Sie sind auf Seite 1von 24

Chapter 2

Models for Claim Frequencies

Contents
2.1 Discrete Distributions . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2 Three Basic Distribution . . . . . . . . . . . . . . . . . . . . . . 3

2.2.1 The Poisson Distribution . . . . . . . . . . . . . . . . . . . . . 3

2.2.2 The Negative Binomial Distribution . . . . . . . . . . . . . . . 4

2.2.3 The Binomial Distribution . . . . . . . . . . . . . . . . . . . . . 5

2.3 The (a, b, 0) Class . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.4 The (a, b, 1) Class . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.4.1 Truncation and Modification at Zero . . . . . . . . . . . . . . . 10

2.4.2 Extended Truncated Negative Binomial (ETNB) . . . . . . . . 14

2.5 Advanced Discrete Distributions . . . . . . . . . . . . . . . . . 15

2.5.1 Compound Frequency Models . . . . . . . . . . . . . . . . . . . 15

2.5.2 Compound Poisson Distribution . . . . . . . . . . . . . . . . . 17

2.5.3 Mixed Poisson Distribution . . . . . . . . . . . . . . . . . . . . 18

2.6 The Impact of Deductible on Claim Frequency . . . . . . . . 20

2.7 Exposure Modifications . . . . . . . . . . . . . . . . . . . . . . . 23

1
CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

2.1 Discrete Distributions


Counting distributions are discrete distributions with probabilities only on the nonneg-
ative integers, Z+ . In an insurance context, counting distributions describe the number
of events such as losses to the insured or claims to the insurance company or claim fre-
quencies. Claim frequencies is the (average) number of claims per time period, usually
per year.

Let N be a random variable representing the number of events.

1. The probability function (pf ) pk denotes the probability that exactly k


events (such as claims or losses) occur. Then

pk = P (N = k), k = 0, 1, 2, . . .

2. The probability generating function (pgf ) of a discrete random variable


N with pf pk is

X
N
P (z) = PN (z) = E(z ) = pk z k .
k=0

3. As is true with the moment generating function, the pgf can be used to generate
moments. In particular, P 0 (1) = E(N ) and P 00 (1) = E[N (N − 1)].

4. Pgf really does generate probabilities,


dm N 
P (m) (z) = E z = E[N (N − 1) · · · (N − m + 1)z N −m ]
dz m

X
= k(k − 1) · · · (k − m + 1)z k−m pk
k=m

(m) P (m) (0)


P (0) = m!pm or pm = .
m!
Note that p0 = P (0).

Example 2.1.1. Suppose N is a discrete distribution with pgf of


1
P (z) = .
1.1 − 0.5z − 0.5z 2
Calculate P (N = 1) and E(N ).

Example 2.1.2. Suppose N is a random variable with pgf of


1 1 1
P (z) = z + z 2 + z 3 .
3 6 2
Calculate the coefficient of variation and skewness of N .

UECM3463 Loss Models May 2019 2


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

2.2 Three Basic Distribution


Three basic distributions are Poisson, negative binomial, and binomial.

2.2.1 The Poisson Distribution


1. The pf of Poisson distribution with parameter λ > 0 is defined by
e−λ λk
pk = , k = 0, 1, 2, . . .
k!
2. The pgf is

X e−λ λk
P (z) = zk
k=0
k!

−λ
X (λz)k
=e
k=0
k!

= e−λ eλz
= eλ(z−1) , λ > 0.

3. Since P 0 (z) = λeλ(z−1) and P 00 (z) = λ2 eλ(z−1) , then the mean and variance can be
computed from the pgf as follows:

E(N ) = P 0 (1) = λ
E[N (N − 1)] = P 00 (1) = λ2
V (N ) = E[N (N − 1)] + E(N ) − [E(N )]2 = λ2 + λ − λ2 = λ.

Theorem 2.2.1. Let N1 , . . . , Nn be independent Poisson variables with parameters


λ1 , . . . , λn . Then, N = N1 + · · · + Nn has a Poisson distribution with parameter
λ1 + · · · + λn .

Proof:
Ni ∼ P oi(λi ), i = 1, 2, . . . , n
λi (z−1)
PNi (z) = e
N = N1 + · · · + Nn
PN (z) = PN1 +···+Nn (z)
= PN1 (z)PN2 (z) · · · PNn (z)
= eλ1 (z−1)
 e 2
λ (z−1)
· · · eλn (z−1)
Pn
= exp i=0 λi (z − 1)
Pn 
=⇒ N ∼ P oi i=0 λi

UECM3463 Loss Models May 2019 3


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Example 2.2.2. Dental Insurance Company sells a policy that covers two types of dental
procedures: root canals and fillings. There is a limit of 1 root canal per year and a separate
limit of 2 fillings per year. The number of root canals a person needs in a year follows
a Poisson distribution with λ = 1, and the number of fillings a person needs in a year
is Poisson with λ = 2. The company is considering replacing the single limits with a
combined limit of 3 claims per year, regardless of the type of claim. Determine the
change in the expected number of claims per year if the combined limit is adopted.

