UECM3463May16 C01 (2p)

May 2016 Chapter 1 Models For Claim Severity
UNIVERSITI TUNKU ABDUL RAHMAN

Department of Mathematics and Actuarial Science
CONTENTS
1 Models for Claim Severities

1.1 Introduction . . . . . . . . . . . .
1.2 Some Parametric Claim Size Distributions . . . . . . . . . . . . . .
1.3 Basic Distributional Quantiles . . .
1.3.1 Moments: . . . . . . . . . .
1.3.2 Special Functions of Moments
1.3.3 Percentiles . . . . . . . . .
1.4 Classifying and Creating Distribution . . . . . . . . . . . . . . . . .
1.4.1 Scaling . . . . . . . . . . .
1.4.2 Multiplication by a Constant
1.4.3 Raising to a Power . . . . .
1.4.4 Exponentiation . . . . . . .
1.4.5 Continuous Mixing . . . . .
1.4.6 Splicing . . . . . . . . . . .
UECM3463 Loss Models
4
4
5
10
10
13
24
30
30
31
35
37
39
43
1.5 Severity with Coverage Modifications . . . . . . . . . . . . . . . . 47

1.5.1 The Limited Loss Variable . 47
1.5.2 Deductibles . . . . . . . . . 57
1.5.3 Franchise Deductible . . . 75
1.5.4 Loss Elimination Ratio (LER) 79
1.5.5 The Effect of Inflation on
Ordinary Deductible . . . . 81
1.5.6 The Effect of Inflation on
Policy Limit . . . . . . . . 83
1.5.7 Coinsurance, Deductible, and
Limits . . . . . . . . . . . 85
1.5.8 Bonus . . . . . . . . . . . . 92
1.6 Reinsurance . . . . . . . . . . . . 97
1.6.1 Excess of loss reinsurance the insurer . . . . . . . . . 97
1.6.2 Excess of loss reinsurance the reinsurer . . . . . . . . 101
1.7 Tails of distributions . . . . . . . . 103
1.7.1 Concept of Tail Weight . . 103
1.7.2 Classification of Tail weight 104
1.8 Measure of Risk . . . . . . . . .

1.8.1 Coherent . . . . . . . . .
1.8.2 Value-at-Risk . . . . . . .
1.8.3 Tail-Value-at-Risk . . . .
1.8.4 Distortion Risk Measures
.
.
.
.
.
108
108
112
116
124
Models for Claim Severities
1.1
Introduction
Given that a claim occurs, the (individual)
claim size X is typically referred to as claim

severity.
While typically this may be of continuous ran-
dom variables, sometimes claim sizes can be

considered discrete.
When modeling claims severity, insurers are
usually concerned with the tails of the distribution. There are certain types of insurance
contracts with what are called long tails.
1.2
Some Parametric Claim Size Distributions
Normal - easy to work with, but careful with
getting negative claims. Insurance claims usually are never negative.

Gamma/Exponential - use this if the tail of
distribution is considered light, applicable

for example with damage to automobiles.
Lognormal - somewhat heavier tails, applica-
ble for example with fire insurance.

Burr/Pareto - used for heavy-tailed business,
such as liability insurance.

Inverse Gaussian - not very popular because
complicated mathematically.
0.8
1.0
Gamma Densities
0.0
0.2
0.4
f(x)
0.6
=5
= 1.5
=1
= 0.5
= 0.25
Inverse Gaussian densities with various and
1.3
1.3.1
Basic Distributional Quantiles

Moments:
Definition 1. k th raw moment

The k th raw moment of X is define as
Z
xk f (x)dx
k = E(X k ) =
if X is continuous
k = E(X k ) =
X
j
if X is discrete.
xkj p(xj )
10
Definition 2. k th central moment

The k th central moment of X is define as
k = E[(X )k ] =
if X is continuous
k = E[(X )k )] =
if X is discrete.
(x )k f (x)dx
X
(xj )k p(xj )
j
11
12
Expectation is linear, so the central moments can

be calculated from the raw moments by binomial expansion. In binomial expansion, the last
2 terms always merge, so we have
2 = 2 2 instead of 2 21 + 2
3 = 3 32 + 23 instead of
3 32 + 312 3
4 = 4 43 + 622 34 instead of
4 43 + 622 413 + 4
1.3.2
13
Special Functions of Moments
Variance: V (X) = 2 and is denoted by 2.
Standard Deviation: = 2.
Coefficient of Variation: CV =
.
The coefficient of variation expresses the standard deviation as a percentage of the sample
mean. This is useful when interest is in the size
of variation relative to the size of the observation, and it has the advantage that the coefficient of variation is INDEPENDENT OF the
UNITS of observation. For example, the value
of the standard deviation of a set of weights
will be different depending on whether they
are measured in kilograms or pounds. The coefficient of variation, however, will be the same
in both cases as it does not depend on the unit
of measurement.
14
Skewness: 1 = 33
Skewness is a measure of symmetry, or more

precisely, the lack of symmetry. A distribution, or data set, is symmetric if it looks the
same to the left and right of the center point.
Negative values for the skewness indicate data
that are skewed left and positive values for
the skewness indicate data that are skewed
right. By skewed left, we mean that the left
tail is long relative to the right tail. Similarly,
skewed right means that the right tail is long
relative to the left tail. The skewness for a
normal distribution is zero, and any symmetric data should have a skewness near zero.
15
Kurtosis: 2 = 44 .
Kurtosis is a measure of whether the data are

peaked or flat relative to a normal distribution.
That is, data sets with high kurtosis tend to
have a distinct peak near the mean, decline
rather rapidly, and have heavy tails. Data sets
with low kurtosis tend to have a flat top near
the mean rather than a sharp peak. A uniform
distribution would be the extreme case. The
kurtosis for a standard normal distribution is
three.
Mode
A mode is x such that f (x) is maximized (or

p(x) for discrete distribution).
Moment Generating Function
The moment generating function (MGF) is defined by

MX (t) = E(etX )
16
Notes:
M (n)(0) = E(X n)
MX1+...Xn (t) = [MX (t)]n when Xis are
identically and independently distributed.
Probability Generating Function
The probability generating function (PGF) is

defined by
PX (z) = E(z X )
It is important to realise that we cannot have
intuition about PGFs because they do not correspond to anything which is directly observable.
Notes:
PGFs make calculations of expectations and
of some probabilities very easy.
* P (1) = E(X)
* P (1) = E[X(X 1)]
* P (3)(1) = E[X(X 1)(X 2)]
17
PGFs make sums of independent random

variables easy to handle. i.e.,
PX1+...+Xn (z) = [PX (z)]n
when Xis are identically and independently
distributed.
Cumulant Generating Function
The cumulant-generating function K(t), is the

natural logarithm of the moment-generating
function:
K(t) = ln E(etX ) = ln MX (t)
The cumulants n are obtained from a power
series expansion of the cumulant generating
function:
X
tn
K(t) =
n .
n!
n=1
This expansion is a Maclaurin series, so that

nth cumulant can be obtained by differentiating the above expansion n times and evaluat-
18
ing the result at zero:

n = K (n)(0).
The first cumulant is the expected value; the
second and third cumulants are respectively
the second and third central moments (the
second central moment is the variance); but
the higher cumulants are neither moments nor
central moments, but rather more complicated
polynomial functions of the moments.
Example 1.
A random variable X has a gamma distribution
with parameters = 5 and = 0.1, calculate the mode, CV, skewness and the kurtosis.
mode = 0.4, CV = 0.4472
Skewness = 0.8945, Kurtosis = 4.2
Sol:
1 x4ex/.1
f (x) = (5).1
5
f (x) = c(4x3ex/.1 10x4ex/.1 = 0
x = 4/10 = 0.4 mode = .4
= E(X) = 0.5
19
p
= 5(.12) = .2236
CV = = .2236
.5 = 0.4472
2 = E(X 2) = .12(5)(6) = .3
3 = E(X 3) = .13(5)(6)(7) = 0.21
3 = 3 32+22 = .213(.3)(.5)+2(.53) =
.01
or K(t) ln(1 t), K (t) = (1 t)1
K (t) = 2(1 t)2,
K (t) = 2 3(1 t)3
3 = K (0) = 2 3 = 2(5)(0.13) = 0.01
.01 = 0.8945
Skewness = 1 = 33 = .2236
3
4 = E(X 4) = .14(5)(6)(7)(8) = .168
4 = 443+62234 = .1684(.21)(.5)+
6(.3)(.52) 3(.54) = 0.0105
.0105 = 4.2
Kurtosis = 2 = 44 = .2236
4
Example 2.
Claim severity has the following distribution:
Claim Size 100 200 300 400 500
Probability 0.05 0.20 0.50 0.20 0.05
20
Determine the distributions skewness and kurtosis. 0,3.125

Sol:
= E(X) = 300; = 89.4427
3 = E(X )3 = 0
E(X)3
1 =
= 0
3
4 = E(X )4 = 2 108
E(X)4
2108 = 3.125
2 =
=
4
89.44274
21
Example 3.
You are given:
(i) For any random variable X with finite first
three moments, the skewness of the distribution of X is denoted Sk(X).
(ii) X and Y are iid random variables with mean
= 0 and finite second and third moments.
E(X + Y )3
= E(X 3) + 3E(X 2)E(Y ) + 3E(X)E(Y 2) + E(Y 3
= E(X 3) + 0 + 0 + E(Y 3)
= 2E(X 3)
2E(X 3)
Sk(X+Y ) = 3/2 3 = 21/2Sk(X) Sk(X).
2 X
So III is true.
Which of the following statements must be true?

