Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
and LifeTable Statistics
Eric V. Slud
Mathematics Department
University of Maryland, College Park
March 22, 2009
c 2009
Eric V. Slud
Statistics Program
Mathematics Department
University of Maryland
College Park, MD 20742
Contents
0.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
1 Basics of Probability & Interest 1
1.1 Probabilities about Lifetimes . . . . . . . . . . . . . . . . . . . 1
1.1.1 Random Variables and Expectations . . . . . . . . . . 6
1.2 Theory of Interest . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Interest Rates and Compounding . . . . . . . . . . . . 9
1.2.2 Present Values and Payment Streams . . . . . . . . . . 14
1.2.3 Principal and Interest, and Discount Rates . . . . . . . 17
1.2.4 Variable Interest Rates . . . . . . . . . . . . . . . . . . 20
1.2.5 Continuoustime Payment Streams . . . . . . . . . . . 24
1.3 Exercise Set 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
1.4 Worked Examples . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5 Useful Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2 Interest & Force of Mortality 33
2.1 More on Theory of Interest . . . . . . . . . . . . . . . . . . . . 33
2.1.1 Annuities & Actuarial Notation . . . . . . . . . . . . . 34
2.1.2 Loan Repayment: Mortgage, Bond, Sinking Fund . . . 39
i
ii CONTENTS
2.1.3 Loan Amortization & Mortgage Reﬁnancing . . . . . . 41
2.1.4 Illustration on Mortgage Reﬁnancing . . . . . . . . . . 42
2.1.5 Computational illustration in R . . . . . . . . . . . . . 44
2.2 Force of Mortality & Analytical Models . . . . . . . . . . . . . 48
2.2.1 Comparison of Forces of Mortality . . . . . . . . . . . . 55
2.3 Exercise Set 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
2.4 Worked Examples . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.5 Useful Formulas from Chapter 2 . . . . . . . . . . . . . . . . . 69
3 Probability & Life Tables 73
3.1 Binomial Variables & Law of Large Numbers . . . . . . . . . . 74
3.1.1 Probability Bounds & Approximations . . . . . . . . . 77
3.2 Simulation of Discrete Lifetimes . . . . . . . . . . . . . . . . . 80
3.3 Expectation of Discrete Random Variables . . . . . . . . . . . 84
3.3.1 Rules for Manipulating Expectations . . . . . . . . . . 87
3.3.2 Curtate Expectation of Life . . . . . . . . . . . . . . . 90
3.4 Interpreting Force of Mortality . . . . . . . . . . . . . . . . . . 91
3.5 Interpolation Between Integer Ages . . . . . . . . . . . . . . . 92
3.5.1 Life Expectancy – Deﬁnition and Approximation . . . 96
3.6 Some Special Integrals . . . . . . . . . . . . . . . . . . . . . . 97
3.7 Exercise Set 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
3.8 Worked Examples . . . . . . . . . . . . . . . . . . . . . . . . . 103
3.9 Appendix on Large Deviation Probabilities . . . . . . . . . . . 109
3.10 Useful Formulas from Chapter 3 . . . . . . . . . . . . . . . . . 112
4 Expected Present Values of Payments 115
CONTENTS iii
4.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.2 Types of Contracts . . . . . . . . . . . . . . . . . . . . . . . . 118
4.2.1 Formal Relations, m = 1 . . . . . . . . . . . . . . . . . 121
4.2.2 Formulas for Net Single Premiums . . . . . . . . . . . 123
4.3 Extension to Multiple Payments per Year . . . . . . . . . . . . 125
4.4 Interpolation Formulas in Risk Premiums . . . . . . . . . . . . 130
4.5 Continuous Risk Premium Formulas . . . . . . . . . . . . . . . 132
4.5.1 Continuous Contracts & Residual Life . . . . . . . . . 133
4.5.2 Integral Formulas . . . . . . . . . . . . . . . . . . . . . 134
4.5.3 Risk Premiums under Theoretical Models . . . . . . . 136
4.5.4 Numerical Calculations of Life Expectancies . . . . . . 140
4.6 Exercise Set 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.7 Worked Examples . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.8 Useful Formulas from Chapter 4 . . . . . . . . . . . . . . . . . 154
A General Features of Duration Data 157
A.1 Survival Data Concepts . . . . . . . . . . . . . . . . . . . . . . 158
A.2 Formal Notion of the Life Table . . . . . . . . . . . . . . . . . 162
A.2.1 The Cohort Life Table . . . . . . . . . . . . . . . . . . 163
A.3 Sample Spaces for Duration Data . . . . . . . . . . . . . . . . 165
A.3.1 Sample Space– Single Life . . . . . . . . . . . . . . . . 166
A.3.2 Sample Space – Cohort Population . . . . . . . . . . . 167
A.3.3 Sample Space – General Case . . . . . . . . . . . . . . 169
Bibliography 171
iv CONTENTS
0.1 Preface
This book is the text for an upperlevel lecture course (STAT 470) at the
University of Maryland on actuarial mathematics, in particular on the basics
of Life Tables, Survival Models, and Life Insurance Premiums and Reserves.
This is a ‘topics’ course, aiming not so much to prepare the students for
speciﬁc Actuarial Examinations – since it cuts across the Society of Actuaries’
Exams FM, M (Segment MLC) and C – as to present the actuarial material
conceptually with reference to ideas from other undergraduate mathematical
studies. Such a focus allows undergraduates with solid preparation in calculus
(not necessarily mathematics or statistics majors) to explore their possible
interests in business and actuarial science. It also allows the majority of such
students — who will choose some other avenue, from economics to operations
research to statistics, for the exercise of their quantitative talents — to know
something concrete and mathematically coherent about the topics and ideas
actually useful in Insurance.
The Insurance material on contingent present values and life tables is
developed as directly as possible from calculus and commonsense notions,
illustrated through word problems. Both the Interest Theory and Proba
bility related to life tables are treated as wonderful concrete applications of
the calculus. The lectures require no background beyond a third semester of
calculus, but the prerequisite calculus courses must have been solidly under
stood. It is a truism of preactuarial advising that students who have not
done really well in and digested the calculus ought not to consider actuarial
studies.
It is not assumed that the student has seen a formal introduction to prob
ability. Notions of relative frequency and average are introduced ﬁrst with
reference to the ensemble of a cohort lifetable, the underlying formal random
experiment being random selection from the cohort lifetable population (or,
in the context of probabilities and expectations for ‘lives aged x’, from the
subset of l
x
members of the population who survive to age x). The cal
culation of expectations of functions of a timetodeath random variables is
rooted on the one hand in the concrete notion of lifetable average, which is
then approximated by suitable idealized failure densities and integrals. Later,
in discussing Binomial random variables and the Law of Large Numbers, the
combinatorial and probabilistic interpretation of binomial coeﬃcients are de
0.1. PREFACE v
rived from the Binomial Theorem, which the student the is assumed to know
as a topic in calculus (Taylor series identiﬁcation of coeﬃcients of a poly
nomial.) The general notions of expectation and probability are introduced,
but for example the Law of Large Numbers for binomial variables is treated
(rigorously) as a topic involving calculus inequalities and summation of ﬁnite
series. This approach allows introduction of the numerically and conceptually
useful largedeviation inequalities for binomial random variables to explain
just how unlikely it is for binomial (e.g., lifetable) counts to deviate much
percentagewise from expectations when the underlying population of trials
is large.
While the basics of actuarial Life Contingencies are treated elsewhere as a
problemsolving method using mortality tables presented in a cohort format,
some eﬀort is devoted in this book to contrasting the form in which the
underlying mortality data are received to the form of the cohort life table
used in calculating premiums and reserves. This allows statistics students to
connect the basic ideas of life table construction – considered by actuaries a
more advanced topic – to the problems of statistical estimation. Accordingly,
some material is included on statistics of biomedical studies and on reliability
which would not ordinarily ﬁnd its way into an actuarial course.
The reader is also not assumed to have worked previously with the The
ory of Interest. These lectures present Theory of Interest as a mathematical
problemtopic, which is rather unlike what is done in typical ﬁnance courses.
Getting the typical Interest problems — such as the exercises on mortgage
reﬁnancing and present values of various payoﬀ schemes — into correct for
mat for numerical answers is often not easy even for good mathematics stu
dents. The approach here is to return to the ﬁrst principles of presentvalue
Equivalence and linear Superposition of payment streams over time. Interest
Theory topics are presented here ﬁrst as a way to learn the skills of apply
ing Equivalence and Superposition principles to real problems, but also as
a way of highlighting the relationship between realized payouts under stan
dard Insurance contracts and instances of standard payment streams with
random duration. In this approach, insurance reserves are seen as natural
generalizations of bond amortization schedules.
While the material in these lectures is presented systematically, it is not
separated by chapters into uniﬁed topics such as Interest Theory, Probability
Theory, Premium Calculation, etc. Instead the introductory material from
vi CONTENTS
probability and interest theory are interleaved, and later, various mathemat
ical ideas are introduced as needed to advance the discussion. No book at
this level can claim to be fully selfcontained, but every attempt has been
made to develop the mathematics to ﬁt the actuarial applications as they
arise logically.
The coverage of the main body of each chapter is primarily ‘theoretical’.
At the end of each chapter is an Exercise Set and a short section of Worked
Examples to illustrate the kinds of word problems which can be solved by
the techniques of the chapter. The Worked Examples sections show how
the ideas and formulas work smoothly together, and they highlight the most
important and frequently used formulas.
Finally, this book diﬀers from other Actuarial texts in its use of compu
tational tools. Realistic problems on present values of payment streams, on
probabilistic survival models related to human lifetimes, and on insurance
contract premiums related to those models, rapidly lead to calculations too
diﬃcult to do by hand or by calculator. Actuarial students often do these
calculations using EXCEL or other spreadsheet programs, but the conceptu
ally based formulas often translate more eﬀectively using mathematical tools
in computing platforms like MATLAB or the statistical language R, especially
where building blocks like rootﬁnders or numerical integration routines are
needed. In this text, we encourage the students to use the free, opensource R
platform because of its powerful tools for numerical integration, rootﬁnding,
lifetable construction, and statistical estimation. Throughout this text, il
lustrations and Exercise solutions and solutions are given in terms of R.
Free web access to the downloadable R platform, including manuals, can
be found at http://www.rproject.org/. There are now many good in
troductory texts on computing with R in statistical applications. One text
which combines a general introduction to R with the speciﬁcs of many sta
tistical data analysis methods, is Venables and Ripley (2002). Some good
free tutorial material on R can also be found on the web, for example at
http://wiener.math.csi.cuny.edu/Statistics/R/simpleR/.
Chapter 1
Basics of Probability and the
Theory of Interest
This ﬁrst Chapter supplies some background on elementary Probability The
ory and basic Theory of Interest. The reader who has not previously studied
these subjects may get an overview here, but will likely want to supplement
this Chapter with reading in any of a number of calculusbased introductions
to probability and statistics, such as Hogg and Tanis (2005) or Devore (2007),
and the basics of the Theory of Interest as covered in the text of Kellison
(2008) or Chapter 1 of Gerber (1997).
1.1 Probability, Lifetimes, and Expectation
In the cohort lifetable model, imagine a number l
0
of individuals born
simultaneously and followed until death, resulting in data d
x
, l
x
for each
integer age x = 0, 1, 2, . . ., where
l
x
= number of lives aged x (i.e. alive at birthday x )
and
d
x
= l
x
−l
x+1
= number dying between ages x, x + 1
Now, allow the agevariable to be denoted by t and to take all real values,
not just whole numbers x, and treat S
0
(t) as the fraction of individuals in a
1
2 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
life table surviving to exact age t. This nonincreasing function S
0
(t) would
be called the empirical ‘survivor’ or ‘survival’ function. Although it takes on
only rational values with denominator l
0
, it can be approximated by a
survival function S(t) which is continuous, decreasing, and continuously
diﬀerentiable (or piecewise continuously diﬀerentiable with just a few break
points) and takes values exactly = l
x
/l
0
at integer ages x. Then for all
positive real y and t, S
0
(y) −S
0
(y +t) is the exact and S(y) −S(y +t)
the approximated fraction of the initial cohort which fails between time y
and y +t, and for integers x, k,
S(x) −S(x +k)
S(x)
=
l
x
−l
x+k
l
x
denotes the fraction of those alive at exact age x who fail before x +k.
What do probabilities have to do with the cohort life table and survival
function ? To answer this, we ﬁrst introduce probability as simply a relative
frequency, using numbers from a cohort lifetable like that of the accompany
ing Illustrative Life Table. In response to a probability question, we supply
the fraction of the relevant lifetable population, to obtain identities like
Pr(life aged 29 dies between exact ages 35 and 41 or between 52 and 60 )
=
S(35) −S(41) +S(52) −S(60)
S(29)
=
_
(l
35
−l
41
) + (l
52
−l
60
)
__
l
29
where our convention is that a life aged 29 is one of the cohort known to
have survived to the 29th birthday. Note that the event of dying between
exact ages 35 and 41 or between 52 and 60 is the union of the nonoverlapping
events of the age random variable having value falling in the interval [35, 41)
with that of falling in [52, 60).
The idea here is that all of the lifetimes covered by the life table are
understood to be governed by an identical “mechanism” of failure, and that
any probability question about a single lifetime is really a question concerning
the fraction of a speciﬁed set of lives, e.g., those alive at age x, whose
lifetimes will satisfy the stated property, e.g., who die either between 35 and
41 or between 52 and 60. This “frequentist” notion of probability of an event
as the relative frequency with which the event occurs in a large population
of (independent) identical units is associated with the phrase “law of large
1.1. PROBABILITIES ABOUT LIFETIMES 3
numbers”, which will be discussed later. For now, remark only that the life
table population should be large for the ideas presented so far to make good
sense. See Table 1.1 for an illustration of a cohort lifetable with realistic
numbers, and for a cohort life table constructed to reﬂect the best estimates
of US male and female mortality rates in 2004, see the Social Security web
page http://www.ssa.gov/OACT/STATS/table4c6.html.
The main ideas arising in the discussion so far are really matters of com
mon sense when applied to relative frequency but require formal axioms when
used more generally:
• Probabilities are numbers between 0 and 1 assigned to subsets of the
entire collection Ω of possible outcomes, with the probability of Ω it
self deﬁned equal to 1. In the examples, the subsets which are assigned
probabilities include subintervals of the interval of possible human life
times measured in years, and also disjoint unions of such subintervals.
These sets in the real line are viewed as possible events summarizing
ages at death of newborns in the cohort population. At this point, we
regard each set A of ages as determining the subset of the cohort
population whose ages at death fall in A.
• The probability Pr(A ∪ B) of the union A ∪ B of disjoint (i.e.,
nonoverlapping) sets A and B is necessarily equal to the sum of the
separate probabilities Pr(A) and Pr(B).
• When probabilities are requested with reference to a smaller universe of
possible outcomes, such as B = lives aged 29, rather than all members
of a cohort population, the resulting conditional probabilities of events
A are written Pr(A B) and calculated as Pr(A∩B)/Pr(B), where
A ∩ B denotes the intersection or overlap of the two events A, B.
The phrase “lives aged 29” deﬁnes an event which in terms of ages at
death says simply “age at death is 29 or larger” or, in relation to the
cohort population, speciﬁes the subset of the population which survives
to exact age 29 (i.e., to the 29th birthday).
• Two events A, B are deﬁned to be independent when Pr(A∩ B) =
Pr(A) · Pr(B) or — equivalently, as long as Pr(B) > 0 — when
the conditional probability Pr(AB) expressing the probability of A
if B were known to have occurred, is the same as the unconditional
probability Pr(A).
4 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
Table 1.1: Illustrative LifeTable, simulated to resemble realistic US Male
lifetable up to age 78. For details of simulation, see Section 3.2 below.
Age x l
x
d
x
x l
x
d
x
0 100000 2629 40 92315 295
1 97371 141 41 92020 332
2 97230 107 42 91688 408
3 97123 63 43 91280 414
4 97060 63 44 90866 464
5 96997 69 45 90402 532
6 96928 69 46 89870 587
7 96859 52 47 89283 680
8 96807 54 48 88603 702
9 96753 51 49 87901 782
10 96702 33 50 87119 841
11 96669 40 51 86278 885
12 96629 47 52 85393 974
13 96582 61 53 84419 1082
14 96521 86 54 83337 1088
15 96435 105 55 82249 1213
16 96330 83 56 81036 1344
17 96247 125 57 79692 1423
18 96122 133 58 78269 1476
19 95989 149 59 76793 1572
20 95840 154 60 75221 1696
21 95686 138 61 73525 1784
22 95548 163 62 71741 1933
23 95385 168 63 69808 2022
24 95217 166 64 67786 2186
25 95051 151 65 65600 2261
26 94900 149 66 63339 2371
27 94751 166 67 60968 2426
28 94585 157 68 58542 2356
29 94428 133 69 56186 2702
30 94295 160 70 53484 2548
31 94135 149 71 50936 2677
32 93986 152 72 48259 2811
33 93834 160 73 45448 2763
34 93674 199 74 42685 2710
35 93475 187 75 39975 2848
36 93288 212 76 37127 2832
37 93076 228 77 34295 2835
38 92848 272 78 31460 2803
39 92576 261
1.1. PROBABILITIES ABOUT LIFETIMES 5
Note: see a basic probability textbook, such as Hogg and Tanis (1997)
or Devore (2007), for formal deﬁnitions and more detailed discussion of the
notions of sample space, event, probability, and conditional probability.
The lifetable, and the mechanism by which members of the population
die, are summarized ﬁrst through the survivor function S(t) which at inte
ger values of t = x agrees with the ratios l
x
/l
0
. Note that S(t) has values
between 0 and 1, and can be interpreted as the probability for a single indi
vidual to survive at least x time units. Since fewer people are alive at larger
ages, S(t) is a decreasing function of the continuous agevariable t, and in
applications S(t) should be continuous and piecewise continuously diﬀeren
tiable (largely for convenience, and because any analytical expression which
would be chosen for S(t) in practice will be piecewise smooth). In addition,
by deﬁnition, S(0) = 1. Another way of summarizing the probabilities of
survival given by this function is to deﬁne the density function
f(t) = −
dS
dt
(t) = −S
(t) (1.1)
as the (absolute) rate of decrease of the function S. Then, by the funda
mental theorem of calculus, for any ages a < b,
Pr( life aged 0 dies between ages a and b )
= S(a) −S(b) =
_
b
a
(−S
(t)) dt =
_
b
a
f(t) dt (1.2)
which has the very helpful geometric interpretation that the probability of
dying within the interval [a, b) is equal to the area under the density curve
y = f(t) over the tinterval [a, b). Note also that the ‘probability’ rule which
assigns the integral
_
A
f(t) dt to the set A (which may be an interval,
a union of intervals, or a still more complicated set) obviously satisﬁes the
ﬁrst two of the bulleted axioms displayed above, namely that P(Ω) = 1
(where Ω is the sample space of all lifetable outcomes) and Pr(A ∪ B) =
Pr(A) + Pr(B) whenever A, B are disjoint or nonoverlapping subsets of
Ω.
The terminal age ω of a life table is an integer value large enough that
S(ω) is negligibly small, but no value S(t) for t < ω is zero. For practical
purposes, no individual lives to the ω birthday. While ω is ﬁnite in real
lifetables and in some analytical survival models, most theoretical forms for
6 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
S(t) have no ﬁnite age ω at which S(ω) = 0, and in those forms ω = ∞
by convention.
In probability theory, the sample space Ω is the set of all detailed out
comes of the underlying datagenerating experiment. Subsets of the sample
space to which probabilities will be assigned are called events. In this book,
all of the interesting events concern lifetimes, or ages at death. Insurance
contract payouts will be expressed as functions of the lifetimes at death of
insured lives, and the average or expected values of these payouts will be used
to calculate a fair equivalent value of the insurance contract to the insured.
The machinery for calculating the average values relates to the concept of
random variable based on the sample space Ω = [0, ∞) of lifetimes.
1.1.1 Random Variables and Expectations
Formally, a random variable is a realvalued mapping X deﬁned on a sample
space Ω, such that {s ∈ Ω : X(s) ∈ (a, b]} is an event with assigned
probability whenever a < b are real numbers. The real number X(s)
is interpreted as the value which would be observed if the detailed outcome
of the underlying random experiment were s ∈ Ω. The most important
feature of a random variable is its probability distribution, which is the
assignment rule of probabilities to all intervals (a, b] of values for X,
denoted for all real numbers a ≤ b by
Pr(a < X ≤ b) ≡ Pr({s ∈ Ω : X(s) ∈ (a, b]})
Remark 1.1 In datasets derived from actual mortality studies or insurance
portfolios, the detailed outcomes can be quite complicated, as discussed in
Appendix A. However, in this and succeeding Chapters, we analyze lifetimes
based on the cohort life table model, also discussed in Appendix A, which
is a simpliﬁed model based on the reduced datastructure, in which numbers
at risk and numbers of observed failures are tabulated on age intervals of one
year.
Now we are ready to deﬁne some terms and motivate the notion of ex
pectation. Think of the age T at which a speciﬁed newly born member of
1.1. PROBABILITIES ABOUT LIFETIMES 7
the population will die as a random variable, which for present purposes
means a variable which takes various values t with probabilities governed
(at integer ages) by the life table data l
x
and the survivor function S(t) or
density function f(t) in a formula like the one just given in equation (1.2).
Suppose there is a contractual amount Y which must be paid (say, to the
heirs of that individual) at the death of the individual at age T, and suppose
that the contract provides a speciﬁc function Y = g(T) according to which
this payment depends on (the wholenumber part of) the age T at which
death occurs. What is the average value of such a payment over all individ
uals whose lifetimes are reﬂected in the lifetable ? Since d
x
= l
x
− l
x+1
individuals (out of the original l
0
) die at ages between x and x + 1,
thereby generating a payment g(x), the total payment to all individuals in
the lifetable can be written as
x
(l
x
−l
x+1
) g(x)
Thus the average payment, at least under the assumption that Y = g(T)
depends only on the largest whole number [T] less than or equal to T, is
x
(l
x
−l
x+1
) g(x) / l
0
=
x
(S(x) −S(x + 1))g(x)
=
x
_
x+1
x
f(t) g(t) dt =
_
∞
0
f(t) g(t) dt
(1.3)
This quantity, the total contingent payment over the whole cohort divided by
the number in the cohort, is called the expectation of the random payment
Y = g(T) in this special case, and can be interpreted as the weighted average
of all of the diﬀerent payments g(x) actually received, where the weights
are just the relative frequency in the life table with which those payments
are received. More generally, if the restriction that g(t) depends only on
the integer part [t] of t were dropped , then the expectation of Y = g(T)
would be given by the same formula
E(Y ) = E(g(T)) =
_
∞
0
f(t) g(t) dt (1.4)
The foregoing discussion of expectations based on lifetime random vari
ables included an interpretation of the expected value of discrete random
variables in terms of weighted averages which holds much more generally. In
this chapter, the averages are taken over all lives tabulated in an underly
ing cohort life table. In Chapter 3, speciﬁcally in Section 3.3, averages are
8 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
taken over large samples of observations of discrete random variables. With
the aid of the Law of Large Numbers, the weightedaverage interpretation of
expectations can be understood as a general mathematical result.
The displayed integral (1.4), like all expectation formulas, can be under
stood as a weighted average of values g(T) obtained over a population,
with weights equal to the probabilities of obtaining those values. Recall from
the Riemannintegral construction in Calculus that the integral
_
f(t)g(t)dt
can be regarded approximately as the sum over very small timeintervals
[t, t + ∆) of the quantities f(t)g(t)∆, quantities which are interpreted as
the base ∆ of a rectangle multiplied by its height f(t)g(t), and the rect
angle closely matches the area under the graph of the function f g over the
interval [t, t + ∆). The term f(t)g(t)∆ can alternatively be interpreted
as the product of the value g(t) — essentially equal to any of the values
g(T) which can be realized when T falls within the interval [t, t +∆) —
multiplied by f(t) ∆. The latter quantity is, by the Fundamental Theorem
of the Calculus, approximately equal for small ∆ to the area under the
function f over the interval [t, t + ∆), and is by deﬁnition equal to the
probability with which T ∈ [t, t +∆). In summary, E(Y ) =
_
∞
0
g(t)f(t)dt
is the average of values g(T) obtained for lifetimes T within small intervals
[t, t +∆) weighted by the probabilities of approximately f(t)∆ with which
those T and g(T) values are obtained. The expectation is a weighted
average because the weights f(t)∆ sum to the integral
_
∞
0
f(t)dt = 1.
Remark 1.2 This way of approximating integrals of continuous integrands
by sums corresponding to the integrals of piecewise constant integrands is
closely related to the construction of the integral in terms of Riemann sums.
For fuller details, see the deﬁnition the Integral via Riemann sums in a cal
culus book like Ellis and Gulick (2002).
The same idea and formula in (1.4) can be applied to the restricted popu
lation of lives aged x. The resulting quantity is then called the conditional
expected value of g(T) given that T ≥ x. The formula will be diﬀerent
in two ways: ﬁrst, the range of integrationis from x to ∞, because of the
resitriction to individuals in the lifetable who have survived to exact age
x; second, the density f(t) must be replaced by f(t)/S(x), the socalled
conditional density given T ≥ x, which is found as follows. From the
1.2. THEORY OF INTEREST 9
deﬁnition of conditional probability, for t ≥ x,
Pr(t ≤ T < t + ∆ T ≥ x) =
Pr( {t ≤ T < t + ∆} ∩ {T ≥ x})
Pr(T ≥ x)
=
Pr(t ≤ T < t + ∆)
Pr(T ≥ x)
=
S(t) −S(t + ∆)
S(x)
Thus the density which can be used to calculate conditional probabilities
Pr(a ≤ T < b  T ≥ x) for x < a < b is
lim
∆→0
1
∆
Pr(t ≤ T < t+∆ T ≥ x) = lim
∆→0
S(t) −S(t + ∆)
S(x) ∆
=
−S
(t)
S(x)
=
f(t)
S(x)
In other words, when it is desired to calculate the expectation of a function
Y = g(T) of the lifetime variable T only within the conditional or restricted
population of individuals with lifetime ≥ x, then the density f(t) in the
expectation formula (1.4) should be replaced by the density which is equal
to f(t)/S(x) for all values of t which are ≥ x, and which is 0 for values
t ∈ [0, x).
The result of all of this discussion of conditional expected values is the
formula, with associated weightedaverage interpretation:
E(g(T)  T ≥ x) =
1
S(x)
_
∞
x
g(t) f(t) dt (1.5)
1.2 Theory of Interest
1.2.1 Interest Rates and Compounding
Since payments based upon unpredictable occurrences or contingencies for
insured lives can occur at diﬀerent times, we study next the Theory of Inter
est, which is concerned with valuing streams of payments made over time.
The general model in the case of constant interest, to which we restrict in
the current subsection, is as follows. Money is deposited in an account like
a bankaccount and grows according to a schedule determined by both the
interest rate and the occasions when interest amounts are compounded, that
10 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
is, deemed to be added to the account. The compounding rules are impor
tant because they determine when new interest interest begins to be earned
on previously earned interest amounts.
The central concept of compound interest is that, over the ﬁxed time
interval of one year, an amount A
0
deposited at the beginning of the interval
accumulates to A
1
= A
0
· (1 + i) which could be withdrawn at the end
of the interval. Since the constant interest rate is quoted as a constant over
the period of one year, we have 1 +i as the accumulation factor by which
an initial deposit is multiplied to ﬁnd the balance at the end of one year.
By convention, interest rates are generally quoted as annualized rates, which
means that the interest rate i
h
applied to a timeinterval [t, t + h] for a
period h of less than one year is prorated down to the interval h to give
hi
h
, which results in an accumulation factor 1 + hi
h
. Thus, for an initial
deposit A
0
at time t which is to be retained in a bank account for the time
h, so that the accumulated amount is compounded (i.e., is calculated by
the bank and owned by the depositor) at time t +h, the balance which the
owner could withdraw at time t +h is A(0) · (1 +hi
h
).
If the quoted interest rate i
h
is annualized, and if interest earned is
to be credited after every successive interval h = 1/m, then we say that
the interest rate is a nominal annualized interest rate with mtimes
yearly compounding or simply the nominal interest rate, and the standard
notation for it is i
(m)
instead of i
1/m
as written above.
Banks are not required to calculate interest from the instant (in practical
terms, the day) of deposit to the instant (i.e., the day) of withdrawal. In
practice, the intervals of compounding are generally fractions h = 1/m of
a year, usually with m = 1, 2, 4, or 12. This means that after a deposit
of A
0
at time t, the depositor wishing to withdraw the full accumulation
or balance at time t +s for 0 < s < h owns only the initial amount A
0
,
because no interest has yet been credited.
The further growth of deposited money over successive time intervals
of length h = 1/m, if compounded at each additional interval of length
1/m, is easily understood inductively. With amount A
0
deposited initially
at time t, the balance as of time t + h is A
0
(1 + i
(m)
/m) and can be
viewed as though it were simultaneously withdrawn and freshly deposited at
time t + h, after which it would accumulate over the succeeding interval
1.2. THEORY OF INTEREST 11
[t + h, t + 2h] by multiplying the deposited amount A
0
(1 + i
(m)
/m) by
the intervalh accumulation factor 1 +i
(m)
/m. Thus the balance as of time
t +2h = t +2/m is A
0
(1 +i
(m)
/m)
2
. Inductively, for k ≥ 2, if the balance
A
0
(1 + i
(m)
/m)
k
owned by the depositor at time t + k/m is regarded as
instantaneously withdrawn and redeposited as an initial balance for the next
interval [t + k/m, t + (k + 1)/m], the balance at time t + (k + 1)/m is
A
0
(1 +i
(m)
/m)
k
multiplied by the intervalh accumulation factor 1 +ih, or
A
0
(1 +i
(m)
/m)
(k+1)
.
The overall result of our reasoning about mtimes yearly compounded
nominal interest is the following:
Proposition 1.1 The accumulated value of an initial bank deposit of A
0
compounded m times yearly at nominal interest rate i
(m)
after a time
k/m+s, where 0 ≤ s < 1/m and k ≥ 0 is an integer, is (1+i
(m)
/m)
k
· A
0
.
Proposition 1.1 with k = m says that at the annualized nominal interest
rate i
(m)
, an initial deposit of A
0
accumulates after exactly one year
to a balance of (1 + i
(m)
/m)
m
A
0
. Since the accumulation from the full
year of deposit has the eﬀect of multiplying the initial deposit by the factor
(1 + i
(m)
/m)
m
, a factor which would have been 1 + i at interest rate i
compounded yearly. This proves that the nominal interest rate i
(m)
with m
timesyearly compounding leads to exactly the same accumulation over whole
years as a deposit account with the onceyearly compounded “eﬀective” rate
i ≡ i
eﬀ
= (1 +i
(m)
/m)
m
− 1
Since any nominal interest rate i
(m)
has its equivalent eﬀective interest
rate i = i
eﬀ
providing the same yearly accumulations, the nominal interest
rates i
(m)
with diﬀerent values of m but the same value of i can also
be regarded as equivalent. These wholeyearequivalent nominal rates are
determined by solving the last equation for i
(m)
in terms of i = i
eﬀ
:
i
(m)
= m
_
(1 +i)
1/m
− 1
_
(1.6)
For example, with i = .05, or 5% eﬀective annual interest, the corre
sponding nominal rates i
(m)
for the most common values of m are obtained
through the R code line:
12 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
mvec = c(1,2,4,12, 365) ; imvec = mvec*(1.05ˆ(1/mvec)  1)
as
i
(1)
= .05 , i
(2)
= .04939 , i
(4)
= .04909 , i
(12)
= .04889 , i
(365)
= .04879
A few simple calculus manipulations allow us to establish the pattern of
the displayed i
(m)
values for all choices of i, m. The righthand side of
equation (1.6) is a function g(h) of h = 1/m, where
g(h) = h
−1
_
(1 +i)
h
− 1
_
= (exp(hln(1 +i)) −1)/h
has the form of a diﬀerence quotient from Calculus. Recall the Taylor series
expansion e
z
= 1 + z + z
2
/2 + z
3
/3! · · · which is valid for all z > 0.
Substitute this series with z = hln(1 + i) into the displayed formula for
g(h) to conclude that
g(h) =
∞
j=1
(ln(1 + i) · h)
j
h · j!
= ln(1 +i) +
∞
j=1
1
j!
(ln(1 +i))
j
h
j−1
is an increasing function of h > 0 and is always greater than its righthand
limit
g(0+) = lim
h0
exp(hln(1 +i)) − 1
h
=
d
dh
_
e
h ln(1+i)
_
h=0
= ln(1 +i)
The information just established concerning the behavior of i
(m)
=
g(1/m) as a function of m for ﬁxed eﬀective interest rate i = i
eﬀ
is
summarized as follows.
Proposition 1.2 When i = i
eﬀ
is ﬁxed, the nominal annual interest rate
i
(m)
for mtimesyearly compounding is a decreasing function of the positive
integer m and tends as m →∞ to the limiting value, deﬁned as the force
of interest,
δ = ln(1 +i
eﬀ
) = lim
m→∞
i
(m)
1.2. THEORY OF INTEREST 13
In the displayed i
(m)
values for i = .05, the dailycompounded nominal
interest rate was i
(365)
= .048973. The corresponding force of interest, also
called the instaneously or continuously compounded nominal interest rate, is
δ = ln(1.05) = .048970.
The eﬀective interest rate i
eﬀ
can be expressed through its nominal
continuously compounded interest rate δ as i = e
δ
, and the other nominal
rates have similar expressions immediately derived from (1.6):
i
(m)
= m(e
δ/m
− 1)
For all durations t which are rational numbers, i.e., are of the form
t = k/m for positive integers k, m, Prop. 1.1 with s = 0 says that the
accumulation factor for duration t = k/m based on mtimesyearly com
pounding at eﬀective interest rate i is (1+i
(m)
/m)
k
= (1+i
(m)
/m)
mt
= e
t δ
.
Since t = k/m is also of the form kl/(ml) for every integer l ≥ 1, the same
reasoning gives e
tδ
as the accumulation factor for duration t under the
same eﬀective interest rate with mltimesyearly compounding. Taking the
limit as l →∞, with t ﬁxed and arbitrary m ≥ 1, says that the accumu
lation factor over duration t for instantaneous or continuous compounding
should be the same. This is essentially a deﬁnition of what accumulation
by continuous compounding should mean, but it is the only deﬁnition under
which continuous compounding is well approximated by compounding arib
trarily (but ﬁnitely) many times per day.
Now it is obvious that the accumulation factor by continuous compound
ing over a duration k/m+s (for 0 ≤ s ≤ 1/m is nondecreasing in s and
must therefore lie within the interval [e
δk/m
, e
δ(k+1)/m
]. By continuity of the
exponential function, there follows:
Proposition 1.3 The accumulated balance of an initial deposit A
0
under
continuous compounding with eﬀective interest rate i, or equivalently with
force of interest δ = ln(1+i), over a duration t > 0 which is not necessarily
a rational number, is exp(δt) · A
0
.
So far, we have described in Props. 1.1 and 1.3 the mechanism of accu
mulation under nominal interest rates applying with either mtimesyearly or
continuous compounding, and in equation (1.6) and Prop. 1.2 the relations
14 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
between nominal interest rates, force of interest, and eﬀective interest rates.
There is a further term for interest rates which must be disclosed to borrow
ers under US contracts, namely the Annual Percentage Rate or APR. Unlike
the interest rate terminology discussed up to this point, APR is a legal term
which refers either (‘eﬀective APR’) to the eﬀective interest rate or (‘nominal
APR’ or simply ‘APR’) to the nominal interest, to which is added in either
case the service fees charged by the lender as a fraction of the beginningof
year loan balance. The APR disclosure is intended as a consumer protection
to the borrower, but may vary across jurisdictions in the way startup fees
(e.g., origination and participation) are required to be reported.
1.2.2 Present Values and Payment Streams
Applications of the theory of interest generally involve comparisons between
streams of payments which may be made at diﬀerent times and may accu
mulate at diﬀerent rates of interest. These payments may be deposits into
a bank or investment account, or loan repayments, or successive payments
designed to accumulate over time at interest to a suﬃcient reserve fund to
meet some future liability.
First, a discrete payment stream is a sequence of (positive) deposit
amounts α
j
made at speciﬁed calendar times t
j
, j = 1, 2, . . . , n and
which are regarded as accumulating from their times of deposit according to
a schedule of interest rates r(t) which remain constant within successive
intervals of calendar time t but which may change from one such interval
to the next.
Two basic principles govern all problems of valuing such payment streams.
• The Principle of Equivalence deﬁnes equivalence at time τ
between two payment streams, one with payments and times (α
j
, t
j
,
j = 1, 2, . . . , n) and interest rate function r(t) and the other with
payments and times (α
∗
j
, t
∗
j
, j = 1, 2, . . . , n
∗
) and interest rate func
tion r
∗
(t), where τ ≥ max
j
t
j
, max
j
t
∗
j
. These streams are called
equivalent at τ if the accumulated values at τ from the two payment
streams under their respective interest rate functions are the same.
1.2. THEORY OF INTEREST 15
• The Principle of Linear Superposition states that the total accu
mulated amount resulting at time τ from a payment stream (α
j
, t
j
, j =
1, 2, . . . , n) under interest rate function r(t) is the same as the sum of
the accumulated values up to time T of n separate deposit accounts
initiated at the respective times t
j
with deposits of α
j
, all under the
same interest rate function r(t).
The two Principles as just stated do not yet tell us how to calculate the
accumulated values at τ under interest rate functions r(t) that vary over
time. However, we can already see that the ﬁrst Principle is a deﬁnition,
while we will see that the second is an essentially obvious restatement of
the commutativity of addition together with the fact that the accumulation
of discrete payment streams is a welldeﬁned linear function of the payment
amounts α
j
.
Consider ﬁrst the case where r(t) ≡ i is constant over the entire time
interval [min
j
t
j
, τ]. Then Prop. 1.3 gives the contribution of the deposit
α
j
at time t
j
to the accumulated value at τ as α
j
· (1 + i)
τ−t
j
. On
the other hand, the direct inductive calculation of the accumulated amounts
at all times t
l
≥ t
j
due to the α
j
deposit, are also given via Prop. 1.3
as α
j
· (1 + i)
t
l
−t
j
, from which (by continuously compounding at interest
rate i from the largest of the times t
l
until τ) the ﬁnal contribution
of the α
j
deposit to the ﬁnal accumulation at τ is again seen to be
α
j
· (1 +i)
τ−t
j
. This argument, with a little more notational eﬀort and an
inductive argument over the successively larger deposit times t
l
, can be made
into a rigorous proof of the Linear Superposition principle in the constant
interestrate environment with continuous compounding. The formula for
the continuously compounded accumulated value of the stream at time τ is
n
j=1
α
j
(1 +i)
τ−t
j
=
n
j=1
α
j
e
δ (τ−t
j
)
(1.7)
If compounding is instead mtimesyearly and all of the timediﬀerences τ−t
l
are integer multiples of 1/m, then we appeal to Propositions 1.1 and 1.3
to conﬁrm that there is no diﬀerence between the accumulated values at
eﬀective interest rate i under continuous or mtimesyearly compounding,
and formula (1.7) again expresses the accumulated value at τ, which is also
16 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
equal to
n
j=1
α
j
(1 +i
(m)
/m)
m(τ−t
j
)
The most important application of the principle of equivalence is in ﬁnd
ing a deposit amount α
PV
at a single ﬁxed time t which is equivalent
to the payment stream (α
j
, t
j
, j = 1, . . . , n) at all times τ ≥ max
j
t
j
, at
the same eﬀective interest rate i = i
eﬀ
with continuous compounding as
is used to accumulate (α
j
, t
j
, j = 1, . . . , n). This amount α
PV
is then
called the present value of the payment stream at time t . To see
why this is possible, consider any ﬁxed τ ≥ max(t, max
j
t
j
) and equate
the accumulated value (1.7) to the accumulated value α
PV
(1 +i)
τ−t
of the
single deposit at time t using interest rate i, yielding:
n
j=1
α
j
(1 +i)
τ−t
j
= α
PV
(1 +i)
τ−t
=⇒ α
PV
=
n
j=1
α
j
(1 +i)
t−t
j
(1.8)
This equation determining α
PV
evidently does not depend upon τ. It tells
ﬁrst (with n = 1, t
1
= 0 in (1.8)) the present value at ﬁxed interest rate i
of a payment of 1 exactly t years in the future, (1 + i)
−t
. This is the
amount which must be put in the bank at time 0 in order to accumulate by
the factor (1 +i)
t
given by Prop. 1.3) to the value 1 at time t. Then, more
generally,
the present value at time 0 under constant interest rate i of
a payment stream consisting of payments α
j
at future times
t
j
, j = 1, . . . , n is equal to the summation
n
j=1
α
j
(1 +i)
−t
j
.
The same phenomenon, that a single deposit α
PV
at time t
0
can be
equivalent at all times τ to a payment stream (α
j
, t
j
, j = 1, . . . , n) turns
out to hold more generally whenever the same timevarying interest rate
function r(t) is used to accumulate both the single deposit and the payment
stream. The proof of this Fact will be left to an Exercise in Section 1.2.4
where accumulation formulas for variable interest rates are discussed. The
magnitude α
PV
is then the general present value of the payment stream at
time t
0
.
1.2. THEORY OF INTEREST 17
1.2.3 Principal and Interest, and Discount Rates
In this Section, we consider the compounding of interest from the point of
view of a borrower of an amount L at time 0, where the interest rate is
constant with i
eﬀ
= i. Initially assume continuous compounding for all
accumulations. If the borrower plans to make payments α
j
, 1 ≤ j ≤ n, at
times 0 < t
1
< t
2
< · · · < t
n
, then by deﬁnition
the principal remaining on the loan as of time t is equal to the
accumulated value at t of the single deposit L at time 0 , minus
the accumulated value at t under continuously compounded
eﬀective rate i of all payments made at times before t, i.e.,
Principal at time t = L(1 +i)
t
−
j: t
j
≤t
α
j
(1 +i)
t−t
j
.
The principal remaining in the loan just after a payment has been made is
the same as the amount the borrower could pay to pay oﬀ the loan completely
at that instant. In addition, if there are fees or late charges due at the times
t
j
when payments are made, then those amounts are added to the Principal
or Balance owed as of t
j
. However, in the present discussion we ignore all
such additional fees or charges.
The principal owed on the loan just after time t reﬂects that as of time
t, the lender must be compensated for the amount (1 +i)
t
L to which the
original loan amount would accumulate; while the accumulated value of the
stream of payments actually made up to time t reduces the debt.
Each payment α
j
made can be broken down into the socalled Interest
and Principal portions by the rule:
Interest Portion of Paymt at t
j
= (Principal at t
j−1
) · ((1 +i)
t
j
−t
j−1
−1)
Principal Portion of Paymt at t
j
= α
j
− Interest Portion of Paymt at t
j
The ﬁrst of these lines is clearly the amount of interest that the principal
just after t
j−1
would have earned at rate i over the time interval t
j
−t
j−1
.
The amount of the payment at t
j
minus the amount of interest at t
j
is the
amount by which the principal decreases from just after t
j−1
to just after
t
j
. This simple Proposition is not quite obvious, but is easily shown by an
algebraic rearrangement of terms, given as an Exercise.
18 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
Exercise 1.A. Show that the foregoing deﬁnitions of Principal and Principal
Portions of payments are compatible by deriving the following identity from
the deﬁnitions. If Π(t) denotes the principal owed just after time t, and
π
j
denotes the principal portion of the payment at t
j
, then
Π(t
j−1
) − π
j
= Π(t
j
) 2
The nominal interest rates i
(m)
for diﬀerent periods of compounding were
seen in Prop. 1.2 to be related by the formulas
(1 +i
(m)
/m)
m
= 1 +i = 1 +i
eﬀ
, i
(m)
= m
_
(1 +i)
1/m
− 1
_
(1.9)
Similarly, interest can be said to be governed by the discount rates d
(m)
for
various compounding periods, deﬁned by
1 − d
(m)
/m = (1 +i
(m)
/m)
−1
Solving the last equation for d
(m)
gives
d
(m)
= i
(m)
/(1 +i
(m)
/m) (1.10)
The idea of discount rates is that if an amount 1 is loaned out at interest,
then the amount d
(m)
/m is the correct amount to be repaid at the beginning
rather than the end of each fraction 1/m of the year, with repayment of
the principal of 1 at the end of the year, in order to amount to the same
eﬀective interest rate. The reason is that, according to the deﬁnition, the
amount 1−d
(m)
/m accumulates at nominal interest i
(m)
to (1−d
(m)
/m) ·
(1 +i
(m)
/m) = 1 after a timeperiod of 1/m.
The quantities i
(m)
and d
(m)
are naturally introduced as the interest
payments which must be made respectively at the ends and the beginnings
of successive timeperiods of length 1/m in order that the principal owed at
each time j/m on an amount 1 borrowed at time 0 will always be 1. To
deﬁne these terms and justify this assertion, consider ﬁrst the simplest case,
m = 1. If 1 is to be borrowed at time 0, then the single payment at time
1 which fully compensates the lender, if that lender could alternatively have
earned interest rate i, is (1+i), which we view as a payment of 1 principal
(the face amount of the loan) and i interest. In exactly the same way, if
1 is borrowed at time 0 for a timeperiod 1/m, then the repayment at
1.2. THEORY OF INTEREST 19
time 1/m takes the form of 1 principal and i
(m)
/m interest. Thus, if
1 was borrowed at time 0, an interest payment of i
(m)
/m at time 1/m
leaves an amount 1 still owed, which can be viewed as an amount borrowed
on the timeinterval (1/m, 2/m]. Then a payment of i
(m)
/m at time
2/m still leaves an amount 1 owed at 2/m, which is deemed borrowed
until time 3/m, and so forth, until the loan of 1 on the ﬁnal timeinterval
((m−1)/m, 1] is paid oﬀ at time 1 with a ﬁnal interest payment of i
(m)
/m
together with the principal repayment of 1. The overall result which we
have just proved intuitively is:
1 at time 0 is equivalent to the stream of m payments of
i
(m)
/m at times 1/m, 2/m, . . . , 1 plus the payment of 1 at
time 1.
Similarly, if interest is to be paid at the beginning of the period of the
loan instead of the end, the interest paid at time 0 for a loan of 1 would
be d = i/(1 + i), with the only other payment a repayment of principal at
time 1. To see that this is correct, note that since interest d is paid at the
same instant as receiving the loan of 1 , the net amount actually received
is 1 − d = (1 + i)
−1
, which accumulates in value to (1 −d)(1 +i) = 1 at
time 1. Similarly, if interest payments are to be made at the beginnings
of each of the intervals (j/m, (j + 1)/m] for j = 0, 1, . . . , m − 1, with
a ﬁnal principal repayment of 1 at time 1, then the interest payments
should be d
(m)
/m. This follows because the amount eﬀectively borrowed
(after the immediate interest payment) over each interval (j/m, (j +1)/m]
is (1−d
(m)
/m), which accumulates in value over the interval of length 1/m
to an amount (1 − d
(m)
/m)(1 +i
(m)
/m) = 1. So throughout the yearlong
life of the loan, the principal owed at (or just before) each time (j +1)/m is
exactly 1. The overall result concerning mperiodyearly discount interest
is
1 at time 0 is equivalent to the stream of m payments of
d
(m)
/m at times 0, 1/m, 2/m, . . . , (m−1)/m plus the payment
of 1 at time 1.
A useful algebraic exercise to conﬁrm the displayed assertions is:
20 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
Exercise 1.B. Verify that the present values at time 0 of the payment
streams with m interest payments in the displayed assertions are respectively
m
j=1
i
(m)
m
(1 +i)
−j/m
+(1 +i)
−1
and
m−1
j=0
d
(m)
m
(1 +i)
−j/m
+(1 +i)
−1
and that both are equal to 1. These identities are valid for all i > 0. 2
1.2.4 Variable Interest Rates
Now we formulate the generalization of these ideas to the case of nonconstant
instantaneously varying, but known or observed, eﬀective interest rate r(t)
at time t , corresponding to the instantaneous continuously compounded
nominal rate, or timevarying force of interest , δ(t) = ln(1+r(t)). Consider
the compounding of interest over successive intervals [b +kh, b + (k + 1)h],
where h = 1/m for large m, there is an essentially constant principal
amount over each interval of length 1/m. Since we assume the functions
r(t) and therefore ln(1 +δ(t)) are uniformly continuous in t, so that over
very short intervals [b +kh, b +(k +1)h] with instantaneous compounding,
the interest rate and its associated force of interest are essentially constant,
with accumulation factor over the interval given by e
hδ(kh)
. Therefore, if an
initial time b and duration τ > 0 are ﬁxed and [mτ] = [τ/h] denotes the
largest integer ≤ mτ, we ﬁnd that the continuous compounding of interest
over the timeinterval [b, b + τ] results in an overall accumulation factor of
approximately
e
hδ(b)
e
hδ(b+h)
e
hδ(b+2h)
· · · e
hδ(b+([τ/h]−1)h)
exp
_
((τ −h[τ/h]) · δ(b +h[τ/h])
_
which has limit as m → ∞ equal to
exp
_
lim
m
1
m
[mτ]−1
k=0
δ(b +k/m)
_
= exp
_
_
t
0
δ(b +s) ds
_
The last step in this chain of equalities relates the concept of continuous
compounding to that of the Riemann integral. To specify continuoustime
varying interest rates in terms of instantaneous eﬀective rates, we would
1.2. THEORY OF INTEREST 21
equate the last displayed formula for the accumulation factor over [b, b +τ]
to
exp
_
_
t
0
ln(1 +r(b +s)) ds
_
Next consider the case of deposits α
0
, α
1
, . . . , α
k
, . . . , α
n
made at times
0, h, . . . , kh, . . . , nh, where h = 1/m is the given compoundingperiod, and
wherenominal annualized instantaneous interestrates δ(kh) (with compounding
period h) apply to the accrual of interest on the interval [kh, (k +1)h). If
the accumulated bank balance just after time kh is denoted by B
k
, then
how can the accumulated bank balance be expressed in terms of α
j
and
δ(jh) ? Clearly
B
k+1
= B
k
· e
δ(kh)/m
+ α
k+1
, B
0
= α
0
The preceding diﬀerence equation can be solved in terms of successive sum
mation and product operations acting on the sequences α
j
and δ(jh), as
follows. First deﬁne a function A
k
to denote the accumulated bank balance
at time kh for a unit invested at time 0 and earning interest with instan
taneous nominal interest rates δ(jh) applying respectively over the whole
compoundingintervals [jh, (j + 1)h), j = 0, . . . , k −1. Then by deﬁnition,
A
k
satisﬁes a homogeneous equation analogous to the previous one, which
together with its solution is given by
A
k+1
= A
k
· e
δ(kh)/m
, A
0
= 1, A
k
=
k−1
j=0
e
δ(jh)/m
We now return to the idea of equivalent investments and present value of
a payment stream, as discussed in Section 1.2.2. Our object is to determine a
single deposit D at time 0 which is equivalent at time τ = nh to a stream
of deposits α
j
, j = 0, 1, 2, . . . , n, where all amounts accumulate according
to the continuously compounded instantaneous eﬀective interest rate r(t)
and associated force of interest δ(t) = ln(1 + r(t)). By approximating the
continuous interest rate function r(t) by one which is constant on intervals
[kh, (k + 1)h), we have just calculated that an amount 1 at time 0
compounds to an accumulated amount A
n
at time τ = nh. Therefore, an
amount D at time 0 accumulates to D · A
n
at time τ, and in particular
D = 1/A
n
at time 0 accumulates to 1 at time τ. Note, as in (1.8)
22 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
of Section 1.2.2, this single equivalent deposit D would be the same if the
accumulations were valued at any other time τ
> nh. Thus the present
value of 1 at time τ = nh is 1/A
n
. Now deﬁne G
k
to be the present
value of the stream of payments α
j
at time jh for j = 0, 1, . . . , k. Since
B
k
was the accumulated value just after time kh of the same stream of
payments, and since the present value at 0 of an amount B
k
at time kh
is just B
k
/A
k
, we conclude
G
k+1
=
B
k+1
A
k+1
=
B
k
exp(δ(km)/m)
A
k
exp(δ(km)/m)
+
α
k+1
A
k+1
, k ≥ 1 , G
0
= α
0
Thus G
k+1
−G
k
= α
k+1
/A
k+1
, and
G
k+1
= α
0
+
k
i=0
α
i+1
A
i+1
=
k+1
j=0
α
j
A
j
In summary, we have simultaneously found the solution for the accumulated
balance B
k
just after time kh and for the present value G
k
at time 0 :
G
k
=
k
i=0
α
i
A
i
, B
k
= A
k
· G
k
, k = 0, . . . , n
The formulas just developed can be used to give the internal rate of return
r over the timeinterval [0, τ] of a unit investment which pays amount α
k
at times t
k
, k = 0, . . . , n, 0 ≤ t
k
≤ τ. This constant (eﬀective) interest
rate r is the one such that
n
k=0
s
k
_
1 +r
_
−t
k
= 1
With respect to the constant interest rate r , the present value of a payment
α
k
at a time t
k
timeunits in the future is α
k
· (1 +r)
−t
k
. Therefore the
stream of payments α
k
at times t
k
, (k = 0, 1, . . . , n) becomes equivalent,
for the uniquely deﬁned interest rate r, to an immediate (time0) payment
of 1.
1.2. THEORY OF INTEREST 23
Exercise 1.C. As an illustration of the notion of eﬀective interest rate, or
internal rate of return, suppose that you are oﬀered an investment option
under which an investment of 10, 000 made now is expected to pay 300
yearly for 5 years (beginning 1 year from the date of the investment), and then
800 yearly for the following ﬁve years, with the principal of 10, 000 returned
to you (if all goes well) exactly 10 years from the date of the investment (at
the same time as the last of the 800 payments. If the investment goes
as planned, what is the eﬀective interest rate you will be earning on your
investment ?
Solution. As in all calculations of eﬀective interest rate, the present value
ofthe paymentstream, at the unknown interest rate r = i
eﬀ
, must be bal
anced with the value (here 10, 000) which is invested. (That is because the
indicated payment stream is being regarded as equivalent to bank interest at
rate r.) The balance equation in the Example is obviously
10, 000 = 300
5
j=1
(1 + r)
−j
+ 800
10
j=6
(1 +r)
−j
+ 10, 000 (1 +r)
−10
The righthand side can be simpliﬁed somewhat, in terms of the notation
x = (1 +r)
−5
, to
300
1 +r
_
1 −x
1 −(1 +r)
−1
_
+
800x
(1 +r)
_
1 −x
1 −(1 +r)
−1
_
+ 10000 x
2
=
1 −x
r
(300 + 800x) + 10000x
2
(1.11)
Setting this simpliﬁed expression equal to the lefthand side of 10, 000 does
not lead to a closedform solution, since both x = (1+r)
−5
and r involve the
unknown r. Nevertheless, we can solve the equation roughly by ‘tabulating’
the values of the simpliﬁed righthand side as a function of r ranging in
increments of 0.005 from 0.035 through 0.075. (We can guess that the
correct answer lies between the minimum and maximum payments expressed
as a fraction of the principal.) This tabulation yields:
r .035 .040 .045 .050 .055 .060 .065 .070 .075
(1.11) 11485 11018 10574 10152 9749 9366 9000 8562 8320
From these values, we can see that the righthand side is equal to 10, 000
for a value of r falling between 0.05 and 0.055. Interpolating linearly to
24 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
approximate the answer yields r = 0.050 +0.005 ∗ (10000 −10152)/(9749 −
10152) = 0.05189, while an accurate equationsolver ﬁnds r = 0.05186. For
example, the rootﬁnding function in R is called uniroot , and the R code for
computing the eﬀective interest rate in this Example is:
Rsolv = function(r) { x = (1+r)^(5)
(1x)*(300+800*x)/r + 10000*x^2  10000 }
uniroot(Rsolv, c(.035,.075))$root
[1] 0.05185676
1.2.5 Continuoustime Payment Streams
There is a completely analogous development for continuoustime deposit
streams with continuous compounding. Suppose D(t) to be the rate per
unit time at which savings deposits are made, so that if we take m to go to
∞ in the previous discussion, we have D(t) = lim
m→∞
mα
[mt]
, where [·]
again denotes greatestinteger. Taking δ(t) to be the timevarying nominal
interest rate with continuous compounding, and B(t) to be the accumulated
balance as of time t (analogous to the quantity B
[mt]
= B
k
from before,
when t = k/m), we replace the previous diﬀerenceequation by
B(t +h) = B(t) (1 +hδ(t)) + hD(t) + o(h)
where o(h) denotes a remainder such that o(h)/h → 0 as h → 0.
Subtracting B(t) from both sides of the last equation, dividing by h, and
letting h decrease to 0, yields a diﬀerential equation at times t > 0 :
B
(t) = B(t) δ(t) +D(t) , A(0) = α
0
(1.12)
The method of solution of (1.12), which is the standard one from diﬀerential
equations theory of multiplying through by an integrating factor, again has
a natural interpretation in terms of present values. The integrating factor
1/A(t) = exp(−
_
t
0
δ(s) ds) is the present value at time 0 of a payment of
1 at time t, and the quantity B(t)/A(t) = G(t) is then the present value
of the deposit stream of α
0
at time 0 followed by continuous deposits at
rate D(t). The ratiorule of diﬀerentiation yields
G
(t) =
B
(t)
A(t)
−
B(t) A
(t)
A
2
(t)
=
B
(t) − B(t) δ(t)
A(t)
=
D(t)
A(t)
1.3. EXERCISE SET 1 25
where the substitution A
(t)/A(t) ≡ δ(t) has been made in the third ex
pression. Since G(0) = B(0) = α
0
, the solution to the diﬀerential equation
(1.12) becomes
G(t) = α
0
+
_
t
0
D(s)
A(s)
ds , B(t) = A(t) G(t)
Finally, the formula can be specialized to the case of a constant unitrate
payment stream ( D(x) = 1, δ(x) = δ = ln(1 + i), 0 ≤ x ≤ T ) with
no initial deposit (i.e., α
0
= 0). By the preceding formulas, A(t) =
exp(t ln(1 +i)) = (1 +i)
t
, and the present value of such a payment stream
is
_
T
0
1 · exp(−t ln(1 +i)) dt =
1
δ
_
1 −(1 +i)
−T
_
Recall that the force of interest δ = ln(1 +i) is the limiting value obtained
from the nominal interest rate i
(m)
using the diﬀerencequotient representa
tion:
lim
m→∞
i
(m)
= lim
m→∞
exp((1/m) ln(1 +i)) − 1
1/m
= ln(1 +i)
The present value of a payment at time T in the future is then
_
1 +
i
(m)
m
_
−mT
= (1 +i)
−T
= exp(−δ T)
1.3 Exercise Set 1
The ﬁrst homework set covers the basic deﬁnitions in two areas:
(i) probability as it relates to events deﬁned from cohort lifetables, including
the theoretical machinery of population and conditional survival, distribu
tion, and density functions and the deﬁnition of expectation; (ii) the theory
of interest and present values, with special reference to the idea of income
streams of equal value at a ﬁxed rate of interest.
(1). For how long a time should $100 be left to accumulate at 5% interest
so that it will amount to twice the accumulated value (over the same time
period) of another $100 deposited at 3% ?
26 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
(2). Use a calculator or computer to answer the following numerically:
(a) Suppose you sell for $6,000 the right to receive for 10 years the amount
of $1,000 per year payable quarterly (starting at the end of the ﬁrst quarter).
What eﬀective rate of interest makes this a fair sale price ? (You will have
to solve numerically or graphically, or interpolate a tabulation, to ﬁnd it.)
(b) $100 deposited 20 years ago has grown at interest to $235. The
interest was compounded twice a year. What were the nominal and eﬀective
interest rates ?
(c) How much should be set aside (the same amount each year) at the
beginning of each year for 10 years to amount to $1000 at the end of the 10th
year at the interest rate of part (b) ?
In the following problems, S(t) denotes the probability for a newborn
in a designated population to survive to exact age t . If a cohort life table
is under discussion, then the probability distribution relates to a randomly
chosen member of the newborn cohort.
(3). Assume that a population’s survival probability function is given by
S(t) = 0.1(100 −t)
1/2
, for 0 ≤ t ≤ 100.
(a) Find the probability that a life aged 0 will die between exact ages 19
and 36.
(b) Find the probability that a life aged 36 will die before exact age 51.
(4). For members of the poulation in Problem (3),
(a) Find the expected age at death of a newborn (life aged 0).
(b) Find the expected age at death of a life aged 20.
(5). Use the Illustrative Lifetable (Table 1.1) to calculate the following
probabilities. (In each case, assume that the indicated span of years runs
from birthday to birthday.) Find the probability
(a) that a life aged 26 will live at least 30 more years;
(b) that a life aged 22 will die between ages 45 and 55;
(c) that a life aged 25 will die either before age 50 or after age 70.
1.3. EXERCISE SET 1 27
(6). In a special population, you are given the following facts:
(i) The probability that two independent lives, respectively aged 25 and
45, both survive 20 years is 0.7.
(ii) The probability that a life aged 25 will survive 10 years is 0.9.
Then ﬁnd the probability that a life aged 35 will survive to age 65.
(7). Suppose that you borrowed $1000 at 6% eﬀective rate, to be repaid
in 5 years in a lump sum, and that after holding the money idle for 1 year
you invested the money and earned 8% eﬀective for theremaining four years.
What is the eﬀective interest rate you earned (ignoring interest costs) over 5
years on the $1000 which you borrowed ? Taking interest costs into account,
what is the present value of your proﬁt over the 5 years of the loan ? Also
redo the problem if instead of repaying all principal and interest at the end
of 5 years, you must make a payment of accrued interest at the end of 3
years, with the additional interest and principal due in a single lumpsum at
the end of 5 years.
(8). Find the total present value at 5% APR of payments of $1 at the end
of 1, 3, 5, 7, and 9 years and payments of $2 at the end of 2, 4, 6, 8, and 10
years.
(9). Find the present value at time 0 at a 6% eﬀective interest rate of a
series payments of 100 at times 1, 2, 3 and of 300 at times 6, 7, 8.
(10). Find the present value at time 0 of payments of 100 at ten successive
times 1, 2, . . . , 10 if the instanteous eﬀective interest rate applying at all
times t in the time interval [0, 10] is r(t) = .07 −(.002)t.
(11). Find the internal rate of return (i.e., the equivaent constant eﬀective
interest rate) over the time interval [0, 7] of an investment which pays bank
interest of 4% at times in [0, 5] if you make deposits of 1000 at each of
the times t = 0, 2, 4, if the interest rate earned on the time interval [5, 7]
is 6%, and if the total balance is withdrawn at time 7.
(12). (i) Find the payment amount K such that a loan of 10, 000 at a 7%
eﬀective annual interest rate is repaid in exactly three payments consisting
of an amount K at times 1 and 3 years and of 2K at 5 years.
28 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
(ii) After ﬁnding K in part (i), decompose each of the three loan repay
ment amounts K, K, 2K at respective times, 1, 3, 5 into their principal
and interest portions.
1.4 Worked Examples
Example 1. How many years does it take for money to triple in value at
interest rate i ?
The equation to solve is 3 = (1 + i)
t
, so the answer is ln(3)/ ln(1 +i),
with numerical answer given by
t =
_
_
_
22.52 for i = 0.05
16.24 for i = 0.07
11.53 for i = 0.10
Example 2. Suppose that a sum of $1000 is borrowed for 5 years at 5%,
with interest deducted immediately in a lump sum from the amount borrowed,
and principal due in a lump sum at the end of the 5 years. Suppose further
that the amount received is invested and earns 7%. What is the value of the
net proﬁt at the end of the 5 years ? What is its present value (at 5%) as
of time 0 ?
First, the amount received by the borrower at time 0 is
1000 (1 − d)
5
= 1000/(1.05)
5
= 783.53, where d = .05/1.05, since the
amount received should compound to precisely the principal of 1000 at 5%
interest in 5 years. Next, the compounded value of 783.53 for 5 years at
7% is 783.53 (1.07)
5
= 1098.94, so the net proﬁt at the end of 5 years, after
paying oﬀ the principal of 1000, is 98.94. The present value of the proﬁt
ought to be calculated with respect to the ‘going rate of interest’, which in
this problem is presumably the rate of 5% at which the money is borrowed,
so is 98.94/(1.05)
5
= 77.52.
Example 3. For the following small cohort lifetable (ﬁrst 3 columns) with 5
agecategories, ﬁnd the probabilities for all values of [T], both uncondition
ally and conditionally for lives aged 2, and ﬁnd the expectation of both [T]
and (1.05)
−[T]−1
.
1.4. WORKED EXAMPLES 29
The basic information in the table is the ﬁrst column l
x
of numbers
surviving. Then d
x
= l
x
− l
x+1
for x = 0, 1, . . . , 4. The random variable
T is the lifelength for a randomly selected individual from the age=0 cohort,
and therefore Pr([T] = x) = Pr(x ≤ T < x + 1) = d
x
/l
0
. The conditional
probabilities given survivorship to agecategory 2 are simply the ratios with
numerator d
x
for x ≥ 2 , and with denominator l
2
= 65.
x l
x
d
x
Pr([T] = x) Pr([T] = xT ≥ 2) 1.05
−x−1
0 100 20 0.20 0 0.95238
1 80 15 0.15 0 0.90703
2 65 10 0.10 0.15385 0.86384
3 55 15 0.15 0.23077 0.82770
4 40 40 0.40 0.61538 0.78353
5 0 0 0 0 0.74622
In terms of the columns of this table, we evaluate from the deﬁnitions and
formula (1.3)
E([T]) = 0 · (0.20) + 1 · (0.15) + 2 · (0.10) + 3 · (0.15) + 4 · (0.40) = 2.4
E([T]  T ≥ 2) = 2 · (0.15385) + 3 · (0.23077) + 4 · (0.61538) = 3.4615
E(1.05
−[T]−1
) = 0.95238 · 0.20 + 0.90703 · 0.15 + 0.86384 · 0.10 +
+0.8277 · 0.15 + 0.78353 · 0.40 = 0.8497
The expectation of [T] is interpreted as the average per person in the cohort
lifetable of the number of completed whole years before death. The quantity
(1.05)
−[T]−1
can be interpreted as the present value at birth of a payment
of 1 to be made at the end of the year of death, and the ﬁnal expectation
calculated above is the average of that presentvalue over all the individuals
in the cohort lifetable, if the going rate of interest is 5%.
Example 4. Suppose that the deathrates q
x
= d
x
/l
x
for integer ages x in
a cohort lifetable follow the functional form
q
x
=
_
4 · 10
−4
for 5 ≤ x < 30
8 · 10
−4
for 30 ≤ x ≤ 55
between the ages x of 5 and 55 inclusive. Find analytical expressions for
S(x), l
x
, d
x
at these ages if l
0
= 10
5
, S(5) = .96.
30 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
The key formula expressing survival probabilities in terms of deathrates
q
x
is:
S(x + 1)
S(x)
=
l
x+1
l
x
= 1 −q
x
or
l
x
= l
0
· S(x) = (1 −q
0
)(1 −q
1
) · · · (1 −q
x−1
)
So it follows that for x = 5, . . . , 30,
S(x)
S(5)
= (1 −.0004)
x−5
, l
x
= 96000 · (0.9996)
x−5
so that S(30) = .940446, and for x = 31, . . . , 55,
S(x) = S(30) · (.9992)
x−30
= .940446 (.9992)
x−30
The deathcounts d
x
are expressed most simply through the preceding
expressions together with the formula d
x
= q
x
l
x
.
1.5. USEFUL FORMULAS 31
1.5 Useful Formulas from Chapter 1
S(x) =
l
x
l
0
, d
x
= l
x
−l
x+1
p. 1
P(x ≤ T < x +k) = S(x) −S(x +k) =
l
x
−l
x+k
l
x
p. 2
f(t) = −S
(t) , S(y) −S(y +t) =
_
y+t
y
f(s) ds
p. 5
E
_
g(T)
¸
¸
¸ T ≥ x
_
=
1
S(x)
_
∞
x
g(t) f(t) dt
p. 9
1 +i = 1 +i
eﬀ
=
_
1 +
i
(m)
m
_
m
=
_
1 −
d
(m)
m
_
−m
= e
δ
p. 18
32 CHAPTER 1. BASICS OF PROBABILITY & INTEREST
Chapter 2
Theory of Interest and
Force of Mortality
The parallel development of Interest and Probability Theory topics continues
in this Chapter. For application in Insurance, we are preparing to value
uncertain payment streams in which times of payment may also be uncertain.
The interest theory allows us to express the present values of deterministic or
certain payment streams compactly, while the probability material prepares
us to ﬁnd and interpret average or expected values of present values expressed
as functions of random lifetime variables.
This installment of the course covers: (a) further formulas and topics in
the pure (i.e., nonprobabilistic) theory of interest, and (b) more discussion
of lifetime random variables, in particular of force of mortality or hazard
rates, and theoretical families of life distributions.
2.1 More on Theory of Interest
In this Section, we deﬁne notations and ﬁnd compact formulas for present
values of some standard payment streams. To this end, newly deﬁned pay
ment streams are systematically expressed in terms of previously considered
ones. There are two primary methods of manipulating one paymentstream
to give another for the convenient calculation of present values:
33
34 CHAPTER 2. INTEREST & FORCE OF MORTALITY
• First, if one paymentstream can be obtained from a second one pre
cisely by delaying all payments by the same amount t of time, then
the present value of the ﬁrst one is v
t
multiplied by the present value
of the second.
• Second, if a paymentstream A can be obtained as the superposition of
payment streams B and C, i.e., can be obtained by paying the sum of
the timed payment amounts deﬁning the streams B and C, then the
present value of stream A is the sum of the present values of B and C.
The following subsection contains several useful applications of these meth
ods. For another simple illustration, see Worked Example 2 at the end of the
Chapter.
2.1.1 Annuities & Actuarial Notation
The general present value formulas above will now be specialized to the case
of constant (instantaneous) interest rate r(t) ≡ i ≡ e
δ
at all times t ≥ 0,
and some very particular streams of payments s
j
at times t
j
, related
to periodic premium and annuity payments. The eﬀective interest rate is
always denoted by i = i
eﬀ
, and as before the mtimesperyear equivalent
nominal interest rate is denoted by i
(m)
. Also, from now on the standard
and convenient notation
v ≡ 1/(1 +i) = 1 /
_
1 +
i
(m)
m
_
m
will be used for the present value of a payment of 1 one year later.
(i) If s
0
= 0 and s
1
= · · · = s
nm
= 1/m in the discrete setting, where
m ≥ 1 denotes the number of payments per year, and t
j
= j/m, then the
paymentstream is called an immediate annuity, and its present value G
n
is given the notation a
(m)
n
and is equal, by the geometricseries summation
formula, to
m
−1
nm
j=1
_
1 +
i
(m)
m
_
−j
=
_
1 +
i
(m)
m
_
−1
1 −(1 +i
(m)
/m)
−nm
m(1 −(1 +i
(m)
/m)
−1
)
2.1. MORE ON THEORY OF INTEREST 35
which shows that
a
(m)
n
=
1 −((1 +i
(m)
/m)
−m
)
n
m(1 +i
(m)
/m−1)
=
1 −v
n
i
(m)
(2.1)
All of these immediate annuity values, for ﬁxed v, n but varying m, are
roughly comparable because all involve a total payment of 1 per year.
Formula (2.1) shows that all of the values a
(m)
n
diﬀer only through the factors
i
(m)
, which diﬀer by only a few percent for varying m and ﬁxed i, as shown
in Table 2.1. Recall from formula (1.9) that i
(m)
= m{(1 +i)
1/m
−1}.
If, instead of the payment stream deﬁning the immediate annuity, s
0
=
1/m but s
nm
= 0, then nm deposits of 1/m are made at an arithmetic
progression of times from 1/m to n inclusive, and the present value notation
changes to ¨a
(m)
n
. The payment stream is then called an annuitydue, and
the present valueis given by any of the equivalent formulas
¨a
(m)
n
= (1 +
i
(m)
m
) a
(m)
n
=
1 −v
n
m
+ a
(m)
n
=
1
m
+a
(m)
n−1/m
(2.2)
The ﬁrst of these formulas recognizes the annuitydue paymentstream as
identical to the annuityimmediate paymentstream shifted earlier by the
time 1/m and therefore worth more by the accumulationfactor (1+i)
1/m
=
1 +i
(m)
/m. The third expression in (2.2) represents the annuitydue stream
as being equal to the annuityimmediate stream with the payment of 1/m
at t = 0 added and the payment of 1/m at t = n removed. The ﬁnal
expression says that if the time0 payment is removed from the annuitydue,
the remaining stream coincides with the annuityimmediate stream consisting
of nm−1 (instead of nm) payments of 1/m.
In the limit as m → ∞ for ﬁxed n, the notation a
n
denotes the
continuous annuity, that is, the present value of an annuity paid instanta
neously at constant unit rate, with the limiting nominal interestrate which
was shown in the previous chapter to be lim
m
i
(m)
= i
(∞)
= δ. The limiting
behavior of the nominal interest rate can be seen rapidly from the formula
i
(m)
= m
_
(1 +i)
1/m
− 1
_
= δ ·
exp(δ/m) −1
δ/m
since (e
z
−1)/z converges to 1 as z →0. Then by (2.1) and (2.2),
a
n
= lim
m→∞
¨ a
(m)
n
= lim
m→∞
a
(m)
n
=
1 −v
n
δ
(2.3)
36 CHAPTER 2. INTEREST & FORCE OF MORTALITY
Table 2.1: Values of nominal interest rates i
(m)
(upper number) and
d
(m)
(lower number), for various choices of eﬀective annual interest rate
i and number m of compounding periods per year.
i = .02 .03 .05 .07 .10 .15
m = 2 .0199 .0298 .0494 .0688 .0976 .145
.0197 .0293 .0482 .0665 .0931 .135
3 .0199 .0297 .0492 .0684 .0968 .143
.0197 .0294 .0484 .0669 .0938 .137
4 .0199 .0297 .0491 .0682 .0965 .142
.0198 .0294 .0485 .0671 .0942 .137
6 .0198 .0296 .0490 .0680 .0961 .141
.0198 .0295 .0486 .0673 .0946 .138
12 .0198 .0296 .0489 .0678 .0957 .141
.0198 .0295 .0487 .0675 .0949 .139
Remark 2.1 The deﬁnition and formulas for the immediate annuity a
(m)
n
and the annuitydue ¨a
(m)
n
remain valid if nm but not necessarily n itself
is an integer. In the limit as m → ∞, the continuous annuity deﬁnition
a
n
and formula remain valid with any positive real number n. 2
A handy formula for annuitydue present values follows easily by recalling
that
1 −
d
(m)
m
=
_
1 +
i
(m)
m
_
−1
implies d
(m)
=
i
(m)
1 +i
(m)
/m
Then, by (2.2) and (2.1),
¨ a
(m)
n
= (1 −v
n
) ·
1 + i
(m)
/m
i
(m)
=
1 −v
n
d
(m)
(2.4)
In case m is 1, the superscript
(m)
is omitted from all of the annuity
notations. In the limit where n → ∞, the notations become a
(m)
∞
and
¨ a
(m)
∞
, and the annuities are called perpetuities (respectively immediate and
due) with presentvalue formulas obtained from (2.1) and (2.4) as:
a
(m)
∞
=
1
i
(m)
, ¨a
(m)
∞
=
1
d
(m)
(2.5)
2.1. MORE ON THEORY OF INTEREST 37
We now build some more general annuityrelated present values out of
the standard functions a
(m)
n
and ¨a
(m)
n
.
