GENERAL J ARTICLE

On a Stochastic Model in Insurance

S Ramasubramanian

Basic aspects of the classical Cramer-Lundberg

insurance model are described.

Introduction

Recent cataclysmic events like the tsunami, torrential

downpour, floods, cyclones, earthquakes, etc. under-
score the fact that everyone would like to be assured The author is with the
that there is some (non-supernatural) agency to bank Statistics and Mathemat-
upon in times of grave need. If the affected parties are ics Unit, Imdian Statisti-
cal Institute, Bangalore.
too poor, then it is the responsibility of the government
His research interests
and the "haves" to come to the rescue. However, there centre around stochastic
are also sizeable sections of the population who are will- processes.
ing to pay a regular premium to suitable agencies during
normal times to be assured of insurance cover to tide
over crises. Insurance has thus become an important A reasonable
aspect of modern society. In such a set-up, a significant mathematical
proportion of the financial risk is shifted to the insur- theory of insurance
ance company. The implicit trust between the insured can possibly
and the insurance company is at the core of the inter- provide a scientific
action. A reasonable mathematical theory of insurance basis for the trust
can possibly provide a scientific basis for this trust. between the
insured and the
Certain types of insurance policies have been prevalent
company.
in Europe since the latter half of the 17th century. But
the foundations of modern actuarial mathematics were
laid only in 1903 by the Swedish mathematician Filip Keywords
Lundberg, and later in the 1930's by the famous Swedish Poisson process, exponential
distribution, arrival times, claim
probabilist Harald Cramer. Insurance mathematics to- size distribution, independent
day is considered a part of applied probability theory, increments, order statistics,
and a major portion of it is described in terms of con- uniform distribution, dis-
tinuous time stochastic processes. counted sum, premium, risk/
surplus process, ruin problem,
This article should be accessible to anyone who has taken ruin probability, random walk,
a course in probability theory. At least statements of the subexponenfiol distribution.

RESONANCE October 2006

 GENERAL J ARTICLE

Insurance various results and the heuristics can be appreciated.

mathematics today
While proofs of some of the basic results are given, for
some others only partial proofs or heuristic arguments
is considered a
are indicated; of course, in a few cases we are content
part of applied
with just citing an appropriate reference. [1,2] are very
probability theory.
good books where an interested reader can find more
information. It is inevitable t h a t a bit of jargon of basic
probability theory is assumed. One may look up [3-6] for
elucidation of terms like random variable, distribution,
density, expectation, independence, independent identi-
cally distributed (i.i.d.) random variables, etc. (Some
of the earlier articles in Resonance compiled in [7] also
contain a few introductory accounts).

Collective Risk Model

We shall mainly look at one model, known as the Cramer-

Lundberg model; it is the oldest and the most important
model in actuarial mathematics. This model is a par-
ticular type of a collective risk model. In a collective
risk model there are a number of anonymous but very
similar contracts or policies for similar risks, like insur-
ance against fire, theft, accidents, floods or crop dam-
age/failure, etc. The main objectives are modelling of
claims that arrive in an insurance business, and decide
how premiums are to be charged to avoid ruin of the
insurance company. Study of probability of ruin and
obtaining estimates for such probabilities are also some
of the interesting aspects of the model.

Main objectives are There are three main assumptions in a collective risk
modelling of claims model:
that arrive in an
1. The total number of claims, say N, occurring in a
insurance business,
given period is random. Claims h a p p e n at times {T~}
and decide how
satisfying 0 < T1 _~ T2 _~ .-- We call them claim
premiums are to be
arrival times (or just arrival times).
charged to avoid ruin
of the insurance 2. The i-th claim arriving at time Ti causes a payment
company. Xi. The sequence {Xi} is assumed to be an i.i.d.

50 RESONANCE J October 2006

 GENERAL I ARTICLE

sequence of nonnegative r a n d o m variables. These ran-

dom variables are called claim sizes.

3. The claim size process {Xi} and the claim arrival

times {Tj} are assumed to be independent. So {Xi}
and N are independent.

T h e first two assumptions are fairly natural, whereas the

third one is more of a m a t h e m a t i c a l convenience.

Take To = 0. Define the claim number process by

N(t) = max{i _> O " Ti <_ t}

= number of claims occurring by time t, t>0.
(1)
Also define the total claim amount process by

N(t)
S ( t ) : X 1 q- X 2 q- . . . Jr- X N ( t ) : E x i ' t ~_ O. (2)
i=1

These two stochastic processes will be central to our

discussions. Note t h a t a sample p a t h of N and the cor-
responding sample p a t h of S have j u m p s at the same
times T/, by 1 for N and by X / f o r S.

