This action might not be possible to undo. Are you sure you want to continue?

# Lecture Notes

on

Actuarial Mathematics

Jerry Alan Veeh

May 9, 2003

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§0. Introduction

The objective of these notes is to present the basic aspects of the theory of

insurance, concentrating on the part of this theory related to life insurance. An

understanding of the basic principles underlying this part of the subject will form a

solid foundation for further study of the theory in a more general setting.

Throughout these notes are various exercises and problems. The reader should

attempt to work all of these.

A calculator, such as the one allowed on the Society of Actuaries examinations,

will be useful in solving some of the problems here. The problems contained here

are not all amenable to solution using only this simple calculator. A computer

equipped with spreadsheet software will sometimes be useful, especially for the

laboratory exercises.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§1. Overview

The central theme of these notes is embodied in the question, “What is the value

today of a random sumof money which will be paid at a randomtime in the future?”

Such a random payment is called a contingent payment.

The theory of insurance can be viewed as the theory of contingent payments.

The insurance company makes payments to its insureds contingent upon the oc-

currence of some event, such as the death of the insured, an auto accident by an

insured, and so on. The insured makes premiumpayments to the insurance company

contingent upon being alive, having sufﬁcient funds, and so on. A natural way to

model these contingencies mathematically is to use probability theory. Probabilistic

considerations will, therefore, play an important role in the discussion that follows.

The other central consideration in the theory of insurance is the time value of

money. Both claims and premium payments occur at various, possibly random,

points of time in the future. Since the value of a sum of money depends on the

point in time at which the funds are available, a method of comparing the value

of sums of money which become available at different points of time is needed.

This methodology is provided by the theory of interest. The theory of interest will

be studied ﬁrst in a non-random setting in which all payments are assumed to be

sure to be made. Then the theory will be developed in a random environment, and

will be seen to provide a complete framework for the understanding of contingent

payments.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§2. Elements of the Theory of Interest

A typical part of most insurance contracts is that the insured pays the insurer

a ﬁxed premium on a periodic (usually annual or semi–annual) basis. Money has

time value, that is, $1 in hand today is more valuable than $1 to be received one year

hence. A careful analysis of insurance problems must take this effect into account.

The purpose of this section is to examine the basic aspects of the theory of interest.

A thorough understanding of the concepts discussed here is essential.

To begin, remember the way in which compound interest works. Suppose an

amount A is invested at interest rate i per year and this interest is compounded

annually. After n years the amount will be A(1+i)

n

. The factor (1+i)

n

is sometimes

called the accumulation factor. If interest is compounded daily after the same n

years the amount will be A(1 +

i

365

)

365n

. In this discussion the interest rate i is called

the nominal annual rate of interest. The effective rate of interest in the example

in which interest is compounded daily is (1 +

i

365

)

365

− 1. This is the rate of interest

which compounded annually would provide the same return.

Exercise 2–1. What is the effective rate of interest corresponding to an interest rate

of 5% compounded quarterly?

It is possible that two different investment schemes with two different nominal

annual rates of interest may in fact be equivalent, that is, may have equal dollar

value at any ﬁxed date in the future. This possibility is illustrated by means of an

example.

Example 2–1. Suppose I have the opportunity to invest $1 in Bank A which pays

5% interest compounded monthly. What interest rate does Bank B have to pay,

compounded daily, to provide an equivalent investment? At any time t in years the

amount in the two banks is given by

1 +

0.05

12

12t

and

1 +

i

365

365t

respectively. It

is now an easy exercise to ﬁnd the nominal interest rate i which makes these two

functions equal.

Exercise 2–2. Find the interest rate i. What is the effective rate of interest?

Situations in which interest is compounded more often than annually will arise

frequently. Some notation will be needed to discuss these situations conveniently.

Denote by i

(m)

the nominal annual interest rate compounded m times per year which

is equivalent to the interest rate i compounded annually. This means that

1 +

i

(m)

m

m

= 1 + i.

Exercise 2–3. Compute 0.05

(12)

.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§2: Elements of the Theory of Interest 5

An important abstraction of the idea of compound interest is the idea of con-

tinuous compounding. If interest is compounded n times per year the amount after

t years is given by

1 +

i

n

nt

. Letting n → ∞ in this expression produces e

it

, and

this corresponds to the notion of instantaneous compounding of interest. In this

context denote by δ the rate of instantaneous compounding which is equivalent

to interest rate i. Here δ is called the force of interest. The force of interest is

extremely important from a theoretical standpoint and also provides some useful

quick approximations.

Exercise 2–4. Show that δ = log(1 + i).

Exercise 2–5. Find the force of interest which is equivalent to 5% compounded

daily.

The converse of the problem of ﬁnding the amount after n years at compound

interest is as follows. Suppose the objective is to have an amount A n years hence.

If money can be invested at interest rate i, how much should be deposited today

in order to achieve this objective? It is readily seen that the amount required is

A(1 + i)

−n

. This quantity is called the present value of A. The factor (1 + i)

−1

is

often called the discount factor and is denoted by v.

Example 2–2. Suppose the annual interest rate is 5%. What is the present value

of a payment of $2000 payable in the year 2014? The present value (in 2004) is

$2000(1 + 0.05)

−10

= $1227.83.

The notion of present value is used to move payments of money through time in

order to simplify the analysis of a complex sequence of payments. In the simple case

of the last example the important idea is this. Suppose you were given the following

choice. You may either receive $1227.83 today or you may receive $2000 in the

year 2014. If you can earn 5% on your money (compounded annually) you should

be indifferent between these two choices. Under the assumption of an interest rate

of 5%, the payment of $2000 in 2014 can be replaced by a payment of $1227.83

today. Thus the payment of $2000 can be moved through time using the idea of

present value. A visual aid that is often used is that of a time diagram which shows

the time and amounts that are paid. Under the assumption of an interest rate of 5%,

the following two diagrams are equivalent.

Two Equivalent Cash Flows

2004 2014 2004 2014

$2000 $1227.83

............................................................................................................................................................................................................................................................................................................. .............................................................................................................................................................................................................................................................................................................

The advantage of moving amounts of money through time is that once all

§2: Elements of the Theory of Interest 6

amounts are paid at the same point in time, the most favorable option is readily

apparent.

Exercise 2–6. What happens in comparing these cash ﬂows if the interest rate is

6% rather than 5%?

In an interest payment setting, the payment of interest of i at the end of the period

is equivalent to the payment of d at the beginning of the period. Such a payment

at the beginning of a period is called a discount. What relationship between i and

d must hold for a discount payment to be equivalent to the interest payment? The

time diagram is as follows.

Equivalence of Interest and Discount

0 1 0 1

i

d

............................................................................................................................................................................................................................................................................................................. .............................................................................................................................................................................................................................................................................................................

The relationship is d = iv follows by moving the interest payment back in time

to the equivalent payment of iv at time 0.

Exercise 2–7. Denote by d

(m)

the rate of discount payable m times per year that is

equivalent to a nominal annual rate of interest i. What is the relationship between

d

(m)

and i? Between d

(m)

and i

(m)

? Hint: Draw the time diagram illustrating the two

payments made at time 0 and 1/ m.

Exercise 2–8. Treasury bills (United States debt obligations) pay discount rather

than interest. At a recent sale the discount rate for a 3 month bill was 5%. What is

the equivalent rate of interest?

The notation and the relationships thus far are summarized in the string of

equalities

1 + i =

1 +

i

(m)

m

m

=

1 −

d

(m)

m

−m

= v

−1

= e

δ

.

§2: Elements of the Theory of Interest 7

Problems

Problem 2–1. Show that if i > 0 then

d < d

(2)

< d

(3)

< ⋅ ⋅ ⋅ < δ < ⋅ ⋅ ⋅ < i

(3)

< i

(2)

< i.

Problem 2–2. Show that lim

m→∞

d

(m)

= lim

m→∞

i

(m)

= δ.

Problem 2–3. Calculate the nominal rate of interest convertible once every 4 years

that is equivalent to a nominal rate of discount convertible quarterly.

Problem 2–4. Interest rates are not always the same throughout time. In theoretical

studies such scenarios are usually modelled by allowing the force of interest to

depend on time. Consider the situation in which $1 is invested at time 0 in an

account which pays interest at a constant force of interest δ. What is the amount

A(t) in the account at time t? What is the relationship between A

′

(t) and A(t)? More

generally, suppose the force of interest at time t is δ(t). Argue that A

′

(t) = δ(t)A(t),

and solve this equation to ﬁnd an explicit formula for A(t) in terms of δ(t) alone.

Problem 2–5. Show that d = 1 − v. Is there a similar equation involving d

(m)

?

Problem 2–6. Show that d = iv. Is there a similar equation involving d

(m)

and i

(m)

?

§2: Elements of the Theory of Interest 8

Solutions to Problems

Problem 2–1. An analytic argument is possible directly from the formulas.

For example, (1 + i

(m)

/ m)

m

= 1 + i = e

δ

so i

(m)

= m(e

δ/ m

− 1). Consider m as a

continuous variable and show that the right hand side is a decreasing function

of m for ﬁxed i. Can you give a purely verbal argument? Hint: How does

an investment with nominal rate i

(2)

compounded annually compare with an

investment at nominal rate i

(2)

compounded twice a year?

Problem 2–2. Since i

(m)

= m((1 + i)

1/ m

− 1) the limit can be evaluated directly

using L’Hopitals rule, Maclaurin expansions, or the deﬁnition of derivative.

Problem 2–3. The relevant equation is

1 + 4i

(1/ 4)

1/ 4

=

1 − d

(4)

/ 4

−4

.

Problem 2–4. In the constant force setting A(t) = e

δt

and A

′

(t) = δA(t). The

equation A

′

(t) = δ(t)A(t) can be solved by separation of variables.

Problem 2–5. 1 − d

(m)

/ m = v

1/ m

.

Problem 2–6. d

(m)

/ m = v

1/ m

i

(m)

/ m.

§2: Elements of the Theory of Interest 9

Solutions to Exercises

Exercise 2–1. The equation to be solved is (1 + 0.05/ 4)

4

= 1 + i, where i is the

effective rate of interest.

Exercise 2–2. Taking t

th

roots of both sides of the equation shows that t plays

no role in determining i and leads to the equation i = 365((1+0.05/ 12)

12/ 365

−1) =

0.04989.

Exercise 2–3. 0.05

(12)

= 12((1 + 0.05)

1/ 12

− 1) = 0.04888.

Exercise 2–4. The requirement for equivalence is that e

δ

= 1 + i.

Exercise 2–5. Here e

δ

= (1 + 0.05/ 365)

365

, so that δ = 0.4999. So as a rough

approximation when compounding daily the force of interest is the same as the

nominal interest rate.

Exercise 2–6. The present value in this case is $2000(1+ 0.06)

−10

= $1116.79.

Exercise 2–7. A payment of d

(m)

/ m made at time 0 is required to be equivalent

to a payment of i

(m)

/ m made at time 1/ m. Hence d

(m)

/ m = v

1/ m

i

(m)

/ m. Since

v

−1/ m

= (1+i)

1/ m

= 1+i

(m)

/ mthis gives d

(m)

/ m = 1−v

1/ m

or 1+i = (1−d

(m)

/ m)

−m

.

Another relation is that d

(m)

/ m − i

(m)

/ m = (d

(m)

/ m)(i

(m)

/ m).

Exercise 2–8. The given information is d

(4)

= 0.05, from which i can be

obtained using the formula of the previous exercise as i = (1 − 0.05/ 4)

−4

− 1 =

0.0410.

§3. Annuities Certain

Many different types of ﬁnancial transactions involve the payment of a ﬁxed

amount of money at regularly spaced intervals of time for a predetermined period.

Such a sequence of payments is called an annuity certain or, more simply, an

annuity. A common example is loan payments. It is easy to use the idea of present

value to evaluate the worth of such a cash stream at any point in time. Here is an

example.

Example 3–1. Suppose you have the opportunity to buy an annuity, that is, for a

certain amount paid by you today you will receive monthly payments of $400, say,

for the next 20 years. How much is this annuity worth to you? Suppose that the

payments are to begin one month from today. Such an annuity is called an annuity

immediate (a truly unfortunate choice of terminology). It is useful to visualize the

cash stream represented by the annuity on a time diagram.

An Annuity Immediate

0 1 2 239 240

$400

↓

$400

↓

. . . . . .

$400

↓

$400

↓

Clearly you would be willing to pay today no more than the present value of the

total payments made by the annuity. Assume that you are able to earn 5% interest

(nominal annual rate) compounded monthly. The present value of the payments is

240

¸

j=1

(1 +

.05

12

)

−j

400.

This sum is simply the sum of the present value of each of the payments using the

indicated interest rate. It is easy to ﬁnd this sum since it involves a very simple

geometric series.

Exercise 3–1. Evaluate the sum.

Since expressions of this sort occur rather often, actuaries have developed some

special notation for this sum. Write a

n

for the present value of an annuity which

pays $1 at the end of each period for n periods. Then

a

n

=

n

¸

j=1

v

j

=

1 − v

n

i

where the last equality follows from the summation formula for a geometric series.

The interest rate per period is usually not included in this notation, but when such

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§3: Annuities Certain 11

information is necessary the notation is a

n i

. The present value of the annuity in the

previous example may thus be expressed as 400a

240 .05/ 12

.

A slightly different annuity is the annuity due which is an annuity in which the

payments are made starting immediately. The notation ¨ a

n

denotes the present value

of an annuity which pays $1 at the beginning of each period for n periods. Clearly

¨ a

n

=

n−1

¸

j=0

v

j

=

1 − v

n

d

where again the last equality follows by summing the geometric series. Note that n

still refers to the number of payments. If the present time is denoted by time 0, then

for an annuity immediate the last payment is made at time n, while for an annuity

due the last payment is made at time n − 1, that is, the beginning of the nth period.

It is quite evident that a

n

= v ¨ a

n

, and there are many other similar relationships.

Exercise 3–2. Show that a

n

= v ¨ a

n

.

The connection between an annuity due and an annuity immediate can be viewed

in the following way. In an annuity due the payment for the period is made at the

beginning of the period, whereas for an annuity immediate the payment for the

period is made at the end of the period. Clearly a payment of 1 at the end of the

period is equivalent to the payment of v = 1/ (1 + i) at the beginning of the period.

This gives an intuitive description of the equality of the previous exercise.

Example 3–2. Suppose that the annuity is paid continuously, that is, that the annu-

itant receives money at a constant rate of σ dollars per unit time. What value of σ

makes this continuous 20 year annuity equivalent to the discrete annuity described

above? Two annuities are said to be equivalent if they have the same present value.

For the continuous annuity the present value of the σ dt dollars received in the time

interval (t, t + dt) is e

−δt

σ dt. The present value of this annuity is therefore

20

0

σe

−δt

dt.

It is now a simple matter to ﬁnd σ.

Exercise 3–3. Find σ. Note that you must ﬁrst ﬁnd δ which is equivalent to an

interest rate of 5% compounded monthly.

A continuous annuity of the type above which is payable at rate σ = 1 for n

periods has present value which is denoted by a

n

.

Exercise 3–4. Show that a

n

=

1 − v

n

δ

.

§3: Annuities Certain 12

Thus far the value of an annuity has been computed at time 0. Another common

time point at which the value of an annuity consisting of n payments of 1 is

computed is time n. Denote by s

n

the value of an annuity immediate at time n, that

is, immediately after the nth payment. Then s

n

= (1 + i)

n

a

n

from the time diagram.

The value s

n

is called the accumulated value of the annuity immediate. Similarly

¨ s

n

is the accumulated value of an annuity due at time n and ¨ s

n

= (1 + i)

n

¨ a

n

.

Here are two examples which further develop skill in the use of these ideas.

Example 3–3. You are going to buy a house for which the purchase price is $100,000

and the downpayment is $20,000. You will ﬁnance the $80,000 by borrowing this

amount from a bank at 10% interest with a 30 year term. What is your monthly

payment? Typically such a loan is amortized, that is, you will make equal monthly

payments for the life of the loan and each payment consists partially of interest and

partially of principal. From the banks point of view this transaction represents the

purchase by the bank of an annuity immediate. The monthly payment, p, is thus the

solution of the equation 80000 = pa

360 0.10/ 12

. In this setting the quoted interest rate

on the loan is assumed to be compounded at the same frequency as the payment

period unless stated otherwise.

Exercise 3–5. Find the monthly payment. What is the total amount of the payments

made?

Example 3–4. Long ago I bought a new car from a local dealer. Let us say the total

cost to me was $15,000. The dealer seemed very anxious that I ﬁnance the purchase

through him, and he presented several arguments as to why I should do so. I could

borrow the entire purchase price at 11.95% ammortized over 60 months. His ﬁrst

sales pitch was as follows. If I ﬁnanced the car I would pay about $5000 in interest.

If I paid cash I would lose about $8000 in interest that I could earn by investing my

money in a savings account at 8.5% interest. Thus I would gain almost $3000 by

ﬁnancing the car. Is this argument correct? If not, what’s wrong with it?

Exercise 3–6. Find the monthly payment on the car if it is ﬁnanced through the

dealer. What is the total interest paid? Is this a relevant fact?

The dealers second argument ran as follows. Suppose that at the end of 5

years the car is worth only half its present value, namely, $7500. Let’s analyze my

available assets under the two alternatives. If I pay cash for the car, at the end of

5 years I will have clear title to a $7500 asset. If I ﬁnance the car, at the end of 5

years I will have clear title to the car ($7500) plus the cash I did not pay originally

($15000) plus the interest on this cash ($8000) for a total of $30500. Obviously

only a fool would pass up this type of opportunity!

Exercise 3–7. What are the ﬂaws, if any, in this second argument?

§3: Annuities Certain 13

Problems

Problem 3–1. Show that a

n

< a

n

< ¨ a

n

. Hint: This should be obvious from the

picture.

Problem 3–2. John borrows $1,000 from Jane at an annual effective rate of interest

i. He agrees to pay back $1,000 after six years and $1,366.87 after another 6 years.

Three years after his ﬁrst payment, John repays the outstanding balance. What is

the amount of John’s second payment?

Problem 3–3. Suppose a loan of amount A is amortized by a series of n payments.

Denote by b

k

the loan balance immediately after the kth payment and write b

0

= A.

Find a relationship that expresses b

k+1

in terms of b

k

, the interest rate i, and P = A/ a

n

.

Problem3–4. There are two common ways of analyzing loans which are amortized.

In the prospective method the loan balance at any point in time is seen to be the

present value of the remaining loan payments. Use the previous problem to show

that this statement is correct.

Problem 3–5. In the retrospective method the loan balance at any point in time is

seen to be the accumulated original loan amount less the accumulated value of the

past loan payments. Show that this formula for the loan balance is correct.

Problem 3–6. An annuity immediate pays an initial beneﬁt of one per year, in-

creasing by 10.25% every four years. The annuity is payable for 40 years. If the

effective interest rate is 5% ﬁnd an expression for the present value of this annuity.

Problem 3–7. You are given an annuity immediate paying $10 for 10 years, then

decreasing by $1 per year for nine years and paying $1 per year thereafter, forever.

If the annual effective rate of interest is 5%, ﬁnd the present value of this annuity.

Problem 3–8. Humphrey purchases a home with a $100,000 mortgage. Mortgage

payments are to be made monthly for 30 years, with the ﬁrst payment to be made one

month from now. The rate of interest is 10%. After 10 years, Humphrey increases

the amount of each monthly payment by $325 in order to repay the mortgage more

quickly. What amount of interest is paid over the life of the loan?

Problem 3–9. On January 1, an insurance company has $100,000 which is due to

Linden as a life insurance death beneﬁt. He chooses to receive the beneﬁt annually

over a period of 15 years, with the ﬁrst payment made immediately. The beneﬁt

he receives is based on an effective interest rate of 4% per annum. The insurance

company earns interest at an effective rate of 5% per annum. Every July 1 the

company pays $100 in expenses and taxes to maintain the policy. How much

money does the company have remaining after 9 years?

§3: Annuities Certain 14

Solutions to Problems

Problem 3–1. Isn’t the chain of inequalities simply expressing the fact that

getting a given amount of money sooner makes it worth more? An analytic

proof should be easy to give too.

Problem 3–2. From Jane’s point of view the equation 1000 = 1000(1 + i)

−6

+

1366.87(1 + i)

−12

must hold. The outstanding balance at the indicated time is

1366.87(1 + i)

−3

, which is the amount of the second payment.

Problem 3–3. b

k+1

= b

k

+ ib

k

− P.

Problem 3–4. The assertion of the prospective method is that b

k

= Pa

n−k

.

Plug in and show that this choice satisﬁes the initial condition b

0

= A and the

recursion of the previous problem.

Problem 3–5. Here the assertion is that b

k

= A(1 + i)

k

− Ps

k

. Show that this

choice satisﬁes the required recursion and initial condition.

Problem 3–6. Each 4 year chunk is a simple annuity immediate. Taking the

present value of these chunks forms an annuity due with payments every 4 years

that are increasing.

Problem3–7. What is the present value of an annuity immediate paying $1 per

year forever? What is the present value of such an annuity that begins payments

k years from now? The annuity desribed here is the difference of a few of these.

Problem 3–8. The initial monthly payment P is the solution of 100, 000 =

Pa

360

. The balance after 10 years is Pa

240

so the interest paid in the ﬁrst 10

years is 120P − (100, 000 − Pa

240

). To determine the number of new monthly

payments required to repay the loan the equation Pa

240

= (P+325)a

x

should be

solved for x. Since after x payments the loan balance is 0 the amount of interest

paid in the second stage can then be easily determined.

Problem 3–9. Since the effective rate of interest for the insurance company is

5%, the rate (0.05)

(2)

should be used to move the insurance company’s expenses

from July 1 to January 1.

§3: Annuities Certain 15

Solutions to Exercises

Exercise 3–1. The sum is the sum of the terms of a geometric series. So

¸

240

j=1

(1+

.05

12

)

−j

400 = 400((1+0.05/ 12)

−1

−(1+0.05/ 12)

−241

)/ (1−(1+0.05/ 12)

−1

) =

60, 610.12.

Exercise 3–2. This follows from the formulas for the present value of the two

annuities and the fact that d = iv.

Exercise 3–3. The force of interest that is equivalent to an interest rate of 5%

compounded monthly is δ = ln(1+0.05/ 12)

12

. Integration then gives the present

value of the continuous annuity as σ(1 − e

−20δ

)/ δ. So σ = 400δ/ i = 4790.03

will make the continuous annuity equivalent to the original discrete annuity.

Exercise 3–4. This formula follows by straightforward integration.

Exercise 3–5. Using the earlier formula gives a

360 0.10/ 12

= 113.95 from which

p = 702.06 and the total amount of the payments is 360p = 252740.60.

Exercise 3–6. The monthly payment is 15, 000/ a

60 .1195/ 12

= 333.29 so that the

total of the payments is 19, 997.27 of which 4, 997.27 is interest. The total of

the interest payments is irrelevant, since the time at which the interest payment

is made is not taken into account.

Exercise 3–7. Where did all those payments go?

§4. Laboratory 1

1. A loan of 10,000 carries an interest rate of 9% compounded quarterly. Equal

loan payments are to be made monthly for 36 months. What is the size of each

payment?

2. An amortization table is a table which lists the principal and interest portions

of each payment for a loan which is being amortized. Construct an amortization

table for the loan of the previous problem. The table should have four columns:

the payment number, the principal part of that payment, the interest part of that

payment, and the loan balance immediately after that payment is made.

3. A loan of 10,000 is to be repaid with equal monthly payments of p. The

interest rate for the ﬁrst year is 1.9%, while the interest rate for the remaining 2

years is 10.9%. What is p? What is the balance after the 6th payment? After the

15th payment? What are the principal and interest components of the 6th payment?

Of the 15th payment?

4. A loan of 10,000 is to be repaid as follows. Payments of p are to be made at

the end of each month for 36 months and a balloon payment of 2500 is to be made

at the end of the 36th month as well. If the interest rate is 5%, what is p? What is

the loan balance at the end of the 12th month? What part of the 15th payment is

interest? Principal?

5. The symbol a

(m)

n

denotes the present value of an annuity immediate that pays

1/ m at the end of each mth part of a year for n years. For example, if m = 12

payments of 1/12 are made at the end of each month. Find a formula for a

(m)

n

.

6. The symbol ¨ a

(m)

n

denotes the present value of an annuity due that pays 1/ m at

the beginning of each mth part of a year for n years. Find a formula for ¨ a

(m)

n

.

7. An increasing annuity immediate with a term of n periods pays 1 at the end

of the ﬁrst period, 2 at the end of the second period, 3 at the end of the third period,

. . . , n at the end of the nth period. Find (Ia)

n

, the present value of such an annuity.

8. A decreasing annuity immediate with a term of n periods pays n at the end

of the ﬁrst period, n − 1 at the end of the second period, n − 2 at the end of the third

period, . . . , 1 at the end of the nth period. Find (Da)

n

, the present value of such an

annuity.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§5. Brief Review of Probability Theory

Another aspect of insurance is that money is paid by the company only if some

event, which may be considered random, occurs within a speciﬁc time frame. For

example, an automobile insurance policy will experience a claim only if there is an

accident involving the insured auto. In this section a brief outline of the essential

material fromthe theory of probability is given. Almost all of the material presented

here should be familiar to the reader. The concepts presented here will play a crucial

role in the rest of these notes.

The underlying object in probability theory is a sample space S, which is simply

a set. This set is sometimes thought of as the collection of all possible outcomes

of a random experiment. Certain subsets of the sample space, called events, are

assigned probabilities by the probability measure (or probability set function)

which is usually denoted by P. This function has a few deﬁning properties.

(1) For any event E ⊂ S, 0 ≤ P[E] ≤ 1.

(2) P[∅] = 0 and P[S] = 1.

(3) If E

1

, E

2

, . . . are events and E

i

∩E

j

= ∅for i ≠j then P[

¸

∞

i=1

E

i

] =

¸

∞

i=1

P[E

i

].

From these basic facts one can deduce all manner of useful computational formulas.

Exercise 5–1. Show that if A ⊂ B are events, then P[A] ≤ P[B].

Another of the basic concepts is that of a random variable. A random variable

is a function whose domain is the sample space of a random experiment and whose

range is the real numbers.

In practice, the sample space of the experiment fades into the background and

one simply identiﬁes the random variables of interest. Once a random variable has

been identiﬁed, one may ask about its values and their associated probabilities. All

of the interesting probability information is bound up in the distribution function of

the random variable. The distribution function of the random variable X, denoted

F

X

(t), is deﬁned by the formula F

X

(t) = P[X ≤ t].

Two types of random variables are quite common. A random variable X with

distribution function F

X

is discrete if F

X

is constant except at at most countably

many jumps. A random variable X with distribution function F

X

is absolutely

continuous if F

X

(t) =

t

−∞

d

ds

F

X

(s) ds holds for all real numbers t.

If X is a discrete random variable, the density of X, denoted f

X

(t) is deﬁned by

the formula f

X

(t) = P[X = t]. There are only countably many values of t for which

the density of a discrete random variable is not 0. If X is an absolutely continuous

random variable, the density of X is deﬁned by the formula f

X

(t) =

d

dt

F

X

(t).

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§5: Brief Review of Probability Theory 18

Example 5–1. A Bernoulli random variable is a random variable which takes on

exactly two values, 0 and 1. Such random variables commonly arise to indicate the

success or failure of some operation. A Bernoulli random variable is discrete.

Exercise 5–2. Sketch the distribution function of a Bernoulli random variable with

P[X = 1] = 1/ 3.

Example 5–2. An exponentially distributed random variable Y with parameter

λ > 0 is a non-negative random variable for which P[Y ≥ t] = e

−λt

for t ≥ 0. Such a

random variable is often used to model the waiting time until a certain event occurs.

An exponential random variable is absolutely continuous.

Exercise 5–3. Sketch the distribution function of an exponential random variable

with parameter λ = 1. Sketch its density function also.

Exercise 5–4. Arandomvariable X is uniformly distributed on an interval (a, b) if

X represents the result of selecting a number at random from (a, b). Find the density

and distribution function of a random variable which is uniformly distributed on the

interval (0, 1).

Exercise 5–5. Draw a picture of a distribution function of a random variable which

is neither discrete nor absolutely continuous.

Another useful tool is the indicator function. Suppose A is a set. The indicator

function of the set A, denoted 1

A

(t), is deﬁned by the equation

1

A

(t) =

1 if t ∈ A

0 if t ∉ A.

Exercise 5–6. Graph the function 1

[0,1)

(t).

Exercise 5–7. Verify that the density of a random variable which is exponential

with parameter λ may be written λe

−λx

1

(0,∞)

(x).

Example 5–3. Random variables which are neither of the discrete nor absolutely

continuous type will arise frequently. As an example, suppose that person has a ﬁre

insurance policy on a house. The amount of insurance is $50,000 and there is a $250

deductible. Suppose that if there is a ﬁre the amount of damage may be represented

by a randomvariable Dwhich has the uniformdistributionon the interval (0, 70000).

(This assumption means that the person is underinsured.) Suppose further that in the

time period under consideration there is a probability p = 0.001 that a ﬁre will occur.

Let F denote the random variable which is 1 if a ﬁre occurs and is 0 otherwise. It is

easy to see that the size X of the claim to the insurer in this setting is given by

X = F

(D − 250)1

[250,50250]

(D) + 500001

(50250,∞)

(D)

.

§5: Brief Review of Probability Theory 19

This random variable X is neither discrete nor absolutely continuous.

Exercise 5–8. Verify the correctness of the formula for X. Find the distribution

function of the random variable X.

Often only the average value of a random variable and the spread of the

values around this average are all that are needed. The expectation (or mean)

of a discrete random variable X is deﬁned by E[X] =

¸

t

t f

X

(t), while the ex-

pectation of an absolutely continuous random variable X is deﬁned by E[X] =

∞

−∞

t f

X

(t) dt. Notice that in both cases the sum (or integral) involves terms of the

form (possible value of X) × (probability X takes on that value). When X is neither

discrete nor absolutely continuous, the expectation is deﬁned by the Riemann–

Stieltjes integral E[X] =

∞

−∞

t dF

X

(t), which again has the same form.

Exercise 5–9. Find the mean of a Bernoulli randomvariable Z with P[Z = 1] = 1/ 3.

Exercise 5–10. Find the mean of an exponential random variable with parameter

λ = 3.

Exercise 5–11. Find the mean and variance of the random variable in the ﬁre

insurance example given above. (The variance of a random variable X is deﬁned

by Var(X) = E[(X − E[X])

2

] and is often computed using the alternate formula

Var(X) = E[X

2

] − (E[X])

2

.)

Computation of conditional probabilities will play an important role. If A and

B are events, the conditional probability of A given B, denoted P[A| B], is deﬁned

by P[A| B] = P[A ∩ B]/ P[B] as long as P[B] ≠ 0.

The events Aand Bare independent if P[A| B] = P[A]. The intuition underlying

the notion of independent events is that the occurance of one of the events does not

alter the probability that the other event occurs.

Similar deﬁnitions can be given in the case of random variables. Intuitively,

the random variables X and Y are independent if knowledge of the value of one of

them does not effect the probabilities of events involving the other. Independence

of random variables is usually assumed based on this intuition. One important fact

is that if X and Y are independent random variables then E[XY] = E[X] E[Y].

§5: Brief Review of Probability Theory 20

Problems

Problem 5–1. Suppose X has the uniform distribution on the interval (0, a) where

a > 0 is given. What is the mean and variance of X?

Problem 5–2. The moment generating function of a random variable X, denoted

M

X

(t), is deﬁned by the formula M

X

(t) = E[e

tX

]. What is the relationship between

M

′ X

(0) and E[X]? Find a formula for Var(X) in terms of the moment generating

function of X and its derivatives at t = 0.

Problem 5–3. Express the Maclaurin expansion of M

X

(t) in terms of the moments

E[X], E[X

2

], E[X

3

],. . . of X. Hint: What is the Maclaurin expansion of e

x

?

Problem 5–4. Find the moment generating function of a Bernoulli randomvariable

Y for which P[Y = 1] = 1/ 4.

Problem5–5. Showthat

d

dt

ln M

X

(t)

t=0

= E[X] and

d

2

dt

2

ln M

X

(t)

t=0

= Var(X). This

is useful when the moment generating function has a certain form.

Problem 5–6. Find the moment generating function of a random variable Z which

has the exponential distribution with parameter λ. Use the moment generating

function to ﬁnd the mean and variance of Z.

Problem 5–7. A double indemnity life insurance policy has been issued to a person

aged 30. This policy pays $100,000 in the event of non-accidental death and

$200,000 in the event of accidental death. The probability of death during the next

year is 0.002, and if death occurs there is a 70%chance that it was due to an accident.

Write a random variable X which represents the size of the claim ﬁled in the next

year. Find the distribution function, mean, and variance of X.

Problem 5–8. In the preceding problem suppose that if death occurs the day of the

year on which it occurs is uniformly distributed. Assume also that the claim will be

paid immediately at death and the interest rate is 5%. What is the expected present

value of the size of the claim during the next year?

Problem 5–9. Show that for a random variable Y which is discrete and takes non–

negative integer values, E[Y] =

∞

¸

i=1

P[Y ≥ i]. Find a similar alternate expression for

E[Y

2

].

Problem 5–10. Suppose Y is a non-negative, absolutely continuous random vari-

able. Show that E[Y] =

∞

0

P[Y > t] dt.

§5: Brief Review of Probability Theory 21

Solutions to Problems

Problem 5–1. E[X] = a/ 2 and Var(X) = a

2

/ 12.

Problem 5–2. M

′ X

(0) = E[X].

Problem 5–3. Since e

x

=

∞

¸

k=0

x

k

/ k!,

M

X

(t) = E[e

tX

]

= E[

∞

¸

k=0

(tX)

k

/ k!]

=

∞

¸

k=0

E[X

k

]t

k

/ k!.

Thus the coefﬁcient of t

k

in the Maclaurin expansion of M

X

(t) is E[X

k

]/ k!.

Problem 5–4. M

Y

(t) = 3/ 4 + e

t

/ 4.

Problem 5–6. M

Z

(t) = λ/ (λ − t) for 0 ≤ t < λ.

Problem 5–7. Let D be a random variable which is 1 if the insured dies in the

next year and 0 otherwise. Let A be a random variable which is 2 if death is due

to an accident and 1 otherwise. Then X = 100000AD.

Problem 5–8. If U is uniformly distributed on the integers from 1 to 365 then

E[100000ADv

U

] is the desired expectation. Here v = 1/ (1 + 0.05

(365)

/ 365).

Problem 5–9. Hint: In the usual formula for the expectation of Y write

i =

¸

i

j=1

1 and then interchange the order of summation.

Problem 5–10. Use a trick like that of the previous problem. Double integrals

anyone?

§5: Brief Review of Probability Theory 22

Solutions to Exercises

Exercise 5–1. Using the fact that B = A ∪ (B \ A) and property (3) gives

P[B] = P[A] + P[B \ A]. By property (1), P[B \ A] ≥ 0, so the inequality

P[B] ≥ P[A] follows.

Exercise 5–2. The distribution function F(t) takes the value 0 for t < 0, the

value 2/ 3 for 0 ≤ t < 1 and the value 1 for t ≥ 1.

Exercise 5–3. The distribution function F(t) is 0 if t < 0 and 1 − e

−t

for t ≥ 0.

The density function is 0 for t < 0 and e

−t

for t ≥ 0.

Exercise 5–4. The distribution function F(t) takes the value 0 for t < 0, the

value t for 0 ≤ t ≤ 1 and the value 1 for t > 1. The density function takes the

value 1 for 0 < t < 1 and 0 otherwise.

Exercise 5–5. The picture should have a jump and also a smoothly increasing

portion.

Exercise 5–6. This function takes the value 1 for 0 ≤ t < 1 and the value 0

otherwise.

Exercise 5–8. The distribution function F(t) takes the value 0 if t < 0, the value

(1−0.001) +0.001×(250/ 70000) for 0 ≤ t < 250 (because X = 0 if either there is

no ﬁre or the loss caused by a ﬁre is less than 250), the value0.999+0.001×t/ 70000

for 250 ≤ t < 50250 and the value 1 for t ≥ 50250.

Exercise 5–9. E[Z] = 0 × (2/ 3) + 1 × (1/ 3) = 1/ 3.

Exercise 5–10. The expectation is

∞

0

tf (t) dt =

∞

0

t3e

−3t

dt = 1/ 3 using

integration by parts.

