Академический Документы
Профессиональный Документы
Культура Документы
Subject CT6
CMP Upgrade 2009/2010
CMP Upgrade
This CMP Upgrade lists all significant changes to the Core Reading and the ActEd
material since last year so that you can manually amend your 2009 study material to
make it suitable for study for the 2010 exams. It includes replacement pages and
additional pages where appropriate. Alternatively, you can buy a full replacement set of
up-to-date Course Notes at a significantly reduced price if you have previously bought
the full price Course Notes in this subject. Please see our 2010 Student Brochure for
more details.
Chapter 6 (new)
Page 1
Page 1
Syllabus objective (iii) 1 has been removed from Subject CT6. Syllabus objectives (iii)
2 to (iii) 7 have been renumbered accordingly.
Page 1
Syllabus objective (iv) has been enhanced. Syllabus objectives (iv) 2 to (iv) 4 are new.
Syllabus objectives (iv) 2 has been renumbered to (iv) 5 accordingly. Syllabus
objectives (iv) 6 to (iv) 8 are new. Overall, syllabus objective (iv) now reads:
(iv) Explain the concept of ruin for a risk model. Calculate the adjustment
coefficient and state Lundberg’s inequality. Describe the effect on the
probability of ruin of changing parameter values and of simple reinsurance
arrangements.
1. Explain what is meant by the aggregate claim process and the cash-flow
process for a risk.
3. Define a compound Poisson process and derive the moments and moment
generating function for such a process.
7. Describe the effect on the probability of ruin, in both finite and infinite
time, of changing parameter values.
Chapter 9 (removed)
Page 1
Syllabus objective (vi) has been removed from Subject CT6. Syllabus objectives (vii)
onwards have been renumbered accordingly.
Page 1
Page 1
Chapter 12
Page 1
Chapter 13
Page 1
Chapter 14
Page 1
Chapter 5
A few references to Core Reading have been amended as a result of the new Core
Reading in the new Chapter 6.
Page 3
Page 8
Page 20
After the first Core Reading paragraph in Section 2.5, a new Core Reading sentence has
been added:
The reason for doing this is that some of the observations will be helpful when
empirical Bayes credibility theory is considered in Chapter 6.
Chapter 6 (new)
All of the Core Reading in this chapter is new. A replacement chapter is provided.
Pages 2 to 4
This section gives an overview of the essential features that distinguish the main
types of general insurance products and so helps to identify the crucial aspects
that influence the nature and extent of the risk to be covered by the insurance.
The range of general insurance products is very wide and continually changing
and therefore it is difficult to set out the features of all types of product.
The sections below provide a general indication and examples of the knowledge
that examiners would expect a candidate to have in relation to the features of the
major types of general insurance product. These topics are covered in more
detail in Subject CA1 – Core Applications Concepts.
Pages 5 to 21
All the Core Reading in Section 1.2 and Section 2 has been moved from Subject CT6 to
Subject CA1. Sections 3 onwards have been renumbered accordingly. A new Core
Reading paragraph has been added to reflect this change:
The range of general insurance products is very wide and continually changing.
They are covered in Subject CA1 – Actuarial Risk Management.
Page 25
At the bottom of this page, a new Core Reading paragraph has been added. It reads:
There are a number of additional elements included when setting the premium to
be charged to policyholders, including the policyholders’ previous claims record
and these are covered in Subject CA1 – Actuarial Risk Management. The
allowance for policyholders’ claim experience could be based on claim
frequency or total claim amounts. This is beyond the scope of CT6 but a specific
example of how policyholder experience can be allowed form on a claim
frequency basis, is covered in Subject CT4 – Models.
The Core Reading in this chapter has been enhanced considerably. A replacement
chapter is provided.
Chapter 9
This chapter has been removed because the Core Reading in it is no longer covered in
Subject CT6.
Chapter 5
A new exam-style question has been added to the end of the chapter. This is provided
in the replacement pages.
Chapter 6 (new)
Section 1.2 and Section 2 has been moved from Subject CT6 to Subject CA1.
Sections 3 onwards have been renumbered accordingly.
Page 2
The introduction has been trimmed to reflect the fact that the chapter has become
smaller.
This chapter has been enhanced considerably due to additional Core Reading. A
replacement chapter is provided.
A new exam-style question has been added to the end of the chapter. This is provided
in any case as this chapter is in the replacement pages.
Chapter 9
This chapter has been removed because the Core Reading in it is no longer covered in
Subject CT6.
Chapters 10 and 11
These two chapters have been swapped round. This is purely to reflect their relative
levels of difficulty when it comes to running 3-Day CT6 tutorials.
A new exam-style question has been added to the end of the chapter. This is provided
in the replacement pages.
Chapter 14
A new exam-style question has been added to the end of the chapter. This is provided
in the replacement pages.
● bookwork questions
● developmental questions or
● exam-style questions.
This change is not sufficient to provide replacement pages but here a summary of the
questions and what type they each are is provided with the replacement pages.
There was a minor typo in the calculations here. The first two paragraphs should now
read:
First we need to obtain random variates from V ~ U (-1,1) . Since the distribution
function is F (v) = 12 (v + 1) , we have v = 2u - 1 using the inverse transform method.
There was a minor typo in the calculations here. The second covariance term should be
7 13
and not . The first two paragraphs should now read:
12 12
cov ( Zt , X t ) = var( Zt , Zt ) = 1
5 1 7
cov ( Zt -1, X t ) = cov( X t -1 , Zt -1 ) - var( Zt -1 ) =
6 4 12
Questions 2.6 to 2.10 are new and attached as replacement pages at the end.
Questions 2.6 is now Question 2.11. Questions 2.10 to 2.24 are now Questions 2.11 to
2.26.
Question 2.25 has been removed. Questions 2.26 to 2.32 have been moved to Q&A
Bank Part 3. They are now Questions 3.1 to 3.7
Questions 3.1 to 3.7 have been replaced as they related to material that is no longer
tested in Subject CT6. They have been replaced by Questions 2.26 to 2.32 from Q&A
Bank Part 2.
Questions 3.8 to 3.12 are new and attached as replacement pages at the end.
Solution X1.5
This solution has been updated to be more efficient, since we observe strategy d3 is
dominated. Replacement pages have been provided.
The notation here has changed. The confusing d A , d B and dC have been replaced
with d1 , d 2 and d 3 .
The minimax solution here should be d 2 and not d3 , d B or dC . This solution now
reads:
The minimax criterion will select the risk function with the smallest maximum loss.
This is d 2 .
Assignment X2
Question X2.1
This has been replaced with a new question. The new question is attached with the
replacement pages.
Question X2.2
This has been replaced with Question X2.6. Questions X2.7 to X2.10 have been
renumbered X2.6 to X2.9
More explanation has been added to this solution after the first paragraph. The new
paragraph reads:
It is also important to note here that S E and SC are NOT independent because the
expenses depend on the claims. Any approach using var [ S ] = var [ SC ] + var [ S E ] is
doomed to fail.
Assignment X3
Question X3.1
This has been removed. Questions X3.2 and X3.3 are now Questions X3.1 and X3.2.
Question X3.4
This has been removed. Questions X3.5 to X3.8 are now Questions X3.3 to X3.6.
Question X3.9
This has been removed. Questions X3.7 and X3.8 are new and replacement pages have
been provided.
Assignment X4
Question X4.1
This question has been rephrased to be more specific as to what is required. It now
reads:
X t = 5 + Z t 1 + ( X t -1 - 5) 2
Describe how X t fluctuates and state the long-term mean of the process.
• Series Y Assignments
• Mock Exam 2009 and Mock Exam 2010
• ASET (ActEd Solutions with Exam Technique) and mini-ASET
• Revision Notes
• Sound Revision
• Flashcards
• Smart Revise.
For further details on ActEd’s study materials, please refer to the 2010 Student
Brochure, which is available from the ActEd website at www.ActEd.co.uk.
5.2 Tutorials
For further details on ActEd’s tutorials, please refer to our latest Tuition Bulletin, which
is available from the ActEd website at www.ActEd.co.uk.
5.3 Marking
You can have your attempts at any of our assignments or mock exams marked by
ActEd. When marking your scripts, we aim to provide specific advice to improve your
chances of success in the exam and to return your scripts as quickly as possible.
For further details on ActEd’s marking services, please refer to the 2010 Student
Brochure, which is available from the ActEd website at www.ActEd.co.uk.
If you have any comments on this course please send them by email to CT6@bpp.com
or by fax to 01235 550085.
The first 4 parts of the Q&A Bank have been split into bookwork questions,
developmental questions and exam-style questions. Part 5 consists only of exam-style
questions. Please note that the following grid refers to the numbering in the 2009 Q&A
Bank.
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.
3 Exam-style questions
Exam-style question 1
q a e -q / m
prior ( m ) = , m >0
m a +1G (a )
q
You are given that the mean of this distribution is .
a -1
(i) Write down the likelihood function for m , based on a random sample of values
x1,… , xn from an exponential distribution. [1]
(ii) Find the form of the posterior distribution for m , and hence show that an
expression for the Bayesian estimate for m under squared error loss is:
q + Â xi
mˆ = [4]
n + a -1
(iii) Show that the Bayesian estimate for m can be written in the form of a credibility
estimate, and write down a formula for the credibility factor. [3]
(iv) The statistician now decides that he will use a prior distribution of this form with
parameters q = 40 and a = 1.5 . His sample data have statistics n = 100 ,
 xi = 9,826 , and  xi2 = 1, 200, 000 . Find the posterior estimate for m , and
the value of the credibility factor in this case. [3]
Exam-style question 2
The annual number of claims from a particular risk has a Poisson distribution with
mean m . The prior distribution for m has a gamma distribution with a = 2 and l = 5 .
(i) Show that the posterior distribution is gamma and determine its parameters. [3]
8
(ii) Given that n = 8 and  xi = 5 determine the Bayesian estimate for m under:
i =1
1 - x1 / m 1 - xn / m e - Â xi / m
L( m ) = e ¥ ¥ e =
m m mn
e - Â xi / m e ( Â i)
- q+ x /m
e -q / m
Post ( m ) μ ¥ =
m a +1 mn m n +a +1
We see that this has the same form as the prior distribution, but with different
parameters. So we have the same distribution as before, but with parameters:
q * = q + Â xi and: a* = n +a
We can now use the formula for the mean of the distribution given in the question:
q* q + Â xi
E(m) = =
a * -1 n + a - 1
We now have to show that this can be written in the form of a credibility estimate.
Splitting the Bayesian estimate into two parts:
q +  xi q  xi
mˆ = = +
n + a -1 n + a -1 n + a -1
=
q
¥
a -1
+
 xi ¥ n
a -1 n + a -1 n n + a -1
We can see from this that m̂ can be written as a weighted average of the prior mean
(
q
) and the maximum likelihood estimator for m (the sample mean
 xi ).
So the
a -1 n
Bayesian estimate m̂ can be written in the form of a credibility estimate, and the
credibility factor is:
n
Z=
n + a -1
We now have:
q + Â xi 40 + 9,826
mˆ = = = 98.1692
n + a - 1 100 + 1.5 - 1
n 100
Z= = = 0.9950
n + a -1 100 + 1.5 - 1
(v) Comments
We see that the value of Z is very close to 1. So our credibility estimate comes out to
be very close to the sample mean (98.26), and takes little account of the prior mean
(80). This is because n is nuch bigger than a .
We would use a prior like this if we are not very sure initially about the true value of
m . We have chosen a prior distribution with a large variance, to reflect our high initial
degree of uncertainty about m .
m x1 m xn
L( m ) = e -m
¥…¥ e - m μ m  xi e - n m
x1 ! xn !
52
f (m) = m e -5 m μ m e -5 m
G (2)
= m e -5 m ¥ m  xi e - nm
The Bayesian estimate for m under squared-error loss is the mean of the posterior:
7
mˆ = 0.538
13
The Bayesian estimate under “all-or-nothing” loss is given by the mode of the posterior.
To find the mode, we need to differentiate the PDF (or equivalently differentiate the log
of the PDF) and equate it to zero. The PDF of a gamma (7,13) is given by:
f ( m ) = constant × m 6e -13 m
ln f ( m ) = ln k + 6ln m - 13 m
d 6
fi ln f ( m ) = - 13 = 0
dm m
6
fi mˆ = 0.462
13
d
f ( m ) = 6 m 5e -13 m - 13m 6e -13 m
dm
= m 5e -13 m (6 - 13m )
a -1
In general the mode of the gamma distribution is (provided that a > 1 ).
l
The Bayesian estimate under absolute error loss is given by the median, M , of the
posterior. Hence P( m < M ) = ½ . Using the gamma-chi squared relationship on page
12 of the Tables, we get:
2
P ( m < M ) = P(2 ¥ 13m < 2 ¥ 13M ) = P ( c14 < 26M ) = 0.5
Using the chi-square percentage points on Page 169 of the Tables, we get:
13.34
26M = 13.34 fi M = 0.513
26
Chapter 6
Emperical Bayes Credibility theory
Syllabus objectives
(v) Explain the fundamental concepts of Bayesian statistics and use these concepts
to calculate Bayesian estimators.
0 Introduction
In this chapter, we will study Empirical Bayes Credibility Theory (EBCT) Models 1
and 2.
The algebra for the EBCT models that we will study in this unit is fairly heavy going.
Try not to get too bogged down in the mathematics and don’t learn the formulae. (Most
of them are in the Tables.) Concentrate on appreciating the differences between the
models and being able to follow the numerical examples.
1.1 Introduction
In this chapter we will discuss two Empirical Bayes Credibility Theory (EBCT) models.
As with the Poisson/gamma and normal/normal models discussed in the previous
chapter, these models are used to estimate the “true” claim frequency or risk premium
based on the total claim amounts in successive periods. Model 1 gives equal weight to
each risk in each year. Model 2 is more sophisticated and takes into account the volume
of business written under each risk in each year.
The main similarities and differences between the Bayesian models and the EBCT
models are outlined below.
Risk parameter
Both approaches use an auxiliary risk parameter θ . With the Bayesian approach, the
quantity we wish to estimate is θ , whereas, for the EBCT models, it is m(θ ) , ie a
function of θ . Unlike the Bayesian approach, the EBCT models don’t assume any
specific statistical distribution for θ .
Both approaches assume that the conditional variables X j θ s are independent and
identically distributed. Unlike the pure Bayesian approach, the EBCT models don’t
assume any specific statistical distribution for X j θ . Instead, formulae are developed
just using the assumption that the mean m(θ ) and variance s2 (θ ) of X j θ can be
expressed as functions of θ . The values of m(θ ) and s2 (θ ) are then estimated by the
models.
Credibility formula
With both approaches, the resulting formula for estimating the claim frequency or risk
premium for a given risk can be expressed using a credibility formula, ie a linear
combination (= weighted average) of the average derived from past claims and an
overall average.
In this section the assumptions for Model 1 of EBCT will be set out. EBCT
Model 1 can be regarded as a generalisation of the normal/normal model. This
point will be considered in more detail later in this section.
The problem of interest is the estimation of the pure premium, or possibly the
claim frequency, for a risk. Let X 1, X 2 ,… denote the aggregate claims, or the
number of claims, in successive periods for this risk. A more precise statement
of the problem is that, having observed the values of X 1, X 2 ,…, X n , the expected
value of X n +1 needs to be estimated. From now on X 1, X 2 ,…, X n will be
denoted by X .
The risk parameter in the Poisson/gamma model is (the Poisson parameter) λ . The risk
parameter for the normal/normal model is θ .
The above assumptions and consequences were all either made or noted for the
normal/normal model of Bayesian credibility in Section 2.5 of the previous
chapter. See (2.17), (2.18), (2.19), (2.23) and (2.24) in the previous chapter.
Note that (1.1) corresponds to (2.17), (1.2) corresponds to (2.18) and (2.19), (1.3)
corresponds to (2.23), and (1.4) corresponds to (2.24).
m (q ) = E ( X j | q )
s 2 ( q ) = var ( X j | q )
Two things should be noticed about m ( q ) and s 2 ( q ) . The first is that since,
given q , the X j ’s are identically distributed, neither m ( q ) nor s 2 ( q ) depends
on j , as their notation suggests. The second is that since q is regarded as a
random variable, both m ( q ) and s 2 ( q ) are random variables.
If the value of q and the distribution of X j given q were known, the obvious
estimate of the expected aggregate claims, or the expected number of claims, in
any future year would be m ( q ) . Since it is assumed that q is not known, the
problem is to estimate:
m ( q ) given X (1.5)
Example
Solution
Here, the risk parameter q represents the “safety coefficient” of a particular driver. It is
assumed that each driver has a constant inherent level of safety, which could (in theory)
be measured by some means. The value of q for a particular driver influences the likely
number of claims that the driver will make. For example, a value of q = 1 may
correspond to a driver who is “100% safe” (ie never makes any claims), while a value of
q = 0 may correspond to a driver who is “0% safe” (ie makes a claim every time the car
is used). The distribution of the values of q , which will vary from one driver to the next,
is assumed to be determined by a definite (but unknown) probability distribution.
Question 6.1
Describe what each of the variables and functions in EBCT Model 1 would represent if
this model was used to study the scores of different golf players on different days, if the
scores for the players are considered to be dependent on their “handicaps”. (For those
of you who are not golfers, a player’s handicap is a measure of his or her golfing ability.
A golfer with a low handicap is a better player than a golfer with a high handicap. A
good golfer will achieve a lower score on a round of golf than a bad golfer.)
The similarities between EBCT Model 1 and the normal/normal model can be
summarised as follows.
(i) The role of θ is the same for both models: it characterises the underlying
distributions of the processes being modelled, eg the aggregate claim
distribution for each year of business. See (2.17) from the previous
chapter and (1.1).
( )
E È s 2 (q ) ˘ corresponds to s 12 ( = var X j | q ).
Î ˚
(iv) The risk parameter, q , is a real number for the normal/normal model but
could be a more general quantity for EBCT Model 1.
In the previous section the assumptions for Model 1 of EBCT were studied and
the similarities between this model and the normal/normal model were
emphasised. In this section a solution to the problem summarised in (1.5) will be
considered, that is, the estimation of m ( q ) given the data X . The derivation of
the credibility premium under this model is beyond the scope of the CT6
syllabus and is not covered here.
Question 6.2
Without looking back to the previous chapter, write down the general credibility
formula.
(1 - Z ) E [m ( q )] + Z X
where:
n
X = ∑Xj /n
j =1
and:
n
Z= (1.6)
n + E È s ( q ) ˘ var ÈÎ m (q )˘˚
2
Î ˚
The first, and most important, point to note about the solution is that it is in the
form of a credibility estimate. In other words, it is in the form of formula (1.1)
from the previous chapter with:
n
 X j / n playing the role of X.
j =1
The second point to note is the similarity between the solution above and the
solution in the normal/normal model, and, in particular, the similarity between
the formulae for the credibility factors, ie (2.16) from the previous chapter and
(1.6). Formula (1.6) is a straight generalisation of formula (2.16) from the
previous chapter, since E È s 2 ( q ) ˘ and var [ m ( q )] can be regarded as
Î ˚
generalisations of s 12 and s 22 , respectively.
The similarities between EBCT Model 1 and the normal/normal model lead to the
similarity in the resulting answers.
The final point to note is that the formula for the credibility estimate given by
(1.6) involves three parameters, E [m ( q )] , E È s 2 ( q ) ˘ and var [m ( q )] , which have
Î ˚
been treated so far as though they were known. EBCT makes no distributional
assumptions about the risk parameter θ , unlike the Bayesian approach to
credibility, but the values of these three parameters need to be known. It can be
noted that these three parameters relate to first and second order moments as
opposed to, say, higher order moments or some other quantities. This is due to
the derivation of the credibility premium and the detail behind this is beyond the
scope of this course. The way in which these parameters would be estimated is
discussed in the next section.
To do this a further assumption is needed: that there are available data from
some other risks similar, but not identical, to the original risk. This requires the
problem to be set up a little more generally, to make some extra assumptions
and to change the notation slightly. An important distinction between a pure
Bayes approach to credibility and EBCT is that no data are required to estimate
parameters in the former case but data are required to estimate the parameters
in the latter.
Table 1
Year
1 2 ... n
Risk 1 X 11 X 12 ... X 1n
Number 2 X 21 X 22 ... X 2n
. . . ... .
