Collective Risk Model

School of Risk and Actuarial Studies
ACTL5106: Insurance Risk Models

Module 2: Collective Risk Modelling1
Vincent Tu
School of Risk & Actuarial Studies

UNSW Business School
July 21, 2018
1
1/47 References: MW 1/ (A 12)
Plan
1 Introduction
2 The Individual Risk Model
Definition
Convolutions of random variables
Using generating functions
3 The Collective Risk Model (Compound distributions)
Definition
The distribution of S
4 Explicit claims count distributions
Introduction
Binomial distribution
Poisson distribution
Mixed Poisson distribution
Negative-binomial distribution
5 Parameter estimation
Introduction
Method of moments
Maximum likelihood estimators
2/47
Introduction
Models for aggregate losses

A portfolio of contracts or a contract will potentially experience a
sequence of losses:
X1 , X2 , X3 , . . .
We are interested in the aggregate sum S of these losses over a
certain period of time.
How many losses will occur? Frequency Freq x Size
if deterministic (n) −→ individual risk model
if random (N) −→ collective risk model <--- Our Focus
How do they relate to each other?
usual assumption: iid
When do these losses occur?
usual assumption: no time value of money
−→ short term models
How big are these losses?
3/47
The Individual Risk Model
Definition
In the Individual Risk Model

n
X
S = X1 + · · · + Xn = Xi ,
i=1
where Xi , i = 1, 2, ..., n, are iid claims. There are several methods

to get probabilities about S:
get the whole distribution of S (if possible)
1. Convolutions convolu n times, too tedious
2. Generating functions better
approximate with the help of the moments of S (Module 4)
4/47
In probability, the operation of determining the distribution of

the sum of two random variables is called a convolution. It is
denoted by
FX +Y = FX ∗ FY . convolute x, y ,z
-> x+y convolute z
The result can then be convoluted with the distribution of
another random variable. For instance,
FX +Y +Z = FZ ∗ FX +Y .
Convolution can be done for both discrete and continuous

random variables. It is also possible for mixed rv’s, but it is
more complicated.
5/47
Formulas
* Review Convolution !
* In short
Discrete case:
P X=x
df: FX +Y (s) = Px FY (s − x) fX (x)
pmf: fX +Y (s) = x fY (s − x) fX (x) Y = s -x
Continuous case:
Rs
cdf: FX +Y (s) = −∞ FY (s − x) fX (x) dx
Rs
pdf: fX +Y (s) = −∞ fY (s − x) fX (x) dx
Examples
discrete case: see [A], Example 2.3.1 on page 35
continuous case: see [A], Example 2.3.2 on page 36
6/47
Numerical example Discrete Claim Size
Consider 3 discrete r.v.’s with probability mass functions

1 1 1
f1 (x) = , , for x = 0, 1, 2
4 2 4
1 1
f2 (x) = , for x = 0, 2
2 2
1 1 1
f3 (x) = , , for x = 0, 2, 4
4 2 4
Calculate the pmf f1+2+3 and the df F1+2+3 of the sum of the
three random variables.
7/47
Solution
x f1 (x) f2 (x) f1+2 (x) f3 (x) f1+2+3 (x) F1+2+3 (x)

0 1/4 1/2 1/8 1/4 1/32 1/32
1 1/2 0 2/8 0 2/32 3/32
wanna 2 1/4 1/2 2/8 1/2 4/32 7/32
cover 3 0 0 2/8 0 6/32 13/32
all
cases 4 0 0 1/8 1/4 6/32 19/32
5 0 0 0 0 6/32 25/32
6 0 0 0 0 4/32 29/32
7 0 0 0 0 2/32 31/32
8 0 0 0 0 1/32 32/32
(0,2) (1,1) (2,0)
f1+2 (2) = 1/4 · 1/2 + 1/2 · 0 + 1/4 · 1/2
f1+2+3 (4) = 1/8 · 1/4 + 2/8 · 0 + 2/8 · 1/2 + 2/8 · 0 + 1/8 · 1/4
8/47
Using generating functions **review pgf
There is a 1-1 relation between a distribution and its mgf or