2.2.2 The Negative Binomial Distribution


1. The negative binomial distribution has been used extensively as an alternative to
the Poisson distribution.

2. Like the Poisson distribution, it has positive probabilities on the nonnegative inte-
gers.

3. Because it has two parameters, it has more flexibility in shape than the Poisson.

4. The pf of negative binomial distribution with parameters r and β is defined by


  r
k+r−1 1 β k
pk =
k 1+β 1+β

k = 0, 1, 2, . . . , r > 0, β > 0.

5. The pgf for the negative binomial distribution is P (z) = [1 − β(z − 1)]−r .
Proof:
∞  
X k+r−1 β
P (z) = (1 − q)r (qz)k , let q = 1+β
k=0
k
∞ 
(1 − q)r X k + r − 1

= r
(1 − qz)r (qz)k
(1 − qz) k=0 k
1
r r
(1 − q)r
 
1+β 1 1+β
= = βz r
 =
(1 − qz)r 1 − 1+β 1+β 1 + β − βz

= [1 − β(z − 1)]−r .

6. The mean and variance of the negative binomial distribution are

E(N ) = rβ and V (N ) = rβ(1 + β).

UECM3463 Loss Models May 2019 4


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

• Because β is positive, so V (N ) > E(N ) for negative binomial distribution. This


relationship is in contrast to the Poisson distribution for which the V (N ) =
E(N ).
• Thus, for a particular set of data, if the observed variance is larger than the
observed mean, the negative binomial might be a better candidate than the
Poisson distribution as a model to be used.

Notes:

(a) Negative binomial distribution is a generalization of the Poisson in the following two
ways:

• mixed Poisson distribution with a gamma mixing distribution, that is


if X|λ ∼ P oi(λ) and Λ ∼ Gamma(α, β), then X ∼ N B(r = α, β).
• a compound Poisson distribution with a logarithmic secondary distribution.
Suppose PN (z) ∼ P oi(λ) and PM (z) ∼ logarithmic(β), then PN [PM (z)] =
[1 − β(z − 1)]−r , where r = λ
ln(1+β)
.

(b) Poisson distribution is a limiting case of the negative binomial distribution.


Let r → ∞, β → 0 and λ = rβ. Then
 −r
λ(z − 1)
lim 1 − = exp[λ(z − 1)].
r→∞ r

(c) Geometric distribution is the special case of negative binomial distribution when
r = 1.

2.2.3 The Binomial Distribution


1. The pf of binomial distribution is defined by
 
m k
pk = q (1 − q)m−k , k = 0, 1, 2, . . . , m and 0 < q < 1.
k
The pgf is
P (z) = [1 + q(z − 1)]m .
Proof:
Let Y ∼ Bin(m = 1, q)
PY (z) = 1k=0 q k (1 − q)1−k · z k = (1 − q) + qz = 1 + q(z − 1).
P

Let N = Y1 + Y2 + · · · + Ym ∼ Bin(m, q).


m
Y
PN (z) = PY1 +Y2 +···+Ym (z) = PYi (z) = [1 + q(z − 1)]m .
i=1

UECM3463 Loss Models May 2019 5


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

2. The mean and variance are

E(N ) = mq and V (N ) = mq(1 − q).

3. Its variance is smaller than its mean, making it useful for data sets in which the
observed sample variance is less than the sample mean.

4. Bernoulli distribution is a special case of binomial distribution with m = 1.

5. For large m and q, binomial probabilities can be approximated as Poisson probabil-


ities with λ = mq.

Example 2.2.3. A portfolio of 10, 000 risks yields the following:

Frequency 0 1 2 3 4
Number of insured 6070 3022 764 126 18

Based on the portfolio’s sample moments, which of the following distributions provide the
best fit to the portfolio’s number of claims?
A. Binomial B. Poisson C. Negative Binomial D. Lognormal E. Pareto

2.3 The (a, b, 0) Class


Given the mean and variance of a distribution, we can determine the parameters and
moreover, we can tell which distribution is appropriate to use by comparing the variance
with the mean, i.e. variance is greater than mean (negative binomial), equal (Poisson) or
less than the mean (binomial). These three frequency distributions are the complete set
of distribution in the (a, b, 0) class.