. . . ... .
N X N1 X N2 ... X Nn
The risk numbers could refer to risks from different companies, insurers, shops etc.
It is helpful to remember that small n represents the number of years while big N
represents the number of risks.
Each row of Table 1 represents observations for a different risk; the first row,
relating to Risk Number 1, is the set of observed values denoted X in
Section 1.2 and Section 1.3. (Notice that X 1, X 2 ,…, X n in Section 1.2 and Section
1.3 have now become X 11, X 12 ,…, X 1n in this section.)
In Section 1.2 two assumptions, (1.1) and (1.2), were made about the connection
between the observed values for the single risk then being considered. In this
section exactly the same assumptions for each of the N risks in the collective
are made. These assumptions are given as (1.7) and (1.8) below.
Notice that the risk parameter, which was denoted θ in Section 1.2 and
Section 1.3, is now denoted θ i , and that an implication of these two
assumptions is that the risk parameters for different risks have different values.
(However, as in Section 1.2 and Section 1.3, the risk parameter for a given risk
does not change value from year to year.) These two assumptions show
something about the relationships within each row of Table 1 but they do not
show anything about the relationship between different rows, ie between
different risks in the collective. The assumption that shows something about the
relationship between different risks in the collective is the following.
This assumption shows that the rows of Table 1 are independent of each other.
The risk parameters θ 1,θ 2 ,…,θ N are independent and identically distributed.
(1.11)
The connection between the different risks, ie rows of the table, is a result of the
assumption that the risk parameters, θ 1,θ 2 ,… ,θ N , are identically distributed.
Intuitively, this means that if, by some means, the values of θ 2 ,θ 3 ,…,θ N were
known, then something about the common distribution of the θ i ’s would be
known and hence something about θ 1 , or, at least, about the distribution it
comes from, would be known.
The functions m ( ) and s 2 ( ) were introduced in Section 1.2. Keeping the same
definitions for these functions and applying them to all the risks in the collective:
(
m ( q i ) = E X ij | q i )
(
s 2 (q i ) = var X ij | q i )
(
Notice that, as in Section 1.2, neither E X ij | q i ) nor var ( X ij | q i ) depends on j
since, given q i , the random variables X i 1 , X i 2 ,… X in are identically distributed.
Notice also that, since q 1 ,q 2 ,… ,q N are identically distributed, E [m ( q i )] ,
n
 X ij / n by X i
j =1
and:
N Ê N n ˆ
 Xi / N Á=  Â
ÁË i = 1 j = 1
X ij /(Nn )˜ by X .
˜¯
i =1
Notice that what is now denoted X i was denoted X in Section 1.2 and
Section 1.3. It is important to recognise the difference between the meanings of
the symbol X in this section and in earlier sections.
X1 + XN
X i denotes the sample mean of the data from Risk Number i and X =
N
denotes the sample mean of the data from all of the risks.
This new notation will be used to reformulate the credibility estimate of the pure
premium, or the number of claims, for the coming year for Risk Number 1 in the
collective as:
(1 - Z ) E [m ( q )] + Z X 1
where:
n
X1 = Â X1j / n
j =1
and:
n
Z= (1.12)
n + E È s ( q ) ˘ var [m (q )]
2
Î ˚
It is important to realise that formula (1.12) is exactly the same as formula (1.6)
but in this case the notation used for the data from the risk itself is X 1 and X 1 j
rather than X and X j .
Each row of Table 1 corresponds to a fixed value of q . Bearing this and the
definitions of m ( q i ) and s 2 ( q i ) in mind, obvious estimators for m ( q i ) and
s 2 ( q i ) are:
n
Xi and (n - 1) -1 Â ( X ij - X i )2
j =1
respectively.
So we estimate m (q i ) by the sample mean of the values for Risk Number i and we
estimate s 2 (q i ) by the sample variance of the values for Risk Number i .
N n
 (n - 1)-1  ( X ij )
2
N -1 - Xi
i =1 j =1
N Ï n 2¸
Ô -1
Note that this is a “nested” sum, ie it is N -1 Â Ì(n - 1) Â X ij - X i ( ) Ô
˝ . It is not two
i =1 Ô
Ó j =1 Ô˛
sums multiplied together.
N
(N - 1)-1 Â ( X i -X )2
i =1
N N n
Â( ) Â ( X ij - X i )
2 2
(N - 1) -1
Xi - X - (Nn )-1 Â (n - 1) -1
i =1 i =1 j =1
This last observation can simplify the working in numerical calculations. Just make
sure that you use 1/ n , where n is the number of data values for each risk, rather than
1/ N in your calculation.
Parameter Estimator
E [ m ( q )] X (1.13)
N n
E È s 2 (q )˘
Î ˚
N -1 Â (n - 1)-1 Â ( X ij - Xi )2 (1.14)
i =1 j =1
N N n
var [m ( q )] (N - 1)-1 Â ( X i -X )2 - (Nn )
-1
 (n - 1)-1  ( X ij - Xi )2
i =1 i =1 j =1
(1.15)
They are on Page 29. We will do some numerical examples shortly to help you see
what is going on.
It could be that the second term works out to be bigger than the first.
Now consider formula (1.6) for the credibility factor for EBCT Model 1. The
different “ingredients” of this formula have the following definitions or
interpretations:
var [m ( q )] is the variability of the average data values for different risks, ie the
variability of the row means in Table 1.
In other words, we include the data for the risk we’re interested in in the calculation of
var [m(q )] and E ÈÎ s 2 (q ) ˘˚ .
There is one further comment to be made about the assumptions for this model.
Assumptions have been made concerning the identical distribution of the X ij ’s
both from different risks and from the same risk. If the X ij ’s come from different
risks it has been assumed that they are (unconditionally) identically distributed
(see (1.10)). If the X ij ’s come from the same risk, i , it has been assumed they
are unconditionally identically distributed (see (1.3)) and conditionally, given q i ,
identically distributed (see (1.2) and (1.8)). A consequence of these assumptions
(
is that neither E X ij | q i ) (
nor var X ij | q i ) depends on j . This consequence
was the only step in the derivation of the credibility estimate (1.12) that required
assumptions about the identical distribution of the X ij ’s. In fact, assumptions
(1.8) and (1.9) (and the corresponding assumption in Section 1.2, (1.2)) could
have been replaced by the following assumptions and it would still have been
possible to derive the same credibility estimate.
(
For each risk, i , i = 1, 2, … , N , neither E X ij | q i ) (
nor var X ij | q i ) depends on
j. (1.18)
Assumptions (1.16) and (1.17) are weakened versions of (1.8) and (1.9).
Assumption (1.18) now has to be included as a separate assumption since it is
not a consequence of (1.16) and (1.17).
Example
The table below shows the aggregate claim amounts (in £m) for an international insurer’s
fire portfolio for a 5-year period, together with some summary statistics. Fill in the
missing entries and calculate E [m(q )] and E ÈÎ s 2 (q ) ˘˚ using EBCT Model 1.
1 5
Year (j) Xi ∑
4 j =1
( X ij − X i ) 2
Country (i) 1 2 3 4 5
1 48 53 42 50 59 50.4 39.3
2 64 71 64 73 70 68.4 17.3
3 85 54 76 65 90 74.0 215.5
4 44 52 69 55 71 ? ?
Solution
X i is the average claims for risk i for the 5-year period. So the missing entry is:
1 5
( ) 1⎡
2
∑ X ij − X i
4 j =1
=
4⎣
(44 − 58.2) 2 + + (71 − 58.2) 2 ⎤ = 132.7
⎦
1 4
X= ∑ X i = (50.4 + 68.4 + 74.0 + 58.2) / 4 = 62.75
4 i =1
E [m(q )] ª X = 62.75
Also:
1 4 1 5
( )
2
E Î s (q ) ˚ ª Â Â X ij - X i
È 2 ˘
4 i =1 4 j =1
Question 6.3
Example
Find the credibility factor for the example above and hence calculate the EBCT premium
for each of the countries for the coming year.
Solution
We can use the estimates we have calculated to find the credibility factor:
n 5
Z= = = 0.8169
n + E ÈÎ s 2 (q ) ˘˚ var [m(q )] 5 + 101.2
90.33
Because we have the same number of time periods for each country, the credibility factor
is the same for each.
We can then use the basic credibility formula P = Z X i + (1 - Z ) E [m(q )] to find the
EBCT premiums for each country:
Question 6.4
(c) None of the random variables or parameters in the model are assumed to have a
normal distribution.
2.1 Introduction
In keeping with the terminology in Section 1, the model in this section will be
referred to as EBCT Model 2.
In Model 1 these quantities were called X j . You will see shortly why a different letter
has been used.
It is assumed that the values of Y1 ,Y2 ,… ,Yn have already been observed and the
expected value of Yn + 1 needs to be estimated.
So far the problem looks exactly like the problem in Section 1. The important
difference between EBCT Model 1 and EBCT Model 2 is that Model 2 involves an
extra parameter known as the risk volume, P j . Intuitively, the value of P j
measures the “amount of business” in year j . For example, P j might represent
the premium income for the risk in year j or the number of separate policies
comprising the risk in year j . An important point to note is that the value of
Pn + 1 at the start of year n + 1 is assumed to be known.
X j = Y j / Pj j = 1, 2, …
These X j ’s, which have been adjusted for the risk volumes, correspond more closely to
the X j ’s in Model 1. In Model 1, we effectively assumed that Pj was always equal to
1, ie the volume of business was the same for each risk group.
( )
E X j | q does not depend on j . (2.3)
( )
P j var X j | q does not depend on j . (2.4)
As in previous sections, q is known as the risk parameter for the risk, and, as for
EBCT Model 1, it could be just a single real valued number or a more general
quantity such as a vector of real valued numbers. Assumption (2.1) is the
standard assumption for all credibility models considered here. See
assumptions (2.2) and (2.9) from the previous chapter, and assumption (1.1).
Assumption (2.2) corresponds to Assumption (1.2) for EBCT Model 1, but notice
that (2.2) is slightly weaker than (1.2). Assumption (2.2) does not require the
X j ’s to be conditionally (given θ ) identically distributed, but only to be
conditionally independent. There is no assumption in EBCT Model 2 that the
X j ’s are unconditionally, or conditionally given θ , identically distributed.
Comparing the above assumptions with assumptions (1.16), (1.17) and (1.18) for
EBCT Model 1, if all the P j ’s are equal to 1, then EBCT Model 2 is exactly the
same as EBCT Model 1.
(
m (q ) = E X j | q )
(
s 2 ( q ) = P j var X j | q )
The definition of m ( q ) corresponds exactly to the definition for EBCT Model 1 in
Section 1 but the definition of s 2 ( q ) is slightly different.
To gain a little more insight into assumptions (2.3) and (2.4), consider the
following example. Suppose the risk being considered is made up of a different
number of independent policies each year and that the number of policies in year
j is P j . Suppose also that the aggregate claims in a single year from a single
Then:
( )
E Y j = Pj m (q )
( )
var Y j = P j s 2 ( q )
( )
E X j = m (q )
( )
P j var X j = s 2 ( q )
The problem was stated rather loosely in the previous section as the estimation
of the expected value of Yn + 1 , given the values of Y1 ,Y2 ,… ,Yn . We can now be
rather more precise about this. The quantity to be estimated is the mean (given
q ) of Yn + 1 . This is given by Pn + 1 m ( q ) . Since in the model Pn + 1 is known at
the start of year n + 1 , the problem is to estimate m ( q ) . The data available are
the values of each Y j and its corresponding P j for j = 1, 2, … , n . As for Model 2
the full derivation of the credibility premium under this model is beyond the
scope of the CT6 syllabus and is not covered here.
n
E [m ( q )] E È s 2 ( q ) ˘ var [m (q )] +
Î ˚ Â Yj
j =1
n
 Pj + E È s 2 ( q ) ˘ var [m (q )]
Î ˚
j =1
Z X + (1 - Z ) E [m ( q )] (2.5)
where:
n n n n
X =  Pj X j  Pj =  Yj  Pj
j =1 j =1 j =1 j =1
and:
n
 Pj
j =1
Z=
n
 Pj + E È s 2 ( q ) ˘ var [m (q )]
Î ˚
j =1
This demonstrates the similarities to and the differences from the Model 1 result.
Question 6.5
Note carefully that the above result tells us how to estimate the value of X for the
coming year. If we want to estimate Yn +1 , the aggregate claim amount for the coming
year, we have to multiply our estimate of X by Pn +1 , the risk volume for the
corresponding year.
(i) If all the Pk ’s were equal to 1, the solution given by (2.5) is exactly the
same as the solution given by (1.6). This is as it should be since if all the
Pk ’s are equal to 1 then EBCT Model 2 is exactly the same as EBCT Model
1.
(ii) As for EBCT Model 1, the solution given by (2.5) involves three
parameters, E [m ( q )] , var [m ( q )] and E È s 2 ( q ) ˘ . The way in which these
Î ˚
parameters are estimated is explained in the next section.
E È s 2 ( q ) ˘ for EBCT Model 2 follows exactly the same steps as the procedure for
Î ˚
EBCT Model 1, which was studied in Section 1.4.
As before, we are now moving to a two-dimensional data set with rows representing
different risks.
It is now assumed that the risk that is of interest is one of a collective of N risks
and that there exist data in the form given in Section 2.2 for each of these N
risks for each of the past n years. These data consist of values for the
aggregate claims, or the number of claims, and the corresponding risk volumes.
Let Yij be a random variable denoting the aggregate claims, or the number of
claims, for risk number i in year j , j = 1, 2, …, n , i = 1, 2, … , N , and let Pij be the
corresponding risk volume.
X ij = Yij / Pij
The data are summarised in the following table, which corresponds to Table 1 for
EBCT Model 1:
Table 2
Year
1 2 … n
1 Y11 , P11 Y12 , P12 … Y1n , P1n
For simplicity it is assumed, as was done in Section 1.4, that the risk that is of
particular interest is Risk Number 1 in this collective. This means that what were
denoted Y j , P j and X j in Section 2.2 and Section 2.3 are now denoted Y1 j , P1 j
and X 1 j respectively in this section. The problem is to estimate the expected
value of X 1, n + 1 and the solution to this problem has already been given by
formulae (2.5), remembering that X j and P j in (2.5) are now denoted X 1 j and
P1 j . The purpose of the data from the other risks in the collective is purely to
These other risks in the collective satisfy assumptions exactly the same as
assumptions (2.1), (2.2), (2.3) and (2.4) for Risk Number 1. These assumptions
are as follows.
(
there exists a function m ( ) such that m ( q i ) = E X ij | q i ) (2.8)
(
there exists a function s ( ) such that s 2 ( q i ) = Pij var X ij | q i .) (2.9)
These four assumptions show that each risk in the collective satisfies the same
assumptions as the particular risk that is of interest. The following two
assumptions, which correspond to (1.9), (1.10) and (1.11) in Section 1.4, show the
connection between different risks in the collective.
n
 Pij by Pi (2.12)
j =1
N
 Pi by P (2.13)
i =1
N
(Nn - 1) Â Pi (1 - Pi / P )
-1
by P * (2.14)
i =1
n
 Pij X ij / Pi by X i (2.15)
j =1
N N n
 Pi X i / P =   Pij X ij / P by X (2.16)
i =1 i =1 j =1
This notation is all given in the Tables (Page 30). Notice that what was denoted
X in formula (2.5) is now denoted X 1 and that X now has a different definition.
(This was exactly what happened in Section 1.4.) Notice also that X i and X are
weighted averages of the X ij ’s, the weights being the risk volumes Pij .
With this new notation the credibility estimate of the pure premium, or number of
claims, per unit of risk volume for the coming year for Risk Number 1 in the
collective originally given by formula (2.5) can be reformulated as:
Z1 X 1 + (1 - Z ) E [m ( q )] (2.17)
where:
n n
X1 =  P1 j X 1 j  P1 j
j =1 j =1
and:
n
 P1 j
j =1
Z1 =
n
 P1 j + E È s 2 ( q ) ˘ var [m ( q )]
Î ˚
j =1
Note that this credibility factor is specific to Risk 1, so is denoted by Z1 . For the other
risks there will be different values of Z .
It is important to realise that this is exactly the same as formula (2.5) but is
written in the notation of this section rather than that of the previous section.
Parameter Estimator
E [ m ( q )] X (2.18)
N n
E È s 2 ( q )˘
Î ˚
N -1 Â (n - 1)-1 Â Pij ( X ij - Xi )2 (2.19)
i =1 j =1
È N n
  ( )
-1 2
var [m ( q )] P * -1 Í(Nn - 1) Pij X ij - X (2.20)
Í i =1 j =1
Î
2˘
N n
-N -1 Â (n - 1)-1 Â Pij ( X ij - Xi ) ˙
˙
i =1 j =1 ˚
(i) They are in exactly the same form as the estimators for EBCT Model 1.
Look back at formulae (1.13), (1.14) and (1.15). In particular, if all the Pij ’s
were equal to 1, then the two sets of estimators would be identical.
(ii) It can happen in practice that formula (2.20) gives a negative value even
though var [m ( q )] must be non-negative. In such a situation the estimate
of var [m ( q )] is taken to be zero. A similar point about the variance
estimates of EBCT Model 1 was discussed in Section 1.4.
(iii) Formulae (2.18), (2.19) and (2.20) are all given in the Tables.
(iv) The proofs that (2.18), (2.19) and (2.20) are unbiased are beyond the scope
of the syllabus.
Example
The table belows show the volumes of business for each country for the insurer in the
example on Page Error! Reference source not found.17.
Volume of business (P ij)
Country (i) Year (j)
1 2 3 4 5 Current
year
1 12 15 13 16 10 20
2 20 14 22 15 30 25
3 5 8 6 12 4 10
4 22 35 30 16 10 12
Solution
The original data for the total claims is Yij and the table above gives values of Pij .
The claims per unit volume X ij = Yij / Pij are shown in the table below.
Country
Pi Pi (1 − Pi / P )
(i)
1 66 52.17
2 101 68.62
3 35 31.11
4 113 72.46
P = 315 P* = 1181
.
n N n
Furthermore, X i and X can be calculated as: X i = ∑ Yij / Pi and X = ∑ ∑ Yij / P .
j =1 i =1 j =1
Country
(i)
Xi ∑ Pij ( X ij − X i ) 2 ∑ Pij ( X ij − X ) 2
1 3.818 57.13 58.94
2 3.386 111.59 147.71
3 10.571 1,237.82 2,756.56
4 2.575 267.73 492.04
X = 3.984
1 4 1 5
( )
2
E ÈÎ s 2 (q ) ˘˚ ª Â Â Pij X ij - X i
4 i =1 4 j =1
= 104.64
Also:
1 È 1 4 5 ˘
( )
2
var [m(q )] ª Í ÂÂ Pij X ij - X - 104.64 ˙
P* ÍÎ 4 ¥ 5 - 1 i =1 j =1 ˙˚
1
= [(58.94 + 147.71 + 2, 756.56 + 492.04) /19 - 104.64]
11.81
= 6.539
n
 P1 j
j =1 66
Z1 = = = 0.8048
n E ÈÎ s 2 (q ) ˘˚ 104.64
66 +
 P1 j + var [m(q )] 6.539
j =1
Since the volume for Country 1 for the coming year is 20 units, the EBCT premium is:
20 × 3851
. = 77.01
Question 6.6
Question 6.7
Explain the effect each of the following changes (acting in isolation) would have on the
value of the credibility factor for EBCT Model 2.
(a) E ÈÎ s 2 (q ) ˘˚ is increased.
The table below shows annual aggregate claim statistics for 3 risks over 4 years.
Annual aggregate claims for risk i , in year j , are denoted by X ij .
1 4 1 4
( )
2
Risk, i Xi = Â X ij
4 j =1
si2 = Â X ij - X i
3 j =1
1 2,517 4,121,280
2 7,814 7,299,175
3 2,920 3,814,001
(i) Calculate the value of the credibility factor for Empirical Bayes Model 1. [3]
(ii) Using the numbers calculated in (i) to illustrate your answer, describe the way in
which the data affect the value of the credibility factor. [4]
[Total 7]
Exam-style question
An actuary wishes to analyse the amounts paid by a group of insurers on their respective
portfolios of commercial property insurance policies using the models of Empirical
Bayes Credibility Theory.