pgf.
Therefore, we can identify the distribution of S by considering
it’s mgf or pgf.
If we know the mgf or pgf function of S,
Sometimes, MS (t) or pS (t) can be recognised: this is the case
for infinitely divisible distributions (Normal, Poisson, Inverse
Gaussian, . . . ) and certain other distributions (Binomial,
Negative binomial)
Otherwise, MS (t) or pS (t) can be expanded numerically to get
moments and/or probabilities
9/47
Calculation of generating functions must be IID
The moment generating function ( mgf ) of S is

h i h i h i
MS (t) = E e tS = E e t(X1 +...+Xn ) = E e tX1 · · · e tXn
and if losses are independent we have

h i h i h i
MS (t) = E e tS = E e tX1 · · · E e tXn = MX1 (t) · · · MXn (t).
The probability generating function (pgf) of S is

h i h i
pS (t) = E t S = E t X1 +...+Xn write down def in exam
and if losses are independent we have

h i h i h i
Ps(t) S X1
MS (t) = E t = E t · · · E t Xn = pX1 (t) · · · pXn (t).
10/47
Example
Consider a portfolio of 10 contracts. The losses Xi ’s for these

contracts are iid rv’s with mean 100 and variance 100. Determine
the distribution, the expected value and the variance of S if these
losses are
10 iid N(100, 100) S~ N(1000, 1000)
1 Normal;
solve alpha & beta by mean, var --> get mgf of x
2 Gamma; --> apply to mgf of 10x
3 Poisson. 10 iid Poi(100) , S~ Poi(1000)
11/47
The Collective Risk Model (Compound distributions)
Definition
Introduction
Two models, depending on the assumption on the number of

losses:
deterministic - n
main focus on the claims of individual policies
(whose number is a priori known)
−→ Individual Risk Model
random - N
main focus on claims of a whole portfolio
(whose number is a priori unknown)
−→ Collective Risk Model
this is another way of separating frequency and severity
In this section we focus on the Collective Risk Model.
12/47
Definition
Definition
In the Collective Risk Model, aggregate losses become

N
X replace n by N
S = X1 + . . . + XN = Xi . 2 sources of randomness
i=1
This is a random sum. We make the following assumptions:

N is the number of claims
Xi is the amount of the ith claim
the Xi ’s are iid with
(c)df G (x)
p(d/m)f p(x)
the Xi ’s and N are mutually independent
13/47
Moments of S
We have see annoted slides
E [S] = E [E [S|N]] = E [NE [X ]] = E [N]p1 ,
and
Var (S) = E [Var (S|N)] + Var (E [S|N])

= E [NVar (X )] + Var (p1 N)
= E [N]Var (X ) + E [X ]2 Var (N) <--- usually stop here
enough
= E [N](E [X 2 ] − E [X ]2 ) + E [X ]2 Var (N)
= E [N]E [X 2 ] + E [X ]2 {Var (N) − E [N]} .
14/47
Moment generating function of S
w.r.t N w.r.t x
It is possible to get MS (t) as a function of MX (t) and MN (t):

h i h h ii
MS (t) = E e tS = E E e t(X1 +X2 +...+XN ) N

h i h i N is treated as
= E Mx (t)N = E e N ln MX (t) Constant
constant
= MN (ln MX (t))
If you know mgf or X and N you know S
15/47
Example (A 12.2.1)
Assume that N is geometric with probability of success p:
Pr[N = n] = pq n , n = 0, 1, . . . ,
where 0 < q < 1 and p = 1 − q. We have then

∞
X p
MN (t) = E [e tN ] = pq n e tn = ,
1 − qe t by Geometric Series
n=0
and thus replace by ln(Mx(t))
p p
MS (t) = MN (ln MX (t)) = ln M (t)
= .
1 − qe X 1 − qMX (t)
16/47
Distribution of S
It is possible to get a fairly general expression for the df of S by

conditioning on the number of claims: condition on N
∞
X ∞
X
FS (x) = Pr[S ≤ x|N = n] Pr[N = n] = P ∗n (x) Pr[N = n],
n=0 n=0
where P ∗n (x) is the n-th convolution of P with the definition