(ii). Consider ﬁrst the case of the increasing perpetual annuitydue,
denoted (I
(m)
¨ a)
(m)
∞
, which is deﬁned as the present value of a stream of
payments (k +1)/m
2
at times k/m, for k = 0, 1, . . . forever. Clearly the
present value is
(I
(m)
¨ a)
(m)
∞
=
∞
k=0
m
−2
(k + 1)
_
1 +
i
(m)
m
_
−k
Here are two methods to sum this series, the ﬁrst purely mathematical, the
second based on actuarial intuition. First, without worrying about the strict
justiﬁcation for diﬀerentiating an inﬁnite series termbyterm,
∞
k=0
(k + 1) x
k
=
d
dx
∞
k=0
x
k+1
=
d
dx
x
1 −x
= (1 −x)
−2
for 0 < x < 1, where the geometricseries formula has been used to sum
the second expression. Therefore, with x = (1 + i
(m)
/m)
−1
and 1 − x =
(i
(m)
/m)/(1 +i
(m)
/m),
(I
(m)
¨a)
(m)
∞
= m
−2
_
i
(m)
/m
1 +i
(m)
/m
_
−2
=
_
1
d
(m)
_
2
=
_
¨a
(m)
∞
_
2
and (2.5) has been used in the last step. Another way to reach the same result
is to recognize the increasing perpetual annuitydue as 1/m multiplied by
the superposition of perpetuitiesdue ¨ a
(m)
∞
paid at times 0, 1/m, 2/m, . . . ,
and therefore its present value must be ¨ a
(m)
∞
· ¨a
(m)
∞
. As an aid in recognizing
this equivalence, consider each annuitydue ¨ a
(m)
∞
paid at a time j/m as
being equivalent to a stream of payments 1/m at time j/m, 1/m at
(j + 1)/m, etc. Putting together all of these payment streams gives a total
of (k+1)/m paid at time k/m, of which 1/m comes from the annuitydue
starting at time 0, 1/m from the annuitydue starting at time 1/m, up
to the payment of 1/m from the annuitydue starting at time k/m.
(iii). The increasing perpetual annuityimmediate (I
(m)
a)
(m)
∞
—
the same payment stream as in the increasing annuitydue, but deferred by
a time 1/m — is related to the perpetual annuitydue in the obvious way
(I
(m)
a)
(m)
∞
= v
1/m
(I
(m)
¨ a)
(m)
∞
= (I
(m)
¨a)
(m)
∞
_
(1 +i
(m)
/m) =
1
i
(m)
d
(m)
38 CHAPTER 2. INTEREST & FORCE OF MORTALITY
(iv). Now consider the increasing annuitydue of ﬁnite duration
n years. This is the present value (I
(m)
¨a)
(m)
n
of the paymentstream of
(k +1)/m
2
at time k/m, for k = 0, . . . , nm−1. Evidently, this payment
stream is equivalent to (I
(m)
¨ a)
(m)
∞
minus the sum of n multiplied by an
annuitydue ¨ a
(m)
∞
starting at time n together with an increasing annuity
due (I
(m)
¨a)
(m)
∞
starting at time n. (To see this clearly, equate the payments
0 = (k + 1)/m
2
− n ·
1
m
− (k − nm + 1)/m
2
received at times k/m for
k ≥ nm.) Thus
(I
(m)
¨a)
(m)
n
= (I
(m)
¨ a)
(m)
∞
_
1 −v
n
_
−n¨ a
(m)
∞
v
n
= ¨ a
(m)
∞
_
¨ a
(m)
∞
−v
n
_
¨a
(m)
∞
+n
_ _
= ¨ a
(m)
∞
_
¨ a
(m)
n
− nv
n
_
where, in the last line, recall that v = (1 +i)
−1
= (1 +i
(m)
/m)
−m
and that
¨ a
(m)
n
= ¨ a
(m)
∞
(1 − v
n
). The latter identity is easy to justify either by the
formulas (2.4) and (2.5) or by regarding the annuitydue payment stream as a
superposition of the paymentstream up to time n−1/m and the payment
stream starting at time n. As an exercise, ﬁll in details of a second, intuitive
veriﬁcation, analogous to the second veriﬁcation in pargraph (ii) above.
(v). The decreasing annuity (D
(m)
¨a)
(m)
n
is deﬁned as (the present
value of) a stream of payments starting with n/m at time 0 and decreasing
by 1/m
2
after every timeperiod of 1/m, with no further payments at or
after time n. The easiest way to obtain the present value is through the
identity
(I
(m)
¨a)
(m)
n
+ (D
(m)
¨ a)
(m)
n
= (n +
1
m
) ¨ a
(m)
n
(2.6)
Again, as usual, the method of proving this is to observe that in the payment
stream whose present value is given on the lefthand side, the payment
amount at each of the times j/m, for j = 0, 1, . . . , nm−1, is
j + 1
m
2
+ (
n
m
−
j
m
2
) =
1
m
(n +
1
m
)
2.1. MORE ON THEORY OF INTEREST 39
2.1.2 Loan Repayment: Mortgage, Bond, Sinking Fund
Perhaps the most common application of interest theory is the calculation
of the payment amounts needed to repay a loan according to a few standard
repayment plans. In ordinary consumer purchases or longterm ﬁxerdrate
loans on the purchase of a house, the usual repayment plan is a series of level
or equal payments made m times yearly (usually with m = 12, and usually
with ﬁrst payment at time 0) for a total duration of n years, so that at the
last payment (the nm’th payment, at time n −1/m if the ﬁrst payment was
made at time 0) the loan has been completely paid oﬀ. We refer to this kind
of repayment schedule with level payments as a mortgage loan.
1
Recall that the present value of a payment stream of amount c per year,
with c/m paid at times 1/m, 2/m, . . . , n − 1/m, n/m, is c a
(m)
n
. Thus,
if an amount L has been borrowed for a term of n years, to be repaid by
equal installments at the end of every period 1/m , at ﬁxed nominal interest
rate i
(m)
, then the level payment c/m or installment payment amount
is obtained by equating L = c a
(m)
n
Mortgage Payment = L/(ma
(m)
n
) =
Li
(m)
m(1 −v
n
)
(2.7)
where v = 1/(1 +i) = (1 +i
(m)
/m)
−m
.
A second kind of plan to repay a loan is to pay equal amounts to cover only
the interest amounts accrued every 1/m year on the principal for a duration
of n years, with ﬁrst payment made at time 1/m and last at time n, with the
principal (the original amount borrowed) also repaid in a lump sum at time
n. This arrangement is used by corporations or government agencies which
issue bonds: the borrowing agency receives the loan amount from investors
at time 0, called the face amount of the bond, and regularly issues interest
payments after every period of 1/m year (usually with m = 2 or 4), and
ﬁnally repays or redeems the face or principal amount of the bond at the end
of the n year term of the bond. (The regular interest payments used to be
called coupon payments because of small paper coupons attached to the paper
bond document, and which the investor would regularly redeem at a bank, at
scheduled times, for the interest payment amount.) The reader should look
1
In legal and historical terms, ‘mortgage’ refers to the way in which the promise to
repay is secured by the house or other property purchased with the amount borrowed.
40 CHAPTER 2. INTEREST & FORCE OF MORTALITY
back to Section 1.2.3 where loan repayment amounts were formally broken
down into interest and principal portions, in order to conﬁrm the sense of this
bond repayment plan. Since each payment by the borrower at times k/m
for k = 1, 2, . . . , nm (apart from the ﬁnal lump sum principal repayment)
consists of interest only (equal to the original face amount multiplied by
(1 + i)
1/m
− 1 = i
(m)
/m), each of the intermediate payments contains 0
principal portion, and the result of Exercise 1.A in Sec. 1.2.3 shows that the
original face amount or principal is also the principal or balance owed on the
loan just after each interest payment, up to the ﬁnal redemption when the
principal is paid.
Now a corporation or governmental agency which borrows money from
investors by issuing a bond, will often obligate itself through a formal legal
arrangement to devote a certain category of income to a socalled sinking
fund, an investment account maintained by a trustee. Apart from the bond
interest payments at a contractual eﬀective interst rate i made directly to
investors (the lenders), the borrowing agency will also pay regular amounts
at intervals of 1/m
year to the sinking fund trustee for the same term of n
years as the bond, with the intention that the sinking fund will accumulate
at its own possibly diﬀerent investment interest rate i
to the principal or face
amount of the issued bond at time n, at which time the principal is repaid
directly to the bond investors. At (or just after) an intermediate times k/m,
the amount built up in the sinking fund is referred to as a reserve toward
the ultimate redemption of the principal of the bond. We will see, later in
this book, that insurance reserves generalize these deterministic reserves to
Insurances, where the future payouts are not determinstic but rather contin
gent on the mortality experience of the portfolio of insured lives.
Remark 2.2 Note that, if a borrowed amount L is repaid by regular interest
payments and a sinking fund, and if the number of payments per year into
the fund is m
= m and the eﬀective interest rate for the sinking fund
is i
= i, then evidently the sum of the regular mtimesyearly interest and
sinking fund payment at each time k/m is precisely the same as the mortgage
payment (2.7). For m
= m or i
= i, a separate calculation is needed,
leading to several exercises at the end of the Chapter. The following Exercise
with sketched solution gives an example.
2.1. MORE ON THEORY OF INTEREST 41
Exercise 2.A. A small city issues a bond for ten million dollars for ten years
at 4% nominal quarterly interest (m = 4) and creates a sinking fund into
which it will make annual deposits (m
= 1), from tax receipts, on which its
ﬁnancial advisors claimit can safely earn an eﬀective annual rate of i
= .055.
Find the amount of the level annual sinking fund deposit.
Solution of Exercise 2.A. Since the interest payments are made at the con
tractual interest rate i (corresponding to i
(4)
= .04), the sinking fund must
appreciate to the loan amount of L = 10
7
at time t = 10. Therefore the
sinking fund present value at time 10 (equal to its present value at t = 0
accumulated by the factor (1 +i
)
10
must be equated to L = 10
7
. So the
annual sinking fund deposit D is found through the equality
L = 10
7
= (1 +i
)
10
· Da
10
= D
1 −(1.055)
−10
.055
1.055
10
where a
indicates that the annuity is calculated at interest rate i
.
2.1.3 Loan Amortization & Mortgage Reﬁnancing
We analyze next the breakdown between principal and interest in repaying
a mortgage loan by level payments (2.7). Of the payment made at time
(k +1)/m, how much can be attributed to interest and how much to princi
pal ? Consider the present value at time 0 of the debt for a unit (L = 1)
loan amount less the accumulated amounts paid through time k/m :
1 − (ma
(m)
k/m
) / (ma
(m)
n
) = 1 −
1 −v
k/m
1 −v
n
=
v
k/m
−v
n
1 −v
n
The remaining debt, per unit of loan amount, valued just after time k/m,
is denoted from now on by B
n, k/m
. It is greater than the displayed present
value at 0 by a factor (1 +i)
k/m
, so is equal to
B
n, k/m
= (1 +i)
k/m
v
k/m
−v
n
1 −v
n
=
1 −v
n−k/m
1 −v
n
(2.8)
The amount of interest for a loan amount of 1 after time 1/m is
(1 +i)
1/m
−1 = i
(m)
/m. Therefore the interest included in the payment at
42 CHAPTER 2. INTEREST & FORCE OF MORTALITY
(k + 1)/m is i
(m)
/m multiplied by the value B
n, k/m
of outstanding debt
just after k/m. Thus the next total payment of i
(m)
/(m(1 −v
n
)) consists
of the two parts
Amount of interest = m
−1
i
(m)
(1 −v
n−k/m
)/(1 −v
n
)
Amount of principal = m
−1
i
(m)
v
n−k/m
/(1 −v
n
)
By deﬁnition, the principal included in each payment is the amount of the
payment minus the interest included in it. These formulas show in particular
that the amount of principal repaid in each successive payment increases
geometrically in the payment number, which at ﬁrst seems surprising. Note
as a check on the displayed formulas that the outstanding balance B
n,(k+1)/m
immediately after time (k + 1)/m is recomputed as B
n, k/m
minus the
interest paid at (k + 1)/m, or
1 −v
n−k/m
1 −v
n
−
i
(m)
m
v
n−k/m
1 −v
n
=
1 −v
n−k/m
(1 +i
(m)
/m)
1 −v
n
=
1 −v
n−(k+1)/m
1 −v
n
=
_
1 − a
(m)
(k+1)/m
_
a
(m)
n
_
v
−(k+1)/m
(2.9)
as was derived above by considering the accumulated value of amounts paid.
The general deﬁnition of the principal repaid in each payment is the excess
of the payment over the interest since the past payment on the total balance
due immediately following that previous payment.
2.1.4 Illustration on Mortgage Reﬁnancing
Suppose that a 30–year, nominalrate 8%, $100, 000 mortgage payable
monthly is to be reﬁnanced at the end of 8 years for an additional 15 years
(instead of the 22 which would otherwise have been remaining to pay it
oﬀ) at 6%, with a reﬁnancing closingcost amount of $1500 and 2 points.
(The points are each 1% of the reﬁnanced balance including closing costs,
and costs plus points are then extra amounts added to the initial balance
of the reﬁnanced mortgage.) Suppose that the new pattern of payments is
to be valued at each of the nominal interest rates 6%, 7%, or 8%, due
to uncertainty about what the interest rate will be in the future, and that
these valuations will be taken into account in deciding whether to take out
the new loan.
2.1. MORE ON THEORY OF INTEREST 43
The monthly payment amount of the initial loan in this example was
$100, 000(.08/12)/(1 −(1 +.08/12)
−360
) = $733.76, and the present value as
of time 0 (the beginning of the old loan) of the payments made through the
end of the 8
th
year is ($733.76) · (12a
(12)
8
) = $51, 904.69. Thus the present
value, as of the end of 8 years, of the payments still to be made under the
old mortgage, is $(100, 000 −51, 904.69)(1 +.08/12)
96
= $91, 018.31. Thus,
if the loan were to be reﬁnanced, the new reﬁnanced loan amount would be
$91, 018.31 + 1, 500.00 = $92, 518.31. If 2 points must be paid in order to
lock in the rate of 6% for the reﬁnanced 15year loan, then this amount
is (.02)92518.31 = $1850.37 . The new principal balance of the reﬁnanced
loan is 92518.31 +1850.37 = $94, 368.68, and this is the present value at a
nominal rate of 6% of the future loan payments, no matter what the term of
the reﬁnanced loan is. The new monthly payment (for a 15year duration) of
the reﬁnanced loan is $94, 368.68(.06/12)/(1 −(1 +.06/12)
−180
) = $796.34.
For purposes of comparison, what is the present value at the current
going rate of 6% (nominal) of the continuing stream of payments under
the old loan ? That is a 22year stream of monthly payments of $733.76,
as calculated above, so the present value at 6% is $733.76 · (12a
(12)
22
) =
$107, 420.21. Thus, if the new rate of 6% were really to be the correct
one for the next 22 years, and each loan would be paid to the end of its
term, then it would be a ﬁnancial disaster not to reﬁnance. Next, suppose
instead that right after reﬁnancing, the economic rate of interest would be
a nominal 7% for the next 22 years. In that case both streams of payments
would have to be revalued — the one before reﬁnancing, continuing another
22 years into the future, and the one after reﬁnancing, continuing 15 years
into the future. The respective present values (as of the end of the 8
th
year) at nominal rate of 7% of these two streams are:
Old loan: 733.76 (12a
(12)
22
) = $98, 700.06
New loan: 796.34 (12a
(12)
15
) = $88, 597.57
Even with these diﬀerent assumptions, and despite closingcosts and points,
it is well worth reﬁnancing.
44 CHAPTER 2. INTEREST & FORCE OF MORTALITY
Exercise 2.B. Suppose that you can forecast that you will in fact sell your
house in precisely 5 more years after the time when you are reﬁnancing. At
the time of sale, you would pay oﬀ the cash principal balance, whatever it
is. Calculate and compare the present values (at each of 6%, 7%, and 8%
nominal interest rates) of your payment streams to the bank, (a) if you
continue the old loan without reﬁnancing, and (b) if you reﬁnance to get
a 15year 6% loan including closing costs and points, as described above.
2.1.5 Computational illustration in R
All of the calculations described above are very easy to program in any lan
guage from Fortran to Mathematica, and also on a programmable calculator;
but they are also very handily organized within a spreadsheet, which seems
to be the way that MBA’s, bankoﬃcials, and actuaries will learn to do them
from now on.
In this section, an R function (cf. Venables & Ripley 2002) is provided to
do some comparative reﬁnancing calculations. Concerning the syntax of R,
the only explanation necessary at this point is that * denotes multiplication,
and
∧
denotes exponentiation.
The function RefExmp given below calculates mortgage payments, bal
ances for purposes of reﬁnancing both before and after application of ad
ministrative costs and points, and the present value under any interest rate
(not necessarily the ones at which either the original or reﬁnanced loans are
taken out) of the stream of repayments to the bank up to and including the
lumpsum payoﬀ which would be made, for example, at the time of selling
the house on which the mortgage loan was negotiated. The output of the
function is a list which, in each numerical example below, is displayed in
‘unlisted’ form, horizontally as a vector. Lines beginning with the symbol #
are commentlines.
The outputs of the function are as follows. Oldpayment is the monthly
payment on the original loan of faceamount Loan at nominal interest i
(12)
=
OldInt for a term of OldTerm years. NewBal is the balance B
n, k/m
of for
mula (2.8) for n = OldTerm, m = 12, and k/m = RefTim, and the
reﬁnanced loan amount is a multiple 1+ Points of NewBal, which is equal
to RefBal + Costs. The new loan, at nominal interest rate NewInt, has
2.1. MORE ON THEORY OF INTEREST 45
R FUNCTION CALCULATING REFINANCE PAYMENTS & VALUES
RefExmp
function(Loan, OldTerm, RefTim, NewTerm, Costs, Points,
PayoffTim, OldInt, NewInt, ValInt)
{
# Function calculates present value of future payment stream
# underrefinanced loan.
# Loan = original loan amount;
# OldTerm = term of initial loan in years;
# RefTim = time in years after which to refinance;
# NewTerm = term of refinanced loan;
# Costs = fixed closing costs for refinancing;
# Points = fraction of new balance as additional costs;
# PayoffTim (no bigger than NewTerm) = time (from refinancing
# time at which new loan balance is to be paid off in
# cash (eg at house sale);
# The three interest rates OldInt, NewInt, ValInt are
# nominal 12timesperyear, and monthly payments
# are calculated.
vold = (1 + OldInt/12)^(12)
Oldpaymt = ((Loan * OldInt)/12)/(1  vold^OldTerm)
NewBal = (Loan * (1  vold^(OldTerm  RefTim)))/
(1  vold^OldTerm)
RefBal = (NewBal + Costs) * (1 + Points)
vnew = (1 + NewInt/12)^(12)
Newpaymt = ((RefBal * NewInt)/12)/(1  vnew^NewTerm)
vval = (1 + ValInt/12)^(12)
Value = (Newpaymt * 12 * (1  vval^PayoffTim))/ValInt +
(RefBal * vval^PayoffTim * (1  vnew^(NewTerm 
PayoffTim)))/(1  vnew^NewTerm)
list(Oldpaymt = Oldpaymt, NewBal = NewBal,
RefBal = RefBal, Newpaymt = Newpaymt, Value = Value)
}
46 CHAPTER 2. INTEREST & FORCE OF MORTALITY
monthly payments Newpaymt for a term of NewTerm years. The loan is to
be paid oﬀ PayoﬀTim years after RefTim when the new loan commences,
and the ﬁnal output of the function is the present value at the start of the
reﬁnanced loan with nominal interest rate ValInt of the stream of payments
made under the reﬁnanced loan up to and including the lump sum payoﬀ.
We begin our numerical illustration by reproducing the quantities calcu
lated in the previous subsection:
> unlist(RefExmp(100000, 30, 8, 15, 1500, 0.02, 15,
0.08, 0.06, 0.06))
Oldpaymt NewBal RefBal Newpaymt Value
733.76 91018 94368 796.33 94368
Note that, since the payments under the new (reﬁnanced) loan are here
valued at the same interest rate as the loan itself, the present value Value of
all payments made under the loan must be equal to the the reﬁnanced loan
amount RefBal.
The comparisons of the previous Section between the original and reﬁ
nanced loans, at (nominal) interest rates of 6, 7, and 8 %, are all recapitulated
easily using this function. To use it, for example, in valuing the old loan at
7%, the arguments must reﬂect a ‘reﬁnance’ with no costs or points for a
period of 22 years at nominal rate 6%, as follows:
> unlist(RefExmp(100000,30,8,22,0,0,22,0.08,0.08,0.07))
Oldpaymt NewBal RefBal Newpaymt Value
733.76 91018 91018 733.76 98701
(The small discrepancies between the values found here and in the previous
subsection are due to the rounding used there to express payment amounts
to the nearest cent.)
We consider next a numerical example showing breakeven point for reﬁ
nancing by balancing costs versus time needed to amortize them.
Suppose that you have a 30year mortage for $100,000 at nominal rate
i
(12)
= 9%, with level monthly payments, and that after 7 years of payments
you reﬁnance to obtain a new 30year mortgage at 7% nominal interest ( =
2.1. MORE ON THEORY OF INTEREST 47
i
(m)
for m = 12), with closing costs of $1500 and 4 points (i.e., 4% of the
total reﬁnanced amount including closing costs added to the initial balance),
also with level monthly payments. Figuring present values using the new
interest rate of 7%, what is the time K (to the nearest month) such that
if both loans — the old and the new — were to be paid oﬀ in exactly K
years after the time (the 7year mark for the ﬁrst loan) when you would have
reﬁnanced, then the remaining paymentstreams for both loans from the time
when you reﬁnance are equivalent (i.e., have the same present value from
that time) ?
We ﬁrst calculate the present value of payments under the new loan.
As remarked above in the previous example, since the same interest rate is
being used to value the payments as is used in ﬁguring the reﬁnanced loan,
the valuation of the new loan does not depend upon the time K to payoﬀ.
(It is ﬁgured here as though the payoﬀ time K were 10 years.)
> unlist(RefExmp(1.e5, 30,7,30, 1500,.04, 10, 0.09,0.07,0.07))
Oldpaymt NewBal RefBal Newpaymt Value
804.62 93640 98946 658.29 98946
Next we compute the value of payments under the old loan, at 7% nominal
rate, also at payoﬀ time K = 10. For comparison, the value under the
old loan for payoﬀ time 0 (i.e., for cash payoﬀ at the time when reﬁnancing
would have occurred) coincides with the New Balance amount of $93640.
> unlist(RefExmp(1.e5, 30,7,23, 0,0, 10, 0.09,0.09,0.07))
Oldpaymt NewBal RefBal Newpaymt Value
804.62 93640 93640 804.62 106042
The values found in the same way when the payoﬀ time K is successively
replaced by 4, 3, 3.167, 3.25 are 99979, 98946, 98593, 98951. Thus, the
payoﬀtime K at which there is essentially no diﬀerence in present value
at nominal 7% between the old loan or the reﬁnanced loan with costs and
points (which was found to have Value 98946), is 3 years and 3 months
after reﬁnancing.
48 CHAPTER 2. INTEREST & FORCE OF MORTALITY
2.2 Force of Mortality & Analytical Models
Up to now, the function S(t) called the survivor or survival function has
been deﬁned to be equal to the lifetable ratio l
x
/l
0
at all integer ages t = x,
and to be piecewise continuously diﬀerentiable for all positive real values of
t. Intuitively, for all positive real y and t, S(y) −S(y +t) is the fraction
of the initial lifetable cohort which dies between ages y and y + t, and
(S(y) −S(y +t))/S(y) represents the fraction of those alive at exact age y
who fail before y +t. An equivalent representation is S(y) =
_
∞
y
f(t) dt ,
where f(t) ≡ −S
(t) is called the failure density. If T denotes the random
variable which is the age at death for a newly born individual governed by
the same causes of failure as the lifetable cohort, then Pr(T ≥ y) = S(y),
and according to the Fundamental Theorem of Calculus,
lim
→0+
P(y ≤ T ≤ y +)
= lim
→0+
1
_
y+
y
f(t) dt = f(y)
as long as the failure density is a continuous function.
Two further useful actuarial notations, often used to specify the theoret
ical lifetime distribution, are:
t
p
y
= P
_
T ≥ y +t  T ≥ y
_
= S(y +t)/S(y)
and
t
q
y
= 1 −
t
p
y
= P
_
T ≤ y +t  T ≥ y
_
= (S(y) −S(y +t))/S(y)
The quantity
t
q
y
is referred to as the agespeciﬁc death rate for periods
of length t. In the most usual case where t = 1 and y = x is an integer,
the notation
1
q
x
is replaced by q
x
, and
1
p
x
is replaced by p
x
. The
rate q
x
would be estimated from the cohort life table as the ratio d
x
/l
x
of
those who die between ages x and x+1 as a fraction of those who reached
age x. The way in which this quantity varies with x is one of the most
important topics of study in actuarial science. For example, one important
way in which numerical analysis enters actuarial science is that one wishes
to interpolate the values
1
q
y
smoothly as a function of y. The topic called
“Graduation Theory” among actuaries is the mathematical methodology of
Interpolation and Splinesmoothing applied to the raw function q
x
= d
x
/l
x
.
2.2. FORCE OF MORTALITY & ANALYTICAL MODELS 49
To give some idea what a realistic set of deathrates looks like, Figure 2.1
pictures the agespeciﬁc 1year deathrates q
x
for the simulated lifetable
given as Table 1.1 on page 4. Additional granularity in the deathrates can
be seen in Figure 2.2, where the logarithms of deathrates are plotted. After
a very high deathrate during the ﬁrst year of life (26.3 deaths per thousand
live births), there is a yearbyyear decline in deathrates roughly from 1.45
per thousand in the second year to 0.34 per thousand in the eleventh year.
(But there were small increases in rate from ages 4 to 7 and from 8
to 9, which are as likely due to statistical irregularity as to real increases
in risk.) Between ages 11 and 40, there is an erratic but roughly linear
increase of deathrates per thousand from 0.4 to 3.0. However, at ages
beyond 40 there is a rapid increase in deathrates as a function of age.
As can be seen from Figure 2.2, the values q
x
seem to increase roughly
as a power c
x
where c ∈ [1.08, 1.10]. (Compare this behavior with the
GompertzMakeham Example (v) below.) This exponential behavior of the
agespeciﬁc deathrate for large ages suggests that the deathrates pictured
could reasonably be extrapolated to older ages using the formula
q
x
≈ q
78
· (1.0885)
x−78
, x ≥ 79 (2.10)
where the number 1.0885 was found as log(q
78
/q
39
)/(78 −39).
Now consider the behavior of
q
x
as gets small. It is clear that
q
x
must also get small, roughly proportionately to , since the probability of
dying between ages x and x + is approximately f(x) when gets
small.
Deﬁnition: The limiting deathrate
q
x
/ per unit time as 0 is
called by actuaries the force of mortality µ(x). In reliability theory or
biostatistics, the same function is called the failure intensity, failure rate, or
hazard intensity.
The reasoning above shows that for small ,
q
y
=
1
S(y)
_
y+
y
f(t) dt −→
f(y)
S(y)
, 0
Thus
µ(y) =
f(y)
S(y)
=
−S
(y)
S(y)
= −
d
dy
ln(S(y))
50 CHAPTER 2. INTEREST & FORCE OF MORTALITY
•
•
•
• • • • • • • • • • • • • •
• • • • • • • • • • • • • • • • •
• • • •
• •
•
•
• •
•
•
•
• •
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
AgeSpecific Death Rates for Illustrative Life Table
Age in Years
A
g
e

S
p
e
c
i
f
i
c
D
e
a
t
h
R
a
t
e
0 20 40 60 80
0
.
0
0
.
0
2
0
.
0
4
0
.
0
6
0
.
0
8
Figure 2.1: Plot of agespeciﬁc deathrates q
x
versus x, for the simulated
illustrative life table given in Table 1.1, page 4.
2.2. FORCE OF MORTALITY & ANALYTICAL MODELS 51
•
•
•
• •
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
• •
•
•
•
•
•
•
•
•
•
•
•
•
• •
• •
•
•
•
•
•
•
•
•
Logarithm of DeathRates versus Age
Age in years
l
o
g
(
D
e
a
t
h
r
a
t
e
)
0 20 40 60 80

8

7

6

5

4

3
for Illustrative Simulated LifeTable
Figure 2.2: Plot of logarithm log(q
x
) of agespeciﬁc deathrates as a function
of age x, for the simulated illustrative life table given in Table 1.1, page 4.
The rates whose logarithms are plotted here are the same ones shown in
Figure 2.1.
52 CHAPTER 2. INTEREST & FORCE OF MORTALITY
where the chain rule for diﬀerentiation was used in the last step. Replacing
y by t and integrating both sides of the last equation between 0 and y,
we ﬁnd
_
y
0
µ(t) dt =
_
−ln(S(t))
_
y
0
= −ln(S(y))
since S(0) = 1. Similarly,
_
y+t
y
µ(s) ds = lnS(y) −lnS(y +t)
Now exponentiate to obtain the useful formulas
S(y) = exp
_
−
_
y
0
µ(t) dt
_
,
t
p
y
=
S(y +t)
S(y)
= exp
_
−
_
y+t
y
µ(s) ds
_
Examples:
(i) If S(t) = (ω −t)/ω for 0 ≤ t ≤ ω (the uniform failure distribution
on [0, ω] ), then µ(t) = (ω −t)
−1
. Note that this hazard function increases
to ∞ as t increases to ω.
(ii) If S(t) = e
−µt
for t ≥ 0 (the exponential failure distribution on
[0, ∞) ), then µ(t) = µ is constant.
(iii) If S(t) = exp(−λt
γ
) for t ≥ 0, then mortality follows the Weibull
life distribution model with shape parameter γ > 0 and scale parameter λ.
The force of mortality takes the form
µ(t) = λγ t
γ−1
This model is very popular in engineering reliability. It has the ﬂexibility
that by choice of the shape parameter γ one can have
(a) (γ > 1) failure rate increasing as a function of x
(b) ( γ = 1) constant failure rate (exponential model), or
(c) (0 < γ < 1) decreasing failure rate.
2.2. FORCE OF MORTALITY & ANALYTICAL MODELS 53
But what one cannot have, in the examples considered so far, is a forceof
mortality function which decreases on part of the timeaxis and increases
elsewhere.
(iv) Two other models for positive random variables which are popular
in various statistical applications are the Gamma, with
S(y) =
_
∞
y
β
α
t
α−1
e
−βt
dt /
_
∞
0
z
α−1
e
−z
dz , α, β > 0
and the Lognormal, with
S(y) = 1 −Φ
_
lny −m
σ
_
, m real, σ > 0
where
Φ(z) ≡
1
2
+
_
z
0
e
−u
2
/2
du
√
2π
is called the standard normal distribution function. In the Gamma case,
the expected lifetime is α/β, while in the Lognormal, the expectation is
exp(m + σ
2
/2). Neither of these last two examples has a convenient or
interpretable forceofmortality function.
Increasing force of mortality intuitively corresponds to aging, where the
causes of death operate with greater intensity or eﬀect at greater ages. Con
stant force of mortality, which is easily seen from the formula S(y) =
exp(−
_
y
0
µ(t) dt) to be equivalent to exponential failure distribution, would
occur if mortality arose only from pure accidents unrelated to age. Decreas
ing force of mortality, which really does occur in certain situations, reﬂects
what engineers call “burnin”, where after a period of initial failures due to
loose connections and factory defects the nondefective devices emerge and
exhibit high reliability for a while. The decreasing force of mortality reﬂects
the fact that the devices known to have functioned properly for a short while
are known to be correctly assembled and are therefore highly likely to have a
standard length of operating lifetime. In human life tables, infant mortality
corresponds to burnin: risks of death for babies decrease markedly after the
oneyear period within which the most severe congenital defects and diseases
of infancy manifest themselves. Of course, human life tables also exhibit an
aging eﬀect at high ages, since the highmortality diseases like heart disease
and cancer strike with greatest eﬀect at higher ages. Between infancy and
54 CHAPTER 2. INTEREST & FORCE OF MORTALITY
late middle age, at least in western countries, hazard rates are relatively ﬂat.
This pattern of initial decrease, ﬂat middle, and ﬁnal increase of the force
ofmortality, seen clearly in Figure 2.1, is called a bathtub shape and requires
new survival models.
As shown above, the failure models in common statistical and reliability
usage either have increasing force of mortality functions or decreasing force of
mortality, but not both. Actuaries have developed an analytical model which
is somewhat more realistic than the preceding examples for human mortalty
at ages beyond childhood. While the standard form of this model does not
accommodate a bathtub shape for deathrates, a simple modiﬁcation of it
does.
Example (v). (GompertzMakeham forms of the force of mortality). Sup
pose that µ(y) is deﬁned directly to have the form A + Bc
y
. (The Bc
y
term was proposed by Gompertz, the additive constant A by Makeham.
Thus the Gompertz forceofmortality model is the special case with A = 0,
or µ(y) = Bc
y
.) By choice of the parameter c as being respectively
greater than or less than 1, one can arrange that the forceofmortality
curve either be increasing or decreasing. Roughly realistic values of c for
human mortality will be only slightly greater than 1: if the Gompertz
(nonconstant) term in forceofmortality were for example to quintuple in
20 years, then c ≈ 5
1/20
= 1.084, which may be a reasonable value except
for very advanced ages. (Compare the comments made in connection with
Figures 2.1 and 2.2: for middle and higher ages in the simulated illustrative
life table of Table 1.1, which corresponds roughly to US male mortality of
around 1960, the ﬁgure of c was found to be roughly 1.09.) Note that in
any case the GompertzMakeham force of mortality is strictly convex (i.e.,
has strictly positive second derivative) when B > 0 and c = 1. The
GompertzMakeham family could be enriched still further, with further ben
eﬁts of realism, by adding a linear term Dy. If D < −B ln(c), with
0 < A < B, c > 1, then it is easy to check that
µ(y) = A + Bc
y
+ Dy
has a bathtub shape, initially decreasing and later increasing.
Figures 2.3 and 2.4 display the shapes of forceofmortality functions (iii)
(v) for various parameter combinations chosen in such a way that the ex
2.2. FORCE OF MORTALITY & ANALYTICAL MODELS 55
pected lifetime is 75 years. This restriction has the eﬀect of reducing the
number of free parameters in each family of examples by 1. One can
see from these pictures that the Gamma and Weibull families contain many
very similar shapes for forceofmortality curves, but that the lognormal and
Makeham families are quite diﬀerent.
Figure 2.5 shows survival curves from several analytical models plotted on
the same axes as the 1959 US male lifetable data from which Table 1.1 was
simulated. The previous discussion about bathtubshaped force of mortality
functions should have made it clear that none of the analytical models pre
sented could give a good ﬁt at all ages, but the Figure indicates the rather
good ﬁt which can be achieved to realistic lifetable data at ages 40 and
above. The models ﬁtted all assumed that S(40) = 0.925 and that for lives
aged 40, T −40 followed the indicated analytical form. Parameters for all
models were determined from the requirements of median age 72 at death
(equal by deﬁnition to the value t
m
for which S(t
m
) = 0.5) and probability
0.04 of surviving to age 90. Thus, all four plotted survival curves have been
designed to pass through the three points (40, 0.925), (72, 0.5), (90, 0.04).
Of the four ﬁtted curves, clearly the Gompertz agrees most closely with the
plotted points for 1959 US male mortality. The Gompertz curve has param
eters B = 0.00346, c = 1.0918, the latter of which is close to the value
1.0885 used in formula (2.10) to extrapolate the 1959 lifetable deathrates
to older ages.
2.2.1 Comparison of Forces of Mortality
What does it mean to say that one lifetime, with associated survival function
S
1
(t), has hazard (i.e. force of mortality) µ
1
(t) which is a constant multiple
κ at all ages of the force of mortality µ
2
(t) for a second lifetime with
survival function S
2
(t) ? It means that the cumulative hazard functions are
proportional, i.e.,
−lnS
1
(t) =
_
t
0
µ
1
(x)dx =
_
t
0
κµ
2
(x)dx = κ(−lnS
2
(t))
and therefore that
S
1
(t) = (S
2
(t))
κ
, all t ≥ 0
56 CHAPTER 2. INTEREST & FORCE OF MORTALITY
Weibull(alpha,lambda)
Age (years)
H
a
z
a
r
d
0 20 40 60 80 100
0
.
0
0
.
0
0
5
0
.
0
1
0
0
.
0
1
5
0
.
0
2
0
0
.
0
2
5
alpha=1
alpha=0.7
alpha=1.3
alpha=1.6
alpha=2
Gamma(beta,lambda)
Age (years)
H
a
z
a
r
d
0 20 40 60 80 100
0
.
0
0
.
0
0
5
0
.
0
1
0
0
.
0
1
5
0
.
0
2
0
0
.
0
2
5
beta=1
beta=0.7
beta=1.3
beta=1.6
beta=2
Figure 2.3: Force of Mortality Functions for Weibull and Gamma Probability
Densities. In each case, the parameters are ﬁxed in such a way that the
expected survival time is 75 years.
2.2. FORCE OF MORTALITY & ANALYTICAL MODELS 57
Lognormal(mu,sigma^2)
Age (years)
H
a
z
a
r
d
0 20 40 60 80 100
0
.
0
0
.
0
0
5
0
.
0
1
0
0
.
0
1
5
0
.
0
2
0
0
.
0
2
5
sigma=0.2
sigma=0.4
sigma=0.6
sigma=1.6
sigma=2
Makeham(A,B,c)
Age (years)
H
a
z
a
r
d
0 20 40 60 80 100
0
.
0
0
.
0
0
5
0
.
0
1
0
0
.
0
1
5
0
.
0
2
0
0
.
0
2
5
A=0.0018, B=0.0010, c=1.002
A=0.0021,B=0.0070, c=1.007
A=0.003, B=0.0041, c=1.014
A=0.0041, B=0.0022, c=1.022
Figure 2.4: Force of Mortality Functions for Lognormal and Makeham Den
sities. In each case, the parameters are ﬁxed in such a way that the expected
survival time is 75 years.
58 CHAPTER 2. INTEREST & FORCE OF MORTALITY
Plots of Theoretical Survival Curves
Age (years)
S
u
r
v
i
v
a
l
P
r
o
b
a
b
i
l
i
t
y
40 50 60 70 80 90 100
0
.
0
0
.
2
0
.
4
0
.
6
0
.
8
• •
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Plotted points from US 1959 male lifetable
Lognormal(3.491, .246^2)
Weibull(3.653, 1.953e6)
Gamma(14.74, 0.4383)
Gompertz(3.46e3, 1.0918)
Figure 2.5: Theoretical survival curves, for ages 40 and above, plotted as
lines for comparison with 1959 US male lifetable survival probabilities plot
ted as points. The four analytical survival curves — Lognormal, Weibull,
Gamma, and Gompertz — are taken as models for ageatdeath minus 40,
so if S
theor
(t) denotes the theoretical survival curve with indicated parame
ters, the plotted curve is (t, 0.925 · S
theor
(t −40)). The parameters of each
analytical model were determined so that the plotted probabilities would be
0.925, 0.5, 0.04 respectively at t = 40, 72, 90.
2.2. FORCE OF MORTALITY & ANALYTICAL MODELS 59
This remark is of especial interest in biostatistics and epidemiology when
the factor κ is allowed to depend (e.g., by a regression model ln(κ) = β·Z )
on other measured variables (covariates) Z. This model is called the (Cox)
ProportionalHazards model and is treated at length in books on survival data
analysis (Cox and Oakes 1984, Kalbﬂeisch and Prentice 1980) or biostatistics
(Lee 1992).
Example. Consider a setting in which there are four subpopulations of the
general population, categorized by the four combinations of values of two
binary covariates Z
1
, Z
2
= 0, 1. Suppose that these four combinations have
respective conditional probabilities for lives aged x (or relative frequencies in
the general population aged x)
P
x
(Z
1
= Z
2
= 0) = 0.15 , P
x
(Z
1
= 0, Z
2
= 1) = 0.2
P
x
(Z
1
= 1, Z
2
= 0) = 0.3 , P
x
(Z
1
= Z
2
= 1) = 0.35
and that for a life aged x and all t > 0,
Pr(T ≥ x +t  T ≥ x, Z
1
= z
1
, Z
2
= z
2
) = exp(−2.5 e
0.7z
1
−.8z
2
t
2
/20000)
It can be seen from the conditional survival function just displayed that the
forces of mortality at ages greater than x are
µ(x +t) = (2.5 e
0.7z
1
−.8z
2
) t/10000
so that the force of mortality at all ages is multiplied by e
0.7
= 2.0138 for
individuals with Z
1
= 1 versus those with Z
1
= 0, and is multiplied by
e
−0.8
= 0.4493 for those with Z
2
= 1 versus those with Z
2
= 0. The eﬀect
on agespeciﬁc deathrates is approximately the same. Direct calculation
shows for example that the ratio of agespeciﬁc death rate at age x+20 for
individuals in the group with (Z
1
= 1, Z
2
= 0) versus those in the group with
(Z
1
= 0, Z
2
= 0) is not precisely e
0.7
= 2.014, but rather
1 −exp(−2.5e
0.7
((21
2
−20
2
)/20000)
1 −exp(−2.5((21
2
−20
2
)/20000)
= 2.0085
Various calculations, related to the fractions of the surviving population at
various ages in each of the four population subgroups, can be performed
easily . For example, to ﬁnd
Pr(Z
1
= 0, Z
2
= 1  T ≥ x + 30)
60 CHAPTER 2. INTEREST & FORCE OF MORTALITY
we proceed in several steps (which correspond to an application of Bayes’
rule, viz. Hogg and Tanis 1997, sec. 2.5, or Devore 2007):
Pr(T ≥ x+30, Z
1
= 0 Z
2
= 1 T ≥ x) = 0.2 exp(−2.5e
−0.8
30
2
20000
) = 0.1901
and similarly
Pr(T ≥ x + 30  T ≥ x) = 0.15 exp(−2.5(30
2
/20000)) + 0.1901 +
+ 0.3 exp(−2.5 ∗ e
0.7
30
2
20000
) + 0.35 exp(−2.5e
0.7−0.8
30
2
20000
) = 0.8795
Thus, by deﬁnition of conditional probabilities (restricted to the cohort of
lives aged x), taking ratios of the last two displayed quantities yields
Pr(Z
1
= 0, Z
2
= 1  T ≥ x + 30) =
0.1901
0.8795
= 0.2162
2
In biostatistics and epidemiology, the measured variables Z = (Z
1
, . . . , Z
p
)
recorded for each individual in a survival study might be: indicator of a spe
ciﬁc disease or diagnostic condition (e.g., diabetes, high blood pressure, spe
ciﬁc electrocardiogram anomaly), quantitative measurement of a riskfactor
(dietary cholesterol, percent caloric intake from fat, relative weighttoheight
index, or exposure to a toxic chemical), or indicator of type of treatment or
intervention. In these ﬁelds, the objective of such detailed models of covari
ate eﬀects on survival can be: to correct for incidental individual diﬀerences
in assessing the eﬀectiveness of a treatment; to create a prognostic index for
use in diagnosis and choice of treatment; or to ascertain the possible risks and
beneﬁts for health and survival from various sorts of lifestyle interventions.
The multiplicative eﬀects of various riskfactors on agespeciﬁc death rates
are often highlighted in the news media.
In an insurance setting, categorical variables for risky lifestyles, occupa
tions, or exposures might be used in riskrating, i.e., in individualizing insur
ance premiums. While riskrating is used routinely in casualty and property
insurance underwriting, for example by increasing premiums in response to
recent claims or by taking location into account, it can be politically sensi
tive in a lifeinsurance and pension context. In particular, while diﬀerences
2.3. EXERCISE SET 2 61
in mortality by gender and to some extgent by family health history can
be used in calculating insurance and annuity premiums, as can certain life
style factors like smoking, it is currently illegal to use racial diﬀerences and
diﬀerences based on genetic testing in this way.
All life insurers must be conscious of the extent to which their policyhold
ers as a group diﬀer from the general population with respect to mortality.
Insurers can collect special mortality tables on special groups, such as em
ployee groups or voluntary organizations, and regressiontype models like the
Cox proportionalhazards model may be useful in quantifying group mortal
ity diﬀerences when the specialgroup mortality tables are not based upon
large enough cohorts for long enough times to be fully reliable. See Chapter
6, Section 4, for discussion about the modiﬁcation of insurance premiums for
select groups.
2.3 Exercise Set 2
(1). The sum of the present value of $1 paid at the end of n years and
$1 paid at the end of 2n years is $1. Find (1 +r)
2n
, where r = annual
interest rate, compounded annually.
(2). Suppose that an individual aged 20 has random lifetime Z with
continuous density function
f
(
t) =
1
360
_
1 +
t
10
_
, for 20 ≤ t ≤ 80
and 0 for other values of t. (The random variable Z in this problem is a
particular type of ageatdeath variable T conditioned on being ≥ 20.)
(a) If this individual has a contract with your company that you must
pay his heirs 10
6
· (1.4 −Z/50) dollars at the exact date of his death if this
occurs between ages 20 and 70, then what is the expected payment ?
(b) If the value of the deathpayment described in (a) should properly be
discounted by the factor exp(−0.08 · (Z −20)), i.e. by the nominal interest
rate of e
0.08
− 1 per year) to calculate the present value of the payment,
then what is the expected present value of the payment under the insurance
contract ?
62 CHAPTER 2. INTEREST & FORCE OF MORTALITY
(3). Suppose that a continuous random variable T has hazard rate function
(= force of mortality)
h(t) = 10
−3
·
_
7.0 −0.5t + 2e
t/20
_
, t > 0
This is a legitimate hazard rate of GompertzMakeham type since its mini
mum, which occurs at t = 20 ln(5), is (17−10 ln(5)) · 10
−4
= 9.1 · 10
−5
> 0.
(a) Construct a cohort lifetable with h(t) as “force of mortality”,
based on integer ages up to 70 and cohortsize (= radix) l
0
= 10
5
. (Give
selected numerical entries, preferably calculated by means of a little computer
program. If you do the arithmetic using handcalculators and/or tables, give
only the values for ages which are multiples of 10.)
(b) Find the probability that the random variable T exceeds 30, given
that it exceeds 3. Hint: ﬁnd a closedform formula for S(t) = P(T ≥ t).
(4). Do the MortgageReﬁnancing exercise given as Exercise 2.B in the
Illustration on mortgage reﬁnancing at the end of Section 2.1.
(5). (a) The mortality pattern of a certain population may be described as
follows: out of every 98 lives born together, one dies annually until there
are no survivors. Find a simple function that can be used as S(x) for this
population, and ﬁnd the probability that a life aged 30 will survive to attain
age 35.
(b) Suppose that for x between ages 12 and 40 in a certain population,
10% of the lives aged x die before reaching age x+1 . Find a simple function
that can be used as S(x) for this population, and ﬁnd the probability that
a life aged 30 will survive to attain age 35.
(6). Suppose that a survival distribution (i.e., survival function based on
a cohort life table) has the property that
1
p
x
= γ · (γ
2
)
x
for some ﬁxed γ
between 0 and 1, for every real ≥ 0. What does this imply about S(x) ?
(Give as much information about S as you can.)
(7). If the instantaneous interest rate is r(t) = 0.01 · t for 0 ≤ t ≤ 3,
then ﬁnd the equivalent single eﬀective rate of interest for money invested at
interest throughout the interval 0 ≤ t ≤ 3.
2.3. EXERCISE SET 2 63
(8). Find the accumulated value of $100 at the end of 15 years if the
nominal interest rate compounded quarterly (i.e., i
(4)
) is 8% for the ﬁrst 5
years, if the eﬀective rate of discount is 7% for the second 5 year interval (i.e.
the interval ranging from time 5 to 10), and if the nominal rate of discount
compounded semiannually (m = 2) is 6% for the third 5 year interval.
(9). Suppose that you borrow $1000 for 3 years at 6% eﬀective rate, to be
repaid in level payments every six months (twice yearly).
(a) Find the level payment amount P.
(b) What is the present value of the payments you will make if you skip
the 2nd and 4th payments ? (You may express your answer in terms of P. )
(10). A survival function has the form S(t) = max(0,
c−t
c+t
). If a mortality
table is derived from this survivalfunction with a radix l
0
of 100,000 at
age 0, and if l
35
= 44, 000 :
(i) What is the terminal age of the table ?
(ii) What is the probability of surviving from birth to age 60 ?
(iii) What is the probability of a person at exact age 10 dying between
exact ages 30 and 45 ?
(11). A separate life table has been constructed for each calendar year of
birth, Y , beginning with Y = 1950. The mortality functions for the
various tables are denoted by the appropriate superscript
Y
. For each Y
and for all ages t
µ
Y
(t) = A · k(Y ) + Bc
t
, p
Y +1
t
= (1 +r) p
Y
t
where k is a function of Y alone and A, B, r are constants (with r > 0).
If k(1950) = 1, then derive a general expression for k(Y ).
(12). A standard mortality table follows Makeham’s Law with force of
mortality
µ(t) = A + Bc
t
at all ages t
A separate, higherrisk mortality table also follows Makeham’s Law with
64 CHAPTER 2. INTEREST & FORCE OF MORTALITY
force of mortality
µ
∗
(t) = A
∗
+ B
∗
c
t
at all ages t
with the same constant c. If for all starting ages the probability of surviving
6 years according to the higherrisk table is equal to the probability of
surviving 9 years according to the standard table, then express each of A
∗
and B
∗
in terms of A, B, c.
(13). A homeowner borrows $100, 000 at eﬀective annual rate 7% from
a bank, agreeing to repay by 30 equal yearly payments beginning one year
from the time of the loan.
(a) How much is each payment ?
(b) Suppose that after paying the ﬁrst 3 yearly payments, the homeowner
misses the next two (i.e. pays nothing on the 4
th
and 5
th
anniversaries of
the loan). Find the outstanding balance at the 6
th
anniversary of the loan,
ﬁgured at 7% ). This is the amount which, if paid as a lump sum at time
6, has present value together with the amounts already paid of $100, 000 at
time 0.
(14). A deposit of 300 is made into a fund at time t = 0. The fund pays
interest for the ﬁrst three years at a nominal monthly rate d
(12)
of discount.
From t = 3 to t = 7, interest is credited according to the force of interest
δ
t
= 1/(3t + 3). As of time t = 7, the accumulated value of the fund is
574. Calculate d
(12)
.
(15). Calculate the price at which you would sell a $10, 000 30year coupon
bond with nominal 6% semiannual coupon (n = 30, m − 2, i
(m)
= 0.06),
15 years after issue, if for the next 15 years, the eﬀective interest rate for
valuation is i
eﬀ
= 0.07.
(16). A 6% ‘zerocoupon’ 30year bond was issued exactly 15 years ago for
a face amount of $10, 000. This bond contractually entitles the bearer to
receive 30 years after the issue date the amount accumulated at i = i
eﬀ
=
0.06 on the face amount. Calculate the fair price at which you would sell
this zerocoupon bond, if for the next 15 years, the eﬀective interest rate will
be i
eﬀ
= 0.07.
2.4. WORKED EXAMPLES 65
(17). Suppose that the borrower of a $100, 000 30year loan with half
yearly payments (m = 2) start in six months from the time of borrowing, and
with nominal interest rate i
(2)
= .04, has made all payments except for two
that he skipped, the 23’rd and 56’th payments. What lumpsum payment
did the borrower have to make at the end of 30 years, in addition to his ﬁnal
payment, in order to pay oﬀ the loan completely ?
(18). In Problem (17), the missed payments did not result in any additional
fees or charges, only in continuing accrued interest on the amounts of the
missed payments. Suppose that the missed payments in (17) actually result
in late fees of $200 each of which is added to the balance at the time(s)
of missed payments. Now answer the same question as in (17) about the
amount of the ﬁnal lumpsum payment required.
(19). One of the curves plotted in the ﬁrst part of Figure 2.3 is the
lognormal(m, σ
2
) hazard intensity where σ = 1.3 is ﬁxed and where m
was determined from it by the requirement that the expectation of survival
time was 75 years. Now ﬁnd the value m associated with σ = 1.3 if the
median survival time is ﬁxed at 72 years, and ﬁnd the force of mortality for
this lognormal at 65 years.
(20). A small city issues a bond for twenty million dollars for ten years
at 5% nominal halfyearly interest (m = 2) and creates a sinking fund into
which it will make twiceyearly deposits (m
= 2), from its tax revenue.
(a) Find the amount of the level payment the city must make into the
sinking fund if the interest it earns on that fund is 5%, and ﬁnd the reserve,
or accumulated balance, in the sinking fund after 6 years.
(b) Answer the same questions if the city knows it can earn 6% on the
money it deposits into its sinking fund.
2.4 Worked Examples
Example 1. How large must a halfyearly payment be in order that the stream
of payments starting immediately be equivalent (in present value terms) at
6% interest to a lumpsum payment of $5000, if the paymentstream is to
last (a) 10 years, (b) 20 years, or (c) forever ?
66 CHAPTER 2. INTEREST & FORCE OF MORTALITY
If the payment size is P, then the balance equation is
5000 = 2 P · ¨a
(2)
n
= 2 P (1 −1.06
−n
)/d
(2)
Since d
(2)
= 2(1 −1/
√
1.06) = 2 · 0.02871, the result is
P = (5000 · 0.02871)/(1 −1.06
−n
) = 143.57/(1 −1.06
−n
)
So the answer to part (c), in which n = ∞, is $143.57. For parts (a) and
(b), respectively with n = 10 and 20, the answers are $325.11, $208.62.
Example 2. Assume m is divisible by 2. Express in two diﬀer
ent ways the present value of the perpetuity of payments 1/m at times
1/m, 3/m, 5/m, . . . , and use either one to give a simple formula.
This example illustrates the general methods enunciated at the beginning
of Section 2.1. Observe ﬁrst of all that the speciﬁed paymentstream is
exactly the same as a stream of payments of 1/m at times 0, 2/m, 4/m, . . .
forever, deferred by a time 1/m. Since this paymentstream starting at 0
is exactly onehalf that of the stream whose present value is ¨a
(m/2)
∞
, a ﬁrst
present value expression is
v
1/m
(1/2) ¨ a
(m/2)
∞
A second way of looking at the paymentstream at odd multiples of 1/m
is as the perpetuitydue payment stream ( 1/m at times k/m for all
k ≥ 0) minus the paymentstream discussed above of amounts 1/m at
times 2k/m for all nonnegative integers k. Thus the present value has the
second expression
¨ a
(m)
∞
− (1/2) ¨ a
(m/2)
∞
Equating the two expressions allows us to conclude that
(1/2) ¨ a
(m/2)
∞
= ¨ a
(m)
∞
_
(1 + v
1/m
)
Substituting this into the ﬁrst of the displayed presentvalue expressions, and
using the simple expression 1/d
(m)
for the present value of the perpetuity
due, shows that that the present value requested in the Example is
1
d
(m)
·
v
1/m
1 + v
1/m
=
1
d
(m)
(v
−1/m
+ 1)
=
1
d
(m)
(2 +i
(m)
/m)
and this answer is valid whether or not m is even.
2.4. WORKED EXAMPLES 67
Example 3. Suppose that you are negotiating a carloan of $10, 000. Would
you rather have an interest rate of 4% for 4 years, 3% for 3 years, 2% for
2 years, or a cash discount of $500 ? Show how the answer depends upon
the interest rate with respect to which you calculate present values, and give
numerical answers for present values calculated at 6% and 8%. Assume that
all loans have monthly payments paid at the beginning of the month (e.g., the
4 year loan has 48 monthly payments paid at time 0 and at the ends of 47
succeeding months).
The monthly payments for an nyear loan at interestrate i is 10000/
(12 ¨ a
(12)
n
) = (10000/12) d
(12)
/(1 − (1 + i)
−n
). Therefore, the present value
at interestrate r of the nyear monthly paymentstream is
10000 ·
1 −(1 +i)
−1/12
1 −(1 +r)
−1/12
·
1 −(1 +r)
−n
1 −(1 +i)
−n
Using interestrate r = 0.06, the present values are calculated as follows:
For 4year 4% loan: $9645.77
For 3year 3% loan: $9599.02
For 2year 2% loan: $9642.89
so that the most attractive option is the cash discount (which would make
the present value of the debt owed to be $9500). Next, using interestrate
r = 0.08, the present values of the various options are:
For 4year 4% loan: $9314.72
For 3year 3% loan: $9349.73
For 2year 2% loan: $9475.68
so that the most attractive option in this case is the 4year loan. (The cash
discount is now the least attractive option.)
Example 4. Suppose that the force of mortality µ(y) is speciﬁed for exact
ages y ranging from 5 to 55 as
µ(y) = 10
−4
· (20 −0.530 −y)
Then ﬁnd analytical expressions for the survival probabilities S(y) for exact
68 CHAPTER 2. INTEREST & FORCE OF MORTALITY
ages y in the same range, and for the (oneyear) deathrates q
x
for integer
ages x = 5, . . . , 54, assuming that S(5) = 0.97.
The key formulas connecting force of mortality and survival function are
here applied separately on the ageintervals [5, 30] and [30, 55], as follows.
First for 5 ≤ y ≤ 30,
S(y) = S(5) exp(−
_
y
5
µ(z) dz) = 0.97 exp
_
−10
−4
(5(y−5)+0.25(y
2
−25))
_
so that S(30) = 0.97 e
−0.034375
= 0.93722, and for 30 ≤ y ≤ 55
S(y) = S(30) exp
_
−10
−4
_
y
30
(20 + 0.5(30 −z)) dz
_
= 0.9372 exp
_
−.002(y −30) + 2.5 · 10
−5
(y −30)
2
_
The deathrates q
x
therefore have two diﬀerent analytical forms: ﬁrst, in
the case x = 5, . . . , 29,
q
x
= S(x + 1)/S(x) = exp
_
−5 · 10
−5
(10.5 +x)
_
and second, in the case x = 30, . . . , 54,
q
x
= exp
_
−.002 + 2.5 · 10
−5
(2(x −30) + 1)
_
2.5. USEFUL FORMULAS FROM CHAPTER 2 69
2.5 Useful Formulas from Chapter 2
v = 1/(1 +i)
p. 34
a
(m)
n
=
1 −v
n
i
(m)
, ¨ a
(m)
n
=
1 −v
n
d
(m)
pp. 35–35
a
(m)
n
= v
1/m
¨a
(m)
n
p. 35
¨ a
(∞)
n
= a
(∞)
n
= a
n
=
1 −v
n
δ
p. 35
a
(m)
∞
=
1
i
(m)
, ¨a
(m)
∞
=
1
d
(m)
p. 36
(I
(m)
¨a)
(m)
n
= ¨ a
(m)
∞
_
¨ a
(m)
n
− nv
n
_
p. 38
(D
(m)
¨ a)
(m)
n
= (n +
1
m
) ¨ a
(m)
n
− (I
(m)
¨ a)
(m)
n
p. 38
nyr m’thly Mortgage Paymt :
Loan Amt
m¨a
(m)
n
70 CHAPTER 2. INTEREST & FORCE OF MORTALITY
p. 39
nyr Mortgage Bal. amt
k
m
+ : B
n,k/m
=
1 −v
n−k/m
1 −v
n
p. 42
t
p
y
=
S(y +t)
S(y)
= exp
_
−
_
t
0
µ(y +s) ds
_
p. 48
t
q
y
= 1 −
t
p
y
p. 48
q
x
=
1
q
x
=
d
x
l
x
, p
x
=
1
p
x
= 1 − q
x
p. 48
µ(y +t) =
f(y +t)
S(y +t)
= −
∂
∂t
lnS(y +t)
p. 49
S(y) = exp( −
_
y
0
µ(t) dt)
p. 52
Unif. Failure Dist.: S(t) =
ω −t
ω
, f(t) =
1
ω
, 0 ≤ t ≤ ω
p. 52
Expon. Dist.: S(t) = e
−µt
, f(t) = µe
−µt
, µ(t) = µ , t > 0
p. 52
2.5. USEFUL FORMULAS FROM CHAPTER 2 71
Weibull. Dist.: S(t) = e
−λt
γ
, µ(t) = λγt
γ−1
, t > 0
p. 52
Makeham: µ(t) = A+Bc
t
, t ≥ 0
Gompertz: µ(t) = Bc
t
, t ≥ 0
S(t) = exp
_
−At −
B
lnc
(c
t
−1)
_
p. 54
72 CHAPTER 2. INTEREST & FORCE OF MORTALITY
Chapter 3
More Probability Theory for
Life Tables
This Chapter introduces several key ideas in Probability Theory which are
essential for an understanding of the book’s core actuarial topics in Chapter
4 and 5. The ﬁrst of these ideas is that survival from one year to the next can
be regarded for each member of a population as a cointoss experiment, with
survival probability p
x
for a life aged x, independently of all other members
of the population. This point of view also provides a convenient vehicle for
conducting computer simulations of population survival experience for large
or small lifetable populations. Since the lifetable summarizes outcomes
on a large number of cointoss experiments, we study next through limit
theorems (law of large numbers and central limit theorem) the high degree of
predictability of these outcomes at the population level. This predictability
will be used in later chapters to justify consideration of expected present
values of contractual payouts to describe an insurer’s liability, so we prepare
the ground by presenting background theory and rules of manipulation for
expectations of discretevalued random variables. Finally, we complete our
probability background with further material on interpreting, approximating
and calculating with probabilities and expectations using theoretical models
of survival between successive years of age.
73
74 CHAPTER 3. PROBABILITY & LIFE TABLES
3.1 Binomial Variables and Limit Theorems
This Section develops basic machinery for the theory of random variables
which count numbers of successes in large numbers of independent biased
cointosses. The motivation is that in large lifetable populations, the number
l
x+t
who survive t timeunits after age x can be regarded as the number
of successes or heads in a large number l
x
of independent cointoss trials
corresponding to the further survival of each of the l
x
lives aged x , which for
each such life has probability
t
p
x
. The one preliminary mathematical result
that the student is assumed to know is the Binomial Theoremstating that
(for positive integers N and arbitrary real numbers x, y, z),
(1 +x)
N
=
N
k=0
_
N
k
_
x
k
, (y +z)
N
=
N
k=0
_
N
k
_
y
k
z
N−k
Recall that the ﬁrst of these assertions follows by equating the k
th
deriv
iatives of both sides at x = 0, where k = 0, . . . , N. The second assertion
follows immediately, in the nontrivial case when z = 0, by applying the ﬁrst
assertion with x = y/z and multiplying both sides by z
N
. This Theo
rem also has a direct combinatorial consequence. Consider the twovariable
polynomial
(y +z)
N
= (y +z) · (y +z) · · · (y +z) N factors
expanded by making all of the diﬀerent choices of y or z from each of
the N factors (y + z), multiplying each combination of choices out to
get a monomial y
j
z
N−j
, and adding all of the monomials together. Each
combined choice of y or z from the N factors (y +z) can be represented
as a sequence (a
1
, . . . , a
n
) ∈ {0, 1}
N
, where a
i
= 1 would mean that y
is chosen and a
i
= 0 would mean that z is chosen in the i
th
factor. Now
this combinatorial fact is immediately deduced from the Binomial Theorem:
since the coeﬃcient
_
N
k
_
is the total number of monomial terms y
k
z
N−k
which are collected when (y +z)
N
is expanded as described, and since these
monomial terms arise only from the combinations (a
1
, . . . , a
N
) of {y, z}
choices in which precisely k of the values a
j
are 1’s and the rest are 0’s,
The number of symbolsequences (a
1
, . . . , a
N
) ∈ {0, 1}
N
such
that
N
j=1
a
j
= k is given, for each k = 0, 1, . . . , N, by
3.1. BINOMIAL VARIABLES & LAW OF LARGE NUMBERS 75
_
N
k
_
= N(N −1) · · · (N −k +1)/ k! . This number
_
N
k
_
, spoken
as ‘N choose k’, therefore counts all of the ways of choosing k
element subsets (the positions j from 1 to N where 1’s occur)
out of N objects.
The random experiment of interest in this Section consists of a number
N of independent tosses of a coin, with probability p of coming up heads
each time. Such cointossing experiments — independently replicated two
outcome experiments with probability p of one of the outcomes, designated
‘success’ — are called Bernoulli (p) trials. The space of possible heads
andtails conﬁgurations, or sample space for this experiment, consists of the
strings of N zeroes and ones, with each string a = (a
1
, . . . , a
N
) ∈ {0, 1}
N
being assigned probability p
a
(1 − p)
N−a
, where a ≡
N
j=1
a
j
. Because
of the ﬁnite additivity axiom of probabilities (saying that Pr(A ∪ B) =
Pr(A) + Pr(B) for disjoint events A, B), the rule by which probabilities
are assigned to sets or events A of more than one string a ∈ {0, 1}
N
is
to add the probabilities of all individual strings a ∈ A. We are particularly
interested in the event (denoted [X = k]) that precisely k of the coin
tosses are heads, i.e., in the subset [X = k] ⊂ {0, 1}
N
consisting of all
strings a such that
N
j=1
a
j
= k. Since each such string has the same
probability p
k
(1 − p)
N−k
, and since, according to the discussion following
the Binomial Theorem above, there are
_
N
k
_
such strings, the probability
which is necessarily assigned to the event of k successes is
Pr( k successes in N Bernoulli(p) trials ) = P(X = k) =
_
N
k
_
p
k
(1−p)
N−k
By virtue of this result, the random variable X equal to the number of suc
cesses in N Bernoulli(p) trials, is said to have the Binomial distribution
with probability mass function p
X
(k) =
_
N
k
_
p
k
(1 −p)
N−k
.
With the notion of Bernoulli trials and the binomial distribution in hand,
we now begin to regard the ideal probabilities S(x + t)/S(x) as true but
unobservable probabilities
t
p
x
= p with which each of the l
x
lives aged x
will survive to age x + t . Since the mechanisms which cause those lives
to survive or die can ordinarily be assumed to be acting independently in a
probabilistic sense, we can regard the number l
x+t
of lives surviving to the
(possibly fractional) age x+t as a Binomial random variable with parameters
N = l
x
, p =
t
p
x
. From this point of view, if derived from an actual cohort
76 CHAPTER 3. PROBABILITY & LIFE TABLES
dataset of size equal to the radix, the observed lifetable counts l
x
would
be treated as random data which reﬂect but do not deﬁne the underlying
probabilities
x
p
0
= S(x) of survival to age x. However, common sense
and experience suggest that, when l
0
is large, and therefore the other life
counts l
x
for moderate values x are also large, the observed ratios l
x+t
/l
x
should reliably be very close to the ‘true’ probability
t
p
x
. In other words,
the ratio l
x+t
/l
x
is a statistical estimator of the unknown constant
t
p
x
.
The good property, called consistency, of this estimator to be close with very
large probability (based upon large lifetable size) to the headsprobability it
estimates, is established in the famous Law of Large Numbers. We state
and prove the result here only in the setting of binomial random variables,
sketching in Section 3.3 how it implies a more general result for ﬁnitevalued
discrete random variables. A more precise quantitative inequality concerning
binomial probabilities, a Large Deviation Inequality which is important in its
own right but more diﬃcult, is stated and proved in the Appendix to this
Chapter, Section 3.9.
Theorem 3.1 (Cointoss Law of Large Numbers) Suppose that X is
a Binomial (N, p) random variable, denoting the number of successes in N
Bernoulli (p) trials. Law of Large Numbers. For arbitrarily small ﬁxed
δ, > 0, not depending upon N, the number N of Bernoulli trials can be
chosen so large that
Pr
_
 X/N −p  ≥ δ
_
≤
Proof. Since the event [ X/N −p ≥ δ ] = [ X −Np ≥ Nδ ] is the union
of the disjoint events [X = k] for k − Np ≥ Nδ, which in turn consist
of all outcomestrings (a
1
, . . . , a
N
) ∈ {0, 1}
N
for which
N
j=1
a
j
= k with
k −Np ≥ Nδ, the subset of the binomial probability mass function values
p
X
(k) with k −Np ≥ Nδ are summed to provide
Pr(X/N−p ≥ δ) =
k: k−Np≥Nδ
Pr(X = k) =
k: k−Np≥Nδ
_
N
k
_
p
k
(1−p)
N−k
This summation is termbyterm less than or equal to
k: k−Np≥Nδ
_
N
k
_
p
k
(1−p)
N−k
(k −Np)
2
(Nδ)
2
≤
N
k=0
_
N
k
_
p
k
(1−p)
N−k
(k −Np)
2
(Nδ)
2
(3.1)
3.1. BINOMIAL VARIABLES & LAW OF LARGE NUMBERS 77
where we have made the second some larger by including more nonnegative
terms in it. However, direct summation shows
N
k=0
k
_
N
k
_
p
k
(1 −p)
N−k
=
N
k=1
kp
N · (N −1)!
k(k −1)!(N −k)!
p
k−1
(1 −p)
N−k
which after replacing k −1 by l, becomes with the aid of the Binomial The
orem
= Np
N−1
l=0
_
N −1
l
_
p
l
(1 −p)
N−1−l
= Np
and similarly (now with j = k −2)
N
k=0
k(k−1)
_
N
k
_
p
k
(1−p)
N−k
=
N
k=2
p
2
N(N −1) (N −2)!
(k −2)!(N −k)!
p
k−2
(1−p)
N−k
= N(N −1)p
2
N−2
j=0
_
N −2
j
_
p
j
(1 −p)
N−2−j
= N(N −1) p
2
Putting together the last calculations, and simplifying algebraically, it is easy
to check via the equality (k −Np)
2
= k(k −1) −(2Np −1)k +(Np)
2
, that
N
k=0
(k −Np)
2
_
N
k
_
p
k
(1 −p)
N−k
= N p(1 −p)
Substituting this ﬁnal relation into (3.1) now shows that
Pr(X/N −p ≥ δ) ≤
Np(1 −p)
(Nδ)
2
=
p(1 −p)
Nδ
2
The assertion of the Theorem now follows by taking N ≥ (p(1−p)/(δ
2
). 2
3.1.1 Probability Bounds & Approximations
Theorem 3.1 provides only a very crude upper bound to the probability with
which X/N − p ≥ δ. A much more accurate upper bound in given in
Theorem 3.2 of the Appendix to the Chapter (Sec. 3.9). To see why more
78 CHAPTER 3. PROBABILITY & LIFE TABLES
accurate bounds are needed, consider the case where N = 1000, p = 0.1,
and δ = 0.03. The exact Binomial(1000, 0.1) probability of (number of
successes in N Bernoulli(p) trials falling in) [0, N(p −δ)] ∪ [N(p +δ), N] =
[0, 70] ∪ [130, 1000] is 0.001916, while the upper bound established in the
proof of Theorem 3.1 is (.1)(.9)/(1000(.03)
2
) = 0.1, more than ﬁfty times
larger ! On the other hand, the upper bound provided by the inequalities of
Theorem 3.2, as in formula (3.30), is 0.0198.
Much closer approximations to the exact probabilities for Binomial(N, p)
random variables to fall in intervals around Np are obtained from the Normal
distribution approximation
Pr(a ≤ X ≤ b) ≈ Φ
_
b −Np
_
Np(1 −p)
_
− Φ
_
a −Np
_
Np(1 −p)
_
(3.2)
where Φ is the standard normal distribution function given explicitly in
integral form in formula (3.29) below. This approximation is the DeMoivre
Laplace Central Limit Theorem (Feller vol. 1, 1957, pp. 16873), which
says precisely that the diﬀerence between the left and righthand sides of
(3.2) converges to 0 when p remains ﬁxed, as n → ∞. Moreover, the
reﬁned form of the DeMoivreLaplace Theorem given in the Feller (1957,
p. 172) reference says that each of the ratios of probabilities
Pr(X < a)
_
Φ
_
a −Np
_
Np(1 −p)
_
, Pr(X > b)
__
1 −Φ
_
b −Np
_
Np(1 −p)
__
converges to 1 if the ‘deviation’ ratios (b − Np)/
_
Np(1 −p) and
(a − Np)/
_
Np(1 −p) are of smaller order than N
−1/6
when N gets
large. This result suggests the approximation
Normal approx. = Φ
_
Nδ
_
Np(1 −p)
_
− Φ
_
−Nδ
_
Np(1 −p)
_
(3.3)
for the true binomial probability Pr(X/N − p ≤ δ). In the example
discussed above, with N = 1000, p = 0.1, δ = 0.03, where the exact
Binomial(N, p) probability of X/N −p ≤ δ was 1−.00192 = .99808, the
normal approximation (3.3) is 0.99843.
To give a feeling for the probabilities with which observed lifetable ra
tios reﬂect the true underlying survivalrates, we have collected in Table 3.1
3.1. BINOMIAL VARIABLES & LAW OF LARGE NUMBERS 79
Table 3.1: Probabilities (in col. 6) with which various Binomial(l
x
,
k
p
x
)
random variables lie within a factor 1 ± of their expectations, together
with lower bounds (in Col. 7) for these probabilities derived from the large
deviation inequalities (3.30)(3.31). The ﬁnal column contains the normal
approximations based on (3.3) to the exact probabilities in column 6.
Cohort Age Time Prob. Toler. Pr. within Lower Normal
n = l
x
x k p =
k
p
x
factor 1 ± bound approx.
10000 40 3 0.99 .003 .9969 .9760 .9972
10000 40 5 0.98 .004 .9952 .9600 .9949
10000 40 10 0.94 .008 .9985 .9866 .9985
1000 40 10 0.94 .020 .9863 .9120 .9877
10000 70 5 0.75 .020 .9995 .9950 .9995
1000 70 5 0.75 .050 .9938 .9531 .9938
10000 70 10 0.50 .030 .9973 .9778 .9973
1000 70 10 0.50 .080 .9886 .9188 .9886
various exact binomial probabilities and their counterparts from the approx
imation of (3.3) and the inequality (3.30) of Section 3.9. The illustration
concerns cohorts of lives aged x of various sizes l
x
, together with ‘theo
retical’ probabilities
k
p
x
with which these lives will survive for a period of
k = 1, 5, or 10 years. The probability experiment determining the size of
the surviving cohort l
x+k
is modelled as the tossing of l
x
independent coins
with common headsprobability
k
p
x
: then the surviving cohortsize l
x+k
is viewed as the Binomial (l
x
,
k
p
x
) random variable equal to the number of
heads in those cointosses. In Table 3.1 are given various combinations of
x, l
x
, k,
k
p
x
which might realistically arise in an insurancecompany life
table, together, with the true and estimated (from Theorem 3.2 and from
(3.3)) probabilities with which the ratios l
x+k
/l
x
agree with
k
p
x
to within
a fraction of the latter. Columns 6 and 7 in the Table show how likely the
lifetable ratios are to be close to the ‘theoretical’ values, but also show that
the lower bounds, while also often close to 1, are still noticeably smaller
than the actual values.
Although the deviationratios in estimating lifetable probabilities are
often close to or larger than N
−1/6
, not smaller as they should be for appli
cability of (3.3), the normal approximations in the ﬁnal column of Table 3.1
80 CHAPTER 3. PROBABILITY & LIFE TABLES
below are sensationally close to the correct binomial probabilities in column
6. A still more reﬁned theorem which justiﬁes this is given by Feller (1972,
section XVI.7 leading up to formula 7.28, p. 553).
