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

# Insurance: Mathematics and Economics 46 (2010) 242–253

**Contents lists available at ScienceDirect
**

Insurance: Mathematics and Economics

journal homepage: www.elsevier.com/locate/ime

Mortality risk modeling: Applications to insurance securitization

$

Samuel H. Cox

a,1

, Yijia Lin

b,∗

, Hal Pedersen

a,2

a

University of Manitoba, Winnipeg, Manitoba, R3T 5V4, Canada

b

University of Nebraska, P.O. Box 880488, Lincoln, NE 68588, USA

a r t i c l e i n f o

Article history:

Received December 2008

Received in revised form

September 2009

Accepted 30 September 2009

a b s t r a c t

This paper proposes a stochastic mortality model featuring both permanent longevity jump and tempo-

rary mortality jump processes. A trend reduction component describes unexpected mortality improve-

ment over an extended period of time. The model also captures the uneven effect of mortality events on

different ages and the correlations among them. The model will be useful in analyzing future mortality de-

pendent cash flows of life insurance portfolios, annuity portfolios, and portfolios of mortality derivatives.

We show how to apply the model to analyze and price a longevity security.

©2009 Elsevier B.V. All rights reserved.

1. Introduction

Over the past half century, and especially in the most re-

cent decades, remarkable mortality improvements have led to the

growthof the populationof older people (BourguignonandMorris-

son, 2002; Lakdawalla and Philipson, 2002; Vaupel, 1998). To the

extent that this progress is unexpected, it has a negative impact

on pension plans and annuity providers. In the US, private defined

benefit pension plans currently have close to $6 trillion in liabilities

for future benefits. In addition, US life insurers hold approximately

$2 trillion in annuity reserves (Salou and Hu, 2006; ACLI, 2006).

Uncertainty of longevity improvements increases risk for pension

funds and annuity insurers, since annuity benefits may need to be

paid longer than expected. In a recent study of pension liabilities

of the companies in the UK’s FTSE100 index, Cowling and Dales

(2008) found overly optimistic longevity assumptions for pension

valuations reported at the end of 2007. These authors believe that

companies underestimate future life expectancy by one to three

years and, therefore, understate the aggregate pension deficit of

these companies by as much £40 billion.

$

The paper was presented at the 4th International Longevity Risk and Capital

Market Solutions Symposium in 2008, the 2008 American Risk and Insurance

Association annual meeting, and the 2008 Financial Management Association

annual meeting. We appreciate the helpful comments andsuggestions fromRichard

MacMinn and participants at these meetings. We are especially grateful to Enrico

Biffis and two anonymous referees for their very useful comments.

∗

Corresponding address: Department of Finance, University of Nebraska, P.O.

Box 880488, Lincoln, NE 68588, USA. Tel.: +1 402 472 0093.

E-mail addresses: sam_cox@umanitoba.ca (S.H. Cox), ylin@unlnotes.unl.edu

(Y. Lin), pedersn@cc.umanitoba.ca (H. Pedersen).

1

Tel.: +1 204 474 7426.

2

Tel.: +1 204 474 9529.

In contrast, in modeling life insurance, one naturally looks at

scenarios which are pessimistic about mortality improvement and

often include threats such as epidemics. Several recent articles

focus on management of such mortality risks, including Hardy

(2005), Rogers (2002), Goss et al. (1998), Cox et al. (2006), Cox

and Lin (2007) and Lin and Cox (2008). Genetic analysis has

confirmed that the virus of ‘‘Spanish flu’’ which killed 40 to 100

million people in 1918 developed in birds and was similar to

today’s ‘‘bird flu’’ (Juckett, 2006). Some public health experts think

that a pandemic is overdue and another will inevitably occur

(Dowdle, 2006). Should a pandemic occur, a life insurer will suffer

financially since it will pay more death benefits than expected

when the policies were issued. It seems reasonable that a major

pandemic event could trigger turbulence in the life insurance

industry. Toole (2007) studied this issue in detail and concluded

that the industry as a whole can withstand a severe pandemic,

as severe as the 1918 pandemic, with a loss of about $64 billion

relative to aggregate risk-based capital (RBC) of about $256 billion

in 2005. While the industry as a whole can sustain a severe flu

pandemic, it would nevertheless be disruptive. Companies holding

less than 100 percent of RBC (only 14 companies in 2005) may

become insolvent. As Toole (2007) notes, if a severe pandemic

were to occur when the financial markets are weak (e.g. when

asset values such as mortgage-backed securities are depressed),

then the financial impact could be much worse and the number

of companies near insolvency could be much greater. Thus, it

seems clear that including pandemic effects is an important issue

in modeling mortality for life insurance liabilities.

In light of the above discussion, the terms mortality risk and

longevity risk are opposites in this paper. Mortality risk is the risk of

more deaths than expected, or the risk that observed death prob-

abilities are higher than expected. We usually think that severe,

short-termed events such as pandemics underlie mortality risk.

0167-6687/$ – see front matter ©2009 Elsevier B.V. All rights reserved.

doi:10.1016/j.insmatheco.2009.09.012

S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253 243

On the other hand, more people may survive than expected, or ob-

served death rates may be lower than expected. Unanticipated im-

provements in medicine and health technology may give rise to

longevity risk, leading to a surge in mortality improvement in a

short period (e.g. one or two years) or excess improvement over a

long period (e.g. more than ten years). For longevity events with

a long-term effect, the slope of the mortality curve may deviate

from that of the base trend with normal deviations, as evident, for

example, in the US population mortality dynamics in the 1970’s.

Actual mortality has been improving, so parametric models esti-

mated with actual data will reflect improving mortality. Models

that have a randomdeviation fromthe expected mortality may re-

flect some longevity risk, but this is not a reflection of a fundamen-

tal change in the trend (as we have in mind). Our model, for which

we will provide a detailed discussion later, allows for such funda-

mental changes.

Marocco and Pitacco (1998), Milevsky and Promislow (2001),

Dahl (2004), Miltersen and Persson (2005), Cairns et al. (2006),

Dahl and Møller (2006), Gründl et al. (2006), Ballotta and Haber-

man (2006), Biffis and Millossovich (2006) and Bauer and Kramer

(2007) focus on mortality risk as deviations from a trend. In those

papers, the base trend may reflect mortality improvements, but

longevity risk is not modeled explicitly. On the other hand, pen-

sion and annuity research has focused on longevity risk. Loise and

Serant (2007) model longevity (and mortality) risk use a stationary

Gaussian process. However, this model may not be appropriate for

modeling longevity (and mortality) shocks which may not followa

stationary Gaussian process. To describe the stochastic longevity

trend, Hári et al. (2008) extend the Lee–Carter model with a

time-varying, stochastic drift. Biffis (2005) captures mortality ran-

dom departures around a time-varying target with a longevity

compound Poisson process, but that model cannot guarantee a

nonnegative force of mortality.

US mortality data has only two extreme events, 1918 flu pan-

demic and the change in the rate of improvement of mortality

around1970. However, experts conjecture that we may experience

extreme events of both types in the future. This means that, in in-

vestigating the impact of possible future events, we cannot sim-

ply calibrate our models to the experience. Therefore, we offer a

flexible, although somewhat complex, method of including expert

opinion in forming future mortality scenarios.

Specifically, we propose a newapproach by introducing a trend

reduction jump component to describe longevity risk. Unexpected

longevity improvement, in general, may be less dramatic than that

of a mortality death shock, but in the long run longevity risk may

be just as important. Most longevity risk events in the past seem

to have a pattern; unexpected survival gains often extended over

a long period of time, leading to a steeper downward sloping force

of mortality curve. The traditional one-time jump models such

as the model that combines a geometric Brownian process and

a compound Poisson jump process, cannot provide this kind of

longevity risk.

Moreover, to obtain relevant results, a stochastic mortality

model must reflect three major features of the current mortality

universe: (a) both mortality improvement and deterioration jump

factors, (b) correlation among different ages and over time, and

(c) uneven effects of a mortality jumps across different ages. Our

model attempts to address these issues. In particular, we are

explicitly combining both mortality and longevity risk in a single,

comprehensive model inorder to make a more realistic assessment

of future survivor dependent cash flows. Finally, we show how to

determine a parsimonious version with historical data.

As an application to mortality risk management, Blake et al.

(2006) examine a wide variety of longevity bonds. We continue in

the same vein, but with explicit mortality derivatives such as op-

tions which we think can be important tools in managing mortality

and longevity risk. Derivatives can be written on indices based on

publicly available data like the LifeMetrics index (Coughlan et al.,

2007) offered by Goldman–Sachs

3

and the Credit Suisse index.

4

Our paper is organized as follows. In Section 2 we describe

mortality dynamics as a combination of a Lee–Carter diffusion

process, a permanent longevity jump process and a temporary

mortality jump process. In Section 3 we discuss capital market so-

lutions for mortality and longevity risk, including new structures,

with our proposed longevity index, respectively. We show how

to price those securities with the indifference pricing method in

Section 4. We conclude the article in Section 5.

2. Dynamic mortality model

Stochastic mortality models usually start with the assumption

that there is an ‘‘initial’’ curve for the force of mortality. For a life

(x) at time 0, the remaining life time random variable is denoted

T(x). The force of mortality at time t > 0 is denoted µ(x + t, t).

In effect, the force acting during the life of a person age x at time

t = 0 is a stochastic process {µ(x +t, t) : t ≥ 0} evolving from an

initial curve µ(x, 0) = µ(x). Mathematically speaking, this is very

complex. Inorder to reduce the complexity, the model specifies the

force of mortality as a parametric function of age and time, with a

small number of parameters.

Statistical methods applied in actuarial practice use this same

approach. The expected value, or forecast value, of the force of

mortality µ(x + t, t) acting on (x + t) at time t, given T(x) > t, is

used to price life insurance and annuity polices. However, actual

mortality experience can change in a nondeterministic way and

deviate from expected mortality rates.

Here is howwe will model these events. The aggregate effect on

(x) over the period (0, h) is a random variable since it is the path

integral over the sample path {µ(x +t, t) : 0 ≤ s ≤ h} of the force

of mortality:

_

h

0

µ(x + t, t)dt.

The probability S(x, h) that (x) will attain age x + h is random

at time 0, because it depends on the sample path of the force of

mortality:

S(x, h) = exp

_

−

_

h

0

µ(x + s, s)ds

_

.

The probability that (x) will attain age x + h is the expected value

of the probability:

Pr [T(x) > h] = E [S(x, h)] .

Among the reasons for changes in the force of mortality are the

following:

• general trends in health and mortality resulting in gradual

improvements to mortality,

• deviations from the general trends,

• relatively rare advances in medical technology resulting in

sudden and permanent improvements in mortality, and

• sudden catastrophic surges in mortality from events such as

pandemics resulting in temporary increases in mortality.

3

http://www.qxx-index.com.

4

http://www.credit-suisse.com/ib/en/fixed_income/longevity_index.html.

244 S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253

Table 1

The estimates of a(x) and b(x) based on the US male population data from 1901 to 2005.

Age a(x) b(x) Age a(x) b(x) Age a(x) b(x)

0 −3.4039 0.1736 35 −5.5990 0.1061 70 −3.0020 0.0392

1 −5.6955 0.2612 36 −5.5442 0.1043 71 −2.9207 0.0398

2 −6.2000 0.2374 37 −5.4851 0.1017 72 −2.8391 0.0400

3 −6.4963 0.2287 38 −5.4259 0.0990 73 −2.7575 0.0396

4 −6.7038 0.2269 39 −5.3661 0.0966 74 −2.6766 0.0393

5 −6.8571 0.2171 40 −5.3015 0.0934 75 −2.5950 0.0384

6 −6.9863 0.2029 41 −5.2354 0.0903 76 −2.5132 0.0378

7 −7.1111 0.1908 42 −5.1666 0.0872 77 −2.4290 0.0377

8 −7.2240 0.1820 43 −5.0960 0.0841 78 −2.3429 0.0381

9 −7.3290 0.1820 44 −5.0242 0.0812 79 −2.2492 0.0371

10 −7.3986 0.1878 45 −4.9509 0.0784 80 −2.1687 0.0391

11 −7.3761 0.1883 46 −4.8750 0.0752 81 −2.0834 0.0391

12 −7.2273 0.1737 47 −4.7995 0.0721 82 −2.0022 0.0384

13 −6.9989 0.1495 48 −4.7227 0.0687 83 −1.9251 0.0370

14 −6.7715 0.1278 49 −4.6471 0.0655 84 −1.8509 0.0352

15 −6.5665 0.1113 50 −4.5699 0.0621 85 −1.7774 0.0335

16 −6.3902 0.0992 51 −4.4928 0.0593 86 −1.7025 0.0319

17 −6.2527 0.0934 52 −4.4151 0.0569 87 −1.6262 0.0309

18 −6.1422 0.0903 53 −4.3380 0.0554 88 −1.5482 0.0303

19 −6.0618 0.0911 54 −4.2612 0.0544 89 −1.4695 0.0298

20 −5.9908 0.0944 55 −4.1827 0.0532 90 −1.3906 0.0295

21 −5.9233 0.0959 56 −4.1041 0.0521 91 −1.3116 0.0290

22 −5.8849 0.0986 57 −4.0249 0.0506 92 −1.2336 0.0282

23 −5.8732 0.1006 58 −3.9452 0.0488 93 −1.1572 0.0274

24 −5.8802 0.1021 59 −3.8663 0.0471 94 −1.0828 0.0265

25 −5.8953 0.1039 60 −3.7859 0.0453 95 −1.0142 0.0269

26 −5.9028 0.1050 61 −3.7063 0.0440 96 −0.9457 0.0265

27 −5.9019 0.1067 62 −3.6267 0.0423 97 −0.8779 0.0265

28 −5.8855 0.1082 63 −3.5503 0.0410 98 −0.8122 0.0267

29 −5.8537 0.1092 64 −3.4739 0.0395 99 −0.7502 0.0268

30 −5.8191 0.1101 65 −3.3970 0.0383 100 −0.6867 0.0270

31 −5.7830 0.1100 66 −3.3189 0.0375 101 −0.6228 0.0277

32 −5.7435 0.1099 67 −3.2396 0.0370 102 −0.5587 0.0262

33 −5.6993 0.1091 68 −3.1614 0.0376 103 −0.5230 0.0222

34 −5.6519 0.1079 69 −3.0822 0.0383

Therefore, it is natural to develop a model for mortality changes

based on the evolution of the force of mortality with a general

trend, diffusion process and jump components. Anumber of recent

studies have sought to model mortality trends involving both

age-dependent and time-dependent terms (Lee and Carter, 1992;

Renshaw et al., 1996; Lee, 2000; Sithole et al., 2000; Milevsky and

Promislow, 2001; Olivieri and Pitacco, 2002; Dahl, 2004; Cairns

et al., 2006). Our model is inspired by, and closely related to, this

stream of research. As an extension, we explicitly model both

mortality and longevity jump processes in a single model (See

Section 2.2).

2.1. Modeling general mortality trends

We adopt the well-known Lee and Carter (1992) mortality

model as a basis. This model captures the evolution of mortality

in mutually exclusive age cohorts, while at the same time a time-

series common risk factor, k(t), links all cohorts together. The force

of mortality µ

LC

(x, t) is modeled as

5

µ

LC

(x, t) = exp [a(x) + b(x)k(t)] . (2.1)

The parameters a(x) and b(x) are age-specific while k(t) is time

varying.

We fit the model to the US male data for ages x = 0, 1, . . . , 103

and times t = 1901, 1902, . . . , 2005, using the technique

5

Lee and Carter use the central death rate rather than the force of mortality. For

reasonable assumptions about the distribution of deaths between age x and x + 1,

the two measures are the same or very close. For example, if the force of mortality

is constant between integral ages, the central death rate at age x and the force of

mortality at age x are equal. The database has values of q(x, t). Under the same

assumption, q(x, t) = 1 − e

−µ(x,t)

so we can easily change from q(x, t) to µ(x, t).

suggested by Lee and Carter (1992). Tables for years 1901 to 1999

are from the Human Life Table Database.