(I) 2Sk(X) = Sk(2X)
(II) Sk(Y ) = Sk(Y )
(III) |Sk(X)| Sk(X + Y )|
Sol:
1. Skewness is dimensionless, so Sk(cX) = Sk(X)
c > 0. I is false.
2. Negating a r.v. negates the moments but not

the denominator since > 0. So II is false.
2 since X and Y are iid.
2
= 2X
3. X+Y
3
3
X+Y
= 23/2X
22
23
Example 4.
Let X be a discrete random variable with probability generating function
PX (z) = 0.44z 180+0.24z 540+0.17z 900+0.07z 1260+0.08z 1620
Calculate the skewness of X.

Sol:
x
180 540 900 1,260 1,620
P (X = x) 0.44 0.24 0.17 0.07 0.08
= E(X) = 579.60; = 455.069
3 = E(X )3 = 89643065.47
= 0.9512
1 = 89643065.47
455.0693
1.3.3
24
Percentiles
Definition 3.
The 100pth percentile of a random variable is any
value p such that F (p) p F (p).
If the distribution function has a value of p for
one and only one x value, then the percentile

is uniquely defined.
If the distribution function jumps from a value
below p to a value above p, then the percentile

is at the location of the jump.
If the distribution function is constant at a
value of p over a range of values, then any value

in that range can be used as the percentile.
25
Example 5.
Suppose
Example 6.
Suppose
x<0
0,
F (x) = 0.01x, 0 x < 100
1,
x 100.
Determine the 50th and 80th percentiles. 50, 80

Sol:
0.01x.5 = .5
.5 = 50
x.5 = .01
0.01x.8 = .8
.8 = 80
x.8 = .01
0,
0.5,
0.75,
F (x) =
0.87,
0.95,
1,
x < 0,
0 x < 1,
1 x < 2,
2 x < 3,
3 x < 4,
x4
Determine the 50th and 80th percentiles. [0,1], 2

Sol:
x.5 = [0,1]
x.8 = 2
26
Example 7.
A random variable X has the following distribution:
x P (X = x)
1
0.20
3
0.25
7
0.45
8
0.10
Calculate the 50th and 90th percentiles of X. 7, [7,8]
Sol:
x.5 = 7
x.9 = [7,8]
0,
0.20,
F (x) = 0.45,
0.90,
1.0,
x<1
1x<3
3x<7
7x<8
x8
27
28
Example 8.
Losses have a Pareto distribution with parameter
and . The 10th percentile is k. The 90th
percentile is 5 3k. Determine the value of .
2
Sol:

= .9

k+
= .9 (1)
2k

= .1
53k+
= .1 (2)
3(2k)
(1)
,3 = 9
(2)
= 2
29
30
Example 9.
You are given the following information about a
study of individual claims:
1. 20th percentile = 18.25 and
1.4
2. 80th percentile = 35.80.

Parameters and of a lognormal distribution
are estimated using percentile matching.Determine
the probability that a claim is greater than 30 using the fitted lognormal distribution. .3446
Definition 4. A parametric distribution is a set

of distribution functions, each member of which
is determined by specifying one or more values
called parameters. The number of parameters is
fixed and finite.
Sol:

ln 18.25
= .2
ln 18.25
= .845 (1)

ln 35.8
= .8
ln 35.8
= .845 (2)
(1) ln 18.25
= 1
,
(2) ln 35.8
= 3.2411
= ln 35.83.2411
= 0.3986
.845

ln
303.2411
= 1(0.4) =
P (X > 30) = 1
0.3986
1 0.6554 = .3446
1.4.1
Classifying and Creating Distribution

Scaling
Definition 5. A parametric distribution is a scale

distribution if, when a random variable from that
set of distribution is multiplied by a positive constant, the resulting random variable is also in that
set of distributions.
Note: All of the continuous distributions in Appendix A (except lognormal and inverse Gaussian) are scale families.
1.4.2
31
Multiplication by a Constant
This transformation is equivalent to applying inflation uniformly across all loss levels and is known
as a change of scale. For example, if this years
losses are given by a random variable X, the uniform inflation of 5% indicates that next years
losses can be modeled with the random variable
Y = 1.05X.
Theorem 1. Let X be a continuous random
variable with pdf fX (x) and cdf FX (x). Let
Y = cX with c > 0. Then
y
1 y
, fY (y) = fX
FY (y) = FX
c
c
c
Example 10.
X is a claim size in 2011, and has a Pareto distribution with parameters = 5 and = 20.
There is an 8% inflation in 2012. Determine the
distribution of the inflated variable.
Sol:
X P areto( = 5, X = 20)
Y = 1.08X, Y = 1.08X = 1.08(20) = 21.6
Y P areto( = 5, Y = 21.6)
Notes:
If X has scale parameter and other parameters, then cX has scale parameter c and the
same other parameters.
If X lognormal(, ), then
cX lognormal( + lnc, ).
32
33
Example 11.
In 2011, the claim amounts for a certain line
of business were normally distributed with mean
= 5000 and variance 2 = 8000. Inflation of
5% impacted all claims uniformly from 2011 to
2012. What is the distribution for claim amounts
in 2012.
Sol:
X N ( = 5000, 2 = 8000)
Y = 1.05X
E(Y ) = 1.05E(X) = 1.05(5000) = 5250
V (Y ) = 1.052V (X) = 1.052(8000) = 8820
Y N (Y = 5250, Y2 = 8820)
34
Example 12.
Losses in 2007 follow the lognormal distribution
with parameters = 5 and = 8. Inflation of
10% impacts all claims uniformly from 2007 to
2008. Determine the probability that losses in
2008 exceed 9. .6406
Sol:
X ln N (X = 5, X = 8)
Y = 1.1X; Y = 5 + ln(1.1) = 5.0953
Y ln N (Y = 5.0953,
= 8)
P (Y > 9) = 1 ln 95.0953
= 1 (.36)
8
= .6406
1.4.3
35
Raising to a Power
Theorem 2. Let X be a continuous random

variable with pdf fX (x) and cdf FX (x) with FX (0) =
1
0. Let Y = X . Then if > 0,
FY (y) = FX (y ), fY (y) = y 1fX (y ), y > 0
while if < 0,
FY (y) = 1 FX (y ), fY (y) = y 1fX (y ).
36
Definition 6.
When raising a distribution to a power, if > 0,
the resulting distribution is called transformed; if
= 1, it is called inverse, and if if < 0 (but
not -1), it is called inverse transformed.
Example 13.
Suppose X exp(1). Determine the cdf of the
inverse, transformed, and inverse transformed exponential distribution.
Sol:
f (x) = ex; F (x) = 1 ex
If Y = X 1, P (Y y) = P (X 1 y) =
P (X > y1 ) = 1 e1/y Y InvExp(1)
If Y = X 1/ , > 0, P (Y y) = P (X 1/
y) = P (X < y ) = 1 e(y)
If Y = X 1/ , < 0, P (Y y) = P (X 1/
y) = P (X > y ) = e(y)
1.4.4
37
Exponentiation
Theorem 3.
Let X be a continuous random variable with pdf
fX (x) and cdf FX (x) with fX (0) > 0 for all real
x. Let Y = exp(X). Then, for y > 0,
1
FY (y) = FX (lny), fY (y) = fX (lny)
y
38
Example 14.
Let X have the normal distribution with mean
and variance 2. Determine the cdf of Y = eX .
Sol:
X
P (Y
y)
= P (e y) = P (X lny) =
ln y
Y LnN (, )
1.4.5
39
Continuous Mixing
Theorem 4.
Let X have pdf fX|(x|) and cdf FX|(x|),
where is a parameter of X, while X may have
other parameters, there are not relevant. Let
be a realization of the random variable with
pdf |(). Then the unconditional pdf of X is
Z
fX (x) = fX|(x|)()d
where the integral is taken over all values of
with positive probability.
Two distributions that are derived from continuous mixing are:

1. Pareto distribution:
If
X| Exp()
and
Inverse Gamma(, )
then
X Pareto(, )
2. Negative Binomial: If
X| Poisson()
and
gamma(, )
then
X Negative Binomial(r = , )
40
41
Example 15. You are given the following:

The amount of an individual claim has an ex-
ponential distribution given by:

1
f (x|) = ex/ x > 0, > 0
The parameter has a probability density function given by:

400
() = 3 e20/ > 0
Determine the mean of the claim severity distribution. 20
Example 16.
The claim count N has a Poisson distribution
with mean . has a gamma distribution with
parameters = 2 and = 0.5. Calculate the
probability that N = 1. 0.2963
Sol:
N | P OI(); gamma( = 2, = 0.5)
N N B(r = 2, = 0.5)
2(.5)
P (N = 1) = 1.53 = 0.2963
Sol:
X| Exp(); InvGamma( = 2, =
20)
X P areto( = 2, = 20)
20 = 20
E(X) = 21
42
1.4.6
43
Splicing
Definition 7. A k-component spliced distribution has a density function that can be expressed as follows:
a1f1(x), c0 < x < c1,
a f (x), c < x < c ,

1
2
fX (x) = . 2 2
a f (x), c
k k
k1 < x < ck .
For j = 1, . . . , k, each aj > 0 and each fj (x)

must be a legitimate density function with all
probability on the interval cj1, cj ). Also a1 +
ak = 1.
Note that the splicing does not ensure that the resulting density function will be continuous. Such
a restriction could be added to the specification.
Example 17.
Create a two-component spliced model using an
exponential distribution with meaN from 0 to c
and a pareto distribution with parameters and
from c to .
Sol:
The basic format is
1 x/
a1 e
, 0<x<c
1ec/
f (x) =
a2 (c+) , c < x < .

(x+)+1
However, we must force the density function to

integrate to 1. All that is needed is to let a1 = v
and a2 = 1 v. The spliced density function
bocomes
1 x/
v e
,
0<x<c
1ec/
f (x) =
(1 v) (c+) , c < x < .

(x+)+1
, , , c > 0, 0 < v < 1.
44
45
Example 18.
An insurance loss is being modeled as a continuous two-spliced distribution as follows:
(
c1ex/100, 0 < x < 100
fX (x) =
c2ex/200, x 100
Calculate the average loss.

Sol:
Rewrite,
(
vf1 (x),
0 < x < 100
fX (x) =
(1 v)f2 (x), x 100
where,
f1 (x) =
1 x/100
e
100
1e1
1 x/200
e
200
e.5
and
c1[100(1 e1 + e.5(200)eA.5] = 1
c1 = 0.007310586
c2 = 0.007310586e.5 = 0.0044334095
R
R
E(X) = 0100 c1xex/100dx + 100
c2xex/200dx
R
= 0.007310586 0100 xex/100dx
R x/200
+0.0044334095 100
xe
dx
= 0.007310586[100(1 e1 e1)]
+0.0044334095[200(e.5 + .5e.5]
= 0.193175 + 0.80685
= 1
f2 (x) =
For continuity,
1 100/100
1 100/200
e
e
v h1001e1
= (1 i v) 200 e.5
1
1
e
= 200
v 100(1e
1 ) + 200
v = 0.46212
R 100
R
E(X) = 0 xvf1 (x)dx + 100 x(1 v)f2 (x)dx
1
R 100
xex/100
= 0 (0.46212) 1001e1 dx
1
R
xex/200
+ 100 (0.53788) 200 e.5
dx
R 100 x/100
0.46212
dx
= 100(1e1 ) 0 xe
R
0.53788
x/200 dx
+ 200e
.5 100 xe
0.46212
= 100(1e1 ) [100(1 e1 e1 )]dx
0.53788
.5 + .5e.5 )]dx
+ 200e
.5 [200(e
= 0.19318 + 0.80682
= 1
1
continuity, c1e1
= c2e.5, c2 = c1e.5
R By
R
100
x/100dx + c e1/200dx = 1
c
e
1
0
100 2
1
c1[100(1 e ] + c2[200e.5] = 1
or
46
47
1.5
1.5.1
Severity with Coverage Modifications

The Limited Loss Variable
The random variable for the amount not paid
due to the deductible is the minimum of X

and d. This random variable is
(
X, X < d,
X d=
d, X d.
It expected value is called the limited expected
value.
E(Xd) =
xf (x)dx+dS(d) =
if the variable is continuous, and
E(X d) =
48
xj P (xj ) + dS(d)
xd
if the variable is discrete.
Definition 8. The limited loss variable (Payment per loss with claims limit)
Z d
Z d
This variable is also called the right censored

variable. The expected value is the expected
amount not paid for loss below d, the integral,
plus the expected amount not paid for above
d, which is dS(d). An insurance phenomena
that related to this variable is the existence
of a policy limit that sets a maximum on the
benefit to be paid. Note that (X d)+ +
X d = X. That is, buying one policy with
limit of d and another with a deductible of d
is equivalent to buying full coverage.
S(x)dx
49
The k th moment of X d is
k
E(Xd) =
x f (x)dx+d S(d) =
kxk1S(x)dx
50
Example 19.
The claim size (X) distribution for an insurance
coverage is modeled as a Pareto distribution with
parameters = 3, = 1000. Calculate E(X
3000) and E(X 3000)2. 468.75, 562500
Sol:
E(X d)k =
xkj P (xj ) + dk S(d)
xd

Remark:
If x , then
Rd k
k
E(X d) = x f (x)dx + dk S(d)
R d k1
k
S(x)dx
= s() + kx
R
E(X 3000) = 03000 S(x)dx
R 3000 1000 3
= 0
x+1000 dx
Let u = x + 1000, du = dx
R 4000
= 1000
10003u3du
h 2 i4000
3
= 1000 u2
1000
h
i
2
2
1000
4000
3
= 1000
2 +
2
= 468.75

2

1000
Using formula, E(X3000) = 1000
21 1 4000
= 468.75
E(X 3000)2
R 3000
1000
2x x+1000
3
dx
= 0
Let u = x + 1000, du = dx
R 4000
3
= 2(1000 ) 1000 (u + 1000)u3du
R 4000 2
3
= 2(1000 ) 1000 (u 1000u3)du
h 1
i4000
u
3
2
= 2(1000 ) 1 + 500u
1000
3
2
= 2(1000 )[500(4000 ) 40001
500(10001) + 10001]
= 562,500
51
Example 20.
The claim size (X) distribution for an insurance
coverage is modeled as a Single Pareto distribution with parameters and , find E(X x).
Sol:

f (x) = x+1 , x > ; S(x) = x , x >

E(X x)
Rx
= + y dy
h +1 ix
= + y+1
h +1 +1 i
= + 1 x1

,x >
= 1
(1)x1
52
53
54
Example 21.
Forty observed losses have been recorded in thousands of dollars and are grouped as follows:
Interval Number of Losses Total Losses
(1, 4/3)
16
20
[4/3, 2)
10
15
[2, 4)
10
35
[4, )
4
20
Example 22.
Let X be a random variable with discrete loss
distribution given by
Sol:
V (X 300) = 7100
Let X be the size of the loss random variable.

