of the number of events given by
-limtoooO(t), supposed to be finite, for U(v) of a cPp i.n.s.
Esscher introduced in 1932 a transform of F(x, t) for a Poisson process [145],
which was in 1963 extended by him to any distribution function F(x), for which the
corresponding characteristic function ~(n/) is known [146]. Let the mean of F(x)
be denoted tq, and its transform, here called the Esscher transform, be denoted
F(x), and the distribution function of the standardized variable, z =(x-l.,t~)/l/~,
F0(z), here g.k is the kth semi-invariant of F(x). h is supposed to be a quantity,
- H~ < h < H2, such that the integral representing ~ ( - ih) converges. The transform is,
then, defined by the following relation

dF(x) = e hx dF(x) /qD(- ih) (7 b)

and, if h is the single root in the interval - H t <h <H~ of the equation

Oq~(- i h )
x = bh oh ' (7 c)

the following relations hold.

IJ(-oo, 0; x) for x<tq,

F(x) = (7d)
I l - J ( O , +oo; x) for x > l q ,

where J(u,, u2; x ) = e - ~ ~o(- ill) f u=e - ~h ~/~ d F 0(u).

t/ U l

By a suitable choice of Fo(z) (7d) can be used for the approximation of F(x).
Before the publication of [146] Esscher prepared a manuscript (not published),
which dealt with the particular case, where F(x) had the form of F(x, t) of a cPp i.n.s.
The transform, F(x, t) say, could be written in the same form as F(x, t) with the sub-
stitution of the following expressions for t, U(v), V(x) respectively.

~ f ° e-U(t-° dU(u) f ~ enUdV(u)

t = t J o enUdV(u); U(v) = fo
r ' ~ _e-U(t-t)
_ _dU(u)
_ ; V ( x ) = f 0 ehUdV(u)
oo '
(7e)

The limits of F0(z, t) for z and t tending to infinity were deducted in the general
case in [65] and [292] respectively, and the limit, when t tends to infinity, for a Polya
process in [67]. With regard to the limits obtained, Fo(z, t) was chosen in the form of
a normal distribution function, and of an incomplete F-function in the application of
(7d) for the approximation of F(x, t) defining a Poisson and a Polya process [67].
Esscher's method of approximation was modified by Pesonen [266] in such a way,
that the solution of (7c) should be independent of ~1, and determined only by x for
an actual case.
Skand. AktuarTidskr. 1968
 Review of the collective theory of risk. I 13

Bohman introduced another method [63, 67], the C-method, for the approximation
of a distribution function F(x) corresponding to a given characteristic function
~001). Let, for v = 1, 2, z,('l)= COl)+(- 1)"0.42iC'01), where C01)=0 for I~1 ~1 and
equal to (l-I,11) cos m l + ( l / n ) sin [zall for Inl <1, and qo,(~)=Z~(~/T)~(~).
Then, ~,('/) correspond to the "improper" distribution functions F,(x), taken to mean
that dF,(x) are, not necessarily, non-negative for all x, and that the integrals .[_+~dF,(x)
are, not necessarily, equal to unity. It has been proved, that the following inequalities
and "conversion"-formulae hold.

F I (x) .<F(x) < F 2(x), (7 f)

F , (x) = ½ -
f _r e-~Uz ¢.p. (u) du 6' - 1, 2), (7 g)
r 2~ iu

which may be evaluated by using numerical integration according to Simpson's

rule. An approximation of F(x, t) is, then, given by the arithmetical mean of (7g)
for v = 1, 2, and the approximation error implied, by the sum of half the difference
of (7g) for v = 2 and 1, and the error involved in the numerical integration. In [67]
F(x, t) of a Poisson, and of a Polya process with the claim distributions refferred to
in § 6 here above were evaluated for different values of t and of x according to
the C-method and according to other methods, the Esscher method inclusive. The
results from the other methods were compared with those from the C-method with
due regard to the estimated errors in the last-mentioned results. Further, according
to similar formulae the corresponding stop loss risk premiums were evaluated by
different methods and compared. In [147] Esscher gave numerical dlustrations for
other cPp i.n.s.
A transform of F(x, t) with the claim distribution defined by the following relation
(7h) shall be given for the general case in the form of (7 i) below [304]. This transform
may be called the com,olution transform.