6

Tables for 2000 to

2005 are from the Human Mortality Database.

7

The estimates of

a(x) and b(x) are in Table 1 and the estimated k(t) is shown in

Fig. 1. There are several generalizations of Lee and Carter (1992)

model in the literature. We believe that our ideas illustrated below

could be applied to them as well by using the original Lee–Carter

model.

Following Lee and Carter (1992), the mortality index k(t)

evolves as

k(t + 1) = k(t) + g

1

+ g

2

× Flu + σz(t),

t = 1901, 1902, . . . , 2005 (2.2)

where g

1

, g

2

, and σ are constants and z(1901), z(1902), . . . , z

(2005) are independent standard normal randomvariables. The flu

in 1918 is identified by a dummy variable, Flu, in Eq. (2.2). Specif-

ically, Flu = 1 if t = 1918, and 0 otherwise. Then we obtain

the estimated g

1

= −0.2244, g

2

= 2.2065, and σ = 0.6123.

Simulating mortality scenarios with the Lee–Carter model, in-

volves simulating future mortality index values k(t) for t > 2005.

For convenience we now re-label the years so that year 2005 cor-

responds to t = 0. For a person age x in 2005, we can generate

6

Data source: Human Life Table Database. Max Planck Institute for Demographic

Research (Germany), University of California, Berkeley (USA) and the Institut

national d’études démographiques (France). Available at www.lifetable.de (data

downloaded on June 8, 2008).

7

Data source: Human Mortality Database. University of California, Berkeley

(USA), and Max Planck Institute for Demographic Research (Germany). Available at

www.mortality.org or www.humanmortality.de(data downloadedonJune 8, 2008).

S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253 245

1901 1911 1921 1931 1941 1951 1961 1971 1981 1991 2001

-15.00

-10.00

-5.00

0.00

5.00

10.00

Fig. 1. Estimated time-series common risk factor k(t) shown in the vertical axis for

year t = 1901, 1902, . . . , 2004, 2005 in the horizontal axis.

Lee–Carter model scenarios for future values of the force of mor-

tality as follows: For x = 0, 1, . . . , 103,

µ

LC

(x, 0) = µ(x) observed in 2005

µ

LC

(x, t) = exp [a(x) + b(x)k(t)] t > 0.

(2.3)

When we use the model to simulate future values of the force

of mortality applying to a life age x at t = 0 over a period of T years

µ

LC

(x, 0), µ

LC

(x + 1, 1), µ

LC

(x + 2, t + 2), . . . , µ

LC

(x + T, T),

we need to provide for the cases for which x+j > 103 and j > 0. In

other words, the way we use the model requires extending Eq. (2.3)

to ages x > 103 after the first year. A natural approach is to simply

extend both functions a(x) and b(x) linearly using the values for

ages 90to 103 as a basis for the extrapolation. This approachresults

in the linear extrapolation of a(x) and b(x),

a(x) = −7.43 + 0.067x

b(x) = 0.055 − 0.0003x

extending Eq. (2.3) to x > 103 for t > 0. This is consistent

with models with linearly changing mortality, as described by

Schoen (2006). In applications for which ‘‘oldest-of-old mortality’’

is critical, other approaches might be necessary.

2.2. Modeling mortality jumps

Our goal is a stochastic mortality model suitable for dynamic

financial analysis, taking into account the complexity of observed

mortality dynamics. We consider two types of mortality shocks:

permanent jump G(x, t) and temporary jump H(x, t).

The unexpected mortality improvement that results from ge-

netic, environmental, behavioral, bio-reliability, and/or hetero-

geneity forces and constraints, often has a long-term effect on

future mortality rates. In contrast, many catastrophe death events,

like the 1918 worldwide flu and the 2008 earthquake in China,

have a more severe but transitory effect on death rates. In many

cases, these short-term events drive up the death rate just for a

couple of years, but the mortality falls back to the normal level

afterwards. Accordingly, we view sharp decreases in mortality as

‘‘permanent’’ events, for instance, associated with aforementioned

medical advances, and view dramatic increases as ‘‘temporary’’

events such as pandemics.

Whenwe addthese two types of mortality jumpevents – G(x, t)

and H(x, t) – to the Lee–Carter force of mortality µ

LC

(x, t), the

initial version of our model takes the form

µ(x, t) = µ

LC

(x, t) × exp (−G(x, t) + H(x, t)) , (2.4)

where µ

LC

(x, t) is the force of mortality fromthe Lee–Carter model

(2.3).

Fig. 2. Medical advancement by age.

2.2.1. Permanent longevity jump G(x, t)

The unexpected mortality improvement process G(x, t) con-

sists of a jump reduction component K(x, t) and a trend reduction

component D(x, t). That is,

G(x, t) = K(x, t) + D(x, t).

The jump reduction component K(x, t) is the one-time long-

evity jump induced by a surge in survival rates over a short period

(for example, a year) with a permanent effect on longevity. It is de-

fined as

K(x, t) =

∞

s=1

y

s

A

s

(x)1

{t≥η

s

}

, (2.5)

where η

s

is the time of jump reduction event s. The time η

s

can

be modeled as the arrival time of the s-th event of a point process,

as described in Daley and Vere-Jones (2003). The simplest version

would be the arrival times of a Poisson process, in which case the

times would have gamma distributions. A second feasible choice

would be a Hawkes process. The primary difference between these

two approaches lies in the ability of a Hawkes process to create

event clustering.

The positive variable y

s

is the maximummortality improvement

of all ages in medical advancement event s. The effect of y

s

is trans-

ferred to the mortality rates through a function A

s

(x) for age x with

0 ≤ A

s

(x) ≤ 1. When A

s

(x) = 0, medical advancement s has no ef-

fect on age x. In contrast, A

s

(x) = 1 means age x enjoys the biggest

mortality improvement (y

s

) among all ages. When 0 < A

s

(x) <

1, it captures the effects that fall between the above two ex-

tremes. That is, the functionA

s

(x) differentially spreads the mortal-

ity improvement across x’s because, when a medical advancement

occurs, the benefits of the technology usually vary by age. For ex-

ample, compared to demography changes of the whole population

in the 1970’s, the annual improvement in mortality of old ages in

the same period is much more dramatic than that of young ages.

In fact, Cutler and Richardson (1998) find that improvement for

the elderly was greater than for the young in the 1970’s as a re-

sult of a decrease in cardiovascular disease deaths. Since cardio-

vascular disease is more prominent later in life than earlier, the

life expectancy gain is greater for the elderly than for the young.

Mortality improvements due to a change in a cause of death may

vary by age.

There is no definitive data set to say what the function A

s

(x)

should look like. Practitioners will use a conservative choice, de-

pending on the products being modeled. If you think the shape

shown in Fig. 2 may be more realistic in the future, you must be-

lieve that there is little mortality improvement for the very young

and very old.

Furthermore, the annual mortality rate over a long period

(e.g. twenty years) may continue to improve at a much higher rate

246 S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253

0.030

0.040

0.050

0.060

0.070

0.080

0.090

0.100

0.110

1901 1921 1941 1961 1981 2001

Fig. 3. Actual force of mortality µ(75, t) for age 75 in year t. The vertical axis stands

for force of mortality and the horizontal axis represents year.

than that in other periods, leading to a steeper downward-sloping

force of mortality curve. For example, the annual percentage de-

crease in death rates for the US population aged 55 and above in

the 1970’s, on average, is about three times that of the same age

cohort in other periods. This type of mortality improvement, in a

certain sense, is more significant than the one-time longevity jump

K(x, t). The data shows no one-time longevity surge across ages

(like a 20 percent decrease in µ(x, t) during one year), but that

does not mean there was no longevity risk. The cumulative effect of

unanticipated mortality improvement over an unexpectedly long

period of time obviously has a devastating effect on the financial

stability of pensions plans and annuity markets.

Therefore, we model this long-term decay in mortality rates

by introducing a trend reduction component D(x, t) defined as

follows:

D(x, t) =

∞

i=1

ζ

i

(t − υ

i

)F

i

(x) exp(−ξ

i

(t − υ

i

))1

{t≥υ

i

}

, (2.6)

where υ

i

is the time of trend reduction event i. The time of the

trend reduction event υ

i

can be modeled like the time η

s

, which

we discussed after Eq. (2.5). The percentage change in µ(x, t) is

the log µ-ratio, log

µ(x,t+1)

µ(x,t)

. When a trend reduction event occurs,

we calculate the difference between its average percentage change

in the trend reduction period and its average level in the whole

observation period. In Eq. (2.6), ζ

i

represents the maximum excess

change across ages given ζ

i

> 0. For instance, consider male

mortality in the 1970’s: age x = 0 had the biggest annual excess

longevity gain, 0.0211 above its average improvement from 1901

to 2005, so ζ

i

= 0.0211.

Like A

i

(x), the function F

i

(x), with 0 ≤ F

i

(x) ≤ 1, distributes age

effects. Given ζ

i

= 0.0211, we have F

i

(0) = 1 in the 1970’s. Based

on the historical data, we can obtain other F

i

(x)’s. For example,

F

i

(75) = 0.3723 for age 75. This means µ(75, t) decreases by

an excess 0.0211 × 0.3723 × 100 = 0.79% per year during this

trend reduction period. Moreover, the factor (t − υ

i

) provides

cumulative mortality improvement as t increases beyond the jump

event time υ

i

.

How long the aforementioned mortality trend reduction will

last in the future is a topic for debate: ten years, twenty years or

even longer? The parameter ξ

i

> 0 specifies the length of trend

reduction event i. Should one believe that history will reflect the

future, when ζ

i

= 0.0211 and F

i

(75) = 0.3723, for example, one

might choose ξ

i

= 0.01. In this case, when an unexpected medical

advancement causes an annual excess decrease in µ(75, t), the

force of mortality curve for age 75 will have a notable steeper

downward slope for about 35–40 years, the pattern similar to the

curve for age 75 after year 1970 in Fig. 3.

Fig. 4. Pandemic effect by age.

2.2.2. Temporary adverse mortality jump H(x, t)

The transitory mortality jump process H(x, t) is defined as

follows:

H(x, t) =

∞

j=1

b

j

B

j

(x) exp(−κ

j

(t − τ

j

))1

{t≥τ

j

}

, (2.7)

where τ

j

is the time of adverse mortality event j. The time of the

adverse mortality event can be modeled like the time η

s

, which we

discussed after Eq. (2.5). The basic effect of a pandemic is modeled

using a Poisson process with a jump size b

j

> 0, where b

j

is

the maximum severity of jump event j. The function B

j

(x), with

0 ≤ B

j

(x) ≤ 1, distributes mortality jump impact across ages. B

j

(x)

can be random to reflect various age effects for different types of

events or it canalsobe deterministic. For example, the specification

of B

j

(x) in Fig. 4 would apply to a pandemic for which the bulk of

the mortality spike is under a certain age (for example, under age

50, as is the case for deaths in the 1918 worldwide flu).

Since modeling the transitory nature of mortality jumps is

important in practice, we introduce a nonnegative deterministic

function exp(−κ

j

(t −τ

j

)) in Eq. (2.7) with κ

j

> 0. The higher κ

j

the

faster the jump effect will die out.

8

We can estimate κ

j

to fit the

historical data and make the mortality jump only have the impact

over a reasonable short period of time.

2.3. Comment on estimation issues

The primary motivation for the model that we have presented is

to provide a flexible and realistic model that can be used to capture

adverse mortality and longevity risks in a business application. In

order to allow for these effects in a manner that can be compared

to the historical data, a fair number of parameters need to be

included. This poses significant estimation challenges. Our method

of estimation/calibration can best be described as follows:

(1) Estimate the original Lee–Carter model based on the available

data;

(2) Calibrate the A, B and F functions to approximate past history.

If the primary focus is on pandemics, one would look at how

the mortality effects were distributed across ages and calibrate

accordingly;

(3) Calibrate the jump frequency for the point processes driving

mortality changes in a range that is consistent with history. For

example, one can use the historical frequency of pandemics as

a guide to setting the jump frequency.

8

In general, κ

j

is much larger than ξ

i

because of the temporary effect of a

catastrophe mortality event and the long-term impact of unexpected mortality

improvement.

S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253 247

0.2

0.4

0.6

0.8

1.0

20 40 60 80 100

0.2

0.4

0.6

0.8

1.0

20 40 60 80 100

Fig. 5. The left figure is the function F(x), the normalized annual excess percentage decrease in µ(x, t) for different ages in the 1970’s. The right figure is the function B(x),

giving the impact of the 1918 worldwide flu on µ(x, t) across ages.

In financial and economic applications of jump-diffusion type

models, one issue that emerges is how to appropriately disentan-

gle the diffusion from the jump dynamics. The problem is acute in

the modelling of stock market returns because one generally intro-

duces jumps into the model with the idea of picking up ‘‘crashes’’

suchas 1929, 1987or 2008, but what canhappenis that the estima-

tion of such a model produces a preponderance of high frequency

lowseverity jumps rather than the intuitively pleasing outcome of

fairly smooth diffusion behavior punctuated by occasional severe

jumps. In our model, we are focused on mortality rather than eq-

uity returns but the same question can be asked. While we do not

have an estimation procedure that simultaneously sorts out the

diffusive and jump parameters of our model, we do feel that the

infrequent occurrence of severe adverse mortality events makes

our imposed reconciliation of the diffusion and adverse mortality

jumps reasonable. In the case of possible longevity jump compo-

nents, we are not able to reliably measure the extent to whichjump

behavior is present. Insofar as the stochastic features of the model

can be ascertained for a given parameterizations, the user of the

model can, at the very least, perform risk management functions

for a range of life insurance products in plausible mortality scenar-

ios.

In sum, while our model can generate a rich set of scenarios,

it does pose a serious estimation challenge. This is especially so

because of the relative paucity of extreme events in mortality data

sets. Indeed our goal is to allow for scenarios reflecting expert

judgment about these events, which may not be the same as the

data. For example, our approach allows one to set the frequency of

a severe 1918-like flu so that it is a 1 in 25 year event, even though

such a flu has been observed only once in 100 years. Therefore,

‘‘estimation’’ of this model requires a combination of traditional

estimation techniques and expert judgment.

2.4. Parsimonious model

For the Lee–Carter base model, substituting in Eq. (2.3) for

µ

LC

(x, t), Eq. (2.4) becomes

µ(x, t) = exp [a(x) + b(t)k(t) − G(x, t) + H(x, t)]

= exp [a(x) + b(t)k(t)] exp

_

−

∞

s=1

y

s

A

s

(x)1

{t≥η

s

}

_

× exp

_

−

∞

i=1

ζ

i

(t − υ

i

)F

i

(x) exp(−ξ

i

(t − υ

i

))1

{t≥υ

i

}

_

× exp

_

∞

j=1

b

j

B

j

(x) exp(−κ

j

(t − τ

j

))1

{t≥τ

j

}

_

. (2.8)

A permanent longevity jump can be driven by the one-time

component K(x, t) that substantially reduces µ(x, t) withina short

period of time, but it can also come from the trend reduction com-

ponent D(x, t) that accumulates the excess mortality improvement

over a long time. As discussed in Section 2.2.1, of these two com-

ponents, D(x, t) seems to play a more important role, especially

for ages above 65—the age range we are very interested in for

longevity risk. Accordingly, we provide a parsimonious version of

our model,

µ(x, t) = exp [a(x) + b(t)k(t)]

× exp

_

−

M(t)

i=1

ζ

i

(t − υ

i

)F

i

(x) exp(−ξ

i

(t − υ

i

))1

{t≥υ

i

}

_

× exp

_

N(t)

i=1

b

j

B

j

(x) exp(−κ

j

(t − τ

j

))1

{t≥τ

j

}

_

. (2.9)

In Eq. (2.9), the permanent longevity jump is only a function of the

trend reduction component. The process M counts mortality trend

reduction jumps; and M(t) is the number of such jumps by time t.