\
Determine E(X
2). 1.575
15
X 2 20
16 = 1.25 10 = 1.5 2 2
16
10
10 4
Prob
40
40
40 10
x
100 200 300 400 500
P (X = x) 0.55 0.20 0.10 0.08 0.07
Calculate V (X 300). 7100
Sol:
x 300 100 200 300 300 300

P (X = x) 0.55 0.20 0.10 0.08 0.07
10
14
E(X 2) = 1.25( 16
40 ) + 1.5( 40 ) + 2( 40 ) = 1.575
55
Example 23.
Claim severity follows a single-parameter Pareto
distribution with = 1 and = 1000. An insurance coverage has a claims limit of 10,000. Determine the mean and variance of the claim severity.
3302.585, 8,092,932
Sol:
R 10,000
S(x)dx
E(X 10, 000) = 0
R 10,000 1000
= 1, 000 + 1000
x dx
10,000
= 1, 000 + 1000[ln x]1000
= 1000 + 1000[ln 10, 000 ln 1, 000]
= 3302.585
R 10000
2
2
E(X 10, 000) = 1000 + 1000 2xS(x)dx
R 10000
2
= 1000 + 1000 2x( 1000
x )dx
R
10000
= 10002 + 1000
2000dx
= 10002 + [2000x]10000
1000
2
= 1000 + 2000[10000 1000]
= 19,000,000
V (X 10, 000) = 19, 000, 000 3302.5852 =
8,092,932
56
Example 24.
Loss X is being modeled as an Inverse Exponential random variable with density as follows:
1
f (x|) = 2 e/xfor x > 0,
x
where the parameter is an Exponential random variable with mean parameter 4. Calculate
E(X 2). 1.6219
Sol:
X| InvExp(); Exp(4)
R
F (x) = 0 e/x( 41 )e/4d
R (x+4)
1
= 4 0 e 4x d
4x ]
= 14 [ x+4
x
= x+4

R2
x
E(X 2) = 0 1
dx
R 2 4 x+4
= 0 x+4 dx
= 4[ln(x + 4)]20
= 4[ln 6 ln 4]
= 1.6219
1.5.2
57
Deductibles
An ordinary deductible d means that the first d

of each claim is not paid. An ordinary deductible
modifies a random variable into either the left
censored and shifted or excess loss variable.
The per-loss variable is
(
0,
X d,
Y L = (X d)+ =
X d, X > d,
For a given value of d with P (X > d) > 0,
the per-payment variable is

(
U ndef ined, X d,
YP =
X d,
X > d,
The corresponding densities are:
f (y + d)
,y > 0
fY P = X
SX (d)
(
FX (d),
y=0
fY L (y) =
fX (y + d), y > 0
The expected value of (X d)+ can be calcu-
lated from
E(Xd)+ =
(xd)f (x)dx =
S(x)dx

X
E(X d)+ =
(xj d)p(xj )
xj >d

The k th moment of (X d)+ is define as
E[(X d)k+] =
Z
d
(x d)k f (x)dx

X
k
(xj d)k p(xj )
E[(X d)+] =
xj >d

58
59
Expected cost per payment(the mean excess
loss)
E(Y P ) = E(X d|X > d) = eX (d)
E(Xd)
= 1F (d)+
R
d (xd)f (x)dx
S(d)
Theorem 5. For an ordinary deductible,

E(Y L) = E(X d)+ = E(X) E(X d)
and
E(Y L)
E(X) E(X d)
P
E(Y ) =
=
1 F (d)
The k th moment of the excess loss variable is

[E(Xd)k+]
k
eX (d)= 1F (d)
R
(xd)k f (x)dx
d
=
S(d)
Special cases
Distribution
eX (d)
Exp()
+d
Pareto(, )
1
d ,d
Single Pareto(, ) 1
60
1 F (d)
61
Example 25.
A random sample of auto glass claims has yielded
the following observed claim amounts:
1000 1, 250 2, 000 2, 500 3, 000
62
Example 26.
Claim severity has the following distribution:
Claim Size 50
150 500 1000 2000 5000 10000
Probability 0.305 0.225 0.220 0.155 0.055 0.030 0.010
Calculate E[(X 1, 500)+], E[(X 1, 500)3+]

and E[X 1, 500]. 600, 9 108, 1,350
Determine E[(X 120)+] and V [(X 120)+].

575.35, 1,705,942
Sol:
Sol:
X
1000 1250
(X 1500)+ 0
0
(X 1500)3+ 0
0
X 1500 1000 1250
2000
500
5003
1500
E[(X 1, 500)+] = 600

E[(X 1, 500)3+] = 9 108
E[X 1, 500] = 1350
2500
1000
10003
1500
3000
1500
15003
1500
Claim Size 50
150 500 1000 2000 5000 10000
(X 120)+ 0
30
380 880 1880 4880 9880
Probability 0.305 0.225 0.220 0.155 0.055 0.030 0.010
E[(X 120)+] = 575.35

V [(X 120)+] = 1,705,942
63
Example 27.
The aggregate losses for an insured, X follows
an exponential distribution with mean $1 million. An insurance policy pays for aggregate
losses that exceed twice the expected value of
X. Calculate the expected loss for the policy.
135,335.3
Sol:
S(x) = ex
R
E(X 2)+ = R2 S(x)dx
= 2 exdx
= e2
The expected value = 1, 000, 000e2 = 135,335.3
64
Example 28.
A random sample of auto glass claims has yielded
the following observed claim amounts:
1, 000 1, 250 2, 000 2, 500 3, 000
What is the value of the empirical mean excess
loss function at x = 1, 500? 1,000
Sol:
(X 1500)+ 0 0 500 1000 1500
E(X 1500)+ = 600
S(1500) = 35
600 = 1000
eX (1500) = 3/5
65
Example 29.
For an insurance, losses, X, has the following
distribution:
Claim Size 50
150 500 1000 2000 5000 10000
Probability 0.305 0.225 0.220 0.155 0.055 0.030 0.010
The insurance has an ordinary deductible of

800 per loss, Y P is the payment per payment
random variable, calculate V (Y P ). 4,256,400
Example 30.
Claim sizes follow a Pareto distribution with
parameters = 2, = 1000.
(a) Determine the mean excess loss at 2000.
3,000
Sol:
X P areto( = 2, = 1, 000)
+d = 1000+2000 = 3000
ex(d) = 1
21
Sol:
Claim Size 50
150 500 1000 2000 5000 10000
(X 800)+ 0
0
0
200 1200 4200 9200
Probability 0.305 0.225 0.220 0.155 0.055 0.030 0.010
E(X 800)+ = 315; S(800) = 0.25, E(X

800)2+ = 1461000
E(X800)+
315 = 1260
= 0.25
S(800)
E(X800)2+
P
2
E(Y ) = S(800) = 1461000
0.25 = 5844000
V (Y P ) = E(Y P )2 E 2(Y P ) = 5844000
12602 = 4,256,400
E(Y P ) =
66
67
(b) Determine the expected claim size for claims

in the interval (10, 000, ). 21,000
(c) Determine the expected claim size for claims

less than 10,000. 833.33
Sol:
Sol:
E(X
R > 10, 000|X > 10, 000)
E(X < 10, 000|X < 10, 000)
10,000 xf (x)dx
s)d)
E(X)[E(X10,000)10,000s(10,000)
=
s(10,000
e(10,000)s(10,000)+10,000s(10,000)
=
s(10,000
= e(10, 000) + 10, 000

=11, 000 + 10, 000
= 21,000
68
R 10,000
xf (x)dx
F (10,000)
E(X10,000)10,000S(10,000)
=
F (10,000)
1000
1000 )2
1000(1 11,000 )10,000( 11,000
= 2500
3
1000 )2
1( 11000
69
Example 31.
Claim sizes follow a Pareto distribution with
parameters = 0.2, = 1000. Determine the
mean excess loss at 2000.
Sol:
R
S(x)dx
E(X 2000)+ = 2000
R 1000 0.2
( x+1000 dx
= 2000
LetR u = x + 1000, du = dx
10000.2u0.2du
= 3000
0.8
= 10000.2[ u0.8 ]
3000
=
Thus, eX (2000) =
E(X2000)+
=
S(d)
70
Example 32.
Calculate the payment per loss for an insurance coverage with deductible of 5 if the loss
distribution is
(i) exponential with mean 10. 6.0653
(ii) Pareto with parameters = 3, = 20. 6.4
(iii) Single-parameter Pareto with parameters =
2, = 1 .2
Sol:
E(X 5)+ = e(5)s(5)
(i) E(X 5)+ = 10e5/10 = 6.0653

3
20+5
20
(ii) E(X 5)+ = 31
= 6.4
20+5

2
5
(iii) E(X 5)+ = 21 51 = 0.2
71
Example 33.
The random variable for a loss X, has the following characteristics:
X F (x) E(X x)
0
0.0
0
100 0.2
91
200 0.6
153
1000 1.0
331
Determine the mean excess loss for a deductible of 100.
300
Sol:
eX (100) = E(X)E(X100)
S(100)
33191
= 0.8
= 300
72
Example 34.
Claim sizes follow a Pareto distribution with parameters = 2 and = 25, 000. Determine the
expected claim size in the interval (25, 000, ).
75,000
Sol:
E(X
R > 25, 000|X > 25, 000)
=
25,000 xf (x)dx
S(25,000)
E(X25,000)+25,000S(25000)
=
S(25,000)
(e(25,000)+25,000)S(25,000)
=
S(25,000)
= e(25, 000) + 25, 000