A function f ( . ) is said to be o(h) if lim ~ = 0; that is,

h--+0
if f decays at a faster rate than h.

Poisson Processes

We first consider the claim number process {N(t) 9 t _>

0}. For each t _> 0, note t h a t N(t, .) is a r a n d o m variable
on the same probability space ( ~ , f , P ) . We list be-
low some of the obvious/desired properties of N (rather
postulates for N ) , which may be taken into account in
formulating a model for the claim n u m b e r process.

9 ( N 1 ) : N(0) ~- 0. For e a c h t >_ O, N(t) i s a n o n -

negative integer-valued r a n d o m variable.

RESONANCE I October 2006 51

 GENERAL I ARTICLE

( N 2 ) : If 0 < s < t t h e n N(s) < N(t). Note

that N(t) - N(s) denotes the n u m b e r of claims in
the time interval (s,t]. So N is a nondecreasing
process.

( N 3 ) : T h e process { N ( t ) : t > 0} has independent

increments; t h a t is, if 0 < tl < t2 < - " < tn < (X)
then g(tl),N(t2) - N ( t l ) , . . - , g ( t n ) - Y(tn-1)
are i n d e p e n d e n t r a n d o m variables, for any n =
1, 2 , - . . . In other words, claims t h a t arrive in dis-
joint time intervals are independent.

( N 4 ) : T h e process {N(t)} has stationary incre-

ments; t h a t is, if 0 _< s < t, h > 0 then the ran-
dom variables N ( t ) - g(s) and N(t + h ) - Y(s +h)
have the same distribution (probability law). This
means t h a t the probability law of the n u m b e r of
claim arrivals in any interval of time depends only
on the length of the interval.

9 ( N h ) : Probability of two or more claim arrivals in

a very short span of time is negligible; t h a t is,

P(N(h) > 2 ) = o ( h ) , a s h $ 0. (3)

9 ( N 6 ) : T h e r e exists A > 0 such t h a t

P(N(h)=l)=Ah+o(h),as h~O. (4)

The n u m b e r )~ is called the claim arrival rate.

T h a t is, in very short time interval the probability
of exactly one claim arrival is roughly proportional
to the length of the interval.

R e m a r k 1. T h e first two postulates are self evident.

The hypothesis (N3) is quite intuitive; it is very reason-
able at least as a first stage approximation to m a n y real
situations. (Nh), (N6) are indicative of the fact t h a t be-
tween two arrivals there will be a gap, but may be very

RESONANCE I October 2006

 GENERAL I ARTICLE

small; (note that bulk arrivals are not considered here). In f o r m u l a t i n g a

(N4) is a time homogeneity assumption; it is not very model it is
crucial. desirable that the
R e m a r k 2. In formulating a model it is desirable that hypotheses are
the hypotheses are realistic and simple. Here 'realistic' realistic and
means that the postulates should capture the essential simple.
features of the phenomenon/problem under study. And
'simple' refers to the mathematical amenability of the
assumptions; once a model is chosen, theoretical prop-
erties and their implications should be considerably rich
and obtainable with reasonable ease. These two aspects
can be somewhat conflicting; so success of a mathemat-
ical model depends very much on the optimal balance
between the two. []

To see what our postulates (N1)-(N6) lead to, put

P,~(t)--P(N(t)=n), t>0, n=0,1,2,-.. (5)

Observe that

Po(t+ h)
= P ( N ( t ) = O , N ( t + h) - N ( t ) -- 0) (by (N1),(N2))
= P ( N ( t ) = 0 ) . P ( N ( t + h) - N ( t ) = 0) (by (N3))
= Po(t). P ( N ( h ) = 0) (by (N4),(N1))
= Po(t). [1 - Ah + o(h)] (by (Nh),(N6))

whence we get (as 0 < Po(t) <_ 1 ),

d
-~Po(t) = -APo(t), t > O. (6) Once a model is
chosen, theoretical
By (N1), note that P0(0) = P(N(O) = 0) = 1. So the
properties and
differential equation (6) and the above initial value give
their implications
Po(t) = P ( N ( t ) : O) = P ( N ( t + s) - g ( s ) : O) should be
=exp(-At), t>0, s>0. (7) considerably rich
and obtainable
Similarly for n _> 1, using (N3)-(N6), we get with reasonable
ease.
Pn(t + h) -~ P ( N ( t + h) = n) = I1 -F 12 -F I3,