V(x, t) = f Vg (x, t) dG(~), (7h)

where the integral shall be taken over the range of ¢, ~ being a random variable
distributed with a distribution function G(~) of the continuous, discontinuous or the
mixed type. Let the distribution functions defining Poisson processes with V(x, t),
V~(x, t) as claim distributions be designated by ~F(x, t), ,F~(x, t) respectively, then,
the convolution transform olaF(x, t) can be written l-l~)eFg(x, triG(i)), where 11~¢) has
been defined in § 1, for # integer valued ]J~'¢) reduces to the convolution of a number
of distribution functions. From this relation, the following expression for the convo-
lution transform ofF(x; t, s) for a cPp i.w.s, with V(x) defined by (7h), is immediately
obtained.

F(x; t, s) = I-I* eft (x, vt dG(~)) d v U(v, s). (7i)

k(¢)

Skand. AktuarTidskr. 1968

14 Carl Philipson

The results in (le), (If) are particular cases of (7h). In the case, where V(x)=
Vl(x) = 1 - e -~, G(O =e(# - I), Cram6r has given an exact expression for ~F(x, t) in
the form of a Bessel function [111]. Esscher deducted in [145] a relation, which leads
to the convolution transform of ~F(x, t) in the particular case, where V(x) is defined
by (7h) with ~ integer valued. Almer [5-6] used this transform for the case where
V~(x) are different exponential functions for different integer values of E, and inserted
thereinafter, Bessel functions according to Cram6r for ~Fg(x, t). He suggested, then,
that these Bessel functions should be approximated by a few terms of their expansions
according to Hankel. The calculation of the convolution of these approximate
expressions can be performed without material computation work. Pesonen [268]
derived approximation formulae for ~F(x, t) with V(x) in the form of exponential
polynomials in a similar way as AImer. Bohman and Esscher discussed a convolution
transform o f f ( x , t), where V(x)was given as the sum of two exponential terms, and
one term defined by the unity distribution e(x - a), which, evidently, is a particular case
of (7h). As an indication for further work, it shall here be remarked that the approxi-
mation methods introduced by Almer and Pesonen might probably be extended by
using (7i) to F(x; t, s) o f a cPp i.w.s, with V(x) defined by (7h).
Pesonen [265, 268], and Hovinen [186, 188] used inter alia a Monte Carlo method
for the approximation of ~F(x, t) with V(x) in a given form. This implies the
simulation of a random sample of x for given values of t, where x is distributed
with the distribution function F(x, t); the estimate of the error involved in the
approximation can also be calculated. Numerical calculations were made ac-
cording to this method, and to other methods, and the results of the latter methods
were compared with those obtained by the Monte Carlo method with due regard to
the approximation error of the Monte Carlo method. A systematic description of the
investigations made by the Finnish school will be found in the book on the risk theory
under preparation, quoted in § 3 here above.
Approximations to F(x, t) with the claim distributions referred to in § 6 here
above were calculated by Cram6r [1 i 1] using the Esscher method, and an Edgeworth
series. The papers [8-13, 189, 199, 218, 260, 262] deal also with numerical illustrations
of approximation methods. In [291] the reviewer derived an expansion of F(x, t)
for a cPp i.w.s., and, particularly, for a Poisson, and for a Polya process, where
V(x) was given in the form of (7 h) with V¢(x) being exponentials for certain values of
~, and in the form e ( x - a ) for other values of L The expansions were intended for
direct computation of a sufficient number of terms in an electronic computer. So far,
the program for the calculation has been considered too complicated for practical use.

8. Applications of F(x, t) to rating problems, and to other problems

If the risk process is considered for smaller groups of insurances, which to a certain
extent are homogeneous, F(x, t) for each such group gives valuable information as
regards the rating a priori. The rating involves the estimation of the risk premiums
from the statistics of such groups, and the decision to which extent the division into
Skand. AktuarTidskr. 1968
 Review of the collective theory of risk. l 15

