Valuation, Hedging and The Risk Management of Insurance Contracts (Doctoral Thesis) - by Jérôme Barbarin, Université Catholique de Louvain, 2008

Université Catholique de Louvain
Institut des Sciences Actuarielles
Valuation, Hedging and the Risk Management of Insurance Contracts.
Membres du jury: Thèse présentée en vue de

l’obtention du grade de
Prof. Pierre Ars Docteur en Sciences
Prof. Griselda Deelstra (orientation sciences actuarielles)
Prof. Michel Denuit par:
Prof. Pierre Devolder Promoteur
Prof. Ragnar Norberg Jérôme Barbarin.
Louvain-La-Neuve
Mai 2008
A mes parents,
"On se lasse de tout sauf de comprendre",

Virgile.
A François et Louise,
"Rêve de grandes choses, cela te permettra d’en faire au moins de toutes petites",
Jules Renard.
Preface.
This dissertation has been prepared in partial fulfillment of the requirements for the Ph.D.
degree at the Institute of Actuarial Sciences, of the Université Catholique de Louvain. The
research was carried out in the period from September 2003 to December 2007 under the
supervision of Pierre Devolder, professor and president of the Institute of Actuarial Sciences.
From September 2005, it was financed by Axa Belgium under the “Axa Belgium Chair in Risk
Management”.
Acknowledgment.
First of all, I would like to thank my supervisor Professor Pierre Devolder, who gave me the
oppportunity to start this PhD thesis. I would like to thank him for his trust, for his constant
support and for his valuable suggestions and advices. I am specially grateful for the freedom
he left me to complete this work.
My gratitude also goes to the members of my doctoral committee, Professors Pierre Ars,
Griselda Deelstra, Michel Denuit, and Ragnar Norberg for the time they spent reading my
thesis. A special thank is due to Professor Ragnar Norberg for coming from London to attend
my PhD defense and for all his detailed remarks and advices on my work.
I would also like to thank Professor Thorsten Rheinlander for inviting me to present my
work at the London School of Economics and for all his advices, encouragements and kind
words.
During the course of this PhD, I spent two years at the “Institut d’Administration et de
Gestion” (IAG) of the Louvain School of Management as a teaching assistant of Professor
Philippe Grégoire. I very much enjoyed this couple of years at the IAG and Professor Gregoire
deserves all my gratitude for welcoming me in the Finance Unit.
For the rest of my PhD, my work was financed by the Chair in Risk Management awarded
by Axa Belgium. I would like to thank Axa Belgium for its financial support and in particular,
Mr. Alfred Bouckaert and Mr. Serge Wibaut for the creation of this chair.
I would also like to acknowledge all my friends from inside the academic world, those of the
Louvain School of Management, Hervé Van Oppens, Olivier Vercruysse and Tanguy de Launois
as well as those of the Institute of Actuarial Sciences, Donatien Hainaut, Jean-Philippe Boucher,
Aurélie Miller, Julien Trufin, Cindy Courtois and Xavier Maréchal. I would always remember
with great pleasure, the passionate discussions that we had Donatien, Jean-Philippe and me
about finance, economics or politics.
Monique Descamps also deserves my warmest thanks for her constant help and availability
as well as for the lively atmosphere she brought to the institute.
I also wish to thank all my friends from outside the academic world: Xavier, Benoît, Laure,
Christophe, Vincent and Valérie.
Last but not least, I would like to thank all my family and, in particular, my parents for
their constant support in the good and the bad times, and for always having been there for me
whenever I needed them. How could I ever acknowledge them enough?
Contents
Preface. 3
Acknowledgment. 3
Introduction. 1
Part 1. Life Insurance Contracts and the Interest Rate Risk. 7
Chapter 1. Risk Measure and Fair Valuation of an Investment Guarantee in Life

Insurance. 9
1. Introduction. 9
2. Cash-Bond-Stock Model (CBS). 10
3. Investment Guarantee and Risk Measures. 17
4. Fair Valuation and Participation Rate. 19
5. Numerical Illustrations 23
6. Conclusions and Extensions. 31
Appendix 1. 32
Appendix 2. 37
Part 2. Life Insurance Contracts and the Surrender Risk. 41
Chapter 2. Risk-Neutral Valuation of Life Insurance Contracts with a Surrender Option. 43

1. Introduction. 43
2. the Theoretical Framework. 44
3. Risk-Neutral Valuation, Absence of Arbitrage and Enlargement of Filtration. 46
4. The Surrender Time. 50
5. The Risk Neutral Valuation Formulas. 56
6. The Brownian Filtration. 59
7. Application to Unit-Linked Contracts. 64
8. Conclusion. 77
Appendix 1. 77
Appendix 2. 79
Appendix 3. 81
Appendix 4. 83
Appendix 5. 86
Appendix 6. 87
5
6
Appendix 7. 89
Appendix 8. 89
Chapter 3. Risk-Minimizing Strategies of Life Insurance Contracts with Surrender

Option. 91
1. Introduction. 91
2. The Theoretical Framework. 92
3. The Risk-Minimization Theory. 94
4. Risk-Minimizing Strategies for a Single Life Insurance Contract. 96
5. Risk-Minimizing Strategies for a Portfolio of Life Insurance Contracts. 107
6. The Risk Process for a Portfolio of Life Insurance Contracts. 116
7. Conclusion. 119
Chapter 4. Locally Risk-Minimizing Strategies of Life Insurance Contracts with Surrender

Option. 121
1. Introduction. 121
3. The Local Risk-Minimization Theory. 125
4. Application to a Life Insurance Contract When the (H)-Hypothesis Holds. 131
5. Application to a Life Insurance Contract When the (H)-Hypothesis Does Not Hold. 144
6. Conclusion. 157
Appendix 1. 157
Part 3. Life Insurance Contracts and the Longevity Risk. 161
Chapter 5. Risk-Minimization with Systematic mortality risk. 163

3. The Risk-Minimization Theory. 169
4. Application to a Life Insurance Contract When the (H)-Hypothesis Holds. 171
5. Application to a Life Insurance Contract When the (H)-Hypothesis Does Not Hold. 179
6. Conclusion. 186
Appendix 1. 186
Appendix 2. 187
Appendix 3. 190
Chapter 6. Heath-Jarrow-Morton Modelling of Longevity Bonds. 193

2. A Critique of the literature. 196
3. The HJM Methodology For Longevity Bonds. 201
4. Asset Allocation with Longevity Bonds. 217
Appendix 1. 229
Appendix 2. 231
Contents 7
Part 4. Non-Life Insurance Contracts and the Inflation Risk. 235

Chapter 7. Risk-Minimization with Inflation and Interest Rate Risk. 237
2. The Model. 239
3. The Risk-Minimizing Strategies. 242
4. Applications in Some Collective Models. 247
5. Applications in Some Individual Models. 253
Appendix 1. 261
Appendix 2. 264
Appendix 3. 266
Appendix 4. 266
Appendix 5. 267
Appendix 6. 267
Appendix 7. 268
Conclusion. 271
Bibliography 273
Introduction.
The past three decades have seen a tremendous development of finance theory based on
the “no-arbitrage” hypothesis. This simple and natural hypothesis proved to have far-reaching
implications in various areas. In particular, it profoundly influenced the actuarial academic
research as well as the actuarial practice. It provided new insights and raised new deeper
questions regarding the valuation of insurance liabilities and the management of the assets
backing these liabilities. This thesis aims at contributing to the study of these questions in
particular and to the study of the interplay between finance and actuarial science in general.
This thesis consists of four parts, each devoted to a specific topic, important for the risk
management of an insurance company. In the first part, we study the impact of the risk
management policy on the pricing, in terms of a guaranteed rate and a participation rate, of
classical single premium life insurance contracts with profit. In the second part, we study the
management, in terms of evaluation and hedging, of the surrender option. The third part is
devoted to the management of life insurance contracts with systematic mortality risk. In the
fourth part, we study the asset allocation problem of a non life insurance company when the
inflation risk and the interest rate risk are taken into account.
One of the main tools applied in this thesis is the Risk-Minimization Theory of Föllmer and
Sondermann [44]. This theory provides, for a given contingent claim, the value of the portfolio
and the associated hedging strategy that minimizes a particular quadratic criterion. This theory
has been applied in numerous papers in the financial literature as well as in the actuarial
literature. Various arguments justify the study of this theory. First, this risk-minimization
theory is interesting in itself because it extends, as shown by Møller in [69] and also in Chapter
3 of this thesis, the traditional actuarial premium principle and reserving principle based on
the law of large numbers. More importantly, this theory is also interesting because it plays a
key role in the solution of various and seemingly unrelated other problems. Firstly, it appears,
for example, in other quadratic hedging theories. More specifically, when the financial asset
prices are continuous, the local risk-minimization theory of Schweizer can be written as a
risk-minimization problem under the so-called minimal martingale measure. Also, finding the
mean-variance hedging strategy of a given contingent claim basically comes down, under the
same condition, to finding the risk-minimizing strategy of this contingent claim under the so-
called variance optimal martingale measure. Secondly, the risk-minimization theory is also a
key element in the indifference pricing theory. This is not surprising when the utility function
is quadratic since, in this case, the indifference pricing theory comes down to finding the mean-
variance value and its hedging strategy. More surprisingly, the risk-minimization theory also
appears in the indifference pricing when the utility function is exponential. Indeed, Mania
1
2
and Schweizer show in [65] that, in this case and for a small risk aversion, the indifference
value of a given contingent claim and its associated hedging strategy are given by, respectively,
the risk-minimizing value of the same contingent claim and its hedging strategy under the
minimal entropy martingale measure. To conclude, we clearly see that this theory constitutes
an important building block in many financial problems. Most of the results derived in this
thesis can thus be used, mutatis mutandis, to solve these various problems.
Each chapter of this thesis can be read independently of the others. Repetitions are thus
unavoidable. We apologize in advance for this inconvenience.
Contribution of the dissertation.
Part I: Chapter 1. The main motivation of this first chapter, is to study the pricing of
a classical single premium life insurance contract with profit. The pricing is here considered in
terms of a guaranteed rate on the premium and a participation rate on the (terminal) financial
surplus.
In this chapter, we argue that these technical parameters should be determined by taking
explicitly into account the risk the insurer accepts to bear. More precisely, given the asset
allocation of the insurer, we should fix these parameters consistently with the risk management
policy of the insurance company, in terms of a risk measure such as the value-at-risk or the
conditional value-at-risk.
In order to answer these points, we suggest to divide this problem in two steps. In the first
one, we consider the insurance contract without profit. Since, given the asset allocation, the
financial risk comes only from the guaranteed rate, we propose to fix the guaranteed rate such
that the value-at-risk or conditional value-at-risk of this modified contract, does not exceed a
certain level chosen by the risk management. In the second step, we fix the participation rate
according to the risk neutral valuation principle and according to the guaranteed rate found in
step one. In this way, we have a contract that is simultaneously fairly priced and that exhibits,
given the asset allocation, a risk consistent with the risk management policy.
We also propose in this chapter, a model for the financial portfolio that includes investment
in cash, stocks and bonds. Different kinds of bond strategies are especially developed in order
to model the influence of the matching policies.
Finally, we illustrate numerically our methodology and study the effects of the matching
policy and the effects of the strategic asset allocation on the technical parameters.
Part II: Chapters 2, 3 and 4. In this second part, we focus on the management of the
surrender option embedded in most life insurance contracts. More precisely, in Chapter 2, we
study the valuation of insurance contracts with a surrender option and in Chapter 3 and 4, we
study the hedging strategies of such insurance contracts.
In the actuarial literature, the time of surrender is usually modelled as an (optimal) stopping
time with respect to the filtration generated by the financial asset prices. In such a model, if
the financial market is initially complete, an insurance contract with a surrender option whose
exercise depends on such a surrender time, can be perfectly hedged. The surrender risk is thus
not a genuine additional risk in this framework.
Introduction. 3
In Chapter 2, we argue that, if we are to model a realistic surrender time, we should avoid
following this traditional way. On the contrary, we argue we should try to model this surrender
time as a random time not adapted to the filtration generated by the financial assets prices.
We follow here the financial literature on the default risk and in particular, the reduced-form
models. We discuss at length the pros and cons of the traditional methodology and justify
the reduced-form approach by the asymmetry of information between the insurer and the
policyholders. We also give a simple illustration of this approach. This second chapter should
be seen as a methodological one; we mainly recall some well-known results from the financial
literature on default-risk that, we think, are tailored for the modelling of the surrender time.
When the surrender time is modelled as described in Chapter 2, an insurance contract
with a surrender option cannot be perfectly hedged even if the financial market is initially
complete. It is thus important for an insurer to study the hedging strategies that would reduce
its exposure to this surrender risk.
In Chapter 3, we study the risk-minimizing strategies of such contracts. More precisely, in
a first step, we describe the risk-minimizing strategies for a single insurance contract. We show
that finding these strategies basically comes down to finding the risk-minimizing strategies
of properly modified purely financial claims. Then, in a second step, we use this result to
describe the risk-minimizing strategies for a portfolio of n insurance contracts. The important
characteristic of the surrender option is that it introduces an additional systematic risk since the
surrender times of the different policyholders depend on the evolution of the financial market
and are thus not independent of each other. However, we show that, even though the surrender
risk represents an additional systematic risk for the insurer, we can combine continuous hedging
and diversification to reduce the overall risk of the portfolio (down to a limit related to the
degree of incompleteness of the financial market). In this way, we generalize previous results
obtained by Møller [68] and Riesner [72] when the insurance payments depend on random
times (the times of death of the policyholders in these papers) independent of the financial
market and independent of each others.
In Chapter 4, we continue our investigation and focus on the “locally risk-minimizing”
strategies. By contrast to the risk-minimization theory described in the previous chapter, the
local risk-minimization theory has only been developed for a single payment at a fixed time.
The first contribution of this chapter is to extend the local risk minimization theory to payment
processes. We show that we can follow the same method than Møller [68] when he extended
the risk-minimization theory to payment processes. We then apply this extended theory to
insurance contracts with surrender option and find the form of the locally risk-minimizing
strategies. We distinguish between two cases. In the first one we assume the so-called (H)-
hypothesis holds and in the second one, we remove this hypothesis. Intuitively, this (H)-
hypothesis implies that the observation of the random time does not alter the dynamics of
the financial market that has been initially defined. As in the previous chapter, we show that
finding the locally risk-minimizing strategies comes down, in both cases, to finding the locally
risk-minimizing strategies of modified purely financial claims, though, these claims are different
according as the (H)-hypothesis holds or not. As a by-product, we also study the impact of a
progressive enlargement of filtration on the so-called “minimal martingale measure”.
4
Part III: Chapters 5 and 6. The third part is devoted to the systematic mortality risk.
Due to its systematic nature, this risk cannot be diversified through increasing the portfolio.
It is thus also important to study the hedging strategies an insurer should follow to mitigate
its exposure to this risk.
In the first chapter, we study the risk-minimizing strategies for a life insurance contract
when no mortality-linked financial assets are traded on the financial market. This problem
has already been studied by Dahl and Møller [39]. Under some simplifying assumptions, they
find closed-form solutions for the risk-minimizing strategies and the cost process (i.e. the
unhedgeable part) of some insurance payments. In this chapter, we extend their results in at
least three respects. Firstly, as far as the financial market is concerned, Dahl and Møller assume
a complete market with only two traded assets, a saving account and a single zero-coupon bond
with maturity T . The instantaneously risk-free rate is modelled as a time-homogeneous single
factor affine model of the CIR type. In this chapter, we only assume the (discounted) financial
assets prices follow an arbitrary s-dimensional local martingale. We do not assume the financial
market is necessarily complete. Secondly, Dahl and Møller assume the insurance payments,
given the times of death of their policyholders, are deterministic. We assume here that the
insurance payments, given the time of death, can be stochastic and depend on the evolution
of the financial market. Since our financial market is not necessarily complete, these payments
cannot necessarily be perfectly hedged. Finally, in Dahl and Møller, the probability distribution
function of the random time of death is described through its instantaneous mortality rate
which is assumed to be stochastic and to follow a time-inhomogeneous single factor model of
the CIR type. In this chapter, we actually start from an (almost) arbitrary stochastic process
that describes the dynamics of the whole survival probability function but without giving the
specific form of this probability function or the specific form of the intensity of the random
time of death. This model allows us to distinguish two situations. In the first one, the (H)-
hypothesis is assumed to hold whereas in the second situations we relax this assumption. To my
knowledge, this (H)-hypothesis actually holds implicitly in every current stochastic mortality
model.
In this general setting, we give the form of the risk-minimizing strategies and the cost
process. The main contribution of this chapter is to provide a clear and intuitive explanation
for the “risk-minimizing” strategies. Indeed, we show that the “risk-minimizing” strategies can
be written as an average of “risk-minimizing” strategies of purely financial claims, weighted by
the stochastic survival probability distribution function. The important corollary of this result
is that, finding the risk-minimizing strategy of a life insurance contract can be separated in
two independent problems, one purely actuarial and related to the modelling of the survival
probabilities and one purely financial related to the hedging of financial claims.
In the second chapter, we introduce mortality-linked financial assets. More precisely, we
want to study the application of the HJM methodology to the modelling of a longevity bonds
market. The idea of using the HJM methodology for modelling the longevity bonds prices is not
new in the actuarial literature. In this second chapter, we give a critical review of this literature
and show that previous models actually do not define properly neither the prices nor the payoffs
of the longevity bonds. Therefore, we try to tackle this problem. The first contribution of this
chapter is to describe a coherent theoretical setting in which we can properly define the longevity
Introduction. 5
bond prices. A second objective is to describe a more realistic longevity bonds market model
than in cited papers. In particular, we introduce an additional effect of the actual mortality on
the longevity bond prices, that does not appear in the literature. We also introduce multiple
term structures of longevity bonds instead of the usual single term structure. Finally, we
introduce non-homogeneous longevity bonds. In this framework, we derive the no-arbitrage
condition for the longevity bonds financial market. We also discuss the links between such
HJM based model and the intensity models for longevity bonds such as those of Dahl [37]
(2004), Biffis[29] (2005), Luciano and Vigna [64] (2005), Schrager [74] (2006) and Hainaut and
Devolder [49] (2007). Our results suggest that the standard pricing formula of these intensity
models could be extended to more general settings.
Finally, as a last contribution of this chapter, we study the asset allocation problem of pure
endowments and annuities portfolios. In order to solve this problem, we again study the “risk-
minimizing” strategies of such portfolios, when some but not all longevity bonds are available
for trading. In this way, we introduce different basis risks. As a by-product, we also found the
fair values of these liabilities when such a longevity bonds market exists.
Part IV: Chapter 7. Finally, the fourth part deals with the design of ALM strategies for
a non-life insurance portfolio. In particular, this chapter aims at studying the asset allocation
problem of a non life insurance company when inflation risk and interest rate risk are taken
into account. The academic literature on this topic is rather scarce. There are however good
arguments to study the asset allocation problem in non life insurance. Indeed, the cost of
claims is highly sensitive to inflation, which, in turn, is itself correlated with the financial
assets returns and in particular, with the interest rates movements. Moreover, the reserve an
insurer should maintain, computed as the expected discounted claims, is subject to interest
rate risk through the variation in the discounting term and indirectly through the variation
in expected outstanding claims due to the inflation. The application of ALM techniques in
non-life insurance is thus justified from a theoretical point of view.
Again, to study this asset allocation problem, we focus on the “risk-minimizing” strategies.
We first derive the general form of these strategies when the cumulative payments of the
insurer are described by an arbitrary increasing process adapted to the natural filtration of a
general marked point process and when the inflation and the term structure of interest rates
are simultaneously described by the HJM model of Jarrow and Yildirim [57].
We then systematically apply this result to four specific models of insurance claims. We
first study two “collective” models. In the first one, we consider a portfolio with a constant
number of policies. In the second one, the number of policies is stochastic and follows a point
process with an inhomogeneous and stochastic intensity. In each case, the cost of claim, for
a given policy, follows a compound Poisson process. We then study two “individual” models
where the claims are notified at a random time and settled through time. In the first model,
the claim, once notified, is settled according to a deterministic function. In the second model,
the claim is settled according to a Beta-Stacy process.
Papers. The chapters of the dissertation give rise to the following list of research papers:
Chapter 1: Barbarin J., Devolder P.“Risk Measure and Fair Valuation of an Investment Guar-
antee in Life Insurance”. (2005) Insurance: Mathematics and Economics, 37, 297–323.
6
Chapter 2: Barbarin J. “Risk Neutral Valuation of Life Insurance Contracts with Surrender
Option. An alternative Approach with Asymmetry of Information”. (2005) Working Paper,
Université Catholique de Louvain.
Chapter 3: Barbarin J. “Risk-Minimizing Strategies of Life Insurance Contracts with Sur-

render Option”. (2006) Submitted to Insurance: Mathematics and Economics.
Chapter 4: Barbarin J. “Local Risk-Minimization for Payment Processes. Application to In-

surance Contracts with Surrender Option or to Default Sensitive Contingent Claims”. (2007)
To be submitted.
Chapter 5: Barbarin J. “Risk-Minimizing Strategies of Life Insurance Contracts with Sys-

tematic Mortality Risk”. (2007) To be submitted.
Chapter 6 is an extended version of: Barbarin J. “Heath-Jarrow-Morton Modelling of Longevity

Bonds and The Risk-Minimization of Life Insurance Portfolios”. (2007) Accepted in Insurance:
Mathematics and Economics.
Chapter 7: Barbarin J., de Launois T., Devolder P. “Risk-Minimization with Inflation and
Interest Rate Risk. Applications to Non-Life Insurance”. (2007) Submitted to the Scandina-
vian Actuarial Journal.
Part 1
Life Insurance Contracts and the Interest

Rate Risk.
CHAPTER 1
Risk Measure and Fair Valuation of an Investment Guarantee in

Life Insurance.
1. Introduction.
Investment guarantees embedded in classical life insurance contracts are surely nowadays
one of the most important challenges for the insurance industry. In order to cope with this risk
two big paradigms seem to be available for the actuary:
(1) the risk neutral approach: under classical assumptions on the market ( completeness,
no arbitrage,..) the price of the guarantee is associated to the price of an option
and the machinery of option pricing can be used in order to valuate the product.
This philosophy has been used with success for equity linked contracts with maturity
guarantee (see for instance Brennan /Schwartz [34], Delbaen [40], Aase/Persson [2],
Nielsen/Sandmann [70]) as well as for life insurances with profit (see for instance
Bacinello [14, 15], Grosen/Jorgensen [47]).In this context computations take place
under a risk neutral probability measure. We can consider these models as directly
inspired by modern finance theory.
(2) the risk management approach: modern tools of simulation permit more and more to
generate a lot of future scenarios and to build the distribution of the final surplus of
the insurer taken into account stochastic financial models. Parameters of the contracts
are then fixed in relation with a certain level of solvency. Value at risk concept or more
generally risk measures are the central tools (Artzner et al. [10, 11], Wirch/Hardy
[83]). Of course computations are done in the real historical world. These models
directly linked with DFA (Dynamic Financial Analysis developed as well in life as
in non life insurance-see for instance Kaufmann et al. [59]) are in fact close to the
classical risk theory and the actuarial well-known concept of probability of ruin.
These two methodologies have already been compared in terms of reserving (Boyle/Hardy [31]).
The purpose of this chapter is to propose a way to combine these two methodologies in order to
fix the technical parameters of a classical life insurance product with profit. In this context we
suppose that the insurer guarantees a certain fixed premium rate. On top of that, participation
is given at maturity if the real investment performances are good. This bonus is expressed as
a rate applied to the eventual final surplus of the insurer. So the decision problems for the
insurer in this kind of product consists of choosing two numbers: the guaranteed rate and the
participation rate. In order to proceed we propose to decompose the problems in two steps:
9
10 Cash-Bond-Stock Model (CBS).
• First step: in our mind the choice of a guaranteed rate is directly linked with solvency
concerns. Perfect hedging for long periods as in life or pensions contracts is a utopia.
Derivative instruments for very long periods as in life or in pension insurance are not
common on the markets and self-hedging is too expensive. A risk neutral price could
give the illusion of an absence of risk like in the Black and Scholes world but in fact risk
is still there at maturity if you do not hedge perfectly as requested by the underlying
pricing process. All this motivates the use of a risk management approach in order to
fix the guaranteed rate for long periods.
• Second step: once this technical rate is chosen and in order to fix the second parameter
- the participation rate- we can try to build a fair contract to the policyholder and the
insurer. Risk neutral fairness is then used to compute the value to the participation
rate.
This way, combination of the two paradigms permits to take into account at the same time
solvency concerns as well as fair valuation principles.
The chapter is organized as follows. Section 2 presents the main assumptions on the financial
market and the investment strategies. In this context we develop a model with cash, bonds and
stocks (Cash-Bond-Stock model - CBS model hereafter) representing better than other simple
models (for instance binomial or geometric Brownian) the fundamental investments used by
the insurers as underlying values for life insurance products with profit. Different kinds of bond
strategies are developed in order to model the influence of matching policies. Section 3 is then
devoted to the risk management analysis in order to select the rate that can be guaranteed.
Explicit solutions in the CBS model are given using a “value at risk” approach; not surprisingly
the rate is a function of the required level of solvency. Formulas are also given if we choose
as criterion the expected shortfall instead of the value at risk. In section 4 fair valuation
principles are then applied in order to determine the equilibrium value of the participation rate
corresponding to the “Risk management” value of the guarantee. Forward risk neutral measure
is used to price the contract. Finally Section 5 presents some numerical illustrations.
2. Cash-Bond-Stock Model (CBS).

We assume a single premium is paid at time t0 = 0. The insurer invests this premium in
assets traded on a financial market. The value of this financial portfolio at date t is denoted
by V (t) and changes according to the evolution of the assets values and the financial strategy
chosen by the insurer. In the next section, we describe the financial market on which the insurer
can trade. In the next one, we describe the financial portfolio.
2.1. The Financial Market. We consider a financial market with three different kinds
of assets: A money market account, a risky asset and zero-coupon bonds of all maturities. The
assumptions related to our financial market are the following:
• Assumption 1: The money market account is a financial asset with no instantaneous
risk. The value of this asset, β(t), grows at the instantaneous risk free rate r(t)
according to the following differential equation:
(2.1) dβ(t) = r(t)β(t)dt
Risk Measure and Fair Valuation of an Investment Guarantee in Life Insurance. 11
with initial value β(0) = 1.

• Assumption 2: We assume the risk free rate follows an Ornstein-Uhlenbeck process.
The risk free rate is then the solution of the following stochastic differential equation:
(2.2) dr(t) = a (b − r(t)) dt + σr dW1 (t)
where W1 (t) is a standard Brownian motion.
• Assumption 3: The value of the risky asset, S(t), follows a geometric Brownian motion,
i.e., it is the solution of the following stochastic differential equation :
h p i
(2.3) dS(t) = S(t) µdt + σS ρdW1 (t) + σS 1 − ρ2 dW2 (t)
where W2 (t) is a standard Brownian motion, independent of W1 (t).The correlation

between the value of the risky asset and the risk free rate is characterized by the
parameter ρ.
• Assumption 4: We assume no arbitrage opportunity.
This last assumption together with assumptions 1 and 2 allows us to find the process followed
by the value of any zero-coupon bond. It is well known that under these assumptions, the value
P (t, T ), at time t of a zero coupon bond with maturity T is given by:
(2.4) dP (t, T ) = P (t, T ) [(r(t) − λσr B(T − t)) dt − σr B(T − t)dW1 (t)]
where
1h i
B(T − t) = 1 − e−a(T −t)
a
and t ≤ T . The parameter λ represents the price of the interest rate risk. The term −λσr B(T −
t) represents the instantaneous risk premium for investing in a zero coupon bond with time-
to-maturity T − t. Thus, −λ can be understood as the price of risk for investing in the bond
market.
Let us point out once more that the values of our assets are all correlated. The values
of the zero-coupon bonds are all perfectly correlated with each other. They are all perfectly
negatively correlated with the risk free rate and the later is also correlated with the risky asset
according to the value of the parameter ρ.
2.2. The financial portfolio. Our financial portfolio is invested in the three different
kinds of assets described above:
• A proportion xa is invested in the risky asset with price S(t) which can be seen as a
stock or a stock index.
• A proportion xb is invested in zero-coupon bonds.
• A proportion xc is invested in the money market account which can be seen as cash
(or almost cash).
Additional assumptions related to the portfolio are made:
• Assumption 5: The portfolio is continuously rebalanced in order to keep these pro-
portions constant over time.
Since the values of the different assets evolve over time, if the portfolio is not rebalanced , the
proportions invested in these assets will fluctuate. In practice, fund managers try indeed to
respect a pre-specified asset allocation.
• Assumption 6: The portfolio is self-financing.
We assume that once the initial premium is paid, no money comes in or comes out of the
portfolio up to the term of the guarantee. One of the main purposes of this chapter is to study
the impact of the financial strategy in general and the bond strategy in terms of matching of
duration in particular, on the value of the portfolio at maturity. So it is important to design a
fairly general framework for the bonds management that embeds different kinds of strategies.
• Assumption 7 : At each time t, Zero-coupon bonds of only one maturity can be
held. We do not allow the portfolio to be invested in zero-coupon bonds of different
maturities at the same time.
This assumption is only made for the sake of simplicity. At the cost of more fastidious calcula-
tions, we could have assumed our bond portfolio is invested in zero coupon bonds with several
maturities.
• Assumption 8: Let s = 0, . . . , n − 1 with t0 = 0 and tn = T . We assume that the ts
are the different dates at which we can invest in new zero coupon bonds with time-
to-maturity Ts ≥ ts+1 − ts . In other words, at time ts the insurer can sell all the
zero-coupon bonds with maturity (ts−1 + Ts−1 ) to invest in zero-coupon bonds with
maturity (ts + Ts ).
We are now ready to describe the evolution of the value of our portfolio over time and in
particular the value at maturity.
2.2.1. The value of the financial portfolio. The value V (t) of the portfolio at time t ∈ [0, T ],
is given by:
n−1
X
(2.5) V (t) = at S(t) + bst It∈[ts ,ts+1 [ P (t, Ts + ts ) + ct β(t)
s=0
where at , bst , ct are respectively, the number of shares of risky assets, the number of zero-coupon
bonds with maturity (ts + Ts ) and finally, the number of shares of money market account. The
assumption 6 (the self-financing assumption) allows us to write the evolution of the value, for
t ∈ [0, T ], of our portfolio under the following differential form:
n−1
X
(2.6) dV (t) = at dS(t) + bst It∈[ts ,ts+1 [ dP (t, Ts + ts ) + ct dβ(t)
s=0
The assumption 5 which states the proportions invested in each type of asset remain constant
over time, gives us the following relations:
xa .V (t)
at =
S(t)
xb .V (t)
bst =
P (t, Ts + ts )
xc .V (t)
ct =
β(t)
Using these relations into Equation (2.6), we find the following differential equation:
n−1
xa .V (t) X xb .V (t) xc .V (t)
dV (t) = dS(t) + I dP (t, Ts + ts ) + dβ(t)
S(t) P (t, Ts + ts ) t∈[ts ,ts+1 [ β(t)
s=0
Or equivalently, we can rewrite this formula as:
dV (t) dS(t) dP (t, Ts + ts ) dβ(t)

(2.7) = xa + xb + xc
V (t) S(t) P (t, Ts + ts ) β(t)
for t ∈ [ts , ts+1 [ . Without loss of generality, we can assume the initial value V (0) of the
portfolio is equal to 1.
2.2.2. The portfolio value at maturity. We are interested in the value of the portfolio at
maturity, V (T ). The self financing assumption allows us to write the return of the portfolio as
the product of the returns over the n different sub periods:
n−1
V (T ) Y V (ts+1 )
V (T ) = =
V (0) V (ts )
s=0
Taking the logarithm, we get:
n−1
X
ln V (T ) − ln V (0) = [ln V (ts+1 ) − ln V (ts )]
s=0
In order to determine the distribution of V (T ), we start by determining the distribution of

each term of this sum conditionally on the information known at time 0 , F0 , (the distribution
of ln V (ts+1 ) − ln V (ts ) |F0 ). In the following, let us fix a time t ∈ [ts , ts+1 [. In this interval,
we know the value of portfolio follows the differential equation (2.7):
dV (t) dS(t) dP (t, Ts + ts ) dβ(t)

= xa + xb + xc
V (t) S(t) P (t, Ts + ts ) β(t)
Inserting Equations (2.1), (2.3) and (2.4), we get:
dV (t) p
= xa µdt + σS ρdW1 (t) + σS 1 − ρ2 dW2 (t)
V (t)
+ xb [(r(t) − λσr B(Ts + ts − t)) dt − σr B(Ts + ts − t)dW1 (t)]
+ xc r(t)dt
= (xa µ − xb λσr B(Ts + ts − t) + (xb + xc ) r(t)) dt
p
+ (xa σS ρ − xb σr B(Ts + ts − t)) dW1 (t) + xa σS 1 − ρ2 dW2 (t)
So that, we have
tZs+1
ln V (ts+1 ) − ln V (ts ) = d ln V (u)

ts
tZs+1 tZs+1
= xa µ − xb λσr B(Ts + ts − u)du + (xb + xc ) r(u)du

ts ts
tZs+1
1
− x2a σS2 + x2b σr2 B(Ts + ts − u)2 − 2xa xb σS σr ρB(Ts + ts − u) du
2
ts
tZs+1 tZs+1
p
+ xa σS ρ − xb σr B(Ts + ts − u)dW1 (u) + xa σS 1 − ρ2 dW2 (u)
ts ts
The difference between the logarithms of the value of the portfolio is then equal to the sum of
four terms, the first one being deterministic and the three others random. Furthermore, it is
easy to see they are all normally distributed. So the difference ln V (ts+1 ) − ln V (ts )also follows
a normal random variable. Since this is equally true for each s = 0, 1, . . . , n − 1, we have the
following proposition:
Proposition 2.1. The logarithm of the value of the portfolio at maturity
n−1
X
ln V (T ) = [ln V (ts+1 ) − ln V (ts )]
s=0
follows a normal random variable with expectation given by:
µP (T ) = E [ln V (T ) |F0 ]
n−1
X
= E [ln V (ts+1 ) − ln V (ts ) |F0 ]
s=0
and variance given by:
σP (T ) = V ar [ln V (T ) |F0 ]
n−1
X
= V ar [ln V (ts+1 ) − ln V (ts ) |F0 ]
s=0
n−2
X n−1
X
+ 2 Covar [ln V (ts+1 ) − ln V (ts ); ln V (tj+1 ) − ln V (tj ) |F0 ]
s=0 j=s+1
So knowing the expectations, the variances and the covariances of the differences of the log-
arithms of the portfolio value over the different sub periods, we can easily find the distribution
of the portfolio value at the term of the guarantee. These expectations, variances and covari-
ances are given by the following formulas. (Detailed calculations can be found in Appendix
1).
Expectation of ln V (ts+1 ) − ln V (ts ):
E [ln V (ts+1 ) − ln V (ts ) |F0 ] = K(Ts , ts+1 − ts )

1 −ats

−a(ts+1 −ts )
+ (xb + xc ) b (ts+1 − ts ) + (b − r(0)) e e −1
a
where
tZs+1
K(Ts , ts+1 − ts ) = xa µ − xb λσr B(Ts + ts − u)du

ts
tZs+1
1
− x2a σS2 + x2b σr2 B(Ts + ts − u)2 − 2xa xb σS σr ρB(Ts + ts − u) du
2
ts

1 2 2
= xa µ − xa σS (ts+1 − ts )
2
− (xb λσr − xa xb σS σr ρ) IB1 (Ts , ts )
1 2 2
− x σ IB2 (Ts , ts )
2 b r
tZs+1
IB1 (Ts , ts ) = B(Ts + ts − u)du

ts
e−aTs

1 a(ts+1 −ts )
= (ts+1 − ts ) + 1−e
a a
tZs+1
IB2 (Ts , ts ) = B(Ts + ts − u)2 du

ts
e−2aTs 2e−aTs

1 2a(ts+1 −ts ) a(ts+1 −ts )
= (ts+1 − ts ) − 1−e + 1−e
a2 2a a
Variance of Rts = ln V (ts+1 ) − ln V (ts ):
σ2 2
V ar [Rts |F0 ] = (xb + xc )2 2r (ts+1 − ts ) + 2a 1
e−a(ts+1 −ts ) − 1 1 − e−2ats
a
σ 2 1 2
2 r −2a(ts+1 −ts ) −a(ts+1 −ts )
+ (xb + xc ) 2 1−e − 1−e
a 2a a
+ (xb σr )2 IB2 (Ts , ts ) − 2xa xb σS σr ρIB1 (Ts , ts ) + (xa σS )2 (ts+1 − ts )
2

+ 2 (xb + xc ) xa σSa σr ρ (ts+1 − ts ) − a1 1 − e−a(ts+1 −ts ) − xbaσr IB1 (Ts , ts )

xb σr2 1
e−aTs
−a(ts+1 −ts ) a(ts+1 −ts ) −a(ts+1 −ts )
+ 2 (xb + xc ) 2 1−e − e −e
a a 2a
Covariance between Rti = ln V (ti+1 ) − ln V (ti ) and Rtj = ln V (tj+1 ) − ln V (tj ):
σ2
Covar Rti ; Rtj |F0 = (xb + xc )2 r3 1 − e−a(ti+1 −ti ) 1 − e−a(tj+1 −tj ) e−a(tj −ti ) − e−a(ti +tj )

2a
σr2
+ (xb + xc ) 3 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(tj −ti )
a
σ2
− (xb + xc ) r3 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(ti+1 −ti ) e−a(tj −ti )
2a
σr x σ ρ
a S

+ (xb + xc ) 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(tj −ti )
a x σa
σr −a(tj+1 −tj ) b r −a(tj −ti+1 ) −a(tj −ti )

− (xb + xc ) 1−e e − e
a a2
σr −aT
e i a(ti+1 −ti ) −a(tj −ti+1 )
+ (xb + xc ) 1 − e−a(tj+1 −tj ) e e − e−a(tj −ti )
a 2
for tj ≥ ti+1 .
2.2.3. The Investment Strategies. The bond strategy described in the preceding section, is
too general to be really useful. We will restrict our attention to a few special cases. In this
section, we present the three bond strategies we choose to analyze.
• Strategy 1: It consists in investing in zero coupon bonds with a time-to-maturity equal
to the term of guarantee. We suppose we keep them up to this term. This strategy
corresponds to the special case where n is equal to one and where T1 is equal to T .
• Strategy 2: It consists in investing initially in zero-coupon bonds with a time-to-
maturity inferior to the term of the guarantee. We assume we keep these bonds up to
their maturity. Then we reinvest in zero coupon bonds with the same time-to-maturity
than the first ones. Again, we keep them up to their maturity and so on to the term
of the guarantee. This strategy corresponds to the special case where Ts is equal to
T1 for all s = 0, 1, . . . , n − 1 and T1 = T /n.
• Strategy 3: It consists in investing initially in zero-coupon bonds with a time-to-
maturity inferior or superior to the term of the guarantee. After a period of time ∆t,
we sell these bonds and reinvest in other zero coupon bonds with the same time-to-
maturity as the first. Again, we keep them during a period of time ∆t, and at this
time, we sell them and reinvest in other zero coupon bonds, and so on up to the term
of the guarantee. This strategy corresponds to the special case where Ts is equal to
T1 for all s = 0, 1, . . . , n − 1 and ∆t = T /n < T1 .
3. Investment Guarantee and Risk Measures.

For the insurer, the risk comes from the possibility of having a financial portfolio value at
time T lower than the liability value L(T ) (“downside risk”). In order to measure this kind of
risk, a number of risk measures have been developed in the literature. In our setting, the value
of a risk measure will depend on the guaranteed rate and the financial strategy. Notice the
participation rate here does not induce any additional risk. The whole risk comes only from
the level of the guaranteed rate and the financial strategy. So given these elements, we can find
the value of the risk measure. But the reverse is also true: given a risk measure, a maximal
accepted value for this risk measure and a financial strategy, we can find the maximal interest
rate we can guarantee on a single premium.
In the next sections, we analyze in more detail these relations for two well-established risk
measures, the value at risk and the expected shortfall.
3.1. Investment Guarantee. As already said, we assume a single premium is paid at

date 0. The insurer guarantees a minimum annual return i on this single premium at time T .
Without loss of generality, we can assume the premium value is 1. Accordingly, the liability
value at date T is equal to L(T ) = (1 + i)T .
3.2. Value at Risk. This is historically the first risk measure described in the literature.
It is defined as the supremum discounted loss we can experience with a probability not less
than a given threshold . Mathematically, for given V (T ), L(T ) and , the value at risk is
defined as the value VaR such that

x
VaR = − inf x Pr V (T ) − L(T ) ≤ ≥
P (0, T )
In our setting, we can easily find an explicit solution for the VaR:
18 Investment Guarantee and Risk Measures.

V aR
Pr ln V (T ) ≤ ln − + L(T ) =
P (0, T )
 
ln − PV(0,T
aR
) + L(T ) − µP (T )
Pr Z ≤ =
σP (T )

V aR
ln − + L(T ) = µP (T ) + σP (T )φ−1 ()
P (0, T )
V aR = P (0, T ) L(T ) − exp µP (T ) + σP (T )φ−1 ()

where Z is a standard normal random variable and φ is the cdf of Z; hence φ−1 () is the
-percentile of the standard normal distribution. Finally, we get:
h i
V aR = P (0, T ) (1 + i)T − exp µP (T ) + σP (T )φ−1 ()

(3.1)
For a given value at risk, a given strategy and a given guaranteed rate, we can also determine
the probability . The two most natural values for the VaR are 0 and the amount of capital set
aside for this contract, assuming we invest this capital in a risk free asset. In the former case,
reduces to the well-known probability of ruin. In our setting, this probability is given by:
 
ln − PV(0,T
aR
) + L(T ) − µP (T )
(3.2) = φ 
σP (T )
The risk measures have been initially developed for risk management purposes. But here, we
could extend their uses to a pricing purpose. Indeed, for a given VaR (or equivalently a given
probability ) and a given financial strategy (on which µP (T ) and σP (T ) depend), we can
determine the corresponding value of the guaranteed rate i. In our setting, this guaranteed
rate is given by the formula:
1
V aR −1
T
(3.3) i = + exp µP (T ) + σP (T )φ () −1
P (0, T )
This has the advantage the guaranteed rate is consistent with the risk management strategy.
3.3. Expected Shortfall or CoValue at Risk. Other well-known risk measures are the
expected shortfall (ES) and the conditional value at risk (CVaR). When the distributions are
continuous, these measures coincide. The CVaR is defined as the discounted expected loss if
the expected loss is higher than a given threshold R. Mathematically, given V (T ), L(T ) and
R, the conditional value at risk is defined as:

R
CV aR = E −P (0, T ) (V (T ) − L (T )) V (T ) − L(T ) ≤ −

P (0, T )
In the literature, R is actually the Value at Risk hence the name conditional value at risk.
The most natural value for R is probably 0. In this case, the CVaR is simply the discounted
expected loss if we are in loss.
In our setting, we can find the following explicit formula for the CVaR:
2
(1 + i)T φ (DP ) − e(µP (T )+0.5σP (T )) φ (DP − σP (T ))
(3.4) CV aR = P (0, T )
φ (D)
where

R T
ln − P (0,T ) + (1 + i) − µP (T )
DP =
σP (T )
Even if this measure has also been initially developed for risk management purposes, we can use
it for a pricing purpose. Given R, the expected shortfall and a financial strategy, we can find a
guaranteed rate which is consistent with the risk management strategy defined by the insurer.
There is no explicit formula for this guaranteed rate but we can easily invert numerically the
last one.
4. Fair Valuation and Participation Rate.

4.1. Participation rate: Classical life insurance products are based on the one hand on
a guaranteed technical rate and on the other hand on the distribution of a part of the surplus
if any ( participation process). If we consider for instance the following assumptions:
• contract with single premium P = 1 paid at time t = 0
• no mortality assumption
• liability to pay at maturity t = T
• guaranteed technical rate i
• participation rate η ≥ 0, on the final surplus (terminal bonus)
then the fair value of the contract at maturity can be written as:
(4.1) F V (T ) = L(T ) + η max(V (T ) − L(T ); 0)

where L(T ) = (1 + i)T is the guarantee and V (T ) is the final value of the underlying portfolio.
Two parameters have to be fixed by the insurer: the guaranteed rate and the participation
rate. As seen before a way to define the guaranteed rate is based on risk management principles,
for instance by using the VaR measure.
How to fix then the participation rate? A possible approach is to rely on fair valuation
principles. Consider for this the market value of the contract at time t = 0 denoted by F V (0).
It seems fair to impose that this value coincides with the premium paid at time t = 0 (classical
actuarial equivalence principle). In order to compute this fair value we can try to use the
classical risk neutral approach; in our setting markets are complete and arbitrage free, hence
we can express the initial pricing of the product as:
20 Fair Valuation and Participation Rate.
(4.2) F V (0) = (1 + i)T P (0, T ) + ηP (0, T )EQT ((V (T ) − (1 + i)T )+ )

where QT is the T -forward risk neutral measure.
So if i is the guaranteed rate determined following the principles of Section 3 and for a
level , then the corresponding equilibrium value of the participation rate is given by:
1 − (1 + i )T P (0, T ) 1 − (1 + i )T P (0, T )
η= =
P (0, T )EQT ((V (T ) − (1 + i )T )+ ) C(0, (1 + i )T )
where C(0, (1+i )T ) is the price at time 0 of a European call option on the underlying portfolio
T
with maturity T and strike (1 + ī ) . Since the price of a call option is always positive, in order
to have the participation rate positive, we should have
(1 + i )T P (0, T ) ≤ 1 ⇔ i ≤ R(0, T )
where R(0, T ) is the annually compounded risk free interest rate with term T . Notice in the
previous section, we did not have this constraint and a guaranteed rate higher than the risk free
rate was acceptable from a risk management perspective. So the guaranteed rate is eventually
given by:
i = min(R(0, T ), i0 )
and the participation rate:
1 − (1 + i )T P (0, T )
(4.3) η=
C(0, (1 + i )T )
4.2. Equilibrium value in the CBS model. We can explicitly compute the value of the
equilibrium participation rate in the framework of the CBS model presented in Section 2. In
order to keep computations as easy as possible we will develop them using only ALM strategy
1 as defined in Section 2.2.3 (zero coupon bonds with maturity corresponding to the duration
of the guarantee). The explicit formula in the general case is then given in Section 4.2.2.
4.2.1. Strategy 1. We first give the processes followed by the different assets under the risk
neutral measure; then we move to the T -forward martingale measure. Under the risk neutral
measure the three assets have the following form:
• Cash:
dβ(t) = r(t)β(t)dt
• Bonds:
dP (t, T ) = P (t, T )(r(t)dt − σr B(T − t)dW 1 (t))
• Stocks:
p
dS(t) = S(t)(r(t)dt + σS ρdW 1 (t) + σS 1 − ρ2 dW 2 (t))
where the new Brownian motions are defined by:
dW 1 (t) = dW1 (t) + λdt
p
dW 2 (t) = dW2 (t) + (µ − r(t) − λρσS )/(σS 1 − ρ2 )dt
= dW2 (t) + θ(t)dt
The equation of portfolio V of Section 2.2.1 becomes in the risk neutral world:
dV (t) p
= r(t)dt + (xa σS ρ − xb σr B(T − t))dW 1 (t) + xa σS 1 − ρ2 dW 2 (t)
V (t)
= r(t)dt + σ 1 (t)dW 1 (t) + σ 2 (t)dW 2 (t)
Finally the risk free rate follows a modified Ornstein-Uhlenbeck process:

σr λ
dr(t) = a(b − − r(t))dt + σr dW 1 (t)
a
= a(b − r(t))dt + σr dW 1 (t)
The T -forward martingale measure is then generated by the change:
dW 1 (t) = dW 1 (t) + σr B(T − t)dt

= dW 1 (t) + σ(t, T )dt
dW 2 (t) = dW 2 (t)
Under this T -forward martingale measure and recalling V (0) = 1 the portfolio value can be
written as:
Zt Zt Zt Zt
1 2
V (t) = exp{ (r(s)−σ 1 (s)σ(s, T )ds+ σ 1 (s)dW 1 (s)+ σ 2 (s)dW 2 (s)− (σ 1 (s)+σ 22 (s))ds}
2
0 0 0 0
In this model, we have then to price a call option:
C(0, K) = P (0, T )EQT ((V (T ) − K)+ )

= P (0, T )EQT (V (T )1V (T )>K ) − KP (0, T )PQT (V (T ) > K)
= I − II
Using standard tools for computing derivatives prices in a stochastic term structure we get
successively:
ln(1/K) − ln P (0, T ) − v 2 (T )/2
II = KP (0, T )Φ( )
v(T )
22 Fair Valuation and Participation Rate.
with
ZT ZT
2 2
v (T ) = (σ 1 (s) + σ(s, T )) ds + σ 22 ds
0 0
= x2a σS2 T + σr2 (1 2
− xb ) B2 (T ) + 2xa (1 − xb )σS σr ρB1 (T )
ZT
B1 (T ) = B(T − u)du
0
T 1
= − 2 (1 − e−aT )
a a
ZT
B2 (T ) = B 2 (T − u)du
0
T 1 2
= 2
+ 3 (1 − e−2aT ) − 3 (1 − e−aT )
a 2a a
and
ln(1/K) − ln P (0, T ) + v 2 (T )/2

I = Φ( )
v(T )
Finally putting
K = (1 + i )T
we get for the price of the call:
−T ln(1 + i ) − ln P (0, T ) + v 2 (T )/2
C(0, (1 + i )T ) = Φ( )
v(T )
−T ln(1 + i ) − ln P (0, T ) − v 2 (T )
− (1 + i )T P (0, T )Φ( )
v(T )
Or
(4.4) C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

Putting the Equation (4.4) in Equation (4.3) we obtain the equilibrium value of the participation
rate:
1 − (1 + i )T P (0, T )
(4.5) η=
Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))
4.2.2. The general case. Similar calculations lead to an explicit formula for the participa-
tion rate in the general case. The detailed calculations can be found in Appendix 2. The
participation rate is still given by:
1 − (1 + i )T P (0, T )
η=
C(0, (1 + i )T )
The price of the call is still a Black-Scholes like formula:
C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

with
−T ln (1 + i ) − ln P (0, T ) + 12 v 2
D+ (i ) =
v
−T ln (1 + i ) − ln P (0, T ) − 12 v 2
D− (i ) =
v
But in the general case the parameter v 2 is given by:
ti+1
n−1
X Z
v 2
= (vV (u, ti + Ti ) + vP (u, T ))2 du
i=0 t
i
where
vV (u, ti + Ti ) = xa vS − xb vP (u, ti + Ti )

σS ρ
vS = p
σ S 1 − ρ2

σr B (ti + Ti − u)
vP (u, ti + Ti ) =
0
In the first expression, we define a2 = a0 · a where a is n × 1 matrix and a0 is its transpose.
5. Numerical Illustrations
We estimated the parameters of our market model on Belgian data’s. The stock parameters
are estimated from January 1991 to March 2004, on the Euronext Belgian All Share Index.
The interest rate parameters are estimated from January 2000 to March 2004 on the 1 to 10
years swap rates. The maximum likelihood estimators of these parameters are given in table
5.1 and 5.2.
Table 5.1. Interest rate

a b σr λ r(0)
0.1577 5.49 % 0.89 % -0.2304 1.23 %
24 Numerical Illustrations
Table 5.2. Stock model

µ σS
9.41 % 14.71 %
In the next subsections, we illustrate the impact of the financial strategy on different vari-
ables, namely the Value at Risk, the probability of ruin, the guaranteed rate, the participation
rate based on the VaR and the conditional value at risk and the guaranteed rate, the participa-
tion rate based on the CVaR. In each case, 4 graphs are given, representing different types of
asset allocation. On the x-axis, we have the number n of times we reinvest in new zero-coupon
bonds and on the y-axis, we have the time-to-maturity T 1 of the bonds we invest in. In all
these illustrations, the term of the guarantee T is set to 15 years and ρ to 0.15.
As already said, the strategy 1 corresponds to the case where n is equal to 1 and T 1 = T =
15. On all these graphs, when it does make sense, this strategy is the first point on the y-axis.
The second strategy corresponds to the case where nT 1 = T . This strategy is represented by
the points on the left edge of the curve. The strategy 3 corresponds to all the other points of
the curve.
5.1. Value at Risk. Figure 5.1 illustrates the effect of the financial strategy on the Value
at Risk for a probability equal to 1%. In this figure, the guaranteed rate is equal to 3.25%.
In table 5.3, we give some numerical examples.
Table 5.3
xa =15%, xb =80%, xc =5% Value at risk
Strategy 1: -0.1865
Strategy 2:
n = 3 and T1 = 5 0.0374
n = 5 and T 1 = 3 0.1039
Strategy 3:
n = 3 and T 1 = 8 -0.1048
n = 5 and T 1 = 5 -0.0282
The first important point to notice is that the value at risk is not necessarily increasing
with the part invested in the risky assets (xa ): the second graph exhibits globally lower values
at risk than the first one. As the third and fourth graph show, this result is reversed the more
we invest in risky assets. With respect to the part invested in the risky asset, the value at risk is
thus at first decreasing then increasing. This result comes from the higher risk premium of the
stocks compared to the bonds: the increase in expected return can offset in a certain extent,
the increase of risk taking into account the necessity of the guarantee. The second important
point is that, in the first graph, we see the value at risk exhibits a minimum with respect to
the term of the bonds. The value at risk is lower for bonds with medium time to maturity than
for short or long time to maturity.
Figure 5.1. Probability = 1%
5.1.1. Probability of ruin: As already said, the probability corresponds to the probability
of ruin if the VaR is set to 0. Figure 5.2 illustrates the effect of the financial strategy on the
probability of ruin for a guaranteed rate equal to 3.25%.
Table 5.4
xa =15%, xb =80%, xc =5% Pr of ruin
Strategy 1: 0.02%
Strategy 2:
n = 3 and T 1 = 5 1.62%
n = 5 and T 1 = 3 3.29%
Strategy 3:
n = 3and T 1 = 8 0.17%
n = 5and T 1 = 5 0.66%
In Figure 5.2, we see the probability of ruin is at first decreasing with the part invested in
the risky asset then globally increasing, as it is with the value at risk. Investing a small part
of the financial portfolio in risky assets can be profitable in terms of probability of ruin.
Figure 5.2. Guaranteed rate = 3.25%
5.1.2. Guaranteed rate and participation rate: Here, we also set the VaR equal to 0. Fig-
ures 5.3 and 5.4 illustrate the effect of the financial strategy on the guaranteed rate and the
participation rate for a probability of ruin equal to 1%.
Table 5.5
xa =15%, xb =80%, xc =5% Guaranteed rate Participation Rate
Strategy 1: 4.00% 96.56%
Strategy 2:
n = 3 and T 1 = 5 3.09% 99.35%
n = 5 and T 1 = 3 2.79% 99.55%
Strategy 3:
n = 3 and T 1 = 8 3.68% 98.23%
n = 5 and T 1 = 5 3.37% 98.98%
Figure 5.3. Probability of Ruin = 1%
Firstly, in Figure 5.3, the guaranteed rate is increasing at first with the part invested in
the risky asset then decreasing. This result is obviously consistent with what we found for the
value at risk. The fourth graph illustrates the fact that when we invest massively in the risky
asset, we can be in a situation where we cannot find a positive guaranteed rate consistent with
our risk policy. Secondly, we also see in Graph 1 that the guaranteed rate exhibits a maximum
with respect to the time to maturity of the zero coupon bonds. Finally, notice the guaranteed
rate is very sensitive to the financial strategy chosen.
In Figure 5.4, we show the participation rates given the financial strategies and the guar-
anteed rates found above. The participation rate depends on both the guaranteed rate and
the risk of the financial strategy in terms of standard deviation. Globally, the participation
rate should be decreasing with the guaranteed rate and decreasing with the standard devia-
tion of the strategy. Both effects can act in the same direction. For example, we find lower
participation rates in Graph 2 than in Graph 1 due to both a higher guaranteed rate and a
higher standard deviation. They can also act on opposite directions. Although we have higher
standard deviations in Graph 3 than in Graph 4, we find higher participation rates in Graph 4
due to offsetting lower guaranteed rates.
Figure 5.4. Probability of Ruin = 1%
5.2. Conditional Value at Risk. Figure 5.5 illustrates the effect of the financial strategy
on the conditional value at Risk for a guaranteed rate equal to 3.25%. Here, R is set to 0.
Table 5.6
xa =15%, xb =80%, xc =5% CvaR
Strategy 1: 1.66%
Strategy 2:
n = 3and T 1 = 5 3.41%
n = 5 and T 1 = 3 4.06%
Strategy 3:
n = 3 and T 1 = 8 2.32%
n = 5and T 1 = 5 2.85%
Firstly, the conditional value at risk is increasing with the part invested in the risky assets
and does not seem to exhibit a first decreasing movement, in contrast with the value at risk.
Figure 5.5. Guaranteed rate = 3.25 %
Secondly, in the first graph, as the value at risk, the conditional value at risk exhibits a minimum
with respect to the term of the bonds.
5.2.1. Guaranteed rate and Participation rate: Figures 5.6 and 5.7 illustrate the effect of
the financial strategy on respectively the guaranteed rate and the participation rate for a
conditional value at risk equal to 4%.
Firstly, in Figure 5.6, we see the guaranteed rate is globally decreasing with the part
invested in the risky asset. The fourth graph illustrates the situation where we cannot find a
positive guaranteed rate consistent with our risk policy. Secondly, here also, the guaranteed
rate exhibits a maximum with respect to the time to maturity of the zero coupon bonds. In
graph one, the guaranteed rates are actually capped for a number of strategies because they
exceed the risk free rate. We can find these strategies in Figure 5.7: these are those exhibiting
a participation rate equal to 0.
Figure 5.6. Conditional Value at Risk = 4%
Table 5.7
xa =15%, xb =80%, xc =5% Guaranteed rate Participation Rate
Strategy 1: 4.69% 45.94%
Strategy 2:
n = 3 and T 1 = 5 3.57% 97.03%
n = 5 and T 1 = 3 3.22% 98.42%
Strategy 3:
n = 3 and T 1 = 8 4.26% 85.87%
n = 5 and T 1 = 5 3.91% 93.98%
As far as the participation rate is concerned, we can make the same comments than with
the Value-at-Risk since it does not depend on the chosen risk measure (apart through the
guaranteed rate). Globally, it is decreasing with the standard deviation of the financial strategy
and the guaranteed rate. Sometimes, these effects act in the same direction, sometimes in
opposite directions.
Figure 5.7. Conditional Value at Risk = 4%
6. Conclusions and Extensions.

In this chapter, we studied the interrelationship between the risk management and the
pricing of a classical single premium life insurance contract with profit. The pricing is considered
in terms of a guaranteed rate on the premium and a participation rate on the financial surplus.
We argued it is a utopia to think we can perfectly hedge the liabilities of long term insurance
contracts. There are two main reasons to this. First of all, derivative instruments with terms of
10, 15 or 20 years are not traded on the markets. Secondly, self-hedging would be too expensive.
So, at least for these very long term liabilities, financial risk is unavoidable. By using only
the risk neutral valuation principle to determine the guaranteed rate and the participation
rate, we cannot properly, take into account this financial risk. Moreover, using this pricing
principle alone, we would have to pick up arbitrarily, a pair of guaranteed and participation
rates among the set of pairs consistent with the no arbitrage assumption. Instead, we argued
these parameters should be determined by taking into account explicitly the risk the insurer
accepts to bear. Moreover, this acceptable level of risk should be set consistently with the risk
management policy of the insurance company, in terms of a risk measure such as the value-at-
risk or the conditional value-at-risk. In order to answer these points, we suggested to divide our
32 Appendix 1.
problem in two parts. In the first one, we consider the insurance contract without profit. Since
the financial risk comes only from the guaranteed rate, we proposed to fix the guaranteed rate
such that the value-at-risk or conditional value-at-risk of this modified contract, does not exceed
a certain level chosen by the risk management. In the second part, we fix the participation rate
according to the risk neutral valuation principle and according to the guaranteed rate found in
step one. In this way, we have a contract that is simultaneously fairly priced and that exhibits
a risk consistent with the risk management policy.
Finally, since the financial risk depends obviously on the financial portfolio, it is important
to model realistically this portfolio. Accordingly, we proposed a more sophisticated model
that includes investment in cash, stocks and bonds and that allows us to study the effect
of the matching policy and the effect of the strategic allocation on the technical parameters.
Furthermore, this model has the advantage to offer closed form solution.
This model can be extended in a variety of ways and still keeps its explicit solutions. First
of all, our portfolio is assumed to be invested in a single zero coupon bond. We could also
have assumed our bond portfolio is invested in several zero coupon bonds. Secondly, we could
also have assumed a more general Gaussian model for our financial market. For example, the
instantaneous risk free rate could follow a multi-factor Gaussian model. In both cases, we still
would have obtained closed form solutions, at the cost of more fastidious calculations.
Thirdly, we could even extend our model to a more general non Gaussian framework. For
example, it certainly would be interesting to model our stock index as a discontinuous Lévy
process since empirically, this family of models is much more justified than the log-normal one.
Moreover, since these kinds of processes exhibit fat tails in contrast with Gaussian model, they
should have an important impact on the risk measure and therefore on the guaranteed rate.
Finally, we could also extend our model to a periodic premium contract. Obviously, in these
cases, we cannot anymore expect to obtain closed form solutions and we should admit to rely
on numerical techniques.
Appendix 1.
In this appendix, we look for the portfolio Value distribution. First, we need a few prelim-
inary results:
Zt
r(t) = e−at r(0) + 1 − e−at b + σr e−a(t−u) dW1 (u)

0
tZs+1 tZs+1 tZs+1Zu

−au −au
σr e−a(u−v) dW1 (v)du

r(u)du = e r(0) + b 1 − e du +
ts ts ts 0

b r(0)
= b (ts+1 − ts ) + − e−a(ts+1 ) − e−a(ts )
a a
 ts+1 tZs+1 tZs+1
Zts
 
Z
+  σr e−a(u−v) du dW1 (v) +  σr e−a(u−v) du dW1 (v)
0 ts ts v
1
= b (ts+1 − ts ) + (b − r(0)) e−ats e−a(ts+1 −ts ) − 1
a
Zts tZs+1
σr −a(ts −v) −a(ts+1 −ts ) σr
− e e − 1 dW1 (v) + 1 − e−a(ts+1 −v) dW1 (v)
a a
0 ts
Expectation of Rts = ln V (ts+1 ) − ln V (ts ):

" #
ts+1
R
E [Rts |F0 ] = E xa µ − xb λσr B(Ts + ts − u) du |F0
ts
" #
ts+1
1
x2a σS2 + x2b σr2 B(Ts + ts − u)2 − 2xa xb σS σr ρB(Ts + ts − u) du |F0
R
− E
2 ts
" #
ts+1
R
+ E (xb + xc ) r(u)du |F0
ts
 ts+1 tZs+1 
Z h p i
+ E [xa σS ρ − xb σr B(Ts + ts − u)] dW1 (u) + xa σS 1 − ρ2 dW2 (u) |F0 
ts ts
Since the two first integrals are deterministic, we have
ts+1
R
E [Rts |F0 ] = xa µ − xb λσr B(Ts + ts − u) du
ts
tZs+1
1
− x2a σS2 + x2b σr2 B(Ts + ts − u)2 − 2xa xb σS σr ρB(Ts + ts − u)du
2
ts
 ts+1 
Z
+ E (xb + xc ) r(u)du |F0 
ts
 ts+1 
Z
= K(Ts , ts+1 − ts ) + (xb + xc ) E  r(u)du |F0 
ts
34 Appendix 1.
Hence
E [Rts |F0 ] = K(Ts , ts+1 − ts )

1 −ats

−a(ts+1 −ts )
+ (xb + xc ) b (ts+1 − ts ) + (b − r(0)) e e −1
a
Variance of Rts = ln V (ts+1 ) − ln V (ts ):
 
ts+1
R
 xa µ − xb λσr B(Ts + ts − u)du 
 ts 
 t s+1
 1 R 2 2

2 2 2

 − xa σS + xb σr B(Ts + ts − u) − 2xa xb σS σr ρB(Ts + ts − u)du 
 2 
 ts 
 ts+1
R 
V ar [Rts |F0 ] = V ar  + (xb + xc ) r(u)du |F0 
 
 tts
 

 s+1 
 + R xa σS ρ − xb σr B(Ts + ts − u)dW1 (u) 
 
 ts 
 ts+1
R p 
+ xa σS 1 − ρ2 dW2 (u)
 
ts
 tZs+1 
= V ar (xb + xc ) r(u)du |F0 

t
| {zs }
(A)
 ts+1 
Z
+ V ar  xa σS ρ − xb σr B(Ts + ts − u)dW1 (u) |F0 
ts
| {z }
(B)
 ts+1 
Z p
+ V ar  xa σS 1 − ρ2 dW2 (u) |F0 
ts
| {z }
(C)
 tZs+1 tZs+1 
+ 2Covar (xb + xc ) r(u)du ; xa σS ρ − xb σr B(Ts + ts − u)dW1 (u) |F0 

ts ts
| {z }
(D)
We have that
σ 2 2 Zts
2 r −a(ts+1 −ts )
(A) = (xb + xc ) e −1 e−2a(ts −u) du
a
0
tZs+1
σ 2 2
r
+ (xb + xc )2 1 − e−a(ts+1 −u) du
a
ts
σ 2
r 2
= (xb + xc )2 e−a(ts+1 −ts ) − 1 1
1 − e−2ats

a 2a
σ 2
r2 1
1 − e−2a(ts+1 −ts ) − 2
1 − e−a(ts+1 −ts )

+ (xb + xc ) (ts+1 − ts ) + 2a a
a
tZs+1
(B) = [xa σS ρ − xb σr B(Ts + ts − u)]2 d(u)

ts
= (xa σS ρ)2 (ts+1 − ts ) + (xb σr )2 IB2 (Ts , ts ) − 2xa xb σS σr ρIB1 (Ts , ts )

tZs+1
h p i2
(C) = xa σS 1 − ρ2 du
ts
= (xa σS )2 1 − ρ2 (ts+1 − ts )

tZs+1
σr
(D) = 2 (xb + xc ) 1 − e−a(ts+1 −u) (xa σS ρ − xb σr B(Ts + ts − u)) du
a
ts
h xb σ2 i
= 2 (xb + xc ) xa σ S σr ρ 1 −a(ts+1 −ts )
a (t s+1 − t s ) − a 1 − e − a r IB1 (Ts , ts )
xb σr2 1
e−aTs
+ 2 (xb + xc ) 2 1−e − e −e
a a 2a
Eventually, we get
σr2 h 2 i
V ar [Rts |F0 ] = (xb + xc )2 (t 1
− ts ) + 2a e −a(t s+1 −ts ) −1 1−e −2at s
a2 s+1
σ2 1 2
+ (xb + xc )2 2r 1 − e−2a(ts+1 −ts ) − 1 − e−a(ts+1 −ts )
a 2a a
+ (xb σr )2 IB2 (Ts , ts ) − 2xa xb σS σr ρIB1 (Ts , ts ) + (xa σS )2 (ts+1 − ts )
h 2
i
+ 2 (xb + xc ) xa σSa σr ρ (ts+1 − ts ) − a1 1 − e−a(ts+1 −ts ) − xbaσr IB1 (Ts , ts )

xb σr2 1
e−aTs
+ 2 (xb + xc ) 1−e − e −e
a2 a 2a
36 Appendix 1.
Covariances between Rti = ln V (ti+1 ) − ln V (ti ) and Rtj = ln V (tj+1 ) − ln V (tj ) with
tj ≥ ti+1 :
 tR tR

i+1 i+1
 (xb + xc ) r(u)du + (xa σS ρ − xb σr B(Ti + ti − u)) dW1 (u) 

 ti ti 
 ti+1 
 R p 
 + x σ 1 − ρ 2 dW (u) ;
a S 2

 
ti
Covar Rti ; Rtj |F0 = Covar  |F
 
tj+1
R t j+1
R 0 
 (xb + xc ) r(u)du + (xa σS ρ − xb σr B(Tj + tj − u)) dW1 (u)
 

t t
 j j

 
 tj+1
R p 
+ xa σS 1 − ρ2 dW2 (u)
 
tj

Rti σr −a(ti+1 −ti ) e−a(ti −u) dW (u)

(x
 b + xc ) a 1−e 1
 0
 tR
i+1
σr
1 − e−a(ti+1 −u) dW1 (u)

 + (xb + xc ) a
ti


 tRi+1 tR
i+1 p
 +
 (xa σS ρ − xb σr B(Ti + ti − u)) dW1 (u) + xa σS 1 − ρ2 dW2 (u) ; |F
 ti ti
= Covar 
 Rtj σr −a(tj+1 −tj ) e−a(tj −u) dW (u)

a 1−e
 (xb + xc ) 1

 0
 tj+1
1 − e−a(tj+1 −u) dW1 (u)
R σr
 + (xb + xc )

a

 tj
 tj+1
R tj+1
R p
+ (xa σS ρ − xb σr B(Tj + tj − u)) dW1 (u) + xa σS 1 − ρ2 dW2 (u)

tj tj
It leads to
Rti σr2
(xb + xc )2 1 − e−a(ti+1 −ti ) 1 − e−a(tj+1 −tj ) e−a(ti −u) e−a(tj −u) du

Covar Rti ; Rtj |F0 = a2
0
tZi+1
σ 2
r
+ (xb + xc ) 1 − e−a(ti+1 −u) 1 − e−a(tj+1 −tj ) e−a(tj −u) du
a
ti
ti+1
Z
σr
+ (xb + xc ) (xa σS ρ − xb σr B(Ti + ti − u)) 1 − e−a(tj+1 −tj ) e−a(tj −u) du
a
ti
σr2
= (xb + xc )2 1 − e−a(ti+1 −ti )
1 − e−a(tj+1 −tj )
e−a(tj −ti )
− e−a(ti +tj )
2a3
σ2
+ (xb + xc ) 3r 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(tj −ti )
a
1 σ2
− (xb + xc ) 3r 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(ti+1 −ti ) e−a(tj −ti )
2 a
σr hx σ ρ
a S
i
+ (xb + xc ) 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(tj −ti )
a a
σr2
− (xb + xc ) xb 3 1 − e−a(tj+1 −tj ) e−a(tj −ti+1 ) − e−a(tj −ti )
a
σ 2 e−aTi
+ (xb + xc ) xb r 3 1 − e−a(tj+1 −tj ) ea(ti+1 −ti ) e−a(tj −ti+1 ) − e−a(tj −ti )
2a
Appendix 2.
In this appendix, we look for the value of C(0, (1 + i )T ) in the general case. Under the
martingale measure Q, we know by using standard results that:
dP (t, T ) = P (t, T ) r(t)dt − vP0 (t, T )dW (t)

dS(t) = S(t) r(t)dt + vS0 dW (t)

where

σr B(T − t) σS ρ W 1 (t)
vP (t, T ) = , vS = p and W (t) =
0 σ S 1 − ρ2 W 2 (t)
The portfolio value for is then given by t ∈ [ts , ts+1 [:
dV (t)
= r(t)dt + xa vS0 − xb vP0 (t, ts + Ts ) dW (t)

V (t)
Put
vV0 (t, ts + Ts ) = xa vS0 − xb vP0 (t, ts + Ts )

Using Ito’s formula, we get:

1 0 0
d ln V (t) = r(t) − vV (t, ts + Ts ) .vV (t, ts + Ts ) dt + vV (t, ts + Ts ) dW (t)
2
t tZs+1

 Zs+1 1 0

0

V (ts+1 ) = V (ts ) exp r(u) − vV (u, ts + Ts ) .vV (u, ts + Ts ) du + vV (u, ts + Ts ) dW (u)
 2 
ts ts
38 Appendix 2.
ti+1 ti+1
 
n−1
XZ 1 0
n−1
X Z
0

V (T ) = V (0) exp r(u) − vV (u, ti + Ti ) .vV (u, ti + Ti ) du + vV (u, ts + Ts ) dW (u)
 2 
i=0 t i=0 t
i i
Under the T -Forward measure QT , we know that W (t) defined by
dW (t) = dW (t) + vP (t, T )dt

is a 2-dimensional Brownian motion. Under this measure, the prices of the different assets are
then given by:
h i
dP (t, ts + Ts ) = P (t, ts + Ts ) r(t)dt + vP0 (t, ts + Ts ).vP (t, T )dt − vP0 (t, ts + Ts )dW (t)
h i
dS(t) = S(t) r(t)dt − vS0 .vP (t, T )dt + vS0 dW (t)
The portfolio value for t ∈ [ts , ts+1 [ is then given by:
dV (t)
= r(t) − xa vS0 − xb vP0 (t, ts + Ts ) .vP (t, T ) dt + xa vS0 − xb vP0 (t, ts + Ts ) dW (t)

V (t)

1 0
d ln V (t) = r(t) − vV (t, ts + Ts ) .vP (t, T ) − vV (t, ts + Ts ) .vV (t, ts + Ts ) dt + vV0 (t, ts + Ts ) dW (t)
0
2
 t 
s+1
r(u) − vV0 (u, ts + Ts ) .vP (u, T ) − 12 vV0 (u, ts + Ts ) .vV (u, ts + Ts ) du

 R 


 

ts
V (ts+1 ) = V (ts ) exp ts+1
 R 0 
 + vV (u, ts + Ts ) dW (u)

 


ts
P tRi+1
 
n−1
r(u) − vV0 (u, ti + Ti ) .vP (u, T ) − 12 vV0 (u, ti + Ti ) .vV (u, ti + Ti ) du

 


 

i=0 ti
V (T ) = V (0) exp
 P tRi+1 0
n−1 
 + vV (u, ti + Ti ) dW (u)

 


i=0 ti
We know that under this T -forward measure, the price of the call is given by:
T
C = P (0, T )E Q [max (FV (T, T ) − K; 0)]
T T
= P (0, T )E Q FV (T, T )1FV (T,T )>K − P (0, T )E Q K1FV (T,T )>K

= P (0, T ) [I − II]
Where FV (t, T ) is the forward price of the portfolio at time t for a maturity T :
V (T )
FV (t, T ) =
P (t, T )
The value P (T, T ) can be expressed as

 T 
Z ZT ZT
1 
P (T, T ) = P (0, T ) exp r(u)du + vP0 (u, T )vP (u, T )du − vP0 (u, T )dW (u)
 2 
0 0 0
Using the formulas for V (T ) and P (T, T ), FV (T, T ) can be expressed as:
P tRi+1 vV0 (u, ti + Ti ) .vP (u, T ) + 12 vV0 (u, ti + Ti ) .vV (u, ti + Ti )

 
n−1

−
 

 1 0 du 

V (0) 
i=0 ti + 2 vP (u, T )vP (u, T ) 
FV (T, T ) = exp
P (0, T )  n−1P tRi+1 0
[vV (u, ti + Ti ) + vP0 (u, T )] dW (u)

 +

 


i=0 ti
P tRi+1 0
 
n−1
1 0 2
 −2
 

 (vV (u, ti + Ti ) + vP (u, T )) du  

i=0 ti
= FV (0, T ) exp
P tRi+1 0
n−1
(vV (u, ti + Ti ) + vP0 (u, T )) dW (u) 
 
 +

 

i=0 ti
We can now compute the value of I and II.

