CT6 CMP Upgrade 09-10

CT6: CMP Upgrade 2009/10 Page 1
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.
This CMP Upgrade contains:

• All changes to the Syllabus objectives and Core Reading.
• Changes to the ActEd Course Notes, Series X Assignments and Question and
Answer Bank that will make them suitable for study for the 2010 exams.
The Actuarial Education Company © IFE: 2010 Examinations

Page 2 CT6: CMP Upgrade 2009/10
1 Changes to the Syllabus objectives and Core Reading
1.1 Syllabus objectives
Chapter 6 (new)
Page 1
Syllabus objectives (v) 8 to (v) 10 are new. These read:
8. Explain the empirical Bayes approach to credibility theory, in particular its

similarities with and its differences from the Bayesian approach.
9. State the assumptions underlying the two models in 8.
10. Calculate credibility premiums for the two models in 8.
Chapter 6 (now Chapter 7)
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.
2. Define a Poisson process, derive the distribution of the number of events

in a given time interval, derive the distribution of inter-event times, and
apply these results.
© IFE: 2010 Examinations The Actuarial Education Company

3. Define a compound Poisson process and derive the moments and moment
generating function for such a process.
4. Define the adjustment coefficient for a compound Poisson process and

for discrete time processes which are not compound Poisson, calculate it
in simple cases and derive an approximation.
5. Define the probability of ruin in infinite/finite and continuous/discrete

time and state and explain relationships between the different
probabilities of ruin.
6. State Lundberg's inequality and explain the significance of the

adjustment coefficient.
7. Describe the effect on the probability of ruin, in both finite and infinite
time, of changing parameter values.
8. Analyse the effect on the adjustment coefficient and hence on the

probability of ruin of simple reinsurance arrangements.
Chapter 9 (removed)
Page 1
Syllabus objective (vi) has been removed from Subject CT6. Syllabus objectives (vii)
onwards have been renumbered accordingly.
Page 1
Syllabus objective (vii) has been renumbered to Syllabus objective (vi).
Page 1
Syllabus objective (viii) has been renumbered to Syllabus objective (vii).

Chapter 12
Page 1
Syllabus objective (ix) has been renumbered to Syllabus objective (viii).
Chapter 13
Page 1
Syllabus objective (ix) has been renumbered to Syllabus objective (viii).
Chapter 14
Page 1
Syllabus objective (x) has been renumbered to Syllabus objective (ix).
1.2 Core Reading
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
The following Core Reading paragraph is no longer in Subject CT6:
This material is covered in detail in Subject ST3 – General Insurance Specialist

Technical.
Page 8
The following Core Reading paragraph is no longer in Subject CT6:
In this course we will study Bayesian credibility. Empirical Bayes credibility

theory is covered in Subject ST3, General Insurance Specialist Technical.

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
The following Core Reading paragraphs are no longer in Subject CT6:
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.
Firstly, it is worth noting some overriding features of all types of insurance.
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.

2 Changes to the ActEd Course Notes
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)
This chapter is new and is provided in the replacement pages.
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.
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.

Chapter 14

3 Changes to the Q&A Bank

The Questions in the Question and Answer Bank have been annotated according to
whether they are:
● 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.
Q&A Bank Part 1
Solution 4.18 (i)
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.
Hence, v1 = 0.1834 , v2 = 0.5872 and s = v12 + v22 = 0.37844 .
Solution 5.12 (ii)(a)
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
First note that:
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

Q&A Bank Part 2
Questions 2.7 to 2.9 have been deleted.
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
Q&A Bank Part 3
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 3.18 are now Questions 3.19 to 3.29.
Questions 3.19 to 3.24 are now Questions 3.13 to 3.18.
Questions 3.8 to 3.12 are new and attached as replacement pages at the end.

4 Changes to the X Assignments

Assignment X1
Solution X1.5
This solution has been updated to be more efficient, since we observe strategy d3 is
dominated. Replacement pages have been provided.
Solution X1.10 (iii) & (iv)
The notation here has changed. The confusing d A , d B and dC have been replaced
with d1 , d 2 and d 3 .
Solution X1.10 (iii)
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

Solution X2.8 (now Solution X2.7)
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:
Consider the following time series:
X t = 5 + Z t 1 + ( X t -1 - 5) 2
where Z t is a white noise process with mean 0 and standard deviation 1.
Describe how X t fluctuates and state the long-term mean of the process.

5 Other tuition services

In addition to this CMP Upgrade you might find the following services helpful with
your study.
5.1 Study Material
We offer the following study material in Subject CT6:
• 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
We offer the following tutorials in Subject CT6:
• a set of Regular Tutorials (lasting two or three full days)

• a Block Tutorial (lasting two or three full days)
• a Revision Day (lasting one full day).
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.

6 Feedback on the study material

ActEd is always pleased to get feedback from students about any aspect of our study
programmes. Please let us know if you have any specific comments (eg about certain
sections of the notes or particular questions) or general suggestions about how we can
improve the study material. We will incorporate as many of your suggestions as we can
when we update the course material each year.
If you have any comments on this course please send them by email to CT6@bpp.com
or by fax to 01235 550085.

Q&A Bank Question Classification Summary
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.
Question Part 1 Part 2 Part 3 Part 4

1 Developmental Developmental Deleted Developmental
2 Developmental Exam-style Deleted Developmental
3 Exam-style Exam-style Deleted Developmental
4 Developmental Exam-style Deleted Exam-style
5 Exam-style Exam-style Deleted Developmental
6 Developmental Bookwork Deleted Developmental
7 Deleted Bookwork Deleted Developmental
8 Deleted Bookwork Bookwork Developmental
9 Deleted Bookwork Developmental Developmental
10 Developmental Developmental Developmental Developmental
11 Exam-style Developmental Developmental Developmental
12 Exam-style Developmental Exam-style Developmental
15 Exam-style Exam-style Exam-style Exam-style
16 Developmental Exam-style Exam-style Developmental
17 Developmental Developmental Exam-style Developmental
18 Exam-style Exam-style Exam-style Developmental
19 Exam-style Exam-style Exam-style Exam-style
20 Exam-style Exam-style Exam-style Developmental
21 Exam-style Exam-style Developmental Developmental
22 Developmental Exam-style Exam-style Bookwork
23 Exam-style Developmental Developmental Exam-style
25 Exam-style Deleted Exam-style
26 Developmental Developmental
27 Developmental Developmental
28 Developmental Exam-style
29 Exam-style Exam-style
31 Developmental Exam-style
33 Exam-style
34 Exam-style

All study material produced by ActEd is copyright and is
sold for the exclusive use of the purchaser. The copyright
is owned by Institute and Faculty Education Limited, a
subsidiary of the Faculty and Institute of Actuaries.
You may not hire out, lend, give out, sell, store or transmit
electronically or photocopy any part of the study material.
You must take care of your study material to ensure that it

is not used or copied by anybody else.
Legal action will be taken if these terms are infringed. In

addition, we may seek to take disciplinary action through
the profession or through your employer.
These conditions remain in force after you have finished

using the course.

CT6-05: Credibility theory Page 25
3 Exam-style questions
Exam-style question 1
A statistician wishes to find a Bayesian estimate of the mean of an exponential

1
distribution with density function f ( x) = e - x / m . He is proposing to use a prior
m
distribution of the form:
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 ,
Â xi = 9,826 , and Â xi2 = 1, 200, 000 . Find the posterior estimate for m , and
the value of the credibility factor in this case. [3]
(v) Comment on the results obtained in part (iv). [2]

[Total 13]

Page 26 CT6-05: Credibility theory
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 .
Claim numbers x1 , … , xn over the last n years have been recorded.
(i) Show that the posterior distribution is gamma and determine its parameters. [3]
8
(ii) Given that n = 8 and Â xi = 5 determine the Bayesian estimate for m under:
i =1
(a) squared-error loss
(b) “all-or-nothing” loss
(c) absolute error loss. [5]

[Total 8]

Solution to exam-style question 1
(i) Likelihood function
The likelihood function is:
1 - x1 / m 1 - xn / m e - Â xi / m
L( m ) = e ¥ ¥ e =
m m mn
(ii) Posterior distribution
The posterior distribution for m is proportional to the prior distribution multiplied by

the likelihood function. So we have (ignoring any terms not involving m ):
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
This is the Bayesian estimate for m under squared error loss.

(iii) Credibility estimate
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 + Â xi q Â xi
mˆ = = +
n + a -1 n + a -1 n + a -1
=
q
¥
a -1
+
Â 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
Â 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
(iv) Posterior estimate
We now have:
q + Â xi 40 + 9,826
mˆ = = = 98.1692
n + a - 1 100 + 1.5 - 1
The value of the credibility factor is:
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 .

(i) Poisson/gamma posterior
The likelihood is given by:
m x1 m xn
L( m ) = e -m
¥…¥ e - m μ m Â xi e - n m
x1 ! xn !
The prior distribution of m ~ gamma(2,5) is given by:
52
f (m) = m e -5 m μ m e -5 m
G (2)
So the posterior is given by:
posterior μ prior ¥ likelihood
= m e -5 m ¥ m Â xi e - nm
= m1+Â xi e - ( n +5) m = m (2+Â xi ) -1e - ( n +5) m
This is the form of a gamma (2 + Â xi , n + 5) distribution.
(ii)(a) Squared-error loss
We have n = 8 and Â xi = 5 , so the posterior is gamma(7,13) .
The Bayesian estimate for m under squared-error loss is the mean of the posterior:
7
mˆ = 0.538
13
(ii)(b) “All-or-nothing” loss
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

Taking logs, differentiating and equating the derivative to zero gives:
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
Alternatively, students may differentiate the original PDF to give:
d
f ( m ) = 6 m 5e -13 m - 13m 6e -13 m
dm
= m 5e -13 m (6 - 13m )
Setting this equal to zero gives m = 0, 13

6
. The m = 0 solution can be discarded as
m ~ gamma(7,13) means that m > 0 .
a -1
In general the mode of the gamma distribution is (provided that a > 1 ).
l
(ii)(c) Absolute error loss
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

CT6-06: Emperical Bayes Credibility theory Page 1
Chapter 6
Emperical Bayes Credibility theory
Syllabus objectives
(v) Explain the fundamental concepts of Bayesian statistics and use these concepts
to calculate Bayesian estimators.
8. Explain the empirical Bayes approach to credibility theory, in particular

its similarities with and its differences from the Bayesian approach.
9. State the assumptions underlying the two models in 8.
10. Calculate credibility premiums for the two models in 8.
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.

Page 2 CT6-06: Emperical Bayes Credibility theory
1 Empirical Bayes Credibility Theory: Model 1
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 .
Conditional claim distribution
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.

Empirical Bayes Credibility Theory is the name given to a particular approach to

the problems in Section 1 of the previous chapter. This approach has led to the
development of a vast number of different models of varying degrees of
complexity. In this chapter, two of these models will be studied. In this section
the simplest possible model will be studied. Although this model, which will be
referred to as Model 1, is not very useful in practice, it does provide a good
introduction to the principles underlying Empirical Bayes Credibility Theory
(EBCT). In particular, it shows the similarities and the differences between the
Empirical Bayes and the pure Bayesian approaches to credibility theory. In
Section 2 an extension of Model 1 will be studied that is far more useful in
practice.
1.2 Model 1: specification
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 following assumptions will be made for EBCT Model 1.
Assumptions for EBCT Model 1
The distribution of each X j depends on a parameter, denoted θ , whose value is

fixed (and the same for all the X j s) but is unknown. (1.1)
Given θ , the X j ’s are independent and identically distributed. (1.2)
The parameter θ is known as the risk parameter. It could, as in Section 2 of the

previous chapter, be a real number or it could be a more general quantity such
as a set of real numbers.
The risk parameter in the Poisson/gamma model is (the Poisson parameter) λ . The risk
parameter for the normal/normal model is θ .

A consequence of these two assumptions is that:
the random variables X j n s are identically distributed. (1.3)
An important point to note is that:
the X j ’s are not (necessarily) unconditionally independent. (1.4)
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).
Next some notation is introduced. Define m ( q ) and s 2 ( q ) as follows:
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.
Note also that they are functions of the unknown parameter θ .
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
A specialist insurer concentrates on providing third party motor insurance to drivers of a

certain type of car who are based in London. Explain what each of the variables and
functions in EBCT Model 1 would represent if this model were used to describe the
numbers of claims made by different drivers in different years. The only risk factor
considered relevant is the safety standard adopted by each individual driver.
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.
X j is a random variable, representing the number of claims made by a particular driver

in year j .
X j q represents the number of claims made by a particular driver whose safety

coefficient is q in year j .
m(q ) = E ( X j q ) is the (theoretical) average number of claims made by a driver whose

safety coefficient is q . In this example, m(q ) would be a decreasing function of q ,
since a higher safety coefficient would reduce the average number of claims.
s 2 (q ) = var( X j q ) is the (theoretical) variance of the number of claims made by a driver

whose safety coefficient is q in different years. s 2 (q ) will take lower values for drivers
with a high safety coefficient, since they are likely to make no claims, or possibly one
claim, each year, whereas the numbers of claims made each year by drivers with low
safety coefficients may range from none to five or more.

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).
(ii) Assumptions concerning the unconditional distribution of the X j ’s are

the same: they are identically distributed in each case. See (2.23) and
(2.24) from the previous chapter, and (1.3) and (1.4).
(iii) Assumptions concerning the (conditional) distribution of the X j ’s given

θ are the same: they are conditionally independent in each case. See
(2.19) and (2.22) from the previous chapter, and (1.2).
EBCT Model 1 can be regarded as a generalisation of the normal/normal model.

The particular points where it differs from, ie generalises, the normal/normal
model are:
(i) E ( X j | q ) is some function of q , m ( q ) , for EBCT Model 1 but is simply q

for the normal/normal model. Hence:
var [m (q )] {EBCT} corresponds to s 22 ( = var ( q ) ) {normal/normal}.
(ii) var ( X j | q ) is a function of q , s 2 ( q ) , for EBCT Model 1 but is a
constant, s 12 , for the normal/normal model. Hence:
( )
E È s 2 (q ) ˘ corresponds to s 12 ( = var X j | q ).
Î ˚

(iii) The normal/normal model makes very precise distributional assumptions

about both X j given q , which is N ( q ,s 12 ) , and q , which is N ( m ,s 22 ) .
EBCT Model 1 makes no such distributional assumptions.
(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.
1.3 Model 1: the credibility premium
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.
Result of EBCT Model 1
The estimate of m ( q ) given X given by EBCT Model 1 is:
(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:
E [m ( q )] playing the role of m and
n
Â 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.
We saw this earlier.
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.