Similarly, the process N(t) counts the number of catastrophe death

events by time t. Evidently, Eq. (2.9) can be further simplified to

this:

µ(x, t) = exp [a(x) + b(x)k(t)]

× exp

_

−ζ F(x)

M(t)

i=1

(t − υ

i

) exp(−ξ(t − υ

i

))1

{t≥υ

i

}

_

× exp

_

bB(x)

N(t)

i=1

exp(−κ(t − τ

j

))1

{t≥τ

j

}

_

, (2.10)

assuming constant mortality and longevity jump effects for age x.

2.5. Example

Belowwe showhowto apply the parsimonious equation (2.10)

to model mortality dynamic process. All of our estimates are based

on the aforementioned US male population mortality tables from

1901 to 2005. As we noted earlier, we begin by using the data to

estimate the Lee–Carter model.

The importance of the impact of medical advancement on

pension plans and annuity insurers should not be underestimated.

Trend reduction factors in our model provide for such mortality

improvements. The 1970’s were the years of accelerating decline

of mortality for most ages, so our trend reduction parameters are

estimated from this period. The left graph of Fig. 5 shows our

estimated F(x) for different ages based on the data from 1970 to

1979. As mentioned in Section 2.2.1, age 0 had the highest average

annual excess mortality decrease rate ζ = 0.0211, so F(0) = 1.

However, we can observe mortality declines at high excess rates

mostly at the older ages. On the other hand, during the 1970s,

248 S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253

5 0 1 2 5 8 0 2 5 6 0 2 5 4 0 2 5 2 0 2 5 0 0 2

0 0 0 . 0

5 0 0 . 0

0 1 0 . 0

Fig. 6. These are simulated sample paths of the force of mortality, based on

Eq. (2.10) and the parameters described earlier. The top path is the simulated force

of mortality µ(60, t), the one just below it is µ(50, t), then µ(40, t) and µ(30, t).

mortality increased for some male young ages, perhaps due to an

increase in homicides, suicides, and accidents. If an age does not

have an excess mortality decrease rate, we set their F(x) = 0. Since

the annuity benefits in general are distributed mostly in the older

age ranges, their unexpected improvement in life expectancy will

have a great bearing on the overall impact of the longevity event.

Furthermore, based on the historical data, we specify the trend

reduction event arrival rate λ

ν

= 0.01 and the duration parameter

ξ = 0.01.

The 1918 worldwide flu is considered the most severe flu

epidemic. Naturally, it serves as the basis for the estimation of our

temporary adverse mortality jump parameters. The right graph of

Fig. 5 presents B(x), the normalized log µ-ratio in 1918 relative to

1917 for different ages. The 1918 fluincreasedthe one-year force of

mortality for different ages very unevenly. The flustruck ages 0–50,

especially ages 20–40, seriously, while there were no excess deaths

at the ages above 53. According to Taubenberger and Morens

(2006), these older people survived earlier flu epidemics and may

have acquired immunity to the flu virus. Among all ages, age 28 has

the highest log µ-ratio, which is the value for b. Specifically, b =

0.9647 and F(28) = 1. Since the 1918 flu is approximately a one-

in-one-hundred-year event, λ

τ

= 0.01. Moreover, this pandemic

only lasted a couple of years so κ = 1 is reasonable.

With the observed force of mortality for age x in year t = 2005

as the base case, sample paths of µ(x, t) for ages 30, 40, 50 and 60

in one simulation iteration are in Fig. 6. They highlight three major

features of our model: first, effects of both mortality improvement

and deterioration jump factors are clear. For example, the curve of

µ(40, t) has a steeper downward slope that starts in year t = 2027

and lasts for about 20 years (its trend reduction longevity jump

D(40, t) illustrated in the left graph of Fig. 7). Moreover, age 40

has a sharp increase in µ(40, t) in year t = 2014 but µ(40, t)

falls back to the normal level afterwards (its temporary jump factor

H(40, t) shown in the right graph of Fig. 7). Second, the movement

of different µ(x, t)’s are correlated among different ages and over

time. This is partially attributed to the functions F(x) and B(x).

Third, F(x) and B(x) also distribute uneven effects of a mortality

jump across different ages. Based on the historical data, in our

model, the adverse mortality jumps mainly strike younger ages

while the older ages enjoy more excess survival gains.

All parameters of the example shown in the above example

are estimated from the historical data. This is a problem that we

will want to carefully consider, as opinions vary widely regarding

future healthcare systemandpandemic. The shape of the mortality

curve often changes in different mortality or longevity jump

events. Toole (2007) concludes that the typical distribution of

excess death rates for seasonal influenza is U-shaped; that is,

‘‘excess deaths are heaped at age 0, quickly decrease to close to

zero until they start to increase again at older ages (typically age

65) with a rapid rise at ages 85 and older’’. However, the 1957

pandemic showed a spike in excess death rates at ages 65 and older

but there were no excess death at age 0 (Luk et al., 2001). Another

exception is the 1918 worldwide flu as mentioned earlier: there

were no excess deaths at ages above 65. It is important to note our

model can flexibly estimate the impact of a particular jump pattern

by adjusting F(x) and B(x) based on different opinions or updated

information.

The parsimonious model (2.10) does not include the jump re-

duction component K(x, t) since there is no clear historical evi-

dence to guide us on its pattern. As long as the estimates of y, A(x)

as well as the jump intensity are developed based on judgment, it

is not difficult to include K(x, t) in a richer model. An illustration

is provided in the next section.

2.6. A simple model

Although we envision the primary role of our model as a risk

management tool for the generation of a robust set of plausible

mortality scenarios, it is possible to calibrate the model to behave

in a manner that is qualitatively similar to the basic Lee–Carter

model. This can be done using jump events in longevity and

adverse mortality and a fixed base mortality rate, resulting in a

simple model of the form

µ(x, t) = µ(x, 0) exp (−K(x, t) + H(x, t)) . (2.11)

We calibrated this model by taking µ(x, 0) as the force of mortality

as of 2005 from the data that we have discussed above. We then

used the forces of mortality for the period 1965 through 2005

to estimate average rates of mortality improvement by age. We

calibrated the adverse mortality effects as was done inthe previous

section. This led to the functions A(x) and B(x) shown in Fig. 8.

One can use the improvement in mortality to identify aver-

age rates but it is not clear how one can identify the frequency

with which longevity events occur. For the purposes of this sim-

ple example, we use a Poisson process with intensity 0.50 to model

0 0 . 0

0 1 . 0

0 2 . 0

0 3 . 0

0 4 . 0

0 5 . 0

0 6 . 0

5 0 1 2 5 8 0 2 5 6 0 2 5 4 0 2 5 2 0 2 5 0 0 2

0 0 . 0

5 0 . 0

0 1 . 0

5 1 . 0

0 2 . 0

5 2 . 0

0 3 . 0

5 3 . 0

0 4 . 0

5 0 1 2 5 8 0 2 5 6 0 2 5 4 0 2 5 2 0 2 5 0 0 2

Fig. 7. This shows the trend reduction longevity jump D(40, t) for age 40 (left) based on Eq. (2.6) and temporary adverse mortality jump H(40, t) (right) based on Eq. (2.7).

These are the ones used in the simulated paths shown in Fig. 6.

S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253 249

0.6

0.8

1.0

0.0

0.2

0.4

0.2

0.4

0.6

0.8

1.0

0 20 40 60 80 100

20 40 60 80 100

Fig. 8. The functions A(x) and B(x) for a very simple calibration process. The left figure is the function A(x), the normalized annual average rates of mortality improvement

in µ(x, t) for different ages. The right figure is the the function B(x), giving the impact of the 1918 worldwide flu on µ(x, t) across ages.

0.000

0.005

0.010

2005 2025 2045 2065 2085 2105

Fig. 9. Evolution of mortality for a simple model, based on Eq. (2.11) and the

parameters described earlier. The top path is the simulated force of mortality

µ(60, t), the one just below it is µ(50, t), then µ(40, t) and µ(30, t).

longevity jumps and a jump size y = 0.06506. Although we know

what the product of these two values ought to be, we cannot iden-

tify them individually. Consequently, one must judge the parame-

ters by the qualitative features of the model.

Fig. 9 shows simulated mortality trajectories for the model

parameters that we have described. If one compares this with the

trajectories from the basic Lee–Carter model, they are similar in

their longevity behavior.

3. Securitization of mortality and longevity risks

3.1. Market development

Financial innovation has led to the development of several

classes of mortality securities. The Swiss Re bond, the first pure

death-risk linked deal, was issued in December 2003 (Stanley,

2003; Swiss Re, 2003; The Actuary, 2004). After successfully issuing

the first-ever pure death-linked security, Swiss Re sold two new

mortality bonds with different tranches in April and December

2006 (Lane, 2006). Following Swiss Re, some life insurers started

to reduce their extreme mortality exposures through financial

markets. For example, in May 2006 Scottish Re sold a mortality

bond with two tranches via a special purpose vehicle called Tartan

Capital Ltd., and in November 2006, AXAissued its first catastrophe

mortality deal—the Osiris bond (Lane and Beckwith, 2007). See

Bauer and Kramer (2007) for more details.

Capital market solutions for unanticipated longevity risk have

been explored relatively recently, first appearing in articles by

Blake and Burrows (2001), Milevsky and Promislow (2001), Lin

and Cox (2005) and others. Possibly inspired by the successful

securitization of catastrophe mortality risks, in November 2004,

the European Investment Bank (EIB) offered the first longevity

bond to provide a solution for pension plans to hedge their long-

term systematic longevity risks. Unlike the Swiss Re mortality

bond, the EIB longevity bond did not sell. The design of the EIB

bond may be problematic. The EIB bond provides ‘‘ground up’’

protection, covering the entire survival payment. Since the plan

can predict the number of survivors to some extent, especially

in the early contract years, a more satisfactory solution may be a

hedge using, for example, J.P. Morganq-forwards witha settlement

date 10 or more years in the future, rather than hedging the whole

benefit liability.

3.2. J.P. Morgan q-forwards

A q-forward contract requires, at maturity, an exchange of a

fixed amount based on the predetermined mortality rate, in return

for a variable amount based on a realized mortality rate published

at a specified future date by J.P. Morgan, in the LifeMetrics system.

9

Coughlan et al. (2007) describe an example with a 10-year q-

forward written on the one-year death rate q

65

for males aged x =

65 in England & Wales. See Table 2 for a summary. The example

provides a longevity hedge for the XYZ pensionplan. The q-forward

was opened on December 31, 2006 with a notional amount of

£50,000,000 and a fixed death rate q

65

= 1.2%. The floating-rate

payer XYZ pays an amount proportional to the mortality rate at

maturity on December 31, 2016, determined by the reference rate

in 2015.

At maturity, the contract is settled at the net amount, which is

the difference between the fixed amount and the floating amount.

For example, if the reference rate is only 1%, then XYZ Pension gets

£10,000,000 at settlement of the q-forward contract, calculated as

the net settlement amount:

Net settlement = Fixed amount − Floating amount

= 50,000,000 × 1.2% × 100 − 50,000,000 × 1% × 100

= £10,000,000.

Better mortality means fewer than 1.2% of those aged 65 in 2015

die within the year, and XYZ will have more than expected pension

benefits to pay if it has a similar experience. If the reference rate is

higher than the fixed rate 1.2%, the settlement is negative and XYZ

pays J.P. Morgan. In this case, mortality increases and it is likely

that the pension plan pays lower pension benefits.

Life insurers canuse q-forwards tohedge mortality risk ina sim-

ilar way. In summary, insurers or pension plans can take a position

9

Available at www.lifemetrics.com.

250 S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253

Table 2

An illustration for a q-forward to hedge longevity risk for a pension plan, described by Coughlan et al. (2007).

Notional amount GBP 50,000,000

Trade date December 31, 2006

Effective date December 31, 2006

Maturity date December 31, 2016

Reference year 2015

Fixed rate 1.2%

Fixed amount Payer J.P. Morgan

Fixed amount GBP 60,000,000 = 0.012 × 50,000,000

Reference rate LifeMetrics male q

65

in the reference year for England & Wales national population

Floating amount Payer XYZ

Floating amount Notional amount × Reference rate × 100

Settlement Net settlement = Fixed amount − Floating amount

(as fixed amount payer or floating amount payer) in q-forwards to

hedge their liability to pay benefits to their clients.

The q-forward contract could be modified to create q-calls

and q-puts. There are two practical problems with the q-forward

(or analogous options). When the q-forward is opened in 2005,

XYZ has to estimate the number of survivors to age 65 in 2015

in order to determine the size of the hedge. While XYZ may be

more concerned about rates 10 years in the future, they also have

longevity risk during the 10year contract periodbeginning in2005.

The example only hedges survivor risk from age 65 to age 66.

The other problem is that it is difficult to extend the q-forward

hedge. The example covers survivor risk only for one year. While

XYZ could opena portfolio of q-forwards on the indices q

64+k,2014+k

for k = 1, 2, . . . , it would also have to forecast the number of

survivors to each age in order to determine the size of each q-

forward contract. This gets more difficult as the hedge is extended

because of the problem of estimating the number of survivors,

exactly the risk the q-forward is supposed to hedge.

3.3. Securities based on a longevity index

Consider the XYZ plan again. As before, XYZ has a liability to

make payments to a plan’s participants who survive to age 66

in 2016. When the longevity security is opened in 2005, those

participants are age 55. XYZ knows the number of participants,

their gender, and projected benefits, so there is no problem in

determining the size of the hedge. Let us consider one participant

with 1 unit of annual benefit paid at age 66 in 2016.

In its funding calculations, XYZ will have an estimate of the

probability that a participant now age 55 survives to age 66,

denoted

11

p

xyz

55

. The index allows XYZ to hedge underestimating

this value (and underfunding its liabilities).

The index for this hedge is based on the government or Lifemet-

rics series of future death probabilities q

55,2005

, q

56,2006

, q

57,2007

,

. . . , q

65,2015

, published each year from 2006 to 2016. The first in-

dex value is p

55,2005

= 1 − q

55,2005

, published in 2006. The second

index value is

2

p

55,2005

=

_

1 − q

56,2006

_

p

55,2005

which is published in 2007. The subsequent index values are

calculated analogously. The index value for hedging the payment

in 2016 is

11

p

55,2005

=

_

1 − q

55,2005

_ _

1 − q

56,2006

_ _

1 − q

57,2007

_

· · ·

_

1 − q

65,2015

_

and it is based on the national population experience during the

hedge period from 2005 to 2015. XYZ can use the index directly

to hedge its payments to the cohort of plan participants age 55

in 2005. Forwards or options can be written on this index just as

easily as they might be written on q

65,2015

but with no problem

in determining the hedge size. Moreover, XYZ could just as easily

extend the hedge to the second benefit payment with an option on

the next index value

12

p

55,2005

.

For example, a call option could pay XYZ when the realized

population index

11

p

55,2005

, published in 2016, exceeds a strike rate

such as

11

p

xyz

55

(which the plan uses in 2005). The population index

is likely to be in the money in the same circumstances that the plan

has experienced unexpected increases in longevity, since the two

groups are subject to some of the same forces of mortality.

Using the population index, based on LifeMetrics or govern-

ment data, reduces moral hazard since the index is transparent

to all investors. However, basis risk may be a problem. Generally

pension plan participants have higher survival rates than the pop-

ulation as a whole. However, XYZ could estimate this relationship

statistically and use this relationship to adjust its hedge. We show

how to price longevity options in the next section.

4. Mortality securitization modeling

Insurance-linked securitization, as an alternative risk manage-

ment method and a newinvestment opportunity, has gained more

and more attention from both scholars and practitioners. Accord-

ingly, to develop this emerging market, actuaries and financial

economists have begun to make considerable efforts to improve its

structure and take on the challenges to connect financial and insur-

ance pricing theories. The well-known capital asset pricing model

(CAPM) lacks the flexibility to be applied appropriately to heavy-

tailed insurance risks. On the other hand, many insurance princi-

ples lack the ability to produce arbitrage-free prices. To address

these problems, we use the indifference pricing method to price

mortality and longevity securities. To illustrate the idea, we com-

bine the indifference pricing method and our proposed stochastic

mortality model. We show how to estimate the risk aversion pa-

rameter, using annuity market data. Then, we use the estimated

parameter to price a longevity option. The same techniques can be

used to price other mortality securities.