P ∗0 (x) = 1 for x ≥ 0.
Note that
N will always be discrete, so this works for any type of rv X
(continuous, discrete or mixed)
however, the type of S will depend on the type of X
17/47
Distribution of S if X is continuous
(1)
If X is continuous, S will generally be mixed:
with a mass at 0 because of Pr[N = 0] (if positive)
continuous elsewhere, but with a density integrating to
1 − Pr[N = 0] <-- conditional on more than 1 claim
Example (A 12.2.3)
Consider the previous example with P(x) = 1 − e −x . Note
that Pr[N = 0] = p.
It is shown in A that the mgf of S can be written as
p h i
= pE e t·0 + (1 − p)E e tY ,

MS (t) = p + q
p−t
P(N=0) P(N>0)
where Y is an exponential rv with parameter p.
18/47
Distribution of S if X is discrete
(2)
For discrete X ’s we can get a similar expression for the pmf of S:
∞
X ∞
X
fS (x) = Pr[S = x|N = n] Pr[N = n] = p ∗n (x) Pr[N = n],
n=0 n=0
where p ∗0 (0) = 1 (and thus 0 anywhere else).

this can be implemented in a table and/or in a program
however, if the range of N goes really to the infinity,
calculating fS (x) may require an infinity of convolutions of X
this formula is more efficient if the number of possible
outcomes for N is small
p ∗n (x) can be calculated using de Pril’s algorithm
(see Module 4)
19/47
Example A 12.2.2 (discrete X , tabular approach)
Assume p(1) = 0.5, p(2) = 0.4, p(3) = 0.1.
The convolution p ∗(n+1) (x) = y p ∗n (x − y )p(y )

P
The number
P of columns depends onP the range of N
fS (x) = ∞ n=0 p ∗n (x) Pr(N = n) = 3 ∗n
n=0 p (x) Pr(N = n),
which is the sumproduct of the row x and row Pr[N = n]:
20/47 e.g. fS (3) = 0.1 · 0.3 + 0.4 · 0.4 + 0.125 · 0.2.
Distribution of S if X is mixed
If X is mixed, S will generally be mixed:

with a mass at 0 because of Pr[N = 0] and Pr[X = 0] (if
positive)
mixed (if X is not continuous for x > 0) or continuous
elsewhere
with a density integrating to something ≤ 1 − Pr[N = 0]
21/47
Explicit claims count distributions
Introduction
There are several popular options for the distribution of N:

Poisson(λ)
E [N] = Var (N) = λ
S is compound Poisson with parameters (λ, P(x))
E [S] = λE [X ] Var (S) = λE [X 2 ] MS (t) = e λ(MX (t)−1)
Negative Binomial(r , p)
E [N] < Var (N)
S is compound Negative Binomial with parameters (r , p, P(x))
Binomial(m, p)
E [N] > Var (N)
S is compound Binomial with parameters (m, p, P(x))
less popular
A summary table is given in [A], Table 12.3.1 on page 376.
22/47
Introduction
Exposure Missed first 10 mins
It makes no sense to talk about frequency in an insurance

portfolio without considering exposure
Werner and Modlin (2010, Chapter 4) define exposure as “the
basic unit that measures a policy’s exposure to loss”
One primary criterion for choosing an exposure base is that it
“should be directly proportional to expected loss”
MW calls exposure ‘volume’, denoted v
In our case, we will assume that it directly affects the
likelihood of a claim to occur - the frequency, such that N/v
is normalised
MW defines
pk = Pr[N = k]
23/47
fixed volume v ∈ N
fixed default probability p ∈ (0, 1) (expected claims frequency)
pmf of N ∼ Binom(v , p) is

v k
pk = Pr[N = k] = p (1 − p)v −k
k
same as a sum of Bernoulli

makes sense for homogenous portfolio with unique possible
events: more a life insurance model (rather than GI)
24/47
fixed volume v > 0

expected claims frequency λ > 0
pmf of N ∼ Poi(λv ) is
(λv )k
pk = Pr[N = k] = e −λv
k!
note: only λv matters (useful)
Lemma 2.9: increase volume while keeping E [N] fixed in a
binomial model leads to a Poisson distribution (more so for
small p compared to v )
25/47
Theorem 2.12: Aggregation property (also A Th 12.4.1)