1.4 Model 1: parameter estimation
In this section the solution to the problem of estimating m ( q ) given X will be

completed by showing how to estimate E [m ( q )] , var [m ( q )] and E È s 2 ( q ) ˘ .
Î ˚
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.
What is of interest is estimating the pure premium, or the expected number of

claims, for a particular risk, just as in Section 1.2 and Section 1.3, and that this
risk is one of N risks in a collective. By a collective is meant a collection of
different risks that are related in a way to be made clearer later in this section.
For simplicity, suppose the particular risk that is of interest is Risk Number 1 in
this collective. It is assumed that the aggregate claims, or claim frequencies, for
each of these N risks for each of the past n years have been observed. Let X ij
denote the aggregate claims, or number of claims, for Risk Number i ,
i = 1, 2, …, N , in year j , j = 1, 2, …, n .
These values are summarised in the following table:
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.
For each risk i , i = 1, 2, … , N :
The distribution of each X ij , j = 1, 2, … , n , depends on the value of a parameter,

denoted q i , whose value is fixed (and the same for each value of j ) but is
unknown. (1.7)
Given θ i , the X ij ’s, j = 1,2,… , n , are independent and identically distributed.

(1.8)
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.
For i π k , the pairs (q i , X ij ) and (q k , X km ) are independent and identically

distributed. (1.9)
This assumption shows that the rows of Table 1 are independent of each other.
Two immediate consequences of this assumption are:
For i π k , X ij and X km are independent and identically distributed. (1.10)
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 )] ,
E È s 2 ( q i ) ˘ and var [m ( q i )] do not depend on i . These are precisely the

Î ˚
parameters denoted E [m ( q )] , E È s 2 ( q ) ˘ and var [m ( q )] in Section 1.2 and
Î ˚
Section 1.3 and which the collective will be used to estimate.
More notation is needed. Denote:
n
Â X ij / n by X i
j =1
and:
N Ê N n ˆ
Â Xi / N Á= Â 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 .
Estimators for E [m ( q )] , E È s 2 ( q ) ˘ and var [m ( q )] can now be produced.

Î ˚
{{X ij } j =1}i =1, whose values will be known

n N
These estimators will be functions of
when the credibility estimate of m ( q1 ) is computed.
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 .

Now, E [m ( q )] is the “average” (over the distribution of q ) of the values of

m ( q ) for different values of q . The obvious estimator for E [m ( q )] is the
average of the estimates of m ( q i ) for i = 1, 2, … , N . In other words, the
estimator for E [m ( q )] is X .
So we estimate E [m (q )] by the average of the sample means X1 ,..., X N . This is

equivalent to the sample mean of the full set of (both direct and collateral) data.
Similarly, E È s 2 ( q ) ˘ is the “average” value of s 2 ( q ) and so an obvious

Î ˚
estimator is the average of the estimates of s 2 ( q i ) , which is:
N n
Â (n - 1)-1 Â ( 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.
For each row of Table 1, X i is an estimate of m ( q i ) , i = 1,2,… , N . So it might be

thought that the observed variance of these values, ie:
N
(N - 1)-1 Â ( X i -X )2
i =1
would be an obvious estimator for var [m ( q )] . Unfortunately, this can be shown

to be a biased estimate of var [m ( q )] .
In fact it is positively biased, ie it tends to overestimate slightly.
It can be shown that an unbiased estimate of var [m ( q )] can be produced by

subtracting a correction term from the above formula. An unbiased estimator for
var [m ( q )] is:
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
The correction term is n -1 times the estimator for E È s 2 ( q ) ˘ .

Î ˚

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.
These estimators can be summarised as follows:
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
Â (n - 1)-1 Â ( X ij - Xi )2
i =1 i =1 j =1
(1.15)
These formulae are given in the Tables.
They are on Page 29. We will do some numerical examples shortly to help you see
what is going on.
An important point to note is that although var [m ( q )] is a non-negative

parameter since it is a variance, the estimator given by (1.15) could be negative.
Formula (1.15) is the difference between two terms. Each of these terms is non-
negative but their difference need not be.
It could be that the second term works out to be bigger than the first.
If, in practice, (1.15) gives a negative value, the accepted procedure is to

estimate var [m ( q )] as 0. Strictly speaking, this means that the estimator for
var [m ( q )] is the maximum of 0 and the value given by (1.15). Although (1.15)
gives an unbiased estimate of var [m ( q )] , taking the maximum of 0 and (1.15)
does not give an unbiased estimate. However, this pragmatic approach avoids a
nonsensical estimate for var [m ( q )] .
The parameter E È s 2 ( q ) ˘ must also be non-negative, but its estimator, given by

Î ˚
formula (1.14), will always be non-negative and so no adjustment to (1.14) is
required.
It can be shown that the estimators for E [m ( q )] , E È s 2 ( q ) ˘ and var [m ( q )] are
Î ˚
unbiased. The proof is beyond the scope of the syllabus.

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:
n is the number of data values in respect of the risk.
Usually this will correspond to the number of years of data available.
E È s 2 ( q )˘ is the average variability of data values from year to year for a

Î ˚
single risk, ie the average variability within the rows of Table 1
var [m ( q )] is the variability of the average data values for different risks, ie the
variability of the row means in Table 1.
Looking at formula (1.6) for Z , the following observations can be made:
(i) Z is always between zero and one.
(ii) Z is an increasing function of n . This is to be expected – the more data

there are from the risk itself, the more it will be relied on when the
credibility estimate of the pure premium or number of claims is calculated.
(iii) Z is a decreasing function of E È s 2 ( q ) ˘ . This is to be expected – the

Î ˚
higher the value of E È s 2 ( q ) ˘ , relative to var [m ( q )] , the more variable,
Î ˚
and hence less reliable, are the data from the risk itself relative to the data
from the other risks in the collective.
(iv) Z is an increasing function of var [m ( q )] . This is to be expected – the
higher the value of var [m ( q )] , relative to E È s 2 ( q ) ˘ , the more variability

Î ˚
there is between the different risks in the collective and hence the less
likely it is that the other risks in the collective will resemble the risk that is
of interest, and the less reliance should be placed on the data from these
other risks.

There appears to be a contradiction in this section. In Section 1.2 of the

previous chapter it was stated that the credibility factor should not depend on
the data from the risk being rated. However, these data have been used to
estimate var [m ( q )] and E È s 2 ( q ) ˘ , whose values are then used to calculate Z .
Î ˚
The explanation is that in principle the credibility factor, Z , as given by formula
(1.12), does not depend on the actual data from the risk being rated.
Unfortunately, the formula for Z involves two parameters, var [m ( q )] and
E È s 2 ( q ) ˘ , whose values are unknown but which can, in practice, be estimated

Î ˚
from data from the risk itself and from the other risks in the collective.
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.
Given q i , the X ij ’s, j = 1, 2, … , n , are independent for each risk i , i = 1, 2, … , N .

(1.16)
For i π k , the pairs (q i , X ij ) and (q k , X kl ) are independent and the risk

parameters q 1 ,q 2 ,… ,q N , are also identically distributed. (1.17)
(
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).

None of the results or formulae in Section 1 would be altered in any way by

making these weaker assumptions. The reason for making the slightly stronger
assumptions, as was done in Section 1.2, is that they make the presentation a
little easier. The reason for pointing out now that weaker assumptions could
have been made is that this will help to link EBCT Model 1 with EBCT Model 2 in
Section 2.
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.
Total claim amount
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
In this example, there are n = 5 years and N = 4 risks (= countries).
X i is the average claims for risk i for the 5-year period. So the missing entry is:
X 4 = (44 + 52 + 69 + 55 + 71) / 5 = 58.2
We can then calculate the other missing entry:
1 5
( ) 1⎡
2
∑ X ij − X i
4 j =1
=
4⎣
(44 − 58.2) 2 + + (71 − 58.2) 2 ⎤ = 132.7
⎦

To find the estimates we need X , which is:
1 4
X= ∑ X i = (50.4 + 68.4 + 74.0 + 58.2) / 4 = 62.75
4 i =1
We can then the evaluate the estimates directly:
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
= (39.3 + 17.3 + 215.5 + 132.7) / 4 = 101.2
Question 6.3
Estimate the value of var [m(q )] .

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:
Country 1: P = 0.8169 × 50.4 + (1 − 0.8169) × 62.75 = 52.66

Country 2: P = 0.8169 × 68.4 + (1 − 0.8169) × 62.75 = 67.37
Country 3: P = 0.8169 × 74.0 + (1 − 0.8169) × 62.75 = 7194
.
Country 4: P = 0.8169 × 58.2 + (1 − 0.8169) × 62.75 = 59.03
Question 6.4
Are the following statements true or false for EBCT Model 1?
(a) θ represents the “true” risk premium for a given risk.
(b) The variance of X j θ doesn’t depend on θ .
(c) None of the random variables or parameters in the model are assumed to have a
normal distribution.

2 Empirical Bayes Credibility Theory: Model 2
2.1 Introduction
Model 2 is a generalisation of Model 1. This section is therefore very similar to

Section 1. The important thing is to note the differences.
In this section the techniques of Empirical Bayes Credibility will be applied to a

second, and slightly more complicated, model. The format will be exactly the
same as in Section 1. First of all, in Section 2.2, the problem will be stated and
the assumptions set out. The problem will be the same as in Section 1, ie to
estimate the pure premium, or the expected number of claims, in the coming
year for a risk. The assumptions will be slightly different from those in Section 1.
Next, in Section 2.3, the credibility estimate for the pure premium or expected
number of claims will be considered. Finally, in Section 2.4, the method of
estimating the values of the parameters that are part of the credibility estimate
will be discussed.
In keeping with the terminology in Section 1, the model in this section will be
referred to as EBCT Model 2.
The examiners will also refer to the model in this way.
2.2 Model 2: specification
The problem is to estimate the expected aggregate claims, or the expected

number of claims, in the coming year for a given risk. Let Y1,Y2 ,…,Yn be random
variables representing the aggregate claims or numbers of claims in successive
years from this risk.
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.
Next a new sequence of random variables, X 1 , X 2 ,… , is defined as follows:
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.
The random variable X j represents the aggregate claims, or the number of

claims, in year j standardised to remove the effect of different levels of
business in different years. The assumptions that specify EBCT Model 2 are as
follows.
Assumptions for EBCT Model 2
The distribution of each X j depends on the value of a parameter, q , whose

value is the same for each j but is unknown. (2.1)
Given q , the X j ’s are independent (but not necessarily identically distributed).

(2.2)
( )
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.
Having made assumptions (2.3) and (2.4), m ( q ) and s 2 ( q ) can be defined as

follows:
(
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.
Remember to include the Pj factor in the definition of s2 (θ ) .
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
policy have mean m ( q ) and variance s 2 ( q ) , where m ( ) and s 2 ( ) are

functions of q and q is the fixed, but unknown, risk parameter for all these
policies. Now let Y j denote the aggregate claims from all the policies in force in
year j .

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 )
This example satisfies assumptions (2.3) and (2.4).
2.3 Model 2: derivation of the credibility premium
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.
The solution to the problem, ie the “best” linear estimate of m ( q ) given X is

given by:
n
E [m ( q )] E È s 2 ( q ) ˘ var [m (q )] +
Î ˚ Â Yj
j =1
n
Â Pj + E È s 2 ( q ) ˘ var [m (q )]
Î ˚
j =1
Remember that Yj is just Pj X j .

This can be written more attractively as follows:
Result of EBCT Model 2
The estimate of m ( q ) given X given by EBCT Model 2 is:
Z X + (1 - Z ) E [m ( q )] (2.5)
where:
n n n n
X = Â Pj X j Â Pj = Â Yj Â Pj
j =1 j =1 j =1 j =1
and:
n
Â Pj
j =1
Z=
n
Â 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
Check that the “attractive” version really is the same.
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.

Here are some additional points of note about this solution:
(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.
2.4 Model 2: parameter estimation
The procedure for estimating the parameters E [ m ( q )] , var [m ( q )] and
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.
For each i and j define:
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
2 Y21 , P21 Y22 , P22 … Y2n , P2n

Risk
Number . . . … .
. . . … .
N YN 1 , PN 1 YN 2 , PN 2 … YNn , PNn
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
help to estimate the parameters E [m ( q )] , var [m ( q )] and E È s 2 ( q ) ˘ that appear

Î ˚
in (2.5).
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.
For each risk i , i = 1, 2, … , N :
the distribution of each X ij , j = 1, 2, … , n , depends on a parameter, q i , whose

value is the same for each j but is unknown (2.6)
given θ i , the X ij ’s are independent (but not necessarily identically distributed)

(2.7)
(
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.
The risk parameters q 1 ,q 2 ,… ,q N , regarded as random variables, are independent

and identically distributed. (2.10)
For i π k , the pairs (q i , X ij ) and (q k , X km ) are independent. (2.11)
Notice that, since the q i ’s are identically distributed, the values of E [m ( q i )] ,
var [m ( q i )] and E È s 2 ( q i ) ˘ do not depend on i so that they can be denoted by

Î ˚
E [m ( q )] , var [m ( q )] and E È s 2 ( q ) ˘ , respectively, as in Section 2.2 and
Î ˚
Section 2.3.
As in Section 1.4, more notation is needed. Denote:
n
Â Pij by Pi (2.12)
j =1
N
Â Pi by P (2.13)
i =1
N
(Nn - 1) Â Pi (1 - Pi / P )
-1
by P * (2.14)
i =1
n
Â Pij X ij / Pi by X i (2.15)
j =1
N N n
Â Pi X i / P = Â Â 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 = Â P1 j X 1 j Â P1 j
j =1 j =1
and:
n
Â P1 j
j =1
Z1 =
n
Â 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.
Unbiased estimators for E [m ( q )] , var [m ( q )] and E È s 2 ( q ) ˘ can be proposed

Î ˚
based on the observed values

RSnYij , Pij sn UVN .
T j =1W
i =1

These estimators are:
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
Â Â ( )
-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 ˚
The points of note about these estimators are:
(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.
And you’ll be pleased to know that …
(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
Calculate the EBCT premium for Country 1 using Model 2.
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.
Total claims per unit volume (Xij)

Country (i) Year (j)
1 2 3 4 5
1 4.000 3.533 3.231 3.125 5.900
2 3.200 5.071 2.909 4.867 2.333
3 17.000 6.750 12.667 5.417 22.500
4 2.000 1.486 2.300 3.438 7.100
We can then calculate Pi , P and P * .

The figures are given 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
So this gives: E [m(q )] ª X = 3.984
From the other columns in the table, we get:
1 4 1 5
( )
2
E ÈÎ s 2 (q ) ˘˚ ª Â Â Pij X ij - X i
4 i =1 4 j =1
= (57.13 + 111.59 + 1, 237.82 + 267.73) /16
= 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
The credibility factor for Country 1 is:
n
Â P1 j
j =1 66
Z1 = = = 0.8048
n E ÈÎ s 2 (q ) ˘˚ 104.64
66 +
Â P1 j + var [m(q )] 6.539
j =1
The risk premium per unit volume is:
Z1 X1 + (1 - Z1 ) E [m(q )] = 0.8048 ¥ 3.818 + (1 - 0.8048) ¥ 3.984 = 3.851
Since the volume for Country 1 for the coming year is 20 units, the EBCT premium is:
20 × 3851
. = 77.01
Question 6.6
Calculate the EBCT premiums for Countries 2, 3 and 4.