4.1. Indifference pricing method

We consider longevity risk in a life-only single premium

immediate annuity (SPIA) for age x. We further assume the annuity

benefit is ¯ s, paid at the end of each period. Thus, the present value

of future annuity benefits for an annuity written on M lives (x

i

) at

time 0, for i = 1, 2, . . . , M is

X

i

= ¯ s

k

e

−kr

k

I

{T(x

i

)≥k}

, (4.1)

where r

k

is the default-free interest yield fromtime 0 to time k and

k runs over the payment times.

10

The aggregate present value is

Y =

M

i=1

X

i

. (4.2)

10

The notation I

A

is the indicator function taking the value 1 if the event A occurs

and the value 0 if not.

S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253 251

The premium P per policy is to be determined from the equation:

E[u(w + MP − Y)] = u(w), (4.3)

where w is the wealth of the insurer prior to accepting the risk to

be priced. We use the exponential utility function,

u(w) =

1 − e

−αw

α

, (4.4)

where α is the risk aversion parameter of the insurer. The random

variable Y depends on the randompath of future force of mortality

values {µ(x + t, t) : t ≥ 0}, which can be simulated with our

model, using Eq. (2.10). The risk aversion parameter α in Eq. (4.3)

can be approximated using known insurance market prices. Under

our exponential utility function, the premiumimplied by (4.3) may

be expressed as

P =

1

Mα

log E

_

e

αY

_

. (4.5)

When calibrating to market data, we need to determine α so that

Eq. (4.5) is satisfied.

In the annuity pricing application we are considering, we are

pricing an aggregate annuity exposure for which the component

risks are identical but not necessarily independent. As we have

noted, the indifference premiumunder exponential utility is given

by (4.5). An approximation for P, which can be found in Gerber and

Pafumi (1998), can be made as

P ≈

1

Mα

_

αE[Y] +

1

2

α

2

Var(Y)

_

.

Since the aggregate loss is composed of M identical risks, E[Y] =

ME[X] and thus

P ≈ E[X] +

α

2M

Var(Y). (4.6)

The mortality is uncertain but it is reasonable to assume that, given

a mortality scenario Θ, the annuitant lives are independent. In this

case, the variance of Y is the sum of two terms:

Var(Y) = E[Var(Y|Θ)] + Var[E(Y|Θ)]

= E[MVar(X|Θ)] + Var[ME(X|Θ)]

= ME[Var(X|Θ)] + M

2

Var[E(X|Θ)].

The approximation becomes

P ≈ E[X] +

α

2

E[Var(X|Θ)] +

Mα

2

Var[E(X|Θ)]. (4.7)

When the mortality scenario is known (as in traditional actuarial

calculations with static tables), there is no variance in E(X|Θ),

the third term is zero, and the premium P per policy does not

depend on the number of polices M. However, in general, mortality

scenarios are uncertain and the premium P will depend on the

portfolio size M as indicated by (4.6). Moreover, even when the

lives are conditionally independent, given the mortality table, the

premium still depends on the size of the portfolio as indicated

by (4.7).

4.2. Estimating α from annuity prices

The Genworth Financial Group sold around $560 million of sin-

gle premiumindividual annuities in the US in 2005 with a monthly

payout rate of $6.40 per $1000 premium(Stern, 2008). We assume

that the policies are identical, issued to males age 65 with a gross

premium of $250,000. Since aggregate premium is $560 million,

then the number of policies is M = 560 million/250,000 = 2240.

0

500

1000

1500

2000

0 1000 2000 3000 4000 5000

Fig. 10. Genworth’s estimated utility function under the exponential utility model

for α = 4.8 × 10

−7

. The vertical axis represents utility and the horizontal axis

represents wealth, both in thousands.

We assume the underlying expense factor is 12.25%, which is

around the industry average, so the net premium per policy to the

insurer is

P = 250, 000 × (1 − 0.1225)

= $219,375.

Eq. (4.1) gives the present value of benefits for one life with ¯ s =

250 × 6.40 = $1600 per month.

We estimate the risk aversion parameter α from Eq. (4.5) by

simulation, with a ‘‘Monte Carlo simulation within a Monte Carlo

simulation’’. We start with a trial value of α = α

0

and estimate the

corresponding value of E(e

α

0

Y

) = E[E(e

α

0

Y

|Θ)]. First we generate

10,000 of scenarios Θ

j

. Then we simulate M = 2240 values of

e

α

0

Y

|Θ

j

, for each j, and estimate E[e

α

0

Y

|Θ

j

] with the sample mean.

The estimate of E

_

e

α

0

Y

_

is the mean of these sample means. This

is a very fast calculation. The corresponding model price, from

Eq. (4.1), is

P

0

=

1

Mα

0

log E

_

e

α

0

Y

_

.

We repeat the calculation with another trial value α

1

, then obtain-

ing P

1

. Then we use the secant method to find a third approxima-

tion,

α

2

= α

1

−

α

1

− α

0

P

1

− P

0

(P

1

− P) .

This process converges rapidly to α = 4.8 × 10

−7

. For the value

of α that was determined above, the unscaled utility function for

Genworth is shown in Fig. 10.

In this example, we used the population mortality tables rather

than annuity tables because we do not have a time series of annu-

ity tables. On a static basis, population mortality tables will tend to

underestimate the longevity risk in annuities relative to the appli-

cable set of annuity tables, thereby tending to create an overesti-

mate of α. However, we estimated α using Lee–Carter projections

for mortality fluctuations which embeds longevity risk. As we do

not knowthe mortality assumptions that Genworth used in setting

their annuity rates, this exercise must be viewed as an illustrative

approximation. The reason that we calibrated α using Lee–Carter

projections rather than the full model is that the full model induces

significantly thicker longevity tail risk than Lee–Carter projections

alone. Indeed, if α is estimated under the full model, a significantly

lower value of α is obtained because of the greater tail risk, and the

Genworth annuity premiums do not appear to contain a material

charge for this tail risk. It seems reasonable to believe that our esti-

mated α = 4.8×10

−7

can serve as an upper bound for Genworth’s

risk aversion to annuity longevity risk. Although we show how to

determine α with an annuity market quote, the same technique

can be applied to estimate life insurance risk aversion parameters.

We will use this value in the next section to estimate what

Genworth would pay for a longevity option.

252 S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253

Table 3

10-year longevity call option prices with data generating process driven by our

model.

Strike rate p 0.775 0.780 0.785 0.790 0.795

Call price $1121,364 $924,288 $772,136 $652,506 $553,345

Table 4

10-year longevity call option prices with data generating process driven by the

Lee–Carter model alone.

Strike rate p 0.775 0.780 0.785 0.790 0.795

Call price $191,962 $77,070 $23,864 $5,522 $937

4.3. Longevity option price

Let us now illustrate the pricing of longevity options using in-

difference pricing with the pricing parameter α that we have pre-

viously derived fromannuity prices. Consider a 10-year call option

issued in December 2005 based on our longevity index

10

p

65,2005

as defined in Section 3.3 with a notional amount of $100,000,000

and a strike price of p. The dollar payoff from this longevity call at

maturity in 2015 may be expressed as

C =

_

100,000,000

_

10

p

65,2005

− p

_

if

10

p

65,2005

> p

0 otherwise.

(4.8)

Although one can attribute a ‘‘cohort’’ to the payoff amount, in

practice the owner of the option is most likely concerned with the

total dollar exposure as it relates to a pension fund or annuity pool

liability.

We apply Eq. (4.3) to price this longevity call option for a

mortality data generating process driven by our mortality model

and by the Lee–Carter model alone. In all cases, we assume that

the market risk aversion parameter is α = 4.8 × 10

−7

, in line

withthe estimate that was made for the GenworthFinancial Group.

Table 3 shows longevity call option prices for a range of strikes for

mortality driven by our model. Table 4 shows longevity call option

prices for a range of strikes for mortality driven by the Lee–Carter

model alone. For example, given the strike level p = 0.775 and the

US treasury discount factors for the 2005 calendar year,

11

our 10-

year longevity call option price equals $1121,364 if the mortality

data generating process is driven by our model.

Evidently, the nature of the mortality data generating process

is critical in assessing the risk in a longevity derivative contract.

In the trading of standardized equity and interest rate derivatives,

one discovers that market prices reflect a ‘‘smile’’. In simple terms,

one finds that is it far more expensive to purchase ‘‘lottery tickets’’

on equity indices or interest rates than basic theory suggests. If

the market for longevity derivatives should evolve to a level of

reasonable liquidity, it would not be surprising if similar effects

were observed. While no one can be certain of the true nature of

longevity risk, our model has shown that the presence of relatively

infrequent mortality improvements poses a significant risk. Fig. 11

shows a histogram of simulated values of

10

p

65,2005

for 1 million

draws using the full model. Fig. 12 shows a histogramof simulated

values of

10

p

65,2005

for 1 million draws using the Lee–Carter model

alone. Evidently, it is the right tail behavior that is driving the

differences between the prices of the longevity calls across the two

mortality data generating processes.

11

To be precise, we require discount factors that are representative of interest

rates for 2005. We take the average US treasury coupon-curve yields for 2005 as

published in the Federal Reserve’s H15 data set and strip them to produce zero-

coupon data. We then fit a Nelson-Siegel yield curve to this zero-coupon data which

can be extrapolated beyond the last observation. At this point we have a set of

discount factors that is representative of market conditions in 2005.

x 10

4

0

2

4

6

8

10

12

14

16

18

0.75 0.8 0.85 0.9

10-year Survival Probability for (65)

0.7 0.95

Fig. 11. Range of outcomes for

10

p

65,2005

under the full model.

x 10

4

0.74 0.76 0.78 0.8 0.72 0.82

10-year Survival Probability for (65)

0

1

2

3

4

5

6

7

8

Fig. 12. Range of outcomes for

10

p

65,2005

under the Lee–Carter model alone.

Reinsurance contracts typically have a cap on the reinsurer’s

liability. It is reasonable to anticipate that mortality traders may

follow this tradition. The typical reinsurance contract is analogous

to a call spread, with two calls having different strike prices.

Consider a call spread written on the survivor index

10

p

65,2005

, with

lower strike level, ‘‘attachment point’’, p

1

and an upper strike level,

or ‘‘detachment point’’, p

2

and a notional amount of $100,000,000.

The dollar payoff fromthis longevity call spreadat maturity in2015

may be expressed as

B =

_

_

_

100,000,000 (p

2

− p

1

) if

10

p

65,2005

≥ p

2

100,000,000

_

10

p

65,2005

− p

1

_

if p

1

<

10

p

65,2005

< p

2

0 if

10

p

65,2005

≤ p

1

.

(4.9)

Table 5 shows the longevity call spread option prices for a range

of strikes for mortality driven by our model. Note that the spreads

in Table 5 have different, decreasing seniorities due to decreasing

trigger and exhaustion levels. In particular, spread A has the lower

attachment and detachment points relative to the other spread

options, and thus it demands a higher price.

5. Summary

We have described a mortality model which provides a rich and

realistic space of sample paths of future mortality rates. The model

evolves as a dynamic process that combines a general mortality

S.H. Cox et al. / Insurance: Mathematics and Economics 46 (2010) 242–253 253

Table 5

Prices for 10-year longevity call spread options.

Spread A Spread B Spread C Spread D

Attachment point p

1

0.775 0.780 0.785 0.790

Detachment point p

2

0.780 0.785 0.790 0.795

Price $122,985 $70,405 $39,337 $25,835

trend, a diffusion process, a permanent longevity jump process,

anda temporary mortality jumpprocess. Furthermore, we describe

a parsimonious version of the model based on judgment and

historical mortality data available through the Human Life Table

Database and the Human Mortality Database. Another novel aspect

is that our model includes uneveneffects of a mortality or longevity

shock to different ages. In addition, the model is tractable enough

to allow path-by-path simulation which can be combined with

pricing techniques, as we illustrate with the indifference pricing

method. The model is well-suited to risk management application

where there is a need to test extreme but plausible mortality

scenarios.

References

ACLI 2006. Life Insurers Fact Book 2006. American Council of Life Insurers.

Ballotta, L., Haberman, S., 2006. The fair valuation problem of guaranteed annuity

options: The stochastic mortality environment case. Insurance: Mathematics

and Economics 38, 195–214.

Bauer, D., Kramer, F.W., 2007. Risk and valuation of mortality contingent

catastrophe bonds. Working Paper, Ulm University, Germany.

Biffis, E., 2005. Affine processes for dynamic mortality and actuarial valuations.

Insurance: Mathematics and Economics 37 (3), 443–468.

Biffis, E., Millossovich, P., 2006. A bidimensional approach to mortality risk.

Decisions in Economics and Finance 29 (2), 1–25.

Blake, D., Burrows, W., 2001. Survivor bonds: Helping to hedge mortality risk.

Journal of Risk and Insurance 68 (2), 339–348.

Blake, D., Cairns, A.J.G., Dowd, K., MacMinn, R., 2006. Longevity bonds: Financial

engineering, valuation, and hedging. Journal of Risk and Insurance 73 (4),

647–672.

Bourguignon, F., Morrisson, C., 2002. Inequality among world citizens: 1820–1992.

American Economic Review 92 (4), 727–744.

Cairns, A.J.G., Blake, D., Dowd, K., 2006. Pricing death: Frameworks for the valuation

and securitization of mortality risk. The ASTIN Bulletin 36 (1), 79–120.

Coughlan, G., Epstein, D., Ong, A., Sinha, A., Hevia-Portocarrero, J., Gingrich, E.,

Khalaf-Allah, M., Joseph, P., 2007. Lifemetrics: A toolkit for measuring and

managing longevity and mortality risks. Technical Report. JPMorgan, March.

URL www.lifemetrics.com.

Cowling, C., Dales, R., 2008. The FTSE 100 and their pension disclosures. Technical

Report. Pension Capital Strategies, January.

Cox, S.H., Lin, Y., 2007. Natural hedging of life and annuity mortality risks. North

American Actuarial Journal 11 (3), 1–15.

Cox, S.H., Lin, Y., Wang, S., 2006. Multivariate exponential tilting and pricing

implications for mortality securitization. Journal of Risk and Insurance 73 (4),

719–736.

Cutler, D.M., Richardson, E., 1998. The value of health: 1970–1990. American

Economic Review 88 (2), 97–100.

Dahl, M., 2004. Stochastic mortality in life insurance: Market reserves and

mortality-linked insurance contract. Insurance: Mathematics and Economics

35, 113–136.

Dahl, M., Møller, T., 2006. Valuation and hedging of life insurance liabilities

with systematic mortality risk. Insurance: Mathematics and Economics 39 (2),

193–217.

Daley, D., Vere-Jones, D., 2003. An Introduction to the Theory of Point Processes, vol.

I. Springer-Verlag, New York.

Dowdle, W.R., 2006. Influenza pandemic periodicity, virus recycling, and the art of

risk assessment. Emerging Infectious Diseases 12, 34–39.

Gerber, H.U., Pafumi, G., 1998. Utility functions: From risk theory to finance. North

American Actuarial Journal 2 (3), 74–92.

Goss, S.C., Wade, A., Bell, F., 1998. Historical and projected mortality for Mexico,

Canada, and the United States. North American Actuarial Journal 2 (4), 108–126.

Gründl, H., Post, T., Schulze, R.N., 2006. To hedge or not to hedge: Managing

demographic risk in life insurance companies. Journal of Risk and Insurance 73

(1), 19–41.

Hardy, M., 2005. Amatter of life and death. Financial Engineering News. Available at

http://www.fenews.com/fen46/topics_act_analysis/topics-act-analysis.htm.

Hári, N., Waegenaere, A.D., Melenberg, B., Nijman, T.E., 2008. Estimating the term

structure of mortality. Insurance: Mathematics and Economics 42 (2), 492–504.