RESONANCE I October 2006

 GENERAL J ARTICLE

The assumptions where

(N1)-(N6) are
I1 ---- P ( g ( t ) = n , N ( t + h) - N ( t ) = 0),
qualitative,
12 = P ( N ( t ) = n - 1, N ( t + h) - N ( t ) = 1),
whereas the
conclusion is 13 = P(N(t) < n- 2, N ( t + h) - N ( t ) > 2),
quantitative.
and hence

Pn(t + h) = Pn(t)[1 -- Ah + o(h)] + Pn-l(t)[/~h + o(h)]

+o(h).
We now get as before

d
-~Pn(t) = - A P n ( t ) q- ~Pn-l(t), t > O. (S)

Using the initial values Pn(0) = P ( N ( O ) = n) = 0,

n > 1 it is fairly easy to inductively solve (8) and get

_~t(/~t) n
Pn(t)=e n! ' n=0,1,2,..-, t_>0.

Thus we have proved the following theorem.

T h e o r e m 1. Let the stochastic process {N(t) 9 t _>

0} satisfy the postulates ( N 1 ) - ( N 6 ) . Then f o r any t >_
0, s _> 0, k = 0, 1 , 2 , . . .

P ( N ( t + s) - N ( s ) = k) = P ( N ( t ) = k)

(~t) k
k--7-- e x p ( - A t ) . (9)

The stochastic process {N(t)} is called a time h o m o -

geneous P o i s s o n process with arrival rate )~ > O.

R e m a r k 3. The assumptions (N1)-(N6) are qualita-

tive, whereas the conclusion is quantitative. Such a
result is usually indicative of a facet of nature; that
is, p h e n o m e n a observed in different disciplines, in un-
related contexts may exhibit the same law/pattern. In

$4 RESONANCE October 2006

 GENERAL [ ARTICLE

fact, Poisson distribution and Poisson process do come Poisson

up in diverse fields like physics, biology, engineering, and distribution and
economics. See [3-5]. [] Poisson process
T h e Poisson arrival model owes its versatility to the fact do come up in
t h a t m a n y natural (and, of course, useful) quantities diverse fields.
connected with the model can be explicitly determined.
We give a few examples which are relevant in the context
of insurance as well.

Interarrival and Waiting Time D i s t r i b u t i o n s

Let {N(t) 9 t _> 0} be a Poisson process with arrival

rate A > 0. Set To - 0. For n = 1 , 2 , . . - define Tn =
inf{t _> 0 9 N ( t ) = n} = time of arrival of n-th claim
(or waiting time until the n-th claim arrival). P u t An =
T~ - Tn-1, n = 1, 2, .. 9 so t h a t A~ = time between (n -
1)-th and n-th claim arrivals. Recall from our initial
c o m m e n t s t h a t we had in fact defined t h e process {N (t)}
starting from {Ti}. T h e r a n d o m variables To, T1, T2,.--
are called claim arrival times (or waiting times); the
sequence {An " n = 1, 2 , . - . } is called t h e sequence of
interarrival times.

For a n y s > 0 n o t e t h a t {TI> s}={N(s)=0}; hence

by (9)

P(A1 > s)= P(T1 > s)= P(N(s)= 0)= exp(--As).

(lo)
So P ( A 1 <_ s) = 1 - e -as, s >_ O. Therefore the r a n d o m
variable A1 has an Exp(A) distribution ( = exponential
distribution with p a r a m e t e r A > 0); t h a t is,
b

P(A1 9 (a,b))=/Ae-~Sds,
a
O<a<b<c~. (11)

Next let us consider the joint distribution of (TI, T2). Let

F(T1,T2) denote the joint distribution function of (T1, T2);
t h a t is, F(r~,T2)(tl,t2) = P(T1 < tx,T2 < t2). As 0 <