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.
(b) var [m(q )] is reduced.
(c) The unit of currency is changed from £ to $.
(d) All the Pj ’s are increased by the same factor.
Does the credibility factor “behave” as you would expect?
We now finish this chapter with two exam-style questions.
Past Exam Question (Subject 106, September 2000, Question 3)
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.
Year 1 Year 2 Year 3 Year 4
Insurer A £14.2m £15.8m £22.7m £19.0m
163 189 252 199
Insurer B £58.6m £63.1m £81.0m £64.2m
4,435 4,761 5,576 4,581
Insurer C £123m £132m £161m £133m
16,184 17,443 20,102 18,000
(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]
∑ P1 j ( X1 j − X1) 2 = 0.014667 ∑ P1 j ( X1 j − X ) 2 = 5106461

.
∑ 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.
The credibility premium can be expressed in terms of a credibility factor, which

depends on the mean and variance of the conditional claim distribution. These
quantities can be estimated based on data derived from a number of different risks.
The table below summarises the models.
Model Quantity Assumptions about the Assumptions about the

estimated underlying distribution conditional claim
of the risk parameter θ distribution X|θ
EBCT Model 1 risk premium a definite but unknown a distribution with
or distribution mean m(θ ) and
2
variance s (θ )
EBCT Model 2 claim
frequency Model 2 adjusts for
volume of business
Empirical Bayes Credibility Theory model 1
Definitions: X j represents the amount (or number) of claims
(
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 ¯
Credibility premium: E [m(q )| X ] = Z X i + (1 - Z ) E [m(q )]

Empirical Bayes Credibility Theory model 2
Definitions: Yj represents the amount (or number) of claims
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.)
X j is a random variable, representing the score obtained by a particular golfer on the

j th day.
X j θ represents the score obtained by a particular golfer whose handicap is θ on the j th

day.
m(θ ) = E ( X j θ ) is the (theoretical) average score for a golfer whose handicap is θ .

Note that, although the average score and the handicap are related, m(θ ) will not be equal
to θ . In this example, m(θ ) will be an increasing function, since golfers with high
handicaps (ie not very good players) will have high average scores (ie they will take lots
of shots to get round).
s 2 (θ ) = var( X j θ ) is the (theoretical) variance in the scores obtained by a golfer whose

handicap is θ . s2 (θ ) will take lower values for players with a high standard of play (a
low θ ), since their scores will be fairly consistent, whereas the scores for novices will
tend to fluctuate over a much wider range. So again, s2 (θ ) will be an increasing
function.
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
The general credibility formula is:
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
Â Â 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 ) ˘˚ .
So, putting in the numbers gives:
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 ˘˚ .
(b) False. The variance of X j q is s 2 (q ) , which is a function of q .
(c) True. In fact, none of the quantities in the model are assumed to have any
specific type of distribution.

Solution 6.5
We start with Z X + (1 - Z ) E [m(q )] . This is equal to:
n n E ÈÎ s 2 (q ) ˘˚
Â Pj Â 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 ) ˘˚
Â Pj + var [m(q )] Â Pj Â 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:
Country Credibility factor Risk premium per EBCT premium

unit volume
1 0.8048 3.851 77.0
2 0.8632 3.468 86.7
3 0.6862 8.504 85.0
4 0.8759 2.750 33.0

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.
Since increasing E ⎡ s 2 (θ ) ⎤ increases the denominator, Z decreases, as expected.

⎣ ⎦
(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.
Since reducing var [m(q )] increases the denominator, Z decreases, as expected.
(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.

Solution to Past Exam Question (Subject 106, September 2000, Question 3)
(i) Credibility factor
E ÈÎ s 2 (q )˘˚ is estimated by the average of the sample variances, ie:
4,121, 280 + 7, 299,175 + 3,814, 001

= 5, 078,152
3
The sample mean of the X i ’s is:
2,517 + 7,814 + 2,920

= 4, 417
3
and the sample variance of the X i ’s is:
(2,517 - 4, 417)2 + (7,814 - 4, 417)2 + (2,920 - 4, 417)2 = 8, 695,309

2
So var ÈÎ m (q ) ˘˚ is estimated by:
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
Z is an increasing function of n , the number of years of past data. If we have more

than 4 years of past data, the credibility factor will increase.
Z is a decreasing function of E ÈÎ s 2 (q )˘˚ . If E ÈÎ s 2 (q )˘˚ increases, eg if the variance of

the claim amounts from one or more of the risks were to increase, then the value of the
credibility factor would fall.
Z is an increasing function of var ÈÎ m (q ) ˘˚ . If var ÈÎ m (q ) ˘˚ increases, eg if there was

greater variation between the individual sample means, then Z would increase.
Solution to Exam-style Question
(i) Analysis using EBCT Model 1
Using EBCT Model 1, we obtain the following numerical values:
Year 1 Year 2 Year 3 Year 4 Xi 1

3 ∑ ( X ij − X i ) 2
Insurer A 14.2 15.8 22.7 19.0 17.925 14.1158
Insurer B 58.6 63.1 81.0 64.2 66.725 96.4358
Insurer C 123 132 161 133 137.25 270.9167
Using the standard formulae for the EBCT estimates, we get:
X = 13 (17.925 + 66.725 + 137.25) = 73.967
E ÈÎ s 2 (q ) ˘˚ = 13 (14.1158 + 96.4358 + 270.9167) = 127.1561
var [m(q )] = 12 ÈÎ (17.925 - 73.967) 2 + (66.725 - 73.967) 2
+(137.25 - 73.967) 2 ˘˚ - 14 ¥ 127.1561
= 3,567.1562

So the credibility factor for the model is:
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:
66.725 × 0.9912 + 73.967 × 0.0088 = £66.79 m
(ii) Analysis using EBCT Model 2
We first need to use the data to calculate the statistics we need. Using the notation
given in the Tables, we have:
P1 = 163 + 189 + 252 + 199 = 803
The same approach for the other insurers gives the values:
P2 = 19,353
P3 = 71,729
For the whole portfolio we have:
P = 803 + 19,353 + 71,729 = 91,885
and:
1È Ê 803 ˆ Ê 19,353 ˆ Ê 71, 729 ˆ ˘

P* = Í803 ÁË1 - + 19,353 Á1 - + 71, 729 Á1 -
11 Î ˜
91,885 ¯ ˜
Ë 91,885 ¯ Ë 91,885 ˜¯ ˙˚
= 2,891.5793
Calculating X i for the first insurer:
14.2 + 158
. + 22.7 + 19
X1 = = 0.089290
803

Similar calculations for the other insurers give:
X 2 = 0.013791
X 3 = 0.007654
For the whole portfolio we have:
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
So for Insurer B, the credibility factor is:
19,353
ZB = = 0.999208
0.002750
19,353 +
0.0001794
So the credibility premium per unit of risk volume for Insurer B is:
0.999208 × 0.013791 + 0.000792 × 0.009660 = 0.013788
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.




using the course.

CT6-09: Ruin theory Page 1
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.
2. Define a Poisson process, derive the distribution of the number of events

in a given time interval, derive the distribution of inter-event times, and
apply these results.
3. Define a compound Poisson process and derive the moments and moment
generating function for such a process.
4. Define the adjustment coefficient for a compound Poisson process and

for discrete time processes which are not compound Poisson, calculate it
in simple cases and derive an approximation.
5. Define the probability of ruin in infinite/finite and continuous/discrete

time and state and explain relationships between the different
probabilities of ruin.
6. State Lundberg's inequality and explain the significance of the

adjustment coefficient.
7. Describe the effect on the probability of ruin, in both finite and infinite
time, of changing parameter values.
8. Analyse the effect on the adjustment coefficient and hence on the

probability of ruin of simple reinsurance arrangements.

Page 2 CT6-09: Ruin theory
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.
In Section 3 we will introduce the adjustment coefficient, a parameter associated with

risk, and Lundberg’s inequality.
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
One technical point needed later in this chapter is that a function f ( x ) is

described as being o( x ) as x goes to zero, if:
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
Which of the following functions are o( x) as x Æ 0 ?
(i) x2 (ii) ex (iii) e- x - 1 + x
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).
In Chapters 7 and 8 the aggregate claims generated by a portfolio of policies

over a single time period were studied. In the actuarial literature, the word “risk”
is often used instead of the phrase “portfolio of policies”. In this chapter both
terms will be used, so that by a “risk” will be meant either a single policy or a
collection of policies. In this chapter this study will be taken a stage further by
considering the claims generated by a portfolio over successive time periods.
Some notation is needed.
N (t ) the number of claims generated by the portfolio in the time

interval [0, t], for all t ≥ 0
Xi the amount of the i-th claim, i = 1, 2, 3, ...
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.
It can be seen that:
N (t )
S (t ) = Â Xi
i =1
with the understanding that S(t) is zero if N(t) is zero.
The stochastic process {S (t )}t ≥0 as defined above is known as the aggregate

claims process for the risk. The random variables N(1) and S(1) represent the
number of claims and the aggregate claims respectively from the portfolio in the
first unit of time. These two random variables correspond to the random
variables N and S, respectively, introduced in Chapter 7.
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:
c the rate of premium income per unit time
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.

1.2 The surplus process
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.
1.3 The probability of ruin in continuous time
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.
Now to be more precise. The following two probabilities are defined:
y (U ) = P [U (t ) < 0, for some t , 0 < t < • ]
y (U , t ) = P [U (t ) < 0, for some t , 0 < t £ t ]

ψ (U ) is the probability of ultimate ruin (given initial surplus U) and ψ (U , t ) is the

probability of ruin within time t (given initial surplus U). These probabilities are
sometimes referred to as the probability of ruin in infinite time and the
probability of ruin in finite time. Here are some important logical relationships
between these probabilities for 0 < t1 ≤ t2 < ∞ and for 0 ≤ U1 ≤ U2:
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 intuitive explanations for these relationships are as follows:
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).
Finally, the probability of ultimate ruin can be approximated by the probability of

ruin within finite time t provided t is sufficiently large, hence (1.4).
Question 9.2
What is lim ψ (u, t ) ?

u→∞
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.
1.4 The probability of ruin in discrete time
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:
ψ h (U ) = P [U (t ) < 0, for some t , t = h, 2h, 3h,… ]
y h (U , t )=P [U (t )<0, for some t , t =h, 2h,… ,t - h, t ]
Note that it is assumed for convenience in the definition of ψ h (U , t ) that t is an

integer multiple of h. Figure 2 shows the same realisation of the surplus process
as given in Figure 1 but assuming now that the process is checked only at
discrete time intervals. The black markers show the values of the surplus
process at integer time intervals (ie h = 1); the black markers together with the
white ones show the values of the surplus process at time intervals of length ½.
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
Explain why Equation 1.9 is true.
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 The Poisson and compound Poisson processes
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.
2.2 The Poisson process
The Poisson process is an example of a counting process. Here the number of

claims arising from a risk is of interest. Since the number of claims is being
l q
counted over time, the claim number process N (t ) t ≥ 0 must satisfy the
following conditions:
(i) N (0) = 0 , ie there are no claims at time 0
(ii) for any t > 0 , N (t ) must be integer valued
(iii) when s < t , N (s ) £ N (t ) , ie the number of claims over time is non-

decreasing
(iv) when s < t , N (t ) - N (s ) represents the number of claims occurring in the

time interval ( s , t ] .
The claim number process lN(t )qt ≥ 0 is defined to be a Poisson process with
parameter l if the following conditions are satisfied:
(i) N (0) = 0 , and N (s ) £ N (t ) when s < t
(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
Explain how motor insurance claims could be represented by a Poisson process.
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:
Let p n (t ) = P (N (t ) = n) . We want to show that:
( 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 .
This will be proved by deriving and solving a “differential-difference” equation.
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:
pn (t + h) = pn -1(t )[ l h + o(h )] + pn (t )[1 - l h + o(h )] + o(h )
= l hpn -1(t ) + [1 - l h ]pn (t ) + o(h)

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)
and this identity holds for n = 1,2,3,… .
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
When n = 0 , an identical analysis yields:
d
p0 (t ) = - l p0 (t ) (2.6)
dt
Solve for pn (t ) by introducing the probability generating function G (s , t ) defined

by:
∞
G( s, t ) = ∑ s n pn (t )
n=0
In Subject CT3 we would have expressed this as GN (t ) ( s ) = E[ s N (t ) ] . Note that we use

s as the dummy variable for the generating function in order to avoid confusion with t
denoting time. We are assuming here that s is a fixed number. By definition the
expected value of a function of a discrete random variable, say g ( X ) , is
E[ g ( X )] = Â g ( x) P( X = x) . Applying this formula with g ( N (t ) ) = s N (t ) gives us the
x
expression used in Core Reading.

So, differentiating with respect to t gives:
∞
d d
dt
G( s, t ) = ∑
sn
dt
p n (t )
n=0
Now multiply (2.5) by sn and sum over all values of n to get:
• • •
d
Â sn
dt
pn (t ) = l Â s n pn -1(t ) - l Â s n pn (t )
n =1 n =1 n =1
Now add (2.6) to the above identity to get:
• • •
d
Â s
dt
n
pn (t ) = l Â
s n pn -1(t ) - l Â
s n pn (t )
n =0 n =1 n =0
which can be written as:
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 Â 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 )
where c(s) is some function of 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 ).

c(s) can be identified by noting that when t = 0 , p0 (t ) = 1 and pn (t ) = 0 for

n = 1,2,3,… , since there are no claims at time zero.
Hence G (s ,0) = 1 and log G (s ,0) = 0 = c ( s ) .
Thus:
G( s, t ) = exp{λt ( s − 1)}
which is the probability generating function for the Poisson distribution with
parameter λt .
This generating function is given on page 7 of the Tables.
Since there is a one-to-one relationship between probability generating functions

and distribution functions (you should recall this from Subject CT3), the
distribution of N(t) is Poisson with parameter λt .
This is what we were trying to prove.
This study of the Poisson process concludes by considering the distribution of

the time to the first claim and the times between claims.
Time to the first claim
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:
P (T1 > t ) = P (N (t ) = 0) = exp{ - l t }
This last step follows from equation (2.3).
And:
P (T1 ≤ t ) = 1 − exp{− λt }
so that T1 has an exponential distribution with parameter λ . This is because the

RHS matches the formula for the distribution function of an exponential distribution.