Exercise 5–11. Notice that the loss random variable X is neither discrete nor

absolutely continuous. The distribution function of X has two jumps: one at

t = 0 of size 0.999 + 250/ 70000 and another at 50250 of size 0.001 − 0.001 ×

50250/ 70000. So E[X] = 0 × (0.999 + 250/ 70000) +

50250

250

t0.001/ 70000 dt +

50250 × (0.001 − 0.001 × 50250/ 70000). The quantity E[X

2

] can be computed

similarly.

§6. Laboratory 2

One iteration of an experiment is conducted as follows. A single six sided die

is rolled and the number D of spots up is noted. Then D coins are tossed and the

number H of heads observed is noted.

1. What are the possible values of the randomvariable D? What are the possible

values of the random variable H?

2. For each relevant value of h and d compute P[H = h, D = d]. This probability

is called the joint density of H and D and is denoted f

H,D

(h, d). Put your computed

values in the form of a rectangular table with rows indexed by h and columns

indexed by d.

3. Find the density and expectation of D. Find the density and expectation of

H.

4. For each value of d ﬁnd the conditional density of H given D = d. Notation-

ally, f

H| D

(h| d) = P[H = h| D = d].

5. For each value of d ﬁnd the conditional expectation of H given D = d.

Notationally, E[H| D = d] =

¸

h

h f

H| D

(h| d).

6. The conditional expectation of H given D, denoted E[H| D], is the random

variable whose value on the event D = d is E[H| D = d]. Find the density of the

random variable E[H| D].

7. Compute E[E[H| D]]. The Theorem of Total Expectation states that for

any two random variables X and Y, E[E[Y| X]] = E[Y]. Do your computations of

E[H] in this problem and problem 3 reﬂect this fact?

8. The computations above show that it is easier to compute E[H] by ﬁrst

ﬁnding E[H| D]. Sometimes probabilistic intuition can be used to ﬁnd a conditional

expectation without computation. The intuition behind the conditional expectation

E[H| D] is that this should be the expected value of H computed after taking the

value of D into account. Give a probabilistic argument (without computation) that

E[H| D] = D/ 2.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§7. Survival Distributions

An insurance policy can embody two different types of risk. For some types

of insurance (such as life insurance) the variability in the claim is only the time at

which the claim is made, since the amount of the claim is speciﬁed by the policy.

In other types of insurance (such as auto or casualty) there is variability in both the

time and amount of the claim. The problems associated with life insurance will be

studied ﬁrst, since this is both an important type of insurance and also relatively

simple in some of its aspects.

The central difﬁculty in issuing life insurance is that of determining the length

of the future life of the insured. Denote by X the random variable which represents

the future lifetime of a newborn. For mathematical simplicity, assume that the

distribution function of X is absolutely continuous. The survival function of X,

denoted by s(x) is deﬁned by the formula

s(x) = P[X > x] = P[X ≥ x]

where the last equality follows from the continuity assumption. The assumption

that s(0) = 1 will always be made.

Example 7–1. In the past there has been some interest in modelling survival func-

tions in an analytic way. The simplest model is that due to Abraham DeMoivre. He

assumed that s(x) = 1 −

x

ω

for 0 < x < ω where ω is the limiting age by which

all have died. The DeMoivre law is simply the assertion that X has the uniform

distribution on the interval (0, ω).

Life insurance is usually issued on a person who has already attained a certain

age x. For notational convenience denote such a life aged x by (x), and denote the

future lifetime of a life aged x by T(x). What is the survival function for (x)? From

the discussion above, the survival function for (x) is P[T(x) > t]. Some standard

notation is now introduced. Set

t

p

x

= P[T(x) > t]

and

t

q

x

= P[T(x) ≤ t].

When t = 1 the preﬁx is ommitted and one just writes p

x

and q

x

respectively.

Generally speaking, having observed (x) some additional information about the

survival of (x) can be inferred. For example, (x) may have just passed a physical

exam given as a requirement for obtaining life insurance. For now this type of

possibility is disregarded. Operating under this assumption

t

p

x

= P[T(x) > t] = P[X > x + t | X > x] =

s(x + t)

s(x)

.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§7: Survival Distributions 25

Exercise 7–1. Write a similar expression for

t

q

x

.

Exercise 7–2. Show that for t ≥ s,

t

p

x

=

t−s

p

x+s s

p

x

.

There is one more special symbol. Set

t| u

q

x

= P[t < T(x) ≤ t + u]

which represents the probability that (x) survives at least t and no more than t + u

years. Again, if u = 1 one writes

t|

q

x

. The relations

t| u

q

x

=

t+u

q

x

−

t

q

x

=

t

p

x

−

t+u

p

x

follow immediately from the deﬁnition.

Exercise 7–3. Prove these two equalitites. Show that

t| u

q

x

=

t

p

x u

q

x+t

.

Exercise 7–4. Compute

t

p

x

for the DeMoivre lawof mortality. Conclude that under

the DeMoivre law T(x) has the uniform distribution on the interval (0, ω − x).

Under the assumption that X is absolutely continuous the random variable T(x)

will be absolutely continuous as well. Indeed

P[T(x) ≤ t] = P[x ≤ X ≤ x + t | X > x] = 1 −

s(x + t)

s(x)

so the density of T(x) is given by

f

T(x)

(t) =

−s

′

(x + t)

s(x)

=

f

X

(x + t)

1 − F

X

(x)

.

Intuitively this density represents the rate of death of (x) at time t.

Exercise 7–5. Use integration by parts to show that E[T(x)] =

∞

0

t

p

x

dt. This

expectation is called the complete expectation of life and is denoted by ˚ e

x

. Show

also that E[T(x)

2

] = 2

∞

0

t

t

p

x

dt.

Exercise 7–6. If X follows DeMoivre’s law, what is ˚ e

x

?

It is often useful to consider then the quantity

µ

x

=

f

X

(x)

1 − F

X

(x)

= −

s

′

(x)

s(x)

which is called the force of mortality. Intuitively the force of mortality is the

instantaneous rate of death of (x). (In component reliability theory this function is

often referred to as the hazard rate.) Integrating both sides of this equality gives the

useful relation

s(x) = exp

−

x

0

µ

t

dt

¸

.

§7: Survival Distributions 26

Exercise 7–7. Derive this last expression.

Exercise 7–8. Show that

t

p

x

= e

−

x+t

x

µ

s

ds

.

Exercise 7–9. Show that the density of T(x) can be written f

T(x)

(t) =

t

p

x

µ

x+t

.

If the force of mortality is constant the life random variable X has an expo-

nential distribution. This is directly in accord with the “memoryless” property of

exponential random variables. This memoryless property also has the interpretation

that a used article is as good as a new one. For human lives (and most manufactured

components) this is a fairly poor assumption, at least over the long term. The force

of mortality usually is increasing, although this is not always so.

Exercise 7–10. Find the force of mortality for DeMoivre’s law.

The curtate future lifetime of (x), denoted by K(x), is deﬁned by the relation

K(x) = [T(x)]. Here [t] is the greatest integer function. Note that K(x) is a discrete

random variable with density P[K(x) = k] = P[k ≤ T(x) < k + 1]. The curtate

lifetime, K(x), represents the number of complete future years lived by (x).

Exercise 7–11. Show that P[K(x) = k] =

k

p

x

q

x+k

.

Exercise 7–12. Show that the curtate expectation of life e

x

= E[K(x)] is given by

the formula e

x

=

¸

∞

i=0

i+1

p

x

. Hint: E[Y] =

¸

∞

i=1

P[Y ≥ i].

§7: Survival Distributions 27

Problems

Problem 7–1. Suppose µ

x+t

= t for t ≥ 0. Calculate

t

p

x

µ

x+t

and ˚ e

x

.

Problem 7–2. Calculate

∂

∂x

t

p

x

and

d

dx

˚ e

x

.

Problem 7–3. A life aged (40) is subject to an extra risk for the next year only.

Suppose the normal probability of death is 0.004, and that the extra risk may be

expressed by adding the function 0.03(1 −t) to the normal force of mortality for this

year. What is the probability of survival to age 41?

Problem 7–4. Suppose q

x

is computed using force of mortality µ

x

, and that q

′x

is

computed using force of mortality 2µ

x

. What is the relationship between q

x

and q

′x

?

Problem 7–5. Show that the conditional distribution of K(x) given that K(x) ≥ k is

the same as the unconditional distribution of K(x + k) + k.

Problem 7–6. Show that the conditional distribution of T(x) given that T(x) ≥ t is

the same as the unconditional distribution of T(x + t) + t.

Problem 7–7. The Gompertz law of mortality is deﬁned by the requirement that

µ

t

= Ac

t

for some constants A and c. What restrictions are there on A and c for this

to be a force of mortality? Write an expression for

t

p

x

and ˚ e

x

under Gompertz’ law.

Problem 7–8. Makeham’s law of mortality is deﬁned by the requirement that

µ

t

= A + Bc

t

for some constants A, B, and c. What restrictions are there on A, B

and c for this to be a force of mortality? Write an expression for

t

p

x

and ˚ e

x

under

Makeham’s law.

§7: Survival Distributions 28

Solutions to Problems

Problem 7–1. Here

t

p

x

= e

−

t

0

µ

x+s

ds

= e

−t

2

/ 2

and ˚ e

x

=

∞

0

t

p

x

dt = √2π/ 2.

Problem 7–2.

∂

∂x

t

p

x

=

t

p

x

(µ

x

− µ

x+t

) and

d

dx

˚ e

x

=

∞

0

∂

∂x

t

p

x

dt = µ

x

˚ e

x

− 1.

Problem7–3. If µ

t

is the usual force of mortality then p

40

= e

−

1

0

µ

40+s

+0.03(1−s) ds

.

Problem 7–4. The relation p

′x

= (p

x

)

2

holds, which gives a relation for the

death probability.

Problem 7–5. P[K(x) ≤ k + l| K(x) ≥ k] = P[k ≤ K(x) ≤ k + l]/ P[K(x) ≥ k] =

l

q

x+k

= P[K(x + k) + k ≤ l + k].

Problem 7–6. Proceed as in the previous problem.

§7: Survival Distributions 29

Solutions to Exercises

Exercise 7–1.

t

q

x

= P[T(x) ≤ t] = P[X ≤ x +t| X > x] = P[x < X ≤ x +t]/ P[X >

x] = (s(x) − s(x + t))/ s(x).

Exercise 7–2.

t

p

x

= s(x +t)/ s(x) = s(x +s +(t −s))/ s(x) = (s(x +s +(t −s))/ s(x +

s))(s(x + s)/ s(x)) =

t−s

p

x+ss

p

x

. What does this mean in words?

Exercise 7–3. For the ﬁrst one,

t| u

q

x

= P[t < T(x) ≤ t + u] = P[x + t < X ≤

t+u+x| X > x] = (s(x+t)−s(t+u+x))/ s(x) = (s(x+t)−s(x)+s(x)−s(t+u+x))/ s(x) =

t+u

q

x

−

t

q

x

. The second identity follows from the fourth term by simplifying

(s(x + t) − s(t + u + x))/ s(x) =

t

p

x

−

t+u

p

x

. For the last one,

t| u

q

x

= P[t < T(x) ≤

t + u] = P[x + t < X ≤ t + u + x| X > x] = (s(x + t) − s(t + u + x))/ s(x) =

(s(x + t)/ s(x))(s(t + x) − s(t + u + x))/ s(x + t) =

t

p

x u

q

x+t

.

Exercise 7–4. Under the DeMoivre law, s(x) = (ω − x)/ ω so that

t

p

x

=

(ω − x − t)/ (ω − x) for 0 < t < ω − x. Thus the distribution function of T(x) is

t/ (ω − x) for 0 < t < ω − x, which is the distribution function of a uniformly

distributed random variable.

Exercise 7–5. E[T(x)] =

∞

0

tf

T(x)

(t) dt = −

∞

0

ts

′

(x + t)/ s(x) dt =

∞

0

s(x +

t)/ s(x) dt =

∞

0

t

p

x

dt. The fact the lim

t→∞

s(x +t) = 0 is assumed, since everyone

eventually dies. The other expectation is computed similarly.

Exercise 7–6. Under DeMoivre’s law, ˚ e

x

= (ω − x)/ 2, since T(x) is uniform on

the interval (0, ω − x).

Exercise 7–7.

x

0

µ

t

dt =

x

0

−s

′

(t)/ s(t) dt = − ln(s(x)) + ln(s(0)) = − ln(s(x)),

since s(0) = 1. Exponentiating both sides to solve for s(x) gives the result.

Exercise 7–8. From the previous exercise, s(x + t) = e

−

x+t

0

µ

s

ds

. Using this

fact, the previous exercise, and the fact that

t

p

x

= s(x + t)/ s(x) gives the formula.

Exercise 7–9. Since f

T(x)

(t) = −s

′

(x + t)/ s(x) and s

′

(x + t) = −s(x + t)µ

x+t

by the

previous exercise, the result follows.

Exercise 7–10. From the earlier expression for the survival function under

DeMoivre’s law s(x) = (ω − x)/ ω, so that µ

x

= −s

′

(x)/ s(x) = 1/ (ω − x), for

0 < x < ω.

Exercise 7–11. P[K(x) = k] = P[k ≤ T(x) < k+1] = (s(x+k)−s(x+k+1))/ s(x) =

((s(x + k) − s(x + k + 1))/ s(x + k))(s(x + k)/ s(x)) =

k

p

x

q

x+k

.

Exercise 7–12. E[K(x)] =

¸

∞

i=1

P[K(x) ≥ i] =

¸

∞

i=1

P[T(x) ≥ i] =

¸

∞

i=1

i

p

x

=

¸

∞

j=0

j+1

p

x

.

§8. Life Tables

In practice the survival distribution is estimated by compiling mortality data in

the form of a life table. An example of a life table appears at the end of these notes.

Here is the conceptual model behind the entries in the table. Imagine that at time

0 there are l

0

newborns. Here l

0

is called the radix of the life table and is usually

taken to be some large number such as 100,000. Denote by l

x

the number of these

original newborns who are still alive at age x. Similarly

n

d

x

denotes the number of

persons alive at age x who die before reaching age x + n. As usual, when n = 1 it is

supressed in the notation.

Exercise 8–1. Show that

n

d

x

= l

x

− l

x+n

.

The ratio

l

x

l

0

is an estimate of s(x) based on the collected data. Assume that in

fact s(x) =

l

x

l

0

for non-negative integer values of x. Since earlier the assumption

was made that the life random variable X is absolutely continuous, the question

arises as to how the values of the survival function will be computed at non-integer

values of x. There are three commonly used methods of doing this, and these

methods produce slightly different numerical results. For the remainder of this

discussion suppose x is ﬁxed (and an integer) and that 0 ≤ t ≤ 1.

Under the assumption of the uniform distribution of deaths in the year of

death, denoted UDD, the survival function is computed by the formula

s(x + t) = (1 − t)s(x) + ts(x + 1).

The UDD assumption is the one most commonly made.

The assumption of a constant force of mortality in each year of age leads to

the formula

s(x + t) = s(x) e

−µt

where µ = − log p

x

.

The Balducci assumption is expressed in the formula

1

s(x + t)

=

1 − t

s(x)

+

t

s(x + 1)

.

Under each of the assumptions an explicit expression for all of the survivor

functions can be found.

Exercise 8–2. Find expressions for

t

q

x

and µ

x+t

, 0 ≤ t ≤ 1, under each of the above

3 assumptions.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§8: Life Tables 31

Having observed (x) may mean more than simply having seen a person aged x.

It may well mean that (x) has just passed a physical exam in preparation for buying

a life insurance policy. One would expect that the survival distribution of such a

person could be different from s(x). If this is believed to be the case the survival

function is actually dependent on two variables: the age at selection (application

for insurance) and the amount of time passed after the time of selection. A life

table which takes this effect into account is called a select table. A family of

survival functions indexed by both the age at selection and time are then required

and notation such as q

[x]+i

denotes the probability that a person dies between years

x + i and x + i + 1 given that selection ocurred at age x. As one might expect it

is reasonable to suppose that after a certain period of time the effect of selection

on mortality is negligable. The length of time until the selection effect becomes

negligable is called the select period. The Society of Actuaries (based in Illinois)

uses a 15 year select period in its mortality tables. The Institute of Actuaries in

Britain uses a 2 year select period. The implication of the select period of 15 years

in computations is that for each j ≥ 0, l

[x]+15+j

= l

x+15+j

.

A life table in which the survival functions are tabulated for attained ages only

is called an aggregrate table. Generally, a select life table contains a ﬁnal column

which constitutes an aggregate table. The whole table is then referred to as a select

and ultimate table and the last column is usually called an ultimate table. With

these observations in mind it is easy to utilize select life tables in computations.

Exercise 8–3. You are given the following extract from a 3 year select and ultimate

mortality table.

x l

[x]

l

[x]+1

l

[x]+2

l

x+3

x + 3

70 7600 73

71 7984 74

72 8016 7592 75

Assume that the ultimate table follows DeMoivre’s law and that d

[x]

= d

[x]+1

=

d

[x]+2

for all x. Find 1000(

2| 2

q

[71]

).

§8: Life Tables 32

Problems

Problem 8–1. Graph µ

x+t

, 0 ≤ t ≤ 1, under each of the 3 assumptions for fractional

years.

Problem 8–2. For each of the 3 assumptions for fractional years ﬁnd a formula for

t

p

x

, 0 ≤ t ≤ 1 in terms of t and p

x

. For each of 20 equally space values of p

x

between

0 and 1, make a plot of

t

p

x

for 0 ≤ t ≤ 1 under each of the 3 assumptions. For what

value(s) of p

x

are the assumptions numerically indistinguishable?

Problem 8–3. Use the life table to compute

1/ 2

p

20

under each of the 3 assumptions

for fractional years.

Problem 8–4. Show that under the assumption of uniform distribution of deaths in

the year of death that K(x) and T(x) − K(x) are independent and that T(x) − K(x) has

the uniform distribution on the interval (0, 1).

Problem 8–5. Show that under UDD ˚ e

x

= e

x

+

1

2

.

§8: Life Tables 33

Solutions to Problems

Problem 8–3. Under UDD,

t

p

x

= (1 − t) + tp

x

so

1/ 2

p

20

= 1/ 2 + 1/ 2p

20

.

Under constant force,

t

p

x

= e

t log p

x

so

1/ 2

p

20

= e

(1/ 2) log p

20

. Under Balducci,

1/

t

p

x

= 1 − t + t/ p

x

so

1/ 2

p

20

= 1/ (1/ 2 + 1/ 2p

20

).

Problem 8–4. For 0 ≤ t < 1, P[K(x) = k, T(x) − K(x) ≤ t] = P[k ≤ T(x) ≤

k + t] =

k

p

x t

q

x+k

=

k

p

x

(t − tp

x+k

) = tP[K(x) = k].

Problem 8–5. Use the previous problem.

§8: Life Tables 34

Solutions to Exercises

Exercise 8–1. Since

n

d

x

is the number alive at age x who die by age x + n, this

is simply the number alive at age x, which is l

x

, minus the number alive at age

x + n, which is l

x+n

.

Exercise 8–2. Under UDD,

t

q

x

= (s(x) −s(x +t))/ s(x) = (ts(x) −ts(x +1))/ s(x) =

tq

x

and µ

x+t

= −s

′

(x+t)/ s(x+t) = (s(x)−s(x+1))/ ((1−t)s(x)+ts(x+1)) = (1−p

x

)/ (1−

t + tp

x

). Under constant force,

t

q

x

= 1 − (p

x

)

t

while µ

x+t

= µ

x

. Under Balducci,

s(x + t) = s(x)s(x + 1)/ ((1 − t)s(x + 1) + ts(x)) so that s(x + t)/ s(x) = p

x

/ (p

x

+ tq

x

).

Thus

t

q

x

= tq

x

/ (p

x

+ tq

x

). Also µ

x+t

= −s

′

(x + t)/ s(x + t) = (1 − q

x

)/ (1 + tq

x

).

Exercise 8–3. The objective is to compute 1000

2| 2

q

[71]

= 1000(

2

p

[71]

−

4

p

[71]

) =

1000(l

[71]+2

− l

[71]+4

)/ l

[71]

= 1000(l

[71]+2

− l

75

)/ l

[71]

, where the effect of the selec-

tion period has been used. To ﬁnd the required entries in the table proceed as

follows. Since 8016−7592 = 424 and using the assumption about the number of

deaths, l

[72]+1

= 8016−212 = 7804 and l

72+3

= 7592−212 = 7380. Since the ulti-

mate table follows DeMoivre’s Law, l

71+3

= (7600 + 7380)/ 2 = 7490. Again us-

ing the assumption about the number of deaths, l

[71]+2

= (7984+7490)/ 2 = 7737

and l

[71]

= 7984 + 247 = 8231. So 1000

2| 2

q

[71]

= 1000(7737 − 7380)/ 8231 =

43.37.

§9. Laboratory 3

1. Below is a table which gives the values of q

x

for ages 1 through 105. Deﬁne

l

1

= 1, 000, 000 and use the values of q

x

to compute l

x

and d

x

for 1 ≤ x ≤ 106.

2. Make a plot of l

x

for 1 ≤ x ≤ 106.

x q

x

x q

x

x q

x

1 0.000637 36 0.000841 71 0.026627

2 0.000430 37 0.000904 72 0.029565

3 0.000357 38 0.000964 73 0.032931

4 0.000278 39 0.001021 74 0.036738

5 0.000255 40 0.001079 75 0.041002

6 0.000244 41 0.001142 76 0.045699

7 0.000234 42 0.001215 77 0.050833

8 0.000216 43 0.001299 78 0.056487

9 0.000209 44 0.001397 79 0.062777

10 0.000212 45 0.001508 80 0.069757

11 0.000219 46 0.001629 81 0.077444

12 0.000228 47 0.001762 82 0.085828

13 0.000240 48 0.001905 83 0.094904

14 0.000254 49 0.002060 84 0.104700

15 0.000269 50 0.002225 85 0.115289

16 0.000284 51 0.002401 86 0.126798

17 0.000301 52 0.002589 87 0.139353

18 0.000316 53 0.002795 88 0.153021

19 0.000331 54 0.003023 89 0.167757

20 0.000345 55 0.003283 90 0.183408

21 0.000357 56 0.003583 91 0.199769

22 0.000366 57 0.003932 92 0.216605

23 0.000373 58 0.004332 93 0.233662

24 0.000376 59 0.004784 94 0.250693

25 0.000376 60 0.005286 95 0.267491

26 0.000378 61 0.005833 96 0.283905

27 0.000382 62 0.006414 97 0.299852

28 0.000393 63 0.007014 98 0.315296

29 0.000412 64 0.007616 99 0.330207

30 0.000444 65 0.008207 100 0.344556

31 0.000499 66 0.008777 101 0.358628

32 0.000562 67 0.009318 102 0.371685

33 0.000631 68 0.009828 103 0.383040

34 0.000702 69 0.010306 104 0.392003

35 0.000773 70 0.010753 105 1.000000

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§10. Status

A life insurance policy is sometimes issued which pays a beneﬁt at a time

which depends on the survival characteristics of two or more people. A status is an

artiﬁcially constructed life form for which the notion of life and death can be well

deﬁned.

Example 10–1. Acommon artiﬁcial life formis the status which is denoted n. This

is the life form which survives for exactly n time units and then dies.

Example 10–2. Another common status is the joint life status which is constructed

as follows. Given two life forms (x) and (y) the joint life status, denoted x : y, dies

exactly at the time of death of the ﬁrst to die of (x) and (y).

Exercise 10–1. If (x) and (y) are independent lives, what is the survival function of

the status x : y?

Exercise 10–2. What is survival function of x : n?

Occasionally, even the order in which death occurs is important. The status

x

1

: n is a status which dies at the time of death of (x) if the death of (x) occurs before

time n. Otherwise, this status never dies.

Exercise 10–3. Under what circumstances does x : n

1

die?

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§10: Status 37

Problems

Problem 10–1. Find a formula for the survival function of x

1

: n in terms of the

survival function of (x).

Problem 10–2. If the UDD assumption is valid for (x), does UDD hold for x

1

: n?

Problem 10–3. Find a formula for the survival function of x : n

1

.

Problem 10–4. If the UDD assumption is valid for (x), does UDD hold for x : n

1

?

Problem 10–5. If the UDD assumption is valid for (x), does UDD hold for x : n?

Problem 10–6. If the UDD assumption is valid for each of (x) and (y) and if (x)

and (y) are independent lives, does UDD hold for x : y?

§10: Status 38

Solutions to Problems

Problem 10–1. P[T(x

1

: n) ≥ t] =

t

p

x

for 0 ≤ t < n and P[T(x

1

: n) ≥ t] =

n

p

x

for t ≥ n.

Problem 10–2. The UDD assuption holds for x

1

: n if and only if P[T(x

1

: n) ≥

k + t] = (1 − t)P[T(x

1

: n) ≥ k] + tP[T(x

1

: n) ≥ k + 1] for all integers k and all

0 ≤ t ≤ 1. Now use the formula for the survival function found in the previous

problem.

§10: Status 39

Solutions to Exercises

Exercise 10–1. The joint life status survives t time units if and only if both (x)

and (y) survive t time units. Using the independence gives s(t) =

t

p

x t

p

y

.

Exercise 10–2. Since a constant random variable is independent of any other

random variable, s(t) =

t

p

x t

p

n

=

t

p

x

if t ≤ n and 0 if t > n, by using the previous

exercise.

Exercise 10–3. The status x : n

1

dies at time n if (x) is still alive at time n,

otherwise this status never dies.

§11. Valuing Contingent Payments

Earlier, the central theme of these notes was asserted to be embodied in the

question, “What is the value today of a random sum of money which will be paid

at a random time in the future?” This question can now be answered. Suppose the

randomamount of money is denoted by Aand the randomtime at which it will be paid

is denoted by T. The value of this payment today is computed in two steps. First, an

expression for the present value of the payment is written. This expression will be a

randomvariable. Here the present value is Av

T

. Then the expectation of this random

variable is computed. This expectation, E[Av

T

], is the value today of the random

future payment. The interpretation of this amount, E[Av

T

], is as the average present

value of the actual payment. Averages are reasonable in the insurance context since,

from the company’s point of view, there are many probabilistically similar policies

for which the company is obliged to pay beneﬁts. The average cost (and income) per

policy is therefore a reasonable starting point fromwhich to determine the premium.

The expected present value is usually referred to as the actuarial present value

in the insurance context. In the next few sections the actuarial present value of

certain standard parts of insurance contracts are computed.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§12. Life Insurance

In the case of life insurance the determination of the value of the insurance

depends on the random time of death of the insured. The amount that is paid at the

time of death is usually ﬁxed by the policy and is non-random. Assume that the

force of interest is constant and known to be equal to δ. Also simply write T = T(x)

whenever clarity does not demand the full notation. The actuarial present value of

an insurance which pays 1 at the time of death is then

E[v

T

]

by the priciple above. Intuitively, the actuarial present value of the beneﬁt is the

single premium payment that an insurance company with no operating expenses

and no desire for proﬁt would charge today in order to provide the beneﬁt payment.

For this reason the actuarial present value of a beneﬁt is also called the net single

premium. The net single premiumwould be the idealized amount an insured would

pay as a lump sum (single premium) at the time that the policy is issued. The case

of periodic premium payments will be discussed later.

Acatalog of the various standard types of life insurance policies and the standard

notation for the associated net single premium follows. In most cases the beneﬁt

amount is assumed to be $1, and in all cases the beneﬁt is assumed to be paid at the

time of death. Keep in mind that a ﬁxed constant force of interest is also assumed

and that v = 1/ (1 + i) = e

−δ

.

Insurances Payable at the Time of Death

Type Net Single Premium

n-year pure endowment A

x:n

1

=

n

E

x

= E[v

n

1

(n,∞)

(T)]

n-year term A

x

1

:n

= E[v

T

1

[0,n]

(T)]

whole life A

x

= E[v

T

]

n-year endowment A

x:n

= E[v

T∧n

]

m-year deferred n-year term

m| n

A

x

= E[v

T

1

(m,n+m]

(T)]

whole life increasing mthly (IA)

(m)

x

= E[v

T

[Tm + 1]/ m]

n-year term increasing annually (IA)

x

1

:n

= E[v

T

[T + 1]1

[0,n)

(T)]

n-year term decreasing annually (DA)

x

1

:n

= E[v

T

(n − [T])1

[0,n)

(T)]

Using this table it is a simple matter to compute the net single premium in the

various cases. The bar is indicative of an insurance paid at the time of death, while

the subscripts denote the status whose death causes the insurance to be paid. These

insurances are now reviewed on a case-by-case basis.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§12: Life Insurance 42

The ﬁrst type of insurance is n-year pure endowment insurance which pays

the full beneﬁt amount at the end of the nth year if the insured survives at least n

years. The notation for the net single premium for a beneﬁt amount of 1 is A

x:n

1

(or

occasionally in this context

n

E

x

). The net single premium for a pure endowment is

just the actuarial present value of a lump sum payment made at a future date. This

differs from the ordinary present value simply because it also takes into account the

mortality characteristics of the recipient.

Exercise 12–1. Show that

n

E

x

= v

n

n

p

x

.

The second type of insurance is n-year terminsurance. The net single premium

with a beneﬁt of 1 payable at the time of death for an insured (x) is denoted A

x

1

:n

.

This type insurance provides for a beneﬁt payment only if the insured dies within n

years of policy inception.

The third type of insurance is whole life in which the full beneﬁt is paid no

matter when the insured dies in the future. The whole life beneﬁt can be obtained

by taking the limit as n → ∞ in the n-year term insurance setting. The notation for

the net single premium for a beneﬁt of 1 is A

x

.

Exercise 12–2. Suppose that T(x) has an exponential distribution with mean 50. If

the force of interest is 5%, ﬁnd the net single premium for a whole life policy for

(x), if the beneﬁt of $1000 is payable at the moment of death.

Exercise 12–3. Show that A

x

= A

x

1

:n

+ v

n

n

p

x

A

x+n

by conditioning on the event

T(x) ≥ n and also by direct reasoning from a time diagram by looking at the

difference of two policies.

The fourth type of insurance, n-year endowment insurance, provides for the

payment of the full beneﬁt at the time of death of the insured if this occurs before

time n and for the payment of the full beneﬁt at time n otherwise. The net single

premium for a beneﬁt of 1 is denoted A

x:n

.

Exercise 12–4. Show that A

x:n

= A

x

1

:n

+ A

x:n

1

.

Exercise 12–5. Use the life table to ﬁnd the net single premium for a 5 year pure

endowment policy for (30) assuming an interest rate of 5%.

The m-year deferred n-year terminsurance policy provides provides the same

beneﬁts as n year term insurance between times m and m + n provided the insured

lives m years.

All of the insurances discussed thus far have a ﬁxed constant beneﬁt. Increasing

whole life insurance provides a beneﬁt which increase linearly in time. Similarly,

§12: Life Insurance 43

increasing and decreasing n-year term insurance provides for linearly increasing

(decreasing) beneﬁt over the term of the insurance.

Corresponding to the insurances payable at the time of death are the same type

of policies available with the beneﬁt being paid at the end of the year of death. The

only difference between these insurances and those already described is that these

insurances depend on the distribution of the curtate life variable K = K(x) instead

of T. The following table introduces the notation.

Insurances Payable the End of the Year of Death

Type Net Single Premium

n-year term A

x

1

:n

= E[v

K+1

1

[0,n)

(K)]

whole life A

x

= E[v

K+1

]

n-year endowment A

x:n

= E[v

(K+1)∧n

]

m-year deferred n-year term

m| n

A

x

= E[v

K+1

1

[m,n+m)

(K)]

whole life increasing annually (IA)

x

= E[v

K+1

(K + 1)]

n-year term increasing annually (IA)

x

1

:n

= E[v

K+1

(K + 1)1

[0,n)

(K)]

n-year term decreasing annually (DA)

x

1

:n

= E[v

K+1

(n − K)1

[0,n)

(K)]

These policies have net single premiums which can be easily computed from

the information in the life table. The primary use for these types of policies is the

computational connection between them and the ‘continuous’ policies described

above. To illustrate the ease of computation when using a life table observe that

from the deﬁnition

A

x

=

∞

¸

k=0

v

k+1

k

p

x

q

x+k

=

∞

¸

k=0

v

k+1

d

x+k

l

x

.

In practice, of course, the sum is ﬁnite. Similar computational formulas are readily

obtained in the other cases.

Exercise 12–6. Show that A

x:n

1

= A

x:n

1

and interpret the result verbally. How would

you compute A

x:n

1

using the life table?

Under the UDD assumption it is fairly easy to ﬁnd formulas which relate the

insurances payable at the time of death to the corresponding insurance payable at

§12: Life Insurance 44

the end of the year of death. For example, in the case of a whole life policy

A

x

= E[e

−δT(x)

]

= E[e

−δ(T(x)−K(x)+K(x))

]

= E[e

−δ(T(x)−K(x))

] E[e

−δK(x)

]

=

1

δ

(1 − e

−δ

)e

δ

E[e

−δ(K(x)+1)

]

=

i

δ

A

x

where the third equality springs from the independence of K(x) and T(x) − K(x)

under UDD, and the fourth equality comes from the fact that under UDD the

random variable T(x) − K(x) has the uniform distribution on the interval (0,1).

Exercise 12–7. Can similar relationships be established for term and endowment

policies?

Exercise 12–8. Use the life table to ﬁnd the net single premium for a 5 year

endowment policy for (30) assuming an interest rate of 5%.

Exercise 12–9. An insurance which pays a beneﬁt amount of 1 at the end of the

mth part of the year in which death occurs has net single premium denoted by A

(m)

x

.

Show that under UDD i

(m)

A

(m)

x

= δ A

x

.

One consequence of the exercise above is that only the net single premiums for

insurances payable at the end of the year of death need to be tabulated, if the UDD

assumption is made. This leads to a certain amount of computational simplicity.

§12: Life Insurance 45

Problems

Problem 12–1. Write expressions for all of the net single premiums in terms of

either integrals or sums. Hint: Recall the form of the density of T(x) and K(x).

Problem 12–2. Show that δA

x

1

:n

= iA

x

1

:n

, but that δA

x:n

≠ iA

x:n

, in general.

Problem 12–3. Use the life table and UDD assumption (if necessary) to compute

A

21

, A

21:5

, and A

21

1

:5

.

Problem 12–4. Show that

dA

x

di

= −v(IA)

x

.

Problem 12–5. Assume that DeMoivre’s law holds with ω = 100 and i = 0.10.

Find A

30

and A

30

. Which is larger? Why?

Problem 12–6. Suppose µ

x+t

= µ and i = 0.10. Compute A

x

and A

x

1

:n

. Do your

answers depend on x? Why?

Problem 12–7. Suppose A

x

= 0.25, A

x+20

= 0.40, and A

x:20

= 0.55. Compute A

x:20

1

and A

x

1

:20

.

Problem 12–8. Show that

(IA)

x

= vq

x

+ v[A

x+1

+ (IA)

x+1

]p

x

.

What assumptions (if any) did you make?

Problem 12–9. What change in A

x

results if for some ﬁxed n the quantity q

x+n

is

replaced with q

x+n

+ c?

Problem 12–10. What is the change in A

x

if δ is replaced by δ +c? If µ is replaced

by µ + c?

§12: Life Insurance 46

Solutions to Problems

Problem 12–1. The densities required are f

T(x)

(t) =

t

p

x

µ

x+t

and f

K(x)

(k) =

k

p

x

q

x+k

respectively.

Problem 12–2. δA

x

1

:n

= δ(A

x

− e

−δn

n

p

x

A

x+n

) = iA

x

− iv

n

n

p

x

A

x+n

= iA

x

1

:n

.

Problem12–3. Use δA

21

= iA

21

, A

21:5

= A

21

1

:5

+

5

E

21

and the previous problem.

Problem 12–4. Just differentiate under the expectation in the deﬁnition of A

x

.

Problem 12–5. Clearly A

30

> A

30

since the insurance is paid sooner in the

continuous case. Under DeMoivre’s law the UDD assumption is automatic and

A

30

=

1

70

70

0

e

−δt

dt.

Problem12–6. The answers do not depend on x since the lifetime is exponential

and therefore ageless.

Problem 12–7. The two relations A

x

= A

x

1

:20

+ v

20

n

p

x

A

x+20

and A

x:20

= A

x

1

:20

+

v

n

n

p

x

along with the fact that A

x:20

1

= v

n

n

p

x

give two equations in the two sought

after unknowns.

Problem12–8. Either the person dies in the ﬁrst year, or doesn’t. If she doesn’t

buy an increasing annually policy for (x + 1) and a whole life policy to make up

for the increasing part the original policy would provide.

Problem 12–9. The new beneﬁt is the old beneﬁt plus a pure endowment

beneﬁt of cv at time n.