3.2 Simulation of Discrete Lifetimes
We began by regarding lifetable ratios l
x
/l
0
in large cohort lifetables as
deﬁning integerage survival probabilities S(x) =
x
p
0
. We said that if the
lifetable was representative of a larger population of prospective insureds,
then we could imagine a newly presented life aged x as being randomly
chosen from the lifetable cohort itself. We motivated the conditional prob
ability ratios in this way, and similarly expectations of functions of lifetable
deathtimes were averages over the entire cohort. Although we found the
calculusbased formulas for lifetable conditional probabilities and expec
tations to be useful, at that stage they were only ideal approximations of
the more detailed but still exact lifetable ratios and sums. At the next
stage of sophistication, we began to describe the (conditional) probabilities
t
p
x
≡ S(x+t)/S(x) based upon a smooth survival function S(x) as a true
but unknown survival distribution, hypothesized to be of one of a number
of possible theoretical forms, governing each member of the lifetable cohort
and of further prospective insureds. Finally, the lifetable can also be viewed
as in Appendix A as an idealized set of data, with each ratio l
x+t
/l
x
equal to
the relative frequency of success among a set of l
x
imagined Bernoulli (
t
p
x
)
trials which Nature performs upon the cohort of lives aged x . With the
mathematical justiﬁcation of the Law of Large Numbers, we come full circle:
these relative frequencies are random variables, but they are not very ran
dom. That is, if the size l
x
of the cohort of surviving lives aged x is large,
the later fractions l
x+t
/l
x
of survivors at x+t to those at x are extremely
likely to lie within a very small tolerance of
t
p
x
. The Law of Large Num
bers applies equally when the agex survivors have been sampled by some
more complicated method than simply watching a cohort from birth. Thus,
in the realistic datacollection scenarios discussed in Appendix A, where the
sizes l
x
of lives under observation at age x are large but the probabilities
p
x
are unknown, the lifetable ratios l
x+1
/l
x
are highly accurate statistical
estimators of the lifetable probabilities .
3.2. SIMULATION OF DISCRETE LIFETIMES 81
Table 3.2: Illustrative Real and Simulated LifeTable Data
Age x l
x
in 195961 LifeTable Simulated l
∗
x
9 96801 96753
19 96051 95989
29 94542 94428
39 92705 92576
49 88178 87901
59 77083 76793
69 56384 56186
79 28814 28657
To make this discussion more concrete, we illustrate the diﬀerence be
tween the entries in a lifetable and the entries one would observe as data
in a randomly generated lifetable of the same size using the initial lifetable
ratios as exact survival probabilities. We used as a source of lifetable counts
the Mortality Table for U.S. White Males 195961 reproduced as Table 2 on
page 11 of C. W. Jordan’s (1967) book on Life Contingencies. That is, using
this Table with radix l
0
= 10
5
, with counts l
x
given for integer ages x from
1 through 80, we treated the probabilities p
x
= l
x+1
/l
x
for x = 0, . . . , 79 as
the correct oneyear survival probabilities for a second, computersimulated
cohort lifetable with radix l
∗
0
= 10
5
. Using simulated random variables
generated in R, we successively generated, as x runs from 1 to 79, random
variables l
∗
x+1
∼ Binomial (l
∗
x
, p
x
). In other words, the mechanism of sim
ulation of the sequence l
∗
0
, . . . , l
∗
79
was to make the variable l
∗
x+1
depend
on previously generated l
∗
1
, . . . , l
∗
x
only through l
∗
x
, and then to generate
l
∗
x+1
as though it counted the heads in l
∗
x
independent cointosses with
headsprobability p
x
. A comparison of the actual and simulated lifetable
counts for ages 9 to 79 in 10year intervals, is given below. The complete
simulated lifetable was given earlier as Table 1.1.
The implication of Table 3.2 is unsurprising: with radix as high as 10
5
,
the agreement between the initial and randomly generated lifetable counts
is extremely good. The Law of Large Numbers guarantees good agreement,
with very high probability, between the ratios l
x+10
/l
x
(which here play
the role of the probability
10
p
x
of success in l
∗
x
Bernoulli trials) and the
corresponding simulated random relative frequencies of success l
∗
x+10
/l
∗
x
. For
example, with x = 69, the ﬁnal simulated count of l
∗
79
= 28657 lives
82 CHAPTER 3. PROBABILITY & LIFE TABLES
aged 79 is the successcount in l
∗
69
= 56186 Bernoulli trials with success
probability 28814/56384 = .51103. With this successprobability, the nor
mal approximation (3.3) says that the simulated count l
∗
79
will diﬀer from
.51103 · 56186 = 28712.8 by 300 or more in either direction with probability
approximately 0.0115. (The exact binomial probabilty of the same event is
0.0113.).
The R code used to generate Table 3.2 is very simple. If lvec denotes a
vector of values (l
0
, l
x(1)
, l
x(2)
, . . . , l
x(K)
) of numbers of surviving lives in a
cohort lifetable with radix l
0
, where the integer ages 0, x(1), x(2), . . . , x(K)
are not necessarily evenly spaced, then the statements
K = length(lvec)1 ; pvec = lvec[2:(K+1)]/lvec[1:K]
create the vector of hypothetical survival probabilities, pvec[j] = l
x(j+1)
/l
x
j
, j =
0, . . . , K − 1, and here is a small function to generate the (K + 1)vector
tt lstar consisting of l
∗
0
≡ l
0
together with the output simulated values
l
∗
x(1)
, . . . , l
x(K)
:
LifTabSim = function(lvec) {
K = length(lvec)1
lstar = c(lvec[1],rep(0,K))
for (j in 1:K) lstar[j+1] =
rbinom(1,lstar[j],lvec[j+1]/lvec[j])
lstar }
The syntax to generate a vector like the third column of Table 3.2 from the
second, where lvec consists of the radix l
0
= 10
5
concatenated with column
2, is: LifTab(lvec)[2:9]. As a further example of such a simulation,
suppose that 1000 individuals aged 40 have successive probabilities
10
p
x
=
0.85, 0.77, 0.70, 0.65, 0.4 for x = 40, 50, 60, 70, 80, then we can simulate
twice, independently and output in R the numbers of surviving lives at these
ages, as follows:
pvec = c(0.85, 0.77, 0.70, 0.65, 0.4)
Lvec = 1000*cumprod(c(1,pvec))
matrix(c(seq(40,80,10), pvec, LifTabSim(Lvec)[2:6],
LifTabSim(Lvec)[2:6]), nrow=4, ncol=5, byrow=T,
3.2. SIMULATION OF DISCRETE LIFETIMES 83
dimnames=list(c("Ages","10_p_x","Sim#1 l_x",
"Sim#2 l_x"), NULL))
[,1] [,2] [,3] [,4] [,5]
Ages 40.00 50.00 60.0 70.00 80.0
10_p_x 0.85 0.77 0.7 0.65 0.4
Sim#1 l_x 842.00 638.00 450.0 296.00 133.0
Sim#2 l_x 854.00 662.00 483.0 344.00 129.0
From small experiments like this, we can see that the variability in the sim
ulated numbers l
∗
x
is considerable for l
0
of 1000 or less.
Exercise 3.A. With the same probabilities
10
p
x
use R to simulate 10 times
independently the numbers of survivors at ages 40, 50, . . . , 80.
(a). What is the spread between the smallest and largest number surviv
ing at each age across your 10 simulations ?
(b). Regarding your 10 sets of simulated numbers of survivors as indepen
dent datasets, if the underlying lifetable probabilities were unknown, what
would be your best estimate of the probability
20
p
40
?
(c). Combining the 10 simulated datasets you generated in (b), what is
your best estimate of the probability
20
p
40
?
84 CHAPTER 3. PROBABILITY & LIFE TABLES
3.3 Expectation of Discrete Random Variables
The Binomial random variables discussed in this Chapter are examples of
socalled discrete random variables, that is, random variables Z with a
discrete (usually ﬁnite) list of possible outcomes z, with a corresponding list
of probabilities or probability mass function values p
Z
(z) with which each of
those possible outcomes occur. These probabilities p
Z
(z) must be positive
numbers which summed over all possible values z add to 1. In an insur
ance context, think for example of Z as the unforeseeable future damage or
liability upon the basis of which an insurer has to pay some scheduled claim
amount c(Z) to fulﬁll a speciﬁc property or liability insurance policy. The
Law of Large Numbers says that we can have a frequentist operational inter
pretation of each of the probabilities p
Z
(z) with which a claim of size c(z)
is presented. In a large population of N independent policyholders, each
governed by the same probabilities p
Z
(·) of liability occurrences, for each
ﬁxed damageamount z we can imagine a series of N Bernoulli (p
Z
(z))
trials, in which the jth policyholder is said to result in a ‘success’ if he
sustains a damage amount equal to z , and to result in a ‘failure’ otherwise.
The Law of Large Numbers (Theorem 3.7) for these Bernoulli trials says that
the number out of these N policyholders who do sustain damage z is for
large N extremely likely to diﬀer by no more than δN from N p
Z
(z).
Returning to a general discussion, suppose that Z is a discrete random
variable with a ﬁnite list of possible values z
1
, . . . , z
m
, and let c(·) be a
realvalued (nonrandom) cost function such that c(Z) represents an eco
nomically meaningful cost incurred when the random variable value Z is
given. Suppose that a large number N of independent individuals give rise
to respective values Z
j
, j = 1, . . . , N and costs c(Z
1
), . . . , c(Z
N
). Here in
dependent means that the mechanism causing diﬀerent individual Z
j
values
is such that information about the values Z
1
, . . . , Z
j−1
does not change the
(conditional) probabilities with which Z
j
takes on its values, so that for all
j, i, and b
1
, . . . , b
j−1
,
P(Z
j
= z
i
 Z
1
= b
1
, . . . , Z
j−1
= b
j−1
) = p
Z
(z
i
)
Then the Law of Large Numbers, applied as above, says that out of the
large number N of individuals it is extremely likely that approximately
p
Z
(k) · N will have their Z variable values equal to k, where k ranges
over {z
1
, . . . , z
m
}. It follows that the average costs c(Z
j
) over the N
3.3. EXPECTATION OF DISCRETE RANDOM VARIABLES 85
independent individuals — which can be expressed exactly as
N
−1
N
j=1
c(Z
j
) = N
−1
m
i=1
c(z
i
) · #{j = 1, . . . , N : Z
j
= z
i
}
— is approximately given by
N
−1
m
i=1
c(z
i
) · (N p
Z
(z
i
)) =
m
i=1
c(z
i
) p
Z
(z
i
)
In other words, the Law of Large Numbers implies that the average cost
per trial among the N independent trials resulting in random variable
values Z
j
and corresponding costs c(Z
j
) has a welldeﬁned approximate
(actually, a limiting) value for very large N
Expectation of cost = E(c(Z)) =
m
i=1
c(z
i
) p
Z
(z
i
) (3.4)
As an application of the formula for expectation of a discrete random
variable, consider the expected value of a costfunction g(T) of a lifetime
random variable which is assumed to depend on T only through the function
g([T]) of the integer part of T. This expectation was interpreted earlier as
the average cost over all members of the speciﬁed lifetable cohort. Now the
expectation can be veriﬁed to coincide with the lifetable average previously
given, if the probabilities S(j) in the following expression are replaced by
the lifetable estimators l
j
/l
0
. Since P([T] = k) = S(k) − S(k + 1), the
general expectation formula (3.4) yields
E(g(T)) = E(g([T]) =
ω−1
k=0
g(k) (S(k) −S(k + 1)) (3.5)
which, after replacing S(k) − S(k + 1) =
_
k+1
k
f(t) dt and [t] = k for
k ≤ t < k + 1, becomes
ω−1
k=0
g(k)
_
k+1
k
f(t) dt =
ω−1
k=0
_
k+1
k
g([t]) f(t) dt =
_
ω
0
g([t]) f(t) dt
agreeing precisely with formula (1.3). Similarly, evaluating the discrete con
ditional expectation given T ≥ x means applying the formula (3.4) to the
86 CHAPTER 3. PROBABILITY & LIFE TABLES
function c(Z) = g(Z) of the discrete random variable Z = [T] using
the conditional probability mass function P(Z = k) = P([T] = k  T ≥
x) = (S(k) −S(k + 1))/S(x) for all integers k ≥ x (and with probability
0 assigned to all integers k < x.) Then the conditional expectation is
E(g([T])  T ≥ x) =
ω−1
k=x
S(k) −S(k + 1)
S(x)
g(k) =
ω−1
k=x
g(k)
S(x)
_
k+1
k
f(t) dt
or
E(g([T]  T ≥ x) =
ω−1
k=x
_
k+1
k
g([t])
S(x)
f(t) dt =
_
ω
0
g([t])
f(t)
S(x
) dt
agreeing precisely with formula (1.5).
The preceding discussion shows that expectations or conditional expec
tations of functions of wholeyear ages can equivalently be calculated using
the expectation formulas for discrete or continuous random variables. In the
discrete case, however, the expressions require knowledge only of the prob
abilities S(y) of survival for wholeyear or integer ages y. Indeed, in the
preceding (discreteversion) formula, for E(g([T])  T ≥ x), let k ≥ x be
replaced by k = x +j, and express
S(k) −S(k + 1)
S(x)
=
S(x +j)
S(x)
_
1 −
S(x +j + 1
S(x +j)
_
=
j
p
x
(1 −p
x+j
)
Then
E(g([T])  T ≥ x) =
ω−1
k=x
S(k) −S(k + 1)
S(x)
g(k) =
ω−x−1
j=0
j
p
x
(1−p
x+j
) g(x+j)
(3.6)
Just as we did in the context of expectations of functions of the life
table waitingtime random variable T , we can interpret the Expectation as a
weighted average of values (costs, in this discussion) which can be incurred in
each trial, weighted by the probabilities with which they occur. There is an
analogy in the continuousvariable case, where Z would be a random variable
whose approximate probabilities of falling in tiny intervals [z, z + dz] are
given by f
Z
(z)dz, where f
Z
(z) is a nonnegative density function integrating
to 1. In this case, the weighted average of costfunction values c(z) which
arise when Z ∈ [z, z +dz], with approximate probabilityweights f
Z
(z)dz,
is written as a limit of sums or an integral, namely
_
c(z) f(z) dz.
3.3. EXPECTATION OF DISCRETE RANDOM VARIABLES 87
3.3.1 Rules for Manipulating Expectations
We have separately deﬁned expectation for continuous and discrete random
variables. In the continuous case, we treated the expectation of a speciﬁed
function g(T) of a lifetime random variable governed by the survival function
S(x) of a cohort lifetable, as the approximate numerical average of the values
g(T
i
) over all individuals i with data represented through observed lifetime
T
i
in the lifetable. The discrete case was handled more conventionally,
along the lines of a ‘frequentist’ approach to the mathematical theory of
probability. First, we observed that our calculations with Binomial (n, p)
random variables justiﬁed us in saying that the sum X = X
n
of a large
number n of independent cointoss variables
1
, . . . , ,
n
, each of which is
1 with probability p and 0 otherwise, has a value which with very high
probability diﬀers from n·p by an amount smaller than δn, where δ > 0 is
an arbitrarily small number not depending upon n. The Expectation p of
each of the variables
i
is recovered approximately as the numerical average
X/n = n
−1
n
i=1
i
of the independent outcomes
i
of independent trials.
This Law of Large Numbers extends to arbitrary sequences of independent
and identical ﬁnitevalued discrete random variables, saying that
if Z
1
, Z
2
, . . . are independent random variables, in the sense
that for all k ≥ 2 and all numbers r,
P(Z
k
≤ r  Z
1
= z
1
, . . . , Z
k−1
= z
k−1
) = P(Z
1
≤ r)
regardless of the precise values z
1
, . . . , z
k−1
, then for each δ > 0,
as n gets large
P
_
n
−1
n
i=1
c(Z
i
) −E(c(Z
1
)) ≥ δ
_
−→ 0 (3.7)
where, in terms of the ﬁnite set S of possible values of Z ,
E(c(Z
1
)) =
z∈S
c(z) P(Z
1
= z) (3.8)
We do not give any further proof here, but the motivating arguments given,
together with straightforward manipulations using the result of Theorem 3.7,
88 CHAPTER 3. PROBABILITY & LIFE TABLES
are an essentially complete proof of (3.7). It is also a fact that the Law of
Large Numbers given in equation (3.7) continues to hold if the deﬁnition
of independent sequences of random variables Z
i
is suitably generalized, as
long as either
Z
i
are discrete with inﬁnitely many possible values deﬁning a
set S, and the expectation is as given in equation (3.8) above
whenever the function c(z) is such that
z∈S
c(z) P(Z
1
= z) < ∞
or
the independent random variables Z
i
are continuous, all with
the same density f(t) such that P(q ≤ Z
1
≤ r) =
_
r
q
f(t) dt,
and expectation is deﬁned by
E(c(Z
1
)) =
_
∞
−∞
c(t) f(t) dt (3.9)
whenever the function c(t) is such that
_
∞
−∞
c(t) f(t) dt < ∞
All of this shows that there really is no choice in devising an appropri
ate deﬁnition of expectations of costfunctions deﬁned in terms of random
variables Z, whether discrete or continuous. For the rest of this book, and
more generally in applications of probability within actuarial science, we are
interested in evaluating expectations of various functions of random variables
related to the contingencies and uncertain duration of life. Many of these
expectations concern superpositions of random amounts to be paid out after
random durations. The following rules for the manipulation of expectations
arising in such superpositions considerably simplify the calculations. Assume
in what follows that all random payments and times are functions of a single
lifetime random variable T.
3.3. EXPECTATION OF DISCRETE RANDOM VARIABLES 89
(1). If a payment consists of a nonrandom multiple (e.g., faceamount
F) times a random amount c(T), then the expectation of the payment is
the product of F and the expectation of c(T):
Discrete case: E(Fc(T)) =
t
F c(t) P(T = t)
= F
t
c(t) P(T = t) = F · E(c(T))
Continuous case: E(Fc(T)) =
_
F c(t)f(t) dt = F
_
c(t)f(t) dt = F·E(c(T))
(2). If a payment consists of the sum of two separate random payments
c
1
(T), c
2
(T) (which may occur at diﬀerent times, taken into account by
treating both terms c
k
(T) as present values as of the same time), then the
overall payment has expectation which is the sum of the expectations of the
separate payments:
Discrete case: E(c
1
(T) +c
2
(T)) =
t
(c
1
(t) +c
2
(t)) P(T = t)
=
t
c
1
(t) P(T = t) +
t
c
2
(t) P(T = t) = E(c
1
(T)) +E(c
2
(T))
Continuous case: E(c
1
(T) +c
2
(T)) =
_
(c
1
(t) +c
2
(t)) f(t) dt
=
_
c
1
(t) f(t) dt +
_
c
2
(t) f(t) dt = E(c
1
(T)) +E(c
2
(T))
Thus, if an uncertain payment under an insurancerelated contract, based
upon a continuous lifetime variable T with density f
T
, occurs only if
a ≤ T < b and in that case consists of a payment of a ﬁxed amount F
occurring at a ﬁxed time h, then the expected present value under a ﬁxed
nonrandom interestrate i with v = (1 +i)
−1
, becomes by rule (1) above,
E(v
h
F I
[a≤T<b]
) = v
h
F E(I
[a≤T<b]
)
where the indicatornotation I
[a≤T<b]
denotes a random quantity which is
1 when the condition [a ≤ T < b] is satisﬁed and is 0 otherwise. Since
90 CHAPTER 3. PROBABILITY & LIFE TABLES
an indicator random variable has the two possible outcomes {0, 1} like the
cointoss variables
i
above, we conclude that E(I
[a≤T<b]
) = P(a ≤ T <
b) =
_
b
a
f
T
(t) dt, and the expected present value above is
E(v
h
F I
[a≤T<b]
) = v
h
F
_
b
a
f
T
(t) dt
(3). The expectation of a nonnegativeintegervalued random variable
can sometimes be simpliﬁed considerably by means of the following useful
Lemma.
Lemma 3.1 Let Z be a nonnegativeintegervalued random variable. Then
EZ =
∞
j=1
P(Z ≥ j) (3.10)
The Lemma is proved using the rule (FubiniTonelli theorem for double
summation) that the order of a double summation of nonnegative summands
can always be reversed:
EZ =
∞
k=0
k p
Z
(k) =
∞
k=1
k
j=1
p
X
(k) =
∞
j=1
∞
k=j
p
Z
(k) =
∞
j=1
P(Z ≥ j)
3.3.2 Curtate Expectation of Life
One example of a function of the number [T] of whole years of life, whose
conditional expectation is useful and interpretable, is the wholeyear residual
life [T]−x for a life aged x, where x is an integer age. The expectation, nec
essarily conditional on the attained age x, is called curtate mean residual
life or curtate life expectancy,
e
x
= E( [T] −x T ≥ x) =
ω−1
t=x
P(t ≤ T < t + 1)
P(T ≥ x)
(t −x) (3.11)
3.4. INTERPRETING FORCE OF MORTALITY 91
Substituting the formula (3.6) with g(t) = t − x gives this formula in
the alternative form
e
x
= E( [T] −x T ≥ x) =
ω−x−1
j=0
j
p
x
(1 −p
x+j
) j (3.12)
A third useful version of this formula can be found by applying formula
(3.10) of Lemma 3.1 to the nonnegative integer valued random variable Z =
[T] −x with probability masses calculated conditionally given T ≥ x. This
yields
e
x
= E( [T] −x T ≥ x) =
ω−x−1
j=1
P([T] −x ≥ j  T ≥ x) =
ω−x−1
j=1
j
p
x
(3.13)
The extension of these expectation formulas to give mean residual life
times which are not truncated to whole years rests on survival function and
density formulas which specify mortality rates between birthdays. The fol
lowing two sections are devoted to a deeper study of continuous mortality
models and interpolation approximations.
3.4 Interpreting Force of Mortality
This Section consists of remarks, relating the force of mortality for a con
tinuously distributed lifetime random variable T (with continuous density
function f ) to conditional probabilities for discrete random variables. In
deed, for m large (e.g. as large as 4 or 12), the discrete random variable
[Tm]/m gives a close approximation to T and represents the attained age
at death measured in wholenumber multiples of fractions h = one m
th
of a year. (Here [·] continues to denote the greatest integer less than or
equal to its real argument.) Since surviving an additional time t = nh can
be viewed as successively surviving to reach times h, 2h, 3h, . . . , nh, and
since (by the deﬁnition of conditional probability)
P(A
1
∩ · · · ∩ A
n
) = P(A
1
) · P(A
2
A
1
) · · · P(A
n
A
1
∩ · · · ∩ A
n−1
)
we have (with the interpretation A
k
= {T ≥ x +kh} )
nh
p
x
=
h
p
x
·
h
p
x+h
·
h
p
x+2h
· · ·
h
p
x+(n−1)h
92 CHAPTER 3. PROBABILITY & LIFE TABLES
The form in which this formula is most often useful is the case h = 1: for
integers k ≥ 2,
k
p
x
= p
x
· p
x+1
· p
x+2
· · · p
x+k−1
(3.14)
Every continuous waitingtime random variable can be approximated by
a discrete random variable with possible values which are multiples of a
ﬁxed small unit h of time, and therefore the random survival time can
be viewed as the (ﬁrst failure among a) succession of results of a sequence
of independent coinﬂips with successive probabilities
h
p
kh
of heads. By
the Mean Value Theorem applied up to seconddegree terms on the function
S(x +h) expanded about h = 0,
S(x+h) = S(x) +hS
(x) +
h
2
2
S
(x+τh) = S(x) −hf(x) −
h
2
2
f
(x+τh)
for some 0 < τ < 1, if f is continuously diﬀerentiable. Therefore, using
the deﬁnition of µ(x) as f(x)/S(x) given on page 49,
h
p
x
= 1 − h ·
_
S(x) −S(x +h)
hS(x)
_
= 1 − h
_
µ(x) +
h
2
f
(x +τh)
S(x)
_
Going in the other direction, the previously derived formula
h
p
x
= exp
_
−
_
x+h
x
µ(y) dy
_
can be interpreted by considering the fraction of individuals observed to reach
age x who thereafter experience hazard of mortality µ(y) dy on successive
inﬁnitesimal intervals [y, y+dy] within [x, x+h). The lives aged x survive
to age x + h with probability equal to a limiting product of inﬁnitesimal
terms (1−µ(y) dy) ∼ exp(−µ(y) dy), yielding an overall conditional survival
probability equal to the negative exponential of accumulated hazard over
[x, x +h).
3.5 Interpolation Between Integer Ages
Cohort lifetable data l
x
and the probability quantities
j
p
x
derived from
them (for integers j) depend on and are determined by the survival func
tion S(k) values only at integer arguments k. Yet many expectations of
3.5. INTERPOLATION BETWEEN INTEGER AGES 93
functions important in actuarial applications necessarily involve the survival
function values between integer ages. It is possible to approximate these only
because, for all but the very youngest and oldest ages, the survival function
for human lives is very smooth within years of age, with derivatives that are
not dramatically large and themselves do not change rapidly. In terms of
calculus concepts, the function value S(x +t) for integer x and 0 ≤ t < 1
is given approximately by the Taylor series formula
S(x +t) = S(x) + t S
(x) +
1
2
t
2
S
(x +θt)
= S(x) − t f
(x) −
1
2
t
2
f
(x +θt) (3.15)
where θ ∈ (0, 1) in the argument of S
in the third term (the meanvalue
type remainder) of the ﬁrst line, and where the second line uses the deﬁnition
f(x +t) = −S
(x +t) valid at all nonnegative integers x and t ∈ [0, 1).
While we will later use Taylor expansions like (3.15) to approximate
expectations E(g(T)) and conditional expectations E(g(T)  T ≥ x), for
now we focus on understanding what (3.15) says about the approximate
probability distribution of the lifetime variable T within years of age. If
S
= −f
is small, as is undoubtedly true for human lifetime between ages
2 and 75 in modern public health conditions, then it is tempting to im
pose the direct modelling assumption S(x + t) ≡ S(x) −tf(x) for integer
x and t ∈ [0, 1), together with continuity of S at all integer points.
This assumption, often called the actuarial approximation, says that for any
0 ≤ a < b < 1,
P(T ∈ [x+a, x+b)  [T] = x) =
S(x +a) −S(x +b)
(S(x) −S(x + 1)
=
(b −a) f(x)
f(x)
= b−a
In other words, the ‘actuarial approximation’ says that failures known to
occur within the year between the x and x+1 birthdays are actually uniformly
distributed (have constant conditional density of 1) within that year.
The ‘actuarialapproximation’ assumption can be understood either as
piecewise linearity, on exactage intervals [x, x + 1), of the continuous sur
vival function S(y) or equivalently as piecewise constancy of the density
function f(y) = −S
(y). This assumption is by far the most commonly
used one in actuarial work. Two other related possible approximations can
94 CHAPTER 3. PROBABILITY & LIFE TABLES
be obtained by Taylor expanding not S(x+t) itself but rather the functions
log S(x + t) or 1/S(x + t). It turns out that the ﬁrst of these alternative
assumptions, piecewise linearity of log S(x + t) or equivalently piecewise
constancy within intervals x, x +1) of
d
dy
log S(y) = −f(y)/S(y) ≡ µ(y)
is well known and has historically been widely used by biostatisticians. Many
models in biostatistics or reliability have been formulated with piecewise con
stant hazards (recall that biostatisticians call µ(y) the hazard function while
actuaries call it force of mortality). The third assumption introduced here,
that of piecewise linearity of 1/S(y), is called the Balducci hypothesis,
and is studied by actuarial students largely for historical reasons and as a
source of examination problems, since it will be seen immediately below for
mula (3.22) to have properties which make it unsuitable as a realistic model
for survival.
To proceed formally, assume that values S(x) for x = 0, 1, 2, . . . have
been speciﬁed or estimated. Approximations to S(y), f(y) and µ(y)
between integers are usually based on one of the following assumptions:
(i) (Piecewiseuniform density) f(x +t) is constant for 0 ≤ t < 1 ;
(ii) (Piecewiseconstant hazard) µ(x +t) is constant for 0 ≤ t < 1 ;
(iii) (Balducci hypothesis) 1/S(x +t) is linear for 0 ≤ t < 1 .
For integers x and 0 ≤ t ≤ 1,
S(x +t)
−lnS(x +t)
1/S(x +t)
_
_
_
is linear in t under
_
_
_
assumption (i)
assumption (ii)
assumption (iii)
(3.16)
Under assumption (i), the slope of the linear function S(x +t) at t = 0 is
−f(x), which implies easily that S(x +t) = S(x) −tf(x), i.e.,
f(x) = S(x) −S(x + 1) , and µ(x +t) =
f(x)
S(x) −tf(x)
(3.17)
so that under (i),
µ(x +
1
2
) = f
T
(x +
1
2
)
_
S
T
(x +
1
2
) (3.18)
Under (ii), where µ(x +t) = µ(x), (3.18) also holds, and
S(x +t) = S(x) e
−t µ(x)
, and p
k
=
S(x + 1)
S(x)
= e
−µ(x)
3.5. INTERPOLATION BETWEEN INTEGER AGES 95
Under (iii), for 0 ≤ t < 1,
1
S(x +t)
=
1
S(x)
+t
_
1
S(x + 1)
−
1
S(x)
_
(3.19)
When equation (3.19) is multiplied through by S(x + 1) and terms are
rearranged, the result is
S(x + 1)
S(x +t)
= t + (1 −t)
S(x + 1)
S(x)
= 1 − (1 −t) q
x
(3.20)
Recalling that
t
q
x
= 1 − (S(x + t)/S(x)), reveals assumption (iii) to be
equivalent to
1−t
q
x+t
= 1 −
S(x + 1)
S(x +t)
= (1 −t)
_
1 −
S(x + 1)
S(x)
_
= (1 −t) q
x
(3.21)
Next diﬀerentiate the logarithm of the formula (3.20) with respect to t, to
show (still under (iii)) that
µ(x +t) = −
∂
∂t
ln S(x +t) =
q
x
1 −(1 −t)q
x
(3.22)
Apart from any other property which the Balducci interpolation assump
tion (iii) might have, formula (3.19) immediately shows that the withinyear
force of mortality µ(x + t), 0 ≤ t < 1, is actually a decreasing function
a feature which seems particularly unrealistic from middle to advanced ages
within human lifetimes. By contrast, the withinyear force of mortality un
der assumption (i) as given in (3.17) is evidently increasing, and almost by
deﬁnition the piecewiseconstant hazard assumption (ii) entails withinyear
constancy of the force of mortality.
The most frequent insurance application for the interpolation assump
tions (i)(iii) and associated survivalprobability formulas is to express prob
abilities of survival for fractional years in terms of probabilities of wholeyear
survival. In terms of the notations
t
p
x
and q
x
for integers x and 0 < t < 1,
the formulas are:
t
p
x
= 1 −
(S(x) −t(S(x + 1) −S(x))
S(x)
= 1 −t q
x
under (i) (3.23)
t
p
x
=
S(x +t)
S(x)
=
_
e
−µ(x)
_
t
= (1 −q
x
)
t
under (ii) (3.24)
96 CHAPTER 3. PROBABILITY & LIFE TABLES
t
p
x
=
S(x +t)
S(x + 1)
S(x + 1)
S(x)
=
1 −q
x
1 −(1 −t)q
x
under (iii) (3.25)
The application of all of these formulas can be understood in terms of the
formula for expectation of a function g(T) of the lifetime random variable T.
(For a concrete example, think of g(T) = (1 +i)
−T
as the present value to
an insurer of the payment of $1 which it will make instantaneously at the
future time T of death of a newborn which it undertakes to insure.) Then
assumptions (i), (ii), or (iii) via respective formulas (3.23), (3.24), and (3.25)
are used to substitute into the ﬁnal expression of the following formulas:
E
_
g(T)
_
=
_
∞
0
g(t) f(t) dt =
ω−1
x=0
_
1
0
g(t +x) f(t +x) dt
=
ω−1
x=0
S(x)
_
1
0
g(t +x)
_
−
∂
∂t
t
p
x
_
dt
3.5.1 Life Expectancy – Deﬁnition and Approximation
In terms of a survival function f and S modelling the distribution of
exact age at death within years of integer age, we can extend the notion of
expected remaining life from the curtate to the complete expectation for a
life aged x of (T −x) :
complete expectation of life = ˚e
x
= E(T −x T ≥ x) (3.26)
This quantity is also called expected residual life or, in demography, life ex
pectancy. This quantity is larger than the the curtate life expectancy e
x
because, for a life just completing its x’th year and surviving to exact age
x + j + t , with x, j integers and 0 ≤ t < 1 the complete residual life
is T − x = j + t while the curtate residual life is [T] − x = j. Thus by
deﬁnition
˚e
x
− e
x
= E( T −[T]  T ≥ x) ∈ [0, 1)
since this diﬀerence is the expectation or weighted average of a quantity
between 0 and 1.
The integral formula for life expectancy can be written in any of the three
3.6. SOME SPECIAL INTEGRALS 97
ways
˚e
x
=
_
ω
x
(y −x)
f(y)
S(x)
dy =
_
ω−x
0
t
t
p
x
µ(x +t) dt =
_
ω−x
0
t
p
x
(3.27)
Of these expressions, the ﬁrst is the basic conditional expectation formula
(1.5) with g(T) = T −x. The second is obtained from it by the change of
variable t = y −x, using the identities
f(x +t)/S(x) = µ(x +t) S(x +t)/S(x) = µ(x +t)
t
p
x
The third is a continuoustime analogue of Lemma 3.1, obtained from the
second expression in (3.27) via integration by parts, using u = t and dv =
µ(x +t)
t
p
x
dt = −(1/S(x)) d(S(x +t)).
Under the ”actuarial approximation” (assumption (i)) of uniform lifetime
distribution within whole years of age, we saw above that T−[T] is a random
variable with values in [0, 1) which is uniformly distributed (with constant
density 1). Therefore, using the formula (1.4) for expectation, we ﬁnd
under (i): ˚e
x
− e
x
=
_
1
0
t dt =
1
2
(3.28)
There are no formulas nearly as simple for the diﬀerence between complete
and curtate life expectancies under interpolation assumptions (ii) or (iii).
3.6 Some Special Integrals
While actuaries ordinarily do not allow themselves to represent real lifetable
survival distributions by simple ﬁniteparameter families of theoretical distri
butions (for the good reason that such distributions never approximate the
real largesample lifetable data well enough), it is important for the student
to be conversant with several integrals which would arise by substituting
some of the theoretical models into formulas for various net single premiums
and expected lifetimes.
Consider ﬁrst the Gamma functions and integrals arising in connection
with Gamma survival distributions. The Gamma function Γ(α) is deﬁned
by
Γ(α) =
_
∞
0
x
α−1
e
−x
dx , α > 0
98 CHAPTER 3. PROBABILITY & LIFE TABLES
This integral is easily checked to be equal to 1 when α = 1, giving
the total probability for an exponentially distributed random variable, i.e.,
a lifetime with constant forceofmortality 1. For α = 2, the integral
is the expected value of such a unitexponential random variable, and it is
a standard integrationbyparts exercise to check that it too is 1. More
generally, the integral Gamma(α + 1) for positive integer α is the α
th
moment of the Exponential distribution with parameters λ = 1. Integration
by parts in the Gamma integral with u = x
α
and dv = e
−x
dx immediately
yields the famous recursion relation for the Gamma integral, ﬁrst derived
by Euler, and valid for all α > 0 :
Γ(α + 1) =
_
∞
0
x
α
e
−x
dx =
_
−x
α
e
−x
_
¸
¸
¸
∞
0
+
_
∞
0
αx
α−1
e
−x
dx = α · Γ(α)
This relation, applied inductively, shows that for all positive integers n,
Γ(n + 1) = n · (n −1) · · · 2 · Γ(1) = n!
The only other simpletoderive formula explicitly giving values for (non
integer) values of the Gamma function is Γ(
1
2
) =
√
π, obtained as follows:
Γ(
1
2
) =
_
∞
0
x
−1/2
e
−x
dx =
_
∞
0
e
−z
2
/2
√
2 dz
Here we have made the integral substitution x = z
2
/2, x
−1/2
dx =
√
2 dz.
The last integral can be given by symmetry, using the change of variable
u = −z and the fact that the integrand is an even function, to show that
_
0
−∞
e
−z
2
/2
dz =
_
∞
0
e
−u
2
/2
du =
1
2
_
∞
−∞
e
−x
2
/2
dx =
1
2
√
2π =
√
π
√
2
where the last equality is equivalent to the fact (proved in most calculus
texts as an exercise in double integration using change of variable to polar
coordinates) that the standard normal distribution
Φ(x) =
1
√
2π
_
x
−∞
e
−z
2
/2
dz (3.29)
is a bonaﬁde distribution function with limit equal to 1 as x → ∞. The
symmetry of the normal density guarantees that half of its probability is
assigned to each of (−∞, 0) and [0, ∞), so that Φ(0) = 1/2.
3.6. SOME SPECIAL INTEGRALS 99
One of the integrals which arises in calculating expected remaining life
times for Weibulldistributed variables is a Gamma integral, after integration
byparts and a changeofvariable. Recall that the Weibull density with pa
rameters λ, γ is
f(t) = λγ t
γ−1
e
−λt
γ
, t > 0
so that S(x) = exp(−λx
γ
). The expected remaining life for a Weibull
distributed life aged x is calculated, via an integration by parts with u =
t −x and dv = f(t)dt = −S
(t)dt, as
_
∞
x
(t −x)
f(t)
S(x)
dt =
1
S(x)
_
−(t −x) e
−λt
γ
¸
¸
¸
∞
x
+
_
∞
x
e
−λt
γ
dt
_
The ﬁrst term in square brackets evaluates to 0 at the endpoints, and the
second term can be reexpressed via the changeofvariable w = λt
γ
with
(1/γ) w
1/γ−1
dw = λ
1/γ
dt, to give in the Weibull example,
E(T −x T ≥ x) = e
λx
γ 1
γ
λ
−1/γ
_
∞
λx
γ
w
(1/γ)−1
e
−w
dw
= Γ(
1
γ
) e
λx
γ 1
γ
λ
−1/γ
_
1 −G
1/γ
(λx
γ
)
_
where we denote by G
α
(z) the Gamma distribution function with shape
parameter α,
G
α
(z) =
1
Γ(α)
_
z
0
v
α−1
e
−v
dv
and the integral on the righthand side is called the incomplete Gamma
function. Values of G
α
(z) can be found either in published tables which
are now quite dated, or among the standard functions of many mathemati
cal/statistical computer packages, such as Matlab or R. One particular case
of these integrals, the case α = 1/2 , can be recast in terms of the standard
normal distribution function Φ(·). We change variables by v = y
2
/2 to
obtain for z ≥ 0,
G
1/2
(z) =
1
Γ(1/2)
_
z
0
v
−1/2
e
−v
dv =
1
√
π
_
√
2z
0
√
2 e
−y
2
/2
dy
=
_
2
π
·
√
2π · (Φ(
√
2z) −Φ(0)) = 2Φ(
√
2z) −1
100 CHAPTER 3. PROBABILITY & LIFE TABLES
One further expectedlifetime calculation with a common type of distri
bution gives results which simplify dramatically and become amenable to
numerical calculation. Suppose that the lifetime random variable T is as
sumed lognormally distributed with parameters m, σ
2
. Then the expected
remaining lifetime of a life aged x is
E( T −x T ≥ x) =
1
S(x)
_
∞
x
t
d
dt
Φ(
log(t) −log(m)
σ
) dt − x
Now change variables by y = (log(t) − log(m))/σ = log(t/m)/σ, so that
t = me
σy
, and deﬁne in particular
x
=
log(x) −log(m)
σ
Recalling that Φ
(z) = exp(−z
2
/2)/
√
2π , we ﬁnd
E( T −x T ≥ x) =
1
1 −Φ(x
)
_
∞
x
m
√
2π
e
σy−y
2
/2
dy − x
The integral simpliﬁes after completing the square σy − y
2
/2 =
σ
2
/2 −(y −σ)
2
/2 in the exponent of the integrand and changing variables
by z = y −σ. The result is:
E( T −x T ≥ x) =
me
σ
2
/2
1 −Φ(x
)
_
1 −Φ(x
−σ)
_
− x , x
=
log(x/m)
σ
3.7 Exercise Set 3
(1). Show that:
∂
∂x
t
p
x
=
t
p
x
· (µ
x
−µ
x+t
) .
(2). For a certain value of x, it is known that
t
q
x
= kt over the time
interval t ∈ [0, 3], where k is a constant. Express µ
x+2
as a function of
k.
(3). Suppose that an individual aged 20 has random lifetime (= exact age
at death) T with continuous density function
f
T
(t) = 0.02 (t −20) e
−(t−20)
2
/100
, t > 20
3.7. EXERCISE SET 3 101
(a) If this individual has a contract with your company that you must
pay his heirs $10
6
· (1.4 − T/50) on the date of his death between ages 20
and 70, then what is the expected payment ?
(b) If the value of the deathpayment described in (a) should properly be
discounted by the factor exp(−0.08(T −20)) (i.e. by the eﬀective interest
rate of e
.08
−1 per year) to calculate the present value of the payment, then
what is the expected present value of the insurance contract ?
Hint for both parts: After a change of variables, the integral in (a)
can be evaluated in terms of incomplete Gamma integrals
_
∞
c
s
α−1
e
−s
ds,
where the complete Gamma integrals (for c=0) are known yield the Gamma
function Γ(α) = (α −1)!, for integer α > 0.
(4). Suppose that a lifetable mortality pattern is this: from ages 20 through
60, twice as many lives die in each 5year period as in the previous ﬁveyear
period. Find the probability that a life aged 20 will die between exact ages 40
and 50. If the force of mortality can be assumed constant over each ﬁveyear
age period (2024, 2529, etc.), and if you are told that l
60
/l
20
= 0.8, then
ﬁnd the probability that a life aged 20 will survive at least until exact age
48.0 .
(5). Obtain an expression for µ
x
if l
x
= k s
x
w
x
2
g
c
x
, where k, s, w, g, c
are positive constants.
(6). Show that:
_
∞
0
l
x+t
µ
x+t
dt = l
x
.
(7). A man wishes to accumulate $50, 000 in a fund at the end of 20 years.
If he deposits $1000 in the fund at the end of each of the ﬁrst 10 years and
$1000+x in the fund at the end of each of the second 10 years, then ﬁnd x
to the nearest dollar, where the fund earns an eﬀective interest rate of 6% .
(8). Express in terms of annuityfunctions a
(m)
N
the present value of an
annuity of $100 per month paid the ﬁrst year, $200 per month for the second
year, up to $1000 per month the tenth year. Find the numerical value of the
present value if the eﬀective annual interest rate is 7% .
(9). Find upper bounds for the following Binomial probabilities, and com
pare them with the exact values calculated via computer (e.g., using a spread
sheet or exact mathematical function such as pbinom in Splus) :
(a). The probability that in 1000 Bernoulli trials with success
102 CHAPTER 3. PROBABILITY & LIFE TABLES
probability 0.4, the number of successes lies outside the (inclusive) range
[364, 446].
(b). The probability that of 1650 lives aged exactly 45, for whom
20
p
45
= 0.72, no more than 1075 survive to retire at age 65.
(10). If the force of mortality governing a cohort lifetable is such that
µ
t
=
2
1 +t
+
2
100 −t
for real t , 0 < t < 100
then ﬁnd the number of deaths which will be expected to occur between ages
1 and 4, given that the radix l
0
of the lifetable is 10, 000.
(11). Find the expected present value at 5% APR of an investment whose
proceeds will with probability 1/2 be a payment of $10, 000 in exactly 5
years, and with the remaining probability 1/2 will be a payment of $20, 000
in exactly 10 years.
Hint: calculate the respective present values V
1
, V
2
of the payments in each
of the two events with probability 0.5, and ﬁnd the expected value of a discrete
random variable which has values V
1
or V
2
with probabilities 0.5 each.
(12). Derive the formula for the 2’nd and 3’rd moments (that is,
_
f(t) g(t) dt
for g(t) = t
r
, r = 2, 3,) of the Gamma(α, λ) density
f(t) = (λ
α
/Γ(α!)) t
α−1
e
−λt
I
[t≥0]
as a function of parameters α and λ. Hint: change variables by y = λt.
(13). Derive the formula for the 2’nd and 3’rd moments of the
Weibull(α, β) density f(t) = β α t
α−1
e
−β t
α
I
[t≥0]
as a function of parameters α and β. Hint: change variables appropriately
and use the Gamma function.
3.8. WORKED EXAMPLES 103
3.8 Worked Examples
Example 1. Assume that a cohort lifetable population satisﬁes l
0
= 10
4
and
d
x
=
_
_
_
200 for 0 ≤ x ≤ 14
100 for 15 ≤ x ≤ 48
240 for 49 ≤ x ≤ 63
(a) Suppose that an insurer is to pay an amount $100· (64−X) (without
regard to interest or present values related to the timedeferral of the payment)
for a newborn in the lifetable population, if X denotes the attained integer
age at death. What is the expected amount to be paid ?
(b) Find the expectation requested in (a) if the insurance is purchased for
a life currently aged exactly 10 .
(c) Find the expected present value at 4% interest of a payment of $1000
to be made at the end of the year of death of a life currently aged exactly 20.
The ﬁrst task is to develop an expression for survival function and density
governing the cohort lifetable population. Since the numbers of deaths are
constant over intervals of years, the survival function is piecewise linear, and
the lifedistribution is piecewise uniform because the the density is piecewise
constant. Speciﬁcally for this example, at integer values y,
l
y
=
_
_
_
10000 −200y for 0 ≤ y ≤ 15
7000 −100(y −15) for 16 ≤ y ≤ 49
3600 −240(y −49) for 50 ≤ y ≤ 64
It follows that the terminal age for this population is ω = 64 for this
population, and S(y) = 1 − 0.02 y for 0 ≤ y ≤ 15, 0.85 − 0.01 y for
15 ≤ y ≤ 49, and 1.536− .024 y for 49 ≤ y ≤ 64. Alternatively, extending
the function S linearly, we have the survival density f(y) = −S
(y) = 0.02
on [0, 15), = 0.01 on [15, 49), and = 0.024 on [49, 64].
Now the expectation in (a) can be written in terms of the random lifetime
variable with density f as
_
15
0
0.02 · 100 · (64 −[y]) dy +
_
49
15
0.01 · 100 · (64 −[y]) dy
104 CHAPTER 3. PROBABILITY & LIFE TABLES
+
_
64
49
0.024 · 100 · (64 −[y]) dy
The integral has been written as a sum of three integrals over diﬀerent ranges
because the analytical form of the density f in the expectationformula
_
g(y)f(y)dy is diﬀerent on the three diﬀerent intervals. In addition, observe
that the integrand (the function g(y) = 100(64−[y]) of the random lifetime
Y whose expectation we are seeking) itself takes a diﬀerent analytical form
on successive oneyear age intervals. Therefore the integral just displayed can
immediately be seen to agree with the summation formula for the expectation
of the function 100(64−X) for the integervalued random variable X whose
probability mass function is given by
P(X = k) = d
k
/l
0
The formula is
E(g(Y )) = E(100(64 −X)) =
14
k=0
0.02 · 100 · (64 −k) +
48
k=15
0.01 · 100 · (64 −k) +
63
k=49
0.024 · 100 · (64 −k)
Thus the solution to (a) is given (after the changeofvariable j = 64 − k),
by
2.4
15
j=1
j +
49
j=16
j + 2
64
j=50
j
The displayed expressions can be summed either by a calculator program or
by means of the easilychecked formula
n
j=1
j = j(j + 1)/2 to give the
numerical answer $3103 .
The method in part (b) is very similar to that in part (a), except that
we are dealing with conditional probabilities of lifetimes given to be at least
10 years long. So the summations now begin with k = 10, or alterna
tively end with j = 64 − k = 54, and the denominators of the conditional
probabilities P(X = kX ≥ 10) are l
10
= 8000. The expectation in (b)
then becomes
14
k=10
200
8000
· 100· (64−k) +
48
k=15
100
8000
· 100· (64−k) +
63
k=49
240
8000
· 100· (64−k)
3.8. WORKED EXAMPLES 105
which works out to the numerical value
3.0
15
1
j + 1.25
49
16
j + 2.5
54
50
j = $2391.25
Finally, we ﬁnd the expectation in (c) as a summation beginning at k =
20 for a function 1000 · (1.04)
−X+19
of the random variable X with
conditional probability distribution P(X = kX ≥ 20) = d
k
/l
20
for k ≥ 20.
(Note that the function 1.04
−X+19
is the present value of a payment of 1
at the end of the year of death, because the end of the age X year for an
individual currently at the 20
th
birthday is X − 19 years away.) Since
l
20
= 6500, the answer to part (c) is
1000
_
48
k=20
100
6500
(1.04)
19−k
+
63
k=49
240
6500
(1.04)
19−k
_
= 1000
_
1
65
1 −1.04
−29
0.04
+
24
650
1.04
−29
1 −(1.04)
−15
0.04
_
= 392.92
Example 2. Find the change in the expected lifetime of a cohort lifetable
population governed by survival function S(x) = 1−(x/ω) for 0 ≤ x ≤ ω
if ω = 80 and
(a) the force of mortality µ(y) is multiplied by 0.9 at all exact ages
y ≥ 40, or
(b) the force of mortality µ(y) is decreased by the constant amount 0.1
at all ages y ≥ 40.
The force of mortality here is
µ(y) = −
d
dy
ln(1 −y/80) =
1
80 −y
So multiplying it by 0.9 at ages over 40 changes leaves unaﬀected the
density of 1/80 for ages less than 40, and for ages y over 40 changes
the density from f(y) = 1/80 to
f
∗
(y) = −
d
dy
_
S(40) exp(−0.9
_
y
40
(80 −z)
−1
dz)
_
106 CHAPTER 3. PROBABILITY & LIFE TABLES
= −
d
dy
_
0.5 e
0.9 ln((80−y)/40)
_
= −0.5
d
dy
_
80 −y
40
_
0.9
=
0.9
80
(2 −y/40)
−0.1
Thus the expected lifetime changes from
_
80
0
(y/80) dy = 40 to
_
40
0
(y/80) dy +
_
80
40
y
0.9
80
(2 −y/40)
−0.1
dy
Using the change of variable z = 2 − y/40 in the last integral gives the
expected lifetime = 10 + .45(80/.9 −40/1.9) = 40.53.
Example 3. Suppose that you have available to you two investment possi
bilities, into each of which you are being asked to commit $5000. The ﬁrst
investment is a riskfree bond (or bank savingsaccount) which returns com
pound interest of 5% for a 10year period. The second is a ‘junk bond’
which has probability 0.6 of paying 11% compound interest and returning
your principal after 10 years, probability 0.3 of paying yearly interest at
11% for 5 years and then returning your principal of $5000 at the end
of the 10
th
year with no further interest payments, and probability 0.1
of paying yearly interest for 3 years at 11% and then defaulting, paying
no more interest and not returning the principal. Suppose further that the
going rate of interest with respect to which present values should properly
be calculated for the next 10 years will either be 4.5% or 7.5%, each
with probability 0.5. Also assume that the events governing the junk bond’s
paying or defaulting are independent of the true interest rate’s being 4.5%
versus 7.5% for the next 10 years. Which investment provides the better
expected return in terms of current (time0) dollars ?
There are six relevant events, named and displayed along with their prob
abilities in the following table, corresponding to the possible combinations
of true interest rate (Low versus High) and payment scenarios for the junk
bond (Full payment, Partial interest payments with return of principal, and
Default after 3 years’ interest payments):
3.8. WORKED EXAMPLES 107
Event Name Description Probability
A
1
Low ∩ Full 0.30
A
2
Low ∩ Partial 0.15
A
3
Low ∩ Default 0.05
A
4
High ∩ Full 0.30
A
5
High ∩ Partial 0.15
A
6
High ∩ Default 0.05
Note that because of independence (ﬁrst deﬁned in Section 1.1), the prob
abilities of intersected events are calculated as the products of the separate
probabilities, e.g.,
P(A
2
) = P(Low) · P(Partial) = (0.5) · (0.30) = 0.15
Now, under each of the events A
1
, A
2
, A
3
, the present value of the ﬁrst
investment (the riskfree bond) is
5000
_
10
k=1
0.05 (1.045)
−k
+ (1.045)
−10
_
= 5197.82
On each of the events A
4
, A
5
, A
6
, the present value of the ﬁrst investment
is
5000
_
10
k=1
0.05 (1.075)
−k
+ (1.075)
−10
_
= 4141.99
Thus, since
P(Low) = P(A
1
∪ A
2
∪ A
3
) = P(A
1
) +P(A
2
) +P(A
3
) = 0.5
the overall expected present value of the ﬁrst investment is
0.5 · (5197.82 + 4141.99) = 4669.90
Turning to the second investment (the junk bond), denoting by PV the
108 CHAPTER 3. PROBABILITY & LIFE TABLES
present value considered as a random variable, we have
E(PV  A
1
)/5000 = 0.11
10
k=1
(1.045)
−k
+ (1.045)
−10
= 1.51433
E(PV  A
4
)/5000 = 0.11
10
k=1
(1.075)
−k
+ (1.075)
−10
= 1.24024
E(PV  A
2
)/5000 = 0.11
5
k=1
(1.045)
−k
+ (1.045)
−10
= 1.12683
E(PV  A
5
)/5000 = 0.11
5
k=1
(1.075)
−k
+ (1.075)
−10
= 0.93024
E(PV  A
3
)/5000 = 0.11
3
k=1
(1.045)
−k
= 0.302386
E(PV  A
6
)/5000 = 0.11
3
k=1
(1.075)
−k
= 0.286058
Therefore, we conclude that the overall expected present value E(PV ) of
the second investment is
6
i=1
E(PV · I
A
i
) =
6
i=1
E(PV A
i
) P(A
i
) = 5000 · (1.16435) = 5821.77
So, although the ﬁrstinvestment is ‘riskfree’, it does not keep up with inﬂa
tion in the sense that its present value is not even as large as its starting value.
The second investment, risky as it is, nevertheless beats inﬂation (i.e., the
expected present value of the accumulation after 10 years is greater than the
initial face value of $5000) although with probability P(Default ) = 0.10
the investor may be so unfortunate as to emerge (in present value terms)
with only 30% of his initial capital.
3.9. APPENDIX ON LARGE DEVIATION PROBABILITIES 109
3.9 Appendix to Chapter 3:
Large Deviation Probabilities
Theorem 3.2 (Large Deviation Inequalities) Suppose that X is a Bi
nomial (N, p) random variable, denoting the number of successes in N
Bernoulli (p) trials. If 1 > b > p > c > 0, then
P(X ≥ Nb) ≤ exp
_
−N
_
b ln
_
b
p
_
+ (1 −b) ln
_
1 −b
1 −p
_
__
P(X ≤ Nc) ≤ exp
_
−N
_
c ln
_
c
p
_
+ (1 −c) ln
_
1 −c
1 −p
_
__
Proof. After the ﬁrst inequality in (a) is proved, the second inequality will
be derived from it. Since the event [X ≥ Nb] is the union of the disjoint
events [X = k] for k ≥ Nb, which in turn consist of all outcomestrings
(a
1
, . . . , a
N
) ∈ {0, 1}
N
for which
N
j=1
a
j
= k ≥ Nb, a suitable subset of
the binomial probability mass function values p
X
(k) are summed to provide
P(X ≥ Nb) =
k:Nb≤k≤N
P(X = k) =
k≥Nb
_
N
k
_
p
k
(1 −p)
N−k
For every s > 1, this probability is
≤
k≥Nb
_
N
k
_
p
k
(1 −p)
N−k
s
k−Nb
= s
−Nb
k≥Nb
_
N
k
_
(ps)
k
(1 −p)
N−k
≤ s
−Nb
N
k=0
_
N
k
_
(ps)
k
(1 −p)
N−k
= s
−Nb
(1 −p +ps)
N
Here extra terms (corresponding to k < Nb) have been added in the next
tolast step, and the binomial theorem was applied in the last step. The trick
in the proof comes now: since the lefthand side of the inequality does not
involve s while the righthand side does, and since the inequality must be
valid for every s > 1, it remains valid if the righthand side is minimized
over s. The calculus minimum does exist and is unique, as you can check by
calculating that the second derivative in s is always positive. The minimum
110 CHAPTER 3. PROBABILITY & LIFE TABLES
occurs where the ﬁrst derivative of the logarithm of the last expression is 0,
i.e., at s = b(1 −p)/(p(1 −b)). Substituting this value for s yields
P(X ≥ Nb) ≤
_
b (1 −p)
p (1 −b)
_
−Nb
_
1 −p
1 −b
_
N
= exp
_
−N
_
b ln
_
b
p
_
+ (1 −b) ln
_
1 −b
1 −p
__
_
as desired.
The second inequality follows from the ﬁrst. Replace X by Y = N−X.
Since Y also is a count of ‘successes’ in Bernoulli (1 −p) trials, where the
‘successes’ counted by Y are precisely the ‘failures’ in the Bernoulli trials
deﬁning X, it follows that Y also has a Binomial (N, q) distribution, where
q = 1 −p. Note also that c < p implies b = 1 −c > 1 −p = q. Therefore,
the ﬁrst inequality applied to Y instead of X with q = 1−p replacing p
and b = 1 −c, gives the second inequality for P(Y ≥ Nb) = P(X ≤ Nc).
Note that for all r between 0, 1, the quantity r ln
r
p
+(1−r) ln
1−r
1−p
as a
function of r is convex and has a unique minimumof 0 at r = p. Therefore
when b > p > c, the upper bound for N
−1
lnP([X ≥ bN] ∪ [X ≤ cN]) is
strictly negative and does not involve N. For an improved estimate of the
probability bounded in Theorem 3.1, let δ ∈ (0, min(p, 1−p)) be arbitrarily
small, choose b = p +δ, c = p −δ, and combine the inequalities of part (a)
to give the precise estimate:
P(
X
N
−p ≥ δ) ≤ 2 · exp(−Na) (3.30)
where
a = min
_
(p +δ) ln(1 +
δ
p
) + (1 −p −δ) ln(1 −
δ
1−p
) ,
(p −δ) ln(1 −
δ
p
) + (1 −p +δ) ln(1 +
δ
1−p
)
_
> 0 (3.31)
This last inequality gives a much stronger and numerically more useful upper
bound than Theorem 3.1 on the probability with which th relative frequency
of success X/N diﬀers from the true probability p of success by as much
as δ. The probabilities of such large deviations between X/N and δ are
in fact exponentially small as a function of the number N. 2
3.9. APPENDIX ON LARGE DEVIATION PROBABILITIES 111
If the probabilities P(X/N −p ≥ δ) in Theorem 3.1 are generally much
smaller than the upper bounds given for them, then why are those bounds
of interest ? (These are 1 minus the probabilities illustrated in Table 1.)
First, they provide relatively quick handcalculated estimates showing that
large batches of independentcointosses are extremely unlikely to yield rela
tive frequencies ofheads much diﬀerent from the true probability or limiting
relative frequency of heads. Another, more operational, way to render this
conclusion of Theorem 3.1 is that two very large insured cohorts with the
same true survival probabilities are very unlikely to have materially diﬀerent
survival experience. However, as Table 1 illustrates, for practical purposes
the normal approximation to the binomial probabilities of large discrepancies
from the expectation is generally much more precise than the large deviation
bounds of Theorem 3.2.
The bounds given in Theorem 3.2 get small with large N much more
rapidly than the simpler bounds based on the Chebychev inequality used
in proving Theorem 3.1 (cf. Hogg and Tanis 1997). We can tolerate the
apparent looseness in the bounds because in actuarial applications involving
really extreme tail probabilities (e.g. Slud and Hoesman 1989), it can be
shown that the exponential rate of decay as a function of N in the true tail
probabilities P
N
= P(X ≥ Nb) or P(X ≤ Nc) in Theorem 3.2 (i.e., the
constants appearing in square brackets in the exponents on the righthand
sides of the bounds) are exactly the right ones: no larger constants replacing
them could give correct bounds.
112 CHAPTER 3. PROBABILITY & LIFE TABLES
3.10 Useful Formulas from Chapter 3
Binomial(N, p) probability P(X = k) =
_
N
k
_
p
k
(1 −p)
N−k
p. 75
Discrete r.v. Expectation E(c(Z)) =
m
i=1
c(z
i
) p
Z
(z
i
)
p. 85
Nonneg. integervalued r.v. Expectation E(Z) =
∞
j=1
P(Z ≥ j)
p. 90
Curtate life expectancy e
x
=
ω−x−1
j=1
j
p
x
p. 91
k
p
x
= p
x
p
x+1
p
x+2
· · · p
x+k−1
, k ≥ 1 integer
p. 92
k/m
p
x
=
k−1
j=0
1/m
p
x+j/m
, k ≥ 1 integer
p. 92
(i) Piecewise Unif. S(x+t) = tS(x+1)+(1−t)S(x) , x integer , t ∈ [0, 1]
p. 94
3.10. USEFUL FORMULAS FROM CHAPTER 3 113
(ii) Piecewise Const. µ : lnS(x +t) = t lnS(x + 1) + (1 −t) lnS(x)
p. 94
(iii) Balducci hypothesis
1
S(x +t)
=
t
S(x + 1)
+
1 −t
S(x)
p. 94
t
p
x
=
S(x) −t(S(x + 1) −S(x))
S(x)
= 1 −t q
x
under (i)
p. 95
t
p
x
=
S(x +t)
S(x)
=
_
e
−µ(x)
_
t
= (1 −q
x
)
t
under (ii)
p. 95
t
p
x
=
S(x +t)
S(x + 1)
S(x + 1)
S(x)
=
1 −q
x
1 −(1 −t)q
x
under (iii)
p. 96
Complete life expectancy ˚e
x
=
_
ω−x
0
s
p
x
ds
p. 97
Γ(α) =
_
∞
0
x
α−1
e
−x
dx
p. 98
Φ(x) =
1
√
2π
_
x
−∞
e
−z
2
/2
dz
p. 98
114 CHAPTER 3. PROBABILITY & LIFE TABLES
Chapter 4
Expected Present Values of
Insurance Contracts
We are now ready to draw together the main strands of the development so
far: (A) expectations of discrete and continuous random variables deﬁned
as functions of a lifetable waiting time T until death, and (B) discounting
of future payment (streams) based on interestrate assumptions. We ﬁrst
deﬁne the contractual terms of and discuss relations between the major sorts
of insurance, endowment and life annuity contracts, and next to use interest
theory to deﬁne the present value of the contractual payment stream by the
insurer as a nonrandom function of the random individual lifetime T. In
each case, this leads to a formula for the expected present value of the payout
by the insurer, an amount called the net single premium or net single
risk premium of the contract because it is the single cash payment by
the insured at the beginning of the insurance period which would exactly
compensate for the average of the future payments which the insurer will
have to make.
The details of the further mathematical discussion fall into two parts:
ﬁrst, the speciﬁcation of formulas in terms of cohort lifetable quantities for
net single premiums of insurances and annuities which pay only at wholeyear
intervals; and second, the application of the various survival assumptions con
cerning interpolation between whole years of age, to obtain the corresponding
formulas for insurances and annuities which have m payment times per year.
We close this Chapter with a discussion of instantaneouspayment insurance,
115
116 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
continuouspayment annuity, and meanresiduallife formulas, all of which
involve continuoustime expectation integrals. We also relate these expecta
tions with their mpaymentperyear discrete analogues, and compare the
corresponding integral and summation formulas.
Similar discussions can be found in the books Life Contingencies by Jor
dan (1967) and Actuarial Mathematics by Bowers et al. (1997). The ap
proach here diﬀers in unifying concepts by discussing together all of the
diﬀerent contracts, ﬁrst in the wholeyear case, next under the interpolation
assumption (i) with m payment periods per year, and ﬁnally in the instan
taneous case.
4.1 Preliminaries
The topic of study in this Chapter is contracts resulting in contingent pay
ment streams depending on the age at death T of a single individual. There
are three major types of contracts to consider: insurance, life annuities, and
endowments. More complicated kinds of contracts — which we do not dis
cuss in detail — can be obtained by combining (superposing or subtracting)
these in various ways. Of course, other types of insurances on lives do exist,
which pay only when a single life terminates due to a speciﬁed cause or set of
causes (insurances based on multiple decrement tables), or which contractu
ally involve more than one life (for example husbandwife pairs), insurances
and annuities on joint lives. For these further topics, we refer the reader to
Bowers et al. (1997). Only single life contracts without distinctions between
causes of mortality are discussed here.
In what follows, we adopt several uniform notations and assumptions.
Let x denote the initial integer age of the holder of the insurance, life
annuity, or endowment contract, assuming for convenience that the contract
is initiated on the holder’s birthday. Fix a nonrandom eﬀective interest rate
i , and retain the notation v = (1 +i)
−1
, together with the other notations
previously discussed for annuities of nonrandom duration. Next, denote by
m the number of paymentperiods per year, all times being measured from
the date of policy initiation. Thus, for given m, an insurance will pay
oﬀ at the end of the fraction 1/m of a year during which death occurs,
and lifeannuities pay regularly m times per year until the annuitant dies.
4.1. PRELIMINARIES 117
The term or duration n of the contract will always be assumed to be an
integer multiple of 1/m. Note that policy durations are all measured from
policy initiation, and therefore are smaller by x than the exact age of the
policyholder at termination. Thus, we refer to policy time for the life aged
x as the time scale with origin at policy initiation, assumed to be the x
birthday of the policyholder, and at chronological age t for the policyholder,
we say the policy age is t −x.
The random exact age at which the policyholder dies is denoted by T,
and all of the contracts under discussion have the property that T is the
only random variable upon which either the amount or time of payment can
depend. In examples based on m payment periods per year, the amount of
the payment will be assumed to depend on T only through the greatest
integer less than or equal to mT.
If
k
m
≤ T −x <
k + 1
m
, then T
m
≡
[mT]
m
= x +
k
m
(4.1)
denotes the attained age at death measured in completed (1/m)’th years.
As before, the survival function of T is denoted S(t), and the density by
f(t). The probabilities of the various possible occurrences under the policy
are therefore calculated using the conditional probability distribution of T
given that T ≥ x, which has continuous probability density f(t)/S(x) at
all times t ≥ x. When the amounts and times of payments under a contract
depend only on the wholeyear age at death (m = 1), all probabilities and
conditional expectations refer only to the discrete random variable [T] = T
1
and are calculated in terms of the conditional probability mass function,
given for nonnegative integers x, k by
P([T] = x +k  [T] ≥ x) = P(k ≤ T −x < k + 1  T ≥ x)
=
k
p
x
−
k+1
p
x
=
k
p
x
q
x+k
(4.2)
and depends only on the cohort lifetable entries, since the displayed condi
tional probability is precisely d
x+k
/l
x
.
In the setting where the insurance and annuity contracts are formulated
in terms of m possible death and payment periods per year, the probability
calculations necessarily involve interpolations of the survival function S(t)
within whole years of age, but only to values t of the form x + k/m, k
118 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
an integer. Then all conditional expecations are calculated in terms of the
probability mass function of the random variable T
m
given as in (4.1):
P(T
m
= x +
k
m
¸
¸
¸ T
m
≥ x) = P(
k
m
≤ T −x <
k + 1
m
¸
¸
¸ T ≥ x)
=
1
S(x)
_
S(x +
k
m
) −S(x +
k + 1
m
)
_
=
k/m
p
x
−
(k+1)/m
p
x
= P(T ≥ x +
k
m
¸
¸
¸ T ≥ x) · P(T < x +
k + 1
m
¸
¸
¸ T ≥ x +
k
m
)
=
k/m
p
x
·
1/m
q
x+k/m
(4.3)
4.2 Insurance & Life Annuity Contracts
An Insurance contract is an agreement to pay a face amount — perhaps
modiﬁed by a speciﬁed function of the time until death — if the insured, a
life aged x, dies at any time during a speciﬁed period, the term of the policy,
with payment to be made at the end of the 1/m year within which the death
occurs. Usually the payment will simply be the face amount F(0), but for
example in decreasing term policies the payment will decrease linearly as a
function of T
m
over the term of the policy. The insurance is said to be a
wholelife policy if the duration n = ∞, and a term insurance otherwise.)
The general form of this contract, for a speciﬁed term n ≤ ∞, payment
amount function F(·), and number m of possible paymentperiods per
year, is to
pay F(T −x) at T
m
−x +
1
m
time units following
policy initiation, if T ∈ [x, x +n). (Ins
m
)
Specializing to the case m = 1, so that T
m
= [T] is the wholeyear age
at death, the present value of the insurance company’s payment under the
contract (Ins) or (Ins
1
) is
_
F([T −x]) v
[T−x]+1
if x ≤ T < x +n
0 otherwise
(4.4)
4.2. TYPES OF CONTRACTS 119
The simplest and most common case of this contract and formula arise
when the faceamount F(0) = F is the constant amount paid whenever
a death within the term occurs. Then the payment is F, with present value
F v
[T]−x+1
, if x ≤ T < x + n, and both the payment and present value
are 0 otherwise. In this case, with F ≡ 1, the net single premium has the
standard notation A
1
x:n
. When the insurance is wholelife (n = ∞), the
subscript n and bracket
n
and superscript
1
over x are dropped, so
that A
1
x:∞
≡ A
x
.
A Life Annuity contract is an agreement to pay a scheduled payment to
the policyholder at every interval 1/m of a year while the annuitant is alive,
up to a maximum number of nm payments. Again the payment amounts are
ordinarily constant, but in principle any nonrandom timedependent schedule
of payments F(k/m) can be used, where F(s) is a ﬁxed function and s
ranges over multiples of 1/m. To avoid ambiguity, we adopt the convention
that in the ﬁniteterm or temporary life annuities, either F(0) = 0 or
F(n) = 0. In this general setting, the life annuity contract requires the
insurer to
pay amounts F(k/m) at policy times
k
m
,
0 ≤
k
m
≤ T −x, at most nm payments. (LifAnn
m
)
As in the case of annuities certain (the nonrandom annuities discussed in
Chapter 1), we refer to life annuities with ﬁrst payment at time 0 as (life)
annuitiesdue and to those with ﬁrst payment at time 1/m (and therefore
last payment at time n in the case of a ﬁnite term n over which the
annuitant survives) as (life) annuitiesimmediate.
Again specialize to the case m = 1: under the contract (LifAnn) or
(LifAnn
1
), up to n payments are made (since F(0) = 0 or F(n) = 0), and
the present value of the insurance company’s payment under the life annuity
contract is
[T−x]
k=0
F(k) v
k
(4.5)
Here the situation is deﬁnitely simpler in the case where the payment amounts
F(k) are level or constant, either F(k) ≡ F for k = 0, 1, . . . , n − 1, or
F(k) ≡ F for k = 1, 2, . . . , n. In the ﬁrst of these cases, the lifeannuity
due payment stream becomes an annuitydue certain (the kind discussed
120 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
previously under the Theory of Interest) as soon as the random variable T
is ﬁxed. Indeed, if we replace F(k) by 1 for k = 0, 1, . . . , n − 1, and
by 0 for larger indices k, then the present value in equation (4.5) is
¨ a
min([T]−x+1, n)
, and its expected present value (= net single premium) is
denoted ¨ a
x:n
. In the case of temporary life annuitiesimmediate, which have
payments commencing at policy time 1 and continuing annually either until
death or for a total of n payments, the present value formula as a function of
[T] is the certain annuity immediate a
min([T]−x, n)
, since this is the present
value of the pattern of annual unit payments starting at policy time 1 up to
[T −x] or n, whichever comes ﬁrst. The expectedpresent value notation for
temporary life annuities immediate is a
x:n
.
The third major type of insurance contract is the Endowment, which
pays a contractual face amount F = F(0) at the end of n policy years if
the policyholder initially aged x survives to age x+n. This contract is the
simplest, since neither the amount nor the time of payment, only whether the
payment is made at all, is uncertain. The pure endowment contract commits
the insurer to
pay an amount F at policy time n if T ≥ x+n (Endow)
The present value of the pure endowment contract payment is
F v
n
if T ≥ x +n, 0 otherwise (4.6)
The net single premium or expected present value for a pure endowment
contract with face amount F = 1 is denoted A
1
x:n
or
n
E
x
and is evidently
equal to
A
1
x:n
=
n
E
x
= v
n
n
p
x
(4.7)
The other contract frequently referred to in actuarial texts is the Endow
ment Insurance, which for a life aged x and term n is simply the sum
of the pure endowment and the term insurance, both with term n and the
same face amount 1. Here the standard contract with m payment periods
per year and unit level face amount calls for the insurer to
pay 1 at policy time
_
T
m
−x + 1/m if T < x +n
n if T ≥ x +n
(EndIns
m
)
4.2. TYPES OF CONTRACTS 121
Simplifying to the case of a single payment per year (m = 1), we express the
present value of this contract as v
n
on the event [T ≥ n] and as v
[T−x]+1
on the complementary event [T < n]. Note that [T −x] +1 ≤ n whenever
T −x < n. Thus, in both cases, the present value is given by
v
min([T−x]+1, n)
(4.8)
The expected present value of the unit endowment insurance (still in the case
m = 1) is denoted A
x:n
. The notations for the net single premium of the
term insurance and of the pure endowment are intended to be mnemonic,
respectively denoting the portions of the endowment insurance net single
premium respectively triggered by the expiration of life — in which case the
superscript 1 is positioned above the x —or by the expiration of the ﬁxed
term, in which case the superscript 1 is positioned above the term n.
Another example of an insurance contract which does not need separate
treatment, because it is built up simply from the contracts already described,
is the nyear deferred insurance. This policy pays a constant face amount
at the end of the current fraction 1/m year containing the policy time
T − x, but only if death occurs after the deferral time n , i.e., after age
x+n for a new policyholder aged precisely x. When the face amount is 1,
the contractual payout is precisely the diﬀerence between the unit wholelife
insurance and the nyear unit term insurance. When m = 1, the notation
and formula for the net single premium is
n
A
x
= A
x
− A
1
x:n
(4.9)
4.2.1 Formal Relations between Risk Premiums, m = 1
In this subsection, we collect a few useful identities connecting the diﬀerent
types of risk premiums for contracts with m = 1 payment period per year.
These identities therefore hold and can be used in computational formulas
without regard to particular lifetable interpolation assumptions. The ﬁrst,
which we have already seen, is the deﬁnition of endowment insurance as the
superposition of a constantfaceamount term insurance with a pure endow
ment of the same face amount and term. In terms of net single premiums,
this identity is
A
x:n
= A
1
x:n
+ A
1
x:n
(4.10)
122 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
Another important identity concerns the relation between expected present
values of endowment insurances and life annuities. The great generality ofthe
identity arises from the fact we saw in the discussion following (4.5) that, for
a ﬁxed value of the random lifetime T, the present value of the life annuity
due payout coincides with the annuitydue certain, and is given by
¨ a
min([T−x]+1, n)
=
1 −v
min([T−x]+1, n)
d
where the second expression follows from the ﬁrst via formula (2.4). Thus,
the unit life annuitydue has present value which is a simple linear function of
v
min([T−x]+1, n)
which we saw in (4.8) is the present value of the unit endow
ment insurance. Taking expectations (over values of the random variable T,
conditionally given T ≥ x) in the present value formula, and substituting
A
(m)
x:n
as expectation of (4.8), then yields:
¨ a
x:n
= E
x
_
1 −v
min([T−x]+1, n)
d
_
=
1 −A
x:n
d
(4.11)
where recall that E
x
( · ) denotes the conditional expectation E( ·  T ≥ x).
A more common and algebraically equivalent form of the identity (4.11) is
d ¨ a
x:n
+ A
x:n
= 1 (4.12)
To obtain a corresponding identity relating net single premiums for life
annuitiesimmediate to those of endowment insurances, we need to relate
the risk premiums for life annuitiesimmediate to those of life annuitiesdue.
Unlike the case of annuitiescertain (i.e., nonrandomduration annuities), one
cannot simply multiply the present value of the life annuitydue for ﬁxed
T by the discountfactor v in order to obtain the corresponding present
value for the life annuityimmediate with the same term n. The diﬀerence
arises because the payment streams (for the life annuitydue deferred 1 year
and the lifeannuity immediate) end at the same time rather than with the
same number of payments when death occurs before time n. The correct
conversionformula is obtained by treating the life annuityimmediate of term
n as paying, in all circumstances, a present value of 1 (equal to the cash
payment at policy initiation) less than the life annuitydue with term n+ 1.