Important information
The time to the first claim in a Poisson process has an exponential distribution with
parameter l .
Time between claims
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 .
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.
2.3 The compound Poisson process
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.
The following three important assumptions are made:
● the random variables { X i } i∞= 1 are independent and identically distributed

●
● the random variables { X i } i∞= 1 are independent of N (t ) for all t ≥ 0
●
● l q
the stochastic process N (t ) t ≥ 0 is a Poisson process whose parameter
is denoted l .
●
It was shown in Section 2.2 that this last assumption means that for any t ≥ 0 ,
the random variable N (t ) has a Poisson distribution with parameter λt , so that:
( 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 .

Note the slight change in terminology here: “Poisson parameter λ ” becomes

“Poisson parameter λt ” when a change is made from the process to the
distribution.
The common distribution function of the X i s will be denoted F ( x ) and it will be

assumed for the remainder of this chapter that F (0) = 0 so that all claims are for
positive amounts.
Remember that F ( x) is defined to be P( X £ x) . In the continuous case we would find

F ( x) by integrating the probability density function (pdf):
x
F ( x) = Ú f (t ) dt
-•
The probability density function of the X i s, if it exists, will be denoted f ( x ) and

the kth moment about zero of the X i s, if it exists, will be denoted mk , so that:
mk = E [ X ik ] for k = 1, 2, 3,…
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 .
Since, for a fixed value of t , S(t ) has a compound Poisson distribution, it

follows from Chapter 7 that the process {S(t )} t ≥ 0 has mean λtm1 , variance
λtm 2 , and moment generating function MS (r ) , where:
MS (r )= exp {l t (M X (r ) - 1)}

Remember from Chapter 7 that if:
S = X1 + X 2 + + XN
where N ~ Poi(l ) , then:
E[ S ] = l E[ X ] = l m1 var[ S ] = l E[ X 2 ] = l m2
M S (r ) = M N [ln M X (r )]
These formulae can be found on page 16 of the Tables.
Here N (t ) ~ Poi (l t ) , therefore E[ S (t )] = l tm1 and var[ S (t )] = l tm2 . Also
( )
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) )
For the remainder of this chapter the following (intuitively reasonable)

assumption will be made concerning the rate of premium income:
c > λm1 (2.8)
so that the insurer’s premium income (per unit time) is greater than the expected
claims outgo (per unit time).
Question 9.5
Why is this intuitively reasonable?
2.4 Probability of ruin in the short term
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:
U (1) = initial surplus + premiums - expenses - claims
= 0.1m + 100 ¥ 5, 000 - 100 ¥ 0.2 ¥ 5, 000 - S (1)

= 0.5m - S (1)
The distribution of S(1) is:
S (1) ~ N (100 × 0.7 × 5,000, 100 × (2.0 × 5,000) 2 ) = N [0.35m,(01

. m) 2 ]
So the probability that the surplus will be negative is:
P[U (1) < 0] = P[ S (1) > 0.5m]
= P ( N [0.35m, (0.1m)2 ] > 0.5m)
Ê 0.5m - 0.35m ˆ
= 1- FÁ ˜¯
Ë 0.1m
= 1 - F(1.5) = 1 - 0.93319 = 0.067
So the required probability is 6.7%.

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):
E [ S (2)] = 60e3+0.55 = 2, 088.80
and: var [ S (2)] = 60e6+ 2.2 = 218, 457
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 ˚
The probability of ruin is approximately 0.25%.

2.5 Premium security loadings
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.
Sometimes c will be written as:
c = (1 + θ )λm1
where θ ( > 0) is the premium loading factor.
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
What security loading is used in the example on page 20?

In summary:
Mean, variance and MGF of the total claim amount
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 )
The moment generating function of the process is given by:
M S (t ) (r ) = exp (l t ( M X (r ) - 1)
2.6 A technicality
In the next section a technical result will be needed concerning M X (r ) (the

moment generating function of the individual claim amount distribution), which,
for convenience, will be presented here. It will be assumed throughout the
remainder of this chapter that there is some number γ (0 < γ ≤ ∞ ) such that
M X (r ) is finite for all r < γ and:
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

This tends to infinity as r tends to α from below.
In the next section the following result will be needed:
lim (l M X (r ) - cr ) = • (2.10)
-
r Æg
If g is finite, (2.10) follows immediately from (2.9).
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.
This probability will be denoted by π . Then:
M X ( r ) ≥ e rε π
This follows by considering the claims below and above ε :
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 +∞ .
lim (l M X (r ) - cr ) = •
r Æ•

3 The adjustment coefficient and Lundberg’s inequality

This section will look at the probability of ruin and introduce the adjustment coefficient,
a parameter associated with risk. The letters R and r will be used interchangeably for
the adjustment coefficient.
3.1 Lundberg’s inequality
Lundberg’s inequality states that:
ψ (U ) ≤ exp{− RU }
where U is the insurer’s initial surplus and y (U ) is the probability of ultimate

ruin. R is a parameter associated with a surplus process known as the
adjustment coefficient. Its value depends upon the distribution of aggregate
claims and on the rate of premium income. Before defining R the importance of
the result and some features of the adjustment coefficient will be illustrated.
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
In the actuarial literature, exp{ -RU } is often used as an approximation to ψ (U ) .
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.
Figure 2 shows a graph of R as a function of the loading factor, θ , when:
(i) the claim amount distribution is exponential with mean 10, and
(ii) all claims are of amount 10.

Figure 2
Note that in both cases, R is an increasing function of θ . This is not surprising

as ψ (U ) would be expected to be a decreasing function of θ , and since
l q
ψ (U ) ≈ exp − RU , any factor causing a decrease in ψ (U ) would cause R to
increase.
Question 9.8
Why would ψ (U ) be expected to be a decreasing function of θ ?
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.

3.2 The adjustment coefficient – compound Poisson processes
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.
The adjustment coefficient, denoted R , is defined to be the unique positive root

of:
l M X (r ) - l - cr = 0 (3.1)
So, R is given by:
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
so that R is independent of the Poisson parameter and simply depends on the

loading factor, θ , and the individual claim amount distribution.
You can see from this equation that all the λ ’s have cancelled.
The adjustment coefficient can be found by solving the equation:
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
The equation for the adjustment coefficient is:
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:
(1 - 500r ) -1 = 1 + 1.3 ¥ 500r = 1 + 650r
Rearranging:
1 = (1 - 500r )(1 + 650r )
1 = 1 - 500r + 650r - 325, 000r 2

0 = 150 - 325, 000r
fi r = 0.000462
From Lundberg’s inequality, y (U ) £ e - rU , so here:
y (U ) £ e -0.000462¥1,000 = 0.630
Note that we didn’t use the Poisson parameter in our solution.
We will see another example of finding the adjustment coefficient shortly.

Question 9.9
In the previous example we ignored the fact that r could be 0. Why?
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.
Define g (r ) = λM X (r ) − λ − cr and consider the graph of g (r ) over the interval

[0, γ ] . Note first that g( 0) = 0 .
This is what we discovered in the previous example.
Further, g (r ) is a decreasing function at r = 0 since:
d d
g (r ) = λ M X (r ) − c
dr dr
so that the derivative of g (r ) at r = 0 is λm1 − c which is less than zero by

assumption (3.8).
Recall that M X′ (0) gives the mean of X , which is E[ X ] or m1 .
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
which is always strictly positive.
The second derivative can be written:
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
Thus there is a unique positive number R satisfying equation (3.1).
Equation (3.1) is an implicit equation for R . For some forms of F ( x ) it is

possible to solve explicitly for R ; otherwise the equation has to be solved
numerically.
Consider the exponential distribution where F ( x ) = 1 − e −αx .
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
since R is the positive root of (3.1).
(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 .
If the equation for R has to be solved numerically, it is useful to have a rough

idea of R ’s value. Equation (3.2) can be used to find a simple upper bound for
R as follows:
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.

Alternatively, we could present this proof as:
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 ˚
So that (c − λm1 )R > 21 λR 2 m 2 , giving:
R < 2(c − λm1 ) / λm 2 (3.4)
so that R < 2θ m1 / m 2 when c = (1 + θ )λm1 . Notice that if the value of R is small,
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.
3.3 A lower bound for R
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

The inequality is proved through the series expansion of exp(RM ) :
•
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 .
Inequality (3.5) can be used to show that:
1
R> log(c / λm1 )
M
when individual claim amounts have a continuous distribution on (0, M ) . This is

the lower bound for R that we are trying to find.
The starting point is the equation defining R (ie Equation 3.2):
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
Other approximations for R can be found, particularly when R is small, by

truncating the series expansion of exp(Rx ) .
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?
An upper bound for the adjustment coefficient:
2(c - l m1 ) 2q m1
R< or R< when c = (1 + q )l m1
l m2 m2
A lower bound for the adjustment coefficient:
1 Ê c ˆ 1
R> log Á or R> log(1 + θ )
M Ë l m1 ¯˜ M
3.4 The adjustment coefficient – general aggregate claims

processes
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.
Let {Si } i∞= 1 be a sequence of independent identically distributed random

variables:
Si ∫ aggregate claims from a risk in time period i .

c is the constant premium charged to insure this risk

The following assumptions are made:
c > E [Si ] (3.6)
There is some number γ > 0 such that:
lim E [e r [Si - c ] ] = • (3.7)

r Æg -
Si has density function h( x ) , -• < x < • (3.8)
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.
Let f (r ) = E [e r (Si − c ) ] for −∞ < r < γ .
Then f (0) = 1 , f ′ (0) = E [Si − c ] < 0 , f ′′ ( x ) > 0 , and:
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

Comments: Suppose Si has a compound Poisson distribution with Poisson

parameter λ and claim size random variable X .
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
which is the same as (3.2).

4 The effect of changing parameter values on finite and

infinite time ruin probabilities
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 ) .
In this section the effect of changing parameter values on ψ (U , t ) and ψ (U ) will

be discussed.
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:
● the Poisson parameter for the number of claims is 1 (4.1)

● the expected value of an individual claim is 1 (4.2)
● individual claims have an exponential distribution. (4.3)
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.
The advantage of using an exponential distribution for individual claims

(Assumption (4.3)) is that both exp( -RU ) and ψ (U ) can be calculated for these
examples. See Section 2 and Section 4.2.

4.2 A formula for ψ (U ) when F(x) is the exponential distribution
The formula for ψ (U ) when individual claims amounts are exponentially

distributed with mean 1, and when the premium loading factor is θ , is given by
the following result.
When F ( x ) = 1 - exp( - x ) :
1 Ê qU ˆ
y (U ) = exp Á - (4.4)
1+q Ë 1 + q ˜¯
The syllabus does not require this result to be derived or memorised.
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
Do you think that y (U , t ) is an increasing or decreasing function of t ?
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
What is the value of ψ (15) ?
Question 9.15
What is the value of e −15 R ?
The features of note in Figure 4 are:
(i) ψ (15, t ) is an increasing function of t ,
(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.
4.4 Ruin probability as a function of initial surplus
Question 9.16
Would you expect y (U , t ) to be an increasing or decreasing function of U ?
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
What is the value of ψ (20) ?

The features of note in Figure 5 are:
(i) the graphs all have the same general shape,
(ii) increasing the value of U decreases the value of ψ (U , t ) for any value of
t,
(iii) each of the three graphs approaches an asymptotic limit as t increases

(as has already been noted for U equal to 15 in the discussion of
Figure 4). Note that ψ (20) = 0.1476 and ψ (25) = 0.0937 .
ψ (U ) is a non-increasing function of U .
In the case of exponentially distributed individual claim amounts, the derivative

with respect to U of ψ (U ) is:
d −θ
ψ (U ) = ψ (U )
dU 1+ θ
which is negative since θ > 0 . Hence ψ (U ) is a decreasing function of U .
It is intuitively clear that ψ (U , t ) (of which ψ (U ) is a special case) should be a

decreasing function of U . An increase in U represents an increase in the
insurer’s surplus without any corresponding increase in claim amounts. Thus,
an increase in U represents an increase in the insurer’s security and hence will
reduce the probability of ruin.

4.5 Ruin probability as a function of premium loading
Question 9.18
Would you expect y (U , t ) to be an increasing or decreasing function of q ?
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

It is clear by general reasoning that ψ (U ) must be a non-increasing function of

θ . In the case of exponential individual claim amounts, ψ (U ) is a decreasing
function of θ .
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 shows ψ (10) as a function of θ .
Figure 7

4.6 Ruin probability as a function of the Poisson parameter
Figure 8 shows ψ (15,10) as a function of λ for three values of the premium

loading factor, q = 0.1, 0.2 and 0.3. This graph is identical to Figure 6 apart from
the labelling of the x-axis. This can be explained by considering the following
two risks.
Figure 8
Risk 1: aggregate claims are a compound Poisson process with Poisson

parameter 1 and F ( x ) = 1 - e - x . The premium income per unit time
to cover this risk is (1 + θ ) .
Risk 2: aggregate claims are a compound Poisson process with Poisson

parameter 0.5 and F ( x ) = 1 - e - x . The premium income per unit
time to cover this risk is 0.5(1 + θ ) .
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 )
but it has no effect on the probability of ultimate ruin (put t = • in the

relationship above) so that:
ψ * (U ) = ψ (U )
The point made in (iii) above was that:
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:
E [S * (1)] = 150 and var[S * (1)] = 300
These are calculated as λm1 = 150 × 1 = 150 and λm2 = 150 × 2 = 300 , where
E ( X 2 ) = 2 for an Exp(1) distribution.

Now, using tables of the normal distribution:
y * (15,1) = P [S * (1) > 15 + 1.2 ¥ 150]
= P [(S * (1) - 150) / 17.32 > 45 / 17.32]
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.
4.7 Concluding remarks
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 .
We’re thinking here of substituting θ = 0 into equation 4.4.
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
is the same as:
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 Reinsurance and ruin
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.
5.2 Proportional reinsurance
Let us consider the idea of a proportional reinsurance approach by way of an example:
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:
RP = (1 + q R )la m1 = 1.5 ¥ 30 ¥ 0.2 ¥ e3.55 = 313.32
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:
E[ Snet (2)] = 60 ¥ 0.8 ¥ e3.55 = 1, 671.04
and: var[ Snet (2)] = 60 ¥ 0.82 ¥ e8.2 = 139,812.49
So ruin will occur if:
Snet (2) > 1, 000 + 2, 400 - 626.64 = 2, 773.36
Using a normal distribution approach as before, we have:
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
Comment on the usefulness of reinsurance in this context.