§12: Life Insurance 47

Solutions to Exercises

Exercise 12–1. Since if the beneﬁt is paid, the beneﬁt payment occurs at time

n,

n

E

x

= E[v

n

1

[n,∞)

(T(x))] = v

n

P[T(x) ≥ n] = v

n

n

p

x

.

Exercise 12–2. Under the assumptions given the net single premium is

E[1000v

T(x)

] =

∞

0

1000e

−0.05t

(1/ 50)e

−t/ 50

dt = 285.71.

Exercise 12–3. For the conditioning argument, break the expectation into two

pieces by writing A

x

= E[v

T

] = E[v

T

1

[0,n]

(T)] +E[v

T

1

(n,∞)

(T)]. The ﬁrst expecta-

tion is exactly A

x

1

:n

. For the second expectation, using the Theorem of Total Ex-

pectation gives E[v

T

1

(n,∞)

(T)] = E[E[v

T

1

(n,∞)

(T)| T ≥ n]]. Now the conditional

distribution of T given that T ≥ n is the same as the unconditional distribution of

T(x+n)+n. Using this fact gives the conditional expectation as E[v

T

1

(n,∞)

(T)| T ≥

n] = E[v

T(x+n)+n

1

(n,∞)

(T(x + n) + n)]1

(n,∞)

(T) = v

n

A

x+n

1

(n,∞)

(T). Taking expecta-

tions gives the result. To use the time diagram, imagine that instead of buying

a whole life policy now, the insured pledges to buy an n year term policy now,

and if alive after n years, to buy a whole life policy at time n (at age x + n). This

will produce the same result. The premium for the term policy paid now is A

x

1

:n

and the premium for the whole life policy at time n is A

x+n

. This latter premium

is only paid if the insured survives, so the present value of this premium is the

second term in the solution.

Exercise 12–4. Using the deﬁnition and properties of expectation gives A

x:n

=

E[v

T

1

[0,n]

(T) + v

n

1

(n,∞)

(T)] = E[v

T

1

[0,n]

(T)] + E[v

n

1

(n,∞)

(T)] = A

x

1

:n

+ A

x:n

1

.

Exercise 12–5. The net single premium for the pure endowment policy is

v

5

5

p

30

= (1.05)

−5

l

35

/ l

30

= (1.05)

−5

95808/ 96477 = 0.778.

Exercise 12–6. A

x:n

1

= E[v

n

1

[n,∞)

(T)] = E[v

n

1

[n,∞)

(K)] = A

x:n

1

= v

n

n

p

x

=

v

n

l

x+n

/ l

x

.

Exercise 12–7. Since term policies can be expressed as a difference of premi-

ums for whole life policies, the answer is yes.

Exercise 12–8. The net single premium for a pure endowment policy is

v

5

5

p

30

= (1.05)

−5

l

35

/ l

30

= (1.05)

−5

95808/ 96477 = 0.778. For the endowment

policy, the net single premium for a 5 year term policy must be added to this

amount. Fromthe relation given earlier, A

30

1

:5

= A

30

−v

5

5

p

30

A

35

. The relationship

between insurances payable at the time of death and insurances payable at the

end of the year of death is used to complete the calculation.

Exercise 12–9. Notice that [mT(x)] is the number of full mths of a year

that (x) lives before dying. (Here [a] is the greatest integer function.) So the

number of mths of a year that pass until the beneﬁt for the insurance is paid

is [mT(x)] + 1, that is, the beneﬁt is paid at time ([mT(x)] + 1)/ m. From here

the derivation proceeds as above. A

(m)

x

= E[v

([mT]+1)/ m

] = E[v

([m(T−K+K)]+1)/ m

] =

E[v

K

]E[v

([m(T−K)]+1)/ m

]. Now T − K has the uniform distribution on the interval

(0, 1) under UDD, so [m(T − K)] has the uniform distribution over the integers

0,. . . , m − 1. So E[v

([m(T−K)]/ m

] =

¸

m−1

j=0

v

j/ m

× (1/ m) = (1/ m)(1 − v)/ (1 − v

1/ m

)

from the geometric series formula. Substituting this in the earlier expression

§12: Life Insurance 48

gives A

(m)

x

= A

x

v

−1

v

1/ m

(1/ m)(1−v)/ (1−v

1/ m

) = A

x

δ/ i

(m)

since i

(m)

= m(v

−1/ m

−1).

§13. Laboratory 4

1. Show that A

x

= vq

x

+ vp

x

A

x+1

. Derive a similar formula for A

x

.

2. The one step recursion formulas derived in problem 1 are especially useful

for computational purposes. The formulas are used to work backwards from large

attained ages to smaller ones, since at large attained ages everyone is dead and

the net premium for the insurance must be zero. Use the values of q

x

given in

Laboratory 3 and i = 5% to compute the values of A

x

and A

x

for x = 1 to x = 106.

Place the result of your computations into a nice table.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§14. Life Annuities

The basic study of life insurance concludes by developing techniques for un-

derstanding what happens when premiums are paid monthly or annually instead of

just when the insurance is issued. In the non–random setting a sequence of equal

payments made at equal intervals in time was referred to as an annuity. Here

interest centers on annuities in which the payments are made (or received) only as

long as the insured survives.

An annuity in which the payments are made for a non–random period of time

is called an annuity certain. From the earlier discussion, the present value of an

annuity immediate (payments begin one period in the future) with a payment of 1

in each period is

a

n

=

n

¸

j=1

v

j

=

1 − v

n

i

while the present value of an annuity due (payments begin immediately) with a

payment of 1 in each period is

¨ a

n

=

n−1

¸

j=0

v

j

=

1 − v

n

1 − v

=

1 − v

n

d

.

These formulas will now be adapted to the case of contingent annuities in which

payments are made for a random time interval.

Suppose that (x) wishes to buy a life insurance policy. Then (x) will pay a

premium at the beginning of each year until (x) dies. Thus the premium payments

represent a life annuity due for (x). Consider the case in which the payment amount

is 1. Since the premiums are only paid annually the term of this life annuity depends

only on the curtate life of (x). There will be a total of K(x) + 1 payments, so the

actuarial present value of the payments is ¨ a

x

= E[¨ a

K(x)+1

] where the left member is

a notational convention. This formula gives

¨ a

x

= E[¨ a

K(x)+1

] = E[

1 − v

K(x)+1

d

] =

1 − A

x

d

as the relationship between this life annuity due and the net single premium for a

whole life policy. A similar analysis holds for life annuities immediate.

Exercise 14–1. Compute the actuarial present value of a life annuity immediate.

What is the connection with a whole life policy?

Exercise 14–2. A life annuity due in which payments are made m times per year

and each payment is 1/ m has actuarial present value denoted by ¨ a

(m)

x

. Show that

A

(m)

x

+ d

(m)

¨ a

(m)

x

= 1.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§14: Life Annuities 51

Example 14–1. The Mathematical Association of America offers the following

alternative to members aged 60. You can pay the annual dues and subscription rate

of $90, or you can become a life member for a single fee of $675. Life members

are entitled to all the beneﬁts of ordinary members, including subscriptions. Should

one become a life member? To answer this question, assume that the interest rate is

5% so that the Life Table at the end of the notes can be used. The actuarial present

value of a life annuity due of $90 per year is

90

1 − A

60

1 − v

= 90

1 − 0.412195

1 − 1/ 1.05

= 1110.95.

Thus one should deﬁnitely consider becoming a life member.

Exercise 14–3. What is the probability that you will get at least your money’s worth

if you become a life member? What assumptions have you made?

Pension beneﬁts often take the form of a life annuity immediate. Sometimes

one has the option of receiving a higher beneﬁt, but only for a ﬁxed number of years

or until death occurs, whichever comes ﬁrst. Such an annuity is called a temporary

life annuity.

Example 14–2. Suppose a life annuity immediate pays a beneﬁt of 1 each year

for n years or until (x) dies, whichever comes ﬁrst. The symbol for the actuarial

present value of such a policy is a

x:n

. How does one compute the actuarial present

value of such a policy? Remember that for a life annuity immediate, payments are

made at the end of each year, provided the annuitant is alive. So there will be a

total of K(x) ∧ n payments, and a

x:n

= E[

¸

K(x)∧n

j=1

v

j

]. A similar argument applies

in the case of an n year temporary life annuity due. In this case, payments are

made at the beginning of each of n years, provided the annuitant is alive. In this

case ¨ a

x:n

= E[

¸

K(x)∧(n−1)

j=0

v

j

] = E[

1−v

(K(x)+1)∧n

d

] where the left member of this equality

introduces the notation.

Exercise 14–4. Show that A

x:n

= 1 − d ¨ a

x:n

. Find a similar relationship for a

x:n

.

Especially in the case of pension beneﬁts it is more realistic to assume that the

payments are made monthly. Suppose payments are made m times per year. In this

case each payment is 1/ m. One could begin from ﬁrst principles (this makes a good

exercise), but instead the previously established facts for insurances together with

the relationships between insurances and annuities given above will be used. Using

§14: Life Annuities 52

the obvious notation gives

¨ a

(m)

x

=

1 − A

(m)

x

d

(m)

=

1 −

i

i

(m)

A

x

d

(m)

=

1 −

i

i

(m)

(1 − d¨ a

x

)

d

(m)

=

id

i

(m)

d

(m)

¨ a

x

+

i

(m)

− i

i

(m)

d

(m)

where at the second equality the UDD assumption was used.

Exercise 14–5. Find a similar relationship for an annuity immediate which pays

1/ m m times per year.

Auseful idealization of annuities payable at discrete times is an annuity payable

continuously. Such an annuity does not exist in the ‘real world’, but it provides a

useful connecting bridge between certain types of discrete annuities. Suppose that

the rate at which the beneﬁt is paid is constant and is 1 per unit time. Then during

the time interval (t, t + dt) the amount paid is dt and the present value of this amount

is e

−δt

dt. Thus the present value of such a continuously paid annuity over a period

of n years is

a

n

=

n

0

e

−δt

dt =

1 − e

−δn

δ

.

A life annuity which is payable continuously will thus have actuarial present value

a

x

= E[a

T(x)

] = E[

1 − e

−δT(x)

δ

].

Exercise 14–6. Show that A

x

= 1 − δa

x

. Find a similar relationship for a

x:n

.

This point of view makes it easy to understand certain modiﬁcations of discrete

annuities. When annuity beneﬁt payments are made at discrete intervals it is

customary to provide a ﬁnal adjustment which takes into account the death of the

annuitant between payment periods. One such modiﬁed annuity is called a complete

annuity immediate whose actuarial present value is denoted by ˚ a

x

. The payment

scheme for such an annuity is $1 at the end of each of the years 1 through K(x) − 1

plus a ﬁnal adjustment payment made at the time of death in order to make this

scheme actuarially equivalent to a continuous annuity. To determine the size of the

adjustment payment, ﬁrst determine the rate at which a continuous annuity must

be paid in order to be equivalent to a discrete annuity immediate. The following

picture illustrates the cash ﬂows.

§14: Life Annuities 53

Equivalence of an Annuity Immediate and a Continuous Annuity

0 1 0 1

1

σ

............................................................................................................................................................................................................................................................................................................. .............................................................................................................................................................................................................................................................................................................

Equating the present value of the two cash streams gives v =

1

0

σ e

−δt

dt from

which σ = δ/ i in order for the streams to be equivalent. It follows that the amount

of the ﬁnal payment, made at time T(x), for the complete annuity immediate must

be

e

δT(x)

T(x)

K(x)

δ

i

e

−δt

dt =

T(x)−K(x)

0

δ

i

e

δt

dt.

Also

˚ a

x

= E[

T(x)

0

δ

i

e

−δt

dt] =

δ

i

a

x

.

Exercise 14–7. When payments are made on an mthly basis (each payment being

1/ m) the actuarial present value of a complete annuity immediate is denoted by ˚ a

(m)

x

.

Find a formula for the adjustment payment and the actuarial present value in this

case.

Premium payments to an insurance company often take the form of an appor-

tioned annuity due. Here the payment scheme consists of $1 immediately and at

the beginning of each year through and including time K(x) less a refund payment

(from the company) at the time of death. The refund payment is made because

(in the premium context) the insured payed for a full year of coverage in advance.

Using techniques similar to those above shows that the refund payment, made at

time T(x), should be in the amount of

e

δT(x)

K(x)+1

T(x)

δ

d

e

−δt

dt =

K(x)+1−T(x)

0

δ

d

e

−δt

dt

and that the actuarial present value of an apportioned annuity due is

¨ a

{1}

x

= E[

T(x)

0

δ

d

e

−δt

dt] =

δ

d

a

x

.

Here, of course, the left most member of the equalities is the notational convention.

Exercise 14–8. In the case of mthly payments (each of size 1/ m) ﬁnd a formula for

¨ a

{m}

x

as well as for the size of the refund payment.

In order to compare apportioned and complete annuities, let us see how pre-

miums paid by a complete annuity immediate would operate. In such a scheme,

§14: Life Annuities 54

premiums would be paid at the end of each year, except that in the year of death

a reduced premium would be paid at the time of death. When viewed from the

insurer’s viewpoint in this way it is obvious that ¨ a

{1}

x

> ˚ a

x

.

There is one other idea of importance. In the annuity certain setting one may be

interested in the accumulated value of the annuity at a certain time. For an annuity

due for a period of n years the accumulated value of the annuity at time n, denoted

by ¨ s

n

, is given by ¨ s

n

= (1 + i)

n

¨ a

n

=

(1+i)

n

−1

d

. The present value of ¨ s

n

is the same as

the present value of the annuity. Thus the cash stream represented by the annuity is

equivalent to the single payment of the amount ¨ s

n

at time n. This last notion has an

analog in the case of life annuities. In the life annuity context

n

E

x

¨ s

x:n

= ¨ a

x:n

where

n

E

x

= v

n

n

p

x

is the actuarial present value (net single premium) of a pure

endowment of $1 payable at time n. Thus ¨ s

x:n

represents the amount of pure

endowment payable at time n which is actuarially equivalent to the annuity.

§14: Life Annuities 55

Problems

Problem 14–1. Show that under UDD

a

x

< a

(2)

x

< a

(3)

x

< ⋅ ⋅ ⋅ < a

x

< ⋅ ⋅ ⋅ < ¨ a

(3)

x

< ¨ a

(2)

x

< ¨ a

x

.

Give an example to show that without the UDD assumption the inequalities may

fail.

Problem 14–2. Show that

˚ a

x

< ˚ a

(2)

x

< ˚ a

(3)

x

< ⋅ ⋅ ⋅ < a

x

< ⋅ ⋅ ⋅ < ¨ a

{3}

x

< ¨ a

{2}

x

< ¨ a

{1}

x

.

Problem 14–3. Show that for any m we have ¨ a

{m}

x

< ¨ a

(m)

x

and that a

(m)

x

< ˚ a

(m)

x

.

Problem 14–4. True or false: A

x

1

:n

= 1 − d ¨ a

x

1

:n

. Hint: When does x

1

: n die?

Problem 14–5. True or false: ¨ s

x:n

≤ ¨ s

n

.

Problem 14–6. Use the life table to calculate the actuarial present value of $1000

due in 30 years if (40) survives.

Problem 14–7. Use the life table to compute a

21

and ¨ a

{4}

21

.

Problem 14–8. Find a general formula for

m| n

¨ a

x

and use it together with the life

table to compute

5| 10

¨ a

20

.

Problem 14–9. Prove a

x:n

=

1

E

x

¨ a

x+1:n

.

Problem 14–10. Show that under UDD

¨ a

(m)

x

= α(m)¨ a

x

− β(m).

Here α(m) =

id

i

(m)

d

(m)

and β(m) =

i − i

(m)

i

(m)

d

(m)

. The functions α(m) and β(m) deﬁned

here are standard actuarial functions.

Problem 14–11. Show that δ(Ia)

T

+ Tv

T

= a

T

.

Problem 14–12. Use the previous problem to show that δ(Ia)

x

+ (IA)

x

= a

x

. Here

(Ia)

x

is the actuarial present value of an annuity in which payments are made at rate

t at time t. Is there a similar formula in discrete time?

Problem 14–13. Show that ¨ a

x

= 1 + a

x

and that

1

m

+ a

(m)

x:n

= ¨ a

(m)

x:n

+

1

m

v

n

n

p

x

.

Problem 14–14. Show that ¨ a

x:n

= ¨ a

x

− v

n

n

p

x

¨ a

x+n

and use this to compute ¨ a

21:5

.

§14: Life Annuities 56

Solutions to Problems

Problem 14–1. As the type of annuity varies from left to right, the annuitant

receives funds sooner and thus the present value is higher.

Problem 14–2. Since ˚ a

(m)

x

= δ/ i

(m)

a

x

the result follows for the earlier relation-

ship between the rates of interest. A similar argument resolves the other half of

the inequalities.

Problem 14–3. The difference betwee the two sides of the inequalities is the

amount of the refund (or extra) payment.

Problem 14–4. The status dies only if (x) dies before time n. The result is true.

Problem 14–5. False.

Problem 14–8.

m| n

¨ a

x

=

m

E

x

¨ a

x+m

−

m+n

E

x

¨ a

x+m+n

.

Problem 14–10. See the text above.

Problem 14–11. Use integration by parts starting with the formula δ(Ia)

T

=

δ

T

0

t e

−δt

dt.

§14: Life Annuities 57

Solutions to Exercises

Exercise 14–1. In this case, E[a

K(x)

] = E[

1−v

K(x)

i

] =

1−v

−1

A

x

i

.

Exercise 14–2. Here there are [mT] +1 payments, so using the geometric series

formula gives ¨ a

(m)

x

= E[

¸

[mT]

j=0

(1/ m)v

j/ m

] = E[(1/ m)(1 − v

([mT]+1)/ m

)/ (1 − v

1/ m

)].

Now m(1 − v

1/ m

) = d

(m)

, which gives the result.

Exercise 14–3. To get your money’s worth, you must live long enough so that

the present value of the annual dues you would pay if you were not a life member

will exceed $675. This gives a condition that K(60) must satisfy if you are to

get your moneys worth.

Exercise 14–4. For the ﬁrst one ¨ a

x:n

= E[

1−v

(K+1)∧n

d

] = E[

1−v

(K+1)∧n

d

1

[0,n−1]

(K)] +

E[

1−v

(K+1)∧n

d

1

[n,∞)

(K)] = E[

1−v

(K+1)

d

1

[0,n−1]

(K)] + E[

1−v

n

d

1

[n,∞)

(K)] = (1/ d)(1 − A

x:n

).

A similar argument shows that a

x:n

= (1/ i)(A

x

1

:n

+

n

p

x

v

n+1

).

Exercise 14–5. The argument proceeds in a similar way, beginning with the

relation a

(m)

x

=

1−v

−1/ m

A

(m)

x

i

(m)

.

Exercise 14–6. The ﬁrst relationship follows directly from the given equation

and the fact that A

x

= E[e

−δT(x)

]. Since T(x : n) = T(x) ∧ n a similar argument

gives a

x:n

= (1/ δ)(1 − A

x:n

).

Exercise 14–7. In this case the payment rate for the corresponding continuous

annuity is δ/ i

(m)

, which gives ˚ a

(m)

x

=

δ

i

(m)

a

x

and the adjustment payment as

e

δT(x)

T(x)

[mT(x)]/ m

(δ/ i

(m)

)e

−δt

dt =

T(x)−[mT(x)]/ m

0

(δ/ i

(m)

)e

δt

dt.

Exercise 14–8. Here the rate is δ/ d

(m)

for the correspondingcontinuous annuity

so that ¨ a

{m}

x

= (δ/ d

(m)

)a

x

and the refund payment is e

δT

[mT]/ m+1/ m

T

δ

d

(m)

e

−δt

dt =

[mT]/ m+1/ m−T)

0

δ

d

(m)

e

−δt

dt.

§15. Laboratory 5

1. Show that ¨ a

x

= 1 + vp

x

¨ a

x+1

. Find formulas expressing ¨ a

(m)

x

and ¨ a

{m}

x

in terms

of ¨ a

x

.

2. The one step recursion formulas for annuities can be used just like the one

step recursions for insurances themselves. Use the q

x

values from Laboratory 3 and

i = 5% and compute ¨ a

x

, ¨ a

(12)

x

, and ¨ a

{12}

x

for x = 1 to x = 106. Place the result of your

computations into a nice table.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§16. Net Premiums

The common types of insurance policies can now be realistically analyzed from

an insurers point of view.

To develop the ideas consider the case of an insurer who wishes to sell a fully

discrete whole life policy which will be paid for by equal annual premiumpayments

during the life of the insured. The terminology fully discrete refers to the fact that

the beneﬁt is to be paid at the end of the year of death and the premiums are to

paid on a discrete basis as well. How should the insurer set the premium? A ﬁrst

approximation is given by the net premium. The net premiumis found by using the

equivalence principle: the premium should be set so that actuarial present value

of the beneﬁts paid is equal to the actuarial present value of the premiums received.

Using the equivalence principle the net premium P should satisfy

E[v

K(x)+1

] = PE[¨ a

K(x)+1

]

or

A

x

− P¨ a

x

= 0.

From here it is easy to determine the net premium, which in this case is denoted P

x

.

Exercise 16–1. Use the life table to ﬁnd the net premium, P

30

, for (30) if i = 0.05.

The notation for other net premiums for fully discrete insurances parallel the

notation for the insurance policies themselves. For example, the net annual premium

for an n year term policy with premiums payable monthly is denoted P

(12)

x

1

:n

.

Exercise 16–2. Use the life table to ﬁnd P

30

1

:10

. What is P

(12)

30

1

:10

? What is P

{1}

30

1

:10

?

Exercise 16–3. An h payment whole life policy is one in which the premiums are

paid for h years, beginnning immediately. Find a formula for

h

P

x

, the net annual

premium for an h payment whole life policy.

Example 16–1. As a more complicated example consider a recent insurance ad-

vertisement which I received. For a ﬁxed monthly premium payment (which is

constant over time) one may receive a death beneﬁt deﬁned as follows:

100000 1

[0,65)

(K(x)) + 75000 1

[65,75)

(K(x))

+ 50000 1

[75,80)

(K(x)) + 25000 1

[80,85)

(K(x)).

What is the net premium for such a policy? Assume that the interest rate is 5% so

that the life table can be used for computations. Using the equivalence principle,

the net annual premium P is the solution of the equation

P¨ a

(12)

x:85−x

= 100000 A

x

1

:65−x

+ 75000

65−x| 10

A

x

+ 50000

75−x| 5

A

x

+ 25000

80−x| 5

A

x

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§16: Net Premiums 60

in terms of certain term and deferred term insurances.

Exercise 16–4. Compute the actual net monthly premium for (21).

The methodology for ﬁnding the net premium for other types of insurance is

exactly the same. The notation in the other cases is now brieﬂy discussed. The most

common type of insurance policy is one issued on a semi-continuous basis. Here

the beneﬁt is paid at the time of death, but the premiums are paid on a discrete basis.

The notation for the net annual premium in the case of a whole life policy is P(A

x

).

The net annual premium for a semi-continuous term policy with premiums payable

mthly is P

(m)

(A

x

1

:n

). The notation for other semi-continuous policies is similar.

Exercise 16–5. What type of policy has net annual premium P

{m}

(A

x:n

)?

Policies issued on a fully continuous basis pay the beneﬁt amount at the time

of death and collect premiums in the form of a continuous annuity. Obviously, such

policies are of theoretical interest only. The notation here is similar to that of the

semi-continuous case, with a bar placed over the P. Thus P(A

x

) is the premium rate

for a fully continuous whole life policy.

§16: Net Premiums 61

Problems

Problem 16–1. Show that if δ = 0 then P(A

x

) = 1/ ˚ e

x

.

Problem 16–2. Arrange in increasing order of magnitude: P

(2)

(A

40:25

), P(A

40:25

),

P

{4}

(A

40:25

), P(A

40:25

), P

{12}

(A

40:25

).

Problem 16–3. If P(A

x

) = 0.03 and if interest is at the effective rate of 5%, ﬁnd

P

{2}

x

.

Problem 16–4. If

15

P

45

= 0.038, P

45:15

= 0.056 and A

60

= 0.625 ﬁnd P

45

1

:15

.

Problem 16–5. Recall that apportionable annuities differ from annuities due only

in the fact that the apportionable annuity offers the additional beneﬁt of a ‘pre-

mium refund’. Let A

PR

x

denote the net single premium for this refund beneﬁt for a

continuous whole life policy with apportioned premiums payable annually. Show

that

A

PR

x

=

P(A

x

)

δ

(A

x

− A

x

).

Problem 16–6. Refer to the previous problem and show that

P(A

PR

x

) =

P(A

x

)

δ ¨ a

x

(A

x

− A

x

).

Problem 16–7. Use the equivalence principle to ﬁnd the net annual premium for

a fully discrete 10 year term policy with beneﬁt equal to $10,000 plus the return,

with interest, of the premiums paid. Assume that the interest rate earned on the

premiums is the same as the interest rate used in determining the premium. Use the

life table to compute the premium for this policy for (21). How does this premium

compare with 10000P

21

1

:10

?

Problem 16–8. A level premium whole life insurance of 1, payable at the end

of the year of death, is issued to (x). A premium of G is due at the beginning

of each year provided (x) survives. Suppose L denotes the insurer’s loss when

G = P

x

, L

∗

denotes the insurer’s loss when G is chosen so that E[L

∗

] = −0.20, and

Var(L) = 0.30. Compute Var(L

∗

).

Problem 16–9. A policy issued to (x) has the following features.

(1) Apportionable premiums are payable annually.

(2) The ﬁrst premium is twice the renewal premium.

(3) Term insurance coverage for $100,000 plus the difference between the ﬁrst

and second premium is provided for 10 years.

§16: Net Premiums 62

(4) An endowment equal to the ﬁrst year premium is paid at the end of 10 years.

(5) Death claims are paid at the moment of death.

Use the equivalence principle to ﬁnd an expression for the renewal net annual

premium.

Problem 16–10. A $1000 whole life policy is issued to (50). The premiums are

payable twice a year, and are calculated on an apportionable basis. The beneﬁt is

payable at the moment of death. Calculate the semi-annual net premium given that

A

50

= 0.3, δ = 0.07, and e

−0.035

= 0.9656.

Problem 16–11. Polly, aged 25, wishes to provide cash for her son Tad, currently

aged 5, to go to college. Polly buys a policy which will provide a beneﬁt in the

form of a temporary life annuity due (contingent on Tad’s survival) in the amount of

$25,000 per year for 4 years commencing on Tad’s 18th birthday. Polly will make 10

equal annual premium payments beginning today. The 10 premium payments take

the formof a temporary life annuity due (contingent on Polly’s survival). According

to the equivalence principle, what is the amount of each premium payment? Use

the life table and UDD assumption (if necessary).

Problem 16–12. Snow White, presently aged 21, wishes to provide for the welfare

of the 7 dwarfs in the event of her premature demise. She buys a whole life policy

which will pay $7,000,000 at the moment of her death. The premium payments for

the ﬁrst 5 years will be $5,000 per year. According to the equivalence principle,

what should her net level annual premium payment be thereafter? Use the life table

and UDD assumption (if necessary).

Problem 16–13. The Ponce de Leon Insurance Company computes premiums for

its policies under the assumptions that i = 0.05 and µ

x

= 0.01 for all x > 0. What

is the net annual premium for a whole life policy for (21) which pays a beneﬁt

of $100,000 at the moment of death and has level apportioned premiums payable

annually?

§16: Net Premiums 63

Solutions to Problems

Problem 16–2. This is really a question about the present value of annuities.

Problem 16–3. Since iA

x

= δA

x

and d

(2)

¨ a

{2}

x

= δa

x

, P

{2}

x

= (d

(2)

/ i)P(A

x

).

Problem 16–4. Use the two equations P

45:15

= P

4

1

5:15

+

15

E

45

/ ¨ a

45:15

and

P

45

1

:15

=

A

45

−

15

E

45

A

60

¨ a

45:15

=

15

P

45

−

15

E

45

A

60

/ ¨ a

45:15

with the given information.

Problem 16–7. The present value of the beneﬁt is 10000v

K+1

1

[0,10)

(K) +

pv

K+1

¨ s

K+1

1

[0,10)

(K) where p is the premium.

Problem 16–8. The loss random variable is (1 + G/ d)v

K+1

− G/ d from which

the mean and variance in the two cases can be computed and compared.

Problem 16–10. The annual premium p satisﬁes p¨ a

{2}

50

= 1000A

50

.

Problem 16–11. The premium p satisﬁes p¨ a

25:10

= 25000

13| 4

¨ a

5

.

Problem 16–12. The premiump satisﬁes 7000000A

21

= 5000¨ a

21:5

+ p

5

E

21

¨ a

26

.

§16: Net Premiums 64

Solutions to Exercises

Exercise 16–1. From the table, P

30

= A

30

/ ¨ a

30

= 0.133866/ 18.189 = 0.0735.

Exercise 16–2. Now P

30

1

:10

= A

30

1

:10

/ ¨ a

30:10

. Also A

30

1

:10

= A

30

− v

10

10

p

30

A

40

=

0.133866 − (1.05)

−10

(94926/ 96477)(0.201506) = 0.01214. Similarly, ¨ a

30:10

=

¨ a

30

− v

10

10

p

30

¨ a

40

= 8.060, giving the premium as 0.001506. The other two

premiums differ only in the denominator, since P

(12)

30

1

:10

= A

30

1

:10

/ ¨ a

(12)

30:10

and

P

{1}

30

1

:10

= A

30

1

:10

/ ¨ a

{1}

30:10

. Now ¨ a

(12)

30:10

= ¨ a

(12)

30

− v

10

10

p

30

¨ a

(12)

40

. Since ¨ a

(12)

30

=

(id/ i

(12)

d

(12)

)¨ a

30

+ (i

(12)

− i)/ i

(12)

d

(12)

and a similar expression holds for ¨ a

(12)

40

,

the value of the annuity can be computed fromthe life table. A similar argument

and the fact that ¨ a

{1}

x

= (i/ δ)¨ a

x

+ (δ − i)/ δd allows the computation of the value

of the apportioned annuity. Note that the UDD assumption has been used here.

Exercise 16–3.

h

P

x

= A

x

/ ¨ a

x:h

.

Exercise 16–4. The net monthly premium is P/ 12 where P = (100000A

21

1

:44

+

75000v

44

44

p

21

A

65

1

:10

+ 50000v

54

54

p

21

A

75

1

:5

+ 25000v

59

59

p

21

A

80

1

:5

)/ ¨ a

(12)

21:64

. These

values can be computed from the life table using the techniques of an earlier

exercise.

Exercise 16–5. This is the premium for a continuous endowment policy with

mthly apportioned premiums.

§17. Laboratory 6

1. A 40 year term insurance policy with semi-annual premiums is issued to

(30). Assume that i = 5% and mortality follows the table given in Laboratory 3.

Compute the semi-annual net premium if the beneﬁt amount is 100,000 and the

beneﬁt is paid at the moment of death.

2. Rework the computation in problem 1 for an interest rate of 4% and 6%.

How much does the change in interest rate change the semi-annual premium?

3. Suppose the beneﬁt amount of the policy in problem 1 is 100,000 for death

in the ﬁrst 30 years (until age 60), and then decreases by 5,000 per year for each of

the remaining years. What is the semi-annual premium?

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§18. Insurance Models Including Expenses

Amore realistic viewof the insurance business includes provisions for expenses.

The proﬁt for the company can also be included here as an expense.

The common method used for the determination of the expense loaded pre-

mium (or the gross premium) is a modiﬁcation of the equivalence principle. Ac-

cording to the modiﬁed equivalence principle the gross premium G is set so that

on the policy issue date the actuarial present value of the beneﬁt plus expenses is

equal to the actuarial present value of the premium income. The premium is usually

assumed to be constant. Under these assumptions it is fairly easy to write a formula

to determine G. Assume that the expenses in policy year k are e

k−1

and are paid at

time k − 1, that is, at the beginning of the year. The actuarial present value of the

expenses is then given by

E[

K(x)

¸

k=0

v

k

e

k

] =

∞

¸

k=0

v

k

e

k k

p

x

.

Typically expenses are dependent on the premium. Also the sales commission is

usually dependent on the policy size.

Example 18–1. Suppose that the ﬁrst year expenses for a $100,000 semi-continuous

whole life policy are 20% of premiums plus a sales commission equal to 0.5% of

the policy amount, and that the expenses for subsequent years are 10% of premium

plus $5. The gross premium G for such a policy satisﬁes

100000A

x

+ (0.20G + 500) + (0.10G + 5)a

x

= G¨ a

x

.

An important, and realistic, feature of the above example is the large amount of

ﬁrst year expense. Expenses are now examined in greater detail.

Example 18–2. Let’s look at the previous example in the case of a policy for a

person aged 21. Assume that the interest rate is 5% and that the life table applies.

Then

G =

100, 000A

21

+ 495 + 5¨ a

21

0.9¨ a

21

− 0.1

= $604.24.

From this gross premium the company must pay $500 in ﬁxed expenses plus 20%

of the gross premium in expenses ($120.85), plus provide term insurance coverage

for the ﬁrst year, for which the net single premiumis 100, 000A

21

1

:1

= $123.97. Thus

there is a severe expected cash ﬂow strain in the ﬁrst policy year! The interested

reader may wish to examine the article “Surplus Loophole” in Forbes, September

4, 1989, pages 44-48.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§18: Insurance Models Including Expenses 67

Expenses typically consist of two parts. The ﬁrst part of the expenses can be

expressed as a fraction of gross premium. These are expenses which depend on

policy amount, such as sales commission, taxes, licenses, and fees. The other part

of expenses consist of those items which are independent of policy amount such

as data processing fees, printing of actual policy documents, clerical salaries, and

mailing expenses.

Studying the gross premium as a function of the beneﬁt provided can be useful.

Denote by G(b) the gross premium for a policy with beneﬁt amount b. The value

G(0) represents the overhead involved in providing the policy and is called the

policy fee. Typically the policy fee is not zero. The ratio G(b)/ b is called the

premium rate for a policy of beneﬁt b and reﬂects (approximately) the premium

change per dollar of beneﬁt change when the beneﬁt amount is b.

Exercise 18–1. In the example above ﬁnd R(b), the premium rate for a policy of

beneﬁt b.

§18: Insurance Models Including Expenses 68

Problems

Problem 18–1. The expense loaded annual premium for an 35 year endowment

policy of $10,000 issued to (30) is computed under the assumptions that

(1) sales commission is 40% of the gross premium in the ﬁrst year

(2) renewal commissions are 5% of the gross premium in year 2 through 10

(3) taxes are 2% of the gross premium each year

(4) per policy expenses are $12.50 per 1000 in the ﬁrst year and $2.50 per 1000

thereafter

(5) i = 0.05

Find the gross premium using the life table.

Problem18–2. Asemi-continuous whole life policy issued to (21) has the following

expense structure. The ﬁrst year expense is 0.4%of the policy amount plus $50. The

expenses in years 2 through 10 are 0.2% of the policy amount plus $25. Expenses in

the remaining years are $25, and at the time of death there is an additional expense

of $100. Find a formula for G(b). Compute G(1) and compare it to A

21

.

Problem 18–3. Your company sells supplemental retirement annuity plans. The

beneﬁt under such a plan takes the form of an annuity immediate, payable monthly,

beginning on the annuitant’s 65th birthday. Let the amount of the monthly beneﬁt

payment be b. The premiums for this annuity are collected via payroll deduction

at the end of each month during the annuitant’s working life. Set up expenses for

such a plan are $100. Subsequent expenses are $5 each month during the premium

collection period, $100 at the time of the ﬁrst annuity payment, and $5 per month

thereafter. Find G(b) for a person buying the plan at age x. What is R(b)?

Problem 18–4. A single premium life insurance policy with beneﬁts payable at the

end of the year of death is issued to (x). Suppose that

(1) A

x

= 0.25

(2) d = 0.05

(3) Sales commission is 18% of gross premium

(4) Taxes are 2% of gross premium

(5) per policy expenses are $40 the ﬁrst year and $5 per year thereafter

Calculate the policy fee that should be charged.

§18: Insurance Models Including Expenses 69

Solutions to Problems

Problem18–1. 10000A

30:35

+0.35G+0.05G¨ a

30:10

+(0.02G+25)¨ a

30:35

+100 =

G¨ a

30:35

.

Problem 18–2. bA

21

+0.002b+ 25 + 0.002b¨ a

21:10

+ 25¨ a

21

+ 100A

21

= G(b)¨ a

21

.