Taking expectations leads to the formula
a
x:n
= ¨ a
x:n+1
− 1 (4.13)
4.2. TYPES OF CONTRACTS 123
Now, combining this conversionformula with the identity (4.11), we ﬁnd
a
x:n
= ¨ a
x:n+1
− 1 =
1 −A
x:n+1
d
−1 =
1
i
−
1
d
A
x:n+1
(4.14)
and
d a
x:n
+ A
x:n+1
=
d
i
= v (4.15)
In these formulas, we have made use of the deﬁnition 1/d = (1 + i)/i,
leading to the simpliﬁcations
1/d = 1/i + 1 , i/d = 1 +i = v
−1
Since the nyear deferred insurance with risk premium (4.9) pays a beneﬁt
only if the insured survives at least n years, it can alternatively be viewed
as an endowment with beneﬁt equal to a whole life insurance to the insured
(then aged x + n) after n years if the insured lives that long. With
this interpretation, the nyear deferred insurance has net single premium
=
n
E
x
· A
x+n
. This expected present value must therefore be equal to
(4.9), providing the identity:
A
x
− A
1
x:n
= v
n
n
p
x
· A
x+n
(4.16)
4.2.2 Formulas for Net Single Premiums
All of the net single premiums (or risk premiums) considered so far are com
putable completely in terms of of lifetable quantities
j
p
x
and q
x+j
. To
emphasize the fact that these risk premiums depend on cohort life table
quantities alone, this subsection collects the formulas for risk premiums of
the insurance, annuity, and endowment contracts deﬁned above, written ex
plicitly as sums for the case m = 1. Recall for this purpose the conditional
probability mass function (4.2) of [T −x] given T ≥ x.
Here and from now on, for an event B depending on the random lifetime
T, the notation I
B
denotes the socalled indicator random variable which is
equal to 1 whenever T has a value such that the condition B is satisﬁed
and is equal to 0 otherwise.
124 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
I
B
= 1 if condition B holds , = 0 if not
First, the expectation of the present value (4.4) of the random term in
surance payment (with level face value F(0) ≡ 1) is
A
1
x:n
= E
x
_
v
[T−x]+1
I
{T≤x+n}
_
=
n−1
k=0
v
(k+1)/m
k
p
x
q
x+k
(4.17)
The index k in the summation formula denotes the year of policy time
within which death occurs, k ≤ T −x < k +1. The summation itself is the
weighted sum, over all indices k such that k < n, of the present values
v
k+1
to be paid by the insurer in the event that the policy age at death falls
in [k, k +1), multiplied by the probability, given in formula (4.2), that this
event occurs.
Putting together the formula (4.17) with the previous identity (4.10) pro
vides us with a formula for the net single premium of the endowment insur
ance,
A
x:n
=
n−1
k=0
v
k+1
_
k
p
x
−
k+1
p
x
_
+ v
n
n
p
x
(4.18)
Next, to ﬁgure the expected present value of the life annuitydue with term
n, note that payments of 1 occur at all policy ages k, k = 0, . . . , n −1,
for which T −x ≥ k. Therefore, since the present values of these payments
are v
k
and the payment at policy time k is made with probability
k
p
x
,
¨ a
x:n
= E
x
_
n−1
k=0
v
k
I
{T−x≥k}
_
=
n−1
k=0
v
k
k
p
x
(4.19)
Finally the pure endowment has present value
n
E
x
= E
x
_
v
n
I
[T−x≥n]
_
= v
n
x
p
n
(4.20)
Most generally of all, a contract which pays G(k) at policy time k if the
insured life initially aged x survives to age x +k, and which pays F(k) at
4.3. EXTENSION TO MULTIPLE PAYMENTS PER YEAR 125
policy time k+1 if the insured life aged x dies at age T ∈ [x+k, x+k+1),
has net single (risk) premium equal to
E
x
_
ω−x−1
k=0
_
G(k) v
k
I
{T≥x+k}
+ F(k) v
k+1
I
{[T−x]=k}
__
=
ω−x−1
k=0
k
p
x
_
G(k) v
k
+ F(k) v
k+1
q
x+k
_
(4.21)
where P([T −x] = k  T ≥ x) has been expressed as in (4.2). This setting,
where both functions G(k) and F(k) could depend on a ﬁnite term param
eter n, encompasses all of the insurances, life annuities, and endowments
introduced so far in this Chapter.
The formulas (4.17) and (4.19) and (4.21) are benchmarks in the sense
that they represent a complete solution to the problem of determining net
single premiums without the need for interpolation of the lifetable survival
function between integer ages. However the insurance, lifeannuity, and
endowmentinsurance contracts payable only at wholeyear intervals are all
slightly impractical as insurance vehicles. In the next section, we approach
the calculation of net single premiums for the more realistic context of m
periodperyear insurances and life annuities, using only the standard cohort
lifetable data collected by integer attained ages.
4.3 Risk Premiums & Relations, m > 1
At this point, we return to the basic deﬁnitions of the standard insurance,
annuity, and endowment contracts deﬁned above, in order to extend the
theoretical formulas, and identities to cover the case of general mpayment
period per year contracts.
The pure endowment contract (Endow) with present value formula (4.6),
and net single premium notation and formula (4.7), does not require any
separate discussion here, since it involved only a single potential payment
at an integer policy time. It is therefore no diﬀerent for general m than for
m = 1.
126 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
Next we consider the pure term insurance (Ins
m
) with term n and m
payment periods per year, and level face amount. Recall that this contract,
with unit face amount, pays 1 at the end of the 1/m year of death, if
death occurs before policy time n. That is, the payment of 1 occurs at
policy time k/m if k/m ≤ T − x < (k + 1)/m, k/m < n. Accordingly,
the present value of the insurer’s payment is
nm−1
k=0
v
(k+1)/m
I
{k/m≤T−x<(k+1)/m}
(4.22)
and the net single premium or expected present value is
A
(m)1
x:n
=
nm−1
k=0
v
(k+1)/m
k/m
p
x 1/m
q
x+k/m
(4.23)
Here k/m in the summation formula denotes the beginning of the 1/m
year of policy time within which death is to occur. Again the risk premium
summation is the weighted sum, over all indices k such that k/m < n, of
the present values v
(k+1)/m
to be paid by the insurer in the event that the
policy age at death falls in [k/m, (k +1)/m) multiplied by the probability,
given in formula (4.3), that this event occurs.
To ﬁgure the expected present value of the life annuitydue with term n,
note that payments of 1/m occur at all policy ages k/m, k = 0, . . . , nm−1,
for which T − x ≥ k/m. Therefore, since the present values of these
payments are (1/m) v
k/m
and the payment at k/m is made with probability
k/m
p
x
,
¨ a
(m)
x:n
= E
x
_
nm−1
k=0
1
m
v
k/m
I
[T−x≥k/m]
_
=
1
m
nm−1
k=0
v
k/m
k/m
p
x
(4.24)
The useful identities described in Section 4.2.1 above, connecting the
diﬀerent types of risk premiums for contracts with m = 1 payment period
per year, all have extensions to general m payment per year contracts.
The ﬁrst extension, analogous to (4.10), is the deﬁnition of endowment
insurance as the superposition of a constantfaceamount term insurance with
a pure endowment of the same face amount and term. Recall that the m
paymentperiod per year endowment insurance with term n and unit face
4.3. EXTENSION TO MULTIPLE PAYMENTS PER YEAR 127
amount pays 1 at policy time (k +1)/m if k/m ≤ T −x < (k +1)/m for
k = 0, 1, . . . , nm− 1, and pays 1 at policy time n if T ≥ x +n. The
present value of the payout clearly has the single epxression v
min(Tm−x+1/m, n)
.
In terms of net single premiums, the notational identity is
A
(m)
x:n
= A
(m)1
x:n
+ A
(m) 1
x:n
=
m(ω−x)−1
k=0
v
min((k+1)/m, n)
k/m
p
x 1/m
q
x+k/m
(4.25)
=
nm−1
k=0
v
(k+1)/m
k/m
p
x 1/m
q
x+k/m
+ v
n
n
p
x
Again we ﬁnd a formula for the endowment insurance by a combining the
identity (4.25) with the formula (4.23) for Insurance:
A
(m)
x:n
=
nm−1
k=0
v
(k+1)/m
_
k/m
p
x
−
(k+1)/m
p
x
_
+ v
n
n
p
x
(4.26)
The general identity (4.11) concerning the relation between expected
present values of endowment insurances and life annuities also extends straight
forwardly. With m payments per year, and the individual payments of 1/m
again totalling 1 per year, the termn life annuitydue payout is given via
formula (2.4) by
¨ a
(m)
min(Tm−x+1/m, n)
= (1 −v
min(Tm−x+1/m, n)
) / d
(m)
Again the unit life annuitydue has present value which is a simple linear func
tion of the present value v
min(Tm−x+1/m, n)
of the unit endowment insurance.
Taking expectations (over values of the random variable T, conditionally
given T ≥ x) in the present value formula, and substituting the notation
A
(m)
x:n
then yields:
¨a
(m)
x:n
= E
x
_
1 −v
min(Tm−x+1/m, n)
d
(m)
_
=
1 −A
(m)
x:n
d
(m)
(4.27)
where recall that E
x
( · ) denotes the conditional expectation E( ·  T ≥ x).
An algebraically equivalent form of the identity (4.27) is
d
(m)
¨ a
(m)
x:n
+ A
(m)
x:n
= 1 (4.28)
128 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
For multiple payment periods per year, the idea for converting from risk
premiums of life annuitiesdue to life annuities immediate is very similar to
the idea behind the conversion formula (4.13) for m = 1. The payment
stream for the unit annuityimmediate to a life aged x with payment of 1 per
year, term n years, and m payments per year consists of payments 1/m
at each of the policy times k/m such that 1 ≤ k ≤ nm and k/n ≤ T −x.
The corresponding payment stream for an annuitydue with term n +1/m
is exaxtly the same, except that the latter omits the initial payment of 1/m
at time 0. Therefore the respective expected present values a
(m)
x:n
and
¨ a
(m)
x:n+1/m
diﬀer by exactly the present value of that initial payment of 1/m,
establishing the identity
a
(m)
x:n
= ¨ a
(m)
x:n+1/m
−
1
m
= E
x
_
¨ a
(m)
min(Tm−x,n)+1/m
_
(4.29)
From this mpaymentperiod conversion formula, we directly obtain an iden
tity relating the net single premium for life annuitiesimmediate with m
payment periods per year to that of the m payment period endowment
insurances. The result is
a
(m)
x:n
= ¨ a
(m)
x:n+1/m
−
1
m
=
1 −A
(m)
x:n+1/m
d
(m)
−
1
m
=
1
i
(m)
−
1
d
(m)
A
(m)
x:n+1/m
(4.30)
and
d
(m)
a
(m)
x:n
+A
(m)
x:n+1/m
=
d
(m)
i
(m)
= v
1/m
(4.31)
In these formulas, we have made use of the deﬁnition
m/ d
(m)
= (1 + i
(m)
/m)
_
(i
(m)
/m)
leading to the simpliﬁcations
m
d
(m)
=
m
i
(m)
+ 1 ,
i
(m)
d
(m)
= 1 +
i
(m)
m
= v
−1/m
We conclude this section with a general formula extending (4.21). A
contract which, for all integers k = 0, 1, . . . , m(ω−x)−1, pays
1
m
G
x
(k/m)
at policy time k/m if the insured life initially aged x survives to age x+k/m,
4.3. EXTENSION TO MULTIPLE PAYMENTS PER YEAR 129
and which pays F
x
(k/m) at policy time (k +1)/m if the insured life aged
x dies within the exactage interval T ∈ [x +k/m, x +(k +1)/m), has net
single (risk) premium equal to
E
x
_
m(ω−x)−1
k=0
_
1
m
G
x
(k/m) v
k/m
I
{T−x≥k/m}
+F
x
(k/m) v
(k+1)/m
I
{Tm−x=k/m}
__
=
m(ω−x)−1
k=0
k/m
p
x
_
1
m
G
x
(k) v
k/m
+ F
x
(k/m) v
(k+1)/m
1/m
q
x+k/m
_
(4.32)
See the Worked Examples (numbers 3 and 4) for illustrations of numerical
calculations with the standard formulas (4.22) and (4.24), as well as the
general formula (4.32).
The idea behind equation (4.32) can also be used to express the net single
premium of a life insurance or annuity in a varying interest rate environment.
Following the ideas of Sections 1.2.4 and 1.2.5, we know that the present
valueof a unit payment at policy time t under a timevarying instantaneous
interest rate r(s) ≡ exp(δ(s)) − 1 (expressed in terms of the policy time
argument s) is 1/A(t) = exp(−
_
t
0
δ(s) ds). Then the present value of a
term insurance of duration n with level payment amount F at policy
time (k + 1)/m if the insured life aged x dies within the exactage interval
T ∈ [x +k/m, x + (k + 1)/m), 0 ≤ k < nm, is given by
F · exp
_
−
_
(k+1)/m
0
δ(s) ds
_
I
{Tm=x+k/m}
=
F
A((k + 1)/m)
I
{k/m≤T−x<(k+1)/m}
Similarly, the present value of a temporary life annuity due of duration n
which makes level payments of amount
1
m
G at all policy times k/m ≤
min(T −x, n), is
G
m
·
nm−1
j=0
exp
_
−
_
j/m
0
δ(s) ds
_
I
{j/m≤T−x}
=
G
m
nm−1
j=0
1
A(k/m)
I
{j/m≤T−x}
Then the net single premium of a contract which pays G/m at policy time
k/m ≤ n if the insured life initially aged x survives to age x + k/m, and
which pays F at policy time (k + 1)/m ≤ n if the insured life aged x
130 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
dies within the exactage interval T ∈ [x + k/m, x + (k + 1)/m), is the
expectation of the sum of the last two displayed expressions, and is given by
nm−1
k=0
k/m
p
x
_
G
m · A(k/m)
+
F
A((k + 1)/m)
1/m
q
x+k/m
_
(4.33)
4.4 Interpolation Formulas in Risk Premiums
A key issue in understanding the special nature of life insurances and annu
ities with multiple payment periods is the calculation of these probabilities
from the underlying probabilities
j
p
y
(for integers j, y) which can be de
duced or estimated from lifetables. In the present Section, we combine the
Actuarial Assumption — (i) of Chapter 3, saying that deaths are uniformly
distributed within whole year of age — with the insurance and (temporary)
life annuitydue risk premium formulas. In this setting, the number m of
payment periods per year is greater than 1, and by formula (3.23) for all
integers j = 0, 1, . . . , m−1:
j/m
p
x
= 1 −
j/m
q
x
= 1 − (j/m) q
x
so that
j/m
p
x 1/m
q
x+j/m
=
j/m
p
x
−
(j+1)/m
p
x
= (1/m) q
x
For any integer k = bm+j, where 0 ≤ l < m−1 and b is an integer,
k/m
p
x 1/m
q
x+k/m
=
b
p
x
(
j/m
p
x+b
−
(j+1)/m
p
x+b
) = (1/m)
b
p
x
q
x+b
(4.34)
Substituting (4.34) into (4.23) with summation indices k = bm+j, gives
A
(m)1
x:n
=
n−1
b=0
m−1
j=0
v
b+(j+1)/m
b+j/m
p
x 1/m
q
x+b+j/m
=
1
m
n−1
b=0
v
b
b
p
x
q
x+b
m−1
j=0
v
(j+1)/m
=
_
1
m
m−1
j=0
v
(j+1)/m
_
· (1 +i)
_
n−1
b=0
v
b+1
b
p
x
q
x+b
_
4.4. INTERPOLATION FORMULAS IN RISK PREMIUMS 131
The two factors in parentheses in the ﬁnal displayed expression are respec
tively the oneyear annuityimmediate present value a
(m)
1
and the one
paymentperyear term insurance riskpremium A
1
x:n
. Since
(1 +i) a
(m)
1
= (1 +i) (1 −v)/i
(m)
= i/i
(m)
it follows that under interpolation assumption (i),
A
(m)1
x:n
= (i/i
(m)
)
n−1
b=0
v
b+1
b
p
x
q
x+b
= (i/i
(m)
) A
1
x:n
(4.35)
Similarly, formula (4.34) substituted into the temporary life annuity for
mula (4.24) with summation index k = bm+j gives
¨a
(m)
x:n
=
1
m
n−1
b=0
m−1
j=0
v
b+j/m
j/m
p
x+b
·
b
p
x
=
1
m
n−1
b=0
v
b
b
p
x
m−1
j=0
v
j/m
(1−
j
m
q
x+b
)
=
1
m
m−1
j=0
v
j/m
¨ a
x:n
−
1 +i
m
2
_
m−1
j=0
j v
j/m
_
A
1
x:n
(4.36)
This formula can be reduced further in either of two ways. First, one can
appeal to the deﬁnition of increasing temporary annuitydue and refer to the
formula given in paragraph (iv) of Section 2.1.1:
1
m
2
m−1
j=0
j v
j/m
=
1
m
2
m−1
j=0
(j + 1) v
j/m
−
1
m
¨ a
(m)
1
= (I
(m)
¨ a)
(m)
1
−
1 −v
md
(m)
=
1
d
(m)
_
1 −v
d
(m)
− v −
1 −v
m
_
(4.37)
Alternatively, using the identity
A
(m)
x:n
= A
(m)1
x:n
+ v
n
n
p
x
within (4.27), and directly substituting (4.35), yields under (i)
¨ a
(m)
x:n
=
1
d
(m)
_
1 −
i
i
(m)
A
1
x:n
− v
n
n
p
x
_
(4.38)
We leave as an Exercise for the interested reader to verify algebraically, using
(4.37) together with (4.11), that the formulas (4.36) and (4.38) are equal.
132 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
4.5 Continuous Risk Premium Formulas
The present chapter has developed formulas for the net single premiums or
risk premiums of the principal life insurance and annuity contracts, ﬁrst in
the setting of one payment period per year (m = 1) and then in the case
of multiple payment periods (m > 1) per year. In the limit as m gets
large, the risk premium formulas become expected present values calculated
as continuous integrals with respect to survival densities. To recall why this
limit exists, note that for any function g(T) which depends on T only
through the last completed 1/m’th year T
m
= [Tm]/m,
E
x
(g(T)) =
m(ω−x)−1
k=0
g(x +k/m)
k/m
p
x 1/m
p
x+k/m
where we have used (4.3) as the probability mass function for T
m
. The
displayed expectation formula is then also valid with m replaced by any
integer multiple M = mn, n ≥ 1, and has the equivalent expression
M(ω−x)−1
k=0
g(x +k/M)
S(x)
_
x+(k+1)/M
x+k/M
f(t) dt
=
1
S(x)
M(ω−x)−1
k=0
_
x+(k+1)/M
x+k/M
g(t) f(t) dt =
_
ω−x
0
g(x +s)
f(x +s)
S(x)
ds
Assume now that the function g(t) is continuous (and therefore uniformly
continuous) on the bounded lifetime interval [0, ω], so that g(t)−g([tM]/M
can be made uniformly small by choice of a suﬃciently large multiple M of
m. Since the displayed expectation formulas are exactly valid when applied
to the function g([tM]/M, it follows also for the general continuous function
g that
E
x
(g(T)) = lim
M→∞
E
x
(g(T
M
)) =
_
ω−x
0
g(x +s)
f(x +s)
S(x)
ds (4.39)
or, with the substitution f(x +s) = µ(x +s) S(x +s),
E
x
(g(T)) =
_
ω−x
0
g(x +s) µ(x +s)
S(x +s)
S(x)
ds (4.40)
4.5. CONTINUOUS RISK PREMIUM FORMULAS 133
Similar justiﬁcations can be given for these expectation formulas also when
ever g is piecewise continuous. These integral formulas can be used either
to calculate the limiting values of expected present values for insurance con
tracts with large m, or to calculate other expectations of demographic and
biostatistical interest, such as life expectancies.
4.5.1 Continuous Insurance Contracts
So far in this Chapter, all of the expectations considered have been associ
ated with the discretized random lifetime variables [T] and T
m
= [mT]/m.
However, Insurance and Annuity contracts can also be deﬁned with re
spectively instantaneous and continuous payments, as follows. First, an
instantaneouspayment or continuous insurance with facevalue F
is a contract which pays an amount F at the instant of death of the in
sured. (In practice, this means that when the actual payment is made at
some later time, the amount paid is F together with interest compounded
from the instant of death.) As a function of the random lifetime T for
the insured life initially with exact integer age x, the present value of the
amount paid is F · v
T−x
for a wholelife insurance and F · v
T−x
· I
[T<x+n]
for an nyear term insurance. The expected present values or net single pre
miums on a life aged x are respectively denoted A
x
for a wholelife contract
and A
1
x:n
for an nyear term insurance. The continuous life annuity is
a contract which provides continuous payments at rate 1 per unit time for
duration equal to the smaller of the remaining lifetime of the annuitant or
the term of n years. Here the present value of the contractual payments,
as a function of the exact age T at death for an annuitant initially of exact
integer age x, is a
min(T−x, n)
where n is the (possibly inﬁnite) duration
of the life annuity. Recall that
a
K
=
_
∞
0
v
t
I
[t≤K]
dt =
_
K
0
v
t
dt = (1 −v
K
)/δ
is the present value of a continuous payment stream of 1 per unit time of
duration K units, where v = (1 +i)
−1
and δ = ln(1 +i) . The actuarial
notation for the net single premium of the temporary continuous life annuity
is ¯ a
x:n
which simpliﬁes to ¯ a
x
when n = ∞.
The objective of this section is to develop and interpret formulas for the
134 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
continuoustime net single premiums, along with one further quantity which
has been deﬁned as a continuoustime expectation of the lifetime variable T,
namely the mean residual life (also called complete life expectancy)
˚e
x
= E
x
(T −x) for a life aged x.
4.5.2 Integral Formulas
We apply the continuous conditional expectation formulas (4.39) or (4.40)
directly for the three choices
g(y) = y −x , v
y−x
, or v
y−x
· I
{y−x<n}
which respectively have the conditional E
x
(·) expectations
˚e
x
, A
x
, A
1
x:n
For easy reference, the integral formulas for these three cases are:
˚e
x
= E
x
(T −x) =
_
∞
0
s µ(x +s)
s
p
x
ds (4.41)
A
x
= E
x
(v
T−x
) =
_
∞
0
v
s
µ(x +s)
s
p
x
ds (4.42)
A
1
x:n
= E
x
_
v
T−x
I
{T−x≤n}
_
=
_
n
0
v
s
µ(x +s)
s
p
x
ds (4.43)
Next, we obtain two additional formulas, for continuous life annuitiesdue
a
x
and a
x:n
which correspond to E
x
{g(T)} for the two choices
g(t) =
_
ω−x
0
v
y
I
{y+x≤t}
dy or
_
n
0
v
y
I
{y+x≤t}
dy
where we naturally assume that x+n ≤ ω in the case of the temporary life
annuity.
4.5. CONTINUOUS RISK PREMIUM FORMULAS 135
After switching the order of the integrals and the conditional expecta
tions, and evaluating the conditional expectation of an indicator as a condi
tional probability, in the form
E
x
_
I
{y≤T−x}
_
= P(T ≥ x +s  T ≥ x) =
S(x +y)
S(x)
=
y
p
x
the resulting two equations become
a
x
= E
x
__
ω−x
0
v
y
I
{y≤T−x}
dt
_
=
_
ω−x
0
v
y
y
p
x
dy (4.44)
a
x:n
= E
x
__
n
0
v
y
I
{y≤T−x}
dy
_
=
_
n
0
v
y
y
p
x
dy (4.45)
As seen above in (4.39), risk premiums for continuous insurance and an
nuity contracts have a close relationship to the corresponding contracts with
m payment periods per year for large m. That is, the term insurance net
single premiums
A
(m)1
x:n
= E
x
_
v
Tm−x+1/m
_
approach the continuous insurance value (4.42) as a limit when m → ∞.
A simple direct proof can be given because the payments at the end of the
fraction 1/m of year of death are at most 1/m years later than the
continuousinsurance payment at the instant of death, so that the following
obvious inequalities hold:
A
1
x:n
≤ A
(m)1
x:n
≤ v
1/m
A
1
x:n
(4.46)
Since the righthand term in the inequality (4.46) also converges for large m
to the leftmost term, the middle term which is sandwiched in between must
converge to the same limit (4.43).
For the continuous annuity, (4.45) can be obtained as a limit of formulas
(4.19) either from (4.39) or by using Riemann sums, as the number m of
payments per year goes to ∞, i.e.,
a
x:n
= lim
m→∞
¨ a
(m)
x:n
= lim
m→∞
nm−1
k=0
1
m
v
k/m
k/m
p
x
=
_
n
0
v
t
t
p
x
ds
The ﬁnal formula coincides with (4.45), according with the intuition that the
limit as m →∞ of the paymentstream which pays 1/m at intervals of time
136 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
1/m between 0 and T
m
− x inclusive is the continuous paymentstream
which pays 1 per unit time throughout the policyage interval [0, T −x).
The limiting argument of the previous paragraph shows immediately that
under interpolation assumption (i), there are simple formulas relating
¯
A
x:n
and ¯ a
x:n
to A
1
x:n
. Indeed, by (4.35), under (i)
¯
A
x:n
= lim
m→∞
A
(m)1
x:n
= lim
m
i
i
(m)
A
1
x:n
=
i
δ
A
1
x:n
(4.47)
and by (4.38), also under (i),
¯ a
x:n
= lim
m→∞
¨ a
(m)
x:n
= lim
m
1
d
(m)
_
1 −
i
i
(m)
A
1
x:n
− v
n
n
p
x
_
or
¯ a
x:n
=
1
δ
_
1 −
i
δ
A
1
x:n
− v
n
n
p
x
_
(4.48)
Finally, it is easy to see by passing to the limit m →∞ in (4.28) that
δ ¯ a
x:n
+
¯
A
x:n
= 1 (4.49)
More elaborate relations will be given in the next Chapter between net
single premiumformulas which do require interpolationassumptions for prob
abilities of survival to times between integer ages to formulas for m = 1,
which do not require such interpolation.
The expressions in formulas (4.41), (4.43), and (4.45) can be contrasted
with the respective expectations (3.12), (4.17) and (4.19) for a function of
the integervalued random variable [T] (taking m = 1). In particular,
the complete life expectancy ˚e
x
in (4.41) is compared to the curtate life
expectancy e
x
under interplation assumption (i) in (3.28). The comparison
between complete and curtate life expectancies under more general within
year survival distributions in subsection 4.5.4 below.
4.5.3 Risk Premiums under Theoretical Models
Let us work out examples of the multiple timeperiod per year and continuous
formulas analytically and numerically, under a particular theoretical survival
model.
4.5. CONTINUOUS RISK PREMIUM FORMULAS 137
Consider ﬁrst the slightly artiﬁcial (but still useful) case where the
residual life T −x for a life aged x is precisely Weibull(γ, λ) distributed,
with force of mortality for T given by µ(x +s) = λγ s
γ−1
, and
s
p
x
= e
−λs
γ
for all s ≥ 0 , q
x+k
= 1 −exp(−λ{(k + 1)
γ
−k
γ
})
Then respectively according to formulas (4.17), (4.23), and (4.42), we ﬁnd
formulas for terminsurance ﬁnitem net single premiums under Weibull
survival as follows:
A
1
x:n
=
n−1
k=0
v
k+1
(e
−λk
γ
−e
−λ(k+1)
γ
)
A
(m)1
x:n
=
n−1
k=0
v
k+1/m
m−1
j=0
v
j/m
(e
−λ(k+j/m)
γ
− e
−λ(k+(j+1)/m)
γ
)
According to formulas (4.19), (4.24), and (4.44), the special Weibulllifetime
temporary life annuitydue risk premiums for ﬁnite m are:
¨a
x:n
=
n−1
k=0
v
k
e
−λk
γ
, ¨ a
(m)
x:n
=
n−1
k=0
m−1
j=0
v
k+j/m
e
−λ(k+j/m)
γ
The continuous cases (m = ∞) of these formulas are as follows:
¯
A
1
x:n
=
_
n
0
v
s
λγ s
γ−1
e
−λs
γ
ds , ¯ a
x:n
=
_
n
0
v
s
e
−λs
γ
ds
Finally, according to formulas (3.11) and (4.41) the curtate and complete
life expectancies at integer ages y ≥ x for Weibull (residual) lifetimes are:
e
x
=
∞
k=0
k (e
−λk
γ
−e
−λ(k+1)
γ
) , ˚e
x
=
_
∞
0
s λγ s
γ−1
e
λs
γ
ds
We next give R code and a table showing some numerical comparisons of
these insurance and lifeannuity risk premiums, for m = 1, 4, ∞. First, we
give an R function for Weibulllifetime term insurance risk premiums:
138 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
> WeibIns = function(lambda, gamma, m, n, i) {
## Function to calculate termn insurance risk prem
### for 1 paymt per year, m paymts, and continuous ins
v = 1/(1+i)
xk = 0:(n1)
Ins1 = v*sum( v^xk*(exp(lambda*xk^gamma)
exp(lambda*(xk+1)^gamma)) )
xkjm = (0:(n*m1))/m
Insm = v^(1/m)* sum( v^xkjm*(exp(lambda*xkjm^gamma) 
exp(lambda*(xkjm+1/m)^gamma)) )
InsC = lambda*gamma*integrate(function(s,.v,.lam,.gam)
.v^s*s^(.gam1)*exp(.lam*s^(.gam)),0, n,
.v=v, .lam=lambda, .gam=gamma)$val
c(termIns.1 = Ins1, termIns.m = Insm, termIns.cont = InsC)
}
To illustrate the use of this function, we create a small Table of values
for a few diﬀerent values of n, with parameters (λ, γ) similar to those
used in Chapter 2, but chosen successively so that (a)
32
p
40
= .5/.925 =
.5405,
50
p
40
= .04/.925 = .04324 as in Figure 2.5, (b)
32
p
40
= .6,
50
p
40
=
.047, or (c)
32
p
40
= .65,
50
p
40
= .05. In each case, the nominal age x is
40, in the R code to produce this Table, we begin by coding a function to
solve for λ, γ) when π
1
=
32
p
40
, π
2
=
50
p
40
are given. The interest rates
considered here are 0.05 or 0.06, and n = 20.
> LamGam = function(pi1,pi2, age1, age2) {
### Function to solve for lambda and gamma Weibull
### parameters, when S(age1)=pi1, S(age2)=pi2 are fixed.
haz1 = log(pi1) ; haz2 = log(pi2)
gam = log(haz2/haz1)/log(age2/age1)
lam = haz1/age1^gam
c(lambda=lam, gamma=gam) }
> LamGam(.5/.925,.04/.925,32,50)
lambda gamma
1.952e06 3.653e+00
4.5. CONTINUOUS RISK PREMIUM FORMULAS 139
> WeibIns(1.952e6,3.653,4,20,.05)
termIns.1 termIns.m termIns.cont
0.04844 0.04930 0.04960
> InsArr = array(0, dim=c(6,6), dimnames=list(NULL, c("lambda",
"gamma", "i", "Ins.1", "Ins.m", "Ins.C")))
intrat = c(.05, .06)
sprobs = rbind(c(.5/.925,.04/.925),c(.6,.047),c(.65,.05))
for(a in 1:2) for (b in 1:3) {
k = 2*(b1)+a
> lamgam=LamGam(sprobs[b,1],sprobs[b,2],32,50)
InsArr[k,] = c(lamgam[1],lamgam[2],intrat[a],
WeibIns(lamgam[1],lamgam[2],4,20,intrat[a])) }
InsArr
lambda gamma i Ins.1 Ins.m Ins.C
[1,] 1.952e06 3.653 0.05 0.04846 0.04932 0.04962
[2,] 1.952e06 3.653 0.06 0.04189 0.04278 0.04309
[3,] 4.715e07 4.009 0.05 0.03396 0.03457 0.03478
[4,] 4.715e07 4.009 0.06 0.02924 0.02986 0.03008
[5,] 1.241e07 4.345 0.05 0.02436 0.02479 0.02494
[6,] 1.241e07 4.345 0.06 0.02091 0.02135 0.02151
The variations among the parameters and interest rates make much more
diﬀerence to the results than does the number m of payment periods per
year. Note that the overall size of the risk premium for a unit 20year term
insurance is reasonable, because the probability it will pay anything at all is
20
q
40
, which (for oddnumbered rows) takes the three values
> 1exp(AnnuArr[c(1,3,5),1]*20^AnnuArr[c(1,3,5),2])
[1] 0.10461 0.07467 0.05435
A similar Table of risk premiums for temporary life annuitiesdue, with
exactly the same parameters, is given below along with the R code for the
lifeannuity function calculation.
140 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
> WeibAnnD = function(lambda, gamma, m, n, i) {
## Function to calculate termn insurance risk prem
### for 1 paymt per year, m paymts, and continuous ins
v = 1/(1+i)
xk = 0:(n1)
Annu1 = sum( v^xk*exp(lambda*xk^gamma) )
xkjm = (0:(n*m1))/m
Annum = sum( v^xkjm*exp(lambda*xkjm^gamma) )/m
AnnuC = integrate(function(s,.v,.lam,.gam)
.v^s*exp(.lam*s^(.gam)),0, n,
.v=v, .lam=lambda, .gam=gamma)$val
c(tmpAnn.1 = Annu1, tmpAnn.m = Annum, tmpAnn.cont = AnnuC)
}
> AnnuArr = InsArr
dimnames(AnnuArr)[[2]][4:6]=c("Annu.1","Annu.m","Annu.C")
for(a in 1:2) for (b in 1:3) {
k = 2*(b1)+a
lamgam=LamGam(sprobs[b,1],sprobs[b,2],32,50)
AnnuArr[k,4:6] = WeibAnnD(lamgam[1],lamgam[2],4,20,intrat[a]) }
AnnuArr
lambda gamma i Annu.1 Annu.m Annu.C
[1,] 1.952e06 3.653 0.05 12.90 12.65 12.56
[2,] 1.952e06 3.653 0.06 11.99 11.72 11.63
[3,] 4.715e07 4.009 0.05 12.96 12.72 12.64
[4,] 4.715e07 4.009 0.06 12.05 11.78 11.69
[5,] 1.241e07 4.345 0.05 13.00 12.76 12.68
[6,] 1.241e07 4.345 0.06 12.09 11.82 11.73
Life expectancy calculations and comparisons are done in the next sub
section, for Gompertz survival. Examples of formula development for other
special parametric distributional families are contained in the Exercises.
4.5.4 Numerical Calculations of Life Expectancies
Formulas (4.41) or (3.12) respectively provide the complete and curtate age
speciﬁc life expectancies, in terms respectively of survival densities and life
4.5. CONTINUOUS RISK PREMIUM FORMULAS 141
table data. Formula (3.28) provides the actuarial approximation for com
plete life expectancy in terms of lifetable data, based upon interpolation
assumption (i) (Uniform mortality within year of age). In this Section, we
illustrate these formulas using the Illustrative simulated and extrapolated
lifetable data of Table 1.1.
Life expectancy formulas necessarily involve life table data and/or sur
vival distributions speciﬁed out to arbitrarily large ages. While life tables
may be based on large cohorts of insured for ages up to the seventies and even
eighties, beyond that they will be very sparse and very dependent on the par
ticular small group(s) of aged individuals used in constructing the particular
table(s). On the other hand, the fraction of the cohort at moderate ages who
will survive past 90, say, is extremely small, so a reasonable extrapolation
of a wellestablished table out to age 105 or so may give suﬃciently accu
rate lifeexpectancy values at ages not exceeding 105. Life expectancies are
in any case forecasts based upon an implicit assumption of future mortality
following exactly the same pattern as recent past mortality. Lifeexpectancy
calculations necessarily ignore likely changes in living conditions and medi
cal technology which many who are currently alive will experience. Thus an
assertion of great accuracy for a particular method of calculation would be
misplaced.
All of the numerical lifeexpectancy calculations produced for the Figure
of this Section are based on the extrapolation (2.10) of the illustrative life
table data from Table 1.1. According to that extrapolation, deathrates q
x
for all ages 78 and greater are taken to grow exponentially, with log(q
x
/q
78
) =
(x − 78) ln(1.0885). This exponential behavior is approximately but not
precisely compatible with a Gompertzform forceofmortality function
µ(78 +t) = µ(78) c
t
in light of the approximate equality µ(x) ≈ q
x
, an approximation which
progressively becomes less valid as the force of mortality gets larger. To see
this, note that under a Gompertz survival model,
µ(x) = Bc
x
, q
x
= 1 −exp
_
−Bc
x
c −1
lnc
_
and with c = 1.0885 in our setting, (c −1)/ ln c = 1.0436.
142 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
Since curtate life expectancy (3.12) relies directly on (extrapolated) life
table data, its calculation is simplest and most easily interpreted. Figure 4.1
presents, as plotted points, the agespeciﬁc curtate life expectancies for in
teger ages x = 0, 1, . . . , 78. Since the complete life expectancy at each age
is larger than the curtate by exactly 1/2 under interpolation assumption
(i), we calculated for comparison the complete life expectancy at all (real
number) ages, under assumption (ii) (from Chapter 3, as in equation 3.16)
of piecewiseconstant force of mortality within years of age. Under this as
sumption, by formula (3.24), mortality within year of age (0 < t < 1) is
t
p
x+k
= (p
x+k
)
t
, and force of mortality is
µ(x +k +t) = −
d
dt
ln
t
p
x+k
= −ln p
x+k
Using formulas (4.41), (3.12), and interpolation assumption (ii), the exact
formula for the diﬀerence between complete and curtate life expectancy be
comes
˚e
x
− e
x
=
ω−x−1
k=0
k
p
x
_
_
k+1
k
s µ(x +s)
s−k
p
x+k
ds − k q
x+k
_
=
ω−x−1
k=0
k
p
x
_
(−ln p
x+k
)
_
1
0
(k +t) e
t lnp
x+k
dt − k q
x+k
_
=
ω−x−1
k=0
k
p
x
_
(−ln p
x+k
)
_
(k + 1)p
x+k
−k
lnp
x+k
−
p
x+k
−1
(lnp
x+k
)
2
_
− k q
x+k
_
=
ω−x−1
k=0
k
p
x
_
−p
x+k
−
q
x+k
lnp
x+k
_
(4.50)
The complete minus curtate life expectancies calculated from this formula
were found range from 0.493 at ages 40 and below, down to 0.485 at age
78 and 0.348 at age 99. (Contrast this result with the constant diﬀerence
of 1/2 under assumption (i).) Thus there is essentially no new information
in the calculated complete life expectancies, and they are not plotted.
The aspect of Figure 4.1 which is most startling to the intuition is the
large expected numbers of additional birthdays for individuals of advanced
ages. Moreover, the large life expectancies shown are comparable to actual
US male mortality circa 1959, so would be still larger today.
4.5. CONTINUOUS RISK PREMIUM FORMULAS 143
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
Expected number of additional whole years of life, by age
Age in years
C
u
r
t
a
t
e
L
i
f
e
E
x
p
e
c
t
a
n
c
y
0 20 40 60 80
1
0
2
0
3
0
4
0
5
0
6
0
7
0
Figure 4.1: Curtate life expectancy e
x
as a function of age, calculated
from the simulated illustrative life table data of Table 1.1, with agespeciﬁc
deathrates q
x
extrapolated as indicated in formula (2.10).
144 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
4.6 Exercise Set 4
(1). For each of the following three lifetime distributions, ﬁnd (a) the
expected remaining lifetime for an individual aged 20, and (b)
7/12
q
40
/q
40
.
(i) Weibull(.00634, 1.2), with S(t) = exp(−0.00634 t
1.2
),
(ii) Lognormal(log(50), 0.325
2
), with S(t) = 1−Φ((log(t)−log(50))/0.325),
(iii) Piecewise exponential with force of mortality given the constant value
µ
t
= 0.015 for 20 < t ≤ 50, and µ
t
= 0.03 for t ≥ 50. In these
integrals, you should be prepared to use integrations by parts, gamma function
values, tables of the normal distribution function Φ(x), and/or numerical
integrations via calculators or software.
(2). (a) Find the expected present value, with respect to the constant
eﬀective interest rate r = 0.07, of an insurance payment of $1000 to be
made at the instant of death of an individual who has just turned 40 and
whose remaining lifetime T −40 = S is a continuous random variable with
density f(s) = 0.05 e
−0.05s
, s > 0.
(b) Find the expected present value of the insurance payment in (a) if
the insurer is allowed to delay the payment to the end of the year in which
the individual dies. Should this answer be larger or smaller than the answer
in (a) ?
(3). If the individual in Problem 2 pays a life insurance premium P at
the beginning of each remaining year of his life (including this one), then
what is the expected total present value of all the premiums he pays before
his death ?
(4). Suppose that an individual has equal probability of dying within each
of the next 40 years, and is certain to die within this time, i.e., his age is x
and
k
p
x
−
k+1
p
x
= 0.025 for k = 0, 1, . . . , 39
Assume the ﬁxed interest rate r = 0.06.
(a) Find the net single wholelife insurance premium A
x
for this person.
(b) Find the net single premium for the term and endowment insurances
A
1
x:20
and A
x:30
.
4.6. EXERCISE SET 4 145
(5). Show that the expected whole number of years of remaining life for a
life aged x is given by
c
x
= E([T] −x  T ≥ x) =
ω−x−1
k=0
k
k
p
x
q
x+k
and prove that this quantity as a function of integer age x satisﬁes the
recursion equation
c
x
= p
x
(1 + c
x+1
)
(6). Show that the expected present value b
x
of an insurance of 1 payable
at the beginning of the year of death (or equivalently, payable at the end of
the year of death along with interest from the beginning of that same year)
satisﬁes the recursion relation (4.51) above.
(7). Prove the identity (4.16) algebraically.
For the next two problems, consider a cohort lifetable population for
which you know only that l
70
= 10, 000, l
75
= 7000, l
80
= 3000, l
85
= 0, and
that the distribution of deathtimes within 5year age intervals is uniform.
(8). Find (a) ˚e
75
and (b) the probability of an individual aged 70 in
this lifetable population dying between ages 72.0 and 78.0.
(9). Find the probability of an individual aged 72 in this lifetable popula
tion dying between ages 75.0 and 83.0, if the assumption of uniform death
times within 5year intervals is replaced by:
(a) a constant force of mortality within 5year ageintervals;
(b) assuming linearity of 1/S(t)) within 5year age intervals.
(10). Suppose that a population has survival probabilities governed at all
ages by the force of mortality
µ
t
=
_
¸
¸
¸
¸
_
¸
¸
¸
¸
_
.01 for 0 ≤ t < 1
.002 for 1 ≤ t < 5
.001 for 5 ≤ t < 20
.004 for 20 ≤ t < 40
.0001 · t for 40 ≤ t
Then (a) ﬁnd
30
p
10
, and (b) ﬁnd ˚e
50
.
146 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
(11). Suppose that a population has survival probabilities governed at all
ages by the force of mortality
µ
t
=
_
_
_
.01 for 0 ≤ t < 10
.1 for 10 ≤ t < 30
3/t for 30 ≤ t
Then (a) ﬁnd
30
p
20
= the probability that an individual aged 20 survives
for at least 30 more years, and (b) ﬁnd ˚e
30
.
(12). Assuming the same force of mortality as in the previous problem, ﬁnd
˚e
70
and A
60
if i = 0.09.
(13). The force of mortality for impaired lives is three times the standard
force of mortality at all ages. The standard rates q
x
of mortality at ages 95,
96, and 97 are respectively 0.3, 0.4, and 0.5 . What is the probability that
an impaired life age 95 will live to age 98 ?
(14). You are given a survival function S(x) = (10−x)
2
/100 , 0 ≤ x ≤ 10.
(a) Calculate the average number of future years of life for an individual
who survives to age 1.
(b) Calculate the diﬀerence between the force of mortality at age 1, and
the probability that a life aged 1 dies before age 2.
(15). An nyear term life insurance policy to a life aged x provides
that if the insured dies within the nyear period an annuitycertain of yearly
payments of 10 will be paid to the beneﬁciary, with the ﬁrst annuity payment
made on the policyanniversary following death, and the last payment made
on the N
th
policy anniversary. Here 1 < n ≤ N are ﬁxed integers. If
B(x, n, N) denotes the net single premium (= expected present value) for
this policy, and if mortality follows the law l
x
= C(ω − x)/ω for some
terminal integer age ω and constant C, then ﬁnd a simpliﬁed expression
for B(x, n, N) in terms of interestrate functions, ω, and the integers
x, n, N. Assume x +n ≤ ω.
(16). The father of a newborn child purchases an endowment and insurance
contract with the following combination of beneﬁts. The child is to receive
$100, 000 for college at her 18
th
birthday if she lives that long and $500, 000
at her 60
th
birthday if she lives that long, and the father as beneﬁciary is
to receive $200, 000 at the end of the year of the child’s death if the child
4.6. EXERCISE SET 4 147
dies before age 18. Find expressions, both in actuarial notations and in
terms of v = 1/(1 + i) and of the survival probabilities
k
p
0
for the child,
for the net single premium for this contract.
(17). Verify algebraically, using (4.37) together with (4.11), that the right
hand sides of formulas (4.36) and (4.38) are equal.
For the next four problems, concerning insurance contract risk premiums
in a variable interest rate environment, apply formula (4.33).
(18). Suppose that a life aged x wants to purchase a 20year term insurance
now, and that interest rates over the next twenty years will be i = .05 for
policy ages in (0, 10] and i = .06 for policy ages in (10, 20]. Find the net
single premium for a unit term insurance (i.e., if the insurance payment is 1)
if survival is governed by the Weibull(2 · 10
−6
, 4) distribution for T−x. (You
may use the formulas and R code of section 4.5.3 to aid in the calculation.)
(19). Pass to the limit m →∞ in formula (4.33) to derive an analogous
formula for the net single premium, in a variable interest rate environment
with instantaneous force of interest δ(t) at policy time t, of a contract
for a life aged x paying a continuoustime stream at rate G at all policy
times t < min(T −x, n) and paying a lumpsum amount F at the instant
of death if death occurs before age x +n.
(20). Suppose that a life aged x wants to purchase a 10year term insurance
or temporary annuitydue, and that interest rates over the next ten years will
be i = .07 for policy ages in (0, 6] and i = .04 for policy ages in (6, 10].
Show that the net single premium for a unit 10year term insurance or a
unit temporary life annuitydue, with one payment period per year (i.e.,
m = 1) depends on the survival distribution only through the cohort life
table quantities
k
p
x
for integers k = 1, 2, . . . , 10.
(21). Suppose that a life aged x wants to purchase a 10year term in
surance, and that it is believed that interest rates will vary over the 10year
interval according to the rule δ(t) = δ · (1 + 0.002 t). Show that the net
single premium for a unit 10year term insurance depends on the continuous
conditional survival probabilities
t
p
x
, 0 ≤ t ≤ 10 and not only on the values
for integer t.
148 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
4.7 Worked Examples
Example 1. Toy LifeTable (assuming uniform failures)
Consider the following lifetable with only six equallyspaced ages. (That
is, assume l
6
= 0.) Assume that the rate of interest i = .09, so that
v = 1/(1 +i) = 0.9174 and (1 −e
−δ
)/δ = (1 −v)/δ = 0.9582.
x Agerange l
x
d
x
e
x
A
x
0 [0, 1) 1000 60 4.2 0.704
1 [1, 2) 940 80 3.436 0.749
2 [2, 3) 860 100 2.709 0.795
3 [3, 4) 760 120 2.0 0.844
4 [4, 5) 640 140 1.281 0.896
5 [5, 6) 500 500 0.5 0.958
Using the data in this Table, and interest rate i = .09, we begin by cal
culating the expected present values for simple contracts for term insurance,
annuity, and endowment. First, for a life aged 0, a 3year term insurance
with payoﬀ amount $1000 has present value given by formula (4.17) as
1000 A
1
0:3
= 1000
_
0.917
60
1000
+ (0.917)
2
80
1000
+ (0.917)
3
100
1000
_
= 199.60
Second, for a life aged 2, a 3year temporary annuitydue of $700 per year
(with last payment at age 4) has present value computed from (4.19) to be
700 ¨ a
2:3
= 700
_
1 + 0.917
760
860
+ (0.917)
2
640
860
_
= 1705.98
For the same life aged 2, the 3year Endowment for $700 has present value
700 A
1
2:3
= 700 · (0.9174)
3
500
860
= 314.26
Thus we can also calculate (for the life aged 2) the present value of the
3year annuityimmediate of $700 per year as
700 ·
_
¨a
2:3
− 1 + A
1
0:3
_
= 1705.98 −700 + 314.26 = 1320.24
4.7. WORKED EXAMPLES 149
We next apply and interpret the formulas of Section 4.5.1, together with
the observation that
j
p
x
· q
x+j
=
l
x+j
l
x
·
d
x+j
l
x+j
=
d
x+j
l
x
to show how the last two columns of the Table were computed. In particular,
by (4.41)
e
2
=
100
860
· 0 +
120
860
· 1 +
140
860
· 2 +
500
860
· 3 +
1
2
=
1900
860
+ 0.5 = 2.709
Moreover: observe that c
x
=
5−x
k=0
k
k
p
x
q
x+k
satisﬁes the “recursion equa
tion” c
x
= p
x
(1 +c
x+1
) (cf. Exercise 5 above), with c
5
= 0, from which
the e
x
column is easily computed by: e
x
= c
x
+ 0.5.
Now apply the present value formula for continuous insurance to ﬁnd
A
x
=
5−x
k=0
k
p
x
q
x+k
v
k
1 −e
−δ
δ
= 0.9582
5−x
k=0
k
p
x
q
x
v
k
= 0.9582 b
x
where b
x
is the expected present value of an insurance of 1 payable at the
beginning of the year of death (so that A
x
= v b
x
) and satisﬁes b
5
= 1
together with the recursionrelation
b
x
=
5−x
k=0
k
p
x
q
x+k
v
k
= p
x
v b
x+1
+ q
x
(4.51)
(Proof of this recursion is Exercise 6 above.)
Example 2. Find a simpliﬁed expression in terms of actuarial expected
present value notations for the net single premium of an insurance on a
life aged x, which pays F(k) = C ¨ a
n−k
if death occurs at any exact ages
between x + k and x + k + 1, for k = 0, 1, . . . , n − 1, and interpret the
result.
Let us begin with the interpretation: the beneﬁciary receives at the end
of the year of death a lumpsum equal in present value to a payment stream
of C annually beginning at the end of the year of death and terminating at
the end of the n
th
policy year. This payment stream, if superposed upon
150 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
an nyear life annuityimmediate with annual payments C, would result in
a certain payment of C at the end of policy years 1, 2, . . . , n. Thus the
expected present value in this example is given by
C a
n
− C a
x:n
(4.52)
Next we rework this example purely in terms of analytical formulas. By
formula (4.52), the net single premium in the example is equal to
n−1
k=0
v
k+1
k
p
x
q
x+k
C ¨a
n−k+1
= C
n−1
k=0
v
k+1
k
p
x
q
x+k
1 −v
n−k
d
=
C
d
_
n−1
k=0
v
k+1
k
p
x
q
x+k
− v
n+1
n−1
k=0
(
k
p
x
−
k+1
p
x
)
_
=
C
d
_
A
1
x:n
− v
n+1
(1 −
n
p
x
)
_
=
C
d
_
A
x:n
− v
n
n
p
x
− v
n+1
(1 −
n
p
x
)
_
and ﬁnally, by substituting expression (4.15) with m = 1 for A
x:n
, we
have
C
d
_
1 − d ¨ a
x:n
− (1 −v) v
n
n
p
x
− v
n+1
_
=
C
d
_
1 − d (1 +a
x:n
− v
n
n
p
x
) − d v
n
n
p
x
− v
n+1
_
=
C
d
_
v − d a
x:n
− v
n+1
_
= C
_
1 −v
n
i
− a
x:n
_
= C {a
n
− a
x:n
}
So the analytically derived answer agrees with the one intuitively arrived at
in formula (4.52).
Example 3. Consider the following cohort life table fragment applicable to
lives aged from 30 to 36. Find the risk premiums for unitfaceamount 6year
duration annuitydue and term insurance, (a) with m=1, (b) with m = 4
and uniform failure density within year of failure, and (c) with m = 4 and
respective failure probabilities .2, .2, .3, .3 of dying within the 4 quarteryears
given the year of failure. Assume interest rate 5% throughout.
4.7. WORKED EXAMPLES 151
x Agerange l
x
d
x
30 [30, 31) 95000 165
31 [31, 32) 94835 150
32 [32, 33) 94685 155
33 [33, 34) 94530 158
34 [34, 35) 94472 172
35 [35, 36) 94300 187
To clarify the diﬀerent calculations in the three parts (a)(c), we provide
R code as well as numerical answers. Parts (a) and (b) respectively make use
of formulas (4.17), (4.19) and (4.35) plus (4.36).
> probmass = c(165,150,155,158,172,187)/95000
> kpx = 1cumsum(c(0,probmass[1:5]))
> probmass
[1] 0.001737 0.001579 0.001632 0.001663 0.001811 0.001968
> kpx
[1] 1.0000 0.9983 0.9967 0.9951 0.9934 0.9916
> Ains = sum( 1.05^((1:6)) * probmass )
AnnDue = sum( 1.05^((0:5)) * kpx )
c(Ains=Ains, AnnDue=AnnDue)
Ains AnnDue
0.008751 5.308505 ### answer to (a)
> i4 = 4*(1.05^.251)
aux1 = sum(1.05^((0:3)/4))/4
aux2 = sum((0:3)*1.05^((0:3)/4))/4^2
c(Ains.m = Ains*i4/.05,
AnnDue.m = aux1*AnnDue  1.05*aux2*Ains)
Ains.m AnnDue.m
0.008592 5.209397 ### answer to (b)
For part (c), we start from formulas (4.23) and (4.24) and calculate by de
composing sums as we did in Section 4.4,
A
(m)1
x:n
=
n−1
b=0
b
p
x
m−1
j=0
v
b+(j+1)/m
(
(j+1)/m
q
x+b
−
j/m
q
x+b
)
152 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
¨a
(m)
x:n
=
n−1
b=0
b
p
x
1
m
m−1
j=0
v
b+j/m
(1 −
j/m
q
x+b
)
Now for m = 4, we have in this problem the special assumption that for
integers b and 0 ≤ j < 4,
(
j/4
q
x+b
−
(j+1)/4
q
x+b
)/q
x+b
=
_
0.2 if j = 0, 1
0.3 if j = 2, 3
It follows that
j/4
q
x+b
/ q
x+b
has respective values 0, 0.2, 0.4, 0.7 for j =
0, 1, 2, 3. Substituting, we ﬁnd
A
(4)1
x:n
=
n−1
b=0
b
p
x
q
x+b
v
b+1
(1 +i)
_
0.2 v
1/4
+ 0.2 v
1/2
+ 0.3 v
3/4
+ 0.3 v
_
= A
1
x:n
·
_
0.2 v
1/4
+ +0.2 v
1/2
+ 0.3 v
3/4
+ 0.3 v
_
(1 +i)
and
¨ a
(4)
x:n
=
1
4
n−1
b=0
b
p
x
v
b
_
1 − v q
x+b
(0.2 v
−3/4
+ 0.4 v
−1/2
+ 0.7 v
−1/4
)
_
= ¨ a
x:n
−
1
4
A
1
x:n
_
0.2 v
−3/4
+ 0.4 v
−1/2
+ 0.7 v
−1/4
_
The numerics in R for part (c) now follow:
> aux3 = .2*(1.05^.75+1.05^.5)+.3*(1.05^.25 + 1)
aux4 = .2*1.05^.75+.4*1.05^.5+.7*1.05^.25
c(Ains.ptc = Ains*aux3, AnnDue.ptc = AnnDue  0.25*Ains*aux4
Ains.ptc AnnDue.ptc
0.008892 5.305604
So the diﬀerent withinyear distribution of failures made roughly a 2% dif
ference in the term insurance and temporary life annuitydue risk premiums.
4.7. WORKED EXAMPLES 153
Example 4. Find the risk premiums for a 24year life annuitydue and insur
ance ﬁgured with m = 1 and m = 4 based on the probability densities, with
parameters as given in Figure 2.5 in Chapter 2, (a) Gamma(14.74, .4383),
and (b) Lognormal(3.491, (.246)
2
).
Here the ideas and formulas are simple: the only issue is how to organize
the numerical calculations. The Gamma distribution function can be directly
called in R, while the Lognormal distribution function values are simply ex
pressed in terms of the normal distribution function. In neither case need we
perform numerical integrations.
kpx1 = 1pgamma(0:24, rate=.4383, shape=14.74)
kpx2 = 1pnorm(log(0:24),mean=3.491, sd =.246)
array(c(sum(diff(kpx1)/1.05^(1:24)),
sum(diff(kpx2)/1.05^(1:24)),
sum(kpx1[1:24]/1.05^(0:23)),
sum(kpx2[1:24]/1.05^(0:23)),
kpx1[12], kpx2[12], kpx1[24], kpx2[24]), dim=c(2,4),
dimnames=list(c("Gamma","Lognormal"),
c("Ains","AnnDue","P(Tx>12)","P(Tx>24)")))
Ains AnnDue P(Tx>12) P(Tx>24)
Gamma 0.04558 14.36 0.9998 0.9001
Lognormal 0.03504 14.41 1.0000 0.9258
The large relative diﬀerence between the terminsurance risk premiums is
due to the more rapid decrease of the Gamma versus the Lognormal survival
function between 12 and 24 years, which can be seen also in Figure 2.5.
154 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
4.8 Useful Formulas from Chapter 4
T
m
= [Tm]/m = x +
k
m
if x +
k
m
≤ T < x +
k + 1
m
p. 117
P(T
m
= x +
k
m
 T ≥ x) =
k/m
p
x
−
(k+1)/m
p
x
=
k/m
p
x
·
1/m
q
x+k/m
p. 118
Endowment A
1
x:n
=
n
E
x
= E
x
_
v
n
I
[T−x≥n]
_
= v
n
n
p
x
p. 120
A
x:n
= A
1
x:n
+ A
1
x:n
= A
1
x:n
+
n
E
x
p. 121
¨ a
x:n
= E
x
_
1 −v
min([T−x]+1, n)
d
_
=
1 −A
x:n
d
p. 122
d ¨ a
x:n
+ A
x:n
= 1 p. 122
Term (temporary) life annuity a
x:n
= ¨ a
x:n+1/m
− 1/m
p. 122
A
(m)
x
− A
(m)1
x:n
= v
n
n
p
x
· A
x+n
p. 123
Term Insurance A
1
x:n
= E
x
_
v
[T−x]+1
I
{T<x+n}
_
=
n−1
k=0
v
k+1
k
p
x
q
x+k
p. 124
4.8. USEFUL FORMULAS FROM CHAPTER 4 155
¨ a
x:n
= E
x
_
¨ a
min([T−x]+1,n)
_
=
n−1
k=0
v
k
k
p
x
p. 124
A
x:n
=
n−1
k=0
v
k+1
_
k
p
x
−
k+1
p
x
_
+ v
n
n
p
x
p. 124
A
1
x:n
= E
x
_
v
Tm−x+1/m
I
{T<x+n}
_
=
nm−1
k=0
v
(k+1)/m
k/m
p
x 1/m
q
x+k/m
p. 126
A
(m)
x:n
= A
(m)1
x:n
+ A
(m) 1
x:n
= A
(m)1
x:n
+
n
E
x
p. 127
A
(m)
x:n
=
nm−1
k=0
v
(k+1)/m
_
k/m
p
x
−
(k+1)/m
p
x
_
+ v
n
n
p
x
p. 127
¨a
(m)
x:n
= E
x
_
1 −v
min(Tm−x+1/m, n)
d
(m)
_
=
1 −A
(m)
x:n
d
(m)
p. 127
d
(m)
¨ a
(m)
x:n
+ A
(m)
x:n
= 1
p. 127
a
(m)
x:n
= ¨ a
(m)
x:n+1/m
− 1/m
p. 128
156 CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
under (i): A
(m)1
x:n
= (i/i
(m)
)
n−1
b=0
v
b+1
b
p
x
q
x+b
= (i/i
(m)
) A
1
x:n
p. 131
under (i): ¨ a
(m)
x:n
=
1
d
(m)
_
1 −
i
i
(m)
A
1
x:n
− v
n
n
p
x
_
p. 131
˚e
x
= E
x
(T −x) =
_
∞
0
s µ(x +s)
s
p
x
ds
p. 134
A
1
x:n
= E
x
_
v
T−x
I
{T−x≤n}
_
=
_
n
0
v
s
µ(x +s)
s
p
x
ds
p. 134
a
x:n
= E
x
__
n
0
v
y
I
{y≤T−x}
dy
_
=
_
n
0
v
y
y
p
x
dy
p. 135
δ ¯ a
x:n
+
¯
A
x:n
= 1
p. 136
under (ii): ˚e
x
− e
x
=
ω−x−1
k=0
k
p
x
_
−p
x+k
−
q
x+k
lnp
x+k
_
p. 142
Appendix A
Duration Data Structures
In this Chapter, we introduce some of the features of real data structures
embodying waitingtime or duration data. Such data arise in a wide variety
of disciplines and applied ﬁelds, including:
• Life Insurance, where payments are made and received as contractually
determined functions of the duration of an insured individual’s lifetime;
• Casualty Insurance, where the durations of interest are the times until
accident, health emergency, or other adverse occurrence resulting in
liability or loss;
• Other Insurance, such as mortgage insurance relating to the waiting
time until a speciﬁed emergency resulting in
• Clinical Trials and other Biomedical studies, where human lives meet
ing speciﬁc criteria are followed between some initiating event (such
as diagnosis of a disease or a speciﬁc treatment or intervention) and
a response of interest (such as alleviation of symptoms, or death, or
tumor recurrence or return of other disease condition);
• Epidemiology, where larger human populations are followed between
recruitment to a study population
• Reliability, where the object of study is either cumulative time or cu
mulative operational loading in an engineered system until failure or
speciﬁed degradation of performance; and
157
158 APPENDIX A. GENERAL FEATURES OF DURATION DATA
• Economics, where the waiting times of interest are generally times of
transition, such as those for individuals from employment to unem
ployment or vice versa, for businesses from inception to proﬁtability or
bankruptcy, for economies between macroeconomic events, etc.
All of these examples involve the analysis of ‘lifetime’ or ‘waitingtime’
or ‘duration’ data, consisting of a waitingtime random variable T ob
served, incompletely, for many individuals of a study population. All of
them also consider probability distributions and expected values for functions
depending on waitingtime random variables, and for many purposes of sta
tistical anlysis and estimation, reduce the complex data as actually recorded
into the idealized format of the Life Table.
A.1 Concepts and Terminology of Duration
(or Mortality, or Survival) Studies
We next deﬁne and discuss some concepts and terminology that will allow us
to identify common versus distinct aspects of duration data in the diﬀerent
subject areas listed above. We restrict attention in our discussion to stud
ies and datasets concerning individuals, usually people, being observed over
chronological time intervals from entry into the study until the occurrence of
an event of interest – the study endpoint – or the end of followup – called
a rightcensoring time – whichever comes ﬁrst.
Study Population. In a formal observational setting, the study popula
tion is deﬁned thorugh qualifying characteristics. For example, one might
recruit into a clinical trial males with the same disease diagnosis at a des
ignated set of hospital centers, who are between 30 and 60 years of age
and otherwise is good health, and who consent to be randomly assigned to
an experimental versus standard treatment. In an epidemiologic context, the
population might consist of those in certain professions or risky occcupations,
ageintervals, and locations, or who have speciﬁed existing medical conditions
(such as high cholesterol) and consent to participate in a study entailing a
number of scheduled medical examinations. In an insurance context, mortal
ity or timetoevent statistics might be compiled for all individuals insured
by one or a group of companies, over a speciﬁed timewindow, or the subset
A.1. SURVIVAL DATA CONCEPTS 159
of such people subject to a particular risk – such as ‘cigarette smokers’. In
some insurance tabulations, data are gathered on special high or lowrisk
populations in order to justify premiums diﬀerent from those paid by the
general population.
Mode of entry. Depending on the purpose for which timetoevent data
are gathered, the initiating event for the interval of length T can have dif
ferent possible relationships to the chronological time at which an individual
is brought under observation. The simplest case is where these are the same,
or where all individuals in the study are entered simultaneously: this kind of
survival or duration study is called a cohort study. For example, a study
in which babies born in a given year are followed for the next period of (3 or
10 or 20) years would be a cohort study, as would a reliability study in which
100 machines of a given type are set running – possibly under heighted load
or stress – and observed until failure. Another example would be a survival
study in which the interesting duration variable T is ‘time from diagnosis to
death’, and data are to be collected by followup over time on a set of subjects
who receive this diagnosis within a short period, say three months.
More broadly, and with a slightly diﬀerent meaning, the term “cohort
study” applies to longitudinal data collected on a set of individuals selected
simultaneously at the outset, for example in a survey, and then followed over
time. In that usage, ‘cohort’ and ‘longitudinal’ study are roughly synony
mous. As used in an actuarial or demographic context, which is the way we
use it in this book, ‘cohort’ refers to a set of individuals who have the same
wholenumber age at the same time and whose waiting time until death or
other failureevent is of interest. In this way, one could refer to the ‘cohort’
of US males in the state of New Jersey who were 50 years old in 1977.
On the other hand, most survival studies and insurance portfolios consist
of individuals who at any single chronological time have widely diﬀering
current ages. Whether in clinical trials or Insurance, entry of individuals
into observation occurs by staggered entry, at diﬀering chronological times
chosen by the individual. In demography or epidemiologic studies, large
populations are studied beginning at a speciﬁed date, so that all entry times
into the study are simultaneous, but the individuals’ ages at entry vary.
Usually in Insurance and demography and epidemiology, the time variable
of interest is age. Thus, birth is the event initiating the individual’s clock,
160 APPENDIX A. GENERAL FEATURES OF DURATION DATA
but at entry into the datacollection, the individual’s age is recorded. When
the entry age is positive, the individual’s data are said to be leftcensored:
the individual could have been observed to experience the study endpoint
only at an age greater than the entry age.
Mode of study termination. Survival and other duration studies are often
conducted over ﬁxed administrative time windows. Subjects enter either
together, in a cohort, or individually, staggered. The study will end and be
reported as of a ﬁxed chronological termination time, so individuals under
study may have a positive age at entry and age of last followup in the study
without ever having experienced the study endpoint. Moreover, in many
studies as in Insurance portfolios, individuals can withdraw from observation
before the study endpoint, for reasons which may or may not be related to
the nearness of that endpoint. For these reasons, data about the individual’s
variable T may be incomplete and are said to be rightcensored: within the
dataset, the individual’s T is known only to be greater than or equal to the
last age of followup, which is also called that individual’s rightcensoring
time.
Based on the examples and discussion above, we can formulate the follow
ing general data structure for a duration or survival study. If the individuals
in a study are indexed administratively by i = 1, 2, . . . , N, then each indi
vidual must come equipped with at least the following information:
E
i
= chronological time of entry of individual i into the study
A
i
= age of individual i at entry into the study
T
i
= age of individual i at last followup or endpoint under the study
D
i
= binary indicator equal to 1 if i experiences endpoint during followup,
and equal to 0 otherwise
In terms of these notations, individual i ﬁrst enters the study at chrono
logical time E
i
and is under active observation, or under followup, for
a total duration of T
i
− A
i
. Thus the chronological interval of followup is
[E
i
, E
i
+T
i
−A
i
]. The individual’s earliest age in the study is A
i
, and the
latest is T
i
. If D
i
= 1, the ﬁnal age T
i
is also the age at which the study
A.1. SURVIVAL DATA CONCEPTS 161
endpoint is observed to occur for individual i, while if D
i
= 0, then individual
i does not experience the study endpoint while under followup.
The terminology ‘under followup’ is the one used in clinical or epidemi
ologic settings. Expressed in terms of age during the followup period, indi
vidual i would be said to be on test — by adoption of an older terminology
from Reliability — on the age interval [A
i
, T
i
]. The biomedical term would be
that individual i is at risk at ages in the interval (A
i
, T
i
], while the Insurance
term is that the individual is exposed (or ‘exposed to risk’, as might be said
also in an epidemiologic or demographic context) on that age interval.
1 2 3 4 5 6 7 8 9 10
1
2
3
4
5
6
7
8
9
Artificial Clinical Trial with Staggered Entry
Right−Censoring, and Dropout
Time
P
a
t
i
e
n
t
#
Figure A.1: Representation of entry times (solid diamonds), atrisk inter
vals (solid line segments), and death (ﬁlled circle) or rightcensoring (hollow
circle) for each of 9 patients in an artiﬁcial clinical trial.
162 APPENDIX A. GENERAL FEATURES OF DURATION DATA
The deﬁnitions for rightcensored staggeredentry survival data are illus
trated in Figure A.1, for 9 patients in an artiﬁcial clinical trial with stagg
gered entry. Each patient’s interval at risk is a horizontal line segment be
ginning just after the entrytime depicted with a solid diamond and ending
at the time indicated with a circle, solid if an observed death and hollow if
a dropout or censoring time. ‘Dropout’ takes place at the hollow circle for
patient 2, while the other censoring events would be called ‘administrative
rightcensoring’ because the clinical trial observation periods all end at the
chronological time 10 indicated by a vertical dashed line.
The central idea of the Life Table, described in Section A.2 below, is to
tabulate over equally spaced intervals of age, or timeontest, the numbers of
study subjects at risk and observed to fail (i.e., to experience the study end
point while under followup). This idea is fundamental to statistical analysis
of duration data in all of the ﬁelds of study listed at the beginning of this
Chapter.
In this Section, we have addressed the most frequently occurring com
plexities of Insurance and other survival data as actually collected. However,
there are still other complexities, some of which we can mention brieﬂy. Sur
vival data, or their underlying study populations are often deﬁned through
information collected in sample surveys, in which some demographic groups
are given heavier weight than their proportion in the general population.
The way in which the data should be analyzed depend on survey inclusion
probabilities or weights, and also on what the target sampled population
was, including whether the criteria for inclusion depend on timedependent
(for example healthrelated) variables. Two books describing many forms of
biomedical survival data, with examples, are those of Klein and Moeschberger
(2003) and Lee (1992).
A.2 Formal Notion of the Life Table
Consider the artiﬁcial clinical trial data summarized in Figure A.1. Such
data might have been collected for the purpose of understanding the rate of
mortality at diﬀerent ages. The Life Table, or more speciﬁcally the cohort
life table, is a simpliﬁed representation which summarizes only the numbers
‘at risk’ and the numbers observed to fail and be censored, in the oneyear
A.2. FORMAL NOTION OF THE LIFE TABLE 163
age intervals [1, 2), [2, 3), . . . , [9, 10). In the notation of the previous Section,
the patient labelled i is said to be at risk (i.e. could potentially be observed
to die) at time t if A
i
< t ≤ T
i
, and is actually observed to die at T
i
only
if D
i
= 1. If we agree to consider only integer times t = k, then we have
Y
k
= # at risk at age k =
i
I
[A
i
<k≤T
i
]
d
k
= # observed to die in [k, k + 1) =
i
D
i
I
[k≤T
i
<k+1]
c
k
= # rightcensored in [k, k + 1) =
i
(1 −D
i
) I
[k≤T
i
<k+1]
However, it is not true that all individuals dying within the age interval
[k, k +1) were necessarily at risk at time k. For this reason, it is important
to tabulate an additional quantity, the total number of subjectyears in the
survival study during which subjects were under followup at exact ages in
[k, k + 1):
τ
k
= Time on test at ages in [k, k+1) =
i
(min(T
i
, k+1) −max(k, A
i
))
Clearly, τ
k
is more informative than simply Y
k
as a denominator against
which to compare the observed number of failures d
k
in order to estimate
the rate of failures within the successive ageintervals. Indeed, we will see in
Chapter 8 that a reasonable estimator of the deathrate within ageintervals
k, k + 1) are the ratios λ
k
= d
k
/τ
k
.
To complete this brief illustration, we tabulate in Table A.1 the quantities
mentioned so far for the data represented in Figure A.1.
A.2.1 The Cohort Life Table
If a cohort of individual subjects were entered into a study simultaneously
with the same agevariable and followed up until they died, then the life
table could have a simpler form, and a simpler interpretation. In that case,
the rightcensored counts c
k
would all be 0, and the table itself would
contain all of the information (E
i
, A
i
, T
i
, D
i
)
n
i=1
for the n subjects. If a
were the common initial age, then the proportions Y
k
/Y
a
would estimate
164 APPENDIX A. GENERAL FEATURES OF DURATION DATA
Table A.1: Life table quantities for the staggeredentry survival data used to
construct Figure A.1.
k Ageint Y
k
d
k
c
k
τ
k
λ
k
1 [1, 2) 0 0 0 1.2 0.0
2 [2, 3) 2 0 0 2.0 0.0
3 [3, 4) 2 0 0 3.0 0.0
4 [4, 5) 4 0 0 4.8 0.0
5 [5, 6) 5 0 0 6.0 0.0
6 [6, 7) 6 1 0 6.8 0.147
7 [7, 8) 5 1 0 6.0 0.167
8 [8, 9) 6 1 0 6.05 0.165
9 [9, 10) 4 3 1 3.0 1.0
the fraction of the population studied which survived to age k, and the
identity Y
k+1
= Y
k
−d
k
would always hold.
However, only in very special applications, not in Insurance, can data
actually be collected in this cohort format. One is a study which follows up
a cohort of newborns, or a cohort of people selected somehow either at the
same age or with the same initiating event (like diagnosis of a disease whose
mortality is of interest). Some longitudinal epidemiologic studies, like the
famous Framingham study [ref ?] which monitored various risk factors for
heart disease, follow large numbers of people – many of whom are of the same
age or fall into narrow age brackets initially – over time. Animal studies can
follow cohorts, e.g. of newborn laboratory rats, which are subjected to the
same diets or survival stresses. But the one area of application in which this
kind of data is very common is Engineering Reliability, where a number of
devices are set running at identical (usually accelerated) stresses, in parallel,
and observed until they fail.
Despite the fact that mortality data for large human populations are
generally not collected in cohorts, the data are often tabulated as though
they were collected that way. Regardless of how the data in a mortality study
were collected, once can ﬁrst estimate the agespeciﬁc death rates directly,
q
k
= number of observed deaths at exact ages in the interval [k, k + 1),
divided by the total number of personyears spent by subjects under
followup in the study
A.3. SAMPLE SPACES FOR DURATION DATA 165
This might be done in practice only after approximating or imputing the
times on test not directly observable in the study. Moreover, in Demography
or Insurance, the estimated deathrates are often altered slightly to enhance
smoothness of the estimated deathrates as a sequence indexed by k. Finally,
in presenting the deathrates for purposes of calculation of insurance premi
ums or population projections, the death rates are presented in the tabular
form which we now deﬁne as the cohort life table.
The cohort life table, brieﬂy, displays the (integerrounded) numbers of
expected survivors at each birthday k and numbers of deaths between
successive birthdays, for a population of large hypothetical size experiencing
exactly the death rates q
k
interpreted as conditional probabilities of dying
at age [k, k + 1) given survival to the k’th birthday. Begin by choosing
a large conventional size l
0
, called the radix of the cohort life table, for a
population cohort of newborns. This number is a power of 10, usually 10
5
.
Denote by [·] the greatestinteger or ﬂoor function. Then the cohort life
table consists of the columns
l
k
=
_
l
0
k−1
j=0
q
j
_
= number of lives aged k
d
k
= l
k
−l
k+1
= number of deaths at ages in [k, k + 1)
for k ranging from 0 up to and including the largest integer age ω − 1,
where ω is the terminal age) seen for any subject of the mortality study.
Next to these columns may also be displayed the deathrates q
k
. Note that,
apart from rounding errors, q
k
= d
k
/l
k
for all k.
This ‘life table’ is an artiﬁcial construction, referring directly to no actu
ally observed population, but containing exactly the same information as the
column of (smoothed, rounded) death rates q
k
. It is the mortality record
of a ﬁctitious population cohort with exactly the same death rates after
smoothing and rounding as those estimated from some actual population.