You should remember that if we have a retained proportion a then:
Y =aX Z = (1 - a ) X
Hence:
E[Y ] = E[a X ] = a E[ X ] E[ Z ] = E[(1 - a ) X ] = (1 - a ) E[ X ]
Question 9.21
Write down expressions for var[Y ] and var[ Z ] .
5.3 Excess of loss reinsurance
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.
If we have a retention limit M , and no upper limit, then:
ÏX X < M Ï 0 X <M
Y =Ì Z =Ì
ÓM X ≥ M ÓX - M X ≥ M
Y can be also be written as min( X , M ) , and Z can also be written as max(0, 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
The formula for an expectation is Ú xf ( x) dx . We have to calculate E[Y ] by carrying

x
out two integrals, to allow for the two different ranges of X :
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
Using integration by parts:
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 .

5.4 Maximising the adjustment coefficient under proportional

reinsurance
First consider the effect of proportional reinsurance with retention α on the

insurer’s adjustment coefficient. Throughout Section 5.4 the insurer’s premium
income per unit time, before payment of the reinsurance premium, will be written
as (1 + θ )λm1 , which represents the expected aggregate claims per unit time for
the compound Poisson process with a loading factor θ . It will also be assumed
that the reinsurance premium is calculated as (1 + ξ )(1 − α )λm1 . Since the
reinsurer pays proportion 1 − α of each claim, (1 − α )λm1 represents the
reinsurer’s expected claims per unit time.
Thus, ξ is the premium loading factor used by the reinsurer. Hence, the
insurer’s premium income, net of reinsurance, is:
[(1 + θ ) − (1 + ξ )(1 − α )]λm1 (5.1)
Question 9.23
Explain this formula.
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.
For the insurer’s premium income, net of reinsurance, to be positive:
1 + θ > (1 + ξ )(1 − α )
ie α > (ξ − θ ) / (1 + ξ )
Question 9.24
What range of values is possible for α if θ = 0.2 and ξ = 0.4 ?
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?
Equation (5.2) specifies the insurer’s minimum retention level since:
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 .
The distribution of the insurer’s individual claims net of reinsurance is

exponential with parameter 0.1 / α . This can be seen by noting that if Y = αX ,
then:
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?

Hence, the equation defining R (see formula (4.2)) is:
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
It can be seen that R is a decreasing function of α . This is sensible as the

smaller the retention, the smaller the risk for the insurer and so ψ (U ) would be
expected to increase, and R to decrease, with α .
Different loadings
Now consider what happens when ξ > θ . For the rest of Section 5.4 assume
that:
● the individual claim amount distribution is F ( x ) = 1 - e -0.1x
● the insurer’s premium loading factor is θ = 0.1 .
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 .
This comes from Equation 5.1.
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
Derive this formula for R .

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
Derive this formula for R .
As when the loading factors were equal, the adjustment coefficient depends on
the retention level.
The value of α that maximises R in (5.5) is sought.
Differentiate R with respect to α (using the quotient rule for differentiation) to

give:
dR 20(12a 2 - a ) - (2a - 1)10(24a - 1)

=
da 100(12a 2 - a )2
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:
20(12α 2 − α ) = (2α − 1)10(24α − 1)
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.

Remember that α must lie in the range (0.5,1).
Consider the following values:
α R
0.5 0
0.9564 0.00911
1.0 0.00909
This shows that R has a maximum in [0.5,1] at 0.9564.
Alternatively you can show that the second derivative is negative.
Figure 10 shows R as a function of α (as given by (5.5)) for values of α greater

than 0.84. This range of α values has been chosen to highlight the important
features of the graph. The dotted line shows the value of R when α = 1 (ie no
reinsurance).
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 .

In terms of maximising the adjustment coefficient, the optimal retention level is

α = 0.9564 . It should be noted, however, that optimality in one sense does not
imply optimality in another. For example, if the insurer does not effect
reinsurance, then the expected profit per unit time is θλm1 (ie λ , since θ = 0.1
and m1 = 10 ).
θλm1 is just the “loading” bit.
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 .
If we put α = 0.9564 , this gives 0.9128λ , as stated above.
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.
The calculations are very similar to the previous case.
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
which leads to:
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:
dR 30(13α 2 − 2α ) − (3α − 2)10(26α − 2)

=
dα 100(13α 2 − 2α ) 2
and the function has a turning point when:
30(13α 2 − 2α ) = (3α − 2)10(26α − 2)
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.
It is not always possible to increase the value of the adjustment coefficient by

effecting reinsurance. Note that when an insurer effects reinsurance, this
reduces the variability of the insurer’s aggregate claims. A reduction in
variability is associated with an increase in the value of the adjustment
coefficient. However, when ξ > θ , the insurer’s premium loading factor, net of
reinsurance, decreases, and the value of the adjustment coefficient is expected
to decrease with the loading factor. When the reinsurer’s premium loading
factor was 0.3, the reduction in the insurer’s security caused by the reduction in
the loading factor has a greater effect on the adjustment coefficient than the
increase resulting from reinsurance for all values of α .
Loading factor (net of reinsurance)
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:
[(1 + θ ) − (1 + ξ )(1 − α )]λm1 − αλm1
This is just an algebraic expression for net premiums less expected net claims.

So the loading factor is:
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.
5.5 Maximising the adjustment coefficient under excess of loss

reinsurance
In this section the effect of excess of loss reinsurance on the adjustment

coefficient will be considered. The following assumptions will be made for
Section 5.5:
● the insurer’s premium income (before reinsurance) per unit time is

(1 + θ )λm1
● the reinsurance premium per unit time is (1 + ξ )λE (Z ) , where ξ ( ≥ θ ) is the

reinsurer’s premium loading factor, and Z = max(0, X - M ) .
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
In Figure 13, R goes to • as M goes to 0.
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:
-1 + 1.2M - 0.03M 2 > M - 0.025M 2
fi M 2 - 40M + 200 < 0
fi 5.8579 < M < 34.1421
Hence, the minimum retention level is 5.8579. Similarly, when ξ = 0.4 , the
minimum retention level is 10.
Figure 12 shows R as a function of M for the following combinations of θ and

ξ:
(a) θ = 0.1 and ξ = 0.2 (solid line)
(b) θ = 0.1 and ξ = 0.4 (dotted line).
Without reinsurance, ie for M = 20 , the insurer’s adjustment coefficient is 0.014

(irrespective of the reinsurer’s loading factor).
From Figure 12, it can be seen that, for ξ = 0.2 :
R (M ) > R (20) for 9.6 < M < 20
R (M ) < R (20) for M < 9.6
and for ξ = 0.4 :
R (M ) < R (20) for R < 20

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
Use the approximation e x ≈ 1 + x + x 2 / 2 to find an approximate numerical value for the

adjustment coefficient for the previous example in the case where β = 0.05 , θ 1 = 0.3 ,
θ 2 = 0.4 and M = 10 .

We conclude this chapter with two exam-style questions:
A general insurance company is planning to set up a new class of travel insurance. It

plans to start the business with £2 million and expects claims to occur according to a
Poisson process with parameter 50. Individual claims are thought to have a gamma
distribution with parameters a = 150 and l = ¼ . A premium loading factor of 30% is
applied.
Explain how each the following changes to the company’s model will affect the
probability of ultimate ruin:
(i) A 28% premium loading factor is applied instead. [1]
(ii) Individual claims are found to have a gamma distribution with parameters
a = 150 and l = ½ . [1]
(iii) The Poisson parameter is now believed to be 60. [1]

[Total 3]
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:
P[U (2) < 0] [5]

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
ψ (u) = P[U (t ) < 0 for some t ] (continuous time)
ψ (u, t0 ) = P[U (t ) < 0 for some t ≤ t0 ]
ψ h (u) = P[U (t ) < 0 for some t = h,2h,3h,… ] (discrete time)
ψ h (u, t0 ) = P[U (t ) < 0 for some t = h,2h,3h,… and t ≤ t0 ]
Poisson process
( λt ) x e − λt
P[ N (t ) = x ] = px (t ) = ( x = 0,1,2,… )
x!
f T (t ) = λe − λt (t > 0)
Compound Poisson process
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 )
Upper and lower bounds for 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 )
If reinsurance is effected this equation becomes:
1 + (1 + θ ) E ( X ) − (1 + ξ ) E ( Z ) r = M Y (r )
where ξ denotes the reinsurer’s security loading, Y represents the amount of an

individual claim paid by the insurer, net of reinsurance, and Z denotes the amount of an
individual claim paid by the reinsurer.
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
(iii) Yes, because if we expand e - x as a power series and simplify, we get:
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:
P( N < 2) = P( N = 0) + P ( N = 1) = e −5 + 5e −5 = 0.040 ie 4.0%
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:
U (2) = initial surplus + premiums - expenses - claims

= 0.1m + (100 + 200) ¥ 5, 000 - (100 + 200) ¥ 0.2 ¥ 5, 000 - S (2)
= 1.3m - S (2)
The distribution of S (2) is:
S (2) ~ N (300 ¥ 0.7 ¥ 5, 000, 300 ¥ (2.0 ¥ 5, 000)2 ) = N [1.05m, (0.173m) 2 ]
The probability that the surplus will be negative at the end of the second year is:
P [U (2) < 0] = P [ S (2) > 1.3m]

= P ( N [1.05m, (0.173m) 2 ] > 1.3m)
Ê 1.3m - 1.05m ˆ
= 1- F Á
Ë 0.173m ˜¯
= 1 - F(1.443) = 0.074
So the probability of insolvency at the end of the second year is 7.4%.
Solution 9.7
Using the information given in the example, we have:
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?
Consider Lundberg’s inequality, y (U ) £ e - RU . If R = 0 , we have an upper bound of 1

for the probability of ruin. That should have been obvious!
In practice, ignore R = 0 , just as we have done in the example.
Solution 9.10
The adjustment coefficient satisfies:
1 + (1 + θ ) m1R = M X ( R)
The distribution of the individual claim sizes X is:
X=
RS10,000 with probability 0.9
T25,000 with probability 0.1
So:
E ( X ) = Â xP( X = x) = 0.9 ¥ 10, 000 + 0.1 ¥ 25, 000 = 11,500
and:
M X ( R) = E (e RX ) = Â e Rx P( X = x) = 0.9 e10,000 R + 0.1 e25,000 R
The security loading is θ = 0.2 .
So the equation for the adjustment coefficient is:
. × 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 :
At R = 0.00002599 : 1 + 13,800 R − (0.9e10,000 R + 01

. e 25,000 R ) = 0.000035
At R = 0.00002601 : 1 + 13,800 R − (0.9e10,000 R + 01

. e 25,000 R ) = −0.000018
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:
E ( X 2 ) = Â x 2 P( X = x) = 0.9 ¥ 10, 0002 + 0.1 ¥ 25, 0002 = 152,500, 000
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
This is probably easiest if we start by writing:
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
Differentiating y (U ) with respect to θ :
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
This is the required expression.
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
var[Y ] = var[a X ] = a 2 var[ X ]
var[ Z ] = var[(1 - a ) X ] = (1 - a ) 2 var[ X ]
Solution 9.22
To find var[ Z ] , we need to find E[ Z 2 ] , since var[ Z ] = E[ Z 2 ] - E 2 [ Z ] .
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
The insurer will charge a premium of:
(1 + q )l E[ X ] = (1 + q )l m1
The reinsurer will charge a premium of:
(1 + x )l E[ Z ]
But E[ Z ] = (1 - a ) E[ X ] = (1 - a )m1 , so the reinsurer’s premium is:
(1 + x )l (1 - a )m1
So the net premium received by the insurer is the difference:
(1 + q )l m1 - (1 + x )l (1 - a )m1 = {(1 + q ) - (1 + x )(1 - a )} l m1
Solution 9.24
0.2
α> = 01429
.
14
.
But since α cannot exceed 1, the possible range of values is 01429

. < α ≤ 1.
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 )
With reinsurance this will become:
l + cnet r = l M Y (r )
for the direct insurer. But we know that cnet = (1 + q )l E[ X ] - (1 + x )l E[ Z ] , so the

equation for R becomes:
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 )
In the case of proportional reinsurance this is:
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:
1 + (1.1 ¥ 10 - 1.2(1 - a ) ¥ 10) r = M Y (r )
fi 1 + (11 - 12(1 - a ) ) r = M Y (r )
fi 1 + (12a - 1)r = M Y (r )
But what about 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
Therefore the equation is:
1
1 + (12a - 1)r =
1 - 10ra
Solution 9.28
Dividing the previous equation through by λ gives:
1
1 + (12α − 1) R =
1 − 10αR
Multiplying through by 1 − 10αR gives:
(1 + 12αR − R)(1 − 10αR) = 1
Multiplying out the left hand side and subtracting 1 from both sides gives:
(10α − 120α 2 ) R 2 + (2α − 1) R = 0
Dividing through by R and rearranging gives the required expression.

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
The adjustment coefficient equation is λ + cR = λ M X ( R) . The net rate of premium

income for the direct insurer equals the rate of premiums charged to the policyholder
minus the rate of premiums paid to the reinsurer:
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
So the equation for the adjustment coefficient is:
1 1
λ + λ [(1 + θ 1 ) − (1 + θ 2 )e − β M ]R = λ [β − R e−( β − R) M ]
β β−R
Cancelling λ ’s and multiplying through by β ( β − R) gives:
β ( β − R ) + ( β − R )[(1 + θ1 ) − (1 + θ 2 )e − β M ]R = β [ β − R e − ( β − R ) M ]
Cancelling the β 2 ’s from both sides gives:
− βR + ( β − R)[(1 + θ 1 ) − (1 + θ 2 )e − β M ]R = − β R e − ( β − R ) M
Cancelling R’s to exclude the trivial solution gives:
− β + ( β − R )[(1 + θ 1 ) − (1 + θ 2 )e − β M ] = − β e − ( β − R ) M
The adjustment coefficient R is the smallest positive solution of this equation.

Solution 9.31
Using the values given, the adjustment coefficient equation becomes:
−0.05 + (0.05 − R)(13 . e −0.5 ) = −0.05e −0.5+10 R

. − 14
Multiplying by −20e 0.5 to clear some of the fractions gives:
e 0.5 − (1 − 20 R)(13
. e 0.5 − 14
. ) = e10 R
Expanding the LHS and applying the approximation to the RHS:
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:
−4.8668 + 4.86682 − 4( −50)( −0.09462)

R= = 0.0268
2( −50)

Solution to Exam-style question 1
(i) Reduction in premium loading factor
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.
(ii) Change in claims distribution
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.
(iii) Change in the Poisson parameter
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.

