Vous êtes sur la page 1sur 96

Insurance Risk Theory

Leda D.Minkova
April 9, 2011
2
Contents
1 Basic Risk Model 7
2 Random Variables 9
2.1 Hazard rate function . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3 Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Probability Generating Functions . . . . . . . . . . . . 12
2.4 Moment Generating Functions . . . . . . . . . . . . . . . . . . 13
2.5 Characteristic Functions . . . . . . . . . . . . . . . . . . . . . 13
2.6 Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . 14
2.7 Counting Random Variables . . . . . . . . . . . . . . . . . . . 15
2.8 Continuous random variables . . . . . . . . . . . . . . . . . . 17
2.8.1 Exponential distribution and lack of memory property 17
2.8.2 Gamma distribution . . . . . . . . . . . . . . . . . . . 18
2.8.3 Beta distribution . . . . . . . . . . . . . . . . . . . . . 19
2.8.4 Weibull distribution . . . . . . . . . . . . . . . . . . . 20
2.8.5 Pareto distribution . . . . . . . . . . . . . . . . . . . . 20
2.8.6 The normal distribution . . . . . . . . . . . . . . . . . 21
2.8.7 Log-normal distribution . . . . . . . . . . . . . . . . . 22
2.8.8 Inverse Gaussian distribution . . . . . . . . . . . . . . 22
2.9 Functions of random variables . . . . . . . . . . . . . . . . . . 23
2.10 Joint density and distribution function . . . . . . . . . . . . . 24
2.11 Conditional distributions . . . . . . . . . . . . . . . . . . . . . 25
2.12 Sum of random variables . . . . . . . . . . . . . . . . . . . . . 26
2.12.1 Negative binomial distribution . . . . . . . . . . . . . . 27
2.12.2 Erlang distribution . . . . . . . . . . . . . . . . . . . . 28
2.13 Mixed distributions . . . . . . . . . . . . . . . . . . . . . . . . 29
2.14 Compound distributions . . . . . . . . . . . . . . . . . . . . . 31
2.14.1 Hiperexponential distribution . . . . . . . . . . . . . . 33
3
4 CONTENTS
3 Counting processes 35
3.1 Poisson process . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.1 Order statistics property . . . . . . . . . . . . . . . . . 39
3.2 Renewal process . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.1 Renewal function . . . . . . . . . . . . . . . . . . . . . 44
3.2.2 Recurrence times of a renewal process . . . . . . . . . . 46
3.3 Delayed Renewal process . . . . . . . . . . . . . . . . . . . . . 48
3.4 Mixed Poisson process . . . . . . . . . . . . . . . . . . . . . . 49
3.5 Compound Poisson process . . . . . . . . . . . . . . . . . . . . 50
3.5.1 Polya - Aeppli process . . . . . . . . . . . . . . . . . . 50
4 Claim Size Models 51
4.1 Heavy tailed distributions . . . . . . . . . . . . . . . . . . . . 53
4.2 Regularly varying functions . . . . . . . . . . . . . . . . . . . 53
4.2.1 Properties . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.2 Regularly varying random variables . . . . . . . . . . . 55
4.3 Subexponential distributions . . . . . . . . . . . . . . . . . . . 57
4.3.1 Properties . . . . . . . . . . . . . . . . . . . . . . . . . 57
5 Cramer - Lundberg model 63
5.1 Ruin probability . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 Integral equation of ruin probability . . . . . . . . . . . . . . . 64
5.3 Cramer - Lundberg approximation . . . . . . . . . . . . . . . 66
5.4 Martingale approximation . . . . . . . . . . . . . . . . . . . . 69
6 Renewal Risk Model 71
6.1 Ordinary renewal risk model . . . . . . . . . . . . . . . . . . . 71
6.1.1 Lundberg exponent . . . . . . . . . . . . . . . . . . . . 72
6.1.2 Pollaczeck - Khinchine formula (Ruin probability as a
compound geometric probability) . . . . . . . . . . . . 74
6.2 Stationary case . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.3 Ruin probability for heavy tailed distributions . . . . . . . . . 76
7 Premium Calculation Principles 79
7.1 Premium calculation principles . . . . . . . . . . . . . . . . . 79
7.1.1 Pure premium principle . . . . . . . . . . . . . . . . . 79
7.1.2 Expected value principle . . . . . . . . . . . . . . . . . 79
7.1.3 The variance principle . . . . . . . . . . . . . . . . . . 80
7.1.4 Standard deviation principle . . . . . . . . . . . . . . . 80
7.1.5 Modied Variance Principle . . . . . . . . . . . . . . . 80
7.1.6 The Principle of Zero Utility . . . . . . . . . . . . . . . 80
CONTENTS 5
7.1.7 The Esscher Principle . . . . . . . . . . . . . . . . . . 82
7.1.8 Risk adjusted premium principle . . . . . . . . . . . . 83
8 Diusion Approximation 85
8.1 Ruin Probability for diusion process . . . . . . . . . . . . . . 87
9 Reinsurance 91
9.1 Proportional Reinsurance . . . . . . . . . . . . . . . . . . . . . 91
9.2 Excess - of - Loss Reinsurance(XL) . . . . . . . . . . . . . . . 93
9.3 Stop - Loss Reinsurance . . . . . . . . . . . . . . . . . . . . . 94
Bibliography 95
6 CONTENTS
Chapter 1
Basic Risk Model
The foundation of the modern risk theory goes back to the works of Filip
Lundberg and Harald Cramer. In 1903 Filip Lundberg proposed the Poisson
process as a simple process in solving the problem of the rst passage time.
Lundbergs work has been extended by Harald Cramer in 1930 for modeling
the ruin of an insurance company as a rst passage time problem. The basic
model is called a Cramer - Lundberg model or classical risk model. Insurance
Risk Theory is a synonym of non-life insurance mathematics.
The basic stochastic elements of the general risk model are:
The times 0
1

2
. . . , of claim arrivals. Suppose that
0
= 0.
The random variables T
n
=
n

n1
, n = 1, 2, . . . , called inter - occurrence
or inter - arrival times are nonnegative.
N(t) = sup{n :
n
t}, t 0 is the number of claims up to time
t. The relations between the times {
0
,
1
, . . .} and the counting process
{N(t), t 0} are given by
{N(t) = n} = {
n
t <
n+1
}, n = 0, 1, . . . .
The sequence {Z
n
, n = 1, 2, . . .} of independent identically distributed
random variables represents the amounts of the successful claims to the in-
surance company. Suppose that the sequence {Z
n
} is independent of the
counting process N(t).
The premiums to the insurance company. Let us denote by (t)
the total amount of the premiums up to time t.
7
8 CHAPTER 1. BASIC RISK MODEL
The accumulated sum of claims up to time t is given by
S(t) =
N(t)

i=1
Z
i
, t 0.
The process S = (S(t))
t0
is dened by the sum S
n
= Z
1
+ . . . + Z
n
, where
n is a realization of the random variable N(t) :
S(t) = Z
1
+ . . . + Z
N(t)
= S
N(t)
, t 0,
or a random sum of random variables. Suppose that S(t) = 0, if N(t) = 0.
The process X(t) = (t)S(t) is called a risk process. The risk reserve
of the insurance company with initial capital u is given by
U(t) = u + (t) S(t), t 0.
Chapter 2
Random Variables
All the random variables are dened on a complete probability space (, F, P).
Denition 2.1 The function
F
X
(x) = P(X x) = P(X (, x]). (2.1)
is called a distribution function of the random variable X.
The distribution function F
X
(x) satises the following:
a) F
X
() = lim
x
F
X
(x) = 0.
b) F
X
() = lim
x
F
X
(x) = 1.
c) F
X
(x) is nondecreasing, i.e., if x < y,
F
X
(x) F
X
(y).
d) F
X
(x) is a right continuous, i.e. for any x,
F
X
(x + 0) = F
X
(x).
The function F
X
= 1 F
X
(x) = P(X > x) is called survival function
for X and represents the tail distribution.
The random variables are characterized by the properties of the distribu-
tion function in three dierent types.
The random variable X is called discrete, if the distribution function is a
stepwise and has countable many jumps.
For the discrete random variable there exists a countable subset E =
{x
0
, x
1
, . . .} of R such that P(X E) = 1. In this case the probability
function is given by
p
k
= P(X = x
k
), k = 0, 1, . . .
9
10 CHAPTER 2. RANDOM VARIABLES
and the distribution function
F
X
(x) =

x
k
x
P(X = x
k
),
where the sum is over all jump points of F
X
(x).
A discrete r.v. X is called arithmetic with a step h, if the values of X
belong to the set {x = ih : i = 0, 1, 2, . . . , }. We use arithmetic distribu-
tion in the case of h = 1. An arithmetic distribution over the nonnegative
values is called a counting distribution. So the counting random variables are
discrete with integer nonegative values, i.e.
p
k
= P(X = k), k = 0, 1, . . . .
The r.v. X is called continuous, if the distribution function F
X
(x) is
continuous. The continuous r.v. X is called absolutely continuous, if F
X
(x)
is dierentiable over all R.
For the continuous r.v. the function
f
X
(x) =
d
dx
F
X
(x),
if it exists, is called probability density function of X. The function f
X
(x) is
nonnegative. It is easy to see that
_

f
X
(x)dx = 1 and F
X
(x) =
_
x

f
X
(t)dt.
In a more general case for any event A,
P(A) =
_
A
f(t)dt.
Intuitively, we could write that P(X = x) = f(x)dx and consider P(X = x)
like the probability P(x < X x + dx) over an innitesimal interval.
The r.v. X is of a mixed type, if it is neither discrete nor continuous. If
x is a jump point of F(x), then we say that X has a mass P(X = x) at the
point x. In the points where F(x) is absolutely continuous we consider the
r.v. like continuous.
2.1 Hazard rate function
Let X be a continuous r.v. with distribution function F and density function
f.
2.1. HAZARD RATE FUNCTION 11
Denition 2.2 The function
(t) =
f(t)
F(t)
is called hazard rate function for the r.v. X.
Depending of the applications, the hazard rate function is called also
failure rate function, force of default, intensity function. Let X be the life
distribution. Then the hazard rate function is equal to the probability that
the individual, survived the time t dies in the additional time dt, i.e.
P(X (t, t + dt)|X > t) =
P(X (t, t + dt), X > t)
P(X > t)
=
P(X (t, t + dt))
P(X > t)

f(t)dt
F(t)
= (t)dt.
The function (t) is interpreted like the intensity of deathes in the set of
individuals at age t.
The function (t) characterizes the distribution of the r.v. Let (t) =
_
t
0
f(s)
F(s)
ds.
Since
(t) =

d
dt
F(t)
F(t)
it follows that
(t) = log F(t).
Consequently
F(t) = exp
_

_
t
0
(s)ds
_
= exp((t)). (2.2)
The function
(t) =
_
t
0
(s)ds
is called a hazard function.
For a discrete random variable with distribution {p
k
}

k=0
the hazard rate
function is given by
(k) =
p
k

j=k
p
j
, k = 0, 1, . . . .
In this case (k) 1.
12 CHAPTER 2. RANDOM VARIABLES
2.2 Moments
Mathematical expectation = EX of the r.v. X is dened by
EX =

k
x
k
p(x
k
),
if X is a discrete and
EX =
_

xf(x)dx,
if X is continuous, conditionally on

k
|x
k
|p(x
k
) < and
_

|x|f(x)dx <
. The mathematical expectation is called also mean value or simply a mean.
In the more general case, for any measurable function g : R R, the
mathematical expectation Eg(X) is dened by the Lebesgue-Stieltjes integral
Eg(X) =
_

g(x)dF(x), providing that


_

|g(x)|dF(x) < .
The mathematical expectation
k
= E(X
k
), for any integer k, is called
k th initial moment of the r.v. X. E(X)
k
is called kth central moment.
The second central moment V ar(X) =
2
= E(X)
2
is called a dispersion
or variance. The variance is often a measure of the risk, modeled by the
random variable. Another measure of the risk is the standard deviation =
_
V ar(X). Essentials in applications are the coecient of variation CV =

and the index of dispersion I =



2

. The skewness coecient is dened by the


third central moment:
1
=
E(X)
3

3
. The kurtosis coecient is
2
=
E(X)
4

4
.
The kth factorial moment is
(k)
= E[X(X 1) . . . (X k + 1)].
2.3 Transforms
2.3.1 Probability Generating Functions
Let X be a nonnegative integer-valued random variable with PMF
p
k
= P(X = k), k = 0, 1, 2, . . . .
Denition 2.3 The Probability generating function (PGF) is dened
as
P
X
(s) = Es
X
=

k=0
s
k
p
k
, (2.3)
provided the expectation is nite.
2.4. MOMENT GENERATING FUNCTIONS 13
It is easy to verify the formulas
p
k
=
1
k!
d
k
dt
k
P
X
(s)|
s=0
=
P
(k)
X
(0)
k!
and the factorial moments

(k)
= E[X(X 1) . . . (X k + 1)] =
d
k
dt
k
P
X
(s)|
s=1
= P
(k)
X
(1).
2.4 Moment Generating Functions
Let X be a real-valued random variable.
Denition 2.4 The function
M
X
(s) = Ee
sX
=
_
e
sx
dF
X
(x), (2.4)
is called a Moment generating function (MGF) of the random variable
X, whenever the expectation is nite.
Note that the MGF is nite if s = 0 and M(0) = 1. The power series of the
exponential function implies that
M
X
(s) =
_

i=0
(sx)
i
i!
dF
X
(s) = 1 +
1
s +
2
s
2
2!
+
3
s
3
3!
+ . . . ,
where

k
= EX
k
=
d
k
dx
k
M
X
(s)|
s=0
, k = 1, 2, . . .
2.5 Characteristic Functions
Denition 2.5 Let X be an R valued random variable. Then
(s) = Ee
isX
=
_

e
sx
dF
X
(x), s R
is called a characteristic function of X.
Apart from a minus sign in the exponent and the factor
1

2
, the characteristic
functions coincide with Fourier transforms in the absolutely continuous case
and with Fourier series in the lattice case.
14 CHAPTER 2. RANDOM VARIABLES
2.6 Laplace Transform
Suppose that the random variable X is nonnegative.
Denition 2.6 The function
LT
X
(s) = Ee
sX
=
_

0
e
sx
dF
X
(x), (2.5)
if it exists is called Laplace transform (LT) of the random variable X or its
distribution.
The Laplace transform is dened for complex values of s with positive
real part. For our purposes it is sucient to assume that s is a positive real
number.
The LT (2.5) of the random variable is called also Laplace - Stijeltes
transform (LST) of the distribution function F
X
(x) and actually is the LT
of the density function f
X
(x), if it exists.
Properties of the Laplace transform
Suppose that the random variable X is nonnegative with density function
f(x). Then
LT
X
(s) = LT
f
(s) =
_

0
e
sx
f(x)dx, (2.6)
where s is a positive real number. It is easy to show that the integral in (2.6)
exists, if f(x) is continuous and there exist numbers
1
,
2
and > 0, such
that, for all x > ,
|f(x)|
1
e

2
x
.
Example 2.1 Find the Laplace transform of f(x) = e
x
, x > 0. According
the denition
LT
f
(s) =
_

0
e
sx
e
x
dx =
_

0
e
(s)x
dx =
1
s
, s > .
Suppose that the LT
f
(s) (2.6) exists.
1. LT
f
(s a) =
_