Juckett, G., 2006. Avian influenza: Preparing for a pandemic. American Family

Physician 74 (5), 783–790.

Lakdawalla, D., Philipson, T., 2002. The rise in old-age longevity and the market for

long-term care. American Economic Review 92 (1), 295–306.

Lane, M.N., 2006. How high is up? The 2006 review of the insurance securitization

market. Technical Report, Lane Financial LLC, Wilmette, IL, April.

Lane, M.N., Beckwith, R.G., 2007. That was the year that was! The 2007 review

of the insurance securitization market. Technical Report, Lane Financial LLC,

Wilmette, IL.

Lee, R.D., 2000. The Lee–Carter method of forecasting mortality, with various

extensions and applications. North American Actuarial Journal 4 (1), 80–93.

Lee, R.D., Carter, L., 1992. Modelling and forecasting the time series of US mortality.

Journal of the American Statistical Association 87 (419), 659–671.

Lin, Y., Cox, S.H., 2005. Securitization of mortality risks in life annuities. Journal of

Risk and Insurance 72 (2), 227–252.

Lin, Y., Cox, S.H., 2008. Securitization of catastrophe mortality risks. Insurance:

Mathematics and Economics 42 (2), 628–637.

Loise, S., Serant, D., 2007. In the core of longevity risk: Hidden dependence in

stochastic mortality models and cut-offs in prices of longevity swaps. Working

Paper. Université Claude Bernard Lyon 1.

Luk, J., Gross, P., Thompson, W.W., 2001. Observations on mortality during the 1918

influenza pandemic. Clinical Infectious Diseases 33 (8), 1375–1378.

Marocco, P., Pitacco, E., 1998. Longevity risk and life annuity reinsurance.

In: Transactions of the 26th International Congress of Actuaries, vol. 6.

Birmingham, England. pp. 453–479.

Milevsky, M.A., Promislow, S.D., 2001. Mortality derivatives and the option to

annuitize. Insurance: Mathematics and Economics 29 (3), 299–318.

Miltersen, K.R., Persson, S.-A., 2005. Is mortality dead? Stochastic forward force of

mortality rate determined by no arbitrage. Working Paper. University of Bergen.

Olivieri, A., Pitacco, E., 2002. Inference about mortality improvements in life annuity

portfolios. In: XXVIIth International Congress of Actuaries.

Renshaw, A., Haberman, S., Hatzoupoulos, P., 1996. The modeling of recent mortality

trends in United Kingdom male assured lives. British Actuarial Journal 2 (2),

449–477.

Rogers, R., 2002. Will mortality improvements continue? National Underwriter 106,

11–13.

Salou, J.-M., Hu, Y.-W., 2006. Overview of the financial wealth accumulated under

funded pension arrangements. Pension Markets in Focus 3, 2–13.

Schoen, R., 2006. Dynamic Population Models. Springer, Dordrecht, The Nether-

lands.

Sithole, T.Z., Haberman, S., Verrall, R.J., 2000. An investigation into parametric

models for mortality projections, with applications to immediate annuitants’

and life office pensioners’ data. Insurance: Mathematics and Economics 27 (3),

285–312.

Stanley, Morgan, 2003. Swiss Re innovative mortality-based security. See the news

release for December 8 2003.

Stern, H.L., 2008. Annuity shopper. Published semi-annually by Web Annuities.com

Inc., January.

Taubenberger, J.K., Morens, D.M., 2006. 1918 influenza: The mother of all

pandemics. Emerging Infectious Diseases 12 (1), 15–22. http://www.cdc.gov/

ncidod/EID/vol12no01/05-0979.htm.

Swiss Re. 2003. Swiss Re and Mitsui Sumitomo arrange USD100 million catastrophe

risk swap. http://www.swissre.com. See the news release for August 4, 2003.

The Actuary. 2004. Swiss Re obtains $400 million of mortality risk coverage. The

Actuary. January/February, P. 16.

Toole, J., 2007. Potential impact of pandemic influenza on the US life insurance

industry. Research report, Society of Actuaries, May. URL http://www.soa.org/

research/life/research-impact-pan-influ-life-ins.aspx.

Vaupel, J.W., 1998. Demographic analysis of aging and longevity. American

Economic Review 88 (2), 242–247.

we cannot simply calibrate our models to the experience. h) is a random variable since it is the path integral over the sample path {µ(x + t . Statistical methods applied in actuarial practice use this same approach. Gründl et al. pension and annuity research has focused on longevity risk. Here is how we will model these events. We conclude the article in Section 5. Marocco and Pitacco (1998). Dynamic mortality model Stochastic mortality models usually start with the assumption that there is an ‘‘initial’’ curve for the force of mortality. t ) : t ≥ 0} evolving from an initial curve µ(x. Dahl (2004). we offer a flexible.4 Our paper is organized as follows. for which we will provide a detailed discussion later. In particular. t )..H.g. Finally. Therefore.html. However. Ballotta and Haberman (2006).g. Specifically. a permanent longevity jump process and a temporary mortality jump process. (2008) extend the Lee–Carter model with a time-varying. but in the long run longevity risk may be just as important. h)] . this is very complex. To describe the stochastic longevity trend. may be less dramatic than that of a mortality death shock. more people may survive than expected. the force acting during the life of a person age x at time t = 0 is a stochastic process {µ(x + t . Our model attempts to address these issues. actual mortality experience can change in a nondeterministic way and deviate from expected mortality rates. we propose a new approach by introducing a trend reduction jump component to describe longevity risk. is used to price life insurance and annuity polices. but with explicit mortality derivatives such as options which we think can be important tools in managing mortality and longevity risk. 1918 flu pandemic and the change in the rate of improvement of mortality around 1970. to obtain relevant results. the model specifies the force of mortality as a parametric function of age and time. as evident. In order to reduce the complexity. The force of mortality at time t > 0 is denoted µ(x + t . Models that have a random deviation from the expected mortality may reflect some longevity risk. 0 The probability S (x. The probability that (x) will attain age x + h is the expected value of the probability: Pr [T (x) > h] = E [S (x. h) = exp − 0 µ(x + s. cannot provide this kind of longevity risk. Moreover. The aggregate effect on (x) over the period (0. experts conjecture that we may experience extreme events of both types in the future. 3 http://www. • relatively rare advances in medical technology resulting in sudden and permanent improvements in mortality. respectively. t ) acting on (x + t ) at time t. including new structures. Most longevity risk events in the past seem to have a pattern. in general. t ) : 0 ≤ s ≤ h} of the force of mortality: h µ(x + t . . although somewhat complex. However. stochastic drift. the remaining life time random variable is denoted T (x). for example. of the force of mortality µ(x + t . In effect. because it depends on the sample path of the force of mortality: h S (x. This means that. Actual mortality has been improving. For a life (x) at time 0. • deviations from the general trends. 2007) offered by Goldman–Sachs3 and the Credit Suisse index. Derivatives can be written on indices based on publicly available data like the LifeMetrics index (Coughlan et al. leading to a steeper downward sloping force of mortality curve. Unanticipated improvements in medicine and health technology may give rise to longevity risk. so parametric models estimated with actual data will reflect improving mortality. this model may not be appropriate for modeling longevity (and mortality) shocks which may not follow a stationary Gaussian process. The traditional one-time jump models such as the model that combines a geometric Brownian process and a compound Poisson jump process. but longevity risk is not modeled explicitly. or observed death rates may be lower than expected. a stochastic mortality model must reflect three major features of the current mortality universe: (a) both mortality improvement and deterioration jump factors. Our model. comprehensive model in order to make a more realistic assessment of future survivor dependent cash flows. On the other hand. The expected value.qxx-index. Cox et al. Miltersen and Persson (2005). one or two years) or excess improvement over a long period (e. Loise and Serant (2007) model longevity (and mortality) risk use a stationary Gaussian process. leading to a surge in mortality improvement in a short period (e. Hári et al. (2006). the slope of the mortality curve may deviate from that of the base trend with normal deviations. Mathematically speaking. 2. We show how to price those securities with the indifference pricing method in Section 4.com/ib/en/fixed_income/longevity_index. Biffis (2005) captures mortality random departures around a time-varying target with a longevity compound Poisson process. or forecast value. In those papers. with a small number of parameters. (2006). (b) correlation among different ages and over time. We continue in the same vein. Dahl and Møller (2006). more than ten years). t )dt . (2006) examine a wide variety of longevity bonds. However. 0) = µ(x). h) that (x) will attain age x + h is random at time 0. US mortality data has only two extreme events. unexpected survival gains often extended over a long period of time. Cairns et al. 4 http://www. but that model cannot guarantee a nonnegative force of mortality. Among the reasons for changes in the force of mortality are the following: • general trends in health and mortality resulting in gradual improvements to mortality. in investigating the impact of possible future events. For longevity events with a long-term effect. we are explicitly combining both mortality and longevity risk in a single. in the US population mortality dynamics in the 1970’s. Biffis and Millossovich (2006) and Bauer and Kramer (2007) focus on mortality risk as deviations from a trend. method of including expert opinion in forming future mortality scenarios. and (c) uneven effects of a mortality jumps across different ages.S. In Section 2 we describe mortality dynamics as a combination of a Lee–Carter diffusion process. In Section 3 we discuss capital market solutions for mortality and longevity risk. Unexpected longevity improvement. with our proposed longevity index. Blake et al. s)ds .credit-suisse. given T (x) > t. Milevsky and Promislow (2001). the base trend may reflect mortality improvements. and • sudden catastrophic surges in mortality from events such as pandemics resulting in temporary increases in mortality. As an application to mortality risk management. allows for such fundamental changes.com. but this is not a reflection of a fundamental change in the trend (as we have in mind). / Insurance: Mathematics and Economics 46 (2010) 242–253 243 On the other hand. we show how to determine a parsimonious version with historical data.

9908 −5. q(x.7227 −4.4290 −2.0398 0.244 S.1422 −6.0960 −5. Modeling general mortality trends We adopt the well-known Lee and Carter (1992) mortality model as a basis.0423 0.0822 −3.3986 −7.0295 0.0298 0.0393 0. The force of mortality µLC (x.9509 −4.4259 −5. This model captures the evolution of mortality in mutually exclusive age cohorts.1820 0. For reasonable assumptions about the distribution of deaths between age x and x + 1.0242 −4.8122 −0.6228 −0. Available at www.4695 −1. Cox et al. We believe that our ideas illustrated below could be applied to them as well by using the original Lee–Carter model.1017 0.8732 −5. .0841 0.3116 −1. Following Lee and Carter (1992).3015 −5.9251 −1.0655 0.0453 0. Simulating mortality scenarios with the Lee–Carter model. 1992.1736 0.8509 −1..0396 0. Available at www.8779 −0.0544 0. the two measures are the same or very close. 1.0934 0.0370 0. .2612 −4.2492 −2.0410 0.0903 0.0687 0. t ).1100 0.0828 −1. Under the same assumption.6262 −1.2240 −7.humanmortality.0593 0.0471 0. University of California.4739 −3.2269 0. 2005 (2. and Max Planck Institute for Demographic Research (Germany).0986 0.0352 0.1687 −2.6519 −5.1041 −4. using the technique 5 Lee and Carter use the central death rate rather than the force of mortality.9207 −2. t ). Renshaw et al.0022 −1.0440 0.3380 −4. .3970 −3.1113 0. (2.3661 −5.0269 0. .1091 0.1572 −1.1. 1902.1) The parameters a(x) and b(x) are age-specific while k(t ) is time varying.2). 7 Data source: Human Mortality Database.1614 −3.0303 0.7025 −1.2287 0. we explicitly model both mortality and longevity jump processes in a single model (See Section 2. Our model is inspired by. For example.6 Tables for 2000 to 2005 are from the Human Mortality Database. 103 and times t = 1901. it is natural to develop a model for mortality changes based on the evolution of the force of mortality with a general trend.4851 −5.1495 0. links all cohorts together. . The flu in 1918 is identified by a dummy variable.0618 −5. There are several generalizations of Lee and Carter (1992) model in the literature. Olivieri and Pitacco. z (1902).1278 0.0384 0. 2004.0621 0.3761 −7.5587 −0.0554 0.8537 −5.8663 −3.1883 0.0488 0. 6 Data source: Human Life Table Database. .9863 −7.0020 −2.0384 0. .2244.5230 Therefore.0812 0.1908 0. 2000.5442 −5. .8953 −5.3290 −7.6471 −4.7063 −3. t ) = 1 − e−µ(x.2336 −1.mortality.0375 0.8849 −5.0752 0. t ) = exp [a(x) + b(x)k(t )] .0270 0.7 The estimates of a(x) and b(x) are in Table 1 and the estimated k(t ) is shown in Fig. if the force of mortality is constant between integral ages.0262 0.0784 0.7575 −2. and closely related to. Cairns et al. .2273 −6.0265 0.1079 Age 35 36 37 38 39 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 a(x) b (x ) 0.1006 0.9457 −0.2000 −6. Berkeley (USA) and the Institut national d’études démographiques (France).0267 0.0277 0.8750 −4. . t ) is modeled as5 suggested by Lee and Carter (1992).2171 0. .1101 0. Age 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 a(x) b(x) 0.0532 0.1878 0.0282 0.4039 −5.0391 0. diffusion process and jump components. A number of recent studies have sought to model mortality trends involving both age-dependent and time-dependent terms (Lee and Carter.0391 0.de(data downloaded on June 8. . and 0 otherwise. t = 1901.0377 0.0506 0.0370 0.1099 0.0992 0. this stream of research.0274 0. .8191 −5.7830 −5.8855 −5. and σ are constants and z (1901).0966 0.1820 0. g2 = 2.0383 0.0721 0.0335 0. Milevsky and Promislow.t ) so we can easily change from q(x.9452 −3.6993 −5. .de (data downloaded on June 8. For convenience we now re-label the years so that year 2005 corresponds to t = 0.1039 0.2029 0.1092 0.0395 0. 1996. 2002.0265 0.5665 −6.1827 −4.0290 0.9989 −6. Tables for years 1901 to 1999 are from the Human Life Table Database.6955 −6. involves simulating future mortality index values k(t ) for t > 2005.3902 −6. For a person age x in 2005. (2. The database has values of q(x.8571 −6.0265 0.9028 −5.5990 −5. k(t ). . while at the same time a timeseries common risk factor.0934 0.6766 −2. 2001.5132 −2. University of California.6867 −0. Flu.8391 −2. We fit the model to the US male data for ages x = 0. t ) to µ(x.0834 −2.org or www.6267 −3. . Specifically.2354 −5.7715 −6. / Insurance: Mathematics and Economics 46 (2010) 242–253 Table 1 The estimates of a(x) and b(x) based on the US male population data from 1901 to 2005.5699 −4.8802 −5.0392 0.4928 −4.4151 −4.1737 0.5482 −1.5503 −3. Then we obtain the estimated g1 = −0.3429 −2. Sithole et al.1021 0.2527 −6.0268 0. 2008). 2005.6123.1666 −5. As an extension.0376 0.0319 0. 2.7502 −0.0944 0. Flu = 1 if t = 1918.0383 Age 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 a(x) b (x ) 0.1082 0.0142 −0.0381 0.0371 0. z (2005) are independent standard normal random variables. and σ = 0. in Eq.2612 0.1067 0.7435 −5.1043 0.0400 0.7774 −1.3906 −1.0521 0.7995 −4. 1.lifetable. 2006).2396 −3.2065.0222 −3.1111 −7. . 1902.9233 −5. g2 .0911 0.9019 −5.3189 −3..7859 −3.7038 −6.0903 0. Dahl.0959 0. we can generate µLC (x.4963 −6.1050 0.2374 0. the central death rate at age x and the force of mortality at age x are equal.0378 0.0249 −3. 2008). the mortality index k(t ) evolves as k(t + 1) = k(t ) + g1 + g2 × Flu + σ z (t ).0569 0.2) where g1 . Lee. 2000..0872 0.1061 0.0990 0. Max Planck Institute for Demographic Research (Germany).5950 −2.0309 0.H. Berkeley (USA).2).