RESONANCE J October 2006

 GENERAL I ARTICLE

T1 _< T2 it is enough to look at F(T1,T2)(tl, t2) for 0 <

tl _~ t2. It is clear that for 0 ~ tl ~ t2,
{T1 _< tl,T2 <_ t2} = {N(tl) >_ 1, N(t2) >_ 2}
= {N(tl) -- 1, N(t2) - N(tl) _> 1} U {N(tl) _> 2},
where the r.h.s, is a disjoint union. So using properties
(N3), (N4) and equation (9) we get
F(T1,T2) (tl, t2)
= P ( N ( t l ) = 1 , N ( t 2 ) - N(tl) >_ 1) + P ( N ( t l ) >_ 2)
= Atle-~t'(1 -- e -~(t2-tl)) + [1 -- (e -~tl + Atle-~tl)]
= - A t l e -~t2 + H(tl),
where H is a function depending only on tl. Conse-
quently the joint probability density function f(T,,T2)
of (T1, T2) is given by
02
f(T1,T2)(tl, t2) a_ Ot_~_otlF(T1,T2)(tl, t2)
_ { A2e-~t2, if O < tl < t2 < oo }
0, otherwise.
(12)
To find the joint distribution of (A1, A2) from the above,
note that
( 0
_ T2 ) . (13)
The linear transformation given by the (2 • 2) matrix
in (13) has determinant I, and transforms the region
{(ix, t2) 9 0 < tl < t2 < ~} in 1 - 1 fashion onto
{(al, a2) " al > 0, a2 :> 0}. So the joint probability den-
sity function f(A1,A2) of (A1, A2) is given by
f(A1,A2) (al, a2)
= f(T~,T~)(al, a~ + a2)

= { (Ae-;~a~)(~e-Aa2)' if al ~ O~a2 ~ O
otherwise.

(14)

$6 RESONANCE [ October 2006

 GENERAL J ARTICLE

Thus A1, A2 are independent random variables each hav- The interarrival
ing an exponential distribution with parameter A. times are
W i t h more effort one can prove the following theorem. independent
random variables
T h e o r e m 2. Let {N(t) : t > 0} be a time homogeneous having exponential
Poisson process with arrival rate A > O. Let A1, A2, 999 distribution.
denote the interarrival times. Then { A n : n = 1 , 2 , . . . }
is a sequence of independent, identically distributed ran-
dom variables (or in other words an i.i.d, sequence)
having Exp(A ) distribution. []

N o t e : As A1 has Exp(A) distribution, its expectation is

given by E(A1) = ~; so $1 is the mean arrival time. Thus
the arrival rate being A is consistent with this conclusion.

N o t e : It is an easy corollary of the theorem that Tn =

A1 -t- A2 -F . . . -~- An has the gamma distribution F(n, A).

R e m a r k 4. One can also go in the other direction.

T h a t is, let 0 = To _< T1 _< T2 _< --. be the claim ar-
rival times; let An = Tn - Tn-1, n > 1. Suppose {An}
is an i.i.d, sequence having Exp(A) distribution. De-
fine {N(t)} by (1). T h e n the stochastic process {N(t) 9
t > 0} can be shown to be time homogeneous Poisson
process with rate A. In the jargon of the theory of sto-
chastic processes, Poisson process is the renewal process
with i.i.d, exponential arrival rates.

Order Statistics Property

This is another i m p o r t a n t property of the Poisson process.

Recall that for events G, H, the conditional probability
of G given H is defined by P ( G I H) =A P(CnI4) P(H) " We first
prove
Order statistics
T h e o r e m 3. N o t a t i o n as earlier. For 0 <_ s <_ t, property is crucial
8 in explicit
P(A1 < s [ N(t)= l)= ~-; (15) computations
involving Poisson
that is, given that exactly one arrival has taken place
processes.
in [0, t], the time of the arrival is uniformly distributed

RESONANCE J October 2006 57

 GENERAL J ARTICLE

over (0, t).

Proof. As t h e P o i s s o n process has i n d e p e n d e n t incre-
ments,

P(A1 < s , N ( t ) = 1)
P(A1 < s [ N ( t ) = 1 ) = P(N(t) =1)
P ( N ( s ) = 1 , N ( t ) - N(s) = O)
P ( N ( t ) = 1)
P ( N ( s ) = 1 ) - P ( N ( t ) - N(s) = 0 )
P ( N ( t ) = 1)
Ase-~Se -~(t-s) s
Ate -~t t '

c o m p l e t i n g t h e proof. []

A n a t u r a l q u e s t i o n is: If N(t) = n, w h a t can one say

a b o u t t h e c o n d i t i o n a l d i s t r i b u t i o n of 7'1, T 2 , . . . , Tn?

Theorem 4. Let { N ( t ) 9 t >_ 0}, T1, :/12,--. be as before.

For any t > O, and any n = 1 , 2 , . - . the conditional
density of ( T 1 , T 2 , . " ,Tn) given g ( t ) = n is
1
fT1,T2,'",Tn((81, S2, " ' " ,8n) I N(t) = n) = n ! . - -t n ' (16)

for 0 < s I < s 2 < . . . < s n < t.