0
e
sx
[e
ax
f(x)]dx;
2. e
as
LT
f
(s) =
_

a
e
sx
f(x a)dx;
3.
1
a
LT
f
(
s
a
) =
_

0
e
sx
f(ax)dx;
2.7. COUNTING RANDOM VARIABLES 15
4. sLT
f
(s) f(0) =
_

0
e
sx
f

(x)dx;
5.
1
s
LT
f
(s) =
_

0
e
sx
__
x
0
f(y)dy

dx;
6.
d
ds
LT
f
(s) =
_

0
e
sx
[xf(x)]dx;
7.
d
n
ds
n
LT
f
(s) =
_

0
e
sx
[(1)
n
x
n
f(x)]dx;
8.
_

s
LT
f
(v)dv =
_

0
e
sx
_
f(x)
x
_
dx;
9. If LT
i
(s) =
_

0
e
sx
f
i
(x)dx, i = 1, 2, then
LT
1
(s)LT
2
(s) =
_

0
e
sx
__
x
0
f
1
(y)f
2
(x y)dy
_
dx.
10. For a distribution function F with density f
LT
f
(s) = LST
F
(s) = sLT
F
(s).
2.7 Counting Random Variables
1. Bernoulli random variables. The random variable X has a Bernoulli
distribution with parameter p, if P(X = 1) = 1 P(X = 0) = p. It is easy
to check that
EX = EX
2
= p and V ar(X) = p(1 p).
2. Geometric distribution. The random variable N has a geometric
distribution with parameter p, (N Ge
1
(p)) if
P(N = k) = p(1 p)
k1
, k = 1, 2, . . . .
The mean value and the distribution function are given by
EN =
1
p
, P(N n) = 1 (1 p)
n
.
The random variable N is interpreted as the number of trials up to the rst
success in a sequence of independent Bernoulli trials. The moment generating
function is
M
N
(s) =
pe
s
1 (1 p)e
s
, s < log(1 p)
16 CHAPTER 2. RANDOM VARIABLES
and the PGF
P
N
(s) =
ps
1 (1 p)s
.
Very important property of the geometric distribution is a lack of memory:
P(N = n + k|N > k) =
P(N = n + k)
P(N > k)
=
(1 p)
n+k1
p
(1 p))
k
= P(N = n),
for any n > 0 and k > 0.
3. Binomial distribution. The random variable N has binomial dis-
tribution with parameters n and p (0, 1), (N Bi(n, p)) if
P(N = k) =
_
n
k
_
p
k
(1 p)
nk
, n = 0, 1, . . . n.
The mean value and the variance are given by
EN = np and V ar(N) = np(1 p).
The MGF is
M
N
(s) = [1 p(1 e
s
)]
n
and the PGF
P
N
(s) = [1 p(1 s)]
n
.
The binomial distributed random variable N is interpreted as the number
of successes in a sequence of n independent Bernoulli trials.
4. Poisson distribution. The random variable N has the Poisson
distribution with parameter , (N Po()) if
P(N = k) =

k
k!
e

, k = 0, 1, . . . .
The mean value and the variance of the Poisson distribution are equal:
EN = V ar(N) = .
The MGF is
M
N
(s) = e
(1e
s
)
and the PGF
P
N
(s) = e
(1s)
.
2.8. CONTINUOUS RANDOM VARIABLES 17
2.8 Continuous random variables
2.8.1 Exponential distribution and lack of memory prop-
erty
The random variable X is exponentially distributed with parameter > 0,
(X exp ()), if
F(x) = 1 e
x
and f(x) = e
x
, x 0.
The MGF of the exponential distribution is
M
X
(s) = Ee
sX
=

s
, t < (2.7)
and the Laplace transform
LT
X
(s) =

+ s
.
All the moments of the random variable could be determine by dieren-
tiation of (2.7).
EX =
1

, and V ar(X) =
1

2
and
EX
n
=
n!

n
.
If = 1, the r.v. X exp(1) is called a standard exponentially dis-
tributed. Let X exp(1) and Y = a +
X

, > 0, < a < . Then


the distribution of Y is determined by the Distribution function and the
probability density function
F(x) = 1 e
(xa)
and f(x) = e
(xa)
, x a
and is noted by Y exp(a, ).
The advantages in applications of the exponential distribution is the lack
of memory property.
The random variable X satises the lack of memory property, if for
any t, s 0 such that P(X t) > 0
P(X t + s|X t) = P(X s). (2.8)
Let X be the lifetime distribution. Then (2.8) means the probability
that the individual will survive another s years, after having attained the
18 CHAPTER 2. RANDOM VARIABLES
age t is equal to the probability that the individual will survive age s. If
the distribution function is not degenerate at zero, the condition (2.8) is
equivalent to
P(X t + s, X t)
P(X t)
= P(X s)
or
P(X t + s) = P(X s)P(X t). (2.9)
The equation (2.9) contains also the degenerated case. The decision is given
by the following theorem [5].
Theorem 2.1 There are only two solutions of the equation (2.9) among dis-
tribution functions. Either F(x) is degenerate at zero, or, with some constant
> 0, F(x) = 1 e
x
, x 0.
Since e
(t+s)
= e
t
e
s
, the only exponentially distributed random vari-
ables are memoryless. If the lifetime distribution is exponential the hazard
rate function is constant:
(t) =
e
t
e
t
= .
Let X and Y be independent exponentially distributed random variables
with respective parameters and . Then the distribution of Z = min(X, Y )
is given by
P(Z z) = 1 P(Z > z) = 1 P(X > z, Y > z) =
= 1 P(X > z)P(Y > z) = 1 e
z
e
z
= 1 e
(+)z
.
Consequently, Z = min(X, Y ) is again exponentially distributed with
parameter + .
2.8.2 Gamma distribution
Let () =
_

0
x
1
e
x
dx, > 0 be the Gamma function.
Properties of the Gamma function
() = ( 1)( 1),
2.8. CONTINUOUS RANDOM VARIABLES 19
(n) = (n 1)!, n 1
and

_
1
2
_
=

.
The random variable X is Gamma distributed with parameters > 0
and > 0, (X (, )), if the density function is given by
f(x) =

()
x
1
e
x
, x > 0.
The parameter is called a shape parameter. Small value of results in a
long tail to the right, or right skewed distribution.
In the special case integer, it is an Erlang distribution.
The MGF is
M(s) =
_
1
s

, t >
and the n th moment
EX
n
=
( + n)
()
n
.
In particular
EX =

and V ar(X) =

2
.
2.8.3 Beta distribution
For > 0 and > 0, the function
B(, ) =
_
1
0
x
1
(1 x)
1
dx
is called a Beta function. The most important property of the Beta function:
B(, ) =
()()
( + )
.
The random variable X is Beta distributed (X B(, )) with parame-
ters and , if the density function is
f(x) =
1
B(, )
x
1
(1 x)
1
, x (0, 1).
The moments of the Beta distributed random variable:
EX
n
=
B(n + , )
B(, )
=
(n + )( + )
(n + + )()
.
20 CHAPTER 2. RANDOM VARIABLES
2.8.4 Weibull distribution
The random variable X has Weibull distribution with parameters > 0 and
> 0, (X W(, )) if the distribution function and density function are
F(x) = 1 e
(
x

, x 0
and
f(x) =

_
x

_
1
e
(
x

, x 0.
The mean value and the variance:
EX = (1 +
1

), V ar(X) =
2
_
(1 +
2

)
_
(1 +
1

)
_
2
_
.
If = 1, W(1, ) = exp().
2.8.5 Pareto distribution
The random variable X has a Pareto distribution with parameters > 0 and
> 0, (X Par(, )), if the distribution function is given by
F(x) = 1
_

+ x
_

, x > 0.
The density function is
f(x) =

( + x)
+1
, x > 0.
The distribution dened by the density
f
1
(x) = f(x ) =

x
+1
, x >
is called a shifted Pareto distribution.
The moments of the random variable with density f(x) are
m
k
=

_

0
x
k
_

+ x
_
+1
dx.
The change of variables y =

+x
leads to
m
k
=
_
1
0

_
(1 y)
y
_
k
y
+1
y
2
dy =
k
_
1
0
y
k1
(1 y)
k
dy,
2.8. CONTINUOUS RANDOM VARIABLES 21
which is a Beta function and
m
k
=
k
( k)(k + 1)
( + 1)
=
k
k!
( k)
()
, > k.
The mean value and the variance are
EX = m
1
=

1
, > 1
and
V ar(X) =

2
( 1)
2
( 2)
, > 2.
It is easy to see that only nite number of moments exist. Even the mean
and the variance not always exist. This means that the Pareto distribution
is a heavy tailed.
2.8.6 The normal distribution
The normal distribution is important in insurance and nance since it appears
like limiting distribution in many cases. The normal density function is
f(x) =
1

2
exp
_

(x )
2
2
2
_
, < x < .
The notation X N(, ) means that X has a normal distribution with
parameters and .
The mean value and the variance are EX = and V ar(x) =
2
. The
most important property is that the random variable Z =
X

is normally
distributed with parameters 0 and 1 (Z N(0, 1)). The distribution function
is denoted and
(x) =
1

2
_
x

x
2
2
dx.
The characteristic function of X is

X
(s) = e
is
s
2

2
2
, s R
and the MGF
M
X
(s) = e
s
s
2

2
2
.
22 CHAPTER 2. RANDOM VARIABLES
2.8.7 Log-normal distribution
The random variable X has a log-normal distribution, if Y = log X has a
normal distribution. If Y N(, ), the density function of X is
f
X
(x) = f
Y
(log x)
1
x
=
1
x

2
e

1
2
(
log x

)
2
, x > 0,
where > 0, < < . The distribution function of X is dened by
F(x) =
_
log x

_
, x > 0,
where (.) is the standard normal distribution function. The moments of
the log-normal distribution can be calculated from the standard normal dis-
tribution function.
m
k
= EX
k
= E(e
kY
) = M
Y
(k) = e
k+

2
k
2
2
,
where M
Y
(k) is the MGF of the normal distribution. Particular, the mean
and variance are
EX = m
1
= e
+

2
2
and
V ar(X) = e
2+
2
[e

2
1].
2.8.8 Inverse Gaussian distribution
If the random variable X is dened by the density
f(x) =

_
2x
3
e

(x)
2
2x
, x 0, (2.10)
where > 0, then X has Inverse Gaussian distribution (X IG(, )).
In the special case of = 1, (2.10) is called Wald distribution. The In-
verse Gaussian distribution has greater skewness and a sharper peak then
Gaussian.
The distribution function is
F(x) =
_
x

x
_
+ e
2

x +

x
_
, x > 0.
The Laplace transform:
LT(s) =
_

0
e
sx

_
2x
3
e

(x)
2
2x
dx = e

[1

1+2s]
, s
1
2
.
The mean value and the variance:
EX = and V ar(X) = .
2.9. FUNCTIONS OF RANDOM VARIABLES 23
2.9 Functions of random variables
Theorem 2.2 Let X be a continuous random variable and (x) is a strictly
increasing function, dened on the set of values of X. Let Y = (X) and
F
X
and F
Y
are the distribution functions of X and Y . Then
F
Y
(y) = F
X
(
1
(y)).
If (x) is strictly decreasing over the set of values of X, then
F
Y
(y) = 1 F
X
(
1
(y)).
Proof. Since (x) is strictly increasing over the set of values of X, the events
{X
1
(y)} and {(X) y} are equivalent. Consequently
F
Y
(y) = P(Y y) = P((X) y) = P(X
1
(y)) = F
X
(
1
(y)).
If (x) is strictly decreasing, then
F
Y
(y) = P(Y y) = P((X) y) =
= P(X >
1
(y)) = 1 P(X
1
(y)) = 1 F
X
(
1
(y)).
2
If f
X
and f
Y
are the density functions of X and Y and (x) is strictly
increasing, then
f
Y
(y) = f
X
(
1
(y))
d
dy

1
(y).
If (x) is strictly decreasing, then
f
Y
(y) = f
X
(
1
(y))
d
dy

1
(y).
The method is applicable in same cases even if the function (x) is neither
increasing nor decreasing. For example, if Y = X
2
, the function is (x) = x
2
.
Then
F
Y
(y) = P(Y y) = P(

y < X

y)
= P(X

y) P(X

y) = F
X
(

y) F
X
(

y).
For the density we obtain
f
Y
(y) =
d
dy
F
Y
(y) =
d
dy
(F
X
(

y) F
X
(

y)) = (f
X
(

y) + f
X
(

y))
1
2

y
.
24 CHAPTER 2. RANDOM VARIABLES
2.10 Joint density and distribution function
Let X = (X, Y ) be a random vector.
Denition 2.7 The joint distribution function of X is the function H :
R
2
[0, 1], dened by
H(x, y) = P(X x, Y y), (x, y) R
2
.
If X and Y absolutely are continuous random variables, the joint density
function satises the equation
H(x, y) =
_
x

_
y

f(t
1
, t
2
)dt
2
dt
1
.
Consequently
f(x, y) =

2
H(x, y)
xy
.
If the random variables X and Y are mutually independent
H(x, y) = F
X
(x)F
Y
(y).
The random variables X and Y are called marginal variables and the
distribution functions F
X
and F
Y
- marginal distributions of (X, Y ). The
joint distribution function is right continuous and satises the properties:
1) lim
y
H(x, y) = F
X
(x) and lim
x
H(x, y) = F
Y
(y),
2) lim
(x,y)(,)
H(x, y) = 1, where (x, y) (, ) means that both
variables x and y tend to innite,
3) lim
x
H(x, y) = lim
y
H(x, y) = 0,
4) for all (x
1
, x
2
) and (y
1
, y
2
), such that x
1
< x
2
and y
1
< y
2
,
H(x
2
, y
2
) H(x
1
, y
2
) H(x
2
, y
1
) + H(x
1
, y
1
) 0.
For the random variables X and Y, the second moment
Cov(X, Y ) = E[(X EX)(Y EY )] = E(XY ) (EX)(EY ),
is called a covariance. Note that the variance V ar(X) of a square integrable
random variable is Cov(X, X). The Pearson correlation coecient is one of
the most useful measures of dependence and is dened by
= Corr(X, Y ) =
Cov(X, Y )
_
V ar(X)
_
V ar(Y )
.
The random variables X and Y are positively correlated if Cov(X, Y ) > 0,
and negatively correlated if Cov(X, Y ) < 0.
2.11. CONDITIONAL DISTRIBUTIONS 25
2.11 Conditional distributions
If X and Y are dened over the same probability space, then the conditional
distribution function of X given Y = y is dened by
F
X|Y
(x|y) = P(X x|Y = y). (2.11)
The following equality gives the relation between the conditional distribution
function and the distribution function of the random variable X
F
X
(x) =
_
F
X|Y
(x|y)dF
Y
(y).
The similar relation exists between the density of X and the conditional
density
f
X
(x) =
_
f
X|Y
(x|y)f
Y
(y)dy.
For discrete random variables:
P(X = x
k
) =

i=0
P(X = x
k
|Y = y
i
)P(Y = y
i
), k = 0, 1, 2, . . . .
If X and Y are independent random variables:
F
X|Y
(x|y) = F
X
(x).
According (2.11) the conditional mean value of X given Y = y:
E(X|Y = y) =
_

0
xdF
X|Y
(x|y),
provided that the random variables are nonnegative. Consequently E(X|Y )
is a random variable. For independent random variables
E(X|Y = y) = EX.
For any function h(x), the conditional mean value of h(X) given Y = y
is:
E[h(X)|Y = y] =
_
h(x)dF
X|Y
(x|y). (2.12)
From (2.12) it follows that
E[h(X)] = E
Y
E
X
[h(X)|Y ],
where on right is taken mathematical expectation of X given x Y and after
that mathematical expectation of Y .
26 CHAPTER 2. RANDOM VARIABLES
Exercise 2.1 Prove that for the nonnegative random variable X :
a) E[E(X|Y )] = EX;
b) V ar(X) = E[V ar(X|Y )] + V ar[E(X|Y )].
2.12 Sum of random variables
Let X and Y be independent random variables. The distribution function of
the sum Z = X + Y, noted
F
Z
(z) = F
X
F
Y
(z)
is called a convolution of F
X
and F
Y
and is dened by
F
Z
(z) =
_
P(Z z|Y = y)dF
Y
(y)
=
_
P(X z y|Y = y)dF
Y
(y)
=
_
F
X|Y
(z y|y)dF
Y
(y) =
_
F
X
(z y)dF
Y
(y).
(2.13)
If X and Y are continuous random variables, the density of Z is obtained
by dierentiation of (2.13) relative to z:
f
Z
(z) = f
X
f
Y
(z) =
_
f
X
(z y)f
Y
(y)dy.
If X and Y are discrete random variables
P(Z = k) =

i=0
P(X = k i)P(Y = i), k = 0, 1, 2, . . . .
In the multivariate case, if X
1
, X
2
, . . . X
n
are random variables, the dis-
tribution of the sum S
n
= X
1
+X
2
+. . . +X
n
is obtained by recursions. Note
S
1
= X
1
and S
j
= S
j1
+ X
j
, j = 2, 3, . . . n. The Laplace transform of the
sum is
LT
S
n
(t) =
n

j=1
LT
X
j
(t), (2.14)
the distribution function is noted by
F
S
n
(x) = F
X
1
F
X
2
. . . F
X
n
(x)
and the density
f
S
n
(x) = f
X
1
f
X
2
. . . f
X
n
(x).
2.12. SUM OF RANDOM VARIABLES 27
If X
1
, X
2
, . . . X
n
are identically distributed with distribution function F
X
(x)
and density function f
X
(x), then
F
S
n
(x) = F
n
X
(x)
and
f
S
n
(x) = f
n
X
(x).
Here n denotes n th convolution of X. In that case (2.14) implies that the
transforms of S
n
= X
1
+X
2
+. . . +X
n
are nth power of the corresponding
transform of X. For example, the PGF is
P
S
n
(t) = [P
X
(t)]
n
.
2.12.1 Negative binomial distribution
Let X
1
, X
2
, . . . X
r
be independent Ge
1
(p) distributed random variables. The
random variable N = X
1
+ X
2
+ . . . X
r
has probability mass function
p
k
=
_
k 1
r 1
_
p
r
(1 p)
kr
, k = r, r + 1, . . .
and is called negative binomial distributed (N NB(r, p)). The interpreta-
tion of the NB distributed random variable is the number of trials up to the
rth success in a sequence of Bernoulli trials. The PGF is given by
M
N
(t) =
_
pe
t
1 (1 p)e
t
_
r
, t < ln(1 p).
The mean value and the variance:
EN =
r
p
V ar(N) =
r(1 p)
p
2
.
An alternative representation of the NBD is by the random variable Y -the
number of failures in Bernoulli trials until r successes. In this case Y = Nr
and the distribution function is
p
k
=
_
r + k 1
k
_
p
r
(1 p)
k
, k = 0, 1, 2, . . .
The MGF is
M
Y
(t) =
_
p
1 (1 p)e
t
_
r
, t < ln(1 p).
The mean value and the variance are
EY =
r(1 p)
p
V ar(Y ) =
r(1 p)
p
2
.
28 CHAPTER 2. RANDOM VARIABLES
2.12.2 Erlang distribution
Let X
1
, X
2
, . . . X
n
be independent, exponentially distributed with parameter
. Find the distribution of the random variable S
n
= X
1
+ . . . + X
n
. It is
known that the exponential distribution is (1, ).
Suppose that
f
X
1
+...+X
n1
(t) = e
t
(t)
n2
(n 2)!
.
For S
n
we obtain
f
X
1
+...+X
n
(t) =
_