groups in the statistics shall be kept also in the tarif. As, however, it must as a rule be
assumed, that the claim frequency and the empirical claim distribution of each sub-
group depend on time, the risk premium of each group depends on time and must be
predicted for the period during which it shall be applied. The principles for the appli-
cation of statistical results to practice have been expressed by Wold in 6.4 of [359].
The risk premiums for life insurance are dependent on age attained and on calendar
time. So far, it seems, that only the classical form of the risk theory has been applied
to life insurance. Recent investigations (e.g.T. Larsson, Mortality in Sweden, Stock-
holm and New York, 1965) have, however, led to the conclusion, that the mortality
intensities of non-overlapping time intervals often are mutually dependent. Further,
the claim distributionsIhere called the distributions of the risk sums, not to be
confounded with the risk distribution defined in § 1, (O(v, T) of (1 d))--depend as a
rule on time as being subject to variation with changes in the economical and social
conditions. Therefore,--in the opinion of the reviewerisimilar view-points shall be
applied to the risk process of life insurance as those discussed in numerous papers for
non-life insurance. The Swedish table of premiums for life insurance has also been
based on a predicted mortality.
It should, thus, be allowed for the variation of both U(v, s) and V(x, t) with the
parameters. It follows, that the risk premium, upon which the tarif rates are based,
will as a rule differ from the risk premium for a later tarif applied to the same group.
The risk premiums used in the tarif, which may be called applied risk premiums,
constitutes, therefore, random processes with discontinuous time parameters, defined
by sample functions in the form of step functions with discontinuity points at each
change of the tarif rates. If also the security loading in the premiums is based on a
prediction (of some measure of the variation of the risk premium) such a loading as
applied in the tarif, is attached to a similar process.
Statistical aspects have been considered by Beard [41], and with particular regard
to mortahty by the same author [44]. Large claims were separately treated by Beard
[43], Depoid & Duchez [129], Franckx [168] and Gumbel [179]. Distribution functions
of the sum of claims, the largest claim excluded, were given in [26, 168], and with the
exclusion of the r largest claims in [27]. Almer (3) introduced two particular statistical
methods. One of these implies a separate calculation of the risk measures: the risk
premium, the claim frequency and the claim distribution for three different claim
groups according to size. This method was called excess claims analysis. The other
enables the estimation of the separate effects on the risk measures of the components
in a parameter vector, the method was called factor analysis (cf. [275, 287]). I f F ( x , t )
refers to an insurance without a clause for self-retention (deductable), the risk premium
for the corresponding insurance with such a clause, can be calculated by the formulae
J'~>s~ (x -s)d~F(x, t), where s is the size of the self-retention, and, ifF(x, t) relates
to the total of the policies of a certain line underwritten by the company, ceded or not
ceded, the risk premium for an excess over s of aggregate loss reinsurance can be
calculated by the same formula [1 I, 274, 276, 283].
Besides the rating a priori, it has been customary in certain branches to account
for the actual experience aposteriori, either by experience rating or by the distribution

Skand. AktuarTidskr. 1968

16 Carl Philipson

o f dividends. A method, particularly used in motor insurance, is the system of no claim

bonus, often combined with some penalty (malus) for a large number of claims [20,
25, 54, 79, 128, 132, 165-167, 171, 178, 181,244-245, 248, 277, 288, 314, 344, 346].
The technical reserves have been considered i.a. in [46, 86, 217, 247, 263, 299, 337].
More or less general applications to different problems in non-life insurance are found
in [1, 13-14, 16, 28, 37, 87, 130, 142-143, 196, 214, 232, 243-244, 270, 273, 278, 287,
312, 314, 325]. In the list of literature some applications of similar models, as those
used for the risk process, to fields outside insurance have been included. [175, 235-
236, 285] deal with the recovery of the human eye after dazzling and [238, 342] with
ecological problems. [220, 300] deal with computer failures, in [220] a branching
Poisson process, and in [300] a branching cPp were used as models. In [300] the rela-
tions of these models to cPp i.w.s., and to the model used in [299] for the extended
risk process (see § 5 here above)_were established.
Problems involved in experience rating have lately been subject to a large interest.
Ammeter treated this problem for the risk process in its classical form [17, 21, 24].
Btihlmann gave at the Astin Colloquium in Lucerne a distribution-free method for
experience rating [81], he proved that the best estimate, in a certain sense, of the risk
premium a posteriori is a conditional mean. He stated, further, that the credibility
method of estimation, implying the weighting of the results of the actual experience
and the results obtained by other experience, e.g. from earlier statistics, could be
explained in terms of the conditional mean. The practice of experience rating used
by American investigators 1 (e.g. [205]) is based on a particular case of Btihlmann's
method. Delaporte proposed in [122-127] the use of the conditional mean as defined
by (I h), on the assumption of a risk process in the form of a modified Polya process,
for experience rating in motor insurance instead of a bonus-malus system. Ove
Lundberg developed [231] a theory for experience rating based on a general cPp i.n.s.,
the conditional mean was also determined according to (1 h) [230]. Derron considered
[I 33] the betterment of credibility by the exclusion of the largest claim, and used for
his deductions [26, 168]. Also Bichsel [55-56] treated at the Colloquium referred to
the experience rating based on the theory of cPp i.n.s., and allowed to a certain extent
for the random variation of the risk with time. The present author stressed--in the
oral discussion at the Colloquium--the necessity of allowing also for a systematic
variation with time. In [302] the theory was, therefore, extended to include cPp
i.w.s., and the relation between the theory based on such processes, and Bi.ihlmann's
general theory was established. Also the connections with modern Bayesian theory
(e.g. Robbins, Rev. de I'Inst. Int. Star. 31, 1963) and with the general decision theory
were discussed in [302]. F o r the estimation of the parameters appearing in the
conditional mean it was in [302] referred to Anscombe (Biometrika, 1950), and to
Grenander [177]. Other papers in the list of literature of this review dealing with
estimations are, i.a., [216, 274--276, 283].
Bohman discussed the experience rating for a company, which aims to increase
the volume of its business [68].