The actuary obtains the following information about the amounts of claim payments
made and the number of policies sold for each of three different insurers. The data
obtained are as follows.
(i) Analyse the data using EBCT Model 1, and calculate the expected total claim
payment to be made by Insurer B in the coming year. [6]
(ii) Analyse the data using EBCT Model 2, and again calculate the expected payout
amount for Insurer B in the coming year, assuming that the expected number of
policies sold for the coming year for Insurer B is 4,800. You may use the
summary statistics given below, which have been calculated using the formulae
and notation given in the Tables, again working in millions of pounds.
Subscripts 1, 2 and 3 refer to Insurers A, B and C respectively. [8]
∑ P2 j ( X 2 j − X 2 ) 2 = 0.006103 ∑ P2 j ( X 2 j − X ) 2 = 0.336408
∑ P3 j ( X 3 j − X 3 ) 2 = 0.003979 ∑ P3 j ( X 3 j − X ) 2 = 0.292641
(iii) Comment on your results. [2]
[Total 16]
Chapter 6 Summary
Empirical Bayes credibility
This approach to credibility theory assumes that the claims for each risk are dependent
on an underlying risk parameter θ . However, no assumptions are made about the form
of the distribution of the claim amounts.
(
m(q ) = E X j | q ) (
s 2 (q ) = var X j | q )
Ê E ÈÎ s 2 (q ) ˘˚ ˆ
Credibility factor: Z = n Án + ˜
Á var [m(q )] ˜
Ë ¯
X j = Yj / Pj ( Pj constant)
(
m(q ) = E X j | q ) (
s 2 (q ) = Pj var X j | q )
n Ê n E ÈÎ s 2 (q ) ˘˚ ˆ
Credibility factor: Z = Â Pj Á Â Pj + ˜
Á j =1 var [m(q )] ˜
j =1 Ë ¯
Credibility premium: E (m(q )| X ) = Z X i + (1 - Z ) E [m(q )]
The formulae for estimating E [m(q )] , E ÈÎ s 2 (q ) ˘˚ and var [m(q )] for the EBCT models
are given in the Tables (Pages 29 & 30).
Chapter 6 Solutions
Solution 6.1
Here, the risk parameter θ represents the handicap for a particular player. The
distribution of the values of θ , which will vary from one player to the next, are assumed
to be determined by a definite (but unknown) probability distribution. The shape of this
distribution could be estimated by finding out the handicaps of a random sample of
players. (In order for the EBCT model to apply here, we’ll have to assume that players
are secretive about their handicaps, so that the actual handicap for a given player isn’t
known with certainty.)
You may find it helpful to keep this example in the back of your mind to help you
understand how the EBCT models work.
Solution 6.2
Z X + (1 - Z ) m
Solution 6.3
We need to find:
1 4 1 4 1 5
( )
2
var [m(q )] ª Â ( X i - X ) -
2
  X ij - X i
3 i =1 4 ¥ 5 i =1 4 j =1
We’ve already worked out the second term (more or less) when we found E ÈÎ s 2 (q ) ˘˚ .
1 4 4
1 5
(
2 1 1
(
X i − X ) − × ∑ ∑ X ij − X i )
2
var [ m(θ )] ≈ ∑
3 i =1 5 4 i =1 4 j =1
1 1
= [(50.4 − 62.75) 2 + + (58.2 − 62.75) 2 ] − ×101.2
3 5
= 90.33
Solution 6.4
(a) False. q is just a risk parameter that reflects the likelihood of claims. The true
risk premium for a given risk is E ÈÎ m(q ) X ˘˚ .
(c) True. In fact, none of the quantities in the model are assumed to have any
specific type of distribution.
Solution 6.5
n n E ÈÎ s 2 (q ) ˘˚
 Pj  Pj X j
j =1 j =1 var [m(q )]
¥ n
+ ¥ E [m(q )]
n E ÈÎ s 2 (q ) ˘˚ n E ÈÎ s 2 (q ) ˘˚
 Pj + var [m(q )]  Pj  Pj + var [m(q )]
j =1
j =1 j =1
We can see that the second term here is equal to the first part of the original expression
n
for the best linear estimator. If we cancel a factor of ∑ Pj in the first term here, and
j =1
n n
remember that ∑ Pj X j = ∑ Yj , we see that the first term here is the same as the
j =1 j =1
second part of the expression for the best linear estimator. So the alternative form is
equivalent to what we got before.
Solution 6.6
The table below shows the figures for all four countries:
Solution 6.7
The formula for the credibility factor for EBCT Model 2 is:
n Ê n E ÈÎ s 2 (q ) ˘˚ ˆ
Z = Â Pj Á Â Pj + ˜
Á j =1 var [m(q )] ˜
j =1 Ë ¯
(a) E ÈÎ s 2 (q ) ˘˚ represents the average deviation from the expected value of the claims
for different risks. Increasing this would mean that claims tended to vary more
from their “true” values. So we should have less confidence in the accuracy of an
estimate based on past claims.
(b) var [m(q )] represents the variation in the average claim amount for different risks.
Reducing this would mean that the actual claims were less strongly influenced by
the true value of the risk parameter. So we would be less concerned about the
actual value of θ indicated by the past claims, and we should put more emphasis
on the overall mean.
(c) Changing the unit of currency should not affect the credibility factor.
Since the quantities E ÈÎ s 2 (q ) ˘˚ and var [m(q )] are measured in units of £ 2 , their
ratio is dimensionless. So changing the unit of currency to $ would not affect Z .
(d) The Pj ’s specify the relative weightings to be put on the claims for each year.
We would not expect a uniform increase applied to all the weightings to affect the
credibility factor.
(
Since the definition E ÈÎ s 2 (q ) ˘˚ = Pj var X j q ) includes a Pj factor, but
var [m(q )] doesn’t, the ratio E ÈÎ s 2 (q ) ˘˚ / var [m(q )] varies in proportion to the
Pj ’s. So any extra factor incorporated in the Pj ’s would cancel out, leaving Z
unchanged.
1
8, 695,309 - ¥ 5, 078,152 = 7, 425, 771
4
n
The credibility factor Z = is then estimated by:
n + E ÎÈ s 2 (q )˚˘ var ÈÎ m (q )˘˚
4
= 0.853997
4 + 5, 078,152 / 7, 425, 771
(ii) How the data affect the value of the credibility factor
= 3,567.1562
n 4
Z= = = 0.9912
E ÈÎ s 2 (q ) ˘˚ 4 + 127.1561
n+ 3,567.1562
var [m(q )]
So the EBCT premium for the coming year for Insurer B is:
We first need to use the data to calculate the statistics we need. Using the notation
given in the Tables, we have:
The same approach for the other insurers gives the values:
P2 = 19,353
P3 = 71,729
and:
= 2,891.5793
14.2 + 158
. + 22.7 + 19
X1 = = 0.089290
803
X 2 = 0.013791
X 3 = 0.007654
14.2 + 158
. + + 161 + 133
X= = 0.009660
91,885
Now using the summary statistics given in the question, we can find our estimates for
E [m(q )] , var [m(q )] and E ÈÎ s 2 (q ) ˘˚ :
E [m(q )] 0.009660
1 È1
var [m(q )] ¥ (5.106461 + 0.336408 + 0.292641)
2,891.5793 ÍÎ 11
1 ˘
- (0.014667 + 0.006103 + 0.003979) ˙
3¥3 ˚
= 0.0001794
1
E ÈÎ s 2 (q ) ˘˚ (0.014667 + 0.006103 + 0.003979) = 0.002750
3× 3
19,353
ZB = = 0.999208
0.002750
19,353 +
0.0001794
So the credibility premium per unit of risk volume for Insurer B is:
So assuming a risk volume in the coming year of 4,800, the risk premium for Insurer B
is £66.18m.
(iii) Comment
The two models give fairly similar results. The estimate in Model 2 will depend on the
prediction of risk volume for the coming year.
In both cases we have used a very high value for the credibility factor. So we are
effectively ignoring the data from the other insurers, and are basing our estimate almost
entirely on the data from Insurer B. This seems pretty sensible, given that both the
volume figures and the average claim amounts appear to be pretty variable between the
three different insurers. So we are tempted to ignore the data from Insurers A and C,
and focus on the information that we have for Insurer B.
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.
Chapter 9
Ruin theory
Syllabus objectives
(iv) Explain the concept of ruin for a risk model. Calculate the adjustment coefficient
and state Lundberg’s inequality. Describe the effect on the probability of ruin of
changing parameter values and of simple reinsurance arrangements.
1. Explain what is meant by the aggregate claim process and the cash-flow
process for a risk.
3. Define a compound Poisson process and derive the moments and moment
generating function for such a process.
7. Describe the effect on the probability of ruin, in both finite and infinite
time, of changing parameter values.
0 Introduction
In the previous two chapters we used the collective risk model to look at the aggregate
claims S arising during a fixed period of time. S was given by the equation
S = X1 + X 2 + + X N , where N denotes the number of claims arising during the period.
In this chapter we will extend this model by treating S (t ) as a function of time. This
gives us the equation S (t ) = X1 + X 2 + + X N (t ) , where N (t ) denotes the number of
claims occurring before time t. N (t ) is called a Poisson process and S (t ) is called a
compound Poisson process. We can use S (t ) to model claims received by an insurance
company and hence consider the probability that this insurance company is ruined. The
notation and the other basic concepts are covered in Section 1.
In Section 2 we will give formal definitions of both the Poisson process and the
compound Poisson process. You may have already met these in Subject CT4. We will
also introduce the concept of a premium security loading. Briefly, this is an additional
amount charged on an insurance premium to reduce the likelihood of an insurance
company becoming ruined.
Section 4 considers the effect of changing parameter values on the probability of ruin
for an insurance company. Finally, in Section 5 we will consider the impact on the
probability of ruin for an insurance company when reinsurance is introduced into the
equation.
1 Basic concepts
1.1 Notation
f (x)
lim =0
x Æ0 x
You can use this notation to simplify your working. For example, the function
g ( x) = 3 x + 0.5 x 2 + 0.004 x3 can be rewritten as g ( x ) = 3 x + o( x ) , since
0.5 x 2 + 0.004 x3
Æ 0 as x Æ 0 . Note that o( x) does not represent an actual number so
x
that co( x) ( c is a constant), -o( x) and o( x) are all equivalent.
Question 9.1
For the purposes of this course, you just need to be able to understand the notation.
However, if you wish to know more, then it’s covered in more detail in the Foundation
ActEd Course (FAC).
S (t ) the aggregate claims in the time interval [0, t], for all t ≥ 0.
{X i }i•=1 is a sequence of random variables. {N (t )}t ≥0 and {S(t )}t ≥0 are both
families of random variables, one for each time t ≥ 0; in other words {N (t )}t ≥0
and {S (t )}t ≥0 are stochastic processes.
You can think of a stochastic process as being a whole family of different random
variables. Consider a time line. On the line there are an infinite number of different
time intervals. For each interval of time, there is a random variable that corresponds to
the aggregate claim amount arising in that time interval. This is what we mean by a
stochastic process.
Of course, if you have previously studied Subject CT4, you should be familiar with
these ideas already.
N (t )
S (t ) = Â Xi
i =1
So we have just taken the idea of a compound distribution and generalised it to cover
different time periods.
The insurer of this portfolio will receive premiums from the policyholders. It is
convenient at this stage to assume, as will be assumed throughout this chapter,
that the premium income is received continuously and at a constant rate. Here is
some more notation:
so that the total premium income received in the time interval [0, t] is ct. It will
also be assumed that c is strictly positive.
Suppose that at time 0 the insurer has an amount of money set aside for this
portfolio. This amount of money is called the initial surplus and is denoted by U.
It will always be assumed that U ≥ 0. The insurer needs this initial surplus
because the future premium income on its own may not be sufficient to cover the
future claims. Here we are ignoring expenses. The insurer’s surplus at any
future time t (> 0) is a random variable since its value depends on the claims
experience up to time t. The insurer’s surplus at time t is denoted by U(t). The
following formula for U(t) can be written:
U (t ) = U + ct - S (t )
In words this formula says that the insurer’s surplus at time t is the initial
surplus plus the premium income up to time t minus the aggregate claims up to
time t. Notice that the initial surplus and the premium income are not random
variables since they are determined before the risk process starts. The above
formula is valid for t ≥ 0 with the understanding that U(0) is equal to U. For a
given value of t, U(t) is a random variable because S(t) is a random variable.
Hence {U (t )}t ≥0 is a stochastic process, which is known as the cash flow
process or surplus process.
Figure 1
Figure 1 shows one possible outcome of the surplus process. Claims occur at
times T1, T2, T3, T4 and T5 and at these times the surplus immediately falls by
the amount of the claim. Between claims the surplus increases at constant rate
c per unit time. The model being used for the insurer’s surplus incorporates
many simplifications, as will any model of a complex real-life operation. Some
important simplifications are that it is assumed that claims are settled as soon as
they occur and that no interest is earned on the insurer’s surplus. Despite its
simplicity this model can give an interesting insight into the mathematics of an
insurance operation.
We are also assuming that there are no expenses associated with the process (or,
equivalently, that S (t ) makes allowance for expense amounts as well as claim
amounts), and that the insurer cannot vary the premium rate c .
We are also ignoring the possibility of reinsurance. Simple forms of reinsurance will be
incorporated into the model later in this chapter.
It can be seen from Figure 1 that the insurer’s surplus falls below zero as a result
of the claim at time T3. Speaking loosely for the moment, when the surplus falls
below zero the insurer has run out of money and it is said that ruin has occurred.
In this simplified model, the insurer will want to keep the probability of this
event, that is, the probability of ruin, as small as possible, or at least below a
predetermined bound. Still speaking loosely, ruin can be thought of as meaning
insolvency, although determining whether or not an insurance company is
insolvent is, in practice, a very complex problem. Another way of looking at the
probability of ruin is to think of it as the probability that, at some future time, the
insurance company will need to provide more capital to finance this particular
portfolio.
ψ (U 2 , t ) ≤ ψ (U1, t ) (1.1)
ψ (U 2 ) ≤ ψ (U1 ) (1.2)
ψ (U , t 1 ) ≤ ψ (U , t 2 ) ≤ ψ (U ) (1.3)
lim y (U , t ) = y (U ) (1.4)
t Æ•
The larger the initial surplus, the less likely it is that ruin will occur either in a
finite time period, hence (1.1), or an unlimited time period, hence (1.2).
For a given initial surplus U, the longer the period considered when checking for
ruin, the more likely it is that ruin will occur, hence (1.3).
Question 9.2
You may be wondering whether it is possible to find numerical values for these ruin
probabilities. In some very simple cases it is. However, for most practical situations,
finding an exact value for the probability of ruin is impossible. In some cases there are
useful approximations to ψ (u) , even if calculation of an exact value is not possible.
The two probabilities of ruin considered so far have been continuous time
probabilities of ruin, so-called because they check for ruin in continuous time. In
practice it may be possible (or even desirable) to check for ruin only at discrete
intervals of time.
For a given interval of time, denoted h, the following two discrete time
probabilities of ruin are defined:
Figure 2
It can be seen from Figure 2 that in discrete time with h = 1, ruin does not occur
for this realisation of the surplus process before time 5, but ruin does occur (at
time 2½) in discrete time with h = ½.
Listed below are five relationships between different discrete time probabilities
of ruin for 0 ≤ U1 ≤ U2 and for 0 ≤ t1 ≤ t2 < ∞. Formulae (1.5), (1.6), (1.7) and (1.8)
are the discrete time versions of formulae (1.1), (1.2), (1.3) and (1.4) above and
their intuitive explanations are similar. The intuitive explanation of (1.9) comes
from Figure 2.
ψ h (U 2 , t ) ≤ ψ h (U1, t ) (1.5)
ψ h (U 2 ) ≤ ψ h (U1 ) (1.6)
ψ h (U , t 1 ) ≤ ψ h (U , t 2 ) ≤ ψ h (U ) (1.7)
lim y h (U , t ) = y h (U ) (1.8)
t Æ•
ψ h (U , t ) ≤ ψ (U , t ) (1.9)
Question 9.3
Intuitively, it is expected that the following two relationships are true since the
probability of ruin in continuous time could be approximated by the probability
of ruin in discrete time, with the same initial surplus, U, and time horizon, t,
provided ruin is checked for sufficiently often, ie provided h is sufficiently small.
lim y h (U , t ) = y (U , t ) (1.10)
h Æ0 +
lim y h (U ) = y (U ) (1.11)
h Æ0 +
Formulae (1.10) and (1.11) are true but the proofs are rather messy and will not
be given here.
2.1 Introduction
In this section some assumptions will be made about the claim number process,
lN(t )qt ≥ 0 , and the claim amounts, lX i q i∞= 1 . The claim number process will be
assumed to be a Poisson process, leading to a compound Poisson process
l q
S(t ) t ≥ 0 for aggregate claims. The assumptions made in this section will hold
for the remainder of this chapter.
The claim number process lN(t )qt ≥ 0 is defined to be a Poisson process with
parameter l if the following conditions are satisfied:
(ii) P (N (t + h ) = r | N (t ) = r ) = 1 - l h + o(h )
P (N (t + h ) = r + 1| N (t ) = r ) = l h + o(h ) (2.1)
P (N (t + h) > r + 1| N (t ) = r ) = o(h )
(iii) when s < t , the number of claims in the time interval (s , t ] is independent
of the number of claims up to time s . (2.2)
Condition (ii) states that in a very short time interval of length h , the only
possible numbers of claims are zero or one. Note that condition (ii) also implies
that the number of claims in a time interval of length h does not depend on
when that time interval starts.
Example
Solution
The events in this case are occurrences of claim events (ie accidents, fires, thefts etc) or
claims reported to the insurer. The parameter λ represents the average rate of
occurrence of claims (eg 50 per day), which we are assuming remains constant
throughout the year and at different times of day. The assumption that, in a sufficiently
short time interval, there can be at most one claim is satisfied if we assume that claim
events cannot lead to multiple claims (ie no motorway pile-ups etc).
The reason why a process satisfying conditions (i) to (iii) is called a Poisson
process is that for a fixed value of t , the random variable N (t ) has a Poisson
distribution with parameter l t . This is proved as follows:
( λt ) n
p n (t ) = exp{− λt } (2.3)
n!
This is the Poisson probability function taken from page 7 of the Tables, where the
parameter is l t .
For a fixed value of t > 0 and a small positive value of h , condition on the
number of claims at time t and write:
What we are saying here is that the event that there are n claims in the interval up to
time t + h can occur in one of two different ways. Either there were n − 1 claims in
total up to time t , followed by an additional claim in the interval (t , t + h) , or there
were n claims already in the interval up to time t , and no additional claim in the
subsequent time interval. The terms l h + o(h) and 1 - l h + o(h) come from condition
(ii) of the Poisson process.
Thus:
p n (t + h) − p n (t ) = λh[ p n − 1 (t ) − p n (t )] + o( h) (2.4)
Now divide (2.4) by h , and let h go to zero from above to get the differential-
difference equation:
d
pn (t ) = l [ pn -1(t ) - pn (t )] (2.5)
dt
df f ( x + h) - f ( x )
Remember that the definition of a derivative is = lim . Note also
dx hÆ0 h
o( h)
from the definition given at the start of the chapter that tends to zero.
h
d
p0 (t ) = - l p0 (t ) (2.6)
dt
∞
G( s, t ) = ∑ s n pn (t )
n=0
∞
d d
dt
G( s, t ) = ∑
sn
dt
p n (t )
n=0
• • •
d
 sn
dt
pn (t ) = l  s n pn -1(t ) - l  s n pn (t )
n =1 n =1 n =1
• • •
d
 s
dt
n
pn (t ) = l Â
s n pn -1(t ) - l Â
s n pn (t )
n =0 n =1 n =0
d
G( s, t ) = λsG( s, t ) − λG( s, t )
dt
Here we have used the definition of G ( s, t ) directly in the first and last sums. If we
•
take a factor of s out of the middle sum it leaves  s n-1 pn-1 (t ) , which is an equivalent
n =1
way of writing G ( s, t ) .
Or, equivalently:
1 d
G (s , t ) = λ (s − 1) (2.7)
G ( s, t ) dt
d f ¢( x )
Recall that ln f ( x) = .
dx f ( x)
Since the left hand side of (2.7) is the same as the derivative with respect to t of
logG(s, t), (2.7) can be integrated to find that:
logG( s, t ) = λt ( s − 1) + c( s )
As we are integrating with respect to t , the “constant” we add when integrating can be
a function of the other variable s (but not of t ).