Calculation of II:
II = KQT (FV (T, T ) > K)
= KQT (ln FV (T, T ) > ln K)
= KΦ(D− (i ))
where Φ(x) is the cumulative probability distribution of a standard normal random variable,
and
−T ln (1 + i ) − ln P (0, T ) − 12 v 2
D− (i ) =
v
and
ti+1
n−1
XZ
2
v = (vV (u, ti + Ti ) + vP (u, T ))2 du
i=0 t
i
In the last expression, we define a2 = a0 · a where a is n × 1 matrix and a0 is its transpose.

Calculation of I:
T
I = E Q FV (T, T )1FV (T,T )>K

In order to find this expectation, we can once again change our measure. Let us consider a
measure Q̃T such that W̃ defined by
ti+1
n−1
X Z
0
vV (u, ti + Ti ) + vP0 (u, T ) 1{u≤t} du

W̃ (t) = W (t) −
i=0 t
i
40 Appendix 2.
is a Brownian motion. Then

n−1
X 0
vV (t, ti + Ti ) + vP0 (t, T ) 1{t∈[ti ,ti+1 [} dt

dW (t) = dW̃ (t) +
i=0
Under this new measure, I is given by:
I = FV (0, T )E Q̃ 1FV (T,T )>K

= FV (0, T )Q̃ [FV (T, T ) > K]

and F V (T, T ) is now expressed as:
n−1 tR
 
i+1
1 P
(vV0 (u, ti + Ti ) + vP0 (u, T ))2 du
 
 +2

 


i=0 ti
FV (T, T ) = FV (0, T ) exp
P tRi+1 0
n−1
(vV (u, ti + Ti ) + vP0 (u, T )) dW̃ (u)
 


 + 


i=0 ti
Hence:
I = FV (0, T )Q̃ [FV (T, T ) > K]
= FV (0, T )Q̃ [ln FV (T, T ) > ln K]
= FV (0, T )Φ [D+ (i )]
V (0)
= Φ [D+ (i )]
P (0, T )
where D+ (i ) = D− (i ) + v.
So eventually,
C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

with D+ (i ) and D− (i ) as given above.
Part 2
Life Insurance Contracts and the Surrender

Risk.
CHAPTER 2
Risk-Neutral Valuation of Life Insurance Contracts with a

Surrender Option.
1. Introduction.
In this chapter, we study the valuation of life insurance contracts with a surrender option.
Two broad approaches are usually distinguished in the literature. In the first, the surrender
time is defined exogenously whereas in the second one, it is defined endogenously. The first
approach consists, roughly, in using a priori fixed probabilities for the surrender time without
trying to find a structural explanation for the surrender decision. These probabilities can be
defined like survival probabilities are, i.e. through a (time-varying) deterministic instantaneous
mortality (here, surrender) rate. In probability theory, these instantaneous rates are more
commonly known as intensities. A very general form of this kind of models can be found
in the literature on disability insurance. See Linnemann [62] for a recent application of this
approach to the valuation of life insurance with a surrender option. However, this approach
has never really been favored by academics. There are mainly two related reasons. The first
one is that the deterministic intensities used in these models imply the independence between
the surrender time and the evolution of economic factors. This is not, obviously, a realistic
assumption. The second reason is that the surrender is a decision taken by the policyholders
and not a purely random event. In order to answer to these drawbacks, the academic literature
on fair valuation of insurance contracts has modelled the surrender time as an optimal stopping
time with respect to the filtration generated by the prices of the financial assets (see Bacinello
[14], [15], [16], [17], Grosen and Jorgensen [47]), Simon and Van Wouwen [73], Tak Kuen
Siu [80] or Weixi Shen and Huiping Xu [81]. In the following, we will called this kind of
optimal stopping time model the “traditional model” because of its widespread acceptance in
the academic literature. Albizzati and Geman [6] described also a kind of ad-hoc model where
both approaches are mixed together.
In this chapter, we argue that, if we are to model a realistic surrender time, we should avoid
using a stopping time with respect to the filtration generated by the financial prices, mainly
because it implicitly assumes that the insurance company and the policyholder have the same
set of information and assumes that the surrender decision is accordingly based on this set
only. On the contrary, we argue that a policyholder takes his surrender decision on a larger
set of information, which is in part, not available to the insurer. Introducing (implicitly or
explicitly) this asymmetry of information allows us to model the surrender time as a random
time that admits a so-called hazard process. This class of random times is extremely large since
essentially any random time not adapted to the filtration generated by the prices of financial
43
44 the Theoretical Framework.
assets, belongs to this class. These random times have been introduced in the default risk
literature by Jeanblanc and Rutkowski [58]. See also the book of Bielecki and Rutkowski [23].
In the financial literature, such models are known as the reduced-form models. We show this
class of random times includes exogenously as well as endogenously defined surrender times, so
that we can reconcile the two main approaches under the asymmetric information assumption.
In particular, this model includes the deterministic intensity models as well as the stochastic
intensity models which have been recently introduced in the fair valuation literature to model
the stochastic mortality (see Dahl [37]).
For the sake of simplicity, we assume there is no mortality risk.
This chapter is organized as follows. In Section 2, we present the theoretical framework.
In particular, in Subsection 2.1, we present the financial market and, in Subsection 2.2, the
life insurance contract we are going to study. In Section 3, we describe how is modelled the
information shared by the policyholder and the insurer and study the impact of introducing a
random time. In particular, in Subsection 3.4, we show the equivalence between the absence of
arbitrage and the so-called (H)-hypothesis when the financial market is complete. In Section
4, we present how the surrender time is usually modelled in the literature on fair valuation and
give its strengths and weaknesses. Then we introduce alternative random time models based
on a hazard process. In Section 5, we show how the general risk neutral valuation formulas can
be simplified in this setting. In Section 6, we study these random time in a Brownian filtration.
In Section 7, we apply this approach to the valuation of a single premium unit-linked contract
in a Gaussian framework.
2. the Theoretical Framework.

2.1. The Financial market. Let (Ω, F, P ) be a complete probability space. A perfect
frictionless financial market is defined on this space. We assume there is one locally risk free
asset denoted by St0 = S 0 (t, ω) and s risky assets Sti = S i (t, ω), i = 1 . . . s following real cádlág
stochastic processes. The price of the locally risk free asset is assumed to follow a strictly
positive, continuous path of finite variation process so that we can always find a continuous,
finite variation process D such that St0 = eDt . The discounted values of these assets are denoted
by
Si
Xti = 0t , i = 0, . . . s
St
Even though this assumption is not necessary, it is often assumed there exists a positive pre-
dictable stochastic process rt = r(t, ω) called the instantaneous risk free rate such that the
value of the locally risk free asset evolves according to the following differential equation
dSt0 = rt St0 dt
Rt
with S00 = 1. In this case, we have Dt = 0 ru du.
Let F = (Ft )t≥0 be the filtration generated by the stochastic processes S i for i = 0, . . . s.
We assume this filtration respects the usual hypotheses. Intuitively, it models the information
one has about the financial market as it evolves over time. In our context, we will assume that
this filtration models the information shared by the insurers and the policyholders.
Risk-Neutral Valuation of Life Insurance Contracts with a Surrender Option. 45
We assume furthermore that the no-arbitrage hypothesis holds in this financial market with
regard to the information set F. This assumption is equivalent to the existence of at least one
probability measure Q, equivalent to P , for which the discounted prices Xti , i = 0 . . . s are
(F, Q)-local martingales. Such a measure Q is called a local martingale measure.
2.2. The Life Insurance Contract. In life insurance, the traditional premium principle
rests on the equivalence between the present value of the insurer’s payments and the present
value of the policyholder’s payments. In this section, we describe the way we model these
payments. Ultimately, the actual payoffs the insurer gets or pays, depend on the surrender
decision. The date at which the surrender occurs is denoted τ . It is a strictly positive random
variable (F-measurable) on Ω. In other words, τ = τ (ω) is a random time. We do not assume
τ is an F-stopping time. In the following, we will use the following notation: t0 will denote
the initial date of the insurance contract, T the term of the contract and t an arbitrary date
between t0 and T . Without loss of generality, we will assume t0 = 0. For the sake of simplicity,
we do not take the mortality into account.
2.2.1. Payments of the insurer. We mainly need three building blocks. The first one is
the payoff the insurer has to pay at the term T of the contract if the policyholder has not
surrendered. This payoff is assumed to be a FT -measurable random variable and is denoted by
g(T, ω). At the term of the contract, the insurer has to pay
g(T, ω)1{τ >T }
For a traditional participating life insurance contract, this could be the guaranteed capital.
For an annuity contract, this could be the present value (at time T ) of the annuities. For a
unit-linked contract, this could be, for example, the value of the units bought, with or without
a guarantee.
The second building block is the amount the insurer has to pay when the policyholder
surrenders before the term T . This amount is denoted by 1{0<τ ≤T } R(τ, ω) where (R(t, ω))t≥0
is assumed to be an F-predictable stochastic process. If we define Ht = Ht (ω) = 1{τ (ω)≤t} , we
can write
Z T
1{0<τ ≤T } R(τ, ω) = R(u, ω)dHu
0
For a traditional participating contract, the policyholder could be allowed, for example, to get
back the premiums he has already paid, with a time-dependent penalty or not, and with or
without a guaranteed interest rate on these premiums. For a unit-linked contract, he could for
example, get back the value of the units he bought with or without penalty, or could be allowed
to get back the premiums he paid to buy these units.
The third building block is the payoffs the insurer has to pay as long as the policyholder has
not surrendered. We model these payoffs through their cumulative value up to time t, C(t, ω),
assumed to be a right-continuous increasing F-adapted process. The cumulative payoff up to
the surrender time is given by
Z T
C(T, ω)1{τ >T } + C(τ− , ω)1{0<τ ≤T } = (1 − Hu )dC(u, ω)
0
46 Risk-Neutral Valuation, Absence of Arbitrage and Enlargement of Filtration.
where we assume C(0, ω) = 0 and C(T, ω) = C(T− , ω). This could be the payments of a
constant or a time-dependent annuity. Notice, this last payoff can be written as a combination
of the two first ones.
2.2.2. Payments of the policyholder. The policyholder is committed to pay premiums pe-
riodically as long as he has not surrendered. This is similar to the second building block of
the previous subsection. We model these premiums through their cumulative value up to time
t, CP (t, ω), assumed to be an F-adapted right-continuous increasing process. The cumulative
payoff up to the surrender time is given by
Z T
CP (T, ω)1{τ >T } + CP (τ− , ω)1{t0 <τ ≤T } = CP (t0 , ω) + (1 − Hu )dCP (u, ω)
t0
We assume CP (T, ω) = CP (T− , ω). For the sake of simplicity, we will usually consider these
premiums are paid at N fixed discrete dates ti with i = 0, . . . N − 1. In this case, CP (·, ω)
is a step-wise function. The premium paid at time ti is denoted by P (ti , ω) and is equal to
P (ti , ω) = CP (ti , ω) − CP (ti− , ω). We assume CP (t0− ω) = 0. P (ti , ω) is then a random
variable Fti -measurable. The payments stream of the policyholder becomes then
N
X −1
P (ti , ω)1{τ >ti }
i=0
This way of describing a life insurance contract is fairly general. As we have seen, it can
accommodate unit-linked products as well as traditional participating contracts.
3. Risk-Neutral Valuation, Absence of Arbitrage and Enlargement of Filtration.

In order to compute the present values of these payoffs, we rely on the risk neutral valuation
principle. In this setting, the value at date t of the liabilities of the insurer or of the policyholder,
is given by the expected discounted payoffs under a risk neutral measure Q, conditionally on
the information known at date t. Since our aim is to determine the fair value of an insurance
contract from the point of view of an insurer, this expectation should be taken conditionally on
the insurer’s information. In our context, this information has a central place. Before studying
the valuation formulas, we first precisely describe how this information is modelled.
3.1. The Information. We already defined the process Ht = 1{τ ≤t} . Let us introduce a
few other definitions.
Definition 3.1. Let H = (Ht )t≥0 be the filtration generated by the process Ht i.e Ht =
σ(Hs , 0 ≤ t ≤ s).
The filtration H models the information an insurance company has over the fact that the
policyholder has already surrendered or not.
Definition 3.2. Let G = (Gt )t≥0 be the filtration defined as Gt = Ft ∨ Ht , ∀ t ≥ 0.
The filtration G models altogether the information from the financial market and if the
policyholder has surrendered or not. F is thus a subfiltration of G, i.e., F ⊆ G. We assume
G is the complete information of the insurer. In particular, this means the insurer has no
private information about its policyholders apart from wether they have already surrendered
or not. According to this assumption, G is thus the filtration under which we should take our
expectations in our valuation formulas.
Notice that if we assume τ is an F-stopping time then H ⊂ F. This means that it is
sufficient to observe the financial market to know if one has surrendered or not. In this case,
the filtration H does not offer any additional information and the filtrations F and G coincide.
Finally, in general, τ is a G-stopping time and Ht is G-adapted.
3.2. Present value of the insurer payment streams. By applying the risk neutral
valuation principle, we have the following proposition for the present value of the total liabilities.
Proposition 3.3. the value at date t, LC
t , of the total liabilities of the insurer is given by
h i
−(DT −Dt )
LCt = E Q
e g(T, ω)1{τ >T } |Gt
Z T
Q −(Du −Dt )
+ E e R(u, ω)dHu |Gt
t
Z T
Q −(Du −Dt )
(3.1) + E e (1 − Hu )dCu |Gt
t
Proof. This is a straightforward application of the risk neutral valuation principle to the
three building blocks described in the previous section.
3.3. Present value of the premium payment streams. As far as the present value At
of the liabilities of the policyholder is concerned, we have according to the risk neutral valuation
principle, the following proposition.
Proposition 3.4. The value at time t, At , of the liabilities of the policyholder is given by
Z T
At = E Q e−(Du −Dt ) (1 − Hu )dCP (u, ω) |Gt
t
If we assume the premiums are paid at N fixed discrete dates, we have
 
N −1
e−(Dti −Dt ) 1{τ >ti } P (ti , ω) |Ft 
X
(3.2) At = E Q 
i=dte
where dte = inf {i |ti > t }.

Proof. Straightforward.
In order to have a fairly valued contract at the initial date t0 of the contract, the present
values of the liabilities of both parties must be equal : Lt0 = At0 . At a later date t ≥ t0 , the
difference Vt = At − Lt gives the fair value of the contract for the insurer.
It is worth noticing that we can imagine alternative ways of modelling the payoff when
one surrenders. In the model we described above, the payment is made at the exact time of
surrender. But we can also consider that the payment is made at discrete times. For example,
it is probably more realistic to consider that the policyholder will get the amount R at the end
48 Risk-Neutral Valuation, Absence of Arbitrage and Enlargement of Filtration.
of the week or the month during which he surrenders. Let us denote tj the different times of
payment with j = 1 · · · K and tK = T . In this case, the payoff at time tj is given by:
R(tj , ω)1{tj−1 <τ ≤tj }
The present value of the commitments of the insurer in the discrete case are then
Z T
Q −(DT −Dt )
LDt = E [e g(T, ω)1τ >T + e−(Du −Dt ) (1 − Hu )dC(u, ω)
t
k
e−(Dtj −Dt ) R(tj , ω)1{tj−1 <τ ≤tj } |Gt ]
X
+
j=dte
This model can be a useful alternative to the one described above. It should also be a good
approximation of LC t as the intervals ]tj−1 , tj ] get shorter. We could also have discretized the
process C(·, ω) as we did for CP (·, ω).
3.4. The Absence of Arbitrage and the Enlargement of Filtration. In Section 2.1,
we assumed that there was no-arbitrage in the financial market with respect to the information
set F. This is a standard assumption in the financial literature. But in the last section, we
actually made a stronger assumption: by using the risk neutral valuation principle, we assumed
the existence of an equivalent measure under which the discounted prices X i are (G, Q)-local
martingales. Accordingly, we implicitly assumed there was no-arbitrage, not only, under the
filtration F but also under the new filtration G. Is this a reasonable economic assumption?
Does the observation of the surrender time affect the no-arbitrage hypothesis in the financial
market? We think the answer to this question is obvious: we cannot expect the observation of
the surrender time will give rise to arbitrage opportunities in the financial market if there is
not initially. Accordingly, the discounted prices X i should also be (G, Q)-local martingales.
In general, going from an initial filtration F to an enlarged one G, (F, Q)-local martingales
are not necessarily (G, Q)-local martingales. Accordingly, the no-arbitrage hypothesis under
the filtration F does not necessarily imply the no-arbitrage hypothesis under the filtration G.
In the following, we derive necessary and sufficient condition for this assumption to hold under
G. As we will see, it implies a constraint on the Q-distribution of the random time τ .
Let us first recall a definition and a couple of results.
Definition 3.5. Let (Ω, F, P ) be a probability space. Let G be a filtration and F be an
arbitrary sub-filtration of G i.e., for every t, Ft ⊆ Gt . We say the (H)-hypothesis holds under
Q between the filtrations F and G if and only if: for every t, F∞ and Gt are conditionally
independent with respect to Ft .
The (H)-hypothesis is equivalent to the invariance of (local) martingales for F and G. More
precisely, according to Brémaud and Yor [32], we have
Lemma 3.6. The following assertions are equivalent.
(1) The (H)-Hypothesis holds.
(2) Every (F, Q)-local martingale is a (G, Q)-local martingale.
In particular, in our setting, we have the following lemma.
Lemma 3.7. The (H)-hypothesis is equivalent to Q (τ ≤ t |F∞ ) = Q (τ ≤ t |Ft ), ∀ t ≥ 0.

Proof. See Jeanblanc and Rutkowski [58] .
We can now prove the relations between the (H)-hypothesis and the no-arbitrage hypothesis
under G.
Proposition 3.8. If there is no-arbitrage in our financial market with respect to F and if
the (H)-hypothesis holds, then there is no-arbitrage with respect to G.
Proof. Straightforward. Since there is no-arbitrage opportunity with respect to F, there
exists at least one equivalent (to P ) measure Q such that the discounted prices X i are (F, Q)-
local martingales. By Lemma 3.6, if the (H)-hypothesis holds, these discounted prices X i are
also (G, Q)-local martingales so that there is no-arbitrage opportunity in the financial market
considering the filtration G.
The following proposition gives almost the converse.
Proposition 3.9. Assume our financial market is arbitrage free and complete under F. If
there is no-arbitrage in the financial market under G then the (H)-hypothesis holds between F
and G.
Proof. See Blanchet-Scalliet and Jeanblanc [30].
If we assume market completeness under F, the (H)-hypothesis is equivalent to the no-
arbitrage hypothesis thanks to Propositions 3.8 and 3.9. Notice though that Proposition 3.8
does not require market completeness; the (H)-hypothesis is thus a sufficient condition for the
absence of arbitrage even under market incompleteness.
Actually, in general, this (H)-hypothesis is not necessarily invariant if we go from the
measure Q to an equivalent one, like the physical measure P for example, see [58] for a counter
example due to Kusuoka. If we nevertheless assume this condition is also true under the
physical measure P , we have P (τ ≤ t|F∞ ) = P (τ ≤ t|Ft ) ∀t, then in our setting, this condition
is intuitively very clear: it tells us the information over the future evolution of the financial
market (after t) does not give us more information on the surrender before t than does the
information on the financial market up to time t. As a consequence, it also tells us that the
fact that one surrenders will not affect the evolution of the financial market. This implication
is very natural. Accordingly, even though, strictly speaking, the (H)-hypothesis is only crucial
under Q, it seems to be harmless to equally impose this (H)-hypothesis under P .
Notice also this setting covers two important special cases:
(1) when τ is an F-stopping time, F and G coincide so that there is no enlargement of
filtration. The risk neutral valuation framework can then be use safely. Indeed, in
this case, the condition Q(τ ≤ t|F∞ ) = Q(τ ≤ t|Ft ) is trivially satisfied.
(2) When τ is Q-independent of F, this (H)-hypothesis directly holds since in this case,
Q(τ ≤ t|F∞ ) = Q(τ ≤ t) = Q(τ ≤ t|Ft ) ∀t. We have this assumption when the
surrender time is exogenously defined as described in the introduction. This assump-
tion also appears in the stochastic mortality literature. The arguments of this section
50 The Surrender Time.
show, as expected, the mortality cannot induce any arbitrage opportunity, under this
independence assumption.
4. The Surrender Time.

Up to now, we have not made any assumption yet on the date τ at which the surrender
occurs. In the first subsection, we describe how τ is traditionally modelled in the literature on
fair valuation of insurance contracts (see Bacinello [14]and [15] or Grosen and Jorgensen [47]).
We give the pros and cons. In the second subsection, we present an alternative model based
on an F-hazard characterization of τ and explain how it could answer to the drawbacks of the
traditional model.
4.1. τ as an optimal F-stopping time. The literature on fair valuation of insurance

contracts has mainly relied on the following definition of τ :
τ = inf {t|R(t, ω) ≥ −V (t, ω)}
where V (t, ·) = At − Lt is the fair value of the insurance contract at time t if one has not
surrendered before t and does not surrender at time t. This definition comes from the optimal
stopping time literature, in general, and from the finance literature on the optimal exercise
time of American style options, in particular. This definition tells us that it is optimal for the
policyholder to surrender when the fair value of his contract is lower or equal to the amount
that he can get by immediately surrendering this contract.
At first, this definition seems very appealing from a theoretical point of view. By defining
the surrender behavior in this way, we endogenously determine the time of surrender. Indeed,
this surrender time actually arises from the very characteristics of the insurance contract and
in the meantime, depends intrinsically on the evolution of the financial market. Unfortunately,
it appears that this definition has also some drawbacks.
Firstly, since τ is defined as an optimal stopping time, the policyholder is assumed to be
able to process all the information, to make all the required complex calculations and is able
to act in a perfectly rational manner. Obviously, this is not the way people behave, probably
not even the best actuaries.
Secondly, policyholders are not only assumed to act rationally but are also assumed to
take their decisions on exactly the same set of information F. As a consequence, it means,
for a given generation, all the policyholders would behave as one: we would observe a period
without any surrender and suddenly all would surrender at the same time. Even though the
surrender decisions among different policyholders are indeed correlated, a perfect correlation is
not realistic. This also means no room is left for idiosyncratic information that could trigger
the decision of surrender. Notice that this criticism has nothing to do with the fact τ is
defined as "optimal" but does come from the fact that τ is an F-stopping time. Notice also,
the information F is assumed to be known by the insurer so that this definition of τ does not
allowed for any asymmetry of information between the policyholders and the insurer.
Thirdly, since to our knowledge, all the continuous-time literature assumes F to be a Brow-
nian filtration, τ is then even an F-predictable stopping time. This means intuitively that, at
each time t, the insurance company is able to predict if its policyholders are going to surrender
or not during the next small interval of time dt. There is never an unexpected surrender. Ob-
viously, in the real world, an insurance company cannot perfectly predict if someone is going to
surrender or not, because the decision is taken, in part, on information that are not accessible
to the insurance company (even though the evolution of the financial market plays a role in
the decision and can give the insurer a valuable information on the probability of surrender).
One could argue we could overcome this "predictability" problem simply by using a model of
financial prices with an unpredictable component, such as a Lévy process with jumps for exam-
ple. In this case, we would have indeed unpredictable surrenders while keeping τ an F-stopping
time. The trouble with this approach is the "unpredictability" of the surrender would come
from the characteristics of the financial market and not from the surrender decision itself.
Fourthly, when τ is an F-stopping time, if the financial market is complete, so is the insur-
ance market. It gives the false impression that the surrender risk can be, at least theoretically,
perfectly hedged away and therefore, that the surrender is not really risky for an insurance
company. As we said, it misses the point that surrender is an unpredictable event and that
for this reason, surrender is risky for an insurance company. From a risk management point
of view, it is clear we cannot rely on and base decisions on this kind of model. For pricing
purposes, one could still accept this definition, arguing it gives us an upper bound of the fair
value of a life insurance contract. The trouble is that using different models for pricing and
risk management, that are by nature, not consistent with each other, is probably not very
appropriate to manage an insurance company.
Finally, the value V (t, ω) that we compare to the surrender value R(t, ω), should be the
value of the contract from the policyholder perspective. Unfortunately, usually, the value
V (t, ω) used in the literature is the value of the contract from the insurer perspective (see also
Bacinello [17] for a discussion of this point). The value of the contract from the policyholder
perspective depends actually on his own information and its own risk aversion. It is thus a
subjective value. By definition, a subjective value is different from one policyholder to another
and is unknown for the insurer. Accordingly, this argument implies once again the existence of
an asymmetry of information.
In conclusion, contrary to the traditional model of τ , we argue that the surrender decision
should be random with respect to the shared information F, even though it should not be
independent of it. This information should actually influence the randomness of the surrender
decision. This leads to the conclusion that the F-measurability of the surrender decision is
really the crucial point. In other words, we argue that we should avoid the hypothesis that τ
is an F-stopping time.
4.2. τ as a random time characterized by an F-hazard process. In the previous

subsection, we saw that using optimal stopping times to model the surrender decisions, does not
allow us to model realistically neither the behavior of the policyholder nor the unpredictable
component of the surrender decision for the insurance company (at least when F is the Brownian
filtration). We also saw that it was crucial to avoid defining τ as an F-stopping time, but that
it was also crucial to allow the probability of surrender to depend on the information F. In this
subsection, we argue that we can introduce an asymmetry of information by modelling τ as a
random time that admits an F-hazard process. Let us first introduce our definition of τ and
the definition of the associated hazard process.
Definition 4.1. Let τ be a non negative random variable defined on (Ω, F, P ). Assume
P (τ = 0) = 0 and P (τ > t) > 0), ∀ t ≥ 0. We denote Ft = P (τ ≤ t |Ft ) ∀ t ≥ 0. We define the
(F, P )-hazard process of τ , denoted Γ by
Ft = 1 − e−Γt
∀ t ≥ 0.
In order to have Γt well-defined for all t, we have to impose the condition Ft < 1 because
Γt = −ln(1 − Ft ). This excludes τ of being an F-stopping time. In other words, an F-stopping
time cannot admit an F-hazard process. Notice that Ft is an F-adapted stochastic process and
admits a right continuous modification. Accordingly, Γt is also an F-adapted càdlàg stochastic
processes with Γt < ∞ ∀t and Γ0 = 0.
As already explained, when τ is an F-stopping time, we can tell from the observation of the
financial market if a policyholder has already surrendered or not. By modelling τ as defined
above, we exclude this situation. Here, the surrender decision depends necessarily on elements
outside the information F. These elements are idiosyncratic information can be different from
a policyholder to another (unless the surrender decision would still be perfectly correlated).
Accordingly, in defining τ as above, we implicitly model an asymmetry of information between
the policyholders and the insurer. For the same reason, the surrender decision is indeed unpre-
dictable for the insurer, even when F is a Brownian filtration. This unpredictability here does
not come from the characteristics of the financial market but from the asymmetry of informa-
tion itself. Notice that even though the surrender decision is unpredictable, the probability of
surrender is stochastic and depends on the evolution of the financial market (through Γt ). Let
us give some examples of such random times.
Example.
(1) Probably the most simple random time having a hazard characterization is an expo-
nential random time τ with constant intensity λ > 0. Indeed we have:
Ft = 1 − P (τ > t|Ft )
= 1 − P (τ > t)
= 1 − e−λt
The hazard process is here an increasing positive deterministic function, equals to
Γt = λt.
(2) The time of the first jump of inhomogeneous Poisson process independent of F, with
hazard function Λ(t) (a deterministic cádlág function), is also such a random time.
Ft = 1 − P (Nt = 0)
= 1 − e−Λ(t)
Again, the hazard process Γ is here an increasing positive deterministic function
Γt = Λ(t). More formally, this function Λ(t) is actually a measure defined on the
measurable space (R, B) where B is the Borel σ-algebra on R. If this measure is abso-
lutely continuous with regard to the Lebesgue measure, we can find the intensity λ(t)
of the Poisson process, which is a positive function, given by
Z t
Λ(t) = λ(u)du
0
(3) The definition of τ given above also includes the time of the first jump of a Cox
process. This process is a counting process that generalizes the Poisson process to the
case where the measure Λ is a random measure on (R, B). Such a counting process Nt
is also known as a doubly stochastic Poisson process. We can define a Cox process in
the following manner. Nt follows a Cox process if and only if conditionally on Ft , Nt
follows an inhomogeneous Poisson process for all t.
The hazard function of the inhomogeneous process can be given, conditionally on
Ft , by the realization of an increasing positive cádlág F-adapted stochastic process
Λt . Since Λt is assumed F-adapted, Λt is indeed, conditionally on Ft , a deterministic
function. Furthermore, since each path is increasing, cádlág and positive, Λt induced,
conditionally on Ft , a deterministic measure on (R, B) and unconditionally, a random
measure on the same space. See Grandell [45] for a more formal description.
According to this definition, conditionally on Ft , the distribution of the time of
the first jump of a Cox process is given by
Ft = 1 − P (Nt = 0|Ft )
= 1 − e−Λt
Once again, this random time admits an increasing, positive F-hazard process Γt given
by Γt = Λt . Unconditionally, we also have
P (τ ≤ t) = E P [1 − e−Λt ]
= 1 − E P [e−Λt ]
If Λt is absolutely continuous with respect to the Lebesgue measure, it exists a unique
positive F-adapted predictable process λt called the (F, P )-intensity process of N, such
that Z t
Λt = λu du
0
A few elements are worth underlining. Firstly, in the two first examples, the hazard process is
a deterministic function and τ cannot depend on the financial market. In the case of the Cox
process, the hazard process is a stochastic process which allows the probability of the surrender
time to depend on the evolution of the financial market.
Secondly, the examples of random times given here are, all three, the time of the first arrival
of a counting process. But we would like to highlight that these are just special cases of our
definition and do not represent all the random times that admit a hazard process. Actually, as
long as we construct a random time with P (τ ≤ t|Ft ) 6= 1, we can find a process Γt .
Thirdly, in these three examples, the hazard functions or the hazard process are increasing
since each three induces a (deterministic or random) measure on the Borel sets. Our definition
of τ does not depend on this assumption. As we said, these examples are just special cases
of our definition and in general we can find random times admitting positive non necessarily
increasing hazard process. So, Γt increasing is not a necessary assumption. However, the (H)-
hypothesis, we have seen in Section 3.4, implies the process F̂t defined as F̂t = Q(τ ≤ t|Ft ), is
an increasing process. The (F, Q)-hazard process Γ̂t defined as Γ̂t = −ln(1 − F̂t ) similarly to
Rt
the Definition 4.1, is then also an increasing process and if λ̂t defined as Γ̂t = 0 λ̂u du, exists,
then it is a positive process. So if we assume the (H)-hypotheis holds then Γ̂t increasing is a
necessary condition for constructing τ but Γt increasing, is not since as we said, the
(H)-hypothesis is not necessarily true under P and does not have to be1. This fact justifies
that we can focus, under Q, on Cox processes to model the surrender time.
4.3. Construction of τ . The hypothesis made on τ in the previous subsection, are very
weak and allows us to construct τ in a great variety of ways. Actually, the two broad approaches
described in the introduction (the exogenous and endogenous one) are compatible with our
definition of τ .
The first one consists in describing explicitly the surrender behavior and the asymmetry
of information between the policyholder and the insurer. In other words, it consists in de-
scribing the surrender from the policyholder’s point of view. For example, we can generalize
the traditional modelling of the surrender time. We saw that the decision of surrender was
taken on a larger set of information that includes idiosyncratic information. Let us denote
this policyholder’s information set by a filtration J with F ⊂ J. We could consider τ to be a
J-optimal stopping time instead of an F-optimal stopping time. For example, we could model
a stochastic mortality which would depend on information related to the health of the policy-
holder and only known by himself and not the insurer. In this case, we could still defined τ as
τ = inf {t|R(t, ω) ≥ −V (t, ω)} but where −V (t, ω) is the value of the insurance contract for the
policyholder but, this time, given his own information J. In this case, τ would be endogenously
defined but still not F-measurable. Accordingly, we could then derive a probability of surrender
F̂t 6= 1 ∀ t such that Γ̂t = −ln(1 − F̂t ) makes sense. Even if this model would probably be too
complex to deal with, it is interesting to realize that this approach does not exclude structural
models of the surrender time. Another simpler example of this approach is the following. Let
us assume that one surrenders when the value of a given financial asset St reaches a certain
level, E. This level at which one surrenders, is specific to each policyholder and is unknown by
the insurer. The insurance contract portfolio is then heterogeneous with respect to this level.
Finally, let assume this level follows a unit exponential distribution under Q and that E is
independent of F. In this case, we have F̂t = 1 − e−sup0<s≤t Ss so that the (F, Q)-hazard process
is given by Γ̂t = sup0<s≤t Ss which is indeed F-adapted, cádlág, positive and increasing. Our
point in giving these examples is to show our definition allows a structural model of τ or in
other words, allows an endogenous definition of the surrender time.
The second approach considers the insurer point of view and is closer to an exogenous
specification of τ . It simply consists in specifying directly the process followed by Γ̂ without
1In Section 3.4, we argued it is probably a reasonable and harmless assumption to assume the (H)- hy-
pothesis holds under P too. In this case, Γt is obviously also an increasing process.
necessarily trying to describe explicitly the actual surrender behavior nor the asymmetry of
information. One way to construct the random time τ given a hazard process, is the following.
Let E be a random variable (F-measurable) defined on Ω, but which is independent of the
filtration (Ft )0≤t≤T . E is assumed to have an unit exponential distribution under Q. Let Λ̂ be
an F-adapted increasing positive process with Λ̂0 = 0 and Λ̂t < ∞. If τ is defined as
τ = inf {t|Λ̂t ≥ E}
then τ is a random time with an F-hazard process under Q given by Γ̂ = Λ̂. Indeed, we have:
Q(τ ≤ t|Ft ) = Q(Λ̂t ≥ E|Ft )
= 1 − e−Λ̂t
Alternatively, we can also start with a positive F-adapted predictable stochastic process λ̂ and
define τ as Z t
τ = inf{t| λ̂u du ≥ E}
u=0
Rt
In this case, τ is a random time with an (F, Q)-hazard process given by Γ̂t = u=0 λ̂u du and with
an (F, Q)-intensity process equals to λ̂. Indeed, in both cases, the random times constructed
can be seen as the time of the first jump of Cox processes. Both of these constructions also
offer an easy way to simulate the random time τ . We simply have to simulate the hazard
process and an independent exponential random variable. The two methods described above
are standard constructions which have the advantage of being very convenient but we can go
further at the cost of constructing random times less tractable. For example, even if Λ̂t is not
an increasing process, we can still define τ in the same way. Let Λ̂ be an F-adapted positive
process with Λ̂0 = 0. We can define τ = inf {t|Λ̂t ≥ E}. In this case, we have
Q(τ ≤ t|Ft ) = Q( sup Λ̂s ≥ E|Ft )
0<s≤t
= 1 − e−sup0<s≤t Λ̂s
The random time τ admits an (F, Q)-hazard process characterization with the hazard process
Γ̂t = sup0<s≤t Λ̂s . This random time can still be seen as the time of the first jump of a
Cox process since sup0<s≤t Λ̂s is an increasing cádlág positive F-adapted process. Obviously,
this form is less tractable than the other two, but can be sometime justified. Notice the
second example of the first approach can be seen as an application of this last construction.
Accordingly, the 2 approaches we gave, are not necessarily antagonist. Rt
In practice, one would probably prefer to specify directly λ̂ defined as Γ̂t = 0 λ̂u du instead
of Γ̂ since the first one has an intuitive meaning the second lacks. It is well known that the
intensity parameter λ or the intensity function λ(t) of a Poisson process corresponds to what
is known in life insurance as the instantaneous mortality rate. When it exists, we have the
well-known following expression:
d ln(1 − P (τ ≤ t)) P (t < τ ≤ t + ∆t|τ > t)
λ(t) = − = lim
dt ∆t↓0 ∆t
56 The Risk Neutral Valuation Formulas.
So informally, λ(t)dt is equal to the probability one surrenders during the next interval dt if
one has not yet surrendered before time t. This conditional probability is here deterministic
for all t. When it exists, we have a similar result for the (F, Q)-intensity process
Q(t < τ ≤ t + ∆t|τ > t ∪ Ft )
λ̂t = lim
∆t↓0 ∆t
So informally, λ̂t dt is again equal to the probability one surrenders, under Q, during the next
interval dt if one has not surrendered before time t and knowing the information Ft . Obviously,
this probability is stochastic and for a given time t, its value depends on the financial market.
To model this instantaneous probability, we can make an explicit reference to the financial
asset prices, by specifying λ̂ as a function of these prices. In most applications, in order to
get tractable formulas, we will restrict ourselves to Markov intensity processes. Let λ̂(·, ·) be a
function from R × Rs+1 → R+ . In this case, we will define the intensity process of τ as:
λ̂t (ω) = λ̂(t, St (ω))
Notice this stochastic process is indeed F-adapted.
To conclude, the definition of τ as a random time characterized by an (F, Q)-hazard process
allows us to specify the surrender time in a great variety of ways and seem to be a valuable
alternative to the traditional modelling. Notice that we constructed the distribution of τ
directly under Q through the specification of Γ̂. To find an explicit functional between Γ̂ and Γ
is not easy task in general. In Section 6, we give a simple functional under certain simplifying
assumptions.
5. The Risk Neutral Valuation Formulas.

In this section, we show how the risk neutral valuation formulas of Sections 3.2 and 3.3
can be modified when we assumed τ is characterized by an F-hazard process or an F-intensity
process. In the first subsection, we introduce a number of useful results that allows us to get
rid of the indicator function by taking conditional expectations with respect to Ft instead of
Gt . In the second and third subsections, we apply these formulas to our valuation problem.
5.1. Conditional expectations with respect to Ft . In this subsection, we assume the
random time τ admits an F-hazard process Γ under the probability measure considered.
Lemma 5.1. Let X be a Fs -measurable random variable with s ≥ t. We have
h i
E 1{τ >s} X |Gt = 1{τ >t} E e−(Γs −Γt ) X |Ft

If moreover, Γ is absolutely continuous, it exists an

R t F-predictable measurable process , referred
to as the F-intensity process of τ , such that Γt = 0 λu du. The last formula becomes
h Rs i
E 1{τ >s} X |Gt = 1{τ >t} E e− t λu du X |Ft

Proof. See Jeanblanc and Rutkowski [58].

In Section 7 when we study unit-linked contracts, we will actually only use this first lemma.
Lemma 5.2. Let Z be an F-predictable process. Then for any t < s ≤ ∞, we have
Z s Z s
Γt
E Zu dHu |Gt = 1{τ >t} e E Zu dFu |Ft
t t
If Ft is a continuous increasing process, then

Z s Z s
Γt −Γu
E Zu dHu |Gt = 1{τ >t} e E Zu e dΓu |Ft
t t
Rt
If moreover, τ admits an F-intensity process such that Γt = 0 λu du, we have
Z s Z s
− tu λv dv
R
E Zu dHu |Gt = 1{τ >t} E λu Zu e du |Ft
t t

Lemma 5.3. Let C be a right continuous F-adapted process. If Ft follows a process of finite
variation then for every t ≤ s
Z s Z s
Γt −Γu
E (1 − Hu )dCu |Gt = 1{τ >t} e E e dCu |Ft
t t
Rt
If moreover, τ admits an F-intensity process such that Γt = 0 λu du, we have
Z s Z s R
Γt − tu λv dv
E (1 − Hu )dCu |Gt = 1{τ >t} e E e dCu |Ft
t t
Proof. A proof can be found in Jeanblanc and Rutkowski for an F-predictable process C
but this assumption is actually not used in their proof.
If the (H)-hypothesis holds under the probability measure considered, it implies Ft is an

increasing process, and Ft is then indeed of finite variation.
5.2. Present value of the insurer’s payments. In order to simplify Equation (3.1), we
can directly apply the 3 lemmas of the last subsection. We have the following proposition.
Proposition 5.4. Let us assume F̂t is an increasing process2, the present value of the
insurer payments is given by
h i
−(DT −Dt ) −(Γ̂T −Γ̂t )
LCt = 1 {τ >t} E Q
e e g(T, ω) |Ft
Z T
Q −(Du −Dt ) Γ̂t
+ 1{τ >t} E e e R(u, ω)dF̂u |Ft
t
Z T
Q −(Du −Dt ) −(Γ̂u −Γ̂t )
(5.1) + 1{τ >t} E e e dCu |Ft
t
2This assumption is necessary if we want the (H)-hypothesis to hold.

58 The Risk Neutral Valuation Formulas.
If F̂t is an increasing continuous process and τ admits an (F, Q)-intensity process λ̂ and if an
instantaneous risk-free rate rt exists, then the present value of the insurer payments is given by
h RT i
LCt = 1{τ >t} E Q e− t (ru +λ̂u )du g(T, ω) |Ft
Z T
− tu (rv +λ̂v )dv
R
Q
+ 1{τ >t} E λ̂u e R(u, ω)du |Ft
t
Z T
− tu (rv +λ̂v )dv
R
Q
(5.2) + 1{τ >t} E e dCu |Ft
t
5.3. Present value of the policyholder’s payments. Again, we can apply Lemma 5.1
to Equation (3.2). We then have the following proposition.
Proposition 5.5. Let us assume the premiums are paid at fixed discrete dates ti with
i = 0, . . . N − 1, then the present value of the policyholder’s payments is given by
 
N −1
e−(Dti −Dt ) e−(Γ̂ti −Γ̂t ) P (ti , ω) |Ft 
X
At = 1{τ >t} E Q 
i=dte
If τ admits an (F, Q)-intensity process λ̂ and if an instantaneous risk-free rate rt exists, we

have
 
NX−1 R ti
At = 1{τ >t} E Q  e− t ru +λ̂u du P (ti , ω) |Ft 
i=dte
where dte = inf {i |ti > t }.

These expressions generalize well known results when τ is stochastic but independent of
F like in the stochastic mortality models. We see Γ̂t plays the same role as Dt or when they
exist, the intensity process λ̂t is similar to the short term rate rt . This means the value of
a life insurance contract with a surrender risk can be considered, except for the second term
in Equations (5.1) and (5.2), to be equal to the value of the same contract without surrender
risk but under a modified stochastic term structure Dt + Γ̂t . This also means all the results
developed in the extensive literature on option pricing models with stochastic interest rates
can be applied, mutatis mutandis, to the valuation of life insurance contracts with stochastic
surrender.
5.4. Change of measure. Change of measure techniques have been very useful to derive
closed form solution of option prices with stochastic term structure, especially in Gaussian mod-
els. We can use a similar change of measure in models with stochastic surrender characterized
by a hazard or an intensity process when we have expectations of the form
h i
1τ >t E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft
where X is a FT -measurable random variable. We can rely on the following proposition to
simplify this kind of expression.
Proposition 5.6. Let QT be the measure equivalent to Q on (Ω, FT ) whose Radon-Nikodym

derivative is given by
dQT e−DT
=
dQ |FT E Q [e−DT |F0 ]
Q-a.s. Let QS(T ) be the measure equivalent to QT on (Ω, FT ) whose Radon-Nikodym derivative
is given by
dQS(T ) e−Γ̂T
=
dQT |FT
h i
E QT e−Γ̂T |F0
QT -a.s.
If X is QS(T ) -integrable, then we have
h i T
h i
1τ >t E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft = 1τ >t P (t, T )E Q e−(Γ̂T −Γ̂t ) X |Ft
T
h i S(T )
= 1τ >t P (t, T )E Q e−(Γ̂T −Γ̂t ) |Ft E Q [X |Ft ]
S(T )
= P (t, T )QT (τ > T |Gt ) E Q [X |Ft ]
Proof. See Appendix 1.
In the financial literature, the measure QT is called the T -forward martingale measure. By
analogy, we called the measure QS(T ) the T -forward surrender measure. In the last proposition,
−(D
Q
P (t, T ) = E e T −Dt ) |F
t is obviously the price of a zero-coupon bond at time t with
maturity T , and QT (τ > T |Gt ) is the probability under the T -forward martingale measure of
no surrender before T conditionally on Gt .
6. The Brownian Filtration.

Up to now, we have made weak assumptions about the distribution of the different processes.
In this section we’re going to make more specific assumptions, namely we assume
(1) the filtration F is the Brownian filtration
(2) τ admits a continuous increasing (F, P )-hazard process Γt .
(3) the (H)-hypothesis holds under P so that the Brownian motion is also a Brownian
motion with respect to G under P .
In this setting, we give a change of measure formula that allows us to go from the real measure
P to an equivalent one. It also provides a simple functional between Γ and Γ̂.
6.1. Change of measure and choice of the local martingale measure. The risk
neutral valuation principle rests on the choice of a local martingale measure Q equivalent to P .
The following lemma gives the set of measures Q equivalent to P , that includes the equivalent
local martingale measure that we should use in our valuation formulas.
60 The Brownian Filtration.
Lemma 6.1. Any probability measure Q equivalent to P has a Radon-Nikodym derivative

with respect to P , ηt , with the following representation
Z t Z t
ηt = 1 + ξu dWu + ζu dMu
0 0
where Wt is a G-Brownian motion under P , Mt = Ht − Γt∧τ is a G-martingale under P and
ξt and ζt are G-predictable processes.
Since ηt is strictly positive, it can also always be written as
Z t Z t
ηt = 1 − ηu− βu dWu + ηu− κu dMu
0 0
where βt and κt > −1 are G-predictable processes. It is well-known that such a stochastic dif-
ferential equation has a unique solution given by the following product of Doleans exponentials
Z · Z ·
ηt = Et κu dMu Et − βu dWu
0 0
Z t∧τ Z t
1 t
Z
2

= 1 + κτ 1{τ ≤t} exp − κu dΓu exp − βu dWu − |βu | du
0 0 2 0
0
Remark 6.2. For any G-predictable process κu , we can find an F-predictable process κu
0
such that κu = κu on {τ > t}. Since in this section, we only use the value of κt on τ > t, we
can, without loss of generality, assume κ is F-predictable.
A Girsanov-like theorem can be formulated in this setting.
Lemma 6.3. Any probability measure Q equivalent to P has a Radon-Nikodym derivative
given by
dQ
= ηt
dP |Gt
= ε1t ε2t
with
Z t∧τ
ε1t

= 1 + κτ 1{τ ≤t} exp − κu dΓu
0
Z t
1 t
Z
ε2t = exp − βu dWu − |βu |2 du
0 2 0
where κ and β are G-predictable processes. Furthermore, the process
Z t
Ŵt = Wt + βu du
0
follows a (G, Q)-Brownian motion, and the process M̂t is given by

Z t∧τ
M̂t = Ht − (1 + κu )dΓu
0
is a (G, Q)-martingale.
Proof. The proofs of Lemmas 6.1 and 6.3 can be found in Jeanblanc and Rutkowski
[58].
However, this set of equivalent measure is too large for our purpose. As we said, it is
crucial that the (H)-hypothesis holds under Q and we know that this property is not necessarily
invariant from an equivalent measure to another. The set of equivalent measures, such that
the (H)-hypothesis holds, forms only a subset of all the equivalent measures. A sufficient
condition for the equivalent measure Q to respect the (H)-hypothesis, is that β should not be
G-predictable but F-predictable processes. We have the following proposition.
Proposition 6.4. Let Q be an equivalent measure to P . If the process β given in Lemma

6.3 is F-predictable then the (H)-hypothesis holds under Q.
Proof. We have to prove
(6.1) Q (τ ≤ t |F∞ ) = Q (τ ≤ t |Ft )
Let us first consider the left-hand term of this equality. We have
E P ε1∞ ε2∞ 1{τ ≤t} |F∞

Q (τ ≤ t |F∞ ) =
E P [ε1∞ ε2∞ |F∞ ]
E P ε1∞ 1{τ ≤t} |F∞

=
E P [ε1∞ |F∞ ]
h R τ ∧∞ i
E P 1 + κτ 1{τ ≤∞} e− 0 κu dΓu 1{τ ≤t} |F∞
=
E P [ε1∞ |F∞ ]
h R τ ∧t i
EP 1 + κτ 1{τ ≤t} e− 0 κu dΓu 1{τ ≤t} |F∞
=
E P [ε1∞ |F∞ ]
h R τ ∧t i
EP 1 + κτ 1{τ ≤t} e− 0 κu dΓu 1{τ ≤t} |Ft
=
E P [ε1∞ |F∞ ]
The second equality comes from the assumption that ε2∞ is F∞ -measurable. The last one comes
from the fact that the (H)-hypothesis holds under P and that this hypothesis implies for any
random variable A, Gt -measurable, we have E P [A |F∞ ] = E P [A |Ft ]. If we now consider the
62 The Brownian Filtration.
denominator of this last equation, we have

Z ∞ Rv
P
(1 + κv ) e− 0 κu dΓu dP (τ ≤ v |F∞ )
1
E ε∞ |F∞ =
Z0 ∞ Rv
= (1 + κv ) e− 0 κu dΓu dP (τ ≤ v |Fv )
Z0 ∞ Rv
(1 + κv ) e− 0 κu dΓu d 1 − e−Γv

=
Z0 ∞ Rv
= (1 + κv ) e− 0 (1+κu )dΓu dΓu
Z0 ∞ Rv
= d 1 − e− 0 (1+κu )dΓu
0
= 1
The first equality comes from the assumption κ is F-adapted. The second equality comes from
the fact the (H)-hypothesis holds under P . The fourth and fifth equalities come from the fact
Γu is a finite variation continuous process. Eventually, we get for the left hand term of Equation
(5.3)
Q (τ ≤ t |F∞ ) = E P ε1t 1{τ ≤t} |Ft

As far as the right-hand side of Equation (5.3) is concerned, we have

E P ε1∞ ε2∞ 1{τ ≤t} |Ft

Q (τ ≤ t |Ft ) =
E P [ε1∞ ε2∞ |Ft ]
E P ε1t ε2t 1{τ ≤t} |Ft

=
E P ε1t ε2t |Ft

E P ε1t 1{τ ≤t} |Ft

=
E P ε1t |Ft

since ε2t is Ft -measurable. P ε1 |F

Since we know E ∞ ∞ = 1, by the law of iterated expectation,
P
1
we can easily show E εt |Ft = 1.
We already said that it is not in general an easy task to find an explicit functional between
Γ and Γ̂. In the case of the Brownian filtration and when β (and κ) is F-predictable, we can
find a simple relation between Γ and Γ̂. We have the following proposition.
Proposition 6.5. Let Q be an equivalent measure to P . If the processes β and κ of Lemma
6.3 are F-predictable then the process Γ̂t defined by
Z t
Γ̂t = (1 + κu )dΓu
0
is the F-hazard process of τ under Q. If τ admits an F-intensity process λ under P , then the
process λ̂t defined by
λ̂t = (1 + κu )λu
is the F-intensity of τ under Q.

Proof. We have
e−Γ̂t = Q (τ > t |Ft )
E P ε1t 1{τ >t} |Ft

=
E P ε1t |Ft

since ε2t is Ft -measurable. From the previous proposition, we know that E P ε1t |Ft = 1. Then

let us first consider the numerator. We have
h R τ ∧t i
E P ε1t 1{τ >t} |Ft = E P 1 + κτ 1{τ ≤t} e− 0 κu dΓu 1{τ >t} |Ft

h Rt i
= E P e− 0 κu dΓu 1{τ >t} |Ft
Rt
= e− 0 κu dΓu E P 1{τ >t} |Ft

Rt
= e− 0 (1+κu )dΓu
The third equality comes from the assumption that κ is F-adapted.

As we said, η is the Radon-Nikodym derivative of Q with respect to P . The problem is
now to find the processes βt and κt that we should use in our valuation formulas.
Let us start with βt first. If we had no surrender risk, η would reduce to ε2 and we would get
the standard case extensively studied in financial economics. In this case, we should choose βt
such that the discounted financial prices are local martingales under the associated equivalent
measure. βt would then correspond to the price of market risk. There is no difference in our
setting; we should simply choose βt as if there was no surrender risk. Let us finally notice that
when the financial market is complete, the local martingale measure of the financial market is
unique and so is βt .
The choice of κt is more problematic. κt is the price of the pure surrender risk. It determines
how the hazard or the intensity process should be adjusted for the pure risk of surrender.
Unfortunately, we do not have any market on which assets affected by surrender risk are
traded. So, even if the financial market is complete, the "life insurance market" is not with
respect to this surrender risk. The equivalent local martingale measure related to this risk
is therefore not unique. We can neither extract the market price of this risk from the values
of assets affected by surrender since as we just said there is no such market. However, this
pure surrender risk correspond to the idiosyncratic risk component of the surrender decision.
If we assume the idiosyncratic component is independent from a policyholder to another, we
can argue that this pure surrender risk can be diversified away. Accordingly, in a competitive
market, no remuneration for this risk should be required and its price could be considered
equal to 0. Another argument pleading for a pure surrender risk equal to 0, is to recognize
that the associated local martingale measure corresponds to the minimal martingale measure
of Schweizer. See Chapter 4 for more details.
If we set κt = 0 ∀t, we see that τ admits the hazard process Γ̂t = Γt or the intensity process
λ̂t = λt under Q. This last equality simply means the processes Γ or λ are not modified if we
64 Application to Unit-Linked Contracts.
set the price of surrender risk equal to 0, but it does not mean the probabilities of surrender are
equal under the local martingale measure and the physical measure. Indeed, these processes
depend on the Brownian motion and are accordingly adjusted under the local martingale mea-
sure, by the price of market risk. In conclusion, by setting κ = 0, the idiosyncratic component
of the surrender decision is not adjusted for the risk, but the common component of the sur-
render risk, which is linked to the financial market and is not diversifiable, is indeed adjusted
for the risk through the market price of risk.
7. Application to Unit-Linked Contracts.

In this section, we apply the framework described in the previous sections to the valuation of
unit-linked contracts. We study single premium contracts and periodic premium contracts. We
do not make any specific assumptions on the distribution followed by our stochastic processes
other than that the discounted prices of the financial assets are, under Q, not only F-martingales
but also G-martingales. This is a slightly stronger assumption than the (H)-hypothesis. In this
section, we derive general formulas for the valuation of unit-linked contracts. You will notice
we only need Lemma 5.1 of Section 5.1 to derive these valuation formulas so that τ does not
need to admit an (F, Q)-intensity process nor does Γ̂ need to be continuous.
7.1. Single premium contract. We assume a single premium P is paid at time t. With
this premium, the policyholder acquires n units of a stock index with value S(t) at time t, as
well as a guarantee g at the term T of the contract. This guarantee could be for example a
guaranteed rate on the initial value of the units, a guaranteed rate on the premium paid, etc.
When the policyholder surrenders, we assume he receives a proportion (1 − α) of the n
units. So, α is here the penalty when one surrenders. We assume the payment is made at the
precise time of surrender τ . Accordingly, using our notation, we have:
R(τ ) = (1 − α)nS(τ )
g(T, ω) = max(nS(T ), g)
The value at date t of the liabilities is given by:
h i
Lt = E Q e−(DT −Dt ) g(T, ω)1τ >T + e−(Dτ −Dt ) R(τ )1{t<τ ≤T } |Gt
We can simplify this equation as followed

h i
Lt = E Q e−(DT −Dt ) 1{τ >T } max (nS(T ), g) |Gt
h i
+ E Q e−(Dτ −Dt ) (1 − α)nS(τ ) 1{τ >t} − 1{τ >T } |Gt
h i
= E Q e−(DT −Dt ) 1{τ >T } max (nS(T ), g) |Gt
| {z }
I
h i
+ E Q (1 − α)nS(τ )e−(Dτ −Dt ) 1{τ >t} |Gt
| {z }
II
h i
− E Q (1 − α)nS(τ )e−(Dτ −Dt ) 1{τ >T } |Gt
| {z }
III
The terms I, II and III can be simplified as follow:

h i
I = E Q e−(DT −Dt ) 1{τ >T } max (nS(T ), g) |Gt
h i
= 1{τ >t} E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) max (nS(T ), g) |Ft
h i
II = E Q (1 − α)nS(τ )e−(Dτ −Dt ) 1{τ >t} |Gt
h i
= 1{τ >t} (1 − α)nE Q e−(Dτ −Dt ) S(τ ) |Gt
= 1{τ >t} (1 − α)nS(t)
In the last equality, we use the hypothesis that all F-martingales are also G-martingales.
h i
III = (1 − α)nE Q e−(Dτ −Dt ) S(τ )1{τ >T } |Gt
h h i i
= (1 − α)nE Q e−(DT −Dt ) E Q S(τ )e−(Dτ −DT ) 1{τ >T } |GT |Gt
h h i i
= (1 − α)nE Q e−(DT −Dt ) 1{τ >T } E Q S(τ )e−(Dτ −DT ) |GT |Gt
h i
= (1 − α)nE Q e−(DT −Dt ) 1{τ >T } S(T ) |Gt
h i
= 1{τ >t} (1 − α)nE Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) S(T ) |Ft
In the last but one equality we use the hypothesis that the discounted price of the financial
asset is, under Q, a G-martingale. Eventually, we have
h i
Lt = 1{τ >t} e(Dt +Γ̂t ) E Q e−(DT +Γ̂T ) max (nS(T ), g) |Ft
n h io
(7.1) + 1{τ >t} (1 − α)n S(t) − e(Dt +Γ̂t ) E Q e−(DT +Γ̂T ) S(T ) |Ft
When an intensity process and an instantaneous risk free rate exist, this last expression becomes
h RT i
Lt = 1{τ >t} E Q e− t (ru +λ̂u )du max (nS(T ), g) |Ft
n h RT io
+ 1{τ >t} (1 − α)n S(t) − E Q e− t (ru +λ̂u )du S(T ) |Ft
These last two expressions are very simple ones that involve the calculation of the price of a
call options with a modified term structure model DtΓ̂ = Dt + Γ̂t or in the second case, with a
modified instantaneous risk free interest rate rtλ̂ = rt + λ̂t . It can be solved, in a number of cases,
in closed-form or in quasi-closed form by Fourier transform for example. Notice, we have not
made any assumption about the form of the processes S(t), Dt or Γ̂t , so that Equation (7.1)
is thus a general valuation formula for single premium unit linked contracts with stochastic
interest rate and stochastic surrender.
7.2. Periodic Premium contract. Here, we assume premiums are paid at time ti with
i = 0 . . . N − 1 if the policyholder has not surrendered. We assume the value of these premiums
is constant and equal to P . As we’ve already seen, the payoff is equal to:
N
X −1
P 1τ >ti
i=0
The value of this payoff at date t, is given by

 
N
X −1
At = E Q  e−(Dti −Dt ) P 1τ >ti |Gt 
i=dte
Using the results of the previous sections, we get

N
X −1 h i
At = 1{τ >t} P E Q e−(Dti −Dt ) e−(Γ̂ti −Γ̂t ) |Ft
i=dte
When an intensity process and an instantaneous risk free rate exist, this last expression becomes
N −1 h R ti i
E Q e− t (ru +λ̂u )du |Ft
X
(7.2) At = 1{τ >t} P
i=dte
As far as the value of the liabilities is concerned, we distinguish two cases. In the first one, the
number of units that the policyholder receives, depends on the value of the units at the time
of payment. In the second case, we assume that the payment of a premium gives the right to
the policyholder, to n units whatever the value of these units at the time of payment. We call
the first kind of contract, a periodic premium contract of type I and the second one, a periodic
premium contract of type II.
7.2.1. Periodic Premiums contract of type I. In this case, we assume a proportion (1 − ρ)

of the premium P paid at time ti is used to buy nti units. This number of unit nti is given by
(1 − ρ)P
nti =
S(ti )
The present value of the liabilities is given by
N −1
" ! #
X
Q −(DT −Dt )
Lt = E e max S(T ) nti , g 1τ >T |Gt
i=0
   
bτ c
X
+ E Q e−(Dτ −Dt ) (1 − α)S(τ )  nti  1{t<τ ≤T } |Gt 
i=0
We show in Appendix 2 that we can rewrite this equation as:

N −1
" ! #
X S(T )
Lt = 1{τ >t} e(Dt +Γ̂t ) E Q e−(DT +Γ̂T ) max (1 − ρ)P , g |Ft
S(ti )
i=0
btc
1 n h io
S(t) − e(Dt +Γ̂t ) E Q S(T )e−(DT +Γ̂T ) |Ft
X
+ 1τ >t (1 − α)(1 − ρ)P
S(ti )
i=0
N −1
X
(Dt +Γ̂t ) Q S(T ) −(Dt +Γ̂t )
(7.3) + 1τ >t (1 − α)(1 − ρ)P e E 1− e i i |Ft
S(ti )
i=dte
When an intensity process and an instantaneous risk free rate exist, we get
N −1
" ! #
Q − tT (ru +λ̂u )du
R X S(T )
Lt = 1{τ >t} E e max (1 − ρ)P , g |Ft
S(ti )
i=0
btc
1 n h RT io
S(t) − E Q S(T )e− t (ru +λ̂u )du |Ft
X
+ 1τ >t (1 − α)(1 − ρ)P
S(ti )
i=0
N −1
X
Q S(T ) − Rtti (ru +λ̂u )du
+ 1τ >t (1 − α)(1 − ρ)P E 1− e |Ft
S(ti )
i=dte
Notice if we take N = 1 and pose (1 − ρ)P = nS(t), we get the same result as for the single
premium contract. The last two expectations can often be found in closed form or in quasi-
closed form. Unfortunately, the first expectation is much harder to compute. No closed form
solution is available for this one, even in the simplest case. However, we can still rely on the
extensive literature on the numerical techniques used for the pricing of Asian options in finance.
7.2.2. Periodic Premiums contract of Type II. We assume the payment of a premium gives
the right to the policyholder, to n units whatever the value of these units at the time of payment.
We assume there are N dates of payment. In this case, we have

bτ c
X
R(τ ) = (1 − α) nS(τ )
i=0
= (1 − α)bτ cnS(τ )
N −1
!
X
g(T, ω) = max nS(T ), g
i=0
= max (N nS(T ), g)
The present value of the liabilities is given by
h i
Lt = E Q e−(DT −Dt ) max(N nS(T ), g)1τ >T |Gt
h i
+ E Q e−(Dτ −Dt ) (1 − α)bτ cnS(τ )1{t<τ ≤T } |Gt
We show in Appendix 3 that we can rewrite this equation as
h i
Lt = 1{τ >t} e(Dt +Γ̂t ) E Q e−(DT +Γ̂T ) max (N nS(T ), g) |Ft
btc
X
+ 1{τ >t} (1 − α)n S(ti )
i=0
 
−1
 NX h i h i
+ 1{τ >t} (1 − α)ne−(Dτ +Γ̂t ) E Q e−(Dτi +Γ̂ti ) S(ti ) |Ft − N E Q e−(DT +Γ̂T ) S(T ) |Ft
 
i=dte
When an intensity process and an instantaneous risk-free rate exist, we have

h RT i
Lt = 1{τ >t} E Q e− t (ru +λ̂u )du max (N nS(T ), g) |Ft
btc
X
+ 1{τ >t} (1 − α)n S(ti )
i=0
 
 NX−1 h R ti i h RT i
(7.4)+ 1{τ >t} (1 − α)n E Q e− t (ru +λ̂u )du S(t ) |F − N E Q e− t (ru +λ̂u )du S(T ) |F
i t t
 
i=dte
7.3. Gaussian Model. In the past subsections, we made very weak assumptions on the
processes followed by St , Dt and Γ̂t . The results derived above are very general and do not
depend on the specific assumptions that one chooses to model these variables. For example,
the value of the stock index could be driven by a jump-diffusion process, a Lévy process, etc. . .
In this section, we now consider a particular model: a Gaussian model. Firstly, we assume
τ admits an F-intensity process λ(t) and there exists an instantaneous risk free rate r(t). Sec-
ondly, we assume they, both, as well as the logarithm of financial asset prices, follow Gaussian
processes.
As far as the surrender decision is concerned, we assume the policyholder does not react
in a perfectly rational manner. Instead, we assume he follows a simple heuristic that consists
of comparing the return of the fund I, he has invested in, to the return of another fund R,
he takes as a reference, so that the higher the return of the reference fund with respect to the
return of the contract fund, the higher the instantaneous surrender rate.
7.3.1. The financial market. We assume the term structure is modelled by a multi-factors
Gaussian affine model. The class of affine models of the term structure has been studied in
details by Dai and Singleton [38] and extended by Duffee [42]. We follow the classification and
notation of Dai and Singleton. We assume the term structure is described by the maximally
affine model denoted by AM0 (3) which corresponds to a three correlated factors Gaussian term
structure model. Under P , the canonical representation of this model is given by
r(t) = δ0 + δ10 X(t)
where δ0 ∈ R, δ1 ∈ R3 . The factors X(t) are solutions of the following system of stochastic
differential equations.
dX(t) = −KX(t)dt + ΣX dW (t)
where W (t) is a 3-dimensional standard Brownian motion.
   