P r o o f . For n o t a t i o n a l s i m p l i c i t y we t a k e n = 2; t h e gen-
eral case is similar. Let 0 < sl < s2 < t; t a k e hi, h2 > 0
small e n o u g h t h a t 0 < S l < S l + h l < s2< s2+h2< t.
T h e n again u s i n g t h e i n d e p e n d e n t increment property
a n d (9), we get

P ( S l < T1 < S l + h l , S 2 < T2 <: s 2 + h 2 I N ( t ) -- 2)

1
P ( N ( t ) = 2)"

N ( S l ) = 0, N ( s l -{- hi) - N ( s l ) = 1,
P N ( s 2 ) - N(sl + h i ) = 0,
N(s2 + h2) - N ( s 2 ) = 1, N ( t ) - N(s2 + h2) = 0 )
S6 RESONANCE I October 2006
 GENERAL I ARTICLE

1
e-at(At)2~2!
9{ e - A S 1 / ~ h l e - A ( s l + h l - s l ) e - A ( s 2 - ( S l + h l ) )
9A h2e-A(s2+h2-s2)e - A ( t - (s2+h2)) }

2!
= -~hlh2.

Dividing by hlh2 and letting hi, h2 ~ 0 we get the desired

result. []

R e m a r k 5. Let V1, 1 / 2 , - ' - , Vn be i.i.d, r a n d o m vari-

ables each having a uniform distribution over (0, t), where
t > 0 is fixed. Note t h a t the probability density function
of each Vi is

1
fvi(s)=7' if O < s < t

= O, otherwise.

Let V(1) <_ V(2) _< . . . _< V(n) denote the order statistics of
P l , g 2 , " " , Vn. T h a t is, V ( 1 ) ( ~ ) , V(2)(od), 9 9 9 , V(n)(a)) de-
notes Vl(w), V2(w), 9 99 , Vn(w) arranged in increasing or-
der for any a; E fL It is not difficult to show t h a t the
joint probability density function of V0), V(2), 999 , V(n) is
given by the r.h.s, of (16). So the preceding theorem
means t h a t

((T1,T2,"" ,rn) I N ( t ) = r t ) d (V(1),V(2)... ,g(n)) '

(17)

where d denotes t h a t two sides have the same proba-

bility distribution. If U1, U 2 , - " , Un are i.i.d. U(0,1)
r a n d o m variables (that is, having uniform distribution
over (0, 1)), then (17) can be expressed as

((T1,T2,... ,r~) I N ( t ) = n) d (tUo),tU(2),.. " ,tU(,)).

(18)
[]

RESONANCE I October 2006 59

 GENERAL I ARTICLE

An important consequence of Theorem 4 and Remark 5

is the following result whose proof is quite involved; see
[1].

T h e o r e m 5. Let { N ( t ) : t _> O} be a time homogeneous

Poisson process with rate ~ > 0; let 0 < T1 < T2 < ...
denote the claim arrival times corresponding to N(.).
Let {Xi : i = 1, 2 , - . - } be an i.i.d, sequence indepen-
dent of the process {N(t)}. Then there exists a sequence
{Uj : j = 1 , 2 , . . - } such that (i) {Uj} is a sequence of
i.i.d, random variables having U(0,1) distribution, (ii)
the families {Uj}, {Xi}, {N(t)} are independent of each
other, (iii) for any reasonable function g of two variables
N(t) N(t)
E g ( T i , Xi ) d E g ( t V i , Xi) ' t>O. (19)
i=1 i=1

[]

The basic strategy for proving Theorem 5 can be easily

stated. Conditioning the 1.h.s. of (19) by {N(t) = n},
we use Theorem 4 to replace 7"/ by tU(i). Then invok-
ing independence of the families {Ui}, {Xj} and the fact
that Xj's are i.i.d.'s, we permute X1, X2, ..- , X~ suit-
ably to facilitate the desired conclusion. Mathematical
justification requires measure theoretic machinery.

Cramer-Lundberg Model

This is the classical and very versatile model in insur-

ance. The claim arrivals {Ti} happen as in a time homo-
geneous Poisson process with rate ~ > 0. The claim sizes
{Xi} are i.i.d, nonnegative random variables. The se-
quences {Xi}, {Tj} are independent of each other. The
total claim amount upto time t in this model is given
by
N(t)
s(t) = Xl + +... + = t>_O.
i=1

60 RESONANCE I October 2006

 GENERAL I ARTICLE

which is the same as (2). Note that {S(t) 9 t > 0} is an

example of a c o m p o u n d P o i s s o n process.

We now look at the discounted s u m corresponding to

the above model. Let r > 0 denote the interest rate.
Define
N(t)

. = N
-So ,(t~ " e-rT~ x ~.,
z__, t >_ o. (20)
i=1

This is the "present value" (at time 0) of the cumulative

claim amount over the time horizon [0, t]. By Theorem 5
for any t > 0

N(t)
So(t) ~ ~ e-~'~'X~, (21)
i=l

where {Ui} is an i.i.d. U(0,1) sequence as in the theorem.