0
f
X
n
(t s)f
X
1
+...+X
n1
(s)ds
=
_

0
e
(ts)
e
s
(s)
n2
(n 2)!
ds =
(t)
n1
(n 1)!
e
t
.
Denition 2.8 The random variable with density function
f
X
(t) =
(t)
n1
(n 1)!
e
t
, t 0, (2.15)
is called Erlang distributed random variable with parameters n and , X
Erl(n, ), n = 1, 2, . . . .
The distribution function is
F(t) = 1
k1

j=0
(t)
j
j!
e
t
, t 0. (2.16)
The is a location parameter, n a shape parameter.
The mean value and the variance are given by
EX =
n

and V ar(X) =
n

2
.
The Laplace - Stieltjes transform:
LS
X
(t) =
_

+ t
_
n
.
It is known that (2.15), with real n is the density of Gamma distributed
random variable (X (n, ).)
2.13. MIXED DISTRIBUTIONS 29
2.13 Mixed distributions
Let X be a random variable with distribution function F
X
(x, ) and density
function f
X
(x, ), where is a parameter.
Suppose that is a realization of a continuous, positive random variable
. For xed we use the notation F
X
(x|), i.e. conditional distribution
function. Let U() and u() be the distribution function and density function
of the random variable . The distribution function of X is dened by
F
X
(x) =
_

F
X
(x|)dU() = E

F
X
(x|), (2.17)
where E

denotes the mathematical expectation relative .


Denition 2.9 The random variable X with a distribution (2.17) is called
mixed by mixing distribution U().
The equation (2.17) gives the unconditional distribution of X. Uncondi-
tional density function of X is dened by
f
X
(x) =
_

f
X
(x|)dU() = E

f
X
(x|).
The moments:
EX
k
=
_

E(X
k
|)dU() = E

E(X
k
|), k = 1, 2, . . .
The unconditional transforms are obtained by the mathematical expec-
tation over the set of all possible values of :
M
X
(z) = E

M
X
(z|)
P
X
(z) = E

P
X
(z|)

X
(z) = E

X
(z|)
LT
X
(z) = E

LT
X
(z|).
Example 2.2 (Poisson mixture) Let X be a Poisson distributed random
variable with MGF
M
X
(s|) = e
(e
s
1)
.
Then, for any distribution (), the unconditional MGF of X is
M
X
(s) = E

M
X
(s|) = E

[e
(e
s
1)
] = M

(e
s
1),
wich is the MGF of with the argument e
s
1.
30 CHAPTER 2. RANDOM VARIABLES
Example 2.3 (NBD) Let X be a Poisson distributed random variable with
probability mass function
P(X = k|) =

k
k!
e

, k = 0, 1, . . . .
The parameter is a realization of Gamma distributed random variable with
density function f() =

r
(r)

r1
e

, > 0, > 0, where is the Gamma


function, r is the shape parameter and the scale parameter. Prove that
P(X = k) =
_

1 +
_
r
_
r + k 1
k
__
1
1 +
_
k
, k = 0, 1, . . . .
The following construction is again a mixture. Let p
1
, p
2
, . . . be a sequence
of nonnegative numbers, such that

i=1
p
i
= 1. Let F
1
(x), F
2
(x), . . . be a
sequence of distribution functions. Then
F
X
(x) =

i=1
p
i
F
i
(x) (2.18)
is again a distribution function and is called a mixture.
Here the weights {p
i
} are not related to any parameter.
Example 2.4 Let us construct the mixture between the degenerate at zero
distribution
F
1
(x) =
_
0, x < 0
1, x 0
and the exponential distribution
F
2
(x) = 1 e
x
, x > 0
by weights and 1 . The distribution function is given by
F
X
(x) = .1 + (1 )(1 e
x
) = 1 (1 )e
x
, x 0.
The Laplace transform of X is
LT
X
(s) = LT
1
(s) + (1 )LT
2
(s) = + (1 )

+ s
.
If all F
i
(x) in (2.18) are the same, the mixture coinsides with (2.17) in
the case of discrete parameter .
2.14. COMPOUND DISTRIBUTIONS 31
2.14 Compound distributions
Let X
1
, X
2
, . . . be a sequence of independent identically distributed random
variables with distribution function F
X
(x), characteristic function
X
(z),
mean value and variance
2
X
. Then the sum X
1
+X
2
+. . . +X
n
, n 1 has
a distribution function F
n
X
(x), characteristic function [
X
(z)]
n
, mean value
n
X
and variance n
2
X
.
Consider the sum
S
N
= X
1
+ X
2
+ . . . X
N
, (2.19)
where N is a discrete random variable and S
N
= 0 if N = 0. Then the the
distribution function of S
N
is
F
S
N
(x) = P(S
N
x) =

n=0
P(S
N
x|N = n)P(N = n) =

n=0
P(N = n)F
n
X
(x),
where
F
0
X
x) =
_
0, x < 0
1, x 0.
Denition 2.10 The distribution of the random sum (2.19) is called com-
pound distribution.
The MGF of the compound distribution is given by
M
S
N
(s) = Ee
sS
N
=

n=0
E[e
s(X
1
+...X
n
)
|N = n]P(N = n)
=

n=0
P(N = n)[M
X
(s)]
n
= E[[M
X
(s)]
N
],
or
M
S
N
(s) = P
N
(M
X
(s)).
Analogously
P
S
N
(s) = P
N
(P
X
(s))

S
N
(s) = P
N
(
X
(s))
LT
S
N
(s) = P
N
(LT
X
(s)).
Exercise 2.2 Show that
ES
N
= E(N)E(X) (2.20)
and
V ar(S
N
) = V ar(N)(EX)
2
+ E(N)V ar(X). (2.21)
32 CHAPTER 2. RANDOM VARIABLES
If N has a Poisson distribution, (2.19) is called compound Poisson distri-
bution.
Theorem 2.3 Let S
i
CPo(
i
, F
i
(x)), i = 1, . . . n be independent Com-
pound Poisson random variables with parameters
i
and Z F
i
(x), x > 0.
Show that S
n
= S
1
+ . . . S
n
is also Compound Poisson with parameters
=

n
i=1

i
and F(x) =

n
i=1

F
i
(x).
Example 2.5 (Compound geometric - exponential distribution) Let X
1
, X
2
, . . .
be independent, identically exponentially distributed with parameter > 0,
density function
f
X
(x) = e
x
x > 0
and
LT
X
(s) =

+ s
.
The Laplace transform of S
N
is
LT
S
(s) =

n=0
P(N = n)
_

+ s
_
n
.
Suppose that N has a geometric distribution with PMF
P(N = n) = p(1 p)
n
, n = 0, 1, 2, . . .
and PGF
P
N
(s) =
p
1 (1 p)s
.
In that case the LT of S
N
is
LT
S
(s) = p
_
1 (1 p)

+ s
_
1
= p
+ s
p + s
= p + (1 p)
p
p + s
.
The LT of S
N
is a mixture of the Laplace transform of degenerate at zero
random variable and the Laplace transform of exp(p) random variable.
According the properties of the LT, the density function of S
N
is
f
S
N
(x) =
_
p, x = 0
(1 p)pe
px
, x > 0
and the distribution function
F
S
N
(x) = 1 (1 p)e
px
, x 0.
2.14. COMPOUND DISTRIBUTIONS 33
Example 2.6 (Polya - Aeppli distribution). Let X
1
, X
2
, . . . be independent
identically Ge
1
(1 ) distributed random variables with parameter [0, 1)
and probability mass function
P(X
1
= i) =
i1
(1 ), i = 1, 2, . . . .
The random variable N Po() is independent of X
i
, i = 1, 2, . . . . The
probability mass function of S
N
= X
1
+ X
2
+ . . . + X
N
is given by
P(S
N
= k) =
_

_
e

, k = 0,
e

i=1
_
k 1
i 1
_
[(1 )]
i
i!

ki
, i = 1, 2, . . .
(2.22)
Denition 2.11 The distribution, dened by the PMF (2.22) is called a
Polya - Aeppli distribution.
2.14.1 Hiperexponential distribution
Let X
i
exp(
i
), i = 1, 2,
1
=
2
be independent. The distribution of the
sum X
1
+ X
2
is given by
f
X
1
+X
2
=
_
t
0
f
X
1
(s)f
X
2
(t s)ds =
_
t
0

1
e

1
s

2
e

2
(ts)
ds
=
1

2
e

2
t
_
t
0
e
(
1

2
)s
ds =

2

1
e

1
t
+

1

2
e

2
t
.
It can be proved by induction that the density function of S
n
= X
1
+
. . . + X
n
, where X
i
exp(
i
), i = 1, . . . , n,
i
=
j
for i = j, is
f
S
n
(t) =
n

i=1
p
i

i
e

i
t
, (2.23)
p
i
=
j=i

j

i
.
Denition 2.12 The random variable dened by density function (2.23) for
some weights p
i
,

n
i=1
p
i
= 1 is called hiperexponential distributed random
variable.
34 CHAPTER 2. RANDOM VARIABLES
The notation is H
n
(p
1
, . . . p
n
;
1
, . . .
n
) or H
n
. That is a random variable
equal to X
i
with probability p
i
, i = 1, . . . n. The distribution (2.23) is a
mixture of exponentially distributed random variables. The mean value is
ES
n
=
n

i=1
p
i

i
and the Laplace transform
LT
S
n
(s) =
n

i=1
p
i

i
+ s
.
Chapter 3
Counting processes
The stochastic process {N(t), t 0} is called a counting process, if N(t) is
equal to the number of events occurred up to time t.
The counting process satises the following:
1. N(t) 0;
2. N(t) has integer values;
3. If s < t, then N(s) N(t);
4. For s < t, N(t) N(s) is the number of events in the time interval
(s, t).
The counting process is called a process with independent increments if
the number of events in disjoint intervals are independent random variables.
The counting process has stationary increments if the distribution of the
number of events occurred in a given time interval depends only on the
length of the interval. This means that for t > 0 and h > 0 the distribution
of N(t + h) N(t) coincides with the distribution of N(h).
3.1 Poisson process
One of the basic counting processes is the Poisson process.
Denition 3.1 The counting process {N(t), t 0} is called a Poisson pro-
cess with intensity rate > 0, if
1. N(0) = 0;
2. The process has stationary independent increments;
3. P(N(h) = 1) = h +(h);
4. P(N(h) 2) = (h).
(The function f is called (h), if lim
h0
f(h)
h
= 0.) It is clear that the
process is integer valued and nondecreasing.
35
36 CHAPTER 3. COUNTING PROCESSES
Notation: P
n
(t) = P(N(t) = n). We will show that the following system
of dierential equations follows from the postulates in the denition.
P

0
(t) = P
0
(t)
P

n
(t) = P
n
(t) + P
n1
(t), n = 1, 2, . . .
(3.1)
For n = 0
P
0
(t + h) = P(N(t) = 0, N(t + h) N(t) = 0) =
= P(N(t) = 0)P(N(t + h) N(t) = 0) = P
0
(t)[1 h +(h)]
and
P
0
(t + h) P
0
(t)
h
= P
0
(t) +
(h)
h
.
By h 0 we obtain the rst equation of (3.1).
Let n 1
P
n
(t + h) = P(N(t) = n, N(t + h) N(t) = 0)
+P(N(t) = n 1, N(t + h) N(t) = 1)+
+

k=2
P(N(t) = n k, N(t + h) N(t) = k).
According the fourth postulate, the sum is (h). Consequently
P
n
(t+h) = P
n
(t)P
0
(h)+P
n1
(t)P
1
(h)+(h) = (1h)P
n
(t)+hP
n1
(t)+(h),
or
P
n
(t + h) P
n
(t)
h
= P
n
(t) + P
n1
(t) +
(h)
h
.
For h 0 we obtain the second equation of (3.1).
To solve (3.1), we use the method of integrating factor. The condition
N(0) = 0 implies the initial conditions
P
0
(0) = 1 and P
n
(0) = 0, n = 1, 2, . . . . (3.2)
The solution of the rst equation of (3.1) is given by
P
0
(t) = C
0
e
t
.
Together with (3.2) this leads to
P
0
(t) = e
t
. (3.3)
3.1. POISSON PROCESS 37
Inserting (3.3) in the second equation of (3.1) for n = 1 yield
P

1
(t) + P
1
(t) = e
t
with the solution
P
1
(t) = (t + C
1
)e
t
.
By the initial condition (3.2) for n = 1, the solution is
P
1
(t) = te
t
.
In the general case we suppose that
P
k
(t) =
(t)
k
k!
e
t
, k = 1, 2, . . . , n 1.
By the second equation of (3.1)
P

n
(t) + P
n
(t) = P
n1
(t) =
(t)
n1
(n 1)!
e
t
with the solution
P
n
(t) =
_
(t)
n
n!
+ C
n
_
e
t
.
Together with the initial condition this gives the solution
P
n
(t) =
(t)
n
n!
e
t
, n = 1, 2, . . . (3.4)
The Poisson process: the number of events in (0, t] has a Poisson
distribution with parameter t, i.e. EN(t) = t.
This leads to the second denition.
Denition 3.2 The counting process {N(t), t 0} is called a Poisson pro-
cess with intensity rate > 0, if
1. N(0) = 0;
2. The process has independent increments;
3. For s < t, the number of claims in the interval (s, t] has a Poisson
distribution with parameter (t s) :
P(N(t) N(s) = n) =
((t s)
n
n!
e
(ts)
, n = 1, 2, . . . .
38 CHAPTER 3. COUNTING PROCESSES
By the law of large number or by the Chebyshevs inequality, it follows
that
N(t)
t
P
, as t . In fact, almost sure convergence holds. This
means that the intensity measures the average frequency or density of claim
arrivals.
According the second denition, if h > 0 is small
P(N(t + h) N(t) = 1) = he
h
= h h(1 e
h
) = h + o(h),
as h 0. The last equality follows from the inequality 0 < 1 e
x
< x, for
x > 0.
Furthermore, for h (0,
1
2
)

n=2
(h)
n
n!