κ11 0 0 1 0 0
K =  κ21 κ22 0  , ΣX =  0 1 0 
κ31 κ32 κ33 0 0 1
In order to take into account the existence of our two funds, we rewrite this model in the
following manner. The instantaneously risk free rate is still given by
r(t) = δ0 + δ10 X(t)
with
dX(t) = −KX(t)dt + ΣdW (t)
but W (t) is this time a 5-dimensional standard Brownian motion and
 
1 0 0 0 0
Σ= 0 1 0 0 0 
0 0 1 0 0
The matrix K remains unchanged
As we said, we have to model two different funds. The first one, SI , is the one the poli-
cyholder has invested in. The second one SR , is the reference fund. We assume the values of
these funds evolve as geometric Brownian motions. Under the real measure P , the values of
these funds are then the solutions of the following differential equations:

1 0
dln(SI (t)) = µI (t) − σI σI dt + σI0 dW (t)
2

1 0 0
dln(SR (t)) = µR (t) − σR σR dt + σR dW (t)
2
where σR and σI ∈ R5 . We do not specify any further the drift coefficients µi and µR since they
will be removed under the local martingale measure and do not play any role in the pricing
formula.
7.3.2. The time of surrender. Here, we use the second of the two approaches that we de-
scribed to define a surrender time i.e. we do not explicitly construct our surrender time but
instead directly specify its intensity process. The simple specification that we choose, is not
based on any empirical work and we do not want to suggest this model corresponds to any
actual surrender behavior. It only aims at illustrating our methodology.
Let h be a positive real function h(·, ·) : R+ × R → R+ , increasing and differentiable in its
second argument. We define the intensity process by
 
SR (t)
SR (0) 
λt = h t, SI (t)
SI (0)
“
S (t)
” “
S (t)
”
ln S R(0) −ln S I (0)
= h t, e R I
The intensity process depends on the time and on the ratio of the funds returns. The higher the
return of the reference fund compare to the return of the fund I, the higher the instantaneous
surrender rate. We approximate this function by the first term of its Taylor expansion in the
second argument and at the point x2 = 1.

SR (t) SI (t)
λt ≈ h(t, 1) + ln − ln hx2 (t, 1)
SR (0) SI (0)
where hx2 (·, ·) is the derivative function of h(·, ·) in its second argument. We obtain a new
intensity process which is linear with respect to difference of the logarithmic returns of the
funds. Since these logarithmic returns follow Gaussian processes, this approximate intensity
process follows also a Gaussian process. Strictly speaking, this process is not a positive process
as it should be. The same kind of problem occurs in Gaussian term structure model where the
instantaneous risk free rate can take negative value whereas it theoretically should be positive.
Gaussian term structure model are although widely used by practitioners and academics as well.
Their use is justified by the argument that, for most realistic specification, the probability of
having a path with negative values should be very low. We make the same implicit assumption
for λt .
We can, for example, assume h(·, ·) is a linear function in its second argument, that is
h(t, x) = α(t) + β(t)x
with α(t) > 0 ∀ t and β(t) > 0 ∀ t. β(t) is then the sensibility of the instantaneous surrender
rate with respect to the difference of the returns of the funds. α(t) represent the instantaneous
surrender rate when a policyholder is not sensitive to the evolution of the financial market.
The time dependence of α(t) and β(t) can express for example, a tax effect on the surrender
probability.
The approximate intensity process is then given by

SR (t) SI (t)
λt ≈ α(t) + β(t) 1 + ln − ln
SR (0) SI (0)
7.3.3. The equivalent martingale measure Q. Lemma 6.3 allows us to separate the specifi-
cation of the price of market risk and the price of the surrender risk. We start with the price
of market risk. Duffee introduced a price of risk Λ(t) with the following form
Λ(t) = γ1 + γ2 X(t)
where γ1 ∈ R5 and γ2 ∈ R5×3 . If we define by Q, the P -equivalent measure that has a

Radon-Nikodym derivatives with respect to P , given by:
Z ·
dQ
= E − Λ(u)du
dP |F· 0
where ε(·) is the Doleans exponential,then

Z t
Q
W (t) = W (t) + Λ(u)du
0
is a Brownian motion under the measure Q. Under this equivalent measure, we have this time
h i
dX(t) = K̃ Θ̃ − X(t) dt + ΣdW Q (t)
where
K̃ = K + Σγ2
Θ̃ = (K + Σγ2 )−1 (−Σγ1 )
In order to Q to be a martingale measure, γ1 and γ2 are to be chosen such that under Q, we

have:

1 0
d ln(SI (t)) = r(t) − σI σI dt + σI0 dW Q (t)
2

1 0 0
d ln(SR (t)) = r(t) − σR σR dt + σR dW Q (t)
2
As far as the price of surrender risk is concerned, we assume the pure surrender risk can be
diversified away in a large portfolio so that the price of surrender risk K is set to 0. We then
have:
λ̂t = λt
7.3.4. The valuation formula. In the following we illustrate the change of measure technique
we explained in Proposition 5.6, by deriving closed form solution in the model described above
for a periodic premiums contract of type II.
By applying the results of Proposition 5.6 to Equations (7.4) and (7.2), we get
T
h RT i S(T )
Lt = 1{τ >t} P (t, T ) E Q e− t λ(u)du |Ft E Q [max (N nSI (T ), g) |Ft ]
| {z } | {z }| {z }
(A) (D)
(B)
btc
X
+ 1{τ >t} (1 − α)n S(ti )
i=0
N −1 h R ti i
t S(t )
X
(7.5) + 1{τ >t} (1 − α)n P (t, ti )E Q i e− t λ(u)du |Ft E Q i [SI (ti ) |Ft ]
i=dte
T
h RT i S(T )
(7.6) − 1{τ >t} (1 − α)nN P (t, T )E Q e− t λ(u)du
|Ft E Q [SI (T ) |Ft ]
| {z }
(C)
and
N −1 h R ti i
t
X
(7.7) At = 1{τ >t} P P (t, ti )E Q i e− t λ(u)du |Ft
i=dte
We derive closed-form solutions for the expressions (A), (B), (C) and (D). These solutions can
then be applied mutatis mutandis for their associated expressions in ti .
Let us begin by solving (A). The price P (t, T ) at time t of a zero coupon bond with maturity
T is given in our 3 factors Gaussian affine model by:
0
P (t, T ) = e−A(t,T )−B(t,T ) X(t)
where
Z T
1
A(t, T ) = δ0 (T − t) + (T − t)δ10 I − B(t, T )0 Θ̃ − 0

B(u, T ) B(u, T )du
2 u=t

1
B(t, T )0 = δ10 N (I − e−Kd (T −t) )N −1
Kd
See Appendix 4 for a proof. In order to value (B), we need to know the distribution of λ under
QT . Here, the Radon-Nikodym derivative of the T -forward martingale measure QT defined in
Proposition 5.6, is given by
dQT 1
RT
B(u,T )0 ΣΣ0 B(u,T )du−
RT
B(u,T )0 ΣdW Q (t)
= e− 2 0 0
dQ |F·
See Appendix 5 for a proof. Let us define the process εt by:

T
Q dQ
εt = E |Ft
dQ
This last process is the solution of the following differential equation
dεt = −εt B(t, T )0 ΣdW Q (t)
T
Thanks to Girsanov’s theorem, we know the process W Q defined by
Z t
QT Q
W (t) = W (t) + Σ0 B(u, T )du, ∀t ∈ [0, T ]
0
is a standard Brownian motion under QT .

λ is given by
SR (u) SI (u)
λ(u) = α(u) + β(u) 1 + ln − ln
SR (0) SI (0)
Under QT , the values of the funds are then solutions of the following stochastic differential
equations:
Z u Z u
1 0 0 0 T
ln SI (u) = ln SI (0) + r(s) − σI σI − σI Σ B(s, T ) ds + σI0 dW Q (s)
0 2 0
Z u Z u
1 0 0 0 0 T
ln SR (u) = ln SR (0) + r(s) − σR σR − σR Σ B(s, T ) ds + σR dW Q (s)
0 2 0
We then have
Z u Z u
1 0 0 0 0 0 0 0 QT
λ(u) = α(u) + β(u) 1 + (σR σR − σI σI ) + (σR − σI )Σ B(s, T )ds + (σI − σR )dW (s)
t 2 t
RT
In Appendix 6, we show the distribution of t λ(u)du under QT is given by
Z T Z T
QT T
λ(u)du = µΓ (t, T ) + σ Γ (u, T )0 dW Q (u)]
t t
where
Z T Z T Z T
T 1 0
µQ
Γ (t, T ) = α(u)du + β(u)du + (σR σR − σI0 σI ) β(u)(u − t)du
t t 2 t
Z T Z T
0 0 0
+ (σR − σI )Σ B(u, T ) β(s)ds du
t u
Z T
σ Γ (t, T )0 = (σI0 − σR
0
) β(s)ds
t
Accordingly, we have for (B)
QT
T
h RT i 1 T Γ
R Γ 0
E Q e− t λ(u)du |Ft = e−µΓ (t,T )+ 2 t σ (u,T )σ (u,T ) du
We can now find (C). We have the following useful equality.

S(T ) S(T ) SI (T )
EQ [SI (T ) |Ft ] = E Q |Ft
P (T, T )
S(T )
= EQ [FSI (T, T ) |Ft ]
where FSI (t, T ) is the forward price at time t of the fund SI for a maturity T . Let us find the
distribution of this forward price under QS(T ) . The Radon-Nikodym derivative of the T -forward
surrender measure QS(T ) defined in Proposition 5.6, is given, in our model, by
0
dQS(T ) 1
RT RT
[(σI0 −σR0 )( β(s)ds)][(σI0 −σR
0 )
R R R T
( uT β(s)ds)] du− 0T (σI0 −σR0 )( uT β(s)ds)dW Q (u)
= e− 2 0 u
dQT
See the Appendix 7 for a proof. Let us define the process εSt by:
" #
dQ S(T )
T
εSt = E Q |Ft
dQT
This last process is the solution of the following differential equation
Z T
S S 0 0 T
dεt = −εt (σI − σR ) β(s)ds dW Q (t)
t
S(T )
Z T Z T
S(T ) T
WQ (t) = W Q (t) + β(s)ds [σI0 − σR
0 0
] du, ∀t ∈ [0, T ]
0 u
is a standard Brownian motion under QS(T ) .

Under QT , the forward price follows a martingale and is the solution of the following differential
equations:
T
d(FSI (t, T )) = FSI (t, T ) σI0 + B(t, T )0 Σ dW Q (t)

See Appendix 8 for a proof. Under QS(T ) , this forward price is then the solution of
Z T
0 0 0 0 QT
d(FSI (t, T )) = FSI (t, T ) − β(s)ds(σI + B(t, T ) Σ)(σI − σR )dt + (σI + B(t, T ) Σ)dW (t)
t
The solution of this equation is given by
Z TZ T
ln FSI (T, T ) = ln FSI (t, T ) − β(s)ds(σI0 + B(u, T )0 Σ)(σI − σR )du
t u
Z T Z T
1 T
− (σI0 + B(u, T ) 0
Σ)(σI0 + B(u, T ) Σ) du +0 0
(σI0 + B(u, T )0 Σ)dW Q (u)
2 t t
or
Z T Z T
T 1 T
ln FSI (T, T ) = ln FSI (t, T ) − µQ
FS (t, T ) − σFS (u, T )σFS (u, T ) du + 0
σFS (u, T )0 dW Q (u)
2 t t
where
T
Z T Z T
µQ
FS (t, T ) = β(s)ds(σI0 + B(u, T )0 Σ)(σI − σR )du
t u
σFS (u, T )0 = σI0 + B(u, T )0 Σ
so that
S(T ) S(T )
EQ [SI (T ) |Ft ] = E Q [FSI (T, T ) |Ft ]
SI (t) − R T R T β(s)ds(σ0 +B(u,T )0 Σ)(σI −σR )du
= e t u I
P (t, T )
Finally, we still have to find the value of
S(T ) S(T )
h g i
EQ [max(N nSI (T ), g) |Ft ] = N nE Q ) |Ft
max(SI (T ),
Nn
QS(T )
h g i
= N nE max(FSI (T, T ), ) |Ft
h Nn i
S(T )
= N n EQ FSI (T, T )1{FS (T,T )> g } |Ft
I Nn
| {z }
I
S(T )
h i
+ g EQ 1{FS (T,T )≤ g } |Ft
I Nn
| {z }
II
Let us start by (II). We have:

S(T )
h i g
EQ 1{FS (T,T )≤ g
} |Ft = QS(T ) FSI (T, T ) ≤ |Ft
I Nn Nn
S(T )
g
= Q lnFSI (T, T ) ≤ ln |Ft
Nn
= N (−d2 )
Where N is the cumulative distribution function of a standard normal random variable, and
T
1 T
−ln Ngn + ln FSI (t, T ) − µQ 0
R
FS (t, T ) − 2 t σFS (u, T ) σFS (u, T )du
d2 = hR i1
T 0 σ (u, T )du 2
u=t σ F S
(u, T ) F S
For (I), we can again change our measure. Let the new measure QS(T )∗ be defined by the
following Radon Nikodym derivative with respect to QS(T ) :
dQS(T )∗ FSI (T, T )
= S(T )
dQS(T ) EQ [FSI (T, T ) |Ft ]
We have the following equality
S(T )
h i S(T ) S(T )∗
h i
EQ FSI (T, T )1{FS (T,T )> g } |Ft = EQ [FSI (T, T ) |Ft ] E Q 1{FS (T,T )> Ngn } |Ft
I Nn I
∗
Let us define the process εSt by:
" #
∗ QS(T ) dQS(T )∗
εtS = E |Ft
dQS(T )
This process is the solution of the following differential equation
∗ ∗ S(T )
dεSt = εSt σI0 + B(t, T )0 Σ dW Q

(t)
S(T )∗
Z t
S(T )∗ S(T ) 0
WQ (t) = W Q (t) − σI0 + B(u, T )0 Σ du, ∀t ∈ [0, T ]
0
is a standard Brownian motion under QS(T )∗ . Under this measure the forward price FSI (t, T )
is given by
1 T
Z Z T
QT 0 T
ln FSI (T, T ) = ln FSI (t, T ) − µFS (t, T ) + σFS (u, T ) σFS (u, T )du + σFS (u, T )0 dW Q (u)
2 t t
We can now find
S(T )∗
h i g
EQ 1{FS (T,T )> Ngn } |Ft = QS(T )∗ FSI (T, T ) > |Ft
I Nn
= P rob(Z > −d1 )
= P rob(Z ≤ d1 )
= N (d1 )
with
T
1 T
R
FS (t, T ) + 2 t σFS (u, T ) σFS (u, T )du
d1 = hR i1
T 0 σ (u, T )du 2
t σ F S
(u, T ) F S
We then find
S(T ) S(T )
EQ [max (N nSI (T ), g) |Ft ] = N nE Q [FSI (T, T ) |Ft ] N (d1 ) + gN (−d2 )
with
S(T ) SI (t) − R T (R T β(s)ds)(σI0 +B(u,T )0 Σ)(σI −σR )du
EQ [FSI (T, T ) |Ft ] = e t u
P (t, T )
T
1 T
R
FS (t, T ) + 2 t σFS (u, T ) σFS (u, T )du
d1 = hR i1
T 0 σ (u, T )du 2
t σ FS (u, T ) FS
T
1 T
R
FS (t, T ) − 2 t σFS (u, T ) σFS (u, T )du
d2 = hR i1
T 0 σ (u, T )du 2
t σ FS (u, T ) FS
8. Conclusion.
In this chapter, we presented an alternative model of the surrender time in life insurance.
Technicalities aside, the main difference with the traditional model of the surrender time, is
the way we model the insurer’s and policyholder’s information. In the traditional model, the
insurer and the policyholder have exactly the same set of information F and all their decisions,
in particular the surrender decision, are based on this set. Accordingly, the surrender time has
to be an F-stopping time. We argue this is not realistic. On the contrary, we assume that the
policyholder takes his surrender decision on a larger set of information. In this situation, the
surrender time can be characterized by its so-called F-hazard process.
We also studied the impact of this framework on the fair valuation of life insurance contract.
In particular, we gave fair valuation formulas for single premium and periodic premiums unit
linked contracts with stochastic term structure and stochastic surrender. We also illustrate
these formulas in a Gaussian framework.
Appendix 1.
Proposition. Let QT be the measure equivalent to Q on (Ω, FT ) whose Radon-Nikodym
derivative is given by
dQT e−DT
=
dQ |FT E Q [e−DT |F0 ]
Q-a.s. Let QS(T ) be the measure equivalent to QT on (Ω, FT ) whose Radon-Nikodym derivative
is given by
dQS(T ) e−Γ̂T
=
dQT |FT
h i
E QT e−Γ̂T |F0
QT -a.s.
If X is QS(T ) -integrable, then we have
h i T
h i
1τ >t E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft = 1τ >t P (t, T )E Q e−(Γ̂T −Γ̂t ) X |Ft
T
h i S(T )
= 1τ >t P (t, T )E Q e−(Γ̂T −Γ̂t ) |Ft E Q [X |Ft ]
S(T )
= P (t, T )QT (τ > T |Gt ) E Q [X |Ft ]
78 Appendix 1.
Proof. We have

e−DT
h i EQ E Q [e−DT |F0 ]
e−(Γ̂T −Γ̂t ) X |Ft
T
EQ e−(Γ̂T −Γ̂t ) X |Ft =
e−DT
EQ E Q [e−DT |F0 ]
|Ft
h i
e−Dt
e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft
EQ
EQ [e−DT |F0 ]
= e−Dt

E Q e−(DT −Dt ) |F
E [e
Q −D T |F0 ] t
h i
E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft
=
E Q e−(DT −Dt ) |Ft
So that we have the first equality of our proposition.
h i h i T
h i
1τ >t E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft = 1τ >t E Q e−(DT −Dt ) |Ft E Q e−(Γ̂T −Γ̂t ) X |Ft
T
h i
(8.1) = 1τ >t P (t, T )E Q e−(Γ̂T −Γ̂t ) X |Ft
Similarly, we have
" #
QT e−Γ̂T
E h iX |Ft
E Q e−Γ̂T |F0
QS(T )
E [X |Ft ] = " #
e−Γ̂T
E QT h i |Ft
E QT e−Γ̂T |F0
−Γ̂ T
h i
QT
he t i EQ e−(Γ̂T −Γ̂t ) X |Ft
E e−Γ̂T |F0
= −Γ̂
h i
he t i E QT e−(Γ̂T −Γ̂t ) |Ft
E QT e−Γ̂T |F0
T
h i
EQ e−(Γ̂T −Γ̂t ) X |Ft
= h i
E QT e−(Γ̂T −Γ̂t ) |Ft
So that
T
h i T
h i S(T )
EQ e−(Γ̂T −Γ̂t ) X |Ft = EQ e−(Γ̂T −Γ̂t ) |Ft E Q [X |Ft ]
Eventually, using this last expression in (8.1), we have the second equality of our proposition.
h i T
h i S(T )
1τ >t E Q e−(DT −Dt ) e−(Γ̂T −Γ̂t ) X |Ft = 1τ >t P (t, T )E Q e−(Γ̂T −Γ̂t ) |Ft E Q [X |Ft ]

Appendix 2.
Proposition. The value of the liabilities of the periodic premium unit-linked contract of
type I is given by
N −1
" ! #
(Dt +Γ̂t ) Q −(DT +Γ̂T )
X S(T )
Lt = 1{τ >t} e E e max (1 − ρ)P , g |Ft
S(ti )
i=0
btc
S(t) X
(Dt +Γ̂t ) Q S(T ) −(DT +Γ̂T )
+ 1τ >t (1 − α)(1 − ρ)P −e E e |Ft
S(ti ) S(ti )
i=0
N −1
X
(Dt +Γ̂t ) Q S(T ) −(Dt +Γ̂t )
+ 1τ >t (1 − α)(1 − ρ)P e E 1− e i i |Ft
S(ti )
i=dte
Proof. The value of these liabilities are given by
N −1
" ! #
X
Lt = E Q e−(DT −Dt ) max S(T ) nti , g 1τ >T |Gt
i=0
   
bτ c
X
+ E Q e−(Dτ −Dt ) (1 − α)S(τ )  nti  1{t<τ ≤T } |Gt 
i=0
where bτ c = sup {i |ti ≤ τ }. We can rewrite this equation as
N −1
" ! #
X
Q −(DT −Dt )
E e max S(T ) nti , g 1τ >T |Gt
Lt = i=0
| {z }
I
   
bτ c
X
E Q e−(Dτ −Dt ) (1 − α)S(τ )  nti  1τ >t |Gt 
+ i=0
| {z }
II
   
bτ c
X
E Q e−(Dτ −Dt ) (1 − α)S(τ )  nti  1τ >T |Gt 
− i=0
| {z }
III
We can study each of these terms separately.

80 Appendix 2.
N −1
" ! #
X
I = E Q e−(DT −Dt ) max S(T ) nti , g 1τ >T |Gt
i=0
N −1
" ! #
X
Q −(DT −Dt ) −(Γ̂T −Γ̂t )
= 1τ >t E e e max S(T ) nti , g 1τ >T |Ft
i=0
N −1
" ! #
Q −(DT −Dt ) −(Γ̂T −Γ̂t )
X S(T )
= 1τ >t E e e max (1 − ρ) P , g 1τ >T |Ft
S(ti )
i=0
   
bτ c
X
II = E Q e−(Dτ −Dt ) (1 − α)S(τ )  nti  1τ >t |Gt 
i=0
   
bτ c
X (1 − ρ)P
= 1τ >t (1 − α)E Q e−(Dτ −Dt ) S(τ )   |Gt 
S(ti )
i=0
N −1
" ! #
Q −(Dτ −Dt )
X 1
= 1τ >t (1 − α)(1 − ρ)P E e S(τ ) 1 |Gt
S(ti ) {τ >ti }
i=0
btc
1 X
Q −(Dτ −Dt )
h i
= 1τ >t (1 − α)(1 − ρ)P E e S(τ )1{τ >ti } |Gt
S(ti )
i=0
N −1
−(Dτ −Dt ) S(τ )
X
Q
+ 1τ >t (1 − α)(1 − ρ)P E e 1 |Gt
S(ti ) {τ >ti }
i=dte
btc
X S(t)
= 1τ >t (1 − α)(1 − ρ)P
S(ti )
i=0
N −1
−(Dτ −Dt ) S(τ )
X
Q Q
+ 1τ >t (1 − α)(1 − ρ)P E E e 1 |Gt |Gt
S(ti ) {τ >ti } i
i=dte
btc
X S(t)
= 1τ >t (1 − α)(1 − ρ)P
S(ti )
i=0
N −1 i
−(Dti −Dt ) 1
X h
Q Q −(Dτ −Dti )
+ 1τ >t (1 − α)(1 − ρ)P E {τ > ti }e E e S(τ ) |Gti |Gt
S(ti )
i=dte
Finally, we have for (II):

btc
X S(t)
II = 1τ >t (1 − α)(1 − ρ)P
S(ti )
i=0
N −1
S(ti )
E Q {τ > ti }e−(Dti −Dt )
X
+ 1τ >t (1 − α)(1 − ρ)P |Gt
S(ti )
i=dte
 
btc N −1
X S(t) h i
E Q e−(Γ̂ti −Γ̂t ) e−(Dti −Dt ) |Ft
X
= 1τ >t (1 − α)(1 − ρ)P +
 S(ti ) 
i=0 i=dte
   
bτ c
X
III = E Q e−(Dτ −Dt ) (1 − α)S(τ )  nti  1τ >T |Gt 
i=0
N −1
" ! #
Q −(Dτ −Dt )
X 1
= 1τ >t (1 − α)(1 − ρ)P E e S(τ ) 1τ >T |Gt
S(ti )
i=0
N −1
" ! #
Q
X 1 −(DT −Dt ) Q
h
−(Dτ −DT )
i
= 1τ >t (1 − α)(1 − ρ)P E e E e S(τ ) |GT 1τ >T |Gt
S(ti )
i=0
N −1
X
Q S(T ) −(DT −Dt ) −(Γ̂T −Γ̂t )
= 1τ >t (1 − α)(1 − ρ)P E e e |Ft
S(ti )
i=0
In the above derivation, we used the fact the discounted price of S is a (Q, G)-martingale.
Appendix 3.
Proposition. The present value of the liabilities of unit-linked contract of type II is given
by
h i
Lt = 1{τ >t} e(Dt +Γ̂t ) E Q e−(DT +Γ̂T ) max (N nS(T ), g) |Ft
btc
X
+ 1{τ >t} (1 − α)n S(ti )
i=0
 
−1
 NX h i h i
+ 1{τ >t} (1 − α)ne−(Dτ +Γ̂t ) E Q e−(Dτi +Γ̂ti ) S(ti ) |Ft − N E Q e−(DT +Γ̂T ) S(T ) |Ft
 
i=dte
Proof. The present value of the liabilities is given by

h i
Lt = E Q e−(DT −Dt ) max(N nS(T ), g)1τ >T |Gt
h i
+ E Q e−(Dτ −Dt ) (1 − α)bτ cnS(τ )1{t<τ ≤T } |Gt
82 Appendix 3.
We can rewrite this equation as
h i
E Q e−(DT −Dt ) max(N nS(T ), g)1τ >T |Gt
Lt = | {z }
I
h i
E Q e−(Dτ −Dt ) (1 − α)bτ cnS(τ )1{τ >t} |Gt
+ | {z }
II
h i
E Q e−(Dτ −Dt ) (1 − α)bτ cnS(τ )1{τ >T } |Gt
− | {z }
III
We can study each of these terms separately.
h i
I = E Q e−(DT −Dt ) max(N nS(T ), g)1τ >T |Gt
h i
= 1{τ >t} e(Dt +Γ̂t ) E Q e−(DT +Γ̂) max(N nS(T ), g) |Ft
N −1
" #
X
II = 1{τ >t} (1 − α)nE Q e−(Dτ −Dt ) 1{τ >ti } S(τ ) |Gt
i=0
btc h i
X
= 1{τ >t} (1 − α)n E Q e−(Dτ −Dt ) 1{τ >ti } S(τ ) |Gt
i=0
N
X −1 h i
+ 1{τ >t} (1 − α)n E Q e−(Dτ −Dt ) 1{τ >ti } S(τ ) |Gt
i=dte
btc h i
X
= 1{τ >t} (1 − α)n 1{τ >ti } E Q e−(Dτ −Dt ) S(τ ) |Gt
i=0
N
X −1 h h i i
+ 1{τ >t} (1 − α)n E Q 1{τ >ti } e−(Dτi −Dt ) E Q e−(Dτ −Dti ) S(τ ) |Gti |Gt
i=dte
 
Xbtc N −1 h i
X
= 1{τ >t} (1 − α)n S(ti ) + e−(Dτ +Γ̂t ) E Q e−(Dτi +Γ̂ti ) S(ti ) |Ft
 
i=0 i=dte
N −1
" #
X
III = (1 − α)nE Q e−(Dτ −Dt ) 1{τ >ti } S(τ )1{τ >T } |Gt
i=0
h i
= (1 − α)nN E Q e−(Dτ −Dt ) S(τ )1{τ >T } |Gt
h h i i
= (1 − α)nN E Q e−(DT −Dt ) E Q e−(Dτ −DT ) S(τ ) |GT 1{τ >T } |Gt
h i
= (1 − α)nN E Q e−(DT −Dt ) S(T )1{τ >T } |Gt
h i
= 1{τ >t} (1 − α)nN e−(Dt +Γ̂t ) E Q e−(DT +Γ̂T ) S(T ) |Ft
In the above derivation, we used the fact the discounted price of S is a (Q, G)-martingale.
Appendix 4.
Proposition. The price at time t of a zero coupon bond with maturity T is given in our
3 factors Gaussian affine model by:
0
P (t, T ) = e−A(t,T )−B(t,T ) X(t)
where
Z T
1
A(t, T ) = δ0 (T − t) + (T − t)δ10 I − B(t, T )0 Θ̃ − 0

B(u, T ) B(u, T )du
2 t

1
B(t, T )0 = δ10 N (I − e−Kd (T −t) )N −1
Kd
Proof. Assume that K̃ can be decomposed as K̃ = N K̃d N −1 where K̃d is a (3 × 3)

diagonal matrix. Let X ∗ (t) = N −1 X(t). We have
dX ∗ (t) = N −1 dX(t)
= N −1 [K̃(Θ̃ − X(t)) + ΣdW Q (t)]
= K̃d N −1 (Θ̃ − X(t))dt + N −1 ΣdW Q (t)
= K̃d (N −1 Θ̃ − X ∗ (t))dt + N −1 ΣdW Q (t)
By Itô’s lemma, we can find the solution of this stochastic differential equation. We have for
each i.
d(e(K̃d )i u Xi∗ (u)) = (K̃d )i e(K̃d )i u Xi∗ (u)du + e(K̃d )i u dXi∗ (u)
= (K̃d )i e(K̃d )i u Xi∗ (u)du
= e(K̃d )i u (K̃d )i (N −1 Θ̃)i du + e(K̃d )i u (N −1 Σ)i,· dW Q (u)
84 Appendix 4.
If we integrate this expression, we have

Z T
e (K̃d )i T
Xi∗ (T ) −e (K̃d )ii t
Xi∗ (t) = e(K̃d )i u (K̃d )i (N −1 Θ̃)i du
t
Z T
+ e(K̃d )i u (N −1 Σ)i,· dW Q (u)
t
e(K̃d )i T Xi∗ (T ) = e(K̃d )i t Xi∗ (t) + [e(K̃d )i T − e(K̃d )i t ](N −1 Θ̃)i

Z T
+ e(K̃d )i u (N −1 Σ)i,· dW Q (u)
t
Finally,
Xi∗ (T ) = e−(K̃d )i (T −t) Xi∗ (t) + [1 − e−(K̃d )i (T −t) ](N −1 Θ̃)i

Z T
+ e−(K̃d )i (T −u) (N −1 Σ)i,· dW Q (u)
t
In matrix form,
X ∗ (T ) = e−K̃d (T −t) X ∗ (t) + [I − e−K̃d (T −t) ]N −1 Θ̃

Z T
+ e−K̃d (T −u) N −1 ΣdW Q (u)
t
To find X(T ), we can multiply this last expression by N.
X(T ) = N e−K̃d (T −t) X ∗ (t) + N [I − e−K̃d (T −t) ]N −1 Θ̃

Z T
+ N e−K̃d (T −u) N −1 ΣdW Q (u)
t
= [N e−K̃d (T −t) N −1 ]X(t) + [I − [N e−K̃d (T −t) N −1 ]]Θ̃

Z T
+ [N e−K̃d (T −u) N −1 ]ΣdW Q (u)
t
RT
To find the distribution of t r(u)du, we need to know
Z T Z T
r(u)du = δ0 + δ10 X(u)du
t t
Z T
= δ0 + δ10 N X ∗ (u)du
t
Z T
= δ0 (T − t) + δ10 N X ∗ (u)du
t
For each i, we have
Z T
1
Xi∗ (u)du = (1 − e−(K̃d )i (T −t) )Xi∗ (t)
t (K̃d )i
1
+ [(T − t) − (1 − e−(K̃d )i (T −t) )](N −1 Θ̃)i
(K̃d )i
Z T
1
+ (1 − e−(K̃d )i (T −u) )(N −1 Σ)i,· dW Q (u)
t (K̃d )i
In matrix form
Z T
1
X ∗ (u)du = (1 − e−(K̃d )(T −t) )X ∗ (t)
t (K̃d )
1
+ [(T − t)I − (I − e−(K̃d )(T −t) )](N −1 Θ̃)
(K̃d )
Z T
1
+ (I − e−(K̃d )(T −u) )(N −1 Σ)dW Q (u)
t (K̃d )
RT
To find t X(u)du, we can multiply the last expression by N
Z T
1
X(u)du = N (1 − e−(K̃d )(T −t) )X ∗ (t)
t (K̃d )
1
+ [(T − t)N I − N (I − e−(K̃d )(T −t) )](N −1 Θ̃)
(K̃d )
Z T
1
+ N (I − e−(K̃d )(T −u) )(N −1 Σ)dW Q (u)
t (K̃d )
1
= [N (1 − e−(K̃d )(T −t) )N 1 ]X(t)
(K̃d )
1
+ [(T − t)I − [N (I − e−(K̃d )(T −t) )N −1 ]]Θ̃
(K̃d )
Z T
1
+ [N (I − e−(K̃d )(T −u) )N −1 ]ΣdW Q (u)
t (K̃ d )
86 Appendix 5.
Finally, we obtain
Z T
1
r(u)du = δ0 (T − t) + δ10 [N (1 − e−(K̃d )(T −t) )N 1 ]X(t)
t (K̃d )
1
+ δ10 [(T − t)I − [N (I − e−(K̃d )(T −t) )N −1 ]]Θ̃
(K̃d )
Z T
1
+ δ10 [N (I − e−(K̃d )(T −u) )N −1 ]ΣdW Q (u)
t (K̃d )
= δ0 (T − t) + B(t, T )0 X(t) + [(T − t)δ10 I − B(t, T )0 ]Θ̃
Z T
+ B(u, T )0 ΣdW Q (u)
t
where
1
B(t, T )0 = δ10 [N (I − e−Kd (T −t) )N −1 ]
Kd
RT
Thus, u=t r(u)du is normally distributed with mean
Z T
E Q
r(u)du |Ft = δ0 (T − t) + B(t, T )0 X(t) + [(T − t)δ10 I − B(t, T )0 ]Θ̃
t
and variance
Z T Z T
V AR Q
r(u)du |Ft = B(u, T )0 ΣΣ0 B(u, T )du
t t
Z T
= B(u, T )0 B(u, T )du
t
RT
Finally since tr(u)du is normally distributed, we have
h RT i
P (t, T ) = E Q e− t r(u)du |Ft
RT
Q r(u)du|Ft ]+ 21 V ARQ [ tT r(u)du||Ft ]
= e−E [
R
t
0 0 0 1
RT
B(u,T )0 B(u,T )du
= e−δ0 (T −t)−B(t,T ) X(t)−[(T −t)δ1 I−B(t,T ) ]Θ̃+ 2 t
The result follows.
Appendix 5.
Proposition. In our model, the Radon-Nikodym derivative of the equivalent T -forward
martingale measure QT with respect to Q is given by
dQT 1
RT
B(u,T )0 ΣΣ0 B(u,T )du−
RT
B(u,T )0 ΣdW Q (t)
= e− 2 0 0
dQ
Proof. The T -forward measure is defined by the following Radon-Nikodym derivative

RT
dQT e− 0 r(u)du
= h RT i , Q-a.s.
dQ E Q e− 0 r(u)du |Ft
Using the results of Appendix 4, we have
0 0 0
RT
B(u,T )0 ΣdW Q (u)
dQT e−δ0 (T −0)−B(0,T ) X(0)−[(T −0)δ1 I−B(0,T ) ]Θ̃− 0
=
dQ e−A(0,T )−B(0,T )0 X(0)
0 0 0
RT
B(u,T )0 ΣdW Q (u)
e−δ0 (T −0)−B(0,T ) X(0)−[(T −0)δ1 I−B(0,T ) ]Θ̃− 0
= 0 0 1
RT
B(u,T )0 B(u,T )du]−B(0,T )0 X(0)
e−δ0 (T −0)−[(T −0)δ1 I−B(0,T ) ]Θ̃+ 2 [ 0
1
RT RT
B(u,T )0 ΣΣ0 B(u,T )du− B(u,T )0 ΣdW Q (t)
= e− 2 0 0
Appendix 6.
RT
Proposition. The distribution of t λ(u)du is given by
Z T Z T
T T
λ(u)du = µQ Γ (t, T ) + σ Γ (u, T )0 dW Q (u)
t t
where
Z T Z T Z T
T 1 0
µQ σR σR − σI0 σI

Γ (t, T ) = α(u)du + β(u)du + β(u)(u − t)du
t t 2 t
Z T Z T
0 0 0

+ σR − σI Σ B(u, T ) β(s)ds du
t u
Z T
σ Γ (t, T )0 = σI0 − σR
0

β(s)ds
t
Proof. We have
Z u Z u
1 0 0 0 T
λ(u) = α(u) + β(u)[1 + r(s) − σI σI − σI Σ B(s, T )ds + σI0 dW Q (s)
t 2 t
Z u Z u
1 0 0 T
− r(s) − σR σR − σR Σ0 B(s, T )ds − σR0
dW Q (s)]
t 2 t
Z u
1 0
= α(u) + β(u)[1 + (σR σR − σI0 σI ) + (σR0
− σI0 )Σ0 B(s, T )ds
t 2
Z u
0 0 QT
+ (σI − σR )dW (s)
t
Z u
1 0
λ(u) = α(u) + β(u)[1 + (σR σR − σI0 σI )(u − t) + 0
(σR − σI0 )Σ0 B(s, T )ds
2 t
Z u
T
+ (σI0 − σR
0
)dW Q (s)
t
88 Appendix 6.
If we integrate this expression, we have

Z T Z T Z T
1 T
Z
0
λ(u)du = α(u)du + [ β(u)du + β(u)(σR σR − σI0 σI )(u − t)du
t t t 2 t
Z T Z u Z T Z u
0 0 0 T
+ β(u) (σR − σI )Σ B(s, T )dsdu + β(u) (σI0 − σR
0
)dW Q (s)du]
t t t t
Using Fubini’s theorem, we have

Z T Z u Z T Z T
0 0 QT T
β(u) (σI − σR )dW (s)du = β(u)(σI0 − σR
0
)dudW Q (s)
t t t s
Z T Z T
T
= (σI0 − 0
σR ) β(u)dudW Q (s)
t s
and
Z T Z u Z T Z T
0
β(u) (σR − σI0 )Σ0 B(s, T )dsdu = 0
β(u)(σR − σI0 )Σ0 B(s, T )duds
t t t s
Z T Z T
0
= (σR − σI0 )Σ0 B(s, T ) β(u)duds
t s
So that we can write

Z T Z T Z T Z T
1 0 0
λ(u)du = α(u)du + [ β(u)du + (σR σR − σI σI ) β(u)(u − t)du
t t t 2 t
Z T Z T Z T Z T
0 T
+ (σR − σI0 )Σ0 B(u, T ) β(s)dsdu + (σI0 − σR
0
) β(s)dsdW Q (u)]
t u t u
To have more concise expressions, we introduce the following notation

Z T Z T
T T
λ(u)du = µQ Γ (t, T ) + σ Γ (u, T )0 dW Q (u)
t t
where
Z T Z T Z T
T 1 0
µQ σR σR − σI0 σI

Γ (t, T ) = α(u)du + β(u)du + β(u)(u − t)du
t t 2 t
Z T Z T
0 0
0
+ σR − σI Σ B(u, T ) β(s)ds du
t u
Z T
σ Γ (t, T )0 = σI0 − σR
0

β(s)ds
t

Appendix 7.
Proposition. In our model, the Radon-Nikodym derivative of the T -forward surrender
measure QS(T ) with respect to QT is given by
dQS(T ) 1
RT
[(σI0 −σR
0 )
RT
β(s)ds][(σI0 −σR
0 )
RT
β(s)ds]0 du−
RT
(σI0 −σR
0 )
RT T
β(s)dsdW Q (u)
= e− 2 0 u u 0 u
dQT
Proof. The T -forward surrender measure is defined by the following Radon-Nikodym
derivative
dQS(T ) e−ΓT
= , QT -a.s.
dQT E QT [e−ΓT |Ft ]
RT
e− 0 λ(u)du
= h RT i , QT -a.s.
− 0 λ(u)du
E QT e |Ft
Using the results of the Appendix 6, we have
Γ (0,T )− T T
σ Γ (u,T )0 dW Q (u)
R
dQS(T ) e−µ 0
= RT
dQT e−µ
Γ (0,T )+ 1
2
σ Γ (u,T )σ Γ (u,T )0 du
0
1 T Γ QT
R Γ 0
RT Γ 0
= e− 2 0 σ (u,T )σ (u,T ) du− 0 σ (u,T ) dW (u)
0 ) T β(s)ds][(σ 0 −σ 0 ) T β(s)ds]0 du− T (σ 0 −σ 0 ) T β(s)dsdW QT (u)
− 21 0T [(σI0 −σR
R R R R R
= e u I R u 0 I R u
Appendix 8.
Proposition. Under QT , the forward price FSI (t, T ) follows a martingale and is solution
of
T
dFSI (t, T ) = FSI (t, T )(σI0 + B(t, T )0 Σ)dW Q (t)
Proof. Under QT , we saw the price of the fund SI is given by:
Z t Z t
1 0 0 0 T
ln SI (t) = ln SI (0) + r(u) − σI σI − σI Σ B(u, T )du + σI0 dW Q (u)
0 2 0
As far as the price of the zero coupon bond is concerned, we have under Q
dP (t, T ) = P (t, T )[r(t) − B 0 (t, T )ΣdW Q (t)]
The solution of this differential equation is given by
Z t Z t
1 0 0 0
ln P (t, T ) = ln P (0, T ) + r(u) − B (u, T )ΣΣ B (u, T )du − B 0 (u, T )ΣdW Q (u)
0 2 0
Under QT , the price of a zero coupon is then given by
Z t Z t
1 0 0 0
ln P (t, T ) = ln P (0, T ) + r(u) + B (u, T )ΣΣ B (u, T )du − B 0 (u, T )ΣdW Q (u)
0 2 0
90 Appendix 8.
Taking the difference between ln SI (t) and ln P (t, T ),

1 t 0
Z
SI (t) SI (0)
ln = ln − B (u, T )ΣΣ0 B 0 (u, T ) + σI0 σI + 2σI0 Σ0 B(u, T )du
P (t, T ) P (0, T ) 2 0
Z t
T
+ σI0 + B 0 (u, T )ΣdW Q (u)
0
Then we get
1
Rt 0 0
Rt T
2 σI0 +B 0 (u,T )ΣdW Q (u)
FSI (t, T ) = FSI (0, T )e− 2 0 (σI +B (u,T )Σ) du+ 0
Which is the solution of the differential equation

T
d(FSI (t, T )) = FSI (t, T )(σI0 + B(t, T )0 Σ)dW Q (t)

CHAPTER 3
Risk-Minimizing Strategies of Life Insurance Contracts with

Surrender Option.
1. Introduction.
This chapter aims to study hedging strategies of insurance contracts with surrender option.
The academic literature on the valuation of surrender options generally considers the surrender
time as an optimal stopping time with respect to the filtration generated by the financial assets
prices. See [14, 15, 47, 81, 80] for examples. A life insurance contract with surrender option
can then be seen as an American option. Even though there are some good arguments in
favor of these models for a valuation purpose, they are not appropriate for a risk management
purpose. Indeed, in these models, the surrender option does not induce any additional source of
risk to the purely financial one. In particular, if the financial market is complete, the insurance
market remains complete and accordingly, any insurance contract with a surrender option can
be perfectly hedged (if we neglect the mortality risk). This is not a realistic assumption in a
risk management perspective. In order to overcome this drawback, we assume the surrender
times are not stopping times with respect to the filtration generated by the financial assets
prices. We refer the reader to the previous chapter for more details.
In this framework, even if the financial market is initially complete, an insurance contract
with a surrender option cannot be perfectly hedged. It is then worthwhile to study the hedging
strategies an insurer should follow to minimize its exposure to this surrender risk. Different
approaches have been developed to tackle the problem of contingent claims hedging in an
incomplete market. One of the most popular approaches are those based on quadratic hedging
strategies as the risk-minimizing, local risk-minimizing and mean-variance hedging strategies.
See Schweizer [78] for a review. In this chapter, we focus on the risk-minimizing hedging
strategies developed by Föllmer and Sondermann in [44]. This approach has been introduced
with success in the insurance literature by Møller [68, 69] and Dahl and Møller [39]. In
particular, Møller in [68] extended the risk-minimization theory to payment processes and
applied this theory to the hedging of a portfolio of unit-linked insurance contracts subject to
non hedgeable mortality risk. In his papers, Møller assumed a complete financial market of
the Black and Scholes type and assumed the policyholders’s dates of death are random times
independent of each other and independent of the financial market. In this setting, Møller
derived the risk-minimizing hedging strategies for portfolios of n policies and showed that, by
following these strategies, the remaining (unhedgeable) risk goes down to zero in relative terms
with the size n of the portfolio. In other words, he showed we could combine diversification
and risk-minimizing hedging to reduce as much as we want the relative risk of such portfolios.
91
92 The Theoretical Framework.
Riesner [72] extended this result by introducing an incomplete financial market driven by a
Lévy process. He showed the relative risk can only be reduced down to a limit related to the
incompleteness of the financial market. In this chapter, we generalize these results in two ways.
Firstly, we do not assume the independence of the random times of payment with the financial
market. It is an imperative assumption if we want to model realistic surrender times. Indeed it
is a well known empirical fact, the surrenders depends on the evolution of the financial market.
This dependence obviously implies the random times are neither independent of each other.
Secondly, we consider a general incomplete financial market (not necessarily driven by a Lévy
process). We only assume the discounted prices of the financial assets are local martingales. In
this setting, we firstly derive the general form of the risk-minimizing hedging strategies for a
single policy. Under an additional assumption, namely the F-independence of the random times,
we extend this result to a portfolio of n policies. Under this assumption, we then generalize
Møller’s conclusion: if the financial market is complete, we show we can reduce as much as we
want the relative risk of a portfolio by combining diversification and continuous hedging. If,
as in Riesner, the financial market is incomplete, by definition, we can only reduce the relative
risk down to a limit related to the degree of incompleteness of the financial market.
In a recent paper, Dahl and Møller [39] generalized these results in a still different way by
introducing a systematic mortality risk. The random times of payment are thus dependent to
each other but independent of the financial market. Our problem is in a sense similar to theirs
but not identical. In both cases, the times of payment are subject to a systematic risk but
in Dahl and Møller, this systematic risk cannot be hedged whereas in our, it can be hedged
perfectly or at least in part whether the financial market is complete or not.
This chapter is organized as follows. In Section 2, we describe the financial market and
the insurance payment processes. In Section 3, we briefly recall the main theoretical results of
Föllmer and Sondermann and Møller on the risk-minimization theory. In Section 4, we apply
these general results to a single insurance policy subject to surrender risk and derive the forms
of the risk-minimizing hedging strategies. In Section 5, we extend the results of the previous
section to a portfolio with n policies. In Section 6, we study the relative risk process and shows
how it evolves with the number n of policies.
2. The Theoretical Framework.

2.1. The financial market. Let (Ω, F, P ) be a complete probability space. A perfect
frictionless financial market is defined in this space. We assume there is one locally risk free
asset denoted by Bt = B(t, ω) and s risky assets Sti = S i (t, ω), i = 1, . . . s following real
càdlàg stochastic processes. The price of the locally risk free asset is assumed to follow a
strictly positive, continuous process of finite variation. The discounted values of these assets
are denoted by
Sti
Xti = , i = 1, . . . s
Bt
Let F = (Ft )0≤t≤T be the filtration generated by the stochastic processes S i for i = 1, . . . s and
Bt . We assume this filtration respects the usual hypothesis.
Risk-Minimizing Strategies of Life Insurance Contracts with Surrender Option. 93
We assume furthermore there is no arbitrage opportunity in this financial market with

regard to the filtration F. This assumption is equivalent to the existence of at least one
probability measure Q, equivalent to P , such that the discounted prices X are (F, Q)-local
martingales i.e. for each i = 1, . . . s, X i ∈ Mloc (Q, F), the space of local martingales with
respect to the filtration F and the measure Q.
2.2. The life insurance contract. The surrender time τ is assumed to be a random
time, i.e. τ = τ (ω) is a strictly positive random variable defined on Ω and F-measurable. We
also defined the process Ht in the following way: Ht = 1{τ ≤t} . The crucial assumption here
is we assume τ is not an F-stopping time. We refer the reader to the previous chapter for a
discussion of this assumption.
2.2.1. The insurer’s payments. A large variety of life insurance contracts can be modelled
as a combination of the following three building blocks. The first one is the payoff the insurer
has to pay at the term of the contract T . This payoff is assumed to be a FT -measurable random
variable and is denoted by g(T, ω). At the term of the contract, the insurer has to pay:
g(T, ω)1{τ >T }

is assumed to be an F-predictable stochastic process. We can write
Z T
1{0<τ ≤T } R(τ, ω) = R(u, ω)dHu
0
not surrendered. We model these payoffs through their cumulative value up to time t, denoted
by C(t, ω). This process is assumed to be a right continuous increasing F- adapted process.
The cumulative payoff up to surrender is then given by:
Z T
0
where we assume C(0, ω) = 0 and C(T, ω) = C(T− , ω).

2.2.2. The policyholder’s payments. We assume a policyholder pays premiums periodically
at N fixed dates ti with i = 0, . . . N − 1, as long as he has not surrendered. We denote by
P (ti , ω) the value of the premium paid at time ti and assume P (ti , ω) is Fti -measurable for
each i = 0, . . . N − 1. The policyholder’s payments is then given by:
N
X −1
P (ti , ω)1{τ >ti }
i=0
94 The Risk-Minimization Theory.
3. The Risk-Minimization Theory.

We will here recall the main results of the theory of risk minimization. These results are
due to Föllmer and Sondermann [44] for a single payoff and were extended by Møller [68] to
payment processes. We mainly follow Schweizer [78] for the presentation.
The financial market is essentially defined as in Section (2.1) but we will here consider an
arbitrary filtration G larger than the filtration F generated by the financial assets prices. We
assume a specific martingale measure Q has been chosen so that X is a (Q, G)-local martingale.
Definition 3.1. A trading strategy ρ is a pair of processes (ε, η) where η is a 1-dimensional
real-valued G-adapted process and ε is an s-dimensional real-valued G-predictable process.
The process εt represents the amount of risky assets held at time t and the process ηt is
the discounted amount invested in the risk free asset.
Definition 3.2. The value process Vt (ρ) of a trading strategy ρ, is defined by Vt (ρ) =
εt Xt + ηt for 0 ≤ t ≤ T .
This value process represents the discounted value of the insurer’s financial portfolio fol-
lowing the trading strategy ρ. This portfolio is not necessarily self-financed. We define by
L2 (X, Q) the following space.
Definition 3.3. L2 (X, Q, G) is the space of Rs - and G-predictable processes ε such that
hR i1/2
T
kεkL2 (X,Q,G) = E Q 0 ε0u d [X, X]u εu < ∞.
We also have the following important lemma due to Schweizer:
Lemma 3.4. Suppose X is a (Q, G)-local martingale. For any ε ∈ L2 (X, Q, G), the process
RT 2 2

εdX ε ∈ L2 (X, Q, G) is a
R
0 εu dXu ∈ M0 (Q, G). Moreover, the space I (X, Q, G) =
stable subspace of M20 (Q, G).
Proof. See Schweizer [78].
Definition 3.5. An RM-strategy ρ is any pair (ε, η) where ε ∈ L2 (X, Q, G)
and η is
a càdlàg adapted process such that the value process Vt (ρ) is right continuous and square
integrable.
The liabilities of the insurer towards a policyholder are modelled as a process At . The
process At is assumed to be càdlàg, G-adapted and square integrable. The process At represents
the discounted value of the cumulative payments up to time t.
Definition 3.6. The cumulative cost process Ct (ρ) of a RM-strategy ρ, is defined by
Rt
Ct (ρ) = Vt (ρ) − 0 εu dXu + At .
The initial cumulative cost process C0 (ρ) is given by C0 (ρ) = V0 (ρ) + A0 . The first term
V0 (ρ) represents the initial amount the insurer has to invest to create his financial portfolio.
The second term is the initial payment the insurer has to pay to the policyholder. Usually,
A0 will be negative and instead of a payment will represent the initial premium paid by the
policyholder.
Definition 3.7. A strategy is said to be mean self financing if its cumulative cost process
is a martingale.
The risk process is defined as follows:
h i
Definition 3.8. The risk process of an RM-strategy ρ is defined by Rt (ρ) = E Q (CT (ρ) − Ct (ρ))2 |Gt .
Following Møller [68], we restrict our attention to strategies which are 0-admissible in the
following sense:
Definition 3.9. A strategy ρ is said to be 0-admissible if and only if: VT (ρ) = 0 , Q-a.s.
The theory of risk minimization aims at finding an 0-admissible RM-strategy that minimizes
the risk process Rt (ρ) in the following sense.
Definition 3.10. An RM-strategy ρ is called risk-minimizing if and only if Rt (ρ) ≤ Rt (ρ̃)
Q-a.s. for every t ∈ [0, T ] and for any RM-strategy ρ̃ which is an admissible continuation of ρ
from t on, in the sense that VT (ρ) = VT (ρ̃) Q-a.s. , ε̃u = εu for u ≤ t and η̃u = ηu for u < t.
To solve this problem, we need the following result:
Lemma 3.11. Any risk-minimizing RM-strategy is also mean self financing.
Proof. See Møller [68] for a proof.
This lemma leads directly to the following one:
Lemma 3.12. The discounted value of the risk-minimizing RM-strategy ρ∗ is given by:
Vt (ρ∗ ) = E Q [AT − At |Gt ]
Proof. Indeed since by Lemma 3.11, the cost process is a martingale, we have
Ct (ρ∗ ) = E Q [CT (ρ∗ ) |Gt ]
Z t Z T
∗ Q
Vt (ρ ) − εu dXu + At = E 0 − εu dXu + AT |Gt
0 0
R
Since the stochastic integral εu dXu is a martingale by Lemma 3.4, we have the result.
The risk-minimizing discounted amount invested in the risk free asset is thus given by:
ηt∗ = E Q [AT − At |Gt ] − ε∗t Xt . We now have to determine the amount of risky assets ε∗ .
Schweizer shows the solution of this problem is closely related to the so-called Galtchouk-
Kunita-Watanabe decomposition of AT . Since I 2 (X, Q, G) is a stable subspace of M20 (Q, G),
any AT square integrable random variable can be uniquely written as:
Z T
AT = E Q [AT |G0 ] + εA A
u dXu + LT
0
Q-a.s., where εA
u∈ L2 (X, Q, G)and LA
∈
u
2 is Q-strongly orthogonal to I 2 (X, Q, G).
M0 (Q, G),
With these notation, we can give the solution of our risk-minimizing problem in the following
lemma:
96 Risk-Minimizing Strategies for a Single Life Insurance Contract.
Lemma 3.13. The unique risk-minimizing RM-strategy ρ∗ = (ε∗ , η ∗ ) of A is given by

ε∗ = εA and ηt∗ = E Q [AT − At |Gt ] − ε∗t Xt . The cumulative cost process of ρ∗ is given by
Ct (ρ∗ ) = E Q [AT |G0 ] + LA ∗ A ∗ ∗
t = C0 (ρ ) + Lt and the value process of ρ is given by Vt (ρ ) =
Q
E [AT − At |Gt ].
Proof. See Schweizer [78] for the single payoff case or Møller [68] for the extension to
payment processes.
As Schweizer noticed in [78], we do not have to assume X is (locally) square integrable for
these results to hold.
4. Risk-Minimizing Strategies for a Single Life Insurance Contract.

In this section, we apply the risk minimization theory to a single insurance contract. In
Section 5, we apply this theory to a portfolio with n insurance contracts.
Henceforth, we will assume a specific equivalent martingale measure Q has been chosen and
we will assume we are in the theoretical framework described in Section 2.
4.1. Preliminary assumptions and results.

4.1.1. The insurer’s information. We assume the insurer observes the financial market and
whether the policyholder has surrendered or not. We define H = (Ht )0≤t≤T with Ht =
σ(Hs , 0 ≤ s ≤ t), the filtration generated by the process Ht . We defined G = (Gt )0≤t≤T
where Gt = Ft ∨ Ht , the filtration corresponding to the progressive enlargement of the filtration
F by H1. Intuitively, the filtration G corresponds to the insurer’s information. Since we assume
τ is not an F-stopping time, the filtration G is strictly larger than F.
4.1.2. The hazard process. In this subsection, we introduce the (F, Q)-hazard process of a
random time τ .
Definition 4.1. Let τ be a non negative random variable defined on (Ω, F). Assume
Q(τ = 0) = 0 and Q(τ > t) > 0 ,∀t∈ R+ . We define the (F, Q)-hazard process Γ̂of τ by
F̂t = 1 − e−Γ̂t
where F̂t = Q(τ ≤ t |Ft ),∀t ∈ R+ . Notice this hazard process is well defined since our surrender
time τ is not an F-stopping time.
Before studying the risk-minimizing strategies, we first need a few preliminary results. The
first results can be found in Jeanblanc and Rutkowski [58] or in Bielecki and Rutkwoski [23]
but for completeness we give them here.
Theorem 4.2. Let the process Lt be defined by Lt = 1{τ >t} eΓ̂t then Lt is a (Q, G)-
martingale.
1We assume G is complete and satisfies the usual hypothesis.

Proof. Let us prove that E Q [ Ls | Gt ] = Lt , ∀ s > t. We have:

h i h i
E Q 1τ >s eΓ̂s Gt = 1τ >t eΓ̂t E Q e−Γ̂s eΓ̂s Ft

= 1τ >t eΓ̂t
= Lt

Theorem 4.3. If the stochastic process Γ̂t is continuous and of finite variation then we
Rt
have Lt = 1 − 0 eΓ̂u dMu where Mu is given by: Hu − Γ̂u∧τ .
Proof. By the integration by parts, we have:
Z t Z t h i
Γ̂u
Lt = 1τ >u− de + eΓ̂u− d1τ >u + eΓ̂· , 1τ >·
0 0 u
The square bracket is equal to

h i X
eΓ̂· , 1τ >· = ∆eΓ̂s ∆1τ >s
u
0≤s≤u
= ∆eΓ̂0 ∆1τ >0

= 1
The second equality comes from the hypothesis that Γ̂u is continuous. The last one comes from
the assumption eΓ̂0− = 1τ >0− = 0 and eΓ̂0 = 1τ >0 = 1. Since Γ̂u is continuous and of finite
variation, we also have
Z t Z t
Γ̂u
1τ >u− de = 1τ >u− eΓ̂u dΓ̂u
0 0
Z t
= eΓ̂u dΓ̂u∧τ
0
and
Z t Z t
Γ̂u−
e d1τ >u = eΓ̂u d1τ >u
0 0
Z t
= − eΓ̂u dHu
0
Finally, we get the result:
Z t Z t
Γ̂u
Lt = 1 + e dΓ̂u∧τ − eΓ̂u dHu
0 0
Z t
= 1− eΓ̂u dMu
0

The corollary of the two previous theorems is that when Γ̂t is continuous and of finite
variation, Γ̂t∧τ corresponds to the (Q, G)-compensator of the counting process Ht . It is not
true when Γ̂t is not continuous. From now on, we assume Condition 2.3 holds.
Condition 4.4. The (Q, F)- hazard process Γt of τ is continuous.
4.1.3. The enlargement of filtration and the (H)-hypothesis. We assumed in Section 2.1,
there is no arbitrage in the financial market with respect to the filtration F. Unfortunately, the
financial market is not necessarily arbitrage free with respect to the enlarged filtration G. A
sufficient condition for the financial market to be arbitrage free with respect to G, is that the
(H)-hypothesis defined below, holds under the martingale measure Q.
Definition 4.5. Let (Ω, F, Q) be a probability space. Let G be a filtration and F an
arbitrary sub-filtration of G i.e. for every t, Ft ⊆ Gt . We say the (H)-hypothesis holds under
Q between the filtrations F and G if and only if: for every t, F∞ and Gt are conditionally
independent with respect to Ft .
From now on, we assume this (H)-hypothesis holds.
Condition 4.6. The (H)-hypothesis holds under Q between F and G.
The following lemma shows this hypothesis is equivalent to the invariance of local martin-
gales from F to G.
Lemma 4.7. The following assertions are equivalent
(1) The (H)-hypothesis holds under Q.
(2) Every (F, Q)-local martingale is a (G, Q)-local martingale.
(3) ∀t ≥ 0, ∀H ∈ F∞ , E Q [H |Gt ] = E Q [H |Ft ]
Proof. See Brémaud and Yor [33].
As we already said, this condition is sufficient for the absence of arbitrage to hold under G. If
furthermore the financial market is complete, it is even a necessary condition. See Blanchet-
Scalliet and Jeanblanc [58] for a proof in a slightly different framework. In our setting, we also
have the following equivalence:
Lemma 4.8. The (H)-hypothesis under Q is equivalent to
Q(τ ≤ t |F∞ ) = Q(τ ≤ t |Ft )
The corollary of this lemma is that under the (H)-hypothesis, F̂t and thus Γ̂t are increasing
and accordingly, are of finite variation. The following proposition is essential.
Proposition 4.9. If the (H)-hypothesis holds for Q and if the (F, Q)-hazard process Γ̂t is
continuous then ∆Uτ = 0 a.s. for any F-adapted càdlàg process Ut .
Proof. First, we have

E Q 1{τ <t} |Ft = E Q lim→0 1{τ ≤t−} |Ft

= lim→0 E Q 1{τ ≤t−} |Ft−

= lim→0 1 − e−Γt−

= 1 − e−Γt−
where > 0. In the third equality, we use the fact that when the (H)-hypothesis holds, we
have for all s > t,
E Q 1{τ ≤t} |Fs = E Q E Q 1{τ ≤t} |F∞ |Fs

= E Q E Q 1{τ ≤t} |Ft |Fs

= E Q 1{τ ≤t} |Ft

where in the second equality we use Lemma 4.2. Since Ut is càdlàg, there is only a countable
number of jumps on the bounded interval [0, T ]. Let us denote Sn the times of jump of Ut in
this interval. We have:
" #
X X
Q
E Q E Q 1{τ =Sn } |FSn

E 1{τ =Sn } =
n n
X
E Q E Q 1{τ ≤Sn } − 1{τ <Sn } |FSn

=
n
X
EQ 1 − e−ΓSn − E Q 1{τ <Sn } |FSn

=
n
X
EQ 1 − e−ΓSn − 1 − e−ΓSn −

=
n
X
E Q e−ΓSn − 1 − e−∆ΓSn

=
n
= 0
In the first equality, we use the factPthe number of jumps is countable. In the last equality, we
use the fact Γ is continuous.
P Since n 1{τ =Sn } is a positive random variable, if its expectation
is equal to zero, it means n 1{τ =Sn } = 0 a.s.
We now apply the theory of risk minimization to the different building blocks we introduced
earlier.
4.2. Payment at the term T of the contract. We want to find the risk-minimizing
strategy of the payoff 1τ >T BgTT where we assume gT is square integrable with respect to Q. We
have the following proposition.
Proposition 4.10. Assume conditions 2.3 and 4.1 hold. Let us denote the (Q,hF)-Galtchouk- i
Kunita-Watanabe decomposition of the purely financial contingent claim Utg = E Q BgTT e−Γ̂T Ft

by
Z t
g
Utg = U0g + φu dXu + LU
t
0
Then the discounted value Vtg (ρ∗ )

of the risk-minimizing strategy for a claim paying at time T ,
1τ >T gT where gT is a FT -measurable random variable, is given by

gT −Γ̂T
Vtg (ρ∗ ) = 1τ >t eΓ̂t E Q e Ft
BT
Its risk-minimizing strategy ρ∗ is given by

ρ∗ = 1τ >t− eΓ̂t φt , Vtg − 1τ >t− eΓ̂t φt Xt
where φt is thus the amount of risky assets of the strategy risk-minimizing the purely financial
contingent claim Utg . The cost process Ctg (ρ∗ )is given by:
Ctg (ρ∗ ) = C0g (ρ∗ ) + Lgt

where

gT −Γ̂T
C0g (ρ∗ ) = E Q
e F0
BT
and
Z t Z t
g
Lgt = Lu− dLU
t − g Γ̂u
Uu− e dMu
0 0
Proof. The cumulative discounted payment process is given by At = 1τ >T BgTT 1{t=T } .
According to Lemma 3.12, the value process Vtg (ρ∗ ) of the risk-minimizing strategy ρ∗ for the
claim At is given by:

g ∗ Q gT
Vt (ρ ) = E 1τ >T Gt
BT

Γ̂t Q gT −Γ̂T
= 1τ >t e E
BT
e Ft
since τ admits an (Q, F)-hazard process Γ̂t .

Finding the amount ε∗ of risky assets comes down to finding the Galtchouk-Kunita-Watanabe
h i
of the square integrable (Q, G)-martingale Jtg defined as Jtg = E Q 1τ >T BgTT Gt 2. Let us de-

h i
note Utg = E Q BgTT e−Γ̂T Ft . It is a (Q, F)-square integrable martingale and thanks to the

2Here, we obviously have J g = V g (ρ∗ ). But this will not be true for the two other payment processes. This
t t
is why we distinguish between the processes.
(H)-hypothesis, it is even a (Q, G)-square integrable martingale. Applying the integration by

parts formula to Jtg , we can write:
Jtg = Lt Utg
Z t Z t Z t
= L0 U0g + Lu− dUug d [L· , U·g ]u
+ g
Uu− dLu +
0 0 0
Z t Z t Z t
= L0 U0g + Lu− dUug − g Γ̂u
Uu− e dMu − eΓ̂u d [M· , U·g ]u
0 0 0
In the third equality, we use the Theorem 4.8 since Γ̂t is, by assumption, continuous and in-
creasing thanks to the (H)-hypothesis. Furthermore thanks to Proposition 4.10, the continuity
of Γ̂t implies Utg does not jump at time τ . We can write the square bracket as
X
[M· , U·g ]t = ∆Ms ∆Usg
0≤s≤t
= M0 U0g
We then have:
Z t Z t
(4.1) Jtg = L0 U0g + Lu− dUug − g Γ̂u
Uu− e dMu
0 0
Rt g Rt g Γ̂u
The (Q, G)-local martingales 0 Lu− dUu and 0 Uu− e dMu are even (Q, G)-square integrable
martingales. Indeed, we have
Z t Z t 2 Z t
g Γ̂u g Γ̂u
[J g , J g ]t = L2u− d [U·g , U·g ]u + Uu− e d [M· , M· ]u − 2 Lu− Uu− e d [U·g , M· ]u
0 0 0
Z t Z t 2
g Γ̂u
= L2u− d [U·g , U·g ]u + Uu− e d [M· , M· ]u
0 0
hR i 2
t 2 Rt
Since E [[J·g , J·g ]t ]
< ∞, we have E < ∞ and E 0 g g
d [M· , M· ]u <
0 Lu− d [U· , U· ]u
g Γ̂u
Uu− e
∞.
We now apply the Galtchouk-Kunita-Watanabe decomposition to the (Q, F)-square inte-
grable martingale Utg , we get
Z t
g
(4.2) Utg = U0g + φu dXu + LU
t
0
g
where φ is an F-predictable stochastic process in L2 (X, Q, F) and LU is in M20 (Q, F) and
strongly (Q, F)-orthogonal to I 2 (X, Q, F). Replacing Equation (4.2) in (4.1) , we then obtain
the following decomposition
Z t
(4.3) Jtg = L0 U0g + Lu− φu dXu + Lgt
0
Rt g R t g Γ̂
where Lgt = 0 Lu− dLU t − 0 Uu− e dMu .
u
Let us prove the Equation (4.3) is indeed the required Galtchouk-Kunita-Watanabe decom-
Rt Rt g
position. Firstly, 0 Lu− φu dXu and 0 Lu− dLU 2
t are obviously in M0 (Q, G). Secondly, we have
g 2
R Lt is strongly
to prove R orthogonal to I(X, Q, G).It Ris equivalent toshowing the following prod-
t t U g Rt t g Γ̂u
ucts 0 ψu dXu 0 Lu− dLt and 0 ψu dXu 0 Uu− e dMu are (Q, G)-martingales for
hR i
· R · g Γ̂
any G-predictable process ψ ∈ L2 (X, Q, G). On one hand, we have 0 ψu dXu , 0 Uu− e u dMu =
Rt t
g Γ̂u
0 ψ u Uu− e d [X · , M · ]u = 0 ∀t ∈ [0, T ] Q-a.s. since M t is a finite variation process and there
is no common jumps between Xt and Mt . On the other hand, knowing that on {τ > t}, any
G-predictable process ψ ∈ L2 (X, Q) can be written as an F-predictable process ψ̂ ∈ L2 (X, Q),
we have
Z · Z · G Z t Z · G
Ug Γ̂u Ug
ψu dXu , Lu− dLu = 1{τ >u−} e d ψ̂u dXu , L·
0 t 0 0 u
Z t Z · F
g
= 1{τ >u−} eΓ̂u d ψ̂u dXu , LU
·
0 0 u
= 0
In the second equality, we use the fact under the (H)-hypothesis, the sharp brackets under
g
G and under F are indistinguishable. The third equality, we use the fact LU is strongly by
(Q, F)-orthogonal to I 2 (X, Q, F).
As far as the cost process is concerned, we immediately have by definition:
Ctg (ρ∗ ) = C0g (ρ∗ ) + Lgt

Z t Z t
g ∗ Ug g Γ̂u
= C0 (ρ ) + Lu− dLt − Uu− e dMu
0 0
where Mt = Hu − Γ̂{τ ∧u} .