Problem18–3. 12b

65−x|

a

(12)

x

+5×12¨ a

(12)

x

+95+100

65+

1

12

−x

E

x

= G(b)×12a

(12)

x:65−x

.

Problem 18–4. G(b) = bA

x

+ 0.18G(b) + 0.02G(b) + 35 + 5¨ a

x

.

§18: Insurance Models Including Expenses 70

Solutions to Exercises

Exercise 18–1. Since the premium is bR(b) when the beneﬁt is b, the modiﬁed

equivalence principle gives bA

x

+ (0.20bR(b) + 0.005b) + (0.20bR(b) + 5)¨ a

x

=

bR(b)¨ a

x

from which R(b) = (bA

x

+ 0.005b + 5¨ a

x

)/ b(0.9¨ a

x

− 0.20).

§19. Multiple Lives

The study of the basic aspects of life insurance is now complete. Two different

but similar directions will now be followed in the ensuing sections. On the one

hand, types of insurance in which the beneﬁt is paid contingent on the death or

survival of more than one life will be examined. On the other hand, the effects of

competing risks on the cost of insurance will be studied.

The ﬁrst area of study will be insurance in which the time of the beneﬁt payment

depends on more than one life. Recall that a status is an artiﬁcially constructed

life form for which there is a deﬁnition of survival and death (or decrement). The

simplest type of status is the single life status. The single life status (x) dies exactly

when (x) does. Another simple status is the certain status n. This status dies at

the end of n years. The joint life status for the n lives (x

1

), . . . (x

n

) is the status

which survives until the ﬁrst member of the group dies. This status is denoted by

(x

1

x

2

. . . x

n

). The last survivor status, denoted by (x

1

x

2

. . . x

n

) is the status which

survives until the last member of the group dies.

When discussing a given status the question naturally arises as to how one

would issue insurance to such a status. If the constituents of the status are assumed

to die independently this problem can be easily solved in terms of what is already

known.

Example 19–1. Consider a fully discrete whole life policy issued to the joint status

(xy). The net annual premium to be paid for such a policy is computed as follows.

Using the obvious notation, the premium, P, must satisfy

A

xy

= P¨ a

xy

.

Using the deﬁnition of the joint life status gives

A

xy

= E[v

K(x)∧K(y)+1

]

and

¨ a

xy

=

1 − A

xy

d

which are obtained as previously.

Exercise 19–1. Obtain an expression for A

xy

in terms that can be computed from

the life table.

Exercise 19–2. Is A

xy

+ A

xy

= A

x

+ A

y

?

A useful technique for writing computational formulas directly is to ask the

question “Under what conditions is a payment made at time t?” The answer will

usually provide a computational formula.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§19: Multiple Lives 72

Example 19–2. What is ¨ a

xy

? This annuity makes a payment of 1 at time k if and

only if both (x) and (y) are alive at time k, and the probability of this is

k

p

xy

=

k

p

x k

p

y

.

Thus ¨ a

xy

=

∞

¸

k=0

v

k

k

p

x k

p

y

.

If one is willing to assume an analytical lawof mortality computations involving

joint lives can be simpliﬁed. Recall that two of the common analytical laws of

mortality are the Gompertz and Makeham laws. The Gompertz law has force of

mortality µ

x

= Bc

x

. It is easily seen that the joint survival of two independent lives

(x) and (y) is probabilistically identical with the survival of a single life (w) if and

only if

µ

(xy)+s

= µ

x+s

+ µ

y+s

= µ

w+s

.

When (x) and (y) have mortality which follows Gompertz’ Law this relation holds if

w satisﬁes c

x

+ c

y

= c

w

. A similar observation applies to Makeham’s law for which

the force of mortality is µ

x

= A+Bc

x

. In this case, however, it is necessary to mimic

the joint life (xy) by using a joint life (ww) at equal ages. Here w is the solution of

2c

w

= c

x

+ c

y

.

Exercise 19–3. Verify these assertions.

A status can also be determined by the order in which death occurs. The idea

here is similar to that used for term insurance earlier in which the status x

1

: n fails

at the time of death of (x) provided (x) dies ﬁrst. As a more complicated example

the status (x : y

2

) dies at the time of death of (y) provided (y) is the second to die.

Hence this status lives forever if (y) dies before (x). An insurance for such a status

is a simple case of what is known as a contingent insurance. Again, if the lives are

assumed to fail independently it is a simple matter to reduce computations involving

contingent insurance to the cases already considered.

§19: Multiple Lives 73

Problems

Problem 19–1. Show

t

p

xy

=

t

p

xy

+

t

p

x

(1 −

t

p

y

) +

t

p

y

(1 −

t

p

x

).

Problem 19–2. Suppose µ

x

= 1/ (110 − x) for 0 ≤ x < 110. Find

10

p

20:30

,

10

p

20:30

,

and ˚ e

20:30

.

Problem 19–3. Find an expression for the actuarial present value of a deferred

annuity of $1 payable at the end of any year as long as either (20) or (25) is living

after age 50.

Problem 19–4. Find the actuarial present value of a 20 year annuity due which

provides annual payments of $50,000 while both (x) and (y) survive, reducing by

25,000 on the death of (x) and by 12,500 on the death of (y).

Problem 19–5. Show that

n

q

x

1

y

=

n

q

xy

2

+

n

q

x n

p

y

.

Problem 19–6. Show that A

x

1

y

− A

xy

2

= A

xy

− A

y

.

Problem 19–7. If µ

x

= 1/ (100 − x) for 0 ≤ x < 100, calculate

25

q

25:50

2

.

Problem 19–8. In a mortality table which follows Makeham’s Law you are given

A = 0.003 and c

10

= 3. Calculate

∞

q

40

1

:50

if ˚ e

40:50

= 17.

Problem 19–9. If the probability that (35) will survive for 10 years is a and the

probability that (35) will die before (45) is b, what is the probability that (35) will

die within 10 years after the death of (45)? Assume the lives are independent.

Problem 19–10. Plot the survival function for Gompertz law of mortality with

A = 0.001 and c = 1.06.

Problem 19–11. Suppose (20) and (30) are independent lives that follow the Gom-

pertz law of mortality given in the previous problem. Plot the survival for the joint

life status 20 : 30. Is there a single age (x) whose survival function is the same as

the survival function of 20 : 30?

Problem 19–12. Plot the survival function for Makeham’s law of mortality with

A = 0.003, B = 0.001, and c = 1.06.

Problem 19–13. Suppose (20) and (30) are independent lives that followthe Make-

ham law of mortality given in the previous problem. Plot the survival for the joint

life status 20 : 30. Is there a single age (x) whose survival function is the same as

the survival function of 20 : 30?

§19: Multiple Lives 74

Solutions to Problems

Problem 19–1.

t

p

xy

= P[[T(x) ≥ t] ∪ [T(y) ≥ t]] = P[T(x) ≥ t, T(y) ≥

t] + P[T(x) ≥ t, T(y) ≤ t] + P[T(x) ≤ t, T(y) ≥ t].

Problem 19–2. From the form of the force of mortality, DeMoivre’s Law

holds.

Problem 19–3.

30|

a

20

+

25|

a

25

−

30|

a

20:25

.

Problem 19–4. The annuity pays 12,500 for 20 years no matter what so the

actuarial present value consists of 3 layers.

Problem 19–7. DeMoivre’s Law holds.

Problem 19–8. Here ˚ e

40:50

=

∞

0

t

p

40:50

dt = 17. Also, by conditioning,

P[T(40) < T(50)] =

∞

0

P[T(50) > t]f

T(40)

(t) dt =

∞

0

t

p

40:50

µ

40+t

dt while by a

symmetric argument P[T(40) > T(50)] =

∞

0

t

p

40:50

µ

50+t

dt. Using the form of

the force of mortality under Makeham’s Law, the fact that these two probabilities

sum to 1, and the given information completes the argument.

Problem 19–9. P[T(35) > T(45) + 10] =

∞

0

P[T(35) > t + 10]

t

p

45

µ

45+t

dt =

∞

0

P[T(35) > t+10| T(35) ≥ 10]P[T(35) ≥ 10]

t

p

45

µ

45+t

dt = a

∞

0

P[T(45)+10 >

t + 10]

t

p

45

µ

45+t

dt = a

∞

0

(

t

p

45

)

2

µ

45+t

dt =

∞

0

−

t

p

45

d

dt

t

p

45

dt = a/ 2. Thus the

desired probability is 1 − a/ 2 − b.

§19: Multiple Lives 75

Solutions to Exercises

Exercise 19–1. Using the independence gives

t

p

xy

=

t

p

xt

p

y

, so that A

xy

=

E[v

K(xy)+1

] =

¸

∞

k=0

v

k+1

(

k

p

xy

−

k+1

p

xy

) =

¸

∞

k=0

v

k+1

(

k

p

xk

p

y

−

k+1

p

xk+1

p

y

).

Exercise 19–2. Intuitively, either (x) dies ﬁrst or (y) dies ﬁrst, so the equation

is true. This can be veriﬁed by writing the expectations in terms of indicators.

Exercise 19–3. Under Makehamthe requirement is that (A+Bc

x+s

)+(A+Bc

y+s

) =

(A + Bc

w+s

) + (A + Bc

w+s

) for all s, and this holds if c

x

+ c

y

= 2c

w

.

§20. Laboratory 7

1. Find an expression for the net single premium for a whole life policy issued

to (x

1

y), a status which fails when (x) dies if T(x) < T(y). Use this expression and

the life table data of Laboratory 3 to compute the premium for a $1,000,000 policy

issued to (30

1

: 40). Use 6% as the interest rate.

2. Find an expression for the net single premiumfor a whole life policy issued to

(xy

2

), where the beneﬁt is paid on the death of (y) if T(x) < T(y). Use this expression

and the life table data of Laboratory 3 to compute the premium for a $1,000,000

policy issued to (30 : 40

2

). Use 6% as the interest rate.

3. Show that if X and Y are independent random variables and one of them is

absolutely continuous then P[X = Y] = 0. Hence under the standard assumptions of

this section no two people can die simultaneously.

4. One model for joint lives which allows for simultaneous death is the common

shock model. The intuition is that the two lives behave almost independently except

for the possibility of death by a common cause. The model is as follows. Let T

∗

(x),

T

∗

(y), and Z be independent random variables. Assume that T

∗

(x) and T

∗

(y) have

the distribution of the remaining lifetimes of (x) and (y) as given by the life table.

The randomvariable Z represents the time of occurrence of the common catastrophe

which will kill any survivors. The common shock model is that the true remaining

lifetimes of (x) and (y) are given as T(x) = min{T

∗

(x), Z} and T(y) = min{T

∗

(y), Z}

respectively. What is the probability that (x) and (y) die simultaneously in this

model? What is the survival function for the joint life status (xy) in this model?

Answer these two questions in general, and then in the special case in which

T

∗

(x), T

∗

(y), and Z have exponential distributions with parameters µ

x

, µ

y

, and µ

z

respectively.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§21. Multiple Decrement Models

In contrast to the case in which a status is deﬁned in terms of multiple lives, the

way in which a single life fails can be studied. This point of view is particularly

important in the context of the analysis of pension plans. In such a setting a person

may withdraw from the workforce (a ‘death’) due to accident, death, or retirement.

Different beneﬁts may be payable in each of these cases. Another common type

of insurance in which a multiple decrement model is appropriate is in the double

indemnity life policy. Here the beneﬁt is twice the face amount of the policy if

the death is accidental. In actuarial parlance the termination of a status is called a

decrement and multiple decrement models will now be developed. These models

also go by the name of competing risk models in other contexts.

To analyze the new situation, introduce the random variable J = J(x) which is a

discrete random variable denoting the cause of decrement of the status (x). Assume

that J(x) has as possible values the integers 1, . . . , m. It is clear that all of the

information of interest is contained in the joint distribution of the random variables

T(x) and J(x). Note that this joint distribution is of mixed type since (as always) T(x)

is assumed to be absolutely continuous while J(x) is discrete. The earlier notation is

modiﬁed in a fairly obvious way to take into account the new model. For example,

t

q

(j)

x

= P[0 < T(x) ≤ t, J(x) = j].

and

t

p

(j)

x

= P[T(x) > t, J(x) = j].

Here

∞

q

(j)

x

gives the marginal density of J(x). To discuss the probability of death

due to all causes the superscript (τ) is used. For example,

t

q

(τ)

x

= P[T(x) ≤ t] =

m

¸

j=1

t

q

(j)

x

and similar expressions for the survival probability and the force of mortality can

be obtained. Although

t

q

(τ)

x

+

t

p

(τ)

x

= 1 a similar equation for the individual causes

of death fails unless m = 1. For the force of mortality

µ

(τ)

x+t

=

f

T(x)

(t)

P[T(x) > t]

and

µ

(j)

x+t

=

f

T(x) J(x)

(t, j)

P[T(x) > t]

.

One must be careful in the use of these formulas. In particular, while

t

p

(τ)

x

= exp{−

t

0

µ

(τ)

x+s

ds}

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§21: Multiple Decrement Models 78

it is also the case that

t

p

(j)

x

≠ exp{−

t

0

µ

(j)

x+s

ds}

as one might not at ﬁrst expect. This latter integral does have an important use

which is explored below.

An important practical problem is that of constructing a multiple decrement life

table. To see how such a problem arises consider the case of a double indemnity

whole life policy. Assume that the policy will pay an amount $1 at the end of the

year of death if death occurs due to non-accidental causes and an amount of $2 if

the death is accidental. Denote the type of decrement as 1 and 2 respectively. The

present value of the beneﬁt is then

v

K(x)+1

1

{1}

(J(x)) + 2v

K(x)+1

1

{2}

(J(x)) = J(x)v

K(x)+1

.

To compute the net premium it remains to compute the expectation of this quantity.

This computation can only be completed if p

(j)

x

is known. Howare these probabilities

calculated?

There are two basic methodologies used. If a large group of people for which

extensive records are maintained is available the actual survival data with the deaths

in each year of age broken down by cause would also be known. It is then very easy

to construct the multiple decrement table. This is seldom the case.

Example 21–1. An insurance company has a thriving business issuing life insurance

to coal miners. There are three causes of decrement (death): mining accidents, lung

disease, and other causes. From the company’s vast experience with coal miners a

decrement (life) table for these three causes of decrement is available. The company

nowwants to enter the life insurance business for salt miners. Here the two causes of

decrement (death) are mining accidents and other. How can the information about

mining accidents for coal miners be used to get useful information about mining

accidents for salt miners?

A simple-minded answer to the question raised in the example would be to

simply lift the appropriate column from the coal miners life table and use it for the

salt miners. Such an approach fails, because it does not take into account the fact

that there are competing risks, that is, the accident rate for coal miners is affected by

the fact that some miners die from lung disease and thus are denied the opportunity

to die from an accident. The death rate for each cause in the absence of competing

risk is needed.

To see how to proceed the multiple decrement process is examined in a bit more

detail. Some auxillary quantities are introduced. Deﬁne

t

p

′

(j)

x

= exp{−

t

0

µ

(j)

x+s

ds}

§21: Multiple Decrement Models 79

and

t

q

′

(j)

x

= 1 −

t

p

′

(j)

x

.

The “probability”

t

q

′

(j)

x

is called the net probability of decrement (or absolute

rate of decrement). It is these “probabilities” that represent the death rates in the

absence of competing risks. To see why this interpretation is reasonable, note that

µ

(j)

x+t

=

f

T(x) J(x)

(t, j)

P[T(x) > t]

=

f

X J(x)

(x + t, j)/ s(x)

s(x + t)/ s(x)

=

f

X J(x)

(x + t, j)

P[X > x + t]

.

This shows that µ

(j)

x+t

represents the rate of death due to cause j among those surviving

up to time x + t.

These “probabilities” may be used to obtain the desired entries in a multiple

decrement table as follows. First

t

p

(τ)

x

=

m

¸

j=1

t

p

′

(j)

x

.

This shows how one can pass from the absolute rate of decrement to total survival

probabilities. Note that this relationship implies that the rates are generally larger

than the total survival probability. Then, under the assumption of constant force of

mortality for each decrement over each year of age in the multiple decrement table,

q

(j)

x

=

1

0

s

p

(τ)

x

µ

(j)

x+s

ds

=

1

0

s

p

(τ)

x

µ

(j)

x

ds

=

µ

(j)

x

µ

(τ)

x

1

0

s

p

(τ)

x

µ

(τ)

x

ds

=

µ

(j)

x

µ

(τ)

x

q

(τ)

x

=

log p

′

(j)

x

log p

(τ)

x

q

(τ)

x

.

This solves the problem of computing the entries in a multiple decrement table

under the stated assumption about the structure of the causes of decrement in that

table.

Exercise 21–1. What happens if p

(τ)

x

= 1?

§21: Multiple Decrement Models 80

Exercise 21–2. Show that the same formula results if one assumes instead that the

time of decrement due to each cause of decrement in the multiple decrement table

has the uniform distribution over the year of decrement. (This assumption means

that

t

q

(j)

x

= t ⋅ q

(j)

x

.)

Exercise 21–3. Assume that two thirds of all deaths at any age are due to accident.

What is the net single premium for (30) for a double indemnity whole life policy?

How does this premium compare with that of a conventional whole life policy?

The previous computations were based on assumptions about the causes of

decrement within the multiple decrement table. In some contexts it is more sensible

to make assumptions about the structure of the individual causes of decrement as

if each were acting independently, that is, to make assumptions about the absolute

rate of decrement in the single decrement tables.

Example 21–2. Suppose we are designing a pension plan and that there are two

causes of decrement: death and retirement. In many contexts (such as teaching) it

is reasonable to assume that retirements all occur at the end of a year, while deaths

can occur at any time. How could we construct a multiple decrement table which

reﬂects this assumption?

One common assumption about a single decrement is the assumption of uniform

distribution of deaths in the year of death. In the multiple decrement context this

translates in the statement that for 0 ≤ t ≤ 1

t

q

′

(j)

x

= t q

′

(j)

x

.

Exercise 21–4. Show that under this assumption we have

t

p

′

(j)

x

µ

(j)

x+t

= q

′

(j)

x

for 0 ≤

t ≤ 1. Hint: Compute

d

dt

t

p

′

(j)

x

in two different ways.

If this uniformity assumption is made for all causes of decrement it is then easy

to construct the multiple decrement table. The computations are illustrated for the

case of 2 causes of decrement. In this setting

q

(1)

x

=

1

0

s

p

(τ)

x

µ

(1)

x+s

ds

=

1

0

s

p

′

(1)

x

s

p

′

(2)

x

µ

(1)

x+s

ds

= q

′

(1)

x

1

0

s

p

′

(2)

x

ds

= q

′

(1)

x

1

0

(1 − sq

′

(2)

x

) ds

= q

′

(1)

x

(1 −

1

2

q

′

(2)

x

)

§21: Multiple Decrement Models 81

with a similar formula for q

(2)

x

. It is easy to see howthis procedure could be modiﬁed

for different assumptions about the decrement in each single decrement table.

Exercise 21–5. Construct a multiple decrement table in which the ﬁrst cause of

decrement is uniformly distributed and the second cause has all decrements occur

at the end of the year. The pension plan described in the example above illustrates

the utility of this technique.

Another approximation which is used to connect single and multiple decrement

tables makes use of the functions

L

x

=

1

0

l

x+t

dt

m

x

=

l

x

− l

x+1

L

x

.

The function m

x

is called the central death rate at age x. The central rate of death

is used in a special technique, called the central rate bridge, in the context of

multiple decrement tables. This technique is now brieﬂy described. Deﬁne

m

(τ)

x

=

1

0

t

p

(τ)

x

µ

(τ)

x+t

dt

1

0

t

p

(τ)

x

dt

and

m

(j)

x

=

1

0

t

p

(τ)

x

µ

(j)

x+t

dt

1

0

t

p

(τ)

x

dt

and

m

′

(j)

x

=

1

0

t

p

′

(j)

x

µ

(j)

x+t

dt

1

0

t

p

′

(j)

x

dt

.

The central rate bridge is based on the following approximation. First, under the

UDD assumption in each single decrement table

m

′

(j)

x

=

q

′

(j)

x

1 −

1

2

q

′

(j)

x

.

Second, under the UDD assumption in the multiple decrement table

m

(j)

x

=

q

(j)

x

1 −

1

2

q

(τ)

x

.

§21: Multiple Decrement Models 82

Thirdly, under the constant force assumption in the multiple decrement table

m

(j)

x

= µ

(j)

x

= m

′

(j)

x

.

Now assume that all of these equalities are good approximations in any case. This

assumption provides a way of connecting the single and multiple decrement tables.

There is no guarantee of the internal consistency of the quantities computed in this

way, since, in general, the three assumptions made are not consistent with each

other. The advantage of this method is that the computations are usually simpler

than for any of the ‘exact’ methods.

Exercise 21–6. Show that each of the above equalities hold under the stated as-

sumptions.

§21: Multiple Decrement Models 83

Problems

Problem 21–1. Assume that each decrement has a uniform distribution over each

year of age in the multiple decrement table to construct a multiple decrement table

from the following data.

Age q

′

(1)

x

q

′

(2)

x

q

′

(3)

x

62 0.020 0.030 0.200

63 0.022 0.034 0.100

64 0.028 0.040 0.120

Problem 21–2. Rework the preceding exercise using the central rate bridge. How

different is the multiple decrement table?

Problem 21–3. In a double decrement table where cause 1 is death and cause 2 is

withdrawal it is assumed that deaths are uniformly distributed over each year of age

while withdrawals between ages h and h + 1 occur immediately after attainment of

age h. In this table one sees that l

(τ)

50

= 1000, q

(2)

50

= 0.24, and d

(1)

50

= 0.06d

(2)

50

. What

is q

′

(1)

50

? How does your answer change if all withdrawals occur at midyear? At the

end of the year?

Problem 21–4. How would you construct a multiple decrement table if you were

given q

′

(1)

x

, q

′

(2)

x

, and q

(3)

x

? What assumptions would you make, and what formulas

would you use? What if you were given q

′

(1)

x

, q

(2)

x

, and q

(3)

x

?

§21: Multiple Decrement Models 84

Solutions to Problems

Problem 21–1. First, p

(τ)

62

= (.98)(.97)(.80) and q

(τ)

62

= 1 − p

(τ)

62

. Also p

′

(j)

62

=

1 − q

′

(j)

62

. From the relation q

(j)

62

=

log p

′

(j)

62

log p

(τ)

62

q

(τ)

62

the ﬁrst row of the multiple

decrement table can be found.

Problem 21–3. From the information d

(2)

50

= 240 and d

(1)

50

= 14. Since with-

drawals occur at the beginning of the year there are 1000 − 240 = 760 people

under observation of whom 14 die. So q

′

(1)

50

= 14/ 760. If withdrawals occur at

year end all 1000 had a chance to die so q

′

(1)

50

= 14/ 1000.

Problem21–4. The central rate bridge could be used. Is there an exact method

available?

§21: Multiple Decrement Models 85

Solutions to Exercises

Exercise 21–1. What would this mean for µ

(τ)

x

and the derivation?

Exercise 21–2. The assumption is that

t

q

(j)

x

= tq

(j)

x

for all j. Hence

t

p

(τ)

x

=

1 − tq

(τ)

x

and µ

(j)

x+s

=

d

ds

s

q

(j)

x

/

s

p

(τ)

x

= q

(j)

x

/

s

p

(τ)

x

. Substitution and integration gives

p

′

(j)

x

= e

−

1

0

µ

(j)

x+s

ds

= (1 − q

(τ)

x

)

q

(j)

x

/ q

(τ)

x

. Since p

(τ)

x

= 1 − q

(τ)

x

, the result follows by

substitution.

Exercise 21–3. The actuarial present value of the beneﬁt is (1/ 3) × 1 × A

x

+

(2/ 3) × 2 × A

x

= (5/ 3)A

x

, from which the premium is easily calculated.

Exercise 21–4. From the deﬁnition,

d

dt

t

p

′

(j)

x

= −

t

p

′

(j)

x

µ

(j)

x+t

, while from the

relation

t

p

′

(j)

x

= 1 − tq

′

(j)

x

,

d

dt

t

p

′

(j)

x

= −q

′

(j)

x

.

Exercise 21–5. Since cause 1 obeys UDD, q

(1)

x

= q

′

(1)

x

1

0

s

p

′

(2)

x

ds as in the

derivation above. For cause 2,

s

p

′

(2)

x

= 1 for s < 1, so q

(1)

x

= q

′

(1)

x

. For cause 2

proceed as in the derivation above to get q

(2)

x

= −

1

0

s

p

′

(1)

x

d

ds

s

p

′

(2)

x

ds. Now

s

p

′

(2)

x

is constant except for a jump of size −q

′

(2)

x

at s = 1. Hence q

(2)

x

= q

′

(2)

x

p

′

(1)

x

=

q

′

(2)

x

(1 − q

′

(1)

x

).

Exercise 21–6. Under UDD in the single decrement table

t

p

′

(j)

x

= 1 − tq

′

(j)

x

and

t

p

′

(j)

x

µ

(j)

x+t

= q

′

(j)

x

so m

′

(j)

x

=

1

0

q

′

(j)

x

dt/

1

0

(1 − tq

′

(j)

x

) dt = q

′

(j)

x

/ (1 −

1

2

q

′

(j)

x

). Under

UDD in the multiple decrement table µ

(j)

x+s

= q

(j)

x

/

s

p

(τ)

x

so that substitution gives

the result. Under the constant force assumption in the multiple decrement table,

µ

(j)

x+s

= µ

(j)

x

for all j and m

(i)

x

= µ

(j)

x

= m

′

(j)

x

by substitution.

§22. Laboratory 8

The service table at the end of these notes contains information about a group

of workers. There are 4 causes of decrement for this population. The ﬁrst cause

is death (d), the second cause is withdrawal (w) (termination of employment), the

third cause is incapacity (i), and the fourth cause is retirement (r).

1. Suppose a concerted effort by the company reduces the rate of on the job

injury (incapacity) by 1/ 3 at all ages. Recompute the entries in the service table.

2. Suppose that in addition to the incapacity improvement an enhanced salary

and beneﬁts plan reduces the withdrawal rate at all ages by 1/ 4. Recompute the

entries in the service table.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§23. Insurance Company Operations

The discussion thus far has been about individual policies. In the next few

sections the operations of the company as a whole are examined. This examination

begins with an overview of the accounting practices of an insurance company. This

is followed by a study of the behavior of the loss characteristics of groups of similar

policies. This last study leads to another method of setting premiums for a policy.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§24. Net Premium Reserves

A realistic model for both insurance policies and the method and amount of

premium payment is now in hand. The next question is how accounting principles

are applied to the ﬁnancial operations of insurance companies.

A basic review of accounting principles is given ﬁrst. There are three broad

categories of items for accounting purposes: assets, liabilities, and equity. Assets

include everything which is owned by the business. Liabilities include everything

which is owed by the business. Equity consists of the difference in the value of the

assets and liabilities. Equity could be negative. In the insurance context liabilities

are referred to as reserve and equity as surplus. When an insurance policy is

issued the insurance company is accepting certain ﬁnancial obligations in return for

the premium income. The basic question is how this information is reﬂected in the

accounting statements of the company. Some of the different accounting procedures

available will now be described. Keep in mind that this discussion only concerns

how the insurance company prepares accounting statements reﬂecting transactions

which have occurred. The method by which gross (or net) premiums are calculated

is not being changed!

Example 24–1. Suppose the following data for an insurance company is given.

Income for Year Ending December 31, 1990

Premiums 341,000

Investment Income 108,000

Expenses 112,000

Claims and Maturities 93,000

Increases in Reserves —

Net Income —

Balance Sheet

December 31, 1989 December 31, 1990

Assets 1,725,000 —

Reserves — 1,433,000

Surplus 500,000 —

The missing entries in the tables can be ﬁlled in as follows (amounts in thou-

sands). Total income is 341 + 108 = 449 while total expenses are 112 + 93 = 205,

so net income (before reserve contributions) is 449 − 205 = 244. Now the reserves

at the end of 1989 are 1, 725 − 500 = 1, 225, so the increase in reserves must be

1, 433 − 1, 225 = 208. The net income is 244 − 208 = 36. Hence the 1990 surplus

is 536 and the 1990 assets are 1,969.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§24: Net Premium Reserves 89

The central question in insurance accounting is “How are liabilities measured?”

The answer to this question has some very important consequences for the operation

of the company, as well as for the ﬁnancial soundness of the company. The general

equation is

Reserve at time t = Actuarial Present Value at time t of future beneﬁts

− Actuarial Present Value at time t of future premiums.

The only accounting assumption required is one regarding the premium to be used

in this formula. Is it the net premium, gross premium, or ???

The only point of view adopted here is that liabilities are measured as the

net level premium reserves. This is the reserve computed under the accounting

assumption that the premiumcharged for the policy is the net level premium. To see

that this might be a reasonable approach, recall that the equivalence principle sets

the premiumso that the actuarial present value of the beneﬁt is equal to the actuarial

present value of the premiums collected. However, it is clear that after the policy

is issued the present value of the beneﬁts and of the un-collected premiums will no

longer be equal, but will diverge in time. This is because the present value of the

unpaid beneﬁts will be increasing in time and the present value of the uncollected

premiums will decrease in time. The discrepency between these two amounts at any

time represents an unrealized liability to the company. To avoid a negative surplus

(technical bankruptcy), this liability must be offset in the accounting statments of

the company by a corresponding asset. Assume (for simplicity) that this asset takes

the form of cash on hand of the insurance company at that time. How does one

compute the amount of the reserve at any time t under this accounting assumption?

This computation is illustrated in the context of an example.

Example 24–2. Consider a fully discrete whole life policy issued to (x) in which

the premium is payable annually and is equal to the net premium. What is the

reserve at time k, where k is an integer? To compute the reserve simply note that

if (x) has survived until time k then the (curtate) remaining life of x has the same

distribution as K(x + k). The outstanding beneﬁt has present value v

K(x+k)+1

while

the present value of the remaining premium income is ¨ a

K(x+k)+1

times the annual

premium payment. Denote by

k

L the random variable which denotes the size of the

future loss at time k. Then

k

L = v

K(x+k)+1

− P

x

¨ a

K(x+k)+1

.

The reserve, denoted in this case by

k

V

x

, is the expectation of this loss variable.

Hence

k

V

x

= E[

k

L] = A

x+k

− P

x

¨ a

x+k

.

This is called the prospective reserve formula, since it is based on a look at the

future performance of the insurance portfolio.

§24: Net Premium Reserves 90

Aword about notation. In the example above the reserve has been computed for

a discrete whole life policy. The notation for the reserves for other types of policies

parallel the notation for the premiums for the policy. Thus

k

V

x

1

:n

is the reserve at

time k for an n year term policy. When discussing general principals the notation

k

V is used to denote the reserve at time k for a general policy.

Exercise 24–1. What types of policies have reserves

t

V(A

x

1

:n

),

k

V(A

x

1

:n

), and

k

V(¨ a

x

)?

Certain timing assumptions regarding disbursements and receipts have been

made in the previous computation. Such assumptions are always necessary, so they

are now made explicit. Assume that a premium payment which is due at time t is

paid at time t+; an endowment beneﬁt due at time t is paid at time t+; a death beneﬁt

payment due at time t is assumed to be paid at time t−, that is, just before time t.

Interest earned for the period is received at time t−. Thus

t

V

x

includes any interest

earned and also the effects of any non-endowment beneﬁt payments but excludes

any premium income payable at time t and any endowment payments to be made at

time t. Also assume that the premium charged is the net level premium. Therefore

the full technical description of what has been computed is the net level premium

terminal reserve. One can also compute the net level premium initial reserve

which is the reserve computed right at time t. This initial reserve differs from the

terminal reserve by the amount of premium received at time t and the amount of the

endowment beneﬁt paid at time t. Ordinarily one is interested only in the terminal

reserve.

In the remainder of this section methods of computing the net level premium

terminal reserve are discussed. For succintness, the term ‘reserve’ is always taken

to mean the net level premium terminal reserve unless there is an explicit statement

to the contrary.

Exercise 24–2. Show that

k

V

x

= 1 −

¨ a

x+k

¨ a

x

. From this lim

k→∞

k

V

x

= 1. Why is this

reasonable?

Exercise 24–3. Use the Life Table to compute the reserve for the ﬁrst ﬁve years

after policy issue for a fully discrete whole life policy to (20). Assume the policy

amount is equal to $100,000 and the premium is the net premium.

The reserve can be viewed in a different way. Typically an insurance company

has many identical policies in force. One may beneﬁt by studying the cash ﬂow

associated with this group of policies on the average. Here is an example.

Example 24–3. Let us examine the expected cash ﬂow associated with a whole life

policy issued to (x). Assume the premium is the net level premium and that the

policy is fully discrete. In policy year k + 1 (that is in the time interval [k, k + 1))

there are the following expected cash ﬂows.

§24: Net Premium Reserves 91

Time Income Cash on Hand

k− (beneﬁts just paid, interest just received)

k

V

x

k P

x k

V

x

+ P

x

k + 1− −q

x+k k

V

x

+ P

x

− q

x+k

k + 1− i(

k

V

x

+ P

x

) (1 + i)(

k

V

x

+ P

x

) − q

x+k

This ﬁnal cash on hand at time k+1− must be equal to the reserve for the policies

of the survivors. Thus

p

x+k k+1

V

x

= (1 + i)(

k

V

x

+ P

x

) − q

x+k

.

This provides an important formula connecting successive reserves.

Exercise 24–4. Show that

1

E

x+k k+1

V

x

=

k

V

x

+ P

x

− vq

x+k

.

The analysis of the previous example illustrates a general argument connecting

the reserves at successive time points.

k

V = Actuarial Present Value at time k of beneﬁts payable in [k, k + 1)

− Actuarial Present Value at time k of premiums payable in [k, k + 1)

+ vp

x+kk+1

V.

Such recursive formulas for reserves are especially useful for computational pur-

poses.

There are other ways to compute the reserve. First the reserve may be viewed as

maintaining the balance between income and expenses. Since at time 0 the reserve is

0 (because of the equivalence principle) the reserve can also be viewed as balancing

past income and expenses. This leads to the retrospective reserve formula

k

E

x k

V

x

= P

x

¨ a

x:k

− A

x

1

:k

.

This formula is derived as follows. Recall that

A

x

= A

x

1

:k

+ v

k

k

p

x

A

x+k

and

¨ a

x

= ¨ a

x:k

+ v

k

k

p

x

¨ a

x+k

.

Since the reserve at time 0 is zero,

0 = A

x

− P

x

¨ a

x

=

A

x

1

:k

+ v

k

k

p

x

A

x+k

− P

x

¨ a

x:k

+ v

k

k

p

x

¨ a

x+k

**where k is an arbitrary positive integer. Rearranging terms and using the prospective
**

formula for the reserve given above produces the retrospective reserve formula.

§24: Net Premium Reserves 92

Exercise 24–5. Sometimes the retrospective reserve formula is written as

h

V

x

= P

x

¨ a

x:h

/

h

E

x

−

h

k

x

= P

x

¨ s

x:h

−

h

k

x

where

h

k

x

is called the accumulated cost of insurance. Find an expression for

h

k

x

.

How would

t

k

x

be computed?

It is relatively straightforward to write expressions for the reserve for any of

the many possible types of insurance policy. Doing this is left as an exercise for

the reader. One should keep in mind that the important point here is to be able to

(ultimately) write a formula for the reserve which one can compute with the data

available in the life table. Hence continuous and/or mthly payment schemes need

to be reduced to their equivalent annual forms. Recursive formulas are also often

used.

§24: Net Premium Reserves 93

Problems

Problem 24–1. True or False: For 0 ≤ k < n,

k

V

x:n

= 1 −

¨ a

x+k:n−k

¨ a

x:n

. What happens at

k = n?

Problem 24–2. Find a formula for the reserve at the end of 5 years for a 10 year

term policy with beneﬁt $1 issued to (30) on a net single premium basis.

Problem 24–3. Show that for 0 ≤ t ≤ n

t

V(A

x:n

) =

P(A

x+t:n−t

) − P(A

x:n

)

a

x+t:n−t

.

This is called the premium difference formula for reserves. Find similar formulas

for the other types of insurance.

Problem 24–4. Show that for 0 ≤ t ≤ n

t

V(A

x:n

) =

1 −

P(A

x:n

)

P(A

x+t:n−t

)

A

x+t:n−t

.

This is called the paid up insurance formula for reserves. Find similar formulas

for the other types of insurance.

Problem 24–5. Find P

x

1

:n

if

n

V

x

= 0.080, P

x

= 0.024 and P

x:n

1

= 0.2.

Problem 24–6. Given that

10

V

35

= 0.150 and that

20

V

35

= 0.354 ﬁnd

10

V

45

.