Therefore using the independence of the three families
of random variables we get

E(So(t))

= ~_, E e-r~U'Xi I X(t) = n 9 P ( N ( t ) = n)

n =O i=1

n=O

= E(N(t)). E(Xl). e-r~Ydy

= ~!(1 - e-~'). E ( X l ) .
r

So we have proved the following theorem.

RESONANCE I October 2006 61

 GENERAL ] ARTICLE

Equation (22) gives Theorem 6. With the notation as above

the average amount
needed to take care of
claims over an initial
period, when premium
income might not be That is, in the Cramer Lundberg model, the average//
sizable. expected amount needed to take care of claims over [0, t]
is given by (22). []

Next let p(t) denote the premium income in the time

interval [0, t]. In the Cramer Lundberg model it is as-
sumed that p(.) is a deterministic linear function; that
is, p(t) = ct, t >_ 0 where c > 0 is a constant called the
premium rate. W i t h the total claim amount S(.) defined
by (2), put for t > 0,

N(t)
u(t) : u + p(t) - s(t) = u + ct - ~ x~. (23)
i=1

The process { U ( t ) : t _> 0} is called the risk process (or

surplus process) of the model; here u is the initial cap-
ital. Note that U(t) is the insurance company's capital
balance at time t. Letting r I 0 in (22) or otherwise,
note that E(S(t)) = I t E ( X 1 ) and hence

E(U(t))= u +ct- E(S(t)) = u + c t - AtE(Xl). (24)

By (24), a minimal requirement in choosing the premium

rate may be taken to be

c > AE(X1) (25)

so that on the average, claim payments are taken care

of by premium income. This somewhat simple criterion
can be justified by other considerations also, as we shall
see later. A more prudent condition is to require that
c > (1+p)AE(X1), where p > 0 is called a safety loading
factor.

RESONANCE I Or 2006
 GENERAL [ ARTICLE

Ruin Problem in the Cramer-Lundberg Model A theoretical

understanding of
As mentioned earlier, in an insurance set-up the finan-
conditions leading to
cial risk is shifted to the insurance c o m p a n y to a large
extent. There have been m a n y instances w h e n insurance ruin of the company,
companies have gone b a n k r u p t unable to cope up with probability of ruin,
claims during m a j o r catastrophes. So a theoretical un- severity of ruin, etc.
derstanding of conditions leading to ruin of the company, will help at least in
probability of ruin, severity of ruin, etc. will help at least avoiding certain
in avoiding certain pitfalls. Study of ruin problems has, pitfalls.
therefore, a central place in insurance m a t h e m a t i c s .

T h e event t h a t the surplus U(-) falls below zero is called

~ i n . Set
T = inf{t > O ' U ( t ) < 0}; (26)

T is called the ruin time; it is the first time the surplus

falls below zero. T h e probability of ruin is t h e n

r c~ I U ( O ) = u ) (27)

for u > 0; it is considered as a function of the initial

capital u. Note t h a t r depends on the p r e m i u m rate
c as well. A very n a t u r a l question is: For w h a t premium
rates c and initial capital u can it happen t h a t r = 1?
T h a t is, when is ruin certain?

By the definition of U(.), note t h a t U(.) increases in the

intervals [Tn, T~+I), n > 0. Therefore ruin can occur only
at s o m e T n . N o w f o r n > 1,
n
U(Tn) = u + cTn - E Xi (because N(Tn) = n)
i=1
n n

=u+E(cA~-Xi) because
i=1 i=1
(28)
n

Put Zi = Xi - cAi, i > 1, So = 0, S n = E Zi, n ~ 1.

i=1
T h e n (28) is just U(Tn) = u - Sn, n > 1. Since "ruin"

RESONANCE J October 2006

 GENERAL J ARTICLE

= {U(Tn) < 0 for some n}, it is now easy to see t h a t

~ ( u ) = P ( s u p Sn > u). (29)
n_>l
Since the families {Ai} and {Xj} are m u t u a l l y indepen-
dent, and each is a sequence of i.i.d.'s, note t h a t {Zi} is
a sequence of i.i.d.'s and hence {S,~ 9 n > 0} is a random
walk on the real line R. The following result concerning
r a n d o m walks on ]~ is known.