Notice the cost process is the sum of two terms which are strongly (Q, G)-orthogonal to
g
each other. If the market is complete LU t = 0 for any t ∈ [0, T ] and the cost process becomes
t g Γ̂u
Ctg (ρ∗ ) = C0g (ρ∗ ) − 0 Uu−
R
e dMu . The first component is thus related to the incompleteness
of the market and the second one is related to the unpredictability of the surrender time.
Notice also, if the market is complete, φ is the self financing replicating portfolio of the modified
contingent claim Utg . The same proposition holds, mutatis mutandis, for the premium payments
P (ti , ω)1{τ >ti } .
4.3. Payments at the surrender time. R t RWe now want to find the risk-minimizing strat-
egy of the discounted payment process At = 0 B u
u
dH u which is assumed to be square integrable
with respect to Q. We have the following proposition.
Proposition 4.11. Assume conditions 2.3 and 4.1 hold. Let us denote the (Q, hF)-Galtchouk- i
R T Ru
Kunita-Watanabe decomposition of the purely financial contingent claim UtR = E Q 0 B d F̂ u Ft

u
by
Z t
R
UtR = U0R + ψu dXu + LU
t
0
Then the discounted value VtR (ρ∗ ) of the risk-minimizing portfolio for the payment Rτ at
the surrender time τ is given by:
T
Z
Ru
VtR (ρ∗ ) Γ̂t Q

= 1τ >t e E dF̂u Ft
t Bu
Its risk-minimizing strategy ρ∗ is given by
T
Z
∗ Γ̂t Γ̂t Q Ru Γ̂t
ρ = 1τ >t− e ψt , 1τ >t e E dF̂u Ft − 1τ >t− e ψt Xt

t Bu
where ψt is the amount of risky assets of the strategy risk-minimizing the purely financial
contingent claim UtR . The cost process CtR (ρ∗ ) of this risk-minimizing strategy is given by
CtR (ρ∗ ) = C0R (ρ∗ ) + LR

t
where
T
Z
Rs
C0R (ρ∗ ) Q

= E dF̂s F0
0 Bs
and
t Z t T
Z Z
R Ru Rs −(Γ̂s −Γ̂u )
LR Lu− dLU − EQ

t = t + e dΓ̂s Fu dMu
0 0 Bu u Bs −
Proof. We essentially follow the same steps as in Proposition 4.10. According to Lemma
3.12, the discounted value of the risk-minimizing portfolio is given by
T Z t
Z
Ru Ru
VtR (ρ∗) Q

= E dHu − dHu Gt
0 Bu 0 Bu
Z T
Ru
= EQ

dHu Gt
t Bu
Z T
Γ̂t Q Ru
= 1τ >t e E dF̂u Ft
t Bu
since τ admits a hazard process Γ̂t .

The amounts of risky assets to hold is given
h R by the Galtchouk-Kunita-Watanabe
i of the
R Q T Ru
square integrable (Q, G)-martingale Jt = E 0 Bu dHu Gt . Using the integration by parts

hR i
T Ru
in the fourth equality and denoting UtR = E Q 0 Bu dF̂u Ft , we have

t T
Z Z
Ru Ru
JtR dHu + E Q

= dHu Gt
0 Bu t Bu
t Z T
Z
Ru Γ̂t Q Ru
= dHu + 1τ >t e E dF̂u Ft
0 Bu t Bu
t Z t Z T
Z
Ru Γ̂t Ru Γ̂t Q Ru
= dHu − 1τ >t e dF̂u + 1τ >t e E dF̂u Ft
0 Bu 0 Bu 0 Bu
Z t Z t Z t Z t Z t
Ru Γ̂t Ru R R R Γ̂u
eΓ̂u M· , U·R u

= dHu − 1τ >t e dF̂u + L0 U0 + Lu− dUu − Uu− e dMu −
0 Bu 0 Bu 0 0 0
Since Γ̂t is continuous, thanks to Proposition 4.10, there is no discontinuity in UtR at time τ ,
the square bracket is thus equal to 0. Furthermore, we have
Z t Z t Z t Z u Z ·
Γ̂t Ru Γ̂u Ru Rs Γ̂u Ru
1τ >t e dF̂u = 1τ >u− e dF̂u − dF̂s e dMu + L· , dF̂u
0 Bu 0 Bu 0 0 Bs − 0 Bu t
Z t Z t Z u
Ru Rs
= 1τ >u− eΓ̂u e−Γ̂u dΓ̂u − dF̂s eΓ̂u dMu
0 Bu 0 0 B s −
Z t Z t Z u
Ru Rs
= dΓ̂u∧τ − dF̂s eΓ̂u dMu
0 B u 0 0 Bs −
In the second equality, we use the fact the square bracket is null since F̂ is continuous. Finally,
since Γ̂t∧τ is the (Q, G)-compensator of Ht , we obtain
Z t Z t Z T
R R R Ru Q Rs −(Γ̂s −Γ̂u )
(4.4) Jt = L0 U0 + Lu− dUu + −E e dΓ̂s Fu dMu
0 0 Bu u Bs −
Thanks to the (H)-hypothesis, the square integrable (Q, F)-martingale UtR is even a square
integrable (Q, G)-martingale.
Using the same argument as in Proposition 4.10, we can eas-
hR i
Rt t Ru T Rs −(Γ̂s −Γ̂u )
ily show 0 Lu− dUuR and 0 B − EQ u B
R
e dΓ̂s Fu dMu are even square inte-

u s −
grable (Q, G)-martingales.
Since UtR is a square integrable (Q, F)-martingale, we can find its Galtchouk-Kunita-
Watanabe decomposition. We have
Z t
R R
(4.5) Ut = U0 + ψu dXu + LU
t
0
R
where ψ is an F-predictable stochastic process in L2 (X, Q, F) and LU is in M20 (Q, F) and
strongly (Q, F)-orthogonal to I 2 (X, Q, F). Replacing Equation (4.5) in Equation (4.4), we
obtain the following decomposition:

T
Z Z t
R Q Rs
Lu− ψu dXu + LR

Jt = E dF̂s F0 + t
0 Bs 0
hR i
Rt R Rt Ru T Rs −(Γ̂s −Γ̂u )
where LR
t =
U
0 Lu− dLt + − EQ u B e d Γ̂s Fu dMu . Using the same ar-

0 Bu s −
gument as in Proposition 4.10, we can easily prove this decomposition is indeed the Galtchouk-
Kunita-Watanabe decomposition.
As far as the cost process is concerned, we immediately have:
CtR (ρ∗ ) = C0R (ρ∗ ) + LR
t
Again, the cost process can be decomposed in two strongly orthogonal components: the
first one is related to the incompleteness of the market and the second one again related to the
unpredictability of the surrender time.
4.4. Payments up to surrender.

Rt We now study the risk-minimizing strategy of the
discounted payment process At = 0 1τ >u B1u dCu where Ct is assumed to be a square integrable
process with respect to Q.
Proposition 4.12. Assume conditions 2.3 and 4.1 hold. Let us denote hthe (Q, F)-Galtchouk-
i
RT
Kunita-Watanabe decomposition of the purely financial claims UtC = E Q 0 B1u e−Γ̂u dCu Ft

by
Z t
C
UtC = U0C + νu dXu + LU
t
0
Then the discounted value VtC (ρ∗ ) of the risk-minimizing portfolio for the payment process
RT
0 1τ >u dCu is given by:
T
Z
1 −Γ̂u
VtC (ρ∗ ) Γ̂t Q

= 1{τ >t} e E e dCu Ft
t Bu
The risk-minimizing strategy ρ∗ is given by
T
Z
∗ Γ̂t Γ̂t Q 1 −Γ̂u Γ̂t
ρ = 1τ >t− e νt , 1τ >t e E e dCu Ft − 1τ >t− e νt Xt

t Bu
where νt is the amount of risky assets of the strategy risk-minimizing the contingent claim UtC .
The cost process CtC (ρ∗ ) of the risk-minimizing strategy is given by:
CtC (ρ∗ ) = C0C (ρ∗ ) + LC
t
where
T
Z
1 −Γ̂u
C0C (ρ∗ ) Q

= E e dCu F0
0 Bu
and
"Z #
t t T
e−(Γ̂s −Γ̂u )
Z Z
UC
LC = Lu− dLt − EQ dCs Fu dMu

t
0 0 u Bs
−
Proof. The discounted value of the risk-minimizing portfolio is given by:

T Z t
Z
1 1
VtC (ρ∗ ) Q

= E 1τ >u dCu − 1τ >u dCu Gt
0 Bu 0 Bu
T
Z
1
= EQ

1τ >u dCu Gt
t Bu
When τ admits a hazard process Γ̂t of finite variation, the last equation can be written as:
T
Z
1 −Γ̂u
VtC (ρ∗ ) Γ̂t Q

= 1{τ >t} e E e dCu Ft
t Bu
As already explained the optimal amounts of risky assets ν to hold is given

h R by the Galtchouk- i
C Q T 1
Kunita-Watanabe of the square integrable (Q, G)-martingale Jt = E 1 dC Gt .

0 τ >u Bu u
We have:
T
Z
1
JtC = EQ

1τ >u dCu Gt
0 Bu
"Z #
Z t T −Γ̂u
1 e
= 1τ >u dCu + 1τ >t eΓ̂t E Q dCu Ft

0 Bu t Bu
Z t −Γ̂u "Z #
Z t T −Γ̂u
1 e e
= 1τ >u dCu − 1τ >t eΓ̂t dCu + 1τ >t eΓ̂t E Q dCu Ft

0 Bu 0 B u 0 Bu
hR i
T e−Γ̂u
Let us denote by UtC the following square integrable (Q, F)-martingale UtC = E Q 0 Bu dCu Ft .

By integration by parts, we have:
t t
e−Γ̂u t t t
Z Z Z Z Z
1
JtC dCu − 1τ >t eΓ̂t Lu− dUuC C Γ̂u
eΓ̂u d M· , U·C t

= 1τ >u dCu + − Uu− e dMu −
0 Bu 0 Bu 0 0 0
Thanks to Proposition 4.10, UtC has no discontinuity at time τ , the square bracket is then equal
to M0 U0C = 0. Furthermore, we have
Z t −Γ̂u !
e
Z t −Γ̂u Z t Z u −Γ̂s
Γ̂t Γ̂u− e e
1τ >t e dCu = 1τ >u− e dCu − dCs eΓ̂u dMu
0 Bu 0 B u 0 0 Bs
−
" Z · −Γ̂u #
e
+ 1τ >· eΓ̂· , dCu
0 Bu
t
!
Z t − Γ̂ Z t Z u −Γ̂s
Γ̂u e e
u
= 1τ >u e dCu − dCs eΓ̂u dMu
0 B u 0 0 Bs
−
Z t Z t Z u −Γ̂s !
1 e
= 1τ >u dCu − dCs eΓ̂u dMu
0 Bu 0 0 Bs
−
We have thus
"Z #
t t T
e−(Γ̂s −Γ̂u )
Z Z
(4.6) JtC = L0 U0C + Lu− dUuC − E Q
dCs Fu dMu

0 0 u Bs
−
From now on, we can easily prove the proposition using the same arguments as in Propositions
4.10 and 4.11. This is left to the reader.
Once again, the cost process
R t can be decomposed in two parts, one related to the incomplete-
C
ness of the financial market 0 Lu− dLU t and the second one related to the unpredictability of
R t Q h R T e−(Γ̂s −Γ̂u ) i
the surrender time − 0 E dCs Fu dMu .

u Bs −
5. Risk-Minimizing Strategies for a Portfolio of Life Insurance Contracts.

We now assume our portfolio consists of n policyholders. We denote by τ i the surrender
time for each policyholder and define Hti = 1{τ i ≤t} with i = 1, . . . n. Møller in [68] studied
a similar situation but he assumed the random times τ i were independent of each other and
he assumed the financial market was the (complete) Black-Scholes model. In this setting, he
showed that by following the risk-minimizing strategy, the risk of such a portfolio can be in
part, hedged away and showed that the unhedgeable part of the risk (the risk process) tends
to zero in relative terms with the size n of the portfolio. In other words, he showed that
by taking a sufficiently large portfolio and by following the risk-minimizing strategy, we can
reduce as much as we want the relative risk of this portfolio. Riesner [72] extended this result
by introducing an incomplete market driven by a Lévy process.
In our framework, the random times τ i are obviously not independent of each other since
they all depend on the financial market. It is not obvious whether or not we can obtain Møller’s
conclusion in our setting. Our aim here is to study how far we can go and keep Møller’s result.
In the first subsection, we give some preliminary assumptions and results. Then in the
next subsections, we will study the risk-minimizing strategies for the different building blocks
108 Risk-Minimizing Strategies for a Portfolio of Life Insurance Contracts.
when there are initially n policyholders. In Section 6, we will study the risk processes of these
strategies.
5.1. Preliminary assumptions and results.
5.1.1. The insurer’s information. We assume here the insurer observes the financial market
and whether the different policyholders have already surrendered or not. This information can
be modelled trough the following filtration Gp = (Gtp )0≤t≤T where Gtp = Ft ∨ Ht1 ∨ Ht2 ∨ · · · ∨ Htn
and Hti = σ(Hti ) for each i = 1, . . . n. We also introduce the following notation.
Notation 5.1. We denote the filtration Gi = Gti 0≤t≤T where Gti = Ft ∨ Hti and the

filtration Gp−i = Gtp−i where Gtp−i = Ft ∨ Ht1 ∨ · · · ∨ Hti−1 ∨ Hti+1 ∨ · · · ∨ Htn .
0≤t≤T
5.1.2. The enlargement of filtration and the (H)-hypothesis. As in Section 4.1.3, we have
to introduce an assumption that avoids the existence of arbitrage opportunities under the
enlarged filtration Gp . As we already know the (H)-hypothesis is a sufficient condition for the
no arbitrage hypothesis to hold.
Condition 5.2. The (H)-hypothesis between the filtration F and Gp holds under Q .
This assumption implies the (H)-hypothesis holds for any subfiltration. Indeed, we have
the following proposition.
Proposition 5.3. If the (H)-hypothesis holds under Q between the filtration F and Gp then
it holds between F and any filtration I such that F ⊆ I ⊆ Gp .
Proof. We can easily prove this proposition thanks to the third condition of Lemma 4.7.
Let H ∈ F∞ and t ≥ 0, we have
E Q [H |It ] = E Q E Q [H |Gtp ] |It

= E Q E Q [H |Ft ] |It

= E Q [H |Ft ]
In the second equality, we use the fact the (H)-hypothesis holds under Q between the filtration
F and Gp
In particular, the (H)-hypothesis holds for the filtrations Gi = Ft ∨ Hti 0≤t≤T for any

i = 1, . . . n.
5.1.3. The F-independence. A well-know property of the risk-minimization theory is its
linearity: the risk-minimizing strategy of the sum of two payoffs is simply the sum of the risk-
minimizing strategies of each separate payoff. It would then be tempting to directly use this
property to conclude the risk-minimizing strategy of a portfolio of n insurance contracts is
simply given by the sum of the risk-minimizing strategies given in Section 4. But this solution
is not true in general. It comes from the fact that when we have n contracts the insurer’s
information set is different from the one of Section 4. Indeed, if we have initially n policyholders,
the filtration we should consider is now the filtration Gp . The linearity property of the risk
minimization indeed holds under the filtration Gp : the risk-minimizing strategy for a portfolio
of n contracts is thus the sum of the risk-minimizing strategies of the individual contracts but
with respect to the filtration Gp . Unfortunately, the risk-minimizing strategy with respect to
Gp , for a single contract is not necessarily the same as the risk-minimizing strategy with respect
to the filtration G (which we derived in the preceding section), for the same single contract.
Intuitively, since in general the different surrender times are not independent of each other,
the fact someone has surrendered or not, does give a valuable information on the probability
someone else has surrendered or not. In establishing our risk-minimizing strategies, we should
accordingly take into account this information. This is not possible if we only observe G.
Deriving these risk-minimizing strategies with respect to Gp seems to be a difficult task
because we could have in general complex dependencies between the different surrender times.
So, we would like to find conditions under which we could use the results we derived in the
previous sections and which are in the meantime, sufficiently realistic. An obvious assumption
is to assume the surrender times are independent of each other. In this case, intuitively, the
fact someone has surrendered should not give any information on the surrender of another. The
individual risk-minimizing strategies with respect to Gp would then be the same as the risk-
minimizing strategies with respect to G. This is the assumption made in Møller’s papers and
Riesner’s and is indeed realistic if we were to model the dates of decease of different individuals,
but is unfortunately too strong if we are to model surrender times. Here we would like to find
an intermediary assumption weaker than the pure independence. Let us introduce the following
condition:
Condition 5.4. For each i, , j = 1, . . . n with i 6= j, τ i and τ j are (Q, F)-independent.
This F-conditional independence assumption has a direct intuitive meaning. As we already
argued, the fact the surrender times are not F-stopping times, means the policyholders take
their surrender decisions partly on financial variables which are observable by the insurer, and
partly on other variables unknown to the insurer. The fact the different surrender times are
F-independent means these unobservable variables which influence the surrender decisions, are
purely idiosyncratic variables. They are not “macro” variables which could have a systematic
influence on the other policyholders decisions. In other word, the only “macro” variables which
have a systematic influence on the surrender decisions comes from the financial market and
are also observable by the insurer. From a mathematical point of view, this assumption has a
number of consequences. In particular, this implies the following equality:
(5.1) E Q [1τ i >T 1τ j >T |FT ] = E Q [1τ i >T |FT ] E Q [1τ j >T |FT ]
We also have the following propositions:
Proposition 5.5. If Condition 5.4 holds then, for t < T and H a GTi -measurable random
variable, we have
E Q [H |Gtp ] = E Q H Ft ∨ Hti

5.1.4. The hazard process. Based on Definition 4.1, we can here define a hazard process
with respect to any sub-filtration of Gp . However, we have the following proposition.

Proposition 5.6. If Condition 5.4 holds then Q τ i ≤ t Gtp−i = Q τ i ≤ t |Ft for each

i = 1, . . . n.
In other words, the (Q, Gp−i )-hazard process of τ i is equal to its (Q, F)-hazard process.
Intuitively, the fact the insurer can observe the financial market and the surrender of the other
policyholders does not give us more information on the surrender of a policyholder than the
observation of the financial market alone. Intuitively, this property should allow us to say the
risk-minimizing strategy for n contracts is given by the sum of the individual risk-minimizing
strategies as derived in Section 4. This is what we will prove formally in the next subsections.
In the rest of the chapter, we will denote by Γ̂it the (Q, F)-hazard process of τ i . For the sake
of simplicity, we will assume the policyholders have homogeneous behaviors with respect to
the evolution of the financial market. This assumption translates into Γ̂it = Γ̂jt ∀ t ∈ [0, T ],
∀i, , j = 1, . . . n. In this case, we will simply write Γ̂t = Γ̂it ∀i = 1, . . . n. As in Section 4, we
also have the following propositions
i
Proposition 5.7. Let the process Lit be defined by Lit = 1{t>τ i } eΓ̂t then if Condition 5.4
holds, Lit is a (Q, Gp )-martingale.
Proof. Let us prove that E Q Lis Gtp = Lit , ∀ s > t. We have:

h i
E Q 1 e−Γ̂s G p−i
h i s>τ i t
i

E Q 1s>τ i eΓ̂s Gtp = 1t>τ i h i
p−i
E Q 1t>τ i | Gt
i
h i i
= 1t>τ i eΓ̂t E Q e−Γ̂s E Q 1s>τ i Gsp−i Gtp−i

= Lit
where in the second and third equalities, we use Proposition 5.6.
Proposition 5.8. If the stochastic process Γ̂it is continuous and of finite variation then we
Rt i
have Lit = 1 − 0 eΓ̂u dMui where Mui is given by: Hui − Γ̂i{u∧τ i } .
Proof. Same proof than Theorem 4.8.
The corollary of the two previous theorems is that when Condition 5.4 holds and Γ̂it is
continuous and of finite variation, Γ̂i{u∧τ i } is the (Q, Gp )-compensator of Hti . If we denote
Nt = ni=1 Hi then the (Q, Gp )-compensator of Nt is given by ni=1 Γ̂i{u∧τ i } or by (n − Nu− ) Γ̂u
P P
if the policyholders are homogeneous.

As in Section 4, we assume from now on the following condition holds:
Condition 5.9. For each i = 1, . . . n, the (Q, F)-hazard process Γ̂it is continuous.
Notice since the (H)-hypothesis holds between Fand Gi for each i = 1, . . . n, then Γ̂it is an
increasing process for each i = 1, ..n thanks to Lemma 4.2, and thus a finite variation process.
As in Proposition 4.10, we can prove there is no jump in the F-adapted càdlàg processes at any
time τ i .
Proposition 5.10. For any i = 1, . . . n, if the (H)-hypothesis holds between F and Gi

under Q and if the (F, Q)-hazard process Γ̂it of τ i is continuous then ∆Uτ i = 0 Q-a.s. for any
F-adapted càdlàg process Ut .
Proof. See Proposition 4.10.
In the three next subsections, we study the risk-minimizing strategies for each of our build-
ing blocks.
5.2. Payment at the term. For a portfolio of n homogeneous policies paying a payoff
gT if the policyholder has not surrender before the term T , we have the following proposition.
Proposition 5.11. If the Conditions 5.2, 5.6 and 5.9 hold, then the discounted value
V g,p (ρ∗ ) of the risk-minimizing n-policyholders portfolio is given by

gT −Γ̂T
Vtg,p (ρ∗ ) = (n − Nt ) e E Γ̂t Q
e Ft
BT
where Nt = ni=0 1τ i ≤t .
P

∗ Γ̂t Γ̂t Q gT −Γ̂T Γ̂t
ρt = (n − Nt− ) e φt , (n − Nt ) e E e Ft − (n − Nt− ) e φt Xt
BT
where φ is the amount of

h risky assets
i of the strategy risk-minimizing the purely financial con-
g Q gT −Γ̂T
tingent claim Ut = E BT e Ft . The cost process is given by:
Ctg,p (ρ∗ ) = C0g,p (ρ∗ ) + Lg,p

t
where

gT −Γ̂T
C0g,p (ρ∗ ) = nE Q
e F0
BT
Z t Z t
g
Lg,p
t = Γ̂u
(n − Nu− ) e dLU
u + g Γ̂u
Uu− e dMup
0 0
and Mtp = Nt − (n − Nt− ) Γ̂t .

Proof. The discounted value of the risk-minimizing n-policyholders portfolio V g,p (ρ∗ ) is
given by:
" n #
g,p ∗ Q
X gT p
Vt (ρ ) = E 1τ i >T G
BT t
i=1
n
X
Q gT i
= E 1τ i >T Ft ∨ Ht
BT
i=1
n
X
Γ̂it Q gT −Γ̂i
= 1τ i >t e E e Ft
T
BT
i=1
where in the second equality, we use Proposition 5.5. For each i = 1, . . . n, we can use the
decomposition given in Proposition 4.10. If for each i = 1, . . . n, we write
h the (Q, F)-Galtchouk-
i
g,i Q gT −Γ̂iT
Kunita-Watanabe decomposition of the (Q, F)-martingale Ut = E BT e Ft as:
Z t
gT −Γ̂i gT −Γ̂i
E Q
e T Ft = E Q
e T F0 + φiu dXu + Lg,i
t
BT BT 0
then, we get:
n n
Z t !
gT −Γ̂i i
Vtg,p (ρ∗ )
X X
Q
= E e T
F0 + 1{τ i >u−} eΓ̂u φiu dXu
BT 0
i=1 i=1
n Z t n Z t
Γ̂iu g,i g,i Γ̂u
X X
+ 1{τ i >u−} e dLU
t + Uu− e dMui
i=1 0 i=1 0
This is indeed the (Q,Gp )-Galtchouk-Kunita-Watanabe

decomposition of Vtg,p (ρ∗ ) since thanks
Γ̂ju
R t Pn Γ̂iu i
R t U g,j are (Q, Gp )-square
to Condition 5.2, 0 i=1 1{τ i >u−} e φu dXu and 0 1{τ j >u−} e dLt
integrable martingales strongly (Q, Gp )-orthogonal for any , j = 1, . . . n and since M j is also a
P
t n Γ̂iu i
(Q, Gp )-square integrable martingale strongly (Q, Gp )-orthogonal to 0
R
i=1 1{τ i >u−} e φu dXu
for any j = 1, . . . n thanks to Proposition 5.10. Accordingly, we can say the (Q, Gp )-Galtchouk-
Kunita-Watanabe decomposition of the payoff of the portfolio is given by the sum of the
(Q, Gi )-Galtchouk-Kunita-Watanabe decomposition of the payoff of each contract.
If furthermore, we assume the policyholders are homogeneous, we have

g,p ∗ Γ̂t Q gT −Γ̂T
Vt (ρ ) = (n − Nt ) e E e Ft
BT
and the decomposition:
Z t
gT −(Γ̂T −Γ̂0 )
Vtg,p (ρ∗ ) = nE Q
e F0 + (n − Nu− ) eΓ̂u φu dXu
BT 0
Z t Z t
Γ̂u Ug g Γ̂u
+ (n − Nu− ) e dLt + Uu− e dMup
0 0
where Mup =
Pn i
i=1 Mu = Nu − (n − Nu− ) Γ̂u .
As an illustration, we can easily show this proposition give the same result than Møller’s.
He made two additional assumptions. He assumed the τ i were independents of the filtration F.
In this case, the process Γ̂t is a continuous deterministic function. Furthermore, he assumed
g
the market was complete. In this case, we have LU t = 0. Eventually, we get
Z t Z t
g,p Q gT −Γ̂T Γ̂u g Γ̂u
Vt (ρ∗) = nE e F0 + (n − Nu− ) e φu dXu + Uu− e dMup
BT 0 0
Z t
g
(n − Nu− ) e−(Γ̂T −Γ̂u ) γu dXu
T
= nQ(τ > T )E Q

F0 +
BT 0
Z t
gT −(Γ̂T −Γ̂u )
EQ dMup

+
B T
Fu e
0 −
Notice γu is F-predictable and is the self financing replication strategy of the discounted payoff
gT . Indeed, we have:

g Q gT −Γ̂T
Ut = E
BT
e Ft

−Γ̂T Q gT
= e E Ft
BT
Z t
−Γ̂T Q gT
= e E F0 + γu dX u
BT 0
Z t
gT
= e−Γ̂T E Q F0 + e−Γ̂T γu dXu
BT 0
For a portfolio of n policyholders, we then have the following strategy

gT
(n − Nu− ) e−(Γ̂T −Γ̂t ) γt , (n − Nu ) e−(Γ̂T −Γ̂t ) E Q Ft − (n − N u− ) e−(Γ̂T −Γ̂t )
γ X
t t
BT
This is exactly Møller’s result.
5.3. Payment at the surrender time. If we have a portfolio with initially n policy-
holders with homogeneous behavior, we have the following proposition.
Proposition 5.12. If the Conditions 5.2, 5.6 and 5.9 hold then the discounted value
VtR,p (ρ∗ )
of the risk-minimizing n-policyholders portfolio is given by:
Z T
R,p ∗ Γ̂t Q Ru
Vt (ρ ) = (n − Nt ) e E dF̂u Ft
t Bu
Pn
where Nt = i=0 1τ i ≤t .
Z T
∗ Γ̂t Γ̂t Q Ru Γ̂t
ρt = (n − Nt− ) e ψt , (n − Nt ) e E dF̂u Ft − (n − Nt− ) e ψt Xt

t Bu
where ψ is the amount of

h Rrisky assets
ofi the strategy risk-minimizing the purely financial con-
R Q T Ru
tingent claim Ut = E 0 Bu dF̂u Ft . The cost process is given by

CtR,p (ρ∗ ) = C0R,p (ρ∗ ) + LR,p

t
where
T
Z
Rs
C0R,p (ρ∗ ) Q

= nE dF̂s F0
0 Bs
t Z t T
Z Z
R Ru Rs −(Γ̂s −Γ̂u )
LR,p Γ̂u
dLU − EQ dMup

t = (n − Nu− ) e t e dΓ̂s Fu
0 0 Bu u Bs −
and Mtp =Nt − (n − Nt− ) Γ̂t .
Proof. The discounted value of the risk-minimizing portfolio is given by

" n Z #
T
R u

VtR,p (ρ∗ ) = E Q dHui Gtp
X
B u
i=1 t

n Z T
X
Q Ru i
i
= E dHu Ft ∨ Ht
t Bu
i=1
n Z T
X i Ru i
= 1τ i >t eΓ̂t E Q dF̂u Ft
i=1 t Bu
where in the second equality, we use Proposition 5.5. To find the risk-minimizing strategy, we
have to find the (Q, Gp )-Galtchouk-Kunita-Watanabe decomposition of
" n Z #
T
R u

JtR,p = E Q dHui Gtp
X
B u
i=1 0

n Z T
X
Q Ru i
i
= E dHu Ft ∨ Ht
0 Bu
i=1
As in the previous proposition, we can easily see the (Q, Gp )-Galtchouk-Kunita-Watanabe de-
composition for the portfolio is given by the sum of the (Q, Gi )-Galtchouk-Kunita-Watanabe
decomposition for each contract. If for each i = 1, . . . n, we write the (Q, F)-Galtchouk-Kunita-
hR i
T Rs −Γ̂is
Watanabe decomposition of the square integrable (Q, F)-martingale UtR,i = E Q 0 B s
e d Γ̂iF
s t
as:
Z T Z T Z t
Rs −Γ̂is i Rs −Γ̂is i
E Q
e dΓ̂s Ft = E Q
e dΓ̂s F0 + ψui dXu + LR,i
t
0 B s 0 B s 0
Then using the results of Proposition 4.11, we have

n n
Z T Z t X !
R,p
X
Q Rs −Γ̂is i Γ̂iu i
Jt = E e dΓ̂s F0 + 1{τ i >u−} e ψu dXu
0 Bs 0
i=1 i=1
n Z t n Z t Z T
X
Γ̂iu U R,i
X Ru Q Rs −(Γ̂is −Γ̂iu ) i
+ 1{τ i >u−} e dLt + −E e dΓ̂s Fu dMui
0 0 Bu u B s −
i=1 i=1
If the policyholders are homogeneous, we have:
T
Z
Ru
VtR,p (ρ∗ ) Γ̂t Q

= (n − Nt ) e E dF̂u Ft
t Bu
and
T
Z Z t
Rs −Γ̂s
JtR,p Q
(n − Nu− ) eΓ̂u ψu dXu

= nE e dΓ̂s F0 +
0 B s 0
Z t Z t Z T
Γ̂u U R R u Q Rs −(Γ̂s −Γ̂u )
dΓ̂s Fu dMup

+ (n − Nu− ) e dLt + −E e
0 0 Bu u Bs −
Pn n p
Where Nt = i=0 Hti = i=0 1τ i ≤u and Mu =Nu − (n − Nu− ) Γ̂u .
P

5.4. Payment up to the surrender time. If we have a portfolio with initially n poli-
cyholders with homogeneous behavior, we obtain this time the following proposition:
Proposition 5.13. If the Conditions 5.2, 5.6 and 5.9 hold, then the discounted value
VtC,p (ρ∗ )
of the risk-minimizing n-policyholders portfolio is given by
Z T
C,p ∗ Γ̂t Q 1 −Γ̂u
Vt (ρ ) = (n − Nt ) e E e dCu Ft
t Bu
Pn
where Nt = i=0 1τ i ≤t .
Z T
∗ Γ̂t Γ̂t Q 1 −Γ̂u
dCu Ft − (n − Nt− ) eΓ̂t νt Xt

ρt = (n − Nt− ) e νt , (n − Nt ) e E e
t Bu
where ν is the amount of
h Rrisky−Γ̂assets of ithe strategy risk-minimizing the purely financial con-
C Q T e u
tingent claim Ut = E 0 Bu dCu Ft . The cost process is given by

CtC,p (ρ∗ ) = C0C,p (ρ∗ ) + LC,p

t
where
T
Z
1 −Γ̂u
C0C,p (ρ∗ ) Q

= E e dCu F0
0 Bu
116 The Risk Process for a Portfolio of Life Insurance Contracts.
"Z #
t t T
e−(Γ̂s −Γ̂u )
Z Z
C
LC,p = Γ̂u−
(n − Nu− ) e dLU − E Q
dCs Fu dMup

t t
0 0 u Bs
−
and Mup =Nu − (n − Nu− ) Γ̂u

Proof. Same proof than the previous proposition.
6. The Risk Process for a Portfolio of Life Insurance Contracts.

In this section, we study the risk process of a portfolio. In particular, we want to study the
behavior of the relative risk process when the number of contract n increases. Before studying
the risk processes in more details, we need the following result.
Proposition 6.1. If Condition 5.4 holds then Lit Ljt is a (Q, Gp )-martingale for each i, j =
1, . . . n with i 6= j.
Proof. For t ≥ s, we have:
h i h i j i
E Q Lit Ljt |Gsp = E Q e(Γ̂t +Γ̂t ) 1τ i >t 1τ j >t |Gsp
h i j i
E Q e(Γ̂t +Γ̂t ) 1τ i >t 1τ j >t 1τ i >s Gsp−i

= 1τ i >s h i
E Q 1τ i >s Gsp−i

h j
i
Q e(Γ̂it +Γ̂t ) 1 p−i−j
E τ i >t 1τ j >t 1τ j >s Gs
i
= 1τ i >s 1τ j >s eΓ̂s h i
E Q 1τ j >s Gsp−i−j

h i j h i i
= Lis Ljs E Q e(Γ̂t +Γ̂t ) E Q 1τ i >t 1τ j >t Gtp−i−j Gsp−i−j

h i j i
= Lis Ljs E Q e(Γ̂t +Γ̂t ) E Q [1τ i >t 1τ j >t |Ft ] Gsp−i−j

h i j i
= Lis Ljs E Q e(Γ̂t +Γ̂t ) E Q [1τ i >t |Ft ] E Q [1τ j >t |Ft ] Gsp−i−j

= Lis Ljs
In the third equality, we use Proposition 5.6. In the fifth equality, we use the independence
of τ i and τ j with respect to Hk for k 6= i, j conditionally on F. In the sixth equality, we use
Equation (5.1).
Lit and Ljt are thus strongly (Q, Gp )-orthogonal. This leads to the following corollary.
Corollary 6.2. If Condition 5.4 holds and if the stochastic processes Γ̂it with i = 1, . . . n,
are continuous and of finite variation then M i and M j are (Q, Gp )-orthogonal for each i, j =
1, . . . n with i 6= j.
Proof. Straightforward thanks to Propositions 5.8 and 6.1.
Now, we can study the risk process of the whole portfolio. For d ∈ {g, R, C}, the risk
process is given by
2
d,p ∗ Q d,p ∗ d,p ∗
R0 (ρ ) = E CT (ρ ) − C0 (ρ )
 !2 
n Z T n Z T
d,i d,i
X X
(6.1) = EQ  Liu− dLUu + Ku− dMui 
i=1 0 i=1 0
h i i
i hR i
T Rs −(Γ̂is −Γ̂iu )
g,i
where Ku− = −E Q e−(Γ̂T −Γ̂u ) BgTT |Fu , Ku−
R,i
= BRu
u
− E Q
u Bs e d Γ̂i |F
s u and
− −
i i
R T −(Γ̂s −Γ̂u )
C,i
Ku− = −E Q u e Bs dCs |Fu . We have the following result.
−
Proposition 6.3. If we assume the Conditions 5.2, 5.4 and 5.9 hold then, for each d ∈
{g, R, C}, the risk process R0d,p (ρ∗ ) for a n-policyholders portfolio is given by
 !2 
n Z T n
" Z
T 2 #
U d,i
R0d,p (ρ∗ ) d,i
X X
= EQ  Liu− dLu + EQ Ku− dMui
i=0 0 i=1 0
h i i
i hR i
T Rs −(Γ̂is −Γ̂iu )
where Ku−g,i
= −E Q e−(Γ̂T −Γ̂u ) BgTT |Fu , Ku− R,i Ru
= B u
− E Q
u Bs e d Γ̂i |F
s u and
− −
C,i R T −(Γ̂is −Γ̂iu )
Ku− = −E Q u e Bs dCs |Fu . Furthermore, if the policyholders are homogeneous then,
−
as n increases, the relative risk process tends to:
q
R0d,p (ρ∗ ) q d
Ud

lim = EQ LU
· , L· T
n→∞ n
Proof. According to Equation (6.1), we can write
 !2 
n Z T
d,i
R0d,p (ρ∗ ) = E Q 
X
Liu− dLU
u

i=1 0
n n Z
" ! Z #
T T
U d,i d,j
X X
+ 2 EQ Liu− dLu Ku− dMuj
j=1 i=1 0 0
n
" Z
T 2 #
d,i
X
Q
+ E Ku− dMui
i=1 0
n−1 n Z T Z T
d,i d,j
X X
Q
+ 2 E Ku− dMui Ku− dMuj
i=1 j=i+1 0 0
118 The Risk Process for a Portfolio of Life Insurance Contracts.
R
T d,i
Thanks to Corollary 6.2, we know the square integrable (Q, Gp )-martingales 0 Ku− dMui
R
T d,j
and 0 Ku− dMuj are strongly (Q, Gp )-orthogonal for i 6= j. We thus have
hR R i
T d,i T d,j j
EQ 0 K u− dM i
u 0 K u− dM u = 0 for i 6= j. We also have
hP R i
n T i d,i T d,j j
EQ U = 0 since these integrals are also strongly (Q, Gp )-
R
i=1 0 L u− dLu 0 K u− dM u
orthogonal thanks to Proposition 5.10.
If furthermore, we assume the n policyholders are homogeneous, the risk process becomes:
n
" Z
T 2 # n−1 n Z T Z T
d,p ∗ Ud Ud j Ud
X X X
Q i Q i
R0 (ρ ) = E Lu− dLu +2 E Lu− dLu Lu− dLu
i=1 0 i=1 j=i+1 0 0
n
" Z
T 2 #
d,i
X
+ EQ Ku− dMui
i=1 0
" Z
T 2 # Z T Z T
d d d
Q
Liu− dLU + n2 − n E Q Liu− dLU Lju− dLU

= nE u u u
0 0 0
" Z
T 2 #
d,i
+ nE Q Ku− dMui
0
for arbitrary i 6= j. The limit of the relative risk process is then given
q s
R0d,p (ρ∗ ) T T
Z Z
lim = EQ Liu− dLU
u
d
Lju− dLU
u
d
n→∞ n 0 0
s Z T
Liu− Lju− d LU
d
Ud

= EQ · , L· u
0
s Z T h i
Liu− Lju− |Fu− d d

= EQ EQ U U
d L· , L· u
0
q d
Ud

= E Q LU· , L· T
where in the third equality we use the predictable projection theorem and in the fourth, the
fact that
h i h i
E Q Liu− Lju− |Fu− = E Q Liu− |Fu− E Q Lju− |Fu−

= 1

As n increases, the risk related to the surrender, goes to 0 which means this risk can be
diversified away, even though the surrender times are not independent. For a sufficiently large
portfolio, the only risk remaining is thus the one related to the incompleteness of the financial
market. This risk could never be diversified away since it is systematic risk pertaining the
d
whole economy. However, if we assume the financial market is complete, the term LU t = 0 ∀t
Q-a.s. and the relative risk process will go to 0 as n increases. This proposition generalizes
Møller’s and Riesner’s results. Notice as n increases, the relative risk process of our portfolio
tends to the relative risk process of a portfolio of modified purely financial claims.
7. Conclusion.
In this chapter, we studied the risk-minimizing strategies for a single life insurance contract
with a surrender option and for a portfolio of such policies. We showed that if the random
times of surrender are F-independent, we can combine diversification and hedging to reduce
the risk of such a portfolio.
Notice that, in this chapter, assuming the (H)-hypothesis is crucial. Otherwise, the fi-
nancial prices are no longer martingales and the theory of risk-minimization does not apply
anymore. In the next chapter, we remove this assumption when studying the theory of local
risk-minimization.
CHAPTER 4
Locally Risk-Minimizing Strategies of Life Insurance Contracts

with Surrender Option.
1. Introduction.
The motivation of this chapter is to extend the local risk-minimization theory initially de-
veloped by Schweizer [76] for a single payoff, to a payment stream, and to apply this extended
theory to insurance contracts with surrender option or equivalently to default sensitive contin-
gent claims. In an independent work [79], Schweizer recently studied the same problem and
even extended the local risk-minimization theory to a multi-dimensional setting.
Unlike the traditional actuarial literature on the fair valuation of surrender option (see
[14, 15, 47, 81, 80] for examples), we do not assume that the surrender time is a(n) (optimal)
stopping time with respect to the filtration generated by the financial assets prices. Instead,
we assume that the surrender time is a random time that admits a hazard process. We refer
to Chapter 2 for a discussion of this assumption. Such random times have been applied to
model the default time in the credit risk literature and they have been thoroughly studied by
Jeanblanc and Rutkwoski in [58]. See also the book of Bielecki and Rutkowski [23]. In the
credit risk literature, such models are known as the reduced-form models.
The assumption that the surrender time admits a hazard process, has an important impact
for an insurer. Indeed, in this case, the surrender time is not predictable and it implies that
an insurance contract with a surrender option cannot be perfectly hedged even if the financial
market is initially complete. In other words, in this case, the surrender option introduces a
genuine surrender risk. It is thus interesting to study the hedging strategies an insurer should
follow to mitigate its exposure to this surrender risk. In the previous chapter, we already studied
the risk-minimizing strategies of such contracts. Here, we focus on the locally risk-minimizing
strategies.
Except for our previous work, the actuarial literature on the hedging of the surrender risk
is extremely limited. However, in our setting, the problem of hedging such insurance contracts
is, technically speaking, essentially the same than the problem of hedging default sensitive
contingent claims. It thus makes sense to review the financial literature. Quadratic hedging
strategies for default sensitive contingent claims have been studied in a number of papers.
Mean-variance hedging strategies have been extensively studied in the context of defaultable
markets in [24, 25, 26, 27, 28, 58]. As far as the local risk-minimization theory is concerned,
it has already been applied in the credit risk literature in Lotz [63] and in Biagini and Cretarola
[21, 22]. Lotz studied the locally risk-minimizing strategies of a defaultable claim when a risk
free bond and a defaultable bond are traded. In [21], Biagini and Cretarola studied the locally
121
risk-minimizing strategies of a defaultable claim with two non defaultable primitive assets, a
money market account and a risky asset, assuming a mutual dependence of the risky asset and
the default time. In [22], they extended their previous work to deal with random recovery at
default time.
This chapter can be seen as an extension of the works of Biagini and Cretarola. Actually,
we extend [22] in at least three different directions. Firstly, in [22], Biagini and Cretarola
showed how to locally risk-minimize a payment at the time of default but they did not give a
local risk-minimization theory for more general payment streams. Here, we show that the local
risk-minimization theory can be extended to arbitrary (square integrable) payment processes
in the same way as Møller did in [68], for the risk-minimization theory. Secondly, Biagini and
Cretarola considered a complete financial market with a single risky asset whose price followed
a (generalized1) geometric Brownian motion. Here, we assume that the discounted financial
assets prices follow a general continuous s-dimensional semimartingale. Thirdly, unlike these
authors, we do not assume that the so-called (H)-hypothesis necessarily holds. In addition to
this, we also give a detailed study of the impact of a progressive enlargement of filtration on
the so-called minimal martingale measure.
The conclusion of this chapter is that, in each case, we can write the (pseudo) locally risk-
minimizing hedging strategies of an insurance contract with surrender option as a function of
the (pseudo) locally risk-minimizing hedging strategies of a properly modified purely financial
claims.
This chapter is organized as follows. In Section 2, we introduce the theoretical framework;
we describe the financial market and the different types of payments of our insurance contract.
In Section 3, we introduce the local risk-minimization theory. In Section 3.1, we briefly present
the local risk-minimization theory for a single payoff as developed in Schweizer [76] and in
Section 3.2, we show how this theory can be extended to payment processes. In Sections 4 and
5, we apply this extended theory to the insurance payments described in Section 2. In Sections
4, we assume that the (H)-hypothesis holds whereas in Section 5, we remove this assumption.

In this section, we introduce the financial market, the surrender time (resp. the default
time) and the payments of the insurance contracts (resp. of the default sensitive contingent
claim) we want to study.
2.1. Description of the financial market. Let (Ω, F, P ) be a complete probability
space. A perfect frictionless financial market is defined in this space. We assume there is one
locally risk free asset denoted by Bt = B(t, ω) and s risky assets Sti = S i (t, ω), i = 1, . . . s
following real càdlàg stochastic processes. The price of the locally risk free asset is assumed to
follow a strictly positive, continuous process of finite variation. The discounted values of these
assets are denoted by
Sti
Xti = , i = 1, . . . s
Bt
1The drift and the diffusion terms can be stochastic in their model.
Locally Risk-Minimizing Strategies of Life Insurance Contracts with Surrender Option. 123
Let F = (Ft )t≥0 be a filtration that satisfies the usual hypothesis. We assume the stochastic
processes Sti for i = 1, . . . s and Bt are adapted to this filtration.
We now recall some standard assumptions about Xt that appear in the mathematical finance
literature. We assume Xt is a special semimartingale in Sloc 2 (P, F). We write its (P, F)-canonical
s
decomposition as Xt = X0 +Nt +At where Nt is R -valued locally square integrable (P, F)-local
martingale null at 0, i.e. Nt ∈ M20,loc (P, F), and At is an Rs -valued F-predictable process of

finite variation null at 0 such that the variation of Ai , Ai t , is assumed to be locally square
integrable for i = 1, . . . s. We assume that, for i = 1, . . . s, Ait N·i , N·i t with predictable

density αti , i.e. we have

Z t
i
αui d N·i , N·i u

At =
0
P -a.s. for all
t and all
i = 1, . . . s. Let B̄t a fixed increasing predictable P
càdlàg
process
null at
0 such that N·i , N·i t B̄t for each i = 1, . . . s (we can choose B̄t = si=1 N·i , N·i t ). The
D E
Kunita-Watanabe inequality implies N·i , N·j B̄t for all i, j = 1, . . . s . We denote by σtij
D E t
i j
the predictable density of N· , N· with respect to B̄t i.e.
t
Z t

i j
N· , N · t = σuij dB̄u
0
P -a.s. for all t and all i, j = 1, . . . s. We can then write
Z t
i
At = γui dB̄u
0
where γti
= αti σtii
P -a.s. for all t and all i = 1, . . . s. Finally, we assume that Xt satisfies the
structure condition. This structure condition is defined as follows:
Definition 2.1. (The structure condition) We say that Xt satisfies the structure condition
if there exists an Rs -valued predictable process λ̂t ∈ L2loc (N ) such that
σt λ̂t = γt
P -a.s. for all t.
Rt
A càdlàg version of the process K̂t := 0 λ̂0u γu dB̄u is called the mean variance trade-off
process of X.
2.2. The random time τ . We now introduce a random time τ , i.e. τ = τ (ω) is a strictly
positive random variable defined on Ω and F-measurable. It describes the time of surrender
of a policyholder (or the time of default of a company in the credit risk literature). We define
the process Ht := 1{τ ≤t} . The filtration generated by this process Ht is denoted by H, i.e.
H = (Ht )t≥0 where Ht = σ(Hs , 0 ≤ s ≤ t). We define G = (Gt )t≥0 where Gt = Ft ∨ Ht , the
filtration corresponding to the progressive enlargement of the filtration F by H2. We assume
that τ satisfies the following condition.
2We assume G is complete and satisfies the usual hypothesis.
Condition 2.2. τ is not an F-stopping time.

See Chapter 2 for a discussion of this assumption for the surrender time. As in Jeanblanc
and Rutkowski [58], we now introduce the so-called (F, P )-hazard process Γt of τ by:
Ft = 1 − e−Γt
where Ft := P (τ ≤ t |Ft ), ∀t ≥ 0. We assume P (τ > t) > 0 for all t ≥ 03. Thanks to Condition
2.2, it implies Ft < 1 for all t ≥ 0 and Γt = − ln(1 − Ft ) is indeed well defined. For simplicity,
we assume F0 = 0.
Ft is a (P, F)-submartingale. Accordingly, it admits a unique Doob-Meyer decomposition
that we write Ft = Dt + Yt where Dt is an F-predictable increasing process and Yt a (P, F)-
martingale.
In [58] or in [23], we can find that the (P, G)-compensator Λt of Ht is given by
Z t∧τ
1
(2.1) Λt = dDu
0 1 − Fu−
We denote the compensated process Mt := Ht − Λt . Throughout this chapter, we will assume
the following condition holds.
Condition 2.3. The (P, F)-hazard process Γt of τ is continuous.
Before describing the payments of our insurance contract, we have to stress a last point.
At the beginning of this subsection, we introduce the enlargement of the filtration F by the
random time τ . This enlargement of filtration deserves some comments. In general, a (local)
martingale with respect to a given filtration is not necessarily a (local) martingale with respect
to a larger one. In our setting, it means the (P, F)-local martingales are not necessarily (P, G)-
local martingales. Let us introduce the following definition.
Definition 2.4. We say the (H)-hypothesis holds under P between F and G if any (P, F)-
local martingale is a (P, G)-local martingale.
In this chapter, if this (H)-hypothesis does not hold, then Nt , in particular, is not necessarily
a (P, G)-local martingale. It implies that the canonical decomposition of Xt is not necessarily
the same if we consider the filtration F or the filtration G. Actually, we do not even know if Xt
is still a semimartingale in the enlarged filtration. Accordingly, without this (H)-hypothesis,
the dynamics of the financial market can be deeply altered by this enlargement of filtration.
We should thus be very cautious and distinguish when this (H)-hypothesis holds and when it
does not. In this chapter, we study in details these two cases: Section 4 deals with the locally
risk-minimizing strategies of life insurance contracts when this (H)-hypothesis is assumed and
Section 5 deals with the same problem when this (H)-hypothesis is removed.
2.3. The insurance contract. We now describe the payments we wish to hedge.
3or at least for all t ∈ [0, T ] where T is the term of the insurance contract.
2.3.1. The insurer’s payments. A large variety of life insurance contracts (or default-
sensitive contingent claims) can be modelled as a combination of the following three building
blocks.
The first one is the payoff the insurer has to pay at the term of the contract T . This payoff
is assumed to be a FT -measurable random variable and is denoted by g(T, ω). At the term of
the contract, the insurer has to pay:
g(T, ω)1{τ >T }

is assumed to be an F-predictable stochastic process. We have
Z T
1{0<τ ≤T } R(τ, ω) = R(u, ω)dHu
0
not surrendered. We model these payoffs through their cumulative value up to time t, denoted
by C(t, ω). This process is assumed to be a right continuous increasing F-adapted process. The
cumulative payoff up to surrender is then given by:
Z T
0
where we assume C(0, ω) = 0 and C(T, ω) = C(T− , ω). Notice this third payoff can be written
as the sum of the two first ones with g(T, ω) = C(T, ω) and R(u, ω) = C(u− , ω).
2.3.2. The policyholder’s payments. We assume a policyholder pays premiums periodically
at N fixed dates ti with i = 0, . . . N − 1, as long as he has not surrendered. We denote by
P (ti , ω) the value of the premium paid at time ti and assume P (ti , ω) is Fti -measurable for
each i = 0, . . . N − 1. The policyholder’s payments is then given by:
N
X −1
P (ti , ω)1{τ >ti }
i=0
We will not refer to these payments in the rest of the chapter since they are similar to the
insurer’s payment at the term of the contract.
3. The Local Risk-Minimization Theory.

In Section 3.1, we briefly review the theory of local risk-minimization developed by Schweizer
in [76], for a single payoff made at a fixed date T . We follow the presentation of Schweizer in
[78]. In Section 3.2, we show how to extend this theory to payment processes.
126 The Local Risk-Minimization Theory.
3.1. A brief review of the local risk-minimization theory. The financial market is
defined as in Section 2.1. Let us start with a few definitions.
Definition 3.1. ΘS is the space of all processes ϑ ∈ L(X) for which the stochastic integral
ϑdX is in the space S 2 (P ) of semimartingales.
R
Definition 3.2. An L2 -strategy ρ is a pair ρ = (ϑ, η) where ϑ ∈ ΘS and η is a real valued

adapted process such that the value process Vt (ρ) = ϑt Xt + ηt is right continuous and square
integrable.
Definition 3.3. The cumulative cost process Ct (ρ) of a strategy ρ = (ϑ, η), is defined by
Rt
Ct (ρ) = Vt (ρ) − 0 ϑu dXu .
Definition 3.4. A strategy ρ is said to be mean-self-financing if its cost process is a

(P, F)-martingale.
Definition 3.5. The risk process Rt (ρ) of a L2 -strategy ρ is defined by

h i
Rt (ρ) = E P (CT (ρ) − Ct (ρ))2 |Gt
We now restrict our attention to the case s = 1 for which the local risk-minimization theory
has been developed.
Definition 3.6. A trading strategy ∆ = (δ, ε) is called a small perturbation if it satisfies
the following conditions:
a) δ is bounded R
b) The variation of δdA is bounded
c) δT = εT = 0
For any subinterval (s, t] of [0, T ], if ∆ = (δ, ε) is a small perturbation then we can define
the small perturbation ∆ (s,t] by

∆ (s,t] := δ1(s,t] , ε1[s,t)
We can now introduce the main definitions:
Definition 3.7. For an L2 -strategy ρ, a small perturbation ∆ and a partition π of [0, T ],
we define the R-quotient rπ [ρ, ∆] (t, ω)

X Rti ρ + ∆ (ti ,ti+1 ] − Rti (ρ)
π
r [ρ, ∆] (t, ω) := h i (ω)1(ti ,ti+1 ] (t)
ti ∈π E hM iti+1 − hM iti |Fti
We can now introduce the definition of a locally risk-minimizing strategy.

Definition 3.8. A strategy ρ is called locally risk-minimizing if
lim inf n→∞ rπn [ρ, ∆] ≥ 0, PM − a.e.

for every small perturbation ∆ and every increasing 0-convergent sequence πn of partitions of
[0, T ].
Schweizer in [76] established the following theorem:
Theorem 3.9. Let H ∈ L2 (FT , P ) be a contingent claim and let ρ be an L2 -strategy with
VT (ρ) = H, P -a.s. If the following assumptions (1-6) hold then ρ is locally risk-minimizing if
and only if ρ is mean-self-financing and the martingale Ct (ρ) is strongly (P, F)-orthogonal to
Nt .
(1) s = 1.
(2) Nt is a (P, F)-square integrable martingale with N0 = 0.
(3) hN· , N· it is strictly increasing P -a.s.
(4) At is continuous.
Rt
(5) A h hN· , N· i with predictable
i density αt i.e. A t = 0 αu d hN· , N· iu P -a.s. for all t.
RT 2
(6) E 0 |αu | d hN· , N· iu < ∞.
Strictly speaking, the local risk-minimization theory has only been developed for s = 14. When
s ≥ 1, inspired by the last theorem, Schweizer defines pseudo locally risk-minimizing strategies
in the following way:
Definition 3.10. Let H ∈ L2 (FT , P ) be a contingent claim. An L2 -strategy ρ with
VT (ρ) = H, P -a.s., is called pseudo-locally risk-minimizing for H if ρ is mean-self-financing
and the martingale Ct (ρ) is strongly (P, F)-orthogonal to Nt .
This definition leads to the following decomposition.
Theorem 3.11. Let H ∈ L2 (FT , P ) be a contingent claim. H admits a pseudo-locally
risk-minimizing L2 -strategy ρ with VT (ρ) = H P -a.s., if and only if H can be written as:
Z T
H = H0 + εH H
u dXu + LT
0
with H0 ∈ L2 (F0 , P ), εH ∈ ΘS and LH 2
t ∈ M0 (P, F) strongly (P, F)-orthogonal to Nt . The
pseudo-locally risk-minimizing strategy ρ = (ϑ, η) is then given by ϑt = εH
t and ηt = Vt (ρ) −
Rt H
εH
t X t where V t (ρ) = H 0 + ε
0 u dXu + LH , ∀t ∈ [0, T ].
t
Proof. See Schweizer [78] or see the proof of Proposition 3.19.

h RT i
Remark 3.12. We can also write the value process as Vt (ρ) = E P H − t εH
u dXu |Ft .
4However, in a recent independent work [79], Schweizer extended the local risk minimization theory to a
multi-dimensional setting.
The decomposition of H described in the previous theorem is called the Föllmer-Schweizer

decomposition of H.
Unfortunately, it is often difficult to directly find this decomposition. However, under some
further assumptions (the continuity of the assets prices for example), the Föllmer-Schweizer
decomposition can be found by finding the Galtchouk-Kunita-Watanabe (GKW) decomposition
of H but with respect of a particular martingale measure P̂ called the minimal martingale
measure.
Theorem 3.13. Assume the structure condition holds and the canonical
decomposition of
R·
Xt is given by Xt = X0 + Nt + At . Define the process Ẑt := Et − 0 λ̂dN . If X is continuous
then Ẑt is a strict martingale density for Xt .
Proof. See Schweizer [77] for examples.
dP̂
If furthermore ẐT ∈ L2 (FT , P ) then the measure P̂ defined as dP := ẐT is called the
minimal martingale measure. When such a minimal martingale measure exists, it is unique.
We then have the following theorem.
Theorem 3.14. Assume the contingent claim H admits a Föllmer-Schweizer decomposi-
tion. Let X be continuous and assume the minimal martingale measure P̂ exists. Then the
P̂ -martingale VtH,P̂ defined as VtH,P̂ = E P̂ [H |Ft ] admits a Galtchouk Kunita Watanabe de-
composition under P̂ with respect to Xt as
Z t
VtH,P̂ = V0H,P̂ + εH,P̂ H,P̂
u dXu + Lt
0
For t = T , this decomposition is the Föllmer-Schweizer decomposition of H with εH,

t
P̂
= εH
t .
We also have H0 = V0H,P̂ = E P̂ [H |F0 ].
Proof. See Schweizer [78] .
3.2. Extension of the local risk-minimization theory for payment processes. In
this section, we extend the local risk-minimization theory described in Section 3.1. Here, we
assume the payoffs are not necessarily paid at a fixed date T but can be paid continuously on the
time interval [0, T ]. In other words, the payments are now modelled as a stochastic process Et .
More precisely, the process Et represents the discounted value of the cumulative payments up to
time t and is assumed to be càdlàg, F-adapted and square integrable. As already explained, the
local risk-minimization has also been independently extended to payment streams in Schweizer
[79]. To derive a local risk-minimization theory for payment processes, we can basically follow
Møller [68] when he extended the risk-minimization theory of Föllmer and Sondermann to
payment processes. As Møller did, we first slightly modify the definition of the cumulative cost
process.
Definition 3.15. The cumulative cost process Ct (ρ) of a strategy ρ, is defined by Ct (ρ) =
Rt
Vt (ρ) − 0 εu dXu + Et .
The initial cumulative cost process C0 (ρ) is given by C0 (ρ) = V0 (ρ) + E0 . The second
term E0 is the initial payment the insurer has to pay to the policyholder. The first term V0 (ρ)
represents the initial amount the insurer has to invest to create his financial portfolio (after
the payment of E0 ). Following Møller [68], we restrict our attention to strategies which are
0-admissible in the following sense:
Definition 3.16. A strategy ρ is said to be 0-admissible if and only if: VT (ρ) = 0 , P -a.s.
All the other definitions introduced in the previous section, remain unchanged.
The following proposition is the counterpart of Theorem 3.9 for payment processes. Here,
we assume s = 1. It is a straightforward extension of Schweizer’s proof.
Proposition 3.17. Let Et be a càdlàg, F-adapted and square integrable payment process
and let ρ be an L2 -strategy 0-admissible. If the 6 conditions of Theorem 3.9 hold then ρ is
locally risk-minimizing if and only if ρ is mean self financing and the (P, F)-martingale Ct (ρ)
is strongly (P, F)-orthogonal to Nt .
Proof. We can follow Schweizer’s proof with the new definition of the cost process and
for a 0-admissible strategy. Let us decompose this proof in three steps:
1) In Lemma 2.1 in [76], Schweizer shows that if a trading strategy ρ is locally risk-minimizing,
it is mean self financing. We can follow the same proof than in [76] if, following the same
notation than Schweizer, we construct the trading strategy ρ̂ = (ε̂, η̂) as:
ε̂ := ε
Z T Z s
η̂s := E VT (ρ) − εu dXu + ET |Fs + εu dXu − Es − εs Xs
0 0
for s ∈ [0, T ]. In this case, we thus have:

Z T Z s
Vs (ρ̂) = E VT (ρ) − εu dXu + ET |Fs + εu dXu − Es
0 0
and in particular VT (ρ̂) = VT (ρ) = 0. Using the modified definition of the cost process, we
have, as in Schweizer, that
Z T
CT (ρ̂) = VT (ρ̂) − ε̂u dXu + ET
0
Z T
= VT (ρ) − εu dXu + ET
0
= CT (ρ)
and that
Z T
E [CT (ρ̂) |Fs ] = E VT (ρ̂) − ε̂u dXu + ET |Fs
0
Z T
= E VT (ρ) − εu dXu + ET |Fs
0
Z s
= Vs (ρ̂) − εu dXu + Es
0
= Cs (ρ̂)
With these results, the rest of the proof is identical to Schweizer’s. A locally risk-minimizing
strategy ρ is thus mean self-financing and the modified version of the cost process Ct (ρ) is a
(P, F)-martingale.
2) In Lemma 2.2 in [76], Schweizer shows that basically we can find a locally risk-minimizing
trading strategy by varying only the ε-component. The proof of this lemma depends on the
fact the cost process Ct (ρ) is a martingale but not on the specific definition of this process. We
can thus directly use Schweizer’s proof.
3) Finally, in Lemma 2.3, Schweizer shows that if ρ is locally risk-minimizing, the martingale
Ct (ρ) is strongly orthogonal to Nt . To prove this lemma, Schweizer uses Theorem 3.2 of [75].
Again, this theorem relies on the fact Ct (ρ) is a martingale and not on its specific definition.
This completes the proof.
When s ≥ 1, we can define pseudo-locally risk-minimizing strategies for payment processes
as in Section 3.1.
Definition 3.18. Let Et be a càdlàg, F-adapted and square integrable payment process.
An 0-admissible L2 -strategy ρ is called pseudo-locally risk-minimizing for Et if ρ is mean-self-
financing and the (P, F)-martingale Ct (ρ) is strongly (P, F)-orthogonal to Nt .
We now have the counterpart of Theorem 3.11 for payment processes:
Proposition 3.19. Let (Et )0≤t≤T ∈ L2 (Ω, P ) be a payment process. (Et )0≤t≤T admits a
pseudo-locally risk-minimizing L2 -strategy ρ with VT (ρ) = 0 P -a.s. if and only if ET admits a
Föllmer-Schweizer decomposition, i.e. if ET can be written as:
Z T
(3.1) ET = J0 + εE E
u dXu + LT
0
with J0 ∈ L2 (F0 , P ), εE ∈ ΘS and LE 2

t ∈ M0 (P, F) strongly (P, F)-orthogonal to Nt .
The pseudo-locally risk-minimizing strategy ρ = (ϑ, η) is then given by ϑt = εEt and ηt =
Vt (ρ) − εE
t Xt where its value process Vt (ρ) is given by
Z t
Vt (ρ) = J0 + εE E
u dXu + Lt − Et
0
∀ t ∈ [0, T ].
Proof. If ρ = (ϑ, η) is pseudo-locally risk-minimizing for Et , we know thanks to the

previous proposition, that LE 2
t := Ct (ρ) − C0 (ρ) ∈ M0 (P, F) and is strongly (P, F)-orthogonal
to Nt . Moreover, by the definition of the cost process, we have that for all t:
LE
t = Ct (ρ) − C0 (ρ)
Z t
= Vt (ρ) − ϑu dXu + Et − C0 (ρ)
0
Rt
so we indeed have Vt (ρ) = C0 (ρ)+ 0 ϑu dXu +LE t −Et . Moreover since by definition VT (ρ) = 0,
this last equation leads, for t = T , to Equation (3.1).
Notice, we also have the following relationships:
Z T
(3.2) Vt (ρ) = E P ET − Et − εE
u dXu |Ft
t
and
J0 = C0 (ρ)
= V0 (ρ) + E0
Z T
P
= E ET − εE
u dXu |F0
0
4. Application to a Life Insurance Contract When the (H)-Hypothesis Holds.