1
2

n=2
(h)
n
=
1
2
(h)
2
1 h
(h)
2
.
This implies that, as h 0
P(at least 2 claims arrive in(t, t + h]) =

n=2
(h)
n
n!
e
h
(h)
2
= o(h),
which is the fourth postulate of the rst denition.
In order to prove that the denitions 3.1 and 3.2 are equivalent, we have
to show that the increments N(t) N(s) Po((t s)), 0 s < t. It
follows from the stationarity condition.
The interarrival times
Let T
1
be the time until the rst claim. For n 2, T
n
is the time between
the (n 1)st and the nth claim. T
2
, T
3
, . . . is a sequence of the interarrival
times.
The distribution of T
n
: Note that the event {T
1
> t} occurs if and
only if, no one event occurs up to time t and
P(T
1
> t) = P(N(t) = 0) = e
t
.
Hence T
1
is exponentially distributed with parameter . Given T
1
, the dis-
tribution of T
2
is
P(T
2
> t) = P(T
2
> t|T
1
= s) =
= P(0 claims in(s, s + t]|T
1
= s) = P(0 claims in(s, s + t]) = e
t
,
(3.5)
where the last two equalities follow from the conditions of independent, sta-
tionary increments. From (3.4) it follows that T
2
is independent of T
1
and is
exponentially distributed with parameter . By the same arguments and by
induction it could be prove the following.
3.1. POISSON PROCESS 39
Theorem 3.1 The interarrival times T
n
, n = 1, 2, . . . are independent, iden-
tically exp() distributed random variables.
Exercise 3.1 Prove that the waiting time
k
= T
1
+. . . +T
k
(k, ), k =
1, 2, . . . , i.e.
f

k
(t) =

(k)
(t)
k1
e
t
, t > 0
and
F

k
(t) = 1
k1

j=0
(t)
j
j!
e
t
.
The properties of Theorem 3.1 and Exercise 3.1 characterize the Poisson
process. This states in the following
Theorem 3.2 Let N(t), t 0 be a stochastic process with N(0) = 0. T
1
is
the time until the rst claim, T
2
, T
3
, . . . the interarrival times. Let T
k
, k =
1, 2, . . . be independent, identically exp() distributed random variables and
N(t) = the number of claims up to time t. Then N(t) is a Poisson process.
Proof. Follows from
P(
k
t) = P(N(t) k), k = 0, 1, . . .
and the lack of memory property of the exponential distribution.
2
The index of dispersion:
I(t) =
V ar(N(t))
EN(t)
= 1.
3.1.1 Order statistics property
Let U
1
, . . . U
n
be independent uniformly distributed random variables (U
i

U([a, b]). The order statistics of this sample are
U
(1)
U
2
. . . U
(n)
,
where U
(n)
= max{U
i
, i = 1, . . . , n}, U
(1)
= min{U
i
, i = 1, . . . n}. The joint
density function of (U
(1)
, . . . U
(n)
) is given by
f(u
1
, . . . u
n
) =
n!
(b a)
n
I
{a<u
1
...<u
n
<b}
and is called the Dirichlet distribution (D
n
([a, b]).
40 CHAPTER 3. COUNTING PROCESSES
Theorem 3.3 Let T
1
, . . . , T
n
be independent exp() distributed random vari-
ables and
k
= T
1
+ . . . T
k
, k = 1, . . . n. Then
a) (
1
, . . .
n
) has a probability density in R
n
given by
f

1
,...,
n
(t
1
, . . . t
n
) =
n
e
t
n
(t
1
, . . . , t
n
), t
1
< t
2
< . . . < t
n
. (3.6)
b)The joint conditional density of (
1
, . . .
n
), given that N(t) = n is
f

1
,...,
n
|N(t)=n
(t
1
, . . . t
n
) =
_
_
_
n!
t
n
, t
1
< t
2
< . . . < t
n
0, otherwise.
(3.7)
Proof. a) The joint distribution of T
1
, . . . , T
n
is given by
f
T
1
,...,T
n
(x
1
, . . . x
n
) =
n
n

i=1
e
x
i
=
n
e

n
i=1
x
i
, x
i
0.
Change of variables
t
1
= x
1
x
1
= t
1
t
2
= x
1
+ x
2
x
2
= t
2
t
1
t
3
= x
1
+ x
2
+ x
3
and x
3
= t
3
t
2
. . . . . .
t
n
= x
1
+ . . . x
n
x
n
= t
n
t
n1
yields (3.6).
b)
P(
1
t
1
, . . .
n
t
n
|N(t) = n) =
P(
1
t
1
, . . .
n
t
n
, N(t) = n)
P(N(t) = n)
=
_
. . .
_
P(N(t) N(s
n
) = 0)f

1
,...
n
(s
1
, . . . s
n
)ds
1
. . . ds
n
P(N(t) = n)
=
_
t
1
0
_
t
2
s
1
. . .
_
t
n
s
n1
e
(ts
n
)

n
e
s
n
ds
n
ds
n1
. . . ds
1
(t)
n
n!
e
t
=
_
t
1
0
_
t
2
s
1
. . .
_
t
n
s
n1
n!
t
n
ds
n
ds
n1
. . . ds
1
.
2
3.2. RENEWAL PROCESS 41
3.2 Renewal process
In terms of point processes: Suppose that 0 =
0

1
. . . are random
points in R
+
, at which a certain event occurs. The counting process is dened
by the number of points in the interval (0, t] and is given by
N(t) =

n=1
I
{
n
t}
, t 0.
Assume this counting process has nite values for each t. This is equivalent
to
n
a.s. as n . The point counting process N(t) is simple if the
occurrence times are distinct: 0 <
1
<
2
< . . . a.s.
Denition 3.3 A simple point process {
n
, n 1} is a renewal process if
the inter-arrival times T
n
=
n

n1
, n 1 are independent, identically
distributed with common distribution function F
T
, F
T
(0) = 0 and
0
= 0.
The points {
n
} are called renewal times.
Suppose that the time between claims are not concentrated at zero, i.e.
P(T = 0) < 1. Let
N(t) = max{n,
n
t} = min{n,
n+1
> t}, (3.8)
or N(t) is equal to the number of renewals in the interval (0, t].
Denition 3.4 The process (3.8) is called a renewal counting process.
The equivalence of the processes {
n
} and {N(t)} follows from the equiva-
lence of the events
{N(t) = n} and {
n
t <
n+1
} (3.9)
and
{N(t) n} = {
n
t}, n = 1, 2, . . . . (3.10)
Note that {N(t), t 0} is dened in continuous time and the sample
paths are continuous on the right.
For the risk model, again (
n
) is a sequence of arrival times and (T
n
) -
the sequence of inter-arrival times.
42 CHAPTER 3. COUNTING PROCESSES
Example 3.1 (Homogeneous Poisson process) Let T
n
exp(). The
renewal process is called a homogeneous Poisson process with intensity . In
this case, by the convolution properties
P(
n
t) = F
n
T
(t) =
_
t
0
e
x
(x)
n1
(n 1)!
dx.
This is gamma distribution with parameters n and and
P(
n
t) = 1
n1

k=0
(t)
k
k!
e
t
.
Using the relation (3.10) we arrive at
P(N(t) n) = 1 P(N(t) n + 1) = 1 P(
n+1
t) =
n

k=0
(t)
k
k!
e
t
.
The motivation for introducing the renewal counting process is that the
homogeneous Poisson process does not describe in an adequate way the claim
arrivals. In many cases it is more natural to model the inter-arrival times by
log-normal or Pareto distribution.
In general the Poisson process is a special case of renewal process, but
many of the asymptotic properties are the same.
Suppose that the inter-arrival times T are dened by the probability
distribution function F
T
(x) and ET =
1

. In some cases the distribution F


T
is defective. This means that lim
x
F
T
(x) < 1, or the random variable
T can be equal to with positive probability 1 F(). Here F() =
lim
x
F(x). We will call in this case the renewal process terminating. We
will show that the limit N() = lim
t
N(t) < with probability 1 and

N()
has a compound geometric distribution.
Theorem 3.4 (Strong law of large numbers) If F() = 1, then the
samples of {N(t), t 0} are increasing with probability 1 and
lim
t
N(t)
t
= = (ET)
1
a.s. (3.11)
Proof. The limit lim
t
N(t) exists, since the sample paths of the process
are increasing.
From (3.9) it follows that lim
t
N(t) = with probability 1. The
strong law of large numbers for (T
n
) yields that lim
n

n
n
=
1

> 0 with
probability 1.
3.2. RENEWAL PROCESS 43
Consequently lim
n

N(t)
N(t)
=
1

with probability 1. Again from (3.9) it


follows that
N(t)
t <
N(t)+1
and

N(t)
N(t)

t
N(t)
<

N(t)+1
N(t) + 1
N(t) + 1
N(t)
.
Let t and obtain (3.10) in the case of > 0. When = 0, ET = .
Note that
N() = lim
t
N(t) = min{n : T
n
= } 1,
i.e. N() is geometrically distributed with parameter P(N() = 0) =
1F(). From the total probability law it follows that the random variable

N()
has a geometric distribution with parameter F().
2
The strong law of large numbers for the renewal process shows that the
mean value EN(t) is approximated by t and plays a key role in the asymp-
totic analysis. Note that in the case of homogeneous Poisson process the
exact mean value EN(t) = t is known.
Theorem 3.5 (Elementary Renewal Theorem) Suppose that F
T
() =
1. Then
lim
t
EN(t)
t
= .
Proof. Since N(t) is not always a stopping time we dene the rst passage
time
(t) = min{n :
n
> t}.
Note that (t) = N(t) + 1. It follows that
EN(t) = E(t) 1 = E
(t)
1 = t + E(
(t)
t) 1,
and
N(t)
t
= +
E(
(t)
t)
t

1
t
. (3.12)
Since
(t)
t 0, we obtain
liminf
t
EN(t)
t
. (3.13)
Note that for some M 0,
P(T
k
M) = 1, k = 1, 2, . . . .
44 CHAPTER 3. COUNTING PROCESSES
Together with (3.12) this implies that
EN(t)
t
+
M
t
and
limsup
t
EN(t)
t
,
which together with (3.13) proves the theorem.
2
The next theorem states the asymptotic behavior of the variance of the
renewal process.
Theorem 3.6 Suppose that V ar(T) < . Then
lim
t
V ar(N(t))
t
=
V ar(T
1
)
(ET
1
)
3
.
Theorem 3.7 (Central Limit Theorem) Let V ar(T
1
) < . Then for
t
N(t) t
_
V ar(T
1
)
(ET
1
)
3
t
Z N(0, 1).
3.2.1 Renewal function
Denition 3.5 The mean value
m(t) = 1 + EN(t) = E(t), t 0
is called a renewal function.
Since {N(t) k} = {
k
t}, k = 1, 2, . . . we have
m(t) = 1 +

k=1
P(N(t) k) = 1 +

k=1
P(
k
t) =

k=0
F
k
T
(t), (3.14)
that is m(t) is the expected number of renewals in [0, t] and is called also a
renewal measure.
Note that m(t) is similar to a distribution function. It is nondecreasing
and right-continuous on R, but has a unit jump at t = 0 and m(t) as
t .
3.2. RENEWAL PROCESS 45
An important property of the renewal function is that it determines
uniquely the distribution F
T
. Taking the Laplace transform of both sides
in (3.14) we have
LT
m(t)
(s) =

k=0
LT
F
k
T
(s) =

k=0
(LT
F
T
(s))
k
=
1
1 LT
F
T
(s)
.
This yields the following
Theorem 3.8 The Laplace transforms of m(t) and F
T
are determine each
other uniquely by the relation
LT
m(t)
(s) =
1
1 LT
F
T
(s)
.
Hence m(t) and F
T
uniquely determine each other.
This result can be used for identication of the renewal process. For example,
the renewal function of the Poisson process is m(t) = t +1 and any renewal
process with this type of renewal function is a Poisson process.
Theorem 3.9 (The Integral Equation for Renewal process) The re-
newal function m(t) satises the integral equation
m(t) = I
[0,)
(t) + F
T
(t) +
_
t
0
m(t s)dF
T
(s). (3.15)
Moreover, m(t) is the unique solution of (3.15), which is bounded on nite
intervals.
Proof. According (3.14) it follows that
m(t) = 1+F
T
(t)+

k=2
F
k
T
(t) = I
[0,)
(t)+F
T
(t)+

k=1
F
(k+1)
T
(t) = 1+F
T
(t)+

k=1
(F
k
T
F
T
)(t),
or
m(t) = I
[0,)
(t) + F
T
(t) + (m F
T
)(t),
which is equivalent to (3.15).
2
The equation (3.15) is called a renewal equation. In the general case the
renewal equation is given by
U(t) = u(t) +
_
t
0
U(t y)dF(y), (3.16)
46 CHAPTER 3. COUNTING PROCESSES
where all functions are dened on [0, ). The function U(t) is unknown, u(t)
is given and F(y) is a distribution function. If F is a defective distribution,
then (3.16) is called a defective renewal equation.
The solution of the equation (3.16) is dened by the next theorem.
Theorem 3.10 If u(t) is bounded on nite intervals, then
U(t) =

k=0
_
t
0
u(t s)dF
k
(s)
=
_

0
u(t s)dm(s), t 0,
is the only solution of the renewal equation (3.16).
Theorem 3.11 (Key renewal theorem) If in addition u(t) is directly Rie-
mann integrable, then
lim
t
U(t) =
_

0
u(s)ds.
3.2.2 Recurrence times of a renewal process
Consider the renewal sequence {
n
, n = 1, 2, . . .
0
= 0} and T
n
> 0. Note
that
{N(t) = n} = {
n
t <
n+1
}.
In particular,

N(t)
t <
N(t)+1
.
For t 0, B(t) = t
N(t)
is the time since the last renewal prior to t and is
called a backward recurrence time of the renewal process (or the age process).
F(t) =
N(t)+1
t is the time to the next renewal after t and is called
forward time (or excess life or residual life).
We will show that for xed 0 x < t, the distribution function P(B(t)
x) satises the renewal equation. Since B(t) t a.s. it is sucient to
consider x < t. Hence for x t, P(B(t) x) = 1.
Start with the identity
P(B(t) x) = P(B(t) x, T
1
t) + P(B(t) x, T
1
> t), x > 0. (3.17)
If T
1
> t, no jumps occur up to t and N(t) = 0. Consequently B(t) = t, and
hence
P(B(t) x, T
1
> t) = (1 F
T
1
(t)), x t.
3.2. RENEWAL PROCESS 47
For T
1
t we will show that
P(B(t) x, T
1
t) =
_
t
0
P(B(t y) x)dF
T
1
(y). (3.18)
According the properties of the renewal process
P(B(t) x, T
1
t) = P(t
N(t)
x, N(t) 1)
=

n=1
P(t
N(t)
x, N(t) = n)
=

n=1
P(t
n
x,
n
t <
n+1
).
For every one of the summands, given {T
1
= y} for y t we have
P(t
n
x,
n
t <
n+1
|T
1
= y)
= P(t [y +

n
i=2
T
i
] x, y +

n
i=2
T
i
t < y +

n+1
i=2
T
i
)
= P(t y
n1
x,
n1
t y
n
)
= P(t y
N(ty)
x, N(t y) = n 1),
and hence
P(B(t) x, T
1
t) =

n=1
_
t
0
P(t y
N(ty)
x, N(t y) = n)dF
T
1
(y)
=
_
t
0
P(B(t y) x)dF
T
1
(y),
which proves (3.18).
Combining (3.17) and (3.18) we get
P(B(t) x) = (1 F
T
1
(t))I
[0,x]
(t) +
_
t
0
P(B(t y) x)dF
T
1
(y).
That is the renewal equation for u(t) = (1 F
T
1
(t))I
[0,x]
(t) and U(t) =
P(B(t) x).
Similarly
P(F(t) > x) =
_
t
0
P(F(t y) > x)dF
T
1
(y) + (1 F
T
1
(t + x)).
It is known that the only solution of the renewal equation is given by
U(t) = P(B(t) x) =
_
t
0
(1 F
T
1
(t y))I
[0,x]
(t y)dm(y).
48 CHAPTER 3. COUNTING PROCESSES
Consider the case of homogeneous Poisson process with intensity . In
this case m(t) = t + 1, 1 F
T
1
(x) = e
x
and
P(B(t) x) = P(t
N(t)
x) =
_
1 e
x
, x < t
1, x t.
Analogously
P(F(t) x) = P(
N(t)+1
t x) = 1 e
x
, x > 0.
3.3 Delayed Renewal process
In many cases the renewal process starts at random point. In this case the
time T
1
has a dierent distribution. The process N(t) is called a delayed
renewal process. The distribution of T
1
is called the delayed distribution.
Denition 3.6 A continuous time stochastic process N(t), t 0 is station-
ary or has stationary increments, if for every points 0 = t
1
< . . . < t
k
and
h > 0
(N(t
1
+ h), . . . , N(t
k
+ h))
d
= (N(t
1
), . . . , N(t
k
))
A basic property of the stationary point process is that the mean value
function is linear.
Theorem 3.12 Let N(t) be a stationary point process and EN(1) is nite.
Then EN(t) = EN(1)t.
Proof. Consider
EN(s + t) = EN(s) + [EN(s + t) EN(s)]
= EN(s) + EN(t).
This is a functional equation f(s + t) = f(s) + f(t), s, t 0. The only
nondecreasing function that satises this equation is f(ct) for some constant
c. In our case c = f(1) = EN(1), and hence EN(t) = EN(1)t.
2
We are ready to characterize the stationary renewal process.
Theorem 3.13 The delayed renewal process N(t) is stationary if and only
if the forward recurrence time process F(t) =
N(t)+1
t is stationary.
3.4. MIXED POISSON PROCESS 49
Proof. Using
n
= inf{s, N(s) = n}, we have
F(t) =
N(t)+1
t = inf{st : N(s) = N(t)+1}
d
= inf{s