A.3 Sample Spaces for Duration Data
The preceding sections have described ﬁrst the actual setting in which ran
dom durations are observed within a realistic mortality study, and then the
166 APPENDIX A. GENERAL FEATURES OF DURATION DATA
idealized presentation of the observed mortality in the form of a cohort life
table, in which mortality of a ﬁctitious population cohort is recorded. From
the viewpoint of Probability Theory, random variables or data are formalized
as measurement functions on the sample space Ω of all possible detailed
outcomes of a survival experiment. It is instructive to deﬁne the sample
spaces needed at three levels of complexity of the probability and statistics
of survival models.
A.3.1 Sample Space for a Single Newly Insured Life
The simplest case is the one studied in the ﬁrst two chapters of this book,
where only integer ages are recorded in the survival experiment, and all
probabilities and expected values related to functions of a single lifetime or
integerageatdeath random variable [T]. Here the sample space and under
lying probability is very easy to describe: Ω = {1, 2, . . . , l
0
} enumerates
the l
0
lives summarized in the cohort life table, with equal assigned proba
bility Pr({i}) = 1/l
0
for each individual labelled i, 1 ≤ i ≤ l
0
. Recalling
that d
k
= l
k
− l
k+1
in the cohort life table, for all k = 0, 1, . . . , ω − 1,
we note that l
0
=
ω−1
k=0
d
k
. Then the single integervalued age random
variable [T] for a new individual being insured can be explicitly constructed
as a function of i ∈ Ω as follows: for k = 0, 1, . . . , ω −1,
[T](i) = k if and only if
k−1
j=0
d
j
< i ≤
k
j=0
d
j
(A.1)
The interpretation of this rule is that if we number the l
0
individuals i in
the cohort life table in order of the wholenumber age k at which they die,
then the ﬁrst
k−1
j=1
d
j
= l
0
− l
k
individuals die at ages less than k, and
the next d
k
individuals die at integer age k.
The underlying random experiment is to select an individual i equiprob
ably from the list of all all l
0
individuals in the cohort ‘population’: that is,
in this simpliﬁed model the lifetime [T] of the newly tobeinsured indi
vidual is modelled as being the same as a randomly selected member of the
cohort population. Then the event [T] = k consists precisely of the subset
of indices i satisfying
k−1
j=0
d
j
< i ≤
k
j=0
d
j
, and therefore has size
d
k
and probability d
k
/l
0
as desired.
A.3. SAMPLE SPACES FOR DURATION DATA 167
Remark A.1 In the next subsection, we consider the sample space appro
priate to a cohort of lives, assumed independent, simultaneously following the
same mortality rates summarized in a cohort life table. There are cases of
intermediate complexity not discussed in detail in this book, cases where each
lifetime is additionally labelled by a cause of death L or where events deﬁned
in terms of the dependent lifetimes T
A
, T
B
of a pair of related lives (such as
a husbandwife pair) have consequences for Insurance. The feature of these
intermediatecomplexity sample spaces is that a single vector (T
1
, . . . , T
K
)
of ﬁnitely many, possibly dependent, lifetime random variables are modelled
simultaneously.
The case of lifetimes (T, L) labelled by cause arises when for each of K
distinct types of mortality, such as death from speciﬁed disease (as in ‘cancer
insurance’) or accident or from other causes, there is an underlying random
variable T
k
, k = 1, 2, . . . , K, giving the age at which the individual would
have died from that cause if not earlier killed from another cause. Then the
actual observed failure age T is min(T
1
, . . . , T
K
) and the random label L
is the integer in {1, 2, . . . , K} for which T
L
= T. This setup is called a
competing risks model (see Gail 1975; David and Moeschberger 1978) in
the biostatistical literature, and relates to multiple decrement (cohort) life
tables in an Insurance context (Gerber 1997 Ch. 7; Jordan 1991 Part II), but
these topics are not treated further in this book.
The joint modelling of pairs (T
A
, T
B
) or larger multilife groups of life
times is important in the calculation of insurance premiums and annuity or
pension values for husbandwife pairs, for example in insurances of both hus
bands and wives or in annuities — possibly variable, like US Social Security,
or with a smaller payment to the survivor — which revert to the surviv
ing member of the pair when one member dies. This topic is treated under
the heading of contingent multilife functions (Jordan 1991 Part II, or
Gerber 1997 Ch. 8). 2
A.3.2 Sample Space for a Full Cohort Population
As mentioned explicitly in Section A.2.1, the cohort population whose mor
tality is summarized in the cohort life table is generally a complete ﬁction.
Nevertheless, the relative frequency ratios deﬁning the survival functions and
168 APPENDIX A. GENERAL FEATURES OF DURATION DATA
death rates which are derived from the life table and used in calculated in
surance premiums could also be viewed as statistical estimators of unknown
statistical parameters based on a set of n independent identically distributed
lifetime random variables T
i
, 1 ≤ i ≤ n. That is, rather than viewing the
cohort data as ﬁxed, we can view them as the realized values of a set of
independent identically distributed lifetime random variables representating
a realistic underlying mechanism of mortality. In particular, while the sam
ple space described in Section A.3.1 is inherently discrete, it is a little more
realistic to treat the possible cohort lifetimes as a set of continuous random
variables, which are independent across individuals and can each take values
anywhere on the positive age axis.
The sample space described in this Section allows us to consider the in
trinsic variability of the estimated death rates q
k
and other statistics derived
as a function of observed ageatdeath random variables if those variable val
ues were lifetime lengths of individuals under followup for their entire lives.
This aspect of the random mortality experiment is still an artiﬁcial idealiza
tion, since we have already argued in this Chapter that realistic mortality
studies generally have a much more complicated and inconvenient pattern
of staggered positive ages at entry and of loss to followup before death for
many subjects under study.
The greater realism of cohorttype survival experiments, whose sample
sizes we now deﬁne, is of particular use in Chapter 3 of this book, where
the quality of deathrate estimates is studied and where the simulation of
new cohort lifetables with speciﬁed survival functions S(t) is described.
Yet it is immediately apparent that this realism comes at a price of greater
mathematical complexity. The sample space itself must be a set of detailed
outcomes not only for a single continuously distributed lifetime, but for a
sequence of n independent lifetimes. The most natural space to use is
Ω = R
n
+
= [0, ∞)
n
, with the vectorvalued mapping given by the identity
mapping on Ω:
T(s) = {T
i
(s)}
n
i=1
= {s
i
}
n
i=1
Unlike the situation in Section A.2.1, where the randomlifetime mapping
was deﬁned in such a way that the probabilities associated with individual
outcomes were equiprobable, now the probabilities are deﬁned by the prop
erty that the lifetimes all follow survival function S(t) and are independent
A.3. SAMPLE SPACES FOR DURATION DATA 169
of one another, a speciﬁcation accomplished by the deﬁnition
Pr
__
s ∈ R
n
: a
1
< s
1
≤ b
1
, a
2
< s
2
≤ b
2
, . . . , a
n
< s
n
≤ b
n
__
=
Pr
__
s ∈ R
n
: T(s) ∈ (a
1
, b
1
]×(a
2
, b
2
]×· · ·×(a
n
, b
n
]
__
=
n
i=1
(S(a
i
)−S(b
i
))
Probabilities of other events concerning the randomvariable components
T
i
(s) are implicitly determined from this deﬁnition on ndimensional recan
gles
n
i=1
(a
i
, b
i
] by means of the probability axioms (ﬁnite or countable
additivity), since a very large class of events can be generated by limits of
increasing unions and decreasing intersections of unions of such rectangles.
Further details of the unique speciﬁcation of probability laws from a gener
ating collection of open sets can be found in more advanced treatments of
Probability Theory, such as Ross (2005) or Billingsley (1995).
A.3.3 Sample Space for the Realistic Mortality Study
The Sample Space Ω needed to accomodate the detailed outcome data
(E
i
, A
i
, T
i
, D
i
, 1 ≤ i ≤ n) described in Section A.1 requires a Cartesian
product of R
3n
+
whose coordinates model the values factors of all of the
random variables E
i
, A
i
, T
i
, 1 ≤ i ≤ n, along with a further space {0, 1}
n
to model the values D
i
, 1 ≤ i ≤ n. The usual assumption of independence
of (E
i
, A
i
, T
i
, D
i
) across diﬀerent individuals i is embodied in a deﬁnition
of Probability on Ω as a socalled ‘product probability’ across n spaces
R
3
+
×{0, 1}. However, the joint probability density of (E
i
, A
i
, T
i
, D
i
), can
have all sorts of diﬀerent realistic dependence structures. We refer to texts on
Survival Analysis (Cox and Oakes 1994; Klein and Moeschberger 2003; David
and Moeschberger 1978) for discussion of such matters. In this book, only in
Chapter 8 do we address a simpliﬁed although typical setting (‘independent
death and censoring’) to introduce maximumlikelihood estimators of survival
in models with piecewise constant hazards and KaplanMeier estimators in
models with general (‘nonparametric’) hazards.
170 APPENDIX A. GENERAL FEATURES OF DURATION DATA
Bibliography
[1] Billingley, P. Probability Theory and Measure, 3rd ed. Wiley
Interscience, New York, 1995
[2] Bowers, N., Gerber, H., Hickman, J., Jones, D. and Nesbitt, C. Actu
arial Mathematics, 2nd ed. Society of Actuaries, Itasca, IL 1997
[3] Cox, D. R. and Oakes, D. Analysis of Survival Data, Chapman and
Hall, London 1984
[4] David, H. and Moeschberger, M. The Theory of Competing Risks.
Griﬃn, Huntington, WV, 1978
[5] Devore, J. Probability and Statistics for Engineering and the
Sciences, 7th ed. Duxbury, Paciﬁc Grove 2007
[6] Ellis, R. and Gulick, D. Calculus, 6th ed, Custom Publishing, 2002
[7] Gail, M. (1975) A review and critique of some models used in competing
risk analysis. Biometrics 31, 209222.
[8] Gerber, H. Life Insurance Mathematics, 3rd ed. SpringerVerlag,
New York, 1997
[9] Hogg, R. V. and Tanis, E. Probability and Statistical Inference,
7th ed. Addison Wesley / PrenticeHall, New York, 2005
[10] Jordan, C. W. Life Contingencies, 2nd ed. Society of Actuaries,
Chicago, 1967, paperback reprinted 1991
[11] Kalbﬂeisch, J. and Prentice, R. The Statistical Analysis of Failure
Time Data, 2nd ed. Wiley, New York 2002
171
172 BIBLIOGRAPHY
[12] Kellison, S. The Theory of Interest, 3rd ed. Irwin/ McGraw Hill,
Homewood, Ill. 2008
[13] Klein, J. and Moeschberger, M. Survival Analysis: Techniques for
Censored and Truncated Data, 2nd ed. Springer, New York, 2003
[14] Lee, E. T. Statistical Models for Survival Data Analysis, 2nd ed.
John Wiley, New York 1992
[15] The R Development Core Team, R: a Language and Environment.
[16] Ross, S. A First Course in Probability, 7th ed. Prentice Hall 2005
[17] Slud, E. and Hoesman, C. (1989) Moderate and largedeviation actuarial
probabilities in actuarial risk theory. Adv. Appl. Prob. 21, 725741.
[18] Spiegelman, M. Introduction to Demography, Revised ed. Univ. of
Chicago, Chicago, 1968
[19] Venables, W. and Ripley, B. Modern Applied Statistics with S, 4th
ed. SpringerVerlag, New York, 2002
c 2009 Eric V. Slud Statistics Program Mathematics Department University of Maryland College Park, MD 20742
Contents
0.1 Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv 1 1 6 9 9
1 Basics of Probability & Interest 1.1 Probabilities about Lifetimes . . . . . . . . . . . . . . . . . . . 1.1.1 1.2 Random Variables and Expectations . . . . . . . . . .
Theory of Interest . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 1.2.2 1.2.3 1.2.4 1.2.5 Interest Rates and Compounding . . . . . . . . . . . .
Present Values and Payment Streams . . . . . . . . . . 14 Principal and Interest, and Discount Rates . . . . . . . 17 Variable Interest Rates . . . . . . . . . . . . . . . . . . 20 Continuoustime Payment Streams . . . . . . . . . . . 24
1.3 1.4 1.5
Exercise Set 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Worked Examples . . . . . . . . . . . . . . . . . . . . . . . . . 28 Useful Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . 31 33
2 Interest & Force of Mortality 2.1
More on Theory of Interest . . . . . . . . . . . . . . . . . . . . 33 2.1.1 2.1.2 Annuities & Actuarial Notation . . . . . . . . . . . . . 34 Loan Repayment: Mortgage, Bond, Sinking Fund . . . 39 i
. . . . . 87 Curtate Expectation of Life . .3 2. . .4 3. . . . . .1 Comparison of Forces of Mortality . . . . .2 3. . . . . . . . . . . . . . . . . . . . . . 109 3. . . . . . . .2 CONTENTS Loan Amortization & Mortgage Reﬁnancing . . . . . . . . . .6 3. . . . .5. . . 69 73 3 Probability & Life Tables 3. .ii 2. .7 3. 112 4 Expected Present Values of Payments 115 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Exercise Set 2 . . . . . . . . . . 91 Interpolation Between Integer Ages . . . . . . . . . . . .1 Life Expectancy – Deﬁnition and Approximation .2. .3. . . . . . . . . . .10 Useful Formulas from Chapter 3 . . . . . . . . . . . . 42 Computational illustration in R . . . . . . . . . . 92 3. 41 Illustration on Mortgage Reﬁnancing . . . . . . .1. . . . . 97 Exercise Set 3 . . 90 3. . . . . . . 100 Worked Examples . . .3. . . . . . . . . . . .2 Rules for Manipulating Expectations . . . . . . . 80 Expectation of Discrete Random Variables . 74 3. . . . . . . . . . 55 2. . . . . . . . . . . . . . . 84 3. . 48 2. . . .4 2. . . . . .3 Simulation of Discrete Lifetimes . . . . . . . . . . . . . . . . . . . . . .1 3. .5 2. . . . . . . . . . .8 3. . . . . . 103 Appendix on Large Deviation Probabilities . . .1. . 61 Worked Examples . 44 Force of Mortality & Analytical Models . . .5 Interpreting Force of Mortality . . . . . . . . . . 77 3. . . .3 2. . . . . . 65 Useful Formulas from Chapter 2 .1 Binomial Variables & Law of Large Numbers . . . . 96 3. . .1. . . . .1. . . . . . . . . . . . . . . . . . . .9 Some Special Integrals . . . . .1 Probability Bounds & Approximations . . . . . . .4 2. .
. .5. . . . . . . 158 A. . . . 118 4.5 Extension to Multiple Payments per Year . . . 132 4. . . . .4 Continuous Contracts & Residual Life . . . . . . . . . . . . . . . .1 Survival Data Concepts . . . .1 The Cohort Life Table . . . . . . . 123 4. . . . . . . .1 4. . . . . . . . .3. m = 1 . 162 A. . . . .2 Sample Space – Cohort Population . . . . . . .2. . . . . . . . . 166 A. .1 4. . . . . . 144 Worked Examples . . . . . . . . . . . . . . . . . . . . .2 4. 116 Types of Contracts . . . . . .2 iii Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Formulas for Net Single Premiums . . . . . . . .2. .8 Exercise Set 4 . . .7 4. . . . . . . . .6 4. . . . .3 Sample Spaces for Duration Data . . . . . .5.3 Sample Space – General Case . . 163 A. . .5.4 4. . . . . . . . . . . . . . . . 133 Integral Formulas . . . . . . . . . .1 4. . 125 Interpolation Formulas in Risk Premiums . . . 136 Numerical Calculations of Life Expectancies . .2 Formal Relations. . . . . 134 Risk Premiums under Theoretical Models . . . . . .2. . . . . 148 Useful Formulas from Chapter 4 . . . . . . . . . . . . . . . 154 157 A General Features of Duration Data A. . . . . . .CONTENTS 4. . . 140 4. . . . . . . . . . . . 167 A. . .3. . . . . . . . . . . 169 Bibliography 171 . . . . .3 4.3. 165 A.1 Sample Space– Single Life . . . . . . . . . . . .5. .3 4. . . . . .2 Formal Notion of the Life Table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Continuous Risk Premium Formulas . . . . . . . .
the underlying formal random experiment being random selection from the cohort lifetable population (or. and Life Insurance Premiums and Reserves. Both the Interest Theory and Probability related to life tables are treated as wonderful concrete applications of the calculus.iv CONTENTS 0.1 Preface This book is the text for an upperlevel lecture course (STAT 470) at the University of Maryland on actuarial mathematics. The Insurance material on contingent present values and life tables is developed as directly as possible from calculus and commonsense notions. illustrated through word problems. in discussing Binomial random variables and the Law of Large Numbers. Notions of relative frequency and average are introduced ﬁrst with reference to the ensemble of a cohort lifetable. from economics to operations research to statistics. in particular on the basics of Life Tables. aiming not so much to prepare the students for speciﬁc Actuarial Examinations – since it cuts across the Society of Actuaries’ Exams FM. The calculation of expectations of functions of a timetodeath random variables is rooted on the one hand in the concrete notion of lifetable average. the combinatorial and probabilistic interpretation of binomial coeﬃcients are de . but the prerequisite calculus courses must have been solidly understood. which is then approximated by suitable idealized failure densities and integrals. Such a focus allows undergraduates with solid preparation in calculus (not necessarily mathematics or statistics majors) to explore their possible interests in business and actuarial science. from the subset of lx members of the population who survive to age x). Survival Models. in the context of probabilities and expectations for ‘lives aged x’. Later. M (Segment MLC) and C – as to present the actuarial material conceptually with reference to ideas from other undergraduate mathematical studies. It is a truism of preactuarial advising that students who have not done really well in and digested the calculus ought not to consider actuarial studies. It is not assumed that the student has seen a formal introduction to probability. The lectures require no background beyond a third semester of calculus. for the exercise of their quantitative talents — to know something concrete and mathematically coherent about the topics and ideas actually useful in Insurance. This is a ‘topics’ course. It also allows the majority of such students — who will choose some other avenue.
some material is included on statistics of biomedical studies and on reliability which would not ordinarily ﬁnd its way into an actuarial course. Premium Calculation. lifetable) counts to deviate much percentagewise from expectations when the underlying population of trials is large.) The general notions of expectation and probability are introduced. but for example the Law of Large Numbers for binomial variables is treated (rigorously) as a topic involving calculus inequalities and summation of ﬁnite series. PREFACE v rived from the Binomial Theorem. etc. These lectures present Theory of Interest as a mathematical problemtopic. which the student the is assumed to know as a topic in calculus (Taylor series identiﬁcation of coeﬃcients of a polynomial. This allows statistics students to connect the basic ideas of life table construction – considered by actuaries a more advanced topic – to the problems of statistical estimation. Interest Theory topics are presented here ﬁrst as a way to learn the skills of applying Equivalence and Superposition principles to real problems. While the basics of actuarial Life Contingencies are treated elsewhere as a problemsolving method using mortality tables presented in a cohort format. but also as a way of highlighting the relationship between realized payouts under standard Insurance contracts and instances of standard payment streams with random duration.1.g. In this approach. The approach here is to return to the ﬁrst principles of presentvalue Equivalence and linear Superposition of payment streams over time. Getting the typical Interest problems — such as the exercises on mortgage reﬁnancing and present values of various payoﬀ schemes — into correct format for numerical answers is often not easy even for good mathematics students. insurance reserves are seen as natural generalizations of bond amortization schedules..0. it is not separated by chapters into uniﬁed topics such as Interest Theory. This approach allows introduction of the numerically and conceptually useful largedeviation inequalities for binomial random variables to explain just how unlikely it is for binomial (e. Accordingly. which is rather unlike what is done in typical ﬁnance courses. While the material in these lectures is presented systematically. The reader is also not assumed to have worked previously with the Theory of Interest. Instead the introductory material from . Probability Theory. some eﬀort is devoted in this book to contrasting the form in which the underlying mortality data are received to the form of the cohort life table used in calculating premiums and reserves.
No book at this level can claim to be fully selfcontained.org/. lifetable construction. One text which combines a general introduction to R with the speciﬁcs of many statistical data analysis methods. . this book diﬀers from other Actuarial texts in its use of computational tools. and statistical estimation. opensource R platform because of its powerful tools for numerical integration. including manuals. Realistic problems on present values of payment streams. Free web access to the downloadable R platform. for example at http://wiener.csi.cuny. especially where building blocks like rootﬁnders or numerical integration routines are needed. and they highlight the most important and frequently used formulas. and later. can be found at http://www. The Worked Examples sections show how the ideas and formulas work smoothly together. is Venables and Ripley (2002). we encourage the students to use the free. There are now many good introductory texts on computing with R in statistical applications. rootﬁnding.rproject. At the end of each chapter is an Exercise Set and a short section of Worked Examples to illustrate the kinds of word problems which can be solved by the techniques of the chapter. Some good free tutorial material on R can also be found on the web.vi CONTENTS probability and interest theory are interleaved. rapidly lead to calculations too diﬃcult to do by hand or by calculator. The coverage of the main body of each chapter is primarily ‘theoretical’. In this text. and on insurancecontract premiums related to those models.edu/Statistics/R/simpleR/. Finally.math. Throughout this text. illustrations and Exercise solutions and solutions are given in terms of R. various mathematical ideas are introduced as needed to advance the discussion. but the conceptually based formulas often translate more eﬀectively using mathematical tools in computing platforms like MATLAB or the statistical language R. on probabilistic survival models related to human lifetimes. Actuarial students often do these calculations using EXCEL or other spreadsheet programs. but every attempt has been made to develop the mathematics to ﬁt the actuarial applications as they arise logically.
x + 1 Now. such as Hogg and Tanis (2005) or Devore (2007). but will likely want to supplement this Chapter with reading in any of a number of calculusbased introductions to probability and statistics. and Expectation In the cohort lifetable model. and the basics of the Theory of Interest as covered in the text of Kellison (2008) or Chapter 1 of Gerber (1997). 2.1 Probability. 1. lx for each integer age x = 0. ..e. resulting in data dx . and treat S0 (t) as the fraction of individuals in a 1 . alive at birthday x ) and dx = lx − lx+1 = number dying between ages x. 1.Chapter 1 Basics of Probability and the Theory of Interest This ﬁrst Chapter supplies some background on elementary Probability Theory and basic Theory of Interest. . The reader who has not previously studied these subjects may get an overview here. . imagine a number l0 of individuals born simultaneously and followed until death. not just whole numbers x. where lx = number of lives aged x (i. allow the agevariable to be denoted by t and to take all real values. Lifetimes.
e. In response to a probability question. using numbers from a cohort lifetable like that of the accompanying Illustrative Life Table. This nonincreasing function S0(t) would be called the empirical ‘survivor’ or ‘survival’ function. Note that the event of dying between exact ages 35 and 41 or between 52 and 60 is the union of the nonoverlapping events of the age random variable having value falling in the interval [35. who die either between 35 and 41 or between 52 and 60.g. those alive at age x. What do probabilities have to do with the cohort life table and survival function ? To answer this. k. we supply the fraction of the relevant lifetable population. and for integers x. we ﬁrst introduce probability as simply a relative frequency.2 CHAPTER 1. whose lifetimes will satisfy the stated property. decreasing. and that any probability question about a single lifetime is really a question concerning the fraction of a speciﬁed set of lives. and continuously diﬀerentiable (or piecewise continuously diﬀerentiable with just a few breakpoints) and takes values exactly = lx /l0 at integer ages x.. S(x) − S(x + k) lx − lx+k = S(x) lx denotes the fraction of those alive at exact age x who fail before x + k. S0 (y) − S0 (y + t) is the exact and S(y) − S(y + t) the approximated fraction of the initial cohort which fails between time y and y + t. to obtain identities like P r(life aged 29 dies between exact ages 35 and 41 or between 52 and 60 ) = S(35) − S(41) + S(52) − S(60) = S(29) (l35 − l41) + (l52 − l60) l29 where our convention is that a life aged 29 is one of the cohort known to have survived to the 29th birthday. The idea here is that all of the lifetimes covered by the life table are understood to be governed by an identical “mechanism” of failure. it can be approximated by a survival function S(t) which is continuous. e. 60).g.. 41) with that of falling in [52. This “frequentist” notion of probability of an event as the relative frequency with which the event occurs in a large population of (independent) identical units is associated with the phrase “law of large . Then for all positive real y and t. Although it takes on only rational values with denominator l0. BASICS OF PROBABILITY & INTEREST life table surviving to exact age t.
• When probabilities are requested with reference to a smaller universe of possible outcomes.gov/OACT/STATS/table4c6. and also disjoint unions of such subintervals. The main ideas arising in the discussion so far are really matters of common sense when applied to relative frequency but require formal axioms when used more generally: • Probabilities are numbers between 0 and 1 assigned to subsets of the entire collection Ω of possible outcomes. to the 29th birthday). and for a cohort life table constructed to reﬂect the best estimates of US male and female mortality rates in 2004. see the Social Security webpage http://www. the subsets which are assigned probabilities include subintervals of the interval of possible human lifetimes measured in years. which will be discussed later. remark only that the life table population should be large for the ideas presented so far to make good sense. the resulting conditional probabilities of events A are written P r(A  B) and calculated as P r(A ∩ B)/P r(B). in relation to the cohort population. PROBABILITIES ABOUT LIFETIMES 3 numbers”. • Two events A. speciﬁes the subset of the population which survives to exact age 29 (i. such as B = lives aged 29. See Table 1. . At this point. with the probability of Ω itself deﬁned equal to 1.1.1. as long as P r(B) > 0 — when the conditional probability P r(AB) expressing the probability of A if B were known to have occurred. These sets in the real line are viewed as possible events summarizing ages at death of newborns in the cohort population. nonoverlapping) sets A and B is necessarily equal to the sum of the separate probabilities P r(A) and P r(B). B.e. • The probability P r(A ∪ B) of the union A ∪ B of disjoint (i. The phrase “lives aged 29” deﬁnes an event which in terms of ages at death says simply “age at death is 29 or larger” or. rather than all members of a cohort population.e.1 for an illustration of a cohort lifetable with realistic numbers. B are deﬁned to be independent when P r(A ∩ B) = P r(A) · P r(B) or — equivalently.ssa.. where A ∩ B denotes the intersection or overlap of the two events A. we regard each set A of ages as determining the subset of the cohort population whose ages at death fall in A.. is the same as the unconditional probability P r(A).html. For now. In the examples.
For details of simulation. BASICS OF PROBABILITY & INTEREST Table 1. see Section 3.4 CHAPTER 1.1: Illustrative LifeTable. Age x 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 lx 100000 97371 97230 97123 97060 96997 96928 96859 96807 96753 96702 96669 96629 96582 96521 96435 96330 96247 96122 95989 95840 95686 95548 95385 95217 95051 94900 94751 94585 94428 94295 94135 93986 93834 93674 93475 93288 93076 92848 92576 dx 2629 141 107 63 63 69 69 52 54 51 33 40 47 61 86 105 83 125 133 149 154 138 163 168 166 151 149 166 157 133 160 149 152 160 199 187 212 228 272 261 x 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 lx 92315 92020 91688 91280 90866 90402 89870 89283 88603 87901 87119 86278 85393 84419 83337 82249 81036 79692 78269 76793 75221 73525 71741 69808 67786 65600 63339 60968 58542 56186 53484 50936 48259 45448 42685 39975 37127 34295 31460 dx 295 332 408 414 464 532 587 680 702 782 841 885 974 1082 1088 1213 1344 1423 1476 1572 1696 1784 1933 2022 2186 2261 2371 2426 2356 2702 2548 2677 2811 2763 2710 2848 2832 2835 2803 .2 below. simulated to resemble realistic US Male lifetable up to age 78.
In addition. for any ages a < b. but no value S(t) for t < ω is zero. no individual lives to the ω birthday. S(0) = 1. and conditional probability.2) which has the very helpful geometric interpretation that the probability of dying within the interval [a. The lifetable. and in applications S(t) should be continuous and piecewise continuously diﬀerentiable (largely for convenience. and the mechanism by which members of the population die. Since fewer people are alive at larger ages. or a still more complicated set) obviously satisﬁes the ﬁrst two of the bulleted axioms displayed above. by the fundamental theorem of calculus. by deﬁnition.1. most theoretical forms for . Note also that the ‘probability’ rule which assigns the integral A f (t) dt to the set A (which may be an interval. While ω is ﬁnite in real lifetables and in some analytical survival models. b). for formal deﬁnitions and more detailed discussion of the notions of sample space. are summarized ﬁrst through the survivor function S(t) which at integer values of t = x agrees with the ratios lx /l0. and can be interpreted as the probability for a single individual to survive at least x time units. and because any analytical expression which would be chosen for S(t) in practice will be piecewise smooth). b) is equal to the area under the density curve y = f (t) over the tinterval [a. namely that P (Ω) = 1 (where Ω is the sample space of all lifetable outcomes) and P r(A ∪ B) = P r(A) + P r(B) whenever A. Note that S(t) has values between 0 and 1. Another way of summarizing the probabilities of survival given by this function is to deﬁne the density function f (t) = − dS (t) = −S (t) dt (1. Then. S(t) is a decreasing function of the continuous agevariable t.1. event. PROBABILITIES ABOUT LIFETIMES 5 Note: see a basic probability textbook. B are disjoint or nonoverlapping subsets of Ω.1) as the (absolute) rate of decrease of the function S. The terminal age ω of a life table is an integer value large enough that S(ω) is negligibly small. such as Hogg and Tanis (1997) or Devore (2007). probability. For practical purposes. P r( life aged 0 dies between ages a and b ) b b = S(a) − S(b) = a (−S (t)) dt = a f (t) dt (1. a union of intervals.
Insurance contract payouts will be expressed as functions of the lifetimes at death of insured lives. However. or ages at death.6 CHAPTER 1. as discussed in Appendix A.1. which is the assignment rule of probabilities to all intervals (a. a random variable is a realvalued mapping X deﬁned on a sample space Ω. which is a simpliﬁed model based on the reduced datastructure. 1. b] of values for X. and the average or expected values of these payouts will be used to calculate a fair equivalent value of the insurance contract to the insured. the sample space Ω is the set of all detailed outcomes of the underlying datagenerating experiment. all of the interesting events concern lifetimes. In this book. The machinery for calculating the average values relates to the concept of random variable based on the sample space Ω = [0. such that {s ∈ Ω : X(s) ∈ (a. BASICS OF PROBABILITY & INTEREST S(t) have no ﬁnite age ω at which S(ω) = 0. in which numbers at risk and numbers of observed failures are tabulated on age intervals of one year. denoted for all real numbers a ≤ b by P r(a < X ≤ b) ≡ P r({s ∈ Ω : X(s) ∈ (a. b]} is an event with assigned probability whenever a < b are real numbers. Think of the age T at which a speciﬁed newly born member of .1 Random Variables and Expectations Formally. Now we are ready to deﬁne some terms and motivate the notion of expectation. in this and succeeding Chapters. we analyze lifetimes based on the cohort life table model. also discussed in Appendix A. the detailed outcomes can be quite complicated. b]}) Remark 1. ∞) of lifetimes. In probability theory.1 In datasets derived from actual mortality studies or insurance portfolios. and in those forms ω = ∞ by convention. The most important feature of a random variable is its probability distribution. Subsets of the sample space to which probabilities will be assigned are called events. The real number X(s) is interpreted as the value which would be observed if the detailed outcome of the underlying random experiment were s ∈ Ω.
1.1. PROBABILITIES ABOUT LIFETIMES
7
the population will die as a random variable, which for present purposes means a variable which takes various values t with probabilities governed (at integer ages) by the life table data lx and the survivor function S(t) or density function f (t) in a formula like the one just given in equation (1.2). Suppose there is a contractual amount Y which must be paid (say, to the heirs of that individual) at the death of the individual at age T , and suppose that the contract provides a speciﬁc function Y = g(T ) according to which this payment depends on (the wholenumber part of) the age T at which death occurs. What is the average value of such a payment over all individuals whose lifetimes are reﬂected in the lifetable ? Since dx = lx − lx+1 individuals (out of the original l0 ) die at ages between x and x + 1, thereby generating a payment g(x), the total payment to all individuals in the lifetable can be written as (lx − lx+1) g(x)
x
Thus the average payment, at least under the assumption that Y = g(T ) depends only on the largest whole number [T ] less than or equal to T , is
x
(lx − lx+1) g(x) / l0 = =
x x+1 x
x
(S(x) − S(x + 1))g(x)
∞ 0
(1.3) f (t) g(t) dt
f (t) g(t) dt =
This quantity, the total contingent payment over the whole cohort divided by the number in the cohort, is called the expectation of the random payment Y = g(T ) in this special case, and can be interpreted as the weighted average of all of the diﬀerent payments g(x) actually received, where the weights are just the relative frequency in the life table with which those payments are received. More generally, if the restriction that g(t) depends only on the integer part [t] of t were dropped , then the expectation of Y = g(T ) would be given by the same formula
∞
E(Y ) = E(g(T )) =
0
f (t) g(t) dt
(1.4)
The foregoing discussion of expectations based on lifetime random variables included an interpretation of the expected value of discrete random variables in terms of weighted averages which holds much more generally. In this chapter, the averages are taken over all lives tabulated in an underlying cohort life table. In Chapter 3, speciﬁcally in Section 3.3, averages are
8
CHAPTER 1. BASICS OF PROBABILITY & INTEREST
taken over large samples of observations of discrete random variables. With the aid of the Law of Large Numbers, the weightedaverage interpretation of expectations can be understood as a general mathematical result. The displayed integral (1.4), like all expectation formulas, can be understood as a weighted average of values g(T ) obtained over a population, with weights equal to the probabilities of obtaining those values. Recall from the Riemannintegral construction in Calculus that the integral f (t)g(t)dt can be regarded approximately as the sum over very small timeintervals [t, t + ∆) of the quantities f (t)g(t)∆, quantities which are interpreted as the base ∆ of a rectangle multiplied by its height f (t)g(t), and the rectangle closely matches the area under the graph of the function f g over the interval [t, t + ∆). The term f (t)g(t)∆ can alternatively be interpreted as the product of the value g(t) — essentially equal to any of the values g(T ) which can be realized when T falls within the interval [t, t + ∆) — multiplied by f (t) ∆. The latter quantity is, by the Fundamental Theorem of the Calculus, approximately equal for small ∆ to the area under the function f over the interval [t, t + ∆), and is by deﬁnition equal to the ∞ probability with which T ∈ [t, t + ∆). In summary, E(Y ) = 0 g(t)f (t)dt is the average of values g(T ) obtained for lifetimes T within small intervals [t, t + ∆) weighted by the probabilities of approximately f (t)∆ with which those T and g(T ) values are obtained. The expectation is a weighted ∞ average because the weights f (t)∆ sum to the integral 0 f (t)dt = 1. Remark 1.2 This way of approximating integrals of continuous integrands by sums corresponding to the integrals of piecewise constant integrands is closely related to the construction of the integral in terms of Riemann sums. For fuller details, see the deﬁnition the Integral via Riemann sums in a calculus book like Ellis and Gulick (2002). The same idea and formula in (1.4) can be applied to the restricted population of lives aged x. The resulting quantity is then called the conditional expected value of g(T ) given that T ≥ x. The formula will be diﬀerent in two ways: ﬁrst, the range of integrationis from x to ∞, because of the resitriction to individuals in the lifetable who have survived to exact age x; second, the density f (t) must be replaced by f (t)/S(x), the socalled conditional density given T ≥ x, which is found as follows. From the
1.2. THEORY OF INTEREST deﬁnition of conditional probability, for t ≥ x, P r(t ≤ T < t + ∆  T ≥ x) = P r( {t ≤ T < t + ∆} ∩ {T ≥ x}) P r(T ≥ x)
9
=
S(t) − S(t + ∆) P r(t ≤ T < t + ∆) = P r(T ≥ x) S(x)
Thus the density which can be used to calculate conditional probabilities P r(a ≤ T < b  T ≥ x) for x < a < b is −S (t) f (t) 1 S(t) − S(t + ∆) P r(t ≤ T < t+∆  T ≥ x) = lim = = ∆→0 ∆ ∆→0 S(x) ∆ S(x) S(x) lim In other words, when it is desired to calculate the expectation of a function Y = g(T ) of the lifetime variable T only within the conditional or restricted population of individuals with lifetime ≥ x, then the density f (t) in the expectation formula (1.4) should be replaced by the density which is equal to f (t)/S(x) for all values of t which are ≥ x, and which is 0 for values t ∈ [0, x). The result of all of this discussion of conditional expected values is the formula, with associated weightedaverage interpretation: E(g(T )  T ≥ x) = 1 S(x)
∞
g(t) f (t) dt
x
(1.5)
1.2
1.2.1
Theory of Interest
Interest Rates and Compounding
Since payments based upon unpredictable occurrences or contingencies for insured lives can occur at diﬀerent times, we study next the Theory of Interest, which is concerned with valuing streams of payments made over time. The general model in the case of constant interest, to which we restrict in the current subsection, is as follows. Money is deposited in an account like a bankaccount and grows according to a schedule determined by both the interest rate and the occasions when interest amounts are compounded, that
the day) of deposit to the instant (i. the day) of withdrawal. which results in an accumulation factor 1 + hih .e. and the standard notation for it is i(m) instead of i1/m as written above. t + h] for a period h of less than one year is prorated down to the interval h to give hih . the balance as of time t + h is A0(1 + i(m)/m) and can be viewed as though it were simultaneously withdrawn and freshly deposited at time t + h.. the balance which the owner could withdraw at time t + h is A(0) · (1 + hih ). which means that the interest rate ih applied to a timeinterval [t. The central concept of compound interest is that. This means that after a deposit of A0 at time t. Since the constant interest rate is quoted as a constant over the period of one year. Banks are not required to calculate interest from the instant (in practical terms. Thus.e. after which it would accumulate over the succeeding interval . an amount A0 deposited at the beginning of the interval accumulates to A1 = A0 · (1 + i) which could be withdrawn at the end of the interval. so that the accumulated amount is compounded (i.. If the quoted interest rate ih is annualized. or 12.10 CHAPTER 1. the depositor wishing to withdraw the full accumulation or balance at time t + s for 0 < s < h owns only the initial amount A0 . 4. is calculated by the bank and owned by the depositor) at time t + h. With amount A0 deposited initially at time t. interest rates are generally quoted as annualized rates. The further growth of deposited money over successive time intervals of length h = 1/m. BASICS OF PROBABILITY & INTEREST is. if compounded at each additional interval of length 1/m. By convention. usually with m = 1. over the ﬁxed time interval of one year. In practice. we have 1 + i as the accumulation factor by which an initial deposit is multiplied to ﬁnd the balance at the end of one year. the intervals of compounding are generally fractions h = 1/m of a year. then we say that the interest rate is a nominal annualized interest rate with mtimesyearly compounding or simply the nominal interest rate. for an initial deposit A0 at time t which is to be retained in a bank account for the time h. deemed to be added to the account. is easily understood inductively. The compounding rules are important because they determine when new interest interest begins to be earned on previously earned interest amounts. because no interest has yet been credited. 2. and if interest earned is to be credited after every successive interval h = 1/m.
the nominal interest rates i(m) with diﬀerent values of m but the same value of i can also be regarded as equivalent. t + (k + 1)/m]. if the balance A0(1 + i(m)/m)k owned by the depositor at time t + k/m is regarded as instantaneously withdrawn and redeposited as an initial balance for the next interval [t + k/m. where 0 ≤ s < 1/m and k ≥ 0 is an integer.2.1. the corresponding nominal rates i(m) for the most common values of m are obtained through the R code line: . or A0(1 + i(m) /m)(k+1) . the balance at time t + (k + 1)/m is A0(1 + i(m)/m)k multiplied by the intervalh accumulation factor 1 + ih. t + 2h] by multiplying the deposited amount A0(1 + i(m)/m) by the intervalh accumulation factor 1 + i(m) /m. THEORY OF INTEREST 11 [t + h. for k ≥ 2. Since the accumulation from the full year of deposit has the eﬀect of multiplying the initial deposit by the factor (1 + i(m)/m)m . This proves that the nominal interest rate i(m) with mtimesyearly compounding leads to exactly the same accumulation over whole years as a deposit account with the onceyearly compounded “eﬀective” rate i ≡ ieﬀ = (1 + i(m)/m)m − 1 Since any nominal interest rate i(m) has its equivalent eﬀective interest rate i = ieﬀ providing the same yearly accumulations. The overall result of our reasoning about mtimes yearly compounded nominal interest is the following: Proposition 1. a factor which would have been 1 + i at interest rate i compounded yearly. is (1+i(m)/m)k · A0 .1 The accumulated value of an initial bank deposit of A0 compounded m times yearly at nominal interest rate i(m) after a time k/m+s. These wholeyearequivalent nominal rates are determined by solving the last equation for i(m) in terms of i = ieﬀ : i(m) = m (1 + i)1/m − 1 (1.6) For example. Inductively. with i = .05.1 with k = m says that at the annualized nominal interest rate i(m) . or 5% eﬀective annual interest. Proposition 1. Thus the balance as of time t + 2h = t + 2/m is A0(1 + i(m) /m)2 . an initial deposit of A0 accumulates after exactly one year to a balance of (1 + i(m) /m)m A0 .
Recall the Taylor series expansion ez = 1 + z + z 2 /2 + z 3 /3! · · · which is valid for all z > 0. Proposition 1. i(2) = .2 When i = ieﬀ is ﬁxed. BASICS OF PROBABILITY & INTEREST mvec = c(1. i(12) = . imvec = mvec*(1.04889 .12 CHAPTER 1.04879 A few simple calculus manipulations allow us to establish the pattern of the displayed i(m) values for all choices of i. m. i(365) = .05 .05 ˆ(1/mvec) . Substitute this series with z = h ln(1 + i) into the displayed formula for g(h) to conclude that ∞ g(h) = j=1 (ln(1 + i) · h)j = ln(1 + i) + h · j! ∞ j=1 1 (ln(1 + i))j hj−1 j! is an increasing function of h > 0 and is always greater than its righthand limit g(0+) = lim h 0 d exp(h ln(1 + i)) − 1 = eh ln(1+i) h dh = ln(1 + i) h=0 The information just established concerning the behavior of i(m) = g(1/m) as a function of m for ﬁxed eﬀective interest rate i = ieﬀ is summarized as follows. deﬁned as the force of interest. δ = ln(1 + ieﬀ ) = lim i(m) m→∞ . where g(h) = h−1 (1 + i)h − 1 = (exp(h ln(1 + i)) − 1)/h has the form of a diﬀerence quotient from Calculus.2.4.12. 365) . i(4) = . the nominal annual interest rate i(m) for mtimesyearly compounding is a decreasing function of the positive integer m and tends as m → ∞ to the limiting value.1) as i(1) = . The righthand side of equation (1.04909 .6) is a function g(h) of h = 1/m.04939 .
are of the form t = k/m for positive integers k. and the other nominal rates have similar expressions immediately derived from (1. Taking the limit as l → ∞. The eﬀective interest rate ieﬀ can be expressed through its nominal continuously compounded interest rate δ as i = eδ . also called the instaneously or continuously compounded nominal interest rate. 1. 1.048973. So far.1 and 1.05. i..1 with s = 0 says that the accumulation factor for duration t = k/m based on mtimesyearly compounding at eﬀective interest rate i is (1+i(m)/m)k = (1+i(m) /m)mt = et δ .048970. with t ﬁxed and arbitrary m ≥ 1. The corresponding force of interest. THEORY OF INTEREST 13 In the displayed i(m) values for i = . Now it is obvious that the accumulation factor by continuous compounding over a duration k/m + s (for 0 ≤ s ≤ 1/m is nondecreasing in s and must therefore lie within the interval [eδk/m. and in equation (1.3 The accumulated balance of an initial deposit A0 under continuous compounding with eﬀective interest rate i.6) and Prop. This is essentially a deﬁnition of what accumulation by continuous compounding should mean. there follows: Proposition 1. eδ(k+1)/m]. but it is the only deﬁnition under which continuous compounding is well approximated by compounding aribtrarily (but ﬁnitely) many times per day. By continuity of the exponential function.1. m.6): i(m) = m (eδ/m − 1) For all durations t which are rational numbers. is exp(δt) · A0. is δ = ln(1. Prop. the dailycompounded nominal interest rate was i(365) = . the same reasoning gives etδ as the accumulation factor for duration t under the same eﬀective interest rate with mltimesyearly compounding. over a duration t > 0 which is not necessarily a rational number.2 the relations .3 the mechanism of accumulation under nominal interest rates applying with either mtimesyearly or continuous compounding. Since t = k/m is also of the form kl/(ml) for every integer l ≥ 1.e. says that the accumulation factor over duration t for instantaneous or continuous compounding should be the same. 1. or equivalently with force of interest δ = ln(1+i).2.05) = . we have described in Props.
to which is added in either case the service fees charged by the lender as a fraction of the beginningofyear loan balance. one with payments and times (αj . n) and interest rate function r(t) and the other with payments and times (α∗ . . 2. . . namely the Annual Percentage Rate or APR. These streams are called j equivalent at τ if the accumulated values at τ from the two payment streams under their respective interest rate functions are the same. and eﬀective interest rates. t∗ . These payments may be deposits into a bank or investment account. . n and which are regarded as accumulating from their times of deposit according to a schedule of interest rates r(t) which remain constant within successive intervals of calendar time t but which may change from one such interval to the next.14 CHAPTER 1. Unlike the interest rate terminology discussed up to this point. origination and participation) are required to be reported.2. First. a discrete payment stream is a sequence of (positive) deposit amounts αj made at speciﬁed calendar times tj . Two basic principles govern all problems of valuing such payment streams. or successive payments designed to accumulate over time at interest to a suﬃcient reserve fund to meet some future liability.. or loan repayments. . The APR disclosure is intended as a consumer protection to the borrower. APR is a legal term which refers either (‘eﬀective APR’) to the eﬀective interest rate or (‘nominal APR’ or simply ‘APR’) to the nominal interest. j = 1. where τ ≥ maxj tj . . 1. . • The Principle of Equivalence deﬁnes equivalence at time τ between two payment streams. BASICS OF PROBABILITY & INTEREST between nominal interest rates. j = 1.2 Present Values and Payment Streams Applications of the theory of interest generally involve comparisons between streams of payments which may be made at diﬀerent times and may accumulate at diﬀerent rates of interest. . 2. . j = 1. There is a further term for interest rates which must be disclosed to borrowers under US contracts. .g. . tj . . n∗) and interest rate funcj j tion r∗ (t). 2. maxj t∗ . force of interest. but may vary across jurisdictions in the way startup fees (e. .
from which (by continuously compounding at interest rate i from the largest of the times tl until τ ) the ﬁnal contribution of the αj deposit to the ﬁnal accumulation at τ is again seen to be αj · (1 + i)τ −tj .7) again expresses the accumulated value at τ . which is also .2. with a little more notational eﬀort and an inductive argument over the successively larger deposit times tl . . all under the same interest rate function r(t). while we will see that the second is an essentially obvious restatement of the commutativity of addition together with the fact that the accumulation of discrete payment streams is a welldeﬁned linear function of the payment amounts αj . 1. 1. However. j = 1.3 gives the contribution of the deposit αj at time tj to the accumulated value at τ as αj · (1 + i)τ −tj .1 and 1. On the other hand.1. τ ]. Consider ﬁrst the case where r(t) ≡ i is constant over the entire timeinterval [minj tj .3 as αj · (1 + i)tl −tj .3 to conﬁrm that there is no diﬀerence between the accumulated values at eﬀective interest rate i under continuous or mtimesyearly compounding. the direct inductive calculation of the accumulated amounts at all times tl ≥ tj due to the αj deposit. are also given via Prop. . . Then Prop. The formula for the continuously compounded accumulated value of the stream at time τ is n n αj (1 + i) j=1 τ −tj = j=1 αj eδ (τ −tj ) (1. 2. n) under interest rate function r(t) is the same as the sum of the accumulated values up to time T of n separate deposit accounts initiated at the respective times tj with deposits of αj . we can already see that the ﬁrst Principle is a deﬁnition. . then we appeal to Propositions 1. tj . THEORY OF INTEREST 15 • The Principle of Linear Superposition states that the total accumulated amount resulting at time τ from a payment stream (αj . This argument. The two Principles as just stated do not yet tell us how to calculate the accumulated values at τ under interest rate functions r(t) that vary over time. and formula (1.7) If compounding is instead mtimesyearly and all of the timediﬀerences τ −tl are integer multiples of 1/m. can be made into a rigorous proof of the Linear Superposition principle in the constant interestrate environment with continuous compounding.
j = 1. This amount αPV is then called the present value of the payment stream at time t .2. .7) to the accumulated value αPV (1 + i)τ −t of the single deposit at time t using interest rate i. more generally. j = 1. . . n). t1 = 0 in (1.3) to the value 1 at time t. Then. tj . maxj tj ) and equate the accumulated value (1. The magnitude αPV is then the general present value of the payment stream at time t0. . . . n) at all times τ ≥ maxj tj . n) turns out to hold more generally whenever the same timevarying interest rate function r(t) is used to accumulate both the single deposit and the payment stream. This is the amount which must be put in the bank at time 0 in order to accumulate by the factor (1 + i)t given by Prop. . . . tj . j = 1. that a single deposit αPV at time t0 can be equivalent at all times τ to a payment stream (αj .8) This equation determining αPV evidently does not depend upon τ .8)) the present value at ﬁxed interest rate i of a payment of 1 exactly t years in the future. yielding: n n αj (1 + i) j=1 τ −tj = αPV (1 + i) τ −t =⇒ αPV = j=1 αj (1 + i)t−tj (1. tj . at the same eﬀective interest rate i = ieﬀ with continuous compounding as is used to accumulate (αj . . j = 1. . It tells ﬁrst (with n = 1. . The proof of this Fact will be left to an Exercise in Section 1. BASICS OF PROBABILITY & INTEREST n αj (1 + i(m)/m)m(τ −tj ) j=1 The most important application of the principle of equivalence is in ﬁnding a deposit amount αPV at a single ﬁxed time t which is equivalent to the payment stream (αj . To see why this is possible. n is equal to the summation j=1 αj (1 + i) The same phenomenon. 1. . . the present value at time 0 under constant interest rate i of a payment stream consisting of payments αj at future times n −tj . . . (1 + i)−t . . consider any ﬁxed τ ≥ max(t.4 where accumulation formulas for variable interest rates are discussed.16 equal to CHAPTER 1. tj .
then those amounts are added to the Principal or Balance owed as of tj . Initially assume continuous compounding for all accumulations.1. However. at times 0 < t1 < t2 < · · · < tn .e. i. Each payment αj made can be broken down into the socalled Interest and Principal portions by the rule: Interest Portion of Paymt at tj = (Principal at tj−1 ) · ((1 + i)tj −tj−1 − 1) Principal Portion of Paymt at tj = αj − Interest Portion of Paymt at tj The ﬁrst of these lines is clearly the amount of interest that the principal just after tj−1 would have earned at rate i over the time interval tj − tj−1 .2. if there are fees or late charges due at the times tj when payments are made. If the borrower plans to make payments αj . but is easily shown by an algebraic rearrangement of terms. we consider the compounding of interest from the point of view of a borrower of an amount L at time 0. The amount of the payment at tj minus the amount of interest at tj is the amount by which the principal decreases from just after tj−1 to just after tj . the lender must be compensated for the amount (1 + i)t L to which the original loan amount would accumulate. where the interest rate is constant with ieﬀ = i. 1 ≤ j ≤ n. . Principal at time t = L (1 + i)t − j: tj ≤t αj (1 + i)t−tj . In addition.3 Principal and Interest. The principal remaining in the loan just after a payment has been made is the same as the amount the borrower could pay to pay oﬀ the loan completely at that instant. THEORY OF INTEREST 17 1.. while the accumulated value of the stream of payments actually made up to time t reduces the debt. in the present discussion we ignore all such additional fees or charges. and Discount Rates In this Section. minus the accumulated value at t under continuously compounded eﬀective rate i of all payments made at times before t.2. given as an Exercise. The principal owed on the loan just after time t reﬂects that as of time t. then by deﬁnition the principal remaining on the loan as of time t is equal to the accumulated value at t of the single deposit L at time 0 . This simple Proposition is not quite obvious.
18
CHAPTER 1. BASICS OF PROBABILITY & INTEREST
Exercise 1.A. Show that the foregoing deﬁnitions of Principal and Principal Portions of payments are compatible by deriving the following identity from the deﬁnitions. If Π(t) denotes the principal owed just after time t, and πj denotes the principal portion of the payment at tj , then Π(tj−1) − πj = Π(tj ) 2
The nominal interest rates i(m) for diﬀerent periods of compounding were seen in Prop. 1.2 to be related by the formulas (1 + i(m)/m)m = 1 + i = 1 + ieﬀ , i(m) = m (1 + i)1/m − 1 (1.9)
Similarly, interest can be said to be governed by the discount rates d(m) for various compounding periods, deﬁned by 1 − d(m) /m = (1 + i(m)/m)−1 Solving the last equation for d(m) gives d(m) = i(m) /(1 + i(m) /m) (1.10)
The idea of discount rates is that if an amount 1 is loaned out at interest, then the amount d(m) /m is the correct amount to be repaid at the beginning rather than the end of each fraction 1/m of the year, with repayment of the principal of 1 at the end of the year, in order to amount to the same eﬀective interest rate. The reason is that, according to the deﬁnition, the amount 1 − d(m) /m accumulates at nominal interest i(m) to (1 − d(m) /m) · (1 + i(m)/m) = 1 after a timeperiod of 1/m. The quantities i(m) and d(m) are naturally introduced as the interest payments which must be made respectively at the ends and the beginnings of successive timeperiods of length 1/m in order that the principal owed at each time j/m on an amount 1 borrowed at time 0 will always be 1. To deﬁne these terms and justify this assertion, consider ﬁrst the simplest case, m = 1. If 1 is to be borrowed at time 0, then the single payment at time 1 which fully compensates the lender, if that lender could alternatively have earned interest rate i, is (1 + i), which we view as a payment of 1 principal (the face amount of the loan) and i interest. In exactly the same way, if 1 is borrowed at time 0 for a timeperiod 1/m, then the repayment at
1.2. THEORY OF INTEREST
19
time 1/m takes the form of 1 principal and i(m) /m interest. Thus, if 1 was borrowed at time 0, an interest payment of i(m)/m at time 1/m leaves an amount 1 still owed, which can be viewed as an amount borrowed on the timeinterval (1/m, 2/m]. Then a payment of i(m)/m at time 2/m still leaves an amount 1 owed at 2/m, which is deemed borrowed until time 3/m, and so forth, until the loan of 1 on the ﬁnal timeinterval ((m − 1)/m, 1] is paid oﬀ at time 1 with a ﬁnal interest payment of i(m)/m together with the principal repayment of 1. The overall result which we have just proved intuitively is: 1 at time 0 is equivalent to the stream of m payments of i(m)/m at times 1/m, 2/m, . . . , 1 plus the payment of 1 at time 1. Similarly, if interest is to be paid at the beginning of the period of the loan instead of the end, the interest paid at time 0 for a loan of 1 would be d = i/(1 + i), with the only other payment a repayment of principal at time 1. To see that this is correct, note that since interest d is paid at the same instant as receiving the loan of 1 , the net amount actually received is 1 − d = (1 + i)−1, which accumulates in value to (1 − d)(1 + i) = 1 at time 1. Similarly, if interest payments are to be made at the beginnings of each of the intervals (j/m, (j + 1)/m] for j = 0, 1, . . . , m − 1, with a ﬁnal principal repayment of 1 at time 1, then the interest payments should be d(m) /m. This follows because the amount eﬀectively borrowed (after the immediate interest payment) over each interval (j/m, (j + 1)/m] is (1 − d(m)/m), which accumulates in value over the interval of length 1/m to an amount (1 − d(m) /m)(1 + i(m)/m) = 1. So throughout the yearlong life of the loan, the principal owed at (or just before) each time (j + 1)/m is exactly 1. The overall result concerning mperiodyearly discount interest is 1 at time 0 is equivalent to the stream of m payments of d(m) /m at times 0, 1/m, 2/m, . . . , (m−1)/m plus the payment of 1 at time 1. A useful algebraic exercise to conﬁrm the displayed assertions is:
20
CHAPTER 1. BASICS OF PROBABILITY & INTEREST
Exercise 1.B. Verify that the present values at time 0 of the payment streams with m interest payments in the displayed assertions are respectively
m
j=1
i(m) (1 + i)−j/m + (1 + i)−1 m
m−1
and
j=0
d(m) (1 + i)−j/m + (1 + i)−1 m 2
and that both are equal to 1. These identities are valid for all i > 0.
1.2.4
Variable Interest Rates
Now we formulate the generalization of these ideas to the case of nonconstant instantaneously varying, but known or observed, eﬀective interest rate r(t) at time t , corresponding to the instantaneous continuously compounded nominal rate, or timevarying force of interest, δ(t) = ln(1+r(t)). Consider the compounding of interest over successive intervals [b + kh, b + (k + 1)h], where h = 1/m for large m, there is an essentially constant principal amount over each interval of length 1/m. Since we assume the functions r(t) and therefore ln(1 + δ(t)) are uniformly continuous in t, so that over very short intervals [b + kh, b + (k + 1)h] with instantaneous compounding, the interest rate and its associated force of interest are essentially constant, with accumulation factor over the interval given by eh δ(kh) . Therefore, if an initial time b and duration τ > 0 are ﬁxed and [mτ ] = [τ /h] denotes the largest integer ≤ mτ , we ﬁnd that the continuous compounding of interest over the timeinterval [b, b + τ ] results in an overall accumulation factor of approximately ehδ(b) ehδ(b+h) ehδ(b+2h) · · · ehδ(b+([τ /h]−1)h) exp ((τ − h[τ /h]) · δ(b + h[τ /h]) which has limit as m → ∞ equal to 1 exp lim m m
[mτ ]−1 t
δ(b + k/m)
k=0
= exp
0
δ(b + s) ds
The last step in this chain of equalities relates the concept of continuous compounding to that of the Riemann integral. To specify continuoustime varying interest rates in terms of instantaneous eﬀective rates, we would
Note. If the accumulated bank balance just after time kh is denoted by Bk . .2. α1 . . kh. αn made at times 0. . 2. . Therefore. (k + 1)h). . as follows.1. then how can the accumulated bank balance be expressed in terms of αj and δ(jh) ? Clearly Bk+1 = Bk · eδ(kh)/m + αk+1 . . we have just calculated that an amount 1 at time 0 compounds to an accumulated amount An at time τ = nh. By approximating the continuous interest rate function r(t) by one which is constant on intervals [kh. B0 = α0 The preceding diﬀerence equation can be solved in terms of successive summation and product operations acting on the sequences αj and δ(jh). (j + 1)h). Our object is to determine a single deposit D at time 0 which is equivalent at time τ = nh to a stream of deposits αj . . A0 = 1. b + τ ] to t exp 0 ln(1 + r(b + s)) ds Next consider the case of deposits α0 . THEORY OF INTEREST 21 equate the last displayed formula for the accumulation factor over [b. as in (1. . First deﬁne a function Ak to denote the accumulated bank balance at time kh for a unit invested at time 0 and earning interest with instantaneous nominal interest rates δ(jh) applying respectively over the whole compoundingintervals [jh. Ak = j=0 eδ(jh)/m We now return to the idea of equivalent investments and present value of a payment stream. 1. k − 1. where all amounts accumulate according to the continuously compounded instantaneous eﬀective interest rate r(t) and associated force of interest δ(t) = ln(1 + r(t)).2. as discussed in Section 1. . . . . . . . which together with its solution is given by k−1 Ak+1 = Ak · e δ(kh)/m . .8) . j = 0. . . an amount D at time 0 accumulates to D · An at time τ . n. where h = 1/m is the given compoundingperiod. . and wherenominal annualized instantaneous interestrates δ(kh) (with compoundingperiod h) apply to the accrual of interest on the interval [kh. j = 0. nh. . h. Ak satisﬁes a homogeneous equation analogous to the previous one. . . . (k + 1)h). . and in particular D = 1/An at time 0 accumulates to 1 at time τ . Then by deﬁnition. αk .2.
. 1. Ai Bk = Ak · Gk . This constant (eﬀective) interest rate r is the one such that n sk 1 + r k=0 −tk =1 With respect to the constant interest rate r .2. . this single equivalent deposit D would be the same if the accumulations were valued at any other time τ > nh. Thus the present value of 1 at time τ = nh is 1/An . n. k = 0. . Therefore the stream of payments αk at times tk . . . for the uniquely deﬁned interest rate r. BASICS OF PROBABILITY & INTEREST of Section 1. Since Bk was the accumulated value just after time kh of the same stream of payments. (k = 0. n The formulas just developed can be used to give the internal rate of return r over the timeinterval [0.2. the present value of a payment αk at a time tk timeunits in the future is αk · (1 + r)−tk . to an immediate (time0) payment of 1. . . k = 0. 1. G0 = α0 Thus Gk+1 − Gk = αk+1 /Ak+1 . and since the present value at 0 of an amount Bk at time kh is just Bk /Ak . Now deﬁne Gk to be the present value of the stream of payments αj at time jh for j = 0. . . k. . 0 ≤ tk ≤ τ . we have simultaneously found the solution for the accumulated balance Bk just after time kh and for the present value Gk at time 0 : k Gk = i=0 αi . we conclude Gk+1 = Bk+1 Bk exp(δ(km)/m) αk+1 + = . τ ] of a unit investment which pays amount αk at times tk . n) becomes equivalent. . . . Ak+1 Ak exp(δ(km)/m) Ak+1 k≥1. . .22 CHAPTER 1. and k Gk+1 = α0 + i=0 αi+1 = Ai+1 k+1 j=0 αj Aj In summary. . .
060 9366 .055 9749 .05 and 0. must be balanced with the value (here 10.1. and then 800 yearly for the following ﬁve years. 000 = 300 j=1 (1 + r)−j + 800 j=6 (1 + r)−j + 10. As in all calculations of eﬀective interest rate. 000 returned to you (if all goes well) exactly 10 years from the date of the investment (at the same time as the last of the 800 payments. (We can guess that the correct answer lies between the minimum and maximum payments expressed as a fraction of the principal. 000 made now is expected to pay 300 yearly for 5 years (beginning 1 year from the date of the investment).) This tabulation yields: r (1. we can see that the righthand side is equal to 10.C. As an illustration of the notion of eﬀective interest rate.065 9000 .035 through 0. (That is because the indicated payment stream is being regarded as equivalent to bank interest at rate r. the present value ofthe paymentstream.11) r Setting this simpliﬁed expression equal to the lefthand side of 10.055. 000 (1 + r)−10 The righthand side can be simpliﬁed somewhat. suppose that you are oﬀered an investment option under which an investment of 10.070 8562 .) The balance equation in the Example is obviously 5 10 10.045 10574 .050 10152 . at the unknown interest rate r = ieﬀ.005 from 0. what is the eﬀective interest rate you will be earning on your investment ? Solution. since both x = (1+r)−5 and r involve the unknown r. with the principal of 10. Interpolating linearly to . Nevertheless.2. or internal rate of return. we can solve the equation roughly by ‘tabulating’ the values of the simpliﬁed righthand side as a function of r ranging in increments of 0. to 300 1+r 1−x 1 − (1 + r)−1 = + 800x (1 + r) 1−x 1 − (1 + r)−1 + 10000 x2 1−x (300 + 800x) + 10000x2 (1. 000) which is invested.075. 000 does not lead to a closedform solution.075 8320 From these values.035 11485 .040 11018 .11) . THEORY OF INTEREST 23 Exercise 1. If the investment goes as planned. 000 for a value of r falling between 0. in terms of the notation x = (1 + r)−5 .
Taking δ(t) to be the timevarying nominal interest rate with continuous compounding. and letting h decrease to 0. dividing by h.075))$root [1] 0. Subtracting B(t) from both sides of the last equation..24 CHAPTER 1.12).10000 } uniroot(Rsolv. BASICS OF PROBABILITY & INTEREST approximate the answer yields r = 0.12) The method of solution of (1.035.5 Continuoustime Payment Streams There is a completely analogous development for continuoustime deposit streams with continuous compounding.05186. c(. we replace the previous diﬀerenceequation by B(t + h) = B(t) (1 + h δ(t)) + h D(t) + o(h) where o(h) denotes a remainder such that o(h)/h → 0 as h → 0. The ratiorule of diﬀerentiation yields G (t) = B(t) A (t) B (t) − B(t) δ(t) D(t) B (t) − = = 2 (t) A(t) A A(t) A(t) .2. yields a diﬀerential equation at times t > 0 : B (t) = B(t) δ(t) + D(t) . and the R code for computing the eﬀective interest rate in this Example is: Rsolv = function(r) { x = (1+r)^(5) (1x)*(300+800*x)/r + 10000*x^2 . For example. Suppose D(t) to be the rate per unit time at which savings deposits are made.05185676 1. again has a natural interpretation in terms of present values. we have D(t) = limm→ ∞ mα[mt]. where [·] again denotes greatestinteger. A(0) = α0 (1. the rootﬁnding function in R is called uniroot . and B(t) to be the accumulated balance as of time t (analogous to the quantity B[mt] = Bk from before. when t = k/m). and the quantity B(t)/A(t) = G(t) is then the present value of the deposit stream of α0 at time 0 followed by continuous deposits at rate D(t).005 ∗ (10000 − 10152)/(9749 − 10152) = 0. The integrating factor t 1/A(t) = exp(− 0 δ(s) ds) is the present value at time 0 of a payment of 1 at time t.050 + 0. so that if we take m to go to ∞ in the previous discussion. while an accurate equationsolver ﬁnds r = 0.05189. which is the standard one from diﬀerential equations theory of multiplying through by an integrating factor.
1.3. EXERCISE SET 1
25
where the substitution A (t)/A(t) ≡ δ(t) has been made in the third expression. Since G(0) = B(0) = α0 , the solution to the diﬀerential equation (1.12) becomes
t
G(t) = α0 +
0
D(s) ds , A(s)
B(t) = A(t) G(t)
Finally, the formula can be specialized to the case of a constant unitrate payment stream ( D(x) = 1, δ(x) = δ = ln(1 + i), 0 ≤ x ≤ T ) with no initial deposit (i.e., α0 = 0). By the preceding formulas, A(t) = exp(t ln(1 + i)) = (1 + i)t, and the present value of such a payment stream is T 1 1 − (1 + i)−T 1 · exp(−t ln(1 + i)) dt = δ 0 Recall that the force of interest δ = ln(1 + i) is the limiting value obtained from the nominal interest rate i(m) using the diﬀerencequotient representation: exp((1/m) ln(1 + i)) − 1 lim i(m) = lim = ln(1 + i) m→∞ m→∞ 1/m The present value of a payment at time T in the future is then 1+ i(m) m
−mT
= (1 + i)−T = exp(−δ T )
1.3
Exercise Set 1
The ﬁrst homework set covers the basic deﬁnitions in two areas: (i) probability as it relates to events deﬁned from cohort lifetables, including the theoretical machinery of population and conditional survival, distribution, and density functions and the deﬁnition of expectation; (ii) the theory of interest and present values, with special reference to the idea of income streams of equal value at a ﬁxed rate of interest. (1). For how long a time should $100 be left to accumulate at 5% interest so that it will amount to twice the accumulated value (over the same time period) of another $100 deposited at 3% ?
26
CHAPTER 1. BASICS OF PROBABILITY & INTEREST
(2). Use a calculator or computer to answer the following numerically: (a) Suppose you sell for $6,000 the right to receive for 10 years the amount of $1,000 per year payable quarterly (starting at the end of the ﬁrst quarter). What eﬀective rate of interest makes this a fair sale price ? (You will have to solve numerically or graphically, or interpolate a tabulation, to ﬁnd it.) (b) $100 deposited 20 years ago has grown at interest to $235. The interest was compounded twice a year. What were the nominal and eﬀective interest rates ? (c) How much should be set aside (the same amount each year) at the beginning of each year for 10 years to amount to $1000 at the end of the 10th year at the interest rate of part (b) ? In the following problems, S(t) denotes the probability for a newborn in a designated population to survive to exact age t . If a cohort life table is under discussion, then the probability distribution relates to a randomly chosen member of the newborn cohort. (3). Assume that a population’s survival probability function is given by S(t) = 0.1(100 − t)1/2, for 0 ≤ t ≤ 100. (a) Find the probability that a life aged 0 will die between exact ages 19 and 36. (b) Find the probability that a life aged 36 will die before exact age 51. (4). For members of the poulation in Problem (3), (a) Find the expected age at death of a newborn (life aged 0). (b) Find the expected age at death of a life aged 20. (5). Use the Illustrative Lifetable (Table 1.1) to calculate the following probabilities. (In each case, assume that the indicated span of years runs from birthday to birthday.) Find the probability (a) that a life aged 26 will live at least 30 more years; (b) that a life aged 22 will die between ages 45 and 55; (c) that a life aged 25 will die either before age 50 or after age 70.
1.3. EXERCISE SET 1 (6). In a special population, you are given the following facts:
27
(i) The probability that two independent lives, respectively aged 25 and 45, both survive 20 years is 0.7. (ii) The probability that a life aged 25 will survive 10 years is 0.9. Then ﬁnd the probability that a life aged 35 will survive to age 65. (7). Suppose that you borrowed $1000 at 6% eﬀective rate, to be repaid in 5 years in a lump sum, and that after holding the money idle for 1 year you invested the money and earned 8% eﬀective for theremaining four years. What is the eﬀective interest rate you earned (ignoring interest costs) over 5 years on the $1000 which you borrowed ? Taking interest costs into account, what is the present value of your proﬁt over the 5 years of the loan ? Also redo the problem if instead of repaying all principal and interest at the end of 5 years, you must make a payment of accrued interest at the end of 3 years, with the additional interest and principal due in a single lumpsum at the end of 5 years. (8). Find the total present value at 5% APR of payments of $1 at the end of 1, 3, 5, 7, and 9 years and payments of $2 at the end of 2, 4, 6, 8, and 10 years. (9). Find the present value at time 0 at a 6% eﬀective interest rate of a series payments of 100 at times 1, 2, 3 and of 300 at times 6, 7, 8. (10). Find the present value at time 0 of payments of 100 at ten successive times 1, 2, . . . , 10 if the instanteous eﬀective interest rate applying at all times t in the time interval [0, 10] is r(t) = .07 − (.002)t. (11). Find the internal rate of return (i.e., the equivaent constant eﬀective interest rate) over the time interval [0, 7] of an investment which pays bank interest of 4% at times in [0, 5] if you make deposits of 1000 at each of the times t = 0, 2, 4, if the interest rate earned on the time interval [5, 7] is 6%, and if the total balance is withdrawn at time 7. (12). (i) Find the payment amount K such that a loan of 10, 000 at a 7% eﬀective annual interest rate is repaid in exactly three payments consisting of an amount K at times 1 and 3 years and of 2K at 5 years.
4 Worked Examples Example 1. What is the value of the net proﬁt at the end of the 5 years ? What is its present value (at 5%) as of time 0 ? First. Example 3. ﬁnd the probabilities for all values of [T ]. For the following small cohort lifetable (ﬁrst 3 columns) with 5 agecategories. where d = . which in this problem is presumably the rate of 5% at which the money is borrowed. the compounded value of 783. . after paying oﬀ the principal of 1000. Suppose that a sum of $1000 is borrowed for 5 years at 5%. with numerical answer given by 22. 1. both unconditionally and conditionally for lives aged 2. Suppose further that the amount received is invested and earns 7%.53 for so the answer is ln(3)/ ln(1 + i). since the amount received should compound to precisely the principal of 1000 at 5% interest in 5 years. 5 into their principal and interest portions. BASICS OF PROBABILITY & INTEREST (ii) After ﬁnding K in part (i). 3. the amount received by the borrower at time 0 is 5 5 1000 (1 − d) = 1000/(1.05/1.05)5 = 77. so is 98. K. and ﬁnd the expectation of both [T ] and (1. 2K at respective times. with interest deducted immediately in a lump sum from the amount borrowed. and principal due in a lump sum at the end of the 5 years.07 i = 0. so the net proﬁt at the end of 5 years.05) = 783. is 98.07)5 = 1098.05.52. 1. How many years does it take for money to triple in value at interest rate i ? The equation to solve is 3 = (1 + i)t.94/(1. The present value of the proﬁt ought to be calculated with respect to the ‘going rate of interest’.94. decompose each of the three loan repayment amounts K.05)−[T ]−1 .10 Example 2.05 i = 0. Next.94.53 for 5 years at 7% is 783.28 CHAPTER 1. i = 0.24 for 11.53 (1.53.52 for t = 16.
61538 0 0 1. dx at these ages if l0 = 105 .10) + 3 · (0.95238 · 0.15385) + 3 · (0.05−[T ]−1 ) = 0.20) + 1 · (0. The conditional probabilities given survivorship to agecategory 2 are simply the ratios with numerator dx for x ≥ 2 .20 + 0. WORKED EXAMPLES 29 The basic information in the table is the ﬁrst column lx of numbers surviving.78353 0.90703 0. 4.15385 0. Suppose that the deathrates qx = dx /lx for integer ages x in a cohort lifetable follow the functional form qx = 4 · 10−4 for 5 ≤ x < 30 8 · 10−4 for 30 ≤ x ≤ 55 between the ages x of 5 and 55 inclusive.15 + 0.61538) = 3.3) E([T ]) = 0 · (0.8497 The expectation of [T ] is interpreted as the average per person in the cohort lifetable of the number of completed whole years before death.4 E([T ]  T ≥ 2) = 2 · (0.86384 0. lx . if the going rate of interest is 5%.10 + + 0.1. Find analytical expressions for S(x). . and the ﬁnal expectation calculated above is the average of that presentvalue over all the individuals in the cohort lifetable.40 0.90703 · 0. and with denominator l2 = 65. x lx 0 100 1 80 2 65 3 55 4 40 5 0 dx 20 15 10 15 40 0 P r([T ] = x) P r([T ] = xT ≥ 2) 0.82770 0. The quantity (1.74622 In terms of the columns of this table.23077) + 4 · (0. . .40) = 2. .15 + 0.96.05−x−1 0.15 0.10 0.05)−[T ]−1 can be interpreted as the present value at birth of a payment of 1 to be made at the end of the year of death.78353 · 0.15) + 4 · (0. and therefore P r([T ] = x) = P r(x ≤ T < x + 1) = dx /l0 .4615 E(1.20 0 0. S(5) = .8277 · 0. we evaluate from the deﬁnitions and formula (1.4.95238 0.40 = 0.15) + 2 · (0. . Example 4. The random variable T is the lifelength for a randomly selected individual from the age=0 cohort.15 0 0. Then dx = lx − lx+1 for x = 0. 1.86384 · 0.23077 0.
9996)x−5 so that S(30) = . . 30.940446 (. . .9992)x−30 The deathcounts dx are expressed most simply through the preceding expressions together with the formula dx = qx lx . . . . BASICS OF PROBABILITY & INTEREST The key formula expressing survival probabilities in terms of deathrates is: lx+1 S(x + 1) = = 1 − qx S(x) lx lx = l0 · S(x) = (1 − q0 )(1 − q1) · · · (1 − qx−1 ) qx or So it follows that for x = 5. and for x = 31. .940446. S(5) lx = 96000 · (0. S(x) = S(30) · (.30 CHAPTER 1. . .0004)x−5 . S(x) = (1 − . 55.9992)x−30 = .
5 E g(T ) T ≥ x = 1 S(x) ∞ g(t) f (t) dt x p. dx = lx − lx+1 p. S(y) − S(y + t) = y f (s) ds p. 18 .5.1.5 Useful Formulas from Chapter 1 S(x) = lx l0 . 1 lx − lx+k lx p. 2 y+t P (x ≤ T < x + k) = S(x) − S(x + k) = f (t) = −S (t) . 9 m −m 1 + i = 1 + ieﬀ = i(m) 1+ m = d(m) 1− m = eδ p. USEFUL FORMULAS 31 1.