For literature the reader is referred to L. H. Longley-Cook, An Introduction to Credibility

Theory, Casualty Act. Soc., New York, 1962.
Skand. AktuarTidskr. 1968
 Review o f the collective theory o f risk. I 17

9. T h e ruin functions

An approach to a model for the gain of an insurance company, accumulated during

a period of time from zero to t on the transformed scale (see § 1), can be based on the
difference Y ( t ) - Z ( t ) , where Y(t) and Z(t) represent the accumulated revenue and
cost respectively. If revenue and cost due to other items than those concerning the pure
risk business are neglected, the gain of this business on the own account of the company
is given by such a difference; a modified such difference by neglecting also the rein-
surers' payments for claims and the cost of the reinsurance, gives the gross gain of the
pure risk business, for certain problems also the gain of the reinsurance may be of
interest. Finally, it is for each problem to be decided whether the payment of dividends
and such alike shall be disregarded; also the interest on revenue and cost may or may
not be considered. These different definitions of the difference, defining the gain, will
here be called modifications. In all the modifications, we have to deal with random
functions attached to stochastic processes. That also Y(t) in a very general approach
must, in all modifications, be considered a random function is a direct consequence
of what has been said in the first two paragraphs of the previous section. Tarifs for
the direct insurance are often subject to amendment at intervals of a few years, this
does, even afortiori, apply to the reinsurance premiums. The applied, continuous
risk premiums and their security loading cannot, therefore, be the same for longer
periods. There are also other causes, than those considered in § 8, for the variation
with time of the continuous premiums collected. Firstly, the market conditions, in-
cluding both the competition between the insurers and between the reinsurers, and
the different interests of a cedent and his reinsurers with respect to the reinsurance
premiums, and other conditions for the reinsurance, shall be observed. As far as the
reinsurance cost is concerned, the reader is referred to the remarks to this effect given
in [287]. Secondly, the premiums may be subject to changes due to the provisions
by law or by the authorities. If the risk reserve at the time point t, Q(t) say, is defined
as the sum of the initial risk reserve, Q(0) = u say, and the accumulated gain, it fulfils
the relatlon

Q(t) = u + Y(t) - Z(t), (9a)

which can be modified as was pointed out here above. As far as the present author
knows, our knowledge of the process constituted by Y(t) is very restricted, at any
rate insufficient for the use of this general model.
At the present stage of our knowledge the ruin theory must be based on very restric-
tive assumptions. Such assumptions are for example, that the continuous risk premium
and the continuous security loading may be considered constant for the whole length
of the period considered, equal to ct the mean of the claim distribution assumed to
be independent of time, and c 12 the continuous security loading subject to the same
condition. This is consistent with the classical form of the risk theory. For a cPp i.n.s.
with t-independent claim distribution c~ is, it is true, a constant, but the continuous
security loading, if based on the standard deviation of the accumulated claims, depend-
ent on time. It has been found that the ruin theory, so far, as a rule, based on these
restrictive assumptions, has in spite of this simplification entailed many difficulties.

2 - 693822 Skand. AktuarTtdskr. 1968

!8 Carl Philipson

In some cases, the theory has been extended by assuming 2 to be a function of Q(t)
at least for a part of the future. It is easily seen, however, that in a very realistic model
even the relation between the security loading and the risk reserve may be changed,
so that the extension does not completely eliminate the restriction on 2. As an example,
the security loading for the Swedish third party liability motor insurance, being
compulsory, and, therefore, strictly controlled by the authorities, was up to 1955
5% and after 1955 3% of the tarif premium; these loadings were determined by the
Registrar General, and applicable to all companies regardless of their risk reserves.
The following context is divided into three parts A, B and C, where A and B deal
with the development up to the publication of [111], [111] inclusive. The context of
A and B is a review of a summary of this development given in [111], A refers to the
theory based on a constant 2, and B to the theory based on a security loading being
a function of Q(t). C refers to the development after the time considered in A and B.