Thus:
G( s, t ) = exp{λt ( s − 1)}
which is the probability generating function for the Poisson distribution with
parameter λt .
This section will show that the time to the first claim has an exponential distribution
with parameter l .
Let the random variable T1 denote the time of the first claim. Then, for a fixed
value of t , if no claims have occurred by time t , T1 > t . Hence:
And:
P (T1 ≤ t ) = 1 − exp{− λt }
Important information
The time to the first claim in a Poisson process has an exponential distribution with
parameter l .
This section will show that the time between claims has an exponential distribution with
parameter l .
For i = 2,3,… , let the random variable Ti denote the time between the (i - 1) th
and the i th claims. Then:
n n +1 n
P (Tn +1 > t | Â Ti = r ) = P ( Â Ti > t + r | Â Ti = r )
i =1 i =1 i =1
= P (N (t + r ) = n | N (r ) = n )
= P (N (t + r ) - N (r ) = 0 | N (r ) = n )
By condition (2.2) (ie using the independence of claim numbers in different time
periods):
P (N (t + r ) - N (r ) = 0ΩN (r ) = n ) = P (N (t + r ) - N (r ) = 0)
Finally:
P (N (t + r ) - N (r ) = 0) = P (N (t ) = 0) = exp{ - l t }
since the number of claims in a time interval of length r does not depend on
when that time interval starts (condition (2.1)). Thus inter-event times also have
an exponential distribution with parameter λ .
Important information
The time between claims in a Poisson process has an exponential distribution with
parameter l .
Note that the inter-event time is independent of the absolute time. In other words the
time until the next event has the same distribution, irrespective of the time since the last
event or the number of events that have already occurred. This is referred to as the
memoryless property of the exponential distribution.
Question 9.4
If reported claims follow a Poisson process with rate 5 per day (and the insurer has a
24 hour hotline), calculate:
(i) the probability that there will be fewer than 2 claims reported on a given day
(ii) the probability that another claim will be reported during the next hour.
In this section the Poisson process for the number of claims will be combined
with a claim amount distribution to give a compound Poisson process for the
aggregate claims.
( λt ) k
P [N (t ) = k ] = exp{− λt } for k = 0,1, 2,…
k!
With these assumptions the aggregate claims process, lS(t )qt ≥ 0 , is called a
compound Poisson process with Poisson parameter l . By comparing the
assumptions above with the assumptions in Section 2.3, it can be seen that the
l q
connection between the two is that if S(t ) t ≥ 0 is a compound Poisson process
with Poisson parameter l , then, for a fixed value of t (≥ 0) , S (t ) has a
compound Poisson distribution with Poisson parameter λt .
x
F ( x) = Ú f (t ) dt
-•
mk = E [ X ik ] for k = 1, 2, 3,…
Important information
mk = E[ X ik ] for k = 1, 2, 3,…
Whenever the common moment generating function of the X i s exists, its value
at the point r will be denoted by M X (r ) .
In case you have forgotten, the definition of a moment generating function is:
M X (r ) = E[e rX ]
Note that we are using r for the dummy variable to avoid confusion with time t .
MS (r )= exp {l t (M X (r ) - 1)}
S = X1 + X 2 + + XN
E[ S ] = l E[ X ] = l m1 var[ S ] = l E[ X 2 ] = l m2
M S (r ) = M N [ln M X (r )]
( )
M N (t ) (r ) = exp l t (er - 1) , so:
( )
M S (t ) (r ) = exp λt (eln M x ( r ) − 1) = exp ( λ t ( M X (r ) − 1) )
so that the insurer’s premium income (per unit time) is greater than the expected
claims outgo (per unit time).
Question 9.5
If we know the distribution of the aggregate claims S (t ) , we can often determine the
probability of ruin for the discrete model over a finite time horizon directly (without
reference to the models), by looking at the cashflows involved.
Example
The aggregate claims arising during each year from a particular type of annual insurance
policy are assumed to follow a normal distribution with mean 0.7 P and standard
deviation 2.0 P , where P is the annual premium. Claims are assumed to arise
independently. Insurers are required to assess their solvency position at the end of each
year.
A small insurer with an initial surplus of £0.1m for this type of insurance expects to sell
100 policies at the beginning of the coming year in respect of identical risks for an annual
premium of £5,000. The insurer will incur expenses of 0.2 P at the time of writing each
policy. Calculate the probability that the insurer will prove to be insolvent for this
portfolio at the end of the coming year. Ignore interest.
Solution
Using the information given, the insurer’s surplus at the end of the coming year will be:
Ê 0.5m - 0.35m ˆ
= 1- FÁ ˜¯
Ë 0.1m
= 1 - F(1.5) = 1 - 0.93319 = 0.067
Question 9.6
If the insurer expects to sell 200 policies during the second year for the same premium
and expects to incur expenses at the same rate, calculate the probability that the insurer
will prove to be insolvent at the end of the second year.
In fact, the normal distribution is probably not a very realistic distribution to use for the
claim amount distribution in most portfolios, as it is symmetrical, whereas many claim
amount portfolios will have skewed underlying distributions. However, it is commonly
used in CT6 exam questions.
Example
The number of claims from a portfolio of policies has a Poisson distribution with
parameter 30 per year. The individual claim amount distribution is lognormal with
parameters m = 3 and s 2 = 1.1 . The rate of premium income from the portfolio is 1,200
per year.
If the insurer has an initial surplus of 1,000, estimate the probability that the insurer’s
surplus at time 2 will be negative, by assuming that the aggregate claims distribution is
approximately normal.
Solution
First we need the mean and variance of the aggregate claims in a two-year period. The
expected number of claims will be 60. So the mean and variance are (using the formulae
for the first two moments of a lognormal distribution):
Ruin will occur if S (2) is greater than the initial surplus plus premiums received. So we
want:
È 3, 400 - 2, 088.80 ˘
P [ S (2) > 1, 000 + 2 ¥ 1, 200] ª P Í N (0,1) > ˙ = 1 - F(2.8053) = 0.0025
Î 218, 457 ˚
So far, we have used c to denote the rate of premium income per unit time, independent
of the claims outgo. In some circumstances it is more useful to think of the rate of
premium income as being related to the rate of claims outgo.
For the insurer to survive, the rate at which premium income comes in needs to be
greater than the rate at which claims are paid out. If this is not true, the insurer is
certain to be ruined at some point.
c = (1 + θ )λm1
The security loading is the percentage by which the rate of premium income exceeds the
rate of claims outgo. So, for the Poisson process outlined above, we have:
c = (1 + q ) E ( S ) = (1 + q )l m1
where q is the security loading. θ is also sometimes called the “relative security
loading”. It might typically be a figure such as 0.2, ie 20%.
The insurer will need to adopt a positive security loading when pricing policies, in order
to cover expenses, profit, contingency margins and so on.
Note that this does not mean that ruin is impossible. It is quite possible for the actual
claims outgo to exceed substantially its expected value. So even in this situation the
insurer’s probability of ruin is non-zero.
Question 9.7
In summary:
For a compound Poisson process S (t ) , the mean and variance of the total claim amount
are given by:
E[ S (t )] = l t E ( X ) var[ S (t )] = l t E ( X 2 )
M S (t ) (r ) = exp (l t ( M X (r ) - 1)
2.6 A technicality
lim M X (r ) = • (2.9)
-
r Æg
(For example, if the X i s have a range bounded above by some finite number,
then γ will be ∞ ; if the X i s have an exponential distribution with parameter α ,
then γ will be equal to α .)
Suppose for example that claim amounts have a continuous uniform distribution on the
interval (0,10) , so that they are bounded above by 10. Then the moment generating
function of the claim distribution is (from the Tables):
e10r − 1
M X (r ) =
10r
This is defined for all positive values of r , and so in this case γ = ∞ . We can see that
as r → ∞ , the limit of the MGF is infinite. If the claim distribution is Exp(α ) , the
MGF (as stated in the Tables) is:
α
M X (r ) = (1 − r / α ) −1 =
α −r
lim (l M X (r ) - cr ) = • (2.10)
-
r Æg
This is because λ , c and r would all have finite values in the limit.
Now it will be shown that (2.10) holds when g is infinite. This requires a little
more care. First note that there is a positive number, ε say, such that:
P[ X i > e ] > 0
The reason for this is that all claim amounts are positive.
So, if we pick a small enough number ( ε = 0.01 maybe), we’re bound to get a
proportion of claims whose amount exceeds this.
M X ( r ) ≥ e rε π
M X (r ) = E (e rX ) = E (e rX | X £ e ) P ( X £ e ) + E (e rX | X > e ) P ( X > e )
≥ 0 + e re p
Hence:
lim (l M X (r ) - cr ) ≥ lim (l e r e p - cr ) = •
r Æ• r Æ•
Here the e rε term is tending to +∞ , while the −cr term is tending to −∞ . Remember
that in such cases the exponential term always “wins”. So the limit is +∞ .
Important information
lim (l M X (r ) - cr ) = •
r Æ•
ψ (U ) ≤ exp{− RU }
The proof of Lundberg’s inequality is not on the syllabus for Subject CT6.
Don’t worry at this stage about what R actually represents. It will be defined shortly.
Until then just think of it as a parameter associated with the surplus process.
Note that if we can find a value for R , then Lundberg’s inequality tells us that we can
find an upper bound for the probability of ruin. This is a very useful result.
Figure 1 shows a graph of both exp{ -RU } and ψ (U ) against U when claim
amounts are exponentially distributed with mean 1, and when the premium
loading factor is 10%. (The solution for R will be found in Section 3.2. The
formula for ψ (U ) is given in Section 3.)
10
In fact it can be shown that the value of R in this case is .
11
It can be seen that, for large values of U , ψ (U ) is very close to the upper bound,
so that ψ (U ) ≈ exp{− RU } .
Figure 1
R can be interpreted as measuring risk. The larger the value of R , the smaller
the upper bound for ψ (U ) will be. Hence, ψ (U ) would be expected to decrease
as R increases. R is a function of the parameters that affect the probability of
ruin, and R ’s behaviour as a function of these parameters can be observed.
Note that R is an inverse measure of risk. Larger values of R imply smaller ruin
probabilities, and vice versa.
(i) the claim amount distribution is exponential with mean 10, and
Figure 2
Question 9.8
Note also that the value of R when claim amounts are exponentially distributed
is less than the value of R when all claim amounts are 10. Again, this result is
not surprising. Both claim amount distributions have the same mean, but the
exponential distribution has greater variability. Greater variability is associated
with greater risk, and hence a larger value of ψ (U ) would be expected for the
exponential distribution, and a lower value of R . This example illustrates that R
is affected by the premium loading factor and by the characteristics of the
individual claim amount distribution. R is now defined and shown, in general, to
encapsulate all the factors affecting a surplus process.
The surplus process depends on the initial surplus, on the aggregate claims
process and on the rate of premium income. The adjustment coefficient is a
parameter associated with a surplus process which takes account of two of
these factors: aggregate claims and premium income. The adjustment
coefficient gives a measure of risk for a surplus process. When aggregate
claims are a compound Poisson process, the adjustment coefficient is defined in
terms of the Poisson parameter, the moment generating function of individual
claim amounts and the premium income per unit time.
l M X (r ) - l - cr = 0 (3.1)
l M X (R ) = l + cR (3.2)
Note that, although R relates to the aggregate claims, the MGF used in the definition is
for the individual claim amount.
It is probably not at all obvious to you at this stage why R is defined in this way. The
reason is bound up with the proof of Lundberg’s inequality, which you are not required
to know. Please accept the definition, so that you can find the value of R in simple
cases.
Note that equation (3.1) implies that the value of the adjustment coefficient
depends on the Poisson parameter, the individual claim amount distribution and
the rate of premium income. However, writing c = (1 + θ )λm1 gives:
M X (r ) = 1 + (1 + θ )m1r
You can see from this equation that all the λ ’s have cancelled.
Important information
λ M X (r ) = λ + cr or M X (r ) = 1 + (1 + q )m1r
Since the rate of claims outgo per unit time is λm1 (the mean of a compound Poisson
distribution), then if a loading factor of θ is used, the rate of premium income will be
c = (1 + θ ) λm1 .
Example
An insurer knows from past experience that the number of claims received per month
has a Poisson distribution with mean 15, and that claim amounts have an exponential
distribution with mean 500. The insurer uses a security loading of 30%. Calculate the
insurer’s adjustment coefficient and give an upper bound for the insurer’s probability of
ruin, if the insurer sets aside an initial surplus of 1,000.
Solution
M X (r ) = 1 + (1 + q )m1r
Ê 1 ˆ
We have X ~ Exp Á ˜ , so that M X (r ) = (1 - 500r ) -1 (this comes from the Tables),
Ë 500 ¯
q = 0.3 , and m1 = E[ X ] = 500 . Substituting these into the equation:
Rearranging:
y (U ) £ e -0.000462¥1,000 = 0.630
Question 9.9
It may not be obvious to you why R does not depend on the Poisson parameter. The
basic reason is that increasing the Poisson parameter speeds up the whole process, so
that claims arise more quickly. This means that ruin, if it is going to happen, will
happen sooner, rather than later. However it does not affect the probability that ruin
does actually occur, when we are considering ruin at any time in the future.
It can be shown that there is indeed only one positive root of (3.1) as follows.
d d
g (r ) = λ M X (r ) − c
dr dr
It can also be shown that if the function g (r ) has a turning point, it must be at
the minimum of the function. The second derivative is:
d2 d2
g (r ) = λ M X (r )
dr 2 dr 2
M X¢¢ (r ) = E ( X 2e rX )
The function in this expectation is made up of two positive factors, and hence the
expectation must have a positive value.
Hence, there can only be one turning point, since any turning point is a
minimum. To show that there is a turning point, note from (3.10) that
lim g (r ) = • . Since g (r ) is a decreasing function at r = 0 , it must have a
r Æg -
minimum turning point and so the graph of g (r ) is as shown in Figure 3.
Figure 3
This is in the Tables, using a as the parameter for the exponential distribution, which
avoids confusion with the Poisson parameter.
-1
α Ê rˆ
For this distribution, M X (r ) = , (the Tables give this as Á1 - ˜ ) so:
α −r Ë a¯
la
l + cR =
a -R
fi la - l R + cRa - cR 2 = la
Ê lˆ l
fi R2 - Áa - ˜ R = 0 fi R = a - (3.3)
Ë c¯ c
(1 + q )l aq 1
If c = , then R = , since the mean of this distribution is m1 = .
a (1 + q ) a
Question 9.10
Write down the equation for the adjustment coefficient for personal accident claims if
90% of claims are for £10,000 and 10% of claims are for £25,000, assuming a
proportional security loading of 20%.
Show that this equation has a solution in the range 0.00002599 < R < 0.00002601 .
λ + cR = λM X (R )
z
∞
= λ e Rx f ( x )dx
0
z
∞
> λ (1 + Rx + 21 R 2 x 2 )f ( x )dx
0
= λ (1 + Rm1 + 21 R 2 m 2 )
The inequality is true because all the terms in the series for e Rx are positive. So e Rx
must always be greater than the total of the first few terms.
l + cR = l E[e RX ]
È 1 ˘
= l E Í1 + RX + R 2 X 2 + ˙
Î 2 ˚
È È1 ˘ ˘
= l Í E[1] + E[ RX ] + E Í R 2 X 2 ˙ + ˙
Î Î2 ˚ ˚
È 1 ˘
> l Í1 + RE[ X ] + R 2 E[ X 2 ]˙
Î 2 ˚
then it should be very close to this upper bound since the approximation to e Rx
should be good.
Question 9.11
Find an upper limit for the adjustment coefficient in the previous self assessment
question, and comment on your answer.
A lower bound for R can be derived when there is an upper limit, say M , to the
amount of an individual claim. For example, if individual claim amounts are
uniformly distributed on (0,100) , then M = 100 . The result is proved in a similar
fashion to Result (3.4). The lower bound is found by applying the inequality:
x x
exp(Rx ) £ exp(RM ) + 1 - for 0 £ x £ M (3.5)
M M
•
x x x (RM ) j x
M
exp{RM } + 1 - =
M M j =0 j !
+ 1-ÂM
•
R j M j - 1x
= 1+ Â j!
j =1
•
(Rx ) j
≥ 1+ Â j!
for 0 £ x £ M
j =1
= exp{Rx }
j -1
since x j £ M x if 0 £ x £ M .
1
R> log(c / λm1 )
M
z
M
λ + cR = λ exp{Rx }f ( x )dx
0
z FGH IJ
M
≤λ
x
M
l q
exp RM + 1 −
x
M K
f ( x )dx
0
λ λ
= exp{RM }m1 + λ − m1
M M
Hence, rearranging:
c 1 RM (RM )2
£ (exp{RM } - 1) = 1 + + +
l m1 RM 2 3!
RM ( RM )
2
<1+ + +
1! 2!
= exp{RM }
1 Ê c ˆ
This gives R > log Á as required.
M Ë l m1 ˜¯
1
If c = (1 + θ ) λm1 , this is just R > log(1 + θ ) .
M
Question 9.12
Find a lower bound for the adjustment coefficient in the example in Question 9.10. Is
the argument valid when the claim distribution is discrete?
Important information
2(c - l m1 ) 2q m1
R< or R< when c = (1 + q )l m1
l m2 m2
1 Ê c ˆ 1
R> log Á or R> log(1 + θ )
M Ë l m1 ¯˜ M
In Section 3.2 the existence of the adjustment coefficient, R, was proved for a
compound Poisson aggregate claims process. In this section the existence of
the adjustment coefficient for a general aggregate claims process is proved.
In the general situation the adjustment coefficient is the positive number R that
can be shown to satisfy the following:
E [e R (Si - c ) ] = 1
The proof that there is one, and only one, positive number R to satisfy this is as
follows.
lim f (r ) = +•
r Æg -
We then use the same argument as for the compound Poisson case, based on the graph
below. This graph forms part of the Core Reading.
f(r)
0 R γ r
Then:
E [e R (Si - c ) ] = 1
fi E [e RSi ] = e Rc
fi M S ( R) = e Rc
fi e l [ M X (R ) - 1] = e Rc
fi l M X (R ) = Rc + l
4.1 Introduction
Recall that y (U ) was defined to be P (U (t ) < 0,t > 0) , and y (U , t ) was defined to be
y (U , t ) = P (U (t ) < 0, 0 < t < t ) .
No new theory will be introduced and the method for obtaining numerical values
for ψ (U , t ) will not be discussed. Features of ψ (U , t ) , and in some cases of
ψ (U ) , will be illustrated by a series of numerical examples. In these examples
the same basic assumptions will be made as in previous sections. In particular,
it will be assumed that the aggregate claims process is a compound Poisson
process. In addition it will be assumed throughout Section 4.3, Section 4.4 and
Section 4.5 that:
In Section 4.6, Assumptions (4.2) and (4.3) will be made, but the Poisson
parameter will be allowed to vary.
The implication of Assumption (4.1) is that the unit of time has been chosen to
be such that the expected number of claims in a unit of time is 1. Hence
ψ (U , 500) is the probability of ruin (given initial surplus U ) over a time period in
which 500 claims are expected. The actual number of claims over this time
period has a Poisson distribution (with parameter 500) and could take any non-
negative integer value.
The implication of Assumption (4.2) is that the monetary unit has been chosen to
be equal to the expected amount of a single claim. Hence ψ (20, 500) is the
probability of ruin (over a time period in which 500 claims are expected) given an
initial surplus equal to 20 times the expected amount of a single claim.
When F ( x ) = 1 - exp( - x ) :
1 Ê qU ˆ
y (U ) = exp Á - (4.4)
1+q Ë 1 + q ˜¯
This result has been stated in order to illustrate how, for this particular
distribution, the ultimate ruin probability is affected by changes in parameter
values.
4.3 y (U , t ) as a function of t
Question 9.13
Figure 4 shows a graph of ψ (15, t ) for 0 £ t £ 500 . The premium loading factor,
θ , is 0.1 so that the premium income per unit time is 1.1. Also shown in Figure 4
are ψ (15) (dotted line) and exp( -15R ) (solid line) for this portfolio. These last
two values are shown as lines parallel to the time axis since their values are
independent of time.