Problem 24–7. Write prospective and retrospective formulas for

40

20

V(A

20

), the re-

serve at time 20 for a semi-continuous 40 payment whole life policy issued to

(20).

Problem 24–8. For a general fully discrete insurance let us suppose that the beneﬁt

payable if death occurs in the time interval (h − 1, h] is b

h

and that this beneﬁt is

paid at time h, that is, at the end of the year of death. Suppose also that the premium

paid for this policy at time h is π

h

. Show that for 0 ≤ t ≤ 1

t

p

x+k k+t

V + v

1−t

t

q

x+k

b

k+1

= (1 + i)

t

(

k

V + π

k

).

This gives a correct way to interpolate reserves at fractional years.

Problem 24–9. In the notation of the preceding problem show that for 0 ≤ t ≤ 1

k+t

V = v

1−t

(

1−t

q

x+k+t

b

k+1

+

1−t

p

x+k+t k+1

V) .

Problem 24–10. Show that under UDD,

h

k

V(A

x:n

) = (i/ δ)

h

k

V

x

1

:n

+

h

k

V

x:n

1

.

§24: Net Premium Reserves 94

Problem 24–11. Show that under UDD,

h

k

V

(m)

x:n

=

h

k

V

x:n

+ β(m)

h

P

(m)

x:n k

V

x

1

:n

. This gives

the reserves for a policy with mthly premium payments in terms of the reserves for

a policy with annual premium payments.

Problem 24–12. Show that

h

k

V

(m)

(A

x:n

) =

h

k

V(A

x:n

) + β(m)

h

P

(m)

(A

x:n

)

k

V

x

1

:n

under

UDD.

Problem 24–13. The amount at risk in year k for a discrete insurance is the

difference between the beneﬁt payment payable at the end of year k and

k

V. Find

the mean and variance of the amount at risk in year 3 of a 5 year term policy issued

to (30) which pays a beneﬁt of 1 at the end of the year of death and has net level

premiums.

Problem 24–14. Suppose that 1000

t

V(A

x

) = 100, 1000P(A

x

) = 10.50, and δ =

0.03. Find a

x+t

.

Problem 24–15. Calculate

20

V

45

given that P

45

= 0.014, P

45:20

1

= 0.022, and

P

45:20

= 0.030.

Problem 24–16. A fully discrete life insurance issued to (35) has a death beneﬁt of

$2500 in year 10. Reserves are calculated at i = 0.10 and the net annual premium

P. Calculate q

44

given that

9

V + P =

10

V = 500.

§24: Net Premium Reserves 95

Solutions to Problems

Problem 24–1. Use the prospective formula and A

x:n

+ d¨ a

x:n

= 1 to see the

formula is true. When k = n the reserve is 1 by the timing assumptions.

Problem 24–2. Prospectively the reserve is A

3

1

5:5

.

Problem 24–3. Use the prospective formula and the premium deﬁnitions.

Problem 24–5. From the retrospective formula

n

E

xn

V

x

= P

x

¨ a

x:n

− A

x

1

:n

. Now

divide by ¨ a

x:n

.

Problem 24–6. Use the prospective formula and the relation A

x

+ d¨ a

x

= 1 to

obtain

k

V

x

= 1 − ¨ a

x+k

/ ¨ a

x

.

Problem 24–7. The prospective and retrospective formulas are

40

20

V(A

20

) =

A

40

− P¨ a

40:20

and

40

20

V(A

20

) = (P¨ a

20:20

− A

20

1

:20

)/

20

E

20

.

Problem 24–8. The value of the reserve, given survival, plus the present value

of the beneﬁt, given death, must equal the accumulated value of the prior reserve

and premium.

Problem 24–13. The amount at risk random variable is 1

{2}

(K(30)) −

3

V

30

1

:5

.

Problem 24–14. Use the prospective reserve formula and the relationship

A

x

+ δa

x

= 1.

Problem 24–15. Use the retrospective formula.

Problem24–16. By the general recursion formula

1

E

4410

V =

9

V+P−2500vq

44

.

§24: Net Premium Reserves 96

Solutions to Exercises

Exercise 24–1.

t

V(A

x

1

:n

) is the reserve at time t for a fully continuous n year

term insurance policy,

k

V(A

x

1

:n

) is the reserve at time k for a semi-continuous n

year term policy, and

k

V(¨ a

x

) is the reserve at time k for a life annuity.

Exercise 24–2. Since A

x

+ d¨ a

x

= 1,

k

V

x

= A

x+k

− (A

x

/ ¨ a

x

)¨ a

x+k

= 1 − d¨ a

x+k

− (1 −

d¨ a

x

)¨ a

x+k

/ ¨ a

x

= 1 − ¨ a

x+k

/ ¨ a

x

.

Exercise 24–3. The reserve amounts are easily computed using the previous

exercise as 100000

1

V

20

= 100000(1 − 19.014/ 19.087) = 382.46, 100000

2

V

20

=

775.40, 100000

3

V

20

= 1184.06, 100000

4

V

20

= 1613.67, and 100000

5

V

20

=

2064.24.

Exercise 24–4. Just multiply the previous equation by v = (1 + i)

−1

.

Exercise 24–5.

h

k

x

= A

x

1

:k

/

k

E

x

, which can be easily computed from the life

table using previous identities.

§25. Laboratory 9

1. Return to Laboratory 6 and ﬁnd the reserve at the end of each policy year

for the term policy of problem 1 at an interest rate of 5%. Begin by deriving a

recurrence relation between the reserves for successive years.

2. Find the reserve at the end of each policy year for the term policy with de-

clining beneﬁts given in problem 3 of Laboratory 6. Begin by deriving a recurrence

relation between the reserves for successive years.

The problems below all concern the following situation. Your company has

just sold 3 year term insurance policies to a group of 20 persons, each aged 30. The

beneﬁt amount for each policy is $1,000,000 and is payable at the moment of death.

Each policy has level premiums payable at the beginning of each policy year. The

mortality characteristics are q

30

= 0.000444, q

31

= 0.000499, q

32

= 0.000562, and

q

33

= 0.000631. The interest rate is 4%.

3. Make a table showing the reserves for this group of policies at the end of

each policy year.

4. Your company charges a premium which is twice the net premium. Make a

table showing the expected free cash ﬂow generated by this group of policies at the

end of each year. Assume that there are no expenses.

5. Estimate the probability that this group of policies will require a cash

infusion from the other business of the company at some point during the life of

these policies. Also estimate the amount of cash required (if any). Carefully explain

your methodology and state your assumptions.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§26. The Individual Risk Model

An insurance company has a large number of policies in force at any given

time. This creates a ﬁnancial risk for the company. There are two aspects to the

problem of analyzing this risk. First, one must be able to estimate the amount of

risk. Secondly, one must be able to model the times at which claims will occur in

order to avoid cash ﬂow difﬁculties. The problem of modeling the amount of risk

will be studied ﬁrst.

In the individual riskmodel the insurers total risk S is assumed to be expressable

in the form S = X

1

+ . . . + X

n

where X

1

, . . . , X

n

are independent random variables

with X

i

representing the loss to the insurer on insured unit i. Here X

i

may be quite

different than the actual damages suffered by insured unit i. In the closed model

the number of insured units n is assumed to be known and ﬁxed. A model in which

migration in and out of the insurance system is allowed is called an open model.

The individual risk model is appropriate when the analysis does not require the

effect of time to be taken into account.

The ﬁrst difﬁculty is to ﬁnd at least a reasonable approximation to the proba-

bilistic properties of the loss random variables X

i

. This can often be done using data

from the past experience of the company.

Example 26–1. For short termdisability insurance the amount paid by the insurance

company can often be modeled as X = cY where c is a constant representing the

daily rate of disability payments and Y is the number of days a person is disabled.

One then is simply interested in modelling the random variable Y. Historical data

can be used to estimate P[Y > y]. In this context P[Y > y], which was previously

called the survival function, is referred to as the continuance function. The same

notion can be used for the daily costs of a hospitalization policy.

The second difﬁculty is to uncover the probabilistic properties of the random

variable S. In theoretical discussions the idea of conditioning can be used to ﬁnd

an explicit formula for the distribution function of a sum of independent random

varibles.

Example 26–2. Suppose X and Y are independent random variables each having

the exponential distribution with parameter 1. By conditioning

P[X + Y ≤ t] =

∞

−∞

P[X + Y ≤ t | Y = y] f

Y

(y) dy

=

t

0

P[X ≤ t − y] f

Y

(y) dy

=

t

0

1 − e

−(t−y)

f

Y

(y) dy

= 1 − e

−t

− te

−t

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§26: The Individual Risk Model 99

for t ≥ 0.

This argument has actually shown that if X and Y are independent and absolutely

continuous then

F

X+Y

(t) =

∞

−∞

F

X

(t − y) f

Y

(y) dy.

This last integral is called the convolution of the two distribution functions and is

often denoted F

X

∗ F

Y

.

Exercise 26–1. If X and Y are absolutely continuous random variables show that

X + Y is also absolutely continuous and ﬁnd a formula for the density function of

X + Y.

Exercise 26–2. Find a similar formula if X and Y are both discrete. Use this formula

to ﬁnd the density of X + Y if X and Y are independent Bernoulli random variables

with the same success probability.

An approach which requires less detailed computation is to appeal to the Central

Limit Theorem.

Central Limit Theorem. If X

1

, . . . , X

n

are independent random variables then the

distribution of

n

¸

i=1

X

i

is approximately the normal distribution with mean

n

¸

i=1

E[X

i

]

and variance

n

¸

i=1

Var(X

i

).

The importance of this theorem lies in the fact that the approximating normal

distribution does not depend on the detailed nature of the original distribution but

only on the ﬁrst two moments. The accuracy of this approximation will be explored

in the exercises and laboratory.

Example 26–3. You are a claims adjuster for the Good Driver Insurance Company

of Auburn. Based on past experience the chance of one of your 1000 insureds being

involved in an accident on any given day is 0.001. Your typical claimis $500. What

is the probability that there are no claims made today? If you have $1000 cash on

hand with which to pay claims, what is the probability you will be able to pay all

of todays claims? How much cash should you have on hand in order to have a

99% chance of being able to pay all of todays claims? What assumptions have you

made? How reasonable are they? What does this say about the solvency of your

company?

It is easily seen, by using the Central Limit Theorem, that if an insurance

company sold insurance at the pure premium not only would the company only

break even (in the long run) but due to random ﬂuctuations of the amount of

claims the company would likely go bankrupt. Thus insurance companies charge

§26: The Individual Risk Model 100

an amount greater than the pure premium. A common methodology is for the

company to charge (1+θ) times the pure premium. When this scheme is followed θ

is called the relative security loading and the amount θ × (pure premium) is called

the security loading. This is a reasonable procedure since the insureds with larger

expected claims pay a proportionate share of the loading. The relative loading θ is

usually adjusted to achieve a certain measure of protection for the company.

Example 26–4. Suppose that a company is going to issue 1,000 ﬁre insurance

policies each having a $250 deductible, and a policy amount of $50,000. Denote

by F

i

the Bernoulli random variable which is 1 if the ith insured suffers a loss, and

by D

i

the amount of damage to the ith insureds property. Suppose F

i

has success

probability 0.001 and that the actual damage D

i

is uniformly distributed on the

interval (0,70000)). What is the relative loading so that the premium income will

be 95% certain to cover the claims made? Using the obvious notation, the total

amount of claims made is given by the formula

S =

1000

¸

i=1

F

i

(D

i

− 250)1

[250,50250]

(D

i

) + 500001

(50250,∞)

(D

i

)

where the F’s and the D’s are independent (why?) and for each i the conditional

distribution of D

i

given F

i

= 1 is uniform on the interval (0,70000). The relative

security loading is determined so that

P[S ≤ (1 + θ) E[S]] = 0.95.

This is easily accomplished by using the Central Limit Theorem.

Exercise 26–3. Compute E[S] and Var(S) and then use the Central Limit Theorem

to ﬁnd θ. What is the probability of bankruptcy when θ = 0?

Another illustration is in connection with reinsurance. It is generally not good

practice for an insurance company to have all of its policy holders homogeneous,

such as all located in one geographical area, or all of the same physical type. A

moments reﬂection on the effect of a hurricane on an insurance company with all

of its property insurance business located in one geographic area makes this point

clear. An insurance company may diversify its portfolio of policies (or just protect

itself from such a concentration of business) by buying or selling reinsurance. The

company seeking reinsurance (the ceding company) buys an insurance policy from

the reinsurer which will reimburse the company for claims above the retention

limit. For stop loss reinsurance, the retention limit applies on a policy-by-policy

basis to those policies covered by the reinsurance. The retention limit plays the

same role here as a deductible limit in a stop loss policy. Usually there is one

reinsurance policy which covers an entire package of original policies. For excess

of loss reinsurance, the retention limit is applied to the total amount of claims for the

package of policies covered by the insurance, not the claims of individual policies.

§26: The Individual Risk Model 101

Example 26–5. You operate a life insurance company which has insured 2,000 30

year olds. These policies are issued in varying amounts: 1,000 people with $100,000

policies, 500 people with $500,000 policies, and 500 people with $1,000,000 poli-

cies. The probability that any one of the policy holders will die in the next year is

0.001. Stop loss reinsurance may be purchased at the rate of 0.0015 per dollar of

coverage. How should the retention limit be set in order to minimize the probabil-

ity that the total expenses (claims plus reinsurance expense) exceed $1,000,000 is

minimized? Let X, Y, and Z denote the number of policy holders in the 3 catagories

dying in the next year. Then X has the binomial distribution based on 1000 trials

each with success probability 0.001, Y has the binomial distribution based on 500

trials each with success probability 0.001, and Z has the binomial distribution based

on 500 trials each with success probability 0.001. If the retention limit is set at r

then the cost C of claims and reinsurance is given by

C = (100000 ∧ r)X + (500000 ∧ r)Y + (1000000 ∧ r)Z

+ 0.0015

1000(100000 − r)

+

+ 500(500000 − r)

+

+ 500(1000000 − r)

+

.

It is then a relatively straightforward, though tedious, task to use the central limit

theorem to estimate P[C ≥ 1, 000, 000].

Exercise 26–4. Verify the validity of the above formula. Use the central limit

theorem to estimate P[C ≥ 1, 000, 000] as a function of r. Find the value(s) of r

which minimize this probability.

§26: The Individual Risk Model 102

Problems

Problem 26–1. The probability of an automobile accident in a given time period is

0.001. If an accident occurs the amount of damage is uniformly distibuted on the

interval (0,15000). Find the expectation and variance of the amount of damage.

Problem 26–2. Find the distribution and density for the sum of three independent

random variables each uniformly distributed on the interval (0,1). Compare the

exact value of the distribution function at a few selected points (say 0.25, 1, 2.25)

with the approximation obtained from the central limit theorem.

Problem26–3. Repeat the previous problemfor 3 independent exponential random

variables each having mean 1. It may help to recall the gamma distribution here.

Problem 26–4. A company insures 1000 essentially identical cars. The probability

that any one car is in an accident in any given year is 0.001. The damage to a car

that is involved in an accident is uniformly distributed on the interval (0,15000).

What relative security loading θ should be used if the company wishes to be 99%

sure that it does not lose money?

§26: The Individual Risk Model 103

Solutions to Problems

Problem 26–1. The amount of damage is BU where B is a Bernoulli variable

with success probability 0.001 and U has the uniform distribution.

Problem 26–4. The loss random variable is of the form

1000

¸

i=1

B

i

U

i

.

§26: The Individual Risk Model 104

Solutions to Exercises

Exercise 26–1. Differentiation of the general distribution function formula

above gives f

X+Y

(t) =

∞

−∞

f

X

(t − y)f

Y

(y) dy.

Exercise 26–2. In the discrete case the same line of reasoning gives f

X+Y

(t) =

¸

y

f

X

(t−y)f

Y

(y). Applying this in the Bernoulli case, f

X+Y

(t) =

¸

m

y=0

n

t−y

p

t−y

(1−

p)

n−t+y

m

y

p

y

(1 − p)

m−y

= p

t

(1 − p)

n+m−t

¸

m

y=0

n

t−y

m

y

=

n+m

t

p

t

(1 − p)

m+n−t

.

Exercise 26–3. The loading θ is chosen so that θE[S]/

√

Var(S) = 1.645, from

the normal table. When θ = 0 the bankruptcy probability is about 1/ 2.

Exercise 26–4. Direct computation using properties of the binomial distribu-

tion gives E[C] = (100000∧r) ×1+(500000∧r) ×(1/ 2) +(1000000∧r) ×(1/ 2) +

0.0015

1000(100000 − r)

+

+ 500(500000 − r)

+

+ 500(1000000 − r)

+

and also

Var(C) = (100000∧r)

2

×0.999+(500000∧r)

2

×0.999/ 2+(1000000∧r)

2

×0.999/ 2.

The probability can now be investigated numerically using the Central Limit

Theorem approximation.

§27. Laboratory 10

Another method of uncovering the probabilistic properties of the random vari-

able S is to use simulation. Suppose that a company insures 10,000 essentially

identical cars. The probability that any one car is in an accident in any given year is

0.001. The damage to a car that is involved in an accident is uniformly distributed

on the interval (0,20000).

1. Write the loss random variable S in this case. Find E[S] and Var(S).

2. Use the Central Limit Theorem to ﬁnd the relative security loading θ that

should be used if the company wishes to be 99% sure that it does not lose money.

3. Consider one of the loss random variables X that occur in the expression for

S. Explain how a random number generator could be used to simulate X.

4. Use the method of problem 3 to simulate 5,000 observations on S. Make

a histogram of these observations. Based on this histogram, what relative security

loading should be used if the company wishes to be 99% sure that it does not lose

money? Compare the result here with that of problem 2.

5. Show that if Y is a random variable with increasing distribution function

F

Y

(t) and if U is uniformly distributed on the interval (0, 1), then F

−1

Y

(U) has the

same distribution as Y. This gives a general method to use for simulating random

variables.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§28. The Collective Risk Model and Ruin Probabilities

Some of the consequences of the collective risk model will now be examined.

In the collective risk model the time at which claims are made is taken to account.

Here the aggregate claims up to time t is assumed to be given by

¸

N(t)

k=1

X

k

where

X

1

, X

2

, . . . are independent identically distributed random variables representing the

sizes of the respective claims, N(t) is a stochastic process representing the number

of claims up to time t, and N and the X’s are independent. The object of interest is

the insurer’s surplus at time t, denoted by U(t), which is assumed to be of the form

U(t) = u + ct −

N(t)

¸

k=1

X

k

where u is the surplus at time t = 0, and c represents the rate of premium income.

Special attention will be given to the problem of estimating the probability that the

insurance company has negative surplus at some time t since this would mean that

the company is ruined.

To gain familiarity with some of the ideas involved, the simpler classical gam-

bler’s ruin problem will be studied ﬁrst.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§29. Stopping Times and Martingales

A discrete time version of the collective risk model will be studied and some

important new concepts will be introduced.

Suppose that a gambler enters a casino with z dollars and plays a game of chance

in which the gambler wins $1 with probability p and loses $1 with probability

q = 1−p. Suppose also that the gambler will quit playing if his fortune ever reaches

a > z and will be forced to quit by being ruined if his fortune reaches 0. The main

interest is in ﬁnding the probability that the gambler is ultimately ruined and also

the expected number of the plays in the game.

In order to keep details to a minimum, the case in which p = q = 1/ 2 will

be examined ﬁrst. Denote by X

j

the amount won or lost on the j

th

play of the

game. These random variables are all independent and have the same underlying

distribution function. Absent any restrictions about having to quit the game, the

fortune of the gambler after k plays of the game is

z +

k

¸

j=1

X

j

.

Nowin the actual game being played the gambler either reaches his goal or is ruined.

Introduce a random variable, T, which marks the play of the game on which this

occurs. Technically

T = inf{k : z +

¸

k

j=1

X

j

= 0 or a}.

Such a random variable is called a random time. Observe that for this speciﬁc

random variable the event [T ≤ k] depends only on the random variables X

1

, . . . , X

k

.

That is, in order to decide at time k whether or not the game has ended it is not

necessary to look into the future. Such special random times are called stopping

times. The precise deﬁnition is as follows. If X

1

, X

2

, . . . are random variables and

T is a nonnegative integer valued random variable with the property that for each

integer k the event [T ≤ k] depends only on X

1

, . . . , X

k

then T is said to be a stopping

time (relative to the sequence X

1

, X

2

, . . .).

The random variable z +

¸

T

j=1

X

j

is the gambler’s fortune when he leaves the

casino, which is either a or 0. Denote by ψ(z) the probability that the gambler

leaves the casino with 0. Then by direct computation E[z +

¸

T

j=1

X

j

] = a(1 − ψ(z)).

A formula for the ruin probability ψ(z) will be obtained by computing this same

expectation in a second way.

Each of the random variables X

j

takes values 1 and −1 with equal probability,

so E[X

j

] = 0. Hence for any integer k, E[

¸

k

j=1

X

j

] = 0 too. So it is at least plausible

that E[

¸

T

j=1

X

j

] = 0 as well. Using this fact, E[z +

¸

T

j=1

X

j

] = z, and equating this

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§29: Stopping Times and Martingales 108

with the expression above gives z = a(1 − ψ(z)). Thus ψ(z) = 1 − z/ a for 0 ≤ z ≤ a

are the ruin probabilities.

There are two important technical ingredients behind this computation. The

ﬁrst is the fact that T is a stopping time. The second is the fact that the gambling

game with p = q = 1/ 2 is a fair game. The notion of a fair game motivates the

deﬁnition of a martingale. Suppose M

0

, M

1

, M

2

, . . . are random variables. The

sequence is said to be a martingale if E[M

k

| M

k−1

, . . . , M

0

] = M

k−1

for all k ≥ 1. In

the gambling context, if M

k

is the gambler’s fortune after k plays of a fair game then

given M

k−1

the expected fortune after one more play is still M

k−1

.

Exercise 29–1. Show that M

k

= z +

¸

k

j=1

X

j

(with M

0

= z) is a martingale.

Example 29–1. The sequence M

0

= z

2

and M

k

=

z +

¸

k

j=1

X

j

2

− k for k ≥ 1 is also

a martingale. This follows from the fact that knowing M

0

, . . . , M

k−1

is the same as

knowing X

1

, . . . , X

k−1

and the fact that the X’s are independent.

Exercise 29–2. Fill in the details behind this example.

The important computational fact is the Optional Stopping Theorem which

states that if {M

k

} is a martingale and T is a stopping time then E[M

T

] = E[M

0

]. In

the gambling context this says that no gambling strategy T can make a fair game

biased.

Example 29–2. Using the martingale M

0

= z

2

and M

k

=

z +

¸

k

j=1

X

j

2

− k for

k ≥ 1 along with the same stopping time T as before can provide information

about the duration of the gambler’s stay in the casino. The random variable M

T

=

z +

¸

T

j=1

X

j

2

−T has an expectation which is easily computed directly to be E[M

T

] =

a

2

(1 − ψ(z)) − E[T]. By the optional stopping theorem, E[M

T

] = E[M

0

] = z

2

.

Comparing these two expressions gives E[T] = a

2

(1 − ψ(z)) − z

2

= az − z

2

as the

expected duration of the game.

The preceding example illustrates the general method. To analyze a particular

problem identify a martingale M

k

and stopping time T. Then compute E[M

T

] in two

ways, directly from the deﬁnition and by using the optional stopping theorem. The

resulting equation will often reveal useful information.

Uncovering the appropriate martingale is often the most difﬁcult part of the

process. One standard method is the following. If X

1

, X

2

, . . . are independent and

identically distributed random variables deﬁne

W

k

=

e

t

¸

k

j=1

X

j

E[e

t

¸

k

j=1

X

j

]

.

§29: Stopping Times and Martingales 109

Notice that the denominator is nothing more than the moment generating function

of the sum evaluated at t. For each ﬁxed t the sequence W

k

is a martingale (here

W

0

= 1). This follows easily from the fact that if X and Y are independent then

E[e

t(X+Y)

] = E[e

tX

] E[e

tY

]. This martingale is called Wald’s martingale (or the

exponential martingale) for the X sequence.

Exercise 29–3. Show that {W

k

: k ≥ 0} is a martingale no matter what the ﬁxed

value of t is.

In many important cases a non-zero value of t can be found so that the denomi-

nator part of the Wald martingale is 1. Using this particular value of t then makes

application of the optional stopping theorem neat and easy.

To illustrate the technique consider the following situation which is closer

to that of the collective risk model. Suppose the insurer has initial reserve z

and that premium income is collected at the rate of c per unit time. Also, X

k

denotes the claims that are payable at time k, and the X’s are independent and

identically distributed random variables. The insurers reserve at time k is then

z + ck −

¸

k

j=1

X

j

= z +

¸

k

j=1

(c − X

j

). Denote by T the time of ruin, so that

T = min{k : z + ck −

k

¸

j=1

X

j

≤ 0}.

The objective is to study the probability ψ(z) that ruin occurs in this setting.

As a ﬁrst step, notice that if E[c − X

j

] ≤ 0, ruin is guaranteed since premium

income in each period is not adequate to balance the average amount of claims in

the period. So to continue, assume that E[c − X

j

] > 0.

Under this assumption, suppose there is a number τ so that E[e

τ(c−X

j

)

] = 1.

This choice of τ in Wald’s martingale makes the denominator 1, and shows that

M

k

= e

τ(z+ck−

¸

k

j=1

X

j

)

is a martingale. Computing the expectation of M

T

using the

Optional Stopping Theorem gives E[M

T

] = E[M

0

] = e

τz

. Computing directly gives

E[M

T

] = E[e

τ(z+cT−

¸

T

j=1

X

j

)

| T < ∞] ψ(z). Hence

ψ(z) = e

τz

/ E[e

τ(z+cT−

¸

T

j=1

X

j

)

| T < ∞].

A problem below will show that τ < 0, so the denominator of this fraction is larger

than 1. Hence ψ(z) ≤ e

τz

. The ruin probability decays exponentially as the initial

reserve increases.

The usual terminology deﬁnes the adjustment coefﬁcient R = −τ. Thus

ψ(z) ≤ e

−Rz

. So a large adjustment coefﬁcient implies that the ruin probability

declines rapidly as the initial reserve increases.

§29: Stopping Times and Martingales 110

Problems

Problem 29–1. By conditioning on the outcome of the ﬁrst play of the game show

that in the gambler’s ruin problem ψ(z) = pψ(z + 1) + qψ(z − 1). Show that if p = q

there is a solution of this equation of the form ψ(z) = C

1

+ C

2

z and ﬁnd C

1

and C

2

by using the natural deﬁnitions ψ(0) = 1 and ψ(a) = 0. Show that if p ≠ q there is a

solution of the form ψ(z) = C

1

+ C

2

(q/ p)

z

and ﬁnd the two constants. This provides

a solution to the gambler’s ruin problem by using difference equations instead of

probabilistic reasoning.

Problem 29–2. In the gambler’s ruin problem, show that if p ≠ q the choice t =

ln(q/ p) makes the denominator of Wald’s martingale 1. Use this choice of t and the

optional stopping theorem to ﬁnd the ruin probability in this case.

Problem 29–3. Suppose p ≠ q in the gambler’s ruin problem. Deﬁne M

0

= z and

M

k

= z +

¸

k

j=1

X

j

− k(p − q) for k ≥ 1. Show that the sequence M

k

is a martingale and

use it to compute E[T] in this case.

Problem 29–4. Suppose that c > 0 is a number and X is a random variable which

takes on only non-negative values. Suppose also that E[c − X] > 0. Show that if

c − X takes on positive and negative values then there is a number τ < 0 so that

E[e

τ(c−X)

] = 1.

§29: Stopping Times and Martingales 111

Solutions to Problems

Problem 29–2. ψ(z) =

(q/ p)

a

−(q/ p)

z

(q/ p)

a

−1

.

Problem 29–3. E[T] =

z

q−p

−

a

q−p

1−(q/ p)

z

1−(q/ p)

a

.

Problem 29–4. Deﬁne a function f (v) = E[e

v(c−X)

]. Then f

′

(v) = E[(c −

X)e

v(c−X)

] and f

′′

(v) = E[(c − X)

2

e

v(c−X)

] > 0. Thus f is a convex function and

the graph of f is concave up. Now f (0) = 1 and f

′

(0) = E[(c − X)] > 0. Thus

the graph of f is above 1 to the right of 0, and below 1 (initially) to the left of

0. Since c − X takes on negative values, lim

v→−∞

f (v) = ∞, so there is a negative

value of v at which f (v) = 1, by continuity.

§29: Stopping Times and Martingales 112

Solutions to Exercises

Exercise 29–1. Knowing M

0

, . . . , M

k−1

is the same as knowingX

1

, . . . , X

k−1

. So

E[M

k

| M

0

, . . . , M

k−1

] = E[M

k

| X

0

, . . . , X

k−1

] = z +

¸

k−1

j=1

X

j

+ E[X

k

| X

0

, . . . , X

k−1

] =

M

k−1

since the last expectation is 0 by independence.

Exercise 29–2. First write M

k

=

z +

¸

k−1

j=1

X

j

+ X

k

2

− k =

z +

¸

k−1

j=1

X

j

2

+

2X

k

(z +

¸

k−1

j=1

X

j

) + X

2

k

− k. Take conditional expectations using the fact that X

k

is independent of the other X’s and E[X

k

] = 0 and E[X

2

k

] = 1 to obtain the result.

Exercise 29–3. Independencegives E[e

t

¸

k

j=1

X

j

] = E[e

t

¸

k−1

j=1

X

j

]×E[e

tX

k

]. Direct

computation of the conditional expectation gives the result.

§30. The Collective Risk Model Revisited

The ideas developed in connection with the gambler’s ruin problem will now

be used to compute the ruin probability in the collective risk model. Since the

processes are now operating in continuous time the details are more complicated

and not every step of the arguments will be fully justiﬁed.

In this setting the claims process is

¸

N(t)

k=1

X

k

where X

1

, X

2

, . . . are independent

identically distributed random variables representing the sizes of the respective

claims, N(t) is a stochastic process representing the number of claims up to time t,

and N and the X’s are assumed to be independent. The insurer’s surplus is given by

U(t) = u + ct −

¸

N(t)

k=1

X

k

, where u > 0 is the surplus at time t = 0 and c > 0 is the rate

at which premium income arrives per unit time. The probability of ruin with initial

surplus u will be denoted by ψ(u).

As in the discrete time setting, the Wald martingale will be used together with the

Optional Stopping Theoremin order to obtain information about the ruin probability.

Here the denominator of the Wald martingale is E[e

νU(t)

], and the ﬁrst step is to ﬁnd

a ν ≠ 0 so that E[e

ν(ct−

¸

N(t)

k=1

X

k

)

] = 1 no matter the value of t.

The new element in this analysis is the random sum

¸

N(t)

k=1

X

k

. Now for each

ﬁxed t, N(t) is a random variable which is independent of the X’s. The moment

generating function of this sumcan be easily computed by conditioning on the value

of the discrete random variable N(t).

E[e

ν

¸

N(t)

k=1

X

k

] = E[E[e

ν

¸

N(t)

k=1

X

k

| N(t)]]

=

∞

¸

j=0

E[e

ν

¸

N(t)

k=1

X

k

| N(t) = j] P[N(t) = j]

=

∞

¸

j=0

E[e

ν

¸

j

k=1

X

k

] P[N(t) = j]

=

∞

¸

j=0

E[e

νX

]

j

P[N(t) = j]

=

∞

¸

j=0

e

j ln(E[e

νX

])

P[N(t) = j]

= M

N(t)

(ln(M

X

(ν))).

Hence there is a ν≠0 so that E[e

ν(ct−

¸

N(t)

k=1

X

k

)

] = 1 if and only if e

νct

M

N(t)

(ln(M

X

(−ν))) =

1 for all t. Suppose for now that there is a number R > 0 so that

e

−Rct

M

N(t)

(ln(M

X

(R))) = 1

for all t. This number R is called the adjustment coefﬁcient. The existence of an

adjustment coefﬁcient will be investigated a bit later. Using −R as the value of ν in

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§30: The Collective Risk Model Revisited 114

Wald’s martingale shows that

W

t

= e

−R(u+ct−

¸

N(t)

k=1

X

k

)

is a martingale.

Deﬁne a stopping time T

a

by T

a

= inf{s : u + cs −

¸

N(s)

k=1

X

k

≤ 0 or ≥ a} where a

is an arbitrary but ﬁxed positive number. It is intuitively clear that T

a

is a stopping

time in an appropriate sense in the newcontinuous time setting. Nowby the Optional

Stopping Theorem, E[W

T

a

] = e

−Ru

. Direct computation gives

E[W

T

a

] = E[e

−R(u+cT

a

−

¸

N(T

a

)

k=1

X

k

)

| u + cT

a

−

N(T

a

)

¸

k=1

X

k

≤ 0] P[u + cT

a

−

N(T

a

)

¸

k=1

X

k

≤ 0]

+ E[e

−R(u+ct−

¸

N(t)

k=1

X

k

)

| u + cT

a

−

N(T

a

)

¸

k=1

X

k

≥ a] P[u + cT

a

−

N(T

a

)

¸

k=1

X

k

≥ a].

Since this equation is valid for any ﬁxed positive a, and since R > 0, it is possible

to take limits as a → ∞. Since lim

a→∞

P[u + cT

a

−

¸

N(T

a

)

k=1

X

k

) ≤ 0] = ψ(u) and

lim

a→∞

e

−Ra

= 0 the following result is obtained.

Theorem. Suppose that in the collective risk model the adjustment coefﬁcient R > 0

satisﬁes e

−Rct

M

N(t)

(ln(M

X

(R))) = 1 for all t. Let T = inf{s : u + cs −

¸

N(s)

k=1

X

k

≤ 0}

be the random time at which ruin occurs. Then

ψ(u) =

e

−Ru

E[e

−R(u+cT−

¸

N(T)

k=1

X

k

)

| T < ∞]

≤ e

−Ru

.

Exercise 30–1. Why is the last inequality true?

As in the discrete time model, the existence of an adjustment coefﬁcient guar-

antees that the ruin probability decreases exponentially as the initial surplus u

increases.

In general there is no guarantee that an adjustment coefﬁcient will exist. For

certain particular types of models the adjustment coefﬁcient can explicitly be found.

Moreover, a more detailed analysis of the claims process can be made in these

special cases.

The more restrictive discussion begins by examining the nature of the process

N(t), the total number of claims up to time t. A common assumption is that this

process is a Poisson process with constant intensity λ > 0. What this assumption

means is the following. Suppose W

1

, W

2

, . . . are independent identically distributed

exponential random variables with mean 1/ λ and common density λe

−λx

1

(0,∞)

(x).

§30: The Collective Risk Model Revisited 115

The W’s are the waiting times between claims. The Poisson process can then

be viewed as the number of claims that arrive up to time t. This means that

N(t) = inf{k :

¸

k+1

j=1

W

j

> t}. It can be shown that for any ﬁxed t the randomvariable

N(t) has the Poisson distribution with parameter λt and that the stochastic process

{N(t) : t ≥ 0} has independent increments, that is, whenever t

1

< t

2

< . . . < t

n

are

ﬁxed real numbers then the random variables N(t

2

) − N(t

1

), . . . , N(t

n

) − N(t

n−1

) are

independent. Using this, direct computation gives

E[e

νN(t)

] =

∞

¸

j=0

e

νj

P[N(t) = j]

=

∞

¸

j=0

eνje

−λt

(λt)

j

/ j!

= e

−λt

∞

¸

j=0

(e

ν

λt)

j

/ j!

= e

λt(e

ν

−1)

.

This simple formula for the moment generating function of N(t) leads to a simple

formula for the adjustment coefﬁcient in this case. The general equation for the

adjustment coefﬁcient was earlier found to be e

−Rct

M

N(t)

(ln(M

X

(R))) = 1. Taking

logarithms and using the form of the moment generating function of N(t) shows that

the adjustment coefﬁcient is the positive solution of the equation

λ + cR = λM

X

(R).

An argument similar to that given in the discrete time case can be used to show that

there is a unique adjustment coefﬁcient in this setting.

Exercise 30–2. Verify that the adjustment coefﬁcient, if it exists, must satisfy this

equation.

Example 30–1. Suppose all claims are for a unit amount. Then M

X

(ν) = e

ν

so the

adjustment coefﬁcient is the positive solution of λ + cR = λe

R

. Note that there is

no solution if c ≤ λ. But in this case the ruin probability is clearly 1.

Exercise 30–3. Show that if c ≤ λE[X] the ruin probability is 1. Show that if

c > λE[X] the adjustment coefﬁcient always exists and hence the ruin probability

is less than 1.