Assume S1 , . . . , Sn are independent with Sj ∼ CompPoi(λj vj , Gj )
for all j. The aggregated claim is also a compound Poisson r.v.:
n
X Gj = pdf of Yj
S= Sj ∼ CompPoi(λv , G ), with
j=1
n n n
X X X λj vj
v= vj , λv = vj λj , G= Gj .
λv
j=1 j=1 j=1
sum of all lambda Weighted Avg of each line of
So what? business
Independent portfolios of losses can be easily aggregated
Total claims paid over n years is compound Poisson, even if
the severity and frequency of losses vary across years (and the
time value of money can be approximated by a change of
scale on Gj for each year)
26/47 Homework 1: Prove the theorem above by using mgf.
Aggregation property- remarks
The aggregation property implies that we can use a

bottom-up modeling approach:
first model each sub-portfolio Sj independently with a
compound Poisson distribution; P
then obtain the total portfolio S = i Si by the aggregation
theorem.
This property is of special importance when we estimate λj
and Gj on the bottom (sub-portfolio) level.
27/47
Example (A Example 12.4.1)
Suppose that N1 , N2 , · · · , Nm are independent random variables.

Further, suppose that Ni follows Poisson(λi ). Let x1 , x2 , · · · , xm be
deterministic numbers. What is the distribution of
x1 N1 + · · · + xm Nm ? ~ Comp( lambda, size dist)
Ni ~ Poi(lambda) Ex. what is dist of 2*N1 + 4*N2 + 6*N3
28/47
Thinning of the Poisson point process

This good property only exist in Comp Poisson, but not in other Comp
Theorem (Thinning of the Poisson point process)

Suppose
S ∼ CompPoi(λv , G ),
PN
for k = 1, · · · , n, Sk = i=1 Yi I {Yi ∈ Ak } where Ak is a
disjoint partition of the total space for Y1 .
Then,
for k = 1, · · · , n, Sk ∼ CompPoi(λk v , Gk ) where
λk = λP (Y1 ∈ Ak ) and Gk (y ) = P(Y1 ≤ y |Y1 ∈ Ak ).
S1 , · · · , Sn are independent.
29/47
Large claims separation-an example of thinning of the

Poisson point process
Example 2.16 (large claims separation) Thinning ->
Suppose
S ∼ CompPoi(λv , G ); Choose a size M to separate the
Claim Size to small/large
A1 = {Y1 ≤ M} and A2 = {Y1 > M};
PN
for k = 1, 2, Sk = i=1 Yi I {Yi ∈ Ak }. -> convolute to get Dist
Then,
for k = 1, 2, Sk ∼ CompPoi(λk v , Gk ) where
λ1 = λP (Y1 ≤ M) = λG (M),
λ2 = λP (Y1 > M) = λ(1 − G (M)),
G1 (y ) = P(Y1 ≤ y |Y1 ≤ M) and G2 (y ) = P(Y1 ≤ y |Y1 > M).
S1 and S2 are independent.
This means that we can model the small and the large claims
layers completely separately and then obtain the total claim
amount distribution by a convolution of the two resulting
30/47 distribution functions.
Large claims separation (Example 2.16)
The large claims separation example (Example 2.16)

demonstrates a very important application of the thinning of
Poisson point process property.
For attritional and catastrophic claims, there often does not
exist one parametric distribution function that applies to the
entire range of possible values of Yi .
Idea here is to divide the process into different layers with
Proof by MGF different distributions
W1 = L1/(L1+L2)
1. Aggregate Comp ( L1, G1)
+ => Comp( L1 + L2, W1*G1+ W2*G2)
Comp ( L2, G2)
31/47
Poisson distribution Freq
A theorem and the sparse vector algorithm Size

Theorem If S ∼ compound Poisson(λ, p(xi ) = πi , i = 1, . . . , m),
then
S = x1 N1 + . . . + xm Nm , Scaled Poisson
where the Ni ’s
represent the number of claims of amount xi
are mutually independent
are Poisson(λi = λπi )
(Homework: Prove the theorem above by using mgf.)
Application:
Sparse vector algorithm: allows to develop an alternative
method for tabulating the distribution of S that is more
efficient as m is small.
S can be used to approximate the Individual Risk Model if
32/47
X = Ib (see section on Approximations)
The sparse vector algorithm: examples
A Example 12.4.2 Suppose S has a compound Poisson distribution

with λ = 0.8 and individual claim amount distribution
x Pr [X = x]
1 0.250
2 0.375
3 0.375
Compute fS (x) = Pr [S = x] for x = 0, 1, ..., 6.