32 CHAPTER 1. BASICS OF PROBABILITY & INTEREST .
There are two primary methods of manipulating one paymentstream to give another for the convenient calculation of present values: 33 . in particular of force of mortality or hazardrates.1 More on Theory of Interest In this Section. we deﬁne notations and ﬁnd compact formulas for present values of some standard payment streams. and theoretical families of life distributions. while the probability material prepares us to ﬁnd and interpret average or expected values of present values expressed as functions of random lifetime variables.e. newly deﬁned payment streams are systematically expressed in terms of previously considered ones. For application in Insurance.. To this end. nonprobabilistic) theory of interest. we are preparing to value uncertain payment streams in which times of payment may also be uncertain.Chapter 2 Theory of Interest and Force of Mortality The parallel development of Interest and Probability Theory topics continues in this Chapter. The interest theory allows us to express the present values of deterministic or certain payment streams compactly. and (b) more discussion of lifetime random variables. This installment of the course covers: (a) further formulas and topics in the pure (i. 2.
if one paymentstream can be obtained from a second one precisely by delaying all payments by the same amount t of time.34 CHAPTER 2. 2. Also. (i) If s0 = 0 and s1 = · · · = snm = 1/m in the discrete setting. related to periodic premium and annuity payments. i. from now on the standard and convenient notation v ≡ 1/(1 + i) = 1 / i(m) 1+ m m will be used for the present value of a payment of 1 one year later. and some very particular streams of payments sj at times tj . The following subsection contains several useful applications of these methods..1 Annuities & Actuarial Notation The general present value formulas above will now be specialized to the case of constant (instantaneous) interest rate r(t) ≡ i ≡ eδ at all times t ≥ 0. where m ≥ 1 denotes the number of payments per year. if a paymentstream A can be obtained as the superposition of payment streams B and C. then the paymentstream is called an immediate annuity. to nm m−1 j=1 1+ i(m) m −j = 1+ i(m) m −1 1 − (1 + i(m)/m)−nm m(1 − (1 + i(m)/m)−1 ) . then the present value of stream A is the sum of the present values of B and C. see Worked Example 2 at the end of the Chapter. INTEREST & FORCE OF MORTALITY • First. and tj = j/m. and as before the mtimesperyear equivalent nominal interest rate is denoted by i(m) . For another simple illustration.1.e. by the geometricseries summation formula. The eﬀective interest rate is always denoted by i = ieﬀ . then the present value of the ﬁrst one is v t multiplied by the present value of the second. and its present value Gn (m) is given the notation an and is equal. • Second. can be obtained by paying the sum of the timed payment amounts deﬁning the streams B and C.
s0 = 1/m but snm = 0.1.2).2) m m m The ﬁrst of these formulas recognizes the annuitydue paymentstream as identical to the annuityimmediate paymentstream shifted earlier by the time 1/m and therefore worth more by the accumulationfactor (1+i)1/m = 1 + i(m) /m.1) shows that all of the values an diﬀer only through the factors i(m). The third expression in (2. The limiting behavior of the nominal interest rate can be seen rapidly from the formula i(m) = m (1 + i)1/m − 1 = δ · exp(δ/m) − 1 δ/m since (ez − 1)/z converges to 1 as z → 0.1) All of these immediate annuity values. with the limiting nominal interestrate which was shown in the previous chapter to be limm i(m) = i(∞) = δ.1.9) that i(m) = m{(1 + i)1/m − 1}. Then by (2. (m) Formula (2. are roughly comparable because all involve a total payment of 1 per year. and a the present valueis given by any of the equivalent formulas i(m) (m) 1 − vn 1 (m) (m) ) an = + an = + an−1/m (2. the present value of an annuity paid instantaneously at constant unit rate. the remaining stream coincides with the annuityimmediate stream consisting of nm − 1 (instead of nm) payments of 1/m. Recall from formula (1. n but varying m.3) an = lim an = lim an = ¨ m→∞ m→∞ δ . 1 − vn (m) (m) (2. which diﬀer by only a few percent for varying m and ﬁxed i. instead of the payment stream deﬁning the immediate annuity. a ¨n (m) = (1 + In the limit as m → ∞ for ﬁxed n. the notation an denotes the continuous annuity. MORE ON THEORY OF INTEREST which shows that an (m) 35 = 1 − vn 1 − ((1 + i(m)/m)−m )n = m (1 + i(m)/m − 1) i(m) (2. that is.2) represents the annuitydue stream as being equal to the annuityimmediate stream with the payment of 1/m at t = 0 added and the payment of 1/m at t = n removed. for ﬁxed v. If.2.1) and (2. and the present value notation (m) changes to ¨ n . The ﬁnal expression says that if the time0 payment is removed from the annuitydue. The payment stream is then called an annuitydue. then nm deposits of 1/m are made at an arithmetic progression of times from 1/m to n inclusive. as shown in Table 2.
0294 .0961 .137 .0675 .0297 .0296 .0485 .03 . INTEREST & FORCE OF MORTALITY Table 2. 2 A handy formula for annuitydue present values follows easily by recalling that 1 − d(m) = m 1 + i(m) m −1 (m) implies d(m) = i(m) 1 + i(m) /m Then.0492 .141 .0942 .0938 .5) .141 .0957 .139 Remark 2.0199 .0669 . In the limit as m → ∞.4) as: a∞ (m) = 1 i(m) .1) and (2. for various choices of eﬀective annual interest rate i and number m of compounding periods per year.138 .0931 .0946 .0198 .0298 .02 .0673 .0671 .0199 .1: Values of nominal interest rates i(m) (upper number) and d(m) (lower number). by (2.143 .0489 .15 .07 .0491 .145 . the continuous annuity deﬁnition an and formula remain valid with any positive real number n.0684 . an ¨ (m) = (1 − v n ) · 1 + i(m) /m 1 − vn = i(m) d(m) (2.0293 .0198 .0295 .1).0198 .1 The deﬁnition and formulas for the immediate annuity an (m) and the annuitydue ¨ n remain valid if nm but not necessarily n itself a is an integer.4) In case m is 1. and the annuities are called perpetuities (respectively immediate and ¨ due) with presentvalue formulas obtained from (2.2) and (2.135 . the notations become a∞ (m) a∞ . In the limit where n → ∞.0682 .0968 .0486 .10 .0490 .0688 .0678 .142 .0949 .0198 .0482 .0297 .05 .0199 .0965 .0294 .0665 . i= m=2 3 4 6 12 . a ¨∞ (m) = 1 d(m) (2.0976 .0680 .0494 .137 .36 CHAPTER 2.0487 .0197 .0296 .0197 .0295 .0484 .0198 . the superscript (m) is omitted from all of the annuity (m) and notations.
1/m from the annuitydue starting at time 1/m. ¨ (m) (m) and therefore its present value must be a∞ · ¨ ∞ . but deferred by a time 1/m — is related to the perpetual annuitydue in the obvious way (I (m)a)∞ (m) (m) = v 1/m (I (m)a)∞ ¨ (m) = (I (m)¨)∞ a (m) (1 + i(m)/m) = 1 i(m) d(m) . Clearly the present value is ∞ (I (m) (m) a)∞ ¨ = k=0 m−2 (k + 1) 1 + i(m) m −k Here are two methods to sum this series.2. . Therefore. First. forever. Consider ﬁrst the case of the increasing perpetual annuitydue. which is deﬁned as the present value of a stream of ¨ payments (k + 1)/m2 at times k/m. the second based on actuarial intuition. (ii). (iii). up to the payment of 1/m from the annuitydue starting at time k/m. . without worrying about the strict justiﬁcation for diﬀerentiating an inﬁnite series termbyterm. 1/m. Putting together all of these payment streams gives a total of (k +1)/m paid at time k/m. . Another way to reach the same result is to recognize the increasing perpetual annuitydue as 1/m multiplied by (m) the superposition of perpetuitiesdue a∞ paid at times 0. for k = 0. where the geometricseries formula has been used to sum the second expression. 1. 1/m at (j + 1)/m. ∞ (k + 1) xk = k=0 d dx ∞ xk+1 = k=0 d x = (1 − x)−2 dx 1 − x for 0 < x < 1. consider each annuitydue a∞ paid at a time j/m as ¨ being equivalent to a stream of payments 1/m at time j/m. MORE ON THEORY OF INTEREST 37 We now build some more general annuityrelated present values out of (m) (m) a the standard functions an and ¨ n . etc. with x = (1 + i(m) /m)−1 and 1 − x = (i(m)/m)/(1 + i(m)/m). (m) denoted (I (m)a)∞ . (I (m) (m) a ¨ )∞ =m −2 i(m)/m 1 + i(m)/m −2 = 1 d(m) 2 = a ¨∞ (m) 2 and (2. . . the ﬁrst purely mathematical. . The increasing perpetual annuityimmediate (I (m)a)∞ — the same payment stream as in the increasing annuitydue. .1. As an aid in recognizing ¨ a (m) this equivalence. 2/m. of which 1/m comes from the annuitydue starting at time 0.5) has been used in the last step.
1. . as usual. nm − 1. ﬁll in details of a second. the payment amount at each of the times j/m. . analogous to the second veriﬁcation in pargraph (ii) above. . equate the payments 1 0 = (k + 1)/m2 − n · m − (k − nm + 1)/m2 received at times k/m for k ≥ nm. This is the present value (I (m) ¨)n a of the paymentstream of 2 (k + 1)/m at time k/m.38 CHAPTER 2. The easiest way to obtain the present value is through the identity (I (m) ¨)n a (m) (m) + (D(m) a)n ¨ (m) = (n + 1 (m) ) an ¨ m (2. Evidently. . (v). the method of proving this is to observe that in the paymentstream whose present value is given on the lefthand side. As an exercise. with no further payments at or after time n. in the last line. INTEREST & FORCE OF MORTALITY (iv).4) and (2. . is j +1 j 1 n 1 − 2) = (n + ) + ( 2 m m m m m . . The latter identity is easy to justify either by the ¨ formulas (2. . for j = 0. . recall that v = (1 + i)−1 = (1 + i(m)/m)−m and that (m) (m) an ¨ = a∞ (1 − v n ).) Thus (I (m)¨)n a (m) = (I (m)a)∞ ¨ = a∞ ¨ = a∞ ¨ (m) (m) (m) (m) (m) 1 − v n − n¨∞ v n a (m) (m) a∞ − v n ¨ ∞ + n ¨ a an ¨ − n vn where. (To see this clearly. Now consider the increasing annuitydue of ﬁnite duration (m) n years.5) or by regarding the annuitydue payment stream as a superposition of the paymentstream up to time n − 1/m and the paymentstream starting at time n. intuitive veriﬁcation. for k = 0. nm − 1. this payment(m) stream is equivalent to (I (m)a)∞ minus the sum of n multiplied by an ¨ (m) annuitydue a∞ starting at time n together with an increasing annuity¨ (m) a due (I (m) ¨)∞ starting at time n. The decreasing annuity (D(m) ¨)n a is deﬁned as (the present value of) a stream of payments starting with n/m at time 0 and decreasing by 1/m2 after every timeperiod of 1/m.6) Again.
2/m. and ﬁnally repays or redeems the face or principal amount of the bond at the end of the n year term of the bond.1.2 Loan Repayment: Mortgage. and usually with ﬁrst payment at time 0) for a total duration of n years. (The regular interest payments used to be called coupon payments because of small paper coupons attached to the paper bond document. . at time n − 1/m if the ﬁrst payment was made at time 0) the loan has been completely paid oﬀ. Thus. and which the investor would regularly redeem at a bank. In ordinary consumer purchases or longterm ﬁxerdrate loans on the purchase of a house. to be repaid by equal installments at the end of every period 1/m .7) . . so that at the last payment (the nm’th payment. with ﬁrst payment made at time 1/m and last at time n. 1 Recall that the present value of a payment stream of amount c per year. 1 (m) L i(m) m (1 − v n ) (2. if an amount L has been borrowed for a term of n years. called the face amount of the bond. MORE ON THEORY OF INTEREST 39 2. at ﬁxed nominal interest rate i(m). is c an . . A second kind of plan to repay a loan is to pay equal amounts to cover only the interest amounts accrued every 1/m year on the principal for a duration of n years. This arrangement is used by corporations or government agencies which issue bonds: the borrowing agency receives the loan amount from investors at time 0. Sinking Fund Perhaps the most common application of interest theory is the calculation of the payment amounts needed to repay a loan according to a few standard repayment plans. ‘mortgage’ refers to the way in which the promise to repay is secured by the house or other property purchased with the amount borrowed. with the principal (the original amount borrowed) also repaid in a lump sum at time n.1. We refer to this kind of repayment schedule with level payments as a mortgage loan. and regularly issues interest payments after every period of 1/m year (usually with m = 2 or 4). the usual repayment plan is a series of level or equal payments made m times yearly (usually with m = 12. at scheduled times. n − 1/m. . for the interest payment amount. (m) with c/m paid at times 1/m. n/m. then the level payment c/m or installment payment amount (m) is obtained by equating L = c an Mortgage Payment = L/(m an ) = where v = 1/(1 + i) = (1 + i(m)/m)−m . Bond.2.) The reader should look In legal and historical terms.
will often obligate itself through a formal legal arrangement to devote a certain category of income to a socalled sinking fund.40 CHAPTER 2. . up to the ﬁnal redemption when the principal is paid. Apart from the bond interest payments at a contractual eﬀective interst rate i made directly to investors (the lenders). Remark 2. For m = m or i = i. At (or just after) an intermediate times k/m. 2.2 Note that.3 where loan repayment amounts were formally broken down into interest and principal portions. where the future payouts are not determinstic but rather contingent on the mortality experience of the portfolio of insured lives. if a borrowed amount L is repaid by regular interest payments and a sinking fund. with the intention that the sinking fund will accumulate at its own possibly diﬀerent investment interest rate i to the principal or face amount of the issued bond at time n. We will see. Now a corporation or governmental agency which borrows money from investors by issuing a bond. . the borrowing agency will also pay regular amounts at intervals of 1/m year to the sinking fund trustee for the same term of n years as the bond. a separate calculation is needed.A in Sec. later in this book. an investment account maintained by a trustee. . 1.3 shows that the original face amount or principal is also the principal or balance owed on the loan just after each interest payment. and the result of Exercise 1. .7). in order to conﬁrm the sense of this bond repayment plan. leading to several exercises at the end of the Chapter. the amount built up in the sinking fund is referred to as a reserve toward the ultimate redemption of the principal of the bond.2. . Since each payment by the borrower at times k/m for k = 1. each of the intermediate payments contains 0 principal portion. then evidently the sum of the regular mtimesyearly interest and sinking fund payment at each time k/m is precisely the same as the mortgage payment (2. and if the number of payments per year into the fund is m = m and the eﬀective interest rate for the sinking fund is i = i.2. nm (apart from the ﬁnal lump sum principal repayment) consists of interest only (equal to the original face amount multiplied by (1 + i)1/m − 1 = i(m)/m). that insurance reserves generalize these deterministic reserves to Insurances. at which time the principal is repaid directly to the bond investors. INTEREST & FORCE OF MORTALITY back to Section 1. The following Exercise with sketched solution gives an example.
Find the amount of the level annual sinking fund deposit. how much can be attributed to interest and how much to principal ? Consider the present value at time 0 of the debt for a unit (L = 1) loan amount less the accumulated amounts paid through time k/m : 1 − (m ak/m ) / (m an ) = 1 − (m) (m) 1 − v k/m v k/m − v n = 1 − vn 1 − vn The remaining debt.A. Solution of Exercise 2. the sinking fund must appreciate to the loan amount of L = 107 at time t = 10.055 where a indicates that the annuity is calculated at interest rate i .055)−10 1. k/m = (1 + i)k/m 1 − v n−k/m v k/m − v n = 1 − vn 1 − vn (2.04). so is equal to Bn.8) The amount of interest for a loan amount of 1 after time 1/m is (1 + i)1/m − 1 = i(m) /m.7).A. Therefore the interest included in the payment at . valued just after time k/m. 2. MORE ON THEORY OF INTEREST 41 Exercise 2. Of the payment made at time (k + 1)/m. from tax receipts.2. per unit of loan amount.1.1. is denoted from now on by Bn. It is greater than the displayed present value at 0 by a factor (1 + i)k/m . So the annual sinking fund deposit D is found through the equality L = 107 = (1 + i )10 · D a10 = D 1 − (1. k/m . on which its ﬁnancial advisors claim it can safely earn an eﬀective annual rate of i = . Therefore the sinking fund present value at time 10 (equal to its present value at t = 0 accumulated by the factor (1 + i )10 must be equated to L = 107 . A small city issues a bond for ten million dollars for ten years at 4% nominal quarterly interest (m = 4) and creates a sinking fund into which it will make annual deposits (m = 1).3 Loan Amortization & Mortgage Reﬁnancing We analyze next the breakdown between principal and interest in repaying a mortgage loan by level payments (2.055. Since the interest payments are made at the contractual interest rate i (corresponding to i(4) = .05510 .
The general deﬁnition of the principal repaid in each payment is the excess of the payment over the interest since the past payment on the total balance due immediately following that previous payment. nominalrate 8%. 7%. $100.) Suppose that the new pattern of payments is to be valued at each of the nominal interest rates 6%. These formulas show in particular that the amount of principal repaid in each successive payment increases geometrically in the payment number.(k+1)/m immediately after time (k + 1)/m is recomputed as Bn. and that these valuations will be taken into account in deciding whether to take out the new loan. and costs plus points are then extra amounts added to the initial balance of the reﬁnanced mortgage.1.4 Illustration on Mortgage Reﬁnancing Suppose that a 30–year. or 8%. Note as a check on the displayed formulas that the outstanding balance Bn. the principal included in each payment is the amount of the payment minus the interest included in it. = 2. 000 mortgage payable monthly is to be reﬁnanced at the end of 8 years for an additional 15 years (instead of the 22 which would otherwise have been remaining to pay it oﬀ) at 6%.42 CHAPTER 2. due to uncertainty about what the interest rate will be in the future. k/m of outstanding debt just after k/m.9) v −(k+1)/m 1 − vn as was derived above by considering the accumulated value of amounts paid. Thus the next total payment of i(m) /(m(1 − v n )) consists of the two parts Amount of interest = m−1 i(m) (1 − v n−k/m )/(1 − v n) Amount of principal = m−1i(m) v n−k/m /(1 − v n ) By deﬁnition. (The points are each 1% of the reﬁnanced balance including closing costs. . with a reﬁnancing closingcost amount of $1500 and 2 points. INTEREST & FORCE OF MORTALITY (k + 1)/m is i(m)/m multiplied by the value Bn. which at ﬁrst seems surprising. or i(m) v n−k/m 1 − v n−k/m (1 + i(m) /m) 1 − v n−k/m − = 1 − vn m 1 − vn 1 − vn 1 − v n−(k+1)/m (m) (m) = 1 − a(k+1)/m an (2. k/m minus the interest paid at (k + 1)/m.
08/12)96 = $91.57 Even with these diﬀerent assumptions. and the present value as of time 0 (the beginning of the old loan) of the payments made through the (12) end of the 8th year is ($733. if the new rate of 6% were really to be the correct one for the next 22 years. of the payments still to be made under the old mortgage.31 + 1850.69. MORE ON THEORY OF INTEREST 43 The monthly payment amount of the initial loan in this example was $100. The respective present values (as of the end of the 8th year) at nominal rate of 7% of these two streams are: Old loan: 733. and the one after reﬁnancing. is $(100. If 2 points must be paid in order to lock in the rate of 6% for the reﬁnanced 15year loan. as of the end of 8 years. continuing another 22 years into the future.00 = $92. 904. if the loan were to be reﬁnanced.34 (12a15 ) = $88. then this amount is (. 904. The new principal balance of the reﬁnanced loan is 92518. 000 − 51. 500. no matter what the term of the reﬁnanced loan is. 368. and this is the present value at a nominal rate of 6% of the future loan payments.37 .37 = $94. (12) (12) . the economic rate of interest would be a nominal 7% for the next 22 years.31. suppose instead that right after reﬁnancing. 597. 018.34. The new monthly payment (for a 15year duration) of the reﬁnanced loan is $94. Next.76) · (12a8 ) = $51. Thus. 018. 518. For purposes of comparison. it is well worth reﬁnancing.68. the new reﬁnanced loan amount would be $91.1. so the present value at 6% is $733.76 · (12a22 ) = $107.2.06 New loan: 796.76.31. Thus the present value.08/12)−360 ) = $733. what is the present value at the current going rate of 6% (nominal) of the continuing stream of payments under the old loan ? That is a 22year stream of monthly payments of $733. and each loan would be paid to the end of its term.68(.06/12)/(1 − (1 + .31 = $1850.31 + 1. 700.02)92518.69)(1 + . then it would be a ﬁnancial disaster not to reﬁnance.76 (12a22 ) = $98. and despite closingcosts and points. (12) as calculated above. 000(.06/12)−180 ) = $796. In that case both streams of payments would have to be revalued — the one before reﬁnancing. 368.21.76. continuing 15 years into the future. Thus.08/12)/(1 − (1 + . 420.
At the time of sale. balances for purposes of reﬁnancing both before and after application of administrative costs and points. In this section. at the time of selling the house on which the mortgage loan was negotiated. which is equal to RefBal + Costs. and the present value under any interest rate (not necessarily the ones at which either the original or reﬁnanced loans are taken out) of the stream of repayments to the bank up to and including the lumpsum payoﬀ which would be made. you would pay oﬀ the cash principal balance. 2. Concerning the syntax of R. Venables & Ripley 2002) is provided to do some comparative reﬁnancing calculations. The outputs of the function are as follows. Calculate and compare the present values (at each of 6%. an R function (cf. m = 12. Lines beginning with the symbol # are commentlines. INTEREST & FORCE OF MORTALITY Exercise 2. The function RefExmp given below calculates mortgage payments. as described above. (a) if you continue the old loan without reﬁnancing. Oldpayment is the monthly payment on the original loan of faceamount Loan at nominal interest i(12) = OldInt for a term of OldTerm years. whatever it is. and ∧ denotes exponentiation. Suppose that you can forecast that you will in fact sell your house in precisely 5 more years after the time when you are reﬁnancing. and the reﬁnanced loan amount is a multiple 1+ Points of NewBal.5 Computational illustration in R All of the calculations described above are very easy to program in any language from Fortran to Mathematica. and k/m = RefTim.44 CHAPTER 2. and 8% nominal interest rates) of your payment streams to the bank. and (b) if you reﬁnance to get a 15year 6% loan including closing costs and points. NewBal is the balance Bn. horizontally as a vector. The new loan. k/m of formula (2. for example.8) for n = OldTerm. the only explanation necessary at this point is that * denotes multiplication. 7%. but they are also very handily organized within a spreadsheet. and also on a programmable calculator. is displayed in ‘unlisted’ form. bankoﬃcials.1. has . which seems to be the way that MBA’s.B. at nominal interest rate NewInt. in each numerical example below. and actuaries will learn to do them from now on. The output of the function is a list which.
# PayoffTim (no bigger than NewTerm) = time (from refinancing# time at which new loan balance is to be paid off in # cash (eg at house sale). NewInt. # OldTerm = term of initial loan in years.vnew^(NewTerm PayoffTim)))/(1 . # RefTim = time in years after which to refinance. # Costs = fixed closing costs for refinancing. # Loan = original loan amount. # Points = fraction of new balance as additional costs.vold^(OldTerm .vnew^NewTerm) list(Oldpaymt = Oldpaymt.vnew^NewTerm) vval = (1 + ValInt/12)^(12) Value = (Newpaymt * 12 * (1 . ValInt are # nominal 12timesperyear. and monthly payments # are calculated.2. NewBal = NewBal. RefTim. NewInt. MORE ON THEORY OF INTEREST 45 R FUNCTION CALCULATING REFINANCE PAYMENTS & VALUES RefExmp function(Loan.vval^PayoffTim))/ValInt + (RefBal * vval^PayoffTim * (1 . RefBal = RefBal.vold^OldTerm) NewBal = (Loan * (1 . ValInt) { # Function calculates present value of future payment stream # underrefinanced loan. OldInt. # The three interest rates OldInt. vold = (1 + OldInt/12)^(12) Oldpaymt = ((Loan * OldInt)/12)/(1 . Newpaymt = Newpaymt. Value = Value) } . PayoffTim. Points.RefTim)))/ (1 . NewTerm. OldTerm. Costs.1.vold^OldTerm) RefBal = (NewBal + Costs) * (1 + Points) vnew = (1 + NewInt/12)^(12) Newpaymt = ((RefBal * NewInt)/12)/(1 . # NewTerm = term of refinanced loan.
06. since the payments under the new (reﬁnanced) loan are here valued at the same interest rate as the loan itself. with level monthly payments. 0.22.8.76 91018 94368 796.33 94368 Note that.06)) Oldpaymt NewBal RefBal Newpaymt Value 733. 0. 7.0. the arguments must reﬂect a ‘reﬁnance’ with no costs or points for a period of 22 years at nominal rate 6%.46 CHAPTER 2.76 91018 91018 733. Suppose that you have a 30year mortage for $100. in valuing the old loan at 7%. 0. INTEREST & FORCE OF MORTALITY monthly payments Newpaymt for a term of NewTerm years. 1500.0. the present value Value of all payments made under the loan must be equal to the the reﬁnanced loan amount RefBal. and that after 7 years of payments you reﬁnance to obtain a new 30year mortgage at 7% nominal interest ( = .22. 0. 15. and the ﬁnal output of the function is the present value at the start of the reﬁnanced loan with nominal interest rate ValInt of the stream of payments made under the reﬁnanced loan up to and including the lump sum payoﬀ. 30.) We consider next a numerical example showing breakeven point for reﬁnancing by balancing costs versus time needed to amortize them.08.0. are all recapitulated easily using this function.08. 8. for example. at (nominal) interest rates of 6.02.0.0. The comparisons of the previous Section between the original and reﬁnanced loans. 15. as follows: > unlist(RefExmp(100000.30.76 98701 (The small discrepancies between the values found here and in the previous subsection are due to the rounding used there to express payment amounts to the nearest cent. and 8 %.000 at nominal rate i(12) = 9%.07)) Oldpaymt NewBal RefBal Newpaymt Value 733. To use it.08. The loan is to be paid oﬀ PayoﬀTim years after RefTim when the new loan commences. We begin our numerical illustration by reproducing the quantities calculated in the previous subsection: > unlist(RefExmp(100000.
. the value under the old loan for payoﬀ time 0 (i. MORE ON THEORY OF INTEREST 47 i(m) for m = 12).29 98946 Next we compute the value of payments under the old loan.07)) Oldpaymt NewBal RefBal Newpaymt Value 804.09. at 7% nominal rate.09. Thus..62 106042 The values found in the same way when the payoﬀ time K is successively replaced by 4. As remarked above in the previous example. > unlist(RefExmp(1.09. 0. the payoﬀtime K at which there is essentially no diﬀerence in present value at nominal 7% between the old loan or the reﬁnanced loan with costs and points (which was found to have Value 98946). .0. then the remaining paymentstreams for both loans from the time when you reﬁnance are equivalent (i. 98593.. For comparison. 98946. 0. have the same present value from that time) ? We ﬁrst calculate the present value of payments under the new loan.25 are 99979. 0.62 93640 98946 658. also with level monthly payments.07.23. the valuation of the new loan does not depend upon the time K to payoﬀ.7.04.) > unlist(RefExmp(1. also at payoﬀ time K = 10.07)) Oldpaymt NewBal RefBal Newpaymt Value 804. Figuring present values using the new interest rate of 7%.0. 1500.. 3.167.0. (It is ﬁgured here as though the payoﬀ time K were 10 years. 98951. 10.e.62 93640 93640 804. 3. with closing costs of $1500 and 4 points (i.e.0.30.1. 30. 30. what is the time K (to the nearest month) such that if both loans — the old and the new — were to be paid oﬀ in exactly K years after the time (the 7year mark for the ﬁrst loan) when you would have reﬁnanced.e5. 4% of the total reﬁnanced amount including closing costs added to the initial balance). for cash payoﬀ at the time when reﬁnancing would have occurred) coincides with the New Balance amount of $93640.0. since the same interest rate is being used to value the payments as is used in ﬁguring the reﬁnanced loan.e5. is 3 years and 3 months after reﬁnancing. 10. 3.2.e.7.
the function S(t) called the survivor or survival function has been deﬁned to be equal to the lifetable ratio lx /l0 at all integer ages t = x. The topic called “Graduation Theory” among actuaries is the mathematical methodology of Interpolation and Splinesmoothing applied to the raw function qx = dx /lx . The way in which this quantity varies with x is one of the most important topics of study in actuarial science. An equivalent representation is S(y) = y f (t) dt .2 Force of Mortality & Analytical Models Up to now. The rate qx would be estimated from the cohort life table as the ratio dx /lx of those who die between ages x and x + 1 as a fraction of those who reached age x. Intuitively. INTEREST & FORCE OF MORTALITY 2. S(y) − S(y + t) is the fraction of the initial lifetable cohort which dies between ages y and y + t. the notation 1 qx is replaced by qx . If T denotes the random variable which is the age at death for a newly born individual governed by the same causes of failure as the lifetable cohort. are: t py = P T ≥ y + t T ≥ y = S(y + t)/S(y) and t qy = 1 − t py = P T ≤ y + t  T ≥ y = (S(y) − S(y + t))/S(y) The quantity t qy is referred to as the agespeciﬁc death rate for periods of length t. and to be piecewise continuously diﬀerentiable for all positive real values of t. for all positive real y and t. In the most usual case where t = 1 and y = x is an integer. . and according to the Fundamental Theorem of Calculus. For example. lim P (y ≤ T ≤ y + ) = lim 1 y y+ →0+ →0+ f (t) dt = f (y) as long as the failure density is a continuous function.48 CHAPTER 2. Two further useful actuarial notations. and (S(y) − S(y + t))/S(y) represents the fraction of those alive at exact age y ∞ who fail before y + t. then P r(T ≥ y) = S(y). and 1 px is replaced by px . often used to specify the theoretical lifetime distribution. one important way in which numerical analysis enters actuarial science is that one wishes to interpolate the values 1qy smoothly as a function of y. where f (t) ≡ −S (t) is called the failure density.
) This exponential behavior of the agespeciﬁc deathrate for large ages suggests that the deathrates pictured could reasonably be extrapolated to older ages using the formula qx ≈ q78 · (1.34 per thousand in the eleventh year.0885 was found as log(q78/q39 )/(78 − 39). which are as likely due to statistical irregularity as to real increases in risk. roughly proportionately to . The reasoning above shows that for small qy Thus µ(y) = = 1 S(y) y+ . there is an erratic but roughly linear increase of deathrates per thousand from 0. As can be seen from Figure 2. Additional granularity in the deathrates can be seen in Figure 2.) Between ages 11 and 40.08. x ≥ 79 (2.0.0885)x−78 .10].2.4 to 3. or hazard intensity.3 deaths per thousand live births). 1.2. (But there were small increases in rate from ages 4 to 7 and from 8 to 9. failure rate.10) where the number 1. f (y) .2. After a very high deathrate during the ﬁrst year of life (26. S(y) 0 f (t) dt −→ y f (y) −S (y) d = = − ln(S(y)) S(y) S(y) dy . Deﬁnition: The limiting deathrate qx / per unit time as 0 is called by actuaries the force of mortality µ(x).45 per thousand in the second year to 0. It is clear that qx Now consider the behavior of qx as must also get small. the same function is called the failure intensity. However.1 pictures the agespeciﬁc 1year deathrates qx for the simulated lifetable given as Table 1. where the logarithms of deathrates are plotted. there is a yearbyyear decline in deathrates roughly from 1.2. Figure 2. since the probability of dying between ages x and x + is approximately f (x) when gets small. at ages beyond 40 there is a rapid increase in deathrates as a function of age.1 on page 4. In reliability theory or biostatistics. FORCE OF MORTALITY & ANALYTICAL MODELS 49 To give some idea what a realistic set of deathrates looks like. the values qx seem to increase roughly as a power cx where c ∈ [1. gets small. (Compare this behavior with the GompertzMakeham Example (v) below.
1: Plot of agespeciﬁc deathrates qx versus x. for the simulated illustrative life table given in Table 1.50 CHAPTER 2.02 0. INTEREST & FORCE OF MORTALITY AgeSpecific Death Rates for Illustrative Life Table • 0.0 • •• ••••••••••••••••••••••••••••••••••••• • 0 20 40 •••• • ••• •• •• • • •• • • 60 80 Age in Years Figure 2.1.06 • AgeSpecific Death Rate • • •• • 0. page 4.04 • • • • • • • • •• 0. .08 • • • 0.
2.1. page 4. for the simulated illustrative life table given in Table 1. The rates whose logarithms are plotted here are the same ones shown in Figure 2. FORCE OF MORTALITY & ANALYTICAL MODELS 51 Logarithm of DeathRates versus Age for Illustrative Simulated LifeTable ••• •• •• • 4 • •• •• •• •• •• ••• ••• 3 log(Death rate) 5 •• 6 • ••• • • • • •• • •• • •• • • 7 •••• • • • 0 8 • • • • ••• • ••••••• •••• •• • • • ••• 20 40 Age in years 60 80 Figure 2. .1.2: Plot of logarithm log(qx ) of agespeciﬁc deathrates as a function of age x.2.
y+t µ(s) ds = ln S(y) − ln S(y + t) y Now exponentiate to obtain the useful formulas y S(y) = exp − 0 µ(t) dt . The force of mortality takes the form µ(t) = λ γ tγ−1 This model is very popular in engineering reliability. INTEREST & FORCE OF MORTALITY where the chain rule for diﬀerentiation was used in the last step. (iii) If S(t) = exp(−λtγ ) for t ≥ 0. ∞) ). then µ(t) = (ω − t)−1 . It has the ﬂexibility that by choice of the shape parameter γ one can have (a) (γ > 1) failure rate increasing as a function of x (b) ( γ = 1) constant failure rate (exponential model). we ﬁnd y y µ(t) dt = 0 − ln(S(t)) 0 = − ln(S(y)) since S(0) = 1. ω] ).52 CHAPTER 2. (ii) If S(t) = e−µt for t ≥ 0 (the exponential failure distribution on [0. Replacing y by t and integrating both sides of the last equation between 0 and y. then µ(t) = µ is constant. Similarly. or (c) (0 < γ < 1) decreasing failure rate. . Note that this hazard function increases to ∞ as t increases to ω. t py = S(y + t) = exp S(y) y+t − y µ(s) ds Examples: (i) If S(t) = (ω − t)/ω for 0 ≤ t ≤ ω (the uniform failure distribution on [0. then mortality follows the Weibull life distribution model with shape parameter γ > 0 and scale parameter λ.
α. Between infancy and . is a forceofmortality function which decreases on part of the timeaxis and increases elsewhere. β > 0 and the Lognormal. since the highmortality diseases like heart disease and cancer strike with greatest eﬀect at higher ages. (iv) Two other models for positive random variables which are popular in various statistical applications are the Gamma. In the Gamma case. human life tables also exhibit an aging eﬀect at high ages. Of course. which really does occur in certain situations. m real. σ > 0 du √ 2π 0 is called the standard normal distribution function. would occur if mortality arose only from pure accidents unrelated to age. reﬂects what engineers call “burnin”. In human life tables. Constant force of mortality.2. with S(y) = 1 − Φ where ln y − m σ 1 + 2 z . the expected lifetime is α/β. where after a period of initial failures due to loose connections and factory defects the nondefective devices emerge and exhibit high reliability for a while. the expectation is exp(m + σ 2/2). with ∞ ∞ S(y) = y β α tα−1 e−βt dt / 0 z α−1 e−z dz . The decreasing force of mortality reﬂects the fact that the devices known to have functioned properly for a short while are known to be correctly assembled and are therefore highly likely to have a standard length of operating lifetime. in the examples considered so far. while in the Lognormal. Neither of these last two examples has a convenient or interpretable forceofmortality function.2. where the causes of death operate with greater intensity or eﬀect at greater ages. which is easily seen from the formula S(y) = y exp(− 0 µ(t) dt) to be equivalent to exponential failure distribution. Decreasing force of mortality. FORCE OF MORTALITY & ANALYTICAL MODELS 53 But what one cannot have. infant mortality corresponds to burnin: risks of death for babies decrease markedly after the oneyear period within which the most severe congenital defects and diseases of infancy manifest themselves. Φ(z) ≡ e−u 2 /2 Increasing force of mortality intuitively corresponds to aging.
but not both.1. at least in western countries. Suppose that µ(y) is deﬁned directly to have the form A + B cy . (GompertzMakeham forms of the force of mortality). initially decreasing and later increasing. ﬂat middle.084. which may be a reasonable value except for very advanced ages.1. Figures 2. (Compare the comments made in connection with Figures 2. by adding a linear term Dy. If D < −B ln(c). and ﬁnal increase of the forceofmortality. hazard rates are relatively ﬂat. Actuaries have developed an analytical model which is somewhat more realistic than the preceding examples for human mortalty at ages beyond childhood. or µ(y) = Bcy .54 CHAPTER 2.4 display the shapes of forceofmortality functions (iii)(v) for various parameter combinations chosen in such a way that the ex . Roughly realistic values of c for human mortality will be only slightly greater than 1: if the Gompertz (nonconstant) term in forceofmortality were for example to quintuple in 20 years.3 and 2. then c ≈ 51/20 = 1. with further beneﬁts of realism. As shown above.) By choice of the parameter c as being respectively greater than or less than 1. Thus the Gompertz forceofmortality model is the special case with A = 0. the ﬁgure of c was found to be roughly 1.2: for middle and higher ages in the simulated illustrative life table of Table 1.) Note that in any case the GompertzMakeham force of mortality is strictly convex (i. The GompertzMakeham family could be enriched still further. c > 1.1 and 2. has strictly positive second derivative) when B > 0 and c = 1. is called a bathtub shape and requires new survival models. the failure models in common statistical and reliability usage either have increasing force of mortality functions or decreasing force of mortality.. which corresponds roughly to US male mortality of around 1960.e.09. seen clearly in Figure 2. a simple modiﬁcation of it does. one can arrange that the forceofmortality curve either be increasing or decreasing. This pattern of initial decrease. While the standard form of this model does not accommodate a bathtub shape for deathrates. Example (v). INTEREST & FORCE OF MORTALITY late middle age. the additive constant A by Makeham. (The Bcy term was proposed by Gompertz. with 0 < A < B. then it is easy to check that µ(y) = A + B cy + Dy has a bathtub shape.
The models ﬁtted all assumed that S(40) = 0. 0.10) to extrapolate the 1959 lifetable deathrates to older ages. The previous discussion about bathtubshaped force of mortality functions should have made it clear that none of the analytical models presented could give a good ﬁt at all ages. Thus.04). Figure 2. but that the lognormal and Makeham families are quite diﬀerent. clearly the Gompertz agrees most closely with the plotted points for 1959 US male mortality.1 Comparison of Forces of Mortality What does it mean to say that one lifetime.5) and probability 0.e.925)..04 of surviving to age 90. FORCE OF MORTALITY & ANALYTICAL MODELS 55 pected lifetime is 75 years. t t − ln S1(t) = 0 µ1 (x)dx = 0 κ µ2 (x)dx = κ(− ln S2 (t)) and therefore that S1 (t) = (S2 (t))κ .0885 used in formula (2. 0.2.00346.e.2. has hazard (i. Of the four ﬁtted curves. Parameters for all models were determined from the requirements of median age 72 at death (equal by deﬁnition to the value tm for which S(tm ) = 0. 0. with associated survival function S1(t).5 shows survival curves from several analytical models plotted on the same axes as the 1959 US male lifetable data from which Table 1.0918. 2. (90.2.925 and that for lives aged 40.5). This restriction has the eﬀect of reducing the number of free parameters in each family of examples by 1. One can see from these pictures that the Gamma and Weibull families contain many very similar shapes for forceofmortality curves. all t ≥ 0 . T − 40 followed the indicated analytical form. i. c = 1.1 was simulated. force of mortality) µ1 (t) which is a constant multiple κ at all ages of the force of mortality µ2 (t) for a second lifetime with survival function S2 (t) ? It means that the cumulative hazard functions are proportional. but the Figure indicates the rather good ﬁt which can be achieved to realistic lifetable data at ages 40 and above. all four plotted survival curves have been designed to pass through the three points (40. the latter of which is close to the value 1. (72. The Gompertz curve has parameters B = 0.
025 0.020 alpha=1 alpha=0.015 Hazard 0.56 CHAPTER 2.0 0 20 40 60 80 100 0.010 Hazard 0.lambda) 0.7 alpha=1.3: Force of Mortality Functions for Weibull and Gamma Probability Densities. .7 beta=1.015 0.lambda) 0.0 0 0.3 beta=1.005 0.020 0.005 20 40 60 80 100 Age (years) Age (years) Figure 2. the parameters are ﬁxed in such a way that the expected survival time is 75 years. INTEREST & FORCE OF MORTALITY Weibull(alpha.025 Gamma(beta. In each case.6 alpha=2 beta=1 beta=0.6 beta=2 0.010 0.3 alpha=1.
025 Makeham(A.2 sigma=0.002 A=0. B=0.B.0 0 20 40 60 80 100 Age (years) Age (years) Figure 2.2.c) 0. FORCE OF MORTALITY & ANALYTICAL MODELS 57 Lognormal(mu. In each case.2.4 sigma=0.sigma^2) 0.020 A=0. c=1. c=1. c=1.022 0.007 A=0.6 sigma=1.003. the parameters are ﬁxed in such a way that the expected survival time is 75 years.014 A=0.0018.015 0. c=1.020 0.0041.0041.010 0.4: Force of Mortality Functions for Lognormal and Makeham Densities.0070.005 sigma=0.B=0.005 0.0010. B=0. . B=0.6 sigma=2 0.025 0.010 Hazard 0.015 Hazard 0.0 0 20 40 60 80 100 0.0022.0021.
04 respectively at t = 40. Gamma.953e6) Gamma(14.74. 72.246^2) Weibull(3.2 Plotted points from US 1959 male lifetable 0.925.5.8 0. The four analytical survival curves — Lognormal.925 · Stheor (t − 40)).4383) Gompertz(3. 0.5: Theoretical survival curves. the plotted curve is (t. 90. 0. Weibull. INTEREST & FORCE OF MORTALITY Plots of Theoretical Survival Curves ••••• ••• ••• •• •• •• •• • • • • Lognormal(3. 0. 0. for ages 40 and above. .0 40 50 60 70 Age (years) 80 90 100 Figure 2.6 • Survival Probability • • • • • • 0.4 • • • • • • 0.0918) • • • • 0. so if Stheor (t) denotes the theoretical survival curve with indicated parameters. plotted as lines for comparison with 1959 US male lifetable survival probabilities plotted as points.491.58 CHAPTER 2.653. and Gompertz — are taken as models for ageatdeath minus 40. .46e3. 1. 1. The parameters of each analytical model were determined so that the plotted probabilities would be 0.
Z2 = 0) = 0.4493 for those with Z2 = 1 versus those with Z2 = 0. related to the fractions of the surviving population at various ages in each of the four population subgroups.35 and that for a life aged x and all t > 0. Px (Z1 = 0. Z2 = z2 ) = exp(−2. This model is called the (Cox) ProportionalHazards model and is treated at length in books on survival data analysis (Cox and Oakes 1984.7 = 2.8 = 0. Example.5e0.0085 1 − exp(−2. The eﬀect on agespeciﬁc deathrates is approximately the same. and is multiplied by e−0.7z1 −.8z2 ) t/10000 so that the force of mortality at all ages is multiplied by e0. Z2 = 0) versus those in the group with (Z1 = 0. by a regression model ln(κ) = β ·Z ) on other measured variables (covariates) Z.5 e0. 1.014. Z1 = z1 .7((212 − 202 )/20000) = 2. For example. P r(T ≥ x + t  T ≥ x.7 = 2. Px (Z1 = Z2 = 1) = 0. Suppose that these four combinations have respective conditional probabilities for lives aged x (or relative frequencies in the general population aged x) Px (Z1 = Z2 = 0) = 0. Direct calculation shows for example that the ratio of agespeciﬁc death rate at age x+20 for individuals in the group with (Z1 = 1.7z1−. Z2 = 1  T ≥ x + 30) .. but rather 1 − exp(−2. Consider a setting in which there are four subpopulations of the general population.8z2 t2 /20000) It can be seen from the conditional survival function just displayed that the forces of mortality at ages greater than x are µ(x + t) = (2.2.3 .5((212 − 202 )/20000) Various calculations. categorized by the four combinations of values of two binary covariates Z1 .5 e0.2 Px (Z1 = 1. to ﬁnd P r(Z1 = 0.0138 for individuals with Z1 = 1 versus those with Z1 = 0. FORCE OF MORTALITY & ANALYTICAL MODELS 59 This remark is of especial interest in biostatistics and epidemiology when the factor κ is allowed to depend (e. Z2 = 0) is not precisely e0.g. Z2 = 1) = 0.2. can be performed easily . Kalbﬂeisch and Prentice 1980) or biostatistics (Lee 1992). Z2 = 0.15 .
i.. Hogg and Tanis 1997. occupations. while diﬀerences 302 ) = 0. .2162 0.8 and similarly P r(T ≥ x + 30  T ≥ x) = 0. by deﬁnition of conditional probabilities (restricted to the cohort of lives aged x).35 exp(−2. for example by increasing premiums in response to recent claims or by taking location into account. the measured variables Z = (Z1 .60 CHAPTER 2.. the objective of such detailed models of covariate eﬀects on survival can be: to correct for incidental individual diﬀerences in assessing the eﬀectiveness of a treatment. categorical variables for risky lifestyles.8795 20000 20000 Thus. The multiplicative eﬀects of various riskfactors on agespeciﬁc death rates are often highlighted in the news media.e.5e−0. INTEREST & FORCE OF MORTALITY we proceed in several steps (which correspond to an application of Bayes’ rule.5. In particular. percent caloric intake from fat.3 exp(−2.7 P r(Z1 = 0.1901 + 302 302 ) + 0.8 ) = 0.15 exp(−2. .2 exp(−2. sec.1901 20000 . . or exposure to a toxic chemical). 2. or exposures might be used in riskrating. in individualizing insurance premiums. it can be politically sensitive in a lifeinsurance and pension context. or indicator of type of treatment or intervention. to create a prognostic index for use in diagnosis and choice of treatment. or Devore 2007): P r(T ≥ x+30. high blood pressure. . viz. Z1 = 0 Z2 = 1 T ≥ x) = 0. quantitative measurement of a riskfactor (dietary cholesterol. In these ﬁelds. speciﬁc electrocardiogram anomaly). taking ratios of the last two displayed quantities yields + 0. Z2 = 1  T ≥ x + 30) = 0.7−0. While riskrating is used routinely in casualty and property insurance underwriting. diabetes. In an insurance setting.5e0.5 ∗ e0. Zp ) recorded for each individual in a survival study might be: indicator of a speciﬁc disease or diagnostic condition (e. or to ascertain the possible risks and beneﬁts for health and survival from various sorts of lifestyle interventions. relative weighttoheight index.1901 = 0.5(302 /20000)) + 0.g.8795 2 In biostatistics and epidemiology.
for 20 ≤ t ≤ 80 and 0 for other values of t. i.e. (The random variable Z in this problem is a particular type of ageatdeath variable T conditioned on being ≥ 20. such as employee groups or voluntary organizations.3 Exercise Set 2 (1).4 − Z/50) dollars at the exact date of his death if this occurs between ages 20 and 70. by the nominal interest rate of e0. compounded annually. Find (1 + r)2n .3. then what is the expected present value of the payment under the insurance contract ? . All life insurers must be conscious of the extent to which their policyholders as a group diﬀer from the general population with respect to mortality. Suppose that an individual aged 20 has random lifetime Z with continuous density function f( t) = 1 360 1+ t 10 . EXERCISE SET 2 61 in mortality by gender and to some extgent by family health history can be used in calculating insurance and annuity premiums. See Chapter 6. (2). and regressiontype models like the Cox proportionalhazards model may be useful in quantifying group mortality diﬀerences when the specialgroup mortality tables are not based upon large enough cohorts for long enough times to be fully reliable.2. it is currently illegal to use racial diﬀerences and diﬀerences based on genetic testing in this way.08 − 1 per year) to calculate the present value of the payment. Insurers can collect special mortality tables on special groups. as can certain lifestyle factors like smoking.) (a) If this individual has a contract with your company that you must pay his heirs 106 · (1. 2. The sum of the present value of $1 paid at the end of n years and $1 paid at the end of 2n years is $1. for discussion about the modiﬁcation of insurance premiums for select groups.08 · (Z − 20)). where r = annual interest rate. then what is the expected payment ? (b) If the value of the deathpayment described in (a) should properly be discounted by the factor exp(−0. Section 4.
give only the values for ages which are multiples of 10.62 CHAPTER 2.) (7). based on integer ages up to 70 and cohortsize (= radix ) l0 = 105 . What does this imply about S(x) ? (Give as much information about S as you can. for every real ≥ 0. (5). (4).1. is (17 − 10 ln(5)) · 10−4 = 9. (a) The mortality pattern of a certain population may be described as follows: out of every 98 lives born together. survival function based on a cohort life table) has the property that 1 px = γ · (γ 2)x for some ﬁxed γ between 0 and 1.01 · t for 0 ≤ t ≤ 3. and ﬁnd the probability that a life aged 30 will survive to attain age 35. then ﬁnd the equivalent single eﬀective rate of interest for money invested at interest throughout the interval 0 ≤ t ≤ 3.1 · 10−5 > 0. If you do the arithmetic using handcalculators and/or tables. one dies annually until there are no survivors.B in the Illustration on mortgage reﬁnancing at the end of Section 2.) (b) Find the probability that the random variable T exceeds 30. 10% of the lives aged x die before reaching age x+1 . . (Give selected numerical entries.. Find a simple function that can be used as S(x) for this population. Hint: ﬁnd a closedform formula for S(t) = P (T ≥ t). and ﬁnd the probability that a life aged 30 will survive to attain age 35.5t + 2et/20 . INTEREST & FORCE OF MORTALITY (3).0 − 0. Suppose that a survival distribution (i. given that it exceeds 3. t>0 This is a legitimate hazard rate of GompertzMakeham type since its minimum. (b) Suppose that for x between ages 12 and 40 in a certain population.e. Suppose that a continuous random variable T has hazard rate function (= force of mortality) h(t) = 10−3 · 7. preferably calculated by means of a little computer program. (a) Construct a cohort lifetable with h(t) as “force of mortality”. Do the MortgageReﬁnancing exercise given as Exercise 2. If the instantaneous interest rate is r(t) = 0. Find a simple function that can be used as S(x) for this population. (6). which occurs at t = 20 ln(5).
pY +1 = (1 + r) pY t t where k is a function of Y alone and A. A survival function has the form S(t) = max(0. (12). (a) Find the level payment amount P .3. beginning with Y = 1950.e.e. and if l35 = 44. If a mortality c+t table is derived from this survivalfunction with a radix l0 of 100. then derive a general expression for k(Y ). If k(1950) = 1. 000 : (i) What is the terminal age of the table ? (ii) What is the probability of surviving from birth to age 60 ? (iii) What is the probability of a person at exact age 10 dying between exact ages 30 and 45 ? (11). For each Y and for all ages t µY (t) = A · k(Y ) + B ct . ) (10). c−t ). to be repaid in level payments every six months (twice yearly). (b) What is the present value of the payments you will make if you skip the 2nd and 4th payments ? (You may express your answer in terms of P . The mortality functions for the various tables are denoted by the appropriate superscript Y . (9). A standard mortality table follows Makeham’s Law with force of mortality µ(t) = A + B ct at all ages t A separate.2. i(4) ) is 8% for the ﬁrst 5 years. higherrisk mortality table also follows Makeham’s Law with . r are constants (with r > 0). EXERCISE SET 2 63 (8). Suppose that you borrow $1000 for 3 years at 6% eﬀective rate. and if the nominal rate of discount compounded semiannually (m = 2) is 6% for the third 5 year interval. the interval ranging from time 5 to 10). Find the accumulated value of $100 at the end of 15 years if the nominal interest rate compounded quarterly (i. A separate life table has been constructed for each calendar year of birth. Y . B. if the eﬀective rate of discount is 7% for the second 5 year interval (i..000 at age 0.
(14). Calculate d(12).06). INTEREST & FORCE OF MORTALITY µ∗ (t) = A∗ + B ∗ ct at all ages t with the same constant c. 000 30year coupon bond with nominal 6% semiannual coupon (n = 30. B. m − 2. 15 years after issue. 000 at eﬀective annual rate 7% from a bank. the eﬀective interest rate will be ieﬀ = 0. Find the outstanding balance at the 6th anniversary of the loan. ﬁgured at 7% ). . (16). the homeowner misses the next two (i. 000. A deposit of 300 is made into a fund at time t = 0. A 6% ‘zerocoupon’ 30year bond was issued exactly 15 years ago for a face amount of $10. if for the next 15 years. the accumulated value of the fund is 574. Calculate the price at which you would sell a $10. if for the next 15 years. agreeing to repay by 30 equal yearly payments beginning one year from the time of the loan. Calculate the fair price at which you would sell this zerocoupon bond. interest is credited according to the force of interest δt = 1/(3t + 3). i(m) = 0. This bond contractually entitles the bearer to receive 30 years after the issue date the amount accumulated at i = ieﬀ = 0. the eﬀective interest rate for valuation is ieﬀ = 0.64 force of mortality CHAPTER 2.07. if paid as a lump sum at time 6. (13). As of time t = 7. 000 at time 0. From t = 3 to t = 7. A homeowner borrows $100. (a) How much is each payment ? (b) Suppose that after paying the ﬁrst 3 yearly payments. then express each of A∗ and B ∗ in terms of A.e. (15). This is the amount which.07. The fund pays interest for the ﬁrst three years at a nominal monthly rate d(12) of discount. c. pays nothing on the 4th and 5th anniversaries of the loan).06 on the face amount. has present value together with the amounts already paid of $100. If for all starting ages the probability of surviving 6 years according to the higherrisk table is equal to the probability of surviving 9 years according to the standard table.
3 is ﬁxed and where m was determined from it by the requirement that the expectation of survival time was 75 years. 000 30year loan with halfyearly payments (m = 2) start in six months from the time of borrowing.3 is the lognormal(m.4. (b) 20 years. (19).3 if the median survival time is ﬁxed at 72 years. How large must a halfyearly payment be in order that the stream of payments starting immediately be equivalent (in present value terms) at 6% interest to a lumpsum payment of $5000. Now answer the same question as in (17) about the amount of the ﬁnal lumpsum payment required.04. 2. (20).4 Worked Examples Example 1. or (c) forever ? . only in continuing accrued interest on the amounts of the missed payments. and ﬁnd the force of mortality for this lognormal at 65 years. Suppose that the missed payments in (17) actually result in late fees of $200 each of which is added to the balance at the time(s) of missed payments. (b) Answer the same questions if the city knows it can earn 6% on the money it deposits into its sinking fund. σ 2) hazard intensity where σ = 1. Now ﬁnd the value m associated with σ = 1. and ﬁnd the reserve. (a) Find the amount of the level payment the city must make into the sinking fund if the interest it earns on that fund is 5%.2. from its tax revenue. if the paymentstream is to last (a) 10 years. the missed payments did not result in any additional fees or charges. in the sinking fund after 6 years. Suppose that the borrower of a $100. in order to pay oﬀ the loan completely ? (18). In Problem (17). WORKED EXAMPLES 65 (17). A small city issues a bond for twenty million dollars for ten years at 5% nominal halfyearly interest (m = 2) and creates a sinking fund into which it will make twiceyearly deposits (m = 2). in addition to his ﬁnal payment. What lumpsum payment did the borrower have to make at the end of 30 years. One of the curves plotted in the ﬁrst part of Figure 2. or accumulated balance. and with nominal interest rate i(2) = . has made all payments except for two that he skipped. the 23’rd and 56’th payments.
. . is $143.57/(1 − 1. Observe ﬁrst of all that the speciﬁed paymentstream is exactly the same as a stream of payments of 1/m at times 0. .62. Since this paymentstream starting at 0 (m/2) is exactly onehalf that of the stream whose present value is ¨ ∞ . Thus the present value has the second expression (m) (m/2) a∞ − (1/2) a∞ ¨ ¨ Equating the two expressions allows us to conclude that (1/2) a∞ ¨ (m/2) (m) = a∞ ¨ (1 + v 1/m) Substituting this into the ﬁrst of the displayed presentvalue expressions. respectively with n = 10 and 20. This example illustrates the general methods enunciated at the beginning of Section 2. forever.11. a ﬁrst a present value expression is ¨ v 1/m (1/2) a∞ (m/2) A second way of looking at the paymentstream at odd multiples of 1/m is as the perpetuitydue payment stream ( 1/m at times k/m for all k ≥ 0) minus the paymentstream discussed above of amounts 1/m at times 2k/m for all nonnegative integers k. $208. 4/m. Assume m is divisible by 2. 5/m. INTEREST & FORCE OF MORTALITY If the payment size is P . .06−n ) = 143.57. Example 2.02871. 3/m.06−n )/d(2) a √ = 2(1 − 1/ 1. shows that that the present value requested in the Example is 1 d(m) · 1 v 1/m 1 = (m) = (m) −1/m 1/m 1 +v d (v + 1) d (2 + i(m)/m) and this answer is valid whether or not m is even. deferred by a time 1/m.1. . .06−n ) So the answer to part (c).02871)/(1 − 1. For parts (a) and (b). . . then the balance equation is 5000 = 2 P · ¨ n = 2 P (1 − 1. and using the simple expression 1/d(m) for the present value of the perpetuitydue. the answers are $325.06) = 2 · 0. in which n = ∞. the result is (2) Since d(2) P = (5000 · 0.66 CHAPTER 2. and use either one to give a simple formula. Express in two diﬀerent ways the present value of the perpetuity of payments 1/m at times 1/m. 2/m.
06.2. 2% for 2 years.. the present values are calculated as follows: For 4year 4% loan: $9645.08.4.02 For 2year 2% loan: $9642.g. 3% for 3 years. Therefore. Next. and give numerical answers for present values calculated at 6% and 8%. The monthly payments for an nyear loan at interestrate i is 10000/ (12) (12 an ) = (10000/12) d(12) /(1 − (1 + i)−n ). using interestrate r = 0. 000.73 For 2year 2% loan: $9475.72 For 3year 3% loan: $9349. or a cash discount of $500 ? Show how the answer depends upon the interest rate with respect to which you calculate present values. the present value ¨ at interestrate r of the nyear monthly paymentstream is 10000 · 1 − (1 + i)−1/12 1 − (1 + r)−n · 1 − (1 + r)−1/12 1 − (1 + i)−n Using interestrate r = 0. (The cash discount is now the least attractive option.) Example 4. Assume that all loans have monthly payments paid at the beginning of the month (e.77 For 3year 3% loan: $9599. Would you rather have an interest rate of 4% for 4 years.530 − y) Then ﬁnd analytical expressions for the survival probabilities S(y) for exact . Suppose that you are negotiating a carloan of $10. the present values of the various options are: For 4year 4% loan: $9314. Suppose that the force of mortality µ(y) is speciﬁed for exact ages y ranging from 5 to 55 as µ(y) = 10−4 · (20 − 0.68 so that the most attractive option in this case is the 4year loan. the 4 year loan has 48 monthly payments paid at time 0 and at the ends of 47 succeeding months). WORKED EXAMPLES 67 Example 3.89 so that the most attractive option is the cash discount (which would make the present value of the debt owed to be $9500).
002 + 2. . . 54.9372 exp − .5 + x) and second. in the case x = 5.97 e−0. . INTEREST & FORCE OF MORTALITY ages y in the same range.97.5(30 − z)) dz = 0. 29. in the case x = 30.93722. .68 CHAPTER 2.5 · 10−5 (y − 30)2 The deathrates qx therefore have two diﬀerent analytical forms: ﬁrst.25(y 2 −25)) so that S(30) = 0. . .034375 = 0. qx = S(x + 1)/S(x) = exp − 5 · 10−5 (10. qx = exp − . . . First for 5 ≤ y ≤ 30. as follows.97 exp −10−4 (5(y−5)+0. and for 30 ≤ y ≤ 55 y S(y) = S(30) exp − 10−4 30 (20 + 0. The key formulas connecting force of mortality and survival function are here applied separately on the ageintervals [5. . y S(y) = S(5) exp(− 5 µ(z) dz) = 0.002(y − 30) + 2. . .5 · 10−5 (2(x − 30) + 1) . . 30] and [30. 55]. and for the (oneyear) deathrates qx for integer ages x = 5. 54. assuming that S(5) = 0.
a ¨∞ (m) = = a∞ ¨ an ¨ − n vn p. 34 1 − vn i(m) 1 − vn d(m) pp. 35 1 − vn δ p. USEFUL FORMULAS FROM CHAPTER 2 69 2.5.2. 35–35 (m) an (m) = . 36 a (I (m) ¨)n (m) (m) (m) an ¨ (∞) = an (∞) = an = a∞ (m) = . an ¨ (m) = an = v 1/m ¨ n a (m) p.5 Useful Formulas from Chapter 2 v = 1/(1 + i) p. 35 1 i(m) 1 d(m) p. 38 Loan Amt m ¨n a (m) nyr m’thly Mortgage Paymt : . 38 ¨ (D(m) a)n (m) = (n + 1 (m) (m) ) an − (I (m)a)n ¨ ¨ m p.
INTEREST & FORCE OF MORTALITY p. ω f (t) = 1 . ω 0≤t≤ω p. Failure Dist. t>0 p. 39 k + : m 1 − v n−k/m 1 − vn p. px = 1px = 1 − qx p. 48 t qy = 1 − tpy p. 52 Expon. 42 S(y + t) = exp − S(y) t nyr Mortgage Bal.k/m = t py = µ(y + s) ds 0 p. 52 Unif. amt Bn. f (t) = µe−µt . µ(t) = µ . Dist.: S(t) = e−µt . 52 .: S(t) = ω−t .70 CHAPTER 2. 48 qx = 1 qx = dx lx . 49 y S(y) = exp( − 0 µ(t) dt) p. 48 µ(y + t) = ∂ f (y + t) = − ln S(y + t) S(y + t) ∂t p.
Dist. 54 .2. t>0 p. USEFUL FORMULAS FROM CHAPTER 2 71 Weibull. t≥0 Gompertz: µ(t) = Bct .: S(t) = e−λt . t≥0 S(t) = exp −At − B t (c − 1) ln c p. 52 Makeham: µ(t) = A + Bct .5. γ µ(t) = λγtγ−1 .
72 CHAPTER 2. INTEREST & FORCE OF MORTALITY .
Finally. This predictability will be used in later chapters to justify consideration of expected present values of contractual payouts to describe an insurer’s liability. 73 . with survival probability px for a life aged x.Chapter 3 More Probability Theory for Life Tables This Chapter introduces several key ideas in Probability Theory which are essential for an understanding of the book’s core actuarial topics in Chapter 4 and 5. The ﬁrst of these ideas is that survival from one year to the next can be regarded for each member of a population as a cointoss experiment. so we prepare the ground by presenting background theory and rules of manipulation for expectations of discretevalued random variables. This point of view also provides a convenient vehicle for conducting computer simulations of population survival experience for large or small lifetable populations. approximating and calculating with probabilities and expectations using theoretical models of survival between successive years of age. independently of all other members of the population. we complete our probability background with further material on interpreting. Since the lifetable summarizes outcomes on a large number of cointoss experiments. we study next through limit theorems (law of large numbers and central limit theorem) the high degree of predictability of these outcomes at the population level.
. . N (1 + x) = k=0 N N k N x . . by applying the ﬁrst assertion with x = y/z and multiplying both sides by z N . . z). . which for each such life has probability t px . The second assertion follows immediately. y. aN ) ∈ {0. . . and since these monomial terms arise only from the combinations (a1. . .74 CHAPTER 3. 1. and adding all of the monomials together. the number lx+t who survive t timeunits after age x can be regarded as the number of successes or heads in a large number lx of independent cointoss trials corresponding to the further survival of each of the lx lives aged x . Consider the twovariable polynomial (y + z)N = (y + z) · (y + z) · · · (y + z) N factors expanded by making all of the diﬀerent choices of y or z from each of the N factors (y + z). aN ) of {y. . z} choices in which precisely k of the values aj are 1’s and the rest are 0’s. N . The one preliminary mathematical result that the student is assumed to know is the Binomial Theorem stating that (for positive integers N and arbitrary real numbers x. where ai = 1 would mean that y is chosen and ai = 0 would mean that z is chosen in the ith factor. k (y + z) = k=0 N N k y k z N −k Recall that the ﬁrst of these assertions follows by equating the k th deriviatives of both sides at x = 0. . . . for each k = 0. Now this combinatorial fact is immediately deduced from the Binomial Theorem: since the coeﬃcient N is the total number of monomial terms y k z N −k k which are collected when (y + z)N is expanded as described. . Each combined choice of y or z from the N factors (y + z) can be represented as a sequence (a1. N . This Theorem also has a direct combinatorial consequence. 1}N . PROBABILITY & LIFE TABLES 3. . . . in the nontrivial case when z = 0. . multiplying each combination of choices out to get a monomial y j z N −j . an ) ∈ {0. The motivation is that in large lifetable populations. where k = 0. . The number of symbolsequences (a1 . 1}N such N is given. . by that j=1 aj = k .1 Binomial Variables and Limit Theorems This Section develops basic machinery for the theory of random variables which count numbers of successes in large numbers of independent biased cointosses.
designated ‘success’ — are called Bernoulli (p) trials. spoken k as ‘N choose k’. . aN ) ∈ {0.3. or sample space for this experiment. we can regard the number lx+t of lives surviving to the (possibly fractional) age x+t as a Binomial random variable with parameters N = lx . if derived from an actual cohort . Since the mechanisms which cause those lives to survive or die can ordinarily be assumed to be acting independently in a probabilistic sense. . This number N .e. the rule by which probabilities are assigned to sets or events A of more than one string a ∈ {0. the probability k which is necessarily assigned to the event of k successes is Pr( k successes in N Bernoulli(p) trials ) = P (X = k) = N k pk (1−p)N −k By virtue of this result. B). 1}N being assigned probability pa (1 − p)N −a . We are particularly interested in the event (denoted [X = k]) that precisely k of the cointosses are heads. there are N such strings. i. we now begin to regard the ideal probabilities S(x + t)/S(x) as true but unobservable probabilities t px = p with which each of the lx lives aged x will survive to age x + t . 1}N is to add the probabilities of all individual strings a ∈ A. k With the notion of Bernoulli trials and the binomial distribution in hand. and since. consists of the strings of N zeroes and ones. is said to have the Binomial distribution with probability mass function pX (k) = N pk (1 − p)N −k . From this point of view. 1}N consisting of all N strings a such that j=1 aj = k. N k 75 The random experiment of interest in this Section consists of a number N of independent tosses of a coin. the random variable X equal to the number of successes in N Bernoulli(p) trials. according to the discussion following probability p (1 − p) the Binomial Theorem above. Such cointossing experiments — independently replicated twooutcome experiments with probability p of one of the outcomes.. with probability p of coming up heads each time. where a ≡ N aj . in the subset [X = k] ⊂ {0. . therefore counts all of the ways of choosing k element subsets (the positions j from 1 to N where 1’s occur) out of N objects. with each string a = (a1. Since each such string has the same k N −k . The space of possible headsandtails conﬁgurations. p = t px . Because j=1 of the ﬁnite additivity axiom of probabilities (saying that Pr(A ∪ B) = Pr(A) + Pr(B) for disjoint events A.1. BINOMIAL VARIABLES & LAW OF LARGE NUMBERS = N (N − 1) · · · (N − k + 1)/ k! . .
We state and prove the result here only in the setting of binomial random variables. is established in the famous Law of Large Numbers. Theorem 3. Law of Large Numbers.1) . called consistency.9.3 how it implies a more general result for ﬁnitevalued discrete random variables. a Large Deviation Inequality which is important in its own right but more diﬃcult. PROBABILITY & LIFE TABLES dataset of size equal to the radix . not depending upon N . sketching in Section 3. > 0. the number N of Bernoulli trials can be chosen so large that Pr  X/N − p  ≥ δ ≤ Proof. 1}N for which j=1 aj = k with k − Np ≥ Nδ. . the ratio lx+t /lx is a statistical estimator of the unknown constant t px .76 CHAPTER 3. The good property. the observed ratios lx+t /lx should reliably be very close to the ‘true’ probability t px .1 (Cointoss Law of Large Numbers) Suppose that X is a Binomial (N. denoting the number of successes in N Bernoulli (p) trials. . For arbitrarily small ﬁxed δ. when l0 is large. common sense and experience suggest that. is stated and proved in the Appendix to this Chapter. A more precise quantitative inequality concerning binomial probabilities. Section 3. Since the event [ X/N − p ≥ δ ] = [ X − Np ≥ Nδ ] is the union of the disjoint events [X = k] for k − Np ≥ Nδ. In other words. of this estimator to be close with very large probability (based upon large lifetable size) to the headsprobability it estimates. p) random variable. . aN ) ∈ {0. the observed lifetable counts lx would be treated as random data which reﬂect but do not deﬁne the underlying probabilities x p0 = S(x) of survival to age x. which in turn consist N of all outcomestrings (a1. However. . the subset of the binomial probability mass function values pX (k) with k − Np ≥ Nδ are summed to provide Pr(X/N −p ≥ δ) = k: k−N p≥N δ Pr(X = k) = k: k−N p≥N δ N k pk (1−p)N −k This summation is termbyterm less than or equal to N k p (1−p) k N −k k: k−N p≥N δ (k − Np)2 ≤ (Nδ)2 N k=0 N k pk (1−p)N −k (k − Np)2 (Nδ)2 (3. and therefore the other lifecounts lx for moderate values x are also large.
it is easy to check via the equality (k − Np)2 = k(k − 1) − (2Np − 1)k + (Np)2 . To see why more .1 Probability Bounds & Approximations Theorem 3. that N (k − Np)2 k=0 N k pk (1 − p)N −k = N p(1 − p) Substituting this ﬁnal relation into (3. and simplifying algebraically.1. becomes with the aid of the Binomial Theorem N −1 N −1 l = Np p (1 − p)N −1−l = Np l l=0 and similarly (now with j = k − 2) N k=0 N k(k − 1) k N p (1 − p) k N −k = k=2 p2 N (N − 1) (N − 2)! k−2 p (1 − p)N −k (k − 2)!(N − k)! N −2 = N (N − 1)p 2 j=0 N −2 j p (1 − p)N −2−j = N (N − 1) p2 j Putting together the last calculations. A much more accurate upper bound in given in Theorem 3.1. 2 3.3. However.1 provides only a very crude upper bound to the probability with which X/N − p ≥ δ.2 of the Appendix to the Chapter (Sec. BINOMIAL VARIABLES & LAW OF LARGE NUMBERS 77 where we have made the second some larger by including more nonnegative terms in it.9). direct summation shows N k k=0 N k N pk (1 − p)N −k = k=1 kp N · (N − 1)! pk−1 (1 − p)N −k k(k − 1)!(N − k)! which after replacing k − 1 by l.1) now shows that Pr(X/N − p ≥ δ) ≤ p(1 − p) Np(1 − p) = 2 (Nδ) Nδ 2 The assertion of the Theorem now follows by taking N ≥ (p(1 − p)/( δ 2 ). 3.
30). The exact Binomial(1000. and δ = 0. the normal approximation (3. p. 70] ∪ [130. the reﬁned form of the DeMoivreLaplace Theorem given in the Feller (1957. Pr(X > b) 1−Φ b − Np Np(1 − p) converges to 1 if the ‘deviation’ ratios (b − Np)/ Np(1 − p) and (a − Np)/ Np(1 − p) are of smaller order than N −1/6 when N gets large. more than ﬁfty times larger ! On the other hand. N] = [0. with N = 1000. 1957. 172) reference says that each of the ratios of probabilities Pr(X < a) Φ a − Np Np(1 − p) . pp. is 0. Much closer approximations to the exact probabilities for Binomial(N.03)2 ) = 0. as n → ∞. 1000] is 0. To give a feeling for the probabilities with which observed lifetable ratios reﬂect the true underlying survivalrates. p) probability of X/N − p ≤ δ was 1 − .2) where Φ is the standard normal distribution function given explicitly in integral form in formula (3. p) random variables to fall in intervals around Np are obtained from the Normal distribution approximation Pr(a ≤ X ≤ b) ≈ Φ b − Np Np(1 − p) −Φ a − Np Np(1 − p) (3.99808. PROBABILITY & LIFE TABLES accurate bounds are needed. as in formula (3.3) is 0. 0.99843. In the example discussed above.78 CHAPTER 3. p = 0.3) for the true binomial probability Pr(X/N − p ≤ δ).2) converges to 0 when p remains ﬁxed.1. consider the case where N = 1000.1) probability of (number of successes in N Bernoulli(p) trials falling in) [0. where the exact Binomial(N. 1.1 .1.2.001916.03.1. the upper bound provided by the inequalities of Theorem 3. p = 0.1)(.29) below. This result suggests the approximation Normal approx. = Φ Nδ Np(1 − p) − Φ −Nδ Np(1 − p) (3. N(p − δ)] ∪ [N (p + δ). 16873).9)/(1000(. δ = 0. Moreover. while the upper bound established in the proof of Theorem 3.0198.00192 = .and righthand sides of (3. This approximation is the DeMoivreLaplace Central Limit Theorem (Feller vol.03.1 is (. we have collected in Table 3. which says precisely that the diﬀerence between the left.
9973 . within factor 1 ± . are still noticeably smaller than the actual values.30) of Section 3.9120 .003 .9995 . the normal approximations in the ﬁnal column of Table 3.2 and from (3.9973 . . BINOMIAL VARIABLES & LAW OF LARGE NUMBERS 79 Table 3. k. lx.75 5 0. k px which might realistically arise in an insurancecompany lifetable.008 .31).94 10 0. while also often close to 1.020 . 5.1 .9188 Normal approx. 6) with which various Binomial(lx.050 .1. but also show that the lower bounds.1: Probabilities (in col.9969 .9778 . Cohort n = lx 10000 10000 10000 1000 10000 1000 10000 1000 Age x 40 40 40 40 70 70 70 70 Time Prob. . Columns 6 and 7 in the Table show how likely the lifetable ratios are to be close to the ‘theoretical’ values.9972 . In Table 3.020 . k p = k px 3 0. not smaller as they should be for applicability of (3.9952 .9877 .9985 .3)) probabilities with which the ratios lx+k /lx agree with k px to within a fraction of the latter.9886 various exact binomial probabilities and their counterparts from the approximation of (3.030 . k px ) random variables lie within a factor 1 ± of their expectations.9.1 are given various combinations of x.9531 .004 .3) to the exact probabilities in column 6.3). or 10 years.9949 .50 Toler.9950 .9938 . The probability experiment determining the size of the surviving cohort lx+k is modelled as the tossing of lx independent coins with common headsprobability k px : then the surviving cohortsize lx+k is viewed as the Binomial(lx .080 Pr. The ﬁnal column contains the normal approximations based on (3.75 10 0.9863 .3) and the inequality (3.99 5 0.98 10 0. k px ) random variable equal to the number of heads in those cointosses. with the true and estimated (from Theorem 3.9760 .94 5 0. Although the deviationratios in estimating lifetable probabilities are often close to or larger than N −1/6.9886 Lower bound . together with ‘theoretical’ probabilities k px with which these lives will survive for a period of k = 1.30)(3.50 10 0. together with lower bounds (in Col.9866 .9600 .9938 . The illustration concerns cohorts of lives aged x of various sizes lx.9995 . 7) for these probabilities derived from the largedeviation inequalities (3.9985 . together.3.