: N((0, s

]+t) = 1}.
Consequently, the stationarity property of N implies F(t)
d
= F(0), t 0.
Then F is stationary process (it is a Markov process). Conversely, since N
counts the number of times F(t) jumps upward,
N(A + t) =

sA
I
{F(u+s)>F((u+t))
.
Therefore the stationarity of F implies N is stationary.
2
Theorem 3.14 Let N(t) be a stationary delayed renewal process. Then
a) EN(t) = t, t 0.
b) F
T
1
(t) =
_
t
0
[1 F
T
2
(s)]ds =
1
ET
2
_
t
0
[1 F
T
2
(s)]ds.
Proof. a) Let N(t) be stationary. Theorem 3.12 ensure that EN(t) =
EN(1)t. Also EN(1) = , since
EN(t)
t
, t . Therefore EN(t) = t.
2
3.4 Mixed Poisson process
The modeling by homogeneous Poisson process is not realistic.
Suppose that the parameter is a realization of the random variable
with distribution function F

. Then
p
k
(t) = P(N(t) = k) =
_

0
(t)
k
k!
e
t
dF

(),
where F

() = P( ) is the distribution function of the mixing distribu-


tion .
For the mixed Poisson process
I(t) = 1 +
V ar()
E
.
Example 3.2 Let (, ) with density function
f

(x) =

()
x
1
e
x
, x > 0.
50 CHAPTER 3. COUNTING PROCESSES
Then
p
k
(t) = P(N(t) = k) =
_
+ k 1
k
__

+ t
_

_
t
+ t
_
k
, k = 0, 1, 2, . . . ,
(3.19)
i.e. the number of claims has negative binomial distribution with parameters
and

+t
(N(t) NBD (,

+t
)). The counting process, dened by (3.18)
is called a Polya process.
3.5 Compound Poisson process
3.5.1 P olya - Aeppli process
Denition 3.7 A counting process {N(t), t 0} is said to be a Polya -
Aeppli process if
a) N(0) = 0;
b) N(t) has independent, stationary increments;
c) for each t > 0, N(t) is Polya - Aeppli distributed.
Theorem 3.15 Suppose that the inter-arrival times {T
k
}
k2
of the station-
ary renewal process are equal to zero with probability and with probability
1 exponentially distributed with parameter . Then the number of re-
newals up to time t, has the Polya - Aeppli distribution with parameters
and .
Chapter 4
Claim Size Models
The claim amounts to the insurance company can be described by discrete
and by continuous random variables. In the case of continuous distributed
claims, the basic is to nd an adequate model for the claim amount. The
probability distributions are separated in two families - light tailed and heavy
tailed distributions. The exponential distribution is the border one.
Denition 4.1 The distribution F is called light tailed, if for > 0,
limsup
x
F(x)
e
x
< .
For the light tailed distribution there are constants a > 0 and > 0, such
that F(x) ae
x
and there is z > 0, such that M
X
(z) < .
According the denition, the exponential distribution is light tailed for
every > 0.
Denition 4.2 If for every > 0,
liminf
x
F(x)
e
x
> 0,
the distribution F has a heavy tail.
For the heavy tailed distributions, for every a > 0 and > 0, F(x) > ae
x
and for every z > 0, M
X
(z) = .
The truncated normal distribution is called a standard distribution. This
is a random variable Z = |Y |, where Y is a normal distributed random
variable with distribution function F(z) = P(|Y | z). If Y has a standard
normal distribution, F(z) = 2((z)
1
2
), x > 0, where (z) is a standard
normal distribution function.
51
52 CHAPTER 4. CLAIM SIZE MODELS
It is easy to show that if (z) is a standard normal density, then
lim
z
z(z)
(z)
= 1,
and hence the truncated normal distribution is a light tailed.
The Pareto distribution with parameters > 0 and > 0
1 F(x) =
_

+ x
_

, x > 0
has a heavy tail.
The Weibull distribution
1 F(x) = e

(
x

, x > 0, > 0
is heavy tailed for < 1 and light tailed for 1.
The most useful light tailed and heavy tailed distributions are given in
the next tables.
Light tailed distributions
Name Parameters Density
Exponential > 0 f
X
(x) = e
x
Gamma > 0, > 0 f
X
(x) =

()
x
1
e
x
Weibull > 0, 1 f
X
(x) = x
1
e
x

Hyperexponential
i
> 0,

n
i=1
p
i
= 1 f
X
(x) =

n
i=1
p
i

i
e

i
x
Heavy tailed distributions
Name Parameters Density
Weibull > 0, 0 < < 1 f
X
(x) = x
1
e
x

Lognormal R, > 0 f
X
(x) =
1
x

2
e

(ln x)
2
2
2
Loggamma > 0, > 0 f
X
(x) =

(ln x)
1
x
+1
()
Pareto > 0, > 0 f
X
(x) =

+x
_

+x
_

Burr > 0, > 0, > 0 f


X
(x) =

x
1
(+x

)
+1
4.1. HEAVY TAILED DISTRIBUTIONS 53
4.1 Heavy tailed distributions
Let
F
= limsup
x
(x)
x
, where (x) = log F(x) is the hazard function
of F. If F is continuously dierentiable, then (x) is also dierentiable
and
d(x)
dx
= (x), where (x) is the intensity rate function. Suppose that
F(0) = 0.
Theorem 4.1 If
F
= 0, then F is heavy tailed distribution function.
Proof. Suppose that
F
= 0, i.e. limsup
x
(x)
x
= 0. Then, for every
> 0, there is x

> 0, such that for x > x

, (x) x. Consequently, there


is a constant c > 0, such that for every x 0, F(x) ce
x
, and therefore
for every t ,
_

0
e
tx
F(x)dx = . (4.1)
Letting 0 yields that (4.1) holds for every t > 0.
2
Remark 4.1 Suppose that F is heavy tailed distribution function. Then for
every t > 0,
lim
x
e
tx
F(x) = . (4.2)
The most popular heavy tailed distributions are the distributions with
regularly varying tails and subexponential distributions.
4.2 Regularly varying functions
Denition 4.3 The positive, measurable function f is called regularly vary-
ing at with index R, if
lim
x
f(tx)
f(x)
= t

for every t > 0. (4.3)


If = 0, f is called slowly varying function.
The family of regularly varying functions with index is denoted by
RV().
Suppose that in (4.3) the limit, if it exists, is nite and positive for every
t > 0. Then the limiting function satises the equation K(ts) = K(t)K(s).
The only solution of this equation is the power function.
54 CHAPTER 4. CLAIM SIZE MODELS
Remark 4.2 If f RV () then
f(x) = x

L(x),
where L(x) is a slowly varying function (L(x) RV (0)).
Remark 4.3 The function f(x) is called regularly varying at 0, if f(
1
x
) is
regularly varying at .
Examples: Slowly varying functions are the positive constants and the func-
tions log(1 + x), log log(e + x).
According the Remark 4.1, the following functions:
x

, x

ln(1 + x), [x ln(1 + x)]

, x

ln(ln(e + x))
are in RV (). Probability distributions whose tails are regularly varying are
1 F(x) = x

, x 1, > 0
and the extreme value distribution

(x) = exp(x

), x 0.
4.2.1 Properties
Theorem 4.2 (Karamatas theorem) Let L be a slowly varying function.
Then there exists t
0
> 0 such that L is locally bounded over [t
0
, ) and
a) For > 1
_
t
t
0
s

L(s)ds ( + 1)
1
t
+1
L(t), t . (4.4)
b) For < 1 or = 1 and
_

0
L(s)
s
ds < ,
_

t
s

L(s)ds ( + 1)
1
t
+1
L(t), t . (4.5)
Conversely, if (4.4) holds with > 1, then L RV (0). If (4.5) holds
with > 1, then L RV (0).
Remark 4.4 Let f RV () and local bounded over [t
0
, ) for some t
0
0.
Then
a) For > 1
lim
t
_
t
t
0
f(s)ds
tf(t)
=
1
+ 1
.
4.2. REGULARLY VARYING FUNCTIONS 55
b) For < 1
lim
t
_

t
f(s)ds
tf(t)
=
1
+ 1
.
If = 1 and for some positive function f, local bounded over [t
0
, )
one of the conditions a) or b) satises, then f RV ().
Theorem 4.3 (Representation theorem) Let f RV (). Then there
exist measurable functions : R
+
R and c : R
+
R, such that
lim
t
(t) = and lim
t
c(t) = c
0
> 0 (4.6)
and t
0
R
+
, such that for t > t
0
f(t) = c(t) exp
__
t
t
0
(s)
s
ds
_
. (4.7)
Conversely, if (4.7) holds with and c satisfying (4.6), then f RV ().
For example, the function L(t) = ln t is slowly varying and the representation
(4.7) holds with t
0
= e, c(t) = 1 and (t) = (ln t)
1
.
Remark 4.5 From the representation theorem, it follows that the functions
of RV () satisfy
lim
t
f(t) =
_
, > 0
0, < 0.
If L is slowly varying, then for every > 0,
lim
t
t

L(t) = 0 and lim


t
t

L(t) = .
4.2.2 Regularly varying random variables
Denition 4.4 The nonnegative random variable X and its distribution are
called regularly varying with index 0, if the right tail distribution F
X
(x)
RV ().
Theorem 4.4 (Regularly varying distributions) Let F be a distribution
function and F(x) < 1 for every x 0.
a) If the sequences (a
n
) and (x
n
) satisfy
a
n
a
n+1
1, x
n
and for
some real function g and all values of some subset of (0, ),
lim
n
a
n
F(x
n
) = g() (0, ),
56 CHAPTER 4. CLAIM SIZE MODELS
then g() =

for some 0 and F RV.


b) Let F be absolutely continuous with density f, such that lim
x
xf(x)
F(x)
=
for some > 0. Then f RV ((1 + )) and F RV ().
c) Let f RV ((1 +)) for > 0. Then lim
x
xf(x)
F(x)
= . This means
that if for > 0, F RV (), then the density f is monotone function.
d) Let X RV () be a nonnegative random variable and > 0. Then
EX

< , if < ;
EX

= , if > .
e) Suppose that F RV (), > 0, . Then
lim
x
x

F(x)
_
x
0
y

dF(y)
=

.
The opposite holds if > . If = , the only that we can say is that
F(x) = (x

L(x)) for some slowly varying function L.


f ) The following are equivalent:
(1)
_
x
0
y
2
dF(y) RV (0);
(2) F(x) = (x
2
_
x
0
y
2
dF(y)), x .
Example 4.1 Distributions similar to Pareto These are the Pareto dis-
tribution, Cauchy distribution, Burr distribution, stable distribution with in-
dex < 2. The right tails of all these distributions are given by
F(x) Kx

, x
for some constants > 0 and K. It is clear that F(x) RV ().
The regularly varying distributions are widely used in practice. The mo-
tivation is given in the next
Lemma 4.1 Let X and Y be independent nonnegative random variables
from RV (), 0. Then X + Y RV () and
P(X + Y > x) P(X > x) + P(Y > x), x .
4.3. SUBEXPONENTIAL DISTRIBUTIONS 57
4.3 Subexponential distributions
Let F(x) be a regularly varying distribution function, dened on (0, ), such
that for every n 2,
lim
x
F
n
(x)
F(x)
= n. (4.8)
F
n
(x) means nth convolution of F(x) = 1 F(x).
The condition (4.8) means that for x
P(X
1
+ . . . + X
n
) > x) nP(X
1
> x).
It is easy to show that (4.7) for n = 2, implies n 2.
Let X
1
, X
2
, . . . be nonnegative identically distributed random variables.
Then for x ,
P(max(X
1
, . . . X
n
) > x) = 1 [F(x)]
n
=
= [1 F(x)]

n1
k=0
[F(x)]
k
n[1 F(x)],
,
and hence (4.8) is equivalent to
lim
x
P(X
1
+ . . . + X
n
) > x)
P(max(X
1
, . . . X
n
) > x)
= 1. (4.9)
This means that the largest claims have the main contribution to the sum
X
1
+. . .+X
n
. This property denes a large family of probability distributions.
Denition 4.5 Subexponential distribution The nonnegative random
variable X and the distribution of X are called subexponential if the inde-
pendent copies X
1
, X
2
, . . . , X
n
satisfy (4.8) and (4.9).
The family of subexponential distributions: SE. It is introduced by V.Chistyakov
in 1964.
4.3.1 Properties
The main properties are given in the next
Lemma 4.2 a) Let F SE. Then for all y R
lim
x
F(x y)
F(x)
= 1. (4.10)
58 CHAPTER 4. CLAIM SIZE MODELS
b) If (4.10) holds, then for every r > 0
lim
x
e
rx
(1 F(x)) =
and
_

0
e
rx
dF(x) = .
c) Let F SE. Then for given > 0, there is a nite constant K, such
that for n 2
1 F
n
(x)
1 F(x)
K(1 + )
n
, x 0. (4.11)
Proof. a) For 0 y x we have
1 F
2
(x)
1 F(x)
= 1 +
F(x) F
2
(x)
1 F(x)
= 1 +
_
y
0
1 F(x t)
1 F(x)
dF(t) +
_
x
y
1 F(x t)
1 F(x)
dF(t)
1 + F(y) +
1 F(x y)
1 F(x)
(F(x) F(y)).
Thus for large x, such that F(x) F(y) = 0,
1
1 F(x y)
1 F(x)

_
1 F
2
(x)
1 F(x)
1 F(y)
_
(F(x) F(y))
1
.
The assertion (4.10) follows by letting x .
If y < 0 then
lim
x
1 F(x y)
1 F(x)
= lim
x
1 F(x y)
1 F((x y) (y))
= lim
v
_
1 F(v (y))
1 F(v)
_
1
= 1,
where v = x y.
b)
The MGF:
_

0
e
rx
dF(x) = 1 +
_

0
_
x
0
re
ry
dydF(x)
= 1 + r
_

0
_

y
e
ry
dF(x)dy = 1 + r
_

0
e
ry
(1 F(y))dy = .
4.3. SUBEXPONENTIAL DISTRIBUTIONS 59
c) Let
C
n
= sup
x0
1 F
n
(x)
1 F(x)
.
Then
1 F
(n+1)
(x)
1 F(x)
= 1 +
F(x) F
(n+1)
(x)
1 F(x)
= 1 +
_
x
0
1 F
n
(x t)
1 F(x)
dF(t)
= 1 +
_
x
0
_
1 F
n
(x t)
1 F(x t)
1 F(x t)
1 F(x)
_
dF(t).
For T < and n 1
C
n+1
1+ sup
0xT
_
x
0
F
n
(x y)
F(x)
dF(y)+sup
xT
_
x
0
F
n
(x y)
F(x y)
F(x y)
F(x)
dF(y)
1 +
1
F(T)
+ C
n
sup
xT
F(x) F
2
(x)
F(x)
,
where
1
F(T)
< .
Since F SE, then for a given > 0, one can choose T = T() such that
sup
xT
F(x) F
2
(x)
F(x)
< 1 + ,
and hence
C
n+1
1 +
1
F(T)
+ C
n
(1 + ).
Note that C
1
= 1. We obtain the assertion recursively
C
n
1 +
1
F(T)
+ C
n1
(1 + ) 1 +
1
F(T)
+ (1 +
1
F(T)
)(1 + ) + C
n2
(1 + )
2
. . .
(1 +
1
F(T)
)

n1
i=0
(1 + )
i

(1+
1
F(T)
)