Solution to Exam-style question 2
The annual premium is 120% of E ( S ) , where S is the aggregate claim amount in a

single year. Since S has a compound Poisson distribution with Poisson parameter 0.2,
we have E ( S ) = 0.2 E ( X ) where:
E ( X ) = 50 ¥ 0.7 + 100 ¥ 0.3 = 65
Hence the annual premium is:
1.2 E ( S ) = 1.2 ¥ 0.2 E ( X ) = 1.2 ¥ 0.2 ¥ 65 = 15.6
The initial surplus is 75, so the surplus at time 2 is:
U (2) = 75 + 2 ¥ 15.6 - S (2) = 106.2 - S (2)
So the probability of ruin is:
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
Hence P [ S (2) < 106.2] = 0.9647 fi P [U (2) < 0] = 0.0353

CT6-10: Generalised linear models Page 35
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.
Exam-style question 1 (Subject 106, April 2003, Question 3)
(i) A random variable Y has density of exponential family form:
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]
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

Page 36 CT6-10: Generalised linear models
You are given that:
10 15 35
Â yi = 11 Â yi = 3 Â yi = 4
i =1 i =11 i =16
It wishes to use a Poisson model to analyse these data.
(i) Show that the Poisson distribution is a member of the exponential family of
distributions. [2]
(ii) The insurer decides to use a model (Model A) for which:
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]
You can assume that f ( y ) = y log y is equal to zero when y = 0 .
(v) Compare Model A directly with Model B, by calculating an appropriate test

statistic. [2]
[Total 15]

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 Â 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.
Solution to exam-style question 1 (Subject 106, April 2003, Question 3)
(i) Mean and variance
We have:
E[Y ] = b ¢(q ) var[Y ] = a (f )b ¢¢(q )
(ii)(a) Exponential form
The PDF of the exponential distribution with mean m is:
1 Ï y¸
f ( y) = exp Ì - ˝
m Ó m˛
This can be written as an exponential:
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
(ii)(b) Natural parameter and variance function

The natural parameter is q , so here the natural parameter is:
1
-
m
The variance function is (by definition) b ¢¢(q ) , so here we find:
1 1
b′(θ ) = − b′′(θ ) = = μ2
θ θ 2
(i) Exponential family
For the Poisson distribution, we have:
f ( y) = e- m m y / y !
We wish to write this in the form:
È yq - b(q ) ˘
g ( y ) = exp Í + c( y, f ) ˙
Î a(f ) ˚
So, rearranging the Poisson formula:
È y log m - m ˘
f ( y ) = exp Í - log y !˙
Î 1 ˚
We can see that this has the correct form if we write:
q = log m b(q ) = m = eq a (f ) = f = 1 c( y, f ) = - log y !

(ii) Maximum likelihood estimates
Using the rearranged form for the Poisson distribution from part (i), we see that the log
of the likelihood function can be written:
log L( m I , m II , m III ) = Â yi log mi - Â mi - Â log yi ! (*)
This now becomes, for Model A:
10 15 35 35
log L = a Â yi + b Â yi + g Â 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
Differentiating this log-likelihood function in turn with respect to a , b and g , we get:
∂
log L = 11 - 10ea
∂a
∂
log L = 3 - 5e b
∂b
and:
∂
log L = 4 - 20eg
∂g
Setting each of these expressions equal to zero in turn, we find that:
aˆ = log1.1 = 0.09531
bˆ = log 0.6 = -0.51083
and: gˆ = log 0.2 = -1.60944
These are our maximum likelihood estimates for a , b and g .

(iii) Simpler model
In this case the log-likelihood function reduces to:
35 35 35
log L = a Â yi - 35ea - Â log yi ! = 18a - 35ea - Â 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 ˜¯
(iv) Scaled deviance for Model A and Model B
The scaled deviance for Model A is given by:
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:
log LS = Â yi log yi - Â yi - Â log yi !
= 4 ¥ 2 log 2 - 18 - 4 log 2 = 4 log 2 - 18 = -15.2274
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
= 11log1.1 + 3log 0.6 + 4 log 0.2 - 11 - 3 - 4 - 4 log 2
= -27.6944

The corresponding value for log LA without the final term is –24.9218.
So the scaled deviance is twice the difference in the log likelihoods:
SD = 2(log LS - log LA ) = 2 (( -15.2274) - ( -27.6944) ) = 24.93
as required.
We now repeat the process for Model B.
Using Equation (***), the log likelihood for Model B is:
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 value without the final term is –29.9696.
The scaled deviance is again twice the difference in the log likelihoods:
SD = 2(log LS - log LB ) = 2 (( -15.2274) - ( -32.7422) ) = 35.03
(v) Comparing A with B
We can use the chi-squared distribution to compare Model A with Model B. We

calculate the difference in the scaled deviances (which is just 2(log LA - log LB ) ):
35.03 - 24.93 = 10.10
This should have a chi-squared distribution with 3 - 1 = 2 degrees of freedom, which

has a critical value at the upper 5% level of 5.991. Our value is significant here, since
10.10 > 5.991 , so that Model A is a significant improvement over Model B. We prefer
Model A here.




using the course.

CT6-14: Monte Carlo simulation Page 27
Relative error
A similar argument can be applied when the relative error qˆ - q q is used to

measure the accuracy. In this case the requirement which must be satisfied is
that:
tˆ
za / 2 <e
|q | n
or, written more accessibly:
n > za2 / 2tˆ 2 /(e 2q 2 )
Of course the exact value of q is unknown, but the estimate qˆ may be used
instead of q in this equation.

Page 28 CT6-14: Monte Carlo simulation
Exam-style question 1 (Subject 103, April 2003, Question 4)
(i) Given a pseudo-random number U uniformly distributed over [0,1] , obtain an

expression in terms of U and θ for a non-negative pseudo-random variable X
which has density function:
f ( x) = q e -q x [2]
(ii) A sequence of simulated observations is required from the density function:
e -q x
g ( x) = k (q ) , x>0
1+ x
where θ is a non-negative parameter and k (θ ) is a constant of integration not

involving x .
(a) Describe a procedure that applies the Acceptance-Rejection method to

obtain the required observations.
(b) Derive an expression involving θ and k (θ ) for the expected number of

pseudo-random variables required to generate a single observation from
the density g using this method. [6]
[Total 8]

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.
(i) Find the value of b . [2]
(ii) Show that this distribution has mean 500, and variance 500,000. [2]
(iii) Find the distribution function of the distribution. [2]
(iv) Comment on the ease with which we could use the inversion method to generate
random numbers from this distribution. [1]
(v) By using a 2-parameter Pareto distribution with a = 3 and l = 500 as a

bounding distribution, generate as many random numbers as possible from this
three-parameter Pareto distribution. You should use the random numbers from a
uniform (0,1) distribution given below (in the order given). [6]
[Total 13]
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

7 Appendix –Linear Congruential Generators (LCGs)

To use an LCG we need to choose three positive integers:
• a, the multiplier
• c, the increment
• m, the modulus m > a, m > c .
To obtain the desired sequence of random numbers u1 , u2 ,… , u N , first generate a

sequence of integers x1 , x2 ,… , xN in the range {0,1, 2,… , m - 1} starting from an initial
value x0 , known as the seed. The sequence of integers is generated using the recursive
rule:
xn = (axn -1 + c ) (mod m ) , n = 1, 2,… , N
ie xn is the remainder when axn -1 + c is divided by m. After that, un is set equal to

xn m .
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
Use this algorithm with parameter values a = 11 , c = 37 , m = 100 and seed x0 = 85 to

generate 5 random numbers from the U (0,1) distribution. How suitable do you think
this method would be for generating random numbers?

Solution 14.4
We can generate 3 variates from Y by using the transformation: Y = 4U + 1 . By using

the first members of the pairs, this gives: 1.5, 4.86, 2.548 . We need to decide
whether to accept or reject each of these.
We have:
3
f ( x ) 32 (1 - x )( x - 5) 1
g ( x) = = = (1 - x )( x - 5)
Ch ( x ) 3 1
¥ 4
2 4
It follows that: g (1.5) = 0.4375, g (4.86) = 0.1351, g ( 2.548) = 0.948924
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:
0.342 < 0.4375, 0.274 ≥ 0.1351, 0.812 < 0.948924
It follows that we accept 1.5, we reject 4.86 and we accept 2.548.
Solution 14.5
Our z values are:
z1 = -2 log 0.802 cos(2p ¥ 0.450) = -0.632
z2 = -2 log 0.802 sin(2p ¥ 0.450) = 0.205
You could have used the two initial values the other way round in which case you would
get different answers.
Using the polar method, we have:
v1 = 2 ¥ 0.802 - 1 = 0.604 v2 = 2 ¥ 0.450 - 1 = -0.1
fi s = 0.6042 + ( -0.1) 2 = 0.374816

So our z values are:
-2 log 0.374816
z1 = ¥ 0.604 = 1.382
0.374816
-2 log 0.374816
and: z2 = ¥ -0.1 = -0.229
0.374816
These are our values from the standard normal distribution.
If we want values from a N (3,8) distribution, then we need to find values x1 and x2
where:
x = 3+ z 8
using the 2 z values we have just calculated.
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
Solution to exam-style question (Subject 103, April 2003, Question 4)
(i) Pseudo-random variable
We can use the inverse transform method here.
The distribution function for this exponential distribution is (from the Tables):
F ( x) = 1 - e -q x
Using fact that F ( X ) ~ U (0,1) , we get:
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.
First, we need to find the scaling constant C , which is:
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
The algorithm for generating pseudorandom values with PDF g ( x) is:
1. Take a value ( = x ) generated from the exponential distribution in part (i).
2. Take a new (independent) pseudorandom value ( = u ) from U (0,1) .
1
3. If u < , accept the value x ; otherwise, reject it.
1+ x
4. If more values are required, go back to step 1. Otherwise, stop.
(ii)(b) Expected number
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

(i) Value for β
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!
We could use integration by parts or by substitution to find b using Ú f ( x)dx = 1 , but it

is clearly much quicker just to match the distribution with one given in the Tables.
(ii) Mean and variance
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
These are the required values.
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.

(iii) Distribution function
This is not in the Tables, so we will have to integrate. We have:
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
where b = 1.5 ¥ 109 as before. [Total 2]
Note that the distribution function tends to 1 as x tends to infinity, as expected.
You do not have to integrate by parts here if you prefer substitution. Substituting
u = 500 + t will also work.
Alternative solution to exam-style question 2

du
Substituting u = 500 + t in the integral, and noting that = 1 , so that du in the
dt
integral can just be replaced by dt :
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).
(iv) Inverting the distribution function
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.
(v) Acceptance-rejection method
First we want the probability density function of the two parameter Pareto distribution:
3 ¥ 5003
h( x ) =
(500 + x) 4

We can now calculate C :
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.
Inverting the distribution function of our two-parameter distribution:
3
Ê 500 ˆ
u ( x) = 1 - Á ˜ fi x = 500 ÈÎ (1 - u ) -1/ 3 - 1˘˚
Ë 500 + x ¯
Substituting in u = 0.74 , we find that x = 283.39 .
We will accept this if u1 < g (283.39) , where u1 = 0.26 . Since

0.26 < g (283.39) = 0.3617 , we accept this random number.
Repeating the process:
x = 500 ÈÎ (1 - 0.82) -1/ 3 - 1˘˚ = 385.55
But u2 = 0.59 > g (385.55) = 0.4354 , so we reject this random number.

Finally:
x = 500 ÈÎ (1 - 0.16) -1/ 3 - 1˘˚ = 29.920
But u3 = 0.71 > g (29.92) = 0.0565 , so we again reject this random number.
So we have a sample containing the single value 283.39.




using the course.

CT6: Q&A Bank Part 2 – Questions Page 3
Question 2.4 (Exam-style)
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.
(i) Derive the maximum likelihood estimator, l̂ , of l . [3]
(ii) Suppose that past experience shows that l has an exponential distribution with
1
mean .
n
(a) Derive the posterior distribution of l .
(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]

Page 4 CT6: Q&A Bank Part 2 – Questions
(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

{ X ij } , i = 1, 2, … , N , j = 1, 2, … , n is a set of data where X ij represents the aggregate

claims in the j th year from the i th risk in a collective. It is assumed that this set of
data satisfies all the assumptions for Model 1 of Empirical Bayes Credibility Theory.
(i) Carefully describe the assumptions common to the distributions of X ij and

X km ( j π m ) in the two cases when:
(a) X ij and X km come from the same risk, so that i = k
(b) X ij and X km come from different risks, so that i π k . [6]
(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 Â ( 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]

Question 2.9 (Developmental)
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 Â ( 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]

An actuarial student is using Empirical Bayes Credibility Theory Model 2 to calculate

credibility premiums for a group of insurers. He has analysed the data for six different
insurers, using 10 years of past data for each insurer. He has obtained the following
figures:
6 10
ÂÂ Pij = 1, 498 P* = 18.24
i =1 j =1
E [m(q )] = 4.00 var [m(q )] = 42.1 E ÈÎ s 2 (q ) ˘˚ = 62.8
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,
Â YKj = 986 and Â 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
Question 2.11 (Bookwork)
(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
Calculate the unconditional mean and standard deviation of Y . [3]
(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)
Obtain the mean and variance of the marginal distribution of U . [2]

[Total 5]
A compound distribution S = X1 + X 2 + + X N has claim number distribution:
P( N = n) = 9(n + 1)4 − n −2 , n = 0, 1, 2,…
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]

CT6: Q&A Bank Part 2 – Solutions Page 9
Solution 2.6
(i) Since the X j ’s are identically distributed, we have:
E ( X ) = E ( X j ) = E[ E ( X j q )] = E[m(q )] [2]
(ii) Since m(q ) is a function of q only, it follows that:
( ) ( )
E ÈÎ Xm(q ) ˘˚ = E E ÈÎ Xm(q ) q ˘˚ = E m(q ) E ÈÎ X q ˘˚ = E ÈÎ m 2 (q ) ˘˚ [2]
(iii) Expressing var( X ) in terms of conditional expectations and variances gives:
var( X ) = E[var( X q )] + var[ E ( X q )]
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
In EBCT Model 1, we assume that:
(a) when i = k :
(1) X ij and X km are identically distributed but not necessarily independent.