A. In this case the risk reserve is defined by the following relation

Q(t) = u + (cl + Z) t - X(t), (9 b)

where cl and 2 are assumed to be constants, it essentially positive, while cl may be

positive or negative. Q(t) is attached to a random process, which is a transform the
risk process, as treated in the previous sections, and X(t) the accumulated claim cost
for a risk process in its classical form, in one modification cl, 2, X(t) refers to the
business on the own account o f the company. If, at some time t, Q(t) becomes negative,
it is said that ruin occurs at t. This is equivalent with the following definitions of the
occurrence of the ruin at t, namely, that ruin is said to occur at t, if a sample function
o f Q(t) crosses the horisontal axis, or, if a sample function o f X(t) crosses the line
x = - u - (cl + 2)t, at the time point t. The ruin functions are defined as probabilities
o f the occurrence of ruin at somepoint t fulfilling certain conditions. In [111] Cram6r
rigorously defined these probabilities, [110], by proving that the events concerned
here, have well-defined probabilities in a discussion similar to that reviewed in § 2
here above, according to which X(t) of the restricted space are associated with prob-
abilities induced from a reference space of an enumerable set of dimensions [111].
This p r o o f was given for the mixed case taken to mean the case, where 0 < V(0) < 1,
the cases, where V(0) = 0 and V(0) = 1, are referred to as the positive case and the nega-
tive case respectively. F o r the definition of the ruin functions (9b) may be considered
as a function either of the continuous parameter t, or of the discontinuous parameter
t =rh, where h is a given positive quantity (e.g. the length of a business period),
and r ~ 1, 2, .... The symbol generally used for the ruin functions, % shall be used in
this section, and be reserved for this purpose (it shall not be confounded with W used
in the previous sections and defined in § 1 here above). The ruin functions, for the
case of t being continuous, are the probabilities for the occurrence of ruin at some
t >0, and at some t in the interval 0 < t < T, where T i s a given value, these functions
are denoted V0(u) and V0(u, T) respectively. The corresponding probabilities, for the
case of t being equal to rh, are denoted Wn(u), yon(u, T) respectively. Let further,
U ( s ) = l + ( c ~ + 2 ) s - f _ ~ e ~ dV(x), where s=tr+i~l, and the integral is the complex

Skand. AktuarTidskr. 1968

 Review o f the collective theory o f risk. I 19

Fourier transform of V(x), for a =0 being the characteristic function corresponding

to V(x).
The ruin functions ~p(u) and Vn(u) were introduced by Filip Lundberg in 1926--1928
[226], and, further, developed by him in 1932, 1934 [228-229]. He obtained for each
of these functions in the positive case an inequality, and an asymptotic relation for
large values of u, given in the following lines.

0 < Vh(u) < r(u) < e -nu, (9c)

~o(u) ~ Ce-nU; Wn(u) ~ Cne -au. (9d)

Apart from Cn being dependent on h, R, C and Ca are constants depending only

on 2 and V(x). In anticipation of the results obtained later, (9c) holds also for the
mixed case, and, for the negative case, if cx +2 <0, where in the last inequality the
sign of equality holds. (In [306] Prabhu has proved the last assertion for a general
additive process with stationary increments.) Under an additional assumption
(see below) (9d) holds also in the mixed case.
Cram6r proved in his papers of 1926 and 1930 [99, 101] that ~(u) in the positive case
is a solution of an integral equation of Volterra type, which can be solved by complex
Fourier transformation. He gave an explicit expression for W(u) in this case, and a proof
of (9d).
In 1941 [105] Cram6r proved that VJ(u) for the mixed case satisfied an integral
equation, not of the Volterra type. Segerdahl gave for this case the first rigorous
proofs of (9c-d). He studied also the positive and the negative cases, and proved for
these cases a number of important results, some of which had, without complete
proofs, been stated by Filip Lundberg. TS.cklind [354] showed, that the integral
equation in the mixed case satisfied by ~,h(u) can be solved by a method due to Wiener
and Hopf [259] involving a complex Fourier transformation combined with argu-
ments from the theory of analytic functions. From this solution a proof of (9d)
for V)h(u) was obtained; he also proved that ~oh(u) tends to a definite limit, for which
an explicit expression was given, as h tends to zero. Cram6r [110] gave a probabllistic
definition for ~0(u), and applied the Wiener-Hopf method directly to the integral
equation for ~,(u) in the mixed case.
Certain preliminary results with respect to W(u, T), and ~on(u, T) were given by Filip
Lundberg [226] and by Segerdahl [317]. In [317] the moments of the time, when ruin
occurs for the first time were calculated. The problem was thoroughly investigated
by Sax6n [315-316], mainly for the negative case, and, by Arfwedson, for the positive
case. F o r these cases explicit expressions for V,(u, T) were given. Arfwedson obtained
also a number of results concerning the asymptotic properties of this function.
In [111] Cram6r rigorously proved, that the limits of ~on(u), Wn(u, T), as h tends to
zero, are equal to W(u), ~p(u, T) respectively. He, further, introduced the basic assump-
tions with respect to V(x) that the means of ix[ over the negative axis, and o f e °x over
the positive axis for some a >0 are finite. F o r the proof of (9d) in the mixed case
an additional assumption was made, implying that, for some a >R, the mean o f e ~*
over the positive axis is finite. The relations between R and certain other constants
2* - 693822 Skand. AktuarTidskr. 1968
20 Carl Philipson