+ 25, 000
= 25,000+25,000
21
= 75,000
73
Example 35.
The cumulative loss distribution for a risk is F (x) =
106 . Determine the average size of a loss
1 (x+10
3 )2
that is less than 1000. 1000/3
Sol:
X P areto( = 2, = 1000)
R 1000
xf (x)dx
F (1000)
E(X1000)1000S(1000)
=
F (1000)
1000
1000(1 2000 )1000( 1000
)2
2000
=
2
1( 1000
2000 )
= 1000
3
E(X < 1000|X < 1, 000) =
Example 36.
The distribution for claim severity follows a singleparameter Pareto distribution of the form

x 4
3
, x > 1000.
f (x) =
1000
1000
Determine the average size of a claim between
10,000 and 100,000, given that the claim is between 10,000 and 100,000. 14,864
Sol:
X SP ( = 3, = 1000)
E(10, 000 < X < 100, 000|10, 000 < X < 100, 000)
R 100,000
xf (x)dx
10,000
= S(10,000)S(100,000)
[(e(10,000)+10,000)S(10,000)][(e(100,000)+100,000)S(100,0
=
S(10,000)S(100,000)
100,000
1000 3
1000 3
[( 10,000
2 +10,000)( 10,000 ) ][( 2 +100,000)( 100,000 ) ]
= 14,864.86
74
1,000 3
1,000 3
) ( 100,000
)
( 10,000
1.5.3
75
Franchise Deductible
A franchise deductible modifies the ordinary deductible by adding the deductible when there is
a positive amount paid.
The per-loss variable is
(
0, X d,
YL =
X, X > d,
The per-payment variable is
YP =
U ndef ined, X d,
X,
X > d,
The corresponding densities are:
f (y)
fY P = X , y > d
SX (d)
(
FX (d), y = 0
fY L (y) =
fX (y), y > d
76
The expected value of Y L for franchise de-
ductible, d, can be calculated from

Z
xf (x)dx
E(Y L) =
d

X
L
(xj )p(xj )
E(Y ) =
xj >d

The k th moment of Y L for franchise deductible,
d is define as
E[(Y L)k ] =
xk f (x)dx

X
L
k
E[(Y ) ] =
xkj p(xj )
xj >d
Notes:
R
R
xf (x)dx 0d xf (x)dx
1. E(Y L) =
= E(X) [E(X d) dS(d)]
= E(X) E(X d) + dS(d)
= E(X d)+ + dS(d)
= e(d)S(d) + dS(d)
= (e(d) + d)S(d)
E(Y L)
2. E(Y P ) = S(d)
E(X)E(Xd)+dS(d)
=
S(d)
E(X)E(Xd)
=
S(d)+d
= e(d) + d
77
78
Example 37.
Losses follow a Pareto distribution with = 3.5,
= 5000. A policy covers losses subjects to a 500
franchise deductible. Determine the average payment per loss and average payment per payment.
1,934, 2,700
Sol:
E(Y L) = [e(500) + 500]S(500)
i3.5
h
ih
500+5000
5000
= 3.51 + 500 500+5000
= 1934.1465
E(Y P ) = e(500)+500 = 500+5000
3.51 +500 = 2,700
1.5.4
79
Loss Elimination Ratio (LER)
Definition 9. The loss elimination ratio is

the ratio of the decrease in the expected payment
with an ordinary deductible to the expected payment without the ordinary deductible. Therefore
LER =
E(X) [E(X) E(X d)] E(X d)

=
E(X)
E(X)
Example 38.
Determine the loss elimination ratio for a Pareto
distribution with = 3, = 2000 with an ordinary deductible of 500. 0.36
Sol:
E(X500)
2
LER(500) = E(X) = 1 [ 2000
2500 ] = 0.36
provided E(X) exists.

LER can be meaningful in evaluating the impact
of a deductible.
Note:
E(X)E(Xd)+
e(d)S(d)
LER =
=
1
E(X)
E(X)
80
1.5.5
81
The Effect of Inflation on Ordinary Deductible
Inflation increase costs, when there is a deductible,

the effect of inflation is magnified.
Theorem 6. For an ordinary deductible of d after uniform inflation of 1 + r,

d
E[(1 + r)X d] = (1 + r)E X
1+r
and

d
E[(1+r)Y L] = (1+r)E(X)(1+r)E X
1+r

d
and, if S 1+r
< 1, then

d
h
i (1 + r)E(X) (1 + r)E X 1+r
P

E (1 + r)Y
=
d
1 F 1+r
82
Example 39.
The underlying loss distribution function for a
certain line of business in 2008 is:
F (x) = 1 x6, x > 1.
From 2008 to 2009, 10% inflation impacts all claims

uniformly. Determine the expected amount not
pay due deductible of 2.2 in year 2009. 1.313125
Sol:
X SP ( = 6, = 1)
E(1.1X 2.2) = 1.1E(X 2.2/1.1)
R
= 1.1[1 + 12 S(x)dx]
R
= 1.1[1 + 12( x1 )6dx]
5
= 1.1[1 + x5 ]21
5
= 1.1[1 + 2 51 ]
1.313125
or Y = 1.1X SP ( = 6, = 1.1)
E(Y 2.2) =
=
=
=
R 2.2
1.1 + 1.1 S(x)dx
R 2.2
)6dx
1.1 + 1.1 ( 1.1
x
5
1.1 + 1.16[ x5 ]2.2
1.1
6 2.251.15
] = 1.313125
1.1 + 1.1 [
5
1.5.6
83
The Effect of Inflation on Policy

Limit
For a policy limit of u, after uniform inflation of

1 + r, the expected cost is

u
E{[(1 + r)X] u} = (1 + r)E X
1+r
84
Example 40.
Impose a limit of 3,000 on a Pareto distribution
with = 3, = 2000. Determine the expected
cost per loss with the limit and after 10% uniform
inflation is applied. 903.2
Sol:
X P areto( = 3, = 2000)
E(1.1X 3000) = 1.1E(X 3000/1.1)
= 1.1E(X 2727.27)
2
= 1.1( 3000
31 )[1 (2000/4727.27) ]
= 903.106
1.5.7
85
Coinsurance, Deductible, and Limits
Coinsurance of means that a portion, , of each

loss is reimbursed by insurance. For example,
80% coinsurance means that insurance will pay
80% of the loss.
When ordinary deductible, limit coinsurance, and
inflation are present, the per loss random variable
is:
d ,
x < 1+r
0,
d X < u ,
Y L = [(1 + r)X d], 1+r
1+r
u .
(u d),
X 1+r
Note that u is the maximum cover loss and (u
d) is policy limit.
The quantities in this definition are applied in
a particular order with the coinsurance applied
last.
86
Theorem 7.

E(Y ) = (1 + r) E X
L
and
u
1+r
d
E X
1+r

E(Y L)

E(Y ) =
d
1 FX 1+r
P
Proof: From the definition of Y L,

d and u = u
let d = 1+r
1+r
Thus,
Y L = (1 + r)[X u X d]
E(Y L) = (1 + r)[E(X u) E(X d)]
87
Theorem 8.

u 2
E[(Y L)2] = 2(1+ r)2{E (X
)
1+r

d
u
r 2
) 2 1+r
E X 1+r
E (X 1+r
d
d
+2 1+r
E X 1+r
}
Proof:
(Y L )2
Example 41.
Determine the mean and standard deviation per
loss for a Pareto distribution with = 3, =
2, 000 with a deductible of 500 and a policy limit
of 2,500. Note that the maximum covered loss
u = 3, 000. 480, 754.7
2 (1+r)2
Sol:
=
=
=
=
0 x < 500
0,
Y L = X 500,
500 x < 3000
2500, x > 3000

= (X 3000) (X 500)
[X u X d]2
(X u)2 + (X d)2 2(X u)(X d)
(X u)2 (X d)2 2(X d)[(X u) (X d]
(X u)2 (X d)2 2d[(X u) (X d)]
Thus
E(Y L)2 = 2(1+r)2[E(X u)2 E(X d)2
2d[E(X u) E(X d)]]
Note:
2(X d)[(X u) (X d)] = 2d[(x u)
(X d)]
To see this, when X < d, both sides equal zero;
when d x u, both sides equal 2d(X d);
and when x u, both sides equal 2d(u d).
88
E(Y L) =
=
=
=
E(X 3000) E(X 500)

2000 )2] 1000[1 ( 2000 )2]
1000[1 ( 5000
2500
840 360
480
E(X d)2
Rd
= 0 2xS(x)dx
Rd
2000 3
= 0 2x( x+2000
dx
Let u = x + 2000
R d+2000
= 2(20003) 2000 (u 2000)u3du
1
= 2(20003)[ u1 + 1000u2]d2000
89
= 2(20003)[1000(d + 2000)2 (d + 2000)1 + 20001 1000(20002)]