Definition 2.3.1. Let pk be the pf of a discrete random variable. It is a member


of the (a, b, 0) class of distributions provided that there exists constants a and b such
that
pk b
=a+ , k = 1, 2, 3, . . .
pk−1 k

This recursion describes the relative size of successive probabilities in the counting distri-
bution. Panjer and Willmot found that negative binomial, Poisson and binomial are the
only possible distribution satisfying this recursive formula.

UECM3463 Loss Models May 2019 6


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

1. (a, b, 0) for the Poisson distribution


e−λ λk
pk = P (N = k) = k!
p1 λe−λ
p0
=a+b= e−λ
= λ ––––(1)
p2 b 2
λ e −λ λ
p1
=a+ 2
= 2λe −λ = 2 ––––(2)

(1) − (2) : 2b = λ2 =⇒ b = λ and a = 0.

2. (a, b, 0) for the Negative Binomial Distribution


 1 r β k
  
k+r−1
pk = k 1+β 1+β
r 
1 β
p1 r 1+β 1+β rβ
p0
=a+b= 1
r = 1+β ––––(1)
1+β
r(r+1)
r 2
1 β
p2 b 2 1+β 1+β (r+1)β
p1
=a+ 2 = 1
r
β
 = 2(1+β) ––––(2)
r 1+β 1+β

(1) − (2) : b
2
= (r−1) β
 2 1+β
β
=⇒ b = (r − 1) , 1+β
   
rβ β β
a= 1+β
− (r − 1) 1+β = 1+β .

3. (a, b, 0) for the Binomial Distribution


pk = m
 k
k
q (1 − q)m−k
p1 mq(1−q)m−1 mq
p0
=a+b= (1−q)m
= 1−q ––––(1)
m! 2 (1−q)m−2
q
p2
p1
= 2!(m−2)!
mq(1−q)m−1
= (m−1)q
2(1−q)
 
b q
(1) − (2) : m − m−1
2
= 2 1−q
 
q
=⇒ b = (m + 1) 1−q ,
 
mq q q
a = 1−q − (m + 1) 1−q = − 1−q

β q
• Since a = 1+β
> 0, a = 0 and a = − 1−q < 0 for the negative binomial, Poisson
and binomial, respectively. Therefore, we can use the sign of a to determine the
distribution belongs to the (a, b, 0) class. The values of a and b are as summarized
in Table 2.1.

• The recursive formula can be rewritten as


pk
k = ak + b, k = 1, 2, 3, . . .
pk−1

• The expression on the left-hand side is a linear function in k.

UECM3463 Loss Models May 2019 7


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Distribution a b p0
Poisson 0 λ e−λ
q q
Binomial − 1−q (m + 1) 1−q (1 − q)m
β β
Negative Binomial 1+β
(r − 1) 1+β (1 + β)−r
β
Geometric 1+β
0 (1 + β)−1

Table 2.1: Formula summary for members of the (a, b, 0) class

• Note from Table 2.1 that the slope a of the straight line is 0 for the Poisson dis-
tribution, is negative for the binomial distribution, and is positive for the negative
binomial distribution, including the geometric.

• This relationship suggests a graphical way of indicating which of the three distribu-
tions might be selected for fitting to data. Begin by plotting
p̂k nk
k =k
p̂k−1 nk−1
against k.

• The observed values should form approximately a straight line if one of these models
is to be selected, and the value of the slope should be an indication of which of the
models should be selected.

• Note that this cannot be done if any of the nk are 0. Hence this procedure is less
useful for a small number of observations.

Example 2.3.2. Consider the accident data in Table 6.2, For the 9,461 automobile insur-
ance policies studied, the number of accidents under the policy is recorded in the table.
Also recorded in the table is the observed value of the quantity that should be linear.
n
Number of accidents, k Number of policies, nk kn k
k−1
0 7840
1 1317 0.17
2 239 0.36
3 42 0.53
4 14 1.33
5 4 1.43
6 4 6.00
7 1 1.75
8+ 0

Table 2.2: Accident profile from Thyrion [173]

UECM3463 Loss Models May 2019 8


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

knk
Figure 2.1: Plot of ratio nk−1
against k

Solution: Figure 2.1 plots the value of the quantity of interest against k, the num-
ber of accidents. It can be seen from the graph that the quantity of interest looks
approximately linear except for the point at k = 6. From the graph, it can be seen
that the slope is positive and the data appear approximately linear, suggesting that
the negative binomial distribution is an appropriate model.
Whether or not the slope is significantly different from 0 is also not easily judged
from the graph. Graphically, it is difficult to distinguish between the Poisson and the
negative binomial distribution because the Poisson requires a slope of 0. However, we
can say that the binomial distribution is probably not a good choice since there is no
evidence of a negative slope. In this case it is advisable to fit both the Poisson and
negative binomial distributions and conduct a more formal test to choose between
them.