00 Fig. 1. The simplest version would be the arrival times of a Poisson process. have a more severe but transitory effect on death rates. µLC (x + 2. we view sharp decreases in mortality as ‘‘permanent’’ events. T ). The unexpected mortality improvement that results from genetic. The jump reduction component K (x.3) to ages x > 103 after the first year. as described by Schoen (2006). compared to demography changes of the whole population in the 1970’s. When we add these two types of mortality jump events – G(x. like the 1918 worldwide flu and the 2008 earthquake in China. behavioral. t ). Furthermore. for instance. the annual mortality rate over a long period (e. .3) to x > 103 for t > 0. . t ). . 0). Accordingly. 1). µLC (x + T . 1902. Modeling mortality jumps Our goal is a stochastic mortality model suitable for dynamic financial analysis. t ) is the force of mortality from the Lee–Carter model (2. t ) = µLC (x. The primary difference between these two approaches lies in the ability of a Hawkes process to create event clustering. This approach results in the linear extrapolation of a(x) and b(x). where ηs is the time of jump reduction event s.3) 2. t ) – to the Lee–Carter force of mortality µLC (x. A second feasible choice would be a Hawkes process. often has a long-term effect on future mortality rates. the way we use the model requires extending Eq.055 − 0. . (2. medical advancement s has no effect on age x. t ). 103.5) µ(x. Cox et al. 2004.1. and view dramatic increases as ‘‘temporary’’ events such as pandemics. these short-term events drive up the death rate just for a couple of years. LC (2. a(x) = −7.43 + 0. t ) = exp [a(x) + b(x)k(t )] t > 0. .067x b(x) = 0.00 1901 1911 1921 1931 1941 1951 1961 1971 1981 1991 2001 -5. . A natural approach is to simply extend both functions a(x) and b(x) linearly using the values for ages 90 to 103 as a basis for the extrapolation. . twenty years) may continue to improve at a much higher rate . t )) .g. When As (x) = 0. .H.00 245 5. t ) and H (x. the life expectancy gain is greater for the elderly than for the young. associated with aforementioned medical advances. When 0 < As (x) < 1. That is. 2005 in the horizontal axis.00 -15. the annual improvement in mortality of old ages in the same period is much more dramatic than that of young ages. taking into account the complexity of observed mortality dynamics. (2. Lee–Carter model scenarios for future values of the force of mortality as follows: For x = 0.00 0. other approaches might be necessary. (2. in which case the times would have gamma distributions. . bio-reliability. the benefits of the technology usually vary by age. t + 2). t ) × exp (−G(x. As (x) = 1 means age x enjoys the biggest mortality improvement (ys ) among all ages. Estimated time-series common risk factor k(t ) shown in the vertical axis for year t = 1901. Fig. Permanent longevity jump G(x. µLC (x + 1. and/or heterogeneity forces and constraints. . t ) The unexpected mortality improvement process G(x. many catastrophe death events. 0) = µ(x) observed in 2005 µLC (x. the function As (x) differentially spreads the mortality improvement across x’s because. t ). but the mortality falls back to the normal level afterwards. a year) with a permanent effect on longevity. . it captures the effects that fall between the above two extremes. (2. The time ηs can be modeled as the arrival time of the s-th event of a point process. .00 -10. 2. t ) is the one-time longevity jump induced by a surge in survival rates over a short period (for example. µLC (x. G(x. 2 may be more realistic in the future. the initial version of our model takes the form K ( x. t ) = s =1 ys As (x)1{t ≥ηs } . There is no definitive data set to say what the function As (x) should look like. For example. Mortality improvements due to a change in a cause of death may vary by age. It is defined as ∞ When we use the model to simulate future values of the force of mortality applying to a life age x at t = 0 over a period of T years µLC (x. The effect of ys is transferred to the mortality rates through a function As (x) for age x with 0 ≤ As (x) ≤ 1. Medical advancement by age. In fact. t ) + H (x. This is consistent with models with linearly changing mortality.3). Since cardiovascular disease is more prominent later in life than earlier. That is.0003x extending Eq. as described in Daley and Vere-Jones (2003). t ) = K (x. we need to provide for the cases for which x + j > 103 and j > 0. t ) consists of a jump reduction component K (x. In contrast. In many cases.2. environmental. In other words. depending on the products being modeled. when a medical advancement occurs. / Insurance: Mathematics and Economics 46 (2010) 242–253 10. In contrast. If you think the shape shown in Fig.S.4) where µ (x. We consider two types of mortality shocks: permanent jump G(x. 1. 2. you must believe that there is little mortality improvement for the very young and very old. Practitioners will use a conservative choice. t ) and temporary jump H (x. t ) and a trend reduction component D(x. The positive variable ys is the maximum mortality improvement of all ages in medical advancement event s.2. t ) + D(x. Cutler and Richardson (1998) find that improvement for the elderly was greater than for the young in the 1970’s as a result of a decrease in cardiovascular disease deaths. In applications for which ‘‘oldest-of-old mortality’’ is critical.

t ) The transitory mortality jump process H (x. In order to allow for these effects in a manner that can be compared to the historical data. when an unexpected medical advancement causes an annual excess decrease in µ(75. is more significant than the one-time longevity jump K (x. t ). The higher κj the faster the jump effect will die out. 3. / Insurance: Mathematics and Economics 46 (2010) 242–253 1921 1941 1961 1981 2001 Fig.8 We can estimate κj to fit the historical data and make the mortality jump only have the impact over a reasonable short period of time. distributes mortality jump impact across ages. Bj (x) can be random to reflect various age effects for different types of events or it can also be deterministic. than that in other periods.7) D(x. The basic effect of a pandemic is modeled using a Poisson process with a jump size bj > 0. Like Ai (x). when ζi = 0. The vertical axis stands for force of mortality and the horizontal axis represents year.100 0. under age 50. the force of mortality curve for age 75 will have a notable steeper downward slope for about 35–40 years. t ) during one year).3723 for age 75.070 0.3723 × 100 = 0. In this case.2. but that does not mean there was no longevity risk. 4 would apply to a pandemic for which the bulk of the mortality spike is under a certain age (for example. t ) for age 75 in year t. Moreover.2. t ) decreases by an excess 0. we calculate the difference between its average percentage change in the trend reduction period and its average level in the whole observation period. t ) = j =1 bj Bj (x) exp(−κj (t − τj ))1{t ≥τj } . Comment on estimation issues The primary motivation for the model that we have presented is to provide a flexible and realistic model that can be used to capture adverse mortality and longevity risks in a business application. (2) Calibrate the A. where τj is the time of adverse mortality event j. the specification of Bj (x) in Fig. Since modeling the transitory nature of mortality jumps is important in practice.040 0. t ) = i =1 ζi (t − υi )Fi (x) exp(−ξi (t − υi ))1{t ≥υi } . one would look at how the mortality effects were distributed across ages and calibrate accordingly. The time of the trend reduction event υi can be modeled like the time ηs . The cumulative effect of unanticipated mortality improvement over an unexpectedly long period of time obviously has a devastating effect on the financial stability of pensions plans and annuity markets. 3. The time of the adverse mortality event can be modeled like the time ηs . Given ζi = 0.0211 × 0. (2. This poses significant estimation challenges. (2.H. Therefore. the factor (t − υi ) provides cumulative mortality improvement as t increases beyond the jump event time υi . distributes age effects. log µ(x. How long the aforementioned mortality trend reduction will last in the future is a topic for debate: ten years.t +1) the log µ-ratio. κ is much larger than ξ because of the temporary effect of a j i catastrophe mortality event and the long-term impact of unexpected mortality improvement.5). with 0 ≤ Fi (x) ≤ 1.3723. we have Fi (0) = 1 in the 1970’s. 0.060 0.79% per year during this trend reduction period. t ) defined as follows: ∞ 2.090 0. 8 In general.0211 above its average improvement from 1901 to 2005. B and F functions to approximate past history. which we discussed after Eq.5). one might choose ξi = 0. in a certain sense. (2. where bj is the maximum severity of jump event j. (2. one can use the historical frequency of pandemics as a guide to setting the jump frequency. for example. Based on the historical data. with 0 ≤ Bj (x) ≤ 1. Cox et al.246 0. The percentage change in µ(x.3. t ) is µ(x. For example. the function Fi (x). The function Bj (x). For instance. The data shows no one-time longevity surge across ages (like a 20 percent decrease in µ(x. For example. In Eq. on average. . Temporary adverse mortality jump H (x. 2. This type of mortality improvement. ζi represents the maximum excess change across ages given ζi > 0. which we discussed after Eq. For example. leading to a steeper downward-sloping force of mortality curve. we introduce a nonnegative deterministic function exp(−κj (t − τj )) in Eq. Fi (75) = 0. Pandemic effect by age. Should one believe that history will reflect the future. consider male mortality in the 1970’s: age x = 0 had the biggest annual excess longevity gain. 4. as is the case for deaths in the 1918 worldwide flu). For example. When a trend reduction event occurs. (3) Calibrate the jump frequency for the point processes driving mortality changes in a range that is consistent with history.0211. is about three times that of the same age cohort in other periods. (2. we can obtain other Fi (x)’s.t ) . Fig. so ζi = 0.0211 and Fi (75) = 0. (2.0211.6).080 0. a fair number of parameters need to be included. the pattern similar to the curve for age 75 after year 1970 in Fig.01. t ).030 1901 S. we model this long-term decay in mortality rates by introducing a trend reduction component D(x. twenty years or even longer? The parameter ξi > 0 specifies the length of trend reduction event i.7) with κj > 0.110 0. the annual percentage decrease in death rates for the US population aged 55 and above in the 1970’s.6) where υi is the time of trend reduction event i. Our method of estimation/calibration can best be described as follows: (1) Estimate the original Lee–Carter model based on the available data. If the primary focus is on pandemics. This means µ(75.050 0. t ) is defined as follows: ∞ H ( x. Actual force of mortality µ(75.

t ).10) to model mortality dynamic process. (2. 5 shows our estimated F (x) for different ages based on the data from 1970 to 1979.6 0.H. age 0 had the highest average annual excess mortality decrease rate ζ = 0.0 1. and M (t ) is the number of such jumps by time t. The left graph of Fig. perform risk management functions for a range of life insurance products in plausible mortality scenarios. we provide a parsimonious version of our model. Example Below we show how to apply the parsimonious equation (2. However. of these two components. (2. the user of the model can.8) A permanent longevity jump can be driven by the one-time component K (x.0 247 0.8 0. Accordingly. (2. we are focused on mortality rather than equity returns but the same question can be asked. Eq. the permanent longevity jump is only a function of the trend reduction component.4 0. µ(x. This is especially so because of the relative paucity of extreme events in mortality data sets. t ) = exp [a(x) + b(t )k(t )] M (t ) × exp − i=1 N (t ) ζi (t − υi )Fi (x) exp(−ξi (t − υi ))1{t ≥υi } .S. we are not able to reliably measure the extent to which jump behavior is present. While we do not have an estimation procedure that simultaneously sorts out the diffusive and jump parameters of our model. The problem is acute in the modelling of stock market returns because one generally introduces jumps into the model with the idea of picking up ‘‘crashes’’ such as 1929. Therefore.8 0. even though such a flu has been observed only once in 100 years. t ) that accumulates the excess mortality improvement over a long time. For example. t ) across ages. 5.9). ‘‘estimation’’ of this model requires a combination of traditional estimation techniques and expert judgment. t ) within a short . Insofar as the stochastic features of the model can be ascertained for a given parameterizations. × exp i=1 bj Bj (x) exp(−κj (t − τj ))1{t ≥τj } (2. In our model. The left figure is the function F (x). In financial and economic applications of jump-diffusion type models.2.1. D(x. t ) that substantially reduces µ(x. substituting in Eq. As mentioned in Section 2.9) can be further simplified to this: µ(x. especially for ages above 65—the age range we are very interested in for longevity risk. which may not be the same as the data. Cox et al. at the very least. Trend reduction factors in our model provide for such mortality improvements. As we noted earlier.2 0. but it can also come from the trend reduction component D(x. The importance of the impact of medical advancement on pension plans and annuity insurers should not be underestimated. Similarly. Parsimonious model For the Lee–Carter base model. t ) = exp [a(x) + b(x)k(t )] × exp −ζ F (x) × exp bB(x) M (t ) (t − υi ) exp(−ξ (t − υi ))1{t ≥υi } . (2.2. we can observe mortality declines at high excess rates mostly at the older ages. the process N (t ) counts the number of catastrophe death events by time t. / Insurance: Mathematics and Economics 46 (2010) 242–253 1. (2. The process M counts mortality trend reduction jumps. µ(x. As discussed in Section 2.4 0. All of our estimates are based on the aforementioned US male population mortality tables from 1901 to 2005. The 1970’s were the years of accelerating decline of mortality for most ages.6 0. one issue that emerges is how to appropriately disentangle the diffusion from the jump dynamics.9) In Eq. t ) for different ages in the 1970’s. t ) + H (x.2 20 40 60 80 100 20 40 60 80 100 Fig.4) becomes LC period of time. giving the impact of the 1918 worldwide flu on µ(x. Eq. the normalized annual excess percentage decrease in µ(x. we do feel that the infrequent occurrence of severe adverse mortality events makes our imposed reconciliation of the diffusion and adverse mortality jumps reasonable.10) i =1 assuming constant mortality and longevity jump effects for age x.0211. our approach allows one to set the frequency of a severe 1918-like flu so that it is a 1 in 25 year event.4. In the case of possible longevity jump components. it does pose a serious estimation challenge. i=1 N (t ) exp(−κ(t − τj ))1{t ≥τj } (2. Evidently. Indeed our goal is to allow for scenarios reflecting expert judgment about these events. t )] ∞ = exp [a(x) + b(t )k(t )] exp − s =1 ys As (x)1{t ≥ηs } ∞ × exp − i =1 ζi (t − υi )Fi (x) exp(−ξi (t − υi ))1{t ≥υi } bj Bj (x) exp(−κj (t − τj ))1{t ≥τj } ∞ × exp j =1 .1. we begin by using the data to estimate the Lee–Carter model. On the other hand. 2. t ) seems to play a more important role. during the 1970s. In sum. so our trend reduction parameters are estimated from this period. while our model can generate a rich set of scenarios. The right figure is the function B(x).5. t ) = exp [a(x) + b(t )k(t ) − G(x.3) for µ (x. 1987 or 2008. but what can happen is that the estimation of such a model produces a preponderance of high frequency low severity jumps rather than the intuitively pleasing outcome of fairly smooth diffusion behavior punctuated by occasional severe jumps. so F (0) = 1. 2.