= 2(20003)[1000(d + 2000)2 (d + 2000)1 + 0.00025]
E(X 500) = 2(20003)[1000(25002) (25001) + 0.00025] = 160, 000

E(X3000) = 2(20003)[1000(50002)(50001)+0.00025] = 1, 440, 000
E(Y L )2 = E(X 3000)2 E(X 500)2 2dE(X 3000) + 2dE(X 500)
= 1, 440, 000 160, 000 2(500)(840) + 2(500)(360)
= 800,000
V (Y L ) = 800, 000 4802 = 569, 600
SD(Y L ) = 569, 600 = 754.7
Example 42.
An individual losses has the Pareto distribution
with parameters = 3 and = 100 with deductible of 55, coinsurance of 75% and a loss limit
of 110 (before application of the deductible and
coinsurance) are applied to each individual loss.
Loss sizes are affected by 10% inflation. Determine the variance of the loss payment on the per
payment basic. 673
Sol:
Y
x < 50
0,
= .75(1.1)(X 50), 50 x < 100
.75(110 50),
x 100
= .75(1.1)[(X 100) (X 50)]
E(Y L) =
=
=
=
.75(1.1)[E(X 100) E(X 50)]

.75(1.1)[50(1 (1/2)2) 50(1 (10/15)2)]
.75(1.1)[37.5 27.7778]
8.0208
E(X d)2
Rd
= 0 2xS(x)dx
Rd
100 3
dx
= 0 2x( x+100
Let u = x + 100
90
=
=
=
=
91
R d+100
2(1003) 100 (u 100)u3du
1
2(1003)[ u1 + 50u2]d+100
100
3
2
2(100 )[50(d + 100) (d + 100)1 + 1001 50(1002)]
2(1003)[50(d + 100)2 (d + 100)1 + 0.005]
E(X 50)2 = 2(1003)[50(150)2 (150)1 + 0.005] =

1111.11
E(X 100)2 = 2(1003)[50(200)2 (200)1 +0.005] = 2500
1.5.8
92
Bonus
An agent receives a bonus if his loss ratio is below

certain amount, or a hospital receives a bonus if
it doesnt submit too many claims.
Bonus = (
max[0, c(rP X)]
0,
c(rP X) < 0
=
c(rP X), c(rP X) > 0
E(Y L)2 = .752(1.1)2[E(X 100)2 E(X 50)2
(
2(50)E(X 100) + 2(50)E(X 50)]
0,
X > rP
=
= (.75 1.1)2[2500 1111.11 2(50)(37.5) + 2(50)(27.7778)]
c(rP X), X < rP
= 283.6
L)
= crP c min(rP, X)
8.0208
E(Y P ) = E(Y
=
=
27.0702
3
S(50)
(100/150)
= crP c(X rP )
E(Y L )2
283.6
P 2
E(Y ) = S(50) = (100/150)3 = 957.15

V (Y P ) = E(Y P )2 E 2(Y P ) = 957.1527.07022 = 228.15
Note: max(0, a b) = a min(a, b)
93
Example 43. An insurance agent will receive a

bonus if his loss ratio is less than 70%. you are
given:
His loss ratio is calculated as incurred losses
divided by earned premium on his block of

business.
The agent will receive a percentage of earned
premium equal tp 1/3 of the difference between 70% and his loss ratio.
B = max[0, 31 (500, 000)(.7 X)]

31 min(X, 350, 000)
= 350,000
3
= 350,000
31 (X 350, 000)
3

2

600,000
1 350,000+650,000
E(X 350, 000) = 600,000
2
= 180,332.41
E(B) = 350,000
31 (180, 332.41) = 56,555.86
3
The agent receive no bonus if his loss ratio is
greater than 70%.

His earned premium is 500,000.
His incurred losses are distributed according
to the Pareto distribution

3
600, 000
F (x) = 1
x + 600, 000
Calculate the expected value of his bonus. 56,555.86
Sol:
94
95
Example 44.
For a particular agent of an insurance company,
his total earned premium is 100. Denote the loss
random variable for his business to be L and define the loss ratio by
L
R=
,
100
which is interpreted as the proportion of earned
premiums from losses. The agent receives a bonus
B if the loss ratio does not exceed 80%. This
bonus is B = 25(.80 R), if positive, otherwise
it is zero. Suppose L has a uniform distribution
on (0, 200). Calculate the expected value of the
bonus for this agent. 4
(
0,
L > 80
=
20 .25L, R < 0.8
L U (0, 200),
1 , 0 < l < 200, 0 otherwise
f (l) = 200
R 80
1 dl
E(B) = 0 (20 .25l) 200
80

2
1 (20.25l)
= 200
2(.25)
0
h 2i
20
1
= 200 .5
= 4
Sol:
L ))
B = max(0, 25(.80 100
=(
max(0, 20 0.25L)
0,
20 .25L < 0
=
L ), 20 .25L < 0
25(.8 100
96
1.6
1.6.1
97
Reinsurance
Excess of loss reinsurance - the
insurer
In excess of loss reinsurance, the insurer will pay

any claim in full up to an amount M , the retention level; any amount above M will be borne
by the reinsurer. The excess of loss reinsurance
arrangement can be written in the following way:
(
X, X M
Y =
M, X > M
i.e. Y = min(X, M ) where X denotes the claim
amount and Y denotes the amount paid by the
insurer.
Z M
E(Y k ) =
xk f (x)dx + M k S(M )
Thus
R
E(Z) = E(X) E(Y )= RM (x M )f (x)dx
= M
s(x)dx
If only a layer between M and 2M is reinsured,

then the amount paid by the insurer is
XM
X,
Y = M,
M < X 2M
X M, X > M
RM
R 2M
E(Y ) = 0R xf (x)dx + M mf (x)dx
+ 2M
(x M )f (x)dx
The reinsurer pays the amount Z = X Y . i.e.

(
0,
XM
.
Z = max(X M, 0) =
X M, X > M
98
99
Example 45.
Loss amounts under a class of insurance policies
follow an exponential distribution with mean 118.
The insurance company wishes to enter into an
individual excess of loss reinsurance arrangement
with retention level M set such that 86 out of 100
claims will not involve the reinsurer. For a given
claim, let XI denote the amount paid by the insurer and XR the amount paid by the reinsurer.
Calculate E(XI ) and E(XR). 101.48, 16.52
Sol:
86
P (X > M ) = 100
R M 1 x/118
= 0.86
0 118 e
M/118
1e
= 0.86
M =
(ln(0.14)(118) = 232.001317052
X, X M
XI =
M, X > M
R 232.001317052 x/118
RM
e
dx = 118(1e232.001317052/118)
E(XI ) = 0 s(x)dx = 0
= 101.48
E(XR ) = E(X) E(XI ) = 118 101.48 = 16.52
Example 46.
For a certain portfolio of insurance policies, claim sizes have
a gamma distribution with mean 100 and variance 5,000.
The insurer is considering purchasing individual excess of
loss reinsurance with retention M from a reinsurer. Let XI
and XR denote the amounts paid by the direct insurer and
the reinsurer, respectively, on an individual claim. Find
E(XR) and E(XI ).
Sol:
X gamma(, )
= 100 (1)
2 = 5, 000 (2)
(2)
, = 50, = 2
(1)
S(x) = P (S2 > x) = P (N (x) < 2) = ex/50 +
x
N (x) P oi( 50
) (
0,
X M
XR = (X M )+
X M, X < M
R
E(XR ) = RM (x M )f (x)dx
= RM S(x)dx
x x/50
e
dx
= M ex/50 + 50
= 50eM/50 + 50P (S2 > M )
= 50eM/50 + 50(eM/50 + M
eM/50)
50
= (100 + M )eM/50
E(XI ) = E(X) E(XR ) = 100 (100M ))eM/50
100
x x/50
e
50
where
1.6.2
101
Excess of loss reinsurance - the

reinsurer
The reinsurer may have a record only of claims
that are greater than M . If a claim is for less

than M the reinsurer may not even know a
claim has occurred.
We write the random variable for the amount
paid by the reinsurer as

Z = X M |X > M.
Suppose that the underlying claim amounts
have PDF f (x) and CDF F (x). Suppose that

the reinsurer is only informed of claims greater
than the retention M and has a record of z =
x M.
P (Z < z) = P (X < z + M |X > M )