Example 2.3.3. Claim frequency follows a distribution in the (a, b, 0) class. You are
given that
• the probability of 4 claims is 0.2734375
•the probability of 5 claims is 0.21875
• the probability of 6 claims is 0.109375
Calculate the probability of no claims.

UECM3463 Loss Models May 2019 9


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Example 2.3.4. For an auto collision coverage, frequency of claims follows a distribution
from the (a, b, 0) class. You are given that
• the probability of 1 claims is 0.03368973
• the probability of 2 claims is 0.08422434
• the probability of 3 claims is 0.1403739.
Calculate the average number of claims.

Example 2.3.5. For a distribution in the (a, b, 0) class, you are given that p2 = 0.25p1
and p4 = 0.225p3 . Determine p2 .

Example 2.3.6. You are given the following claim frequency data:
Frequency 0 1 2 3 4
Number of insureds 20 24 10 6 3
Which probability distribution is suggested by this data based on (i) successive ratios of
probabilities and (ii) moments?
A. (i) Binomial, (ii) Binomial B. (i) Poisson, (ii) Binomial
C. (i) Negative Binomial, (ii) Poisson D. (i) Negative Binomial, (ii) Negative Binomial.
E. None of the above.

Example 2.3.7. For a random variable in the (a, b, 0) class


p3 p8
= 1.6; = 0.975.
p2 p7
Determine the mode of this random variable.

2.4 The (a, b, 1) Class

2.4.1 Truncation and Modification at Zero


• For insurance count data, the probability at zero is the probability that no claims
occur during the period under study.

• For applications in insurance where the probability of occurrence of a loss is low,


the probability at zero has the largest value.

• Thus, it is important to pay special attention to the fit at this point. However, the
three discrete distributions of (a, b, 0) class are inadequate because they do not give
appropriate probability to 0 claims.

UECM3463 Loss Models May 2019 10


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Definition 2.4.1. Let pk be the pf of a discrete random variable. It is a member


of the (a, b, 1) class of distributions provided that there exists constants a and b such
that
pk b
=a+ , k = 2, 3, 4, . . .
pk−1 k

Note that the only difference from the (a, b, 0) class is that the recursion begins at p1
rather than p0 .

• To obtain a distribution in this class is to take one from the (a, b, 0) class and
truncate it at 0, i.e., make the probability of 0 equal to zero, and then scale all the
other probabilities so that they add up to 1.

• We distinguish between the situations in which p0 = 0 and those where p0 > 0. The
first subclass is called the zero-truncated distributions. The second subclass is
referred to as the zero-modified distributions.

1. Let pk = P (N = k). When referring to a zero-truncated distribution, we use


pTk , and when referring to a zero-modified distribution, we use pM
k .

2. Let P (z) denote the pgf of a member of the (a, b, 0) class.


Let P M (z) denote the pgf of the corresponding member of the (a, b, 1) class;
that is,
pM
k = cpk , k = 1, 2, 3, . . .

and pM
0 is an arbitrary number.

3. Then

X ∞
X
M
P (z) = pM
k z
k
= pM
0 + pM
k z
k

k=0 k=1

X
= pM
0 +c pk z k = pM
0 + c[P (z) − p0 ].
k=1

4. Because P M (1) = P (1) = 1,

1 = pM
0 + c(1 − p0 ),

resulting in
1 − pM
0
c= or pM
0 = 1 − c(1 − p0 ).
1 − p0

UECM3463 Loss Models May 2019 11


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

5. Then, we have

1 − pM
0
P M (z) = pM
0 + [P (z) − p0 ]
1 − p0
1 − pM 1 − pM
 
0 0
= 1− 1+ P (z).
1 − p0 1 − p0

This is a weighted average of the pgfs of the degenerate distribution and the
corresponding (a, b, 0) member.

6. Furthermore,
1 − pM
0
pM
k = pk , k = 1, 2, . . .
1 − p0

7. Let P T (z) denote the pgf of the zero-truncated distribution corresponding to


an (a, b, 0) pgf P (z).
By setting pM
0 = 0, we have

pk
pTk = , k = 1, 2, . . .
1 − p0
which implies that
pM M T
k = (1 − p0 )pk

and
∞ ∞
T
X X pk k
P (z) = pTk z k = z
k=1 k=1
1 − p0
1
= [P (z) − p0 ].
1 − p0

We can also substituting pM


0 = 0 to

1 − pM
 
M 0
P (z) = pM
0 + [P (z) − p0 ],
1 − p0

thus
1
P T (z) = [P (z) − p0 ]
1 − p0
which implies that
P M (z) = pM M T
0 + (1 − p0 )P (z).