Theorem 7. Let {Zi}, {Sn} be as above; assume that

Zi is not identically zero, and E(Zi) exists.
(a) If E(Z1) > 0, then P ( l i m Sn = + c o ) = 1.
n----*O0

(b) If E ( Z l ) < 0, then g ( l i m Sn = - o c ) = 1.

n---* Oo

(c) If E(Z1) = O, then

P ( l i m sup Sn = + c o , lim inf Sn = --cxD) ---- 1.
n---* (x) n---+cx)

[]

N o t e : While (a), (b) above are i m m e d i a t e consequences

of the strong law of large numbers, assertion (c) requires
a lengthy proof; see [2,3]. []

From (29) and the above theorem it follows t h a t r =

1 for all u > 0, if E(X1) - cE(A1) >_ 0; note t h a t
E(A1) = $1 as A1 has an exponential distribution with
p a r a m e t e r A; so in the C r a m e r - L u n d b e r g model ruin is
certain if (25) is not satisfied. The condition (25) is called
the net profit condition, which is generally assumed to
be satisfied.

If (25) holds, the above does not imply that ruin is

In the Cramer-
avoided; it only means t h a t one m a y hope to have r (u) <
Lundberg model, ruin
1, u > O. In this direction we have the following result.
is certain if equation
(25) is not satisfied. Theorem 8.In the Cramer-Lundberg model assume
The condition (25) is that the net profit condition (25) holds. Assume also
called the net profit that there exists r > 0 such that
condition. E[exp(rZ1)] = E [ e x p ( r ( X l - c A 1 ) ) ] - 1. (30)

RESONANCE J October 2006

 GENERAL J ARTICLE

Then

r < exp(-ru), for u > 0. (31)

[]

The constant r in (30) is called the adjustment coeffi-

cient; in the Cramer-Lundberg model, this coefficient
exists if (25) holds and X1 has moment generating func-
tion (or Laplace transform) in a neighbourhood of 0. The
inequality (31) is known as Lundberg inequality. See [2]
for proof and extensions.

N o t e : An elegant way of proving Theorem 8 is through

martingale methods. Martingale theory is a powerful
tool in a probabilist's kit. One may see [6,8] for some of
the elementary aspects and applications of martingales,
and [2] for applications to insurance.

R e m a r k 6. In addition to the above bound one can

also derive an integral equation for the ruin probability.
Suppose E(X1) < oo and that the net profit condition
holds. Then one can get a distribution function G (ex-
plicitly in terms of the distribution function of X1) such
that
t

r - AE(X1)[1 - G(t)] + AE(X1) / r - y)dG(y).

c c
0
(32)
Equation (32) is a renewal type equation; however it
is a defective renewal equation because AE(X1)/c < 1
(as the net profit condition holds). Still following the
methods of renewal theory one can get a series solution
to (32). See [1,21 for details. []

Claim Size Distribution

The common distribution of the i.i.d, sequence {Xi} is

called the claim size distribution. With the exception

RESONANCE J October 2006

 GENERAL t ARTICLE

Risks regarding of Theorem 8, we have not m a d e any specific reference

insurance of to the claim size distribution so far. A conventional as-
airplanes, sumption is t h a t Xi have an exponential distribution.
skyscrapers, In such a case P ( X i > x) = e -;~x,x > 0, where ~ > 0;
dams, bridges, etc. t h a t is, the (right) tail of the claim size distribution
are very high. decays at an exponential rate. Most of the distribu-
tions used for modelling in statistics have this property.
Companies have
The ubiquitous normal (or Gaussian) distribution de-
also faced ruin due
cays even at a faster rate. Such distributions are called
to a very small
light tailed distributions; for these distributions the mo-
number of very
ment generating functions exist in a neighbourhood of 0.
huge claims.
An i m p o r t a n t development of late is to consider claim
sizes that are not necessarily light tailed. Risks regard-
ing insurance of airplanes, skyscrapers, dams, bridges,
etc. are very high. In recent years, companies have also
faced ruin or near ruin due to a very small number of
very huge claims; in some instances, a single massive
claim has done the damage. T h e r e are quite a few no-
tions of heavy tailed distributions; invariably the mo-
merit generating function does not exist in any neigh-
bourhood of 0 for these distributions. A versatile notion
of heavy-tailedness in the insurance context is given be-
low.