The previous exercises suggest that only the case in which c > λE[X] is of

interest. Henceforth write c = (1 + θ)λE[X] for some θ > 0. Here θ is the relative

security loading.

Even more detailed information can be obtained when N(t) is a Poisson process.

To do this deﬁne a stopping time T

u

= inf{s : U(s) < u} to be the ﬁrst time that

§30: The Collective Risk Model Revisited 116

the surplus falls below its initial level and denote by L

1

= u − U(T

u

) the amount by

which the surplus falls below its initial level. Then

P[T

u

< ∞, L

1

≥ y] =

1

(1 + θ)E[X]

∞

y

(1 − F

X

(x)) dx.

The proof of this fact is rather technical.

proof : Let h > 0 be small. Then P[N(h) = 0] = e

−λh

≈ 1, P[N(h) = 1] = λhe

−λh

≈ λh and

P[N(h) ≥ 2] ≈ 0. Denote by R(u, y) the probability that with an initial surplus of u the ﬁrst

time the surplus drops below 0, the surplus actually drops below −y. Conditioning on the

value of N(h) gives

R(u, y) ≈ (1 − λh)R(u + ch, y) + λh

u

0

R(u − x, y)f

X

(x) dx +

∞

u+ch+y

f

X

(x) dx

.

Re-arranging gives

R(u, y) − R(u + ch, y)

ch

= −

λ

c

R(u + ch, y) +

λ

c

u

0

R(u − x, y)f

X

(x) dx +

λ

c

u+ch+y

f

X

(x) dx.

Now take limits as h → 0 to obtain

−R

′

(u, y) = −

λ

c

R(u, y) +

λ

c

u

0

R(u − x, y)f

X

(x) dx +

λ

c

u+y

f

X

(x) dx.

Since R(u, y) ≤ ψ(u) ≤ e

−Ru

, both sides can be integrated with respect to u from 0 to ∞.

Doing this gives

R(0, y) = −

λ

c

∞

0

R(u, y) du +

λ

c

∞

0

u

0

R(u − x, y)f

X

(x) dx du +

λ

c

∞

0

∞

u+y

f

X

(x) dx du.

Interchanging the order of integration in the double integrals shows that the ﬁrst double

integral is equal to

∞

0

R(u, y) du, while the second double integral is equal to

∞

y

x−y

0

f

X

(x) du dx =

∞

y

(x − y)f

X

(x) dx

=

∞

y

xf

X

(x) dx − yP[X ≥ y]

=

∞

y

(1 − F

X

(x)) dx

after integration by parts. Substitution nowcompletes the proof after using c = (1+θ)λE[X].

This formula has two useful consequences. First, by taking y = 0, the probability

that the surplus ever drops below its initial level is 1/ (1 + θ). Second, an explicit

formula for the size of the drop below the initial level is obtained as

P[L

1

≤ y| T

u

< ∞] =

1

E[X]

y

0

(1 − F

X

(x)) dx.

§30: The Collective Risk Model Revisited 117

This expression can be evaluated in certain cases.

Exercise 30–4. Derive this expression for P[L

1

≤ y| T

u

< ∞].

Exercise 30–5. What is the conditional distribution of L

1

given T

u

< ∞ if the claim

size has an exponential distribution with mean 1/ δ?

Exercise 30–6. Show that the conditional moment generating function of L

1

given

T

u

< ∞ is (M

X

(t) − 1)/ (tE[X]).

This information can also be used to study the random variable L which repre-

sents the maximum aggregrate loss and is deﬁned by L = max

t≥0

{

¸

N(t)

k=1

X

k

− ct}.

Note that P[L ≤ u] = 1−ψ(u) fromwhich it is immediately seen that the distribution

of L has a discontinuity at the origin of size 1 − ψ(0) = θ/ (1 + θ), and is continuous

otherwise. In fact a reasonably explicit formula for the moment generating function

of L can be obtained.

Theorem. If N(t) is a Poisson process and L = max

t≥0

{

¸

N(t)

k=1

X

k

− ct} then

M

L

(ν) =

θE[X]ν

1 + (1 + θ)E[X]ν − M

X

(ν)

.

proof : Note from above that the size of each new deﬁcit does not depend on the initial starting

point of the surplus process. Thus

L =

D

¸

j=1

A

j

where A

1

, A

2

, . . . are independent identically distributed random variables each having the

same distribution as the conditional distribution of L

1

given T

u

< ∞, and D is a random

variable independent of the A’s which counts the number of times a new deﬁcit level is

reached. From here it is a simple matter to compute the moment generating function of

L.

Exercise 30–7. Complete the details of the proof.

This formula for the moment generating function of L can sometimes be used

to ﬁnd an explicit formula for the distribution function of L, and hence ψ(u) =

1 − P[L ≤ u].

There are some other interesting consequences of the assumption that the claim

number process N(t) is a Poisson process.

First, a bit of notation. If A and B are random variables, write A

d

= B to denote

that A and B have the same distribution.

§30: The Collective Risk Model Revisited 118

A random variable S is said to have the compound Poisson distribution

with Poisson parameter λ and mixing distribution F(x), denoted S

d

= CP(λ, F),

if S

d

=

¸

N

j=1

X

j

where X

1

, X

2

, . . . are independent identically distributed random vari-

ables with common distribution function F and N is a random variable which is

independent of the X’s and has a Poisson distribution with parameter λ.

Example 30–2. For each ﬁxed t, the aggregate claims process CP(λt, F

X

).

Example 30–3. If S

d

= CP(λ, F) then the moment generating function of S is

M

S

(ν) = exp{λ

∞

−∞

(e

uν

− 1) dF(u)}.

This follows from the earlier general derivation of the moment generating function

of a random sum.

Exercise 30–8. Suppose that S

d

= CP(λ, F) and T

d

= CP(δ, G) and that S and T are

independent. Show that S + T

d

= CP(λ + δ,

λ

λ+δ

F +

δ

λ+δ

G).

This last property is very useful in the insurance context. Because of this

property the results of the analysis of different policy types can be easily combined

into one grand analysis of the company’s prospects as a whole. Acompound Poisson

distribution can also be decomposed.

Example 30–4. Suppose each claim is either for $1 or $2, each event having

probability 0.5. If the number of claims is Poisson with parameter λ then the

amount of total claims, S, is compound Poisson distributed with moment generating

function

M

S

(ν) = exp{0.5λ(e

ν

− 1) + 0.5λ(e

2ν

− 1)}.

Hence S

d

= Y

1

+2Y

2

where Y

1

and Y

2

are independent Poisson random variables with

mean λ/ 2. Thus the number of claims of each size are independent!

Example 30–5. The collective risk model can be used as an approximation to

the individual risk model. In the individual risk model the claim amount is often

represented by a product B

j

X

j

in which B is a Bernoulli random variable which

represents whether a claim is paid or not and X is the amount of the claim. Then

BX

d

=

B

¸

j=1

X

′

j

≈

N

¸

j=1

X

′

j

where N has a Poisson distribution with parameter P[B = 1] and X

′ 1

, X

′ 2

, . . . are

independent random variables each having the same distribution as X. Thus the

distribution of BX may be approximated by the CP(P[B = 1], F

X

) distribution.

§30: The Collective Risk Model Revisited 119

Problems

Problem 30–1. If N has a Poisson distribution with parameter λ express P[N = k]

in terms of P[N = k − 1]. This gives a recursive method of computing Poisson

probabilities.

Problem 30–2. Show that if X takes positive integer values and S

d

= CP(λ, F

X

)

then x P[S = x] =

¸

∞

k=1

λkP[X = k]P[S = x − k] for x > 0. This is called Panjer’s

recursion formula. Hint: First show, using symmetry, that E[X

j

| S = x, N = n] =

x/ n for 1 ≤ j ≤ n and then write out what this means.

Problem 30–3. Suppose in the previous problem that λ = 3 and that X takes on the

values 1, 2, 3, and 4 with probabilities 0.3, 0.2, 0.1, and 0.4 respectively. Calculate

P[S = k] for 0 ≤ k ≤ 40.

Problem 30–4. Suppose S

1

has a compound Poisson distribution with λ = 2 and

that the compounded variable takes on the values 1, 2, or 3 with probabilities 0.2,

0.6, and 0.2 respectively. Suppose S

2

has a compound Poisson distribution with

parameter λ = 6 and the compounded variable takes on the values 3 or 4 with

probabilities 1/2 each. If S

1

and S

2

are independent, what is the distribution of

S

1

+ S

2

?

Problem 30–5. The compound Poisson distribution is not symmetric about its

mean, as the normal distribution is. One might therefore consider approximation of

the compound Poisson distribution by some other skewed distribution. A random

variable G is said to have the Gamma distribution with parameters α and β if G has

density function

f

G

(x) =

β

α

Γ(α)

x

α−1

e

−βx

1

(0,∞)

(x).

It is useful to recall the deﬁnition and basic properties of the Gamma function in this

connection. One easily computes the moments of such a random variable. In fact

the moment generating function is M

G

(ν) = (β/ β − ν)

α

. The case in which β = 1/ 2

and 2α is a positive integer corresponds to the chi–square distribution with 2α

degrees of freedom. Also the distribution of the sum of n independent exponential

random variables with mean 1/ β is a gamma distribution with parameters n and β.

For approximation purposes the shifted gamma distribution is used to approximate

the compound Poisson distribution. This means that an α, β, and x is found so that

x + G has approximately the same distribution as the compound Poisson variate.

The quantities x, α, and β are found by using the method of moments. The ﬁrst

three central moments of both random variables are equated, and the equations are

then solved. Show that when approximating the distribution of a compound Poisson

§30: The Collective Risk Model Revisited 120

random variable S the method of moments leads to

α =

4[Var(S)]

3

[E[(S − E[S])

3

]]

2

=

4λ

E[X

2

]

3

E[X

3

]

2

β =

2Var(S)

E[(S − E[S])

3

]

=

2E[X

2

]

E[X

3

]

x = E[S] −

2[Var(S)]

2

E[(S − E[S])

3

]

= λE[X] −

2λ

E[X

2

]

2

E[X

3

]

.

Problem 30–6. A random variable X is a mixture of exponential random variables

if the value of X is determined in the following way. Fix a number 0 < p < 1.

Perform a two stage experiment. In the ﬁrst stage, select a number U at random in

the interval (0, 1). For the second stage, proceed as follows. If U < p select the

value of X to be the value of an exponential random variable with parameter λ

1

. If

U > p select the value of X to be the value of an exponential random variable with

parameter λ

2

. Show that the density of X is f

X

(x) = pλ

1

e

−λ

1

x

+ (1 − p)λ

2

e

−λ

2

x

for

x ≥ 0. Show that the moment generating function of X is E[e

tX

] =

p

λ

1

−t

+

1−p

λ

2

−t

.

Problem 30–7. What is the density of a randomvariable X with moment generating

function E[e

tX

] = (30 − 9t)/ 2(5 − t)(3 − t) for 0 < t < 3?

Problem 30–8. In the continuous time model, if the individual claims X have

density f

X

(x) = (3e

−3x

+ 7e

−7x

)/ 2 for x > 0 and θ = 1, ﬁnd the adjustment coefﬁcient

and ψ(u).

Problem30–9. In the continuous time model, if the individual claims X are discrete

with possible values 1 or 2 with probabilities 1/ 4 and 3/ 4 respectively, and if the

adjustment coefﬁcient is ln(2), ﬁnd the relative security loading.

Problem 30–10. Use integration by parts to show that the adjustment coefﬁcient in

the continuous time model is the solution of the equation

∞

0

e

rx

(1−F

X

(x)) dx = c/ λ.

Problem 30–11. In the continuous time model, use integration by parts to ﬁnd

M

L

1

(t). Find expressions for E[L

1

], E[L

2

1

] and Var(L

1

). Here L

1

is the random

variable which is the amount by which the surplus ﬁrst falls below its initial level,

given that this occurs.

Problem 30–12. Find the moment generating function of the maximum aggregate

loss random variable in the case in which all claims are of size 5. What is E[L]?

Hint: Use the Maclaurin expansion of M

X

(t) to ﬁnd the Maclaurin expansion of

M

L

(t).

§30: The Collective Risk Model Revisited 121

Problem 30–13. If ψ(u) = 0.3e

−2u

+ 0.2e

−4u

+ 0.1e

−7u

, what is the relative security

loading?

Problem 30–14. If L is the maximum aggregate loss random variable, ﬁnd expres-

sions for E[L], E[L

2

], and Var(L) in terms of moments of X.

Problem 30–15. In the compound Poisson continous time model suppose that

λ = 3, c = 1, and X has density f

X

(x) = (e

−3x

+ 16e

−6x

)/ 3 for x > 0. Find the relative

security loading, the adjustment coefﬁcient, and an explicit formula for the ruin

probability.

Problem 30–16. In the compound Poisson continous time model suppose that

λ = 3, c = 1, and X has density f

X

(x) =

9x

25

e

−3x/ 5

for x > 0. Find the relative security

loading, the adjustment coefﬁcient, and an explicit formula for the ruin probability.

What happens if c = 20?

Problem 30–17. The claim number random variable is sometimes assumed to have

the negative binomial distribution. A random variable N is said to have the n

¯

egative

binomial distribution with parameters p and r if N counts the number of failures

before the rth success in a sequence of independent Bernoulli trials, each having

success probability p. Find the density and moment generating function of a random

variable N with the negative binomial distribution. Deﬁne the compound negative

binomial distribution and ﬁnd the moment generating function, mean, and variance

of a random variable with the compound negative binomial distribution.

Problem 30–18. In the case of ﬁre insurance the amount of damage may be quite

large. Three common assumptions are made about the nature of the loss variables in

this case. One is that X has a lognormal distribution. This means that X

d

= e

Z

where

Z

d

= N(µ, σ

2

). A second possible assumption is that X has a Pareto distribution.

This means that X has a density of the form αx

0

/ x

α+1

1

[x

0

,∞)

(x) for some α > 0. Note

that a Pareto distribution has very heavy tails, and the mean and/or variance may

not exist. A ﬁnal assumption which is sometimes made is that the density of X is a

mixture of exponentials, that is,

f

X

(t) = (0.7)λ

1

e

−λ

1

t

+ (0.3)λ

2

e

−λ

2

t

for example. After an assumption is made about the nature of the underlying

distribution one may use actual data to estimate the unknown parameters. For each

of the three models ﬁnd the maximum likelihood estimators and the method of

moments estimators of the unknown parameters.

Problem 30–19. For automobile physical damage a gamma distribution is often

postulated. Find the maximum likelihood and method of moments estimators of the

unknown parameters in this case.

§30: The Collective Risk Model Revisited 122

Problem 30–20. One may also examine the beneﬁts, in terms of risk reduction,

of using reinsurance. Begin by noting the possible types of reinsurance available.

First there is proportional reinsurance. Here the reinsurer agrees to pay a fraction

α, 0 ≤ α ≤ 1, of each individual claim amount. Secondly, there is stop–loss

reinsurance, in which the reinsurer pays the amount of the individual claim in

excess of the deductible amount. Finally, there is excess of loss reinsurance in

which the reinsurer pays the amount by which the claims of a portfolio of policies

exceeds the deductible amount. As an example, the effect of stop–loss reinsurance

with deductible d on an insurer’s risk will be analyzed. The amount of insurer’s

risk will be measured by the ruin probability. In fact, since the ruin probability is

so difﬁcult to compute, the effect of reinsurance on the adjustment coefﬁcient will

be measured. Recall that the larger the adjustment coefﬁcient, the smaller the ruin

probability. Initially (before the purchase of reinsurance) the insurer’s surplus at

time t is

U(t) = u + ct −

N(t)

¸

j=1

X

j

where c = (1 + θ)λE[X] and N(t) is a Poisson process with intensity λ. The

adjustment coefﬁcient before the purchase of reinsurance is the positive solution of

λ + cr = λM

X

(r).

After the purchase of stop loss reinsurance with deductible d the insurer’s surplus is

U

′

(t) = u + c

′

t −

N(t)

¸

j=1

(X

j

∧ d)

where c

′

= c − reinsurance premium. Note that this process has the same structure

as the original one. The new adjustment coefﬁcient is therefore the solution of

λ + c

′

r = λM

X∧d

(r).

By examining the reinsurance procedure from the reinsurer’s standpoint it is clear

that the reinsurer’s premium is given by

(1 + θ

′

)λE[(X − d)1

[d,∞)

(X)]

where θ

′

is the reinsurer’s relative security loading. With this information the new

adjustment coefﬁcient can be computed. Carry out these computations when λ = 2,

θ = 0.50, θ

′

= 0.25, d = 750, and X has a exponential distribution with mean 500.

Problem 30–21. Repeat the previous problem for the case of proportional reinsur-

ance.

Problem 30–22. The case of excess of loss reinsurance leads to a discrete time

model, since the reinsurance is applied to a portfolio of policies and the reinsurance

§30: The Collective Risk Model Revisited 123

is paid annually (say). The details here a similar to those in the discussion of the

discrete time gamblers ruin problem. Analyze the situation described in the previous

problem if the deductible for an excess of loss policy is 1500 and the rest of the

assumptions are the same. Which type of reinsurance is better?

Problem 30–23. In the discrete time model, suppose the X’s have the N(10, 4)

distribution and the relative security loading is 25%. A reinsurer will reinsure a

fraction f of the total portfolio on a proportional basis for a premium which is 140%

of the expected claimamount. Find the insurers adjustment coefﬁcient as a function

of f . What value of f maximizes the security of the ceding company?

§30: The Collective Risk Model Revisited 124

Solutions to Problems

Problem 30–2. By symmetry, E[X

1

| S = x, N = n] = x/ n while direct com-

putation gives E[X

1

| S = x, N = n] =

¸

∞

k=1

k P[X

1

= k| S = x, N = n]. Now

P[X

1

= k, S = x, N = n] = P[X

1

= k,

¸

n

j=1

X

j

= x, N = n] = P[X

1

=

k] P[

¸

n

j=2

X

j

= x − k] P[N = n] = P[X

1

= k] P[

¸

n

j=2

X

j

= x − k] P[N = n −

1]λ/ n = P[X

1

= k] P[S = x − k, N = n − 1]λ/ n. Making this substitution gives

xP[S = x, N = n] =

¸

∞

k=1

λkP[X

1

= k] P[S = x − k, N = n − 1]. Summing both

sides on n from 0 to ∞ gives the result.

Problem 30–4. The sum has a compound Poisson distribution with λ = 8.

Problem 30–6. Compute the distribution function of X by conditioning on

U to obtain F

X

(x) = P[X

1

≤ x] p + P[X

2

≤ x] (1 − p) for x ≥ 0 where X

1

and

X

2

are exponentially distributed random variables with parameters λ

1

and λ

2

respectively.

Problem 30–7. Use partial fractions and the previous problem to see that X

is a mixture of two exponentially distributed random variables with parameters

λ

1

= 3 and λ

2

= 5 and p = 1/ 4.

Problem 30–8. Here M

X

(t) = (5t − 21)/ (t − 3)(t − 7) and E[X] = 5/ 21. This

leads to R = 1.69. Also M

L

(t) = 1/ 2 − 0.769/ (t − 1.69) − 0.280/ (t − 6.20) using

partial fractions. Hence the density of L is f

L

(t) = 0.769e

−1.69t

+ 0.280e

−6.20t

together with a jump of size 1/ 2 at t = 0. (Recall that L has both a discrete and

absolutely continuous part.) Thus ψ(u) = 0.454e

−1.69u

+ 0.045e

−6.20u

.

Problem 30–9. Here θ = 10/ 7 ln(2) − 1 = 1.0609.

Problem 30–11. The density of L

1

is f

L

1

(t) = (1 − F

X

(t))/ E[X] for t ≥ 0.

Integration by parts then gives M

L

1

(t) = (M

X

(t) − 1)/ tE[X]. Using the Maclaurin

expansion of M

X

(t) = 1+tE[X] +t

2

E[X

2

]/ 2+. . . then gives M

L

1

(t) = 1+

E[X

2

]

2E[X]

t +

E[X

3

]

6E[X]

t

2

+ . . ., from which the ﬁrst two moments of L

1

can be read off.

Problem 30–12. M

L

(t) = 5θt/ (1 + 5(1 + θ)t − e

5t

) = 1 + t

E[X

2

]

2θE[X]

+ . . ..

Problem 30–13. Here θ = 2/ 3 since ψ(0) = 1/ (1 + θ).

Problem 30–14. Substitute the Maclaurin expansion of M

X

(t) into the expres-

sion for moment generating function of L in order to get the Macluarin expansion

of M

L

(t).

Problem30–15. Here θ = 4/ 5 and R = 2. Also M

L

(t) = 4/ 9+(8/ 9)

1

2−t

+(4/ 9)

1

4−t

so that ψ(u) = (4/ 9)e

−2u

+ (1/ 9)e

−4u

.

Problem30–16. Here M

X

(t) =

9

25

(3/ 5−t)

−2

so that E[X] = 10/ 3 and θ = −9/ 10

when c = 1. Since θ < 0, there is no adjustment coefﬁcient and the ruin

probability is 1. When c = 20, θ = 1 and R = 0.215. Also M

L

(t) =

1

2

−0.119/ (t −

§30: The Collective Risk Model Revisited 125

0.215) + 0.044/ (t − 0.834) by partial fractions. The density of the absolutely

continuous part of L is f

L

(t) = 0.119e

−.0.215t

− 0.044e

−0.834t

, and the distribution

of L has a jump of size 1/ 2 at the origin. So ψ(u) = 0.553e

−0.215u

− 0.053e

−0.834u

.

Problem 30–17. The compound negative binomial distribution is the distribu-

tion of the random sum

N

¸

i=1

X

i

where N and the X’s are independent, N has the

negative binomial distribution, and the X’s all have the same distribution. Now

P[N = k] =

k+r−1

r−1

p

r

(1 − p)

k

for k ≥ 0 and M

N

(t) = p

r

(1 − (1 − p)e

t

)

−r

. Now use

the general result about the moment generating function of a random sum.

Problem 30–20. Here M

X

(t) = (1 − 500t)

−1

for t < 1/ 500. The adjustment

coefﬁcient before reinsurance is then R = 1/ 1500. The reinsurance premium is

(1 + 0.25)2E[(X − 750)1

(750,∞)

(X)] = 278.91 and the insurer’s new adjustment

coefﬁcient is the solution of 2 + (1500 − 278.91)R = M

X∧750

(R) which gives

R = 0.00143.

Problem 30–21. As in the preceding problem, R = 1/ 1500 before reinsurance.

Suppose the insurer retains 100(1−α)%of the liability. The reinsurance premium

is then (1 + θ

′

)λαE[X] = 1250α. The adjustment coefﬁcient after reinsurance

is then the solution of 2 + (1500 − 1250α)R = 2M

(1−α)X

(R) = 2M

X

((1 − α)R). So

R = (2−α)/ 500(5α

2

−11α+6) which is always at least 1/ 1500. Notice that since

θ

′

< θ here, the insurer should pass off all of the risk to the reinsurer. By using

α = 1 the insurer collects the difference between the original and reinsurance

premiums, and has no risk of paying a claim.

Problem 30–22. The computational details here are quite complicated. In

a time interval of unit length the total claims are C =

¸

N

j=1

X

j

where N is a

Poisson random variable with parameter λ. Now recall that in the discrete time

setting the adjustment coefﬁcient is the solution of the equation E[e

R(c−C)

] = 1.

As before c = 1500. Also M

C

(t) = e

λ(M

X

(t)−1)

. So the adjustment coefﬁcient

before reinsurance is 1/ 1500. The reinsurance premium with deductible 1500

is λ(1 + θ

′

)E[(C − 1500)1

(1500,∞)

(C)] = 568.12. This is obtained numerically

by conditioning on the value of N and using the fact that conditional on N = k,

C has a gamma distribution with parameters α = k and β = 1/ 500. The new

adjustment coefﬁcient solves E[e

R(1500−568.12−C∧1500)

] = 1.

Problem 30–23. Here M

X

(t) = e

10t+2t

2

, the premium income is 12.5 for each

time period, and the adjustment coefﬁcient is the solution of e

−12.5t

M

X

(t) = 1

which gives R = 1.25. The reinsurance premium is 14f so that after reinsurance

the adjustment coefﬁcient satisﬁes e

−(12.5−14f )t

M

X

((1 − f )t) = 1, which gives

R = (5 − 8f )/ 4(1 − 2f + f

2

). The value f = 1/ 4 produces the maximum value of

R, namely 4/ 3.

§30: The Collective Risk Model Revisited 126

Solutions to Exercises

Exercise 30–1. Since R > 0 and u + cT −

¸

N(T)

k=1

X

k

≤ 0 when T < ∞ the

denominator expectation is at least 1.

Exercise 30–2. M

U(t)−u

(−R) = 1 holds if and only if −ctR − λt(M

X

(R) − 1) = 0,

which translates into the given condition.

Exercise 30–3. If c ≤ λE[X] premium income is less than or equal to the

average rate of the claim process. So eventually the company will be ruined

by a run of above average size claims. By Maclaurin expansion, M

X

(R) =

1 + E[X]R + E[X

2

]R

2

/ 2 + . . . and all of the coefﬁcients are positive since X is a

positive randomvariable. So λ +cR−λM

X

(R) = (c −λE[X])R−E[X

2

]R

2

/ 2 −. . .

is a function which is positive for R near 0 and negative for large values of R.

Thus there is some positive value of R for which this function is zero.

Exercise 30–4. From the deﬁnition of conditional probability, P[L

1

≤ y| T

u

<

∞] = P[L

1

≤ y, T

u

< ∞]/ P[T

u

< ∞] and the result follows from the previous

formula and the fact that P[T

u

< ∞] = 1/ (1 + θ).

Exercise 30–5. Since in this case F

X

(t) = 1 − e

−δt

for t > 0, direct substitution

gives P[L

1

≤ y| T

u

< ∞] = 1 − e

−δy

for y > 0.

Exercise 30–6. Given T

u

< ∞the density of L

1

is (1−F

X

(y))/ E[X] for y > 0. Us-

ing integration by parts then gives the conditional moment generating function of

L

1

as

∞

0

e

ty

(1−F

X

(y))/ E[X] dy = e

ty

(1 − F

X

(y))/ tE[X]

∞

0

+

∞

0

e

ty

f

X

(y)/ tE[X] dy =

(M

X

(t) −1)/ tE[X]. Notice that the unconditional distribution of L

1

has a jump of

size θ/ (1 + θ) at the origin. The unconditional moment generating function of

L

1

is θ/ (1 + θ) + (M

X

(t) − 1)/ (1 + θ)tE[X].

Exercise 30–7. Since P[D = k] = (θ/ (1 + θ))(1/ (1 + θ))

k

for k = 0, 1, 2, . . .,

conditioning gives M

L

(t) = E[e

t

¸

D

j=1

A

j

] = E[M

A

(t)

D

] =

¸

∞

k=0

M

A

(t)

k

(θ/ (1 +

θ))(1/ (1 + θ))

k

= (θ/ (1 + θ))/ (1 − M

A

(t)/ (1 + θ)) = θ/ (1 + θ − M

A

(t)) and this

simpliﬁes to the desired result using the formula of the previous exercise.

Exercise 30–8. Using the independence, M

S+T

(ν) = M

S

(ν)M

T

(ν) and the result

follows by substituion and algebraic rearrangement.

§31. Related Probability Models

In the next few sections some probability models are discussed which can be

used as models for transactions other than life insurance.

Discrete and continuous time Markov chains are often used as models for

a sequence of random variables which are dependent. One application of such

stochastic processes is as a model for the length of stay of a patient in a nursing

home.

The Brownian motion process is often used as a building block for a model of

stock prices. The deﬁnition and simple properties of Brownian motion are developed

here.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§32. Discrete Time Markov Chains

In many situations the random variables which serve naturally as a model are

not independent. The simplest kind of dependence allows future behavior to depend

on the present situation.

Example 32–1. Patients in a nursing home fall into 3 categories, and each category

of patient has a differing expense level. Patients who can care for themselves with

minimal assistance are in the lowest expense category. Other patients require some

skilled nursing assistance on a regular basis and are in the next higher expense

category. Finally, some patients require continuous skilled nursing assistance and

are in the highest expense category. One way of modeling the level of care a

particular patient requires on a given day is as follows. Denote by X

i

the level of

care this patient requires on day i. Here the value of X

i

would be either 1, 2, or

3 depending on which of the 3 expense categories is appropriate for day i. It is

intuitively clear that the random variables {X

i

} are not independent.

Possibly the simplest type of dependence structure for a sequence of random

variables is that in which the future probabilistic behavior of the sequence depends

only on the present value of the sequence and not on the entire history of the

sequence. A sequence of random variables {X

n

: n = 0, 1, . . .} is said to be a

Markov chain if

(1) P[X

n

∈ {0, 1, 2, 3, . . .}] = 1 for all n and

(2) for any real numbers a < b and any ﬁnite sequence of non-negative integers

t

1

< t

2

< ⋅ ⋅ ⋅ < t

n

< t

n+1

,

P[a < X

t

n+1

≤ b| X

t

1

, . . . , X

t

n

] = P[a < X

t

n+1

≤ b| X

t

n

].

The second requirement is referred to as the Markov property.

The possible values of the chain are called states.

Exercise 32–1. Show that any sequence of independent discrete random variables

is a Markov chain.

Because of the simple dependence structure a vital role is played by the tran-

sition probabilities P[X

n+1

= j| X

n

= i]. In principle, this probability depends not

only on the two states i and j, but also on n. A Markov chain is said to have

stationary transition probabilities if the transition probabilities P[X

n+1

= j| X

n

= i]

do not depend on n. In the discussion here, the transition probabilities will always

be assumed to be stationary, and the notation P

i,j

= P[X

n+1

= j| X

n

= i] will be used.

The transition probabilities together with the distribution of X

0

determine com-

pletely the probabilistic behavior of the Markov chain. This is seen in the following

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§32: Discrete Time Markov Chains 129

computation.

P[X

n

= i

n

, . . . , X

0

= i

0

]

= P[X

n

= i

n

| X

n−1

= i

n−1

, . . . , X

0

= i

0

] × P[X

n−1

= i

n−1

, . . . , X

0

= i

0

]

= P[X

n

= i

n

| X

n−1

= i

n−1

] × P[X

n−1

= i

n−1

, . . . , X

0

= i

0

]

= P

i

n−1

,i

n

P[X

n−1

= i

n−1

, . . . , X

0

= i

0

]

= . . .

= P

i

n−1

,i

n

P

i

n−2

,i

n−1

. . . P

i

0

,i

1

P[X

0

= i

0

].

Exercise 32–2. Justify each of the steps here completely. Where was the Markov

property used?

Exercise 32–3. Showthat for a Markov chain with stationary transition probabilities

P[X

4

= 3, X

3

≠ 3| X

2

= 3, X

1

≠ 3, X

0

= 3] = P[X

2

= 3, X

1

≠ 3| X

0

= 3]. Generalize.

For a Markov chain with stationary transition probabilities it is useful to collect

the transition probabilities into the transition matrix P = [P

i,j

] of the chain. This

matrix may be inﬁnite in extent.

Exercise 32–4. Show that if P is a transition matrix then

¸

j

P

i,j

= 1 for each i.

Example 32–2. In the previous nursing home example, suppose the transition

matrix is P =

¸

¸

0.9 0.05 0.05

0.1 0.8 0.1

0 0.05 0.95

¸

**. Then using conditioning it is easy to compute
**

P[X

3

= 2| X

0

= 1].

Example 32–3. The gambler’s ruin problem illustrates many of the features of a

Markov chain. A gambler enters a casino with $z available for wagering and sits

down at her favorite game. On each play of the game, the gambler wins $1 with

probability p and loses $1 with probability q = 1 − p. She will happily leave the

casino if her fortune reaches $a > 0, and will deﬁnitely leave, rather unhappily,

if her fortune reaches $0. Denote by X

n

the gambler’s fortune after the nth play.

Clearly {X

n

} is a Markov chain with P[X

0

= z] = 1. The natural state space here is

{0, 1, . . . , a}.

Exercise 32–5. Find the (a + 1) × (a + 1) transition matrix.

Even with the simplifying assumption of stationary transition probabilities the

formula for the joint distribution of the values of the chain is unwieldy, especially

since in most cases it is the long term behavior of the chain that is of interest.

Fortunately, it is possible to ﬁnd relatively simple answers to the following central

questions.

§32: Discrete Time Markov Chains 130

(1) If {X

n

} is a Markov chain with stationary transition probabilities, what is the

limiting distribution of X

n

?

(2) If s is a state of a Markov chain with stationary transition probabilities how

often is the process in state s?

As a warm up exercise for studying these questions the n step transition

probabilities deﬁned by P

n

i,j

= P[X

n+m

= j| X

m

= i] and the corresponding n step

transition probability matrix P

(n)

will now be computed.

Exercise 32–6. Show that P[X

n+m

= j| X

m

= i] does not depend on m.

Theorem. The n step transition probability matrix is given by P

(n)

= P

n

where P is

the transition probability matrix.

proof : The case n = 1 being clear, the induction step is supplied.

P

n

i,j

= P[X

n+m

= j| X

m

= i]

= P[[X

n+m

= j] ∩

∞

¸

k=0

[X

n+m−1

= k]

]| X

m

= i]

=

∞

¸

k=0

P[[X

n+m

= j, X

n+m−1

= k]| X

m

= i]

=

∞

¸

k=0

P[X

n+m

= j| X

n+m−1

= k, X

m

= i] P[X

n+m−1

= k| X

m

= i]

=

∞

¸

k=0

P[X

n+m

= j| X

n+m−1

= k] P[X

n+m−1

= k| X

m

= i]

=

∞

¸

k=0

P

k,j

P

(n−1)

i,k

.

The induction hypothesis together with the formula for the multiplication of matrices

conclude the proof.

Using this lemma gives the following formula for the density of X

n

in terms of

the density of X

0

.

( P[X

n

= 0] P[X

n

= 1] . . . ) = ( P[X

0

= 0] P[X

0

= 1] . . . )P

n

.

Exercise 32–7. Verify that this formula is correct.

Consequently, if X

n

converges in distribution to Y as n → ∞ then

( P[Y = 0] P[Y = 1] . . . ) = lim

n→∞

( P[X

n

= 0] P[X

n

= 1] . . . )

= lim

n→∞

( P[X

0

= 0] P[X

0

= 1] . . . )P

n

= lim

n→∞

( P[X

0

= 0] P[X

0

= 1] . . . )P

n+1

= ( P[Y = 0] P[Y = 1] . . . )P

§32: Discrete Time Markov Chains 131

which gives a necessary condition for Y to be a distributional limit for the chain,

namely, the density of Y must be a left eigenvector of P corresponding to the

eigenvalue 1.

Example 32–4. For the nursing home chain given earlier there is a unique left

eigenvector of P corresponding to the eigenvalue 1, after normalizing so that the

sum of the coordinates is 1. That eigenvector is (0.1202, 0.1202, 0.7595). Thus a

patient will, in the long run, spend about 12% of the time in each of categories 1

and 2 and about 76% of the time in category 3.

Exercise 32–8. Find the left eigenvectors corresponding to the eigenvalue 1 of the

transition matrix for the gambler’s ruin chain.

Example 32–5. Consider the Markov chain with transition matrix P =

0 1

1 0

.

This chain will be called the oscillating chain. The left eigenvector of P corre-

sponding to the eigenvalue 1 is ( 1/ 2 1/ 2 ). If the chain starts in one of the states,

there is clearly no limiting distribution.

Exercise 32–9. Show that this last chain does not have a limiting distribution.

The oscillating chain example shows that a Markov chain need not have a

limiting distribution. Even so, this chain does spend half the time in each state, so

the entries in the left eigenvector do have an intuitive interpretation as properties of

the chain. To explore this possiblilty further, some terminology is introduced. Let

P be the transition probability matrix of a Markov chain X with stationary transition

probabilities. A vector π = ( π

0

, π

1

, . . . ) is said to be a stationary distribution for

the chain X if

(1) π

i

≥ 0 for all i,

(2)

¸

∞

i=0

π

i

= 1, and

(3) πP = π.

If a limiting distribution for the chain X exists then that limiting distribution will

be a stationary distribution. In the example above, ( 1/ 2 1/ 2 ) is a stationary

distribution even in though the chain has no limiting distribution.

From a relative frequency viewpoint a stationary distribution should arise as the

limit of the occupation times: π

i

= lim

L→∞

1

L+1

¸

L

n=0

1

{i}

(X

n

). Indeed, this is the case

even in the rather poorly behaved example above. It will be easier to differentiate

the possible cases that can arise by studying the states of the chain more closely.