In this section, we apply the (extended) theory of local risk-minimization to the payments
described in Section 2.3.1, assuming the (H)-hypothesis holds. More precisely, we study their
pseudo locally risk-minimizing strategies.
In Section 5.1, we first give some preliminary results and in Section 2.1, we briefly study
the impact of the enlargement of filtration on the financial market. In Section 4.3, we first
study the minimal martingale measure and its properties. Then, assuming the continuity of
Xt , we derive the forms of the pseudo locally risk-minimizing strategies by finding the GKW
decomposition of our payments under this minimal martingale measure. In Section 4.4, we
remove the continuity of Xt and give the forms of the pseudo locally risk-minimizing strategies
by directly finding the Föllmer-Schweizer decomposition.
Our main conclusion is that finding the pseudo locally risk-minimizing strategies of any
of these payments comes down to finding the pseudo locally risk-minimizing strategies of a
properly modified purely financial claims.
Throughout this section, we assume the following condition holds.
Condition 4.1. The (H)-hypothesis holds under P between F and G.
An alternative definition of the (H)-Hypothesis often appears in the literature.
Definition 4.2. We say the (H)-hypothesis holds under P between F and G if and only
if: for every t, F∞ and Gt are conditionally independent with respect to Ft .
See Brémaud and Yor [33] for a proof of the equivalence between this definition and Defi-
nition 2.4.
132 Application to a Life Insurance Contract When the (H)-Hypothesis Holds.
4.1. Preliminary results. Before studying the Föllmer-Schweizer decomposition of our

payments or their Galtchouk-Kunita-Watanabe decomposition under the minimal martingale
measure, we introduce some preliminary results.
Proposition 4.3. Let mt and nt be two F-adapted processes such that hm· , n· iP,F t exists.
P,G P,F
Then, if the (H)-hypothesis holds under P , hm· , n· it and hm· , n· it are P -indistinguishable.
Proof. By definition, hm· , n· iP,F
t is the (P, F)-compensator of [m· , n· ]t , i.e. [m· n· ]−hm· , n· iP,F
t
is a (P, F)-local martingale. Thanks to the (H)-hypothesis, [m· n· ] − hm· , n· iP,F t is thus also a
(P, G)-local martingale. By the uniqueness of the compensator, hm· , n· iP,F t is thus also the
(P, G)-compensator of [m· n· ]t .
Also, Condition 4.1 ensures all (P, F)-strongly orthogonal martingales will also be (P, G)-
strongly orthogonal martingales.
Corollary 4.4. Let mt and nt be two square integrable (P, F)-martingales strongly (P, F)-
orthogonal. Then mt and nt are (P, G)-strongly orthogonal.
Proof. (P, F)-strong orthogonality is equivalent to hm· , n· iP,F
t = 0. Since under the (H)-
P,F P,G
hypothesis, hm· , n· it = hm· , n· it , we thus also have the (P, G)-strong orthogonality.
In our setting, i.e. when a filtration F is enlarged with the observation of a random time
τ , we have also the following equivalence:
Lemma 4.5. The (H)-hypothesis under P is equivalent to
P (τ ≤ t |F∞ ) = P (τ ≤ t |Ft )
The corollary of this lemma is that under the (H)-hypothesis, Ft and thus Γt are increasing
and accordingly, are of finite variation. Recall that the Doob-Meyer decomposition of the
submartingale Ft is written Ft = Yt + Dt . As far as Yt is concerned, we have the following
lemma:
Lemma 4.6. If Conditions 4.1 and 2.3 hold then the (P, F)-martingale Yt is equal to 0 for
all t ≥ 0 P -a.s.
Proof. Conditions (4.1) and (2.3) imply Ft is increasing and continuous. Ft is thus a
predictable finite variation process. By the uniqueness of the Doob-Meyer decomposition, it
leads to Ft = Dt for all t ≥ 0 P -a.s.
As far as the (P, G)-compensator Λt of Ht is concerned, we have the following lemma.
Lemma 4.7. If Γt is continuous and increasing, the (P, G)-compensator Λt of Ht is given
by Λt = Γt∧τ .
Proof. This proof can be found in Jeanblanc and Rutkowski [58] or in Bielecki and
Rutkowski [23]. We give it here for the sake of completeness. Injecting Dt = Ft in Equa-
tion (2.1), we have:
Z t∧τ
eΓu d 1 − e−Γu

Λt =
0
Z t∧τ
= dΓu
0

Under these assumptions, the compensated process Mt can be written Mt = Ht − Γt∧τ . We
will also need the following lemma.
Lemma 4.8. If the (P, F)-hazard process Γt is continuous and of finite variation then the
(P, G)-martingale Lt := 1{τ >t} eΓt can be written as
Z t
Lt = 1 − eΓu dMu
0
4.2. The financial market. We now study how the enlargement of filtration affects the
financial market. Fortunately, Condition 4.1 ensures this enlargement does not modify the
main structure of the financial market.
2 (P, G) and its (P, G)-canonical decomposition is still given by
Proposition 4.9. Xt ∈ Sloc
X t = X 0 + N t + At .
Proof. Straightforward. By definition of the (H)-hypothesis, the (P, F)-local martingale
Nt is a (P, G)-local martingale. We even have Nt ∈ M2loc (P, G) since square integrability does
not depend on the filtration. At is still a finite variation process with locally square integrable
variation, since these properties do not depend on the filtration.
The following proposition is essential.
Proposition 4.10. If the (H)-hypothesis holds for P and if the (F, P )-hazard process Γt
is continuous then ∆Uτ = 0 a.s. for any F-adapted càdlàg process Ut .
Proof. See the previous chapter.
4.3. The minimal martingale measure approach. We saw in Section 3 that under
some further assumptions (for example when Xt is continuous) we can find the Föllmer-
Schweizer decomposition of a contingent claim by finding its Galtchouk-Kunita-Watanabe de-
composition under the minimal martingale measure denoted by P̂ . In this section, we use this
approach to find the pseudo-locally risk-minimizing strategies. We first give some preliminary
results on the minimal martingale measure in Section 4.3.1. Then in the subsequent sections,
we study the locally risk-minimizing strategies of the different payments. From now on, we
assume the minimal martingale measure P̂ exists for the financial market restricted to the
filtration F.
4.3.1. The minimal martingale measure and the enlargement of the filtration. In the first
proposition, we study the impact of the enlargement of the filtration on the minimal martingale
measure when the (H)-hypothesis holds under P between the filtrations F and G. In general,
the minimal martingale measure depends on the choice of the filtration. When they exist, we
denote P̂ F and P̂ G the minimal martingale measures for, respectively, the filtration F and G.
G
dP̂ F dP̂
We denote by ẐtF := dP |Ft (resp. ẐtG := dP |Gt ) the restriction on the filtration F (resp. G) of
the Radon-Nikodym derivative of P̂ F (resp. P̂ G ) with respect to P .
Proposition 4.11. (Invariance of the minimal martingale measure P̂ ) Let us assume the
minimal martingale measure P̂ F exists. If the (H)-hypothesis holds under P , then the minimal
martingale measure for G, P̂ G , exists and ẐtG = ẐtF , P -a.s. ∀t.
R
·
Proof. If P̂ F exists then its Radon-Nikodym derivative is given by ẐtF := Et − 0 λ̂u dNu =
Rt F dN (see Ansel and Stricker [7, 8] or Schweizer [77]). By the (H)-hypothesis under
1− 0 λ̂u Ẑu− u
P , the (P, F)-local martingale Nt is also a (P, G)-local martingale and its still represents the
martingale part of the canonical decomposition of Xt . Accordingly, ẐtF is a (P, G)-local mar-
tingale. Since by assumption, it is square integrable, ẐtF is even a uniformly square integrable
(P, G)-martingale. By definition, we thus have ẐtG = ẐtF .
When the (H)-hypothesis holds, the minimal martingale measures are indistinguishable
and we can thus talk about the minimal martingale measure P̂ . In the next proposition, we
study the hazard process under the measure P̂ .
Proposition 4.12. (Invariance of the hazard process) Let Γt be the (P, F)-hazard process
of τ , and Γ̂t be the (P̂ , F)-hazard process of τ . If the (H)-hypothesis holds under P , we have
Γt = Γ̂t ∀t P -a.s.
Proof. By Bayes rule, we have:
h i
E P Ẑ∞
F 1
{τ >t} |Ft
E P̂ 1{τ >t} |Ft =

ẐtF
h i
EP E P 1{τ >t} |Ft
F |F

Ẑ∞ t
=
ẐtF
= E P 1{τ >t} |Ft

In the second equality, we use Definition 4.2 of the (H)-hypothesis. By definition of the hazard
process, we thus have that Γ̂t = Γt ∀t P -a.s.
Notice it does not mean the probabilities of surrender

−Γ −Γ h i are equal under P and P̂ . Indeed,
P P̂ P̂ − Γ̂
we have P (τ > t) = E e t 6= E e t = E e t = P̂ (τ > t).
Remark 4.13. We know that if the hazard process is continuous then Γt∧τ is equal to the
(P, G)-compensator Λt of τ . So here, Γt∧τ is also equal to the (P̂ , G)-compensator Λ̂t of τ .
Proposition 4.14. (Invariance of the (H)-hypothesis) If the (H)-hypothesis holds under
P between the filtration F and G then the (H)-hypothesis holds under the minimal martingale
measure P̂ .
Proof. To show that the (H)-hypothesis holds under P̂ , we have to show that any (P̂ , F)-
local martingale is also a (P̂ , G)-local martingale. For any (P̂ , F)-local martingale Rt , we have
that ẐtF Rt is (P, F)-local martingale. Thanks to the (H)-hypothesis under P , ẐtF Rt is also a
G
(P, G)-local martingale. Since according to Proposition 4.11, ẐtF = ẐtG then Ẑt Rt is also a
(P, G)-local martingale. Since ẐtG is the G-density of P̂ , Rt is a (P̂ , G)-local martingale.
We now study the pseudo-locally risk-minimizing strategies for the payments described in
Section 2.3.1. We assume the minimal martingale measure under the filtration F exists.
4.3.2. Payment at the term of the contract. We want to find the pseudo-locally risk-minimizing
strategy of the payment process5 1{T =t} 1{τ >T } BgTT where we assume gT is FT -adapted and
square integrable with respect to P . The following proposition gives the Föllmer-Schweizer
decomposition of this payment.
Proposition 4.15. Let us assume Conditions 2.3 and 4.1 hold.h If the (P̂ , F)-Galtchouk-
i
Kunita-Watanabe decomposition of the (P̂ , F)-martingale Ûtg := E P̂ BgTT e−ΓT |Ft is written
Z t
gT −ΓT Ug
E P̂
e |Ft = Û0g + φ̂u dXu + L̂t
BT 0
then the (P, G)-Föllmer-Schweizer decomposition of 1{τ >T } BgTT is given by

Z T
gT g
1{τ >T } = Û0 + 1{τ >u−} eΓu φ̂u dXu + L̂gT
BT 0
where
Z t Z t
g g
L̂gt = 1{τ >u−} e Γu
dL̂U
t − U
eΓu V̂u− dMu
0 0
and
g
V̂tU = Ûtg
Proof. According to Theorem 3.14, we know that when Xt is continuous, the G-Föllmer-
Schweizer decomposition of 1{τ >T } BgTT under the measure P , is given by the G-Galtchouk-
h i
Kunita-Watanabe decomposition under the minimal martingale measure P̂ of E P̂ 1{τ >T } BgTT |Gt .
5Since we only have a single payoff at the date T , we do not really need here the extension of the local
risk-minimization theory to payment processes described above.
Since the (P̂ F , F)-hazard process of τ is Γt , we have

gT gT
E P̂ 1{τ >T } |Gt = 1{τ >t} eΓt E P̂ e−ΓT |Ft
BT BT
h i
Let us define Ûtg := E P̂ e−ΓT BgTT Ft . By hypothesis, its (P̂ , F)-Galtchouk-Kunita-Watanabe

is given by
Z t
g g g
Ût = Û0 + φ̂u dXu + L̂U
t
0
Using the integration by parts formula, we can write

gT
E P̂ 1{τ >T } |Gt = 1{τ >t} eΓt Ûtg
BT
Z t Z t h i
g Γu
= 1{τ >u−} eΓu dÛug − Ûu− e dMu + Û·g , 1{τ >·} eΓ·
0 0 t
where Mu = 1{τ ≤u} − Γu∧τ . Thanks to the (H)-hypothesis, we know Γt is increasing and
Lt = 1{τ >} eΓ is thus a finite variation process. It implies
g
h i
= Û0g L0 +
X
Û·g , 1{τ >} eΓ ∆Ûs ∆Ls
t
0<s≤t
Since Γ̂t is continuous, thanks to Proposition 4.10, there is no discontinuity in Ûtg at time τ ,
the square bracket is thus equal to Û0g L0 = Û0g . We thus have
Z t
gT g
P̂
E 1{τ >T } |Gt = Û0 + 1{τ >u−} eΓu φ̂u dXu + L̂gt
BT 0
g Rt Γ U g Rt Γ g
where L̂t = 0 1{τ >u−} e u dL̂u − 0 e u Ûu− dMu .
h i
Let us now prove this decomposition is indeed the (P̂ , G)-GKW decomposition of E P̂ 1{τ >T } BgTT |Gt .
First, notice that the (P, G)-martingale Mu is a (P̂ , G)-martingale thanks to Remark 4.13 and
g
also that the (P̂ , F)-martingales Xt and L̂U t are (P̂ , G)-martingales thanks to Proposition 4.14.
D EP̂ ,G
We still have to prove X· , L̂g· = 0. We have
D EP̂ ,G Z t D EP̂ ,G Z t
Ug Ug
g
X· , L̂· = Γu
1{τ >u−} e d X· , L̂· − eΓu V̂u− d hX· , M· iP̂ ,G
0 0
= 0
We use the fact hX· , M· iP̂ ,G = 0 since M is a finite variation process and Xt is continuous.
Moreover, according to Proposition 4.3 and since the (H)-hypothesis holds under P̂ thanks to
g P̂ ,G g P̂ ,F
D E D E
Proposition 4.14, we know that X· , L̂U · = X· , L̂U
· = 0.
We can now give the form of the pseudo-locally risk-minimizing portfolio of 1{τ >T } BgTT .
Corollary 4.16. If the Conditions 2.3 and 4.1 hold, then the discounted value process
V̂tg (ρ̂∗ ) of the locally risk-minimizing portfolio for the claim 1{τ >T } BgTT is given by:
g
V̂tg (ρ̂∗ ) = 1{τ >t} eΓt V̂tU (ρ̂)
and, on {τ > t}, the locally risk-minimizing strategy ρ̂∗ = (φ̂∗ , η̂ ∗ ) is given by:

ρ̂∗t = eΓt φ̂t , η̂t
g

where V̂tU (ρ̂) and ρ̂ = φ̂, η̂ are respectively the discounted value process and the strategy of
gT −ΓT
the locally risk-minimizing portfolio of the purely financial claim BT e .
The cost process Ctg (ρ̂∗ ) is given by
Ctg (ρ̂∗ ) = C0g (ρ̂∗ ) + L̂gt

where C0g (ρ̂∗ ) = Û0g .
Proof. We simply have

gT
V̂tg (ρ̂∗ ) P̂
= E 1{τ >T } |Gt
BT

Γt P̂ −ΓT gT
= 1{τ >t} e E e |Ft
BT
g
= 1{τ >t} eΓt V̂tU (ρ̂)
where in third equality, we use the fact the restriction of P̂ on F and G are indistinguishable.
∗
According to the previous theorem, on {τ > t}, we obviously have φ̂t = eΓt φ̂t . As far as η̂t∗ is
concerned, on {τ > t}, we have
η̂t∗ = V̂tg (ρ∗ ) − φ̂∗t Xt
g
= eΓt V̂tU (ρ̂) − eΓt φ̂t Xt
= eΓt η̂t

4.3.3. Payment at the surrender time. We now study R Tthe pseudo locally risk-minimizing
Ru
strategies of the discounted cumulative payment process 0 Bu dHu .
Proposition 4.17. Let us assume Conditions 2.3 and 4.1 hold.h If the (P̂ , F)-Galtchouk-
i
R T Ru
Kunita-Watanabe decomposition of the (P̂ , F)-martingale ÛtR := E P̂ 0 B u
dFu |Ft is written
Z T Z t
P̂ Ru UR
E dFu |Ft = Û0R + ψ̂u dXu + L̂t
0 Bu 0
RTRu
then the (P, G)-Föllmer-Schweizer decomposition of 0 Bu dHu is given by
Z T Z T
Ru
dHu = Û0R + 1{τ >u−} eΓu ψ̂u dXu + L̂R
T
0 Bu 0
where
Z t Z t
R Ru Γu U R
L̂R
T = Γu
1{τ >u−} e dL̂U
t + − e V̂u− dMu
0 0 Bu
and
Z T
R Ru
V̂tU = E P̂
dFu |Ft
t Bu
R T Ru 3.14, we know that when Xt is continuous, the G-Föllmer-

Proof. According to Theorem
Schweizer decomposition of 0 Bu dHu under the measure P , is given by the G-Galtchouk-
hR i
P̂ T Ru
Kunita-Watanabe decomposition under the minimal martingale measure P̂ of E 0 Bu dHu |Gt .
We have
hR i
P̂ T Ru
Z T
Ru
Z t
Ru E t Bu dH u |F t
E P̂ dHu |Gt = dHu + 1{τ >t}
0 Bu 0 Bu

E P̂ 1{τ >t} |Ft
Z t Z T
Ru Γt P̂ Ru
= dHu + 1{τ >t} e E dFu |Ft
0 Bu t Bu
Z t Z t
Ru Γt Ru
= dHu − 1{τ >t} e dFu
0 Bu 0 Bu
Z T
Ru
+ 1{τ >t} eΓt E P̂ dFu |Ft
0 Bu
In the second equality, we use the fact that

h R according toi (4.12), the (P̂ , F)-hazard process of τ
R P̂ T Ru
is equal to Γt . Let us define Ût := E 0 Bu dFu Ft . By hypothesis, its (P̂ , F)-Galtchouk-

Kunita-Watanabe is given by
Z t
R
R
Ût = Û0 + R
ψˆu dXu + L̂U
t
0
Using the integration by parts formula and the continuity of Ft , we can show that
Z t Z t Z t Z u
Γt Ru Ru Rs
1{τ >t} e dFu = dΓu∧τ − dFs eΓu dMu
0 Bu 0 Bu 0 0 Bs −
We also have
Z t t
Z h i
Γt
1{τ >t} e ÛtR = 1{τ >t−} e Γt
dÛuR
− R Γu
Ûu− e dMu + 1{τ >·} eΓ· , Û·R
0 0 t
Z t Z t
= Û0R + 1{τ >t−} eΓt dÛuR − R Γu
Ûu− e dMu
0 0
h i
since 1{τ >·} eΓ· , Û·R = Û0R L0 + R = Û0R . We thus obtain
P
0<s≤t ∆Ûs ∆Ls
t
T Z t Z t
Z
P̂ Ru Ru
dHu Gt = Û0R + 1{τ >t−} eΓt dÛuR −

E dMu
0 Bu 0 0 Bu
Z t Z u
R Rs
− Ûu− − dFs eΓu dMu
0 0 B s −

t Rs R
Since ÛtR − 0 B dFs = V̂tU , we eventually get
R
s
Z T Z t
P̂ Ru R
E dHu |Gt = Û0 + 1{τ >u−} eΓu ψ̂u dXu + L̂R
t
0 Bu 0
Rt
R t Ru
where L̂R Γu UR Γu U R dM .
t = 0 1{τ >u−} e dL̂u + 0 Bu − e V̂u− u
Following the same arguments than in Proposition 4.15, hR we can easily i prove this decompo-
P̂ T Ru
sition is indeed the (P̂ , G)-GKW decomposition of E 0 Bu dHu |Gt .
R T Ru
We can easily find the form of the pseudo-locally risk-minimizing portfolio of 0 B u
dHu as
in Corollary 4.16. R
T
The payment 0 1{τ >u} dCu can be studied in a similar fashion. This is left to the reader
or see Proposition 4.20.
4.4. The Föllmer-Schweizer decomposition approach. When the (H)-hypothesis

holds, we can actually even remove the assumption that Xt is continuous, by studying di-
rectly the Föllmer-Schweizer decomposition. We only study the second and third payments.
This is left to the reader.
4.4.1. Payment at the surrender time. We have the following proposition.
Proposition 4.18. If the Conditions 2.3 and 4.1 hold and R T Rif, for Ru , F-predictable and
P -square integrable, the Föllmer-Schweizer decomposition of 0 Bu dFu exists and is written
u
Z T Z T
Ru UR R
dFu = J0 + ψu dXu + LU T
0 Bu 0
R T Ru
then the Föllmer-Schweizer decomposition of 0 Bu dHu is given by
Z T Z T
Ru UR
dHu = J0 + 1{τ >u−} eΓu ψu dXu + LR T
0 Bu 0
where
Z T Z T
R Ru U R Γu
LR
T = 1{τ >u−} e Γu
dLU
u + − Vu− e dMu
0 0 Bu
and
T Z T
Z
R Ru
VtU P P

= E dFu Ft − E
ψt dXu Ft
t Bu t
hR i
T Ru
Proof. Let us define UtR := E P 0 B dFu Ft . We have

u
Z t Z T
R UR UR P

Ut = J0 + ψu dXu + Lt + E ψu dXu Ft
0 t
R R
= JtU+ WtU
Rt hR i
R R R R T
where JtU := J0U U U
+ 0 ψu dXu + Lt and Wt := E P
t ψu dXu Ft . On the other hand,

we can write
Z T Z t Z T

P Ru Ru Ru
dHu + E P

E dHu Gt =
dHu Gt
0 Bu 0 Bu t Bu
Z t Z t Z T
Ru Γt Ru Γt P Ru
= dHu − 1{τ >t} e dFu + 1{τ >t} e E dFu Ft
0 Bu 0 Bu 0 Bu
Z t Z t
Ru Ru R R
= dHu − 1{τ >t} eΓt dFu + 1{τ >t} eΓt JtU + 1{τ >t} eΓt WtU
0 Bu 0 Bu
Using the same arguments than in Proposition 4.17, we can write that
T Z t Z t
Z
P Ru UR Γt UR Ru
E dHu Gt = J0 +
1{τ >t−} e dJt − dMu
0 Bu 0 0 Bu
Z t Z u
UR Rs R
− Ju− − dFs eΓu dMu + 1{τ >t} eΓt WtU
0 0 B s −
R R t Rs R R
Since JtU − 0 Bs dFs = VtU and since WTU = 0 then for t = T , we can write the last
equation as
Z T Z T Z t
Ru R R Ru U R Γu
dHu = J0U + 1{τ >u−} e Γu
dJuU + − Vu− e dMu
0 Bu 0 0 Bu
R
Replacing the expression of we have JU ,
Z T Z T
Ru UR
dHu = J0 + 1{τ >u−} eΓu ψu dXu + LR
T
0 Bu 0
RT
R T Ru
where LR Γu UR U R Γu dM .
T := 0 1{τ >u−} e dLu + 0 Bu − Vu− e u
RWe still have to prove this decomposition is indeed the Föllmer-Schweizer decomposition
T Ru
of 0 Bu dHu . Since, thanks to the (H)-hypothesis, the canonical decomposition of Xt =
X0 + Nt + At does not vary with the enlargement of filtration, we just have to show that Lgt is
P,G
strongly (P, G)-orthogonal to Nt or equivalently that hLg· , N· it = 0. We have
Z · Z · P,G
g P,G Γu Ug U g Γu
hL· , N· it = 1{τ >u−} e dLu − Vu− e dMu , N·
0 0 t
Z t P,G
Z t
g g
1{τ >u−} eΓu d LU U
d hL· , N· iP,G

= · , N· u
− Vu− u
0 0
On the one hand, we have hM· , N· iP,G

= 0 since we know [M· , N· ] = 0. On the other
t
hand, thanks to the definition of the Föllmer-Schweizer decomposition, we know the (P, F)-
g
g P,F
martingales LUt and Nt are strongly (P, F)-orthogonal so that LU
· , N· t = 0. Thanks to

Ug P,G
Ug P,F
Proposition 4.3, we know that L· , N· t = L· , N· t = 0 for all t ≥ 0. We thus have
P,G
hLg· , N· it = 0.
RT Ru
We can now give the form of the pseudo-locally risk-minimizing portfolio of 0 Bu dHu .
Corollary 4.19. If the Conditions 2.3 and 4.1 hold then the discounted value VtR (ρ∗ ) of
R T Ru
the locally risk-minimizing portfolio for the payment 0 B u
dHu are respectively given by
R
VtR (ρ∗ ) = 1{τ >t} eΓt VtU (ρ)
and, on {τ > t}, the locally risk-minimizing strategy ρ∗ is given by:
ρ∗t = eΓt (ψt , ηt )

R
where VtU (ρ) and ρt = (ψt , ηt ) are respectively the discounted value and the strategy of the
R T Ru
locally risk-minimizing portfolio of the purely financial claim 0 B u
dFu .
R ∗
The cost process Ct (ρ ) is given by
CtR (ρ∗ ) = C0R (ρ∗ ) + LR
t
R
where C0R (ρ∗ ) = J0U .
Proof. According to Equation (3.2), we have
Z T Z T
R ∗ P Ru Γu
Vt (ρ ) = E dHu − 1{τ >u−} e ψu dXu |Gt
t Bu t
Z T Z T
Γt P Ru Γt P Γu
= 1{τ >t} e E dFu |Ft − 1{τ >t} e E 1{τ >u−} e ψu dAu |Ft
t Bu t
Z T Z T
Γt P Ru Γt P
= 1{τ >t} e E dFu |Ft − 1{τ >t} e E ψu dXu |Ft
t Bu t
R
= 1{τ >t} eΓt VtU
In the third equality, we use the predictable projection theorem and the continuity of Γt . For
the rest of the proof, see Corollary 4.16.
4.4.2. Payments up to the surrender time. We now study

R T the pseudo-locally risk-minimizing
strategies of the discounted cumulative payment process 0 1{τ >u} B1u dCu where Ct is assumed
to be an F-adapted square integrable process with respect to P .
Proposition 4.20. If the Conditions 2.3 and 4.1 hold and if, for Ct F-adapted and P -
R T −Γu
square integrable, the Föllmer-Schweizer decomposition of 0 eBu dCu exists and is written
Z T −Γu Z T
e UC C
dCu = J0 + νu dXu + LU T
0 Bu 0
RT
then the Föllmer-Schweizer decomposition of 0 1{τ >u} B1u dCu is given by
Z T Z T
1 UC
1{τ >u} dCu = J0 + 1{τ >u−} eΓu νu dXu + LC
T
0 Bu 0
where
Z T Z T
C C
LC
T = 1{τ >u−} eΓu dLU
u −
U
Vu− eΓu dMu
0 0
and
T Z T

Z
C 1 −Γu
VtU P P

= E e dCu Ft − E
νt dXu Ft
t Bu t
hR i
T
Proof. Let us define UtC := E P t B1u e−Γu dCu Ft . We have

t T
Z Z
C C
UtC J0U LU P

= + νu dXu + t +E νu dXu Ft
0 t
C C
= JtU + WtU
UC
Rt hR i
C C C T
where JtU:= J0U + νu dXu + LU and Wt := EP νu dXu Ft . On the other hand,

0 t t
we can write:
T
e−Γu
Z Z t Z t
P 1 1 Γt
E 1{τ >u} dCu Gt =
1{τ >u} dCu − 1{τ >t} e dCu
0 Bu 0 Bu 0 Bu
Z T −Γu
e
+ 1{τ >t} eΓt E P

dCu Ft
0 Bu
e−Γu
Z t Z t
1 Γt
= 1{τ >u} dCu − 1{τ >t} e dCu
0 Bu 0 Bu
C
+ 1{τ >t} eΓt JtU + 1{τ >t} eΓt WtC
Using the integration by parts formula, we can show

Z t −Γu Z t −Γu Z t Z u −Γs
Γt e Γu− e e
1{τ >t} e dCu = 1{τ >u−} e dCu − dCs eΓu dMu
0 Bu 0 B u 0 0 Bs −
Z · −Γu
e
+ 1{τ >·} eΓ , dCu
0 Bu t
Z t −Γu Z t Z u −Γs
Γu e e
= 1{τ >u} e dCu − dCs eΓu dMu
0 B u 0 0 Bs −
Z t Z t Z u −Γs
1 e
= 1{τ >u} dCu − dCs eΓu dMu
0 B u 0 0 Bs −
As in other propositions, we also have
Z t Z t
Γt C C C C
1{τ >t} e JtU = J0U + Lu− dJuU − U
Ju− eΓu dMu
0 0
We thus obtain
Z T Z t
P 1 C C
dCu Gt = J0U + 1{τ >u−} eΓu dJuU

E 1{τ >u}
0 Bu 0
Z t Z t −Γs
U C e
− Jt − dCs eΓu dMu + 1{τ >t} eΓt WtC
0 0 B s
C R −Γ
U t e s U C
Since Jt − 0 Bs dCs = Vt and since WT = 0 then for t = T , we can write
Z T Z T Z T
1 UC Γu UC U C Γu
1{τ >u} dCu = J0 + 1{τ >u−} e dJu − Vu− e dMu
0 Bu 0 0
C
Replacing the expression of JtU , we obtain
Z T Z T
1 UC
1{τ >u} dCu = J0 + 1{τ >u−} eΓu νu dXu + LC
T
0 Bu 0
RT UC
R T UC Γ
where LC Γu
T = 0 1{τ >u−} e dLu − 0 Vu− e dMu . Following the same arguments than in
u
Proposition 4.18,
R Twe can easily prove this equation corresponds to the Föllmer-Schweizer de-
composition of 0 1{τ >u} dCu .
RT
We can easily find the form of the pseudo-locally risk-minimizing portfolio of 0 1{τ >u} dCu
as in Corollary 4.19. This is left to the reader.
Notice that, when Xt is continuous, the two methods obviously leads to the same decom-
position. To see this, compare Propositions 4.17 and 4.18, for examples. Indeed, the integrand
R
ψ and the
R T martingale LU in the Föllmer Schweizer decomposition under P of the F-adapted
Ru R
process 0 B u
dFu , are respectively, the same than the integrand ψ̂ and the martingale L̂U of
the GKW decomposition under P̂ since according to Theorem 4.11, the restriction of the mini-
mal martingale P̂ (defined initially for the filtration G) corresponds to the minimal martingale
measure defined for the filtration F.
144 Application to a Life Insurance Contract When the (H)-Hypothesis Does Not Hold.
5. Application to a Life Insurance Contract When the (H)-Hypothesis Does Not

Hold.
In this section, we study the pseudo local risk-minimizing strategies of the payments de-
scribed in Section 2.3.1 when we remove the (H)-hypothesis.
In Section 5.1, we first introduce some preliminary results. In Section 2.1, we study the
impact of the enlargement of filtration on the financial market and in Section 5.3, its impact
on the minimal martingale measure. Finally, in Section 5.4, we study the pseudo locally risk-
minimizing strategies in themselves.
As in the previous section, we assume the (P, F)-hazard process Γt of τ is continuous. We
also introduce the following condition.
6
Condition 5.1. For any F-stopping time θ, P (τ = θ) = 0.
According to Blanchet-Scalliet and Jeanblanc [30], under this condition, for any (P, F)-
martingale mt , the stopped process
Z t∧τ
(5.1) m̃t∧τ = mt∧τ + eΓs d [m· , Y· ]s
0
is a (P, G)-martingale. We call m̃t∧τ the (P, G)-martingale part of mt∧τ . We keep this notation
in the rest of the chapter. In particular, here, the (P, G)-martingale part of the square integrable
(P, F)-martingale Y can be written
Z t∧τ
Ỹt∧τ = Yt∧τ + eΓs d [Y· , Y· ]s
0
5.1. Preliminary results. Before studying the Föllmer-Schweizer decomposition of our
payments, we introduce some preliminary results.
Proposition 5.2. Let mt be an (P, F)-martingale such that hm· , m· iP,F
t exists then hm· , m· iP,G
t∧τ
P,F
and hm· , m· it∧τ are P -indistinguishable .
Proof. By definition of the sharp bracket, we have that rt := [m· , m· ]t − hm· , m· iP,F
t is a
(P, F)-(local) martingale. We thus have that
Z t∧τ
r̃t∧τ = rt∧τ + eΓs d [r· , Y· ]s
0
is locally a (P, G)-martingale. Since rt is a finite variation process and Yt is continuous, we have
[r· , Y· ]t = 0≤s≤t ∆rs ∆Ys = 0. Accordingly, rt∧τ = [m· , m· ]t∧τ − hm· , m· iP,F
P
t∧τ is a (P, G)-(local)
martingale. By definition and uniqueness of the sharp bracket, we thus have hm· , m· iP,F t∧τ =
P,G
hm· , m· it∧τ P -a.s.
Proposition 5.3. Let mt be a (P, F)-martingale such that hm· , m· iP,F
t exists, then
P,G P,G
hm̃·∧τ , m̃·∧τ it and hm·∧τ , m·∧τ it are P -indistinguishable.
6Notice that in the previous section, this condition was implied by the continuity of the hazard process and
the (H)-hypothesis. See Proposition 4.10.
R t∧τ
Proof. we have [m̃·∧τ , m̃·∧τ ] = [m·∧τ , m·∧τ ] since 0 eΓs d [m· , Y· ]s is a continuous finite
variation process. By the definition of the sharp bracket and its uniqueness, we have the
result.
Corollary 5.4. Let mt be a (P, F)-martingale such that hm· , m· iP,F
t exists, then
hm·∧τ , m·∧τ iP,F
t and hm̃ ·∧τ , m̃ iP,G
·∧τ t are P -indistinguishable.
Proof. Combine Proposition 5.3 and Proposition 5.2.
Corollary 5.5. Let mt and nt be two square integrable (P, F)-martingales strongly (P, F)-
orthogonal. Then m̃t∧τ and ñt∧τ are (P, G)-strongly orthogonal.
Proof. We have to show that hm̃·∧τ , ñ·∧τ iP,G

t = 0 P -a.s. for all t. Since, by hypothesis,
mt and nt are strongly (P, F)-orthogonal, we know that hm· , n· iP,F
t = 0.Thanks to the previous
corollary, we have that hm̃·∧τ , m̃·∧τ it = hm·∧τ , m·∧τ it = hm· , m· iP,F
P,G P,F
t∧τ = 0.
5.2. The financial market. In this section, we want to check that on {τ > t}, the
dynamics of the financial market is not “too much” altered by the enlargement of filtration.
First, we want to check that Xt∧τ ∈ Sloc 2 (P, G). Second, we have to study if the structure
condition still holds under the new filtration.

Proposition 5.6. Xt∧τ ∈ Sloc 2 (P, G) and its (P, G)-canonical decomposition is given by
Xt∧τ = X0 + Ñt∧τ + Ãt∧τ where

Z t∧τ
Ñt∧τ = Nt∧τ + eΓs d [N· , Y· ]s
0
Z t∧τ
Ãt∧τ = At∧τ − eΓs d [N· , Y· ]s
0
R t∧τ
Proof. Since Nt is locally a martingale, we know that Ñt∧τ = Nt∧τ + 0 eΓs d [N· , Y· ]s is
locally a (P, G)-martingale. Furthermore, since Y is continuous, [N· , Y· ]t is continuous and so
R t∧τ
is 0 eΓs d [N· , Y· ]s . Since any continuous process is also locally square integrable and since Nt
is locally square integrable, it implies Ñt∧τ is (locally) a locally square integrable martingale
and thus a locally square integrable
R t∧τ martingale.
On the other hand, since 0 eΓs d [N· , Y· ]s is continuous, its (finite) variation is also con-
tinuous and thus locally square integrable. Since the variation of At is locally square integrable,
it implies Ãt∧τ is a predictable finite variation process whose variation is locally square inte-
grable.
Before studying if the structure condition holds, we need the following proposition.
D EP,G
Proposition 5.7. Ãt∧τ Ñ· , Ñ· .
t∧τ
Proof. Since Yt is a square integrable (P, F)-martingale, its (P, F)-Galtchouk-Kunita-

Watanabe decomposition exists with respect to each Nti for i = 1, . . . s. In other words, for
each i = 1, . . . s, there exists an (P, F)-predictable process hit ∈ L2 (N i ) such that:

Z t
Yt = Y0 + hiu dNui + LY,i
t
0
for each t and such that, for each i = 1, . . . s, LY,i 2

t ∈M0 (P, F) and is (P, F)-strongly orthogonal
P,F t P,F
to Nti . We thus have N·i , Y· t = 0 hiu d N·i , N·i u .

R
On the other hand, by definition of Ãi t∧τ , we can write

Z t∧τ
Ait∧τ eΓu d N·i , Y· u

Ãi t∧τ = −
0
Z t∧τ Z t∧τ
P,F P,F
αui d N·i , N·i u − eΓu d N·i , Y· u

=
0 0
Z t∧τ
P,F
αui − eΓs his d N·i , N·i s

=
0
P,F
P,G
Thanks to Proposition 5.2, we know that, on {τ > t}, N·i , N·i t = N·i , N·i t . So eventu-

ally, we have
Z t∧τ
P,G
αui − eΓs his d N·i , N·i s

i
Ã t∧τ =
0

P,G
P,F P,G
Since N·i , N·i t∧τ = N·i , N·i t∧τ , we can easily see that N i , N j t∧τ B̄t∧τ for all i, j =

P,G
1, . . . s and that the predictable density of N i , N j t∧τ with respect to B̄t∧τ is equal to σtij i.e.

Z t∧τ

i j P,G
N , N t∧τ = σuij dB̄u
0
P -a.s. for all t and all i, j = 1, . . . s. We can then write

Z t∧τ
i
Ãt∧τ = γ̃ui dB̄u
0
where γ̃ i = α̃i σ ii P -a.s. for all t and all i = 1, . . . s. We denote by γ̃ the s-dimensional vector
whose elements are γ˜i for i = 1, . . . s.
We can now show that the structure condition still holds after the enlargement of filtration
(at least on {τ > t}).
Proposition 5.8. If Xt satisfies the structure condition for F and if Conditions 2.3 and
5.1 hold, then Xt∧τ satisfies the structure condition for G i.e. there exists an Rs -valued G
predictable process λ̃ ∈ L2loc (Ñ·∧τ ) such that, on {τ > t},
σt λ̃t = γ̃t
Proof. Since Yt is a square integrable martingale, we can find its (P, F)-Galtchouk-Kunita-
Watanabe decomposition with respect to the s-dimensional vector Nt , i.e. there exists an
s-dimensional predictable process gt ∈ L2 (N ) such that
Z t
0
Yt = Y0 + gu dNu + LYt
0
s Z
X t
= Y0 + guj dNuj + LYt
j=1 0
P -a.s. for all t, and such that LYt ∈ M20 (P, F) and is (P, F)-strongly orthogonal to I 2 (N ), the
stable subspace generated by N . We can thus write for each i = 1, . . . s:
s Z t
i P,F
X P,F
P,F
guj d N·j , N·i t + LY· , N·i t

Y· , N· t =
j=1 0
s Z t
X P,F
guj d N·j , N·i t

=
j=1 0
s
!
Z t X
= guj σuji dB̄u
0 i=1
On the other hand, we already saw Yt could be decomposed in the following way for each
i = 1, . . . s:
Z t
Yt = Y0 + hiu dNui + LY,i
t
0
where ∈hit L2 (N i )
and LY,i
t ∈ M20 (P, F) and is (P, F)-strongly orthogonal to Nti . We can thus
write for each i = 1, . . . s:
P,F
Z t P,F
Y· , N·i t hiu d N·i , N·i t

=
0
Z t
= hiu σ ii dB̄t
0
i
Accordingly, if we denote by h̃t the s−predictable process whose elements are h̃t = hit σtii for
i = 1, . . . s, we can write h̃t = σt gt .
Finally, on {τ > t}, since γ̃ti = α̃i σ ii = αti σtii − eΓt hit σtii , we have
γ̃ = σt λ̂t − eΓt σt gt
= σt λ̃t
where λ̃t = λ̂t − eΓt gt .
It remains to show that λ̃ ∈ L2loc (Ñ·∧τ ). By hypothesis λ̂t ∈ L2loc (N ). By definition of
the GKW decomposition, gt ∈ L2loc (N ) and since eΓt is locally bounded, we can even say
that eΓt gt ∈ L2loc (N ). Accordingly, λ̃t ∈ L2loc (N ). Now, thanks to Corollary 5.4, we know
P,F D EP,G
i , Ni i , Ñ i
N·∧τ ·∧τ t
= Ñ·∧τ ·∧τ for each i = 1, . . . s. We can then conclude that λ̃t ∈
t
L2loc (Ñ·∧τ ).
5.3. The minimal martingale measure. As in Section 4.3, we study the impact of the
enlargement of filtration on the minimal martingale measure. When they exist, we denote by P̂ F
and P̂ G the minimal martingale measures for, respectively, the filtration F and G. We denote
P̂ F
by ẐtF := ddP |Ft
, the restriction, to the filtration F, of the Radon-Nikodym derivative of P̂ F
G
dP̂
with respect to P and by ẐtG := dP |Gt the restriction to the filtration G, of the Radon-Nikodym
derivative of P̂ G with respect to P . When the (H)-hypothesis does not hold, these densities do
not necessarily coincide. We have the following candidate for the minimal martingale measure
with respect to the enlarged filtration G.
Proposition 5.9. When the minimal martingale measure P̂ G exists, its density is given,
on {τ > t}, by
Z ·
G 0
Ẑt∧τ = Et − λ̃u dÑu∧τ
0
Furthermore, if Xt is continuous, we have
F
Ẑt∧τ
G
Ẑt∧τ = 0
R·
Et∧τ − 0 λu − λ̃u dNu
Proof. Since Xt∧τ satisfies the structure condition, we directly know that if the minimal
martingale measure exists, then on {τ > t}, we have
Z ·
G 0
Ẑt = Et − λ̃u dÑu
0
As far as the second equation is concerned, when Xt is continuous, we can write
Rt P,G Rt 0
1
λ̃0u dhÑ·∧τ ,Ñ·∧τ iu λ̃u −
G
Ẑt∧τ = e− 2 0 0 λ̃u dÑu∧τ
Rt 0 Rt 0
1 0 P,F
λ̃u dNu∧τ − 0t∧τ eΓu λ̃u dhN,Y iP,F
R
= e− 2 R0 λ̃u dhN·∧τ ,N·∧τ iu λ̃u − 0 u
− 21 0t λ̂0u dhN·∧τ ,N·∧τ iP,F 1 t Γu 0

R P,F Γu Rt Γ 0 P,F
λ̂ −
u 2 0 e g dhN ·∧τ ,N i
·∧τ u e gu + 0 e u gu dhN·∧τ ,N·∧τ iu λ̂u
= e u u
Rt 0 Rt Γ 0 R t∧τ Γ 0 P,F R t∧τ Γ2 0 P,F
e− 0 λ̂u dNu∧τ + 0 e gu dNu∧τ − 0 e λ̂u dhN,Y iu + 0 e u gu dhN,Y iu
u u
1 t 0
R P,F Rt 0 1 t Γu 0
R P,F Γu Rt Γ 0 P,F
e− 2 0 λ̂u dhN·∧τ ,N·∧τ iu λ̂u − 0 λ̂u dNu∧τ − 2 0 e gu dhN·∧τ ,N·∧τ iu e gu + 0 e gu dhN·∧τ ,N·∧τ iu λ̂u
u
=
Rt Γ 0 R t∧τ Γ 0 P,F R t∧τ Γ2 0 P,F
e+ 0 e gu dNu∧τ − 0 e λ̂u dhN,N iu gu + 0 e u gu dhN,N iu gu
u u
1 t Γu 0
R P,F Γu R t Γu 0
= F
Ẑt∧τ e 2 0 e gu dhN·∧τ ,N·∧τ iu e gu + 0 e gu dNu∧τ
D EP,G
In the second equality, we use the fact that hN·∧τ , N·∧τ iP,F
t = Ñ ·∧τ , Ñ ·∧τ and in the
Rt 0 t
fourth, the fact that Yt can be written as Yt = Y0 + 0 gu dNu + LYt . Eventually, we get the
result if we notice
Z · Rt Γ 0
1 t Γu 0 P,F Γu
Γ 0
R
= e− 2 0 e gu dhN·∧τ ,N·∧τ iu e gu − 0 e gu dNu∧τ
u
Et − e gu dNu∧τ
0
From now on, we assume Ẑt∧τG ∈ L2 (P ) so that, on {τ > t}, Ẑ G is indeed the density of
t
the minimal martingale measure P̂ G . An important point to notice is that the restriction on F
of this minimal martingale measure is not equal to ẐtF . The minimal martingale measure P̂ G
is thus not simply an extension of P̂ F but a genuine different martingale measure.
We now introduce the following definition.
Definition 5.10. When Xt is continuous, we define the F-adapted process Υt by

Z · 0
Υt
e = Et − λu − λ̃u dNu
0

R· 0
Notice that since Et − 0 λu − λ̃u dNu > 0 when Xt is continuous, the process Υt is
here well-defined.
5.4. The pseudo locally risk-minimizing strategies. We now study the pseudo-locally
risk-minimizing strategies of our insurance payments when Xt is continuous. We still need the
following result.
Proposition 5.11. When Γt and Xt are continuous, the process 1{τ >t} e(Γt +Υt ) can be
written as
Z t Z t
2
(Γt +Υt )
1{τ >t} e = 1− e (Γu +Υu )
dMu + 1{τ >u−} e(Γu +Υu ) dỸu∧τ
0 0
Z t
2 0
(5.2) − 1{τ >u−} e(Γu +Υu ) gu dÑu∧τ
0
or
Z t Z t
2
(5.3) 1{τ >t} e(Γt +Υt ) = 1 − e(Γu +Υu ) dMu + 1{τ >u−} e(Γu +Υu ) dL̃Yu∧τ
0 0
R t∧τ
where L̃Yt∧τ = LYt∧τ + eΓu d Y· , LY· u .

0
Proof. On one hand, we can prove (or see Blanchet-Scalliet) that, when Γt is continuous,
we have
Z t Z t
Γt Γu 2
1{τ >t} e = 1− e dMu + 1{τ >u−} eΓu dỸu
0 0
On the other hand, by definition, we have that
Z · 0
eΥt = Et − λu − λ̃u dNu
0
Z t
0
= 1− e(Γu +Υu ) gu dNu
0
Using the integration by parts formula, we can find Equation (5.2)
Z t Z t Z t
2 2 0
1{τ >t} e(Γt +Υt )
= 1− e (Γu +Υu )
dMu + 1{τ >u−} e(Γu +Υu ) gu dNu
1{τ >u−} e(Γu +Υu ) dỸu∧τ −
0 0 0
Z t h i
2 0
− 1{τ >u−} eΓu e(Γu +Υu ) gu d Ỹ·∧τ , N·
0 u
Z t Z t Z t
Γ2u +Υu ) 2 0
= 1− e(Γu +Υu )
dMu + 1{τ >u−} e ( dỸu∧τ − 1{τ >u−} e(Γu +Υu ) gu dNu∧τ
0 0 0
Z t Z u∧τ
2 0
− 1{τ >u−} e(Γu +Υu ) gu d eΓs d [Y· , N· ]s
0 0
Z t Z t Z t
2 2 0
= 1− eΓu +Υu dMu + 1{τ >u−} e(Γu +Υu ) dỸu∧τ − 1{τ >u−} e(Γu +Υu ) gu dÑu∧τ
0 0 0
hR i
u∧τ
where in the second equality, we use the fact 0 eΓs d [Y· , N· ]s , N = 0 since [Y· , N· ] is a
continuous finite variation process.
To find Equation (5.3), we have to notice that
Z t
0
Ỹt∧τ = Y0 + gu dÑu∧τ + L̃Yt∧τ
0
5.4.1. Payment at the term of the contract. We first need the following result.
Proposition 5.12. If we denote by P̂ G and P̂ F the minimal martingale measure with respect
to G and to F, we have the following relationship

P̂ G gT (Γt +Υt ) P̂ F −(ΓT +ΥT ) gT
E 1{τ >T } |Gt = 1{τ >t} e E e |Ft
BT BT
Proof. We have
h i

gT
E P ẐTG 1{τ >T } BgTT |Gt
P̂ G
E 1{τ >T } |Gt =
BT ẐtG
h i
E P ẐTF e−ΥT 1{τ >T } BgTT |Ft
= 1{τ >t}
ẐtG e−Γt
h i
E P ẐTF e−ΥT 1{τ >T } BgTT |Ft
= 1{τ >t}
ẐtF e−(Υt +Γt )

gT
= 1{τ >t} e(Γt +Υt ) E P̂ e−(ΓT +ΥT )
F
|Ft
BT

The Föllmer-Schweizer decomposition of 1{τ >T } BgTT is given in the following proposition.
Proposition 5.13. If, for gT , FT -adapted and P -square integrable, the

h (P̂ F , F)-Galtchouk-
i
Kunita-Watanabe decomposition of the (P̂ F , F)-martingale Utg := E P̂ BgTT e−(ΓT +ΥT ) |Ft is
F
written
Z T
P̂ F gT −(ΓT +ΥT ) g g
E e |Ft = U0 + φu dXu + LU T
BT 0
Then the (P, G)-Föllmer-Schweizer decomposition of 1{τ >T } BgTT is given by

Z T
gT
1{τ >T } = U0g + 1{τ >u−} e(Γu +Υu ) φu dXu + L̃gT
BT 0
where
Z t Z t
L˜gt = U g (Γ2u +Υu )
(Γu +Υu ) g
1{τ >u−} e dL̃U
u + 1{τ >u−} Vu− e dL̃Yu
0 0
Z t
g
U (Γu +Υu )
− Vu− e dMu
0
and
g
VtU = Utg
What is interesting here is that, again, finding the locally risk-minimizing strategies of our
contingent claims comes down to finding the locally risk-minimizing strategy of a modified
financial claims. Notice that, somewhat surprisingly, the solution involved the Galtchouk-
Kunita-Watanabe decomposition of a modified claim but under the minimal martingale measure
P̂ F i.e. under the minimal martingale measure of the non-modified financial market.
h i
1{τ >T } BgTT |Gt .
G
Proof. Our goal is to find the (P̂ G , G)-Galtchouk-Kunita-Watanabe of E P̂
According to Proposition 5.12, we have

P̂ G gT (Γt +Υt ) P̂ F −(ΓT +ΥT ) gT
E 1{τ >T } |Gt = 1{τ >t} e E e |Ft
BT BT
h i
Let us define Utg := E P̂ e−(ΓT +ΥT ) BgTT |Ft . By hypothesis, its (P̂ F , F)-Galtchouk-Kunita-
F
Watanabe decomposition with respect to Xt is given by

Z t
g
Utg = U0g + εgu dXu + LU
t
0
Using the integration by parts formula and the integral representation of 1{τ >t} e(Γt +Υt ) given
in Proposition 5.11, we have

gT
E P̂ G
1{τ >T } |Gt = 1{τ >t} e(Γt +Υt ) Utg
BT
Z t
= 1{τ >u−} e(Γu +Υu ) dUug
0
Z t Z t
g (Γu +Υu ) g (Γ2u +Υu )
− Uu− e dMu + 1{τ >u−} Uu− e dL̃Yu∧τ
h0 i 0
g (Γ· +Υ· )
+ U· , 1{τ >·} e
t
As far as the square bracket is concerned, we have
h i Z t h i
2
U·g , 1{τ >·} e(Γ· +Υ· ) = U0g + 1{τ >u−} e(Γu +Υu ) εgu d X· , L̃Y·∧τ
t 0 u
Z t h i
2
1{τ >u−} e(Γu +Υu ) d LU
g Y
+ · , L̃·∧τ
0 u
Z t h i
2
g
= U0g + · , L̃Y
·∧τ
0 u
Z t h i
2
g
= U0g + · , Ỹ·∧τ
0 u
Z t Z ·
Γ2u +Υu ) 0
− 1{τ >u−} e( Ug
d L· , gs dÑs∧τ
0 0 u
Z t Z τ ∧u
g (Γu +Υu ) Γs Ug

= U0 + 1{τ >u−} e d e d L· , Y· s
0 0
h i P,F
In the second equality, we use the fact X· , L̃Y·∧τ = N· , LY· u = 0 and in the fourth that

h g R 0 i D g R 0 EP,F u
U · U ·
L· , 0 gs dÑs∧τ = L· , 0 gs dNs = 0. Eventually, we get
u u
Z t
gT g
E P̂ G
1{τ >T } |Gt = U0 + 1{τ >u−} e(Γu +Υu ) εgu dXu + L̃gt
BT 0
Rt g R t g (Γ +Υ ) Rt g (Γ2u +Υu )
where L̃gt = 0 1{τ >u−} e(Γu +Υu ) dL̃Uu∧τ − 0 Uu− e
u u dM +
u 0 1{τ >u−} Uu− e dL̃Yu∧τ .
g
Notice L̃t is indeed a (P̂ G , G)-martingale since by definition of the minimal martingale measure
all (P, G)-martingales (P, G)-strongly orthogonal to Ñt remain (P̂ G , G)-martingales. We still
have to show that L̃gt is (P̂ G , G)-strongly orthogonal to Xt on {τ > t}. We have to show that
D EP̂ G ,G
L̃g· , X· = 0 . We have
t∧τ
D EP̂ G ,G Z t G ,G
g (Γu +Υu )
L̃g· , X· = − Uu− e d hM· , X· iP̂u
t∧τ 0
Z t D g EP̂ G ,G
+ 1{τ >u−} e(Γu +Υu ) d L̃U
·∧τ , X ·
0 u
Z t D EP̂ G ,G
g (Γ2u +Υu )
+ 1{τ >u−} Uu− e d L̃Y·∧τ , X·
0 u
G
On one hand, since Xt is continuous, we obviously have hM· , X· iP̂u ,G = 0. On the other hand,
D EP̂ G ,G D EP̂ G ,G D EP,G
L̃Y·∧τ , X· = L̃Y·∧τ , Ñ· = L̃Y·∧τ , Ñ· = 0. We can follow the same argument for
t t t
D g EP̂ G ,G
L̃U·∧τ , X· .
u
5.4.2. Payment at the surrender time. We first need the following result.
Proposition 5.14. Let us denote by P̂ G and P̂ F the minimal martingale measure with re-
spect to G and to F. If ht is a (bounded) F-predictable process, we have the following relationship
Z T Z T
G F
E P̂ hu dHu |Gt = 1{τ >t} e(Γt +Υt ) E P̂ hu dF̃u |Ft
t t
where F̃t = 1 − e−(Γt +Υt ) .

Proof. We have
h RT i
G
Z T E P ẐTG t hu dHu |Gt
E P̂ hu dHu |Gt =
t ẐtG
h RT i
E P ẐTG t hu dHu |Ft
= 1{τ >t} eΓt
ẐtG
h RT i
E P ẐTG t hu dHu |Ft
= 1{τ >t} e(Γt +Υt )
ẐtF
Since on {τ > t}, ẐtG = ẐtF e−Υt .

Pn Let us now assume hu is a bounded stepwise F-predictable process. We can write hu =
i=0 hti 1{ti <u≤ti+1 } where t = t0 < t1 < · · · < u < · · · < tn = T and hti is a bounded
Fti -measurable random variable for i = 0, . . . n. We have in this case
h RT i h i
E P ẐTG t hu dHu |Ft E P ẐTG ni=0 hti 1{ti <τ ≤ti+1 } |Ft
P
=
ẐtF ẐtF
Pn h i Pn h i
P Ẑ G h 1 P Ẑ G h 1
i=0 E T ti {τ >ti } |Ft i=0 E t
T i {τ >ti+1 } |Ft
= −
ẐtF ẐtF
Pn h h i i
P hti 1{τ >ti } E P ẐTG |Gti |Ft
i=0 E
=
ẐtF
Pn h h i i
E P h 1 E P Ẑ G G |F
i=0 ti {τ >ti+1 } T ti+1 t
−
ẐtF
Pn h i Pn h i
P h 1 P h 1
ti {τ >ti } Ẑti |Ft
G G |F
i=0 E i=0 E Ẑ
ti {τ >ti+1 } ti+1 t
= −
ẐtF ẐtF
Pn h i
i=0 E
P hti 1{τ >ti } e−Υti ẐtFi |Ft
=
ẐtF
h i
Pn P h 1 −Υti+1 F
i=0 E ti {τ >ti+1 } e Ẑti+1 |Ft
−
ẐtF
n n h i
E P̂ hti 1{τ >ti+1 } e−Υti+1 |Ft
X X
P̂ F −Υti F

= E hti 1{τ >ti } e |Ft −
i=0 i=0
n
" #
n o
hti e−(Γti +Υti ) − e−(Γti+1 +Υti+1 ) |Ft
X
P̂ F
= E
i=0
Z T
P̂ F
= E hu dF̃u |Ft
t
When hu is a general (bounded) F-predictable process, we can approximate this process by

a suitable sequence of stepwise processes. In this case, we would obtain, underR the sign of
T
conditional expectation, a sequence of sums that would converge to the integral t hu dF̃u in
the sense of Itô. At the price of a much more cumbersome proof, we could even generalize this
proposition to an unbounded hu (F-predictable) process.
We can now study the

R t pseudo locally risk-minimizing strategies of the discounted cumulative
Ru
payment process At = 0 Bu dHu which is assumed to be square integrable with respect to P .
Proposition 5.15. If, for Ru , F-predictable and P -square integrable,

hR the (P̂iF , F)-Galtchouk-
F T R
Kunita-Watanabe decomposition of the (P̂ F , F)-martingale E P̂ 0 Bu dF̃u |Ft is written
u
Z T Z t
P̂ F Ru R
E dF̃u |Ft = U0R + ψu dXu + LU
t
0 Bu 0
RTRu
then the (P, G)-Föllmer-Schweizer decomposition of 0 Bu dHu is given by
Z T Z T
Ru
dHu = U0R + 1τ >u− e(Γu +Υu ) ψu dXu + L̃R
T
0 Bu 0
where
Z T
Ru U R (Γu +Υu )
L̃R
T = − Vu− e dMu
0 Bu
Z T Z t
(Γu +Υu ) UR Γu Ru (Γu +Υu ) U R
+ 1τ >u− e dL̃u − 1{τ >u−} e −e Vu− dL̃Yu
0 0 Bu
and
T
Z
R P̂ F Ru
VtU

= E dF̃u Ft
t Bu
h
G RT R
i
Proof. We have to find the (P̂ G , G)-Galtchouk-Kunita-Watanabe of E P̂ 0 Bu
u
dHu |Gt .
Thanks to the previous proposition, we can write
Z T Z t Z T
P̂ G Ru Ru P̂ G Ru
E dHu |Gt = dHu + E dHu |Gt
0 Bu 0 Bu t Bu
Z t Z T
Ru (Γt +Υt ) P̂ F Ru ˜
= dHu + 1{τ >t} e E dFu |Ft
0 Bu t Bu
where F̃t = 1 − e−(Γt +Υt ) . We thus have

Z T Z T
G Ru F Ru
E P̂ dHu |Gt = 1{τ >t} e(Γt +Υt ) E P̂ dF̃u |Ft
0 Bu Bu
| {z 0 }
(I)
Z t Z t
Ru Ru
+ dHu − 1{τ >t} e(Γt +Υt ) dF̃u
Bu Bu
|0 {z 0
}
(II)
As far as the term (I) is concerned, we can follow the same argument as in Proposition 5.13.
It leads to
Z t Z t
R (Γu +Υu ) R
(I) = U0 + 1{τ >u−} e εu dXu + 1{τ >u−} e(Γu +Υu ) dL̃R
u
0 0
Z t Z t
R (Γu +Υu ) R (Γ2u +Υu )
− Uu− e dMu + 1{τ >u−} Uu− e dL̃Yu
0 0
hR i
P̂ F Ru T
where UtR = E 0 Bu dF̃t |Ft .
As far as the term (II) is concerned, in Appendix 1, we show that it can be written as
Z t Z u
Ru (Γu +Υu ) Rs
(II) = +e dF̃s dMu
0 Bu 0 Bs −
Z t Z u
Γu Ru (Γu +Υu ) Rs
− 1{τ >u−} e +e dF̃s dL̃Yu
0 Bu 0 Bs −
U R R
R t Ru
Since Vt = Ut − 0 Bu dF̃u . So eventually, we have
Z T Z t Z t
P̂ G Ru R (Γu +Υu ) R
E dHu |Gt = U0 + 1{τ >u−} e εu dXu + 1{τ >u−} e(Γu +Υu ) dL̃R
u
0 Bu 0 0
Z t
Ru UR
+ − e(Γu +Υu ) Vu− dMu
0 B u
Z t
Γu Ru (Γu +Υu ) U R
− 1{τ >u−} e −e Vu− dL̃Yu
0 B u
As in Proposition 5.13, we can easily show L̃R G

t is (P̂ , G)-strongly orthogonal to Xt .
RT
We can find similar results for the payment 0 1{τ >u} dCu if we notice we have the following
equality:
Z T Z T
P̂ G (Γt +Υt ) P̂ G −(Γu +Υu )
E 1{τ >u} dCu |Gt = 1{τ >t} e E e dCu |Ft
t t
This is left to the reader.
6. Conclusion.
In this chapter, we studied the locally risk-minimizing strategies for insurance contracts
with a surrender option. We distinguished two different cases. In the first one, we assumed the
(H)-hypothesis held. In the second case, we removed this assumption. In each case, we studied
the impact of an enlargement of filtration on the financial market and the minimal martingale
measure.
Prior to that, an another important contribution of this chapter consisted of showing how
the local risk-minimization theory could be extended to payment processes.
Appendix 1.
Proposition. We have
Z t Z t Z t Z u
Ru (Γt +Υt ) Ru Ru (Γu +Υu ) Rs
dHu − 1{τ >t} e dF̃u = +e dF̃s dMu
0 Bu 0 Bu 0 Bu 0 Bs −
Z t Z u
Γu Ru (Γu +Υu ) Rs
− 1{τ >u−} e +e dF̃s dL̃Yu
0 Bu 0 Bs −
Proof. First, recall we have the different representations

Z t Z t
2
1{τ >t} e(Γt +Υt )
= 1− e(Γu +Υu )
dMu + 1{τ >u−} e(Γu +Υu ) dỸu∧τ
0 0
Z t
2 0
− 1{τ >u−} e(Γu +Υu ) gu dÑu∧τ
0
and
Z t
0
eΥt = 1− eΓu +Υu gu dNu
0
We need to know the integral representation of e−Υt and F̃t . Using Itô’s formula for continuous
semimartingales on the function f (x) = x−1 , we have
Z t Z t
−Υt −Υ2u Υu 3
e−Υu d eΥ· , eΥ· u

e = 1− e de +
0 0
Z t Z t
0 2 0 0
(6.1) = 1+ eΓu −Υu gu dNu + eΓu −Υu gu d hN· , N· i gu
0 0
As far as F̃t is concerned, we have

158 Appendix 1.
F̃t = 1 − e−(Γu +Υu )

Z t Z t
−Υu −Γu
e−Γu de−Υu − e−Γ· , e−Υ· t

= − e de −
0 0
Z t Z t
e−Υu dFu − e−Γu de−Υu + Y· , e−Υ· t

=
0 0
Rt 0 0
Replacing Equation (6.1) in the square bracket, we get Y· , e−Υ· t = 0 eΓu −Υu gu d hN· , N· i gu .

Thus it leads to
Z t Z t
−Υu 0
F̃t = = e dFu − e−Υu gu dNu
0 0
Since Ft = Yt + Dt , we can also write
Z t Z t Z t
−Υu −Υu 0
F̃t = e dYu + e dDu − e−Υu gu dNu
0 0 0
Z t Z t
(6.2) = e−Υu dDu + e−Υu dLYu
0 0
Rt Ru
Let us now study the term 1{τ >t} e(Γt +Υt ) 0 Bu dF̃u .Using the integration by parts formula, we
have
Z t Z t Z u
(Γt +Υt ) Ru Rs
1{τ >t} e dF̃u = − dF̃s e(Γu +Υu ) dMu
0 Bu 0 0 Bs −
Z t Z u
( Γ2u +Υu ) Rs
+ 1{τ >u−} e dF̃s dL̃Yu∧τ
0 0 Bs −
Z t
R u
+ 1{τ >u−} e(Γu +Υu ) dF̃u
0 Bu
Z t
Ru h i
+ d F̃· , 1{τ >·} e(Γ· +Υ· )
0 Bu u
Replacing Equation (6.2), we get

Z t Z t Z u
(Γt +Υt ) Ru Rs
1{τ >t} e dF̃u = − dF̃s e(Γu +Υu ) dMu
0 Bu 0 0 B s −
Z t Z u
( 2
Γu +Υu ) Rs
+ 1{τ >u−} e dF̃s dL̃Yu∧τ
0 0 B s −
Z t Z t
Ru Ru Y
+ 1{τ >u−} eΓu dDu + 1{τ >u−} eΓu dL
0 Bu 0 Bu u
Z t Z t Z t
Ru −Υu Y (Γ2u +Υu )
+ d e dLu , 1{τ >u−} e dYu∧τ
0 Bu 0 0 u
Z t Z t Z t
Ru −Υu Y ( 2
Γu +Υu ) 0
− d e dLu , 1{τ >u−} e gu dNu∧τ
0 Bu 0 0 u
Since the last square bracket is null, eventually, we have
Z t Z t Z t Z u
(Γt +Υt ) Ru Ru Rs
1{τ >t} e dF̃u = dΛu − dF̃s e(Γu +Υu ) dMu
0 Bu 0 B u 0 0 Bs −
Z t Z u
Γu R u (Γu +Υu ) Rs
+ 1{τ >u−} e +e dF̃s dL̃Yu∧τ
0 Bu 0 Bs −
Part 3
Life Insurance Contracts and the Longevity

Risk.
CHAPTER 5
Risk-Minimization with Systematic mortality risk.
1. Introduction.
Systematic mortality risk is, by now, widely recognized as one of the important risks a life
insurance company faces. Unfortunately, the management of this risk poses serious problems.
On one hand, its systematic nature does not allow an insurer to diversify this risk away. On
the other hand, it is, at this time, still difficult to buy an efficient protection from a reinsurer
or to transfer this risk to the financial market. Accordingly, this systematic risk represents an
important source of incompleteness and it is important to closely study the hedging strategies
an insurer should follow to mitigate its effects.
Different approaches have been developed to tackle the problem of hedging in an incomplete
market. In this chapter, we focus on the risk-minimizing hedging strategies developed by
Föllmer and Sondermann in [44] and extended in Møller [68]. Actually, the application of
the risk-minimization theory to insurance contracts with systematic mortality risk has already
been studied in Dahl and Møller [39]. Under some simplifying assumptions, these authors
found closed-form solutions for the risk-minimizing strategies and the cost process (i.e. the
unhedgeable part) of some insurance payments. As far as the cost process is concerned, they
found it was mainly due to the unpredictability of the random time of death and to the random
evolution of a single point of the stochastic survival probability distribution function. As far
as the risk-minimizing strategies are concerned, Dahl and Møller showed they are given by
relatively complex expressions without, at first sight, any particular meaning. However, a close
look at their solutions reveals that these strategies actually take a very simple and intuitive
form. Indeed, in their model, simple algebraic manipulations show that the risk-minimizing
strategies can be written as the average of risk-minimizing strategies of purely financial claim,
weighted by the stochastic survival probabilities. The main motivation of this chapter is to
investigate if these conclusions would hold in a more general setting or if they are intrinsically
related to the particular assumptions made by Dahl and Møller in [39].
Our theoretical setting is more general than Dahl and Møller’s in, at least, three respects.
Firstly, as far as the financial market is concerned, Dahl and Møller assume that only two
assets are traded, a saving account and a single zero-coupon bond with maturity T . The
instantaneously risk-free rate is modelled as a time-homogeneous single factor affine model
of the CIR type. Since the term structure is modelled through a single factor model, bonds
of all maturities are perfectly correlated to the traded T -bond. Bonds of any maturities can
then be perfectly synthesized with the help of the two traded assets. Their financial market,
given the price of the T -bond, is thus a complete market. In this chapter, we only assume the
163
164 Introduction.
(discounted) financial assets prices follow an arbitrary s-dimensional local martingale. We do

not assume the financial market is complete.
Secondly, Dahl and Møller assume the insurance payments, given the times of death of their
policyholders, are deterministic. Since the financial market is complete, these deterministic
payments can be perfectly hedged with the help of two traded assets. In this chapter, the
insurance payments, given the time of death, can be stochastic and depend on the evolution
of the financial market. Furthermore, since our financial market is not necessarily complete,
these payments can not necessarily be perfectly hedged.
Finally, in Dahl and Møller, the probability distribution function of the random time of
death is described through its intensity. Indeed, in their model, the (stochastic) survival prob-
ability distribution function can be derived from this intensity, basically as bond prices of all
maturities can be derived from the short rate. In Dahl and Møller, this intensity is assumed
to be stochastic and to follow a time-inhomogeneous single factor model of the CIR type also1.
In such a model, the (stochastic) survival probability distribution function is thus driven by
a single factor as the term structure of interest rate is. In other words, the survival probabil-
ities at two different points are perfectly correlated with each others as interest rates at two
different maturities are. Put again differently, the evolution of the whole probability distri-
bution function can be summarized through the evolution of a single point of this probability
distribution function. We follow a different course in this thesis. We actually describes the
whole survival probability distribution function as a stochastic process that takes its values in
the space of finite measure but without giving the specific form of this probability function or
the specific form of the intensity of the random time of death. This method is very general
since it allows, for examples, for non-Brownian models, for multi-factors (and even possibly for
an uncountable infinite number of factors) models and even for discontinuities in the process
describing the survival probability distribution function. Moreover, our approach differs from
Dahl and Møller’s in that it allows to distinguish two situations. In the first one, the so called
(H)-hypothesis is assumed to hold whereas in the second situations we relax this assumption.
To my knowledge, this (H)-hypothesis actually holds implicitly in every current stochastic
mortality models, including Dahl and Møller’s, of course.
These very weak assumptions on the financial market and the dynamics of the probability
distribution function of the random time of death, are actually already sufficient to derive
interesting results that generalize Dahl and Møller’s. Our main conclusions are the following.
As far as the risk-minimizing strategies are concerned, we show that we can always write them as
an average of risk-minimizing strategies of purely financial claims, weighted by the stochastic
survival probabilities. The important corollary of this is that, finding the risk-minimizing
strategy of a life insurance contract can be separated in two independent problems, one related
to the modelling of the survival probabilities and one related to the hedging of purely financial
claims. As far as the cost process is concerned, the conclusion depends on whether or not
we assume the (H)-hypothesis holds. When this hypothesis holds, we find a result similar to
Dahl and Møller’s. The differences come from the fact that, firstly, in our model, the cost
process depends on the movement of the whole survival probability distribution function and
1Under the real measure as well as under the martingale measure.

Risk-Minimization with Systematic mortality risk. 165
not on the movement of a single point of this function. This is explained by the fact we do
not necessarily assume a single factor model for the dynamics of the probability distribution
function. Secondly, we have an additional term related to the incompleteness of the financial
market which does not appear in Dahl and Møller since their financial market is complete.
When the (H)-hypothesis does not hold, we have in addition, a new systematic term related
to the stochastic mortality which does not appear in the cost process of Dahl and Møller.
Technically speaking, the crucial point is to notice that the expected insurance payments
and the stochastic probability distribution of death can be understood as Banach space-valued
processes. We show that the stochastic integral defined in Gravereaux and Pellaumail [46] with
respect to such Banach space-valued processes, can be useful in life insurance. In particular,
we give an integration-by-parts formula for Banach space-valued martingales that appears to
be one of the key element to find the risk-minimizing strategies.
The present chapter is organized as follows. In Section 2, we introduce the general theo-
retical setting. More precisely, in Section 2.1, we describe our financial market model and in
Section 2.2, we describe our model for the random time of death and its stochastic probability
distribution function. In Section 2.3, we describe the combined model. Finally, in Section 2.4,
we present the insurance payments we wish to hedge. In Section 3, we give a brief review of the
risk-minimization theory. In Section 4 and Section 5, we study in details the risk-minimizing
strategies of our insurance payments. In the first section, we assume the (H)-hypothesis holds;
in the second one, we remove this assumption.