8. The mean and variance of the zero-truncated distributions are


1 X E(N )
E(N T ) = kpk =
1 − p0 1 − p0

UECM3463 Loss Models May 2019 12


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

1 X 2 E(N 2 )
E[(N T )2 ] = k pk =
1 − p0 1 − p0

V (N T ) = E[(N T )2 ] − E 2 (N T )
2
E(N 2 )

E(N )
= −
1 − p0 1 − p0
V (N ) − p0 E(N 2 )
=
(1 − p0 )2

9. As pM M T
k = (1 − p0 )pk , the mean and variance for zero-modified distributions
are:
E(N M ) = (1 − pM T
0 )E(N )

and
E[(N M )2 ] = (1 − pM T 2
0 )E[(N ) ]

Thus,

V (N M ) = E[(N M )2 ] − E 2 (N M )
= (1 − pM T 2 M 2 2 T
0 )E[(N ) ] − (1 − p0 ) E (N )

= (1 − pM T 2 M 2 T
0 )[E[(N ) ] − (1 − p0 )E (N )]

= (1 − pM T M 2 T
0 )[V (N ) + p0 E (N )]

= (1 − pM T M M 2 T
0 )V (N ) + p0 (1 − p0 )E (N )

• For example, the mean and variance of Poisson distribution are:


p0 = e−λ
λ
E(N T ) = 1−e−λ
−λ (λ+λ2 ) −λ (1+λ)]
V (N T ) = λ−e
(1−e−λ )2
= λ[1−e
(1−e−λ )2
λ
E(N M ) = (1 − pM 0 ) 1−e−λ
λ[1−e−λ (1+λ)] λ2
V (N M ) = (1 − pM 0 ) (1−e−λ )2 + pM0 (1 − pM
0 ) (1−e−λ )2

Notes:
pT
• For zero-truncated distribution, pT
k
= a + kb , k = 2, 3, . . . .
k−1
pM
• For zero-modified distribution, pM
k
= a + kb , k = 2, 3, . . . .
k−1

Example 2.4.2. For distribution from (a, b, 1) class, p1 = 0.4, p2 = 0.2, p3 = 0.1. Deter-
mine p0 .

Example 2.4.3. For a zero-modified Poisson distribution, p1 = 0.25, p2 = 0.1, calculate


the probability of 0.

UECM3463 Loss Models May 2019 13


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Example 2.4.4. A zero-truncated Geometric distribution has a mean of 3. Calculate the


probability of 5.

Example 2.4.5. A claim count distribution has zero-truncated binomial distribution


with m = 4, q = 0.2. Determine the probability of 2 or more claims.

Example 2.4.6. You are given pk denotes the probability that the number of claims
pn m!
equals k for k = 0, 1, 2, . . . . Suppose pm
= n!
, m ≥ 0, n ≥ 0. Using corresponding
zero-modified claim count distribution with pM
0 = 0.1. Calculate pM
1 .

Example 2.4.7. For a discrete distribution, you are given


•p0 = 0.8
pk−1
•pk = 4k
, for k > 1.
Calculate the mean and variance of the distribution.

Example 2.4.8. For a zero-modified distribution from (a, b, 1) class, you are given

p3 = 0.18737; p4 = 0.06246; p5 = 0.00833.

Determine the mode of distribution.

2.4.2 Extended Truncated Negative Binomial (ETNB)


1. Although we have only discussed the zero-modified distributions of the (a, b, 0) class,
the (a, b, 1) class admits additional distributions.

2. The (a, b) parameter space can be expanded to admit an extension of the negative
binomial distribution to include cases where −1 < r < 0.

3. For the (a, b, 0) class, r > 0 is required. By adding the additional region to the
sample space, the “extended” truncated negative binomial (ETNB) distribution
has parameter restrictions β > 0, r > −1, r 6= 0.

4. When r → 0, the limiting case of the ETNB is the logarithmic distribution with
[β/(1 + β)]k
pTk = , k = 1, 2, 3, . . .
k ln(1 + β)
The pgf of the logarithmic distribution is
ln[1 − β(z − 1)]
P T (z) = 1 −
ln(1 + β)

UECM3463 Loss Models May 2019 14


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

5. The zero-modified logarithmic distribution is created by assigning an arbitrary prob-


ability at zero and reducing the remaining probabilities.

Proof:

X
P T (z) = pTk z k
k=1
∞  k
1 X β 1 k
= ·z
ln(1 + β) k=1 1 + β k
  
1 zβ
= − ln 1 −
ln(1 + β) 1+β
− ln( 1+β−zβ
1+β
) ln[1 − β(z − 1)]
= =1− .
ln(1 + β) ln(1 + β)

Example 2.4.9. Determine the first four probabilities for an ETNB distribution with
r = −0.5 and β = 1. Do this for both truncated and modification versions with pM
0 = 0.6.