Here:
● ψ (15, t ) has been worked out using a numerical method not described here
● ψ (15) has been calculated using equation 4.4
● l q
exp −15R is the upper bound given by Lundberg’s inequality.
Figure 4
Question 9.14
Question 9.15
(ii) for small values of t , ψ (15, t ) increases very quickly (its value doubles as
t increases from 25 to 50 and doubles again as t increases from 50 to
100),
(iii) for larger values of t , ψ (15, t ) increases less quickly and approaches
asymptotically the value of ψ (15) .
General reasoning should help you to understand (ii) and (iii). You would expect a
much higher probability of ruin before time 50 than before time 25 since the overall
performance of the fund could easily change in such a short time period. However, if
premium rates are expected to be profitable in the long term, then at time 400, say, a
significant surplus will have built up in most cases and so the probability of ruin at time
425 won’t be that much higher than at time 400. We are assuming here that
accumulated surpluses stay in the fund and are not, for example, distributed to
shareholders.
Question 9.16
Figure 5 shows values of ψ (U , t ) for 0 £ t £ 500 and for three values of the initial
surplus, U = 15, 20, and 25. The premium loading factor is 0.1 as in Figure 4.
For U = 15 the graph of ψ (U , t ) is as in Figure 4.
Figure 5
Question 9.17
(ii) increasing the value of U decreases the value of ψ (U , t ) for any value of
t,
ψ (U ) is a non-increasing function of U .
d −θ
ψ (U ) = ψ (U )
dU 1+ θ
Question 9.18
Figure 6 shows values of ψ (15, t ) for 0 £ t £ 500 and for three values of the
premium loading factor, q = 0.1, 0.2 and 0.3. The graph of ψ (15, t ) for q = 0.1 is
the same as the graph in Figure 4 and the same as one of the graphs in Figure 5.
Figure 6 is, in many respects, similar to Figure 5. The features of note in Figure 6
are:
(i) the graphs of ψ (15, t ) all have the same general shape,
(ii) increasing the value of θ decreases the value of ψ (15, t ) for any given
value of t ; this is in fact true for any value of U , and is an obvious result
since an increase in θ is equivalent to an increase in the rate of premium
income with no change in the aggregate claims process,
(iii) it can be seen that when q = 0.2 and 0.3, ψ (15, t ) is more or less constant
for t greater than about 150. For t1 £ t2 , the difference ψ (15, t 2 ) − ψ (15, t 1 )
represents the probability that ruin occurs between times t 1 and t 2 .
Hence for these values of θ , 0.2 and 0.3, (and for this value of the initial
surplus, 15, and for this aggregate claims process) ruin, if it occurs at all,
is far more likely to occur before time 150, ie within the time period for 150
claims to be expected, than after time 150. This point will be discussed
further in Section 4.6.
Figure 6
d
ψ (U ) = − (1 + θ ) −1ψ (U ) − U (1 + θ ) −2 ψ (U )
dθ
Question 9.19
d
Verify this expression for ψ (U ) .
dθ
This is clearly negative since θ , U and ψ (U ) are all positive quantities. Since
the derivative is less than zero for all values of θ , ψ (U ) is a decreasing function
of θ .
Figure 7
Figure 8
The unit of time is taken to be one year. The only difference between these risks
is that twice as many claims are expected each year under Risk 1. This is
reflected in the two premiums.
Consider Risk 2 over a new time unit equivalent to two years. Then the
distribution of aggregate claims and the premium income per unit time are now
identical to the corresponding quantities for Risk 1. Hence, the probability of
ruin over an infinite timespan is the same for both risks.
The solid line in Figure 9 shows an outcome of the surplus process for Risk 1
when θ = 0.1 . The dotted line shows the same surplus process when the unit of
time is two years. This illustrates that any outcome of the surplus process that
causes ultimate ruin for Risk 1 will also cause ultimate ruin for Risk 2. There is
thus no difference in the probability of ultimate ruin for these two risks. It is only
the time (in years) until ruin that will differ. Measuring times in years, the
probability of ruin by time 1 for Risk 1 is the same as the probability of ruin by
time 2 for Risk 2. This explains why Figures 6 and 8 show the same functions.
For example, the value of ψ (15,10) when λ = 50 (Figure 8) is the same as the
value of ψ (15, 500) when λ = 1 (Figure 6).
Figure 9
Point (iii) in Section 4.5 will now be investigated, where it was noted that values
of ψ (15, t ) were more or less constant for values of t greater than 150 when
q = 0.2 and 0.3. In particular, the situation will be considered when the premium
loading factor is 0.2.
Consider a second aggregate claims process, which is the same as the process
considered throughout this section except that its Poisson parameter is 150 and
not 1. (This second process is really identical to the original one; all that has
happened is that the time unit has been changed.) Use ψ * to denote ruin
probabilities for the second process and ψ to denote, as before, ruin
probabilities for the original process. The change of time unit means that for any
t ≥ 0:
ψ * (U , t ) = ψ (U ,150t )
ψ * (U ) = ψ (U )
ψ (15,150) ≈ ψ (15)
From this and the previous two relations it can be seen that:
y * (15,1) ª y * (15)
In words this relation says that for the second process, starting with initial
surplus 15, the probability of ruin within one time period is almost equal to
(actually a little less than) the probability of ultimate ruin. This conclusion
depends crucially on the fact that ψ * (15,1) is a continuous time probability of
ruin. To see this, consider ψ * (15,1) , which is just the probability that for the
second process the surplus at the end of one time period is negative. ψ * (15,1)
can be calculated approximately by assuming that the aggregate claims in one
time period, which will be denoted S*(1), have a normal distribution. Recall that
individual claims have an exponential distribution with mean 1 and that the
number of claims in one time period has a Poisson distribution with mean 150.
From this:
These are calculated as λm1 = 150 × 1 = 150 and λm2 = 150 × 2 = 300 , where
E ( X 2 ) = 2 for an Exp(1) distribution.
ª 0.005
X -m
Recall that if X ~ N ( m , s 2 ) , then Z = has a standard normal distribution,
s
ie Z ~ N (0,1) . Probabilities for this distribution can be looked up in the Tables.
From Figure 6 it can be seen that the value of ψ (15,150) , and hence of ψ * (15,1) ,
is about 0.07 which is very different from the (approximate) value of the discrete
time probability of ruin ψ * (15,1) calculated above.
When individual claim amounts are exponentially distributed with mean 1, first
note that if θ = 0 , then ψ (U ) = 1 irrespective of the value of U .
This result is in fact true for any form of F ( x ) . (It trivially follows that if θ < 0 ,
then ψ (U ) = 1.) In other words a positive premium loading is essential if ultimate
ruin is not to be certain.
Also note that throughout this section it has been assumed that individual claim
amounts are exponentially distributed with mean 1. This mean could be
measured in units of £100, £1,000 or perhaps £1,000,000. The parameter of the
exponential distribution can be set to 1 without loss of generality provided that
the monetary unit is correctly specified. In simple terms, the probability of ruin
when U is £1 is the same as the probability of ruin when U is 100 pence. It can
be said that:
ψ (U ) when F ( x ) = 1 − e −αx
ψ (α U ) when F ( x ) = 1 − e − x
In other words, if the expected claims per unit time increase by a factor α , so
too must the initial surplus if the probability of ultimate ruin is to be unchanged.
5.1 Introduction
One of the options open to an insurer who wishes to reduce the variability of
aggregate claims from a risk is to effect reinsurance. A reduction in variability
would be expected to increase an insurer’s security, and hence reduce the
probability of ruin. A reinsurance arrangement could be considered optimal if it
minimises the probability of ruin. As it is difficult to find explicit solutions for the
probability of ruin, the effect of reinsurance on the adjustment coefficient will be
considered instead. If a reinsurance arrangement can be found that maximises
the value of the adjustment coefficient, the upper bound for the probability of
ultimate ruin will be minimised. As the adjustment coefficient is a measure of
risk, it seems a reasonable objective to maximise its value. In the following, the
effect on the adjustment coefficient of proportional and of excess of loss
reinsurance arrangements will be considered.
Throughout this section we will use the notation X = individual claim amount,
Y = amount paid by the direct insurer and Z = amount paid by the reinsurer.
Example
Consider the insurer in the example on page 19. This insurer is investigating the
possibility of using proportional reinsurance. He has approached a reinsurer, who uses
a security loading of 50% to calculate his reinsurance premiums. If the insurer decides
to reinsure 20% of each risk in the portfolio, estimate the effect the reinsurance will
have on his probability of ruin at Time 2. Again you can assume that the aggregate
claim distribution is approximately normal.
Solution
We first need to calculate the reinsurance premium. Since the reinsurer takes
responsibility for 20% of each risk, and uses a loading factor of 50%, the reinsurance
premium (per annum) will be:
So over a two year period, the insurer will pay 626.64 for the reinsurance.
We now use Snet (2) for the insurers aggregate payments (net of reinsurance). We need
the mean and variance of Snet (2) , which are, using the formulae for a compound
Poisson distribution as before:
È 2, 773.36 - 1, 671.04 ˘
P [ Snet (2) > 2, 773.36] = P Í N (0,1) > ˙
Î 139,812.49 ˚
= 1 - F(2.9481) = 0.0016
or about 0.16%.
Question 9.20
Y =aX Z = (1 - a ) X
Hence:
Question 9.21
We can apply the same type of logic if the insurer decides to buy excess of loss
reinsurance. You might like to think about the effect on the probability of ruin if the
insurer in the previous example purchases excess of loss reinsurance with an individual
retention of 2,000, say, and a security loading of 50% as before.
ÏX X < M Ï 0 X <M
Y =Ì Z =Ì
ÓM X ≥ M ÓX - M X ≥ M
Example
Calculate E[Y ] if X has an exponential distribution with parameter 0.01, and the
insurer has an excess of loss reinsurance arrangement with retention limit M .
Solution
M •
E[Y ] = Ú xf ( x) dx + Ú Mf ( x) dx
0 M
M •
-0.01x -0.01x
= Ú 0.01xe dx + Ú 0.01Me dx
0 M
M M ∞
⎡ 0.01x −0.01x ⎤ 0.01 −0.01x ⎡ 0.01 −0.01x ⎤
= ⎢−
⎣ 0.01
e ⎥ −
⎦0
∫ −
0.01
e dx + ⎢ − M
⎣ 0.01
e ⎥
⎦M
0
M
M ⎡ 1 ⎤ ∞
= ⎡ − xe−0.01x ⎤ − ⎢ e −0.01x ⎥ + ⎡ − Me −0.01x ⎤
⎣ ⎦ 0 ⎣ 0.01 ⎦0 ⎣ ⎦M
1 −0.01M 1
= − Me −0.01M − e + + Me−0.01M
0.01 0.01
=
1
0.01
(
(1 − e −0.01M ) = 100 1 − e −0.01M )
Question 9.22
Calculate var[ Z ] (in terms of M ) if X ~ U (0,100) , where the insurer has an excess of
loss reinsurance arrangement with retention limit M , 0 < M < 100 .
Thus, ξ is the premium loading factor used by the reinsurer. Hence, the
insurer’s premium income, net of reinsurance, is:
Question 9.23
It will also be assumed that ξ ≥ θ . If this were not true, it would be possible for
the insurer to pass the entire risk on to the reinsurer and to make a certain profit.
This of course ignores commission, expenses and other adjustments to the theoretical
risk premium.
1 + θ > (1 + ξ )(1 − α )
ie α > (ξ − θ ) / (1 + ξ )
Question 9.24
There is, however, a more important constraint on the insurer. The insurer’s net
of reinsurance premium income per unit time must exceed the expected
aggregate claims per unit time. Otherwise ultimate ruin is certain (as noted in
Section 4.7). Net of reinsurance, the insurer’s expected aggregate claims per
unit time are αλm1 .
Thus:
(1 + θ ) − (1 + ξ )(1 − α ) > α
or:
θ
α > 1− (5.2)
ξ
Question 9.25
So what is the range of possible values of α now, given the figures in the previous
question?
1 − θ / ξ ≥ (ξ − θ ) / (1 + ξ ) when θ ≤ ξ
the only case of interest. If the premium loading factors are equal, then
inequality (5.2) becomes α > 0 . In this case there exists a risk sharing
arrangement and any retention level is possible. If, however, ξ > θ then the
insurer has to retain part of the risk.
Same loadings
First consider the case where both the insurer and the reinsurer use θ as the
premium loading factor. The adjustment coefficient will be found as a function
of the retention level α , when F ( x ) = 1 - e -0.1x .
P [Y ≤ y ] = P [ X ≤ y / α ] = 1 − exp{−0.1y / α }
Note that the assumptions for the claims process and the adjustment coefficient equation
apply equally well in the presence of reinsurance, provided that we use the net premium
and the net claim amounts in the adjustment coefficient equation.
Question 9.26
What will be the general equation for R , the direct insurer’s adjustment coefficient,
when there is reinsurance?
λ + (1 + θ )λ 10α R = λ z
∞ Rx
0
e (0.1 / α )e −0.1x / α dx
1
⇒ 1 + (1 + θ )10α R = (5.3)
1 − 10αR
θ
⇒R= for 0 < α ≤ 1 (5.4)
(1 + θ )10α
Different loadings
Now consider what happens when ξ > θ . For the rest of Section 5.4 assume
that:
Case A – x = 0.2
Suppose first that the reinsurer’s premium loading factor is ξ = 0.2 , so that the
insurer’s (net) premium income per unit time is (12α − 1)λ .
Equation (5.2) shows that the insurer must retain at least 50% of each claim.
Hence, a value of α will be sought in the interval [0.5,1] that maximises the
value of R . The equation defining R is:
λ
λ + (12α − 1)λR =
1 − 10αR
Question 9.27
This follows from (5.3) as only the premium is different − which leads to:
2a - 1
R= for 0.5 < α ≤ 1 (5.5)
10(12a 2 - a )
The right hand side is based on the MGF of the net claim amounts which have an
Exp(01. / α ) distribution.
Question 9.28
As when the loading factors were equal, the adjustment coefficient depends on
the retention level.
du dv
v -u
d Êuˆ da da .
The quotient rule is ÁË ˜¯ =
da v v 2
1 1
Alternatively the algebra is a little easier if you write R = − .
10α 12α − 1
Now the denominator is always positive for values of α in [0.5,1] , so there will
be a turning point of the function when:
ie when:
24α 2 − 24α + 1 = 0
The roots of this quadratic are 0.9564 and 0.0436, and so the turning point which
is of interest is 0.9564.
α R
0.5 0
0.9564 0.00911
1.0 0.00909
Figure 10
It can be seen from Figure 10 that there is a range of values for α , β < α < 1,
such that if the retention level is in this range, the value of the adjustment
coefficient exceeds the value when α = 1 . The value of β can be calculated
from (5.5) by setting the value of R at β equal to the value of R at 1, giving
b = 0.9167 . (You can check for yourself that β = 11 / 12 .) The arrow in Figure 10
indicates the value of β .
If the insurer effects reinsurance with retention level 0.9564, then the expected
profit per unit time is 0.9128λ (ie premium income, from (5.1), less expected
claims).
The expected profit per unit time is now found in terms of α and λ .
It has already been calculated from (5.1), that with θ = 0.1 , ξ = 0.2 and m1 = 10 ,
the insurer’s net premium income is (12α − 1)λ . The insurer’s expected claims
per unit time are 10αλ . Hence, the expected profit per unit time is (2α − 1)λ .
This shows that expected profit per unit time is an increasing function of α , and
if the insurer were to choose α to maximise the expected profit per unit time, the
choice would be α = 1 . This example illustrates a general point – the level of
reinsurance is a trade-off between security and profit.
Case B – x = 0.3
The value of α is now found that maximises R when the reinsurer’s premium
loading factor is 0.3.
From (5.1), the insurer’s net premium income is (13α − 2)λ , so that the equation
defining R is:
λ
λ + (13α − 2)λR =
1 − 10α R
3α − 2
R= for 0.67 < α ≤ 1
10(13α 2 − 2α )
1 1
Or R = − , adopting the same approach as before.
10α 13α − 2
By (5.2), the insurer must retain at least 2/3 of each claim so the value of a in the
range [2/3, 1] that maximises R is sought. Differentiation gives:
ie when:
39α 2 − 52α + 4 = 0
The roots of this quadratic are 0.0820 and 1.2514, so there are no turning points
in the interval [2 3 ,1] and R as a function of α in this interval increases from 0
at α = 2 / 3 to 0.00909 at α = 1 . Thus, the value of α which maximises the
adjustment coefficient is 1.
The insurer’s premium loading factor, net of reinsurance, implied by (5.1) is now
found, and shown to be an increasing function of α .
The loading factor is found by dividing the expected profit per unit time by the
expected claims per unit time. The expected profit per unit time is:
This is just an algebraic expression for net premiums less expected net claims.
q ¢ = [(1 + q ) - (1 + x )(1 - a ) - a ] / a
= x - (x - q ) / a
dθ ′
Now = (ξ − θ ) / α 2 which is positive since ξ > θ , so that θ ′ is an increasing
dα
function of α . Thus, the net loading factor increases as the retention level
increases.
The insurer’s individual net claim payments are distributed as Y = min( X , M ) , and
the insurer’s premium income, net of reinsurance, is:
c * = (1 + θ )λm1 − (1 + ξ )λE (Z )
(you may see this referred to as cnet ) which gives the equation defining R as:
ÈM ˘
Í Ú
l + c * R = l Í e Rx f ( x )dx + e RM [1 - F (M )]˙
˙
Î0 ˚
Question 9.29
Explain where the right hand side of this equation comes from.
This is formula (4.2) with a truncated claim amount distribution as a result of the
excess of loss reinsurance. To illustrate ideas, look at the situation when
X ~ U (0,20) , so that f ( x ) = 0.05 for 0 < x < 20 . Then for 0 < M £ 20 :
20
Ú ( x - M ) 0.05 dx = 10 - M + 0.025M
2
E (Z ) =
M
and:
z
M
MY (R ) = e Rx 0.05dx + e RM (1 − 0.05M )
0
0.05 RM
= (e − 1) + e RM (1 − 0.05M )
R
The equation for R must be solved numerically for given values of θ and ξ .
Figure 11 shows R as a function of M when θ = ξ = 0.1 . As in Section 5.2, any
retention level is possible when the premium loading factors are equal. R is a
decreasing function of M .
Figure 11
When θ < ξ , there is a minimum retention level for the same reason as in the
previous section.
Recall that the lower limit for α given by Equation 5.2 applied when we were
considering proportional reinsurance.
For example, when θ = 0.1 and ξ = 0.2 the insurer’s net premium income, c * , is
11λ − 1.2λ (10 − M + 0.025M 2 ) and this must exceed the insurer’s expected claims,
net of reinsurance. The insurer’s expected net claims equal λE ( X ) − λE (Z ) ,
which gives λ (M − 0.025M 2 ) . Thus:
Hence, the minimum retention level is 5.8579. Similarly, when ξ = 0.4 , the
minimum retention level is 10.
Figure 12
Hence, for ξ = 0.2 it is possible for the insurer to increase the value of the
adjustment coefficient by effecting reinsurance, provided that the retention level
is above 9.6. However, when ξ = 0.4 , the insurer should retain the entire risk in
order to maximise the value of the adjustment coefficient. As in the case of
proportional reinsurance, the insurer’s expected profit per unit time is reduced if
reinsurance is effected.
Question 9.30
Claims occur as a Poisson process with rate λ and individual claim sizes X follow an
Exp( β ) distribution. The office premium includes a security loading θ 1 . An
individual excess of loss arrangement operates under which the reinsurer pays the
excess of individual claims above an amount M in return for a premium equal to the
reinsurer’s risk premium increased by a proportionate security loading θ 2 . Derive and
simplify as far as possible an equation satisfied by the adjustment coefficient for the
direct insurer.
Question 9.31
6 Exam-style questions
We conclude this chapter with two exam-style questions:
Exam-style question 1
Explain how each the following changes to the company’s model will affect the
probability of ultimate ruin:
(ii) Individual claims are found to have a gamma distribution with parameters
a = 150 and l = ½ . [1]
Exam-style question 2
Claims occur according to a compound Poisson process at a rate of 0.2 claims per year.