2.1. The financial market. Let (ΩF , F F , P F ) be a complete probability space. A perfect
frictionless financial market is defined on this space. We assume there is one locally risk free
asset denoted by Bt = B(t, ω) and s risky assets Sti = S i (t, ω), i = 1, . . . s following real-valued
càdlàg stochastic processes. The price of the locally risk free asset is assumed to follow a
strictly positive, continuous process of finite variation. The discounted values of these assets
are denoted by
Sti
Xti = , i = 1, . . . s
Bt
Let F = (Ft )t≥0 be the filtration generated by the stochastic processes Sti for i = 1, . . . s and
Bt . We assume this filtration respects the usual hypothesis.
We assume furthermore there is no arbitrage opportunity in this financial market with
respect to the filtration F. This assumption is equivalent to the existence of at least one
probability measure QF , equivalent to P F , such that the discounted prices X i are (F, QF )-
local martingales for each i = 1, . . . s.
2.2. The random time of death. Let us introduce another probability space (ΩI , F I , QI ).
We study a portfolio consisting of a single policyholder. The random time of death of this pol-
icyholder is defined on ΩI and denoted by τ = τ (ω). We define the stochastic process Ht by
Ht = 1{τ ≤t} . The filtration generated by the process Ht is denoted by H = (Ht )t≥0 where
Ht = σ (Hs , 0 ≤ s ≤ t).
We assume that the insurer can observe an additional set of information relevant to the
estimation of the distribution of τ . We assume this information comes from the observation
of a set of some unspecified stochastic processes also defined on ΩI . We denote J = (Jt )t≥0
the filtration generated by these unspecified stochastic processes and we assume this filtration
respects the usual hypothesis. and assume the following condition holds.
Condition 2.1. τ is not a J-stopping time.
This filtration could be generated, for example, by the number and the timing of deaths in
a given population or generated by some socio-economic factors related to the evolution of the
mortality.
As in Jeanblanc and Rutkowski [58], we can define the so called (Q, J)-hazard process Γt
of τ in the following way:
Ft := 1 − e−Γt
where Ft = Q (τ ≤ t |Jt ). For simplicity, we assume F0 = 0. We also assume Q(τ > t) > 0 for
all t ≥ 02. Thanks to Condition 2.1, it implies Ft < 1 for all t ≥ 0 and Γt = − ln(1 − Ft ) is
thus indeed well defined.
Ft is a (Q, J)-submartingale and, accordingly, Ft admits a unique Doob-Meyer decompo-
sition. We write Ft = Dt + Zt where Dt is an J-predictable increasing process and Zt a
(Q, J)-martingale.
The information related to the mortality is denoted by I = (It )t≥0 where It = Ht ∨ Jt . The
I
(Q , I)-compensator Λt of Ht is given by
Z t∧τ
1
(2.1) Λt = dDu
0 1 − Fu−
We denote the compensated process MtQ = Ht − Λt .
For any time t ≥ 0, we can also define the regular conditional distribution of τ (with respect
to J) Qt (du) = Qt (ω, du) as a modification of the following conditional expectation:
I
Qt (u) := E Q 1{τ ≤u} |Jt

= Q (τ ≤ u |Jt )
On one hand, for a given t ≥ 0, Qt (·) defines a random probability measure on ([0, T ], B ([0, T ]))
(where T is the term of an insurance contract). On the other hand, for a given u ∈ [0, T ], the
stochastic process Qt (u) is a (QI , J)-martingale. Let us also denote
St (u) := Q (τ > u |Jt )
= 1 − Qt (u)
Alternatively, Qt (·) and St (·) can also be seen as stochastic processes taking their values in the
space of probability measures or, more generally, in the space of signed and finite measures on
([0, T ], B ([0, T ])). Endowed with the appropriate norms, these spaces are Banach spaces. Qt (·)
2or at least to the term of the contract T .

and St (·) can thus be seen as Banach space-valued stochastic processes. Let us now introduce
the following definition.
Definition 2.2. Let E be a Banach space and let E ∗ be the dual of E. We say that an
E-valued process Gt is an E-valued martingale if and only if, for any (deterministic) h ∈ E ∗ ,
the action of the operator Th associated with h, on Gt , denoted by Th (Gt ) = (h, Gt ), is a
real-valued martingale.
Actually, Qt (·) and St (·) are even Banach space-valued martingales. Indeed, the dual of the
space of signed and finite measures is the space BM ([0, T ]) of bounded measurable functions
on [0, T ]. If h(u) ∈ BM ([0, T ]), then the action of the operator associated with h(u), on Qt (·)
is given by
Th (Qt ) = (h, Qt )
Z T
= h(u)Qt (du)
0
which is a real-valued martingale.
In [46], Gravereaux and Pellaumail defined a stochastic integral for Banach space-valued
stochastic processes. In Appendix 1, based on the definition of these authors, we give an
integration-by-parts formula for Banach space-valued processes. This result will be at the
hearth of our results on the risk-minimizing strategies. The remark that Qt (·) can be seen as
a Banach space-valued martingale, is thus essential in this chapter.
Before continuing, we have a last point to underline. In this section, we introduced a first
enlargement of filtration by adding H to J to form the filtration I. In general, a local martingale
with respect to a given filtration is not necessarily a local martingale with respect to a larger
one. In our setting, it means that the (local) (QI , J)-martingales are not necessarily (local)
(QI , I)-martingales. In particular, the process Zt or, for a given u ∈ [0, T ], the process Qt (u)
are not necessarily (QI , I)-martingales. Let us introduce the following condition:
Condition 2.3. τ avoids all J-stopping times i.e. for any J-stopping time θ, Q(τ = θ) = 0.
According to Blanchet-Scalliet and Jeanblanc [30], under this condition, for any (QI , J)-
martingale mt , the stopped process
Z t∧τ
(2.2) m̂t∧τ = mt∧τ + eΓs d [m· , Z· ]s
0
is a (QI , I)-martingale. We call m̂t∧τ the I
(Q , I)-martingalepart of mt∧τ . In particular, the
(QI , I)-martingale part of the (QI , J)-martingale Z can be written
Z t∧τ
Ẑt∧τ = Zt∧τ + eΓs d [Z· , Z· ]s
0
and for each u ∈ [0, T ], the (QI , I)-martingale part of the (QI , J)-martingale Qt (u) := Q (τ ≤ u |Jt )can
be written:
Z t∧τ
Q̂t∧τ (u) = Qt∧τ (u) + eΓs d [Q· (u), Z· ]s
0
We also have
Z t∧τ
Ŝt∧τ (u) = St∧τ (u) + eΓs d [S· (u), Z· ]s
0
Let us introduce the following definition.
Definition 2.4. We say the (H)-hypothesis holds under QI between J and I if any (QI , J)-
local martingale is a (QI , I)-local martingale.
In this chapter, we will distinguish two situations. Section 4 deals with the risk-minimization
of life insurance contracts when this (H)-hypothesis holds and Section 5 deals with the same
problem when this (H)-hypothesis does not hold.
2.3. The combined model. The combined model consists in the following probability
space ΩF × ΩI , F F ⊗ F I , Q where Q is the product measure of QF and QI . We define the

filtration G = (Gt )t≥0 where Gt = Ft ∨ It . Under the product measure Q, the filtration F and
I are independent.
We have here a second enlargement of filtration from F to G. Fortunately, if the inde-
pendence of F and I under Q holds, there is no real issue. Indeed, this independence implies
the (local) martingale properties of the F-adapted and I-adapted processes are preserved un-
der this enlargement of filtration. In other words, all the local Q F , F -martingales and all
the local QI , I -martingales are local (Q, G)-martingales. In particular, it means the QI , I -

compensator of the process Ht is indistinguishable of its (Q, G)-compensator. It also means the
discounted prices of the financial securities are local (Q, G)-martingales. If the financial market
is initially arbitrage-free then it remains arbitrage-free after this enlargement of filtration.
2.4. The payments. In this section, we describe the payments of the life insurance con-
tracts, we wish to hedge.
2.4.1. The insurer’s payments. A large variety of life insurance contracts can be modelled
as a combination of the following three building blocks. The first one is the payoff the insurer
has to pay at the term of the contract T . This payoff is assumed to be a FT -measurable random
variable and is denoted by g(T, ω). At the term of the contract, the insurer has to pay:
g(T, ω)1{τ >T }

The second building block is the amount the insurer has to pay when the policyholder dies
before the term T . This amount is denoted by 1{0<τ ≤T } R(τ, ω) where (R(t, ω))t≥0 is assumed
to be an F-predictable stochastic process. We have
Z T
1{0<τ ≤T } R(τ, ω) = R(u, ω)dHu
0
The third building block is the payoffs the insurer has to pay as long as the policyholder
is alive. We model these payoffs through their cumulative value up to time t, denoted by
Ct = C(t, ω). This process is assumed to be a right-continuous increasing square integrable
F-adapted process. This cumulative payoff is then given by:
Z T
0
where we assume C(0, ω) = 0 and C(T, ω) = C(T− , ω). hR i

u 1
For each u ∈ [0, T ], let us consider the conditional expectation ŨtC (u) = E Q 0 Bs dCs |Ft .
It can be shown that there exists a càdlàg process UtC (·) taking
its values in the space of
B([0, T ])-measurable function and such that, for each u, UtC (u) is a
version of ŨtC (u). Further-
C C
more, since U· (u) is increasing with u, Ut (·) can also be seen as a stochastic process taking
its value in the space of finite (signed) measures on ([0, T ], B([0, T ])). In other words, UtC (·) is
a Banach space-valued process. We can also easily see that it is even a Banach space-valued
martingale since, for any u ∈ [0, T ], the process ŨtC (u) is a (Q, G)-martingale. hIn the following,
i
Ru
we will always consider the version UtC (u) of the conditional expectation E Q 0 B1s dCs |Ft .
2.4.2. The policyholder’s payments. We assume the policyholder pays premiums periodi-
cally at N fixed dates ti with i = 0, . . . N − 1, as long as he is alive. We denote by P (ti , ω)
the value of the premium paid at time ti and assume P (ti , ω) is Fti -measurable for each
i = 0, . . . N − 1. The policyholder’s payments is then given by:
N
X −1
P (ti , ω)1{τ >ti }
i=0
3. The Risk-Minimization Theory.

We will here recall the main results of the theory of risk-minimization. These results are
due to Föllmer and Sondermann [44] for a single payoff and were extended by Møller [68] to
payment processes. We mainly follow Schweizer [78] for the presentation.
The financial market is defined as in Section 2.1. We will here consider an arbitrary filtration
G larger than the filtration F generated by the financial assets prices. We assume a specific
martingale measure Q has been chosen so that X is an (Q, G)-local martingale. This condition
holds indeed in the enlarged model described in Section 2.3.
Definition 3.1. A trading strategy ρ is a pair of processes (ε, η) where η is a 1-dimensional
real-valued G-adapted process and ε is an s-dimensional real-valued G-predictable process.
The process εt represents the amount of risky assets held at time t and the process ηt is
the discounted amount invested in the risk free asset.
Definition 3.2. The value process Vt (ρ) of a trading strategy ρ, is defined by Vt (ρ) =
εt Xt + ηt for 0 ≤ t ≤ T .
This value process represents the discounted value of the insurer’s financial portfolio fol-
lowing the trading strategy ρ. This portfolio is not necessarily self-financed. We define by
L2 (X, Q) the following space.
170 The Risk-Minimization Theory.
Definition 3.3. L2 (X, Q, G) is the space of Rs - and G-predictable processes ε such that
hR i1/2
T
kεkL2 (X,Q,G) = E Q 0 ε0u d [X, X]u εu < ∞.
We also have the following important lemma due to Schweizer:
Lemma 3.4. Suppose X is a (Q, G)-local martingale. For any ε ∈ L2 (X, Q, G), the process
RT 2 (Q, G). Moreover, the space I 2 (X, Q, G) =

ε ∈ L2 (X, Q, G) is a
R
0 ε u dX u ∈ M 0 εdX
stable subspace of M20 (Q, G).
Proof. See Schweizer [78].
Definition 3.5. An RM-strategy ρ is any pair (ε, η) where ε ∈ L2 (X, Q, G) and η is
a càdlàg adapted process such that the value process Vt (ρ) is right continuous and square
integrable.
The liabilities of the insurer towards a policyholder are modelled as a process At . The
process At is assumed to be càdlàg, G-adapted and square integrable. The process At represents
the discounted value of the cumulative payments up to time t.
Definition 3.6. The cumulative cost process Ct (ρ) of a RM-strategy ρ, is defined by
Rt
Ct (ρ) = Vt (ρ) − 0 εu dXu + At .
The initial cumulative cost process C0 (ρ) is given by C0 (ρ) = V0 (ρ) + A0 . The first term
V0 (ρ) represents the initial amount the insurer has to invest to create his financial portfolio.
The second term is the initial payment the insurer has to pay to the policyholder. Usually,
A0 will be negative and instead of a payment will represent the initial premium paid by the
policyholder.
Definition 3.7. A strategy is said to be mean self financing if its cumulative cost process
is a martingale.
The risk process is defined as follows:
h i
Definition 3.8. The risk process of an RM-strategy ρ is defined by Rt (ρ) = E Q (CT (ρ) − Ct (ρ))2 |Gt .
Following Møller [68], we restrict our attention to strategies which are 0-admissible in the
following sense:
Definition 3.9. A strategy ρ is said to be 0-admissible if and only if: VT (ρ) = 0 , Q-a.s.
The theory of risk minimization aims at finding an 0-admissible RM-strategy that minimizes
the risk process Rt (ρ) in the following sense.
Definition 3.10. An RM-strategy ρ is called risk-minimizing if and only if Rt (ρ) ≤ Rt (ρ̃)
Q-a.s. for every t ∈ [0, T ] and for any RM-strategy ρ̃ which is an admissible continuation of ρ
from t on, in the sense that VT (ρ) = VT (ρ̃) Q-a.s. , ε̃u = εu for u ≤ t and η̃u = ηu for u < t.
To solve this problem, we need the following result:
Lemma 3.11. Any risk minimizing RM-strategy is also mean self financing.
Proof. See Møller [68] for a proof.

This lemma leads directly to the following one:
Lemma 3.12. The discounted value of the risk minimizing RM-strategy ρ∗ is given by:
Vt (ρ∗ ) = E Q [AT − At |Gt ]
Proof. Indeed since by Lemma 3.11, the cost process is a martingale, we have
Ct (ρ∗ ) = E Q [CT (ρ∗ ) |Gt ]
Z t Z T
Vt (ρ∗ ) − εu dXu + At = E Q 0 − εu dXu + AT |Gt
0 0
R
Since the stochastic integral εu dXu is a martingale by Lemma 3.4, we have the result.
The risk-minimizing discounted amount invested in the risk free asset is thus given by:
ηt∗ = E Q [AT − At |Gt ] − ε∗t Xt
We now have to determine the amount of risky assets ε∗ . Schweizer shows the solution of this
problem is closely related to the so-called Galtchouk-Kunita-Watanabe decomposition of AT .
Since I 2 (X, Q, G) is a stable subspace of M20 (Q, G), any AT square integrable random variable
can be uniquely written as:
Z T
AT = E Q [AT |G0 ] + εA A
u dXu + LT
0
where εA
u ∈ L2 (X, Q, G) and LA
u∈ M20 (Q, G),
is Q-strongly orthogonal to I 2 (X, Q, G). With
these notation, we can give the solution of our risk minimizing problem in the following lemma:
Lemma 3.13. The unique risk minimizing RM-strategy ρ∗ = (ε∗ , η ∗ ) of A is given by
ε∗ = εA and ηt∗ = E Q [AT − At |Gt ] − ε∗t Xt . The cumulative cost process of ρ∗ is given by
Ct (ρ∗ ) = E Q [AT |G0 ] + LA ∗ A ∗ ∗
Q
E [AT − At |Gt ].
Proof. See Schweizer [78] for the single payoff case or Møller [68] for the extension to
payment processes.
As Schweizer noticed in [78], we do not have to assume X is (locally) square integrable for
these results to hold.
4. Application to a Life Insurance Contract When the (H)-Hypothesis Holds.

In this section, we assume the following hypothesis holds:
Condition 4.1. The (H)-hypothesis holds under QI between the filtrations J and I.
In our setting, we have the following equivalence:
Lemma 4.2. The (H)-hypothesis under QI is equivalent to
Q(τ ≤ t |J∞ ) = Q(τ ≤ t |Jt )
for any t ≥ 0.
The corollary of this lemma is that, under the (H)-hypothesis, Ft and thus Γt are increasing
and accordingly, of finite variation.
By definition, when this (H)-hypothesis holds then any square integrable (QI , J)-martingale
mt is a square integrable (QI , I)-martingale. If, in addition to the (H)-hypothesis, we assume
Condition 2.3 holds then m̂t∧τ the (QI , I)-martingale part of mt∧τ given by Equation (2.2) is
indistinguishable of mt∧τ . In particular, we have:
Q̂t∧τ (u) = Qt∧τ (u)
As far as Zt is concerned, we have the following proposition:
Proposition 4.3. If Conditions 4.1 and 2.3 hold then the (QI , J)-martingale part Zt of
the Doob-Meyer decomposition of the (QI , J)-submartingale Ft is equal to 0 for all t ≥ 0 a.s.
Proof. Conditions 2.3 and 4.1 implies hZ· , Z· it = 0 for all t > 0 a.s. since Ẑt∧τ = Zt∧τ .
In turn, this implies Zt = 0 for all t > 0 a.s.
In addition of being increasing, Ft is even J-predictable since Ft = Dt . We can even say a

little more thanks to the following proposition.
Proposition 4.4. If Condition 4.1 holds then Condition 2.3 holds if and only if the (Q, J)-
hazard process Γt is continuous.
Proof. First, we have
E Q 1{τ <t} |Jt = E Q lim→0 1{τ ≤t−} |Jt

= lim→0 E Q 1{τ ≤t−} |Jt−

= lim→0 1 − e−Γt−

= 1 − e−Γt−
where > 0. In the third equality, we use the fact that when the (H)-hypothesis holds, we
have for all s > t,
E Q 1{τ ≤t} |Js = E Q E Q 1{τ ≤t} |J∞ |Js

= E Q E Q 1{τ ≤t} |Jt |Js

= E Q 1{τ ≤t} |Jt

where in the second equality we use Lemma 4.2. Let us denote by θ an arbitrary J-stopping
time. We have:
Q(τ = θ) = E Q 1{τ =θ}

= E Q E Q 1{τ ≤θ} − 1{τ <θ} |Jθ

= E Q 1 − e−Γθ − 1 − e−Γθ−

= E Q e−Γθ− 1 − e−∆Γθ

On one hand, if Γt is continuous, we have Q(τ = θ) = 0 for any J-stopping time θ. On the
other hand, since Γt is increasing, e−Γθ− − e−Γθ is a positive random variable and Q(τ = θ) = 0
only if e−Γθ− − e−Γθ = 0. We thus have ∆Γθ = 0 for any J-stopping time θ.
Accordingly, when Conditions 4.1 and 2.3 hold Ft is continuous and increasing. Thanks to
the uniqueness of the Doob-Meyer decomposition, it confirms the fact Ft = Dt .
As far as the (QI , I)-compensator Λt of Ht is concerned, we have the following proposition:
Proposition 4.5. If Γt is continuous and increasing, the (QI , I)-compensator Λt of Ht is
given by Λt = Γt∧τ .
Proof. The proof can be found in Jeanblanc and Rutkowski [58]. We give it here for the
sake of completeness. Replacing Dt = Ft in Equation (2.1), we have:
Z t∧τ
eΓu d 1 − e−Γu

Λt =
0
Z t∧τ
= dΓu
0
The second equality depends on Γt being a finite variation continuous process.
We denote the compensated process MtQ = Ht − Γt∧τ . We can now study the risk-
minimizing strategy of each payment described in Section 2.4.
4.1. Payment at the term of the contract. We assume the insurer pays an amount
gT at the term T of the contract if the policyholder is still alive. The discounted payoff is then
given by 1{τ >T } BgTT where gT is assumed to be FT -measurable and QF -square integrable.
Proposition 4.6. The value of the risk minimizing portfolio Vtg (ρ∗ ) is given by

St (T ) Q gT
Vtg (ρ∗ )
= 1{τ >t} E |Ft
St (t) BT
and the risk-minimizing strategy ρ∗,g = ε∗,g ∗,g
t , ηt is given by
Z
gT
(4.1) E Q
1 |Gt = V0g (ρ∗ ) + ε∗,g g
u dXu + Lt
BT {τ >T } u:]0,t]
where
εu∗,g = 1{τ >u−} eΓu− Su− (T )εgu
and
Z
gT
Lgt = − eΓu− Q
Su− (T )E |Fu dMuQ
u:]0,t] B T −
Z
gT
+ 1{τ >u−} eΓu− E Q |Fu dSu (T )
u:]0,t] B T −
Z
g
(4.2) + 1{τ >u−} eΓu− Su− (T )dLU u
u:]0,t]
Ug
The processes εgt and Lt are given by the Galtchouk-Kunita-Watanabe decomposition of the
purely financial contingent claim:
Z
Q gT Q gT g
E |Ft = E |F0 + εgu dXu + LU
t
BT BT u:]0,t]
The discounted amount invested in risk free asset ηt∗,g is given by:

∗,g St (T ) Q gT
ηt = 1{τ >t} E |Ft − ε∗,g
t Xt
St (t) BT
The risk-minimizing strategy of this insurance payment is given by the product of the risk-
minimizing strategy of a purely financial contingent claim BgTT and the conditional probability
at time t of survival up to time T . Also the part that cannot be hedged through this risk
minimizing strategy (the cost process) is equal to the sum of three terms. The first one is
related to the unpredictability of the time of death, the second one is related to the random
evolution of the T -survival probability and the third one is related to the incompleteness of the
financial market.
Proof. The discounted value of the risk-minimizing strategy is given by:

g ∗ Q gT
Vt (ρ ) = E 1{τ >T } |Gt
BT

I F gT
= E Q 1{τ >T } |Jt ∨ Ht E Q

|Ft
BT

QI
QF gT
= 1{τ >t} E 1{τ >T } |Jt ∩ {τ > t} E |Ft
BT

St (T ) QF gT
= 1{τ >t} E |Ft
St (t) BT
The risk-minimizing strategy ρ∗,g is given
h by the (Q, G)-Galtchouk-Kunita-Watanabe
i of
Q gT
the (Q, G)-square integrable martingale E BT 1{τ >T } |Gt . Using Theorem 1 of Blanchet-
Scalliet and Jeanblanc [30] and using the fact that (1) Zt = 0, (2) Γt is continuous and (3)
mgt∧τ = m̂gt∧τ , we can write the following decomposition:

Z t
Q gT Q gT
E 1 |Gt = E 1 |G0 + 1{τ >u−} eΓu− dmgu
BT {τ >T } BT {τ >T } 0
Z
Γu− g Q
(4.3) − e mu− dMu
u:]0,t]
h i F
h i
where mgt := E Q 1{τ >T } BgTT |Ft ∨ Jt . Let us denote Utg := E Q BgTT |Ft . Thanks to the
Q-independence between F and J and using the integration-by parts formula, we have:
mgt = St (T )Utg
Z Z
g g
= Su− (T )dUu + Uu− dSu (T ) + [S· (T ), U·g ]t
u:]0,t] u:]0,t]
Z Z
g g g
= m0 + Su− (T )dUu + Uu− dSu (T )
u:]0,t] u:]0,t]
Since the process Utg is a square integrable (QF , F)-martingale, we can write its (QF , F)-GKW
decomposition:
Z
g
Utg = U0g + εgu dXu + LU
t
u:]0,t]
g
where εgt
is an F-predictable process in L2 (X, QF , F)
and LUt is a square integrable (QF , F)-
martingale strongly (Q , F)-orthogonal to I (X, Q , F). Replacing this equation in mgt , we
F 2 F
have:
Z Z
g g g
mt = m0 + Su− (T )εu dXu + Su− (T )dLU
u
u:]0,t] u:]0,t]
Z
g
(4.4) + Uu− dSu (T )
u:]0,t]
Replacing Equation (4.4) in Equation (4.3),
h we have the
i result. Notice Equation (4.1) gives us
Q gT
the Galtchouk-Kunita-Watanabe of E BT 1{τ >T } |Gt . Indeed, thanks to the (H)-hypothesis
between J and I and the Q-independence between F and I, Lgt defined by Equation (4.2) is a
(Q, G)-square integrable martingale and is Q-strongly orthogonal to I 2 (X, Q, G).
4.2. Payment at the time of death. We assume that the insurer pays an amount Rτ
R T Ru τ if it happens
at the time of death
Rt
before the term of the contract T . The discounted payoff is
then given by 0 Bu dHu where Bt is assumed to be F-predictable and QF -square integrable.
Proposition 4.7. The value of the risk minimizing portfolio VtR (ρ∗ )is given by
Z
1 Q Ru
VtR (ρ∗ )
= 1{τ >t} E |Ft St (du)
St (t) u:]t,T ] Bu

and the risk-minimizing strategy ρ∗,R = ε∗,R
t , η ∗,R
t is given by
"Z # Z
Ru
(4.5) E Q
dHu |Gt = V0R (ρ∗ ) + ε∗,R R
u dXu + Lt
u:]0,T ] Bu u:]0,t]
where
Z
R
ε∗,R
u = 1{τ >u−} e Γu−
εU
u (s)Qu− (ds)
s:]u,T ]
and
Z "Z # !
Ru R s
LR
t = − eΓu− E Q dHs |Fu ∨ Ju dMuQ
u:]0,t] Bu s:]u,T ] Bs
−
Z Z
Rs
+ 1{τ >u−} eΓu− EQ |Fu dQu (ds)
u:]0,t] s:]u,T ] Bs −
Z Z
R
(4.6) + 1{τ >u−} eΓu− Qu− (ds)dLU u (s)
u:]0,t] s:]u,T ]
R R
where εU U
u (s) and Lu (s) for s > u, are given by the following infiniteh family
i of G-K-W
R Q Rs
decompositions of the purely financial contingent claims Ut (s) := E Bs |Ft :
Z
R R
UtR (s) = U0R (s) + εU U
u (s)dXu + Lt (s)
u:]0,t]
The discounted amount invested in risk free asset is given by:

Z
∗,R 1 Q Ru
ηt = 1{τ >t} E |Ft St (du) − ε∗,R
t Xt
St (t) u:]t,T ] Bu
Here the risk-minimizing strategy is given by the weighted average of the risk minimizing-
Rs
strategies of the purely financial contingent claim B s
. The weights are given by the conditional
probabilities of death. Also the part that cannot be hedged through this risk-minimizing
strategy (the cost process) is equal to the sum of three terms. The first one is related to the
unpredictability of the time of death, the second one is related to the random evolution of the
whole stochastic probability distribution function of the time of death and the third one is
related to the incompleteness of the financial market. Notice in the paper of Dahl and Møller,
a term similar to the second one appears. But in their paper, this term depends on the random
evolution of a single probability (i.e. a single maturity). Here this term depends on the random
evolution of the whole probability measure (i.e. on all maturities from t to T ). This difference
comes from the fact their model of the stochastic probability measure is a single factor model
whereas our is a possibly infinite factors model. In a single factor model, the probabilities
associated to different maturities are perfectly correlated and the movement of the whole curve
can be summarized through the movement of a probability of a single maturity.
Proof. The discounted value of the risk minimizing strategy is given by:
"Z #
R ∗ Q Ru
Vt (ρ ) = E dHu |Gt
u:]t,T ] Bu
hR i
Ru
E Q u:]t,T ] B u
dH u |Ft ∨ J t
= 1{τ >t}
St (t)
Z
1 Ru
= 1{τ >t} EQ |Ft St (du)
St (t) u:]t,T ] Bu
In the second and third equalities, we use the Q-independence between F and J.
The risk-minimizing strategyhρ∗,R is given by theiGaltchouk-Kunita-Watanabe of the (Q, G)-
Ru
square integrable martingale E Q u:]0,T ] B
R
u
dHu |Gt . Using Theorem 1 of Blanchet-Scalliet and
Jeanblanc [30] and using the fact that (1) Zt = 0 and (2) Γt is continuous and mR R
t∧τ = m̂t∧τ ,we
can write the following decomposition:
"Z # "Z # Z
t
R u R u
EQ dHu |Gt = EQ dHu |G0 + 1{τ >u−} eΓu− dmR
u
u:]0,T ] Bu u:]0,T ] Bu 0
Z "Z # !
Ru R s
(4.7) + − eΓu− E Q dHs |Fu ∨ Ju dMuQ
u:]0,t] B u s:]u,T ] B s
−
hR i
where mR Q Ru
1 − e−Γu |Ft ∨ Jt . We show in Appendix 2, the process mR

t =E ]0,T ] Bu d t can
be written as:
Z "Z #
UR
mR
t = mR
0 + εs (u)Qs− (du) dXs
s:]0,t] u:]s,T ]
Z Z Z Z
Q Ru R
(4.8) + E |Fs dQs (du) + Qs− (du)dLU
s (u)
s:]0,t] u:]s,T ] Bu − s:]0,t] u:]s,T ]
Replacing Equation (4.8) in Equation (4.7), we have the result. Thanks to the (H)-hypothesis
between J and I and the Q-independence between F and I, LR t defined by Equation (4.6) is a
(Q, G)-square integrable martingale and is Q-strongly orthogonal to I 2 (X, Q, G). Thus,
hREqua- i
Q Ru
tion (4.5) gives us indeed the (Q, G)-Galtchouk-Kunita-Watanabe decomposition of E u:]0,T ] Bu dH u |Gt .

4.3. Payment up to the time of death. We assume the insurer pays an accumulated
amount Cu up to the time of death of the policyholder. The discounted payoff is then given by
R 1{τ >u} R 1
u:]0,T ] Bu dCu where Ct is assumed F-adapted and where u:]0,T ] Bu dCu is assumed square
Q-integrable.
Proposition 4.8. The value of the risk-minimizing portfolio VtC (ρ∗ ) is given by
Z
1
VtC (ρ∗ )
= 1{τ >t} St (u)UtC (du)
St (t) u:]t,T ]

and the risk minimizing strategy ρ∗,C = ε∗,C
t , ηt
∗,C
is given by
"Z #
1{τ >u}
Z
E Q
dC |Gt = V0C (ρ∗ ) + ε∗,C C
u dXu + Lt
u:]0,T ] Bu u:]0,t]
where
Z
C
ε∗,C
u = 1{τ >u−} e Γu−
Su− (s)εU
u (ds)
s:]u,T ]
and
"Z #
1{τ >s}
Z
LR
t =− eΓu−
E Q
dCs |Fu ∨ Ju dMuQ
u:]0,t] s:]u,T ] Bs
−
Z Z
+ 1{τ >u−} eΓu− C
Uu− (ds)dSu (s)
u:]0,t] s:]u,T ]
Z Z
C
+ 1{τ >u−} eΓu− Su− (s)dLU
u (ds)
u:]0,t] s:]u,T ]
C C
and where εU U
u (s) and Lu (s) for s > u, are given by the infinite
hR family of G-K-W i decomposi-
C Q 1
tions of the purely financial contingent claims Ut (s) := E ]0,s] Bu dCu |Ft
Z
C C
UtC (s) = U0C (s) + εU U
u (s)dXu + Lt (s)
u:]0,t]

Z
∗,C 1
ηt = 1{τ >t} St (u)UtC (du) − ε∗,C
t Xt
St (t) u:]t,T ]
Here, the risk-minimizing strategy is given by a weighted average of survival probabilities

where the weights are given by the risk-minimizing strategies of the purely financial contin-
1
R
gent claim ]0,s] Bu dCu . Again, the part that cannot be hedged through this risk minimizing
strategy (the cost process) is equal to the sum of three terms. The first one is related to the
unpredictability of the time of death, the second one is related to the random evolution of the
whole stochastic survival probability distribution function and the third one is related to the
incompleteness of the financial market.
Proof. The discounted value of the risk-minimizing strategy is given by:

"Z #
1 {τ >u}
VtC (ρ∗ ) = E Q dC |Gt
u:]t,T ] Bu
"Z #
1 1 {τ >u}
= 1{τ >t} EQ dC |Ft ∨ Jt
St (t) u:]t,T ] Bu
Z
1
= 1{τ >t} St (u)UtC (du)
St (t) u:]t,T ]
hR i
where UtC (u) = E Q ]0,u] B1s dCs |Ft . In the second and third equalities, we use the Q-
independence between F and J.
The risk-minimizing strategy ρ∗,C his given by the Galtchouk-Kunita-Watanabei of the
Q
R 1{τ >u}
(Q, G)-square integrable martingale E u:]0,T ] Bu dCu |Gt . Using Theorem 1 of Blanchet-
Scalliet and Jeanblanc [30] and using the fact that (1) Zt = 0, (2) Γt is continuous and (3)
mC C
t∧τ = m̂t∧τ , we can write the decomposition:
"Z # "Z # Z
1 1 t
{τ >u} {τ >u}
EQ dCu |Gt = EQ dCu |G0 + 1{τ >u−} eΓu− dmC
u
u:]0,T ] Bu u:]0,T ] Bu 0
" # !
1{τ >s}
Z Z
Γu− Q
(4.9) − e E dCs |Fu ∨ Ju dMuQ
u:]0,t] s:]u,T ] B s
−
hR i
e−Γu
where mC t = E
Q
]0,T ] Bu dCu |Ft ∨ Jt . We show in Appendix 3 that the process mC
t can
be written as:
Z "Z #
C
mC
t = J0C + Ss− (u)εU
s (du) dXs
s:]0,t] u:]s,T ]
Z Z Z Z
C C
(4.10) + Us− (du)dSs (u) + Ss− (u)dLU
s (du)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
Replacing Equation (4.10) in Equation (4.9),

hR we have the result.i Again, we have indeed the
Q 1{τ >u}
(Q, G)-Galtchouk Kunita Watanabe of E u:]0,T ] Bu dCu |Gt .
5. Application to a Life Insurance Contract When the (H)-Hypothesis Does Not

Hold.
In this section, we drop the assumption that the (H)-hypothesis holds under QI between
the filtration J and I but we keep Condition 2.3. In this case, the process Zt in the Doob-Meyer
decomposition of Ft is not necessarily equal to 0. We are in the general setting described in
Section 2.2.
5.1. Payment at the term of the contract. We assume the insurer pays an amount
gT at the term T of the contract if the policyholder is still alive. The discounted payoff is then
given by 1{τ >T } BgTT where gT is assumed to be FT -measurable and Q-square integrable.
Proposition 5.1. The value of the risk minimizing portfolio Vtg (ρ∗ ) is given by

St (T ) Q gT
Vtg (ρ∗ )
= 1{τ >t} E |Ft
St (t) BT
and the risk-minimizing strategy ρ∗,g = ε∗,g ∗,g

t , ηt is given by
Z
gT
E Q
1 |Gt = V0g (ρ∗ ) + ε∗,g g
u dXu + L̂t
BT {τ >T } u:]0,t]
where
εu∗,g = 1{τ >u−} eΓu− Su− (T )εgu
Z
gT
L̂gt = − Γu
e Su− (T )E Q
|Fu dMuQ
u:]0,t] B T −
Z
gT
+ 1{τ >u−} e2Γu− Su− (T )E Q |Fu dẐu
u:]0,t] BT −
Z
gT
+ 1{τ >u−} eΓu− E Q |Fu dŜu (T )
u:]0,t] BT −
Z
g
+ 1{τ >u−} eΓu− Su− (T )dLU u
u:]0,t]
Ug
where εgu and Lu , are given by the G-K-W decomposition of the purely financial contingent
claim:
Z
Q gT Q gT g
E |Ft = E |F0 + εgu dXu + LU
t
BT BT u:]0,t]
The discounted amount invested in risk free asset ηt∗,g is given by:

St (T ) Q gT
ηt∗,g = 1{τ >t} E |Ft − ε∗,g
t Xt
St (t) BT
Proof. The value of the risk-minimizing strategy Vtg (ρ∗ ) remains unchanged when we
drop the (H)-hypothesis. We can thus follow the same proof
h than in Proposition
i 4.6. As far
Q gT
as the GKW decomposition of the (Q, G)-martingale E BT 1{τ >T } |Gt is concerned, using
Theorem 1 of Blanchet-Scalliet and Jeanblanc [30], we can write
Z t
gT gT
EQ 1 |Gt = EQ 1{τ >T } |G0 + 1{τ >u−} eΓu− dm̂gu
BT {τ >T } BT 0
Z
g
− eΓu mu− dMuQ
u:]0,t]
Z
+ 1{τ >u−} e2Γu− mgu− dẐu
u:]0,t]
where, on {τ > t}, m̂gu is the I-martingale part of

g Q gT
mt = E 1 |Ft ∨ Jt
BT {τ >T }
= St (T )Utg
given , on {τ > t}, by
Z t
m̂gt = mgt + eΓs d [m· , Z· ]s
0
We saw in the proof of Proposition 4.6 that

Z Z
mgt = mg0 + Su− (T )dUtg + g
Uu− dSu (T )
u:]0,t] u:]0,t]
On {τ > t}, we thus have

Z t Z t
g g Γu g g Γu
m̂t = mt + Su− (T )e d [U· , Z· ]u + Uu− e d [S· (T ), Z· ]u
0 0
Z t
g g Γu
= mt + Uu− e d [S· (T ), Z· ]u
Z0 Z Z u
g g g Γs
= m0 + Su− (T )dUu + Uu− d Su (T ) + e d [S· (T ), Z· ]s
u:]0,t] u:]0,t] 0
Z Z
g g g
= m0 + Su− (T )dUu + Uu− dŜu (T )
u:]0,t] u:]0,t]
In the second hequality,i we use the fact U g and Z are strongly orthogonal. Using the decompo-
sition for E Q BgTT |Fu as in Proposition 4.6, we get the result.
The important point to notice here is that the risk minimizing strategy ρ∗,g = ε∗,g ∗,g
t , ηt
remains unchanged when we drop the (H)-hypothesis. Only the cost process is modified.
5.2. Payment at the time of death. The risk minimizing strategy of the payment at
the time of death is given by
Proposition 5.2. The value of the risk minimizing portfolio VtR (ρ∗ )is given by
Z
1 Q Ru
VtR (ρ∗ )
= 1{τ >t} E |Ft St (du)
St (t) u:]t,T ] Bu

and the risk minimizing strategy ρ∗,R = ε∗,R
t , η ∗,R
t is given by
"Z # Z
Ru
E Q
dHu |Gt = V0R (ρ∗ ) + ε∗,R R
u dXu + L̂t
u:]0,T ] Bu u:]0,t]
where
Z
R
ε∗,R
u = 1{τ >u−} eΓu− εU
u (s)Qu− (ds)
s:]u,T ]
and
Z "Z # !
Ru Rs
L̂R
t = − eΓu− E Q dHs |Fu ∨ Ju e∆Γu dMuQ
−
Z "Z # !
Γu− Ru Γu− Q Rs
− 1{τ >u−} e −e E dHs |Fu ∨ Ju dẐu
−
Z Z
R s
+ 1{τ >u−} eΓu− EQ |Fu dQ̂u (ds)
u:]0,t] s:]u,T ] Bs −
Z Z
R
+ 1{τ >u−} eΓu− Qu− (ds)dLU u (s)
u:]0,t] s:]u,T ]
R R
where εU U
u (s) and Lu (s) for s > u, are given by the following infiniteh family
i of G-K-W
R Q Rs
decompositions of the purely financial contingent claims Ut (s) := E Bs |Ft :
Z
R R
UtR (s) = U0R (s) + εU U
u (s)dXu + Lt (s)
u:]0,t]

Z
∗,R 1 Q Ru
ηt = 1{τ >t} E |Ft St (du) − ε∗,R
t Xt
St (t) u:]t,T ] Bu
Proof. The value of the risk-minimizing strategy VtR (ρ∗ ) remains unchanged when we
remove the (H)-hypothesis. We can thus follow basically the same proof than in Propo-
hR 4.7. As far asi the GKW decomposition of the square integrable (Q, G)-martingale
sition
Q Ru
E u:]0,T ] Bu dHu |Gt is concerned, using Theorem 1 of Blanchet-Scalliet and Jeanblanc [30],
we can write
"Z # "Z # Z
t
Q Ru Q Ru
E dHu |Gt = E dHu |G0 + 1{τ >u−} eΓu− dm̂R u
B
u:]0,T ] u B
u:]0,T ] u 0
Z "Z # !
Ru R s
+ − eΓu− E Q dHs |Fu ∨ Ju e∆Γu dMuQ
−
Z "Z # !
Γu− Ru Γu− Q Rs
(5.1) − 1{τ >u−} e −e E dHs |Fu ∨ Ju dẐu
−
hR i
where m̂t R is the (Q, I)-martingale part of mR Q Ru −Γu |F ∨ J

t = E ]0,T ] Bu d 1 − e t t given by
Equation (2.2). In Appendix 2, we show that:
Z "Z #
R R UR
mt = m0 + εs (u)Qs− (du) dXs
s:]0,t] u:]s,T ]
Z Z Z Z
R R
+ Us− (u)dQs (du) + Qs− (du)dLU
s (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
So, on {τ > t}, we have

Z t
R R
eΓs d mR

m̂t = mt + · , Z· s
0
Z tZ
= mR t + Us−R
(u)eΓs d [Q· (du), Z· ]s
0 u:]s,T ]
Z "Z # Z Z
R R
= mR
0 + εU
s (u)Qs− (du) dXs + Qs− (du)dLU
s (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
Z Z Z Z "Z #
R R
+ Us− (u)dQs (du) + Us− (u)d eΓv d [Q· (du), Z· ]v
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ] v:]0,s]
Z "Z # Z Z
UR R
= mR
0 + εs (u)Qs− (du) dXs + Qs− (du)dLU
s (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
Z Z " Z #
R
+ Us− (u)d Qs (du) + eΓv d [Q· (du), Z· ]v
s:]0,t] u:]s,T ] v:]0,s]
Eventually, on {τ > t}, the decomposition of m̂R

t is thus given by:
Z "Z # Z Z
R R U R R
m̂t = m0 + εs (u)Qs− (du) dXs + Qs− (du)dLU
s (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
Z Z
R
+ Us− (u)dQ̂s (du)
s:]0,t] u:]s,T ]
Replacing this last equation in Equation (5.1), we get the result.
As in the previous proposition, the risk-minimizing strategy remains unchanged when we

drop the (H)-hypothesis.
5.3. Payment up to the time of death.

Proposition 5.3. The value of the risk-minimizing portfolio VtC (ρ∗ ) is given by
Z
1
VtC (ρ∗ ) = 1{τ >t} St (u)UtC (du)
St (t) u:]t,T ]

and the risk minimizing strategy ρ∗,C = ε∗,C ∗,C
t , ηt is given by
"Z #
1{τ >u}
Z
E Q
dC |Gt = V0C (ρ∗ ) + ε∗,C C
u dXu + L̂t
u:]0,T ] Bu u:]0,t]
with
Z
C
ε∗,C
u = 1{τ >u−} e Γu−
Su− (s)εU
u (ds)
s:]u,T ]
and
"Z #
1
Z
{τ >s}
L̂R
t =− eΓu− E Q dCs |Fu ∨ Ju e∆Γu dMuQ
u:]0,t] s:]u,T ] Bs
−
"Z #
1{τ >s}
Z
Γu− Q
+ 1{τ >u−} e E dCs |Fu ∨ Ju eΓu− dẐu
u:]0,t] s:]u,T ] B s
−
Z Z
+ 1{τ >u−} eΓu− C
Uu− (ds)dŜu (s)
u:]0,t] s:]u,T ]
Z Z
C
+ 1{τ >u−} eΓu− Su− (s)dLU u (ds)
u:]0,t] s:]u,T ]
C C
and where εU U
u (s) and Lu (s) for s > u, are given by the infinite
hR family of G-K-W i decomposi-
C Q 1
tions of the purely financial contingent claims Ut (s) := E ]0,s] Bu dCu |Ft
Z
C C C UC
Ut (s) = U0 (s) + εU
u (s)dXu + Lt (s)
u:]0,t]

Z
∗,C 1
ηt = 1{τ >t} St (u)UtC (du) − ε∗,C
t Xt
St (t) u:]t,T ]
Proof. Again, as for the other payments, the value of the risk-minimizing strategy VtC (ρ∗ )
remains unchanged when we remove the (H)-hypothesis. We can thus follow basically the same
hR than in Propositioni 4.8. As far as the GKW decomposition of the (Q, G)-martingale
proof
1
E Q u:]0,T ] {τB>u} u
dCu |Gt is concerned, using Theorem 1 of Blanchet-Scalliet and Jeanblanc
[30], we can write
"Z # "Z # Z
1 1 t
{τ >u} {τ >u}
EQ dCu |Gt = EQ dCu |G0 + 1{τ >u−} eΓu− dm̂Cu
u:]0,T ] Bu u:]0,T ] Bu 0
"Z #
1
Z
{τ >s}
− eΓu− E Q dCs |Fu ∨ Ju e∆Γu dMuQ
u:]0,t] s:]u,T ] Bs
−
"Z #
1{τ >s}
Z
Γu− Γu− Q
(5.2) + 1{τ >u−} e e E dCs |Fu ∨ Ju dẐu
u:]0,t] s:]u,T ] Bs
−
hR i
e−Γu
where m̂t C is the I-martingale part of mC t = E Q
]0,T ] Bu dC u |F t ∨ J t given by Equation
(2.2). In Appendix 3, we show that
Z "Z #
C C UC
mt = m0 + Ss− (u)εs (du) dXs
s:]0,t] u:]s,T ]
Z Z
C
+ Ss− (u)dLU
s (du)
s:]0,t] u:]s,T ]
Z Z
C
+ Us− (du)dSs (u)
s:]0,t] u:]s,T ]
So on {τ > t}, we have

Z
C C
eΓs d mC

m̂t = mt + · , Z· s
s:]0,t]
"Z #
Z t Z
= mC
t + eΓs d C
Uv− (du)dSv (u), Z·
s:]0,t] v:]0,·] u:]v,T ]
s
Z t Z
= mC
t + eΓs C
Us− (du)d [S· (u), Z· ]s
s:]0,t] u:]s,T ]
Z Z "Z #
= mC
t + C
Us− (du)d Γv
e d [S· (u), Z· ]v
s:]0,t] u:]s,T ] v:]0,s]
Z "Z # Z Z
C C
= mC
0 + Ss− (u)εU
s (du) dXs + Ss− (u)dLU
s (du)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
Z Z " Z #
C
+ Us− (du)d Ss (u) + eΓv d [S· (u), Z· ]v
s:]0,t] u:]s,T ] v:]0,s]
186 Appendix 1.
Eventually, on {τ > t}, the decomposition of m̂C

t is given by:
Z " Z #
C
m̂C
t = mC 0 + Ss− (u)εU
s (du) dXs
s:]0,t] u:]s,T ]
Z Z
C
+ Ss− (u)dLU
s (du)
s:]0,t] u:]s,T ]
Z Z
C
+ Us− (du)dŜs (u)
s:]0,t] u:]s,T ]
Replacing this equation in Equation (5.2), we get the result.
As in the previous propositions, the risk-minimizing strategy remains unchanged when we

drop the (H)-hypothesis.
6. Conclusion.
In this chapter, we studied the risk-minimizing strategies of a single life insurance policy.
Our main conclusion is that these strategies can be written as an average of purely financial
strategies, weighted by the stochastic survival probabilities.
The fact that we focused on a single life insurance policy might seem rather limited. How-
ever, we can easily find the risk-minimizing strategies for a portfolio of policies if we assume
that the random times of death of the different policyholders are independent conditionally on
J. In this case, the risk-minimizing strategies for the portfolio are simply given by the sum of
the risk-minimizing strategies of each policy as described here (i.e as if they were considered
individually).
Appendix 1.
Proposition. Integration by parts formula.
Let us consider two Banach spaces F and E and a bilinear bounded application (·, ·) : F ×E → R.
Let Gt (resp. Ht ) be an E-valued (resp. F - valued) càdlag stochastic process. We assume
(∆Ht , ∆Gt ) = 0, ∀t ≥ 0 Q-a.s.
If Gt is a E-valued local Q-martingale, Ht an F -valued local Q-martingale, then, if (Ht , Gt )
is a real-valued local Q-martingale, we have:
Z Z
(Ht , Gt ) = (H0 , G0 ) + (Hs− , dGs ) + (dHs , Gs− )
]0,t] ]0,t]
Q-a.s., where the stochastic integrals are defined as in Gravereaux and Pellaumail [46].
Proof. In [46], Gravereaux and Pellaumail defined a stochastic integral for Banach space-
valued processes. More precisely, for three Banach spaces E, F, B and a bilinear bounded
application (·, ·) : F × E → B, they define a B-valued stochastic integral for E-valued càdlàg
stochastic integrators and F -valued càglàd stochastic integrands, as a sequence of elementary
stochastic integrals converging in ucp. For Xt a E-valued càdlàg stochastic process and Yt a
F -valued càglàd stochastic process, they denote their B-valued càdlàg stochastic process Zt as
Z
Zt = (Ys , dXs )
]0,t]
In our proposition, we assume B = R equipped with the usual norm. Let us define the
real-valued process Kt :
Z Z
Kt := (Ht , Gt ) − (Hs− , dGs ) − (dHs , Gs− )
]0,t] ]0,t]
where the stochastic integrals are defined as in [46]. We have to show that Kt = 0 ∀t ≥ 0,
Q-a.s. We prove this in 4 steps.
1) Thanks to the bilinearity of (·, ·) and the definition of Graveraux and Pellaumail of the
stochastic integrals as converging sequence in ucp of elementary stochastic integrals, we can
show this process Kt is the limit in ucp of
X n n n n

Kt = (H0 , G0 ) + lim H Ti+1 − H Ti , GTi+1 − GTi
n→∞
i
where (σn )n≥1 is a sequence of random partitions 0 = T0n ≤ T1n ≤ · · · ≤ Tin ≤ · · · ≤ Tknn
tending to the identity. We can show, as in the usual integration by parts formula for real-
valued stochastic processes, that this limit tends to a finite variation process.
2) Furthermore, we have that ∆Kt = ∆ (Ht , Gt ) − (Ht− , ∆Gt ) − (∆Ht , Gt− ). Thanks to the
bilinearity of the operator (·, ·), simple algebraic manipulations lead to
∆Kt = (∆Ht , ∆Gt )
Since by assumption (∆Ht , ∆Gt ) = (H0 , G0 ) ∀t ≥ 0 Q-a.s., Kt is thus, a.s., a continuous
process.
3) We can prove that a stochastic integral with respect to a Banach-valued local martingale
integrator and for a càglàd integrand, is a real-valued local martingale. Since (Ht , Gt ) is, by
assumption, a local Q-martingale, Kt is thus a real-valued local Q-martingale.
4) Since a predictable finite-variation real-valued local martingale is constant, we have that
Kt = (H0 , G0 ) and we can thus conclude that
Z Z
(Ht , Gt ) = (H0 , G0 ) + (Hs− , dGs ) + (dHs , Gs− )
]0,t] ]0,t]
Appendix 2.
Proposition. The Galtchouk-Kunita-Watanabe decomposition of
"Z #
R u
mR = EQ d 1 − e−Γu |Ft ∨ Jt

t
]0,T ] Bu
188 Appendix 2.
is given by:
Z "Z #
R
mR
t = mR
0 + εU
s (u)Qs− (du) dXs
s:]0,t] u:]s,T ]
Z Z Z Z
R R
s (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
Proof. We can rewrite mR t as:

Z Z "Z #
R Ru −Γu
Ru Q Ru
mt = d 1−e − Qt (du) + E Qt (du) |Ft ∨ Jt
]0,t] Bu ]0,t] Bu ]0,T ] Bu
hR i
Ru
Let us focus on JtR := E Q ]0,T ] B u
Qt (du) |F t ∨ Jt . Thanks to the Q-independence between
F and J, we can write:
Z
JtR = UtR (u)Qt (du)
u:]0,T ]
The crucial point here is to notice Qt (·) can be seen as a process taking its values in the space
of finite signed measures on ([0, T ], B([0, T ])). This space equipped with the usual norm is a
Banach space. The process UtR (·) can also be seen as a process taking its values in the space
of bounded Borel measurable functions on [0, T ], BM ([0, T ]), equipped with the supremum
norm. This space is also a Banach space. For h(·) ∈ BM ([0, T ]) and µ(·) a finite measure
RT
on ([0, T ], B([0, T ])), the application (h, µ) defined as (h, µ) = 0 h(u)µ(du), is thus a bilinear
bounded application. Accordingly, we have that
JtR = (UtR , Qt )
and the result of Appendix 1 applies. We can thus write the process JtR as:
Z Z
R R R R

Jt = J0 + dUs , Qs− + Us− , dQs
]0,t] ]0,t]
From now on, we use the obvious following notation:

Z Z Z Z
JtR = J0R + Qs− (du)dUsR (u) + R
Us− (u)dQs (du)
s:]0,t] u:]0,T ] s:]0,t] u:]0,T ]
Since for each u ∈ [0, T ], the process UtR (u) is a (Q, F)-square integrable martingale, we can
find its (Q, F)-GKW decomposition for each u ∈ [0, T ]. We can write:
Z
R R R UR
Ut (u) = U0 (u) + εU
s (u)dXs + Lt (u)
s:]0,t]
UR
where for each u ∈ [0, T ], εt (u) is an F-predictable process in L2 (X, QF ) and LU
R
t (u) is a
square integrable (QF , F)-martingale strongly orthogonal to I 2 (X, QF , F) and where for all
R R R
u < t, we have εU U U R
t (u) = 0 and Lt (u) = Lu (u). Replacing this decomposition in Jt , we
expect to get
Z Z
R
JtR = J0R + Qs− (du)εU
s (u)dXs
s:]0,t] u:]0,T ]
Z Z
R
+ Qs− (du)dLU s (u)
s:]0,t] u:]0,T ]
Z Z
R
+ Us− (u)dQs (du)
s:]0,t] u:]0,T ]
Z "Z #
R
= J0R + εU
s (u)Qs− (du) dXs
s:]0,t] u:]s,T ]
Z Z
R
+ Qs− (du)dLU
s (u)
s:]0,t] u:]s,T ]
Z Z
R
+ Us− (u)dQs (du)
s:]0,t] u:]s,T ]
Z Z
R
+ Us− (u)dQs (du)
s:]0,t] u:]0,s]
R R R
since εU U U
s (u) = 0 and Ls (u) = Lu (u) for all u ≤ s. The last term can also be written as:
Z Z Z Z
R
Us− (u)dQs (du) = UuR (u)dQs (du)
s:]0,t] u:]0,s] s:]0,t] u:]0,s]
Z Z
R
= Uu (u) dQs (du)
u:]0,t] s:]u,t]
Z
Ru
= [Qt (du) − Qu (du)]
u:]0,t] Bu
R (u) = U R (u) for any u ≤ s. In the second equality,

In the first equality, we use the fact that Us− u
we assume that we can safely interchange the order of integration. Eventually, we have for mR t :
Z "Z #
UR
mR
t = mR
0 + εs (u)Qs− (du) dXs
s:]0,t] u:]s,T ]
Z Z Z Z
R R
s (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
This completes the proof.

190 Appendix 3.
Appendix 3.
Proposition. The Galtchouk-Kunita-Watanabe decomposition of
"Z #
e−Γu
mCt = EQ dCu |Ft ∨ Jt
]0,T ] Bu
is given by:
Z "Z #
C
mC
t = mC
0 + Ss− (u)εU
s (du) dXs
s:]0,t] u:]s,T ]
Z Z Z Z
C
+ Ss− (u)dLU
s (du) +
C
Us− (du)dSs (u)
s:]0,t] u:]s,T ] s:]0,t] u:]s,T ]
hR i
1
Proof. Let us denote UtC (u) = E Q ]0,u] Bs dCs |Ft . For a given u ∈ [0, T ], UtC (u)
is a square integrable (QF , F)-martingale. For each t ∈ [0, T ], UtC (·) is a finite measure on
([0, T ], B([0, T ])). We can rewrite mC
t as:
Z t −Γu "Z #
e e−Γu
mC t = dCu + E Q dCu |Ft ∨ Jt
0 Bu ]t,T ] Bu
Z t −Γu "Z #
e Q St (u)
= dCu + E dCu |Ft ∨ Jt
0 Bu ]t,T ] Bu
Z t −Γu Z t "Z #
e St (u) S t (u)
= dCu − dCu + E Q dCu |Ft ∨ Jt
0 Bu 0 Bu u:]0,T ] Bu
Let us first study the two first terms, we have:

Z t −Γu Z t Z
e St (u)
dCu − dCu = − [St (u) − Su (u)] UtC (du)
0 Bu 0 Bu u:]0,t]
hR i
C Q St (u)
Let us now focus on Jt := E ]0,T ] Bu dCu |Ft ∨ Jt . Thanks to the Q-independence be-
tween Fand J, we can write
Z
JtC = St (u)UtC (du)
]0,T ]
The crucial point here is to notice that St (·) is a process taking its value in the space of bounded
Borel measurable functions BM ([0, T ]) and UtC (·) is a process taking its values in the space of
finite signed measures on [0, T ]. As explained in the previous appendix, these spaces, equipped
RT
with the appropriate norms, are Banach spaces and the application (h, µ) = 0 h(u)µ(du) for
h ∈ BM ([0, T ]) and µ a finite measure, is thus a bilinear bounded application. We can thus
write
JtC = St , UtC

Thanks to Appendix 1, we have:
JtC St , UtC

=
Z Z
C
Ss− , dUsC C

= J0 + + dSs , Us−
s:]0,t] s:]0,t]
From now on, we use the following obvious notation:

Z Z Z Z
JtC = J0C + Ss− (u)dUsC (du) + C
Us− (du)dSs (u)
s:]0,t] u:]0,T ] s:]0,t] u:]0,T ]
Since, for any B ∈ B([0, T ]), the process UtC (B) is a (QF , F)-square integrable martingale,
we can find its (QF , F)-GKW decomposition. For any B ∈ B([0, T ]), we can write
Z
C C
UtC (B) = U0C (B) + εU U
s (B)dXs + Lt (B)
s:]0,t]
C C
where εU 2 F U
t (B) is an F-predictable process in L (X, Q ) and Lt (B) is a square integrable
C Uc
(QF , F)-strongly orthogonal to I 2 (X, QF , F). The processes εUt (·) and Lt (·) are stochastic
processes taking their values in the space of finite and signed measures. Accordingly, we expect
to get
Z "Z #
UC
JtC = J0C + Ss− (u)εs (du) dXs
s:]0,t] u:]0,T ]
Z Z
C
+ Ss− (u)dLU
s (du)
s:]0,t] u:]0,T ]
Z Z
C
+ Us− (du)dSs (u)
s:]0,t] u:]0,T ]
Z "Z #
C
= J0C + Ss− (u)εU
s (du) dXs
s:]0,t] u:]s,T ]
Z Z
C
+ Ss− (u)dLU
s (du)
s:]0,t] u:]s,T ]
Z Z
C
+ Us− (du)dSs (u)
s:]0,t] u:]s,T ]
Z Z
C
+ Us− (du)dSs (u)
s:]0,t] u:]0,s]
192 Appendix 3.
C C C
In the second equality, we use the fact that εU U U
s (u) = 0 and Ls (u) = Lu (u) for all u ≤ s.
As far as the last term is concerned, we have:
Z Z Z Z
C
Us− (du)dSs (u) = dSs (u)UtC (du)
s:]0,t] u:]0,s] s:]0,t] u:]0,s]
Z Z
= dSs (u)UtC (du)
u:]0,t] s:]u,t]
Z
= [St (u) − Su (u)] UtC (du)
u:]0,t]
In the first equality, we use the fact that for t ≥ s, UsC (u) = UtC (u) for all u ≤ s. In the second
equality, we assume that we can safely interchange the order of integration. Eventually, we
have for mC t :
Z "Z #
UC
mC
t = mC
0 + Ss− (u)εs (du) dXs
s:]0,t] u:]s,T ]
Z Z
C
+ Ss− (u)dLU
s (du)
s:]0,t] u:]s,T ]
Z Z
C
+ Us− (du)dSs (u)
s:]0,t] u:]s,T ]
since J0C = mC
0.
CHAPTER 6
Heath-Jarrow-Morton Modelling of Longevity Bonds.
1. Introduction.
A zero-coupon longevity bond is a financial security whose single payout occurs at maturity
and is equal to the value at this time, of a so-called survivor index. The value at any times
t, of a survivor index is given by the proportion of people still alive at time t in an initially
given population. We can distinguish two broad approaches in the literature to the modelling
of longevity bond prices. A first approach is known as the “intensity” approach. In a sense, it
is the counterpart to the “short-rate” approach in the interest rate term structure literature.
Typical examples of these intensity models are, for examples, Biffis [29] (2005), Dahl [37]
(2004), Luciano and Vigna [64] (2005), Schrager [74] (2006) and Hainaut and Devolder [49]
(2007). The second approach consists of adapting the Heath-Jarrow-Morton (HJM) (see [50]
(1992)) methodology to the modelling of longevity bond prices. Cairns et al. [36] (2006)
is, to our knowledge, the first paper to exploit this idea. These authors studied different
specifications for the prices of the longevity bonds. In particular, they described the term
structure of longevity bond prices as the product of two independent HJM models, one related
to the the term structure of risk-free zero-coupon bonds and and a second one related to the
survival probabilities. Miltersen and Persson [67] (2005) extended Blake et al.’s results by
removing this independence. However, they assumed the payouts of their longevity bonds at
maturity were defined on the survival of a single individual. Bauer [18] (2006) introduced the
fact that the actual payouts of the longevity bonds depend on the proportion of individuals
alive in a given population and not on a single individual.
Unfortunately, none of these papers properly defines the prices of the longevity bonds they
are supposed to study. The trouble comes from the fact that they fail to take into account
the effect of the actual mortality in the population on the prices of their longevity bonds.
The first and, probably, the main goal of this chapter is, therefore, to describe a coherent
theoretical setting to which we can rigorously apply the HJM methodology to the modelling of
longevity bond prices. For this, it is necessary to introduce an explicit model of the (actual)
number of deaths in a given population, through a more or less general (marked) point process.
Modelling this mortality explicitly is fundamental to offering a coherent and rigorous definition
of the longevity bond prices.
A second important objective of this chapter is to describe a more realistic financial market.
We extend the existing literature in various directions. Firstly, we take into account an addi-
tional effect of the actual mortality on the longevity bond prices. We can distinguish two main
effects of this actual mortality. The first is a purely mechanical effect. If 5% of the population
193
194 Introduction.
of the survivor index on which a longevity bond is defined dies, then, everything else being
equal, the price of the longevity bond will also decline by 5%. This simple mechanical and
proportional effect clearly appears for example, in the pricing formula of longevity bonds in
the intensity models. This first effect can obviously only depend on the actual mortality in the
population of the survivor index on which the longevity bond is defined. However, there is also
a possibe much more subtle second effect which has (to my knowledge) never been taken into
account either in the intensity models or in the HJM models. We know from financial econom-
ics theory that the price at time t of a longevity bond with maturity T , should depend on the
survival probabilities up to time T , possibly adjusted for the risk, in the population on which
the survivor index is defined. A priori, there is no reason to believe these (risk-adjusted) sur-
vival probabilities might not also depend on the actual mortality in the population. Investors,
if they are rational, should update their estimates of the survival probabilities (even under the
real measure) if they observe a mortality which is not in line with their previous expectations.
They could even modify their correction for risk if the actual mortality makes them think they
are bearing more (or less) systematic risk than they previously thought. Unlike the first effect,
this second effect does not necessarily only depend on the actual mortality in the survivor
index on which the longevity bond is defined; it may also depend on the mortality in other
parts of the population. To be more concrete, let us assume we have a longevity bond defined
on 40-year-old individuals and assume the time to maturity is 20 years. The price of this bond
should depend not only on the actual current mortality of 40-year-old individuals, but also on
the actual current mortality of 45, 50, 55,etc -year-old individuals because investors can use
this information to update their expected future survival probabilities of the current 40-year-old
population. This second effect does not appear in the existing intensity models because these
models assume, at least implicitly, the random times of death are conditionally independent
with respect to a certain filtration (which does not include the actual mortality) and to the
risk-adjusted measure (the risk-neutral martingale measure). To me, these assumptions are not
obvious.
Secondly, we simultaneously consider different survivor indices, each defined on a different
age group of the overall population. We can thus define a longevity bond term structure on
each of these indices. In other words, unlike the existing literature, we do not consider a single
longevity bond term structure but multiple longevity bond term structures. This is a natural
extension since longevity bonds defined on several indices would allow the insurance companies
and pension funds to better tailor their assets to their liabilities. As the longevity bonds market
matures, more and more longevity bonds defined on different survivor indices, will naturally
be issued. Since each longevity bond depends on the actual mortality in the whole population,
complex dependencies could appear between the prices of the different longevity bonds. In
particular, it is interesting to study the condition of absence of arbitrage between the different
longevity bond term structures.
Finally, we extend the existing literature by allowing the different survivor indices to be
heterogeneous. In the “intensity” models, the survivor index is always defined on individuals of
the same age and who are homogeneous (identical forces of mortality). We do not make these
assumptions here. In particular, our survivor indices can be defined on individuals of different
ages. As we said, it is very likely that longevity bonds defined on different survivor indices
HJM Modelling of Longevity Bonds. 195
will be issued in the future but the number of different survivor indices will probably remain
relatively low. In order to cover a large proportion of the insured population with a limited
number of indices, it might be interesting to simply define larger survivor indices. For example,
one longevity bond could be defined on the population currently between 20 and 35 years old,
another on the population currently between 35 and 45 years old, etc. Survivor indices could
be defined on larger or smaller age groups according to the needs of the insurance companies or
the pension funds. Obviously, the population of such survivor indices would be heterogenous
from the point of view of mortality. Our model allows such “heterogenous” longevity bonds to
be dealt with.
The third aim is to study the asset allocation problem for portfolios of endowments and
annuities. In particular, we choose to study the “risk-minimizing” strategies of these liabilities.
This theory has already been applied with success to life insurance by Møller [68](2001), Riesner
[72] (2006) and Dahl and Møller [39] (2006), for example. In particular, Dahl and Møller deal
with the risk-minimizing strategies of life insurance liabilities with longevity risk. However,
these authors assume only risk-free zero-coupon bonds are available for trading. In this thesis,
we allow for trading in longevity bonds. An important point to underline is that we do not
assume that the traded longevity bonds are defined on the same age groups than the insurer’s
liabilities. We thus introduce different types of basis risks.
Finally, as a by-product of the asset allocation problem, we also study the market value
of portfolios of endowments and annuities when longevity bonds are traded in the financial
market. In this situation, the market values of these liabilities can be directly inferred from the
observable prices of the longevity bonds, without assuming the completeness of the financial
market.
This chapter is organized as follows. In Section 2, we give a critical review of the literature
and show that previous models actually do not define properly either the prices or the payoffs
of the longevity bonds. In Section 3, we study in detail how to apply the HJM framework to
the modelling of longevity bond prices. More precisely, Section 3.1 introduces the theoretical
setting on which this whole chapter rests. In particular, we describe the random times of death
in the population as a marked point process. The assumptions about the distribution of these
random times are very weak. We only assume they cannot occur simultaneously and that the
times of death are totally inaccessible stopping times. In particular, our model encompasses
all the “intensity” models defined on a Brownian filtration, as in Dahl [37] (2004), Biffis [29]
(2005), Dahl and Møller [39] (2006), Luciano and Vigna [64] (2005), Schrager [74] (2006).
Notice it does not, however, encompass Hainaut and Devolder’s model [49] (2007) since this
model is defined on a more general “Lévy” filtration. In Section 3.2, we define the survivor
indices on which the longevity bonds will depend, and in Section 3.3, the prices of the risk-free
zero-coupon bonds. For the sake of generality, we assume these prices can also depend on
the development of the actual mortality in the overall population. In this way, we introduce,
another possible dependence between the risk-free zero-coupon bonds and the longevity bonds
that does not appear in the existing literature. In Section 3.4, we define the longevity bond
prices. We believe our definition is more coherent than the previous ones appearing in the
literature. In particular, in order to introduce the second effect of the actual mortality on the
longevity bond prices described above, we assume the forward rates of the longevity bonds
196 A Critique of the literature.
depend on the actual mortality in the whole population. In Section 3.5, we then derive the
no-arbitrage condition in this general financial market. As in the traditional HJM model, this
no-arbitrage condition can be expressed as a constraint on the drift of the different forward
rates. Since we are modelling multiple term structures, this condition ensures there is not
only no arbitrage opportunity for each term structure but also between the different term
structures. In Section 3.6, we give alternative representations of the zero-coupon bond prices
and longevity bond prices. In this way, we make the connection between the HJM approach
and the “intensity” approach for modelling the longevity bonds. The results suggest that the
basic pricing formula of the “intensity” approach could be extended to more general intensity
models. In Section 3.7, we study the well-known forward martingale measures and recall their
main properties. Then we describe the exact counterpart to these forward martingale measures
for the longevity bonds that we call the longevity forward martingale measures. We show the
main properties of the classical forward martingale measures are preserved for the longevity
forward martingale measures.
In Section 4, we consider the asset allocation problem for endowments and annuities port-
folios using the risk minimization theory. We recall the main results of this theory in Section
4.1. In Section 4.2 and Section 4.3, we study the risk-minimizing strategies for these portfolios
when some, but not all, the longevity bonds are traded. Sections 4.2 and 4.3 differ in the sense
that in the second one, we assume the policyholders of a given insurance company, represent
only a fraction of the population on which a survivor index is defined, whereas in the first
one, we assume the policyholders represents the whole of this population. As a by-product
of the risk-minimizing strategies, we, in Sections 4.2 and 4.3, also give the market values of
endowments and annuities portfolios when longevity bonds are traded.
2. A Critique of the literature.