((1 + )
n
.
2
Remark 4.6 The assertion b) explains the name of the sub-exponential dis-
tributions F SE. Since for every > 0
_

y
e
x
dF(x) e
y
F(y), y 0,
MGF of the distributions F SE, doesnt exist. Consequently, the Laplace -
Stieltjes also doesnt exist.
60 CHAPTER 4. CLAIM SIZE MODELS
The following lemma gives a sucient condition for subexponentiality.
Lemma 4.3 Let for z (0, 1] the limit
(z) = lim
x
1 F(zx)
1 F(x)
exists and is left-continuous at 1. Then F is a subexponential distribution
function.
Proof. Note that
F
2
(x) = P(X
1
+ X
2
x) P(X
1
x, X
2
x) = [F(x)]
2
.
Assume that F(0) = 0. Hence
liminf
x
1 F
2
(x)
1 F(x)
liminf
x
1 [F(x)]
2
1 F(x)
= liminf
x
[1 + F(x)] = 2.
For xed n 1
limsup
x
1 F
2
(x)
1 F(x)
= 1 + limsup
x
_
x
0
1 F(x y)
1 F(x)
dF(y)
1 + limsup
x
n

k=1
1 F(x
kx
n
)
1 F(x)
_
F
_
kx
n
_
F
_
(k 1)x
n
__
= 1 + (1
1
n
).
Since n is arbitrary, is left continuous at 1.
2
Example 4.2 Consider the Pareto(, ) distribution.
1 F(zx)
1 F(x)
=
_

+zx
_

_

+x
_

=
_
+ x
+ zx
_

,
as x . It follows that Pareto distribution is subexponential.
Remark 4.7 The condition (4.10) gives another denition of the heavy tailed
distributions. For the random variable X with distribution F, the condition
(4.10) can be written as
lim
x
P(X > x + y|X > x) = lim
x
F(x + y)
F(x)
= 1, y > 0.
4.3. SUBEXPONENTIAL DISTRIBUTIONS 61
The light tail of the exponential distribution F(x) = e
x
, x > 0, for > 0,
satises the condition
lim
x
F(x + y)
F(x)
= e
y
, y > 0.
62 CHAPTER 4. CLAIM SIZE MODELS
Chapter 5
Cramer - Lundberg model
5.1 Ruin probability
The ruin probability is a measure for the risk related to the some company.
For convenience we consider an insurance company.
Consider the usual risk model. Let (, F, P) be a complete probability
space with
a) Counting process N(t), N(0) = 0;
b) A sequence {Z
k
}

1
of independent identically distributed random vari-
ables with distribution function F such that F(0) = 0, mean value and
variance
2
.
The risk process is given by
X(t) = ct
N(t)

k=1
Z
k
(
0

k=1
def
= 0),
where c is a positive constant, representing the premium incomes per unit
time. The constant c is called gross risk premium rate.
Suppose that N(t) has an intensity and EN(t) = t. The prot of this
risky business in (0, t] is X(t). The expected prot is
EX(t) = ct EN(t)EZ
k
= (c )t.
Dene the safety loading coecient :
=
c

=
c

1.
The risk process X(t) has a positive safety loading, if > 0, i.e.. c > .
In this case X(t) has a trend to + and we say that there is a net prot
condition, (NPC).
63
64 CHAPTER 5. CRAM

ER - LUNDBERG MODEL
The ruin probability (u) for the insurance company with initial capital
u is dened by
Denition 5.1
(u) = P{u + X(t) < 0 for some t > 0}.
Sometimes it is more convenient to use the probability of non ruin (u) =
1 (u).
From the denition it follows that (u) = 1 for u < 0.
Denition 5.2 If N(t) is a homogeneous Poisson process with intensity ,
i.e. N(t) Po(t), X(t) is called a classical risk model, or Cramer -
Lundberg model.
5.2 Integral equation of ruin probability
Consider the Poisson process as a renewal process. At rst we will derive the
equation for the non-ruin probability (u).
Let T
1
be the time to the rst claim. Then X(T
1
) = cT
1
Z
1
. Conditioning
on the no claim in (0, T
1
), we obtain
(u) = E[(u + cT
1
Z
1
)] =
=
_

0
e
s
_
u+cs
0
(u + cs z)dF(z)ds.
The change of variables x = u + cs leads to
(u) =

c
e

u
c
_

u
e

x
c
_
x
0
(x z)dF(z)dx,
and hence is a dierentiable and

(u) =

c
(u)

c
_
u
0
(u z)dF(z). (5.1)
5.2. INTEGRAL EQUATION OF RUIN PROBABILITY 65
Integrating from 0 to t :
(t) (0) =

c
_
t
0
(u)du +

c
_
t
0
_
u
0
(u z)d(1 F(z))du
=

c
_
t
0
(u)du +

c
_
t
0
_
(0)(1 F(u) (0) +
_
u
0

(u z)(1 F(z))dz
_
du
=

c
(0)
_
t
0
(1 F(u))du +

c
_
t
0
(1 F(z))dz
_
t
z

(u z)du
=

c
(0)
_
t
0
(1 F(u))du +

c
_
t
0
(1 F(z))((t z) (0))dz
leads to the equation
(u) = (0) +

c
_
u
0
(u z)(1 F(z))dz. (5.2)
It follows that for u , () = (0) +

c
(). The law of large
numbers:
lim
t
X(t)
t
= c with probability 1.
The NPC says that > 0 and c > . Consequently there is a dependent of
N and {Z
k
} random variable T, such that X(t) > 0 for every t > T.
Since only a nite number of claims can arrive up to T,
inf
t>0
X(t) < with probability 1
and () = 1. Consequently
1 = 1 (0) +

c
,
i.e.
(0) =

c
=
1
1 +
.
Example 5.1 Suppose that the claims are exponentially distributed with mean
value , i.e. F(x) = 1 e

, x 0. The equation (5.1) is given by

(u) =

c
(u)

c
_
u
0
(u z)e

dz =

c
(u)

c
_
u
0
(z)e

uz

dz.
66 CHAPTER 5. CRAM

ER - LUNDBERG MODEL
The second derivative:

(u) +
1

1 +

(u) = 0.
Together with the initial conditions
() = 1 and (0) =

1 +
the solutions is
(u) = 1
1
1 +
e

1+
u
.
From the equation (5.2) and (u) = 1 (u) it follows that
(u) =

c
_

u
(1 F(z))dz +

c
_
u
0
(u z)(1 F(z))dz. (5.3)
5.3 Cramer - Lundberg approximation
From the denition of the safety loading coecient:

c
=
1
(1+)
. Denote:
F
I
(z) =
1

_
z
0
(1 F(x))dx,
called an integrating tail distribution. In terms of the safety loading the
equation (5.2) for the nonruin probability is
(u) =

1 +
+
1
1 +
_
u
0
(u z)dF
I
(z).
Denote q =
1
1+
and rewrite the equation for (u).
(u) = qF
I
(u) +
_
u
0
(u x)d(qF
I
(x)). (5.4)
This is a renewal equation relative the measure qF
I
(x). It is easy to see
that lim
x
qF
I
(x) = q < 1, i.e. qF
I
(x) is not a probability measure. The
equation (5.4) is a defective renewal equation.
For some r and x > 0, we dene
F
(r)
(x) =
_
x
0
e
rz
d(qF
I
(z)) =
q

_
x
0
e
rz
(1 F(z))dz.
5.3. CRAM

ER - LUNDBERG APPROXIMATION 67
F
(r)
(x) is a distribution function. It is nondecreasing and lim
x
F
(r)
(x) =
1, i.e.
q

_

0
e
rz
(1 F(z))dz = 1. (5.5)
The distribution, generated by F
(r)
(x) is called Esscher transform of F.
The equation (5.5) is called a Cramer condition.
From (5.5) it follows that
f
(r)
(x) =
q

e
rx
(1 F(x)), x > 0.
is a density of a proper probability distribution.
Multiplying the equation (5.4) by e
ru
:
e
ru
(u) = qe
ru
F
I
(u) +
_
u
0
e
r(ux)
(u x)e
rx
d(qF
I
(x))
= qe
ru
F
I
(u) +
_
u
0
e
r(ux)
(u x)dF
(r)
(x)). (5.6)
This is a renewal equation
U(t) = u(t) +
_
t
0
U(t y)dF(y),
for u(t) = qe
ru
F
I
(u), F = F
(r)
and U(t) = e
rt
(t). The function U(t) is
bounded on nite intervals. By the Theorem 3.10, the solution is:
U(t) = e
rt
(t) =
_
t
0
u(t y)dm
(r)
(y) = q
_
t
0
e
r(ty)
F
I
(t y)dm
(r)
(y),
where m
(r)
is the renewal function, corresponding to the renewal process with
F
(r)
distributed interarrival times. In general, the function m
(r)
is unknown.
The Key renewal theorem gives the asymptotic solution of the equation for
u .
If the integrals C
1
= q
_

0
e
rz
F
I
(z)dz and C
2
= q
_

0
ze
rz
F
I
(z)dz exist,
then for the solution of (5.6):
lim
u
e
ru
(u) =
C
1
C
2
. (5.7)
From (5.5) it follows that the Cramer condition is given by
M
Z
(r) 1
r
= (1 + ),
68 CHAPTER 5. CRAM

ER - LUNDBERG MODEL
where M
Z
(r) is the MGF of Z. We can see that the adjustment coecient
is independent of the Poisson parameter .
It can be shown that the function M
Z
(r) 1 r(1 +) is convex. The
equation
M
Z
(r) 1 = r(1 + ) (5.8)
has a solution equal to zero. If no zero solution exists, it is positive. Let R
be the positive solution of (5.8). Then for C
1
and C
2
we get
C
1
=
1
R

1 +
and
C
2
=
1
R
1
1 +
1

(M

Z
(R) (1 + ))
and the limit (5.7)
lim
u
e
Ru
(u) =

M

Z
(R) (1 + )
. (5.9)
Denition 5.3 (5.9) is called Cramer - Lundberg approximation. The
constant R, the non negative solution of the equation (5.5) is called a Lund-
berg exponent or adjustment coecient.
Example 5.2 We continue the Example 5.1 for exponentially distributed
claims. The MGF is given by M
Z
(r) =
1

0
e
rz
e

dz =
1
1r
. The con-
stant R is the positive solution of the equation
r
1 r
= r(1 + ),
and hence
R =
1

1 +
.
The derivative of the MGF is M

Z
(r) = (1 + )
2
. (5.9) is given by
lim
u
e
Ru
(u) =
1
1 +
.
This result and the ruin probability of Example 5.1 show that the Cramer -
Lundberg approximation for exponentially distributed claims is exact.
5.4. MARTINGALE APPROXIMATION 69
5.4 Martingale approximation
Remember two theorems, important for the approximation.
Theorem 5.1 (Stopping Time Theorem) Let be a nite stoping time
(markov moment), i.e. t
0
< and M is a right continuous F mar-
tingale (supermartingale). Then
E[M()|F
0
] = ()M(0), P a.s.
Theorem 5.2 Let X(t) be a continuous process such that
1. X(0) = 0 P a.s.;
2. X has stationary independent increments;
3. EX(t) = t, > 0;
4. Ee
rX(t)
< for r > 0.
Then Ee
rX(t)
= e
g(r)t
for some function g(.).
If X(t) is a classical risk model with NPC, then X(t) has stationary
independent increments with = c and
Ee
rX(t)
= e
rct
Ee
r(Z
1
+...+Z
N(t)
)
= e
rct

k=0
(t)
k
k!
e
t
[M
Z
(r)]
k
= e
rct
e
M
Z
(r)t
e
t
= e
[(M
Z
(r)1)cr]t
.
In this case
g(r) = (M
Z
(r) 1) cr.
It is easy to show that
M
u
(t) =
e
r(u+X(t))
e
g(r)t
is a martingale relative the algebra, generated by the process X.
Let T
u
= inf{t 0 : u + X(t) < 0} be the time to ruin for a company
with initial capital u. T
u
is a stoping time relative the ltration F
X
and the
ruin probability is
(u) = P(T
u
< ).
For t
0
< , t
0
T
u
is nite F
X
- stoping time. F
X
0
is a trivial
algebra and M
u
> 0. According the Theorem 5.1
e
ru
= M
u
(0) = E[M
u
(t
0
T
u
)] =
= E[M
u
(t
0
T
u
)|T
u
t
0
]P(T
u
t
0
) + E[M
u
(t
0
T
u
)|T
u
> t
0
]P(T
u
> t
0
)
E[M
u
(t
0
T
u
)|T
u
t
0
]P(T
u
t
0
) = E[M
u
(T
u
)|T
u
t
0
]P(T
u
t
0
).
70 CHAPTER 5. CRAM

ER - LUNDBERG MODEL
Since u + X(T
u
) 0 for T
u
< , then e
r(u+X(T
u
))
1. Consequently
P(T
u
t
0
)
e
ru
E[M
u
(T
u
)|T
u
t
0
]

e
ru
E[e
g(r)T
u
)
|T
u
t
0
]
e
ru
sup
0tt
0
e
g(r)t
.
Let t
0
and obtain
(u) e
ru
sup
t0
e
g(r)t
.
Let R = sup{r : g(r) 0}, i.e. R is the positive solution of the equation
g(r) = 0. This is just the equation (5.8) and R is the Lundberg exponent.
We obtain the inequality
(u) e
Ru
, (5.10)
named Lundberg inequality.
This approximation gives an interpretation of the Lundberg coecient.
Consider the surplus process U(t) = u+X(t). We will show that the process
e
rU(t)
is a martingale:
Ee
rU(t)
= Ee
r(u+ctS(t))
= e
ru
e
rct
e
[M
Z
(r)1]t
= e
ru
.
So, R is the unique positive number, such that e
rU(t)
is a martingale. It is
known that the martingale property is related to the fair game.
We are ready to give the basic theorem.
Theorem 5.3 (Cramer - Lundberg)
1. If > 0, then
(u) =

1 +

n=0
_
1
1 +
_
n
F
n
I
(u); (5.11)
2. If the constant R > 0 exists and
_

0
e
Rx
dF
I
(x) = 1 + ,
then
(u) e
Ru
,
where R is the Lundberg exponent.
3. Suppose that the integral
_

0
xe
Rx
F
I
(x)dx = C is nite. Then
lim
u
(u) =

RC
e
ru
.
Chapter 6
Renewal Risk Model
Consider the surplus process
U(t) = u + ct S(t),
where S(t) =

N(t)
k=1
Z
k
, the counting process N(t) is a renewal process,
0 =
0

1

2
. . . are the claim arrival times. The inter-arrival
times T
k
=
k

k1
, k = 1, 2, . . . are independent identically distributed
random variables with mean value ET
k
=
1

. The claims Z
k
are independent
identically distributed random variables, independent of N(t).
Let F
T
1
(t) be the distribution function of the time to the rst claim and
F
T
(t) be the distribution function of T
2
, T
3
, . . . . If F
T
1
(t) = F
T
(t), the count-
ing process is called an ordinary renewal process and the risk model is called
a Sparre Andersen model.
If the stationarity condition
F
T
1
(t) =
_
t
0
[1 F
T
(x)]dx (6.1)
holds, the counting process is stationary and the risk model is called a sta-
tionary renewal risk model.
6.1 Ordinary renewal risk model
In the ordinary case we use
0
(u) and
0
(u) = 1
0
(u) for ruin and non
ruin probability. According the argument that ruin can occur only at claim
times, the ruin probability for this model
0
(u) = P(U(t) 0, t 0) can
be given by

0
(u) = P(u + c
n
S(
n
) 0, n 1)
= P (u +

n
k=1
(cT
k
Z
k
) 0, n 1) = P
_
sup
n1

n
k=1
(Z
k
cT
k
) u
_
.
71
72 CHAPTER 6. RENEWAL RISK MODEL
Notation: X
k
= Z
k
cT
k
, W
n
=

n
k=1
X
k
and
M
n
= sup
n1
n

k=1
X
k
= sup
n1
W
n
. (6.2)
Then
0
(u) = 1
0
(u) = P(M
n
u). The dened random variables
X
k
, k = 1, 2, . . . are independent identically distributed. A sum of indepen-
dent identically distributed random variables is called a random walk. This
proves the following
Proposition 1 The ruin probability for the zero-delayed case can be repre-
sented as
0
(u) = P(M
n
> u), where M
n
is given by (6.2) with W
n
a discrete
time random walk with increments distributed as the dierence Z cT be-
tween claims Z and the interarrival time cT.
The relative safety loading =
cET
1
EZ
1
1 =
c

1 > 0, i.e. the premium


received per unit time exceeds the expected claim payments per unit time.
The NPC implies that
EX
1
= E(Z
1
cT
1
) =
c