sample paths of µ(x.50 to model 0 . suicides. As long as the estimates of y. A(x) as well as the jump intensity are developed based on judgment. For example.01.2 5 0 .0 0 2005 2025 2045 2065 2085 2105 Fig. t ) has a steeper downward slope that starts in year t = 2027 and lasts for about 20 years (its trend reduction longevity jump D(40.1 5 0 . t ). (2.. Based on the historical data. these older people survived earlier flu epidemics and may have acquired immunity to the flu virus.0 0 5 0 . which is the value for b. we specify the trend reduction event arrival rate λν = 0. It is important to note our model can flexibly estimate the impact of a particular jump pattern by adjusting F (x) and B(x) based on different opinions or updated information.3 0 0 . their unexpected improvement in life expectancy will have a great bearing on the overall impact of the longevity event. λτ = 0. in our model. Cox et al. An illustration is provided in the next section. the movement of different µ(x. The parsimonious model (2. If an age does not have an excess mortality decrease rate. Since the 1918 flu is approximately a onein-one-hundred-year event. Another exception is the 1918 worldwide flu as mentioned earlier: there were no excess deaths at ages above 65. . However. ‘‘excess deaths are heaped at age 0. F (x) and B(x) also distribute uneven effects of a mortality jump across different ages. This shows the trend reduction longevity jump D(40. We then used the forces of mortality for the period 1965 through 2005 to estimate average rates of mortality improvement by age.1 0 0 .0 0 2005 time.3 5 0 . The 1918 flu increased the one-year force of mortality for different ages very unevenly. The shape of the mortality curve often changes in different mortality or longevity jump events. 7). With the observed force of mortality for age x in year t = 2005 as the base case. Moreover. the adverse mortality jumps mainly strike younger ages while the older ages enjoy more excess survival gains. 2001). the curve of µ(40. quickly decrease to close to zero until they start to increase again at older ages (typically age 65) with a rapid rise at ages 85 and older’’. t ) for ages 30. then µ(40. t ) shown in the right graph of Fig. t ) = µ(x. t )) . effects of both mortality improvement and deterioration jump factors are clear. This is a problem that we will want to carefully consider.10) and the parameters described earlier.6 0 0 . These are the ones used in the simulated paths shown in Fig. it serves as the basis for the estimation of our temporary adverse mortality jump parameters. 6. Toole (2007) concludes that the typical distribution of excess death rates for seasonal influenza is U-shaped.9647 and F (28) = 1. Since the annuity benefits in general are distributed mostly in the older age ranges. based on the historical data. while there were no excess deaths at the ages above 53. it is possible to calibrate the model to behave in a manner that is qualitatively similar to the basic Lee–Carter model. / Insurance: Mathematics and Economics 46 (2010) 242–253 0 .11) We calibrated this model by taking µ(x.6. t ). t ) and µ(30. 5 presents B(x). 6. t )’s are correlated among different ages and over 0 . and accidents. age 40 has a sharp increase in µ(40. All parameters of the example shown in the above example are estimated from the historical data. We calibrated the adverse mortality effects as was done in the previous section. that is.01. 2. t ) falls back to the normal level afterwards (its temporary jump factor H (40.0 1 0 0 . resulting in a simple model of the form µ(x. (2. t ) + H (x. the normalized log µ-ratio in 1918 relative to 1917 for different ages. 0) exp (−K (x. The top path is the simulated force of mortality µ(60. The flu struck ages 0–50.H. 40. mortality increased for some male young ages. the 1957 pandemic showed a spike in excess death rates at ages 65 and older but there were no excess death at age 0 (Luk et al. The 1918 worldwide flu is considered the most severe flu epidemic.0 5 2025 2045 2065 2085 2105 0 . based on Eq.3 0 0 . t ) since there is no clear historical evidence to guide us on its pattern. this pandemic only lasted a couple of years so κ = 1 is reasonable. Moreover. t ) (right) based on Eq. t ) in a richer model.01 and the duration parameter ξ = 0. we use a Poisson process with intensity 0. 0) as the force of mortality as of 2005 from the data that we have discussed above. 50 and 60 in one simulation iteration are in Fig.248 S. especially ages 20–40. Third. perhaps due to an increase in homicides. This is partially attributed to the functions F (x) and B(x).0 0 0 2005 2025 2045 2065 2085 2105 Fig. 7). 6.5 0 0 . This led to the functions A(x) and B(x) shown in Fig. According to Taubenberger and Morens (2006). b = 0. t ) illustrated in the left graph of Fig. 7.4 0 0 .10) does not include the jump reduction component K (x. age 28 has the highest log µ-ratio.6) and temporary adverse mortality jump H (40. seriously. 8. For the purposes of this simple example. t ) for age 40 (left) based on Eq.7). (2. (2. They highlight three major features of our model: first. One can use the improvement in mortality to identify average rates but it is not clear how one can identify the frequency with which longevity events occur. Specifically.2 0 0 . A simple model Although we envision the primary role of our model as a risk management tool for the generation of a robust set of plausible mortality scenarios. it is not difficult to include K (x. Among all ages. we set their F (x) = 0. t ). Furthermore. Naturally. the one just below it is µ(50. as opinions vary widely regarding future health care system and pandemic. These are simulated sample paths of the force of mortality. Second.4 0 0 . The right graph of Fig.1 0 0 . This can be done using jump events in longevity and adverse mortality and a fixed base mortality rate. t ) in year t = 2014 but µ(40.2 0 0 .

Swiss Re. AXA issued its first catastrophe mortality deal—the Osiris bond (Lane and Beckwith. At maturity. The example provides a longevity hedge for the XYZ pension plan. and in November 2006. they are similar in their longevity behavior. The design of the EIB bond may be problematic. the one just below it is µ(50. the EIB longevity bond did not sell. Possibly inspired by the successful securitization of catastrophe mortality risks. In summary. then XYZ Pension gets £10. Cox et al. covering the entire survival payment. For example. Lin and Cox (2005) and others. t ) and µ(30. Although we know what the product of these two values ought to be.6 0. The left figure is the function A(x).2%. in November 2004. Unlike the Swiss Re mortality bond. The right figure is the the function B(x).010 0. For example. the first pure death-risk linked deal.S.H. Market development Financial innovation has led to the development of several classes of mortality securities.2%. at maturity. 3. one must judge the parameters by the qualitative features of the model. (2. The floating-rate payer XYZ pays an amount proportional to the mortality rate at maturity on December 31.000. t ).000 × 1. Securitization of mortality and longevity risks 3.1. Life insurers can use q-forwards to hedge mortality risk in a similar way. Milevsky and Promislow (2001).000 at settlement of the q-forward contract.000 × 1% × 100 = £10.2% of those aged 65 in 2015 die within the year.0 0 20 40 60 80 100 20 40 60 80 100 Fig. Capital market solutions for unanticipated longevity risk have been explored relatively recently. / Insurance: Mathematics and Economics 46 (2010) 242–253 1.P. 2003. Fig.P. Evolution of mortality for a simple model. 2025 2045 2065 2085 2105 Fig. If one compares this with the trajectories from the basic Lee–Carter model. t ). If the reference rate is higher than the fixed rate 1.6 0. 2007). J. The Actuary. 2004).2 0. the European Investment Bank (EIB) offered the first longevity = 50. See Bauer and Kramer (2007) for more details. The functions A(x) and B(x) for a very simple calibration process.06506.000.4 0. we cannot identify them individually. a more satisfactory solution may be a hedge using. some life insurers started to reduce their extreme mortality exposures through financial markets. See Table 2 for a summary. Better mortality means fewer than 1. in May 2006 Scottish Re sold a mortality bond with two tranches via a special purpose vehicle called Tartan Capital Ltd. Consequently.2% × 100 − 50.8 0.P. for example.000. (2007) describe an example with a 10-year qforward written on the one-year death rate q65 for males aged x = 65 in England & Wales. and XYZ will have more than expected pension benefits to pay if it has a similar experience. In this case.11) and the parameters described earlier.000 2005 bond to provide a solution for pension plans to hedge their longterm systematic longevity risks. especially in the early contract years. in the LifeMetrics system.005 0.P. The top path is the simulated force of mortality µ(60.4 249 0. 9 shows simulated mortality trajectories for the model parameters that we have described. which is the difference between the fixed amount and the floating amount. Morgan q-forwards with a settlement date 10 or more years in the future.000. the normalized annual average rates of mortality improvement in µ(x. an exchange of a fixed amount based on the predetermined mortality rate. insurers or pension plans can take a position 9 Available at www.2 0. Following Swiss Re. 3. was issued in December 2003 (Stanley.0 1. The Swiss Re bond.com. The q-forward was opened on December 31.000. the contract is settled at the net amount. if the reference rate is only 1%. The EIB bond provides ‘‘ground up’’ protection. rather than hedging the whole benefit liability. first appearing in articles by Blake and Burrows (2001). 2006 with a notional amount of £50. t ) across ages.. giving the impact of the 1918 worldwide flu on µ(x. based on Eq. After successfully issuing the first-ever pure death-linked security.000 and a fixed death rate q65 = 1.2. Morgan q-forwards A q-forward contract requires. Morgan. t ). the settlement is negative and XYZ pays J. 8. Since the plan can predict the number of survivors to some extent.lifemetrics. in return for a variable amount based on a realized mortality rate published at a specified future date by J.8 0. Morgan. . Swiss Re sold two new mortality bonds with different tranches in April and December 2006 (Lane. 9. then µ(40. 2003.9 Coughlan et al. 2016.0 0.000. 2006). calculated as the net settlement amount: Net settlement = Fixed amount − Floating amount longevity jumps and a jump size y = 0. 0. J. determined by the reference rate in 2015. t ) for different ages. mortality increases and it is likely that the pension plan pays lower pension benefits.

2) 10 The notation I is the indicator function taking the value 1 if the event A occurs A and the value 0 if not. The first index value is p55. we use the estimated parameter to price a longevity option. For example.000. Thus. XYZ has to estimate the number of survivors to age 65 in 2015 in order to determine the size of the hedge. There are two practical problems with the q-forward (or analogous options). 2006 December 31. q65. based on LifeMetrics or government data. Generally pension plan participants have higher survival rates than the population as a whole. We further assume the annuity ¯ benefit is s. (4. we use the indifference pricing method to price mortality and longevity securities. To address these problems. The example covers survivor risk only for one year. .2% J.2006 .H.2015 1 − q56. However. Then.2005 = 1 − q55.10 The aggregate present value is M and it is based on the national population experience during the hedge period from 2005 to 2015. XYZ knows the number of participants. The q-forward contract could be modified to create q-calls and q-puts. When the longevity security is opened in 2005. (2007).2015 but with no problem in determining the hedge size. the present value of future annuity benefits for an annuity written on M lives (xi ) at time 0. While XYZ may be more concerned about rates 10 years in the future. This gets more difficult as the hedge is extended because of the problem of estimating the number of survivors. Morgan GBP 60.012 × 50.000 December 31. Forwards or options can be written on this index just as easily as they might be written on q65. as an alternative risk management method and a new investment opportunity. However. . Let us consider one participant with 1 unit of annual benefit paid at age 66 in 2016. basis risk may be a problem. their gender.000 LifeMetrics male q65 in the reference year for England & Wales national population XYZ Notional amount × Reference rate × 100 Net settlement = Fixed amount − Floating amount (as fixed amount payer or floating amount payer) in q-forwards to hedge their liability to pay benefits to their clients.2006 1 − q57. those participants are age 55. The well-known capital asset pricing model (CAPM) lacks the flexibility to be applied appropriately to heavytailed insurance risks. When the q-forward is opened in 2005. Cox et al. a call option could pay XYZ when the realized population index 11 p55. 3. / Insurance: Mathematics and Economics 46 (2010) 242–253 Table 2 An illustration for a q-forward to hedge longevity risk for a pension plan. The index allows XYZ to hedge underestimating 55 this value (and underfunding its liabilities).2005 . for i = 1. has gained more and more attention from both scholars and practitioners.2005 extend the hedge to the second benefit payment with an option on the next index value 12 p55.000.2005 ¯ Xi = s k e−krk I{T (xi )≥k} . we combine the indifference pricing method and our proposed stochastic mortality model. exactly the risk the q-forward is supposed to hedge. The subsequent index values are calculated analogously.1.000 = 0. published in 2016. . q57.P. 4. .2015 . . . 4. As before. Moreover. . 2. XYZ can use the index directly to hedge its payments to the cohort of plan participants age 55 in 2005. . The example only hedges survivor risk from age 65 to age 66.250 S. denoted 11 pxyz .2005 which is published in 2007. Notional amount Trade date Effective date Maturity date Reference year Fixed rate Fixed amount Payer Fixed amount Reference rate Floating amount Payer Floating amount Settlement GBP 50.000. The population index 55 is likely to be in the money in the same circumstances that the plan has experienced unexpected increases in longevity. since the two groups are subject to some of the same forces of mortality. While XYZ could open a portfolio of q-forwards on the indices q64+k. The same techniques can be used to price other mortality securities. Indifference pricing method We consider longevity risk in a life-only single premium immediate annuity (SPIA) for age x. XYZ will have an estimate of the probability that a participant now age 55 survives to age 66. . paid at the end of each period. XYZ has a liability to make payments to a plan’s participants who survive to age 66 in 2016. Accordingly. We show how to estimate the risk aversion parameter. and projected benefits. described by Coughlan et al. it would also have to forecast the number of survivors to each age in order to determine the size of each qforward contract. q56. using annuity market data. Securities based on a longevity index Consider the XYZ plan again. Using the population index.2014+k for k = 1. The index for this hedge is based on the government or Lifemetrics series of future death probabilities q55. In its funding calculations. published each year from 2006 to 2016. 2006 December 31. (4. . . to develop this emerging market.3. many insurance principles lack the ability to produce arbitrage-free prices. We show how to price longevity options in the next section.2005 1 − q65. M is = 1 − q56.2005 .2007 . exceeds a strike rate such as 11 pxyz (which the plan uses in 2005).1) = 1 − q55. The index value for hedging the payment in 2016 is 11 p55.2005 . XYZ could just as easily Y = i =1 Xi . 2016 2015 1. On the other hand. To illustrate the idea. reduces moral hazard since the index is transparent to all investors. so there is no problem in determining the size of the hedge. . 2. they also have longevity risk during the 10 year contract period beginning in 2005. XYZ could estimate this relationship statistically and use this relationship to adjust its hedge.2007 · · · where rk is the default-free interest yield from time 0 to time k and k runs over the payment times.2006 p55. . published in 2006. The other problem is that it is difficult to extend the q-forward hedge. The second index value is 2 p55. actuaries and financial economists have begun to make considerable efforts to improve its structure and take on the challenges to connect financial and insurance pricing theories.2005 . Mortality securitization modeling Insurance-linked securitization.