In this section, we study how the Heath-Jarrow-Model methodology (see [50] (1992)) has
been applied to the modelling of longevity bonds prices. To our knowledge, three papers treat
this topic: Miltersen and Persson’s [67] (2005), Cairns et al.’s [36] (2006) and Bauer’s [18]
(2006). Unfortunately, a detailed study of these papers shows that none of them properly
tackles the problem they are supposed to study. All these papers actually do not properly
define either the prices or the payoffs of the longevity bonds. In the following subsections, we
study each of the three papers mentioned above, and underline their main troubles.
2.1. Miltersen and Persson’s Paper. In [67], Miltersen and Persson study longevity
bonds defined on a single individual. In other words, a longevity bond with maturity T pays 1
at time T if a given individual is alive at time T and pays 0 if this individual dies before time
T . They denote by τ the time of death of this individual and P e(t, T, x + t − t0 ) the price at
time t of a longevity bond with maturity T . For the sake of simplicity, we will simply write
P e(t, T ) in the following. In this paper, these authors define for all t ≤ T , the price of their
longevity bonds by the following equation (see middle of page 7 in [67]):
RT
P e(t, T ) = e− t f (t,s)+µ(t,s;x+s−t0 )ds
where f (t, s) is the s-(risk-free) forward rate at time t and µ(t, s; x + s − t0 ) is a so-called
(stochastic) s-forward force of mortality at time t for an individual of age x at time t (see the
original paper for more details).
This definition implies that the price of such a longevity bond is always strictly positive.
However, the price of this longevity bond should be null for all t ≥ τ since in this case, we know
with certainty that the payoff is null at maturity. We thus have: P e(t, T ) = 0 for all t ≥ τ .
Accordingly, Miltersen and Persson should have defined the price of such a longevity bond as
RT
(2.1) P e(t, T ) = 1{τ >t} e− t f (t,s)+µ(t,s;x+s−t0 )ds
At first sight, this simple modification seems innocuous but we will see that it is actually
fundamental. Intuitively, by neglecting this indicator function, they miss a potential jump of
size -100%. When the payoff of the longevity bonds depends on a single individual as in their
model, the actual death of this person is actually the main factor of risk, probably much more
important than the risk coming from the Brownian motion that drives the forward rates and
the forward forces of mortality.
Since all further results rest on this definition, it could lead us to wrong conclusions. For
example, Miltersen and Persson derive the no-arbitrage condition in their model. If the no-
arbitrage condition holds under Q, the discounted price P e∗ (t, T ) = P e(t,T Bt
)
should be a (local)
Q-martingale for all T . Top of the page 8 in [67], the authors give the dynamics of this
discounted price based on their definition of P e(t, T ):
 RT 
µ(t, t; x + t − t0 ) − t (νf (t, s) + νµ (t, s; x + s − t0 )) ds
dP e∗ (·, T )t = P e∗ (t, T )   dt
 
R 2
1 T
+ 2 t σf (t, s) + σµ (t, s; x + s − t0 )ds
R
(2.2) − P e∗ (t, T ) T
t σf (t, s) + σµ (t, s; x + s − t0 )ds
dWt
where Wt is a Brownian motion under the pricing measure Q. To be a (local) martingale, the
drift of this SDE should be equal to 0. Based on the drift of this equation, they should have
written their no-arbitrage condition as
!2
Z T Z T
1
(νf (t, s) + νµ (t, s; x + s − t0 )) ds − µ(t, t, x + t − t0 ) = (σf (t, s) + σµ (t, s; x + s − t0 )) ds
t 2 t
But actually, they did not. After soad-hoc argument (they study the price of another contingent
claim), they eventually conclude the no-arbitrage condition is not given by the previous formula
but by the following one (see the last but one formula from the bottom of page 8 in [67]):
!2
Z T Z T
1
(2.3) (νf (t, s) + νµ (t, s; x + s − t0 )) ds = σf (t, s) + σµ (t, s; x + s − t0 )ds
t 2 t
This is indeed the correct no arbitrage condition but they should have found it directly from
the condition on the drift of P e(t,T )
Bt . By using the modified definition of P e(t, T ), they would
have directly found the right condition since, in this case, the drift of P e∗ (t, T ) would have
directly given the correct no arbitrage condition. Indeed, if we consider the modified definition
in Equation (2.1), we can write using the integration-by-parts formula
RT
∗ e− t f (t,s)+µ(t,s;x+s−t0 )ds
dP e (t, T ) = d1{τ >t}
Bt
RT
e− t f (t,s)+µ(t,s;x+s−t0 )ds
+ 1{τ >t− } d
Bt
" RT #
e− t f (t,s)+µ(t,s;x+s−t0 )ds
+ 1{τ >t} ,
Bt
Rt
The square bracket is null since the exponential term is continuous. Denote 0 1{τ >s− } µ(s, x)ds
the Q-compensator1 of 1{τ ≤t} , we can then write:
 RT 
µ(t, t; x + t − t0 ) − µ(t, x) − t
(νf (t, s) + νµ (t, s; x + s − t0 )) ds
dP e∗ (t, T ) = P e∗ (t, T )−   dt
 
R 2
T
+ 21 σf (t, s) + σµ (t, s; x + s − t0 )ds
t
R
T
− P e∗ (t, T )− t
σf (t, s) + σµ (t, s; x + s − t0 )ds dWt
− Pe ∗
(t, T )− dMtQ
Rt
where MtQ = 1{τ ≤t} − 0 1{τ >s−} µ(s, x)ds is a Q-martingale by definition of the compensator.
For the no-arbitrage condition to hold, we need:
!2
Z T Z T
1
+ νµ (t, s; x + s − t0 )) ds + µ(t; x) − µ(t, t, x + t − t0 )
(νf (t, s)(2.4) = (σf (t, s) + σµ (t, s; x + s − t0 )) ds
t 2 t
If we let T → t, this condition leads to µ(t; x) − µ(t, t, x + t − t0 ) = 0. Replacing this equality

in Equation (2.4), we directly find the correct no-arbitrage condition (Equation (2.3)) without
having to rely on other arguments. This simple example shows that, though the authors do
correctly model the actual mortality (here in their simple setting whether the individual is
alive or not), they do not properly take it into account in their definition of the longevity
bonds prices.
2.2. Cairns, Blake and Dowd’s Paper. Let us now study Cairns et al.’s paper [36].
In this paper, the payoff of a longevity bond at maturity is given by the value, at this time, of
a survivor index, denoted by S(T, x). Indeed, they explicitly state (see Section 2.1.2 in [36])
that “ the (T, x)-bond pays the amount S(T, x) at time T ”. On one hand, in Equation (2.1) of
their paper, the authors define the survivor index as
RT
(2.5) S(T, x) = e− 0 µ(t,x+t)dt
where µ(t, x + t) is a (possibly stochastic) sport force of mortality. The value of such a survivor
index is necessarily strictly positive. On another hand, in Equation (2.2) of their paper, the
1In other words the intensity process of 1
{τ ≤t} is given by 1{τ >t− } µ(t, x).
authors write
EP [Yx (T ) |MT ] = S(T, x)
where MT is the information “generated by the evolution of the term structure of mortality [...]
up to time t” and where Yx (T ) is a random variable taking the value 1 if a given individual is
alive at time T and 0 if not. The authors are not really prolix on how the filtration (Mt )0≤t≤T
is defined but we can safely say that the authors, at least implicitly, assume the random variable
Yx (T ) is not measurable with respect to MT . Indeed, otherwise, the index could take the value
0, which would contradict the first definition (Equation (2.5)). In other words, the information
set MT does not include whether the given individual is alive or not at time T .
Alternatively, if we consider a population with initially n0 (x) individuals with all the same
age, we can denote the actual proportion of individual alive in this population at time T by
Pn0 (x)
Y i (T )
IT (x) := i=1n0 (x)x where Yxi (T ) is a random variable taking the value 1 if the ith individual
is alive at time T and 0 if not. We can then write:
hP i
n0 (x) i
EP i=1 Yx (T ) |MT
EP [IT (x) |MT ] =
n0 (x)
Pn0 (x) i
i=1 EP Yx (T ) |MT
=
n0 (x)
n0 (x)S(T, x)
=
n0 (x)
= S(T, x)
where in the third equality we assume that the
h population
i is homogeneous from the point of
i j
view of mortality i.e. EP Yx (T ) |MT = EP Yx (T ) |MT = S(T, x) for all i, j = 1, . . . n0 (x).
Since we know the random variable Yxi (T ) are not measurable with respect to MT , we also know
S(T, x) cannot be equal to IT (x). These equations clearly show S(T, x) can be understood as
the expected (given some information MT at time T ) proportion of individuals alive at time
T.
But why choose as a payoff, the EXPECTED proportion at time T when, at time T , you
know the ACTUAL proportion of individuals alive up to time T ? Moreover, this definition of
the payoff of a longevity bond contradicts the one made in the intensity models literature (see
for example Dahl [37] (2004), Biffis [29] (2005) , Luciano and Vigna [64] (2005), Schrager [74]
(2006), etc.) where they explicitly consider the actual proportion of the population still alive
at time T as their payoff, i.e. IT (x) instead of S(T, x). Notice that we have
E [IT (x)] = E [E [IT (x) |MT ]]
= E [S(T, x)]
However, it does not mean IT (x) = S(T, x).
Again, one may say it is just a matter of definition. But actually it is not true since here,
Cairns et al. ’s definition leads to non-sensical conclusions. For example, assume the survival
probabilities are deterministic, then in this case S(T, x) is deterministic. According to their
definition, the payoff of a longevity bond is also deterministic (since it is equal to this S(T, x)).
In this case, according to their definition, a longevity bond is thus just a simple risk free zero-
coupon bond. It is not realistic: the payoff of a longevity bond is random whether the survival
probabilities are deterministic or not.
To conclude, we clearly see the payoffs of Cairns et al.’s longevity bonds are not correctly
defined. The trouble comes from the fact the authors do not properly take into account the
actual mortality in the definition of their longevity bond payoffs. Actually, they do not even
model this actual mortality at all.
2.3. Bauer’s Paper. As far as Bauer’s paper [18] is concerned, even though he claims
to take into account, in his longevity bonds payoffs, the actual proportion of individuals alive
in a given population it is not true. In order to explicitly model this actual mortality, one
has to necessarily introduce a more or less general point process (as Miltersen and Persson did
actually). However, nowhere does Bauer explicitly describe such point process and accordingly
nowhere does he explicitly model the actual proportion of the population still alive.
(T )
Moreover, page 4 in [18], Bauer denotes by T −τ pxτ the “realized proportion of the popula-
tion still alive at time τ and xτ years old at time τ , who are still alive at time T ”. We can for
(T )
example consider T px0 which is, according to this definition, the actual proportion in a given
population of x0 years old at time 0 , who are still alive at time T . But at p.12 in [18], Bauer
writes
∂ (t)
t px0 = −µ̃t (0, t, x0 )t p(t)
x0
∂t
The solution of this differential equation is then given by
Rt
(t) − µ̃s (0,s,x0 )ds
t px0 = e 0
where µ̃s (0, s, x0 ) is simply the spot force of mortality at time s (see the second point in
Definition 3.4 page 7 in [18]). Accordingly, Bauer uses the exact same definition than in Cairns
et al’s paper (see Equation (2.5) above). Again, Bauer makes the same confusion than Cairns
et al.
By the way, Bauer himself recognizes his notation are very “cryptic”. Actually, all these
complicated notation are not necessary and they simply translate the fact that his model is not
well-defined from the start.
To conclude, we clearly see none of these three papers properly applies the HJM methodol-
ogy to the modelling of longevity bonds prices. As far as Cairns et al.’s and Bauer ’s papers are
concerned, the authors failed to define the payoff of a longevity bond as the actual proportion
of individual alive in a given population at the bond’s maturity. The reason for this is simple:
they do not model this actual mortality at all but instead model its expectation. As far as
Miltersen and Persson’s paper is concerned, the authors are well aware they have to model the
actual mortality. Unfortunately, they did not properly take it into account in their definition
of the longevity bond prices.
3. The HJM Methodology For Longevity Bonds.

3.1. The theoretical setting. As explained in the introduction, the payout of a longevity
bond at maturity T is given by the value at time T , of a given survivor index. The value of
a survivor index at any time t, is given by the proportion of people still alive at time t, in
an initial given age group of an overall population. In this chapter, we study survivor indices
defined for different age groups of the same overall population. In order to model the values
of these survivor indices, in a realistic way, we have to explicitly model the number of deaths
in the overall population and, more precisely, the number of deaths in each of the different age
groups. Moreover, in Section 4, we study the asset allocation problem of an insurance company
in the presence of different basis risks. In particular, there exists an important basis risk coming
from the fact that, even for a given age group, the policyholders of a given insurance company
represent only a fraction of the population on which the survivor index is defined. In order to
study, in a rigorous manner, the asset allocation problem in the presence of such basis risk, we
have to explicitly model the number of deaths not only in a given age group but also in a given
insurance company. We formalize these ideas in this section.
We consider an overall population consisting of policyholders of different ages and splitted
up among different insurance companies. Let us define a measurable space (K, K) where K
represents the set of all these insurance companies. We assume the size of this set is finite.
K is the set P(K) of all subsets of K. Let us also define a measurable space (X, X ) where
X is the subset of R+ representing the range of admissible values at time 0, for the age of
the policyholders. For example, we can consider a finite space X = {x1 , x2 , . . . , xa } where the
xi are the different ages of the policyholders at time 0, or we can consider an infinite space
X = [xmin , xmax ]. The σ-algebra X is assumed to be the Borel σ-algebra: X = B(X). We
denote by (E, E) the product space of these two spaces, i.e. E = X × K and E = X ⊗ K.
Let us define a probability space (Ω, F, P ). We assume a d-dimensional process Wt and
a E-marked point process (Tn , Zn )n≥1 are defined on this space. See, for example, Brémaud
[33] (1981) for an introduction to marked point processes. (Tn )n≥1 is a sequence of real-valued
random variables representing the times of death of the policyholders in the whole population.
(Zn )n≥1 is a sequence of random variables taking their values in the space E. Given the death
of a policyholder at time Tn , the random variable Zn describes the age of this policyholder and
the insurance company to which he or she belongs.
For each A ∈ E, we define the counting process N (t, A) by
∞
X
N (t, A) = 1{Tn ≤t} 1{Zn ∈A}
n=1
for t ≥ 0. It represents the number of deaths up to time t associated with the set A. For each
A ∈ E, we denote by n0 (A) the initial number of policyholders belonging to the set A. Since
we necessarily have N (t, A) ≤ n0 (A) < ∞, this marked point process is bounded and is thus
nonexplosive and integrable.
We define the filtration G = (Gt )t≥0 as the one generated by Wt and (Tn , Zn )n≥1 i.e.
Gt = σ (Ws : 0 ≤ s ≤ t) ∨ σ (N (s, A) : 0 ≤ s ≤ t, ∀A ∈ E). We assume Wt is a d-dimensional
(P, G)-Brownian motion. For each A ∈ E, we denote by Λ(t, A) the (P, G)-compensator of the
202 The HJM Methodology For Longevity Bonds.
counting process N (t, A) i.e. Λ(t, A) is the unique (P, G)-predictable increasing process such
that q(t, A) := N (t, A) − Λ(t, A) is a (P, G)-martingale. For each A ∈ E, the compensator
Λ(t, A)(ω) = 0 a.s. on the set {(t, ω) : N (t, A)(ω) = n0 (A)}.
We also assume there exists a (P, G)-kernel intensity for N (t, ·) i.e. there exists a measure-
RtR
valued process νt (de) is such that Λ(t, A) = 0 A νs (de)ds for each A ∈ E and each t ≥ 0.
Notice that the assumptions on the distribution of the times of death in the population are
rather weak since we only assume that they are totally inacessible stopping times and there
is no simultaneous death in the population. For example, we do not assume the policyholders
in a given set A ∈ E, necessarily have homogeneous intensities of mortality. To the best of
my knowledge, it generalizes all the existing intensity models defined on a Brownian filtration,
described in the literature. In these models, for all k ∈ K and for a given age x ∈ X, the
intensity kernel νt ((x, k)) can be written as
νt ((x, k)) = [n0 ((x, k)) − N (t, (x, k))] µ(t, x + t)
where µ(t, x + t) is the “force of mortality” at time t for an individual of age x + t. For example,
in Dahl [37] (2004), this force of mortality follows an extended Cox-Ingersoll-Ross model
p
dµ(t, x + t) = (β µ (t, x) − γ µ (t, x)µ(t, x + t)) dt + ρµ (t, x) µ(t, x + t)dWt
Luciano and Vigna [64] (2005) suggest various non-reverting models such as2
dµ(t, x + t) = aµ(t, x + t)dt + σdWt
or
p
dµ(t, x + t) = aµ(t, x + t)dt + σ µ(t, x + t)dWt
3.2. The survivor indices. In this chapter, we want to introduce multiple survivor in-
dices and define a term structure of longevity bonds on each of these survivor indices. As
already explained, the payout of a longevity bond at maturity, is given by the value of a sur-
vivor index at this time. The value of this survivor index at any times t, is given by the
proportion of people still alive at time t, from an initial age group of the overall population.
Accordingly, we are particularly interested in counting processes of the form N (t, B × K) where
B ∈ X , since it represents the number of deaths in the age group B of the overall population.
In order to define several survivor indices, we consider a finite number of disjoints subsets of X
denoted Bm with m = 1, . . . M , such that each Bm ∈ X . For each Bm , m = 1, . . . M , we can
now define the value of a survivor index It (Bm ) as
n0 (Bm × K) − N (t, Bm × K)
It (Bm ) =
n0 (Bm × K)
where n0 (Bm × K) is the initial number of policyholders in the age group Bm of the overall
population. We thus have defined a M -dimensional survivor index. At time 0, we obviously
have I0 (Bm ) = 1 for each m = 1, . . . M .
In the following sections, we will define a term structure of longevity bonds for each index.
We will then derive not only the no-arbitrage condition between the risk-free zero-coupon bonds
2They also suggest to adding a jump component to these models.
and the longevity bonds but also between the longevity bonds associated with different survivor
indices.
3.3. The risk-free zero-coupon bonds. In this section, we define the prices of the risk-
free zero-coupon bonds. As usual in the HJM framework, we assume there is a zero-coupon
bond for each maturity T ∈ [0, T ∗ ]. The price at time t of a zero-coupon bond of maturity
T ≥ t, is denoted B(t, T ). Following the HJM methodology, we assume the dynamics of the
zero-coupon bond prices can be described through the dynamics of the instantaneous forward
rates fn (t, s) defined by
∂lnB(t, y)
(3.1) fn (t, s) = −
∂y |y=s
So for each t, we have

RT
B(t, T ) = e− t fn (t,s)ds
In this chapter, we assume the forward rates follow, for each s ≥ t, the stochastic differential
equation
Z t Z t
fn (t, s) = fn (0, s) + αn (u, s)du + σn (u, s)dWu
0 0
Z tZ
(3.2) + ζn (u, s, e)N (du, de)
0 E
where αn (t, s), σn (t, s) are respectively a 1-dimensional and a d-dimensional G-predictable
processes and where ζn (t, s, e) is a 1-dimensional G-predictable E-marked process. We assume
all the required technical integrability conditions hold such that the bond prices are well defined.
Here, we assume the prices of these bonds not only depend on the (P, G)-Brownian motion
Wt as it is usual in the literature, but can also depend on the actual mortality (through the
integral with respect to N (t, e) in the dynamics of the forward rates). This assumption is
maybe not very realistic but we prefer to keep it for two reasons. First of all, we will need
similar calculations for the longevity bonds and it is easier for us to introduce these calculations
at this point. Secondly, removing this assumption does not lead to significantly easier technical
results. If this is considered as clearly too unrealistic, the reader can very easily remove this
assumption by setting ζn (t, s, e) = 0 for all t, s and e.
The following proposition gives the dynamics of the zero-coupon bond prices:
Proposition 3.1. If the forward rates follow Equation (3.2), then the price of a zero-coupon
bond with maturity T is given by the solution of the stochastic differential equation
 h h ii 
0
 fn (t, t) − αn∗ (t, T ) + 21 σn∗ (t, T ) σn∗ (t, T ) dt 
 
(3.3) dB(t, T ) = B(t− , T )
−σn∗ (t, T )dWt + E ψn∗ (t, T, e)N (dt, de)

 R 

where
Z T
αn∗ (t, T ) = αn (t, s)ds
t
Z T
σn∗ (t, T ) = σn (t, s)ds
t
∗ (t,T,e)
ψn∗ (t, T, e) = e −ζn
−1
Z T
∗
ζn (t, T, e) = ζn (t, s, e)ds
t
3.4. The longevity bonds. In this section, we define the prices of the longevity bonds.
In addition to the risk-free term structure, we assume that there are M longevity bond term
structures, one for each survivor index. A zero-coupon (T, Bm )-longevity bond is a financial
security that pays, at time T , the value of the survivor index associated with the age group
Bm :
n0 (Bm × K) − N (T, Bm × K)
IT (Bm ) =
n0 (Bm × K)
As usual in the HJM setting, we assume there is, for each Bm , m = 1, . . . M , a zero-coupon
longevity bond for each maturity T ∈ [0, T ∗ ]. The price at time t of such a zero-coupon
longevity bond with maturity T , is P (t, T, Bm ). Following the HJM methodology, we assume
the dynamics of zero-coupon longevity bond prices can be described through the dynamics of
some forward rates. More precisely, we define the price at time t of a (T, Bm )-longevity bond
as
RT
(3.4) P (t, T, Bm ) = It (Bm )e− t fl (t,s,Bm )ds
where the processes fl (t, s, Bm ) are called longevity forward rates and are described below
in more detail. Notice that these longevity forward rates depend on the survivor index on
which the longevity bond is defined. This definition has several interesting features. First, it
ensures that the price at maturity of any longevity bond is equal to the value of its payout i.e.
P (T, T, Bm ) = IT (Bm ) for each T and each Bm . Secondly, we have
P (t, U, Bm ) RU
= e− T fl (t,s,Bm )ds
P (t, T, Bm )
Accordingly, these forward rates can then be written as
∂lnP (t, y, Bm )
f (t, s, Bm ) = −
∂y |y=s
This equation shows that the definition of the longevity forward rates is the exact counterpart
to the definition of the risk-free forward rates (see Equation (3.1)). Thirdly, it allows to seperate
the effects of the actual mortality on the price of a longevity bond, that we discussed in the
introduction. As we already explained, the first is a purely mechanical and proportional effect
due to the actual mortality in the associated survivor index. The term It (Bm ) in Equation (3.4)
translates this first effect. We canR then introduce, independently of the first, an additional effect
T
by allowing the second term e− t fl (t,s,Bm )ds to depend on the actual mortality in the whole
population. More precisely, we assume the longevity forward rates fl (t, s, Bm ) are described
for each s ≥ t, by the equation
Z t Z t
fl (t, s, Bm ) = fl (0, s, Bm ) + αl (u, s, Bm )du + σl (u, s, Bm )dWu
0 0
Z tZ
(3.5) + ζl (u, s, Bm , e)N (du, de)
0 E
where, for each Bm , (m = 1 . . . M ) , αl (t, s, Bm ) and σl (t, s, Bm ) are respectively a 1-dimensional
and a d-dimensional G-predictable process, and where ζl (t, s, Bm , e) is a 1-dimensional G-
predictable E-marked process. Notice that we let the forward rates depend on the actual
mortality in the whole set E and not only in the subset Bm × K. In other words, this defini-
tion also allows the price of a longevity bond to depend on the actual mortality in the whole
population and not only on the actual mortality in the associated survivor index.
Here again, we assume that all the required technical integrability conditions hold such that
− tT fl (t,s,Bm )ds
R
the longevity bond prices are well defined. In the following, we will denote e by
Z(t, T, Bm ).
The following proposition gives the dynamics of the zero-coupon longevity bond prices.
Proposition 3.2. If the longevity forward rates follow Equation (3.5) then the price of a
zero-coupon (T, Bm )-longevity bond is given by the solution of the stochastic differential equation
[fl (t, t, Bm ) − αl∗ (t, T, Bm )] dt
 

 


 

 
1 ∗ 0 ∗
(3.6)dP (t, T, Bm ) = P (t− , T, Bm ) + 2 σl (t, T, Bm ) σl (t, T, Bm )dt

 

 
∗ ∗
 R 
−σl (t, T, Bm )dWt + E ψl (t, T, Bm , e)N (dt, de)
 
where
Z T
αl∗ (t, T, Bm ) = αl (t, s, Bm )ds
t
Z T
σl∗ (t, T, Bm ) = σl (t, s, Bm )ds
t
−ζl∗ (t,T,Bm ,e)
ψZ∗ (t, T, Bm , e) = e −1
∗
e−ζl (t,T,Bm ,e)
ψl∗ (t, T, Bm , e) = ψZ∗ (t, T, Bm , e) − 1{e∈Bm ×K}
(n0 (Bm × K) − N (t− , Bm × K))
Z T
ζl∗ (t, T, Bm , e) = ζl (t, s, Bm , e)ds
t
Proof. Following the same lines as in the proof of Proposition 3.1, we can easily show
that for each Bm , m = 1 . . . M and each T ∈ [0, T ∗ ], the process Z(t, T, Bm ) is the solution of
the stochastic differential equation

[fl (t, t, Bm ) − αl∗ (t, T, Bm )] dt
 

 


 

 
0
dZ(t, T, Bm ) = Z(t− , T, Bm ) + 12 σl∗ (t, T, Bm ) σl∗ (t, T, Bm )dt

 

 
−σl∗ (t, T, Bm )dWt ∗
 R 
+ E ψZ (t, T, Bm , e)N (dt, de)
 
∗
where ψZ∗ (t, T, Bm , e) = e−ζl (t,T,Bm ,e) − 1. Using the integration-by-parts formula for stochastic
processes, we can find the price of a zero-coupon longevity bond P (t, T, Bm ) is given by
dP (t, T, Bm ) = It− (Bm )dZ(t, T, Bm )
+ Z(t− , T, Bm )dIt (Bm )
+ d [I· (Bm ), Z(·, T, Bm )]t
= It− (Bm )dZ(t, T, Bm )
1
− Z(t− , T, Bm )N (dt, Bm × K)
n0 (Bm × K)
1
− d [N (·, Bm × K), Z(·, T, Bm )]t
n0 (Bm × K)
The square bracket is given by
Z
d [N (·, Bm × K), Z(·, T, Bm )]t = Z(t− , T, Bm ) ψZ∗ (t, T, Bm , e)N (dt, de)
Bm ×K
Eventually we get
[fl (t, t, Bm ) − αl∗ (t, T, Bm )] dt
 

 


 

 
0
 
+ 12 σl∗ (t, T, Bm ) σl∗ (t, T, Bm )dt

 


 

dP (t, T, Bm ) = P (t− , T, Bm )
−σl∗ (t, T, Bm )dWt + E ψZ∗ (t, T, Bm , e)N (dt, de)
R

 


 


 

−ζl∗ (t,T,Bm ,e)

 

− Bm ×K (n0 (Bme×K)−N
R
(t− ,Bm ×K)) N (dt, de)

 

3.5. The no-arbitrage condition. In this section, we examine the condition for the no-
arbitrage hypothesis to hold in the financial market described above. This condition ensures
there is no-arbitrage (1) between the risk-free zero-coupon bonds, (2) between the longevity
bonds associated with the same survivor index, (3) between the longevity bonds and the risk-
free zero-coupon bonds and (4) between the longevity bonds associated with different survivor
indices. We assume there is a locally risk-free asset Bt which grows at rate rt = fn (t, t). We
first need the following theorem.
Theorem 3.3. Any probability measure Q equivalent to P on (Ω, F) has a density ηt =

dQ
dP |Gt that is solution of
Z
w N

dηt = ηt− −βt dWt + β (t, e) − 1 q(dt, de) , η0 = 1
E
for a d-dimensional G-predictable process βtW and a 1-dimensional G-predictable E-marked

process β N (t, e) > 1 and such that E P [ηT ∗ ] = 1. The processes WtQ and q Q (t, A) defined below,
are respectively a d−dimensional (Q, G)-Brownian motion and, for each A ∈ E, a (Q, G)-local
martingale:
Z t
Q
Wt = Wt + βuW du
0
Z tZ
q Q (t, A) = q(t, A) − β N (u, e) − 1 νu (de)du

0 A
Z tZ
= N (t, A) − β N (u, e)νu (de)du
0 A
for all A ∈ E.
Proof. See Aase [1] (1988).
The next proposition gives the no-arbitrage condition.
Proposition 3.4. There is no arbitrage if there exist an G-predictable process βtW and an
G-predictable E-marked process β N (t, e) > 1, such that for each T and each Bm , m = 1, . . . M ,
we have simultaneously
1h ∗ 0
i
αn∗ (t, T ) = σn (t, T ) σn∗ (t, T ) + σn∗ (t, T )βtW
Z2
+ ψn∗ (t, T, e)β N (t, e)νt (de)
E
and
1h ∗ 0
i
αl∗ (t, T, Bm ) = σl (t, T, Bm ) σl∗ (t, T, Bm ) + σl∗ (t, T, Bm )βtW
Z2
+ ψl∗ (t, T, Bm , e)β N (t, e) νt (de)
E
N
R
Bm ×K β (t, e) νt (de)
+
(n0 (Bm × K) − N (t− , Bm × K))
Proof. For the no-arbitrage condition to hold, we have to find the conditions of existence
of an equivalent measure, Q, such that B(t, T )/Bt and P (t, T, Bm )/Bt are (Q, G)-local mar-
tingales for each T ∈ [t, T ∗ ] and each Bm simultaneously. We can write the discounted prices
of these assets as
 h h ii 
f (t, t) − r − α ∗ (t, T ) + 1 σ ∗ (t, T )0 σ ∗ (t, T ) dt


 n t n 2 n n




 

B(t, T ) B(t− , T )  
d = + σn∗ (t, T )βtW + E ψn∗ (t, T, e)β N (t, e) νt (de) dt
R
Bt Bt 
 


 

−σn∗ (t, T )dWtQ + E ψn∗ (t, T, e)q Q (dt, de)

 R 

and
[fl (t, t, Bm ) − rt − αl∗ (t, T, Bm )] dt
 

 


 

 0

+ 21 σl∗ (t, T, Bm ) σl∗ (t, T, Bm )dt

 

 
P (t, T, Bm ) P (t− , T, Bm )  
d =
Bt Bt 
 + σl∗ (t, T, Bm )βtW + E ψl∗ (t, T, Bm , e)β N (t, e) νt (de) dt
R 


 


 

 
−σl∗ (t, T, Bm )dWtQ + E ψl∗ (t, T, Bm , e)q Q (dt, de)

 R 

For these prices to be (Q, G)-local martingales, the drift should be equal to 0 P -a.s for all T
and all Bm . Since fn (t, t) = rt , the first condition is immediate. As far as the second condition
is concerned, βtW and β N (t, e) should be such that
1h ∗ 0
i
αl∗ (t, T, Bm ) = fl (t, t, Bm ) − rt + σl (t, T, Bm ) σl∗ (t, T, Bm )
2
Z
∗
+ σl (t, T, Bm )βt +W
ψl∗ (t, T, Bm , e)β N (t, e) νt (de)
E
where
∗
e−ζl (t,T,Bm ,e)
ψl∗ (t, T, Bm , e) = ψZ∗ (t, T, Bm , e) − 1{e∈Bm ×K}
(n0 (Bm × K) − N (t− , Bm × K))
If we let T → t, the drift condition leads to the equality
Z
fl (t, t, Bm ) − rt = − ψl∗ (t, t, Bm , e)β N (t, e) νt (de)
E
N
R
(3.7) =
(n0 (Bm × K) − N (t, Bm × K)− )
By rearranging the different terms, we get the result.
Remark 3.5. The assumption that the drifts of the forward rates fn (·, ·) and the drifts of
the longevity forward rates fl (·, ·, Bm ) are absolutely continuous with respect to the Lebesgue
measure, implies the (P, G)-compensator of N (t, ·) is itself absolutely continuous. In other
words, this assumption implies the (P, G)-kernel intensity of N (t, ·), νt (de) exists. It justifies
the assumption we made in Section 3.1 on the existence of this kernel intensity.
Notice in the previous proof, we show the absence of arbitrage leads to the equality
N
R
fl (t, t, Bm ) − rt =
(n0 (Bm × K) − N (t− , Bm × K))
νtQ (Bm × K)
=
(n0 (Bm × K) − N (t− , Bm × K))
The right-hand term is the average (Q, G)-intensity of mortality for the individuals in the
category age Bm at time t. The left-hand term is the spread of the longevity forward rates
over the instantaneous risk-free rate when T → t. This equality generalizes the result on page
4 of Miltersen and Persson [67] (2005) where, when T → t, they show that what they called
the forward force of mortality rate is equal to the spot force of mortality rate.
The no-arbitrage condition leads to the following representation for the forward rates:
Proposition 3.6. If there is no arbitrage, the forward rates of the zero-coupon bonds are
given by
Z s
0
fn (s, u) = fn (0, u) + σn (v, u) σn∗ (v, u)dv
0
Z sZ
− ζn (v, u, e)ψn∗ (v, u, e)β N (v, e) νv (de)dv
Z0 s E Z sZ
Q
+ σn (v, u)dWv + ζn (v, u, e)q Q (dv, de)
0 0 E
and the longevity forward rates of the longevity bonds are given for each Bm , by
Z s
0
fl (s, u, Bm ) = fl (0, u, Bm ) + σl (v, u, Bm ) σl∗ (v, u, Bm )dv
0
Z sZ
− ζl (v, u, Bm , e)ψl∗ (v, u, Bm , e)β N (v, e)νv (de)dv
0 E
Z s Z sZ
Q
+ σl (v, u, Bm )dWv + ζl (v, u, Bm , e)q Q (dv, de)
0 0 E
Proof. Deriving the no-arbitrage condition with respect to T , we also find the equivalent
set of conditions:
0
αn (t, T ) = σn (t, T ) σn∗ (t, T ) + βtW

Z
∗
− ζn (t, T, e)e−ζn (t,T,e) β N (t, e)νt (de)
E
0
αl (t, T, Bm ) = σl (t, T, Bm ) σl∗ (t, T, Bm ) + βtW

Z
∗
− ζl (t, T, Bm , e)e−ζl (t,T,Bm ,e) β N (t, e) νt (de)
E
∗
ζl (t, T, Bm , e)e−ζl (t,T,Bm ,e)
Z
+ β N (t, e) νt (de)
Bm ×K (n0 (Bm × K) − N (t− , Bm × K))
Replacing these conditions in Equations (3.2) and (3.5), we get the result.
The no-arbitrage condition leads also to the following representation for the risk-free zero-
coupon bond prices and the zero-coupon longevity bond prices:
Proposition 3.7. If there is no arbitrage, the price of a risk-free zero-coupon bond with
maturity T is given by:
 rt − σn∗ (t, T )βtW − E ψn∗ (t, T, e) β N (t, e) − 1 νt (de) dt 
 R 
dB(t, T ) = B(t− , T )
−σn∗ (t, T )dWt + E ψn∗ (t, T, e)q(dt, de)
 R 
or
n o
dB(t, T ) = B(t− , T ) rt dt − σn∗ (t, T )dWtQ + E ψn∗ (t, T, e)q Q (dt, de)
R
and the price of a zero-coupon (T, Bm )-longevity bond is given by:

 rt − σl∗ (t, T, Bm )βtW − E ψl∗ (t, T, Bm , e) β N (t, e) − 1 νt (de) dt 
 R 
dP (t, T, Bm ) = P (t− , T, Bm )
−σl∗ (t, T, Bm )dWt + E ψl∗ (t, T, Bm , e)q(dt, de)
 R 
or
n o
dP (t, T, Bm ) = P (t− , T, Bm ) rt dt − σl∗ (t, T, Bm )dWtQ + E ψl∗ (t, T, Bm , e)q Q (dt, de)
R
Proof. Straightforward. Replace the no-arbitrage conditions of Proposition 3.4 and Equa-
tion (3.7) in Equations (3.3) and (3.6).
These equations give the equilibrium risk premiums for the risk-free zero-coupon bonds and
the zero-coupon longevity bonds. These risk premiums consist of two parts: one is related to
the risk coming from the Brownian motion and the other to the risk coming from the actual
mortality. In particular, for the longevity bonds, we can study in more detail, the risk premium
related to the actual mortality. Since we know the actual mortality can have two effects on the
price of the longevity bonds, we can expect to also have two separate risk premiums for this
mortality risk. Indeed, we can write:
N
R
Bm ×K β (t, e) − 1 νt (de)
Z
∗ N

− ψl (t, T, Bm , e) β (t, e) − 1 νt (de) =
(n (B × K) − N (t− , Bm × K))
E
Z 0 m
ψZ∗ (t, T, Bm , e)δ(t, Bm ) β N (t, e) − 1 νt (de)

(3.8) −
E

1{e∈Bm ×K}
where δ(t, Bm ) = 1 − (n0 (Bm ×K)−N (t− ,Bm ×K)). If, as in the intensity models, we assume the
actual mortality has only a mechanical effect on the prices of the longevity bonds or in other
words, if we assume the longevity forward rates do not depend on the actual mortality, then
we can put ζl (u, s, Bm , e) = 0. In this case, the second term vanishes. Accordingly, the first
term is the risk premium due to the mechanical effect of the actual mortality. Intuitively, this
risk premium simply corresponds to the average (Q, G)-intensity of mortality for the age group
on which the longevity bond is defined. It also appears in the existing intensity models. The
second term is thus an additional risk premium due to the second effect of the actual mortality,
which does not appear in the existing intensity models. It is important to notice that in the
general case, we have this additional risk premium. If, in order to find the value of β N (·, ·) , the
risk premium Rof a longevity bond is estimated, the estimate does not necessarily correspond
m ×K
(β N (t,e) −1)νt (de)
to the term (nB0 (B m ×K)−N (t− ,Bm ×K))
, as the “intensity” models suggest, but to a much more
complex expression.
3.6. Alternative representations of zero-coupon bond prices and longevity bond

prices. This section aims to underline the connections between the intensity models and our
HJM model. In Proposition 3.8, we first recall a well-known result in the traditional HJM
model, namely, that the price of a risk-free zero-coupon bond, in our HJM framework, can
also be written as in the short-rate models. In Proposition 3.9, we show that we can derive
a similar result for the longevity bonds: the price of a longevity bond, in our HJM model,
can also be written as in the intensity models. We obtain a more general form that suggests
current intensity models could be extended to more general settings. As a by-product, it also
strengthens our definition of the longevity bond prices. We assume here the discounted prices
are (Q, G)-martingales.
Proposition 3.8. The price at time t of a zero-coupon bond of maturity T , B(t, T ) is
given by
h RT i
B(t, T ) = E Q e− t rs ds |Gt
Proof. The proof is straightforward and can be found in any textbook. We give it for the
sake of completeness. Using the no-arbitrage condition, the discounted value of a zero-coupon
bond with maturity T is a (Q, G)-martingale that can be written as

B(t, T ) B(t− , T ) n o
−σn∗ (t, T )dWtQ + E ψn∗ (t, T, e)q Q (dt, de)
R
d =
Bt Bt
So we have

B(t, T ) Q B(T, T )
= E |Gt
Bt BT
where B(T, T ) = 1.
The price of a longevity bond can also be written as in the intensity models. We have the
following proposition:
Proposition 3.9. The price at time t of a longevity bond with maturity T associated with
the index IT (Bm ) is given by
h RT i
(3.9) P (t, T, Bm ) = It (Bm )E Q e− t ru +κu (Bm )du |Gt
∗
e−ζl (u,T,Bm ,e) β N (u,e) νu (de)
R
Bm ×K
where κu (Bm ) = (n0 (Bm ×K)−N (u− ,Bm ×K)) .
Proof. Using the no-arbitrage condition, for each Bm and each T , the process Z(t, T, Bm )
can be written as
 [rt + κt (Bm )] dt − σl∗ (t, T )dWtQ 
 
dZ(t, T, Bm ) = Z(t− , T, Bm )
+ E ψZ∗ (t, T, Bm , e)q Q (dt, de)
 R 
Rt
Z(t,T,Bm ) − κu (Bm )du
Accordingly, the process Bt e 0 is a (Q, G)-martingale and we thus have

Z(t, T, Bm ) − R t κu (Bm )du Q Z(T, T, Bm ) − 0T κu (Bm )du
R
e 0 = E e |Gt
Bt BT
Since Z(T, T, Bm ) = 1 and P (t, T, Bm ) = It (Bm )Z(t, T, Bm ), we have the result.
This pricing formula looks like the existing ones in the literature on intensity models (see
Biffis [29] (2005), Dahl [37] (2004), etc.). In these models, the price of a longevity bond is
given by the proportion of individuals alive in the survivor index multiplied by an expectation,
under a martingale measure Q, of a modified discounting term rs + µQ (s, x + s), where µQ is
the Q-force of mortality. However, there are some substantial differences.
First of all, in the existing intensity models, the random times of death are assumed to
be conditionally Q-independent with respect to a filtration that does not include these times
of death and which is, accordingly, strictly smaller than our filtration G. Mathematically
speaking, this assumption implies that the expectations that appear in the pricing formula of
these intensity models are taken with respect to this smaller filtration and not with respect to
G, as happens here. Intuitively speaking, as we explained in the introduction, these models are
unable to take into account both effects of the actual mortality on the price of the longevity
bonds. In these models, the actual mortality can only affect the prices of the longevity bonds
through the proportion It (Bm ).
Secondly, if, as in the intensity models, we assume the longevity forward rates cannot
depend on the actual mortality, we can put ζl∗ (u, T, Bm , e) = 0. In this case, we have
N
R
Bm ×K β (u, e) νu (de)
κu (Bm ) =
(n0 (Bm × K) − N (u− , Bm × K))
Intuitively, κu (Bm ) represents the average (Q, G)-intensity of mortality in the age group Bm .
This is similar to the existing “intensity” models except for two things. First, in these intensity
models, the policyholders are assumed to be homogeneous (identical force of mortality) but, as
we already explained, we do not make this assumption in this chapter: the members of the age
group Bm are not necessarily homogeneous. For example, the group Bm could represent the
policyholders between 30 and 40 years old. Even for a given age, we do not actually assume that
the population is necessarily homogeneous. Accordingly, our formula extends the existing one
to the pricing of longevity bonds on heterogeneous population and shows that what is important
is the average intensity of mortality in the chosen population. For example, if we assume that
in the set Bm × K there is a finite number of (disjoint) homogeneous groups of policyholders
(Bm × K)i where Bm × K = ∪i (Bm × K)i with P -intensity of mortality λu ((Bm × K)i ), and
if β N (u, e) is constant on each of these sets, then the Q-intensity of mortality is given by
λQ N
u ((Bm × K)i ) = β (u, (Bm × K)i ) λu ((Bm × K)i ). κu (Bm ) can then be written as the
weighted average of these Q−intensities:
X n0 ((Bm × K) ) − N (u− , (Bm × K) )
κu (Bm ) = i i
λQ
u ((Bm × K)i )
(n0 (Bm × K) − N (u− , Bm × K))
i
Second, in these models, the P -intensity of mortality of a policyholder, given he or she is alive
(the force of mortality), and the change of measure parameter β(·) are not allowed to depend
on the past actual mortality. We do not make these assumptions here.
∗
Thirdly, the term e−ζl (u,T,Bm ,e) does not appear in the existing intensity models literature.
It plays the role of an additional change of measure specific to each longevity bond category
and each maturity. Indeed, in the proof of the last proposition, we showed that the process
Z(t,T,Bm ) − t κu (Bm )du
R
Bt e 0 is a (Q, G)-martingale. For each T and each Bm , we can define a measure
T,Bm
T,Bm T,Bm dQZ
QZ equivalent to Q such that its density ηtZ = dQ |Gt for t ≤ T , is given by
Rt
− 0 κu (Bm )du
T,Bm Z(t, T, Bm )e
ηtZ =
Z(0, T, Bm )
We can also write this density as
Z
T,Bm Z T,Bm
dηtZ = ηt− −σl∗ (t, T )dWtQ + ψZ∗ (t, T, Bm , e)q Q (dt, de)
E
with η0 Z T,Bm = 1. According to Theorem 3.3, for each A ∈ E, the process

T,Bm
Z tZ
QZ
q (t, A) = N (t, A) − (1 + ψZ∗ (u, T, Bm , e)) β N (u, e) νu (de)du
0 A
Z tZ
∗
= N (t, A) − e−ζl (uT,Bm ,e) β N (u, e) νu (de)du
0 A
T,Bm
is a (QZ , G)-localmartingale. Accordingly, κu (Bm ) should be understood as the average
Z T,Bm
(Q , G)-intensity of mortality in the age group Bm . This additional change of measure
translates the additional risk premium we discussed in Equation (3.8).
Notice that, in the HJM methodology, the prices of the bonds are given a priori. In other
words, we do not use Equation (3.9) to derive the prices of the longevity bonds. It is only
a mathematical identity implied by the no-arbitrage condition. However, it suggests current
intensity models could be extended to much more general models without losing the form of
their basic pricing formula.
3.7. Forward martingale measures and longevity forward martingale measures.
In the pricing of interest rate derivatives, it is often useful to rewrite a problem under, what is
called in the financial literature, a forward martingale measure instead of the spot martingale
measure Q. In Section 3.7.1, we give the form of these forward martingale measures in our
setting and show some of their well-known properties. In Section 3.7.2, we present new mar-
tingale measures that are the exact counterparts to these forward measures, but for longevity
bonds. We call these new measures the longevity forward martingale measures. We show they
basically exhibit similar properties to the traditional forward measures. We assume here the
discounted prices of the zero-coupon bonds and the discounted prices of the longevity bonds
are (Q, G)-martingales.
3.7.1. T -forward martingale measure. As in the traditional HJM model, we can, for each
T
T, define a measure QT , equivalent to Q, such that its density ηtT = dQ
dQ |Gt is given for t ≤ T ,
by
Rt
e− 0 ru du
B(t, T )
ηtT =
B(0, T )
We can also write this density as a solution of the differential equation
1 Rt
dηtT = d e− 0 ru du B(t, T )
B(0, T )
Rt
B(t− , T )e− 0 ru du
Z
∗ Q ∗ Q
= −σn (t, T )dWt + ψn (t, T, e)q (dt, de)
B(0, T ) E
Z
= ηt− T
−σn∗ (t, T )dWtQ + ψn∗ (t, T, e)q Q (dt, de)
E
with η0T = 1. In other words, the density ηtT

can be written as the following Doléans exponential:
Z
∗ Q
T
ηt = E(−σn (t, T )dWt + ψn∗ (t, T, e)q Q (dt, de))
E
As in the traditional HJM model, we have the two following well known properties:
Proposition 3.10. Under the forward martingale measure QT , the forward rate f (t, T ) is
a (QT , G)−martingale.
T
Proof. Using Theorem 3.3, we know the process WtQ is a d-dimensional (QT , G)-Brownian
T
motion and q Q (t, A) is a (QT , G)-local martingale where these processes are defined by:
Z t
QT Q
Wt = Wt + σn∗ (u, T )du
0
Z tZ
T
q Q (t, A) = q Q (t, A) − ψn∗ (u, T, e)β N (u, e)νu (de)du
0 A
Z tZ
= N (t, A) − (1 + ψn∗ (u, T, e)) β N (u, e)νu (de)du
0 A
The forward rate f (t, T ) can the be written as
Z t Z tZ
QT T
fn (t, T ) = fn (0, T ) + σn (v, T )dWv + ζn (v, T, e)q Q (dv, de)
0 0 E
which is indeed a T
(Q , G)-martingale.
It is well-known from financial economics that the usefulness of these forward martingale
measures resides in the following proposition:
Proposition 3.11. For any GT -measurable random variable HT , we have

Q Bt T
E HT |Gt = B(t, T )E Q [HT |Gt ]
BT
Proof. We have
E Q ηTT HT |Gt

QT
E [HT |Gt ] =
E Q ηTT |Gt

RT
Q e− 0 ru du B(T,T )
E B(0,T ) HT |Gt
= Rt
e− 0 ru du B(t,T )
B(0,T )
Thanks to the previous proposition, we can obtain another formula for the price of a
(T, Bm )-longevity bond.
Proposition 3.12. The price at time t of a (T, Bm )-longevity bond can be written as
T
P (t, T, Bm ) = B(t, T )E Q [IT (Bm ) |Gt ]
or
T
h RT i
P (t, T, Bm ) = It (Bm )B(t, T )E Q e− t κu (Bm )du
|Gt
Proof. Using Proposition 3.11 with HT = IT (Bm ), we have

h RT i
E Q I (B )e− t ru du |G
T T m t
E Q [IT (Bm ) |Gt ] =
B(t, T )
RT
For the second equality, using Proposition 3.9 , we have for HT = e− t κu (Bm )du
h RT i
P (t, T, Bm ) = It (Bm )E Q e− t ru +κu (Bm )du |Gt
T
h RT i
= It (Bm )B(t, T )E Q e− t κu (Bm )du |Gt
3.7.2. (T, Bm )-longevity forward martingale measure. In this section, we give the counter-
part to the forward martingale measures but for the longevity bonds. Similarly to the previous
subsection, for each Bm and each T , we can define a measure QT,Bm equivalent to Q such that
T,Bm
its density ηtT,Bm = dQdQ |Gt is given for t ≤ T by
Rt
e− 0 ru du
P (t, T, Bm )
ηtT,Bm =
P (0, T, Bm )
We can also write this density as a solution of the following differential equation
1 Rt
dηtT,Bm = d e− 0 ru du P (t, T, Bm )
P (0, T, Bm )
Rt
P (t− , T, Bm )e− 0 ru du
Z
∗ Q ∗ Q
= −σl (t, T, Bm )dWt + ψl (t, T, Bm , e)q (dt, de)
P (0, T, Bm ) E
Z
T,Bm ∗ Q ∗ Q
= ηt− −σl (t, T, Bm )dWt + ψl (t, T, Bm , e)q (dt, de)
E
with η0T,Bm = 1. Again, the density ηtT,Bm

can be written as the following Doléans exponential:
Z
T,Bm ∗ Q ∗ Q
ηt = E −σl (t, T, Bm )dWt + ψl (t, T, Bm , e)q (dt, de)
E
We have the following property which is the counterpart to Proposition 3.10 for longevity
forward rates:
Proposition 3.13. Under the forward martingale measure QT,Bm , the longevity forward
rate fl (t, T, Bm ) is a (QT,Bm , G)-martingale.
T,Bm
Proof. Using Theorem 3.3, we know the process WtQ is a d-dimensional (QT,Bm , G)-
T,B m
Brownian motion and q Q (t, A) is a (QT,Bm , G)-local martingale where these processes are
defined by:
Z t
QT,Bm Q
Wt = Wt + σl∗ (u, T, Bm )du
0
Z tZ
T,B
ψl∗ (u, T, Bm , e)β N (u, e)νu (de)du
m
qQ (t, A) = q Q (t, A) −
0 A
Z tZ
= N (t, A) − (1 + ψl∗ (u, T, Bm , e)) β N (u, e)νu (de)du
0 A
The longevity forward rates of the (T, Bm )-longevity bond can then be written as
Z t Z tZ
QT,Bm T,Bm
fl (t, T, Bm ) = fl (0, T, Bm ) + σl (v, T, Bm )dWv + ζl (v, T, Bm , e)q Q (dv, de)
0 0 E
which shows fl (t, T, Bm ) is indeed a (QT,Bm , G)-martingale.
Notice that the previous proposition was not obvious a priori since the definition of a
longevity bond price given by Equation (3.4) is different from the traditional HJM definition
of a zero-coupon bond price.
The following proposition is the counterpart to Proposition 3.11 for the longevity forward
martingale measures:
Proposition 3.14. For any GT -measurable random variable HT
h RT i T,,Bm
E Q HT (n0 (Bm × K) − N (T, Bm × K)) e− t ru du |Gt = n0 (Bm × K)P (t, T, Bm )E Q [HT |Gt ]
Proof. We have
RT
e− 0 ru du P (T,T,Bm )
EQ P (0,T,Bm ) HT |Gt
T,Bm
EQ [HT |Gt ] = Rt
e− 0 ru du P (t,T,Bm )
P (0,T,Bm )
h RT i
1 E Q (n0 (Bm × K) − N (T, Bm × K)) e− t ru du HT |Gt
=
n0 (Bm × K) P (t, T, Bm )

This proposition shows that the problem of finding the value of a security paying a stochastic
payout HT to each surviving member of a survivor index basically comes down to finding the
expectation of this payout under the associated longevity forward martingale measure.
4. Asset Allocation with Longevity Bonds.

The aim of this section is to study the asset allocation problem for endowment and annuity
portfolios when longevity bonds are traded in the financial market. Longevity bonds give
insurers the opportunity, at least theoretically, to perfectly match their liabilities and reduce
their risk to zero. However, this situation would seldom occur in real life. Most of the time,
an insurer would bear some basis risks since the payouts of the longevity bonds traded in the
financial market would not exactly match either the payouts of the liabilities the insurer wants
to hedge, or their timings. We can distinguish at least 3 sources of possible basis risks:
(1) the maturities of the traded longevity bonds do not match the maturities of the in-
surance contracts;
(2) the age groups of the survivor indices on which the traded longevity bonds are defined,
do not match the age of the policyholders;
(3) the policyholders of an insurance company is only a fraction of the whole population
on which the survivor indices are defined.
To study the asset allocation problem with such basis risks, we chose to apply the risk min-
imization theory. This theory was developed by Föllmer and Sondermann [44] (1986) for a
single payoff and extended by Møller [68] (2001) for payment processes. It has already been
applied in life insurance by Møller [68] (2001), Riesner [72] (2006) and Dahl and Møller [39]
(2006) for example. In particular, Dahl and Møller only studied the risk-minimizing strategies
of life insurance portfolios with longevity risk, when risk-free zero-coupon bonds are traded.
Unlike them, we assume longevity bonds are also traded.
In Section 4.1, we give a brief review of the main results of risk minimization theory (see
the original papers for more details). In Section 4.2, we study the asset allocation for portfolios
of pure endowments and annuities when the set of policyholders of these portfolios exactly
corresponds to the population of one of the survivor indices on which a longevity bond is
defined. As a by-product of the risk-minimizing strategies calculation, we derive the market
values of such portfolios of pure endowments and annuities in, respectively, Propositions 4.2
and 4.5. In Section 4.3, under an additional assumption, we study the asset allocation for
218 Asset Allocation with Longevity Bonds.
portfolios of pure endowments and annuities when the set of policyholders of these portfolios
only represents a fraction of the population of one of these survivor indices. We also give the
market values of such portfolios in Propositions 4.11 and 4.14.
4.1. A review of the risk-minimization theory. Let us consider an arbitrary filtration

G. X is the s-dimensional vector of the discounted prices of the financial securities used for
trading. X corresponds to a vector of s1 discounted prices of risk-free zero-coupon bonds and of
s2 zero-coupon longevity bonds. We assume a specific martingale measure Q has been chosen
so that X is a (local) martingale under Q.
A trading strategy ρ is a pair of processes (ε, η) where η is a 1-dimensional real-valued
G-adapted process and ε is an s-dimensional real-valued G-predictable process. The process
εt represents the amount of risky assets held at time t, and the process ηt is the discounted
amount invested in the instantaneously risk-free asset. The discounted value process Vt (ρ) of a
trading strategy ρ, is defined by Vt (ρ) = εt Xt + ηt for 0 ≤ t ≤ T ∗ . This value process represents
the discounted value of the insurer’s financial portfolio following the trading strategy ρ. This
portfolio is not necessarily self-financed.
The liabilities of the insurer are modelled as a process (At )0≤t≤T ∗ . The process At is as-
sumed to be càdlàg, G-adapted and square integrable. The process At represents the discounted
value of the cumulative payments up to time t. Rt
The cumulative cost process Ct (ρ) of a strategy ρ, is defined by Ct (ρ) = Vt (ρ) − 0 εu dXu +
At . It represents the discounted value of the portfolio reduced by the discounted trading gains
and including the discounted net payments to the policyholders. The initial cumulative cost
process C0 (ρ) is given by C0 (ρ) = V0 (ρ) + A0 . The first term V0 (ρ) represents the initial
amount an insurer needs to create a financial portfolio. The second term is the initial payment
the insurer has to pay to the policyholders. A strategy
h ρ is called risk-minimizing,
i if it minimizes
Q 2
the risk process Rt (ρ) defined by Rt (ρ) = E (CT ∗ (ρ) − Ct (ρ)) |Gt , for every t and for every
strategy such that VT ∗ (ρ) = 0 , Q − a.s.
Föllmer and Sondermann [44] (1986) for a single payoff and Møller [68] (2001) for a payment
process, show that the solution of this problem is closely related to the so-called Galtchouk-
Kunita-Watanabe decomposition of the square integrable martingale E [AT ∗ |Gt ]:
Z t
Q
(4.1) E [AT ∗ |Gt ] = E [AT ∗ |G0 ] + εA A
u dXu + Lt Q − a.s.
0
where εA A
u is an s-dimensional predictable process and Lu is a zero-mean martingale strongly
orthogonal to (the stable subspace generated by) X. With these notation, we can give the
solution of the risk minimization problem in the following theorem:
Theorem 4.1. The unique risk-minimizing RM-strategy ρ∗ = (ε∗ , η ∗ ) of At is given by

ε∗ = εA and ηt∗ = E Q [AT ∗ − At |Gt ] − ε∗t Xt . The cumulative cost process of ρ∗ is given by
Ct (ρ∗ ) = E Q [AT ∗ |G0 ] + LA ∗ A ∗ ∗
Q
E [AT ∗ − At |Gt ].
From a more practical point of view, the predictable integrand εA

t of the Galtchouk-Kunita-
Watanabe can be found in the following way. If we denote Jt (ρ∗ ) = E Q [AT ∗ |Gt ], we can write
s Z .
* +
X
j Q i i j

J. , X. t = E [AT ∗ |G0 ] + εu dXu + L. , X.
i=1 0
t
s Z t
X
εiu d X.i , X.
j

= u
i=1 0
Accordingly, the integrand εA 1 2 s

t = εt , εt , . . . , εt can be found by solving the system of equations
d
X·1 , X·1 t d
X·2 , X·1 t . . . d
X·s , X·1 t

d
J· , X·1 t

ε1t
    
 d J· , X 2 1 2 2 2 . . . d X·s , X·2 t   ε2t 
 =  d X· , X · t d X· , X · t
 
 · t  
 ... 
...
...
... ... ...
  
d hJ· , X·s it d X·1 , X·s t d X·2 , X·s t . . . d hX·s , X·s it εst
4.2. Risk-minimization for the population of a survivor index. In this section, we

study the risk-minimizing strategies of a portfolio of pure endowments and annuities, subscribed
by a set of policyholders that exactly corresponds to the population of a survivor index. In
Section 4.3, we will study the risk-minimizing strategies for portfolios of endowments and
annuities subscribed by the policyholders of a given insurance company k. Mathematically
speaking, this means that, in this section, we want to hedge the mortality risk of a set of
policyholders of the type Bm × K whereas in the next section, we want to hedge the mortality
risk of a set of policyholders of the type Bm × k where k ∈ K.
4.2.1. Pure endowments. In this section, we want to find the risk-minimizing hedging strat-
egy of a portfolio of pure endowments with maturity T . We assume this portfolio is subscribed
by a set of policyholders of the type Bm × K where Bm is the age group on which a survivor
index is defined. The discounted payment at time T is given by
gT (n0 (Bm × K) − N (T, Bm × K))
AT =
BT
where gT is deterministic. To derive the risk-minimizing strategies, we first need the following
propositions. The first one gives the market value of such an endowments portfolio.
Proposition 4.2. The discounted value of the risk-minimizing portfolio is given by
Vtg (ρ∗ ) = n0 (Bm × K)gT P (t, T, Bm )
Proof. Straightforward since gT is deterministic.
h RT i
Vtg (ρ∗ ) = E Q gT (n0 (Bm × K) − N (T, Bm × K)) e− 0 rs ds |Gt

1 Q (n0 (Bm × K) − N (T, Bm × K)) − R T rs ds
= n0 (Bm × K)gT E e t |Gt
Bt n0 (Bm × K)
P (t, T, Bm )
= n0 (Bm × K)gT
Bt
Proposition 4.3. The martingale representation of Jtg (ρ∗ ) is given by
Jtg (ρ∗ ) = n0 (Bm × K)gT P (0, T, Bm )

Z t
P (s− , T, Bm ) ∗
− n0 (Bm × K)gT σl (s, T, Bm )dWsQ
0 Bs
Z tZ
P (s− , T, Bm ) ∗
+ n0 (Bm × K)gT ψl (s, T, Bm , e)q Q (ds, de)
0 E B s
Proof. We have Jtg (ρ∗ ) = Vtg (ρ∗ ) and
P (t, T, Bm )
Vtg (ρ∗ ) = n0 (Bm × K)gT
Bt
Z t
P (s, T, Bm )
= n0 (Bm × K)gT P (0, T, Bm ) + n0 (Bm × K)gT d
0 Bu
We get the result since

Z
∗ Q ∗ Q
dP (s, T, Bm ) = P (s− , T, Bm ) rs ds − σl (s, T, Bm )dWs + ψl (s, T, Bm , e)q (ds, de)
E
The risk-minimizing strategy depends on the available securities. If we assume that a

(T, Bm )-longevity bond exists then we can perfectly hedge this portfolio with a simple “buy
and hold” strategy by buying n0 (Bm × K)gT longevity bonds with maturity T . So as already
explained, the interesting situations are when:
(1) there are longevity bonds defined on the same age group Bm but not for the maturity
T;
(2) there are longevity bonds with the same maturity T but not defined on the same age
group;
(3) there are no longevity bonds defined on the same age group or the same maturity.
Obviously, even though we assume we cannot directly trade (T, Bm )-longevity bonds, we nev-
ertheless assume we can find good estimates of their values and volatilities. This is a standard
assumption (at least implicitly) when the HJM methodology is applied. From now on, all the
propositions will directly treat the third case above.
We assume we can trade in s1 risk-free zero-coupon bonds of different maturities denoted
B(t, Ti ) with i = 1, . . . s1 , and in s2 longevity bonds of different maturities and different age
groups denoted P (t, Ti , Bi ) with i = 1, . . . s2 , such that Bi ∈ (B1 , . . . , BM ). s1 and s2 are such
that s1 + s2 = s . The risk-minimizing strategies are given by the following propositions:
strategy ρ∗ = (ε∗t , ηt∗ ) = ε1t , . . . , εst , ηt∗ is given

Proposition 4.4. The risk-minimizing
by the solution of the system of equations

1 1
d
X·2 , X·1 t . . . d
X·s , X·1 t

d
J· , X·1 t


ε1t
 
d
X · , X · t
 d J· , X 2   d X·1 , X·2 d X·2 , X·2 t . . . d X·s , X·2 t   ε2t 
 · t  =  t  
...  ... 
...
... ... ...
  

where
σn∗ (t, Ti )σn∗ (t, Tj )0 dt+
 
B(t− , Ti ) B(t− , Tj )  
d X·i , X· j

t
=
Bt Bt
ψn∗ (t, Ti , e)ψn∗ (t, Tj , e)β N (t, e)νt (de)dt
 R 
E
for all i, j = 1, . . . , s1 ,
 0 
P (t− , Ti , Bi ) P (t− , Tj , Bj )  σl∗ (t, Ti , Bi )σl∗ (t, Tj , Bj ) dt+ 
d X·i , X· j

t
=
Bt Bt
ψl∗ (t, Ti , Bi , e)ψl∗ (t, Tj , Bj , e)β N (t, e)νt (de)dt
 R 
E
for all i, j = s1 + 1, . . . , s ,
 0 
B(t− , Ti ) P (t− , Tj , Bj )  σn∗ (t, Ti )σl∗ (t, Tj , Bj ) dt+ 
d X·i , X· j

t
=
Bt Bt
ψn∗ (t, Ti , e)ψl∗ (t, Tj , Bj , e)β N (t, e)νt (de)dt
 R 
E
for all i = 1, . . . , s1 and j = s1 + 1, . . . , s
P (t− , T, Bm ) B(t− , Ti )
d J· , X·i t = n0 (Bm × K)gT

Φ(t, T, Bm , i)
Bt Bt
where
 0 
 σl∗ (t, T, Bm )σn∗ (t, Ti ) dt+ 
Φ(t, T, Bm , i) =
ψl∗ (t, T, Bm , e)ψn∗ (t, Ti , e)β N (t, e)νt (de)dt
 R 
E
for all i = 1, . . . , s1 , and

P (t− , T, Bm ) P (t− , Ti , Bi ) i
d J· , X·i t = n0 (Bm × K)gT

Ψ (t, T, Bm , i)
Bt Bt
where
 0 
 σl∗ (t, T, Bm )σl∗ (t, Ti , Bi ) dt+ 
Ψ(t, T, Bm , i) =
ψl∗ (t, T, Bm , e)ψl∗ (t, Ti , Bi , e)β N (t, e)νt (de)dt
 R 
E
for all i = s1 + 1, . . . , s.
The discounted amount to invest in the risk-free asset is given by

s1 s
B(t, Ti ) P (t, Ti , Bi )
K)gT P (t,T,B m)
X X
ηt∗ = n0 (Bm × Bt − εit − εit
Bt 1
Bt
i=1 i=s +1
Proof. Use Theorem 4.1 and Propositions 4.2 and 4.3 as well as the rules of the sharp
bracket.
4.2.2. Annuities. In this section, we want to find the risk-minimizing strategy of a portfolio
of continous annuities paid, up to a given time T , to the surviving members of a survivor index.
If we assume the survivor index is defined on the age group Bm , then the cumulated discounted
payment at time T , AT , is given by
Z T
(n0 (Bm × K) − N (u, Bm × K))
AT = a(u) du
0 Bu
Again, to find the risk-minimizing strategies, we need the following propositions. The first one
gives the market value of these annuities.
Proposition 4.5. The discounted value Vta (ρ∗ ) of the risk-minimizing portfolio is
Z T
a ∗ P (t, u, Bm )
Vt (ρ ) = n0 (Bm × K) a(u) du
t Bt
Proof. We have
Vta (ρ∗ ) = E Q [AT − At |Gt ]
Z T
Q (n0 (Bm × K) − N (u, Bm × K))
= E a(u) du |Gt
t Bu
Z T
1 Q Bt (n0 (Bm ) − N (u, Bm × K))
= n0 (Bm × K) a(u) E |Gt du
t Bt Bu n0 (Bm )
Z T
P (t, u, Bm )
= n0 (Bm × K) a(u) du
t Bt

Proposition 4.6. The martingale representation of Jta (ρ∗ ) is
Z T
P (0, u, Bm )
Jta (ρ∗ ) = n0 (Bm × K) a(u) du
0 B0
Z t
− n0 (Bm × K) σa∗ (s, T, Bm )dWsQ
0
Z tZ
+ n0 (Bm × K) ψa∗ (s, T, Bm , e)q Q (ds, de)
0 E
where
Z T
P (s− , u, Bm ) ∗
σa∗ (s, T, Bm ) = a(u) σl (s, u, Bm )du
s Bs
Z T
P (s− , u, Bm ) ∗
ψa∗ (s, T, Bm , e) = a(u) ψl (s, u, Bm , e)du
s Bs
The risk-minimizing strategies are then given by:
Proposition 4.7. The risk-minimizing strategy ρ∗ = (ε∗t , ηt∗ ) = ε1t , . . . , εst , ηt∗ is given

by Proposition 4.4 where
B(t− , Ti )
d J· , X·i t = n0 (Bm × K)

Φa (t, T, Bm , i)
Bt
where
 0 
 σa∗ (t, T, Bm )σn∗ (t, Ti ) dt+ 
Φa (t, T, Bm , i) =
ψa∗ (t, T, Bm , e)ψn∗ (t, Ti , e)β N (t, e)νt (de) dt
 R 
E
for all i = 1, . . . , s1 , and

P (t− , Ti , Bi )
d J· , X·i t = n0 (Bm × K)

Ψa (t, T, Bm , i)
Bt
where
σa∗ (t, T, Bm )σl∗ (t, Ti , Bi )dt+
 
 
Ψa (t, T, Bm , i) =
ψa∗ (t, T, Bm , e)ψl∗ (t, Ti , Bi , e)β N (t, e)νt (de)dt
 R 
E
for all i = s1 + 1, . . . , s.
The discounted amount to invest in the risk-free asset is
Z T s1 s
∗ P (t, u, Bm ) X B(t, Ti ) X P (t, Ti , Bi )
ηt = n0 (Bm × K) a(u) du − εit − εit
t Bt Bt 1
Bt
i=1 i=s +1
Proof. Use Theorem 4.1 and Propositions 4.5 and 4.6 as well as the rules of the sharp
bracket.
4.3. Risk-minimization for the portfolio of an insurance company. In the previous

section, we showed how to hedge pure endowments and annuities associated with a survivor
index. However, in real life, the portfolio of an insurance company represents only a fraction
of the population of a survivor index. In this section, we study the risk-minimizing strategies
for the portfolio of a given insurance company k ∈ K.
In order to apply the risk-minimization theory as in the previous section, we need to know
the dynamics of the process P k (t, T, Bm ) defined by

k Q n0 (Bm × k) − NT (Bm × k) − R T ru du
P (t, T, Bm ) = E e t |Gt
n0 (Bm × k)
In the general model described above, we cannot say much about this process and cannot go
much further. However, by introducing the following assumption, we can find interesting and
useful relationships between the longevity bond price P (t, T, Bm ), and the process P k (t, T, Bm ).
Condition 4.8. For each t, we assume the following equality holds:
N (t, e) ν (de) N
R R
Bm ×k β t Bm ×K β (t, e) νt (de)
=
(n0 (Bm × k) − N (t− , Bm × k)) (n0 (Bm × K) − N (t− , Bm × K))
Intuitively, this condition means that the average Q-intensity of mortality of the policyhold-
ers of the company k in the age group Bm is equal to the average Q-intensity of mortality of the
whole population in the same age group. Notice also, this condition holds, at least implicitly,
in any of the existing intensity models. Indeed, for a given age x ∈ X, this condition becomes
PK N
β N (t, (x × k)) νt ((x × k)) i=1 β (t, (x × k)) νt ((x × k))
=
and since in these intensity models, we have β N (t, (x × k)) = β N (t, x) and νt ((x × k)) =
(n0 (Bm × k) − N (t− , Bm × k)) µ(t, x), we finally get:
PK N
(n0 (Bm × k) − N (t− , Bm × k)) β N (t, x) µ(t, x) i=1 β (t, x) (n0 (Bm × k) − N (t− , Bm × k)) µ(t, x)
=
β N (t, x) µ(t, x) = β N (t, x) µ(t, x)
It must be emphasized that only the average Q-intensity intervenes in Condition 4.8. We
have now the following proposition:
Proposition 4.9. If Condition 4.8 holds then
P k (t, T, Bm ) = Itk (Bm )Z(t, T, Bi )
or
Itk (Bm )
P k (t, T, Bm ) = P (t, T, Bi )
It (Bm )
Proof. Let us assume longevity bonds of all maturities are defined on an additional index
n0 (Bm × k) − NT (Bm × k)
Itk (Bm ) =
n0 (Bm × k)
As in Equation (3.4), we define the forward rates flk (t, s, Bm )as
RT
flk (t,s,Bm )ds
P k (t, T, Bm ) = Itk (Bm )e− t
where we assume
Z t Z t
flk (t, s, Bm ) = flk (0, s, Bm ) + αlk (u, s, Bm )du + σlk (u, s, Bm )dWu
0 0
Z tZ
+ ζlk (u, s, Bm , e)N (du, de)
0 E
We can write
RT RT
fl∆k (t,s,Bm )ds
P k (t, T, Bm ) = Itk (Bm )e− t fl (t,s,Bm )ds− t
where fl∆k (t, s, Bm ) = flk (t, s, Bm ) − fl (t, s, Bm ). We could extend Proposition 3.4 with this
additional term structure. We would obtain a similar no-arbitrage condition. In particular, we
would obtain the equalities
N
R
fl (t, t, Bm ) − rt =
(n0 (Bm × K) − N (t− , Bm × K))
and
β N (t, e) νt (de)
R
Bm ×k
flk (t, t, Bm ) − rt =
(n0 (Bm × k) − N (t− , Bm × k))
If Condition 4.8 holds then fl (t, t, Bm ) − rt = flk (t, t, Bm ) − rt which implies fl∆k (t, t, Bm ) = 0
for all t ∈ [0, T ∗ ]. In turn, this leads to fl∆k (s, t, Bm ) = 0 for all s ≤ t. Eventually, we have
RT
P k (t, T, Bm ) = Itk (Bm )e− t fl (t,s,Bm )ds
= Itk (Bm )Z(t, T, Bm )

Itk (Bm )
= P (t, T, Bm )
It (Bm )

Intuitively, due to Condition 4.8, the price of a security defined on a fraction of a given
survivor index is equal to the price of the associated longevity bond times the ratio of the
current proportions of survivors. Without Condition 4.8, this simple proportional rule does
not necessarily hold. Since Condition 4.8 holds in all the existing intensity models, this simple
rule also works well in these cases. However, Proposition 4.9 suggests we could extend these
intensity models to more general settings and preserve this property.
As a by-product, Condition 4.8 also leads to the following result. We give it here because
we believe it is interesting in its own right but it will not be used later.
Proposition 4.10. If Condition 4.8 holds then the ratio
Itk (Bm )
It (Bm )
is a (QT,Bm , G)-martingale.
I k (B )
Proof. On one hand, using Proposition 3.14 with HT = ITT (Bm m)
, we have:
h RT i
P k (t, T, Bm ) = E Q e− t ru du ITk (Bm ) |Gt
k
QT,Bm IT (Bm )
= P (t, T, Bm )E |Gt
IT (Bm )
Itk (Bm )
On the other hand, we have thanks to the previous proposition, P k (t, T, Bm ) = It (Bm ) P (t, T, Bm ).
This gives
k
Itk (Bm )

QT,Bm IT (Bm )
= E |Gt
It (Bm ) IT (Bm )

We can now focus on the risk-minimizing strategies of an insurer’s portfolio. Frow now on,
we will assume Condition 4.8 holds.
4.3.1. Pure endowments. The discounted payment at time T is now
gT (n0 (Bm × k) − N (T, Bm × k))
AT =
BT
where gT is deterministic. As in the previous section, we have several propositions. The first
one also gives the market value of the insurer’s portfolio when Condition 4.8 holds.
Proposition 4.11. The discounted value of the risk-minimizing portfolio is given by
Itk (Bm ) P (t, T, Bm )

Vtg (ρ∗ ) = n0 (Bm × k)gT
It (Bm ) Bt
Proof. Since gT is deterministic, we have
h RT i
Vtg (ρ∗ ) = E Q gT (n0 (Bm × k) − N (T, Bm × k)) e− 0 rs ds |Gt

1 Q (n0 (Bm × k) − N (T, Bm × k)) − R T rs ds
= n0 (Bm × k)gT E e t |Gt
Bt n0 (Bm × k)
Itk (Bm ) P (t, T, Bm )
= n0 (Bm × k)gT
It (Bm ) Bt

Proposition 4.12. The martingale representation of Jtg (ρ∗ ) is given by
Jtg (ρ∗ ) = n0 (Bm × k)gT P (0, T, Bm )
!
Isk− (Bm ) P (s− , T, Bm )
Z t
− n0 (Bm × K)gT σl∗ (s, T, Bm )dWsQ
0 Is− (B m ) B s
!
Z tZ k
Is− (Bm ) P (s− , T, Bm )
+ n0 (Bm × K)gT ψl∗,k (s, T, Bm , e)q Q (ds, de)
0 E Is− (Bm ) Bs
where
∗
e−ζl (s,T,Bm ,e)
ψl∗,k (s, T, Bm , e) = ψZ∗ (s, T, Bm , e) − 1{e∈Bm ×k}
(n0 (Bm × k) − N (s− , Bm × k))
Proof. Again, we have here Jtg (ρ∗ ) = Vtg (ρ∗ ) and since P k (0, T, Bm ) = P (0, T, Bm ), we
can write
Vtg (ρ∗ ) = n0 (Bm × k)gT P (0, T, Bm )
Z t k
P (s, T, Bm )
+ n0 (Bm × k)gT d
0 Bs
Following the same argument as for P (s, T, Bm ), we can prove
Z
P k (s, T, Bm ) = P k (s− , T, Bm ) rs ds − σl∗ (s, T, Bm )dWsQ + ψl∗,k (s, T, Bm , e)q Q (ds, de)
E
where ψl∗,k (s, T, Bm , e) is given above.