Example 2.4.10. For a zero-modified ETNB distribution, you are given p1 = 0.72,
p2 = 0.06 and p3 = 0.01.

(i) Determine the probability of 0.

(ii) Determine the variance of the distribution.

2.5 Advanced Discrete Distributions

2.5.1 Compound Frequency Models


• A larger class of distributions can be created by the processes of compounding
any two discrete distributions.

• The term compounding reflects the idea that the pgf of the new distribution
PS (z) is written as
PS (z) = PN [PM (z)],

where PN (z) and PM (z) are called the primary and secondary distributions,
respectively.

• Let N be a counting random variable with pgf PN (z).

• Let M1 , M2 , . . . be identically and independently distributed counting random

UECM3463 Loss Models May 2019 15


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

variables with pgf PM (z).


Assuming that the Mj s do not depend on N , the pgf of the random sum
S = M1 + M2 + · · · + MN is PS (z) = PN [PM (z)].
Proof:

X ∞ X
X ∞
k
PS (z) = P (S = k)z = P (S = k | N = n)P (N = n)z k
k=0 k=0 n=0

X ∞
X
= P (N = n) P (M1 + M2 + · · · + MN = k | N = n)z k
n=0 k=0
X∞
= P (N = n)[PM (z)]n
n=0

= PN [PM (z)].

Example 2.5.1. Demonstrate that any zero-modified distribution is a compound distri-


bution.

• The probability of exactly k claims can be written as



X
P (S = k) = P (S = k | N = n)P (N = n)
n=0

X
= P (M1 + M2 + · · · + MN = k | N = n)P (N = n)
n=0

X
= P (M1 + M2 + · · · + MN = k)P (N = n).
n=0

• Letting gn = P (S = n), pn = P (N = n) and fn = P (M = n), this is rewritten


as ∞
X
gk = pn fk∗n ,
n=0

where fk∗n , k = 0, 1, . . . is the “n-fold convolution” of the function fk , k =


0, 1, . . . .

• When PN (z) is chosen to be a member of the (a, b, 0) class,


 
b
pk = a + pk−1 , k = 1, 2, . . . ,
k

UECM3463 Loss Models May 2019 16


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

then a simple recursive formula can be used. This formula avoids the use of
convolutions and thus reduces the computations considerably.

Theorem 2.5.2. If the primary distribution is a member of the (a, b, 0) class, the
recursive formula is
k  
1 X bj
gk = a+ fj gk−j , k = 1, 2, 3, . . .
1 − af0 j=1 k

Theorem 2.5.3. For the (a, b, 1) class, the recursive formula is


Pk
[pM M
1 − (a + b)p0 ]fk + j=1 (a + bj/k)fj gk−j
fS (k) = gk = , k = 1, 2, 3, . . . .
1 − af0

Theorem 2.5.4. For any compound distribution,g0 = PN (f0 ), where PN (z) is the pgf
of the primary distribution and f0 is the probability that the secondary distribution
takes on the value zero.

Example 2.5.5. An actuary has created a compound claims frequency model with the
following properties:
• The primary distribution is the negative binomial with pgf

PN (z) = [1 − 3(z − 1)]−2 ,

• The secondary distribution is the Poisson with pgf

PM (z) = eλ(z−1) ,

• The probability of no claims equals 0.067.


Calculate λ.

2.5.2 Compound Poisson Distribution


• This model arises from the fact that the Poisson distribution is often a good model
to describe the number of claim-causing accidents, and the number of claims from
an accident is often itself a random variable.

UECM3463 Loss Models May 2019 17


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

• There is a close connection between the compound Poisson distributions and the
mixed Poisson frequency distributions

• The compound Poisson pgf may be expressed as

P (z) = exp{λ[Q(z) − 1]}

where Q(z) is the pgf of the secondary distribution.

Example 2.5.6. Obtain the pgf for the Poisson-ETNB distribution and show that it
looks like the pgf of a Poisson-negative binomial distribution.

Example 2.5.7. Calculate g3 for the Poisson-ETNB distribution where λ = 3 for the
Poisson distribution and r = −0.5, β = 1 for the ETNB distribution.

2.5.3 Mixed Poisson Distribution


Many compound distributions can arise in a way that is very different from compounding.
In this section, we examine mixture distributions by treating one or more parameters as
being ”random” in some sense.

Mixed Frequency Distribution


• Let P (z | θ) denote the pgf of the number of events (e.g., claims) if the risk parameter
is known to be θ.