We said that if the lifetable was representative of a larger population of prospective insureds. Finally.80 CHAPTER 3. The Law of Large Numbers applies equally when the agex survivors have been sampled by some more complicated method than simply watching a cohort from birth. but they are not very random. where the sizes lx of lives under observation at age x are large but the probabilities px are unknown.28. At the next stage of sophistication. and similarly expectations of functions of lifetable deathtimes were averages over the entire cohort. governing each member of the lifetable cohort and of further prospective insureds. . 553). section XVI. A still more reﬁned theorem which justiﬁes this is given by Feller (1972. With the mathematical justiﬁcation of the Law of Large Numbers. in the realistic datacollection scenarios discussed in Appendix A. Thus. if the size lx of the cohort of surviving lives aged x is large.2 Simulation of Discrete Lifetimes We began by regarding lifetable ratios lx /l0 in large cohort lifetables as deﬁning integerage survival probabilities S(x) = x p0 . p. then we could imagine a newly presented life aged x as being randomly chosen from the lifetable cohort itself. the lifetable can also be viewed as in Appendix A as an idealized set of data. 3. with each ratio lx+t /lx equal to the relative frequency of success among a set of lx imagined Bernoulli (t px ) trials which Nature performs upon the cohort of lives aged x .7 leading up to formula 7. That is. we come full circle: these relative frequencies are random variables. We motivated the conditional probability ratios in this way. PROBABILITY & LIFE TABLES below are sensationally close to the correct binomial probabilities in column 6. Although we found the calculusbased formulas for lifetable conditional probabilities and expectations to be useful. the later fractions lx+t /lx of survivors at x + t to those at x are extremely likely to lie within a very small tolerance of t px . we began to describe the (conditional) probabilities t px ≡ S(x + t)/S(x) based upon a smooth survival function S(x) as a true but unknown survival distribution. hypothesized to be of one of a number of possible theoretical forms. at that stage they were only ideal approximations of the more detailed but still exact lifetable ratios and sums. the lifetable ratios lx+1 /lx are highly accurate statistical estimators of the lifetable probabilities .
we treated the probabilities px = lx+1 /lx for x = 0.3. The Law of Large Numbers guarantees good agreement. . with counts lx given for integer ages x from 1 through 80. Jordan’s (1967) book on Life Contingencies. 79 as the correct oneyear survival probabilities for a second. For ∗ example. l79 was to make the variable lx+1 depend ∗ ∗ ∗ on previously generated l1 .S. . . The implication of Table 3. px ).2.1.2: Illustrative Real and Simulated LifeTable Data Age x 9 19 29 39 49 59 69 79 lx in 195961 LifeTable 96801 96051 94542 92705 88178 77083 56384 28814 ∗ Simulated lx 96753 95989 94428 92576 87901 76793 56186 28657 81 To make this discussion more concrete. Using simulated random variables generated in R. the ﬁnal simulated count of l79 = 28657 lives . as x runs from 1 to 79. . . . . . and then to generate ∗ ∗ lx+1 as though it counted the heads in lx independent cointosses with headsprobability px . we illustrate the diﬀerence between the entries in a lifetable and the entries one would observe as data in a randomly generated lifetable of the same size using the initial lifetable ratios as exact survival probabilities. . SIMULATION OF DISCRETE LIFETIMES Table 3. with x = 69. . between the ratios lx+10/lx (which here play ∗ the role of the probability 10 px of success in lx Bernoulli trials) and the ∗ ∗ corresponding simulated random relative frequencies of success lx+10/lx . W. with very high probability. . we successively generated. The complete simulated lifetable was given earlier as Table 1. That is. the agreement between the initial and randomly generated lifetable counts is extremely good. In other words. A comparison of the actual and simulated lifetable counts for ages 9 to 79 in 10year intervals. lx only through lx. White Males 195961 reproduced as Table 2 on page 11 of C.2 is unsurprising: with radix as high as 105 . . computersimulated ∗ cohort lifetable with radix l0 = 105 . the mechanism of sim∗ ∗ ∗ ulation of the sequence l0 . is given below. random ∗ ∗ variables lx+1 ∼ Binomial (lx. using this Table with radix l0 = 105 . We used as a source of lifetable counts the Mortality Table for U.
pvec[j] = lx(j+1)/lxj .77. .80.70.8 by 300 or more in either direction with probability approximately 0. is: LifTab(lvec)[2:9].82 CHAPTER 3. If lvec denotes a vector of values (l0 . x(1). x(2). . . . lx(K) : LifTabSim = function(lvec) { K = length(lvec)1 lstar = c(lvec[1].77. . 70. With this successprobability.lstar[j]. (The exact binomial probabilty of the same event is 0.0113. 0. . byrow=T. . 0.lvec[j+1]/lvec[j]) lstar } The syntax to generate a vector like the third column of Table 3.65. lx(2).rep(0. lx(1). . 0. .70.2 from the second. . .85.51103 · 56186 = 28712. LifTabSim(Lvec)[2:6]).85.0115. LifTabSim(Lvec)[2:6]. then the statements K = length(lvec)1 . . 80. The R code used to generate Table 3.K)) for (j in 1:K) lstar[j+1] = rbinom(1. PROBABILITY & LIFE TABLES ∗ aged 79 is the successcount in l69 = 56186 Bernoulli trials with successprobability 28814/56384 = . 0.65. 0.). ncol=5. pvec = lvec[2:(K+1)]/lvec[1:K] create the vector of hypothetical survival probabilities.3) says that the simulated count l79 will diﬀer from .4) Lvec = 1000*cumprod(c(1. then we can simulate twice. where the integer ages 0. lx(K)) of numbers of surviving lives in a cohort lifetable with radix l0 . As a further example of such a simulation. 0. independently and output in R the numbers of surviving lives at these ages. x(K) are not necessarily evenly spaced. the nor∗ mal approximation (3. 60. K − 1.51103. . 0. . j = 0. . as follows: pvec = c(0.10). and here is a small function to generate the (K + 1)vector ∗ tt lstar consisting of l0 ≡ l0 together with the output simulated values ∗ lx(1).4 for x = 40.pvec)) matrix(c(seq(40. nrow=4. . pvec. suppose that 1000 individuals aged 40 have successive probabilities 10px = 0. where lvec consists of the radix l0 = 105 concatenated with column 2.2 is very simple. . 50. 0.
"10_p_x". (a).77 0. .00 450.00 80.0 344. What is the spread between the smallest and largest number surviving at each age across your 10 simulations ? (b).3. if the underlying lifetable probabilities were unknown.00 50. we can see that the variability in the sim∗ ulated numbers lx is considerable for l0 of 1000 or less. what is your best estimate of the probability 20p40 ? .7 0. 80.00 662.65 0.00 60.4] [.0 83 From small experiments like this. 50.00 483. Regarding your 10 sets of simulated numbers of survivors as independent datasets.0 10_p_x 0.0 296.0 Sim#2 l_x 854. what would be your best estimate of the probability 20 p40 ? (c).3] [.0 70.4 Sim#1 l_x 842.2] [.5] Ages 40. With the same probabilities 10px use R to simulate 10 times independently the numbers of survivors at ages 40. SIMULATION OF DISCRETE LIFETIMES dimnames=list(c("Ages".85 0. Combining the 10 simulated datasets you generated in (b).00 638. Exercise 3."Sim#1 l_x". .00 129.A.2. "Sim#2 l_x"). .1] [.00 133. . NULL)) [.
and let c(·) be a realvalued (nonrandom) cost function such that c(Z) represents an economically meaningful cost incurred when the random variable value Z is given. think for example of Z as the unforeseeable future damage or liability upon the basis of which an insurer has to pay some scheduled claim amount c(Z) to fulﬁll a speciﬁc property or liability insurance policy. i. It follows that the average costs c(Zj ) over the N . P (Zj = zi  Z1 = b1. . . in which the j th policyholder is said to result in a ‘success’ if he sustains a damage amount equal to z . zm }. and to result in a ‘failure’ otherwise. c(ZN ). . Returning to a general discussion. . PROBABILITY & LIFE TABLES 3. . These probabilities pZ (z) must be positive numbers which summed over all possible values z add to 1. zm . . . . . . The Law of Large Numbers (Theorem 3. so that for all j. . In a large population of N independent policyholders. . and b1.84 CHAPTER 3. . . . applied as above. . . bj−1 . N and costs c(Z1 ). . . Here independent means that the mechanism causing diﬀerent individual Zj values is such that information about the values Z1 .7) for these Bernoulli trials says that the number out of these N policyholders who do sustain damage z is for large N extremely likely to diﬀer by no more than δN from N pZ (z). each governed by the same probabilities pZ (·) of liability occurrences. . j = 1. .3 Expectation of Discrete Random Variables The Binomial random variables discussed in this Chapter are examples of socalled discrete random variables. with a corresponding list of probabilities or probability mass function values pZ (z) with which each of those possible outcomes occur. . . Zj−1 does not change the (conditional) probabilities with which Zj takes on its values. Suppose that a large number N of independent individuals give rise to respective values Zj . . The Law of Large Numbers says that we can have a frequentist operational interpretation of each of the probabilities pZ (z) with which a claim of size c(z) is presented. that is. . for each ﬁxed damageamount z we can imagine a series of N Bernoulli (pZ (z)) trials. In an insurance context. suppose that Z is a discrete random variable with a ﬁnite list of possible values z1 . where k ranges over {z1. random variables Z with a discrete (usually ﬁnite) list of possible outcomes z. says that out of the large number N of individuals it is extremely likely that approximately pZ (k) · N will have their Z variable values equal to k. . Zj−1 = bj−1 ) = pZ (zi) Then the Law of Large Numbers. . .
.3).3. Since P ([T ] = k) = S(k) − S(k + 1). .4) yields ω−1 E(g(T )) = E(g([T ]) = k=0 g(k) (S(k) − S(k + 1)) k+1 k (3. This expectation was interpreted earlier as the average cost over all members of the speciﬁed lifetable cohort.4) to the . Now the expectation can be veriﬁed to coincide with the lifetable average previously given. evaluating the discrete conditional expectation given T ≥ x means applying the formula (3. N : Zj = zi } — is approximately given by m m N −1 i=1 c(zi ) · (N pZ (zi )) = i=1 c(zi ) pZ (zi ) In other words. becomes ω−1 k+1 ω−1 k+1 f (t) dt and [t] = k for ω g(k) k=0 k f (t) dt = k=0 k g([t]) f (t) dt = 0 g([t]) f (t) dt agreeing precisely with formula (1.3. the Law of Large Numbers implies that the average cost per trial among the N independent trials resulting in random variable values Zj and corresponding costs c(Zj ) has a welldeﬁned approximate (actually. Similarly. . the general expectation formula (3. a limiting) value for very large N m Expectation of cost = E(c(Z)) = i=1 c(zi) pZ (zi ) (3. consider the expected value of a costfunction g(T ) of a lifetime random variable which is assumed to depend on T only through the function g([T ]) of the integer part of T . after replacing S(k) − S(k + 1) = k ≤ t < k + 1.5) which. .4) As an application of the formula for expectation of a discrete random variable. EXPECTATION OF DISCRETE RANDOM VARIABLES independent individuals — which can be expressed exactly as N m 85 N −1 j=1 c(Zj ) = N −1 i=1 c(zi ) · #{j = 1. if the probabilities S(j) in the following expression are replaced by the lifetable estimators lj /l0 .
the expressions require knowledge only of the probabilities S(y) of survival for wholeyear or integer ages y.) Then the conditional expectation is ω−1 E(g([T ])  T ≥ x) = k=x S(k) − S(k + 1) g(k) = S(x) ω−1 k+1 k ω−1 k=x g(k) S(x) ω k+1 f (t) dt k or E(g([T ]  T ≥ x) = k=x g([t]) f (t) dt = S(x) g([t]) 0 f (t) ) dt S(x agreeing precisely with formula (1. for E(g([T ])  T ≥ x). and express S(k) − S(k + 1) S(x + j) S(x + j + 1 = 1− S(x) S(x) S(x + j) Then ω−1 = j px (1 − px+j ) E(g([T ])  T ≥ x) = k=x S(k) − S(k + 1) g(k) = S(x) ω−x−1 j px j=0 (1−px+j ) g(x+j) (3. The preceding discussion shows that expectations or conditional expectations of functions of wholeyear ages can equivalently be calculated using the expectation formulas for discrete or continuous random variables. In the discrete case. There is an analogy in the continuousvariable case. with approximate probabilityweights fZ (z)dz. PROBABILITY & LIFE TABLES function c(Z) = g(Z) of the discrete random variable Z = [T ] using the conditional probability mass function P (Z = k) = P ([T ] = k  T ≥ x) = (S(k) − S(k + 1))/S(x) for all integers k ≥ x (and with probability 0 assigned to all integers k < x. .5). in the preceding (discreteversion) formula. Indeed. namely c(z) f (z) dz. In this case. where fZ (z) is a nonnegative density function integrating to 1. z + dz] are given by fZ (z)dz. let k ≥ x be replaced by k = x + j. in this discussion) which can be incurred in each trial.86 CHAPTER 3. however. the weighted average of costfunction values c(z) which arise when Z ∈ [z. weighted by the probabilities with which they occur. where Z would be a random variable whose approximate probabilities of falling in tiny intervals [z. is written as a limit of sums or an integral. z + dz].6) Just as we did in the context of expectations of functions of the lifetable waitingtime random variable T . we can interpret the Expectation as a weighted average of values (costs.
. then for each δ > 0.8) We do not give any further proof here. . but the motivating arguments given. . together with straightforward manipulations using the result of Theorem 3. First. . .7) where. .1 Rules for Manipulating Expectations We have separately deﬁned expectation for continuous and discrete random variables. . we treated the expectation of a speciﬁed function g(T ) of a lifetime random variable governed by the survival function S(x) of a cohort lifetable. The discrete case was handled more conventionally. EXPECTATION OF DISCRETE RANDOM VARIABLES 87 3.3. P (Zk ≤ r  Z1 = z1 . In the continuous case. where δ > 0 is an arbitrarily small number not depending upon n. . . are independent random variables. . . . . . has a value which with very high probability diﬀers from n·p by an amount smaller than δn. E(c(Z1 )) = z∈S c(z) P (Z1 = z) (3. Zk−1 = zk−1 ) = P (Z1 ≤ r) regardless of the precise values z1 . . . in the sense that for all k ≥ 2 and all numbers r. p) random variables justiﬁed us in saying that the sum X = Xn of a large number n of independent cointoss variables 1 . The Expectation p of each of the variables i is recovered approximately as the numerical average X/n = n−1 n i of independent trials. . in terms of the ﬁnite set S of possible values of Z . n .3. as the approximate numerical average of the values g(Ti) over all individuals i with data represented through observed lifetime Ti in the lifetable. each of which is 1 with probability p and 0 otherwise. zk−1 . along the lines of a ‘frequentist’ approach to the mathematical theory of probability. as n gets large n P n−1 i=1 c(Zi ) − E(c(Z1 )) ≥ δ −→ 0 (3. we observed that our calculations with Binomial(n.3. saying that if Z1 .7. Z2 . i=1 i of the independent outcomes This Law of Large Numbers extends to arbitrary sequences of independent and identical ﬁnitevalued discrete random variables.
we are interested in evaluating expectations of various functions of random variables related to the contingencies and uncertain duration of life. The following rules for the manipulation of expectations arising in such superpositions considerably simplify the calculations. . For the rest of this book. Assume in what follows that all random payments and times are functions of a single lifetime random variable T .8) above whenever the function c(z) is such that c(z) P (Z1 = z) < ∞ z∈S or the independent random variables Zi are continuous. It is also a fact that the Law of Large Numbers given in equation (3. as long as either Zi are discrete with inﬁnitely many possible values deﬁning a set S. and more generally in applications of probability within actuarial science. all with r the same density f (t) such that P (q ≤ Z1 ≤ r) = q f (t) dt. PROBABILITY & LIFE TABLES are an essentially complete proof of (3. and the expectation is as given in equation (3.7). whether discrete or continuous.88 CHAPTER 3.9) whenever the function c(t) is such that ∞ c(t) f (t) dt < ∞ −∞ All of this shows that there really is no choice in devising an appropriate deﬁnition of expectations of costfunctions deﬁned in terms of random variables Z. Many of these expectations concern superpositions of random amounts to be paid out after random durations. and expectation is deﬁned by ∞ E(c(Z1 )) = −∞ c(t) f (t) dt (3.7) continues to hold if the deﬁnition of independent sequences of random variables Zi is suitably generalized.
occurs only if a ≤ T < b and in that case consists of a payment of a ﬁxed amount F occurring at a ﬁxed time h. becomes by rule (1) above. c2 (T ) (which may occur at diﬀerent times.g. then the expected present value under a ﬁxed nonrandom interestrate i with v = (1 + i)−1 . Since .. then the expectation of the payment is the product of F and the expectation of c(T ): Discrete case: E(F c(T )) = t F c(t) P (T = t) = F t c(t) P (T = t) = F · E(c(T )) F c(t)f (t) dt = F c(t)f (t) dt = F ·E(c(T )) Continuous case: E(F c(T )) = (2). If a payment consists of a nonrandom multiple (e. faceamount F ) times a random amount c(T ). then the overall payment has expectation which is the sum of the expectations of the separate payments: Discrete case: E(c1 (T ) + c2 (T )) = t (c1 (t) + c2 (t)) P (T = t) = t c1(t) P (T = t) + t c2(t) P (T = t) = E(c1(T )) + E(c2(T )) Continuous case: E(c1 (T ) + c2 (T )) = = c1 (t) f (t) dt + (c1 (t) + c2 (t)) f (t) dt c2(t) f (t) dt = E(c1 (T )) + E(c2 (T )) Thus. If a payment consists of the sum of two separate random payments c1(T ). EXPECTATION OF DISCRETE RANDOM VARIABLES 89 (1). if an uncertain payment under an insurancerelated contract. taken into account by treating both terms ck (T ) as present values as of the same time).3. E(v h F I[a≤T <b]) = v h F E(I[a≤T <b]) where the indicatornotation I[a≤T <b] denotes a random quantity which is 1 when the condition [a ≤ T < b] is satisﬁed and is 0 otherwise.3. based upon a continuous lifetime variable T with density fT .
where x is an integer age. is the wholeyear residual life [T ]−x for a life aged x. is called curtate mean residual life or curtate life expectancy. we conclude that E(I[a≤T <b]) = P (a ≤ T < b b) = a fT (t) dt.1 Let Z be a nonnegativeintegervalued random variable.2 Curtate Expectation of Life One example of a function of the number [T ] of whole years of life.11) .3. 1} like the cointoss variables i above. The expectation.90 CHAPTER 3. PROBABILITY & LIFE TABLES an indicator random variable has the two possible outcomes {0. Lemma 3.10) The Lemma is proved using the rule (FubiniTonelli theorem for double summation) that the order of a double summation of nonnegative summands can always be reversed: ∞ ∞ k ∞ ∞ ∞ EZ = k=0 k pZ (k) = k=1 j=1 pX (k) = j=1 k=j pZ (k) = j=1 P (Z ≥ j) 3. ω−1 ex = E( [T ] − x  T ≥ x ) = t=x P (t ≤ T < t + 1) (t − x) P (T ≥ x) (3. whose conditional expectation is useful and interpretable. and the expected present value above is b E(v h F I[a≤T <b]) = v h F a fT (t) dt (3). The expectation of a nonnegativeintegervalued random variable can sometimes be simpliﬁed considerably by means of the following useful Lemma. Then ∞ EZ = j=1 P (Z ≥ j) (3. necessarily conditional on the attained age x.
.g. INTERPRETING FORCE OF MORTALITY 91 Substituting the formula (3.10) of Lemma 3. . nh. .12) A third useful version of this formula can be found by applying formula (3. .6) with g(t) = t − x gives this formula in the alternative form ω−x−1 ex = E( [T ] − x  T ≥ x ) = j=0 j px (1 − px+j ) j (3.13) The extension of these expectation formulas to give mean residual lifetimes which are not truncated to whole years rests on survival function and density formulas which specify mortality rates between birthdays. as large as 4 or 12). 3h.4 Interpreting Force of Mortality This Section consists of remarks. (Here [·] continues to denote the greatest integer less than or equal to its real argument. the discrete random variable [T m]/m gives a close approximation to T and represents the attained age at death measured in wholenumber multiples of fractions h = one mth of a year.) Since surviving an additional time t = nh can be viewed as successively surviving to reach times h. for m large (e. relating the force of mortality for a continuously distributed lifetime random variable T (with continuous density function f ) to conditional probabilities for discrete random variables. The following two sections are devoted to a deeper study of continuous mortality models and interpolation approximations.1 to the nonnegative integer valued random variable Z = [T ] − x with probability masses calculated conditionally given T ≥ x.4. 3. Indeed.3. 2h. This yields ω−x−1 ω−x−1 ex = E( [T ] − x  T ≥ x ) = j=1 P ([T ] − x ≥ j  T ≥ x) = j=1 j px (3. and since (by the deﬁnition of conditional probability) P (A1 ∩ · · · ∩ An ) = P (A1 ) · P (A2A1) · · · P (An A1 ∩ · · · ∩ An−1 ) we have (with the interpretation Ak = {T ≥ x + kh} ) nh px = h px · h px+h · h px+2h · · · h px+(n−1)h .
x+h).92 CHAPTER 3. if f is continuously diﬀerentiable. Therefore. and therefore the random survival time can be viewed as the (ﬁrst failure among a) succession of results of a sequence of independent coinﬂips with successive probabilities h pkh of heads. yielding an overall conditional survival probability equal to the negative exponential of accumulated hazard over [x. By the Mean Value Theorem applied up to seconddegree terms on the function S(x + h) expanded about h = 0. the previously derived formula x+h h px = exp − x µ(y) dy can be interpreted by considering the fraction of individuals observed to reach age x who thereafter experience hazard of mortality µ(y) dy on successive inﬁnitesimal intervals [y. 3. h px =1 − h· S(x) − S(x + h) hS(x) = 1 − h µ(x) + h f (x + τ h) 2 S(x) Going in the other direction. Yet many expectations of . h2 h2 S(x+h) = S(x) + hS (x) + S (x+τ h) = S(x) − hf (x) − f (x+τ h) 2 2 for some 0 < τ < 1. PROBABILITY & LIFE TABLES The form in which this formula is most often useful is the case h = 1: for integers k ≥ 2.5 Interpolation Between Integer Ages Cohort lifetable data lx and the probability quantities j px derived from them (for integers j) depend on and are determined by the survival function S(k) values only at integer arguments k. using the deﬁnition of µ(x) as f (x)/S(x) given on page 49. x + h). y+dy] within [x.14) k px = px · px+1 · px+2 · · · px+k−1 Every continuous waitingtime random variable can be approximated by a discrete random variable with possible values which are multiples of a ﬁxed small unit h of time. The lives aged x survive to age x + h with probability equal to a limiting product of inﬁnitesimal terms (1−µ(y) dy) ∼ exp(−µ(y) dy). (3.
1). 1) in the argument of S in the third term (the meanvalue type remainder) of the ﬁrst line. then it is tempting to impose the direct modelling assumption S(x + t) ≡ S(x) − tf (x) for integer x and t ∈ [0.5. x + 1). for now we focus on understanding what (3. It is possible to approximate these only because. This assumption. the ‘actuarial approximation’ says that failures known to occur within the year between the x and x+1 birthdays are actually uniformly distributed (have constant conditional density of 1) within that year. P (T ∈ [x+a.15) says about the approximate probability distribution of the lifetime variable T within years of age. says that for any 0 ≤ a < b < 1. While we will later use Taylor expansions like (3.15) to approximate expectations E(g(T )) and conditional expectations E(g(T )  T ≥ x). INTERPOLATION BETWEEN INTEGER AGES 93 functions important in actuarial applications necessarily involve the survival function values between integer ages. 1). for all but the very youngest and oldest ages. the survival function for human lives is very smooth within years of age. as is undoubtedly true for human lifetime between ages 2 and 75 in modern public health conditions. x+b)  [T ] = x) = (b − a) f (x) S(x + a) − S(x + b) = = b−a (S(x) − S(x + 1) f (x) In other words. with derivatives that are not dramatically large and themselves do not change rapidly. This assumption is by far the most commonly used one in actuarial work.15) where θ ∈ (0. together with continuity of S at all integer points. If S = −f is small. In terms of calculus concepts. The ‘actuarialapproximation’ assumption can be understood either as piecewise linearity. and where the second line uses the deﬁnition f (x + t) = −S (x + t) valid at all nonnegative integers x and t ∈ [0. on exactage intervals [x. the function value S(x + t) for integer x and 0 ≤ t < 1 is given approximately by the Taylor series formula S(x + t) = S(x) + t S (x) + 1 2 t S (x + θt) 2 1 = S(x) − t f (x) − t2 f (x + θt) 2 (3. often called the actuarial approximation. of the continuous survival function S(y) or equivalently as piecewise constancy of the density function f (y) = −S (y).3. Two other related possible approximations can .
For integers x and 0 ≤ t ≤ 1. Approximations to S(y). piecewise linearity of log S(x + t) or equivalently piecewise d constancy within intervals x. Many models in biostatistics or reliability have been formulated with piecewise constant hazards (recall that biostatisticians call µ(y) the hazard function while actuaries call it force of mortality). (iii) (Balducci hypothesis) 1/S(x + t) is linear for 0 ≤ t < 1 . i. (ii) (Piecewiseconstant hazard) µ(x + t) is constant for 0 ≤ t < 1 . 1.22) to have properties which make it unsuitable as a realistic model for survival. the slope of the linear function S(x + t) at t = 0 is − f (x). .e.18) and µ(x + t) = f (x) S(x) − tf (x) (3. and µ(x + S(x + t) = S(x) e−t µ(x) .18) also holds. and pk = S(x + 1) = e−µ(x) S(x) (3.17) . x + 1) of dy log S(y) = −f (y)/S(y) ≡ µ(y) is well known and has historically been widely used by biostatisticians. where µ(x + t) = µ(x).. since it will be seen immediately below formula (3. PROBABILITY & LIFE TABLES be obtained by Taylor expanding not S(x + t) itself but rather the functions log S(x + t) or 1/S(x + t). It turns out that the ﬁrst of these alternative assumptions. have been speciﬁed or estimated. that of piecewise linearity of 1/S(y).94 CHAPTER 3. is called the Balducci hypothesis.16) Under assumption (i). 1 1 1 ) = fT (x + ) ST (x + ) 2 2 2 Under (ii). f (y) and µ(y) between integers are usually based on one of the following assumptions: (i) (Piecewiseuniform density) f (x + t) is constant for 0 ≤ t < 1 . which implies easily that S(x + t) = S(x) − tf (x). 2. so that under (i). . To proceed formally. assume that values S(x) for x = 0. f (x) = S(x) − S(x + 1) . S(x + t) is linear in t under − ln S(x + t) 1/S(x + t) assumption (i) assumption (ii) assumption (iii) (3. . The third assumption introduced here. and is studied by actuarial students largely for historical reasons and as a source of examination problems. (3.
24) = S(x + t) = S(x) = (1 − qx)t . the withinyear force of mortality under assumption (i) as given in (3. is actually a decreasing function a feature which seems particularly unrealistic from middle to advanced ages within human lifetimes.5. the result is S(x + 1) S(x + 1) = t + (1 − t) = 1 − (1 − t) qx S(x + t) S(x) Recalling that equivalent to 1−t qx+t t qx (3.21) Next diﬀerentiate the logarithm of the formula (3. INTERPOLATION BETWEEN INTEGER AGES Under (iii). 1 1 1 1 = +t − S(x + t) S(x) S(x + 1) S(x) 95 (3.3.20) with respect to t. By contrast.19) When equation (3. for 0 ≤ t < 1.20) = 1 − (S(x + t)/S(x)). formula (3. reveals assumption (iii) to be =1− S(x + 1) S(x + 1) = (1 − t) 1 − S(x + t) S(x) = (1 − t) qx (3. to show (still under (iii)) that µ(x + t) = − ∂ qx ln S(x + t) = ∂t 1 − (1 − t)qx (3.23) (3.17) is evidently increasing. the formulas are: t px = 1 − t px (S(x) − t(S(x + 1) − S(x)) = 1 − t qx S(x) e−µ(x) t under (i) under (ii) (3.22) Apart from any other property which the Balducci interpolation assumption (iii) might have.19) immediately shows that the withinyear force of mortality µ(x + t). 0 ≤ t < 1. and almost by deﬁnition the piecewiseconstant hazard assumption (ii) entails withinyear constancy of the force of mortality. The most frequent insurance application for the interpolation assumptions (i)(iii) and associated survivalprobability formulas is to express probabilities of survival for fractional years in terms of probabilities of wholeyear survival.19) is multiplied through by S(x + 1) and terms are rearranged. In terms of the notations t px and qx for integers x and 0 < t < 1.
in demography. The integral formula for life expectancy can be written in any of the three .26) This quantity is also called expected residual life or. Thus by deﬁnition ˚x − ex = E( T − [T ]  T ≥ x ) ∈ [0.96 t px CHAPTER 3. or (iii) via respective formulas (3.24).23).) Then assumptions (i). j integers and 0 ≤ t < 1 the complete residual life is T − x = j + t while the curtate residual life is [T ] − x = j. This quantity is larger than the the curtate life expectancy ex because. we can extend the notion of expected remaining life from the curtate to the complete expectation for a life aged x of (T − x) : complete expectation of life = ˚x = E(T − x  T ≥ x) e (3. for a life just completing its x’th year and surviving to exact age x + j + t . (ii).25) are used to substitute into the ﬁnal expression of the following formulas: ∞ ω−1 1 E g(T ) = 0 ω−1 g(t) f (t) dt = x=0 1 0 g(t + x) f (t + x) dt ∂ t px dt ∂t = x=0 S(x) 0 g(t + x) − 3. and (3. think of g(T ) = (1 + i)−T as the present value to an insurer of the payment of $1 which it will make instantaneously at the future time T of death of a newborn which it undertakes to insure.25) The application of all of these formulas can be understood in terms of the formula for expectation of a function g(T ) of the lifetime random variable T . (For a concrete example. life expectancy.5. (3. with x. PROBABILITY & LIFE TABLES = 1 − qx S(x + t) S(x + 1) = S(x + 1) S(x) 1 − (1 − t)qx under (iii) (3.1 Life Expectancy – Deﬁnition and Approximation In terms of a survival function f and S modelling the distribution of exact age at death within years of integer age. 1) e since this diﬀerence is the expectation or weighted average of a quantity between 0 and 1.
the ﬁrst is the basic conditional expectation formula (1. 3.28) There are no formulas nearly as simple for the diﬀerence between complete and curtate life expectancies under interpolation assumptions (ii) or (iii). Therefore.27) via integration by parts. Under the ”actuarial approximation” (assumption (i)) of uniform lifetime distribution within whole years of age. 1) which is uniformly distributed (with constant density 1). The Gamma function Γ(α) is deﬁned by ∞ Γ(α) = 0 xα−1 e−x dx . obtained from the second expression in (3.3. we ﬁnd 1 under (i): ˚x − ex = e 0 t dt = 1 2 (3. using the identities f (x + t)/S(x) = µ(x + t) S(x + t)/S(x) = µ(x + t) t px The third is a continuoustime analogue of Lemma 3. α>0 . we saw above that T −[T ] is a random variable with values in [0.4) for expectation.27) Of these expressions.5) with g(T ) = T − x.6. SOME SPECIAL INTEGRALS ways ω 97 ˚x = e x (y − x) f (y) dy = S(x) ω−x ω−x t t px µ(x + t) dt = 0 0 t px (3.6 Some Special Integrals While actuaries ordinarily do not allow themselves to represent real lifetable survival distributions by simple ﬁniteparameter families of theoretical distributions (for the good reason that such distributions never approximate the real largesample lifetable data well enough). The second is obtained from it by the change of variable t = y − x. Consider ﬁrst the Gamma functions and integrals arising in connection with Gamma survival distributions. it is important for the student to be conversant with several integrals which would arise by substituting some of the theoretical models into formulas for various net single premiums and expected lifetimes.1. using the formula (1. using u = t and dv = µ(x + t) tpx dt = −(1/S(x)) d(S(x + t)).
The last integral can be given by symmetry. so that Φ(0) = 1/2. 0) and [0. The symmetry of the normal density guarantees that half of its probability is assigned to each of (−∞. giving the total probability for an exponentially distributed random variable. and it is a standard integrationbyparts exercise to check that it too is 1. shows that for all positive integers n. applied inductively. For α = 2. Γ(n + 1) = n · (n − 1) · · · 2 · Γ(1) = n! The only other simpletoderive formula explicitly√ giving values for (noninteger) values of the Gamma function is Γ( 1 ) = π.e. More generally. ∞). to show that √ 0 ∞ 1 ∞ −x2 /2 1√ π −z2 /2 −u2 /2 e dz = e du = e dx = 2π = √ 2 −∞ 2 2 −∞ 0 where the last equality is equivalent to the fact (proved in most calculus texts as an exercise in double integration using change of variable to polar coordinates) that the standard normal distribution 1 Φ(x) = √ 2π x e−z −∞ 2 /2 dz (3. using the change of variable u = −z and the fact that the integrand is an even function. obtained as follows: 2 Γ( 1 ) = 2 ∞ ∞ x−1/2 e−x dx = 0 0 e−z 2 /2 √ 2 dz √ Here we have made the integral substitution x = z 2 /2. x−1/2 dx = 2 dz. the integral is the expected value of such a unitexponential random variable. Integration by parts in the Gamma integral with u = xα and dv = e−x dx immediately yields the famous recursion relation for the Gamma integral. . the integral Gamma(α + 1) for positive integer α is the αth moment of the Exponential distribution with parameters λ = 1. PROBABILITY & LIFE TABLES This integral is easily checked to be equal to 1 when α = 1. i. and valid for all α > 0 : ∞ Γ(α + 1) = 0 x e α −x dx = −x e α −x ∞ ∞ + 0 0 α xα−1 e−x dx = α · Γ(α) This relation.98 CHAPTER 3.29) is a bonaﬁde distribution function with limit equal to 1 as x → ∞. ﬁrst derived by Euler. a lifetime with constant forceofmortality 1..
z 1 Gα (z) = v α−1 e−v dv Γ(α) 0 and the integral on the righthand side is called the incomplete Gamma function. One particular case of these integrals. SOME SPECIAL INTEGRALS 99 One of the integrals which arises in calculating expected remaining lifetimes for Weibulldistributed variables is a Gamma integral. via an integration by parts with u = t − x and dv = f (t)dt = −S (t)dt. to give in the Weibull example. or among the standard functions of many mathematical/statistical computer packages. γ t>0 so that S(x) = exp(−λ xγ ). the case α = 1/2 . such as Matlab or R. We change variables by v = y 2/2 to obtain for z ≥ 0.3. after integrationbyparts and a changeofvariable. The expected remaining life for a Weibulldistributed life aged x is calculated. Recall that the Weibull density with parameters λ. E(T − x  T ≥ x) = eλx = Γ( γ 1 −1/γ λ γ ∞ w(1/γ)−1 e−w dw λ xγ 1 λ xγ 1 −1/γ )e λ 1 − G1/γ (λ xγ ) γ γ where we denote by Gα (z) the Gamma distribution function with shape parameter α. Values of Gα (z) can be found either in published tables which are now quite dated.6. as ∞ (t − x) x 1 f (t) γ dt = − (t − x) e−λt S(x) S(x) ∞ ∞ + x x e−λt dt γ The ﬁrst term in square brackets evaluates to 0 at the endpoints. γ is f (t) = λ γ tγ−1 e−λ t . and the second term can be reexpressed via the changeofvariable w = λ tγ with (1/γ) w1/γ−1 dw = λ1/γ dt. √ G1/2 (z) = = z 2z √ 1 1 2 v −1/2 e−v dv = √ 2 e−y /2 dy Γ(1/2) 0 π 0 √ √ 2 √ · 2π · (Φ( 2z) − Φ(0)) = 2Φ( 2z) − 1 π . can be recast in terms of the standard normal distribution function Φ(·).
Suppose that an individual aged 20 has random lifetime (= exact age at death) T with continuous density function fT (t) = 0. 1 − Φ(x ) 2 x = log(x/m) σ 3. where k is a constant. PROBABILITY & LIFE TABLES One further expectedlifetime calculation with a common type of distribution gives results which simplify dramatically and become amenable to numerical calculation. Show that: = tpx · (µx − µx+t ) . Express µx+2 as a function of k. Suppose that the lifetime random variable T is assumed lognormally distributed with parameters m. and deﬁne in particular log(x) − log(m) σ √ Recalling that Φ (z) = exp(−z 2/2)/ 2π .100 CHAPTER 3. we ﬁnd x = 1 E( T − x  T ≥ x ) = 1 − Φ(x ) ∞ x m 2 √ eσy−y /2 dy − x 2π The integral simpliﬁes after completing the square σy − y 2/2 = σ 2 /2 − (y − σ)2/2 in the exponent of the integrand and changing variables by z = y − σ. (2). so that t = m eσy . Then the expected remaining lifetime of a life aged x is E( T − x  T ≥ x ) = 1 S(x) ∞ t x d log(t) − log(m) Φ( ) dt − x dt σ Now change variables by y = (log(t) − log(m))/σ = log(t/m)/σ.7 Exercise Set 3 ∂ p ∂x t x (1). (3). t > 20 . For a certain value of x.02 (t − 20) e−(t−20) 2 /100 . The result is: meσ /2 E( T − x  T ≥ x ) = 1 − Φ(x − σ) − x . σ 2. 3]. it is known that tqx = kt over the timeinterval t ∈ [0.
the integral in (a) ∞ can be evaluated in terms of incomplete Gamma integrals c sα−1 e−s ds. twice as many lives die in each 5year period as in the previous ﬁveyear period. If he deposits $1000 in the fund at the end of each of the ﬁrst 10 years and $1000 + x in the fund at the end of each of the second 10 years. (8). Express in terms of annuityfunctions aN the present value of an annuity of $100 per month paid the ﬁrst year. If the force of mortality can be assumed constant over each ﬁveyear age period (2024.). Find the numerical value of the present value if the eﬀective annual interest rate is 7% . then ﬁnd x to the nearest dollar.e. (4).3.08 − 1 per year) to calculate the present value of the payment. (9).g. Show that: ∞ 0 2 x lx+t µx+t dt = lx . Find upper bounds for the following Binomial probabilities. EXERCISE SET 3 101 (a) If this individual has a contract with your company that you must pay his heirs $106 · (1. The probability that in 1000 Bernoulli trials with success(m) . for integer α > 0.7. then ﬁnd the probability that a life aged 20 will survive at least until exact age 48. s.8. (7).08(T − 20)) (i. 2529. then what is the expected payment ? (b) If the value of the deathpayment described in (a) should properly be discounted by the factor exp(−0. $200 per month for the second year. etc. (6). A man wishes to accumulate $50. w. up to $1000 per month the tenth year. g.0 . Suppose that a lifetable mortality pattern is this: from ages 20 through 60.. and compare them with the exact values calculated via computer (e. and if you are told that l60/l20 = 0. where k. c are positive constants. Find the probability that a life aged 20 will die between exact ages 40 and 50. where the fund earns an eﬀective interest rate of 6% . using a spreadsheet or exact mathematical function such as pbinom in Splus) : (a). 000 in a fund at the end of 20 years.4 − T /50) on the date of his death between ages 20 and 70. then what is the expected present value of the insurance contract ? Hint for both parts: After a change of variables. Obtain an expression for µx if lx = k sx wx g c . (5). where the complete Gamma integrals (for c=0) are known yield the Gamma function Γ(α) = (α − 1)!. by the eﬀective interest rate of e.
the number of successes lies outside the (inclusive) range [364. 3.) of the Gamma(α. 000. for g(t) = tr . Derive the formula for the 2’nd and 3’rd moments of the Weibull(α. r = 2. given that the radix l0 of the lifetable is 10. no more than 1075 survive to retire at age 65. V2 of the payments in each of the two events with probability 0.5 each. Hint: calculate the respective present values V1 . for whom p45 = 0.72.4. 000 in exactly 5 years. 446]. Find the expected present value at 5% APR of an investment whose proceeds will with probability 1/2 be a payment of $10. β) density f (t) = β α tα−1 e−β t I[t≥0] α f (t) g(t) dt as a function of parameters α and β. . Hint: change variables appropriately and use the Gamma function. 000 in exactly 10 years. The probability that of 1650 lives aged exactly 45. and ﬁnd the expected value of a discrete random variable which has values V1 or V2 with probabilities 0. Derive the formula for the 2’nd and 3’rd moments (that is. λ) density f (t) = (λα /Γ(α!)) tα−1 e−λ t I[t≥0] as a function of parameters α and λ. PROBABILITY & LIFE TABLES probability 0. 20 (10). If the force of mortality governing a cohort lifetable is such that µt = 2 2 + 1+t 100 − t for real t . (12). and with the remaining probability 1/2 will be a payment of $20. Hint: change variables by y = λt. (13). 0 < t < 100 then ﬁnd the number of deaths which will be expected to occur between ages 1 and 4.5. (11).102 CHAPTER 3. (b).
01 · 100 · (64 − [y]) dy . (c) Find the expected present value at 4% interest of a payment of $1000 to be made at the end of the year of death of a life currently aged exactly 20. and (a) Suppose that an insurer is to pay an amount $100· (64 − X) (without regard to interest or present values related to the timedeferral of the payment) for a newborn in the lifetable population.024 y for 49 ≤ y ≤ 64.85 − 0. the survival function is piecewise linear. The ﬁrst task is to develop an expression for survival function and density governing the cohort lifetable population. 0. 49). Speciﬁcally for this example. Alternatively. if X denotes the attained integer age at death.01 y for 15 ≤ y ≤ 49. and 1. extending the function S linearly.02 · 100 · (64 − [y]) dy + 0 15 0.8. = 0.02 on [0.02 y for 0 ≤ y ≤ 15. Now the expectation in (a) can be written in terms of the random lifetime variable with density f as 15 49 0.8 Worked Examples Assume that a cohort 200 dx = 100 240 lifetable population satisﬁes l0 = 104 for 0 ≤ x ≤ 14 for 15 ≤ x ≤ 48 for 49 ≤ x ≤ 63 Example 1. WORKED EXAMPLES 103 3. we have the survival density f (y) = −S (y) = 0.024 on [49. 64]. for 0 ≤ y ≤ 15 10000 − 200y ly = 7000 − 100(y − 15) for 16 ≤ y ≤ 49 3600 − 240(y − 49) for 50 ≤ y ≤ 64 It follows that the terminal age for this population is ω = 64 for this population.536 − . and the lifedistribution is piecewise uniform because the the density is piecewise constant. What is the expected amount to be paid ? (b) Find the expectation requested in (a) if the insurance is purchased for a life currently aged exactly 10 . and = 0. Since the numbers of deaths are constant over intervals of years. and S(y) = 1 − 0. at integer values y.01 on [15. 15).3.
So the summations now begin with k = 10. by 15 49 64 2. PROBABILITY & LIFE TABLES 64 0.02 · 100 · (64 − k) + 0. In addition.01 · 100 · (64 − k) + k=15 k=49 0. except that we are dealing with conditional probabilities of lifetimes given to be at least 10 years long.104 + CHAPTER 3. or alternatively end with j = 64 − k = 54. Therefore the integral just displayed can immediately be seen to agree with the summation formula for the expectation of the function 100(64−X) for the integervalued random variable X whose probability mass function is given by P (X = k) = dk /l0 The formula is 14 E(g(Y )) = E(100(64 − X)) = k=0 48 63 0. and the denominators of the conditional probabilities P (X = kX ≥ 10) are l10 = 8000. observe that the integrand (the function g(y) = 100(64 − [y]) of the random lifetime Y whose expectation we are seeking) itself takes a diﬀerent analytical form on successive oneyear age intervals. The method in part (b) is very similar to that in part (a).024 · 100 · (64 − [y]) dy 49 The integral has been written as a sum of three integrals over diﬀerent ranges because the analytical form of the density f in the expectationformula g(y)f (y)dy is diﬀerent on the three diﬀerent intervals. The expectation in (b) then becomes 14 k=10 200 100 240 · 100 · (64 − k) + · 100 · (64 − k) + · 100 · (64 − k) 8000 8000 8000 k=15 k=49 48 63 .024 · 100 · (64 − k) Thus the solution to (a) is given (after the changeofvariable j = 64 − k).4 j=1 j + j=16 j + 2 j=50 j The displayed expressions can be summed either by a calculator program or n by means of the easilychecked formula j=1 j = j(j + 1)/2 to give the numerical answer $3103 .
04)−15 + 1.X year for an individual currently at the 20th birthday is X − 19 years away. (Note that the function 1.9 dy y (80 − z)−1 dz) 40 .) Since l20 = 6500.04)19−k + 6500 63 k=49 240 (1.25 16 j + 2.25 Finally.9 at ages over 40 changes leaves unaﬀected the density of 1/80 for ages less than 40.8. WORKED EXAMPLES which works out to the numerical value 15 49 54 105 3.04 650 0. because the end of the age. The force of mortality here is µ(y) = − d 1 ln(1 − y/80) = dy 80 − y So multiplying it by 0.04−29 65 0.04−29 1 − (1. we ﬁnd the expectation in (c) as a summation beginning at k = 20 for a function 1000 · (1.9 at all exact ages y ≥ 40.5 50 j = $2391. and for ages y over 40 changes the density from f (y) = 1/80 to f ∗ (y) = − d S(40) exp(−0. or (b) the force of mortality µ(y) is decreased by the constant amount 0. the answer to part (c) is 48 1000 k=20 100 (1.04 Example 2. Find the change in the expected lifetime of a cohort lifetable population governed by survival function S(x) = 1 − (x/ω) for 0 ≤ x ≤ ω if ω = 80 and (a) the force of mortality µ(y) is multiplied by 0.04)19−k 6500 = 392.3.04)−X+19 of the random variable X with conditional probability distribution P (X = kX ≥ 20) = dk /l20 for k ≥ 20.1 at all ages y ≥ 40.92 = 1000 24 1 1 − 1.04−X+19 is the present value of a payment of 1 at the end of the year of death.0 1 j + 1.
Partial interest payments with return of principal. The second is a ‘junk bond’ which has probability 0.53. Suppose that you have available to you two investment possibilities.6 of paying 11% compound interest and returning your principal after 10 years. named and displayed along with their probabilities in the following table.5%.9 (2 − y/40)−0. The ﬁrst investment is a riskfree bond (or bank savingsaccount) which returns compound interest of 5% for a 10year period.9 ln((80−y)/40) dy = −0. Example 3. Also assume that the events governing the junk bond’s paying or defaulting are independent of the true interest rate’s being 4. into each of which you are being asked to commit $5000.5% versus 7.9) = 40.3 of paying yearly interest at 11% for 5 years and then returning your principal of $5000 at the end of the 10th year with no further interest payments.45(80/. Suppose further that the going rate of interest with respect to which present values should properly be calculated for the next 10 years will either be 4. Which investment provides the better expected return in terms of current (time0) dollars ? There are six relevant events. probability 0.1 dy 80 Using the change of variable z = 2 − y/40 in the last integral gives the expected lifetime = 10 + . PROBABILITY & LIFE TABLES d 0. each with probability 0.9 (2 − y/40)−0. corresponding to the possible combinations of true interest rate (Low versus High) and payment scenarios for the junk bond (Full payment.5 d dy 80 − y 40 0.5.5% for the next 10 years.1 80 80 0 Thus the expected lifetime changes from 40 80 (y/80) dy = 40 to (y/80) dy + 0 40 y 0.5 e0.9 − 40/1.106 = − CHAPTER 3.1 of paying yearly interest for 3 years at 11% and then defaulting. paying no more interest and not returning the principal.5% or 7. and probability 0. and Default after 3 years’ interest payments): .9 = 0.
the present value of the ﬁrst investment (the riskfree bond) is 10 5000 k=1 0. A5.1).075)−k + (1.05 (1.82 On each of the events A4.5) · (0.15 0.075)−10 = 4141.8.90 Turning to the second investment (the junk bond).05 (1.045)−10 = 5197.05 0.99 Thus.05 107 Note that because of independence (ﬁrst deﬁned in Section 1.30 0. A2 . A3.5 · (5197.30 0.30) = 0.99) = 4669.82 + 4141.3.15 0.5 the overall expected present value of the ﬁrst investment is 0.045)−k + (1.15 Now.g. denoting by P V the .. A6. under each of the events A1. the probabilities of intersected events are calculated as the products of the separate probabilities. P (A2 ) = P (Low) · P (P artial) = (0. e. WORKED EXAMPLES Event Name A1 A2 A3 A4 A5 A6 Description Low ∩ Full Low ∩ Partial Low ∩ Default High ∩ Full High ∩ Partial High ∩ Default Probability 0. the present value of the ﬁrst investment is 10 5000 k=1 0. since P (Low) = P (A1 ∪ A2 ∪ A3) = P (A1) + P (A2 ) + P (A3) = 0.
075)−k + (1. although the ﬁrstinvestment is ‘riskfree’.075)−k = 0.51433 (1.302386 (1. nevertheless beats inﬂation (i.11 k=1 5 (1. it does not keep up with inﬂation in the sense that its present value is not even as large as its starting value. we conclude that the overall expected present value E(P V ) of the second investment is 6 6 E(P V · IAi ) = i=1 i=1 E(P V Ai) P (Ai ) = 5000 · (1.e.11 E(P V  A2)/5000 = 0.108 CHAPTER 3.24024 k=1 5 E(P V  A4)/5000 = 0. we have 10 E(P V  A1)/5000 = 0.11 Therefore.075)−10 = 1.77 So.045)−10 = 1.11 k=1 10 (1.045)−k + (1.286058 k=1 E(P V  A6)/5000 = 0. the expected present value of the accumulation after 10 years is greater than the initial face value of $5000) although with probability P (Default ) = 0. . PROBABILITY & LIFE TABLES present value considered as a random variable.045)−10 = 1.93024 k=1 3 E(P V  A5)/5000 = 0. The second investment. risky as it is.045)−k + (1..045)−k = 0.16435) = 5821.075)−k + (1.11 E(P V  A3)/5000 = 0.075)−10 = 0.10 the investor may be so unfortunate as to emerge (in present value terms) with only 30% of his initial capital.12683 (1.11 k=1 3 (1.
it remains valid if the righthand side is minimized over s. The minimum . APPENDIX ON LARGE DEVIATION PROBABILITIES 109 3. aN ) ∈ {0. then P (X ≥ Nb) ≤ exp − N b ln b p c p + (1 − b) ln 1−b 1−p 1−c 1−p P (X ≤ Nc) ≤ exp − N c ln + (1 − c) ln Proof. . denoting the number of successes in N Bernoulli (p) trials. The calculus minimum does exist and is unique. Since the event [X ≥ Nb] is the union of the disjoint events [X = k] for k ≥ Nb. .9 Appendix to Chapter 3: Large Deviation Probabilities Theorem 3. the second inequality will be derived from it. a suitable subset of the binomial probability mass function values pX (k) are summed to provide P (X ≥ Nb) = k:N b≤k≤N P (X = k) = k≥N b N k pk (1 − p)N −k For every s > 1. If 1 > b > p > c > 0. . p) random variable. The trick in the proof comes now: since the lefthand side of the inequality does not involve s while the righthand side does. as you can check by calculating that the second derivative in s is always positive. 1}N for which j=1 aj = k ≥ Nb. . and since the inequality must be valid for every s > 1. this probability is ≤ k≥N b N k pk (1 − p)N −k sk−N b = s−N b k≥N b N N k (ps)k (1 − p)N −k ≤ s−N b k=0 N k (ps)k (1 − p)N −k = s−N b (1 − p + ps)N Here extra terms (corresponding to k < Nb) have been added in the nexttolast step. which in turn consist of all outcomestrings N (a1. and the binomial theorem was applied in the last step.2 (Large Deviation Inequalities) Suppose that X is a Binomial (N. After the ﬁrst inequality in (a) is proved.3.9.
Therefore when b > p > c.30) a = min (p + δ) ln(1 + δ ) + (1 − p − δ) ln(1 − p (p − δ) ln(1 − δ ) + (1 − p + δ) ln(1 + p δ ) 1−p δ ). it follows that Y also has a Binomial(N. the upper bound for N −1 ln P ([X ≥ bN] ∪ [X ≤ cN]) is strictly negative and does not involve N .110 CHAPTER 3. gives the second inequality for P (Y ≥ Nb) = P (X ≤ Nc). 2 .1 on the probability with which th relative frequency of success X/N diﬀers from the true probability p of success by as much as δ. let δ ∈ (0. For an improved estimate of the probability bounded in Theorem 3. r 1−r Note that for all r between 0. Substituting this value for s yields P (X ≥ Nb) ≤ b (1 − p) p (1 − b) −N b 1−p 1−b N = exp −N b ln as desired. Therefore. min(p. i. Note also that c < p implies b = 1 − c > 1 − p = q. 1. Replace X by Y = N − X. where the ‘successes’ counted by Y are precisely the ‘failures’ in the Bernoulli trials deﬁning X. The probabilities of such large deviations between X/N and δ are in fact exponentially small as a function of the number N . where q = 1 − p. the quantity r ln p +(1 − r) ln 1−p as a function of r is convex and has a unique minimum of 0 at r = p. the ﬁrst inequality applied to Y instead of X with q = 1 − p replacing p and b = 1 − c. 1−b b + (1 − b) ln p 1−p The second inequality follows from the ﬁrst. 1 − p)) be arbitrarily small. choose b = p + δ. and combine the inequalities of part (a) to give the precise estimate: P ( where X − p ≥ δ) ≤ 2 · exp(−Na) N (3.31) This last inequality gives a much stronger and numerically more useful upper bound than Theorem 3. Since Y also is a count of ‘successes’ in Bernoulli (1 − p) trials. q) distribution. at s = b(1 − p)/(p(1 − b)). c = p − δ..e. 1−p > 0 (3.1. PROBABILITY & LIFE TABLES occurs where the ﬁrst derivative of the logarithm of the last expression is 0.
APPENDIX ON LARGE DEVIATION PROBABILITIES 111 If the probabilities P (X/N − p ≥ δ) in Theorem 3. However. Hogg and Tanis 1997). Another. it can be shown that the exponential rate of decay as a function of N in the true tailprobabilities PN = P (X ≥ Nb) or P (X ≤ Nc) in Theorem 3. more operational. The bounds given in Theorem 3..e.g. then why are those bounds of interest ? (These are 1 minus the probabilities illustrated in Table 1.3.2 (i. for practical purposes the normal approximation to the binomial probabilities of large discrepancies from the expectation is generally much more precise than the large deviation bounds of Theorem 3. . We can tolerate the apparent looseness in the bounds because in actuarial applications involving really extreme tail probabilities (e. they provide relatively quick handcalculated estimates showing that large batches of independentcointosses are extremely unlikely to yield relative frequencies ofheads much diﬀerent from the true probability or limiting relative frequency of heads.1 (cf.2 get small with large N much more rapidly than the simpler bounds based on the Chebychev inequality used in proving Theorem 3.1 is that two very large insured cohorts with the same true survival probabilities are very unlikely to have materially diﬀerent survival experience. the constants appearing in square brackets in the exponents on the righthand sides of the bounds) are exactly the right ones: no larger constants replacing them could give correct bounds.1 are generally much smaller than the upper bounds given for them.2.9. Slud and Hoesman 1989). as Table 1 illustrates.) First. way to render this conclusion of Theorem 3.
92 k−1 k/m px = j=0 1/m px+j/m . Expectation E(c(Z)) = i=1 c(zi ) pZ (zi) p. 85 ∞ Nonneg. 94 . 1] p.112 CHAPTER 3.v. t ∈ [0. 90 ω−x−1 Curtate life expectancy ex = j=1 j px p. 75 m Binomial(N.10 Useful Formulas from Chapter 3 P (X = k) = N k pk (1 − p)N −k p. Expectation E(Z) = j=1 P (Z ≥ j) p. p) probability Discrete r.v. 92 (i) Piecewise Unif. k ≥ 1 integer p. S(x+t) = tS(x+1)+(1−t)S(x) . k ≥ 1 integer p. PROBABILITY & LIFE TABLES 3. x integer . integervalued r. 91 k px = px px+1 px+2 · · · px+k−1 .
94 t px = S(x) − t(S(x + 1) − S(x)) = 1 − t qx S(x) under (i) p. 97 ∞ Γ(α) = 0 xα−1 e−x dx p. 95 t px = 1 − qx S(x + t) S(x + 1) = S(x + 1) S(x) 1 − (1 − t)qx under (iii) p. 96 ω−x Complete life expectancy ˚x = e 0 s px ds p.3. 95 t px = S(x + t) = S(x) e−µ(x) t = (1 − qx)t under (ii) p. 94 (iii) Balducci hypothesis t 1−t 1 = + S(x + t) S(x + 1) S(x) p. 98 . USEFUL FORMULAS FROM CHAPTER 3 113 (ii) Piecewise Const. 98 1 Φ(x) = √ 2π x e−z −∞ 2 /2 dz p.10. µ : ln S(x + t) = t ln S(x + 1) + (1 − t) ln S(x) p.
PROBABILITY & LIFE TABLES .114 CHAPTER 3.
and next to use interest theory to deﬁne the present value of the contractual payment stream by the insurer as a nonrandom function of the random individual lifetime T . The details of the further mathematical discussion fall into two parts: ﬁrst. 115 . and second. this leads to a formula for the expected present value of the payout by the insurer. In each case. and (B) discounting of future payment (streams) based on interestrate assumptions. the application of the various survival assumptions concerning interpolation between whole years of age. endowment and life annuity contracts.Chapter 4 Expected Present Values of Insurance Contracts We are now ready to draw together the main strands of the development so far: (A) expectations of discrete and continuous random variables deﬁned as functions of a lifetable waiting time T until death. an amount called the net single premium or net single risk premium of the contract because it is the single cash payment by the insured at the beginning of the insurance period which would exactly compensate for the average of the future payments which the insurer will have to make. We close this Chapter with a discussion of instantaneouspayment insurance. the speciﬁcation of formulas in terms of cohort lifetable quantities for net single premiums of insurances and annuities which pay only at wholeyear intervals. We ﬁrst deﬁne the contractual terms of and discuss relations between the major sorts of insurance. to obtain the corresponding formulas for insurances and annuities which have m payment times per year.
assuming for convenience that the contract is initiated on the holder’s birthday. together with the other notations previously discussed for annuities of nonrandom duration. all times being measured from the date of policy initiation. we adopt several uniform notations and assumptions. Of course. and compare the corresponding integral and summation formulas. We also relate these expectations with their mpaymentperyear discrete analogues. an insurance will pay oﬀ at the end of the fraction 1/m of a year during which death occurs. There are three major types of contracts to consider: insurance. and ﬁnally in the instantaneous case. and endowments. and retain the notation v = (1 + i)−1 . which pay only when a single life terminates due to a speciﬁed cause or set of causes (insurances based on multiple decrement tables). all of which involve continuoustime expectation integrals. . More complicated kinds of contracts — which we do not discuss in detail — can be obtained by combining (superposing or subtracting) these in various ways. or endowment contract. Thus. insurances and annuities on joint lives. Fix a nonrandom eﬀective interest rate i . The approach here diﬀers in unifying concepts by discussing together all of the diﬀerent contracts. for given m. life annuity. (1997). other types of insurances on lives do exist. In what follows. and meanresiduallife formulas. For these further topics. (1997). Let x denote the initial integer age of the holder of the insurance. Similar discussions can be found in the books Life Contingencies by Jordan (1967) and Actuarial Mathematics by Bowers et al.1 Preliminaries The topic of study in this Chapter is contracts resulting in contingent payment streams depending on the age at death T of a single individual. Only single life contracts without distinctions between causes of mortality are discussed here. EXPECTED PRESENT VALUES OF PAYMENTS continuouspayment annuity. next under the interpolation assumption (i) with m payment periods per year. or which contractually involve more than one life (for example husbandwife pairs). denote by m the number of paymentperiods per year. life annuities. Next. and lifeannuities pay regularly m times per year until the annuitant dies.116 CHAPTER 4. ﬁrst in the wholeyear case. we refer the reader to Bowers et al. 4.
all probabilities and conditional expectations refer only to the discrete random variable [T ] = T1 and are calculated in terms of the conditional probability mass function. The random exact age at which the policyholder dies is denoted by T . we say the policy age is t − x. and therefore are smaller by x than the exact age of the policyholder at termination. If k k+1 ≤T −x< m m .1. The probabilities of the various possible occurrences under the policy are therefore calculated using the conditional probability distribution of T given that T ≥ x. When the amounts and times of payments under a contract depend only on the wholeyear age at death (m = 1). and all of the contracts under discussion have the property that T is the only random variable upon which either the amount or time of payment can depend. the amount of the payment will be assumed to depend on T only through the greatest integer less than or equal to mT . PRELIMINARIES 117 The term or duration n of the contract will always be assumed to be an integer multiple of 1/m.2) and depends only on the cohort lifetable entries. Note that policy durations are all measured from policy initiation. we refer to policy time for the life aged x as the time scale with origin at policy initiation. then Tm ≡ k [mT ] = x+ m m (4. and at chronological age t for the policyholder. since the displayed conditional probability is precisely dx+k /lx . given for nonnegative integers x. In examples based on m payment periods per year.1) denotes the attained age at death measured in completed (1/m)’th years.4. and the density by f (t). k by P ([T ] = x + k  [T ] ≥ x) = P (k ≤ T − x < k + 1  T ≥ x) = k px − k+1 px = k px qx+k (4. assumed to be the x birthday of the policyholder. the probability calculations necessarily involve interpolations of the survival function S(t) within whole years of age. Thus. the survival function of T is denoted S(t). In the setting where the insurance and annuity contracts are formulated in terms of m possible death and payment periods per year. which has continuous probability density f (t)/S(x) at all times t ≥ x. k . As before. but only to values t of the form x + k/m.
118 CHAPTER 4. and number m of possible paymentperiods per year. if T ∈ [x. a life aged x. paymentamount function F (·).4) .2 Insurance & Life Annuity Contracts An Insurance contract is an agreement to pay a face amount — perhaps modiﬁed by a speciﬁed function of the time until death — if the insured.3) 4. (Insm ) Specializing to the case m = 1.1): P (Tm = x + = k+1 k k Tm ≥ x) = P ( ≤T −x< T ≥ x) m m m = k/m px 1 k k+1 S(x + ) − S(x + ) S(x) m m − (k+1)/m px = P (T ≥ x + k+1 k k T ≥ x) · P (T < x + T ≥x+ ) m m m = k/m px · 1/m qx+k/m (4. the present value of the insurance company’s payment under the contract (Ins) or (Ins1 ) is F ([T − x]) v [T −x]+1 0 if x ≤ T < x + n otherwise (4. so that Tm = [T ] is the wholeyear age at death. for a speciﬁed term n ≤ ∞. but for example in decreasing term policies the payment will decrease linearly as a function of Tm over the term of the policy.) The general form of this contract. EXPECTED PRESENT VALUES OF PAYMENTS an integer. the term of the policy. Then all conditional expecations are calculated in terms of the probability mass function of the random variable Tm given as in (4. with payment to be made at the end of the 1/m year within which the death occurs. and a term insurance otherwise. dies at any time during a speciﬁed period. x + n). The insurance is said to be a wholelife policy if the duration n = ∞. Usually the payment will simply be the face amount F (0). is to 1 pay F (T − x) at Tm − x + m time units following policy initiation.
we refer to life annuities with ﬁrst payment at time 0 as (life) annuitiesdue and to those with ﬁrst payment at time 1/m (and therefore last payment at time n in the case of a ﬁnite term n over which the annuitant survives) as (life) annuitiesimmediate. In this general setting. In this case. A Life Annuity contract is an agreement to pay a scheduled payment to the policyholder at every interval 1/m of a year while the annuitant is alive. . Again the payment amounts are ordinarily constant. . .2. . . n. up to n payments are made (since F (0) = 0 or F (n) = 0). When the insurance is wholelife (n = ∞). 2. k 0 ≤ m ≤ T − x. TYPES OF CONTRACTS 119 The simplest and most common case of this contract and formula arise when the faceamount F (0) = F is the constant amount paid whenever a death within the term occurs. n − 1. we adopt the convention that in the ﬁniteterm or temporary life annuities.4. (LifAnnm ) As in the case of annuities certain (the nonrandom annuities discussed in Chapter 1). . . 1. at most nm payments. so 1 that Ax:∞ ≡ Ax . the subscript n and bracket n and superscript 1 over x are dropped. . where F (s) is a ﬁxed function and s ranges over multiples of 1/m. Then the payment is F . with F ≡ 1. and both the payment and present value are 0 otherwise. either F (0) = 0 or F (n) = 0. either F (k) ≡ F for k = 0.5) Here the situation is deﬁnitely simpler in the case where the payment amounts F (k) are level or constant. if x ≤ T < x + n. the net single premium has the 1 standard notation Ax:n . and the present value of the insurance company’s payment under the life annuity contract is [T −x] F (k) v k k=0 (4. the life annuity contract requires the insurer to k pay amounts F (k/m) at policy times m . To avoid ambiguity. but in principle any nonrandom timedependent schedule of payments F (k/m) can be used. or F (k) ≡ F for k = 1. In the ﬁrst of these cases. up to a maximum number of nm payments. with present value F v [T ]−x+1. Again specialize to the case m = 1: under the contract (LifAnn) or (LifAnn1). the lifeannuitydue payment stream becomes an annuitydue certain (the kind discussed .
only whether the payment is made at all. since neither the amount nor the time of payment. which have ¨ payments commencing at policy time 1 and continuing annually either until death or for a total of n payments. the present value formula as a function of [T ] is the certain annuity immediate amin([T ]−x. 0 otherwise (4. The third major type of insurance contract is the Endowment. This contract is the simplest. which pays a contractual face amount F = F (0) at the end of n policy years if the policyholder initially aged x survives to age x + n. . both with term n and the same face amount 1. whichever comes ﬁrst.120 CHAPTER 4.5) is amin([T ]−x+1. . and by 0 for larger indices k. The pure endowment contract commits the insurer to pay an amount F at policy time n if T ≥ x+n (Endow) The present value of the pure endowment contract payment is F vn if T ≥ x + n. n − 1. 1. . and its expected present value (= net single premium) is ¨ denoted ax:n . which for a life aged x and term n is simply the sum of the pure endowment and the term insurance. if we replace F (k) by 1 for k = 0. n) . . since this is the present value of the pattern of annual unit payments starting at policy time 1 up to [T − x] or n. then the present value in equation (4. In the case of temporary life annuitiesimmediate. EXPECTED PRESENT VALUES OF PAYMENTS previously under the Theory of Interest) as soon as the random variable T is ﬁxed. Indeed. Here the standard contract with m payment periods per year and unit level face amount calls for the insurer to pay 1 at policy time Tm − x + 1/m if T < x + n (EndInsm ) n if T ≥ x + n .6) The net single premium or expected present value for a pure endowment 1 contract with face amount F = 1 is denoted Ax:n or n Ex and is evidently equal to 1 Ax:n = n Ex = v n n px (4. The expectedpresent value notation for temporary life annuities immediate is ax:n . is uncertain. n) .7) The other contract frequently referred to in actuarial texts is the Endowment Insurance.
This policy pays a constant face amount at the end of the current fraction 1/m year containing the policy time T − x. in which case the superscript 1 is positioned above the term n. m = 1 In this subsection.9) 4.e. this identity is 1 1 Ax:n = Ax:n + Ax:n (4. In terms of net single premiums. because it is built up simply from the contracts already described. we collect a few useful identities connecting the diﬀerent types of risk premiums for contracts with m = 1 payment period per year. The ﬁrst. is the deﬁnition of endowment insurance as the superposition of a constantfaceamount term insurance with a pure endowment of the same face amount and term.2.4. When m = 1.2. is the nyear deferred insurance. we express the present value of this contract as v n on the event [T ≥ n] and as v [T −x]+1 on the complementary event [T < n]. These identities therefore hold and can be used in computational formulas without regard to particular lifetable interpolation assumptions.. in both cases. The notations for the net single premium of the term insurance and of the pure endowment are intended to be mnemonic. When the face amount is 1. the notation and formula for the net single premium is n Ax 1 = Ax − Ax:n (4. respectively denoting the portions of the endowment insurance net single premium respectively triggered by the expiration of life — in which case the superscript 1 is positioned above the x —or by the expiration of the ﬁxed term. Another example of an insurance contract which does not need separate treatment. n) (4. the contractual payout is precisely the diﬀerence between the unit wholelife insurance and the nyear unit term insurance. Thus. Note that [T − x] + 1 ≤ n whenever T − x < n. but only if death occurs after the deferral time n . i. TYPES OF CONTRACTS 121 Simplifying to the case of a single payment per year (m = 1). the present value is given by v min([T −x]+1. after age x + n for a new policyholder aged precisely x.1 Formal Relations between Risk Premiums.10) .8) The expected present value of the unit endowment insurance (still in the case m = 1) is denoted Ax:n . which we have already seen.
in all circumstances.8). we need to relate the risk premiums for life annuitiesimmediate to those of life annuitiesdue.11) is d ax:n + Ax:n = 1 ¨ (4. The diﬀerence arises because the payment streams (for the life annuitydue deferred 1 year and the lifeannuity immediate) end at the same time rather than with the same number of payments when death occurs before time n.8) is the present value of the unit endowment insurance.11) where recall that Ex ( · ) denotes the conditional expectation E( ·  T ≥ x). the present value of the life annuitydue payout coincides with the annuitydue certain. A more common and algebraically equivalent form of the identity (4. n) = ¨ 1 − v min([T −x]+1. Taking expectations (over values of the random variable T . for a ﬁxed value of the random lifetime T .13) . The great generality ofthe identity arises from the fact we saw in the discussion following (4. Taking expectations leads to the formula ax:n = ax:n+1 − 1 ¨ (4. nonrandomduration annuities). EXPECTED PRESENT VALUES OF PAYMENTS Another important identity concerns the relation between expected present values of endowment insurances and life annuities. and substituting (m) Ax:n as expectation of (4.4). n) d where the second expression follows from the ﬁrst via formula (2.e. The correct conversionformula is obtained by treating the life annuityimmediate of term n as paying. then yields: ax:n = Ex ¨ 1 − v min([T −x]+1. n) which we saw in (4.. Unlike the case of annuitiescertain (i. one cannot simply multiply the present value of the life annuitydue for ﬁxed T by the discountfactor v in order to obtain the corresponding present value for the life annuityimmediate with the same term n. conditionally given T ≥ x) in the present value formula. and is given by amin([T −x]+1. Thus. a present value of 1 (equal to the cash payment at policy initiation) less than the life annuitydue with term n + 1.12) To obtain a corresponding identity relating net single premiums for life annuitiesimmediate to those of endowment insurances.5) that.122 CHAPTER 4. the unit life annuitydue has present value which is a simple linear function of v min([T −x]+1. n) d = 1 − Ax:n d (4.
annuity. we ﬁnd ax:n = ax:n+1 − 1 = ¨ and d ax:n + Ax:n+1 = d = v i (4.9). i/d = 1 + i = v −1 Since the nyear deferred insurance with risk premium (4.15) 1 − Ax:n+1 1 1 − 1 = − Ax:n+1 d i d (4. the notation IB denotes the socalled indicator random variable which is equal to 1 whenever T has a value such that the condition B is satisﬁed and is equal to 0 otherwise. TYPES OF CONTRACTS 123 Now.9) pays a beneﬁt only if the insured survives at least n years. With this interpretation.16) 4. the nyear deferred insurance has net single premium = n Ex · Ax+n .14) In these formulas. for an event B depending on the random lifetime T . This expected present value must therefore be equal to (4. written explicitly as sums for the case m = 1.2. Here and from now on. providing the identity: 1 Ax − Ax:n = v n n px · Ax+n (4. and endowment contracts deﬁned above.11). combining this conversionformula with the identity (4. .4. leading to the simpliﬁcations 1/d = 1/i + 1 . To emphasize the fact that these risk premiums depend on cohort life table quantities alone. Recall for this purpose the conditional probability mass function (4. we have made use of the deﬁnition 1/d = (1 + i)/i.2. this subsection collects the formulas for risk premiums of the insurance.2 Formulas for Net Single Premiums All of the net single premiums (or risk premiums) considered so far are computable completely in terms of of lifetable quantities j px and qx+j .2) of [T − x] given T ≥ x. it can alternatively be viewed as an endowment with beneﬁt equal to a whole life insurance to the insured (then aged x + n) after n years if the insured lives that long.
17) The index k in the summation formula denotes the year of policy time within which death occurs. for which T − x ≥ k. Putting together the formula (4.124 CHAPTER 4. n − 1. note that payments of 1 occur at all policy ages k.4) of the random term insurance payment (with level face value F (0) ≡ 1) is n−1 1 Ax:n = Ex v [T −x]+1 I{T ≤x+n} = k=0 v (k+1)/m k px qx+k (4. since the present values of these payments are v k and the payment at policy time k is made with probability k px . a contract which pays G(k) at policy time k if the insured life initially aged x survives to age x + k. of the present values v k+1 to be paid by the insurer in the event that the policy age at death falls in [k.10) provides us with a formula for the net single premium of the endowment insurance. that this event occurs. the expectation of the present value (4. .17) with the previous identity (4. multiplied by the probability.19) Finally the pure endowment has present value n Ex = Ex v n I[T −x≥n] = v n x pn (4.18) Next. to ﬁgure the expected present value of the life annuitydue with term n. k = 0. over all indices k such that k < n. . EXPECTED PRESENT VALUES OF PAYMENTS IB = 1 if condition B holds .20) Most generally of all. Therefore. k + 1). and which pays F (k) at . given in formula (4.2). n−1 Ax:n = k=0 v k+1 k px − k+1 px + v n n px (4. n−1 n−1 ax:n = Ex ¨ k=0 v k I{T −x≥k} = k=0 v k k px (4. . . The summation itself is the weighted sum. = 0 if not First. k ≤ T − x < k + 1.
m > 1 At this point. and net single premium notation and formula (4.3 Risk Premiums & Relations.19) and (4. where both functions G(k) and F (k) could depend on a ﬁnite term parameter n.21) where P ([T − x] = k  T ≥ x) has been expressed as in (4. and endowment contracts deﬁned above. lifeannuity. EXTENSION TO MULTIPLE PAYMENTS PER YEAR 125 policy time k + 1 if the insured life aged x dies at age T ∈ [x + k. However the insurance. In the next section.3.2). . we return to the basic deﬁnitions of the standard insurance. and endowmentinsurance contracts payable only at wholeyear intervals are all slightly impractical as insurance vehicles. encompasses all of the insurances.6). It is therefore no diﬀerent for general m than for m = 1. and identities to cover the case of general mpaymentperiod per year contracts. x + k + 1).7).21) are benchmarks in the sense that they represent a complete solution to the problem of determining net single premiums without the need for interpolation of the lifetable survival function between integer ages. and endowments introduced so far in this Chapter. using only the standard cohort lifetable data collected by integer attained ages. since it involved only a single potential payment at an integer policy time. This setting. does not require any separate discussion here. life annuities.17) and (4. The formulas (4. has net single (risk) premium equal to ω−x−1 Ex k=0 G(k) v k I{T ≥x+k} + F (k) v k+1 I{[T −x]=k} ω−x−1 = k=0 k px G(k) v k + F (k) v k+1 qx+k (4. we approach the calculation of net single premiums for the more realistic context of mperiodperyear insurances and life annuities. in order to extend the theoretical formulas. The pure endowment contract (Endow) with present value formula (4. annuity.4. 4.
over all indices k such that k/m < n. . . k = 0. that this event occurs. (k + 1)/m) multiplied by the probability. given in formula (4. nm−1. for which T − x ≥ k/m. analogous to (4.24) The useful identities described in Section 4. Therefore. That is. Accordingly. since the present values of these payments are (1/m) v k/m and the payment at k/m is made with probability k/m px . connecting the diﬀerent types of risk premiums for contracts with m = 1 payment period per year. note that payments of 1/m occur at all policy ages k/m. The ﬁrst extension. . the payment of 1 occurs at policy time k/m if k/m ≤ T − x < (k + 1)/m. k/m < n.126 CHAPTER 4. Again the risk premium summation is the weighted sum. all have extensions to general m payment per year contracts. To ﬁgure the expected present value of the life annuitydue with term n. of the present values v (k+1)/m to be paid by the insurer in the event that the policy age at death falls in [k/m.3). nm−1 (m) ax:n ¨ = Ex k=0 1 k/m v I[T −x≥k/m] m 1 = m nm−1 v k/m k/m px k=0 (4. EXPECTED PRESENT VALUES OF PAYMENTS Next we consider the pure term insurance (Insm ) with term n and m payment periods per year. Recall that the mpaymentperiod per year endowment insurance with term n and unit face . pays 1 at the end of the 1/m year of death. and level face amount. is the deﬁnition of endowment insurance as the superposition of a constantfaceamount term insurance with a pure endowment of the same face amount and term.2.10).1 above. if death occurs before policy time n.23) Here k/m in the summation formula denotes the beginning of the 1/m year of policy time within which death is to occur. . the present value of the insurer’s payment is nm−1 v (k+1)/m I{k/m≤T −x<(k+1)/m} k=0 (4. Recall that this contract. with unit face amount.22) and the net single premium or expected present value is nm−1 A(m)1 x:n = k=0 v (k+1)/m k/m px 1/m qx+k/m (4.