< 0
and that W
n
, a.s. It is easy to proof that 0 M
n
< .
The safety loading coecient is the same like in the Cramer - Lundberg
model. The dierence is that here EN(t) = t, and hence excluding the case
T exp(),
E (U(t) u) = (c )t.
According the Law of large numbers
lim
t
E(U(t) u)
t
= c ,
and (u) 0 for u .
6.1.1 Lundberg exponent
Let U(t) be an ordinary renewal risk model.
Lemma 6.1 Suppose that M
Z
(r) < and for r 0, g(r) is the unique
solution of the equation
M
Z
(r)M
T
(g(r) cr) = 1. (6.3)
Then the discrete time process e
rU

k
g(r)
k
is a martingale.
6.1. ORDINARY RENEWAL RISK MODEL 73
Proof. If r 0, then M
Z
(r) 1. M
T
(r) is increasing continuous func-
tion, dened for r 0. Consequently M
T
(r) 0 for r and there is
an unique solution g(r) of the equation (6.3). Then
E
_
e
rU

k+1
g(r)
k+1
|F

k
_
= E
_
e
r[c(
k+1

k
)Z
k+1
]g(r)(
k+1

k
)
|F

e
rU

k
g(r)
k
= E
_
e
rZ
k+1
e
(cr+g(r))(
k+1

k
)
|F

e
rU

k
g(r)
k
= E [M
Z
(r)M
T
(cr g(r))] e
rU

k
g(r)
k
= e
rU

k
g(r)
k
.
2
Example 6.1 For exponentially distributed interarrival times F(t) = 1
e
t
. The equation (6.3) is given by
M
Z
(r)

+ g(r) + cr
= 1.
As in the classical case, the function g(r) is convex and g(0) = 0. There
exists a nonnegative solution R to the equation g(R) = 0. This solution is
called again adjustment coecient or Lundberg exponent.
Recall that R is the only positive solution of the equation
M
Z
(r)M
T
(cr) = 1. (6.4)
Example 6.2 Let U(t) be a renewal risk model with exp(1) distributed claims,
premium rate c = 4 and iterarrival time distribution F(t) = 1
1
2
(e
2t
+ e
3t
) .
It follows that M
Z
((r) exists for r < 1, M
T
((r) exists for r < 2 and = 2.4.
The NPC 4 > 2.4 is fullled. The equation to solve is
1
1 r
1
2
_
2
2 + 4r
+
3
3 + 4r
_
= 1.
Thus
2(3 + 4r) + 3(2 + 4r) = 2(1 r)(2 + 4r)(3 + 4r)
or equivalently
4r
3
+ r
2
r = 0.
We nd the solutions r = 0 and
r
1,2
=
1

17
8
,
such that r
1
=
1+

17
8
> 0 and r
2
=
1

17
8
< 0. We proved that there is only
no negative solution. Why do we get r
2
< 0. Obviously M
Z
(r
2
) < 1 < .
But cr =
1+

17
2
> 2 and thus M
T
(cr
2
) = and r
2
is not a solution of
(6.4).
74 CHAPTER 6. RENEWAL RISK MODEL
6.1.2 Pollaczeck - Khinchine formula (Ruin probability
as a compound geometric probability)
The Poisson process is an ordinary renewal process and by the renewals
arguments we can proof formula (5.11) of the Cramer - Lundberg theorem.
Let
0
(u) be the probability of no ruin in the ordinary case. From the
equation of the non ruin probability

0
(u) =
0
(0) +
1
1 +
_
u
0

0
(u z)dF
I
(z) (6.5)
by Laplace transform follows (5.11). Denote by
LT

0(s) =
_

0
e
sz

0
(z)dz
the Laplace transform and
LST

0(s) =
_

0
e
sz
d
0
(z)
the Laplace - Stieltjes transform of
0
. Recall the relation LT

0(s) =
1
s
LST

0(s). Taking the Laplace transform in both sides of the equation (6.5)
gives
LT

0(s) =

0
(0)
s
+
1
1+
_

0
e
sz
_
z
0

0
(z t)dF
I
(t)dz
=

0
(0)
s
+
1
1+
_

0
_

t
e
sz

0
(z t)dzdF
I
(t)
=

0
(0)
s
+
1
1+
_

0
_

0
e
s(x+t)

0
(x)dxdF
I
(t)
=

0
(0)
s
+
1
1+
_

0
e
st
__

0
e
sx

0
(x)dx

dF
I
(t)
=

0
(0)
s
+
1
1+
LT

0(s)
_

0
e
st
dF
I
(t)
=

0
(0)
s
+
1
1+
LT

0(s)LST
F
I
(s).
Hence, for the Laplace transform we have
LT

0(s) =

0
(0)
s
_
1
1
1+
LST
F
I
(s)
. (6.6)
6.2. STATIONARY CASE 75
The Laplace - Stieltjes transform:
LST

0(s) =

0
(0)
1
1
1+
LST
F
I
(s)
=
0
(0)

n=0
_
1
1 +
LST
F
I
(s)
_
n
.
According the inversion formula with initial condition
0
(0) =

1+
, the prob-
ability of non ruin
0
(u) in the ordinary case is given by

0
(u) =

1 +

n=0
_
1
1 +
_
n
F
n
I
(u), (6.7)
called Pollaczeck - Khinchine formula.
It is easy to see that this formula is a compound geometric sum, i.e. a
geometric sum of independent, identically distributed random variables with
distribution function F
I
(z).
Exercise 6.1 Show that the ruin probability for the ordinary renewal risk
model is given by

0
(u) =

1 +

n=1
_
1
1 +
_
n
F
n
I
(u),
6.2 Stationary case
Note that the ruin probability for the delayed case with T
1
= s can be
expressed in terms of the zero-delayed case as
(u) = 1 F(u + cs) +
_
u+cs
0

0
(u + cs y)dF(y).
Indeed, the rst term represents the probability P(U
1
cs > u) of ruin at
the time s of the rst claim. The second term is P((u) < , U
1
cs u),
as follows easily by noting that the evolution of the risk process after time s
is that of a renewal risk model with initial reserve U
1
cs.
Proposition 2 The non-ruin probability (u) and the ruin probability (u)
in the stationary case satisfy the integral representations
(u) = (0) +

c
_
u
0

0
(u z)(1 F(z))dz (6.8)
and
(u) =

c
__

u
(1 F(z))dz +
_
u
0

0
(u z)(1 F(z))dz
_
. (6.9)
76 CHAPTER 6. RENEWAL RISK MODEL
Since () =
0
() = 1 when c > , we have
(0) = 1

c
.
Taking the Laplace transform of (6.8) and applying (6.6) we have
LT

(s) =
(0)
s
+

c
LST
F
I
(s)

0
(0)
s[1 LST
F
I
(s)]
.
Again, the standard properties of the transforms lead to
LST

(s) = 1

c
+

c

0
(0)LST
F
I
(s)

n=0
_
1
1 +
_
n
[LST
F
I
(s)]
n
.
So, the ruin probability in the stationary case is given by
(u) =

c
_
F
I
(u) + [1

c
] F
I
(u)

n=1
_

c
_
n
F
n
I
(u)
_
, (6.10)
where F
I
(u) = 1 F
I
(u).
In terms of the relative safety loading the ruin probability is given by
(u) =
1
1 +
_
F
I
(u) +

1 +
F
I
(u)

n=1
_
1
1 +
_
n
F
n
I
(u)
_
.
Example 6.3 Again, consider the case in which the claim amount distribu-
tion is exponential with mean value . Applying the argument of the ordinary
case we obtain the ruin probability
(u) =
1
1 +
exp
_

1 +
u
_
. (6.11)
6.3 Ruin probability for heavy tailed distri-
butions
Recall that in the Cramer - Lundberg model the following relation for (u)
holds
(u) =

1 +

n=1
_
1
1 +
_
n
F
n
I
(u),
6.3. RUIN PROBABILITY FOR HEAVY TAILED DISTRIBUTIONS 77
where F
I
(u) =
1

_
u
0
[1 F(x)]dx is the integrated tail distribution. Under
the condition that F
I
RV () for some 0 we might hope that the
following asymptotic estimate holds:
(u)
F
I
(u)
=

1 +

n=1
_
1
1 +
_
n
F
n
I
(u)
F
I
(u)
(6.12)


1 +

n=1
_
1
1 +
_
n
n =
1

, u . (6.13)
(6.13) is a natural estimate of ruin probability whenever F
I
is regularly
varying. We shall show that a similar estimate holds true for much wider
class of distribution functions. (6.13) can be reformulated as follows.
Proposition 3 For claim size distributions with regularly varying tails, the
ruin probability (u) for large initial capital u is essentially determined by
the tail F(z) of the claim size distribution for large values of z, i.e.
(u)
1

_

u
F(z)dz, u .
The main step in obtaining (6.13) is the property of the subexponential
distributions
F
n
I
(u) nF
I
(u) for any arbitrary n 2 and u .
This leads to
Theorem 6.1 (Cramer - Lundberg theorem for large claims, I) Con-
sider the Cramer - Lundberg model with NPC and F
I
(z) SE. Then
(u)
1

F
I
(u), u . (6.14)
Proof. Since
1
1+
< 1, there exists an > 0 such that
1
1+
(1 + ) < 1.
Together with the basic property
F
n
(u)
F(u)
K(1 + )
n
, u 0
it follows that
78 CHAPTER 6. RENEWAL RISK MODEL
1
(1 + )
n
F
n
(u)
F(u)

1
(1 + )
n
K(1 + )
n
, u 0,
which allows by dominated convergence the interchange of limit and sum in
(6.12), yielding the result.
2
For claim size distributions with subexponential integrated tail distribu-
tion, ultimate ruin probability (u) is given by (6.14).
From mathematical point of view this result can be substantially im-
proved.
Theorem 6.2 (Cramer - Lundberg theorem for large claims, II)
Consider the Cramer - Lundberg model with NPC. Then the following as-
sertions are equivalent.
a) F
I
(u) SE.
b) 1 (u) SE.
c)
(u)
F
I
(u)

, u .
Consequently, the estimate (6.14) is only possible under the condition
F
I
(u) SE.
Chapter 7
Premium Calculation
Principles
Denote by
X
the premium that an insurer charges to cover the risk X. The
risk X means that claims from this risk are described by the random variable
X and the distribution of X. The premium
X
is a function X, for example

X
= (X). The rule that assigns the numerical value of
X
is referred to
as a premium calculation principle.
Properties of premium principles:
1.
X
EX (nonnegative loading);
2. If X
1
and X
2
are independent, then
X
1
+X
2
=
X
1
+
X
2
(additivity);
3. If Z = aX, where a > 0, then
Z
= a
X
(scale invariance);
4. If Y = X + c, where c > 0, then
Y
=
X
+ c (consistency);
5. If there is a nite maximum value of the claim amount x
m
then

X
x
m
.
7.1 Premium calculation principles
7.1.1 Pure premium principle

X
= EX.
The pure premium is not very attractive.
7.1.2 Expected value principle

X
= (1 + )EX,
79
80 CHAPTER 7. PREMIUM CALCULATION PRINCIPLES
where > 0 is the safety loading factor. EX is the loading in the premium.
The premium is easy to calculate. It assigns the same premium to all risks
with the same mean value and is not sensible to heavy tailed distributions.
7.1.3 The variance principle

X
= EX + V ar(X),
where > 0. The loading is proportional to V ar(X). This principle counts
two characteristics of the risk - the mean value and the variance and is more
sensible to higher risks.
7.1.4 Standard deviation principle

X
= EX +
_
V ar(X),
where > 0. The loading is proportional to the standard deviation of X.
The loss can be written as

X
X =
_
V ar(X)
_

X EX
_
V ar(X)
_
,
or the loss is equal to the loading parameter minus a random variable with
mean value 0 and variance 1.
7.1.5 Modied Variance Principle
In the Variance Principle, the changing in monetary unite changes the secu-
rity loading. The following modication

X
=
_
EX +
V ar(X)
EX
, EX > 0
0, EX = 0
for > 0 can change this.
7.1.6 The Principle of Zero Utility
The worst thing that may happen for the company is a very high accumulated
sum of claims. Therefore high losses should be weighted stronger then small
losses. Hence, the company chooses a utility function v, which should have
the following properties:
1) v(0) = 0.
2) v(x) is strictly increasing.
7.1. PREMIUM CALCULATION PRINCIPLES 81
3) v(x) is strictly concave.
The rst property is for convenience. The second means that less losses
are preferred. The last condition gives stronger weights for higher losses.
The premium is dened by the equation
v(u) = E[v(u +
X
X)], (7.1)
where u is the insurers surplus. This means that the expected utility is the
same whether the insurance contract is taken or not. In general, the premium
depends on the surplus.
Lemma 7.1 1. If the solution of (7.1) exists, it is unique.
2. If for every x < , v

(x) < 0, then > EX;


3. The premium is independent of u if and only if v

(x) = 0 or v(x) =
A(1 e
x
), A > 0, > 0.
Proof.
1. Let
1
>
2
be two solutions of (7.1). Since v

(x) > 0, then


v(u) = E[v(u +
1
X)] > E[v(u +
2
X)] = v(u),
which is a contradiction.
2. The Jensens inequality:
v(u) = E[v(u + X)] < v(u + EX)
and since v

(x) > 0, we obtain u + EX > u.


3. It is easy to see that if v

(x) = 0 or v(x) = A(1 e


x
), then the
premium principle is independent of u.
Suppose that the premium is independent of u. Let P(X = 1) = 1
P(X = 0) = q and (q) is the premium. Then
qv(u + (q) 1) + (1 q)v(u + (q)) = v(u). (7.2)
Dierentiation in (7.2) relative to u leads to
qv

(u + (q) 1) + (1 q)v

(u + (q)) = v

(u). (7.3)
The derivative of (7.2) relative to q is
v(u+(q)1)v(u+(q))+

(q)(qv

(u+(q)1)+(1q)v

(u+(q)) = 0.
(7.4)
Insert (7.4) in (7.2) and obtain
82 CHAPTER 7. PREMIUM CALCULATION PRINCIPLES
v(u + (q) 1) v(u + (q)) +

(q)v

(u) = 0. (7.5)
Note that

(q) > 0. The derivative of (7.5) relative to u is


v

(u + (q) 1) v

(u + (q)) +

(q)v

(u) = 0,
and relative to q

(q)(v

(u+(q)1)v

(u+(q)))+

(q)v

(u) = [

(q)]
2
v

(u)+

(q)))v

(u) = 0.
Consequently

(q) 0 and for 0,


v

(u)
v

(u)
=
which proves the statement.
2
Remark 7.1 In the case of exponential utility function from the equation
(7.1) we obtain

X
=
1
log Ee
X
(7.6)
and is called an exponential principle.
7.1.7 The Esscher Principle

X
=
E[Xe
hX
]
E[e
hX
]
,
where h > 0.
The Esscher premium can be interpreted as a pure premium for Y, related
to X as follows. Let X be a continuous nonnegative random variable with
density function f. Dene the function g, such that
g(x) =
e
hX
f(x)
_

0
e
hx
f(x)dx
. (7.7)
The function g, dened by (7.7) is the density of the random variable Y.
The distribution function is given by
G(x) =
_
t
0
e
hy
f(y)dy
M
X
(h)
7.1. PREMIUM CALCULATION PRINCIPLES 83
and is called Esscher transform for the function F with parameter h. G(x) is
related to the distribution function of the risk X, but does give more weight
to larger losses. From the equation The MGF of Y is
M
Y
(t) =
M
X
(t + h)
M
X
(h)
.
Example 7.1 Let F(x) = 1 exp(x), x 0 be the distribution function
of the random variable X. Find the Esscher transform with parameter h < .
The MGF: M
X
(t) =

t
, hence
M
Y
(t) =
M
X
(t + h)
M
X
(h)
=
h
h t
.
The Esscher transform of the function F is
G(x) = 1 e
(h)x
.
The density g is an weighted version of f. From (7.7) it follows that
g(x) = w(x)f(x), where w(x) =
e
hx
M
X
(h)
. Since h > 0, then the weights increase
when x increases. This means that the transforms are useful for the heavy
tailed distributions. The expected value of Y :
EY =
_

0
xe
hx
f(x)dx
_

0
e
hx
f(x)dx
=
E[Xe
hX
]
E[e
hX
]
=
X
.
Example 7.2 Let X exp(1). Find the premium by the Esscher principle
with parameter h < 1.
7.1.8 Risk adjusted premium principle

X
=
_

0
[1 F(x)]
1

dx,
where 1 is called a risk index.
This principle is dened for nonnegative random variable X with distri-
bution function F.
Let Z be a random variable with distribution function H, dened by the
equation
1 H(x) = [1 F(x)]
1