Individual claim amounts, X , have probability function:
P ( X = 50) = 0.7
P ( X = 100) = 0.3
The insurer’s surplus at time 0 is 75 and the insurer charges a premium of 120% of the
expected annual aggregate claim amount at the beginning of each year. The insurer’s
surplus at time t is denoted U (t ) . Find:
This page has been left blank so that you can keep the chapter summaries
together as a revision tool.
Chapter 9 Summary
The surplus of a general insurer at future times can be modelled using aggregate claims
process models, which can be used to find the probability of ruin over a finite or infinite
time horizon in discrete or continuous time.
Claim numbers can be modelled using a Poisson process. Total claim amounts can be
modelled using a compound Poisson process.
For the continuous time model with an infinite time horizon, Lundberg’s inequality,
which uses a parameter r called the adjustment coefficient, provides a good
approximation for the probability of ultimate ruin.
The upper bound for the probability of ultimate ruin given by Lundberg’s inequality is a
decreasing function of r . As a result, an insurer may want to effect a system of
reinsurance that maximises the value of r .
The probability of ultimate ruin also decreases if the insurer’s security loading is
increased or if the insurer’s initial surplus is increased.
An increase in the value of the Poisson parameter will not affect the probability of
ultimate ruin. However, it will reduce the time it takes for ruin to occur.
An equation for the adjustment coefficient can also be obtained when reinsurance is used,
by considering the direct insurer’s net premium income and net claims outgo.
In practice, the security loading used by the reinsurer will be greater than the security
loading used by the direct insurer.
Chapter 9 Formulae
Surplus process
U (t ) = u + ct − S (t ) , t ≥ 0 (continuous time)
Ruin probabilities
Poisson process
( λt ) x e − λt
P[ N (t ) = x ] = px (t ) = ( x = 0,1,2,… )
x!
f T (t ) = λe − λt (t > 0)
M S ( t ) (u) = e λt [ M X ( u) −1]
E [ S ( t )] = λ tE ( X ) var[ S ( t )] = λ tE ( X 2 )
Adjustment coefficient
For a compound Poisson process, the adjustment coefficient r is the unique positive root
of the equation:
λ + cr = λ M X (r )
2[c / λ − E ( X )] 1
r< r> log(c / λm1 )
E( X ) M
If the insurer uses a security loading of θ , the equation for the adjustment coefficient is:
1 + (1 + θ )m1r = M X (r )
1 + (1 + θ ) E ( X ) − (1 + ξ ) E ( Z ) r = M Y (r )
For a general aggregate claims process, the adjustment coefficient r is the unique
positive root of the equation: E[e r ( Si −c) ] = 1 , where Si denotes the aggregate claims
from a risk in time period i and c is the constant premium charged to insure this risk.
Lundberg’s inequality
ψ (u) ≤ e − ru
This page has been left blank so that you can keep the chapter summaries
together as a revision tool.
Chapter 9 Solutions
Solution 9.1
(i) Yes
(ii) No
x 2 x3
e- x = 1 - x + - +
2! 3!
e- x - 1 + x 1
This gives = x + terms in x 2 and higher powers.
x 2
Solution 9.2
As the amount of initial surplus increases, ruin will become less and less likely. So the
limit is zero.
Solution 9.3
y (U , t ) involves checking for ruin at all possible times. Since the more often we check
for ruin, the more likely we are to find it, we would expect that y (U , t ) would be
greater than y h (U , t ) .
Solution 9.4
(i) The expected number of claims reported on a given day is 5. So the number of
claims reported on a given day has a Poisson(5) distribution and the probability
that there will be fewer than 2 claims is:
Here we have used the formula for the Poisson probability, but alternatively,
using page 176 of the Tables, P ( N < 2) = P ( N £ 1) = 0.04043 .
(ii) The waiting time until the next event has an Exp(5) distribution. We need to
find a probability using the exponential distribution. To do this, we can use the
cumulative distribution function:
P (T < t ) = 1 - e - lt
1
So the probability that there will be a claim (or several claims) during the next hour ( 24
of a day) is:
5
- 24
P (T < 1) = 1- e = 0.1881 ie 18.8%
24
Note that it is unlikely that the rate would be constant over time in reality.
Solution 9.5
Otherwise the insurer would be charging premiums that were less than the amount it
expected to pay out in claims.
In the real world this assumption may not always be true, especially during periods of
competitive pressure when premium rates are soft.
Solution 9.6
The insurer’s surplus at the end of the second year will be:
The probability that the surplus will be negative at the end of the second year is:
Solution 9.7
1, 200 = (1 + q ) ¥ 30e3.55
The solution of this equation is q = 0.14899 . So the security loading is about 14.9%.
Solution 9.8
θ is the security loading, and ψ (U ) is the probability of ruin for a fixed level of
surplus U . If θ increases the premiums we charge will increase, and we should
become more secure, ie the probability of ruin should fall.
Solution 9.9
The first reason is that Core Reading defines R to be the unique positive root, so R
cannot be zero.
The second reason is that R = 0 should always be a solution to the equation. Why?
Consider the LHS. M X (0) = E[e0 ] = 1 . Consider the RHS. 1 + (1 + q )m1 ¥ 0 = 1 . So
R = 0 is always a solution, but why do we ignore it?
Solution 9.10
1 + (1 + θ ) m1R = M X ( R)
X=
RS10,000 with probability 0.9
T25,000 with probability 0.1
So:
and:
. × 11,500 R = 0.9e10,000 R + 01
1 + 12 . e25,000 R
ie 1 + 13,800 R = 0.9e10,000 R + 01
. e25,000 R
We can show that there is a solution in the range stated by looking at the values of
LHS − RHS :
Since there is a reversal of signs (and we are dealing with a continuous function), the
difference must be zero at some point between these two values, ie there is a solution of
the equation in the range 0.00002599 < R < 0.00002601 .
Solution 9.11
Here:
E ( X ) = 11,500
Also:
So:
2θ m1 2 × 0.2 × 11,500
R< = = 0.0000302
m2 152,500,000
So this is a reasonable initial estimate, compared with the correct value of approximately
0.000026.
Solution 9.12
1 1 1
R> log(c / λm1 ) = log(1 + θ ) = . = 0.00000729
log 12
M M 25,000
The steps used in the proof are equally valid for a discrete claims distribution.
However, note that the lower bound obtained here is not very close to the accurate value
for R .
Solution 9.13
y (U , t ) is the probability of ruin at some point between times 0 and t . This should
increase with time since the longer the time period, the more chance there is of ruin. It
should be intuitively obvious that y (U , t1 ) < y (U , t2 ) for t1 < t2 , since if a scenario
produces ruin before time t1 , then ruin has also occurred before time t2 .
Solution 9.14
15
1 −0.1× 11.
ψ (15) = e = 0.23248
11
.
Solution 9.15
αθ 1 × 01
.
R is worked out from R = = .
1+θ 11
.
So:
0.1
-15¥ 1.1
e -15 R = e = 0.2557
Solution 9.16
Decreasing. The bigger the initial surplus, the less chance there should be of ruin.
Solution 9.17
1 -0.1¥ 1.1
20
y (20) = e = 0.14756
1.1
This is the limit to which the middle of the three lines is tending to as t tends to ∞ .
Solution 9.18
Decreasing. If everything else remains unchanged, then increasing the premium income
will reduce the probability of ruin.
Solution 9.19
1
ψ (U ) =
1+θ
o
exp U (1 + θ ) −1 − U t
We need to differentiate this using the product rule:
d dv du
(uv) = u +v
dq dq dq
d
dθ
o
ψ (U ) = (1 + θ ) −1 −U (1 + θ ) −2 exp U (1 + θ ) −1 − U t − (1 + θ ) −2
o
exp U (1 + θ ) −1 − U t
Now substituting back in each term for ψ (U ) :
d
ψ (U ) = −U (1 + θ ) −2 ψ (U ) − (1 + θ ) −1ψ (U )
dθ
Solution 9.20
The reinsurance has reduced the probability of ruin to some extent, ie from about 0.25%
to about 0.16%. However, this result is probably quite sensitive to the assumptions
made (we are near the tail of the normal distribution), and slightly different assumptions
might give us very different results.
We will also want to look at the effect of reinsurance on profitability. As we are paying
a reinsurance premium, it is likely that the overall effect on profitability is negative
(although the effect on security is positive, as we have seen). There is likely to be a
trade off between security and profitability here.
Solution 9.21
Solution 9.22
M 100
Ú0 Ú
2 2
E[ Z ] = f ( x) dx + ( x - M ) 2 f ( x) dx
0 M
1
The pdf of the U (0,100) distribution is (see page 13 of the Tables), so:
100
100 100
( x - M )2 È ( x - M )3 ˘ (100 - M )3
Ú
2
E[ Z ] = dx = Í ˙ =
M
100 ÍÎ 300 ˙˚ 300
M
We now need E[ Z ] :
100 100
(x - M ) È ( x - M )2 ˘ (100 - M ) 2
E[ Z ] = Ú 100
dx = Í ˙ =
M ÎÍ 200 ˚˙ M
200
So:
2
(100 - M )3 Ê (100 - M ) 2 ˆ
var[ Z ] = -Á ˜
300 Ë 200 ¯
(100 - M )3 (100 - M ) 4
= -
300 40, 000
Solution 9.23
(1 + q )l E[ X ] = (1 + q )l m1
(1 + x )l E[ Z ]
(1 + x )l (1 - a )m1
Solution 9.24
0.2
α> = 01429
.
14
.
Solution 9.25
0.2
α > 1− = 0.5 . So 0.5 < α ≤ 1 .
0.4
Solution 9.26
From the previous work, we know that the equation for R is:
l + cr = l M X (r )
l + cnet r = l M Y (r )
l + ((1 + q )l E[ X ] - (1 + x )l E[ Z ]) r = l M Y (r )
or:
1 + ((1 + q ) E[ X ] - (1 + x ) E[ Z ]) r = M Y (r )
1 + ((1 + q ) E[ X ] - (1 + x )(1 - a ) E[ X ]) r = M Y (r )
Solution 9.27
1
Using the equation from the previous question where q = 0.1 , E[ X ] = = 10 and
0.1
x = 0.2 , we get:
fi 1 + (11 - 12(1 - a ) ) r = M Y (r )
fi 1 + (12a - 1)r = M Y (r )
-1
rY ra X Ê ra ˆ 1
M Y (r ) = E[e ] = E[e ] = M X (a r ) = Á1 - =
Ë 0.1˜¯ 1 - 10ra
1
1 + (12a - 1)r =
1 - 10ra
Solution 9.28
1
1 + (12α − 1) R =
1 − 10αR
Multiplying out the left hand side and subtracting 1 from both sides gives:
Solution 9.29
ÏX X < M
We need l M Y (r ) , where Y = Ì and X has pdf f ( x) . By definition,
ÓM X ≥ M
M Y (r ) = E[e rY ] , but we need to express this as two separate integrals to take into
account the different ranges of X :
M •
E[erY ] = Ú erx f ( x) dx + Úe
rM
f ( x) dx
0 M
M •
Úe Ú
rx rM
= f ( x) dx + e f ( x) dx
0 M
But the second integral is just integrating the pdf from M to • . This is the same as
P( X > M ) , which can be written as 1 - F ( M ) . So the right hand side of the equation
is:
ÈM ˘
l Í Ú erx f ( x) dx + erM (1 - F ( M )) ˙
ÍÎ 0 ˙˚
Solution 9.30
z
∞
1
c = (1 + θ 1 ) λ − (1 + θ 2 ) λ ( x − M )β e − β x dx
β M
The second term can be integrated using the substitution y = x − M , and identifying the
integral as the mean of an Exp( β ) distribution. This gives:
1
c=λ [(1 + θ 1 ) − (1 + θ 2 )e − β M ]
β
The individual net claims are the claims paid to policyholders minus the recoveries from
the reinsurer. So the MGF (which is valid for all values of R < β ) is:
z z
M ∞
1
M X ( R) = e R x β e − β x dx + e RM β e − β x dx = [β − R e−(β − R) M ]
0 M
β−R
1 1
λ + λ [(1 + θ 1 ) − (1 + θ 2 )e − β M ]R = λ [β − R e−( β − R) M ]
β β−R
β ( β − R ) + ( β − R )[(1 + θ1 ) − (1 + θ 2 )e − β M ]R = β [ β − R e − ( β − R ) M ]
− βR + ( β − R)[(1 + θ 1 ) − (1 + θ 2 )e − β M ]R = − β R e − ( β − R ) M
− β + ( β − R )[(1 + θ 1 ) − (1 + θ 2 )e − β M ] = − β e − ( β − R ) M
Solution 9.31
e 0.5 − (1 − 20 R)(13
. e 0.5 − 14
. ) = e10 R
0.90538 + 14.8668 R = 1 + 10 R + 50 R 2
ie −0.09462 + 4.8668 R − 50 R 2 = 0
Solving this using the quadratic formula (taking the smallest positive root) gives:
Since there is a smaller loading factor, the premiums will be reduced even though
claims remain the same. Hence, the probability of ultimate ruin will increase.
The mean of the distribution has decreased from 600 to 300, and the variance has
decreased from 2,400 to 600. Therefore the claims are smaller on average and less
uncertain. Both of these factors will decrease the probability of ultimate ruin.
The Poisson parameter has increased so claims occur more often (but their size is
unchanged). However, the premium received will also increase proportionally (as
c = (1 + q )l m1 ). Hence, the timing at which ruin may occur will be earlier, but not the
probability of it occurring in the first place. Therefore, the probability of ultimate ruin
will be unchanged.
P [U (2) < 0] = P [106.2 - S (2) < 0] = P [ S (2) > 106.2] = 1 - P [ S (2) £ 106.2]
Considering P [ S (2) £ 106.2] , and remembering that N (2) ~ Poi (2 ¥ 0.2) we have the
following scenarios where the aggregate claim amount by time 2 is less than 106.2:
amount of
number of claims probability
claim(s)
0 claims 0 e -0.4 = 0.67032
50 0.4e -0.4 ¥ 0.7 = 0.18769
1 claim
100 0.4e -0.4 ¥ 0.3 = 0.08044
2 claims 50, 50 0.42 e -0.4 ¥ 0.7 2 = 0.02628
2
5 Exam-style questions
We conclude this chapter with two exam-style questions. The first is a short question
on exponential families. The second is a longer question involving MLEs.
Ê yq - b(q ) ˆ
f ( y ) = exp Á + c( y , f )˜
Ë a (f ) ¯
State the mean and variance of Y in terms of b(q ) and its derivatives and a(f ) .
[1]
(ii)(a) Show that an exponentially distributed random variable with mean μ has a
density that can be written in the above form.
(ii)(b) Determine the natural parameter and the variance function. [3]
[Total 4]
Exam-style question 2
An insurer wishes to use a generalised linear model to analyse the claim numbers on its
motor portfolio. It has collected the following data on claim numbers yi ,
i = 1, 2, ..., 35 from three different classes of policy:
Class I 1 2 0 2 1 0 0 2 2 1
Class II 1 0 1 1 0
Class III 0 0 0 0 0 1 0 1 0 0
1 0 1 0 0 0 0 0 0 0
10 15 35
 yi = 11  yi = 3  yi = 4
i =1 i =11 i =16
(i) Show that the Poisson distribution is a member of the exponential family of
distributions. [2]
Ïa i = 1, 2, ..., 10
Ô
log mi = Ì b i = 11, 12, ..., 15
Ôg i = 16, 17, ..., 35
Ó
where mi is the mean of the relevant Poisson distribution. Derive the likelihood
function for this model, and hence find the maximum likelihood estimates for
a , b and g . [4]
(iii) The insurer now analyses the simpler model log mi = a , i = 1, 2, ..., 35 , for all
policies. Find the maximum likelihood estimate for a under this model
(Model B). [2]
(iv) Show that the scaled deviance for Model A is 24.93, and find the scaled
deviance for Model B. [5]
Solution 10.18
yi - mi
If Yi ~ N ( mi ,s 2 ) , then the Pearson residuals are .
s
n
( yi - mi ) 2 n
From page 27, the scaled deviance is  s2
= Â di . The deviance residuals are
i =1 i =1
given by:
yi - mi yi - mi
sign( yi - mi ) di = sign( yi - mi ) =
s s
Hence the Pearson residuals and the deviance residuals are the same.
We have:
1 Ï y¸
f ( y) = exp Ì - ˝
m Ó m˛
Ï 1 y¸
f ( y ) = exp Ìln - ˝
Ó m m˛
Comparing this to the standard form given in part (i), we can define:
1 1
q =- , a(f ) = 1, b(q ) = - ln = - ln( -q ), c( y, f ) = 0
m m
1
-
m
1 1
b′(θ ) = − b′′(θ ) = = μ2
θ θ 2
f ( y) = e- m m y / y !
È yq - b(q ) ˘
g ( y ) = exp Í + c( y, f ) ˙
Î a(f ) ˚
È y log m - m ˘
f ( y ) = exp Í - log y !˙
Î 1 ˚
Using the rearranged form for the Poisson distribution from part (i), we see that the log
of the likelihood function can be written:
10 15 35 35
log L = a  yi + b  yi + g  yi - 10e a
- 5e - 20e - Â log yi !
b g
i =1 i =11 i =16 i =1
35
= 11a + 3b + 4g - 10ea - 5e b - 20eg - Â log yi ! (**)
i =1
∂
log L = 11 - 10ea
∂a
∂
log L = 3 - 5e b
∂b
and:
∂
log L = 4 - 20eg
∂g
aˆ = log1.1 = 0.09531
35 35 35
log L = a  yi - 35ea -  log yi ! = 18a - 35ea -  log yi ! (***)
i =1 i =1 i =1
Differentiating this with respect to a , and setting the result equal to zero, we find that:
Ê 18 ˆ
18 - 35ea = 0
ˆ
fi aˆ = log Á = -0.66498
Ë 35 ˜¯
SD = 2(log LS - log LA )
where log LS is the value of the log likelihood function for the saturated model, and
log LA is the value of the log-likelihood function for Model A.
For the saturated model, we replace the mi ’s with the yi ’s in Equation (*) – as it fits
the observed data perfectly – so the expected results are the observed results. So:
We use the hint in the question here. yi log yi is zero when y = 0 , and also when
y = 1 . So the only contribution to the first term is when y = 2 , giving 4 lots of 2 log 2 .
For the log likelihood for Model A, we replace the parameters a , b and g with their
estimates â , bˆ and gˆ in Equation (**):
35
ˆ
log LA = 11aˆ + 3bˆ + 4gˆ - 10ea - 5e b - 20eg - Â log yi !
ˆ ˆ
i =1
= -27.6944
The corresponding value for log LA without the final term is –24.9218.
as required.
35
log LB = 18aˆ - 35ea - Â log yi !
ˆ
i =1
Ê 18 ˆ
= 18log Á ˜ - 18 - 4 log 2 = -32.7422
Ë 35 ¯
The scaled deviance is again twice the difference in the log likelihoods:
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.
Relative error
tˆ
za / 2 <e
|q | n
Of course the exact value of q is unknown, but the estimate qˆ may be used
instead of q in this equation.
6 Exam-style questions
f ( x) = q e -q x [2]
e -q x
g ( x) = k (q ) , x>0
1+ x
Exam-style question 2
An insurer is using the random variable X to model its claim amounts. The probability
density function of X is:
bx
f ( x) = x≥0
(500 + x)5
where b is a constant.
(ii) Show that this distribution has mean 500, and variance 500,000. [2]
(iv) Comment on the ease with which we could use the inversion method to generate
random numbers from this distribution. [1]
For generating values from the 2-parameter Pareto: 0.74 0.82 0.16
For making the acceptance-rejection decision: 0.26 0.59 0.71
This method generates random numbers on the interval [0,1) rather than [0,1] . (When
writing ranges, mathematicians use square brackets to indicate that the endpoint is
included in the range and round brackets to indicate that it is not. So these ranges
correspond to 0 £ < 1 and 0 £ £ 1 .)
In fact, it is clear that it can only give numbers of the form i m where
i = 0,1, 2,… , m - 1 . In practice this doesn’t matter, assuming m is large enough, so we
won’t distinguish between the two in this chapter.
Example 14.9
Solution 14.4
We have:
3
f ( x ) 32 (1 - x )( x - 5) 1
g ( x) = = = (1 - x )( x - 5)
Ch ( x ) 3 1
¥ 4
2 4
We only accept a value x if u £ g ( x ) . In this case therefore (using the second members
of the pairs of random numbers) we have:
Solution 14.5
You could have used the two initial values the other way round in which case you would
get different answers.