This can be done in two ways:
Basic method (seen earlier in the lecture): requires to
calculate up to the 6th convolution of X
Sparse vector algorithm: requires no convolution of X
33/47
Solution - Basic Method

Convolution of x 6 convolution ( because 6 claims)
(1) (2) (3) (4) (5) (6) (7) (8) (9)
x p ∗0 (x) p (x) p ∗2 (x) p ∗3 (x) p ∗4 (x) p ∗5 (x) p ∗6 (x) fS (x)
0 1 - - - - - - 0.4493
1 - 0.250 - - - - - 0.0899
2 - 0.375 0.0625 - - - - 0.1438
3 - 0.375 0.1875 0.0156 - - - 0.1624
4 - - 0.3281 0.0703 0.0039 - - 0.0499
5 - - 0.2813 0.1758 0.0234 0.0010 - 0.0474
6 - - 0.1406 0.2637 0.0762 0.0073 0.0002 0.0309
P(N=n) n
(0.8)n
0 1 2 3 4 5 6
e −0.8 n!
0.4493 0.3595 0.1438 0.0383 0.0077 0.0012 0.0002
The convolutions are done the usual way

The fS (x) are the sumproduct of the row x and row Pr[N = n]
The number of convolutions (and thus of columns) will
increase by 1 for each new value of fS (x), until the infinity!
34/47
Convert S into sum of Scaled Poi

Solution - Sparse vector algorithm Transform Collective to Individual
Thanks to Theorem 12.4.2, we can write S = N1 + 2N2 + 3N3

(1) (2) (3) (4) (5) (6)
x Pr [N1 = x] Pr [2N2 = x] Pr [3N3 = x] Pr [N1 + 2N2 = x] fS (x)
multiple of 2 = (2)*(3) = (4)*(5)
0 0.818731 0.740818 0.740818 0.606531 0.449329
1 0.163746 0 0 0.121306 0.089866
2 0.016375 0.222245 0 0.194090 0.143785
3 0.001092 0 0.222245 0.037201 0.162358
4 0.000055 0.033337 0 0.030974 0.049906
5 0.000002 0 0 0.005703 0.047360
6 0.000000 0.003334 0.033337 0.003288 0.030923
xi 1 2 3
λi = λπi 0.2 0.3 0.3
(0.2)x (0.3)x/2 (0.3)x/3
Pr[Ni = x/i] e −0.2 x!
e −0.3 (x/2)! e −0.3 (x/3)!
Only two convolutions are needed: columns (5) and (6)

(5)[3] = .818731 · 0 + .163746 · .222245 + .016375 · 0 + .001092 · .740818
(6)[3] = .740818 · .037201 + 0 · .194090 + 0 · .121306 + .222245 · .606531
number of values needed (for fS (x)) doesn’t matter

35/47

N follows a mixed Poisson distribution if
N|Λ = λ be Poisson(λv ), and
Λ ∼ H with H(0) = 0, E [Λ] = λ and var (Λ) > 0.
We have then
∞ ∞
e −λv (λv )n
Z Z
Pr[N = n] = Pr[N = n|Λ = λ]dH(λ) = dH(λ)
0 0 n!
E [N] = E [E [N|Λ]] = E [Λv ]
Var (N) = E [Var (N|Λ)] + Var (E [N|Λ]) = E [Λv ] + Var (Λv )
h i h h ii h t
i
MN (t) = E e tN = E E e tN |Λ = E e Λv (e −1) = MΛ (v (e t − 1)
Interpretation:
Distribution of Λ represents different classes of risk
(inhomogeneous portfolio)
Same idea in greatest accuracy credibility
36/47
E [N] < Var (N)
Examples for N|Λ ∼ Poisson(Λv )
Λ ∼ gamma(α, β) and v = 1 (A Example 12.3.1)
MN (t) = MΛ (e t − 1)
α
β
=
β − (e t − 1)
α
β/(β + 1)
=
1 − [1 − β/(β + 1)]e t
which means N ∼ Negative Binomial(r = α, p = β/(β + 1)).