Let F be a distribution function supported on (0, oc);

(this corresponds to a positive r a n d o m variable). Let
X 1 , X 2 , - ' . be an i.i.d, sequence with c o m m o n distrib-
n

ution function F . Set Sn = ~ Xi, Mn = max{X1, X2,

i=1
,Xn}. If

P ( S n > x)
lim -1, for n > 2, (33)
=--+ooP ( M n > a~)

then F is said to be subexponential. If F is subexpo-

nential then it can be shown t h a t ea=P(X >_ x) ---+ oo,
for any a > 0, where X is a r a n d o m variable with distri-
bution function F. So moment generating function does
not exist in any neighbourhood of 0 for such distribu-

66 RESONANCE I October 2006

 GENERAL I ARTICLE

tions. Two classes of subexponential distributions are Suggested Reading

given below.
[1] T Mikosch, Non-life Insur-
(i) Weibul distribution: In this case T ( z ) a__ 1 - F(z) = ance Mathematics: an Intro-

e x p ( - c S ) , i f z > 0, and F(z) = 0, i f z _< 0, where duction with Stochastic Pro-

cesses, Springer, New Delhi,
c > 0, ~- > 0 are constants. This family of distributions
2004.
has been useful in reliability theory, besides insurance. [2] T Rolski, H Schmidli, V
If 0 < ~- < 1, then F is subexponential. See [1,2]. Schmidt and J Teugels, Sto-
chastic Processes for Insur-
(ii) Pareto distribution: Again it is convenient to de- ance andFinance, Wiley, New
fine in terms of the right tail of the distribution; here York, 2001.
--fi(z) a= 1 - F(z) = ~ / ( ~ + z)'~,z > 0, where ~,c~ > 0 [3] W Feller, An Introduction to
Probability Theory and its Ap-
are constants. This class is subexponential; (even expec-
plications, 2 vols, Wiley-East-
tation exists only when c~ > 1.) This family has also ern, New Delhi, 1991.
been used in economics to describe income distributions. [4] P G Hoel, S C Port and C J
Stone, Introduction to Prob-
As the moment generating function does not exist for ability Theory, Universal
heavy tailed distributions, note that Theorem 8 is not Book Stall, New Delhi, 1991.
applicable. In fact, when the claim size distribution be- [5] S M Ross, Introduction to
Probability Models, 8th edi-
longs to an appropriate subclass of subexponential dis-
tion. Elsevier India, New
tributions, it can be established that the ruin probabil- Delhi, 2005.
ity decays only at a power rate, viz. r behaves like [6] S M Ross, Stochastic Pro-
K u -~ for large u, where K, ~i > 0 are suitable constants. cesses, 2nd edition, (Wiley
Contrast this to the exponential rate e -ru, where r > 0 Student Edition), John
Wiley, Singapore, 2004.
in Theorem 8. So ruin is much more formidable if the
[7] M Delampady, T Krishnan
claim size distribution is heavy tailed. See [1,2]. and S Ramasubramanian
(eds.), Probability and Sta-
Assorted Comments tistics. A volume in Echoes
from Resouance, Universities
We have dealt with a few elementary aspects of just one
Press, Hyderabad, 2001.
model. Comments below are meant to give a flavour of [8] S Karlin and H MTaylor, A
some other aspects/models. First Course in Stochastic Pro-
cesses, Academic Press, New
1. A more general model is the renewal risk model (also York, 1975.
called Sparre Andersen model). In this model, the in-
terarrival times A1, A 2 , " " are just i.i.d, nonnegative
random variables (not necessarily exponentially distrib-
uted). The net profit condition is given by an analogue
of (25), viz. c > E(X1)/E(A1). Lundberg inequality
holds provided that the net profit condition is satisfied

RESONANCE I October 2006 67

 GENERAL J ARTICLE

and that the adjustment coefficient exists. Renewal risk

model with subexponential claim sizes continue to be
objects of research.

2. Life insurance/pension insurance models are gen-

erally described in terms of continuous time Markov
processes with state space having only a finite number
of elements; at least one state is absorbing, and cer-
tain transitions may be disallowed. For example, in the
simplest life insurance model there are only two states,
one signifying "alive" and the absorbing state indicating
"dead", reflecting the status of the insured.

3. In addition to the basic insurance aspects, more com-

plex models can be considered. For example, an insur-
ance company can invest part of its surplus in bonds
giving returns at fixed rates, and another part in stocks
which are subject to the volatility of the market. Some
problems of interest are how optimally should these in-
vestments be made so that the ruin probability is mini-
mized, or so that the dividend payment by the company
is maximized.

4. We have not touched upon any statistical aspect like

estimation of claim arrival rate, parameters of the claim
size distribution, or when does claim size data indicate
heavy tailed behaviour, etc.

[1,2] and the references therein deal with the above is-
sues and more.

Acknowledgement
This is an expanded version of the Professor Wazir Hasan
Abdi memorial lecture given on October 1, 2005 at the
Department of Mathematics, Cochin University of Sci-
ence and Technology, Kochi.

68 RESONANCE J October 2006