with Yl(s) were analysed. R was defined as the least upper bound of q, subject to the
conditions that for 0 <~r .~q the complex Fourier transform of V(x) is analytic and
regular, and 17(o) > 0. For the mixed case the integral equations satisfied by the ruin
functions and by their complex Fourier transforms and by certain other functions,
are discussed by means of the Wiener-Hopf method [259]. One of these equations,
satisfied by the complex Fourier transform of V0(u, T), leads to explicit expressions
for W(u), W(u, T), and to certain results for the asymptotic properties of these functions.
Some results of such properties due to Segerdahl [317] were proved; also an inequality
for ~(u) -~(u, T)was deducted. In the positive case (9d) was proved, and the inequality
for the difference W(u)-~v(u, T) was strengthened; an asymptotic relation for the
difference for this case (stated without proof by Arfwedson, later proved by him)
was proved in [Ili].
In the mixed case the form of V(x) indicated by T/icklind (see § 6 here above)
the following expression for ~(u) was obtained in [111].

N
W(u) = Z C, e- ~"~, (9 e)
u l

where N = r, the number of terms in the exponential polynomial, if cl +2 >0, and

N=r + 1, ifct +2 <0, Cn is given by an expression analogous to the expression for C
in the equality for W(u) under more general conditions (5.10, [111]), and Rn are the
zeros of 17(s) in a rectangle formed by ~r+iT, Z + iT (H(s) is on the contour of this
rectangle positive for sufficiently great values of .E and T). In the positive case
Cn = 2][ - II'(R,)], (N = r), and, if, particularly, r - 1, C~ = 1/(I + 2), and R~ = 1 - C,.

B. In this case a never increasing function 2[Q(t)] is substituted in (9b) for 2, if Q(t) <a,
where a is equal to a finite constant, or to infinity. This problem was already treated
by Filip Lundberg in 1926-1928 [226], and, further, by Laurin [215], Tacklind [354]
and Davidson [121]. For the particular case, where V(0) =0, V(x) = 1 - e -~, and a is
given by a finite constant, Davidson gave the following relation for W(u), where
H(u) is defined by

H ' ( u ) ~ e x p { - fuJo~2(v) dr},

H(u) + H'(u)
~(u) ~ 1 (9 f)
i+2
H(a) + H'(a)
2

If in (9f), a =0, i.e. 2(Q(t)) is equal to 2 independently of Q(t), the assumptions in B,

reduce to the assumptions in the last paragraph of A, with r = 1. In fact, for a =0,
(9f), reduces to (9e) with

CI = 1/(1 +2), and R~ ~1 - C t.