R z+M f (x)
dx
= M
1F (M )
Example 47.
Loss amounts under a class of insurance policies follow a Pareto distribution with parameters = 2 and = 1000. The insurance company wishes to enter into an individual excess of
loss reinsurance arrangement with retention level
M = 1200. Suppose that the reinsurer is only
informed of claims greater than the retention M
and has a record of z = x M . Determine f (z)
and expected amount paid by the reinsurer.
Sol: X P areto(2, 1000)
2(1000)2
f (x) = (x+1000)3 ; F (x) = 1
Z = X M |X > M
2(1000)2
(z+2000)3
f (z) = 0.5
E(Z) = 2000
F (z+M )F (M )
1F (M )
f (z+M )
E(XM )
Thus, f (z) = 1F (M ) and E(Z) = S(M ) +
102
4(10002)
= (z+2000)3
1000
x+1000
2
1.7
103
Tails of distributions
The tail of a distribution is the portion of the

distribution corresponding to large values of the
random variable. Random variables that tend
to assign higher probabilities to larger values are
said to be heavier-tailed.
1.7.1
Concept of Tail Weight
1. Relative: Model A has a heavier tail the Model

B.
2. Absolute: Distributions with certain property
are classified a heavy-tail.
When using a parametric distribution to model
loss size, it is important to provide for the possibility of very high claims, which may occur rarely.
The bigger the tail weight of the distribution, the
more provision for high claims.
1.7.2
104
Classification of Tail weight
1. Classification based on moments: The more

positive raw moment exist, the less tail weight.
2. Comparison based on limiting tail behavior: if
S (x)
f (x)
lim 1
= lim 1
=
x S2(x)
x f2(x)
then X1 has heavier tail.
3. Classification based on the hazard rate function: An increasing hazard rate function (w.r.t
x) have light tail.
4. Classification based on mean excess loss
function: If the mean excess loss function
(eX (d)) is increasing in d, the distribution is
considered to have a heavy tail. If the mean
excess loss function (eX (d)) is decreasing in d,
the distribution is considered to have a light
tail.
105
Example 48.
Random variable X1 with distribution function
F1 and probability density function f1 has a heavier tail than random variable X2 with distribution
function F2 and probability density function f2.
Which of the following statements is true?
1. X1 will tend to have fewer positive moments
than X2.
106
Example 49.
Which of the following are true based on the existence of moments test?
1. The Loglogistic Distribution has a heavier tail
than the Gamma Distribution.
2. The Paralogistic Distribution has a heavier tail
than the Lognormal Distribution.
2. The limiting ratio of the density functions, ff1 ,

2
will go to infinity.
3. The Inverse Exponential has a heavier tail than

the Exponential Distribution.
3. The hazard rate of X1 will increase more rapidly

than hazard rate of X2.
Sol:
4. The mean residual life of X1 will increase more

rapidly than the mean residual life of X2.
Sol:
2. True
1. True
3. True
1. True
2. True
3. False
4. True
107
Example 50.
You are given:
500,000
X has an density f (x), where f (x) =
,
x3
for x > 500.
Y has density g(y) = yey/500/500, 000.
Which of the following are true?

1. X has an increasing mean residual life function.
2. Y has an increasing hazard rate.
3. X has a heavier tail than Y based on the hazard rate test.
Sol:
1. True
2. True
3. True
1.8
108
Measure of Risk
A risk measure is a mapping from the random

variable representing the loss associated with the
risks to the real line. A risk measure gives a single
number that is intended to quantify the risk exposure. Risk measures are denoted by the function
(X).
1.8.1
Coherent
Definition 10.
A coherent risk measure is a risk measure (X) that has
the following four properties for any two loss random variables X and Y :
1. Subadditivity: (X + Y ) (X) + (Y ).
2. Monotonicity: if X Y for all possible outcomes, then

(X) (Y ).
3. Positive homogeneity: For any positive constant c, (cX) =

c(X).
4. Translation invariance: For any positive constant c, (X+
c) = (X) + c.
109
Subadditivity means that the risk measure
for two risks combined will not be greater than

for the risks treated separately. Subadditivity
reflects the fact that there should be some diversification benefit from combining risks.
Monotonicity means that if one risk always
has greater losses than another risk under all

circumstances, the risk measure should always
be greater.
Positive homogeneity means that the risk
measure is independent of the currency in which

risk is measured. Equivalently, it means that,
for example, doubling the exposure to a particular risk requires double the capital.
Translation invariance means that adding
a constant to a random variable should add the

same constant to risk measure.
Risk measures satisfying these four criteria are
deemed to be coherent. Coherent is a measure of
what is a good risk measure.
110
Example 51.
Which of the following risk measures are coherent?
(i) (X) = E(X)
(ii) (X) = E(X)(1 + ), > 0
(iii) (X) = E(X) + c, c > 0
(iv) (X) = E(X) + SD(X), > 0
Sol:
(i) Yes
(ii) Fail translation.
(X + c) = E(X + c)(1 + )
= (1 + )E(X) = c(1 + )
= (X) + c(1 )
6= (X) + c
(iii) Fail homogeneity
(cX) = E(cX) + c
= cE(X) + c
= c(E(X) + 1)
6 c(X)
111
(iv) Fail montonicity. Since a constant Y woud

have no variance while X < Y nay have
variance, a higher (X) as a result.
1.8.2
112
Value-at-Risk
Definition 11. Let X denote a loss random

variable. The value-at-Risk of X at the 100p%
level, denoted V aRp(X) or p, is the 100p percentile (or quantile) of the distribution of X.
For continuous distributions, we can simply write
V aRp(X) for the random variable X as the value
of pJ satisfying P (X > p) = 1 p.
For a discrete distribution of loss, the percentile
may not be unique. We define the V aR as the
lowest percentile.
p = min(|P (X ) p)
Note: VaR does not satisfy subadditivity requirement.
113
Example 52.
Losses have a lognormal distribution with mean
10 and variance 300. Calculate the VaR at 95%
and 99%. 34.68, 77.33
Example 53.
Losses have the following distribution:
Loss size Probability
0
0.5
5
0.3
10
0.15
20
0.05
Sol:
2
e+ /2 = 10 + 2/2 = ln 10 (1)
2
e2+2 = 300 + 102 2 + 2 2 = ln 400- (2)
(2) (1) 2, 2 = ln(400/100) = ln 4
= ln 10 ln 4/2 = ln 5
( ln .95ln 5 = .95
ln 4
ln
ln
5
.95
= 1.645
ln 4
.95 = 34.68
ln
.99ln 5 = 2.326
ln 4
.99 = 77.33
Calculate the VaR at 95% and 99%. 10,20

Sol:
0.5
F (x) = 0.8
0.95
1
V aR.99 = 20
x<0
0x<5
5 x < 10 V aR.95 = 10
10 x < 20
x 20
114
Example 54.
Which of the following statements are true?
(i) VaR satisfies Translation Invariance
(ii) VaR satisfies Monotonicity
(iii) VaR satisfies Subadditivity
(iv) VaR satisfies Positive Homogeneity
Sol:
115
1.8.3
Tail-Value-at-Risk
Definition 12. Let X denote a loss random

variable. The Tail-Value-at-Risk of X at the
100p% level, denoted T V aRp(X) or Conditional
Tail Expextation (CT Ep), is the expected loss
given that the loss exceeds the 100p percentile of
the distribution.
T V aRp(X) = E(X|X > p)
1. Yes
2. Yes
3. No
4. Yes
116
For continuous distribution,

T V aRp(X) =
R
p
xf (x)dx
R S(p)
xf (x)dx
= p 1p
117
Note that
R
R p
p xf (x)dx = 0 xf (x)dx 0 xf (x)dx