For each state i deﬁne the randomvariable N

i

=

¸

∞

n=0

1

{i}

(X

n

). Clearly N

i

counts

the total number of visits of the Markov chain X to the state i. It is possible that

N

i

= ∞. A state i for which P[N

i

= ∞| X

0

= i] = 1 is a state which is sure to be

§32: Discrete Time Markov Chains 132

revisited inﬁnitely many times. Such a state is said to be recurrent. Anon-recurrent

state, that is, a state i for which P[N

i

= ∞| X

0

= i] < 1 is said to be transient. It is a

rather amazing fact that for a transient state i, P[N

i

= ∞| X

0

= i] = 0. Thus for each

state i the random variable N

i

is either always inﬁnite or never inﬁnite.

Exercise 32–10. Show that if i is a transient state then N

i

is a geometric random

variable.

Checking each state to see whether that state is transient or recurrent is clearly

a difﬁcult task with only the tools available now. Some other useful notions can

greatly simplify the job.

One key notion is that of accessibility. The state j is said to be accessible from

the state i if there is a positive probability that the chain can start in state i and reach

state j. Technically, the requirement for j to be accessible from i is that P

n

i,j

> 0 for

some n ≥ 0. Two states i and j are said to communicate, denoted i↔j, if each is

accessible from the other.

Example 32–6. Consider the Markov chain X in which X

n

denotes the outcome of

the nth toss of a fair coin in which 1 corresponds to a head and 0 to a tail. Clearly

0↔1.

Example 32–7. In the gambler’s ruin problem it is intuitively clear that the states

0 and a are accessible from any other state but do not communicate with any state

except themselves. Such states are absorbing. The other states all communicate

with each other.

Exercise 32–11. Prove that the intuition of the preceding example is correct.

Example 32–8. In the nursing home example, all states communicate with each

other.

It is simple to check that the communication relation between states is an equiv-

alence relation, and therefore partitions the state space of the chain into equivalence

classes. If the chain has but a single equivalence class the chain is said to be

irreducible.

It is an important fact that if i↔j then i is recurrent if and only if j is recurrent.

(Said brieﬂy, recurrence is a class property.)

Exercise 32–12. For the coin tossing chain, is the state 1 recurrent?

Exercise 32–13. What are the recurrent states for the gambler’s ruin chain?

Because of the possible existence of transient states, the discussion of the

limiting behavior of the chain is a bit complicated. Denote by f

i,j

the probability that

§32: Discrete Time Markov Chains 133

the chain ever enters state j given that the chain is currently in state i. Denote by µ

i,i

the expected number of time steps between visits to state i. (For a transient state,

µ

i,i

= ∞ and it is possible for µ

i,i

= ∞ even for a recurrent state.) With this notation,

the central result in the theory of Markov chains can be stated. For any two states i

and j, given that the chain begins at time zero in state i,

lim

L→∞

1

L + 1

L

¸

n=0

1

{j}

(X

n

) = f

i,j

/ µ

j,j

.

This is the result that was anticipated based on previous examples.

This result can be used to provide an interpretation of the entries in a stationary

distribution. To do this, take expectations of both sides of the above limit to obtain

lim

L→∞

1

L + 1

L

¸

n=0

P

n

i,j

= f

i,j

/ µ

j,j

. This can be expressed in matrix terms by deﬁning S to be

the matrix with entries f

i,j

/ µ

j,j

. The foregoing statement is then lim

L→∞

1

L + 1

L

¸

n=0

P

n

= S.

If π is a stationary distribution, πP = π, and by multiplication of the limiting

statement by π, πS = π too. So every stationary distribution is in the row space of

the matrix S.

Now consider the case in all states i and j are recurrent and communicate. Then

f

i,j

= 1 for all i and j, and all of the rows of S are the same. Thus there is only one

stationary distribution and the value π

j

= 1/ µ

j,j

, which is the expected fraction of the

time the chain spends in state j. Since there is only one stationary distribution in this

case, there is also at most one limiting distribution in this case. The oscillating chain

example shows that the stationary distribution need not be a limiting distribution.

If some of the states are transient, the rows of S may not all be the same. This

is in accord with the behavior of the gambler’s ruin chain which shows that the

limiting behavior may depend on the initial state.

Exercise 32–14. Show that if a Markov chain is irreducible and every state is

transient then there is no stationary distribution for the chain.

Even for an irreducible and recurrent chain a limiting distribution need not exist

for another reason. If µ

i,i

= ∞ there is no stationary distribution. Such a chain is

null recurrent. If µ

i,i

< ∞, then as above there is one stationary distribution. In

this case the chain is strongly ergodic (or positive recurrent).

To identify the cases in which the chain has a limiting distribution one additional

idea is needed. The period of the state i, denoted by d(i), is the greatest common

divisor of the set {n ≥ 1 : P

n

i,i

> 0}. If this set is empty, the period is deﬁned to be

0. If d(i) = 1 for all states of the chain then the chain is said to be aperiodic. If a

§32: Discrete Time Markov Chains 134

Markov chain is aperiodic, a limiting distribution exists (given that the chain starts

in a particular state) and lim

n→∞

P[X

n

= j| X

0

= i] = f

i,j

/ µ

j,j

. If all states communicate

and are recurrent, this limiting distribution will not depend on the initial state.

Exercise 32–15. What is the period of each state in the nursing home chain? The

gambler’s ruin chain? The oscillating chain?

Example 32–9. For the gambler’s ruin chain ( c, 0, . . . , 0, 1 − c ) is a stationary

distribution for any 0 ≤ c ≤ 1. One easily sees that only the two recurrent classes

contribute non-zero probabilities here. A formula for f

i,0

was found earlier.

§32: Discrete Time Markov Chains 135

Problems

Problem 32–1. Show that P

n

i,i

= 0 if d(i) does not divide n. Show that if i↔j then

d(i) = d(j). (Said brieﬂy, period is a class property.)

Problem 32–2. Suppose the chain has only ﬁnitely many states all of which com-

municate with each other. Are any of the states transient?

Problem 32–3. Suppose N

i,j

is the total number of visits of the chain to state j given

that the chain begins in state i. Show that for i ≠ j, E[N

i,j

] =

∞

¸

k=0

E[N

k,j

] P

i,k

. What

happens if i = j?

Problem 32–4. Suppose the chain has both transient and recurrent states. Relabel

the states so that the transient states are listed ﬁrst. Partition the transition matrix in

to blocks P =

P

T

Q

0 P

R

**. Explain why the lower left block is a zero matrix. Show
**

that the T × T matrix of expectations E[N

i,j

] as i and j range over the transient states

is (I − P

T

)

−1

.

Problem 32–5. Show that if i ≠ j and both are transient, f

i,j

= E[N

i,i

]/ E[N

j,j

]. What

happens if i = j?

Problem 32–6. In addition to the 3 categories of expenses in the nursing home ex-

ample, consider also the possibilities of withdrawal from the home and death. Sup-

pose the corresponding transition matrix is P =

¸

¸

¸

¸

¸

¸

0.8 0.05 0.01 0.09 0.05

0.5 0.45 0.04 0.0 0.01

0.05 0.15 0.70 0.0 0.10

0.0 0.0 0.0 1.0 0.0

0.0 0.0 0.0 0.0 1.0

¸

**where the states are the 3 expense categories in order followed by withdrawal and
**

death. Find the stationary distribution(s) of the chain. Which states communicate,

which states are transient, and what are the absorption probabilities given the initial

state?

Problem 32–7. An auto insurance company classiﬁes insureds in 2 classes: (1)

preferred, and (2) standard. Preferred customers have an expected loss of $400 in

any one year, while standard customers have an expected loss of $900 in any one

year. A driver who is classiﬁed as preferred this year has an 85% chance of being

classiﬁed as preferred next year; a driver classiﬁed as standard this year has a 40%

chance of being classiﬁed as standard next year. Losses are paid at the end of each

year and i = 5%. What is the net single premium for a 3 year term policy for an

entering standard driver?

§32: Discrete Time Markov Chains 136

Solutions to Problems

Problem 32–1. Since i↔j there are integers a and b so that P

a

i,j

> 0 and

P

b

j,i

> 0. As shown earlier, P

a+n+b

j,j

≥ P

b

j,i

P

n

i,i

P

a

i,j

. If P

n

i,i

> 0 this inequality shows

that P

a+b+n

j,j

> 0 too and therefore d(j) divides a + b + n. But since P

2n

i,i

≥

P

n

i,i

2

a

similar inequality shows that d(j) divides a + b + 2n as well. Hence d(j) divides

a + b + 2n − (a + b + n) = n, and so d(j) ≤ d(i). Interchanging the roles of i and j

shows that d(i) ≤ d(j).

Problem 32–2. No. Since all states communicate, either all are transient or

all are recurrent. Since there are only ﬁnitely many states they can not all be

transient. Hence all states are recurrent.

Problem 32–3. The formula follows by conditioning on the ﬁrst step leaving

state i. When i = j the formula is E[N

i,i

] = 1 +

∞

¸

k=0

E[N

k,i

] P

i,k

, by the same

argument.

Problem32–4. It is not possible to go froma recurrent state to a transient state.

Express the equations of the previous problem in matrix form and solve.

Problem 32–5. f

i,i

= (E[N

i,i

] − 1)/ E[N

i,i

].

Problem32–6. The 3 expense category states communicate with each other and

are transient. The other 2 states are recurrent and absorbing. The probabilities

f

i,j

satisfy f

0,4

= 0.8f

0,4

+ 0.05f

1,4

+ 0.05f

2,4

+ 0.09 and 2 other similar equations,

from which f

0,4

= 0.382, f

1,4

= 0.409, and f

2,4

= 0.601.

Problem 32–7. From the given information the transition matrix is P =

.85 .15

.6 .4

**. The two year transition probabilities for an entering standard
**

driver are found from the second row of P

2

to be ( .75 .25 ). The premium is

900v + (.6 × 400 + .4 × 900)v

2

+ (.75 × 400 + .25 × 900)v

3

= 1854.90.

§32: Discrete Time Markov Chains 137

Solutions to Exercises

Exercise 32–1. Because of the independence both of the conditional probabil-

ities in the deﬁnition are equal to the unconditional probability P[a < X

t

n+1

< b].

Exercise 32–2. This is just the deﬁnition of conditional probability together

with the use of the Markov property to simplify each of the conditional proba-

bilities.

Exercise 32–3. Write P[X

4

= 3, X

3

≠ 3| X

2

= 3, X

1

≠ 3, X

0

= 3] =

¸

j≠3

P[X

4

=

3, X

3

= j| X

2

= 3, X

1

≠ 3, X

0

= 3] =

¸

j≠3

P[X

4

= 3, X

3

= j, X

2

= 3, X

1

≠ 3, X

0

=

3]/ P[X

2

= 3, X

1

≠ 3, X

0

= 3] =

¸

j≠3

¸

k≠3

P[X

4

= 3, X

3

= j, X

2

= 3, X

1

= k, X

0

=

3]/ P[X

2

= 3, X

1

≠ 3, X

0

= 3] =

¸

j≠3

¸

k≠3

P[X

4

= 3, X

3

= j| X

2

= 3, X

1

= k, X

0

=

3]P[X

2

= 3, X

1

= k, X

0

= 3]/ P[X

2

= 3, X

1

≠ 3, X

0

= 3] =

¸

j≠3

¸

k≠3

P[X

4

=

3, X

3

= j| X

2

= 3]P[X

2

= 3, X

1

= k, X

0

= 3]/ P[X

2

= 3, X

1

≠ 3, X

0

= 3] =

¸

j≠3

P[X

4

= 3, X

3

= j| X

2

= 3] =

¸

j≠3

P[X

2

= 3, X

1

= j| X

0

= 3] = P[X

2

=

3, X

1

≠ 3| X

0

= 3], where the Markov property and stationarity have been used.

Exercise 32–4.

¸

j

P

i,j

=

¸

j

P[X

1

= j| X

0

= i] = P[X

1

∈ R| X

0

= i] = 1.

Exercise 32–5.

¸

¸

¸

¸

¸

¸

¸

1 0 0 0 0 . . . 0

q 0 p 0 0 . . . 0

0 q 0 p 0 . . . 0

0 0 q 0 p . . . 0

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

0 0 0 0 0 . . . 1

¸

.

Exercise 32–6. Use induction on n. The case n = 1 is true from the deﬁnition

of stationarity. For the induction step assume the result holds when n = k. Then

P[X

k+1+m

= j| X

m

= i] =

¸

b

P[X

k+1+m

= j, X

k+m

= b| X

m

= i] =

¸

b

P[X

k+1+m

=

j| X

k+m

= b]P[X

k+m

= b| X

m

= i] =

¸

b

P[X

k+1

= j| X

k

= b]P[X

k

= b| X

0

= i] =

P[X

k+1

= j| X

0

= i], as desired.

Exercise 32–7. P[X

n

= k] =

¸

i

P[X

n

= k| X

0

= i]P[X

0

= i] =

¸

i

P

n

i,k

P[X

0

= i]

which agrees with the matrix multiplication.

Exercise 32–8. Matrix multiplication shows that the left eigenvector condition

implies that the left eigenvector x = (x

0

, . . . , x

a

) has coordinates that satisfy

x

0

+ qx

1

= x

0

, qx

2

= x

1

, px

k−1

+ qx

k+1

= x

k

for 2 ≤ k ≤ a − 2, px

a−2

= x

a−1

and

px

a−1

+ x

a

= x

a

. From these equations it follows that only x

0

and x

a

can be

non-zero, and that these two values can be arbitrary. Hence all left eigenvectors

corresponding to the eigenvalue 1 are of the form (c, 0, 0, . . . , 0, 1 − c) for some

0 ≤ c ≤ 1.

Exercise 32–9. P[X

n

= 1] is 0 or 1 depending on whether n is odd or even, so

this probability has no limit.

Exercise 32–10. Let p = P[N

i

< ∞| X

0

= i]. Then P[N

i

= k| X

0

= i] = p

k

(1 − p)

for k = 0, 1, . . ..

Exercise 32–11. If the current fortune is i, and i is not 0 or a, then it is possible

§32: Discrete Time Markov Chains 138

to obtain fortune j is | j − i| plays of the game by having | j − i| wins (or losses)

in a row.

Exercise 32–12. Yes, 1 is recurrent since state 1 is sure to be visited inﬁnitely

often.

Exercise 32–13. The only recurrent states are 0 and a.

Exercise 32–14. Under these assumptions, f

i,j

= 0 so the limit above is always

zero. Hence there is no stationary distribution.

Exercise 32–15. For the nursing home chain and the gambler’s ruin chain

each state has period 1. For the oscillating chain each state has period 2. This

explains why the oscillating chain does not have a limiting distribution.

§33. Continuous Time Markov Chains

The next step in the study of Markov chains retains a discrete state space but

allows time to vary continuously.

A continuous time Markov chain is a stochastic process {X

t

: t ≥ 0} with state

space {0, 1, 2, . . .} which satisﬁes the Markov property: whenever n > 0 and 0 ≤

t

1

< t

2

< . . . < t

n

< t then P[X

t

= x

t

| X

t

n

= x

t

n

, . . . , X

t

1

= x

t

1

] = P[X

t

= x

t

| X

t

n

= x

t

n

].

As in the discrete time case the discussion here assumes that the Markov chain has

stationary transition probabilities, that is, P[X

t+s

= i| X

t

= j] does not depend on t.

A discrete time chain always spent one time unit in each state before the next

transition. For a continuous time chain the time spent in each state before a transition

will be random. Intuitively, this is the only difference between discrete time and

continuous time Markov chains. The bulk of the work consists of verifying that this

intuition is indeed correct and in identifying the probabilistic properties of the times

between transitions for the chain.

As for discrete time chains the transition probabilities P

i,j

(t) = P[X

t

= j| X

0

= i]

and the associated transition probability matrix P(t) = [P

i,j

(t)] play an important

role. By convention, P(0) = I, the identity matrix.

The ﬁrst result parallels a result for discrete time chains.

Transition Semi-group Property. The transition matrices {P(t) : t ≥ 0} form a

semi-group under matrix multiplication, that is, P(s + t) = P(s)P(t) for all s, t ≥ 0.

Exercise 33–1. Prove the theorem.

For a discrete time chain, the one step transition probability matrix determined

all of the interesting properties of the chain. For continuous time chains, a similar

role is played by the matrix P

′

(0), which is called the inﬁnitesimal generator of

the Markov chain (or of the transition semi-group {P(t) : t ≥ 0}). Since typically

the inﬁnitesimal generator is speciﬁed during model building the central question is

how the inﬁnitesimal generator determines properties of the chain.

Example 33–1. The Poisson process provides a typical example of the way in which

the transition probabilities are speciﬁed for continuous time Markov chains. As a

speciﬁc example let X

t

denote the number of calls that have arrived at a telephone

switchboard by time t. In a very short time interval essentially either 0 or 1 call can

arrive. Denote by λ > 0 the average rate at which calls arrive. A model for the

behavior of the chain X

t

can be expressed by saying that for small h > 0

(1) P[X

t+h

− X

t

= 1| X

t

= i] ≈ λh,

(2) P[X

t+h

− X

t

= 0| X

t

= i] ≈ 1 − λh, and

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§33: Continuous Time Markov Chains 140

(3) P[X

t+h

− X

t

≥ 2| X

t

= i] ≈ 0

or equivalently in terms of the transition probabilities

(1) P

i,i

(h) ≈ 1 − λh,

(2) P

i,i+1

(h) ≈ λh, and

(3) P

i,j

(h) ≈ 0 if j ≠ i, i + 1.

Since P(0) = I these assumptions show that the derivative of P(t) exists at t = 0 and

is given by

P

′

(0) =

¸

¸

¸

¸

−λ λ 0 0 . . .

0 −λ λ 0 . . .

0 0 −λ λ . . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

¸

.

In the case of the Poisson process, the row sums of the generator are zero. An

inﬁnitesimal generator for which the row sums are all zero is conservative. Also,

the off-diagonal entries of the generator are non-negative, while the diagonal entries

of the generator are non-positive.

Exercise 33–2. Show that if the chain has only ﬁnitely many states then the in-

ﬁnitesimal generator must be conservative.

Exercise 33–3. How could a generator fail to be conservative?

Exercise 33–4. Are the off-diagonal entries of an inﬁnitesimal generator always

non-negative? Are the diagonal entries of an inﬁnitesimal generator always non-

positive?

Suppose P(t) is a Markov semi-group which is continuous at 0. Then using

results from real analysis, the inﬁnitesimal generator A = P

′

(0) exists. Using the

semi-group property gives

P

′

(t) = lim

h→0

P(t + h) − P(t)

h

= lim

h→0

P(h)P(t) − P(t)

h

=

lim

h→0

P(h) − I

h

P(t)

= AP(t).

This equation is known as Kolmogorov’s backward equation. The same argument

leads also to Komogorov’s forward equation P

′

(t) = P(t) A.

Exercise 33–5. Why aren’t these derivations rigorous?

§33: Continuous Time Markov Chains 141

These two equations, especially the forward equation, are very useful in appli-

cations. Unfortunately there are no simple general conditions which guarantee that

the forward equation holds.

In the case of chains with only a ﬁnite number of states the theory is very simple:

the transition semi-group is uniquely determined by the inﬁnitesimal generator

(which must be conservative) and both the forward and backward equations hold.

Also, the semi-group can be obtained fromthe generator by the formula P(t) = e

tP

′

(0)

.

In this formula the exponential of a matrix is computed from the Maclaurin series

for the exponential function.

The case in which the chain has inﬁnitely many states is more complex. The

main result in this case is just stated here.

Theorem. Suppose the inﬁnitesimal generator is conservative. If P(t) is the unique

solution of either the forward or the backward equation and if

¸

∞

j=0

P

i,j

(t) = 1 for all

i then P(t) is the unique solution to both the forward and the backward equation and

is the unique transition semi-group with the given generator.

Example 33–2. Assume for the moment that the forward equation holds for the

Poisson process. (Later this will be shown to be true.) Recall that the inﬁnitesimal

generator in this case is

A = P

′

(0) =

¸

¸

¸

¸

−λ λ 0 0 . . .

0 −λ λ 0 . . .

0 0 −λ λ . . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

¸

.

Translating the matrix form of the forward equation P

′

(t) = P(t)A into statements

about the individual transition probabilities gives

P

′

i,j

(t) = −λP

i,j

(t) + λP

i,j−1

(t).

To compute M(t) = E[X

t

| X

0

= i] use the deﬁnition of expectation and this equation

to obtain

M

′

(t) =

∞

¸

j=0

jP

′

i,j

(t)

=

∞

¸

j=0

j

−λP

i,j

(t) + λP

i,j−1

(t)

= −λ

∞

¸

j=0

jP

i,j

(t) + λ

∞

¸

j=0

(j − 1)P

i,j−1

(t) + λ

= λ.

So M(t) = λt + i. Thus the forward equation can be used to ﬁnd the conditional

expectation without ﬁrst ﬁnding the transition matrices.

§33: Continuous Time Markov Chains 142

Exercise 33–6. Try the same computation using the backward equations.

Amethod will nowbe developed to directly interpret the entries of the generator

in terms of the probabilistic behavior of the chain.

Deﬁne T

0

= 0 and inductively set T

n

= inf{t ≥ T

n−1

: X

t

≠ X

T

n−1

}. The T’s are

the times of changes of state for the chain.

To study the behavior of the chain let t > 0 and i ≠ j be states. Then

P[T

1

> t, X

T

1

= j| X

0

= i]

= lim

n→∞

∞

¸

l=0

P[X

k/ 2

n = i, 1 ≤ k ≤ [2

n

t] + l, X

([2

n

t]+l+1)/ 2

n = j| X

0

= i]

= lim

n→∞

∞

¸

l=0

P

i,i

(1/ 2

n

)

[2

n

t+l]

P

i,j

(1/ 2

n

)

= lim

n→∞

P

i,j

(1/ 2

n

)

1/ 2

n

P

i,i

(1/ 2

n

)

2

n

[2

n

t]/ 2

n

∞

¸

l=0

1

2

n

P

i,i

(1/ 2

n

)

l

= a

i,j

e

a

i,i

t

lim

n→∞

1/ 2

n

1 − P

i,i

(1/ 2

n

)

=

−a

i,j

a

i,i

e

a

i,i

t

.

(Recall that a

i,i

≤ 0.) Hence P[X

T

1

= j| X

0

= i] =

−a

i,j

a

i,i

and P[T

1

> t | X

0

= i] = e

a

i,i

t

for

t > 0.

Exercise 33–7. Fill in the details for all of these calculations.

This computation means that the time spent in state i until a transition occurs has

an exponential distribution with mean −1/ a

i,i

and that the probability upon leaving

state i of entering state j is −a

i,j

/ a

i,i

.

Exercise 33–8. Extend the argument above to show that

P[T

l

> t

l

, X

T

l

= i

l

, 0 ≤ l ≤ n| X

0

= i

0

] =

n−1

¸

l=0

−a

i

l−1

,i

l

a

i

l−1

,i

l−1

e

a

i

l−1

,i

l−1

t

l

.

The exercise justiﬁes the following view of continuous time Markov chains.

Begin at time 0 in state i

0

. Stay in state i

0

for a random time T

1

, where T

1

has the

exponential distribution with mean −1/ a

i

0

,i

0

. Then move to state i

1

with probability

−a

i

0

,i

1

/ a

i

0

,i

0

. Remain in state i

1

an exponentially distributed time, etc.

If interest is only in the states that are visited by the chain and not in the time

spent in each state one may as well study the embedded discrete time chain with

§33: Continuous Time Markov Chains 143

transition matrix P = [−a

i,j

/ a

i,i

(1 − δ

i,j

)]. Note that if a

i,i

= 0 the corresponding

transition probability in the embedded chain is P

i,i

= 1 since state i is obviously an

absorbing state for the continuous time chain.

Example 33–3. For the Poisson process the embedded chain has transition matrix

P =

¸

¸

¸

¸

0 1 0 0 0 . . .

0 0 1 0 0 . . .

0 0 0 1 0 . . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

¸

.

Hence all states are transient and there is no stationary distribution.

The relationship between the inﬁnitesimal generator and the chain itself has been

made rather precise. The question of the behavior of the chain is now considered.

Suppose I = ( P[X

0

= 0] P[X

0

= 1] . . . ) is the initial distribution of the chain.

If a limiting distribution π for the chain exists then

π = lim

t→∞

I P(t) = lim

t→∞

I P(t + s) = lim

t→∞

I P(t)P(s) = πP(s)

for all s ≥ 0.

Exercise 33–9. Verify that this computation is correct.

Once again the notion of a stationary distribution will play an important role.

In the continuous time context π = ( π

0

π

1

. . . ) is said to be a stationary

distribution if

(1) π

i

≥ 0 for all i,

(2)

¸

∞

i=0

π

i

= 1, and

(3) πP(s) = π for all s ≥ 0.

Because of the close relationship between the inﬁnitesimal generator and the

transition semi-group it should be possible to ﬁnd the stationary distribution of the

chain using only the generator.

Theorem. Suppose the inﬁnitesimal generator A of the chain is conservative and

that π

i

≥ 0 for all i and

¸

∞

i=0

π

i

= 1. Then π is a stationary distribution if and only

if πA = 0.

There is no notion of period in the continuous time setting. Thus the distinction

between a stationary distribution and a limitingdistribution depends only on whether

there are transient states, and the mean recurrence time for the recurrent states. This

makes the continuous time case somewhat simpler than the discrete time case.

§33: Continuous Time Markov Chains 144

Some additional examples of continuous time chains will now be given. One of

the main features is the way in which the behavior of the chain is determined from

the inﬁnitesimal generator.

Example 33–4. Look once again at the Poisson process. Here the forward equation

is P

′i,j

(t) = −λP

i,j

(t) + λP

i,j−1

(t). This can be rewritten using an integrating factor

as

d

dt

e

λt

P

i,j

(t)

= λe

λt

P

i,j−1

(t). Thus P

0,0

(t) = e

−λt

and by induction P

0,n

(t) =

(λt)

n

e

−λt

/ n!. Similarly, P

i,0

(t) = 0 if i > 0 and by induction P

i,j

(t) = (λt)

j−i

e

−λt

/ (j −

i)!1

[0,∞)

(j − i). This solution is unique and

¸

∞

j=0

P

i,j

(t) = 1 for all i. From the general

theory P

i,j

(t) is therefore the unique solution to the backward equation as well and

is also the unique transition semi-group with this inﬁnitesimal generator.

Exercise 33–10. Fill in the details of the induction arguments above.

Example 33–5. The next example is a pure birth process. This is a variant of the

Poisson process in which the probability of additional calls depends on the number

of calls received. Speciﬁcally assume that

(1) P[X

t+h

− X

t

= 1| X

t

= k] ≈ λ

k

h,

(2) P[X

t+h

− X

t

= 0| X

t

= k] ≈ 1 − λ

k

h, and

(3) P[X

t+h

− X

t

≥ 2| X

t

= k] ≈ 0.

As before this leads to the generator

A =

¸

¸

¸

¸

−λ

0

λ

0

0 0 . . .

0 −λ

1

λ

1

0 . . .

0 0 −λ

2

λ

2

. . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

¸

.

Using the forward equation and the integrating factor e

λ

j

t

yields

P

i,j

(t) = δ

i,j

e

−λ

j

t

+ λ

j−1

e

−λ

j

t

t

0

e

λ

j

t

P

i,j−1

(s) ds.

This shows inductively that P

i,j

(t) ≥ 0 and that the solution is unique. The remaining

question is whether or not

¸

∞

j=0

P

i,j

(t) = 1 for all i. Fix i and deﬁne S

n

(t) =

¸

n

j=0

P

i,j

(t).

Using the forward equation gives S

′ n

(t) = −λ

n

P

i,n

(t) and so 1 − S

n

(t) =

t

0

−S

′ n

(s) ds =

λ

n

t

0

P

i,n

(s) ds. Now from the deﬁnition of S

n

(t), 1 −

¸

∞

j=0

P

i,j

(t) ≤ 1 − S

n

(t) ≤ 1 so

1

λ

n

¸

1 −

∞

¸

j=0

P

i,j

(t)

¸

≤

t

0

P

i,n

(s) ds ≤

1

λ

n

.

Summing on n gives

∞

¸

n=0

1

λ

n

¸

1 −

∞

¸

j=0

P

i,j

(t)

¸

≤

t

0

∞

¸

n=0

P

i,n

(s) ds ≤

∞

¸

n=0

1

λ

n

.

§33: Continuous Time Markov Chains 145

The right hand part of this inequality shows that if

¸

∞

n=0

1

λ

n

< ∞ then

¸

∞

j=0

P

i,j

(t) can

not be 1 for all t, while the left hand part of this inequality shows that if

¸

∞

n=0

1

λ

n

= ∞

then

¸

∞

j=0

P

i,j

(t) = 1 for all t ≥ 0. Thus the pure birth process exists if and only if

¸

∞

n=0

1

λ

n

= ∞. Note that this condition is obviously satisﬁed for the Poisson process.

Exercise 33–11. Explain intuitively what goes wrong with the pure birth process

when

¸

∞

n=0

1

λ

n

< ∞.

Example 33–6. The Yule process is a pure birth process for which λ

n

= nβ, that

is, the birth rate is proportional to the number present. This process clearly meets

the criteria for existence which was established in the previous example. If M

i

(t) =

E[X

t

| X

0

= i] then the forward equation can be used to show that M

′ i

(t) = βM

i

(t).

Hence M

i

(t) = ie

βt

.

Exercise 33–12. Fill in the details of how the forward equation is used in this

computation. What is the conditional variance of X

t

?

Example 33–7. The ﬁnal example is the birth and death process. Suppose X

t

is the size of a population at time t. Then in a short time period there will be

(essentially) a single birth or a single death or neither. Formally the model is

(1) P[X

t+h

− X

t

= 1| X

t

= k] ≈ λ

k

h,

(2) P[X

t+h

− X

t

= 0| X

t

= k] ≈ 1 − (λ

k

+ µ

k

)h,

(3) P[X

t+h

− X

t

= −1| X

t

= k] ≈ µ

k

h, and

(4) P[| X

t+h

− X

t

| ≥ 2| X

t

= k] ≈ 0.

Here λ

k

is birth rate and µ

k

is the death rate when the population size is k.

As before this leads to the generator

A =

¸

¸

¸

¸

¸

−λ

0

λ

0

0 0 . . .

µ

1

−µ

1

− λ

1

λ

1

0 . . .

0 µ

2

−µ

2

− λ

2

λ

2

. . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

¸

.

As in the case of the pure birth process the condition

¸

∞

n=0

1

λ

n

= ∞ will guarantee

that the process is well deﬁned.

Example 33–8. For the birth and death process the embedded chain has transition

matrix

P =

¸

¸

¸

0 1 0 0 . . .

µ

1

µ

1

+λ

1

0

λ

1

µ

1

+λ

1

0 . . .

.

.

.

.

.

.

.

.

.

.

.

.

.

.

.

¸

if λ

0

> 0 so this chain behaves in a manner similar to the gambler’s ruin chain,

except that when the gambler reaches fortune 0, she begins again with 1 at the

§33: Continuous Time Markov Chains 146

next play. If λ

0

= 0 and 0 < λ

i

for i > 0 then ﬁnding the absorption probability

for the birth and death chain is exactly the same problem as ﬁnding the absorption

probability at 0 for the gambler’s ruin chain.

Exercise 33–13. How does the ﬁrst row of P differ from that shown when λ

0

= 0?

For a continuous time Markov chain with conservative generator Athe stationary

distributions π are the solutions of πA = 0. Also −1/ a

i,i

is the mean of the

exponentially distributed time that the chain spends in state i prior to a transition

and −a

i,j

/ a

i,i

is the probability that when a transition occurs from state i the chain

will move to state j ≠i. The probability of a transition from state i to itself is 0 unless

a

i,i

= 0 in which case i is an absorbing state.

One consequence of the Markov property is that the time spent in each state is

exponential. In many models this is not realistic because the exponential distribution

is memoryless.

§33: Continuous Time Markov Chains 147

Problems

Problem 33–1. Consider a continuous time version of the nursing home chain in

which the states are (1) minimal care, (2) some skilled assistance required, (3)

continuous skilled assistance required, and (4) dead. Suppose the inﬁnitesimal

generator A of the chain has entries a

1,2

= 0.12, a

1,3

= 0.05, a

1,4

= 0.08, a

2,1

= 0.05,

a

2,3

= 0.07, a

2,4

= 0.12, a

3,4

= 0.20 and all other non-diagonal entries are zero.

What is the matrix A? What are the stationary distributions of the chain? What is

the expected time until absorption as a function of the initial state?

Problem 33–2. For the birth-death process show that a stationary distribution must

satisfy π

j+1

=

λ

j

µ

j+1

π

j

for j ≥ 0 and that π

0

=

1

1 +

¸

∞

n=1

λ

0

⋅⋅⋅λ

n−1

µ

1

⋅⋅⋅µ

n

. Thus the chain has a

limiting distribution as long as

∞

¸

n=1

λ

0

⋅ ⋅ ⋅ λ

n−1

µ

1

⋅ ⋅ ⋅ µ

n

< ∞. What happens if λ

0

= 0?

Problem 33–3. Find the conditional moment generating function E[e

vX

t

| X

0

= i] for

a Poisson process.

Problem 33–4. A continuing care retirement community is modeled by a contin-

uous time Markov chain with ﬁve classiﬁcations for the status of a patient: (1)

Individual Living Unit, (2) Skilled Nursing Facility (Temporary), (3) Skilled Nurs-

ing Facility (Permanent), (4) Dead, and (5) Withdrawn from Care. The non-zero

off-diagonal entries of the generator are

¸

¸

¸

¸

¸

¸

0.3 0.1 0.1 0.2

0.5 0.3 0.3

0.4

¸

.

If a patient is currently in the Individual Living Unit, what is the probability of more

than one visit to the Skilled Nursing Facility? What is the expected duration of a

visit to the Skilled Nursing Facility?

Problem 33–5. An HMO has 3 classes of patients: (1) healthy, (2) moderately

imparied, and (3) severely impaired. The status of a patient is modeled by a

continuous time Markov chain with generator whose non-zero off-diagonal entries

are

¸

¸

1/ 2 1/ 2

1/ 3 1/ 3

1/ 4 1/ 4

¸

**. What is the stationary distribution of this chain? If a
**

healthy person arrives today, about how long will it be until this person becomes

severely impaired?

§33: Continuous Time Markov Chains 148

Solutions to Problems

Problem 33–1. A =

¸

¸

−0.25 0.12 0.05 0.08

0.05 −0.24 0.07 0.12

0 0 −0.20 0.20

0 0 0 0

¸

**, from which the sta-
**

tionary distribution is easily found to be ( 0 0 0 1 ). If e

i

is the expected

time until absorption starting from state i then e

1

= 4 + (.12/ .25)e

2

+ (.05/ .25)e

3

and two other similar equations also hold. Solving gives e

1

= 8.55, e

2

= 7.40,

and e

3

= 5.

Problem 33–2. The relation πA = 0 gives λ

0

π

0

= µ

1

π

1

and λ

i−1

π

i−1

− (µ

i

+

λ

i

)π

i

+ µ

i+1

π

i+1

= 0 for i ≥ 1. Sum this last equality on i from 1 to j and use

telescoping and the ﬁrst equation to simplify. The rest follows by normalization.

Problem 33–3. If g(t) = E[e

vX

t

| X

0

= i] the forward equation shows that

g

′

(t) = −λg(t) + λe

v

g(t). Now solve this differential equation and use the initial

condition g(0) = e

vi

.

Problem 33–4. The probability of no visits to the Skilled Nursing Facility is

3/ 7. The probability of exactly one visit is 3/ 7(6/ 11+(5/ 11)(3/ 7)) +1/ 7. Similar

reasoning applies when computing the expected duration, keeping in mind that

the probabilities are conditional on visiting the Skilled Nursing Facility.

Problem33–5. The stationary distribution is π

1

= 2/ 9, π

2

= 3/ 9, and π

3

= 4/ 9.

If e

1

and e

2

are the expected waiting times until the ﬁrst entry into state 3 given

the starting state, then e

1

= 1/ 2 + 1/ 2(1 + e

2

) and e

2

= 1/ 2(3/ 2) + 1/ 2(3/ 2 + e

1

).

Hence e

1

= 7/ 3 (and e

2

= 8/ 3).

§33: Continuous Time Markov Chains 149

Solutions to Exercises

Exercise 33–1. Hint: Look at the proof of the parallel result for discrete time

chains.