-2 log 0.374816
z1 = ¥ 0.604 = 1.382
0.374816
-2 log 0.374816
and: z2 = ¥ -0.1 = -0.229
0.374816
If we want values from a N (3,8) distribution, then we need to find values x1 and x2
where:
x = 3+ z 8
Solution 14.6
za2 2tˆ 2
We need n > 2
with za2 2 = 1.96 , e = 100, 000 and tˆ = 107 . Therefore:
e
1.962 ¥ 1014
n> = 38, 416
1010
The distribution function for this exponential distribution is (from the Tables):
F ( x) = 1 - e -q x
1
U = F ( x) = 1 - e -q X fi X =- log(1 - U )
q
(ii)(a) Procedure
We have seen in part (i) how to generate values from the distribution with PDF f ( x) .
The PDF of g ( x) is similar, so we can apply the acceptance-rejection method.
g ( x) k (q ) 1 k (q ) 1 k (q )
C = sup = sup = sup =
x>0 f ( x) x > 0 q 1 + x q x > 01+ x q
g ( x) 1
So we can use the function = to decide whether to accept or reject each of
Cf ( x) 1 + x
the values generated from the exponential distribution in part (i). We accept a particular
1
value x with probability .
1+ x
1
3. If u < , accept the value x ; otherwise, reject it.
1+ x
1
The probability of accepting each exponential value is . So the expected number of
C
k (q )
exponential values required until we accept one is C , which equals .
q
For each cycle of the algorithm we also need a pseudorandom number for step 2. So in
fact, the total expected number of pseudorandom numbers required is twice this
2k (q )
number, ie .
q
We will use the Tables. We see that this is an example of the three-parameter Pareto
distribution (see Page 15 of the Tables). Here we have:
k = 2 , a = 3 and l = 500
So the value of b is given by the constant term in the PDF given in the Tables:
G (a + k ) l a G (5) 5003
b= =
G (a ) G (k ) G (3) G (2)
Using the fact that G (n) = (n - 1)! from Page 5 of the Tables, we find that:
4! 5003
b= = 1.5 ¥ 109
2! 1!
Again we will use the formulae in the Tables. Using the parameter values given above:
kl 2 ¥ 500
E( X ) = = = 500
a -1 2
k (k + a - 1)l 2 2 ¥ 4 ¥ 5002
and: var( X ) = = = 500, 000
(a - 1) 2 (a - 2) 22 ¥ 1
Again, integration will work here, but using the Tables will be quicker. Also, we can
see from the number of marks on offer that the examiners are not expecting quantities of
heavy algebra. So go for the quick approach.
x x
bt
F ( x) = P( X £ x) = Ú f (t )dt = Ú dt
0 0
(500 + t )5
Integrating by parts, using the formula on page 5 of the Tables with u = b t and
dv
= (500 + t ) -5 :
dt
x x x
bt È bt -4 ˘ b
F ( x) = Ú 5
dt = Í - (500 + t ) ˙ + Ú (500 + t ) -4 dt
0
(500 + t ) Î 4 ˚0 4 0
x
bx -4 bÈ 1 -3 ˘
=- (500 + x ) + Í - (500 + t ) ˙
4 4Î 3 ˚0
b bx b
= 3
- 4
-
12 ¥ 500 4(500 + x) 12(500 + x)3
bx b
= 1- 4
-
4(500 + x) 12(500 + x)3
You do not have to integrate by parts here if you prefer substitution. Substituting
u = 500 + t will also work.
x u =500 + x 500 + x
bt b (u - 500) b 500 b
Ú (500 + t )5 dt = Ú u 5
du = Ú u 4
-
u5
du
0 u =500 500
500 + x
È b 500 b ˘
= Í- 3 + ˙
Î 3u 4u 4 ˚500
Ê b 500 b ˆ Ê b 500 b ˆ
= Á- + - -
4˜ Á
+ ˜
Ë 3(500 + x) 3
4(500 + x) ¯ Ë 3 ¥ 5003
4 ¥ 5004 ¯
b 500 b
= 1- 3
+
3(500 + x) 4(500 + x) 4
Note to markers – this expression looks different from (but is equivalent to) the
expression obtained using integration by parts. There are a number of slightly different
looking but equally acceptable expressions that students might have here:
bx b
F ( x) = 1 - 4
-
4(500 + x) 12(500 + x)3
b 500 b
or: F ( x) = 1 - 3
+
3(500 + x) 4(500 + x) 4
È 4 b x + 500 b ˘
or: F ( x) = 1 - Í 4˙
Î 12(500 + x) ˚
All of these should be given full credit. There may be other equivalent expressions too
(particularly if students have not kept b for the original constant).
In order to use the inversion method to generate random values from this distribution,
we will need to invert the distribution function. This will not be easy to do. We would
need some form of computer assistance to solve the resulting equations.
First we want the probability density function of the two parameter Pareto distribution:
3 ¥ 5003
h( x ) =
(500 + x) 4
f ( x) ÏÔ bx (500 + x) 4 Ô¸ Ê 4x ˆ
C = max = max Ì ¥ ˝ = max Á =4
h( x ) ÔÓ (500 + x)
5 3
3 ¥ 500 Ô˛ Ë 500 + x ˜¯
f ( x) x
and: g ( x) = =
4h( x) 500 + x
We can now simulate the required values. We will use 0.74 to simulate a value from
our two-parameter Pareto distribution, and then use the random number 0.26 to decide
whether to accept or reject it.
3
Ê 500 ˆ
u ( x) = 1 - Á ˜ fi x = 500 ÈÎ (1 - u ) -1/ 3 - 1˘˚
Ë 500 + x ¯
But u3 = 0.71 > g (29.92) = 0.0565 , so we again reject this random number.
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.
The number of claims from one group of drivers in a year has a Poisson distribution
with mean l , and the number of claims from a second group of drivers has a Poisson
distribution with mean 2l . In one year, there are n1 claims from group 1 and n2
claims from group 2.
(ii) Suppose that past experience shows that l has an exponential distribution with
1
mean .
n
(b) Show that the Bayesian estimate of l under quadratic loss may be
written in the form of a credibility estimate combining the prior mean of
l with the maximum likelihood estimate l̂ in (i). State the credibility
factor. [4]
[Total 7]
Claim amounts X are believed to follow a gamma distribution with parameters a and
l , where a is a known constant. An insurer wishes to carry out a Bayesian analysis,
using a gamma prior distribution for l with parameters b and d .
(i) A random sample from the distribution gives the individual sample values
x1 , x2 ,… , xn . Show that the posterior distribution for l is also a gamma
distribution, and write down its parameters in terms of a , b , d and the sample
data. [4]
(ii) Write down the formula for the Bayesian estimate for l under squared error
loss. [1]
(iii) Show that this formula can be written in the form of a credibility estimate:
lˆ = Z l + (1 - Z ) m
where l is the maximum likelihood estimator for l calculated from the sample
data, and m is the mean of the prior distribution. Write down the formula for
the credibility factor Z . [3]
(iv) Explain how the value of Z changes if the sample size increases, and explain
why it is reasonable to expect Z to change in this way. [1]
[Total 9]
Chapter 6
Show that, in the usual notation for Empirical Bayes Credibility Model 1:
(i) E ( X ) = E[m(q )]
(ii) E ÈÎ X m(q ) ˘˚ = E ÈÎ m 2 (q ) ˘˚
1 È 2 ˘
(iii) var( X ) = E s (q ) ˚ + var [m(q )] [6]
n Î
(ii) The table below gives values of X ij for each of five risks in the collective over
each of the past five years, together with some summary statistics. Use this set
of data to calculate the empirical Bayes credibility premium for risk number 1
for the coming year.
Year j
1 5
1 2 3 4 5 Xi  ( X ij - X i )2
4 j =1
1 85 73 91 88 70 81.4 87.3
2 62 107 104 71 68 82.4 456.3
Risk i 3 122 144 99 80 150 119.0 879.0
4 72 78 90 88 98 85.2 105.2
5 170 81 153 140 175 143.8 1425.7
[8]
[Total 14]
An insurer uses Empirical Bayes Credibility Model 1 to calculate the annual premiums
for a collective of risks. The insurer has data for this collective from the last five years
and estimates E[ s 2 (q )] to be 620 and var[m(q )] to be 145. Calculate the credibility
factor Z for the coming year. [2]
The following table gives the aggregate amounts paid out (in millions of pounds) by
four companies under a certain type of fire insurance over a period of five years,
together with some summary statistics for Empirical Bayes Credibility Theory Model 1:
Year 1 n
Insurer
1 2 3 4 5 Xi  ( X ij - X i )2
4 j =1
1 37 43 44 50 53 45.4 39.3
2 12 17 22 23 30 20.8 45.7
3 53 58 60 59 57 57.4 7.3
4 82 81 85 89 93
Complete the table by calculating the two missing entries, and use these data to find
estimates for E[m(q )] , E[ s 2 (q )] and var[m(q )] . Hence estimate the empirical Bayes
credibility premium for each of the insurers for the coming year. [9]
6 10
ÂÂ Pij = 1, 498 P* = 18.24
i =1 j =1
He has just received the following information relating to a 7th insurer (Insurer I), and
he wishes to update his estimates using the past ten years of claims data for Insurer I
given in the table below:
Year i 1 2 3 4 5
Aggregate
Yi 100 85 90 102 109
claims
Volume Pi 22 24 26 20 25
Year i 6 7 8 9 10
Aggregate
Yi 106 128 132 150 131
claims
Volume Pi 30 29 35 40 36
(i) Calculate his updated estimates for E [m(q )] , E ÈÎ s 2 (q ) ˘˚ and var [m(q )] , and
hence find the credibility premium for Insurer I for the coming year, given that
Insurer I is expected to have a volume figure for the coming year of 38. [20]
(ii) The student also needs a credibility estimate for Insurer K, one of the six
insurers included in the original analysis. He knows that, for Insurer K,
 YKj = 986 and  PKj = 327 . Explain whether the credibility premium for
Insurer K (based on the full analysis of the seven insurers) will be greater or less
than the corresponding figure for Insurer I (per unit of risk volume), and give
reasons for your answer. [3]
[Total 23]
Chapter 7
(i) State the two conditions that must hold for a risk to be insurable. [2]
(ii) List five other risk criteria that would be considered desirable by a general
insurer. [5]
[Total 7]
(i) The random variable X has a gamma distribution with parameters a = 3 and
l = 2 . Y is a related variable with conditional mean and variance of:
E (Y | X = x) = 3 x + 1 var(Y | X = x) = 2 x 2 + 5
(ii) The random variable V has a Poisson distribution with mean 5. For a given
value of V , the random variable U is distributed as follows:
U | (V = v) ~ U (0, v)
If the individual claim size distribution X is exponential with a mean of 2, what are the
values of E ( S ) and var( S ) ? [5]
Solution 2.6
E ( X ) = E ( X j ) = E[ E ( X j q )] = E[m(q )] [2]
( ) ( )
E ÈÎ Xm(q ) ˘˚ = E E ÈÎ Xm(q ) q ˘˚ = E m(q ) E ÈÎ X q ˘˚ = E ÈÎ m 2 (q ) ˘˚ [2]
È1 ˘ 1
= E Í var( X j q ) ˙ + var[ E ( X j q )] = E ÈÎ s 2 (q ) ˘˚ + var [ m(q )]
În ˚ n
[2]
Solution 2.7
(i) Assumptions
(a) when i = k :
(b) when i π k :
If q i and q k are the values of the risk parameters associated with the two risks,
( X ij ,q i ) and ( X km, q k ) are independent and identically distributed. [2]
1
X = (81.4 + 82.4 + + 143.8) = 102.36 [1]
5
1 5 È1 5 ˘ 1
Â Í Â ( X ij - X i ) 2 ˙ = (87.3 +
5 i =1 Í 4 j =1 ˙˚ 5
+ 1425.7) = 590.7 [1]
Î
Thirdly we calculate the sum of the squared deviations between the individual and
overall sample means:
1 5 1È 5 ˘
 i
4 j =1
( X - X ) 2
= Í Â i
4 Í j =1
X 2
- 5 ¥ X 2
˙
˙˚
Î
1
= (81.42 + + 143.82 - 5 ¥ 102.362 ) = 781.59 [2]
4
n 5
Z= = = 0.84885 [1]
2
E[ s (q )] 590.70
n+ 5+
var[m(q )] 663.45
So the premium for the coming year for the first risk is:
Solution 2.8
n
Z=
n + E[ s 2 (q )] / var[m(q )]
we get:
5
Z= = 0.539 [2]
5 + 620 /145
Solution 2.9
430
Xi = = 86 [1]
5
1 1 1
and S ( X ij - X i ) 2 = ÈÎ S X ij2 - 5 X i2 ˘˚ = (37, 080 - 36,980) = 25 [1]
4 4 4
45.4 + + 86
E[m(q )] = X = = 52.4 [1]
4
1 N 1 n
E[ s 2 (q )] = Â Â
N i =1 n - 1 j =1
( X ij - X i ) 2
1
= (39.3 + + 25) = 29.325 [1]
4
N N n
1 1 1
var[m(q )] =
N - 1 i =1
 ( X i - X )2 - Â
Nn n - 1
 ( X ij - X i ) 2
i =1 j =1
1 1
= (45.42 + + 862 - 4 ¥ 52.42 ) - ¥ 29.325 = 727.975 [2]
3 5
n 5
Z= = = 0.9920 [1]
Ê E( s 2 (q )) ˆ Ê 29.325 ˆ
Á n + ÁË 5 + ˜
Ë Var(m(q )) ˜¯ 727.975 ¯
Solution 2.10
6 10
  Pij X ij 6 10
i =1 j =1
4.00 = 6 10
fi  Pij X ij = 4 ¥ 1, 498 = 5,992 [1]
  Pij i =1 j =1
i =1 j =1
10 10 10
 YIj =  PIj X Ij = 1,133 and  PIj = 287 [1]
j =1 j =1 j =1
7 10
  Pij X ij
i =1 j =1 5,992 + 1,133
X = = = 3.9916 [1]
7 10 1, 498 + 287
  Pij
i =1 j =1
1 N ÏÔ 1 n ¸
2Ô
Â Ì Â ij ij i ˝
N i =1 Ô n - 1 j =1
P ( X - X ) [1]
Ó ˛Ô
6 10
  Pij ( X ij - X i )2 = 62.8 ¥ 9 ¥ 6 = 3,391.2 [1]
i =1 j =1
N n
We now need to add on the contribution to   Pij ( X ij - X i )2 from Insurer I. This
i =1 j =1
10
is  PIj ( X Ij - X I )2 . We can calculate this by rewriting it as:
j =1
10 10 10 10
 PIj ( X Ij - X I )2 =  PIj X Ij2 - 2 X I  PIj X Ij + X I2  PIj
j =1 j =1 j =1 j =1
10
 PIj X Ij
j =1
But since X I = 10
, the last two terms can be added together to give:
 PIj
j =1
10 10 10
 PIj ( X Ij - X I )2 =  PIj X Ij2 - X I2  PIj
j =1 j =1 j =1
Ê Ê 100 ˆ
2
Ê 131ˆ ˆ Ê 1,133 ˆ
2 2
= Á 22 ¥ Á + + 36 ¥ Á - ¥ 287
Ë Ë 22 ˜¯ Ë 36 ˜¯ ˜¯ ÁË 287 ˜¯
7 10
So the new total for   Pij ( X ij - X i )2 is:
i =1 j =1
1 1
¥ ¥ 3, 457.5034 = 54.881 [1]
7 9
We now want the estimate for var [m(q )] . The formula for the estimator for var [m(q )]
is:
1 Ê 1 N n
1 N 1 n ˆ
Á Â
P * Ë Nn - 1 i =1
 Pij ( X ij - X )2 -   ij ij i
N i =1 n - 1 j =1
P ( X - X ) 2
˜ [1]
j =1 ¯
1 Ê 1 6 10 ˆ
42.1 = Á Â Â Pij ( X ij - X )2 - 62.8˜
18.24 Ë 59 i =1 j =1 ¯
6 10 6 10 6 10
  Pij ( X ij - X ) 2
=Â Â Pij X ij2 -X 2
  Pij
i =1 j =1 i =1 j =1 i =1 j =1
6 10
  Pij X ij2 = 49, 011.536 + 42 ¥ 1, 498 = 72,979.536 [1]
i =1 j =1
We can now get our new estimate for this statistic by adding on the contribution from
Insurer I:
7 10
  Pij X ij2 = 72,979.536 + 4,539.0874 = 77,518.6234 [1]
i =1 j =1
So:
7 10
  Pij ( X ij - X )2 = 77,518.6234 - 3.99162 ¥ (1, 498 + 287)
i =1 j =1
We now need our new value for P * . Using the formula for P * , we know that:
1 N Ê Pi ˆ 1 ÊN N
Pi 2 ˆ
P* =  i P ˜¯ Nn - 1 Á  i  P ˜
P
Nn - 1 i =1 ÁË
1 - = P -
Ë i =1 i =1 ¯
6
Now we know that  Pi = P = 1, 498 , so:
i =1
Ê 6 ˆ
Á
1 Á
 Pi 2 ˜
18.24 = 1, 498 - i =1 ˜
59 Á 1, 498 ˜
ÁÁ ˜˜
Ë ¯
6
This gives a value for  Pi 2 of 631,916.32.
i =1
7
So  Pi 2 is equal to 631,916.32 + 287 2 = 714, 285.32 . We can now find our new P * :
i =1
Ê 7
2ˆ
Á
1 Á 7
 Pi ˜
˜ = 1 Ê (1, 498 + 287) - 714, 285.32 ˆ = 20.07015
P* = Â
69 Á i =1
Pi - i =1
P ˜ 69 ËÁ (1, 498 + 287) ¯˜
[2]
ÁÁ ˜˜
Ë ¯
1 Ê 1 ˆ
ÁË ( 49, 078.449) - 54.881˜¯ = 32.70533 [1]
20.07015 69
E [m(q )] = 3.9916
E ÈÎ s 2 (q ) ˘˚ = 54.881
var [m(q )] = 32.7053
We can now find the EBCT premium for the coming year for Insurer I.
287
ZI = = 0.994187 [1]
54.881
287 +
32.705
1,133
CP = 0.994187 ¥ + 0.005813 ¥ 3.9916 = 3.94799 [1]
287
(ii) Insurer K
986
The estimate for X K is = 3.0153 , which is lower than the corresponding estimate
327
for X I , and is also lower than X . The credibility factor for Insurer K is:
327
ZK = = 0.99489
E ÈÎ s 2 (q ) ˘˚
327 +
var [m(q )]
which is bigger than the corresponding factor for Insurer I. So we will be placing more
emphasis on the individual mean for Insurer K, and this will pull the estimate down
more. So the credibility premium per unit of volume will be lower for Insurer K than
for Insurer I. [3]
Individual claim amounts in a Poisson claims process with a frequency of 50 claims per
year for the whole portfolio have mean £5,000 and standard deviation £2,500. If the
annual premium rate is £300,000, calculate the tightest upper bound for the adjustment
coefficient. [2]
An insurer calculates the annual premiums for fire insurance of flats by increasing the
risk premium by 30% and adding a £30 loading. The claim frequency is 3% and
individual claim amounts can be assumed to be:
● £2,000 with probability 0.9
● £15,000 with probability 0.1.
Calculate the insurer’s adjustment coefficient for these policies, to 2 significant figures.
[8]
A Poisson claims process has security loading q = 2 / 5 and claim size density function:
f ( x) = 3
2
e -3 x + 72 e -7 x , x > 0
(i) Derive the moment generating function (MGF) for the claim size distribution,
and state the values of t for which it is valid. [3]
(i)(a) Show that the adjustment coefficient for a compound Poisson claims process
satisfies the inequality:
2[c / l - E ( X )]
r<
E( X 2 )
(i)(b) Explain how this inequality provides information about the probability of
ultimate ruin for the process. [10]
(ii) An insurer considers that claims of a certain type occur in accordance with a
compound Poisson process. The claim frequency for the whole portfolio is 100
per annum and individual claims have an exponential distribution with a mean of
£8,000.
(a) Calculate the adjustment coefficient if the total premium rate for the
portfolio is £1,000,000 per annum.