Skand. AktuarTtdskr. 1968
 Review o f the collective theory o f risk. t 21

C. After the pubhcation of [111] Arfwedson published the second part of [35] with
numerous results in the ruin theory. Segerdahl published also new studies into this
theory particularly dealing with the time point at which ruin occurs for the first time
[321, 322]. This has also been treated by Prabhu [306], who used queuing theory in
his developments. Arfwedson has recently given some notes on [306], unpublished,
showing the relation between the proofs of [306] and [35]. In [323] Segerdahl gave,
for a great number of particular cases, explicit expressions for V,(u) and for V(u, T),
in one of these cases the interest accrued on Q(t) was accounted for; he referred also
to cases treated by Arfwedson [31, 33, 35], where Z was allowed to take zero or even
negative values. Segerdahl, further, derived an expression for V&(u) under the asump-
tions of one of Ammeter's models described in § 5 here above, including the assump-
tion that ty~/~ is bounded even for t ~ oo. Ammeter has, however, in [9] arrived to a
similar expression for VJ~(u) where this last-mentioned assumption seems not to have
been used. As far as it is known to the reviewer, no other deduction of the ruin
functions, based on other forms of the risk process than the classical form, have
been published so far. Almer [3] indicates, however, that the deduction of approxi-
mate expressions for w(u, T) should be possible under wider assumptions, if based
on his approximations of F(x, t). In 1966-1967 Segerdahl lectured on the risk theory
at Stockholm University; in these lectures [l l l] was reviewed with new proofs for
particular cases; one of the problems treated in the lectures was, further, studied by
Thorin [343]. Segerdahl discussed also in these lectures by the methods used in [l 1l]
a rigorous extension of the ruin theory to a Polya process, which will be published later.
An interesting contribution to the ruin theory for the classical form of the risk
process was published in 1966 by Beekman [47]. According to his developments
1-~(u, T) could for the mixed case, be determined by the conversion of a double
Laplace transform of the probabdity for the occurrence of the event Max ( - Q(t)) < ~,
by a conversion method described by Widder (The Laplace transform, Princeton
University Press, 1946). In the positive case, the probability mentioned, with c~ +2,
particularly, replaced by zero, is equal to F(~, t), which, thus, can be deducted either
by a limit passage of the said probability, or by the conversion of the corresponding
double Laplace transform. The theory has been illustrated by a few simple numerical
examples; the application to more realistic models shall be subject for future research.
Beekman's paper has been discussed by Thorin in a recent manuscript to the Lund-
berg Symposium.

10. Application of the ruin theory and of other theories to decision problems,
and references to studies into reinsurance problems

In a great part of the literature criteria for decision problems in insurance companies,
particularly for decisions related to the reinsurance policy, have been based on some
ruin funchon implying that a decision shall be chosen, which entails a reduction of
the ruin function to a fixed predetermined level. It seems evident, that it must be con-
sidered more realistic for this purpose to use ,p(u, T) than ~(u). In many cases such

Skand. AktuarTidskr. 1968

22 Carl Philipson

criteria are to be applied to other decision problems such as those regarding the
magnitude of the risk reserve and of the security loading. In fact, these decisions are
connected with the choice of reinsurance policy, and ought, therefore, to be simul-
taneously considered. Also the choice of a system for the distribution of dividends
is connected with the decisions, just mentioned. A very interesting application of the
ruin functions is the solvency control of insurance companies according to the Finnish
Act of Insurance, originally suggested and drawn up by Pentikainen. The eager
interest for suitable methods for the approximation of F(x, t) and VJ(u, T), shown
by the authors of the Finnish school, is a consequence of the legal provision just
mentioned [186-188, 199, 264, 265, 268, 269].
Such criteria, based solely on the ruin theory as reviewed in the previous section,
has been subject to criticism by two groups of critiques, referred to in § 2 of this
review. The first group can here be exemplified by Campagne [87], Campagne &
Driebergen [88], de Finetti [155-161], Giuseppe [254-257] and Giovanni Ottavian
[258] and Tedeschi [333-335]. One of the arguments given in some of these papers,
is that it seems unnatural, that the criterion of the previous paragraph becomes
gradually more and more severe (quoted from [19]). The papers of this group were
published from 1940 to 1957. In [158, 161] de Finetti suggested that the reduction oi
the ruin function to a fixed level should be combined with an auxiliary condition,
which implied a maximisation of future gains.
Even if it should become possible to give a realistic definition of the gains according
to the view-points on this problem given in the introduction of the previous section,
it is uncertain, whether a discussion based only on the gains will be found sufficient
in all cases. Business enterprises in general have very often other aims besides pure
profitableness; this seems to be particularly true for insurance companies (cf. e.g.
[68]). In the preference theory of economics tools for measuring the preference
have been given, which have been called utility functions. By the application of this
theory combined with the theory of games it is possible to account for different aims
of the company and for that part of the variation of the continuous premiums collected
which is connected with the competition in the markets and between the interests
of the cedent and his reinsurers, as referred to in the first paragraphs of § 9 here
above. The application of the theories just mentioned has been introduced by the
second group of critiques e.g. Borch [69-75], Kahn [197], Ohlin [253] and Wolff
[361]. These papers were published from 1962 to 1967. Particularly Borch's contri-
butions are to a wide extent based on the theories mentioned, as given by Neumann-
Morgenstern (Theory of Games and Economic Behaviour, Princeton, 1944). As was
mentioned in § 2 of this review, Borch has, however, in some papers used an un-
realistic model for the distribution of the claim cost, without considering the extensive
research on such models accomplished before Borch's first contribution was published,
and referred to in the preceding sections of this review.1
The last remark applies to the last term, Z(t), in (9a). With respect to the middle
term, Y(t), it is evident, that in Borch's approach only a part of the variation in the