−→ S is compound Negative Binomial
Λ ∼ inverse Gaussian(α, β) and v = 1 (A Example 12.3.2)
−→ S is compound Poisson-inverse Gaussian
37/47
Negative Binomial distribution

Suppose N is a negative binomial distribution with
fixed volume v > 0
expected claims frequency λ > 0 and dispersion parameter
γ > 0,
Then,
the pdf of N is

k +γ−1 k
pk = Pr[N = k] = p (1 − p)γ
k
λv
with p = γ+λv .
N follows the following mixed Poisson distribution
Θ ∼ Gamma(γ, γ), and
conditionally, given Θ, N ∼ Poi(Θλv ).
38/47
Negative Binomial distribution-Interpretation
Recall that if N is a negative binomial distribution with

fixed volume v > 0
expected claims frequency λ > 0 and dispersion parameter
γ > 0,
then N follows the following mixed Poisson distribution
Θ ∼ Gamma(γ, γ), and
conditionally, given Θ, N ∼ Poi(Θλv ).
The Θ above reflects the uncertainty about the ‘true’
parameter of the Poisson distribution.
this is not diversifiable (remains even for large v ):
q
N
Vco = (λv )−1 + γ −1 → γ −1/2 > 0 for v → ∞.
v
39/47
Parameter estimation
Introduction
Estimation methods
You should be familiar with the main estimation methods:
Method of moments: Suppose for a r.v. X , the first kth
moments are finite and can be expressed as functions of the
parameters, θ1 , θ2 , · · · , θk :
µ1 := E [X ] = g1 (θ1 , · · · , θk )
..
.
µk := E [X k ] = g1 (θ1 , · · · , θk ).
Then the method of moments estimator, θ̂1 , · · · , θ̂k , is the
solution to the following equations:
µ̂1 = g1 (θ̂1 , · · · , θ̂k )
..
.
40/47 µ̂k = g1 (θ̂1 , · · · , θ̂k ).
Introduction
Lemma 2.26
Assume that (N1 , . . . , NT )0 is the vector of observations.
Here the problem is slightly complicated because our
observations may not be directly comparable due to varying
exposures v ’s.
Assume there exist strictly positive volumes v1 , . . . , vT such
that the components of (N1 /v1 , . . . , NT /vT ) are independent
with

Nt 2 Nt
λ=E and τt = Var ∈ (0, ∞), t = 1, . . . , T .
vt vt
Then the unbiased, linear estimator for λ with minimal
variance is:
T
!−1 T
MV
X 1 X Nt /vt
λ
b
T = .
t=1 t
τ2 t=1
τt2
41/47
Method of moments
Binomial case
Suppose Nt ∼ Binomial(vt , p)
Need to estimate p:
Note E [Nt /vt ] = p
τt2 = Var (Nt /vt ) = p(1 − p)/vt
Then, the unbiased, minimal variance estimator for p:
T
!−1 T
X 1 X Nt /vt
pbTMV =
τ2
t=1 t t=1
τt2
T
1 X
= PT Nt
s=1 vs t=1
T
X vt Nt
= PT
v
s=1 vs t
t=1
42/47
Method of moments
Poisson case
Poisson case for λ:

T T
bMV = P 1 vt Nt
X X
λT T
Nt = PT
s=1 vs t=1
v
s=1 vs t
t=1
43/47
Method of moments
Negative binomial case
More complicated, because:

Nt Nt
E = λ and Var = λ/vt + λ2 /γ = τt2 ,
vt vt
Unbiased (but not guaranteed minimal variance):

T T
bNB = P 1 vt Nt
X X
λ T T
Nt = PT
s=1 vs t=1
v
s=1 vs t
t=1
44/47
Method of moments
We need a sense of the dispersion for estimating the dispersion

parameter γ.
Let the weighted sample variance
T 2
1 X Nt bNB
VbT2 = vt − λT .
T −1 vt
t=1
Then we have
T PT !
(λbNB )2 1 X vt2
bTNB
γ = T
vt − Pt=1
T
,
V −λ T −1
b 2 b NB
T T t=1 t=1 vt
ONLY if VbT2 > λ

bNB . Otherwise use Poisson or binomial.
T
45/47
Binomial and Poisson cases
Estimators are identical to method of moments estimators. Or

conversely, the MLE estimators are actually unbiased.
binomial case for p:
T T
1 X X vt Nt
pbTMLE = PT Nt = PT = pbTMV
s=1 vs t=1 v
s=1 s
v t
t=1
Poisson case for λ:

T T
bMLE = P 1 vt Nt
X X
λ Nt = bMV
=λ
T T PT T
s=1 vs t=1
v
s=1 vs t
t=1
46/47
Negative binomial case
Assume N1 , . . . , NT are independent and NegBin(λvt , γ). The

bMLE , γ
MLE (λ bTMLE ) are the solution of
T
T
∂ X Nt + γ − 1
log + γ log(1 − pt ) + Nt log pt = 0,
∂λ∂γ Nt
t=1
with pt = λvt /(γ + λvt ) ∈ (0, 1).
47/47

Collective Risk Model

Uploaded by

Copyright:

Available Formats

Collective Risk Model

Uploaded by

Document Information

Original Description:

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

Collective Risk Model

Uploaded by

Copyright:

Available Formats

School of Risk and Actuarial Studies

ACTL5106: Insurance Risk Models

School of Risk & Actuarial Studies

July 21, 2018

Models for aggregate losses

The Individual Risk Model

In the Individual Risk Model

where Xi , i = 1, 2, ..., n, are iid claims. There are several methods

Convolutions of random variables

In probability, the operation of determining the distribution of

Convolution can be done for both discrete and continuous

Numerical example Discrete Claim Size

Consider 3 discrete r.v.’s with probability mass functions

x f1 (x) f2 (x) f1+2 (x) f3 (x) f1+2+3 (x) F1+2+3 (x)

Using generating functions **review pgf

There is a 1-1 relation between a distribution and its mgf or

Calculation of generating functions must be IID

The moment generating function ( mgf ) of S is

and if losses are independent we have

The probability generating function (pgf) of S is

and if losses are independent we have

Consider a portfolio of 10 contracts. The losses Xi ’s for these

Two models, depending on the assumption on the number of

In the Collective Risk Model, aggregate losses become

This is a random sum. We make the following assumptions:

We have see annoted slides

E [S] = E [E [S|N]] = E [NE [X ]] = E [N]p1 ,

Var (S) = E [Var (S|N)] + Var (E [S|N])

Moment generating function of S

It is possible to get MS (t) as a function of MX (t) and MN (t):

Assume that N is geometric with probability of success p:

where 0 < q < 1 and p = 1 − q. We have then

and thus replace by ln(Mx(t))

It is possible to get a fairly general expression for the df of S by

where P ∗n (x) is the n-th convolution of P with the definition

where p ∗0 (0) = 1 (and thus 0 anywhere else).

Example A 12.2.2 (discrete X , tabular approach)

Assume p(1) = 0.5, p(2) = 0.4, p(3) = 0.1.

The convolution p ∗(n+1) (x) = y p ∗n (x − y )p(y )

If X is mixed, S will generally be mixed:

There are several popular options for the distribution of N:

E [S] = λE [X ] Var (S) = λE [X 2 ] MS (t) = e λ(MX (t)−1)

Exposure Missed first 10 mins

It makes no sense to talk about frequency in an insurance

same as a sum of Bernoulli

fixed volume v > 0

Theorem 2.12: Aggregation property (also A Th 12.4.1)

Aggregation property- remarks

The aggregation property implies that we can use a

Example (A Example 12.4.1)

Suppose that N1 , N2 , · · · , Nm are independent random variables.

x1 N1 + · · · + xm Nm ? ~ Comp( lambda, size dist)

Ni ~ Poi(lambda) Ex. what is dist of 2*N1 + 4*N2 + 6*N3

Thinning of the Poisson point process

Theorem (Thinning of the Poisson point process)

Large claims separation-an example of thinning of the

Large claims separation (Example 2.16)

The large claims separation example (Example 2.16)

A theorem and the sparse vector algorithm Size

Ni ~ Poi(lambda) Ex. what is dist of 2N1 + 4N2 + 6*N3