Exercise 33–2. Since there are only ﬁnitely many states, the derivative of

the sum is the sum of the derivatives, so the equation

¸

j

P

i,j

(t) = 1 can be

differentiated to show that the generator is conservative.

Exercise 33–3. If there are inﬁnitely many states, the derivative and the sum

in the previous exercise might not be interchangeable, and this could make the

generator not be conservative.

Exercise 33–4. Yes in both cases. Look at the sign of the difference quotients

used to compute the entries of the generator.

Exercise 33–5. If there are inﬁnitely many states there could be problems since

these ‘derivations’ involve passing a limit through the inﬁnite sum occurring in

the matrix multiplication.

Exercise 33–6. The backward equation gives P

′i,j

(t) = λP

i−1,j

(t)−λP

i,j

(t) which

does not produce a differential equation for M(t).

Exercise 33–10. The inductive step is

d

dt

(e

λt

P

i,j

(t)) = λ

j

t

j−1

/ (j −1)! fromwhich

integration gives e

λt

P

i,j

(t) = (λt)

j

/ j!, as desired.

Exercise 33–11. If the process is currently in state k the expected waiting time

until a birth is 1/ λ

k

. If the sum is ﬁnite then there is a ﬁnite expected waiting

time for inﬁnitely many births to occur, that is, the population size will become

inﬁnite in a ﬁnite amount of time.

Exercise 33–12. The forward equation gives P

′i,j

(t) = βjP

i,j

(t) +β(j −1)P

i,j−1

(t).

Now multiply both sides of this equation by j and sum on j = j − 1 + 1 to obtain

the equation.

Exercise 33–13. When λ

0

= 0 the ﬁrst row is ( 1 0 0 0 . . . ).

§34. Introduction to Brownian Motion

The Brownian motion model originated as a description of the movement of

microscopic particles in liquid. Applications to modeling stock prices and other

situations quickly followed. Brownian motion is stochastic process for which both

time and state space are continuous. To make this transition from earlier studies

a simpler discrete time and state space model will be constructed which mimicks

some of the important features of Brownian motion. A suitable passage to the limit

will enable this simpler model to become a Brownian motion.

The discrete time and state space model will be constructed to model the motion

of a particle in one dimension. Let S

n

denote the position of the particle at time n and

assume S

0

= 0. The change in position of the particle at time i will be denoted D

i

.

The random variables D

1

, D

2

, . . . will be assumed to be independent and identically

distributed random variables each taking the values 1 and −1 with equal probability.

Hence S

n

=

¸

n

i=0

D

i

and {S

n

: n ≥ 0} is nothing more than a random walk model.

A convenient visual and conceptual aid is the path generated by D

1

, . . . , D

n

which consists of the vertices {(j, S

j

) : 0 ≤ j ≤ n} together with the line segments

joining the vertices. Many interesting probabilistic problems can be turned into path

counting problems by using this device.

Proposition. The number of paths joining the origin to a point (n, x) is

n

n+x

2

if

(n + x)/ 2 is a non-negative integer.

proof : Denote by P

n

the number of displacements D

1

, . . . , D

n

which are positive and let N

n

= n−P

n

denote the number of negative displacements. The D’s generate a path from the origin to

(n, x) if and only if n = P

n

+ N

n

and x = P

n

− N

n

. This shows that there are

n

P

n

such paths.

From the two equations P

n

= (n + x)/ 2 and the result follows.

The reﬂection principle is another path inspired device. The notation corre-

sponds to that of the picture below, which also provides the proof.

Reﬂection Principle. The number of paths from A to B which touch or cross the

level T is equal to the number of all paths from A

′

to B.

..................................................................................................................................................................................................................................................................................................................

.............

.............

.............

.............

.............

.............

.............

.............

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

•

S

n

T

n

A

B

A

′

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§34: Introduction to Brownian Motion 151

As an application consider an election in which candidate A wins with a votes

compared to the b < a votes for candidate B. If the ballots are counted one at a time,

what is the probability that candidate A is always ahead in the count?

Ballot Theorem. Suppose n and x are positive integers. The number of paths from

the origin to (n, x) which do not touch the n axis is

x

n

n

n+x

2

.

proof : The number of paths meeting the requirement is the same as the number of paths from

(1, 1) to (n, x) which do not touch or cross the n axis. Now the total number of paths from

(1, 1) to (n, x) is the same as the total number of paths from (0, 0) to (n − 1, x − 1), which is

n−1

n−1+x−1

2

**by the proposition. The number of paths from (1, 1) to (n, x) which touch or cross
**

the n axis can be computed using the reﬂection principle. By the reﬂection principle this

is the same as the number of paths from (1, −1) to (n, x) which in turn is the same as the

number of paths from (0, 0) to (n − 1, x + 1). Again using the proposition gives this number

as

n−1

n+x

2

**. Subtraction gives the desired result.
**

Exercise 34–1. Use the Ballot Theorem to show that the probability that candidate

A is always ahead in the tally is (a − b)/ (a + b). Thus if A wins 70 votes to 30 the

probability that A was always ahead is 0.40.

Some of the basic properties of Brownian motion will now be developed by

passing to the limit, in an appropriate way, in the discrete time and state space

model.

One of the ﬁrst applications of the Brownian motion model was to describe the

displacement of a small particle suspended in ﬂuid. As a ﬁrst approximation to the

behavior of the particle assume that the particle moves only because of collisions

with other particles. Assume also that collisions are as likely to come from the

left as from the right. To simplify even further suppose that each collision causes

a displacement of magnitude ∆ and that the time between collisions is τ. If the

particle begins at the origin at time 0 the position X

t

of the particle at time t is

then ∆S

[t/ τ]

in the notation above. The idea is to see what happens as both ∆

and τ go to 0. This passage to the limit makes the approximate model become

more realistic. Note that E[X

t

] = 0 and Var(X

t

) ≈ ∆

2

t/ τ. To avoid disaster some

relationship between ∆

2

and τ must be maintained in the passage to the limit. The

assumption made here is that ∆

2

/ τ = σ

2

for some σ > 0. After passing to the limit

under this assumption the Central Limit Theorem shows that X

t

d

= N(0, σ

2

t). The

process X

t

obtained in this way is also a process with independent increments.

This means that if t

1

< t

2

< . . . < t

n

then the randomvariables X

t

n

−X

t

n−1

, . . . , X

t

2

−X

t

1

are independent. The increments are also stationary which means that for t > s

X

t

− X

s

d

= X

t−s

. Intuition further suggests that the sample paths of the process X

t

should be continuous functions. This means that for each point ω in the underlying

probability space the function of t deﬁned by t → X

t

(ω) should be continuous. This

intuitive motivation leads to the following more formal deﬁnition. A stochastic

process {X

t

: t ≥ 0} is said to be a Brownian motion process (or Wiener process)

§34: Introduction to Brownian Motion 152

with parameter σ

2

if

(1) X

t

is a process with stationary independent increments,

(2) X

t

d

= N(0, σ

2

t) for each t > 0, and

(3) X

0

= 0.

The term standard Brownian motion refers to the case in which σ

2

= 1.

Exercise 34–2. Use the independent increments property to show Cov(X

t

, X

s

) =

σ

2

(t ∧ s).

The important but difﬁcult to prove fact about Brownian motion is that the

sample paths of the process can be assumed to be continous. This result can be

intuitively deduced from the approximation argument given earlier. The sample

paths can also be shown to be nowhere differentiable. Again this fact is intuitively

apparent from the earlier construction since the sample paths of the approximating

sums have a lot of corners.

For some purposes the notion of a Brownian motion starting from x will be

useful. Such a process is a Brownian motion as above except for the fact that X

0

= x

is assumed.

Some of the basic properties of Brownian motion are derived here in an intuitive

way. The continuous time analog of the reﬂection principle plays an important role.

Example 34–1. Suppose a > 0 and denote by T

a

the ﬁrst time at which the standard

Brownian motion X

t

takes the value a. Then T

a

= inf{s > 0 : X

s

= a}. The

random variable T

a

is well deﬁned because of the continuity of the sample paths

of the process X. Let t > 0 be ﬁxed. Now by the Theorem of Total Probability

P[X

t

≥ a] = P[X

t

≥ a| T

a

≤ t] P[T

a

≤ t] + P[X

t

≥ a| T

a

> t] P[T

a

> t]. Clearly

P[X

t

≥ a| T

a

> t] = 0. If T

a

≤ t the independent increment property of X suggests

that after time T

a

the Brownian motion restarts as though from scratch except that

the intial value is a rather than 0. Hence P[X

t

≥ a| T

a

≤ t] = P[X

t

≤ a| T

a

≤ t] = 1/ 2.

So P[T

a

≤ t] = 2 P[X

t

≥ a] =

2

√2π

∞

a/

√

t

e

−x

2

/ 2

dx. From this formula P[T

a

< ∞] = 1

but E[T

a

] = ∞.

Exercise 34–3. Show that P[T

a

< ∞] = 1 and E[T

a

] = ∞.

Exercise 34–4. What happens if a < 0?

Example 34–2. Again let a > 0 and ﬁx t > 0. Then P[max

0≤x≤t

X

s

≥ a] = P[T

a

≤ t].

Exercise 34–5. Provide the details of this argument.

§34: Introduction to Brownian Motion 153

Example 34–3. Suppose 0 < t

1

< t

2

. What is the probability that X

t

is 0 at least

once in the interval (t

1

, t

2

)? Let Z denote the event that X

t

is zero for at least one

value of t in (t

1

, t

2

).

P[Z] = E[P[Z| X

t

1

]]

=

∞

−∞

P[T

| x|

≤ t

2

− t

1

] dF

X

t

1

(x)

=

∞

−∞

P[T

| x|

≤ t

2

− t

1

]

1

√

2πt

1

e

−x

2

/ 2t

1

dx.

Exercise 34–6. Simplify this expression to obtain the Arcsine Law.

Martingales can also be exploited to obtain some results about Brownian motion.

The veriﬁcation that {X

t

: t ≥ 0}, {X

2

t

− t : t ≥ 0} and {exp{λX

t

− λ

2

t/ 2} : t ≥ 0}

are martingales is quite straightforward.

Exercise 34–7. Verify that these are martingales.

Example 34–4. Suppose a < 0 < b and let T = inf{s : X

s

= a or X

s

= b}. Applying

the Optional Stopping Theorem gives 0 = E[X

0

] = E[X

T

] = a P[X

T

= a] + b P[X

T

=

b]. Solving gives P[X

T

= a] = b/ (b − a) and P[X

T

= b] = −a/ (b − a).

Exercise 34–8. Verify this computation. Why does the Optional Stopping Theorem

apply?

Exercise 34–9. Use the martingale {X

2

t

− t : t ≥ 0} to compute E[T].

Example 34–5. A Brownian motion with drift is a stochastic process B

t

of the

form B

t

= X

t

+ µt where X is a standard Brownian motion and µ is a nonzero

constant. Again a simple computation shows that {exp{λB

t

− µλt − λ

2

t/ 2} : t ≥ 0}

is a martingale for each ﬁxed value of λ. Choose for simplicity λ = −2µ so that

e

−2µB

t

is a martingale. Let T = inf{s : B

s

= a or B

s

= b} where a < 0 < b are ﬁxed.

The Optional Stopping Theorem gives P[B

T

= b] = (1 − e

−2µa

)/ (e

−2µb

− e

−2µa

). This

is comparable to the result for the gambler’s ruin process when the game is biased.

Exercise 34–10. Fill in the details of the computation.

Example 34–6. Brownian motion in several dimensions is deﬁned in way which

imitates the deﬁnition in dimension 1. A d dimensional Brownian motion process

{X

t

: t ≥ 0} is a process with stationary independent increments for which X

0

= 0

and for each t > 0 the random vector X

t

has the d variate normal distribution with

mean vector 0 and covariance matrix tΣ for some positive deﬁnite matrix Σ. The

standard d dimensional Brownian motion has Σ as the identity matrix. This implies

that the coordinates of a standard d dimensional Brownian motion are independent.

§34: Introduction to Brownian Motion 154

Example 34–7. In the multidimensional setting the distance of a standard Brownian

motion process X

t

from the origin is often of interest. The process Y

t

= || X

t

|| is

called a radial Brownian motion or a Bessel process.

Exercise 34–11. Find the expected time required for a standard multidimensional

Brownian motion to leave the ball of radius a > 0 centered at the origin. Hint:

Show that || X

t

||

2

− dt is a martingale.

§34: Introduction to Brownian Motion 155

Solutions to Exercises

Exercise 34–1. The number of paths from the origin to (a + b, a − b) which

don’t touch the axis is the number of counting paths in which A is always ahead.

The Ballot Theorem gives the number of these paths. The total number of paths

from the origin to (a + b, a − b) is the total number of ways the ballots could be

counted. This number is given by the Proposition.

Exercise 34–2. If t > s write X

t

= X

t

−X

s

+X

s

and use the independent increments

property to compute as follows: Cov(X

t

, X

s

) = E[X

t

X

s

] = E[(X

t

−X

s

)X

s

]+E[X

2

s

] =

0 + Var(X

s

) = σ

2

s. When s > t a similar argument shows that the covariance is

σ

2

t.

Exercise 34–3. Let t → ∞ to obtain P[T

a

< ∞] = (2/

√

2π)

∞

0

e

−x

2

/ 2

dx = 1.

Differentiation gives the density of T

a

as e

−a

2

/ 2t

/ √2πt

3

for t > 0, from which it

is easy to see that E[T

a

] = ∞.

Exercise 34–4. The Reﬂection Principle gives T

a

d

= T

| a|

.

Exercise 34–5. If the maximumexceeds a then the time required to reach level

a must have been smaller than t.

Exercise 34–6. Substitute fromthe earlier formula for P[T

| x|

≤ t

2

−t

1

] and then

integrate by parts.

Exercise 34–7. Just write X

t

= X

t

−X

s

+X

s

and use the independent increments

property.

Exercise 34–9. 0 = E[X

2

T

− T] = a

2

P[X

T

= a] + b

2

P[X

T

= b] − E[T] so that

E[T] = −ab after substitution for the probabilities from the example above.

Exercise 34–11. Here || X

t

||

2

− dt is a martingale so if T

a

is the time at which

the Brownian motion hits the surface of the ball of radius a, E[T

a

] = a

2

/ d by the

Optional Stopping Theorem.

§35. Laboratory 11

One theory about the behavior of stock price P

t

over time is that P

t

should

behave like a stochastic process with independent ratios, that is, if t

1

< . . . < t

n

then P

t

n

/ P

t

n−1

, . . . , P

t

2

/ P

t

1

should be independent. A simple model for this sort of

behaviour is given by a geometric Brownian motion process P

t

= e

B

t

where B

t

is

a Brownian motion with drift. One use of this model of stock prices is to construct

a theoretical pricing model for stock options. A call option is a contract which

gives the owner of the contract the right to buy one share of the underlying stock at

a particular price K called the strike price of the option. The call option contract

expires after a ﬁxed time T. Thus a call that is not used by time T becomes worthless.

1. If the drift and variance parameters of the underlying Brownian motion with

drift are 0.10 and 0.005 respectively, what is the probability that a stock with a price

of $20 today has a price exceeding $25 one year from now? What is the probability

that the stock price exceeds $25 at some time during the next year?

2. The economist Robert Merton has presented an economic argument that call

options will not be exercised before the expiration time T. This implies that the

value of a call option today depends only on the price of the underlying stock at

time T. Since the value of the call at time T is max{P

T

− K, 0}, the value of the call

today (at time 0) is just the actuarial present value of this amount. Thus the value

of a call today is E[v

T

max{P

T

− K, 0}]. This is the Black–Scholes option pricing

formula. Simplify this expression to obtain a form suitable for computation.

3. Explain how stock price data could be used in conjunction with the formula

of the previous question in order to value a call option on a particular stock. Choose

a stock for which call options are traded on the exchange and estimate the parameters

governing the stock price process. Compare the option price with that given by the

Black–Scholes formula. Most traders believe that the interest rate on 3 month U.S.

Treasury bills is the interest rate to use when computing the present value.

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§36. Utility Functions

In this concluding section a fundamental question will be examined. Why would

an insurer assume a risk which the insured is unwilling to assume? To answer this

question, a basic understanding of how value is attached to quantities of money will

be required.

The concept of utility was invented to provide a mathematical methodology

for at least formally measuring the value of money (that is the utility of money).

Naively, one may think that a dollar is a dollar and that’s it. A moments reﬂection

shows that your desire (or your measure of worth) of a payment of, say, $10,000

might differ considerably from that of the person next to you. Consider also the

importance of the same $10,000 to Bill Gates.

These ideas will now be formalized. In this discussion quantities of money

will always be measured in dollars. Denote by u(w) the utility of w dollars. It

is convenient to make a strictly mathematical assumption. Practically speaking, a

utility function can only be deﬁned on a certain discrete set (the multiples of $.01).

However for mathematical tractibility it is assumed that utility functions are deﬁned

on the entire real line or an interval.

As already noted above, different people may well have different utility func-

tions. What features should utility functions have in common? One typical as-

sumption is that for any individual the utility function should be non–decreasing.

This property is the mathematical expression of the fact that having more money

is ‘better.’ Furthermore one would also expect that as one’s wealth increases the

utility of an additional dollar should decrease. This expresses the notion that to

someone having only $10 the prospect of gaining an additional dollar is greater than

the prospect of gaining an additional dollar if one already has $1,000,000.

Exercise 36–1. Argue that if the utility function is sufﬁciently smooth then u

′

(w) ≥

0 and u

′′

(w) ≤ 0. Thus a utility function will be concave (down).

A real valued function u(w) is said to be a risk averse utility function if u(w) is

non–decreasing and concave down.

The reason for the terminology risk averse will be explained shortly.

Example 36–1. It is straightforward to check that each of the functions −e

−αw

,

log w, and αw+β is a risk averse utility function as long as α ≥ 0. A utility function

of the form −e

−αw

is called an exponential utility function.

Example 36–2. Suppose a person with a risk averse utility function has current

wealth w and is given a choice between two investment schemes. In the ﬁrst

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§36: Utility Functions 158

scheme the person will receive an amount b outright. In the second scheme the

person will receive a random amount W where E[W] = b. How does the person

decide between these two alternatives?

To answer this question, decision makers will be assumed to act according to the

Expected Utility Principle: a decision maker always chooses the available option

with the largest expected utility.

Example 36–3. In the previous example the expected utility of the ﬁrst investment

scheme is E[u(w+b)] = u(w+b), while the expected utility of the second investment

scheme is E[u(w + W)].

Generally it is not easy to perform the actual computations required by the

expected utility principal. The following theorem can provide some helpful infor-

mation.

Jensen’s Inequality. If X is a random variable and f is a function which is concave

then

E[f (X)] ≤ f (E[X]).

proof : This proof is valid under the additional assumption that the function f is twice continuously

differentiable. For any ﬁxed x and a, the Fundamental Theorem of Calculus and an

integration by parts gives

f (x) = f (a) +

x

a

f

′

(t) dt

= f (a) + f

′

(a) (x − a) +

x

a

(x − t) f

′′

(t) dt

≤ f (a) + f

′

(a) (x − a)

since f is concave. Substituting X for x and E[X] for a yields

f (X) ≤ f (E[X]) + f

′

(E[X]) (X − E[X]).

Taking expectations of both sides of this last inequality proves the theorem.

Exercise 36–2. Under what conditions does equality hold in Jensen’s inequality?

Continuing the example, from Jensen’s inequality and the fact that a risk averse

utility function is concave

E[u(w + W)] ≤ u(E[w + W]) = u(w + b)

so that such a person will always select the sure payoff b! Would such a person buy

a lottery ticket?

Having discussed the methodology by which a measure of value is assigned to

a given quantity of money, the question of why insurance exists can be addressed.

§36: Utility Functions 159

Consider a situation in which a person with utility function u and current fortune

w faces the prospect of a potential ﬁnancial loss of amount A. For concreteness

suppose that the person is insuring a house against the possibility of a loss due to

ﬁre. It is natural to view the amount of loss A as a random variable since the amount

of loss depends on the severity of the ﬁre, the speed with which the ﬁre department

responds, and other such factors which are unpredictable in nature. Suppose that

for a ﬁxed non–random premium amount P an insurer will indemnify the insured

against such a loss. What premium would the person be willing to pay? If the

offer of insurance is accepted, the expected utility is u(w − P). If the person does

not buy insurance then E[u(w − A)] is the persons expected utility. Because of the

expected utility principle, the person would be willing to buy insurance whenever

the premium satisﬁes

u(w − P) ≥ E[u(w − A)].

This analysis applies to the person seeking insurance. Now analyze the position

of the insurer. Assume that the insurer has utility function u

i

and current wealth

w

i

. Reasoning as before shows that the insurance company will offer insurance at a

premium P if

u

i

(w

i

) ≤ E[u

i

(w

i

+ P − A)].

If there is a set of values Pwhich satisfy simultaneously both of these inequalities

then there is the possibility of insurance being issued.

Exercise 36–3. What is the situation if both the person and the insurer have linear

utility functions?

Exercise 36–4. What happens if the insurer and the person have the same expo-

nential utility function −e

−w/ 1000

, the persons wealth is $5,000, the insurers wealth is

$5,000,000 and the loss variable A has an exponential distribution with mean $500?

Exercise 36–5. What happens in the previous exercise if the utility function is

log w?

Jensen’s inequality can provide some interesting information about the condi-

tions under which a person will purchase insurance. Consider the largest premium

that an individual would pay for insurance. This premium P must satisfy

u(w − P) = E[u(w − A)].

Using Jensen’s inequality on the right member gives

u(w − P) ≤ u(w − E[A])

and since the utility function is non–decreasing, P ≥ E[A]. Thus a risk averse

decision maker would be willing to pay a premium greater than the pure premium

§36: Utility Functions 160

(expected loss) in order to obtain insurance. Similarly the premium charged by the

insurance company must also exceed the pure premium. These two facts reinforce

intuition and lend a certain credibility to the analysis and assumptions.

As a ﬁnal note, consider a type of insurance policy that is typically issued.

A standard type of insurance policy is the stop loss policy in which the insurer

pays only the amount of the claim which exceeds a certain pre–arranged deductible

amount d. In such a policy the insurers payment when the loss is x is given by

(x − d) 1

[d,∞)

(x). There are many other conceivable types of payment policies, and

the following discussion shows why such policies do not generally occur. Suppose

there is another type of payout procedure which for a claim of amount x would pay

R(x). Naturally, 0 ≤ R(x) ≤ x, but even more is true.

Theorem. Suppose a person with risk averse utility function u is to be insured

against a loss of amount A. If there is a stop loss policy with expected payout

E[R(A)] which is offered for the same premium as the policy with payout R(A) then

the stop loss policy has greater expected utility than the policy with payout R(A).

proof : Let d be the deductible for the stop loss policy which exists by hypothesis, let the current

wealth of the person be w, and let the premium be P. Arguing as in the proof of Jensen’s

inequality gives

u(w+R(A) − A − P) − u(w + (A − d) 1

[d,∞)

(A) − A − P)

≤ u

′

(w + (A − d) 1

[d,∞)

(A) − A − P)

R(A) − (A − d) 1

[d,∞)

(A)

≤ u

′

(w − d − P)

R(A) − (A − d) 1

[d,∞)

(A)

**where the fact that u
**

′

is decreasing has been used. Taking expectations gives the result,

since the expectation of the last term is zero (why?).

§36: Utility Functions 161

Problems

Problem36–1. Consider a game of chance in which a fair coin is tossed until a head

apprears. Let N denote the toss on which this occurs. Find the density, expectation,

and variance of N. Suppose a prize of X = 2

N

is paid when a head ﬁrst appears.

What is E[X]? If your utility function is u(w) = log(w) what is the expected utility

of the prize?

Problem 36–2. In a typical state lotto game a person who buys a $1 ticket wins

$1,000,000 with probability 10

−8

. Find a possible utility function for a person who

plays the lotto. Can this person have a risk averse utility function?

Problem 36–3. Suppose a person has a utility function which is increasing and

concave up. Give an example to explain why such a person might be called a risk

lover.

Problem36–4. Showthat if the amount of randomloss X has a uniformdistribution

on the interval (0, 1) and if the insured has utility function u(w) = log(w) then the

maximum amount the insured will pay as a premium is

w −

w

w

e(w − 1)

w−1

.

Problem 36–5. Repeat the previous problem if the utility function is u(w) = −e

−αw

and the loss random variable X has a chi square distribution with n degrees of

freedom.

Problem 36–6. Use the Taylor expansion of u(w − x) about x = µ to show that the

maximum premium an insured will pay is approximately

µ −

1

2

u

′′

(w − µ)

u

′

(w − µ)

σ

2

.

Here µ and σ

2

are the mean and variance of the loss random variable.

Problem 36–7. A group medical insurance policy pays $D each time a member of

the group is hospitalized. The group consists of g distinct subgroups, which differ in

the rate of hospitalization. Suppose that the annual number of hospital admissions

for subgroup i has a Poisson distribution with parameter λ

i

. Find an expression

for the expected claims payments in one year, and also ﬁnd the distribution of the

number of admissions in one year.

Problem 36–8. Suppose the loss random variable X has an exponential distribution

with mean 10. Suppose a premium of $5 will be paid. Show that the propor-

tional insurance policy with beneﬁt X/ 2 and the stop loss policy with beneﬁt

§36: Utility Functions 162

(X −10 log2)1

[10 log 2,∞)

(X) are both feasible insurance policies. Which policy would

the insured choose, and why?

Problem 36–9. True or False: A person with an exponential utility function, −e

−αw

,

considers wealth irrelevent when deciding the maximum premium to pay for com-

plete protection against a random loss.

Problem 36–10. An insurer with wealth w insures a loss X which has the following

probability distribution:

P[X = 0] = P[X = 16] =

1

2

.

The insurer’s utility function is u(x) = log x. The insurer is willing to pay a maximum

of 6 to a reinsurer who accepts 50 percent of the loss. Find w.

Problem 36–11. An insurer and an individual have exponential utility functions

with parameters α and 2α respectively. The individual faces a random loss that is

normally distributed with mean 100. The insurer assumes that the variance is σ

2

.

The individual assumes that the variance is 25. Determine the largest value of σ

2

for which the insurer can charge a premium acceptable to the individual.

Problem 36–12. A decision maker has utility function u(x) = −e

−3x

and initial

wealth w. The decision maker faces two random losses. The loss X has a normal

distribution with mean α and variance 4. The loss Y has a normal distribution with

mean 10 and variance 8. Determine the maximumvalue of α for which the decision

maker prefers X to Y.

Problem36–13. Three insurers have identical utility functions u(w) = log w, w > 0,

and a wealth of 36, 25, and 16 respectively. All three companies insure the same

risk. In the event of a loss each insurer will pay 11. The probability of loss is

0.5. Another company offers to reinsure each company’s complete risk at the same

premium π. Each company is willing to accept the reinsurance at the premium π

if its expected utility is maximized. Determine π such that the reinsurer maximizes

its total expected proﬁt.

§36: Utility Functions 163

Solutions to Problems

Problem 36–1. Here P[N = n] = 2

−n

for n = 1, 2, . . .. Hence E[X] = +∞, and

E[log(X)] =

∞

¸

n=1

n log(2)2

−n

≈ 1.386.

Problem 36–3. See the previous problem.

Problem 36–7. Denote by N

i

the number of hospitalizations for group i. Then

the amount of claims payments in one year is D

g

¸

i=1

N

i

. Also

g

¸

i=1

N

i

has a Poisson

distribution with parameter

g

¸

i=1

λ

i

.

Problem 36–9. True. The variable w cancels from both sides of the equation.

§36: Utility Functions 164

Solutions to Exercises

Exercise 36–1. Since more wealth is better, u(w) is increasing. Thus u

′

(w) ≥ 0.

Since the utility of an additional dollar decreases as wealth increases, u(w+ 1) −

u(w) ≈ u

′

(w) is decreasing. Thus u

′′

(w) = (u

′

(w))

′

≤ 0.

Exercise 36–2. Equality holds if f

′′

(x) = 0 for all x, which means f (x) is a

linear function of x.

Exercise 36–3. The person requires P ≤ E[A] while the insurer requires

P ≥ E[A] so that the only possible premium is P = E[A].

Exercise 36–4. In this case the wealth of each party is irrelevant and the only

common premium is P = 1000 lnE[e

A/ 1000

] = 2000.

Exercise 36–5. The person requires the premium to satisfy ln(5000 − P) ≥

E[ln(5000−A)] while the insurer requires ln(5000000) ≤ E[ln(5000000+P−A)].

To compute the expectations the utility function u(w) is taken to have the value

0 when w ≤ 0. Then the requirement for the person is P ≤ 536.46 while for

the company, P ≥ 500.03 by numerical computation using a computer algebra

package.

§37. Life Table at 5% Interest

Age l

x

d

x

¨ a

x

1000 A

x

1000

2

A

x

1000 A

xx

0 100000 1260 19.922 51.329 17.924 82.836

1 98740 92 20.122 41.823 7.253 63.528

2 98648 64 20.097 43.022 7.071 64.962

3 98584 49 20.064 44.553 7.152 67.000

4 98535 40 20.028 46.307 7.392 69.426

5 98495 36 19.987 48.236 7.747 72.144

6 98459 33 19.944 50.301 8.178 75.075

7 98426 30 19.898 52.498 8.684 78.211

8 98396 26 19.848 54.835 9.272 81.562

9 98370 23 19.796 57.327 9.961 85.156

10 98347 19 19.741 59.974 10.751 88.988

11 98328 19 19.681 62.792 11.662 93.087

12 98309 24 19.619 65.751 12.666 97.393

13 98285 37 19.555 68.811 13.724 101.824

14 98248 52 19.490 71.902 14.759 106.242

15 98196 67 19.425 75.008 15.751 110.613

16 98129 82 19.359 78.129 16.695 114.936

17 98047 94 19.293 81.268 17.585 119.212

18 97953 102 19.226 84.453 18.447 123.493

19 97851 110 19.158 87.726 19.316 127.852

20 97741 118 19.087 91.091 20.195 132.295

21 97623 124 19.014 94.552 21.083 136.826

22 97499 129 18.939 98.134 22.002 141.488

23 97370 130 18.861 101.852 22.964 146.305

24 97240 130 18.779 105.751 24.015 151.356

25 97110 128 18.693 109.849 25.173 156.670

26 96982 126 18.602 114.174 26.470 162.297

27 96856 126 18.507 118.737 27.921 168.252

28 96730 126 18.406 123.534 29.520 174.518

29 96604 127 18.300 128.576 31.284 181.112

30 96477 127 18.189 133.866 33.220 188.034

31 96350 130 18.072 139.426 35.355 195.319

32 96220 132 17.950 145.244 37.680 202.935

33 96088 137 17.822 151.342 40.226 210.918

34 95951 143 17.688 157.708 42.984 219.240

35 95808 153 17.549 164.348 45.968 227.902

36 95655 163 17.404 171.242 49.162 236.861

37 95492 175 17.254 178.404 52.586 246.138

38 95317 188 17.098 185.832 56.247 255.719

39 95129 203 16.936 193.533 60.159 265.611

40 94926 220 16.768 201.506 64.328 275.804

Copyright 2003 Jerry Alan Veeh. All rights reserved.

§37: Life Table at 5% Interest 166

Age l

x

d

x

¨ a

x

1000 A

x

1000

2

A

x

1000 A

xx

41 94706 241 16.595 209.750 68.764 286.290

42 94465 264 16.417 218.248 73.454 297.031

43 94201 288 16.233 227.000 78.408 308.020

44 93913 314 16.044 236.014 83.643 319.265

45 93599 343 15.849 245.292 89.171 330.761

46 93256 374 15.649 254.826 94.994 342.489

47 92882 410 15.443 264.618 101.127 354.445

48 92472 451 15.232 274.647 107.553 366.588

49 92021 495 15.017 284.891 114.257 378.874

50 91526 540 14.798 295.346 121.241 391.286

51 90986 584 14.574 306.018 128.527 403.836

52 90402 631 14.344 316.935 136.156 416.563

53 89771 684 14.110 328.092 144.138 429.454

54 89087 739 13.871 339.464 152.455 442.463

55 88348 797 13.628 351.054 161.123 455.591

56 87551 856 13.380 362.858 170.152 468.830

57 86695 919 13.127 374.890 179.571 482.196

58 85776 987 12.870 387.137 189.384 495.670

59 84789 1063 12.609 399.586 199.586 509.225

60 83726 1145 12.344 412.195 210.141 522.795

61 82581 1233 12.076 424.941 221.027 536.340

62 81348 1324 11.806 437.794 232.219 549.814

63 80024 1415 11.534 450.744 243.712 563.202

64 78609 1502 11.260 463.800 255.529 576.519

65 77107 1587 10.983 476.997 267.729 589.820

66 75520 1674 10.702 490.357 280.360 603.143

67 73846 1764 10.419 503.878 293.435 616.486

68 72082 1864 10.132 517.547 306.957 629.837

69 70218 1970 9.843 531.305 320.858 643.109

70 68248 2083 9.553 545.108 335.091 656.247

71 66165 2193 9.263 558.900 349.587 669.173

72 63972 2299 8.974 572.682 364.352 681.895

73 61673 2394 8.684 586.454 379.395 694.420

74 59279 2480 8.395 600.260 394.790 706.822

75 56799 2560 8.103 614.130 410.598 719.156

76 54239 2640 7.810 628.073 426.851 731.452

77 51599 2721 7.517 642.054 443.518 743.680

78 48878 2807 7.224 656.017 460.530 755.786

79 46071 2891 6.933 669.858 477.742 767.655

80 43180 2972 6.647 683.490 495.023 779.196

§37: Life Table at 5% Interest 167

Age l

x

d

x

¨ a

x

1000 A

x

1000

2

A

x

1000 A

xx

81 40208 3036 6.367 696.796 512.188 790.281

82 37172 3077 6.096 709.717 529.133 800.856

83 34095 3083 5.834 722.208 545.770 810.886

84 31012 3052 5.580 734.292 562.116 820.423

85 27960 2999 5.334 746.010 578.225 829.545

86 24961 2923 5.097 757.276 593.938 838.167

87 22038 2803 4.873 767.969 609.034 846.154

88 19235 2637 4.659 778.150 623.584 853.585

89 16598 2444 4.452 787.993 637.853 860.684

90 14154 2246 4.251 797.588 651.990 867.598

91 11908 2045 4.057 806.812 665.785 874.229

92 9863 1831 3.875 815.461 678.882 880.384

93 8032 1608 3.707 823.460 691.127 886.009

94 6424 1381 3.554 830.749 702.385 891.054

95 5043 1159 3.416 837.313 712.595 895.509

96 3884 945 3.294 843.126 721.669 899.329

97 2939 754 3.184 848.396 729.931 902.714

98 2185 587 3.084 853.138 737.372 905.648

99 1598 448 2.992 857.517 744.245 908.260

100 1150 335 2.907 861.589 750.614 910.548

101 815 245 2.825 865.483 756.669 912.543

102 570 177 2.740 869.539 762.975 914.484

103 393 126 2.649 873.840 769.651 916.294

104 267 88 2.549 878.615 777.064 917.908

105 179 60 2.426 884.468 786.271 919.465

106 119 41 2.253 892.737 799.737 921.800

107 78 27 2.006 904.455 819.532 925.261

108 51 18 1.616 923.036 852.465 933.393

109 33 33 1.000 952.381 907.029 952.381

§38. Service Table

x l

(τ)

x

d

(d)

x

d

(w)

x

d

(i)

x

d

(r)

x

30 100000 100 19990

31 79910 80 14376

32 65454 72 9858

33 55524 61 5702

34 49761 60 3971

35 45730 64 2693 46

36 42927 64 1927 43

37 40893 65 1431 45

38 39352 71 1181 47

39 38053 72 989 49

40 36943 78 813 52

41 36000 83 720 54

42 35143 91 633 56

43 34363 96 550 58

44 33659 104 505 61

45 32989 112 462 66

46 32349 123 421 71

47 31734 133 413 79

48 31109 143 373 87

49 30506 156 336 95

50 29919 168 299 102

51 29350 182 293 112

52 28763 198 259 121

53 28185 209 251 132

54 27593 226 218 143

55 27006 240 213 157

56 26396 259 182 169

57 25786 276 178 183

58 25149 297 148 199

59 24505 316 120 213

60 23856 313 3552

61 19991 298 1587

62 18106 284 2692

63 15130 271 1350

64 13509 257 2006

65 11246 204 4448

66 6594 147 1302

67 5145 119 1522

68 3504 83 1381

69 2040 49 1004

70 987 17 970

Copyright 2003 Jerry Alan Veeh. All rights reserved.