1 Segerdahl kindly drew the reviewer's attention to a paper on the same topic by Klinger in 1965
[199"], where the stringent developments lead to some results, later pubhshed also by Borch.
Skand. AktuarTtdskr. 1968
 Review of the collective theory of risk. 1 23

continuous premium collected is accounted for. It is for example difficult to see any
possibility of accounting for the influence of the provisions by law and by the authori-
ties on the rating. Further, the part of the variation in the continuous premium
collected, which was referred to in the introduction of § 8 in this review, is not ac-
counted for in Borch's models. This variation is due to the fact, that, at least, if the risk
distribution and the claim distribution are dependent on time, every new tarif must
be based on new statistics, and on new predictions for the Z(t)-process. For an ideal
decision theory it seems, thus, necessary to combine Borch's ideas with a deeper study
of the Z(t)-process and the dependence of ¥(t) on the trends in the risk measures,
which determine the Z(t)-process. Such studies must be based on the ideas, which
led to the extensions of the classical form, and which were reviewed in § 5 here above.
Many items in the reference list deal with reinsurance problems [8, 11-12, 16, 21-22,
42, 49-53, 58-59, 67, 76, 84-85, 90, 98, 134, 139-141, 147, 155, 157-158, 160, 182,
190, 197, 200, 211-213, 218, 222-226, 246, 249, 253, 255-256, 258, 261-262, 267,
276, 283, 311,320, 328, 338-340, 353, 357,361-363]. A part of these papers have been
commented upon earlier in this review. Many of the papers referred to are based on
the classical form of the risk process, in some papers, however, e.g. [12, 16, 362], a
Polya process has been used in the model. Modern forms of reinsurance have been
discussed in [52, 338] and other papers.

Biihlmann suggested in [82], that--with a new terminology--the risk theory should

be divided into three such theories with regard to the different approaches in the
theory of decisions in an insurance company as described in the first, second and third
paragraph respectively of § 10 here above. Bi.ihlmann's argumentation for such
a terminology does not seem the reviewer very convincing. In all the approaches the
claim cost ought to be based on the knowledge of the Z(t)-process gained so far.
Arguments for an extension of this experience, and for a study of the influence of
this process on the Y(t)-process have been given in the fourth paragraph of § 10.
Therefore, the decision theory must be based on all contributions to the risk theory
reviewed in this paper, and on a further study of the risk process. Most of the develop-
ment of our knowledge of the risk process is made under the assumptions leading to
the classical form, and as far as the ruin functions are concerned, on the restrictive
assumption, that Y(t) is proportional to t eventually modified by the assumption that
the security loading depends on the magnitude of the risk reserve. This has led to
possibihties for the approximation, and for the application of F(x, t) under very
realistic particular assumptions with respect to the form of V(x). For much wider,
and, in fact, very mild conditions with respect to this function, and under the assump-
tion just mentioned for Y(t), the probabilities for the occurrence of ruin for the clas-
sical form have been completely treated as is seen from § 9 here above. The assump-
tions of the classical form involve, in fact, that the part of the variation in the con-
tinuous premium collected, which was described in the introduction in § 8, should
be reduced to a variation due to the sampling errors of the estimated premiums. The
assumptions of A and B in § 9 are, thus, connected with the assumptions for the
Skand. AktuarTidskr. 1968
24 Carl Philipson

classical form, as far as the influence of the provisions by law and authorities, and of
the competition is neglected. From the development of the classical form two lines of
development have branched out, one refers to the generalization of the fundamental
assumptions, as reviewed in § 4, followed by the extensions reviewed in § 5, and referred
to in several remarks in §§ 6-10. The other line refers to the extensions of the decision
theory, as reviewed in the second and third paragraph of § 10. The fourth paragraph
of § 10 points finally to a union of these two lines in future research. It is evident, that
a division of the risk theory according to Bi.ihlmann's suggestions implies the necessity
of using force against the strong connections between the three different view-points
in the decision theory, and the remaining part of the risk theory.
The development of the risk theory in its classical form has been accomplished by
Filip Lundberg, and by Cram6r and many others. The first extension of this form
was introduced by Ove Lundberg--Filip's s o n - - a n d by Ammeter, who were followed
by many others. The new ideas in the decision theory were introduced by de Finetti,
and by Borch, and studied by other authors. These lines of development are, how-
ever, all based on the fundamental conception of the collective risk theory, which was
created by Filip Lundberg.

Skand. AktuarTidskr. 1968