• Let U (θ) = P (Θ ≤ θ) be the cdf of Θ, where Θ is the risk parameter, which is


viewed as a random variable.

• Let u(θ) be the pf or pdf of Θ. Then,


Z X
P (z) = P (z | θ)u(θ)dθ or P (z) = P (z | θj )u(θj )

is the unconditional pgf of the number of events (where the formula selected depends
on whether Θ is discrete or continuous)

• The corresponding probabilities are denoted by


Z X
pk = pk (θ)u(θ)dθ or pk = pk (θj )u(θj )

• The mixing distribution denoted by U (θ) may be of the discrete or continuous


type or even a combination of discrete and continuous types.
Example 2.5.8. Demonstrate that the zero-modified distributions may be created by
using a two-point mixture.

UECM3463 Loss Models May 2019 18


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Mixed Poisson Distribution

• Let pk (θ) have the Poisson distribution.

• Let P(z) be the pgf of a mixed Poisson distribution with arbitrary mixing distribu-
tion U (θ).

• By introducing a scale parameter λ, we have


Z Z
λθ(z−1)
P (z) = e u(θ)dθ = [eλ(z−1) ]θ u(θ)dθ
 
λ(z−1) θ
= E [e ] = MΘ [λ(z − 1)],

where MΘ (z) is the mgf of the mixing distribution.

• Therefore, P 0 (z) = λMΘ0 [λ(z − 1)] and with z = 1 we obtain E(N ) = λE(Θ), where
N has the mixed Poisson distribution.

• Also, P 00 (z) = λ2 MΘ00 [λ(z − 1)], implying that E[N (N − 1)] = λ2 E(Θ2 ) and, there

V (N ) = E[N (N − 1)] + E(N ) − [E(N )]2


= λ2 E(Θ2 ) + E(N ) − λ2 [E(Θ)]2
= λ2 V (Θ) + E(N )
> E(N ),

and thus for mixed Poisson distributions the variance is always greater than the
mean.

• There is also an important connection between mixed Poisson distributions and


compound Poisson distributions.

Definition 2.5.9. A distribution is said to be infinitely divisible if for all values


of n = 1, 2, 3, . . . its characteristic function ϕ(z) can be written as

ϕ(z) = [ϕn (z)]n ,

where ϕn (z) is the characteristic function of some random variable.

• “Characteristic function” could have been replaced by “moment generating


function” or “probability generating function” or some other transform. That
is, if the definition is satisfied for one of these transforms, it will be satisfied for all
others that exist for the particular random variable.

UECM3463 Loss Models May 2019 19


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Theorem 2.5.10. Suppose P (z) is a mixed Poisson pgf with an infinitely divisible
mixing distribution. Then P (z) is also a compound Poisson pgf and may be expressed
as
P (z) = eλ[P2 (z)−1] ,

where P2 (z) is a pgf. If we insist that P2 (0) = 0, then P2 (z) is unique.

If one chooses any infinitely divisible mixing distribution, the corresponding mixed
Poisson distribution can be equivalently described as a compound Poisson distribution.

Example 2.5.11. Demonstrate that a gamma mixture of Poisson variables is negative


binomial.

2.6 The Impact of Deductible on Claim Frequency


An important component in analyzing the effect of policy modifications pertains to the
change in the frequency distribution of payments when the deductible (ordinary or fran-
chise) is imposed or changed. When a deductible is imposed or increased, there will
be fewer payments per period, while if a deductible is lowered, there will be more pay-
ments.

• Suppose Xj , the severity, represents the ground-up loss on the jth such loss
and there are no coverage modifications. Let N L denote the number of losses.

• Consider a coverage modification such that v is the probability that a loss will
result in a payment. For example, if there is a deductible of d, v = P (X > d).

• Define the indicator random variable Ij by



 1, if the j th loss results in a payment
Ij =
 0, otherwise

Then, Ij ∼ Bernoulli(v) and pgf of Ij is

PIj (z) = 1 − v + vz.

• Then,
N P = I1 + · · · + IN L

represents the number of payments.

UECM3463 Loss Models May 2019 20


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

• If I1 , I2 , . . . are mutually independent and are also independent of N L , then


N P has a compound distribution with N L as the primary distribution and a
Bernoulli secondary distribution. Thus

PN P (z) = PN L [PIj (z)] = PN L [1 + v(z − 1)].

• In the important special case in which the distribution of N L depends on a


parameter θ such that

PN L (z) = PN L (z; θ) = B[θ(z − 1)],

where B(z) is functionally independent of θ. Then,

PN P (z) = B[θ(1 − v + vz − 1)] = B[vθ(z − 1)] = PN L (z; vθ).

• This result implies that N L and N P are both from the same parametric family
and only the parameter θ need be changed.