= E(X) [E(X p) pS(p)]
= E(X) E(X p) + (1 p)p
T V aRp can also be written as

T V aRp = E(X|X
R > p )
p (xp)f (x)dx
= p +
1p
= p + e(p)
Example 55.
Losses have a lognormal distribution with mean
10 and variance 300. Calculate T V aR0.95. 63.84
Sol:
From Example 52, = ln 5, 2 = ln 4, .95 =
34.68
E(X)E(X,95)+.05.95
CT E.95 =
.05
5ln 4 )+34.68(.05)]+.05(34.68)
10[10( ln 34.68ln
ln 4
=
.05
1010(0.47)
=
0.05
1010(0.6808)
=
0.05
= 63.84
This formula should be use for Pareto and exponential distributions which have simple formulas
for e(p).
118
119
120
Example 56.
Losses have a Pareto distribution with mean 10
and variance 300. Calculate T V aR0.95. 61.43
Example 57. Losses have an exponential distribution with mean 10. Calculate T V aR0.95.
39.96
Sol:
Sol:
e.95/10 = .05
.95/10 = ln .05
.95 = 29.96
T V aR.95 = .95+e(.95) = 29.96+10 = 39.9694
1 = 10 (1)
2 2
= 400 (2)
(1)(2)
2
(1)2
(1)(2)
1
,
=
2
4
(2) (1)2
2
4( 2) = 2( 1)
4 8 = 2 2
=3
/2 = 10
= 20
3
20
= .05
.95+20
.95 = 34.29
+ 34.29 =
T V ar.95 = e(.95) + .95 = 34.29+20
2
61.43
121
Example 58.
Losses have a normal distribution with mean 100
and variance 400. Calculate T V aR0.98. 2.9659
Example 59.
Losses have the following distribution:
Loss size Probability
0
0.5
5
0.3
10
0.15
20
0.05
Sol:
.98 = 2.055(20) + 100 = 141.1
T VR aR.98
xf (x)dx
= 1.41.11p
R
x100
1
e 2(400) dx
= 141.1 x 2(20)
Let z = x100
20
R
2
1
= 2.055(20z + 100) 2(20)
ez /2dz
i
hR
R R
2 /2
1
1
z
dz
= 20 2.055 2ze z2/2 dz + 100 2.055 2.055 2 e
= 20( 12 )[ez
2 /2
]
2.055 + 100[1 (2.055)]
= 202 e2.055 /2 + 100(0.2)

= 2.9659
122
Calculate T V aR0.90 and T V aR0.95. 15,20

Sol:
The partial expectation of the top 10% is 0.05(10)+
0.05(20) = 1.5
1.5 = 15
T V aR0.90 = 0.1
The partial expectation of the topm 5% is 0.05(20) =
1
1 = 20
T V aR0.95 = .05
123
Example 60.
A loss random variable X has a Pareto distribution with parameters = 4, and satisfying:
V aRp = 7.4320
T V aRp = 17.4773
Determine p. .6778
Sol:
7.432+ = 17.4773
3
= 22.7039

4
22.7039
22.7039+7.432
p = 0.6778
=1p
1.8.4
Distortion Risk Measures
A distortion function g(x) is defined on [0, 1] such

that g(0) = 0, g(1) = 1, and g(x) is monotonically non-decreasing.
The risk measure is then
R
(x) = 0 g(S(x))dx.Note that the higher survival, the higher risk measure, so usually g(x)
x.
V aRp(X) can be defined with distortion func-
tion
(
0, 0 x 1 p
g(x) =
1, 1 p < x < 1
Sol: Proof:
(
g(S(x)] =
124
0, 0 S(x) 1 p
1, 1 p S(x) 1
(X) = 0 g[S(x)]dx
R S(x)=1
= S(x)=1p 1dx
R x=S 1(1p)
= x=S (1)
1dx
125
R
= 0 p 1dx
= p
= V aRp
T V aRp(X) can be defined with distortion func-
tion
g(x) =
Sol: Proof:(
g(S(x)] =
x
1p ,
1,
1p<x<1
1,
S(x)
1p ,
0x1p
0 S(x) 1 p
1 p S(x) 1
R
(X) = 0 g[S(x)]dx
R S(x)=1
R S(x)=1p S(x)
= S(x)=0
1p dx + S(x)=1p 1dx
R x=S 1(0) S(x)
= x=S (1p) 1p dx + p
R1
= p dx + p
Example 61.
Losses follow an inverse exponential distribution with = 1, 000. Suppose the distortion
risk function is given by
(
0, 0 s(x) 0.1
g[s(x)] =
1, 0.1 < s(x) < 1
Calculate the distortion risk measure. 9491.2

Sol:
1000
1 e .9 = .9
= 9491.2216
.9 = 1000
ln(.9)
(X) = .9 = 9491.2216
E(X
= 1p p + p
= e(p) + p
= T V aRp
126
127
Example 62.
Losses follow a Pareto distribution with = 2,
= 10. To calculate the risk, the distortion
function
g[s(x)] =
s(x)
0.01 ,
1,
0 s(x) 0.01
0.01 < s(x) < 1
Calculate the distortion risk measure. 190

Sol:
10 )2 = 1
( 10+
.99
.99 = 90
CT E.99 = e(90) + 90 = 90+10
21 + 90 = 190
The Proportional Hazard Transform
g(x) = x1/k , k
Notes:
1. For X exp(), S(x) = ex/ , and
Z
Z
ex/k dx = k
g[S(x)]dx =
(x) =
0
2. For X W EI(, ), S(x) = e(x/) ;

R
(x)= R0 g[S(x)]dx
1/k
) dx
= Z0 (e(x/)

e k1/
=
dx
0

1
= k 1 + 1

3. For X Pareto(, ); S(x) = +x
R
(x) = 0 g[S(x)]dx
/k
Z
dx
=
+x
0
= /k1
for
128
> 1, otherwise it is undefined.
129
Example 63.
Losses follow a Pareto distribution with = 4,
= 10. Calculate the proportional hazard
transform risk measure at k = 3. 30
Sol:
10 = 30
(x) = 4/31
Example 64.
You are given the distortion function:
g(x) = x
Calculate the risk measure for losses that follow the Pareto distribution with = 10000
and = 4. 10,000
Sol:
10,000
= 10,000
(x) = 4/21
130
131
Wangs Transform(WT Transform)
g(x) = (1(x) + )
where = 1(p)
Notes:
1. This is hard to calculate, except for a lognormal distribution and normal distribution.
2. For lognormal or normal distribution, the
effect is to send to + .
3. For X lognormal(, ),
2
W T (p) = e++ /2
4. For X N ormal(, 2),
This is the survival function of a lognormal

distribution with = + , = .
R
ln x(+ 2 )
++ 2/
])dx
=
e
(X) = 0(1[
g(S(x)] = [1(S(x)) + ]
= [1(1 x
) + ]
= (( x
) + )
= 1 (( x
) )
)
= 1 ( x
R
(X) = 0(1 ( x
)dx = +
W T (p) = + .
Sol:
g[S(x)] = [1(1 ( ln x
)) + ]
= [( ln x
) + ]
= 1 [ ln x
]
ln x(+ 2)
= 1 [
]
132
133
Example 65.
Losses follow a lognormal distribution with =
10, = 0.5. Calculate the risk measure using
Wangs Transform with p = 0.975. 66502.84
Sol:
= 1(0.975) = 1.96
2
2
W T (0.975) = e++ /2 = e10+1.96(0.5)+0.5 /2
= 66502.84

UECM3463May16 C01 (2p)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

UECM3463May16 C01 (2p)

Uploaded by

Copyright:

Available Formats

May 2016 Chapter 1 Models For Claim Severity

UNIVERSITI TUNKU ABDUL RAHMAN

1 Models for Claim Severities

May 2016 Chapter 1 Models For Claim Severity

1.5 Severity with Coverage Modifications . . . . . . . . . . . . . . . . 47

May 2016 Chapter 1 Models For Claim Severity

1.8 Measure of Risk . . . . . . . . .

May 2016 Chapter 1 Models For Claim Severity

Models for Claim Severities

Given that a claim occurs, the (individual)

claim size X is typically referred to as claim

dom variables, sometimes claim sizes can be

UECM3463 Loss Models

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

May 2016 Chapter 1 Models For Claim Severity

Some Parametric Claim Size Distributions

Normal - easy to work with, but careful with

getting negative claims. Insurance claims usually are never negative.

distribution is considered light, applicable

ble for example with fire insurance.

such as liability insurance.

UECM3463 Loss Models

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

May 2016 Chapter 1 Models For Claim Severity

UECM3463 Loss Models

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

Inverse Gaussian densities with various and

May 2016 Chapter 1 Models For Claim Severity

Basic Distributional Quantiles

Definition 1. k th raw moment

UECM3463 Loss Models

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

Definition 2. k th central moment

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

Expectation is linear, so the central moments can

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

Special Functions of Moments

Variance: V (X) = 2 and is denoted by 2.

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

Skewness is a measure of symmetry, or more

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

Kurtosis is a measure of whether the data are

A mode is x such that f (x) is maximized (or

The moment generating function (MGF) is defined by

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

The probability generating function (PGF) is

May 2016 Chapter 1 Models For Claim Severity

PGFs make sums of independent random

The cumulant-generating function K(t), is the

This expansion is a Maclaurin series, so that

UECM3463 Loss Models

May 2016 Chapter 1 Models For Claim Severity

ing the result at zero:

May 2016 Chapter 1 Models For Claim Severity