An approximation for P. In this example. and estimate E[eα0 Y |Θj ] with the sample mean.6). given a mortality scenario Θ . we used the population mortality tables rather than annuity tables because we do not have a time series of annuity tables. u(w) = 1 − e−αw 1500 1000 α .000 of scenarios Θj .40 = $1600 per month.2. even when the lives are conditionally independent. and the Genworth annuity premiums do not appear to contain a material charge for this tail risk. the premium still depends on the size of the portfolio as indicated by (4. Then we use the secant method to find a third approximation. The risk aversion parameter α in Eq. (4. using Eq. the indifference premium under exponential utility is given by (4.40 per $1000 premium (Stern. The vertical axis represents utility and the horizontal axis represents wealth. Since aggregate premium is $560 million.000 = 2240. which can be simulated with our model. in general. As we have noted. both in thousands. a significantly lower value of α is obtained because of the greater tail risk. It seems reasonable to believe that our estimated α = 4. We assume that the policies are identical.1). As we do not know the mortality assumptions that Genworth used in setting their annuity rates. P1 − P0 α 2 E[Var(X |Θ )] + Mα 2 Var[E(X |Θ )]. In this case. Although we show how to determine α with an annuity market quote. (4. from Eq. Then we simulate M = 2240 values of eα0 Y |Θj . 10. (4. population mortality tables will tend to underestimate the longevity risk in annuities relative to the applicable set of annuity tables. α E[Y ] + α 2 Var(Y ) . t ) : t ≥ 0}. the unscaled utility function for Genworth is shown in Fig. there is no variance in E(X |Θ ). then the number of policies is M = 560 million/250. (2. . (4. In the annuity pricing application we are considering. (4. the same technique can be applied to estimate life insurance risk aversion parameters.8 × 10−7 . We start with a trial value of α = α0 and estimate the corresponding value of E(eα0 Y ) = E[E(eα0 Y |Θ )].375. so the net premium per policy to the insurer is P = 250. This is a very fast calculation.8 × 10−7 .7). / Insurance: Mathematics and Economics 46 (2010) 242–253 251 The premium P per policy is to be determined from the equation: E[u(w + MP − Y )] = u(w). given the mortality table.H.1225) = $219. 4.4) 500 where α is the risk aversion parameter of the insurer. mortality scenarios are uncertain and the premium P will depend on the portfolio size M as indicated by (4. Genworth’s estimated utility function under the exponential utility model for α = 4. (4. Under our exponential utility function. For the value of α that was determined above. The estimate of E eα0 Y is the mean of these sample means. 2008). 000 × (1 − 0. (4. thereby tending to create an overestimate of α . Indeed. which is around the industry average. we need to determine α so that Eq. is P0 = 1 M α0 log E eα0 Y . When calibrating to market data.000.S. and the premium P per policy does not depend on the number of polices M.3) 2000 where w is the wealth of the insurer prior to accepting the risk to be priced. This process converges rapidly to α = 4. We will use this value in the next section to estimate what Genworth would pay for a longevity option. The reason that we calibrated α using Lee–Carter projections rather than the full model is that the full model induces significantly thicker longevity tail risk than Lee–Carter projections alone. with a ‘‘Monte Carlo simulation within a Monte Carlo simulation’’. ¯ Eq.8 × 10−7 can serve as an upper bound for Genworth’s risk aversion to annuity longevity risk.6) The mortality is uncertain but it is reasonable to assume that. the third term is zero.3) can be approximated using known insurance market prices. The random variable Y depends on the random path of future force of mortality values {µ(x + t . the annuitant lives are independent. However. (4. which can be found in Gerber and Pafumi (1998).3) may be expressed as P = 1 Mα log E eα Y . The corresponding model price. (4.5) by simulation. then obtaining P1 .5) 0 0 1000 2000 3000 4000 5000 Fig. On a static basis. The approximation becomes P ≈ E[X ] + α2 = α1 − α1 − α0 (P1 − P ) .25%. Cox et al. However. if α is estimated under the full model. Estimating α from annuity prices The Genworth Financial Group sold around $560 million of single premium individual annuities in the US in 2005 with a monthly payout rate of $6. We use the exponential utility function. We estimate the risk aversion parameter α from Eq. 10. 2 1 Since the aggregate loss is composed of M identical risks. we estimated α using Lee–Carter projections for mortality fluctuations which embeds longevity risk. for each j.5). = E[MVar(X |Θ )] + Var[ME(X |Θ )] = ME[Var(X |Θ )] + M 2 Var[E(X |Θ )]. can be made as P ≈ 1 Mα We assume the underlying expense factor is 12.1) gives the present value of benefits for one life with s = 250 × 6. we are pricing an aggregate annuity exposure for which the component risks are identical but not necessarily independent.7) When the mortality scenario is known (as in traditional actuarial calculations with static tables).10). the variance of Y is the sum of two terms: Var(Y ) = E[Var(Y |Θ )] + Var[E(Y |Θ )] We repeat the calculation with another trial value α1 . E[Y ] = ME[X ] and thus P ≈ E[X ] + α 2M Var(Y ). the premium implied by (4. First we generate 10. (4.5) is satisfied. Moreover. this exercise must be viewed as an illustrative approximation. issued to males age 65 with a gross premium of $250.

For example.76 0. In all cases.H. Reinsurance contracts typically have a cap on the reinsurer’s liability.780 $77. p1 and an upper strike level. Range of outcomes for 10 p65. Note that the spreads in Table 5 have different.2005 > p otherwise.2005 . (4. Fig.864 0.252 S. (4.364 0. . The dollar payoff from this longevity call at maturity in 2015 may be expressed as C = 2 0 0. It is reasonable to anticipate that mortality traders may follow this tradition.2005 < p2 if 10 p65. Table 4 shows longevity call option prices for a range of strikes for mortality driven by the Lee–Carter model alone. or ‘‘detachment point’’.9 0. decreasing seniorities due to decreasing trigger and exhaustion levels.790 $5. 5. (4.000. Cox et al. Evidently. p2 and a notional amount of $100. ‘‘attachment point’’. Range of outcomes for 10 p65. / Insurance: Mathematics and Economics 46 (2010) 242–253 x 10 4 Table 3 10-year longevity call option prices with data generating process driven by our model.85 0.8 0. given the strike level p = 0.2005 under the Lee–Carter model alone.78 0. Strike rate p Call price 0. In the trading of standardized equity and interest rate derivatives.775 and the US treasury discount factors for the 2005 calendar year.000.3 with a notional amount of $100.75 0.000.3) to price this longevity call option for a mortality data generating process driven by our mortality model and by the Lee–Carter model alone.2005 for 1 million draws using the full model.8 0.000.8 × 10−7 . While no one can be certain of the true nature of longevity risk. Consider a 10-year call option issued in December 2005 based on our longevity index 10 p65. In particular.364 if the mortality data generating process is driven by our model.3. one discovers that market prices reflect a ‘‘smile’’.2005 ≥ p2 if p1 < 10 p65. spread A has the lower attachment and detachment points relative to the other spread options.000 0 10 p65. x 10 4 8 7 100.136 0.7 0.288 0.775 $1121.785 $772.506 0. Summary We have described a mortality model which provides a rich and realistic space of sample paths of future mortality rates.522 0. Fig.795 $937 10 8 6 4 4. in practice the owner of the option is most likely concerned with the total dollar exposure as it relates to a pension fund or annuity pool liability. Strike rate p Call price 0. it is the right tail behavior that is driving the differences between the prices of the longevity calls across the two mortality data generating processes. 11 shows a histogram of simulated values of 10 p65.8) 6 5 4 3 2 1 0 0. We then fit a Nelson-Siegel yield curve to this zero-coupon data which can be extrapolated beyond the last observation. 11.000 and a strike price of p.2005 − p1 if 10 p65.82 10-year Survival Probability for (65) Fig.795 $553.9) Table 5 shows the longevity call spread option prices for a range of strikes for mortality driven by our model. and thus it demands a higher price.2005 under the full model. in line with the estimate that was made for the Genworth Financial Group.2005 as defined in Section 3. Longevity option price Let us now illustrate the pricing of longevity options using indifference pricing with the pricing parameter α that we have previously derived from annuity prices. We take the average US treasury coupon-curve yields for 2005 as published in the Federal Reserve’s H15 data set and strip them to produce zerocoupon data. The dollar payoff from this longevity call spread at maturity in 2015 may be expressed as B= 100. If the market for longevity derivatives should evolve to a level of reasonable liquidity. 12.000 0 10 p65.2005 −p if 10 p65.785 $23. one finds that is it far more expensive to purchase ‘‘lottery tickets’’ on equity indices or interest rates than basic theory suggests. the nature of the mortality data generating process is critical in assessing the risk in a longevity derivative contract. The typical reinsurance contract is analogous to a call spread.72 Although one can attribute a ‘‘cohort’’ to the payoff amount.070 0.000. we require discount factors that are representative of interest rates for 2005. At this point we have a set of discount factors that is representative of market conditions in 2005.2005 for 1 million draws using the Lee–Carter model alone. Consider a call spread written on the survivor index 10 p65.95 10-year Survival Probability for (65) Fig.000.11 our 10year longevity call option price equals $1121. with lower strike level.345 18 16 14 12 Table 4 10-year longevity call option prices with data generating process driven by the Lee–Carter model alone. 0.790 $652. The model evolves as a dynamic process that combines a general mortality 11 To be precise. Evidently. In simple terms.962 0. our model has shown that the presence of relatively infrequent mortality improvements poses a significant risk.780 $924. it would not be surprising if similar effects were observed. with two calls having different strike prices. We apply Eq.74 0. we assume that the market risk aversion parameter is α = 4.775 $191.2005 ≤ p1 .000 (p2 − p1 ) 100. 12 shows a histogram of simulated values of 10 p65. Table 3 shows longevity call option prices for a range of strikes for mortality driven by our model.

2008. Ballotta.780 0. C. MacMinn. 2006.. T. 79–120.. Financial Engineering News. 108–126. In: Transactions of the 26th International Congress of Actuaries. Pension Capital Strategies.. Bourguignon. 1998.. URL www. Avian influenza: Preparing for a pandemic. Goss. H. Gingrich.L. 193–217. 2005. D. Cox. Lee..N. J. See the news release for December 8 2003. Morrisson...D. North American Actuarial Journal 11 (3).. 285–312. 1996. Pitacco.. 2001.. Biffis. IL.D.. 1375–1378.H.. Schoen. K. The fair valuation problem of guaranteed annuity options: The stochastic mortality environment case. A.. Spread A Attachment point p1 Detachment point p2 Price 0. Emerging Infectious Diseases 12 (1). 2005. That was the year that was! The 2007 review of the insurance securitization market. E. Bauer. 2003. L. Vere-Jones. Available at http://www... valuation. J. Dowd.G.H.. Vaupel. Insurance: Mathematics and Economics 29 (3)..J..W. 647–672. D. S.G.N.. British Actuarial Journal 2 (2).H. Will mortality improvements continue? National Underwriter 106. 299–318. An Introduction to the Theory of Point Processes. Lin.. 2007. P. American Council of Life Insurers. 2006. 2007. 2004.785 0. 2002.. Beckwith.. E..R.N.. Toole. Blake.. Dynamic Population Models. 2006. I. 2003. S. England. Insurance: Mathematics and Economics 39 (2). Daley.K. D.. Lifemetrics: A toolkit for measuring and managing longevity and mortality risks.fenews. 227–252. Sithole.G. N.. Multivariate exponential tilting and pricing implications for mortality securitization. 2002.C. Cowling. Longevity risk and life annuity reinsurance. R.. January/February. D. Insurance: Mathematics and Economics 42 (2).. P. A. J.J. Millossovich. American Family Physician 74 (5). Lane.. S.. To hedge or not to hedge: Managing demographic risk in life insurance companies.. Hevia-Portocarrero... In the core of longevity risk: Hidden dependence in stochastic mortality models and cut-offs in prices of longevity swaps. Richardson. M. Wang. North American Actuarial Journal 2 (3). Cox et al. W.lifemetrics.-W.. P. 74–92. S. Y. Swiss Re obtains $400 million of mortality risk coverage.. Kramer. 2007. B. Miltersen. Risk and valuation of mortality contingent catastrophe bonds..835 253 trend.M.. The ASTIN Bulletin 36 (1). S. Clinical Infectious Diseases 33 (8). How high is up? The 2006 review of the insurance securitization market. URL http://www. Loise. 6.Z. JPMorgan. P. Haberman. Ong.790 0. Cutler. Schulze. P. Philipson. Estimating the term structure of mortality. Coughlan.. 2004. Wilmette. S. Cox. 727–744. 2003. Khalaf-Allah. Swiss Re and Mitsui Sumitomo arrange USD 100 million catastrophe risk swap. F. Germany. E... University of Bergen. R. Annuity shopper.. Securitization of catastrophe mortality risks. 2001. Insurance: Mathematics and Economics 27 (3). S.. 295–306.. Y. S. Cox. 2006. E. 15–22... Natural hedging of life and annuity mortality risks. Renshaw.-A. W.. A.. M. Observations on mortality during the 1918 influenza pandemic. Pitacco.A... Insurance: Mathematics and Economics 35. American Economic Review 92 (1). 2008. http://www. A. .com. Journal of Risk and Insurance 73 (4).. E.. T. 2006. Lin. 97–100. S.. Ulm University. H. Lane... R. The model is well-suited to risk management application where there is a need to test extreme but plausible mortality scenarios.soa. Juckett. Swiss Re innovative mortality-based security. 1918 influenza: The mother of all pandemics. Milevsky. Springer. Valuation and hedging of life insurance liabilities with systematic mortality risk.. 1–25. Université Claude Bernard Lyon 1. 2003. Technical Report. F. Working Paper. Hári. Lin. R. J. 16. F.-M. 2000... Historical and projected mortality for Mexico. References ACLI 2006. 19–41. L. Gerber.780 $122.785 $70.. W.. January. Marocco. Blake. Blake. 11–13.. 2006.org/ research/life/research-impact-pan-influ-life-ins. S.. Journal of Risk and Insurance 73 (1). K. Burrows....337 Spread D 0. Birmingham. K. See the news release for August 4. Life Insurers Fact Book 2006. North American Actuarial Journal 4 (1). Carter. D..com. D. Olivieri. 2008. 1998. 2006.. A.. American Economic Review 88 (2). Canada. Lane Financial LLC. S. Journal of the American Statistical Association 87 (419).J. In: XXVIIth International Congress of Actuaries. Modelling and forecasting the time series of US mortality. Research report.swissre. 2000. The rise in old-age longevity and the market for long-term care. E. with various extensions and applications. S.. Taubenberger. Dahl. Joseph. Pafumi. M. J. M. G. vol. Stochastic mortality in life insurance: Market reserves and mortality-linked insurance contract. Promislow. Furthermore.405 Spread C 0. Rogers. Longevity bonds: Financial engineering.W. The Netherlands. Mortality derivatives and the option to annuitize. 2006. 2005. R. Springer-Verlag. 453–479.. J. North American Actuarial Journal 2 (4). http://www. T. R. Wilmette. 1998. 1998. 719–736.. Inequality among world citizens: 1820–1992. pp. 2008. Sinha. A. Technical Report. 2007. The Actuary. Journal of Risk and Insurance 72 (2). 1992. Thompson. A. Dowd. 659–671. A. we describe a parsimonious version of the model based on judgment and historical mortality data available through the Human Life Table Database and the Human Mortality Database. April. 2–13. 443–468.htm.. T. American Economic Review 92 (4). R.. Pension Markets in Focus 3.H. Insurance: Mathematics and Economics 37 (3). 80–93.. 1–15. Stern. D. 195–214. Influenza pandemic periodicity... Morens. D. 2007. Inference about mortality improvements in life annuity portfolios. Dales. 2006. and the art of risk assessment. D. Journal of Risk and Insurance 73 (4). Haberman. Y. with applications to immediate annuitants’ and life office pensioners’ data. Insurance: Mathematics and Economics 38.. The modeling of recent mortality trends in United Kingdom male assured lives. 2007. Survivor bonds: Helping to hedge mortality risk. Melenberg.S. 2006. Stanley. vol. and hedging.U..gov/ ncidod/EID/vol12no01/05-0979.. Hardy.795 $25. and a temporary mortality jump process.. Nijman. Y. M. as we illustrate with the indifference pricing method. Dordrecht.R. M. A bidimensional approach to mortality risk. Emerging Infectious Diseases 12. Persson.. 339–348. American Economic Review 88 (2). The Actuary. Epstein. Securitization of mortality risks in life annuities.. Y. 2006. 113–136. Wade. Post.. the model is tractable enough to allow path-by-path simulation which can be combined with pricing techniques. 2005. Bell.. D. IL.D. 242–247..E. January. A matter of life and death.htm.. Is mortality dead? Stochastic forward force of mortality rate determined by no arbitrage. The FTSE 100 and their pension disclosures. Luk. Verrall.775 0.. Hu. Lakdawalla. Overview of the financial wealth accumulated under funded pension arrangements. T. 2001. Salou. 2002. Hatzoupoulos. The value of health: 1970–1990. Morgan. Waegenaere. Serant. Utility functions: From risk theory to finance.cdc. 783–790. 628–637. 449–477. C. M. 492–504.com Inc. New York. Gründl. / Insurance: Mathematics and Economics 46 (2010) 242–253 Table 5 Prices for 10-year longevity call spread options. May. 1998. H. virus recycling.. Haberman. a diffusion process. 2002. Lee. 2006. Working Paper. Technical Report. Dahl.. Dowdle. Lin.. Cox. D... Lane Financial LLC.H. Biffis. Society of Actuaries. Working Paper.com/fen46/topics_act_analysis/topics-act-analysis.. Published semi-annually by Web Annuities. P. An investigation into parametric models for mortality projections. R. Pricing death: Frameworks for the valuation and securitization of mortality risk. Affine processes for dynamic mortality and actuarial valuations.. Møller.W. Cairns. Demographic analysis of aging and longevity.985 Spread B 0. Swiss Re. 34–39. Another novel aspect is that our model includes uneven effects of a mortality or longevity shock to different ages.aspx. G. In addition.. Technical Report. and the United States.D. G.790 $39. Decisions in Economics and Finance 29 (2). 2006. R. Journal of Risk and Insurance 68 (2).. a permanent longevity jump process.M. Gross. Potential impact of pandemic influenza on the US life insurance industry. Insurance: Mathematics and Economics 42 (2). Cairns. March. The Lee–Carter method of forecasting mortality.