.
The expected value of Z :
EZ =
_

0
[1 H(x)]dx,
and hence
X
= EZ.
84 CHAPTER 7. PREMIUM CALCULATION PRINCIPLES
Example 7.3 Let X exp(
1

). Find the risk adjusted premium


X
.
Here
1 F(x) = e
x
and
1 H(x) = e

.
Consequently Z exp(

) and
X
=

.
Example 7.4 Let X Par(, ). Find the risk adjusted premium
X
.
In this case
1 F(x) =
_

+ x
_

and
1 H(x) =
_

+ x
_

.
So Z Par(

, ) and
X
=

, < .
If X is a continuous random variable with density function f, then the
density of Z is h, dened by
h(x) =
1

[1 F(x)]
1

1
f(x). (7.8)
This means that the density of Z is an weighted version of f.
Chapter 8
Diusion Approximation
Consider the surplus process
U(t) = u + ct S
t
, t 0,
where S
t
= Z
1
+ . . . + Z
N(t)
is the accumulated loss process. N(t) is a
homogeneous Poisson process with intensity and S
t
= 0 for N(t) = 0. The
individual losses Z
1
, Z
2
, . . . are positive, independent, identically distributed
random variables, independent of N(t). Suppose that M
Z
(r) exists.
Here we will give an approximation of the surplus process by a Wiener
process with trend, called a generalized Wiener process.
The process U(t) increases continuously with slot c, which is the premium
per unit time and in random times
1
,
2
, . . . has jumps equal to Z
1
, Z
2
, . . . .
Remember that the risk process X(t) = ct S
t
, t 0 satises the
properties:
1. X(0) = 0;
2. EX(t) = ct tEZ;
3. V ar(X(t)) = tEZ
2
.
The goal is to construct a continuous time process with the same prop-
erties.
Remember the denition of the Wiener process.
Denition 8.1 Continuous time stochastic process {W
t
, t 0} is called a
Wiener process, if
1. W
0
= 0;
2. {W
t
, t 0} has stationary independent increments;
3. Foe every t > 0, W
t
N(0,
2
t), where > 0 is a constant.
Denition 8.2 The continuous time stochastic process {W
t
, t 0} is called
a generalized Wiener process (Wiener process with drift), if it is a Wiener
85
86 CHAPTER 8. DIFFUSION APPROXIMATION
process, with EW
t
= t, i.e. W
t
N(t,
2
t). This process is called also a
diusion process.
We shall prove that the surplus process {U(t), t 0} can be approxi-
mated by a Generalized Wiener process. Consider the limit of the surplus
process U(t), conditioning on large expected number of jumps with small
sizes. Suppose that the expected values and the variances of the processes are
the same. Under these conditions, the surplus process with Poisson counting
process can be approximated by a generalized Wiener process.
Let
= c EZ
and

2
= EZ
2
be the mean value and the variance of the generalized Wiener process. Then
c = +
2
EZ
EZ
2
.
Suppose that the claim size Z is given by Z = Y, for some random
variable Y with arbitrary mean and variance. Then
=

2
EY
2
1

2
and
c = +
2
EY
EY
2
1

.
Let 0, then . Since the processes {S
t
, t 0}, U(t) and X(t), t
0 have stationary independent increments, then the limiting process is the
same. X(0) = 0, consequently we have to prove only that for every t, the
limit of X(t) is normally distributed with parameters t and
2
t.
Let
M
X(t)
(r) = Ee
r(ctS
t
)
= e
[rc+(M
Z
(r)1)]t
be the MGF. Then
log M
X(t)
(r)
t
= rc + [M
Z
(r) 1]
= r[ + EZ] + [1 rEZ +
r
2
2!
EZ
2

r
3
3!
EZ
3
+ . . . 1]
= r +
r
2
2
EZ
2
[
r
3
3!
EZ
3

r
4
4!
EZ
4
+ . . .]
= r +
r
2
2

2

2
[
r
3
3!
EY
3

2 r
4
4!
EY
4
+ . . .].
8.1. RUIN PROBABILITY FOR DIFFUSION PROCESS 87
For 0 we obtain
lim
0
log M
X(t)
(r)
t
= r +
r
2
2

2
,
and then
lim
0
M
X(t)
(r) = e
(r+
r
2
2

2
)t
.
This is the MGF of N(t,
2
t) - distributed random variable, consequently
the limiting process is a generalized Wiener process with mean value t.
From the denition of the surplus process U(t) it follows that the sample
paths are dierentiable everywhere except in the jump points. Since in the
limiting case the number of points increases, the sample paths of the limiting
process are not dierentiable. Also, for 0, the jumps size tends to zero,
consequently the sample paths of the limiting process are continuous with
probability 1.
8.1 Ruin Probability for diusion process
We proved that if W
t
is N(t,
2
t) - distributed Wiener process, U(t) = u+W
t
is a risk process with initial capital U(0) = u. Consider the ruin probability
in a nite time interval (0, ) and let . The ruin probability up to
time is given by
(u, ) = 1(u, ) = P(T
u
< ) = P( min
0<t<
U(t) < 0) = P( min
0<t<
W
t
< u).
Theorem 8.1 The ruin probability of the dened diusion process is given
by
(u, ) =
_

u +

_
+ e

2
u

_
,
where is the standard normal distribution function.
From this result, letting , we obtain
Corrolary 8.1 The ultimate ruin probability is given by
(u) = 1 (u) = P(T
u
< ) = e

2
u
.
Corrolary 8.2 The distribution of the time to ruin, given that ruin occurs
is
(u, )
(u)
= P(T
u
< |T
u
< ) = e
2

2
u

u +

_
+
_

_
, > 0.
(8.1)
88 CHAPTER 8. DIFFUSION APPROXIMATION
Corrolary 8.3 Dierentiation in (8.1) relative to u gives the probability
density function of the time to ruin
f
T
u
() =
u

3
2
e

(u)
2
2
2

, > 0. (8.2)
Substituting
u
2

2
= and
u

= a in the density function we obtain


f
T
u
() =
_

2
3
_1
2
e


2
(
a

)
2
, > 0,
which is the standard Inverse Gaussian distribution.
Hence, the time to ruin conditioning that ruin occurs has an Inverse
Gaussian distribution with expected value
u

and variance
u
2

3
. If = 0 ruin
occurs with probability 1 and the density function is obtain from (8.2) with
= 0, i.e.
f
T
u
() =
u

3
2
e

u
2
2
2

, > 0.
The distribution function is
F
T
u
() = 2
_

_
, > 0.
This is one-sided stable distribution with index
1
2
.
These results could be applied like approximation to the arbitrary risk
process U(t), with Poisson counting process. For the approximated risk pro-
cess we obtain
(u, ) =
_

u + EZ

EZ
2
_
+ e

2EZ
EZ
2
u

u EZ

EZ
2
_
, u > 0, > 0,
(u) = e

2EZ
EZ
2
u
, u > 0
and
f
T
u
() =
u

2EZ
2

3
2
e

(uEZ)
2
2EZ
2
, > 0.
Here is the safety loading factor and c = (1 + )EZ.
Similarly, for a given risk process with Poisson counting process there
exists a simple numerical approximation.
For example, the expected value of the time to ruin if ruin occurs is given
by
ET
u
=
u

=
u
EZ
.
8.1. RUIN PROBABILITY FOR DIFFUSION PROCESS 89
It is easy to see that it depends on four parameters. If the initial capital
is large, the time to ruin increases. The increasing in all the other three
parameters causes a decreasing of the time to ruin.
90 CHAPTER 8. DIFFUSION APPROXIMATION
Chapter 9
Reinsurance
In many cases the premiums to the insurance company are not enough to
carry the risk. This is the case of large claims. In that cases the insurer shares
a part of the risk with other companies. Sharing the risk as well as the premi-
ums is done by reinsurance contracts, which are mutual agreements between
insurance companies. Sometimes the insurance companies have agreements
about reinsuring certain parts of the portfolios.
We consider reinsurance that applies to the individual claims. If the claim
size is z the insurer retains a part h(z), where h : R
+
R
+
is an increasing
function, such that h(0) = 0 and 0 h(z) z for all z 0. The reinsurer
covers the remain part z h(z). We assume that reinsurance premiums are
payable continuously and that the reinsurer pays its share of a claim as soon
as that claim occurs. The function h(x) determines the rule of reinsurance.
The aggregate sum of claims for insurer is S
I
t
=

N(t)
i=1
h(Z
i
). The sum of
claims for reinsurer is S
R
t
= S
t
S
I
t
.
9.1 Proportional Reinsurance
Suppose the insurer chooses proportional reinsurance with retention level
b [0, 1]. The function h is h(z) = bz. The premium rate for the reinsurance
is given by
(1 + )(1 b),
where > 0 is the relative safety loading, dened by the reinsurance com-
pany. We consider the case > . The premium rate for the insurer is:
[(1 + ) (1 + )(1 b)] = [b(1 + ) ( )],
and the surplus process becomes
91
92 CHAPTER 9. REINSURANCE
U(t, b) = u + [b(1 + ) ( )]t
N(t)

k=1
bZ
k
. (9.1)
In order that the net prot condition is fullled we need
[b(1 + ) ( )]
b
> 1,
i.e.
b > 1

.
Let M
Z
(r) be the moment generating function of the individual claim
amount distribution evaluated at r. Then the adjustment coecient R(b)
under proportional reinsurance is the unique positive solution of the equation
[M
Z
(br) 1] [b(1 + ) ( )]r = 0. (9.2)
Let (u, b) denote the probability of ultimate ruin when the proportional
reinsurance is chosen. Then
(u, b) = P(U(t, b) < 0 for some t > 0)
Our objective is to nd the retention level that minimizes (u, b). Ac-
cording the Lundberg inequality, the retention level will be optimal, if the
corresponding Lundberg exponent R is maximal. We know that there is a
unique b [0, 1] where the maximum is attained. If the maximizer b > 1, it
follows from the uni-modality that the optimal b is 1, i.e. no reinsurance is
chosen.
The next result gives the optimal retention level b and maximal adjust-
ment coecient R(b) (see [9]).
Lemma 9.1 The solution of equation (9.2) is given by
R(b(r)) =
(1 + )r [1 M
Z
(r)]
( )
, (9.3)
where b r(b) is invertible.
Proof. Assume that r(b) = bR((b)), where R(b) will be the maximal value of
the adjustment coecient and r(b) is invertible. If we consider the function
r b(r), it follows that
b(r) =
( )r
(1 + )r [1 M
Z
(r)]
. (9.4)
9.2. EXCESS - OF - LOSS REINSURANCE(XL) 93
Now R(b(r)) =
r
b(r)
in details is given by (9.3).
2
Theorem 9.1 Assume that M
Z
(r) < . Suppose there is a unique solution
r to
M

Z
(r) (1 + ) = 0. (9.5)
Then r > 0, the maximal value of R(b(r)) and the retention level b(r) are
given by (9.3) and (9.4).
Proof. The necessary condition for maximizing the value of the adjustment
coecient is given by equation (9.5).
Since R

(b(0)) =

> 0, the function R(b(r)) is strictly increasing at


0. The second derivative in zero R

(b(0)) =
1
()
EZ
2
< 0 shows that
R(b(r)) is strictly concave. Consequently, the function R(b(r)) has an unique
maximum in r, which is the solution of (9.3). The retention level is given by
(9.4).
2
Remark 9.1 Note that the value of the adjustment coecient does not de-
pend on c but on the relative safety loadings only.
9.2 Excess - of - Loss Reinsurance(XL)
The Excess-of-Loss reinsurance is nonproportional type of reinsurance. The
insurer covers each individual claim up to a certain retention level M, i.e.
when a claim of size Z occurs the insurer pays Z
M
= min(Z, M) = h(z)
and the reinsurer Z
R
= Z Z
M
= max(Z M, 0) = (Z M)
+
so that
Z = Z
M
+ Z
R
. Suppose the number of claims N(t) follows an ordinary
renewal process. Hence the insurer risk process at time t is
X
M
(t) = (c c
M
)t
N(t)

i=1
min(Z
i
, M),
where c
M
is the XL reinsurance premium. For a given M, the adjustment
coecient R
M
is the unique positive root of
g
M
(r) = 1,
if it exists with
g
M
(r) = E[e
rZ
M
]E[e
(cc
M
)rT
].
94 CHAPTER 9. REINSURANCE
9.3 Stop - Loss Reinsurance
Stop - Loss reinsurance works similarly to the XL reinsurance, but it cov-
ers the total amount of claims. For stop-loss contract with retention level
(deductible) d, the amount paid by reinsurer to the insurer is
I
d
=
_
0, if S d
S d, if S > d.
Sometimes: I
d
= (S d)
+
. Note that I
d
as a function of the aggregate
claims S is also a random variable. The amount of claims retained by the
insurer is
min(S, d) = S I
d
=
_
S, if S d
d, if S > d.
Thus, the amount retained is bounded by d, which explains the name stop-
loss contract.
The expected claims paid by the reinsurer:
EI
d
=
_

d
(x d)f
S
(x)dx (9.6)
= ES d +
_
d
0
(d x)f
S
(x)dx (9.7)
=
_

d
[1 F
S
(x)]dx (9.8)
= ES
_
d
0
[1 F
S
(x)]dx. (9.9)
When ES is available (9.7) and (9.9) are preferable by numerical integration.
(9.8) and (9.9) hold for general distribution, including discrete and mixed. If
the distribution is given, (9.6) is the most tractable formula.
Example 9.1 Let S Gamma(, ). Then
EI
d
=
_

d
xf(x)dx d[1 F
S
(d)]
=
_

d

()
x

e
x
dx d[1 G(d; , )]
=

[1 G(d; + 1, )] d[1 G(d; , )],


where G(d; , ) =
1
()
_
d
0
x
1
e
x
dx.
Bibliography
[1] Asmussen S. (2000). Ruin Probabilities, World Scientic, Singapure.
[2] Bowers N.L., Gerber H.U., Hhickman J., Jones D. and Nesbit C. (1997).
Actuarial Mathematics, Society of Actuaries, Itasca.
[3] Daykin C.D., Pentik amen T. and Pesonen M. (1994). Practical Risk
Theory for Actuaries, Chapman & Hall, London.
[4] Embrechts P., Kluppelberg C. and Mikosh T. (1997). Modelling extremal
events, Appl.Math. 33, Springer, Berlin.
[5] Galambos J. and Kotz S. (1978). Characterizations of Probability Dis-
tributions, Springer - Verlag.
[6] Gerber H.U. (1979). An Introduction to Mathematical Risk Theory,
Richard D.Irwin, Homewood.
[7] Grandell, J. (1991). Aspects of Risk Theory, Springer-Verlag, New York.
[8] Grandell, J. (1997). Mixed Poisson Processes, Chapman & Hall, London.
[9] Hald M. and Schmidli H. (2004). On the maximisation of the adjustment
coecient under proportional reinsurance, ASTIN Bulletin, 34(1), 75-
83.
[10] Kaas R., Goovaerts M., Dhaene J. and Denuit M. (2001). Modern Ac-
tuarial Risk Theory, Klawer, Boston.
[11] Klugman S.A., Panjer H.H. and Willmot G.E. (1998). Loss Models. From
data to decisions, Wiley.
[12] Johnson N.L., Kotz S. and Kemp A.W. (1992). Univariate Discrete Dis-
tributions, 2nd ed., Wiley Series in Probability and Mathematical Statis-
tics.
95
96 BIBLIOGRAPHY
[13] Mikosh T. (2004). Non - life insurance mathematics, Springer.
[14] Minkova L.D. (2001). A Family of Compound Discrete Distributions, J.
Bulg. Acad. Sci., 54(2), 11-14.
[15] Minkova L.D. (2002). A generalization of the classical discrete distribu-
tions, Comm. Statist.Theory and Methods, v.31(6), 871-888.
[16] Minkova L.D. (2004). The Polya - Aeppli process and ruin problems, J.
Appl. Math. Stoch. Analysis, 3, 221-234.
[17] Rolski, T., Schmidli, H., Schmidt, V. and Teugels, J. (1999). Stochastic
Processes for Insurance and Finance, John Wiley & Sons, Chichester.
[18] Ross S.M. (2007). Introduction to Probability Models, 9th edition, El-
sevier Inc.
[19] Serfozo R. (2009). Basic of Applied Stochastic Processes, Springer.
[20] Willmot G.E. and Lin X.S. (2001). Lundberg Approximations for Com-
pound Distributions with Insurance Applications, Springer.

Vous aimerez peut-être aussi