(b) Verify that the value calculated in (ii)(a) satisfies the inequality in (i)(a).
(c) The insurer decides to take out excess of loss reinsurance for this
portfolio. The reinsurer has agreed to pay the excess of any individual
claim above £20,000 in return for an annual premium of £80,000.
Calculate the adjustment coefficient for the direct insurer when the
reinsurance is in operation.
(d) Estimate the direct insurer’s probability of ultimate ruin with and without
the reinsurance arrangement, assuming that the initial surplus is £20,000
and that future premiums remain at the same level.
l + cr = l M X (r )
(i) Using the above equation for R , or otherwise, show that, provided R is small,
an approximation to R is R̂ , where:
2(c / l - m )
Rˆ =
s 2 + m2
(ii) Describe how the adjustment coefficient can be used to assess reinsurance
arrangements on the basis of security. [3]
(iii) The Poisson parameter, l , for this portfolio is 20 and all individual claims are
for a fixed amount of £5,000. The insurer’s premium loading factor, q , is 0.15
and proportional reinsurance can be purchased from a reinsurer who calculates
premiums using a loading factor of 0.25.
Calculate the maximum proportion of each claim that could be reinsured so that
the insurer’s security, measured by R̂ , is greater than the insurer’s security
without reinsurance. [9]
[Total 16]
Chapter 10
In the context of Generalised Linear Models, consider the exponential distribution with
density function f ( x ) , where:
1
f ( x) = e− x/ μ ( x > 0) .
μ
1
(ii) Show that the canonical parameter, θ , is given by θ = − . [1]
μ
(iii) Determine the variance function and the dispersion parameter. [3]
[Total 5]
The random variable Zi has a binomial distribution with parameters n and mi , where
0 < mi < 1 . A second random variable, Yi , is defined as Yi = Zi / n .
(i) Show that Yi is a member of the exponential family, stating clearly the natural
and scale parameters and their functions a(j ) , b(q ) and c( y,j ) . [4]
S - E (S )
But is approximately standard normal. So we must have:
var( S )
2
So: i = 10499
. − 1 = 01023
. ie the required interest rate is 10.23% per annum.
[3]
Solution 3.8
Since we don’t know the precise distribution of the individual claim amounts, this is the
best we can do.
Solution 3.9
l + cr = l E (e rX )
5.29 - 3.3
So: Rª = 0.1525 [1]
13.05
So R lies between 0.094 and 0.0945. So the value of the adjustment coefficient correct
to 2SF is 0.000094 (in units of £ -1 ). [1]
[Total 8]
Solution 3.10
• • •
E (etX ) = Ú etx ÈÎ 32 e -3 x + 72 e -7 x ˘˚ dx = 3
2
(t -3) x (t - 7) x
Ú e dx + 72 Ú e dx
0 0 0
• •
3 È e(t -3) x ˘ 7 È e(t -7) x ˘ 3 7
= Í ˙ + Í ˙ = + [2]
2 ÍÎ t - 3 ˙˚ 2 ÍÎ t - 7 ˙˚ 2(3 - t ) 2(7 - t )
0 0
The first integral converges if t < 3 and the second if t < 7 . So the MGF is valid for
values of t < 3 . [1]
[Total 3]
1 + (1 + q ) E ( X )r = M X (r )
M X (t ) = 32 (3 - t ) -1 + 72 (7 - t ) -1
M X¢ (t ) = 32 (3 - t ) -2 + 72 (7 - t ) -2
3 1 5
M X¢ (0) = + = [1]
18 14 21
7 5r 3 7 3(7 - r ) + 7(3 - r ) 21 - 5r
1+ ¥ = + = = [1]
5 21 2(3 - r ) 2(7 - r ) 2(3 - r )(7 - r ) (3 - r )(7 - r )
Multiplying through by 3:
63 - 15r
3+ r =
(3 - r )(7 - r )
63 - 9r - 7 r 2 + r 3 = 63 - 15r
6r - 7 r 2 + r 3 = 0 [1]
Factorising:
r (6 - r )(1 - r ) = 0
which gives r = 1 as the positive solution that satisfies the inequality t < 3 . So the
value of the adjustment coefficient in this case is 1. [1]
[Total 4]
Solution 3.11
l + cr = l E (erX ) [1]
The expected claim frequency l is the expected number of claims occurring per unit of
time. [1]
The premium rate c is the constant amount of premium actually received per unit of
time. [1]
The claim size X is a random variable representing the amount of an individual claim.
[1]
Expanding the RHS of the equation defining the adjustment coefficient gives:
r2
l + cr = l E (erX ) = l [1 + r E ( X ) +
E ( X 2 ) + "] [1]
2
Since the individual claim sizes X take positive values, the terms on the RHS are all
positive. So, ignoring terms in powers higher than X 2 gives:
r2
l + cr > l [1 + r E ( X ) + E ( X 2 )] [1]
2
r2
c r > l [r E ( X ) + E ( X 2 )] [1]
2
r
c > l [E( X ) + E ( X 2 )] [1]
2
2[c / l - E ( X )]
r< [1]
E( X 2 )
The inequality in (i)(a) gives an upper limit for the adjustment coefficient. The
probability of ultimate ruin starting with a surplus of u is y (u ) ª e - ru . Combining
these results gives an approximate lower limit for the probability of ultimate ruin. [1]
[Total 10]
l + cr = l E (erX )
Dividing by 100:
The adjustment coefficient is the smallest positive solution ie r = 1/ 40, 000 = 0.000025
[1]
E ( X ) = 8, 000
and:
Since the direct insurer will still have to pay a part of every claim, the claim frequency
l is unchanged.
The net rate of premium income (after paying the reinsurance costs) for the direct
insurer is:
The MGF for the net claim amount X net paid by the direct insurer is:
20,000 •
E (erX net ) = Ú e rx e - x / 8000 / 8, 000 dx + Ú e 20,000 r e - x / 8000 / 8, 000 dx
0 20,000
20,000 •
1 - x (1/ 8000 - r ) e 20,000 r
= Ú e dx + Ú e - x / 8000 dx
8, 000 0
8, 000 20,000
1 - e -20,000(1/ 8000- r )
= + e 20,000 r e -20,000 / 8,000
8, 000 (1/ 8, 000 - r )
1 - e 20,000 r - 2.5
= + e 20,000 r - 2.5 [2]
1 - 8, 000r
È 1 - e 20,000r - 2.5 ˘
100 + 920, 000r = 100 Í + e 20,000 r - 2.5 ˙ [1]
ÍÎ 1 - 8, 000r ˙˚
Dividing by 100r :
Evaluating the function f (r ) = 80e 20,000 r - 2.5 - 736, 000r + 12 for trial values of r , we
find: [1]
f (0.000025) = 4.43
f (0.00003) = 1.89
f (0.00004) = -2.83
f (0.000033) = 0.42
f (0.000034) = -0.06
So the probability of ruin with and without the reinsurance are approximately:
(ii)(e) Comment
Solution 3.12
(i) Approximation to R
tX (tX ) 2 t2
M X (t ) = E (e ) = E (1 + tX + + ") = 1 + t E ( X ) + E ( X 2 ) + "
2! 2!
R2
So: M X ( R) = 1 + R E ( X ) + E ( X 2 ) +" [1]
2!
Assuming that R is small enough for terms in R3 and higher powers to be neglected,
we can write:
R2 2
M X ( R) ª 1 + R m + (m + s 2 ) [1]
2!
R2 2
l + cR = l[1 + R m + ( m + s 2 )]
2!
l R2
(c - lm ) R = (m 2 + s 2 ) [1]
2!
2(c - lm ) 2(c / l - m )
R= = [1]
l(m 2 + s 2 ) s 2 + m2
Note however that the adjustment coefficient cannot tell us anything about the relative
profitability of different reinsurance arrangements. This will need to be assessed using
other means.
First we consider the insurer’s security without reinsurance. The equation for the
adjustment coefficient is:
1 + (1 + q )m1r = M X (r )
So, using the approximation derived for the adjustment coefficient in part (i), we have:
2(115000 / 20 - 5000)
Rª 2
= 6 ¥ 10 -5 [1]
5000
Now assume that a proportion k of each risk is reinsured. The net premium income is
now:
[1.15 ¥ 5000 - 1.25 ¥ k ¥ 5000] l = (5750 - 6250k ) l [1]
M X (t ) = e5000(1- k ) R [1]
If we want the insurer’s security with reinsurance to be greater than without reinsurance,
we want the adjustment coefficient with reinsurance to be larger, ie:
3 - 5k
2
> 6 ¥ 10 -5 [1]
50000(1 - k )
3k 2 - k < 0 [1]
Solution X1.5
The total profit (in £000s) corresponding to each level of client-days is:
q1 q2 q3
d1 1,360 1,520 1,760
d2 1,407 1,541 1,742
d3 1,250 1,350 1,500
[1]
Notice that strategy d3 is dominated by the two other strategies so we can discard it.
The expected total profit (in £000s) for each level of client-days is:
E (profit under d1 ) = (1,360 ¥ 0.1) + (1,520 ¥ 0.6) + (1, 760 ¥ 0.3) = 1,576
E (profit under d 2 ) = (1, 407 ¥ 0.1) + (1,541 ¥ 0.6) + (1, 742 ¥ 0.3) = 1,587.9
The Bayes criterion solution selects the client-days with the highest expected profit.
Thus the solution is d 2 , with an expected profit of £1,587,900. [2]
Remember that strategy d3 is dominated by the two other strategies so we can discard
it. The worst possible outcomes associated with d1 and d 2 are:
1,360 1,407
The minimax solution is the best worst case scenario (ie maximises the minimum gain)
which is 1,407, so the minimax solution is d 2 . [1]
1,760 1,742
The maximax solution is the best best case scenario (ie maximises the maximum gain)
which is 1,760, so the maximax solution is d1 . [1]
Solution X1.6
Since the prior distribution for m is uninformative, we take the prior distribution to be
constant. Alternatively, we could say f ( m ) = 1 k where k Æ • . [½]
We now want to write this as the PDF of a normal distribution, with m as the variable.
So we want it to look like:
2
1 Ê m - m* ˆ 1
- Á - ( m 2 - 2 m* m + m*2 )
2 Ë s *2 ˜¯ 2s *2
f ( m | x) μ e =e (*)
So, expanding the bracket in our posterior and summing the terms (ignoring any
expressions that do not involve m , since these can be absorbed into the constant term):
 ((ln xi )2 - 2 m ln xi + m 2 )
1 1
 (ln xi - m )
2
- -
f ( m | x) μ e 2s 2 =e 2s 2
-
1
2s 2
( (ln x ) -2 m  ln x +nm )
i
2
i
2
=e [½]
-
1
2s 2
(-2 m  ln x +nm ) i
2
μe
m -2 m  i ˜
n Ê 2 ln x ˆ
- 2 ÁË
2s n ¯
=e [1]
Completing the square (the missing term is absorbed into the constant):
2
1 Ê Â ln xi ˆ
- ÁË m - n ˜¯
2(s 2 n)
f ( m | x) μ e [1]
Comparing this with equation (*) , we see that we have a normal distribution with
parameters:
m* =
 ln xi and s *2 =
s2
[½]
n n
Question X2.1
(i) In the context of Empirical Bayes Credibility Theory Model 1, the credibility
factor Z is given by:
n
Z=
E[ s 2 (q )]
n+
var[m(q )]
(ii) The table below shows the aggregate claim statistics for each of four risks over
three years:
3 3
Risk i  X ij  ( X ij - X i )2
j =1 j =1
1 2,184 22,344
2 2,721 20,294
3 3,450 21,800
4 3,099 23,994
Use EBCT Model 1 to calculate the credibility premium for the coming year for
Risk 3. [4]
(iii) Without carrying out any calculations, determine with reference to part (i) what
would happen to the credibility factor if a fifth risk was added with:
3 3
 X 5 j = 3,999 and  ( X 5 j - X 5 )2 = 21, 734 [2]
j =1 j =1
[Total 9]
Question X2.2
Claims occur according to a compound Poisson process at a rate of ¼ claim per year.
Individual claim amounts, X , have PF:
P( X = 50) = 0.8
P( X = 100) = 0.2
The insurer charges a premium at the beginning of each year using a 20% loading
factor. The insurer’s surplus at time t is U (t ) . Find P[U (2) < 0] if the insurer starts at
time 0 with a surplus of 100. [4]
Question X2.3
Question X2.4
On a portfolio of insurance policies, the claim size, Y is assumed to depend on the age
of the policyholder, X . Suppose that the conditional mean and variance of Y are:
E (Y | X = x) = 2 x + 400
x2
and var(Y | X = x) =
2
The distribution of X over the portfolio is assumed to be normal with mean 50 and
standard deviation 14.
Assignment X2 Solutions
Solution X2.1
(i) Effect on Z
credibility premium = Z ¥ X i + (1 - Z ) ¥ X
E[ s 2 (q )] is a measure of the overall variability within each of the different risks in the
group. As E[ s 2 (q )] increases, Z decreases. So more emphasis is put on the collateral
data. This makes sense, since the more variable the individual risk’s experience is, the
less reliable it is. So we would expect to rely more on the collateral data. [1]
var[m(q )] is a measure of the variability between the different risks in the group. As
var[m(q )] increases, Z increases. So more emphasis is put on the particular risk. This
makes sense, since the larger the variability between the different risks, the less relevant
the other risks are in assessing the premium of our particular risk. So we want to rely
more on the direct data. [1]
[Total 3]
3 4
E[ s 2 (q )] ª 1
4 Â Â ( X ij - X i )2
1
2
i =1 j =1
= 28,962.3 [1]
3
Z= = 0.88714
11, 054
3+
28,962.3
This risk has a much higher mean, so the variance of the means, var[m(q )] , will
increase. It has a similar variance to the other risks, so the mean variance of each risk,
E[ s 2 (q )] , will remain similar. [1]
Hence the proportionally larger var[m(q )] will lead to a larger Z, because more
emphasis will be put on the direct data. [1]
Question X3.6
The table below shows 4 models considered by the analyst and their scaled deviances
for the data set.
Parameterised form
Number of
Model of the linear Scaled deviance
parameters
predictor
SA a +bx 2 238.4
SA + PT 206.7
SA + PT + SA i PT 178.3
SA * PT + NB 166.2
SA * PT * NB a ij¢¢ + b ij¢¢x 120 58.9
(ii) On the basis of scaled deviance which model should the analyst choose? [5]
(iii) What further information should the analyst consider before making her
recommendation about an appropriate choice of model? [2]
[Total 10]
Question X3.7
Claims for a particular risk arrive in a Poisson process rate l . The claim sizes are
independent and identically distributed with density f ( x) and are independent of the
claims arrival process. Assume there is a constant g (0 < g < •) such that
lim M (r ) = • where M (r ) is the moment generating function of a claim. Premiums
r Æg
are received continuously at constant rate with premium loading factor q > 0 .
(b) Define the surplus process and the probability y (u ) of ruin with initial
surplus u > 0 .
x
-
1 m
(ii) Derive the adjustment coefficient if f ( x) = e , x > 0 , and q = 0.25 . [4]
m
(c) Compare your answers to (iii)(b) with those obtained if the claim sizes
are mistakenly assumed to be exponentially distributed with mean m ,
and comment briefly. [11]
[Total 18]
Question X3.8
Within each tier claim amounts are uniformly distributed. All policies are charged the
same annual premium, which assumes a claim frequency of 2% per annum and
incorporates a 20% security loading.
The company maintains a portfolio of 10,000 policies and currently has a surplus of
£1m in respect of this class of business.
(ii) Show that, if all amounts are expressed in units of £000s, the adjustment
coefficient r satisfies an equation of the form:
where the p ’s are constants whose values you should specify. [7]
(iii) By expanding the right hand side of the equation in (ii) as a series and discarding
terms containing r 4 or higher powers, find an approximate value for r . [4]
(iv) Given that the accurate value for r is 5.56 × 10 −3 , estimate the probability of
ultimate ruin. [2]
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.
2 2
We have to interpolate in the tables between c 90 and c100 to get the figure of 118.7.
Solution X3.7
The adjustment coefficient R is the smallest strictly positive solution of the equation:
1 + (1 + q ) m1 R = M ( R )
•
where m1 = Ú x f ( x) dx is the mean of the claim amount distribution. [1]
0
You may just prefer to remember the equation l + cR = l M X ( R ) , and the fact that
c = (1 + q )l m1 . Putting these together gives the equation quoted above.
U (t ) = u + (1 + q ) l m1 t - S (t )
The probability of ruin y (u ) is the probability that the surplus process becomes
negative at any point in the infinite time interval 0 < t < • .
y (u ) £ e - Ru
(ii) Derivation
M ( R) = (1 - m R) -1
1 + 1.25 m R = (1 - m R ) -1 [1]
1
R= [2]
5m
• • •
m = Ú x ½ e (1 + 2e ) dx = ½ Ú x e dx + ½ Ú x 2e -2 x dx
-x -x -x
0 0 0
Looking at these two integrals, we see that the first one represents the mean of an
Exp(1) distribution, and the second is the mean of an Exp (2) distribution. We know
that the mean of an Exp(l ) distribution is 1/ l , so:
m = ½ ¥1+ ½ ¥ ½ = ¾ [2]
To find the adjustment coefficient, we first need the MGF of this distribution. Using a
similar logic, we can express the MGF here as the average of the MGF’s of two
exponential distributions. We then use the formula for the MGF of an exponential
distribution from the Tables:
•
M X ( R ) = E (e RX
) = Ú e Rx ½ e - x (1 + 2e - x ) dx
0
• •
= ½Ú e Rx
e dx + ½ Ú e Rx 2e -2 x dx
-x
0 0
= ½ (1 - R) -1 + ½ (1 - R / 2) -1
[2]
We can now write down the equation for the adjustment coefficient:
1 + 1.25 ¥ ¾ R = ½ (1 - R) -1 + ½ (1 - R / 2) -1 [½]
4 - 3R
1 + 0.9375R =
(2 - 2 R)(2 - R )
So the adjustment coefficient is R = 0.2356 , and the upper bound for the probability of
ruin is:
1
R= = 0.26667 [½]
5 ¥ 3/ 4
In the exponential case the upper bound is lower, and the adjustment coefficient is
higher. These both suggest that the exponential case is more secure. This is not
surprising, since claims are likely to be less variable in the situation where a single
claim distribution is used than in the case where the claim distribution is a mixture of
two distributions. [1]
[Total 11]
Solution X3.8
(i)(a) Premium
FG 22,000 220,000 IJ
H
0.02 × 0.95 ×
2
+ 0.05 ×
2 K
= £319 [1½]
c = 12
. × 319 = £382.80 [½]
(i)(b) Surplus
The premium for each policy contains a loading of 0.2 × 319 = £63.80
During the next year the surplus will be expected to increase by the total amount of the
loadings for all policies in the portfolio, which will give:
λ + cr = λ M X (r )
We are told that λ = 0.02 and we have calculated that c = 0.3828 (in units of £000s).
M X (r ) = E (e rX ) = 0.95 z2
20 e rx
18
dx + 0.05 z
20
200 e rx
180
dx [2]
F
0.95 e 20r − e 2r I
0.05 e 200r − e 20r F I
M X (r ) =
18 GH r
+
180 JK r GH JK
[1]
1 RSFG 0.05IJ e 200r FG 0.95 − 0.05IJ e 20 r F 0.95IJ e UV
+ G− 2r
=
r TH 180 K +
H 18 180 K H 18 K W
So now let:
0.95
p1 = − = −0.05278
18
0.95 0.05
p2 = − = 0.0525
18 180
0.05
and p3 = = 2.778 × 10 −4
180
1
M X (r ) =
r
o
p1e 2r + p2 e 20r + p3e 200r t [2]
Using the numerical values for the p ’s, and summing down the columns, we find that:
As we would expect, r = 0 is a root of this equation. So, to find the nontrivial root, we
cancel the r terms and divide by r 2 , which gives:
From Lundberg’s inequality the probability of ultimate ruin with an initial surplus u is
ψ ( u) ≈ e − ru . So, with an initial surplus of 1,000 (in units of £000s), we have:
The probability of ultimate ruin is very low. This is because there are relatively few
large claims and, as we saw in part (i)(b), the surplus can be expected to grow quite
quickly because of the 20% loading incorporated in the premiums. [1]
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.