Notes:

1. For Poisson distribution


N L ∼ P oi(λ)
PN L (z) = eλ(z−1)
PN P (z) = PN L (z; vθ) = evλ(z−1)
∴ N P ∼ P oi(vλ).

2. For Binomial distribution


N L ∼ Bin(m, q)
PN L (z) = (1 + q(z − 1))m
PN P (z) = PN L (z; vθ) = (1 + vq(z − 1))m
∴ N P ∼ Bin(m, vq).

3. For Negative Binomial distribution


N L ∼ N B(r, β)
PN L (z) = (1 + β(z − 1))−r
PN P (z) = PN L (z; vθ) = (1 + vβ(z − 1))−r
∴ N P ∼ N B(r, vβ).

UECM3463 Loss Models May 2019 21


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Example 2.6.1. You are given

n P (N = n) x P (X = x)
1 0.8 100 0.2
2 0.2 200 0.7
500 0.1

An insurance for the losses has an ordinary deductible of 100 per loss. Calculate the
probability of 2 or more losses occur.

Example 2.6.2. The number of annual losses has a Poisson distribution with a mean of
5. The size of each loss has a two-parameter Pareto distribution with θ = 10 and α = 2.5.
An insurance for the losses has a ordinary deductible of 5 per loss. Calculate the expected
value of the number of losses.

Example 2.6.3. The number of students of taking an actuarial exam has a negative
binomial distribution with parameters r = 10, β = 1.5. Each student has a 0.4 probability
of passing. Determine the probability that 3 or more students pass.

Example 2.6.4. The frequency distribution of the number of losses when there is no
deductible is Binomial with m = 3 and q = 0.8. Loss amounts have a Weibull distribu-
tion with τ = 0.3 and θ = 1000. Determine the expected number of payments when a
deductible of 200 is applied.

Example 2.6.5. The losses on an auto comprehensive coverage have a Pareto distribution
with parameters α = 2, θ = 1000. The number of losses has a geometric distribution
with an average of 0.2 losses per year. Loss sizes are affected by 10% inflation. A 250
deductible is imposed. Calculate the variance of the frequency of claims after inflation
and the deductible.

UECM3463 Loss Models May 2019 22


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

The results can be generalized for zero-modified and zero-truncated distributions.

• Suppose N L depends on paramters α and θ such that

PN L (z) = PN L (z; θ, α)
B[θ(z − 1)] − B(−θ)
= α + (1 − α)
1 − B(−θ)

• Note that α = PN L (0) = P (N L = 0), and so the modified probability at zero.

B[vθ(z − 1)] − B(−θ)


PN P (z) = α∗ + (1 − α∗ )
1 − B(−θ)
= PN L (z; vθ, α∗ )

where α∗ = P (N P = 0) = PN P (0).

• In particular, B(z) = 1 + ln(1 − z) yield the zero-modified logarithmic distri-


bution, and B(z) = (1 − z)−r for −1 < r < 0 yield the ETNB distribution.

Example 2.6.6. Let the frequency distribution be zero-modified negative binomial with
r = 2, β = 3 and pM
0 = 0.4. Illustrate the effect when a deductible of 250 is imposed on a
Pareto distribution with α = 3 and θ = 1000.

2.7 Exposure Modifications


• Assume that the current portfolio consists of n entities, each of which could produce
claims. Let Nj be the number of claims produced by the jth entity.

• Then N = N1 + · · · + Nn . If we assume that the Nj s are independent and identically


distributed, then
PN (z) = [PN1 (z)]n .

• Now suppose the portfolio is expected to expand to n∗ entities with frequency N ∗ .


Then
∗ ∗ /n
PN ∗ (z) = [PN1 (z)]n = [PN (z)]n .

• Thus, if N is infinitely divisible, the distribution of N ∗ will have the same form as
that of N , but with modified parameters.

UECM3463 Loss Models May 2019 23


CHAPTER 2. MODELS FOR CLAIM FREQUENCIES

Example 2.7.1. It has been determined from past studies that the number of workers
compensation claims for a group of 300 employees in a certain occupation class has nega-
tive binomial distribution with β = 0.3, r = 10. Determine the frequency distribution for
a group of 500 such individuals.

Example 2.7.2. Suppose claims on a portfolio of 500 policies are independent and iden-
tically distributed. Each has a geometric distribution with parameter β = 0.001. If 250
additional policies are put on the books, what is the new probability of no claims during
the years?

Example 2.7.3. For an employee health coverage got 50 individuals, the aggregate claims
frequency distribution is negative binomial with mean 10 and variance 20. The group then
expands to 60. Calculate the probability of 10 claims for the expanded group.

UECM3463 Loss Models May 2019 24