. n) d(m) = 1 − Ax:n d(m) (m) (4. With m payments per year.26) The general identity (4. n) = (1 − v min(Tm −x+1/m.27) where recall that Ex ( · ) denotes the conditional expectation E( ·  T ≥ x). . the notational identity is m(ω−x)−1 (m) Ax:n = A(m)1 x:n +A (m) 1 x:n = k=0 v min((k+1)/m. and substituting the notation (m) Ax:n then yields: a ¨ x:n = Ex (m) 1 − v min(Tm−x+1/m. An algebraically equivalent form of the identity (4. the termn life annuitydue payout is given via formula (2.11) concerning the relation between expected present values of endowment insurances and life annuities also extends straightforwardly.25) with the formula (4. conditionally given T ≥ x) in the present value formula. Taking expectations (over values of the random variable T . n) of the unit endowment insurance.27) is d(m) ax:n + Ax:n = 1 ¨ (m) (m) (4.4) by amin(T ¨ (m) m −x+1/m.3. .4. . and the individual payments of 1/m again totalling 1 per year. EXTENSION TO MULTIPLE PAYMENTS PER YEAR 127 amount pays 1 at policy time (k + 1)/m if k/m ≤ T − x < (k + 1)/m for k = 0. n) ) / d(m) Again the unit life annuitydue has present value which is a simple linear function of the present value v min(Tm−x+1/m.23) for Insurance: nm−1 Ax:n = k=0 (m) v (k+1)/m k/m px − (k+1)/m px + v n n px (4. The present value of the payout clearly has the single epxression v min(Tm −x+1/m. nm − 1. n) . In terms of net single premiums. 1. and pays 1 at policy time n if T ≥ x + n.28) . n) k/m px 1/m qx+k/m (4.25) nm−1 = k=0 v (k+1)/m k/m px 1/m qx+k/m + v n n px Again we ﬁnd a formula for the endowment insurance by a combining the identity (4.
(m) d i i(m) i(m) = v −1/m = 1+ d(m) m We conclude this section with a general formula extending (4.128 CHAPTER 4. and m payments per year consists of payments 1/m at each of the policy times k/m such that 1 ≤ k ≤ nm and k/n ≤ T − x. . the idea for converting from risk premiums of life annuitiesdue to life annuities immediate is very similar to the idea behind the conversion formula (4. .n)+1/m m m (4. term n years. ¨ establishing the identity ax:n = ax:n+1/m − ¨ (m) (m) 1 (m) ¨ = Ex amin(T −x. for all integers k = 0. EXPECTED PRESENT VALUES OF PAYMENTS For multiple payment periods per year.31) In these formulas. pays m Gx (k/m) at policy time k/m if the insured life initially aged x survives to age x+k/m.30) − − (m) m d m i d d(m) ax:n + Ax:n+1/m = (m) (m) (m) and d(m) = v 1/m i(m) (4. we have made use of the deﬁnition m/ d(m) = (1 + i(m)/m) (i(m)/m) leading to the simpliﬁcations m m = (m) + 1 . m(ω−x)−1. we directly obtain an identity relating the net single premium for life annuitiesimmediate with m payment periods per year to that of the m payment period endowment insurances.21). 1. The corresponding payment stream for an annuitydue with term n + 1/m is exaxtly the same. . The result is (m) ax:n = (m) ax:n+1/m ¨ 1 − Ax:n+1/m 1 1 1 1 (m) = = (m) − (m) Ax:n+1/m (4. . The payment stream for the unit annuityimmediate to a life aged x with payment of 1 per year. . A 1 contract which. Therefore the respective expected present values ax:n and (m) ax:n+1/m diﬀer by exactly the present value of that initial payment of 1/m. except that the latter omits the initial payment of 1/m (m) at time 0.29) From this mpaymentperiod conversion formula.13) for m = 1.
4.3. EXTENSION TO MULTIPLE PAYMENTS PER YEAR
129
and which pays Fx (k/m) at policy time (k + 1)/m if the insured life aged x dies within the exactage interval T ∈ [x + k/m, x + (k + 1)/m), has net single (risk) premium equal to
m(ω−x)−1
Ex
k=0 m(ω−x)−1
1 Gx (k/m) v k/m I{T −x≥k/m} + Fx (k/m) v (k+1)/m I{Tm−x=k/m} m 1 Gx (k) v k/m + Fx(k/m) v (k+1)/m 1/m qx+k/m m
=
k=0
k/m px
(4.32)
See the Worked Examples (numbers 3 and 4) for illustrations of numerical calculations with the standard formulas (4.22) and (4.24), as well as the general formula (4.32). The idea behind equation (4.32) can also be used to express the net single premium of a life insurance or annuity in a varying interest rate environment. Following the ideas of Sections 1.2.4 and 1.2.5, we know that the present valueof a unit payment at policy time t under a timevarying instantaneous interest rate r(s) ≡ exp(δ(s)) − 1 (expressed in terms of the policy timet argument s) is 1/A(t) = exp(− 0 δ(s) ds). Then the present value of a term insurance of duration n with level payment amount F at policy time (k + 1)/m if the insured life aged x dies within the exactage interval T ∈ [x + k/m, x + (k + 1)/m), 0 ≤ k < nm, is given by
(k+1)/m
F · exp −
0
δ(s) ds I{Tm=x+k/m} =
F I{k/m≤T −x<(k+1)/m} A((k + 1)/m)
Similarly, the present value of a temporary life annuity due of duration n 1 which makes level payments of amount m G at all policy times k/m ≤ min(T − x, n), is G · m
nm−1 j/m
exp −
j=0 0
δ(s) ds I{j/m≤T −x}
G = m
nm−1
j=0
1 I{j/m≤T −x} A(k/m)
Then the net single premium of a contract which pays G/m at policy time k/m ≤ n if the insured life initially aged x survives to age x + k/m, and which pays F at policy time (k + 1)/m ≤ n if the insured life aged x
130
CHAPTER 4. EXPECTED PRESENT VALUES OF PAYMENTS
dies within the exactage interval T ∈ [x + k/m, x + (k + 1)/m), is the expectation of the sum of the last two displayed expressions, and is given by
nm−1 k/m px k=0
G F + 1/m qx+k/m m · A(k/m) A((k + 1)/m)
(4.33)
4.4
Interpolation Formulas in Risk Premiums
A key issue in understanding the special nature of life insurances and annuities with multiple payment periods is the calculation of these probabilities from the underlying probabilities j py (for integers j, y) which can be deduced or estimated from lifetables. In the present Section, we combine the Actuarial Assumption — (i) of Chapter 3, saying that deaths are uniformly distributed within whole year of age — with the insurance and (temporary) life annuitydue risk premium formulas. In this setting, the number m of payment periods per year is greater than 1, and by formula (3.23) for all integers j = 0, 1, . . . , m − 1:
j/m px
= 1 − j/m qx = 1 − (j/m) qx =
j/m px
so that
j/m px 1/m qx+j/m
−
(j+1)/m px
= (1/m) qx
For any integer k = bm + j, where 0 ≤ l < m − 1 and b is an integer,
k/m px 1/m qx+k/m
= b px ( j/m px+b −
(j+1)/m px+b )
= (1/m) b px qx+b
(4.34)
Substituting (4.34) into (4.23) with summation indices k = bm + j, gives
n−1 m−1
A(m)1 x:n
=
b=0 j=0
v b+(j+1)/m b+j/m px 1/m qx+b+j/m 1 m
n−1 m−1
= 1 m
m−1
v b b px qx+b
b=0 j=0
v (j+1)/m
n−1
=
v (j+1)/m
j=0
· (1 + i)
b=0
v b+1 b px qx+b
4.4. INTERPOLATION FORMULAS IN RISK PREMIUMS
131
The two factors in parentheses in the ﬁnal displayed expression are respec(m) and the onetively the oneyear annuityimmediate present value a1 1 paymentperyear term insurance riskpremium Ax:n . Since (1 + i) a1
(m)
= (1 + i) (1 − v)/i(m) = i/i(m)
it follows that under interpolation assumption (i),
n−1
A(m)1 x:n
= (i/i
(m)
)
b=0
1 v b+1 b px qx+b = (i/i(m)) Ax:n
(4.35)
Similarly, formula (4.34) substituted into the temporary life annuity formula (4.24) with summation index k = bm + j gives
(m) ¨ x:n a
1 = m
n−1 m−1
v
b=0 j=0
b+j/m
j/m px+b
· b px
1 = m
n−1
m−1
v b px
b=0 m−1 j=0
b
v j/m (1−
j qx+b ) m (4.36)
1 = m
m−1
v
j=0
j/m
ax:n ¨
1+i − m2
1 j v j/m Ax:n j=0
This formula can be reduced further in either of two ways. First, one can appeal to the deﬁnition of increasing temporary annuitydue and refer to the formula given in paragraph (iv) of Section 2.1.1: 1 m2 = (I (m)a)1 ¨
m−1
jv
j=0
j/m
1 = 2 m
m−1
(j + 1) v j/m −
j=0
1 (m) a ¨ m 1 (4.37)
1−v 1 = (m) (m) md d Alternatively, using the identity
(m)
−
1−v 1−v −v − (m) d m
Ax:n = A(m)1 + v n n px x:n within (4.27), and directly substituting (4.35), yields under (i) ax:n = ¨
(m)
(m)
1 d(m)
1 −
i i(m)
1 Ax:n − v n n px
(4.38)
We leave as an Exercise for the interested reader to verify algebraically, using (4.37) together with (4.11), that the formulas (4.36) and (4.38) are equal.
In the limit as m gets large. EXPECTED PRESENT VALUES OF PAYMENTS 4. n ≥ 1. Since the displayed expectation formulas are exactly valid when applied to the function g([tM ]/M. ω]. so that g(t)−g([tM ]/M can be made uniformly small by choice of a suﬃciently large multiple M of m. it follows also for the general continuous function g that ω−x Ex (g(T )) = lim Ex (g(TM )) = M →∞ 0 g(x + s) f (x + s) ds S(x) (4.40) .132 CHAPTER 4.3) as the probability mass function for Tm .39) or. m(ω−x)−1 Ex (g(T )) = k=0 g(x + k/m) k/m px 1/mpx+k/m where we have used (4. ﬁrst in the setting of one payment period per year (m = 1) and then in the case of multiple payment periods (m > 1) per year. with the substitution f (x + s) = µ(x + s) S(x + s). note that for any function g(T ) which depends on T only through the last completed 1/m’th year Tm = [T m]/m. the risk premium formulas become expected present values calculated as continuous integrals with respect to survival densities. The displayed expectation formula is then also valid with m replaced by any integer multiple M = mn. ω−x Ex (g(T )) = 0 g(x + s) µ(x + s) S(x + s) ds S(x) (4.5 Continuous Risk Premium Formulas The present chapter has developed formulas for the net single premiums or risk premiums of the principal life insurance and annuity contracts. To recall why this limit exists. and has the equivalent expression M (ω−x)−1 k=0 g(x + k/M) S(x) x+(k+1)/M x+(k+1)/M f (t) dt x+k/M ω−x 1 = S(x) M (ω−x)−1 g(t) f (t) dt = k=0 x+k/M 0 g(x + s) f (x + s) ds S(x) Assume now that the function g(t) is continuous (and therefore uniformly continuous) on the bounded lifetime interval [0.
4. However. or to calculate other expectations of demographic and biostatistical interest. 4. CONTINUOUS RISK PREMIUM FORMULAS 133 Similar justiﬁcations can be given for these expectation formulas also whenever g is piecewise continuous. where v = (1 + i)−1 and δ = ln(1 + i) . These integral formulas can be used either to calculate the limiting values of expected present values for insurance contracts with large m. is amin(T −x. an instantaneouspayment or continuous insurance with facevalue F is a contract which pays an amount F at the instant of death of the insured. as a function of the exact age T at death for an annuitant initially of exact integer age x. n) where n is the (possibly inﬁnite) duration of the life annuity. such as life expectancies.) As a function of the random lifetime T for the insured life initially with exact integer age x. Here the present value of the contractual payments. Recall that ∞ K aK = 0 v t I[t≤K] dt = 0 v t dt = (1 − v K )/δ is the present value of a continuous payment stream of 1 per unit time of duration K units. Insurance and Annuity contracts can also be deﬁned with respectively instantaneous and continuous payments. this means that when the actual payment is made at some later time.1 Continuous Insurance Contracts So far in this Chapter. The expected present values or net single premiums on a life aged x are respectively denoted Ax for a wholelife contract 1 and Ax:n for an nyear term insurance. First.5.5. ¯ ¯ The objective of this section is to develop and interpret formulas for the . The continuous life annuity is a contract which provides continuous payments at rate 1 per unit time for duration equal to the smaller of the remaining lifetime of the annuitant or the term of n years. (In practice. the amount paid is F together with interest compounded from the instant of death. all of the expectations considered have been associated with the discretized random lifetime variables [T ] and Tm = [mT ]/m. the present value of the amount paid is F · v T −x for a wholelife insurance and F · v T −x · I[T <x+n] for an nyear term insurance. The actuarial notation for the net single premium of the temporary continuous life annuity is ax:n which simpliﬁes to ax when n = ∞. as follows.
Ax . .39) or (4. namely the mean residual life (also called complete life expectancy) ˚x = Ex (T − x) for a life aged x.5. v y−x . e 4. the integral formulas for these three cases are: ∞ ˚x = Ex (T − x) = e 0 ∞ s µ(x + s) s px ds v s µ(x + s) s px ds 0 n (4. or v y−x · I{y−x<n} which respectively have the conditional Ex (·) expectations ˚x e .43) = Ex v T −x I{T −x≤n} = 0 v s µ(x + s) s px ds Next. 1 Ax:n For easy reference.2 Integral Formulas We apply the continuous conditional expectation formulas (4.134 CHAPTER 4.40) directly for the three choices g(y) = y−x . we obtain two additional formulas. along with one further quantity which has been deﬁned as a continuoustime expectation of the lifetime variable T . EXPECTED PRESENT VALUES OF PAYMENTS continuoustime net single premiums.42) (4. for continuous life annuitiesdue ax and ax:n which correspond to Ex {g(T )} for the two choices ω−x n g(t) = 0 v y I{y+x≤t} dy or 0 v y I{y+x≤t} dy where we naturally assume that x + n ≤ ω in the case of the temporary life annuity.41) Ax = Ex (v T −x) = 1 Ax:n (4.
45) ax:n = Ex = As seen above in (4.43).e. in the form Ex I{y≤T −x} = P (T ≥ x + s  T ≥ x) = S(x + y) = y px S(x) the resulting two equations become ω−x ω−x ax = Ex 0 n v y I{y≤T −x} dt v I{y≤T −x} dy 0 y = 0 n v y y px dy v y y px dy 0 (4.19) either from (4.46) also converges for large m to the leftmost term. the term insurance net single premiums A(m)1 = Ex v Tm −x+1/m x:n approach the continuous insurance value (4. and evaluating the conditional expectation of an indicator as a conditional probability. risk premiums for continuous insurance and annuity contracts have a close relationship to the corresponding contracts with m payment periods per year for large m.45) can be obtained as a limit of formulas (4.39) or by using Riemann sums. For the continuous annuity.42) as a limit when m → ∞. nm−1 ax:n = lim m→∞ (m) ax:n ¨ = lim m→∞ k=0 1 k/m v k/m px = m n v t t px ds 0 The ﬁnal formula coincides with (4. i. (4.4.44) (4. That is. as the number m of payments per year goes to ∞.46) Since the righthand term in the inequality (4. so that the following obvious inequalities hold: 1 1 Ax:n ≤ A(m)1 ≤ v 1/m Ax:n x:n (4. A simple direct proof can be given because the payments at the end of the fraction 1/m of year of death are at most 1/m years later than the continuousinsurance payment at the instant of death. CONTINUOUS RISK PREMIUM FORMULAS 135 After switching the order of the integrals and the conditional expectations.5. the middle term which is sandwiched in between must converge to the same limit (4.39).. according with the intuition that the limit as m → ∞ of the paymentstream which pays 1/m at intervals of time .45).
T − x). there are simple formulas relating Ax:n 1 and ax:n to Ax:n . the complete life expectancy ˚x in (4. The comparison between complete and curtate life expectancies under more general withinyear survival distributions in subsection 4.38).48) Finally. (4.45) can be contrasted with the respective expectations (3.5.28).5.17) and (4. EXPECTED PRESENT VALUES OF PAYMENTS 1/m between 0 and Tm − x inclusive is the continuous paymentstream which pays 1 per unit time throughout the policyage interval [0. under (i) ¯ ¯ Ax:n = lim A(m)1 = lim x:n m→∞ m i i(m) 1 Ax:n = i 1 A δ x:n (4.4 below.28) that ¯ δ ax:n + Ax:n = 1 ¯ (4.43). . The expressions in formulas (4.41).3 Risk Premiums under Theoretical Models Let us work out examples of the multiple timeperiod per year and continuous formulas analytically and numerically.136 CHAPTER 4. also under (i).35). ¨ ax:n = lim ax:n = lim ¯ m→∞ m (m) 1 d(m) 1 − i i(m) 1 Ax:n − v n n px or ax:n = ¯ i 1 1 1 − Ax:n − v n n px δ δ (4. and (4.49) More elaborate relations will be given in the next Chapter between net single premium formulas which do require interpolationassumptions for probabilities of survival to times between integer ages to formulas for m = 1. which do not require such interpolation.47) and by (4. by (4. Indeed. The limiting argument of the previous paragraph shows immediately that ¯ under interpolation assumption (i).41) is compared to the curtate life e expectancy ex under interplation assumption (i) in (3. it is easy to see by passing to the limit m → ∞ in (4. (4. 4. under a particular theoretical survival model.12). In particular.19) for a function of the integervalued random variable [T ] (taking m = 1).
γ n ax:n = ¯ 0 v s e−λs ds γ Finally.42). (m) ax:n ¨ = k=0 j=0 v k+j/m e−λ (k+j/m) γ The continuous cases (m = ∞) of these formulas are as follows: n ¯1 Ax:n = 0 v s λγ sγ−1 e−λs ds . the special Weibulllifetime temporary life annuitydue risk premiums for ﬁnite m are: n−1 n−1 m−1 a ¨ x:n = k=0 v e k −λ k γ .41) the curtate and complete life expectancies at integer ages y ≥ x for Weibull (residual) lifetimes are: ∞ ∞ ex = k=0 k (e−λk − e−λ(k+1) ) . ∞.11) and (4.4. and (4. with force of mortality for T given by µ(x + s) = λ γ sγ−1 . (4. according to formulas (3.44).19). CONTINUOUS RISK PREMIUM FORMULAS 137 Consider ﬁrst the slightly artiﬁcial (but still useful) case where the residual life T − x for a life aged x is precisely Weibull(γ. and (4. 4.5. for m = 1.17). qx+k = 1 − exp(−λ{(k + 1)γ − k γ }) Then respectively according to formulas (4. (4. we give an R function for Weibulllifetime term insurance risk premiums: . and s px = e−λ s γ for all s ≥ 0 .23). we ﬁnd formulas for terminsurance ﬁnitem net single premiums under Weibull survival as follows: n−1 1 Ax:n = k=0 v k+1 (e−λk − e−λ(k+1) ) γ γ n−1 m−1 A(m)1 = x:n k=0 v k+1/m j=0 v j/m (e−λ(k+j/m) − e−λ(k+(j+1)/m) ) γ γ According to formulas (4. λ) distributed. First.24). γ γ ˚x = e 0 s λγ sγ−1 eλs ds γ We next give R code and a table showing some numerical comparisons of these insurance and lifeannuity risk premiums.
925 = . n.05.925. we begin by coding a function to solve for λ. termIns.65. . in the R code to produce this Table.50) lambda gamma 1.6. gamma.5405.05 or 0. age2) { ### Function to solve for lambda and gamma Weibull ### parameters. . n. 50 p40 = .v^s*s^(.pi2.m = Insm.. with parameters (λ. when S(age1)=pi1. S(age2)=pi2 are fixed.lam.04324 as in Figure 2.32. and continuous ins v = 1/(1+i) xk = 0:(n1) Ins1 = v*sum( v^xk*(exp(lambda*xk^gamma)exp(lambda*(xk+1)^gamma)) ) xkjm = (0:(n*m1))/m Insm = v^(1/m)* sum( v^xkjm*(exp(lambda*xkjm^gamma) exp(lambda*(xkjm+1/m)^gamma)) ) InsC = lambda*gamma*integrate(function(s. 50p40 = . γ) when π1 = 32p40 .0.5/. age1.5. > LamGam = function(pi1..v=v..925.653e+00 . we create a small Table of values for a few diﬀerent values of n. γ) similar to those used in Chapter 2. or (c) 32p40 = .04/. The interest rates considered here are 0. and n = 20.047.138 CHAPTER 4.gam1)*exp(. m paymts. i) { ## Function to calculate termn insurance risk prem ### for 1 paymt per year.gam=gamma)$val c(termIns. 50p40 = . π2 = 50p40 are given.06.lam*s^(.v.1 = Ins1. EXPECTED PRESENT VALUES OF PAYMENTS > WeibIns = function(lambda.952e06 3. (b) 32p40 = .5/. haz1 = log(pi1) . haz2 = log(pi2) gam = log(haz2/haz1)/log(age2/age1) lam = haz1/age1^gam c(lambda=lam. the nominal age x is 40.. termIns.cont = InsC) } To illustrate the use of this function.925 = . gamma=gam) } > LamGam(.lam=lambda. .gam) .04/. In each case.gam)). m. but chosen successively so that (a) 32p40 = .
lamgam[2].5).04962 [2. "Ins..c(.715e07 4.009 0.02924 0.6.20.653 0. .715e07 4.04/.6).06 0.653 0.lamgam[2].009 0. "i".] 4.04960 139 > InsArr = array(0.04846 0.952e6.2]) [1] 0.02479 0.cont 0.] 1.] = c(lamgam[1].m termIns.5/. WeibIns(lamgam[1].3.05)) for(a in 1:2) for (b in 1:3) { k = 2*(b1)+a > lamgam=LamGam(sprobs[b..1 termIns.03396 0. "Ins.047).] 1. "gamma". .07467 0.653.05) termIns.04278 0.952e06 3.06) sprobs = rbind(c(.02151 The variations among the parameters and interest rates make much more diﬀerence to the results than does the number m of payment periods per year.sprobs[b.05435 A similar Table of risk premiums for temporary life annuitiesdue. which (for oddnumbered rows) takes the three values > 1exp(AnnuArr[c(1.5).1].03008 [5.C"))) intrat = c(.05 0.3.04189 0. c("lambda". dimnames=list(NULL. Note that the overall size of the risk premium for a unit 20year term insurance is reasonable.intrat[a])) } InsArr lambda gamma i Ins. "Ins.02494 [6.1". dim=c(6.02436 0.02986 0.06 0.1 Ins.4.32.] 4.04844 0. with exactly the same parameters.345 0.02091 0.] 1.04932 0.50) InsArr[k.02135 0.c(..65.10461 0.3.06 0.03457 0.05 0.04930 0..2].925).] 1.925.4.m".345 0.241e07 4.05 0. because the probability it will pay anything at all is 20 q40 .05.20.1]*20^AnnuArr[c(1.04309 [3.m Ins.952e06 3.C [1.4.5.intrat[a].03478 [4. is given below along with the R code for the lifeannuity function calculation. CONTINUOUS RISK PREMIUM FORMULAS > WeibIns(1.241e07 4.
m.140 CHAPTER 4.12) respectively provide the complete and curtate agespeciﬁc life expectancies.4.lam*s^(.] 1.72 11.cont = AnnuC) } > AnnuArr = InsArr dimnames(AnnuArr)[[2]][4:6]=c("Annu.64 [4."Annu.06 11.] 4.0.gam=gamma)$val c(tmpAnn. n.05 12. gamma.82 11.intrat[a]) } AnnuArr lambda gamma i Annu.v^s*exp(.952e06 3.69 [5.65 12.78 11.00 12.41) or (3.009 0.2].1].C [1.06 12. and continuous ins v = 1/(1+i) xk = 0:(n1) Annu1 = sum( v^xk*exp(lambda*xk^gamma) ) xkjm = (0:(n*m1))/m Annum = sum( v^xkjm*exp(lambda*xkjm^gamma) )/m AnnuC = integrate(function(s.06 12.sprobs[b..32.gam) .v.009 0.1 = Annu1.] 4.345 0. ..96 12.m = Annum.50) AnnuArr[k. .] 1."Annu.C") for(a in 1:2) for (b in 1:3) { k = 2*(b1)+a lamgam=LamGam(sprobs[b.4:6] = WeibAnnD(lamgam[1].] 1.1".241e07 4. n.952e06 3.241e07 4.99 11. . i) { ## Function to calculate termn insurance risk prem ### for 1 paymt per year.63 [3.4 Numerical Calculations of Life Expectancies Formulas (4. in terms respectively of survival densities and life .90 12.] 1.20.56 [2. Examples of formula development for other special parametric distributional families are contained in the Exercises.v=v.68 [6.. tmpAnn.lam.653 0. 4.05 13.5.m Annu.715e07 4.72 12.73 Life expectancy calculations and comparisons are done in the next subsection. EXPECTED PRESENT VALUES OF PAYMENTS > WeibAnnD = function(lambda. m paymts.m".lam=lambda.1 Annu.715e07 4.gam)).09 11.76 12.345 0.05 11. for Gompertz survival.05 12. tmpAnn.653 0.lamgam[2].
28) provides the actuarial approximation for complete life expectancy in terms of lifetable data. According to that extrapolation.1. based upon interpolationassumption (i) (Uniform mortality within year of age).4. beyond that they will be very sparse and very dependent on the particular small group(s) of aged individuals used in constructing the particular table(s).0436. This exponential behavior is approximately but not precisely compatible with a Gompertzform forceofmortality function µ(78 + t) = µ(78) ct in light of the approximate equality µ(x) ≈ qx. the fraction of the cohort at moderate ages who will survive past 90.10) of the illustrative life table data from Table 1.0885). so a reasonable extrapolation of a wellestablished table out to age 105 or so may give suﬃciently accurate lifeexpectancy values at ages not exceeding 105.0885 in our setting. . To see this. All of the numerical lifeexpectancy calculations produced for the Figure of this Section are based on the extrapolation (2. (c − 1)/ ln c = 1. Life expectancies are in any case forecasts based upon an implicit assumption of future mortality following exactly the same pattern as recent past mortality. Thus an assertion of great accuracy for a particular method of calculation would be misplaced. In this Section. deathrates qx for all ages 78 and greater are taken to grow exponentially. note that under a Gompertz survival model.1. is extremely small. Lifeexpectancy calculations necessarily ignore likely changes in living conditions and medical technology which many who are currently alive will experience. an approximation which progressively becomes less valid as the force of mortality gets larger. Life expectancy formulas necessarily involve life table data and/or survival distributions speciﬁed out to arbitrarily large ages. Formula (3. CONTINUOUS RISK PREMIUM FORMULAS 141 table data. On the other hand. we illustrate these formulas using the Illustrative simulated and extrapolated lifetable data of Table 1. µ(x) = Bcx . say. qx = 1 − exp −Bcx c−1 ln c and with c = 1.5. with log(qx/q78 ) = (x − 78) ln(1. While life tables may be based on large cohorts of insured for ages up to the seventies and even eighties.
24).50) The complete minus curtate life expectancies calculated from this formula were found range from 0. down to 0. so would be still larger today. 1.142 CHAPTER 4. as plotted points. Figure 4. its calculation is simplest and most easily interpreted. and interpolation assumption (ii).16) of piecewiseconstant force of mortality within years of age. . and force of mortality is d ln t px+k = − ln px+k dt Using formulas (4. the exact formula for the diﬀerence between complete and curtate life expectancy becomes µ(x + k + t) = − ω−x−1 k+1 k px k=0 ω−x−1 k 1 k px (− ln px+k ) k=0 ω−x−1 0 ˚x − ex = e s µ(x + s) s−k px+k ds − k qx+k = (k + t) et ln px+k dt − k qx+k = k=0 k px (− ln px+k ) ω−x−1 (k + 1)px+k − k px+k − 1 − ln px+k (ln px+k )2 k px − k qx+k = k=0 − px+k − qx+k ln px+k (4. mortality within year of age (0 < t < 1) is t t px+k = (px+k ) .485 at age 78 and 0. under assumption (ii) (from Chapter 3. The aspect of Figure 4.348 at age 99. Since the complete life expectancy at each age is larger than the curtate by exactly 1/2 under interpolation assumption (i).) Thus there is essentially no new information in the calculated complete life expectancies. (3.1 presents. the agespeciﬁc curtate life expectancies for integer ages x = 0. .12) relies directly on (extrapolated) lifetable data. EXPECTED PRESENT VALUES OF PAYMENTS Since curtate life expectancy (3. (Contrast this result with the constant diﬀerence of 1/2 under assumption (i). by formula (3. 78. Under this assumption. we calculated for comparison the complete life expectancy at all (realnumber) ages. as in equation 3. and they are not plotted. . the large life expectancies shown are comparable to actual US male mortality circa 1959. . Moreover.12).493 at ages 40 and below.1 which is most startling to the intuition is the large expected numbers of additional birthdays for individuals of advanced ages. .41).
10). with agespeciﬁc deathrates qx extrapolated as indicated in formula (2. by age 70 •••• 60 •• •• •• •• •• •• •• 50 •• Curtate Life Expectancy •• •• •• •• 40 •• •• •• •• •• 30 •• •• •• •• •• •• 20 •• •• •• •• •• •• •• •• 10 •• ••• ••• ••• •• 80 0 20 40 Age in years 60 Figure 4.5. .1: Curtate life expectancy ex as a function of age.4. calculated from the simulated illustrative life table data of Table 1.1. CONTINUOUS RISK PREMIUM FORMULAS 143 Expected number of additional whole years of life.
.325). (a) Find the net single wholelife insurance premium Ax for this person.03 for t ≥ 50.025 Assume the ﬁxed interest rate r = 0. (a) Find the expected present value. 1. i.00634. with respect to the constant eﬀective interest rate r = 0. (b) Find the expected present value of the insurance payment in (a) if the insurer is allowed to delay the payment to the end of the year in which the individual dies.05 e−0.144 CHAPTER 4. and µt = 0.07. and is certain to die within this time. and/or numerical integrations via calculators or software.6 Exercise Set 4 (1). s > 0. 1. If the individual in Problem 2 pays a life insurance premium P at the beginning of each remaining year of his life (including this one).00634 t1. . you should be prepared to use integrations by parts. then what is the expected total present value of all the premiums he pays before his death ? (4). (2). with S(t) = 1 − Φ((log(t) − log(50))/0. gamma function values. of an insurance payment of $1000 to be made at the instant of death of an individual who has just turned 40 and whose remaining lifetime T − 40 = S is a continuous random variable with density f (s) = 0. (i) Weibull(.05 s . . In these integrals.2 ). Suppose that an individual has equal probability of dying within each of the next 40 years. 39 k px − k+1 px = 0.e.06. 0.3252 ). . and (b) 7/12q40/q40 . For each of the following three lifetime distributions. EXPECTED PRESENT VALUES OF PAYMENTS 4. with S(t) = exp(−0.015 for 20 < t ≤ 50. (iii) Piecewise exponential with force of mortality given the constant value µt = 0. . 1 Ax:20 (b) Find the net single premium for the term and endowment insurances and Ax:30 . tables of the normal distribution function Φ(x). his age is x and for k = 0. ﬁnd (a) the expected remaining lifetime for an individual aged 20. . (ii) Lognormal(log(50). Should this answer be larger or smaller than the answer in (a) ? (3).2).
Find the probability of an individual aged 72 in this lifetable population dying between ages 75. payable at the end of the year of death along with interest from the beginning of that same year) satisﬁes the recursion relation (4. e .4.0. (8).51) above. Show that the expected present value bx of an insurance of 1 payable at the beginning of the year of death (or equivalently.0 and 83.0.16) algebraically. 000.0001 · t for 40 ≤ t Then (a) ﬁnd 30 p10 . For the next two problems. Prove the identity (4. and (b) ﬁnd ˚50. EXERCISE SET 4 145 (5).001 for 5 ≤ t < 20 µt = .0 and 78. l85 = 0. (b) assuming linearity of 1/S(t)) within 5year age intervals. l80 = 3000. Find (a) ˚75 and (b) the probability of an individual aged 70 in e this lifetable population dying between ages 72. Show that the expected whole number of years of remaining life for a life aged x is given by ω−x−1 cx = E([T ] − x  T ≥ x) = k=0 k k px qx+k and prove that this quantity as a function of integer age x satisﬁes the recursion equation cx = px (1 + cx+1 ) (6). (10).6. l75 = 7000. Suppose that a population has survival probabilities governed at all ages by the force of mortality for 0 ≤ t < 1 .002 for 1 ≤ t < 5 . if the assumption of uniform deathtimes within 5year intervals is replaced by: (a) a constant force of mortality within 5year ageintervals. (9). and that the distribution of deathtimes within 5year age intervals is uniform. (7). consider a cohort lifetable population for which you know only that l70 = 10.004 for 20 ≤ t < 40 .01 .
e (12). 96. n. and (b) ﬁnd ˚30 . 000 for college at her 18th birthday if she lives that long and $500. An nyear term life insurance policy to a life aged x provides that if the insured dies within the nyear period an annuitycertain of yearly payments of 10 will be paid to the beneﬁciary. (a) Calculate the average number of future years of life for an individual who survives to age 1. Assume x + n ≤ ω. and 97 are respectively 0.146 CHAPTER 4. and if mortality follows the law lx = C(ω − x)/ω for some terminal integer age ω and constant C. The child is to receive $100. (16).01 .1 µt = 3/t Then (a) ﬁnd 30p20 = the probability that an individual aged 20 survives for at least 30 more years. Suppose that a population ages by the force of mortality .09. 0 ≤ x ≤ 10. You are given a survival function S(x) = (10 − x)2 /100 . The father of a newborn child purchases an endowment and insurance contract with the following combination of beneﬁts. and the father as beneﬁciary is to receive $200. and the probability that a life aged 1 dies before age 2. and 0. N) denotes the net single premium (= expected present value) for this policy. Assuming the same force of mortality as in the previous problem.3. n. 000 at her 60th birthday if she lives that long.4. 000 at the end of the year of the child’s death if the child . What is the probability that an impaired life age 95 will live to age 98 ? (14). ω. with the ﬁrst annuity payment made on the policyanniversary following death. and the last payment made on the N th policy anniversary. N. ﬁnd ˚70 and A60 if i = 0. The standard rates qx of mortality at ages 95. EXPECTED PRESENT VALUES OF PAYMENTS has survival probabilities governed at all for 0 ≤ t < 10 for 10 ≤ t < 30 for 30 ≤ t (11).5 . The force of mortality for impaired lives is three times the standard force of mortality at all ages. Here 1 < n ≤ N are ﬁxed integers. n. N) in terms of interestrate functions. e (13). then ﬁnd a simpliﬁed expression for B(x. and the integers x. If B(x. (b) Calculate the diﬀerence between the force of mortality at age 1. (15). 0.
for the net single premium for this contract. 10] and i = . (18). . 4) distribution for T −x.4. EXERCISE SET 4 147 dies before age 18. Verify algebraically. 0 ≤ t ≤ 10 and not only on the values for integer t. in a variable interest rate environment with instantaneous force of interest δ(t) at policy time t.6. 10]. .06 for policy ages in (10. that the righthand sides of formulas (4. and that it is believed that interest rates will vary over the 10year interval according to the rule δ(t) = δ · (1 + 0.5.33) to derive an analogous formula for the net single premium. concerning insurance contract risk premiums in a variable interest rate environment. m = 1) depends on the survival distribution only through the cohort life table quantities k px for integers k = 1. Suppose that a life aged x wants to purchase a 20year term insurance now. For the next four problems.e.002 t).3 to aid in the calculation. Pass to the limit m → ∞ in formula (4.37) together with (4. Suppose that a life aged x wants to purchase a 10year term insurance or temporary annuitydue. Find expressions. and that interest rates over the next twenty years will be i = .38) are equal. . apply formula (4. (20). Find the net single premium for a unit term insurance (i. Show that the net single premium for a unit 10year term insurance or a unit temporary life annuitydue. using (4. both in actuarial notations and in terms of v = 1/(1 + i) and of the survival probabilities k p0 for the child. (21). 20].. of a contract for a life aged x paying a continuoustime stream at rate G at all policy times t < min(T − x. if the insurance payment is 1) if survival is governed by the Weibull(2 · 10−6 .) (19). (You may use the formulas and R code of section 4.e. .07 for policy ages in (0. n) and paying a lumpsum amount F at the instant of death if death occurs before age x + n.. with one payment period per year (i.04 for policy ages in (6.11). Suppose that a life aged x wants to purchase a 10year term insurance. (17). and that interest rates over the next ten years will be i = .36) and (4. 10. Show that the net single premium for a unit 10year term insurance depends on the continuous conditional survival probabilities t px . 6] and i = . 2. .05 for policy ages in (0.33).
98 − 700 + 314.) Assume that the rate of interest i = . (That is.09. 5) [5.749 0.844 0. Toy LifeTable (assuming uniform failures) Consider the following lifetable with only six equallyspaced ages.9174 and (1 − e−δ )/δ = (1 − v)/δ = 0. 2) [2.17) as 1000 A 1 = 1000 0.2 3.148 CHAPTER 4. for a life aged 0. First.26 = 1320.958 Using the data in this Table.5 Ax 0.917)2 860 860 = 1705. EXPECTED PRESENT VALUES OF PAYMENTS 4. and interest rate i = . so that v = 1/(1 + i) = 0.917)3 1000 1000 1000 = 199.896 0. 6) lx 1000 940 860 760 640 500 dx 60 80 100 120 140 500 ex 4.7 Worked Examples Example 1. the 3year Endowment for $700 has present value 1 700 A2:3 = 700 · (0. a 3year term insurance with payoﬀ amount $1000 has present value given by formula (4.26 860 Thus we can also calculate (for the life aged 2) the present value of the 3year annuityimmediate of $700 per year as 1 700 · ¨ 2:3 − 1 + A0:3 a = 1705.9174)3 500 = 314.917 760 640 + (0. for a life aged 2. a 3year temporary annuitydue of $700 per year (with last payment at age 4) has present value computed from (4. and endowment. we begin by calculating the expected present values for simple contracts for term insurance. x 0 1 2 3 4 5 Agerange [0.19) to be 700 a2:3 = 700 ¨ 1 + 0.917)2 + (0. 1) [1.09.24 .917 0:3 60 80 100 + (0. 4) [4.436 2. 3) [3.98 For the same life aged 2.60 Second.0 1. assume l6 = 0. annuity.795 0.709 2.9582.281 0.704 0.
This payment stream.9582 bx where bx is the expected present value of an insurance of 1 payable at the beginning of the year of death (so that Ax = v bx ) and satisﬁes b5 = 1 together with the recursionrelation 5−x bx = k=0 k px qx+k v k = px v bx+1 + qx (4.1. together with the observation that j px · qx+j = lx+j dx+j dx+j · = lx lx+j lx to show how the last two columns of the Table were computed. Now apply the present value formula for continuous insurance to ﬁnd 5−x Ax = k=0 1 − e−δ = 0. .9582 k px qx+k v δ k 5−x k px qx k=0 v k = 0.709 860 860 860 860 2 860 Moreover: observe that cx = 5−x k k px qx+k satisﬁes the “recursion equak=0 tion” cx = px (1 + cx+1 ) (cf. WORKED EXAMPLES 149 We next apply and interpret the formulas of Section 4.4.41) e2 = 120 140 500 1 1900 100 ·0+ ·1+ ·2+ ·3+ = + 0. 1. .7. which pays F (k) = C an−k if death occurs at any exact ages ¨ between x + k and x + k + 1.5. In particular. with c5 = 0. for k = 0. and interpret the result. if superposed upon . Exercise 5 above). n − 1. .51) (Proof of this recursion is Exercise 6 above. .) Example 2.5 = 2. from which the ex column is easily computed by: ex = cx + 0.5. Let us begin with the interpretation: the beneﬁciary receives at the end of the year of death a lumpsum equal in present value to a payment stream of C annually beginning at the end of the year of death and terminating at the end of the nth policy year. Find a simpliﬁed expression in terms of actuarial expected present value notations for the net single premium of an insurance on a life aged x. by (4.
Thus the expected present value in this example is given by C an − C ax:n (4. . Find the risk premiums for unitfaceamount 6year duration annuitydue and term insurance. n.3 of dying within the 4 quarteryears given the year of failure. .3. . . Assume interest rate 5% throughout. (b) with m = 4 and uniform failure density within year of failure.52).2. (a) with m=1. the net single premium in the example is equal to n−1 n−1 v k+1 k px qx+k C ¨ n−k+1 = C a k=0 n−1 k=0 v k+1 k px qx+k n−1 1 − v n−k d C = d = v k=0 k+1 k px qx+k −v n+1 k=0 (k px − k+1 px ) C 1 Ax:n − v n+1 (1 − n px ) d C = Ax:n − v n n px − v n+1 (1 − n px ) d and ﬁnally. 2. Example 3. . would result in a certain payment of C at the end of policy years 1. .150 CHAPTER 4. Consider the following cohort life table fragment applicable to lives aged from 30 to 36. and (c) with m = 4 and respective failure probabilities . . By formula (4. . by substituting expression (4. EXPECTED PRESENT VALUES OF PAYMENTS an nyear life annuityimmediate with annual payments C.15) with m = 1 for Ax:n .2. we have C 1 − d ax:n − (1 − v) v n n px − v n+1 ¨ d C = 1 − d (1 + ax:n − v n n px ) − d v n n px − v n+1 d C 1 − vn = v − d ax:n − v n+1 = C − ax:n d i = C {an − ax:n } So the analytically derived answer agrees with the one intuitively arrived at in formula (4.52).52) Next we rework this example purely in terms of analytical formulas.
1.0000 0.187)/95000 > kpx = 1cumsum(c(0.19) and (4. 34) [34. 31) [31. 33) [33.001632 0. > probmass = c(165.05^((0:5)) * kpx ) c(Ains=Ains. 36) lx 95000 94835 94685 94530 94472 94300 dx 165 150 155 158 172 187 151 To clarify the diﬀerent calculations in the three parts (a)(c).m = Ains*i4/.9916 > Ains = sum( 1. 35) [35.7.m = aux1*AnnDue .001737 0.001811 0. n−1 m−1 b px b=0 j=0 A(m)1 = x:n v b+(j+1)/m ((j+1)/m qx+b − j/m qx+b ) . AnnDue.9951 0. AnnDue=AnnDue) Ains AnnDue 0. 32) [32. (4.008751 5.4.9934 0.05^((0:3)/4))/4 aux2 = sum((0:3)*1.24) and calculate by decomposing sums as we did in Section 4.36).05*aux2*Ains) Ains.probmass[1:5])) > probmass [1] 0.05^((0:3)/4))/4^2 c(Ains.001579 0.001663 0.05.23) and (4.155. we provide R code as well as numerical answers.251) aux1 = sum(1.05^.17). WORKED EXAMPLES x 30 31 32 33 34 35 Agerange [30.05^((1:6)) * probmass ) AnnDue = sum( 1.209397 ### answer to (b) For part (c).172. we start from formulas (4.008592 5.9983 0.150.158.4. Parts (a) and (b) respectively make use of formulas (4.35) plus (4.m AnnDue.9967 0.308505 ### answer to (a) > i4 = 4*(1.m 0.001968 > kpx [1] 1.
75+.05^.2 v 1/4 + 0.152 CHAPTER 4.2 v −3/4 + 0. Substituting.3 if j = 2.3 v 1 = Ax:n · 0.ptc = Ains*aux3.2 v 1/4 + + 0.3 v 3/4 + 0.05^.ptc AnnDue.2*1.4 v −1/2 + 0. EXPECTED PRESENT VALUES OF PAYMENTS n−1 (m) a ¨ x:n = b=0 1 b px m m−1 v b+j/m (1 − j/m qx+b ) j=0 Now for m = 4. .05^.7 v −1/4 ) 1 1 Ax:n 0.4*1.05^.7 for j = 0. we ﬁnd n−1 A(4)1 x:n = b=0 b px qx+b v b+1 (1 + i) 0.5+.2 if j = 0. AnnDue.2*(1.5)+. 3.4. 0.ptc = AnnDue .7 v −1/4 4 = ax:n − ¨ The numerics in R for part (c) now follow: > aux3 = .75+1.305604 So the diﬀerent withinyear distribution of failures made roughly a 2% difference in the term insurance and temporary life annuitydue risk premiums. 0.2 v 1/2 + 0. 1.25 c(Ains.3 v (1 + i) and (4) ax:n ¨ 1 = 4 n−1 b px b=0 v b 1 − v qx+b (0.7*1.3*(1. we have in this problem the special assumption that for integers b and 0 ≤ j < 4.05^.3 v 3/4 + 0.008892 5.25 + 1) aux4 = .2 v 1/2 + 0. (j/4qx+b − (j+1)/4qx+b )/qx+b = 0. 0.25*Ains*aux4 Ains. 2.4 v −1/2 + 0.0. 3 It follows that j/4qx+b / qx+b has respective values 0.05^.2 v −3/4 + 0.2.ptc 0. 1 0.
"AnnDue".491.74) kpx2 = 1pnorm(log(0:24).5. (a) Gamma(14. Find the risk premiums for a 24year life annuitydue and insurance ﬁgured with m = 1 and m = 4 based on the probability densities."P(Tx>24)"))) Ains AnnDue P(Tx>12) P(Tx>24) Gamma 0.mean=3. which can be seen also in Figure 2.4.41 1.246) array(c(sum(diff(kpx1)/1. while the Lognormal distribution function values are simply expressed in terms of the normal distribution function.5 in Chapter 2.4). kpx2[24]).05^(1:24)).9258 The large relative diﬀerence between the terminsurance risk premiums is due to the more rapid decrease of the Gamma versus the Lognormal survival function between 12 and 24 years.9001 Lognormal 0.7. sd =. .4383. kpx1[12]. with parameters as given in Figure 2. . In neither case need we perform numerical integrations. sum(kpx2[1:24]/1.491.4383). kpx1[24]. and (b) Lognormal(3. sum(kpx1[1:24]/1.05^(0:23)).74.04558 14.9998 0. Here the ideas and formulas are simple: the only issue is how to organize the numerical calculations.246)2 ). dim=c(2. kpx2[12].03504 14. WORKED EXAMPLES 153 Example 4. shape=14. The Gamma distribution function can be directly called in R. (.36 0. sum(diff(kpx2)/1. kpx1 = 1pgamma(0:24. c("Ains"."Lognormal").0000 0. dimnames=list(c("Gamma"."P(Tx>12)". rate=.05^(0:23)).05^(1:24)).
8 Useful Formulas from Chapter 4 Tm = [T m]/m = x + k m if x + k k+1 ≤T <x+ m m p. 117 P (Tm = x + k  T ≥ x) = m k/m px − (k+1)/m px = k/m px · 1/m qx+k/m p. 122 Term (temporary) life annuity ax:n = ax:n+1/m − 1/m ¨ p. 123 n−1 1 Term Insurance Ax:n = Ex v [T −x]+1 I{T <x+n} = k=0 v k+1 k px qx+k p. 118 1 Endowment Ax:n = n Ex = Ex v n I[T −x≥n] = v n n px p. 122 A(m) − A(m)1 x:n x = v n n px · Ax+n p. 124 . n) d = 1 − Ax:n d p. 121 ax:n = Ex ¨ 1 − v min([T −x]+1. EXPECTED PRESENT VALUES OF PAYMENTS 4.154 CHAPTER 4. 122 d ax:n + Ax:n = 1 ¨ p. 120 1 1 1 Ax:n = Ax:n + Ax:n = Ax:n + n Ex p.
127 a ¨ x:n = Ex (m) 1 − v min(Tm−x+1/m.n) ¨ = k=0 v k k px p.4. n) d(m) = 1 − Ax:n d(m) (m) p. 124 nm−1 1 Ax:n = Ex v Tm−x+1/m I{T <x+n} = k=0 v (k+1)/m k/m px 1/m qx+k/m p.8. 127 nm−1 (m) Ax:n = k=0 v (k+1)/m k/m px − (k+1)/m px + v n n px p. 124 n−1 Ax:n = k=0 v k+1 k px − k+1 px + v n n px p. 127 d(m) ax:n + Ax:n = 1 ¨ p. 126 1 Ax:n = A(m)1 + A(m) x:n = A(m)1 + n Ex x:n x:n (m) p. USEFUL FORMULAS FROM CHAPTER 4 155 n−1 ¨ ax:n = Ex amin([T −x]+1. 127 ax:n = ax:n+1/m − 1/m ¨ p. 128 (m) (m) (m) (m) .
131 ∞ ˚x = Ex (T − x) = e 0 s µ(x + s) s px ds p. 135 ¯ δ ax:n + Ax:n = 1 ¯ p. 131 under (i): ax:n = ¨ (m) 1 d(m) 1 − i i(m) 1 Ax:n − v n n px p. EXPECTED PRESENT VALUES OF PAYMENTS n−1 under (i): A(m)1 x:n = (i/i (m) ) b=0 1 v b+1 b px qx+b = (i/i(m)) Ax:n p. 134 n n ax:n = Ex 0 v I{y≤T −x} dy y = 0 v y y px dy p. 142 . 134 n 1 Ax:n = Ex v T −x I{T −x≤n} = 0 v s µ(x + s) s px ds p.156 CHAPTER 4. 136 ω−x−1 under (ii): ˚x − ex = e k=0 k px − px+k − qx+k ln px+k p.
where the object of study is either cumulative time or cumulative operational loading in an engineered system until failure or speciﬁed degradation of performance. we introduce some of the features of real data structures embodying waitingtime or duration data. or death. • Epidemiology.Appendix A Duration Data Structures In this Chapter. where the durations of interest are the times until accident. health emergency. where larger human populations are followed between recruitment to a study population • Reliability. including: • Life Insurance. or tumor recurrence or return of other disease condition). and 157 . where human lives meeting speciﬁc criteria are followed between some initiating event (such as diagnosis of a disease or a speciﬁc treatment or intervention) and a response of interest (such as alleviation of symptoms. • Casualty Insurance. where payments are made and received as contractually determined functions of the duration of an insured individual’s lifetime. Such data arise in a wide variety of disciplines and applied ﬁelds. such as mortgage insurance relating to the waiting time until a speciﬁed emergency resulting in • Clinical Trials and other Biomedical studies. • Other Insurance. or other adverse occurrence resulting in liability or loss.
such as those for individuals from employment to unemployment or vice versa. the population might consist of those in certain professions or risky occcupations. where the waiting times of interest are generally times of transition. usually people. ageintervals. the study population is deﬁned thorugh qualifying characteristics. and locations. All of them also consider probability distributions and expected values for functions depending on waitingtime random variables. In an epidemiologic context. and who consent to be randomly assigned to an experimental versus standard treatment. being observed over chronological time intervals from entry into the study until the occurrence of an event of interest – the study endpoint – or the end of followup – called a rightcensoring time – whichever comes ﬁrst. A. For example. mortality or timetoevent statistics might be compiled for all individuals insured by one or a group of companies. incompletely.1 Concepts and Terminology of Duration (or Mortality. reduce the complex data as actually recorded into the idealized format of the Life Table. In an insurance context. or the subset . Study Population. who are between 30 and 60 years of age and otherwise is good health. consisting of a waitingtime random variable T observed. In a formal observational setting. and for many purposes of statistical anlysis and estimation. for economies between macroeconomic events. or who have speciﬁed existing medical conditions (such as high cholesterol) and consent to participate in a study entailing a number of scheduled medical examinations. or Survival) Studies We next deﬁne and discuss some concepts and terminology that will allow us to identify common versus distinct aspects of duration data in the diﬀerent subject areas listed above. for many individuals of a study population. All of these examples involve the analysis of ‘lifetime’ or ‘waitingtime’ or ‘duration’ data. We restrict attention in our discussion to studies and datasets concerning individuals. etc. for businesses from inception to proﬁtability or bankruptcy. GENERAL FEATURES OF DURATION DATA • Economics. one might recruit into a clinical trial males with the same disease diagnosis at a designated set of hospital centers.158 APPENDIX A. over a speciﬁed timewindow.
and data are to be collected by followup over time on a set of subjects who receive this diagnosis within a short period. ‘cohort’ refers to a set of individuals who have the same wholenumber age at the same time and whose waiting time until death or other failureevent is of interest. one could refer to the ‘cohort’ of US males in the state of New Jersey who were 50 years old in 1977. at diﬀering chronological times chosen by the individual. entry of individuals into observation occurs by staggered entry. so that all entry times into the study are simultaneous. In some insurance tabulations.A. the term “cohort study” applies to longitudinal data collected on a set of individuals selected simultaneously at the outset. Mode of entry. On the other hand. In this way. large populations are studied beginning at a speciﬁed date. For example. More broadly. and with a slightly diﬀerent meaning.or lowrisk populations in order to justify premiums diﬀerent from those paid by the general population. The simplest case is where these are the same. as would a reliability study in which 100 machines of a given type are set running – possibly under heighted load or stress – and observed until failure. which is the way we use it in this book. In demography or epidemiologic studies. for example in a survey. ‘cohort’ and ‘longitudinal’ study are roughly synonymous. birth is the event initiating the individual’s clock. and then followed over time. the initiating event for the interval of length T can have different possible relationships to the chronological time at which an individual is brought under observation. but the individuals’ ages at entry vary. a study in which babies born in a given year are followed for the next period of (3 or 10 or 20) years would be a cohort study. Depending on the purpose for which timetoevent data are gathered. In that usage. Another example would be a survival study in which the interesting duration variable T is ‘time from diagnosis to death’. Usually in Insurance and demography and epidemiology. SURVIVAL DATA CONCEPTS 159 of such people subject to a particular risk – such as ‘cigarette smokers’.1. . or where all individuals in the study are entered simultaneously: this kind of survival or duration study is called a cohort study. most survival studies and insurance portfolios consist of individuals who at any single chronological time have widely diﬀering current ages. say three months. As used in an actuarial or demographic context. Whether in clinical trials or Insurance. the time variable of interest is age. data are gathered on special high. Thus.
Survival and other duration studies are often conducted over ﬁxed administrative time windows.160 APPENDIX A. . The study will end and be reported as of a ﬁxed chronological termination time. then each individual must come equipped with at least the following information: Ei = chronological time of entry of individual i into the study Ai = age of individual i at entry into the study Ti = age of individual i at last followup or endpoint under the study Di = binary indicator equal to 1 if i experiences endpoint during followup. we can formulate the following general data structure for a duration or survival study. which is also called that individual’s rightcensoring time. Based on the examples and discussion above. the individual’s data are said to be leftcensored: the individual could have been observed to experience the study endpoint only at an age greater than the entry age. and the latest is Ti. the individual’s T is known only to be greater than or equal to the last age of followup. Ei + Ti − Ai ]. for a total duration of Ti − Ai . GENERAL FEATURES OF DURATION DATA but at entry into the datacollection. For these reasons. Moreover. 2. individuals can withdraw from observation before the study endpoint. for reasons which may or may not be related to the nearness of that endpoint. . individual i ﬁrst enters the study at chronological time Ei and is under active observation. N . in many studies as in Insurance portfolios. . and equal to 0 otherwise In terms of these notations. When the entry age is positive. or under followup. in a cohort. . data about the individual’s variable T may be incomplete and are said to be rightcensored: within the dataset. the ﬁnal age Ti is also the age at which the study . staggered. so individuals under study may have a positive age at entry and age of last followup in the study without ever having experienced the study endpoint. or individually. Subjects enter either together. The individual’s earliest age in the study is Ai. Mode of study termination. If the individuals in a study are indexed administratively by i = 1. Thus the chronological interval of followup is [Ei . the individual’s age is recorded. If Di = 1.
Expressed in terms of age during the followup period. The biomedical term would be that individual i is at risk at ages in the interval (Ai . atrisk intervals (solid line segments).A. as might be said also in an epidemiologic or demographic context) on that age interval. and death (ﬁlled circle) or rightcensoring (hollow circle) for each of 9 patients in an artiﬁcial clinical trial. individual i would be said to be on test — by adoption of an older terminology from Reliability — on the age interval [Ai .1: Representation of entry times (solid diamonds). Ti]. while the Insurance term is that the individual is exposed (or ‘exposed to risk’. The terminology ‘under followup’ is the one used in clinical or epidemiologic settings. Artificial Clinical Trial with Staggered Entry Right−Censoring. then individual i does not experience the study endpoint while under followup. . while if Di = 0. SURVIVAL DATA CONCEPTS 161 endpoint is observed to occur for individual i.1. and Dropout Patient # 1 1 2 3 4 5 6 7 8 9 2 3 4 5 6 Time 7 8 9 10 Figure A. Ti ].
for 9 patients in an artiﬁcial clinical trial with stagggered entry. A. and also on what the target sampled population was. is to tabulate over equally spaced intervals of age. there are still other complexities. described in Section A. solid if an observed death and hollow if a dropout or censoring time. are those of Klein and Moeschberger (2003) and Lee (1992).e. in the oneyear . ‘Dropout’ takes place at the hollow circle for patient 2.1. This idea is fundamental to statistical analysis of duration data in all of the ﬁelds of study listed at the beginning of this Chapter. Each patient’s interval at risk is a horizontal line segment beginning just after the entrytime depicted with a solid diamond and ending at the time indicated with a circle.2 below. including whether the criteria for inclusion depend on timedependent (for example healthrelated) variables. Two books describing many forms of biomedical survival data. we have addressed the most frequently occurring complexities of Insurance and other survival data as actually collected. or their underlying study populations are often deﬁned through information collected in sample surveys. while the other censoring events would be called ‘administrative rightcensoring’ because the clinical trial observation periods all end at the chronological time 10 indicated by a vertical dashed line. Survival data. or more speciﬁcally the cohort life table.162 APPENDIX A. in which some demographic groups are given heavier weight than their proportion in the general population. to experience the study endpoint while under followup). The Life Table. The way in which the data should be analyzed depend on survey inclusion probabilities or weights.. However. is a simpliﬁed representation which summarizes only the numbers ‘at risk’ and the numbers observed to fail and be censored. the numbers of study subjects at risk and observed to fail (i.2 Formal Notion of the Life Table Consider the artiﬁcial clinical trial data summarized in Figure A.1. or timeontest. The central idea of the Life Table. In this Section. some of which we can mention brieﬂy. Such data might have been collected for the purpose of understanding the rate of mortality at diﬀerent ages. with examples. GENERAL FEATURES OF DURATION DATA The deﬁnitions for rightcensored staggeredentry survival data are illustrated in Figure A.
we will see in Chapter 8 that a reasonable estimator of the deathrate within ageintervals k.A. To complete this brief illustration. In the notation of the previous Section. If a were the common initial age. the total number of subjectyears in the survival study during which subjects were under followup at exact ages in [k. the patient labelled i is said to be at risk (i. τk is more informative than simply Yk as a denominator against which to compare the observed number of failures dk in order to estimate the rate of failures within the successive ageintervals. . A.1. it is important to tabulate an additional quantity. k + 1) = ck = # rightcensored in [k. we tabulate in Table A. k + 1) are the ratios λk = dk /τk .2. 3). FORMAL NOTION OF THE LIFE TABLE 163 age intervals [1. Indeed. k+1) = i (min(Ti .1 The Cohort Life Table If a cohort of individual subjects were entered into a study simultaneously with the same agevariable and followed up until they died. then the proportions Yk /Ya would estimate . k + 1): τk = Time on test at ages in [k. and a simpler interpretation. then we have Yk = # at risk at age k = i I[Ai <k≤Ti ] Di I[k≤Ti<k+1] i dk = # observed to die in [k. Ai )) Clearly. For this reason. and is actually observed to die at Ti only if Di = 1. and the table itself would contain all of the information (Ei .e. . it is not true that all individuals dying within the age interval [k. Ti. k + 1) = i (1 − Di ) I[k≤Ti <k+1] However. 2). In that case. . [2.1 the quantities mentioned so far for the data represented in Figure A. k + 1) were necessarily at risk at time k. the rightcensored counts ck would all be 0. 10). k+1) − max(k.2. could potentially be observed to die) at time t if Ai < t ≤ Ti. If we agree to consider only integer times t = k. then the life table could have a simpler form. [9. Ai . . Di )n i=1 for the n subjects.
which are subjected to the same diets or survival stresses.g. Despite the fact that mortality data for large human populations are generally not collected in cohorts. Regardless of how the data in a mortality study were collected. Animal studies can follow cohorts. But the one area of application in which this kind of data is very common is Engineering Reliability.0 6.0 0. the data are often tabulated as though they were collected that way. divided by the total number of personyears spent by subjects under followup in the study . However. can data actually be collected in this cohort format. and observed until they fail. 9) [9.05 3. qk = number of observed deaths at exact ages in the interval [k. 6) [6.165 1.0 3. 5) [5. 4) [4. follow large numbers of people – many of whom are of the same age or fall into narrow age brackets initially – over time. of newborn laboratory rats.0 the fraction of the population studied which survived to age k.8 6.1: Life table quantities for the staggeredentry survival data used to construct Figure A. 3) [3. k 1 2 3 4 5 6 7 8 9 Ageint [1.0 6. 10) Yk 0 2 2 4 5 6 5 6 4 dk 0 0 0 0 0 1 1 1 3 ck 0 0 0 0 0 0 0 0 1 τk 1.0 0.0 λk 0. GENERAL FEATURES OF DURATION DATA Table A.0 0.0 0. in parallel. 8) [8. One is a study which follows up a cohort of newborns. 7) [7. e. or a cohort of people selected somehow either at the same age or with the same initiating event (like diagnosis of a disease whose mortality is of interest). once can ﬁrst estimate the agespeciﬁc death rates directly. k + 1).2 2.0 4. and the identity Yk+1 = Yk − dk would always hold.8 6. where a number of devices are set running at identical (usually accelerated) stresses.0 0.167 0. Some longitudinal epidemiologic studies. 2) [2. only in very special applications.147 0. not in Insurance.1. like the famous Framingham study [ref ?] which monitored various risk factors for heart disease.164 APPENDIX A.
rounded) death rates qk . k + 1) given survival to the k’th birthday. This ‘life table’ is an artiﬁcial construction. but containing exactly the same information as the column of (smoothed. Denote by [·] the greatestinteger or ﬂoor function. the estimated deathrates are often altered slightly to enhance smoothness of the estimated deathrates as a sequence indexed by k. brieﬂy. Next to these columns may also be displayed the deathrates qk .3 Sample Spaces for Duration Data The preceding sections have described ﬁrst the actual setting in which random durations are observed within a realistic mortality study. Begin by choosing a large conventional size l0 . in presenting the deathrates for purposes of calculation of insurance premiums or population projections. displays the (integerrounded) numbers of expected survivors at each birthday k and numbers of deaths between successive birthdays. referring directly to no actually observed population. usually 105 . k + 1) for k ranging from 0 up to and including the largest integer age ω − 1. in Demography or Insurance. where ω is the terminal age) seen for any subject of the mortality study. Moreover. Finally. SAMPLE SPACES FOR DURATION DATA 165 This might be done in practice only after approximating or imputing the times on test not directly observable in the study. It is the mortality record of a ﬁctitious population cohort with exactly the same death rates after smoothing and rounding as those estimated from some actual population. apart from rounding errors. Then the cohort life table consists of the columns k−1 lk = l0 j=0 qj = number of lives aged k dk = lk − lk+1 = number of deaths at ages in [k. This number is a power of 10.A. the death rates are presented in the tabular form which we now deﬁne as the cohort life table. called the radix of the cohort life table. A. for a population of large hypothetical size experiencing exactly the death rates qk interpreted as conditional probabilities of dying at age [k.3. and then the . for a population cohort of newborns. The cohort life table. Note that. qk = dk /lk for all k.
.166 APPENDIX A. . l0 } enumerates the l0 lives summarized in the cohort life table. . GENERAL FEATURES OF DURATION DATA idealized presentation of the observed mortality in the form of a cohort life table. . Then the single integervalued age random variable [T ] for a new individual being insured can be explicitly constructed as a function of i ∈ Ω as follows: for k = 0.3. . . and all probabilities and expected values related to functions of a single lifetime or integerageatdeath random variable [T ]. Here the sample space and underlying probability is very easy to describe: Ω = {1. for all k = 0. . From the viewpoint of Probability Theory. Recalling that dk = lk − lk+1 in the cohort life table. . ω − 1. 2. random variables or data are formalized as measurement functions on the sample space Ω of all possible detailed outcomes of a survival experiment.1 Sample Space for a Single Newly Insured Life The simplest case is the one studied in the ﬁrst two chapters of this book. ω−1 we note that l0 = k=0 dk . . . 1. 1 ≤ i ≤ l0. with equal assigned probability P r({i}) = 1/l0 for each individual labelled i. . . 1. The underlying random experiment is to select an individual i equiprobably from the list of all all l0 individuals in the cohort ‘population’: that is. ω − 1. It is instructive to deﬁne the sample spaces needed at three levels of complexity of the probability and statistics of survival models. Then the event [T ] = k consists precisely of the subset k−1 k of indices i satisfying j=0 dj < i ≤ j=0 dj . and therefore has size dk and probability dk /l0 as desired. k−1 k [T ](i) = k if and only if j=0 dj < i ≤ j=0 dj (A. A. and the next dk individuals die at integer age k.1) The interpretation of this rule is that if we number the l0 individuals i in the cohort life table in order of the wholenumber age k at which they die. . in which mortality of a ﬁctitious population cohort is recorded. in this simpliﬁed model the lifetime [T ] of the newly tobeinsured individual is modelled as being the same as a randomly selected member of the cohort population. k−1 then the ﬁrst j=1 dj = l0 − lk individuals die at ages less than k. where only integer ages are recorded in the survival experiment.
.1 In the next subsection. but these topics are not treated further in this book. . . . possibly dependent. cases where each lifetime is additionally labelled by a cause of death L or where events deﬁned in terms of the dependent lifetimes T A . Then the actual observed failure age T is min(T1. lifetime random variables are modelled simultaneously. . . K. . There are cases of intermediate complexity not discussed in detail in this book. or with a smaller payment to the survivor — which revert to the surviving member of the pair when one member dies.1. Nevertheless. David and Moeschberger 1978) in the biostatistical literature. 2. we consider the sample space appropriate to a cohort of lives. for example in insurances of both husbands and wives or in annuities — possibly variable. or Gerber 1997 Ch. such as death from speciﬁed disease (as in ‘cancer insurance’) or accident or from other causes. . This topic is treated under the heading of contingent multilife functions (Jordan 1991 Part II.2 Sample Space for a Full Cohort Population As mentioned explicitly in Section A.A. assumed independent. 8). . . K} for which TL = T .2. TK ) and the random label L is the integer in {1.3. The feature of these intermediatecomplexity sample spaces is that a single vector (T1. T B of a pair of related lives (such as a husbandwife pair) have consequences for Insurance. there is an underlying random variable Tk . . TK ) of ﬁnitely many. Jordan 1991 Part II). giving the age at which the individual would have died from that cause if not earlier killed from another cause. T B ) or larger multilife groups of lifetimes is important in the calculation of insurance premiums and annuity or pension values for husbandwife pairs. The case of lifetimes (T. L) labelled by cause arises when for each of K distinct types of mortality. simultaneously following the same mortality rates summarized in a cohort life table. k = 1. This setup is called a competing risks model (see Gail 1975. . . the cohort population whose mortality is summarized in the cohort life table is generally a complete ﬁction. 2 A. like US Social Security.3. SAMPLE SPACES FOR DURATION DATA 167 Remark A. . and relates to multiple decrement (cohort) life tables in an Insurance context (Gerber 1997 Ch. . . 7. The joint modelling of pairs (T A . 2. the relative frequency ratios deﬁning the survival functions and .
but for a sequence of n independent lifetimes. since we have already argued in this Chapter that realistic mortality studies generally have a much more complicated and inconvenient pattern of staggered positive ages at entry and of loss to followup before death for many subjects under study. now the probabilities are deﬁned by the property that the lifetimes all follow survival function S(t) and are independent . is of particular use in Chapter 3 of this book. In particular. Yet it is immediately apparent that this realism comes at a price of greater mathematical complexity. with the vectorvalued mapping given by the identity + mapping on Ω: T (s) = {Ti(s)}n = {si}n i=1 i=1 Unlike the situation in Section A. it is a little more realistic to treat the possible cohort lifetimes as a set of continuous random variables. which are independent across individuals and can each take values anywhere on the positive age axis.1 is inherently discrete. while the sample space described in Section A. The most natural space to use is Ω = Rn = [0.168 APPENDIX A. ∞)n . where the quality of deathrate estimates is studied and where the simulation of new cohort lifetables with speciﬁed survival functions S(t) is described. The sample space itself must be a set of detailed outcomes not only for a single continuously distributed lifetime. The sample space described in this Section allows us to consider the intrinsic variability of the estimated death rates qk and other statistics derived as a function of observed ageatdeath random variables if those variable values were lifetime lengths of individuals under followup for their entire lives. we can view them as the realized values of a set of independent identically distributed lifetime random variables representating a realistic underlying mechanism of mortality. rather than viewing the cohort data as ﬁxed.2. whose sample sizes we now deﬁne. 1 ≤ i ≤ n.1.3. GENERAL FEATURES OF DURATION DATA death rates which are derived from the life table and used in calculated insurance premiums could also be viewed as statistical estimators of unknown statistical parameters based on a set of n independent identically distributed lifetime random variables Ti . The greater realism of cohorttype survival experiments. This aspect of the random mortality experiment is still an artiﬁcial idealization. That is. where the randomlifetime mapping was deﬁned in such a way that the probabilities associated with individual outcomes were equiprobable.
the joint probability density of (Ei . David and Moeschberger 1978) for discussion of such matters. SAMPLE SPACES FOR DURATION DATA of one another. bn ] = i=1 (S(ai)−S(bi )) Probabilities of other events concerning the randomvariable components Ti(s) are implicitly determined from this deﬁnition on ndimensional recann gles i=1 (ai . Ai . b2]×· · ·×(an . a speciﬁcation accomplished by the deﬁnition Pr Pr s ∈ Rn : n 169 a1 < s1 ≤ b1. Ai .3. an < sn ≤ bn n = s ∈ R : T (s) ∈ (a1. However.3. bi ] by means of the probability axioms (ﬁnite or countable additivity). Further details of the unique speciﬁcation of probability laws from a generating collection of open sets can be found in more advanced treatments of Probability Theory. Ai . 1 ≤ i ≤ n) described in Section A. The usual assumption of independence of (Ei . . 1 ≤ i ≤ n. only in Chapter 8 do we address a simpliﬁed although typical setting (‘independent death and censoring’) to introduce maximum likelihood estimators of survival in models with piecewise constant hazards and KaplanMeier estimators in models with general (‘nonparametric’) hazards. 1 ≤ i ≤ n. since a very large class of events can be generated by limits of increasing unions and decreasing intersections of unions of such rectangles. Ti . In this book. We refer to texts on Survival Analysis (Cox and Oakes 1994.1 requires a Cartesian 3n product of R+ whose coordinates model the values factors of all of the random variables Ei . Klein and Moeschberger 2003. Di . Di ) across diﬀerent individuals i is embodied in a deﬁnition of Probability on Ω as a socalled ‘product probability’ across n spaces R3 × {0. 1}. b1]×(a2. . such as Ross (2005) or Billingsley (1995). 1}n to model the values Di . can + have all sorts of diﬀerent realistic dependence structures. . . Ai. Ti. Ti. Di ). Ti.A. along with a further space {0. . a2 < s2 ≤ b2 .3 Sample Space for the Realistic Mortality Study The Sample Space Ω needed to accomodate the detailed outcome data (Ei . A.
GENERAL FEATURES OF DURATION DATA .170 APPENDIX A.
SpringerVerlag. R. H. 209222. The Theory of Competing Risks. and Moeschberger. Life Contingencies. H. and Prentice. 2nd ed. Gerber. Calculus. H. V. J. Society of Actuaries. Paciﬁc Grove 2007 [6] Ellis. New York. Griﬃn. Duxbury. Probability and Statistical Inference. and Oakes.. R. P. M. D. Huntington. London 1984 [4] David. Analysis of Survival Data. W. and Nesbitt. J. New York 2002 171 . 1967.. Actuarial Mathematics. Probability and Statistics for Engineering and the Sciences. Jones.. M. paperback reprinted 1991 [11] Kalbﬂeisch. Probability Theory and Measure. D. C.Bibliography [1] Billingley. 1978 [5] Devore. 3rd ed. 7th ed. 1995 [2] Bowers. and Tanis. J. 3rd ed. Chicago. 2005 [10] Jordan. E. R. Life Insurance Mathematics. WileyInterscience. New York. 2002 [7] Gail. 1997 [9] Hogg. Addison Wesley / PrenticeHall. Biometrics 31. C. Hickman. and Gulick. 2nd ed. Society of Actuaries. Chapman and Hall. New York. N. WV. IL 1997 [3] Cox. Custom Publishing. Wiley. The Statistical Analysis of Failure Time Data. Itasca. D. 7th ed. R. [8] Gerber. D. 6th ed. (1975) A review and critique of some models used in competing risk analysis. 2nd ed.
S. New York 1992 [15] The R Development Core Team. Revised ed. 2nd ed. 1968 [19] Venables. Modern Applied Statistics with S. 3rd ed. 7th ed. Appl. Prob. Irwin/ McGraw Hill. New York. SpringerVerlag. C. 2008 [13] Klein. W. 725741. E. M. Statistical Models for Survival Data Analysis. (1989) Moderate and largedeviation actuarial probabilities in actuarial risk theory. Survival Analysis: Techniques for Censored and Truncated Data. Chicago. New York. [18] Spiegelman. and Moeschberger. A First Course in Probability. The Theory of Interest. S.172 BIBLIOGRAPHY [12] Kellison. T. of Chicago. R: a Language and Environment. E. and Ripley. John Wiley. Introduction to Demography. 21. J. Springer. 2002 . M. B. 2003 [14] Lee. Adv. [16] Ross. 2nd ed. Univ. Homewood. and Hoesman. Ill. 4th ed. Prentice Hall 2005 [17] Slud.
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.