[2]
(2) If q i is the risk parameter associated with risk i , X ij q i and X km q i are

independent and identically distributed. [2]
(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]

Page 10 CT6: Q&A Bank Part 2 – Solutions
(ii) Empirical Bayes credibility premium
First we calculate the overall mean:
1
X = (81.4 + 82.4 + + 143.8) = 102.36 [1]
5
Secondly we calculate the mean of the individual sample variances:
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 ˘
Â 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
So our estimators are:
E[m(q )] = 102.36 E[ s 2 (q )] = 590.70 [1]

1
var[m(q )] = 781.59 - ¥ 590.7 = 663.45 [1]
5
So the credibility factor is:
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:
P1 = Z X1 + (1 - Z ) E[m(q )] = 0.84885 ¥ 81.4 + 0.15115 ¥ 102.36 = 84.57

[1]
[Total 8]

Solution 2.8
Using the formula:
n
Z=
n + E[ s 2 (q )] / var[m(q )]
we get:
5
Z= = 0.539 [2]
5 + 620 /145
Solution 2.9
For Insurer 4 we have:
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
So we have the following estimates:
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
Â ( X i - X )2 - Â
Nn n - 1
Â ( X ij - X i ) 2
i =1 j =1
1 1
= (45.42 + + 862 - 4 ¥ 52.42 ) - ¥ 29.325 = 727.975 [2]
3 5

So the credibility factor is given by:
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 ¯
So the EBCT premiums are given by:

Insurer 1: 0.9920 ¥ 45.4 + 0.0080 ¥ 52.4 = £45.46m
Insurer 2: 0.9920 ¥ 20.8 + 0.0080 ¥ 52.4 = £21.05m
Insurer 3: 0.9920 ¥ 57.4 + 0.0080 ¥ 52.4 = £57.36m
Insurer 4: 0.9920 ¥ 86.0 + 0.0080 ¥ 52.4 = £85.73m [2]
[Total 9]
Solution 2.10
(i) Updated estimates and the credibility premium
Since X is our estimate of E [m(q )] , we have:
6 10
Â Â Pij X ij 6 10
i =1 j =1
4.00 = 6 10
fiÂ Â Pij X ij = 4 ¥ 1, 498 = 5,992 [1]
Â Â Pij i =1 j =1
i =1 j =1
From the data given for Insurer I:
10 10 10
Â YIj = Â PIj X Ij = 1,133 and Â PIj = 287 [1]
j =1 j =1 j =1
So our new estimate for E [m(q )] is:
7 10
Â Â Pij X ij
i =1 j =1 5,992 + 1,133
X = = = 3.9916 [1]
7 10 1, 498 + 287
Â Â Pij
i =1 j =1

Now consider E ÈÎ s 2 (q ) ˘˚ . The formula for the estimator for E ÈÎ s 2 (q ) ˘˚ is:
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
Since N = 6 (initially) and n = 10 , we know that:
6 10
Â Â 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 Â Â Pij ( X ij - X i )2 from Insurer I. This
i =1 j =1
10
is Â PIj ( X Ij - X I )2 . We can calculate this by rewriting it as:
j =1
10 10 10 10
Â PIj ( X Ij - X I )2 = Â PIj X Ij2 - 2 X I Â PIj X Ij + X I2 Â PIj
j =1 j =1 j =1 j =1
10
Â PIj X Ij
j =1
But since X I = 10
, the last two terms can be added together to give:
Â PIj
j =1
10 10 10
Â PIj ( X Ij - X I )2 = Â PIj X Ij2 - X I2 Â 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 ˜¯
= 4,539.0874 - 4, 472.7840 = 66.3034 [2]
7 10
So the new total for Â Â Pij ( X ij - X i )2 is:
i =1 j =1
3,391.2 + 66.3034 = 3, 457.5034 . [1]

Our new estimate for E ÈÎ s 2 (q ) ˘˚ is:
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
Â Pij ( X ij - X )2 - Â Â ij ij i
N i =1 n - 1 j =1
P ( X - X ) 2
˜ [1]
j =1 ¯
Substituting in the numbers we know, we get:
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 ¯
This tells us that:
6 10 6 10 6 10
Â Â Pij ( X ij - X ) 2
=Â Â Pij X ij2 -X 2
Â Â Pij
i =1 j =1 i =1 j =1 i =1 j =1
= 59 (42.1 ¥ 18.24 + 62.8) = 49, 011.536 [1]
It follows from this that:
6 10
Â Â 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
Â Â Pij X ij2 = 72,979.536 + 4,539.0874 = 77,518.6234 [1]
i =1 j =1

So:
7 10
Â Â Pij ( X ij - X )2 = 77,518.6234 - 3.99162 ¥ (1, 498 + 287)
i =1 j =1
= 49, 078.449 [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* = Â i P ˜¯ Nn - 1 Á Â i Â P ˜
P
Nn - 1 i =1 ÁË
1 - = P -
Ë i =1 i =1 ¯
6
Now we know that Â Pi = P = 1, 498 , so:
i =1
Ê 6 ˆ
Á
1 Á
Â Pi 2 ˜
18.24 = 1, 498 - i =1 ˜
59 Á 1, 498 ˜
ÁÁ ˜˜
Ë ¯
6
This gives a value for Â Pi 2 of 631,916.32.
i =1
7
So Â 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
Â 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 ¯
So our new estimate for var [m(q )] is:
1 Ê 1 ˆ
ÁË ( 49, 078.449) - 54.881˜¯ = 32.70533 [1]
20.07015 69

So our three parameter estimates are:
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.
The credibility factor for Insurer I is:
287
ZI = = 0.994187 [1]
54.881
287 +
32.705
So the credibility premium per unit volume is:
1,133
CP = 0.994187 ¥ + 0.005813 ¥ 3.9916 = 3.94799 [1]
287
So the credibility premium for the coming year is:
3.94799 ¥ 38 = 150.024 [1]

[Total 20]
(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]
(ii) Calculate the value of the adjustment coefficient. [4]

[Total 7]

(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 )
and define what each of the symbols represents.
(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.
(e) Comment briefly on the effect of the reinsurance on the probability of

ruin.
[14]
[Total 24]

Claims occur on a portfolio of insurance policies according to a Poisson process with

Poisson parameter l . Claim amounts, X1 , X 2 ,… , are assumed to be identically
distributed with moment generating function M X (t ) . The insurer calculates premiums
using a loading factor q (> 0) . The insurer’s adjustment coefficient, R , is defined to be
the smallest positive root of the equation:
l + cr = l M X (r )
where c is the insurer’s premium income rate.
(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
where m = E[ X i ] and s 2 = var[ X i ] . [4]
(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
(i) Show that f ( x ) can be written in the form of an exponential family of

distributions. [1]
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]
(ii) Determine the variance function of Yi . [2]

[Total 6]

Rearranging this inequality gives:
P S < 800,000(1 + i )½ − 52,500(1 + i ) −½ = 0.90
Standardising, we now have:
È S - E ( S ) 800, 000(1 + i )½ - 52,500(1 + i ) -½ - 748,103.05 ˘

PÍ < ˙ [2]
ÍÎ var( S ) 1.0639755 ¥ 109 ˙˚
S - E (S )
But is approximately standard normal. So we must have:
var( S )
800, 000(1 + i )½ - 52,500(1 + i ) -½ - 748,103.05

= 1.28155
9
1.0639755 ¥ 10
800, 000(1 + i )½ - 52,500(1 + i ) -½ - 748,103.05 = 41,802.415
800, 000(1 + i ) - 789,905.46(1 + i )½ - 52,500 =0 [2]
Solving this quadratic in (1 + i )½ gives:
½ 789,905.46 ± 789,905.462 − 4 × 800,000 × ( −52,500)

(1 + i ) =
2 × 800,000
= 10499
. (or − 0.0625)
2
So: i = 10499
. − 1 = 01023
. ie the required interest rate is 10.23% per annum.
[3]

Solution 3.8
The adjustment coefficient satisfies the inequality:
2[c / l - E( X )] 2[300, 000 / 50 - 5, 000]

r< 2
= = 0.000064 [2]
E( X ) 5, 0002 + 2,5002
Since we don’t know the precise distribution of the individual claim amounts, this is the
best we can do.
Solution 3.9
The adjustment coefficient equation is:
l + cr = l E (e rX )
The average claim size is:
E ( X ) = 0.9 ¥ 2, 000 + 0.1 ¥ 15, 000 = £3,300 [1]
So the annual risk premium is:
l E ( X ) = 0.03 ¥ 3,300 = £99
So the annual office premium is c = 1.3 ¥ 99 + 30 = £158.70 . [1]
So the adjustment coefficient equation is:
0.03 + 158.70r = 0.03(0.9e 2,000 r + 0.1e15,000 r ) [1]
Simplifying by dividing by 0.03 and writing R = 1, 000r gives:
1 + 5.29 R = 0.9e2 R + 0.1e15 R [1]
Expanding the RHS as a series to get a first approximation:
1 + 5.29 R = 0.9(1 + 2 R + 2 R 2 + ") + 0.1(1 + 15 R + 112.5 R 2 + ")
= 1 + 3.3R + 13.05 R 2 + "

5.29 - 3.3
So: Rª = 0.1525 [1]
13.05
Evaluating f ( R) = 0.9e2 R + 0.1e15 R - 1 - 5.29 R for different values of R , we get:
f (0.15) = 0.37 f (0.1) = 0.018 f (0.09) = - 0.013
f (0.094) = - 0.0015 f (0.095) = 0.00156 f (0.0945) = 0.000011 [2]
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
(i) Moment generating function
The moment generating function is given by:
• • •
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]
(ii) Adjustment coefficient
For the adjustment coefficient, we require:
1 + (1 + q ) E ( X )r = M X (r )

To find E ( X ) , we can differentiate the MGF:
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
Alternatively we can note that this distribution is a mixture of two exponential

distributions. So the mean is given by a weighted average of two exponential means, ie
1 ¥ 1 + 1 ¥ 1 = 5 . Or we could integrate from first principles.
2 3 2 7 21
So E ( X ) = 5 and the equation for r is:

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 )
Multiplying through by (3 - r )(7 - r ) and multiplying out brackets, we obtain:
63 - 9r - 7 r 2 + r 3 = 63 - 15r
Gathering up the terms, we obtain:
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
(i)(a) Adjustment coefficient
The adjustment coefficient r is the smallest positive solution of the equation:
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
Subtracting a l from both sides:
r2
c r > l [r E ( X ) + E ( X 2 )] [1]
2
Dividing by r (which must be a positive number):
r
c > l [E( X ) + E ( X 2 )] [1]
2
Rearranging to get an inequality for r gives:
2[c / l - E ( X )]
r< [1]
E( X 2 )

(i)(b) Probability of ruin
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]
(ii)(a) Adjustment coefficient
The adjustment coefficient equation is:
l + cr = l E (erX )
Since individual claims have an exponential distribution with mean 8,000:
E (e rX ) = M X (r ) = 1/(1 - 8000r ) ( t < 1/ 8000 )
So the adjustment coefficient satisfies:
100 + 1, 000, 000 r = 100 /(1 - 8000 r ) [1]
Dividing by 100:
1 + 10, 000 r = 1/(1 - 8000 r )
So: (1 + 10, 000 r )(1 - 8000 r ) = 1
1 + 2, 000 r - 80, 000, 000 r 2 = 1 [1]
Cancelling the 1’s and factorising:
2, 000 r (1 - 40, 000 r ) = 0
The adjustment coefficient is the smallest positive solution ie r = 1/ 40, 000 = 0.000025
[1]

(ii)(b) Verify inequality
Since X has an exponential distribution:
E ( X ) = 8, 000
and:
E ( X 2 ) = var( X ) + [ E ( X )]2 = 8, 0002 + 8, 0002 = 2 ¥ 8, 0002
So the inequality in (i)(a) states that:
2[1, 000, 000 /100 - 8, 000]

r< = 0.00003125 [1]
2 ¥ 8, 0002
The exact value of r (ie 0.000025) is indeed less than this.
(ii)(c) Adjustment coefficient with reinsurance
The adjustment coefficient equation (with the reinsurance in effect) is:
l + cnet r = l E (er X net ) [1]
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:
cnet = 1, 000, 000 - 80, 000 = 920, 000 [1]

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
So the adjustment coefficient equation becomes:
È 1 - e 20,000r - 2.5 ˘
100 + 920, 000r = 100 Í + e 20,000 r - 2.5 ˙ [1]
ÍÎ 1 - 8, 000r ˙˚
Multiplying through by 1 - 8, 000r and dividing through by 100 gives:
(1 + 9, 200r )(1 - 8, 000r ) = 1 - e 20,000 r - 2.5 + e 20,000 r - 2.5 (1 - 8, 000r )
Expanding and cancelling the 1’s:
1, 200r - 73, 600, 000r 2 = -8, 000 r e20,000 r - 2.5
Dividing by 100r :
12 - 736, 000r = -80 e 20,000 r - 2.5

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
f (0.000035) = -0.54 [1]
So r is approximately 0.000034 (measured in units of £ -1 ). [1]
(ii)(d) Probability of ruin
So the probability of ruin with and without the reinsurance are approximately:
y without ª e -0.000025¥ 20,000 = 0.61 and y with ª e -0.000034¥ 20,000 = 0.51

[1]
(ii)(e) Comment
So the reinsurance reduces the probability of ultimate ruin. [1]

[Total 14]

Solution 3.12
(i) Approximation to R
The MGF of X can be written:
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!
Substituting this into the defining equation for R :
R2 2
l + cR = l[1 + R m + ( m + s 2 )]
2!
Subtracting l from both sides and rearranging, we get:
l R2
(c - lm ) R = (m 2 + s 2 ) [1]
2!
Dividing through by R and simplifying, we get:
2(c - lm ) 2(c / l - m )
R= = [1]
l(m 2 + s 2 ) s 2 + m2
This is the required expression. [Total 4]

(ii) Assessment of reinsurance
The adjustment coefficient can be used to assess the effectiveness of different

reinsurance arrangements, using Lundberg’s inequality to find an upper bound for the
probability of ruin for the insurer under different reinsurance arrangements. An
arrangement that produces a lower upper bound for the probability of ruin is in some
sense more secure for the insurer than an arrangement that has a higher upper bound for
the probability of ruin. [3]
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.
(iii) Maximum reinsurance
First we consider the insurer’s security without reinsurance. The equation for the
adjustment coefficient is:
1 + (1 + q )m1r = M X (r )
Substituting in q = 0.15 , m1 = 5000 and M X (r ) = e5000 r , we get:
1 + 5750 r = e5000 r [1]
The rate of premium income is:
c = (1 + q )l m1 = 1.15 ¥ 20 ¥ 5000 = 115, 000
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]
The MGF of the insurer’s net claim payments is now:
M X (t ) = e5000(1- k ) R [1]

So the equation for the insurer’s adjustment coefficient is now:
1 + (5750 - 6250k ) R = e5000(1- k )t [1]
Using the same approximation to R as before, we have:
2[(5750 - 6250k ) - 5000(1 - k )] 3 - 5k

Rª 2 2
= [1]
5000 (1 - k ) 50000(1 - k ) 2
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 )
Rearranging this inequality, we get:
3k 2 - k < 0 [1]
The solution of this is 0 < k < 1/ 3 . So the maximum proportion to be reinsured is

33 13 %. [1]
[Total 9]

CT6: Assignment X1 Solutions Page 5
Solution X1.5
(i) Bayes criterion solution
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]
(ii) Minimax and maximax solutions
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]
The best possible outcomes associated with d1 and d 2 are:
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]

Page 6 CT6: Assignment X1 Solutions
Solution X1.6
The likelihood function for m is given by:

2 2 2
1 Ê ln xn - m ˆ 1 Ê ln xn - m ˆ 1 Ê ln xi - m ˆ
1 - Á
2Ë s ˜¯ 1 - Á
2Ë s ˜¯ - Â
2 ÁË s ˜¯
L( m ) = e ¥ ¥ e μe [1]
x1s 2p xns 2p
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]
So the PDF of the posterior distribution for m is given by:

2 2
1 Ê ln xi - m ˆ 1 Ê ln xi - m ˆ
- ÂÁ
2 Ë s ¯
˜ - ÂÁ
2 Ë s ¯
˜
f ( m | x ) μ prior ¥ likelihood = constant ¥ e μe [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):
Â ((ln xi )2 - 2 m ln xi + m 2 )
1 1
Â (ln xi - m )
2
- -
f ( m | x) μ e 2s 2 =e 2s 2
-
1
2s 2
(Â (ln x ) -2 m Â ln x +nm )
i
2
i
2
=e [½]
-
1
2s 2
(-2 m Â ln x +nm ) i
2
μe
m -2 m Â 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* =
Â ln xi and s *2 =
s2
[½]
n n

CT6: Assignment X2 Questions Page 1
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 )]
Explain how changes in the values of n, E[ s 2 (q )] and var[m(q )] affect the

value of Z and comment on why Z behaves in this way. [3]
(ii) The table below shows the aggregate claim statistics for each of four risks over
three years:
3 3
Risk i Â X ij Â ( 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
Â X 5 j = 3,999 and Â ( X 5 j - X 5 )2 = 21, 734 [2]
j =1 j =1
[Total 9]

Page 2 CT6: Assignment X2 Questions
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
S is a compound Poisson random variable with Poisson parameter 3 and individual

claim size distribution P( X = 1) = P( X = 2) = P( X = 3) = 1 / 3. T is a compound
Poisson random variable with Poisson parameter 2 and individual claim size distribution
P( X = 1) = P( X = 2) = 1 / 2 . S and T are independent. If U = S + T , determine the
individual claim size distribution of U . [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.
Calculate the unconditional mean and standard deviation of Y . [5]

Assignment X2 Solutions
Solution X2.1
(i) Effect on Z
The credibility factor Z is calculated as:
credibility premium = Z ¥ X i + (1 - Z ) ¥ X
where n is the number of years of past data. As n increases, Z increases. So more

emphasis is put on the particular risk. This makes sense, since the more information we
have from the relevant risk, the less emphasis we will wish to place on the collateral
data. [1]
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]
(ii) Credibility premium
The means for each of the risks are:
X1 = 728 X 2 = 907 X 3 = 1,150 X 4 = 1, 033

The estimators for the parameters are:
E[m(q )] ª X = 14 (728 + 907 + 1,150 + 1, 033) = 954.5 [1]
3 4
E[ s 2 (q )] ª 1
4 Â Â ( X ij - X i )2
1
2
i =1 j =1
= 14 (11,172 + 10,147 + 10,900 + 11,997)
= 11, 054 [1]

3 È 3 4 ˘
var[m(q )] ª 1
3 Â i ( X - X ) 2
- 3 4 Â 2Â
1 Í1
Í
1 ( X ij - X i ˙
) 2
˙˚
i =1 Î i =1 j =1
= 13 ÈÎ (728 - 954.5) 2 + (907 - 954.5) 2 + (1,150 - 954.5) 2
+ (1, 033 - 954.5) 2 ˘˚ - 13 ¥ 11, 054
= 13 ¥ 97,941 - 13 ¥ 11, 054
= 28,962.3 [1]
So, the credibility factor is:
3
Z= = 0.88714
11, 054
3+
28,962.3
Thus the credibility premium for risk 3 is:
0.88714 ¥ 1,150 + 0.11286 ¥ 954.5 = 1,128 [1]

[Total 4]
(iii) Addition of a fifth risk
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
An analyst at a general insurance company is examining claims data on a portfolio of

home insurance policies in a particular region. An exponential distribution models the
claim amounts and the following rating factors are used:
SA sum assured, x (as a continuous variable)

PT property type, Ti (as a factor with i = 1, 2, … ,10 )
NB number of bedrooms, B j (as a factor with j = 1, 2, … ,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
(i) Complete the table. [3]
(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]

Page 8 CT6: Assignment X3 Questions
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 .
(i) (a) Define the adjustment coefficient, R .
(b) Define the surplus process and the probability y (u ) of ruin with initial
surplus u > 0 .
(c) Write down Lundberg’s inequality. [3]
x
-
1 m
(ii) Derive the adjustment coefficient if f ( x) = e , x > 0 , and q = 0.25 . [4]
m
(iii) Consider the case where f ( x) = ½ e - x (1 + 2 e - x ) , x > 0 , and q = 0.25 .
(a) Calculate the expected claim size m .
(b) Calculate the corresponding adjustment coefficient, and determine an

upper bound for y (15) .
(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
The underwriters at a general insurance company assume that household buildings

insurance claims fall into one of two tiers: “small” and “large”. The ranges for the tiers
and the proportion of claims falling in each range are as follows:
Range Proportion of claims

“small” (£2,000, £20,000) 95%
“large” (£20,000, £200,000) 5%
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.
(i) (a) Calculate c , the premium charged for each policy.

(b) Calculate the expected size of the surplus in one year’s time. [4]
(ii) Show that, if all amounts are expressed in units of £000s, the adjustment
coefficient r satisfies an equation of the form:
r + 19.14r 2 = p1e 2 r + p2 e 20r + p3e 200r
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]
(v) Comment on the magnitude of your probability in (iv). [1]

[Total 18]




using the course.

(ii) Comparing models
Comparing SA+PT with SA

The difference in the scaled deviances is 238.4 - 206.7 = 31.7
2 2
This is greater than 16.92, the upper 5% critical value of a c11- 2 = c 9 distribution.
So SA + PT is a significant improvement over the SA model. [1]
Alternatively, using the D(deviance) > 2 ¥ D(parameters) approximation, we get

31.7 > 2 ¥ 9 so the SA+PT model is a significant improvement over the SA model.
Comparing SA*PT with SA+PT

2 2
This is greater than 16.92, the upper 5% critical value of a c 20 -11 = c 9 distribution.
So SA * PT is a significant improvement over the SA + PT model. [1]

28.4 > 2 ¥ 9 so the SA * PT model is a significant improvement over the SA+PT model.
Comparing SA*PT+NB with SA*PT

2 2
This is greater than 11.07, the upper 5% critical value of a c 25 - 20 = c 5 distribution.
So SA * PT + NB is a significant improvement over the SA * PT model. [1]

12.1 > 2 ¥ 5 so the SA * PT model is a significant improvement over the SA * PT+NB
model.
Comparing SA*PT*NB with SA*PT+NB

2 2
This is less than 118.7, the upper 5% critical value of a c120 - 25 = c 95 distribution.
2 2
We have to interpolate in the tables between c 90 and c100 to get the figure of 118.7.
So SA * PT * NB is not a significant improvement over the SA * PT+NB model. [1]

107.3 > 2 ¥ 95 so the SA * PT * NB model is not a significant improvement over the
SA * PT+NB model.
So the analyst should choose the SA * PT+NB model. [1]

(iii) Further information
The analyst should also check
• that the SA * PT+NB model is a significant improvement when the order is

different, eg add the NB factor before the PT factor [½]
• other models involving these rating factors, eg SA * NB+PT [½]
• the residuals of the proposed model (to ensure that it is a good fit to the data) [½]
• the significance of the parameters of the proposed model (to ensure that all the
estimated parameters are significantly different from zero) [½]
Solution X3.7
(i)(a) Adjustment coefficient
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.
(i)(b) Surplus process and probability of ruin
The surplus process is:
U (t ) = u + (1 + q ) l m1 t - S (t )
Here, S (t ) is the aggregate amount of claims received up to time t . The term

(1 + q )l m 1 t represents the total premiums received to time t . [½]
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 < • .
ie: y (u ) = P(U (t ) < 0 for some t , 0 < t < •) [½]

(i)(c) Lundberg’s inequality
Lundberg’s inequality states that:
y (u ) £ e - Ru
where R is the adjustment coefficient and u is the initial surplus. [1]

[Total 3]
(ii) Derivation
The moment generating function (MGF) of this distribution is:
M ( R) = (1 - m R) -1
So the equation for R in this case is:
1 + 1.25 m R = (1 - m R ) -1 [1]
Multiplying through by 1 - m R , we get:
(1 + 1.25 m R)(1 - m R) = 1 [1]
Multiplying out the brackets and simplifying, we find that:
1
R= [2]
5m
This is the value of the adjustment coefficient in this case. [Total 4]
(iii)(a) Expected claim size
The expected claim size m is now:
• • •
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]
Alternatively we can integrate by parts to obtain the same answer.
(iii)(b) Upper bound
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]
We can write this as:
4 - 3R
1 + 0.9375R =
(2 - 2 R)(2 - R )
Multiplying through by (2 - 2 R)(2 - R) , gathering terms and simplifying, we get:
1.875 R3 - 3.625 R 2 + 0.75R = 0 [3]
Dividing through by R and solving the resulting quadratic equation gives us

R = 0.235610 (or a number greater than one, which cannot be the solution since one of
the MGF’s is not defined for values greater than one). [1]

So the adjustment coefficient is R = 0.2356 , and the upper bound for the probability of
ruin is:
e -15 R = 0.0292 [½]
(iii)(c) Exponential distribution
The corresponding results for the exponential distribution are:
1
R= = 0.26667 [½]
5 ¥ 3/ 4
This gives an upper bound for the probability of ruin of:
e -15 R = 0.0183 [½]
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
The risk premium per policy is:
FG 22,000 220,000 IJ
H
0.02 × 0.95 ×
2
+ 0.05 ×
2 K
= £319 [1½]
Incorporating the security loading gives:
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:
1,000,000 + 10,000 × 6380

. = £1,638,000 [2]
[Total 4]
(ii) Adjustment coefficient
The adjustment coefficient is found from the equation:
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).
The moment generating function is (again working in £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]
Evaluating the integrals:
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
We can write the MGF as:
1
M X (r ) =
r
o
p1e 2r + p2 e 20r + p3e 200r t [2]

So we find that r is the solution of the equation:

1
0.02 + 0.3828r = 0.02 ×
r
o
p1e 2r + p2 e 20r + p3e 200r t [1]
Dividing by 0.02 and multiplying by r gives the required equation:
r + 19.14r 2 = p1e 2r + p2e 20r + p3e 200r [1]

[Total 7]
(iii) Approximate value for r
Expanding the RHS as a series gives:
r + 19.14r 2 = p1[1 + (2r ) + 12 (2r ) 2 + 16 (2r )3 ]
+ p2 [1 + (20r ) + 12 (20r ) 2 + 16 (20r )3 ]
+ p3[1 + (200r ) + 12 (200r ) 2 + 16 (200r )3 ] + o(r 4 ) [1]
Using the numerical values for the p ’s, and summing down the columns, we find that:
r + 19.14r 2 = 0 + r + 15.95r 2 + 440.3r 3 + o(r 4 ) [1]
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:
19.14 = 15.95 + 440.3r + o(r 2 ) ⇒ r ≈ 7.245 × 10 −3 [2]

[Total 4]
(iv) Probability of ultimate ruin
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:
. × 10−3 × 1,000) = e −5.56 = 0.0038 or 0.38%

ψ (u) ≈ exp( −556 [2]
(v) Comment on part (iv)
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]




using the course.

CT6 CMP Upgrade 09-10

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

CT6 CMP Upgrade 09-10

Uploaded by

Copyright:

Available Formats

CT6: CMP Upgrade 2009/10 Page 1

This CMP Upgrade contains:

The Actuarial Education Company © IFE: 2010 Examinations

1 Changes to the Syllabus objectives and Core Reading

1.1 Syllabus objectives

Syllabus objectives (v) 8 to (v) 10 are new. These read:

8. Explain the empirical Bayes approach to credibility theory, in particular its

9. State the assumptions underlying the two models in 8.

10. Calculate credibility premiums for the two models in 8.

Chapter 6 (now Chapter 7)

Chapter 8 (now Chapter 9)

2. Define a Poisson process, derive the distribution of the number of events

© IFE: 2010 Examinations The Actuarial Education Company

4. Define the adjustment coefficient for a compound Poisson process and

5. Define the probability of ruin in infinite/finite and continuous/discrete

6. State Lundberg's inequality and explain the significance of the

8. Analyse the effect on the adjustment coefficient and hence on the

Chapter 10 (now Chapter 11)

Syllabus objective (vii) has been renumbered to Syllabus objective (vi).

Chapter 11 (now Chapter 10)

Syllabus objective (viii) has been renumbered to Syllabus objective (vii).

The Actuarial Education Company © IFE: 2010 Examinations

Syllabus objective (ix) has been renumbered to Syllabus objective (viii).

Syllabus objective (ix) has been renumbered to Syllabus objective (viii).

Syllabus objective (x) has been renumbered to Syllabus objective (ix).

1.2 Core Reading

The following Core Reading paragraph is no longer in Subject CT6:

This material is covered in detail in Subject ST3 – General Insurance Specialist

The following Core Reading paragraph is no longer in Subject CT6:

In this course we will study Bayesian credibility. Empirical Bayes credibility

© IFE: 2010 Examinations The Actuarial Education Company

Chapter 6 (now Chapter 7)

The following Core Reading paragraphs are no longer in Subject CT6:

Firstly, it is worth noting some overriding features of all types of insurance.

The Actuarial Education Company © IFE: 2010 Examinations

Chapter 8 (now Chapter 9)

© IFE: 2010 Examinations The Actuarial Education Company

2 Changes to the ActEd Course Notes

This chapter is new and is provided in the replacement pages.

Chapter 6 (now Chapter 7)

Chapter 8 (now Chapter 9)

Chapter 8 (now Chapter 9)

Chapter 11 (now Chapter 10)

The Actuarial Education Company © IFE: 2010 Examinations

© IFE: 2010 Examinations The Actuarial Education Company

3 Changes to the Q&A Bank

Q&A Bank Part 1

Solution 4.18 (i)

Hence, v1 = 0.1834 , v2 = 0.5872 and s = v12 + v22 = 0.37844 .

Solution 5.12 (ii)(a)

First note that:

The Actuarial Education Company © IFE: 2010 Examinations

Q&A Bank Part 2

Questions 2.7 to 2.9 have been deleted.

Q&A Bank Part 3

Questions 3.8 to 3.18 are now Questions 3.19 to 3.29.

Questions 3.19 to 3.24 are now Questions 3.13 to 3.18.

© IFE: 2010 Examinations The Actuarial Education Company

4 Changes to the X Assignments

Solution X1.10 (iii) & (iv)

Solution X1.10 (iii)

The Actuarial Education Company © IFE: 2010 Examinations

Solution X2.8 (now Solution X2.7)