We can now give the risk-minimizing strategy. As in the previous section, we assume we
can trade in risk-free zero-coupon bonds of s1 different maturities and in longevity bonds of s2
different maturities with s1 + s2 = s.
Proposition 4.13. If Condition 4.8 holds, the risk-minimizing strategy ρ∗ = (ε∗t , ηt∗ ) =
ε1t , . . . , εst , ηt∗ is given by Proposition 4.4 except we have

Itk− (Bm ) P (t− , T, Bm ) B(t− , Ti ) k

J· , X·i t

d = n0 (Bm × k)gT Φ (t, T, Bm , i)
It− (Bm ) Bt Bt
where
0
σl∗ (t, T, Bm )σn∗ (t, Ti ) dt+
 
 
Φk (t, T, Bm , i) =
ψl∗,k (t, T, Bm , e)ψn∗ (t, Ti , e)β N (t, e)νt (de)dt
 R 
E
for all i = 1, . . . , s1 , and
Itk (Bm ) P (t− , T, Bm ) P (t− , Ti , Bi ) k
d J· , X·i t = n0 (Bm × k)gT −

Ψ (t, T, Bm , i)
It− (Bm ) Bt Bt
where
0
σl∗ (t, T, Bm )σl∗ (t, Ti , Bi ) dt+
 
 
Ψk (t, T, Bm , i) =
ψl∗,k (t, T, Bm , e)ψl∗ (t, Ti , Bi , e)β N (t, e)νt (de)dt
 R 
E
for all i = s1 + 1, . . . , s.
The discounted amount to invest in the risk-free asset is
s1 s
I k (Bm ) P (t,T,Bm )
X B(t, Ti ) X P (t, Ti , Bi )
ηt∗ = n0 (Bm × k)gT Itt (Bm ) Bt − εit − εit
Bt 1
Bt
i=1 i=s +1
Proof. Again, use Theorem 4.1 and Propositions 4.11 and 4.12 as well as the rules of the
sharp bracket.
4.3.2. Annuities. In this section, we study the risk-minimizing strategy of a portfolio of
continous annuities paid, up to a given time T , to each surviving policyholder in the age group
Bm of a given insurer k. The cumulated discounted payments At at time t, is then given by
Z t
(n0 (Bm × k) − N (u, Bm × k))
At = a(u) du
0 Bu
The following proposition gives the market value of the annuity portfolio when Condition 4.8
holds:
Proposition 4.14. The discounted value Vta (ρ∗ ) of the risk-minimizing portfolio is given
by
T
I k (Bm )
Z
P (t, u, Bm )
Vta (ρ∗ ) = n0 (Bm × k) t a(u) du
It (Bm ) t Bt
Proof. We have
Vta (ρ∗ ) = E Q [AT − At |Gt ]
Z T
Q (n0 (Bm × k) − N (u, Bm × k))
= E a(u) du |Gt
t Bu
Z T
1 Q Bt (n0 (Bm ) − N (u, Bm × k))
= n0 (Bm × k) a(u) E |Gt du
t Bt Bu n0 (Bm )
Z T
I k (Bm ) P (t, u, Bm )
= n0 (Bm × k) a(u) t du
t It (Bm ) Bt

Proposition 4.15. The martingale representation of Jta (ρ∗ ) is
Z T
a ∗ P (0, u, Bm )
Jt (ρ ) = n0 (Bm × k) a(u) du
0 B0
Z t k
Is− (Bm ) ∗,k
− n0 (Bm × k) σa (s, T, Bm )dWsQ
0 Is− (Bm )
Z tZ k
Is− (Bm ) ∗,k
+ n0 (Bm × k) ψa (s, T, Bm , e)q Q (ds, de)
0 E Is − (Bm )
where
Z T
P (s− , u, Bm ) ∗
σa∗,k (s, T, Bm ) = a(u) σl (s, u, Bm )du
s Bs
Z T
P (s− , u, Bm ) ∗,k
ψa∗,k (s, T, Bm , e) = a(u) ψl (s, u, Bm , e)du
s Bs
Proof. The proof is similar that in Appendix 2.
As far as the risk-minimizing strategy is concerned, we have:

∗ ∗ ∗
∗ 4.16. If Condition 4.8 holds, the risk-minimizing strategy ρ = (εt , ηt ) =
Proposition
ε1t , . . . , εst
, ηt is given by Proposition 4.7 except we have
Itk (Bm ) B(t− , Ti ) k

d J· , X·i t = n0 (Bm × k) −

Φa (t, T, Bm , i)
It− (Bm ) Bt
where
σa∗,k (t, T, Bm )σn∗ (t, Ti ) dt+
 0 
 
Φka (t, T, Bm , i) =
ψa∗,k (t, T, Bm , e)ψn∗ (t, Ti , e)β N (t, e)νt (de) dt
 R 
E
for all i = 1, . . . , s1 , and
Itk− (Bm ) P (t− , Ti , Bi ) k

J· , X·i t

d = n0 (Bm × k) Ψa (t, T, Bm , i)
It− (Bm ) Bt
where
σa∗,k (t, T, Bm )σl∗ (t, Ti , Bi )dt+
 
 
Ψka (t, T, Bm , i) =
ψa∗,k (t, T, Bm , e)ψl∗ (t, Ti , Bi , e)β N (t, e)νt (de)dt
 R 
E
for all i = s1 + 1, . . . , s.
The discounted amount to invest in the risk-free asset is given by:
s 1 s
T
I k (Bm )
Z
P (t, u, Bm ) X
i B(t, Ti )
X P (t, Ti , Bi )
ηt∗ = n0 (Bm × k) t a(u) du − εt − εit
It (Bm ) t Bt Bt 1
Bt
i=1 i=s +1
Proof. Again, it suffices to use Theorem 4.1 and Propositions 4.14 and 4.15 as well as the
rules of the sharp bracket.
Appendix 1.
Proposition. If the forward rates follow Equation (3.2), then the price of a zero-coupon
bond with maturity T is given by the solution of the differential equation
 h h ii 
0
 fn (t, t) − αn∗ (t, T ) + 12 σn∗ (t, T ) σn∗ (t, T ) dt
 

dB(t, T ) = B(t− , T )
−σn∗ (t, T )dWt + E ψn∗ (t, T, e)N (dt, de)

 R 

230 Appendix 1.
where
Z T
αn∗ (t, T ) = αn (t, s)ds
t
Z T
σn∗ (t, T ) = σn (t, s)ds
t
∗ (t,T,e)
ψn∗ (t, T, e) = e −ζn
−1
Z T
∗
ζn (t, T, e) = ζn (t, s, e)ds
t
RT
Proof. Let us first find the dynamics of Jn (t, T ) = t fn (t, s)ds. We have
Z T Z T Z t Z T Z t
Jn (t, T ) = fn (0, s)ds + αn (u, s)du ds + σn (u, s)dWu ds
t t 0 t 0
Z T Z tZ
+ ζn (u, s, e)N (du, de) ds
t 0 E
Z T Z t Z T Z t Z T
= fn (0, s)ds + αn (u, s)ds du + σn (u, s)ds dWu
t 0 t 0 t
Z tZ Z T
+ ζn (u, s, e)ds N (du, de)
0 E t
Z T Z t Z T Z t Z T
0 0 u 0 u
Z tZ Z T
0 E u
Z t Z t Z t Z t Z t
− fn (0, s)ds − αn (u, s)ds du − σn (u, s)ds dWu
0 0 u 0 u
Z tZ Z t
− ζn (u, s, e)ds N (du, de)
0 E u
On the other hand,

Z t Z t Z t Z s Z t Z s
fn (s, s)ds = fn (0, s)ds + αn (u, s)du ds + σn (u, s)dWu ds
0 0 0 0 0 0
Z tZ sZ
+ ζn (u, s, e)N (du, de) ds
0 0 E
Z t Z t Z t Z t Z t
0 0 u 0 u
Z tZ Z t
0 E u
which leads to
Z T Z t Z t Z t
Jn (t, T ) = fn (0, s)ds − fn (s, s)ds + αn∗ (u, T )du + σn∗ (u, T )dWu
0 0 0 0
Z tZ
+ ζn∗ (u, T, e) N (du, de)
0 E
Using Itô’s lemma for semimartingale on the function f (x) = exp(−x), we can find the dynamic
of B(t, T ) = e−Jn (t,T ) . We obtain
1
dB(t, T ) = −e−Jn (t− ,T ) dJn (t, T )c + e−Jn (t− ,T ) d hJn (t, T ), Jn (t, T )ic + e−Jn (t,T ) − e−Jn (t− ,T )
2
1h ∗ 0 ∗
i
−Jn (t− ,T ) c −∆Jn (t,T )
= −e dJn (t, T ) + σ (t, T ) σn (t, T ) dt + e −1
2 n
n h 0
i ∗
o
= B(t− , T ) −dJn (t, T )c + 12 σn∗ (t, T ) σn∗ (t, T ) dt + E e−ζn (t,T,e) − 1 N (dt, de)
R
So eventually, we get
( h h 0
ii )
fn (t, t) − αn∗ (t, T ) + 12 σn∗ (t, T ) σn∗ (t, T ) dt
dB(t, T ) = B(t− , T ) ∗
−σn∗ (t, T )dWt + E e−ζn (t,T,e) − 1 N (dt, de)
R
Appendix 2.
Proposition. The martingale representation of Jta (ρ∗ ) is given by
Z T
P (0, u, Bm )
Jta (ρ∗ ) = a(u)n0 (Bm × K) du
0 B0
Z t Z T
P (s− , u, Bm ) ∗
− n0 (Bm × K) a(u) σl (s, u, Bm )du dWsQ
0 s B s
Z t Z T Z
P (s− , u, Bm ) ∗
+ n0 (Bm × K) a(u) ψl (s, u, Bm , e)du q Q (ds, da)
0 s B s E
232 Appendix 2.
Proof. We have
Z T
1 (n0 (Bm × K) − N (u, Bm × K))
Jta = n0 (Bm × K)E Q
a(u) du |Gt
0 Bu n0 (Bm × K)
Z t
1 (n0 (Bm × K) − N (u, Bm × K))
= n0 (Bm × K) a(u) du
0 Bu n0 (Bm × K)
Z T
1 Q Bt (n0 (Bm × K) − N (u, Bm × K))
+ n0 (Bm × K) a(u) E |Gt du
t Bt Bu n0 (Bm × K)
Z t
1 (n0 (Bm × K) − N (u, Bm × K))
= n0 (Bm × K) a(u) du
0 B u n0 (Bm × K)
Z T
P (t, u, Bm )
+ n0 (Bm × K) a(u) du
t Bt
RT
Let us focus on the second term (I) = n0 (Bm × K) t a(u) P (t,u,B
Bt
m)
du. Since the price of a
longevity bond is given by
P (t, u, Bm )
= P (0, u, Bm )
Bt
Z t
P (s− , u, Bm ) ∗
− σl (s, u, Bm )dWsQ
0 B s
Z t Z
P (s− , u, Bm )
+ ψl∗ (s, u, Bm , e)q Q (ds, de)
0 Bs E
∗ −ζl∗ (s,u,Bm ,e)

where ψl∗ (s, u, Bm , e) = e−ζl (s,u,Bm ,e) − 1 − 1{e∈Bm ×K} (n0 (Bme×K)−N

(s− ,Bm ×K)) , we can write
Z T
P (0, u, Bm )
(I) = a(u)n0 (Bm × K) du
t B0
Z T Z t
P (s− , u, Bm ) ∗ Q
− a(u)n0 (Bm × K) σl (s, u, Bm )dWs du
t 0 Bs
Z T Z t Z
P (s− , u, Bm ) ∗ Q
+ a(u)n0 (Bm × K) ψl (s, u, Bm , e)q (ds, de) du
t 0 Bs E
Z T
P (0, u, Bm )
(I) = a(u)n0 (Bm × K) du
0 B0
Z t Z T
P (s− , u, Bm ) ∗
0 s B s
Z t Z T Z
P (s− , u, Bm ) ∗
+ n0 (Bm × K) a(u) ψl (s, u, Bm , e)du q Q (ds, de)
0 s B s E
Z t
P (0, u, Bm )
− a(u)n0 (Bm × K) du
0 B0
Z t Z t
P (s− , u, Bm ) ∗
+ n0 (Bm × K) a(u) σl (s, u, Bm )du dWsQ
0 s Bs
Z t Z t Z
P (s− , u, Bm ) ∗
− n0 (Bm × K) a(u) ψl (s, u, Bm , e)du q Q (ds, de)
0 s Bs E
Then, by changing the order of integration again, the three last terms (II) can be written as
Z t
P (0, u, Bm )
(II) = − a(u)n0 (Bm × K) du
0 B0
Z t Z u
P (s− , u, Bm ) ∗ Q
+ n0 (Bm × K)a(u) σl (s, u, Bm )dWs du
0 0 Bs
Z t Z u Z
P (s− , u, Bm ) ∗ Q
− n0 (Bm × K)a(u) ψl (s, u, Bm , e)q (ds, de) du
0 0 Bs E
Z t
P (u, u, Bm )
= − a(u)n0 (Bm × K) du
0 Bu
Z t
n0 (Bm × K) − N (u, Bm × K)
= − a(u) du
0 Bu
Eventually, we get
Z T
a P (0, u, Bm )
Jt = a(u)n0 (Bm × K) du
0 B0
Z t Z T
P (s− , u, Bm ) ∗
0 s Bs
Z t Z T Z
P (s− , u, Bm ) ∗
+ n0 (Bm × K) a(u) ψl (s, u, Bm , e)du q Q (ds, de)
0 s Bs E

Part 4
Non-Life Insurance Contracts and the

Inflation Risk.
CHAPTER 7
Risk-Minimization with Inflation and Interest Rate Risk.
1. Introduction.
This chapter aims at studying the asset allocation problem of a non-life insurance company.
The academic literature on this topic is rather scarce. Indeed, most portfolio selection problems
in actuarial mathematics concern the optimal asset allocation for life insurance portfolios or
pension plans. This is due to the fact that these contracts are long-term contracts more
subject to investment risk than the short-term non-life contracts. There are however good
arguments to study the asset allocation problem in non-life insurance as well. Indeed, the cost
of claims is highly sensitive to inflation, which, in turn, is itself correlated with the financial
assets returns and in particular, with the interest rates movements. Moreover, the reserve an
insurer should maintain, computed as the expected discounted claims, is subject to interest
rate risk through the variation in the discounting term and indirectly through the variation
in expected outstanding claims due to the inflation. The application of ALM techniques in
non-life insurance is thus justified from a theoretical point of view.
The ALM methods based on duration and convexity have been studied in non-life insur-
ance by Ahlgrim and al [5]. More precisely, they studied the effective durations and convexities
of non-life insurance liabilities. These notions extend the classical concepts of duration and
convexity when the future cash flows are interest rates sensitive. In their paper, they assume
the interest rates and the inflation are stochastic and correlated. Furthermore, they assume a
fraction of the claim costs is inflation sensitive. They showed that the effective durations and
convexities based on these assumptions are substantially lower than those measured with tradi-
tional approaches i.e. without stochastic inflation and interest rates. Their analysis underlines
the importance of introducing stochastic interest rates and inflation when dealing with ALM
in non-life insurance.
Even though Ahlgrim et al. [5] introduce stochastic interest rates and inflation, their ap-
proach relies nonetheless on static ALM techniques. Dynamic portfolio selection in continuous
time for non-life insurance businesses is rather rare. We can cite the paper of Browne [35],
those of Hipp and Plum [52, 53] and Korn [60]. In [35], Browne derives the financial policy
that minimizes the probability of ruin and the one that maximizes the expected exponential
utility of terminal wealth when the cumulative cost of claims follows a Brownian motion with
drift. For each objective, he derives the optimal policy when the insurer can invest in a single
risky asset and when he can invest in a risky asset and a risk free asset. The risky asset fol-
lows a geometric Brownian motion and the risk free asset grows at a constant risk free rate.
Hipp and Plum [52] derive the asset allocation that minimizes the probability of ruin when
237
238 Introduction.
the cost of claims follows a compound Poisson process and when the insurer can invest in a
single risky asset following a geometric Brownian motion. In [53], Hipp and Plum study the
accumulated discounted expected utility of wealth as the objective function. They find the
optimal financial policy of an insurer with a cost of claims following a compound Poisson pro-
cess and a compound Cox process (generated by a finite state space Markov process) when the
insurer can invest in a risky financial asset following a geometric Brownian motion. In [60],
Korn extends the work of Browne by introducing the possibility of a market crash. He finds
the best deterministic financial policy that maximizes the worst-case exponential utility from
terminal wealth. As in Browne, the cost of claims follows a Brownian motion with drift and
the financial market consists of a risky asset following a geometric Brownian motion subject to
a single sudden downward jump, and a risk free asset. In a very interesting paper, Delong [41]
introduces another objective. He defines a linear-quadratic function that penalizes deviations
of the insurance surplus process below a reserve for outstanding claims (computed under pru-
dential basis) and rewards deviations above the reserve. He assumes the cost of claims follows
a compound Poisson process and considers a financial market with a risk free asset (constant
rate) and n risky assets following geometric Brownian motions. Unfortunately, these papers
do not take into account the inflation risk and in most case, they do not assume a stochastic
interest rates term structure.
In order to try to overcome these shortcomings, we choose to follow Møller who applied
the risk-minimization theory in [68] to non-life insurance problems. This theory has been
developed by Föllmer and Sondermann [44] for a single payoff and extended by Møller [68]
for a payment process. We extend the work of Møller in two directions. Firstly, we introduce
a financial market with stochastic interest rates and stochastic inflation. Secondly, as far as
the cumulative cost of claims is concerned, we study an arbitrary (increasing) process (non
necessarily Markov) adapted to the filtration of a general marked point process. However, we
are also slightly less general than Møller in the sense that he allows the distribution of the cost
of the claims to depend on the value of a financial index whereas in our model, the size of
claims depends on the financial market through the mechanism of inflation only.
This chapter is organized as followed. In Section 2.1, we describe the financial market
developed by Jarrow and Yildirim [57]. It is a general continuous time HJM model with
inflation. In Section 2.2, we describe the claims model. It is based on a general marked point
process (with absolutely continuous compensator though) as suggested by Arjas [9]. In Section
3.1, we give a brief review of the risk minimization theory and in Section 3.2, we solve the
risk minimization problem in the general framework described in Section 2. In Section 4 and
Section 5, we then apply systematically this result to 4 specific models of insurance claims.
Section 4 treats two collective models based on the compound Poisson process. In the first one,
the number of policies is constant whereas in the second one the number of policies is stochastic
and follows a point process with an inhomogeneous and stochastic intensity. Section 5 treats
two individual models where claims are notified at a random time and settled through time
according to a deterministic function in the first one and according to a Beta-Stacy process in
the second one.
Risk-Minimization with Inflation and Interest Rate Risk. 239
2. The Model.
2.1. The financial market. We use the model developed by Jarrow and Yildirim [57].
Hughston [55] also described a similar model. In their paper, these authors exploit an analogy
between the inflation and an exchange rate. They distinguish two “economies”. One is a
“nominal economy” and the other is a “real economy”. For each one, an interest rate curve of
the HJM-type is described. The inflation plays the role of an exchange rate between these two
economies. Let us describe briefly this model. See the original paper for more details.
Let W be a d−dimensional Brownian motion defined on a probability space ΩF , F F , P F .

We assume F = (Ft )0≤t<T ∗ is the P F -completed version of the filtration generated by W .
The nominal interest rate term structure. We assume there exists for each maturity
T ∈ [0, T ∗ ], a nominal 0-coupon bond whose price at time t ∈ [0, T ] is denoted Pn (t, T ). As in
the traditional HJM model, we assume the nominal forward rates follow:
dfn (t, T ) := αn (t, T )dt + σn (t, T )dW (t)
where αn (t, T ) and σn (t, T ) are some F-adapted stochastic processes. By definition, the price
Pn (t, T ) at time t of a nominal 0-coupon bond with maturity T is given by
RT
Pn (t, T ) := e− t fn (t,u)du
We assume we can invest in an instantaneously risk free asset (in nominal terms) whose price
at time t is denoted Bn (t). The price of this asset evolves as
dBn (t) = n(t)Bn (t)dt
with Bn (0) = 1 and where n(t) is an instantaneously nominal risk free rate. It can be shown
that we have n(t) = fn (t, t).
The real interest rate term structure. Similarly, we assume there exists a real 0-
coupon bond for each maturity T ∈ [0, T ∗ ] whose price is denoted Pr (t, T ). The real interest
rates curve is described through the following real forward rates:
dfr (t, T ) := αr (t, T )dt + σr (t, T )dW (t)

The price Pr (t, T ) at time t of a real 0-coupon bond with maturity T is then given by
RT
Pr (t, T ) := e− t fr (t,u)du
We assume there is an instantaneously risk free asset (in real terms) whose price is denoted by
Br (t). The price of this asset is given by
dBr (t) = r(t)Br (t)dt
where r(t) is the instantaneously real risk free rate. We also have r(t) = fr (t, t). The difference
n(t) − r(t) is a risk premium for bearing the inflation risk.
240 The Model.
The price index. As far as the value of the price index I(t) is concerned, its evolution is
described by the following SDE:
dI(t) := I(t) [µI (t)dt + σI (t)dW (t)]

where µI (t) and σI (t) are some arbitrary F-adapted stochastic processes.
The no-arbitrage opportunity condition. As usual in financial economics, Jarrow and

Yildirim assume that the no-arbitrage hypothesis holds. This assumption implies there exists
at least one measure QF equivalent to P F , under which the following discounted asset prices
are (QF , F)-martingales:
I(t)Br (t) Pn (t, T ) I(t)Pr (t, T )
, ,
Bn (t) Bn (t) Bn (t)
Such a measure QF is called a martingale measure.
Let us define
Z T Z T
σn∗ (t, T ) := σn (t, u)du αn∗ (t, T ) := αn (t, u)du
t t
Z T Z T
σr∗ (t, T ) := σr (t, u)du αr∗ (t, T ) := αr (t, u)du
t t
Jarrow and Yildirim showed that the no arbitrage condition implies the following constraints
on the drift of the inflation and the drifts of the forward rates:
µI (t) = n(t) − r(t) + σI (t)β(t)

∗ ∗ 1 ∗ 0
αn (t, T ) = σn (t, T ) σn (t, T ) + β(t)
2

1
αr∗ (t, T ) = σr∗ (t, T ) σr∗ (t, T )0 + β(t) − σI (t)0
2
where β(t) is the d-dimensional market price of risk associated to the Brownian motion W . In
other words, for any stochastic process β(t) such that these constraints hold, the measure QF
F R·
defined by dQ := ET ∗ − 0 β(u)du is an equivalent martingale measure. The process W Q (t)

dP F Rt
defined by W Q (t) = W (t) + 0 β(u)du is a QF , F -Brownian motion.

Jarrow and Yildirim gave the following dynamics:
dI(t) = I(t) [(n(t) − r(t) + σI (t)β(t)) dt + σI (t)dW (t)]

dPn (t, T ) = Pn (t, T ) [(n(t) − σn∗ (t, T )β(t)) dt − σn∗ (t, T )dW (t)]
dPr (t, T ) = Pr (t, T ) r(t) + σr∗ (t, T )σI (t)0 − σr∗ (t, T )β(t) dt − σr∗ (t, T )dW (t)

We can also write:

dI(t) = I(t) (n(t) − r(t)) dt + σI (t)dW Q (t)

dPn (t, T ) = Pn (t, T ) n(t)dt − σn∗ (t, T )dW Q (t)

dPr (t, T ) = Pr (t, T ) r(t) + σr∗ (t, T )σI (t)0 dt − σr∗ (t, T )dW Q (t)

where W Q (t) is a d-dimensional Brownian motion under QF .
Treasury Inflation Protected securities (TIPS). In Section 3.2, we will assume that
we can invest in two different categories of securities. The first category is the nominal 0-coupon
bonds. The second category is the 0-coupon Treasury Inflation Protected Securities (TIPS). A
0-coupon TIPS of maturity T pays one real unit at time T . In other words, it pays I(T ) units
in nominal terms. Jarrow and Yildirim showed the price T p(t, T ) at time t of such security is
given by

QF Bn (t)
T p(t, T ) = E I(T ) Ft
Bn (T )
= I(t)Pr (t, T )
Using the results of the previous subsection, we can easily show the discounted price of a
nominal 0-coupon bond with maturity T and the discounted price of a 0-coupon TIPS with
maturity T follow the SDEs:

Pn (t, T ) Pn (t, T ) ∗
d = − σ (t, T )dW Q (t)
Bn (t) Bn (t) n

T p(t, T ) T p(t, T ) ∗
d = − σ (t, T )dW Q (t)
Bn (t) Bn (t) T p
where σT∗ p (t, T ) = σr∗ (t, T ) − σI (t).
2.2. The Insurance model. We follow some ideas developed in Arjas [9]. Let us consider
another probability space ΩI , F I , QI . We assume a marked point process (Tn , Zn )n≥1 is
defined on this space. (Tn )n≥1 is a sequence of random times. (Zn )n≥1 is a sequence of random
variables taking their values in a Borel space (E, E) and associated to the sequence (Tn )n≥1 .
For a detailed description of marked point processes, see Brémaud [33], Jacobsen [56] or Last
and Brandt [61]. For each B ∈ E, we define the counting process N (t, B) = ∞
P
n=1 Tn ≤t 1Zn ∈B .
1
The filtration I is assumed to be the natural filtration of this marked point process i.e. I =
(It )0≤t≤T ∗ where It = σ(N (s, B), 0 ≤ s ≤ t, ∀B ∈ E). For each B ∈ E, we denote by p(t, B)
the (QI , I)-compensator of N (t, B) and q(t, B) the compensated counting process defined as
q(t, B) := N (t, B) − p(t, B).
The cumulative payments paid by the insurer to the policyholders are described by a process
(St )0≤t≤T ∗ defined on the space ΩI . The payments are expressed in real units. We denote by
ΛSt the QI , I -compensator St and assume the following conditions hold:

242 The Risk-Minimizing Strategies.
Condition 2.1. St is a square integrable, increasing1, right-continuous stochastic process

adapted to the filtration I.
‌
Condition 2.2. ΛSt is absolutely continuous with respect to the Lebesgue measure i.e.
t
ΛSt = 0 λSu du for an I-predictable process λSu .
R
We note MtS the compensated process: MtS = St − ΛSt .

We do not assume QI is necessarily the physical measure.
2.3. The combined model. The combined model consists in the following probability
space Ω × Ω , F ⊗ F , Q where Q is the product measure of QF and QI . We define the
F I F I

filtration G = (Gt )0≤t≤T ∗ where Gt = Ft ∨ It . Under the product measure Q, the filtration F
and I are independent. It implies the martingale property of the different
processes is preserved
under this enlargement of filtration. In other words, all the QF , F -martingales and all the
QI , I -martingales are (Q, G)-martingales2. Accordingly, the discounted prices of the financial
securities are (Q, G)-martingales and the QI , I -compensators of the processes N (t, B) and St

are also their (Q, G)-compensators (for any B ∈ E).

In the previous section, we expressed the cumulative insurance payments in real terms. We
can now express the cumulative discounted nominal payments process At . This process is given
by
Z t
I(u)
At = dSu
0 Bn (u)
3. The Risk-Minimizing Strategies.

3.1. A review of the risk-minimization theory. Let us consider an arbitrary filtration
G. We denote by X an s-dimensional vector of discounted prices of the financial securities used
for trading. X corresponds to a vector of discounted prices of 0-coupon nominal bonds and of
0-coupon TIPS. We assume a specific martingale measure Q has been chosen so that X is a
(local) martingale under Q.
A trading strategy ρ is a pair of processes (ε, η) where η is a 1-dimensional real-valued
G-adapted process and ε is an s-dimensional real-valued G-predictable process. The process
εt represents the amount of risky assets held at time t and the process ηt is the discounted
amount invested in the instantaneously risk free asset. The discounted value process Vt (ρ) of a
trading strategy ρ, is defined by Vt (ρ) = εt Xt + ηt for 0 ≤ t ≤ T ∗ . This value process represents
1We could easily extend our results to a finite variation process (not necessarily increasing) since it can
be written as the difference of two increasing processes. Thanks to the linearity of the Galtchouk Kunita
Watanabe, the risk-minimizing strategies of this process is also the difference of the risk-minimizing strategies
of the increasing processes.
2In particular, the sharp bracket with respect to the filtration F, of a given process is indistinguishable of
the sharp bracket with respect to the filtration G of the same process. In the following, we will not distinguish
these different versions of the sharp bracket.
the discounted value of the insurer’s financial portfolio following the trading strategy ρ. This
portfolio is not necessarily self-financed.
The liabilities of the insurer are modelled as a process (At )0≤t≤T ∗ . The process At is as-
sumed to be càdlàg, G-adapted and square integrable. The process At represents the discounted
value of the cumulative payments up to time t. Rt
The cumulative cost process Ct (ρ) of a strategy ρ, is defined by Ct (ρ) = Vt (ρ) − 0 εu dXu +
At . It represents the discounted value of the portfolio reduced by the discounted trading gains
and added discounted net payments to the policyholders. The initial cumulative cost process
C0 (ρ) is given by C0 (ρ) = V0 (ρ) + A0 . The first term V0 (ρ) represents the initial amount
the insurer needs to create his financial portfolio. The second term is the initial payment the
insurer has to pay to the policyholder. Usually, A0 will be negative and instead of a payment will
represent the initial premium paid by the policyholder. A strategy
h ρ is called risk-minimizing,
i
if it minimizes the risk process Rt (ρ) defined by Rt (ρ) = E (CT ∗ (ρ) − Ct (ρ))2 |Gt , for every
Q
t and for every strategy such that VT ∗ (ρ) = 0 , Q-a.s.

Föllmer and Sondermann [44] for a single payoff and Møller [68] for a payment process,
showed that the solution of this problem is closely related to the so-called Galtchouk-Kunita-
Watanabe decomposition of the square integrable martingale E Q [AT ∗ |Gt ]:
Z t
(3.1) E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ] + εA A
u dXu + Lt
0
where εA A
t is an s-dimensional predictable process and Lt is a zero-mean martingale strongly
orthogonal to (the stable subspace generated by) X. With these notation, we can give the
solution of the risk minimization problem in the following theorem:
Theorem 3.1. The unique risk-minimizing RM-strategy ρ∗ = (ε∗ , η ∗ ) of At is given by

ε∗ = εA and ηt∗ = E Q [AT ∗ − At |Gt ] − ε∗t Xt . The cumulative cost process of ρ∗ is given by
Ct (ρ∗ ) = E Q [AT ∗ |G0 ] + LA ∗ A ∗ ∗
Q
E [AT ∗ − At |Gt ].
From a more practical point of view, the predictable integrand εA

t of the Galtchouk-Kunita-
Watanabe can be found in the following way. If we denote Jt = E Q [AT ∗ |Gt ], we can write:
s Z
* +
X .
J. , X.j t Q
εiu dXui L. , X.j

= E [AT ∗ |G0 ] + +
i=1 0
t
s Z
X t
εiu d X.i , X.j u

=
i=1 0
Accordingly, the integrand εA 1 2 s

t = εt , εt , . . . , εt can be found by solving the following ordinary
system of equations:
d
X·1 , X·1 t d
X·2 , X·1 t . . . d
X·s , X·1 t

d
J· , X·1 t

ε1t
    
 d J· , X 2 1 2 2 2 . . . d X·s , X·2 t   ε2t 
 =  d X· , X · t d X· , X · t
 
 · t  
 ... 
...
...
... ... ...
  
3.2. Risk-Minimization in non-life insurance. In this section we give the risk-

minimizing strategy of the cumulative payments process At described in Section 2.3.
Proposition 3.2. The value process Vt (ρ∗ ) is given by

Z T∗
T p(t, u) QI S
Vt (ρ∗ ) =

(3.2) E λu |It du
t Bn (t)
Proof. Indeed we have
"Z #
T∗
I(u)
Vt (ρ∗ ) = E Q dSu |Gt
t Bn (u)
"Z #
T∗
I(u)
= EQ λS du |Gt
t Bn (u) u
Z T∗
Q I(u) Q
S
= E Gt E λ u |Gt du
t Bn (u)
here we use Fubini Theorem and h the Q-independence
i F
h between i the financial market and the
Q I(u) Q I(u) T p(t,u) Q
S
insurance payments. Since E Bn (u) |Gt = E Bn (u) |Ft = Bn (t) and E λ u |Gt =
QI S
E λu |It , we get the result.
The next proposition is the heart of this chapter. It gives the martingale representation of
E Q [AT ∗ |Gt ] which will lead directly to the Galtchouk-Kunita Watanabe decomposition.
Proposition 3.3. The martingale representation of E Q [AT ∗ |Gt ] is given by
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]

Z t
I(u)
+ dMuS
0 B n (u)
Z tZ Z T∗ !
T p(u, s)
(3.3) + h(s, u, z) ds q(du, dz)
0 E u Bn (u)
Z t Z T∗ !
T p(u, s) ∗ I
σ (u, s)E Q λSs |Iu − ds dWuQ

−
0 u Bn (u) T p
or
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]
Z tZ Z T∗ !
I(u) T p(u, s)
(3.4) + k(u, z) + h(s, u, z) ds q(du, dz)
0 E Bn (u) u Bn (u)
Z t Z T∗ !
T p(u, s) ∗ QI
S
λs |Iu − ds dWuQ

− σT p (u, s)E
0 u B n (u)
I
where h(s, u, z) is the predictable integrand in the martingale representation E Q [λSs |It ] =
RtR
E[λSs |I0 ] + 0 E h(s, u, z)q(du, dz), k(u, z) is the predictable integrand in the martingale rep-
RtR
resentation MtS = M0S + 0 E k(u, z)q(du, dz).
Proof. The somewhat technical proof can be found in Appendix 1. It mainly relies on the
absolute continuity of the compensator of St and on the martingale representation property of
a marked point process in its natural filtration.
Notice the integral with respect to q(du, dz) in Equation (3.3) does not appear in Møller’s
paper. We now have all the tools to derive the risk-minimizing strategy. We assume we can
trade in nominal 0-coupon bonds of s1 different maturities and in TIPS of s2 different maturities
with s1 + s2 = s .
Proposition 3.4. The risk-minimizing strategy ρ∗ = (ε∗t , ηt∗ ) = ε1t , . . . , εst , ηt∗ is given

by the solution of the following system of equations:
d
X·1 , X·1 t d
X·2 , X·1 t . . . d
X·s , X·1 t

d
J· , X·1 t

ε1t
    
 d J· , X 2 1 2 2 2 . . . d X·s , X·2 t   ε2t 
 =  d X· , X · t d X· , X · t
 
 · t  
 ... 
...
...
... ... ...
  
where
Pn (t, Ti )Pn (t, Tj ) h ∗ ∗ 0
i
d X·i , X·j t =

σ n (t, T )σ
i n (t, Tj ) dt
Bn2 (t)
for all i, j = 1, . . . , s1 ,
T p(t, Ti )T p(t, Tj ) h ∗ ∗ 0
i
d X·i , X·j t =

σ Tp (t, T )σ
i Tp (t, Tj ) dt
Bn2 (t)
for all i, j = s1 + 1, . . . , s, and
Pn (t, Ti )T p(t, Tj ) h ∗ ∗ 0
i
d X·i , X·j t =

σ n (t, Ti )σ T p (t, Tj ) dt
Bn2 (t)
for all i = 1, . . . , s1 and j = s1 + 1, . . . , s. Furthermore, we have
" Z T∗ ! #
Pn (t, Ti ) I 0
(3.5) d J· , X·i t = T p(t, s)σT∗ p (t, s)E Q λSs |It − ds σn∗ (t, Ti ) dt

Bn2 (t) t
for all i = 1, . . . , s1 . and

" ! #
Z T∗
T p(t, Ti ) I 0
i
T p(t, s)σT∗ p (t, s)E Q ∗

S
(3.6) d J· , X· t = λs |It − ds σT p (t, Ti ) dt
Bn2 (t) t
for all i = s1 + 1, . . . , s.
The discounted amount invested in the risk free asset is given by:
s 1 s
Z T∗
T p(t, u) QI S X
i Pn (t, Ti )
X T p(t, Ti )
ηt∗ εit

= E λu |It du − εt −
t Bn (t) Bn (t) 1
Bn (t)
i=1 i=s +1
As far as the cost process is concerned, we have:
Proposition 3.5. The cost process Ct (ρ∗ ) is given by
Ct (ρ∗ ) = E Q [AT ∗ |G0 ]

Z tZ Z T∗ !
I(u) T p(u, s)
(3.7) + k(u, z) + h(s, u, z) ds q(du, dz)
0 E Bn (u) u Bn (u)
Z t Z T∗ !
T p(u, s) ∗ QI
S ∗
λs |Iu − ds − σRM (u) dWuQ

− σT p (u, s)E
0 u B n (u)
and the risk process Rt (ρ∗ ):

 !2
Z T∗ Z Z T∗

I(u) T p(u, s)
Rt (ρ∗ ) = E Q 

k(u, z) + h(s, u, z) ds p(du, dz) Gt 
t E Bn (u) u Bn (u)
2
h R ∗ R ∗ i
T T T p(u,s) ∗
+ EQ QI λS |I ∗ (u) du G

(3.8) t u σ
Bn (u) T p (u, s)E s u − ds− σ RM t
∗ (u) =
Ps1 i Pn (u,Ti ) ∗
Ps i T p(u,Ti ) ∗
where σRM i=1 εu Bn (u) σn (u, Ti ) + i=s1 +1 εu Bn (u) σT p (u, Ti ) and k·k is the Eu-
clidean norm.
It is worth noticing that in Equation (3.7), the cost process consists in the sum of two
stochastic integrals. The first one measures the part that cannot be hedge due to the policy-
holder’s claims. The second stochastic integral measures the part that cannot be hedge due to
the incompleteness of the financial market (inflation and interest rates).
Rt
Proof. The cost process is defined by Ct (ρ∗ ) = Vt (ρ∗ ) − 0 ε∗u dXu + At = E Q [AT ∗ |Gt ] −
Rt ∗
0 εu dXu . We immediately get the result thanks to Proposition 3.3 and the dynamics of the
asset prices.
As far as the risk process is concerned, using Equation (3.7), we immediately get:

R T ∗ R I(u) R T ∗ T p(u,s) 2 
t E B (u)
k(u, z) + u Bn (u) h(s, u, z) ds q(du, dz)
Rt (ρ∗ ) = E Q   R T ∗ R T ∗ Tnp(u,s)  Gt 

I
∗ ∗

(u) dWuQ
Q

− t u Bn (u) σT p (u, s)E λSs |Iu − ds − σRM

2
Q R T ∗ R I(u) R T ∗ T p(u,s)
= E t E Bn (u)
k(u, z) + u Bn (u) h(s, u, z) ds p(du, dz) Gt
"Z ∗ #
T R ∗ 2
Q T T p(u,s) ∗ I ∗
+ E u Bn (u) σT p (u, s)E Q λSs |Iu − ds − σRM (u) du Gt

t
since the integrals with respect to q(t, ·) and WtQ are Q-strongly orthogonal.
4. Applications in Some Collective Models.

In this section and the following one, we study different models for the payments process
St . In particular in this section, we study two collective models of risk. In the first one, we
assume that the payments process follows the well-known compound Poisson process. This
model has already been studied in Møller [68]. He even studied a more general model since the
size of claims could depend on the value of a financial index. In the second one, we study an
extension of this simple model where we assume that the number of policies in the portfolio is
stochastic.
4.1. Compound Poisson process. In this section, we assume that the payment process
St follows a square integrable compound Poisson process3 i.e.
Nt
X
St = Zi
i=1
where Nt ∼ P oi(λ) is a Poisson process and (Zn )n≥0 are i.i.d. square integrable positive
I
random variables with expectation E Q [Zn ] = µZ . If we denote by Tn the random times
of jump of P the process Nt , we can introduce for each B ∈ B(R+ ), the counting process
Nc (t, B) = ∞ i=1 1{Tn ≤t} 1{Zn ∈B} . We obviously have Nt = N (t, R+ ). For each B ∈ B(R+ ),
R t Rc
the compensator of Nc (t, B) can be written as pc (t, B) = 0 B λ QZ (dz)du where QZ (dz) is
the distribution of the random variables Zn under QI . We denote by qc (t, B) the compensated
process: qc (t, B) = Nc (t, B) − pc (t, B). We can then write:
Z tZ
St = z Nc (du, dz)
0 R+
3We could extend the results of this subsection to more general increasing Lévy (additive) processes. Since
these processes can be described as a constant (deterministic) drift and a real-marked point process with
absolutely continuous compensator and a Lévy measure independent (dependent) of time, these processes fit
into the framework defined in Section 2.2.
248 Applications in Some Collective Models.
The filtration I is the one generated by the marked point process Nc (t, B). St is thus in-
Rt
deed I-adapted. Moreover, its compensator is absolutely continuous and given by 0 λSu du =
RtR Z
Rt Z
0 R+ z λQ (dz)du = 0 λµ du.
I
Proposition 4.1. The conditional expectation E Q λSu |It is given by

I
E Q λSu |It = λµZ

(4.1)
Proof. Obvious since λSu = R+ z λQZ (dz) = λµZ is a constant.
R


Z T∗
T p(t, u)
Vt (ρ∗ ) = λµZ du
t Bn (t)
Proof. Put Equation (4.1) in Equation (3.2).
Proposition 4.3. The martingale representation of E Q [AT ∗ |Gt ] is given here by
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]
Z tZ
I(u)
+ z qc (du, dz)
0 R+ Bn (u)
Z t Z T∗ !
T p(u, s)
− λµZ σ ∗ (u, s)ds dWuQ
0 u Bn (u) T p
I
Proof. The integrand h(s, u, z) is equal to 0 since E Q [λSs |Gt ] = λ µZ is a constant. The
integrand k(u, z) is obviously equal to z since
Z tZ Z tZ
S
Mt = z Nc (du, dz) − z λQZ (dz)du
0 R+ 0 R+
Z tZ
= z qc (du, dz)
0 R+

Proposition 4.4. The risk-minimizing strategy is given by Proposition 3.4 where
" Z T∗ ! #
P n (t, Ti ) 0
d J· , X·i t = λµZ T p(t, s)σT∗ p (t, s)ds σn∗ (t, Ti ) dt

Bn2 (t) t
for all i = 1, . . . , s1 . and

" ! #
Z T∗
T p(t, Ti ) 0
J· , X·i t Z
T p(t, s)σT∗ p (t, s) ds σT∗ p (t, Ti )

d = λµ dt
Bn2 (t) t
for all i = s1 + 1, . . . , s.
s 1 s
Z T∗
T p(t, u) X Pn (t, Ti ) X T p(t, Ti )
ηt∗ = λµ Z
du − εit − εit
t Bn (t) Bn (t) 1
Bn (t)
i=1 i=s +1
Proof. Put Equation (4.1) in Equations (3.5) and (3.6).
Notice that, in this simple setting, the financial strategies do not depend on the liabilities (or
only up to the constant λµZ ). Mathematically speaking, the financial strategy ρ∗ is F-adapted
and not G-adapted.
Proposition 4.5. The cost process is given by
Z tZ
∗ Q I(u)
Ct (ρ ) = E [AT ∗ |G0 ] + z qc (du, dz)
0 R+ Bn (u)
Z t
Z
R T ∗ T p(u,s) ∗
− λµ u σ
Bn (u) T p (u, s) ds − σ ∗ (u)
RM dWuQ
0
and the risk process Rt (ρ∗ ):

Z T h 2 i
∗ QI 2 QF I(u)
Rt (ρ ) = λE [Z ] E Bn (u)
Ft du
t
Z T∗ 2
QF Z R T∗ T p(u,s) ∗ ∗ (u)
+ E λµ Bn (u) σT p (u, s) − σRM Ft du

u
t
The first stochastic integral in the cost process translates the fact that the insurer cannot
hedge perfectly its liabilities due to the unpredictable arrival and size of the claims. The second
stochastic integral is related to the incompleteness of the financial market.
4.2. Stochastic portfolio size. In the previous section, we could have alternatively as-
sumed that the number of policies in the portfolio was a constant k and that the claims of each
policyholder followed independent compound Poisson processes with an intensity of arrival of
0
claims λ = λk . We now extend this model by allowing the number kt of policies to be stochastic.
We model kt as a point process in the following way:
Z t
kt = k0 + It − 1{ku− >0} dOu
0
where It and Ot are simple independent point processes. It describes

Rt the cumulative number
of new policyholders in the portfolio on the interval ]0, t] and 0 1{ku− >0} dOu the cumulative
number of policyholders leaving the portfolio on the interval ]0, t]. We study a model where
the size of the portfolio evolves heterogeneously with respect to time. We assume It admits a
deterministic intensity λI (t) and Ot a stochastic intensity kt− λO (t) where λO (t) is a determin-
istic function. Accordingly, we assume that the number of policyholders leaving the portfolio
depends on its size whereas the arrival of new policyholders does not. We denote by Λkt the
compensator of kt and the compensated process Mtk = kt − Λkt . The compensator Λkt is given
by
Z t
Λkt =
I
λ (u) − ku− λO (u) du

0
For each policyholder in the portfolio, we assume that the cost of claims follows an in-
dependent compound Poisson process. Let us (Sn )n≥1 denote the random times of arrival of
a claim in the portfolio and let us (Zn )n≥1 denote the i.i.d square integrable positive ran-
dom variables describing the sizePof the claims. We can then introduce for each B ∈ B(R+ ),
the counting process Nc (t, B) = ∞ i=1 1{Sn ≤t} 1{Zn ∈B} . Its compensator pc (t, B) is written as
RtR
pc (t, B) = 0 B λku− Q (dz)du where QI (dz) is the distribution of the random variables Zn .
I
We can then describe the payment process as:

Z tZ
St = z Nc (du × dz)
0 R+
Notice that we can rewrite our three point processes Nc (t, B), It and Ot in the form of a
single general point process as in Section 2.2. Consider the following measurable space (E, E)
where E = {c, i, o} × R+ and E = B(E). Let us define a point process (Tn , Yn )n≥1 where Yn
0
are E-valued random variables associated to the random times Tn . For each B ∈ E where
0
B = S × B, we can then define the general point process:
∞
0
X
N (t, B ) = 1{Tn ≤t} 1{Yn ∈B 0 }
n=1
whose compensator is given by
Z tXZ
0
p(t, B ) = msu (dz)λu (s)du
0 s∈S B
where λu (c) = λku− , mcu (dz) = QI (dz), λu (i) = λI (u), mi (dz) = δ{0} (dz), λu (o) = ku− λO (u),
mO (dz) = δ{0} (dz). In particular, we can write
It = N (t, (i, R+ ))
∞
X
= 1{Tn ≤t} 1{Yn ∈(i,R+ )}
n=1
Ot = N (t, (o, R+ ))
∞
X
= 1{Tn ≤t} 1{Yn ∈(o,R+ )}
n=1
Nc (t, B) = N (t, (c, B))

∞
X
= 1{Tn ≤t} 1{Yn ∈(c,B)}
n=1
0
for all B ∈ B(R+ ). The filtration I is the one generated by this process N (t, B ). St is indeed
I-adapted.
In order to determine the risk-minimizing strategy and its value process, we first need the
following result:
I
Proposition 4.6. The conditional expectation E Q λSu |It is given by:

I I
E Q λSu |It = λµZ E Q [ku |It ]

(4.2)
and where
Ru Rv
λO (w)dw
!
QI
Ru
λO (v)dv λI (v)e t dv Ru
λ0 (v)dv

(4.3) E [ku |It ] = kt e− t + R tu Rv
λO (w)dw
1 − e− t
t λO (v)e t dv
Proof. Since λSu = λku− µZ , we have

I I
E Q λSu |It = E Q λku− µZ |It

I
= λµZ E Q [ku |It ]
I
For the expression of E Q [ku |It ], see Appendix 2.
Z T∗
∗ Z T p(t, u) QI
Vt (ρ ) = λµ E [ku |It ] du
t Bn (t)
Z tZ
I(u)
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ] + z qc (du, dz)
0 R+ Bn (u)
!
Z t Z T∗
T p(u, s) − R s λO (v)dv
+ λµZ e u ds dMuk
0 u Bn (u)
!
Z t Z T∗
Z T p(u, s) ∗ QI
− λµ σ (u, s)E [ks |Iu ]− ds dWuQ
0 u Bn (u) T p
I
Proof. In Appendix 3, we show that the martingale representation of E Q
S
λs |Iu is
given by:
I
Rs O
dE Q λSs |Iu = λµZ e− u λ (w)dw dMuk

The integrand h(s, u, z) is then given by

Rs
λO (w)dw
h(s, u, z) = λµZ e− u
Moreover the integrand k(u, z) in the martingale representation of MtS is obviously equal to z
since
Z tZ Z tZ
S
Mt = z Nc (du, dz) − z λku− P (dz)du
0 R+ 0 R+
Z tZ
= z qc (du, dz)
0 R+

" Z T∗ ! #
P n (t, Ti ) I 0
d J· , X·i t = λµZ T p(t, s)σT∗ p (t, s)E Q [ks |It ]− ds σn∗ (t, Ti ) dt

Bn2 (t) t
for all i = 1, . . . , s1 . and

" ! #
Z T∗
T p(t, Ti ) QI 0
i
= λµZ T p(t, s)σT∗ p (t, s)E [ks |It ]− ds σT∗ p (t, Ti )

d J· , X· t
dt
Bn2 (t) t
for all i = s1 + 1, . . . , s.
s s 1
Z T∗
T p(t, u) QI X Pn (t, Ti ) X T p(t, Ti )
ηt∗ = λµ Z
E [ku |It ] du − εit − εit
t Bn (t) Bn (t) 1
Bn (t)
i=1 i=s +1
The interesting feature here is that the financial strategy ρ∗ is not anymore F-adapted
but well G-adapted. Indeed, the risk-minimizing strategy depends on the size of the portfolio
I
kt− (through E Q [ks |It ]− ).
Proposition 4.10. The cost process is given by

Z tZ
I(u)
Ct (ρ∗ ) = E Q [AT ∗ |G0 ] + z qc (du, dz)
0 R+ Bn (u)
!
Z t Z T∗
T p(u, s) − R s λO (v)dv
(4.4) + λµZ e u ds dMuk
0 u Bn (u)
Z t
Z
R T∗ T p(u,s) ∗ QI

− λµ u Bn (u) σT p (u, s)E
∗ (u)
[ks |Iu ]− ds − σRM dWuQ
0
The risk process Rt (ρ∗ ) is given by:

Z T∗ " 2 #
I F I(u) I
Rt (ρ∗ ) = λE Q Z 2 EQ Ft E Q [ku |It ] du

t Bn (u)
Z T∗

Z T∗ !2 
2 F T p(u, s) Rs O I
h i
+ λµZ EQ  e− u λ (v)dv ds Ft  E Q λku |It du

t u Bn (u)
Z T∗ 2
Q Z
R T ∗ T p(u,s) ∗ Q I ∗
+ E λµ u Bn (u) σT p (u, s)E [ks |Iu ]− ds − σRM (u) Gt du

t
QI I
k
λu |It = λI (u) − E Q [ku |It ] λO (u).

where E
In the cost process of the previous model, we had a term similar to the first stochastic
integral of Equation (4.4) but the second integral was absent. Intuitively, these terms mean
that the insurer cannot hedge perfectly its liabilities due to the unpredictable arrival and size
of the claims (for a given size of his portfolio), and due to the unpredictable evolution of the
number of policies in his portfolio. The third stochastic integral is related to the incompleteness
of the financial market.
5. Applications in Some Individual Models.

In this section, we study two different models. In both cases, we consider a single policy
and assume that a single claim can occur. We also assume that the claim is not paid up-front
when reported but is settled through time. In the first model, we assume the claim is settled
according to a deterministic function. In the second model, we assume the claim is settled
according to a stochastic process. The fact that we study a single policy seems rather limited
but keep in mind the risk-minimizing strategy of a portfolio of independent policies4 is given by
the sum of the risk-minimizing strategies of each policy but with respect to their own particular
filtration (not the whole filtration generated by the whole portfolio). Under this independence
assumption, we can thus easily extend the following results to a portfolio.
5.1. Deterministic claim settlement. We consider a single policy. Let T be the ran-
dom time of occurrence of a claim. The insurer covers the claims if it occurs in the deterministic
interval [0, L]. We assume the policyholder notifies the claims at the time of occurrence. The
claim amount is a square integrable random variable Z with distribution QZ (dz). This amount
is settled through time according to an increasing deterministic function F (t) taking its value
between [0, 1] with F (t) = 0 for t ≤ 0. We assume there is no upfront payment at the time of
notification since F (0) = 0.
Here, the general point process of Section 2.2 comes down to N (t, B) = 1{T ≤t} 1{Z∈B} for
all B ∈ B(R+ ). The filtration is here It = σ(N (s, B) : 0 ≤ s ≤ t, B ∈ B(R+ )). We assume
4By independent policies, we mean that the payment processes expressed in real units and associated to
the different policies are independent of each other under QI . Obviously, expressed in nominal units, they are
not anymore independent due to the inflation.
254 Applications in Some Individual Models.
the random time T has an absolutely continuous distribution QI (T ≤ t) = G(t). Let us define
λ(t) = −d[ln(1−G(t))]
dt , then the process
Z tZ
q(t, B) = N (t, B) − 1{T >u−} λ(u)QZ (dz)du
0 B
is a (QI , I)-martingale.
At time t, the cumulative payments process St is then given by
St = 1{T ≤L} 1{T ≤t} Z F (t − T )
where St is indeed I-adapted. Since F (T − T ) = g(0) = 0, we have in differential form:
dSt = 1{T ≤L} 1{T ≤t−} Z dF (t − T )
Since St is increasing and predictable, theR compensator ΛSt of St is St itself. Let us assume
t
there is a function f (t) such that F (t) = 0 f (u)du. We then have an absolutely continuous
compensator:
dΛSt = λSt dt
= 1{T ≤L} 1{T ≤t−} Z f (t − T )dt
following result:
I

I
E Q λSu |It = 1{t≤L} 1{T >t} µZ f ∗ (t, u)

(5.1) + 1{T ≤L} 1{T ≤t} Z f (u − T )
R L∧u Rv
where f ∗ (t, u) = t f (u − v)λ(v)e− t λ(w)dw
dv
Proof. See Appendix 4 for a proof.

Z T∗
∗ Z T p(t, u) ∗
Vt (ρ ) = 1{T >t} 1{t≤L} µ f (t, u) du
t Bn (t)
Z T∗
T p(t, u)
+ 1{T ≤t} 1{T ≤L} Z f (u − T ) du
t Bn (t)
R L∧u Rv
where f ∗ (t, u) = t f (u − v)λ(v)e− t λ(w)dw dv.
We can now apply the general result of Proposition 3.3 to find the martingale representation:
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]

!
Z t Z Z T∗
T p(u, s)
z f (s − u) − µZ f ∗ (u, s) ds q(du, dz)

+ 1{u≤L}
0 R+ u Bn (u)
Z t Z T∗ !
Z T p(u, s) ∗ ∗
− 1{u≤L} 1{T >u−} µ σT p (u, s) f (u, s) ds dWuQ
0 u B n (u)
Z t Z T∗ !
T p(u, s) ∗
− 1{T ≤L} 1{T ≤u−} Z σT p (u, s) f (s − T ) ds dWuQ
0 u B n (u)
R L∧s Rv
where f ∗ (u, s) = u f (s − v)λ(v)e− u λ(w)dw
dv.
I
Proof. In Appendix 5, we show the martingale representation of E Q λSt |Is is given by

Z
I
dE Q λSt |Is = 1{s≤L} z f (t − s) − µZ f ∗ (s, t) q(ds, dz)

R+
so the integrand h(s, u, z) is given by
h(s, u, z) = 1{u≤L} z f (s − u) − µZ f ∗ (u, s)

Rt
Moreover, the integrand k(u, z) is equal to 0 because 0 BI(u)
n (u)
dMuS = 0 since St = ΛSt . We get
immediately the result if we put the integrands h(s, u, z) and k(u, z) in Equation (3.4).
The risk-minimizing strategy is given by the following proposition.

" Z T∗ ! #
P n (t, Ti ) 0
d J· , X·i t = 1{t≤L} 1{T >t−} µZ T p(t, s) σT∗ p (t, s) f ∗ (t, s) ds σn∗ (t, Ti ) dt

Bn2 (t) t
" Z T∗ ! #
Pn (t, Ti ) ∗ ∗ 0
+ 1{T ≤L} 1{T ≤t−} Z T p(t, s) σT p (t, s) f (s − T ) ds σn (t, Ti ) dt
Bn2 (t) t
for all i = 1, . . . , s1 . and

" ! #
Z T∗
T p(t, Ti ) 0
i
= 1{t≤L} 1{T >t−} µZ T p(t, s) σT∗ p (t, s) f ∗ (t, s) ds σT∗ p (t, Ti )

d J· , X· t
dt
Bn2 (t) t
" ! #
Z T∗
T p(t, Ti ) 0
= 1{T ≤L} 1{T ≤t−} Z T p(t, s)σT∗ p (t, s) f (s − T ) ds σT∗ p (t, Ti ) dt
Bn2 (t) t
for all i = s1 + 1, . . . , s.
The discounted amount to invest in the risk free asset is given by:
s s 1
Z T∗
T p(t, u) QI S X
i Pn (t, Ti )
X T p(t, Ti )
ηt∗ εit

= E λu |It du − εt −
t Bn (t) Bn (t) 1
Bn (t)
i=1 i=s +1
The interesting feature here is that we have two different financial strategies depending on
whether the claim has already been reported or not. Accordingly, we distinguish the RBNI
(Reserved but not incurred) claims and the RBNS (Reserved but not settled) claims. In other
words, the asset allocation of our non-life insurance will depend on the structure of its liabilities.
Proposition 5.5. The cost process Ct (ρ∗ ) is given by
Ct (ρ∗ ) = E Q [AT ∗ |G0 ]

!
Z t Z Z T∗
T p(u, s) Z ∗

+ 1{u≤L} z f (s − u) − µ f (u, s) ds q(du, dz)
0 R+ u Bn (u)
Z t Z T∗ !
T p(u, s)
− 1{u≤L} 1{T >u−} µZ σ ∗ (u, s) f ∗ (u, s) ds − σRM
∗
(u) dWuQ
0 u Bn (u) T p
Z t Z T∗ !
T p(u, s) ∗ ∗
− 1{T ≤L} 1{T ≤u−} Z σ (u, s) f (s − T ) ds − σRM (u) dWuQ
0 u Bn (u) T p
The risk process Rt (ρ∗ ) is given by:

Z T∗ Z h i Ru
F
Rt (ρ∗ ) = 1{T >t} 1{u≤L} EQ H1 (u, z, T ∗ )2 |Ft QZ (dz) e− t λ(v)dv λ(u)du
t R+
h R ∗ R ∗ 2 i
T T
+ EQ T p(u,s) ∗ QI ∗
S
Bn (u) σT p (u, s)E λs |Iu − ds − σRM (u) du Gt

t u
R T∗ T p(u,s)
where H1 (u, z, T ∗ ) = z f (s − u) − µZ f ∗ (u, s) ds.

u Bn (u)
Proof. Straightforward due to Proposition 3.5.
As in the general model, the cost process consists of two parts. The first one corresponds
to the stochastic integral with respect to q(du, dz) and represents the part that cannot be
hedged due to the unpredictable arrival and unpredictable size of the policyholder’s claim. The
second part corresponds to the two stochastic integrals with respect to the Brownian motion
and represents the part of the claim process that cannot be hedged due the incompleteness of
the financial market. Notice the impact of the incompleteness of the financial market depends
on whether the claims has already happened or not.
5.2. Stochastic claim settlement. Again, we consider a single policy. We assume the
claim occurrs and is reported at the random time T (described as in Section 5.1). The insurer
covers the claim if it occurs in the deterministic interval [0, L]. The total claim amount is a
square integrable random variable Z with distribution QZ (dz). This amount is settled through
time according to a stochastic process F (t) taking its value between [0, 1] . In other words, the
payment process St is given by
St = 1{T ≤L} 1{t≥T } Z F (t − T )
We assume the total amount Z is known at the time of notification; Z is thus GT -measurable.
As in the previous example, for each B ∈ B(R+ ), we can define a point process NZ (t, B) =
RtR
1{T ≤t} 1{Z∈B} with a compensator pz (t, B) = 0 B 1{T >u−} λ(u)QZ (dz)du
We describe the stochastic process F (t) as a Beta-Stacy process. See Walker and Muliere
[82] and Hjort [54] for more details about beta-Stacy process. We describe here the model
used by Hjort [54] except that we assume At contains no fixed point of discontinuity.
Following Hjort [54], we define F (t) in the following way:
Y
F (t) = 1 − (1 − ∆As )
0≤s≤t
with At a beta process: At ∼ beta(c(·), A0 (·)) where A0 = 0, c(t) is a piecewise continuous

non negative function and A0 (·) is an increasing right-continuous function with A0 (0) = 0.
According to Walker and Muliere [82], F (t), defined in this way, is a Beta-Stacy process.
A Beta process At ∼ beta(c(·), A0 (·)) can also be described as an integral with respect to
a real-marked point process NA (t, B), whose compensator pA (t, B) is given by pA (t, B) =
R (1−a)c(t)−1
B c(t) a da A0 (dt) for each B ∈ B([0, 1]), in the following way:
Z tZ
At = a NA (du, da)
0 [0,1]
Since we assume At contains no fixed points of discontinuity, the function A0 (·) is continuous.
Rt 0
If furthermore, we assume the A0 (t) is absolutely continuous i.e. A0 (t) = 0 A0 (u)du, the beta
process At can also be described as an increasing additive process with (an inhomogeneous)
c(t)−1 0
Lévy measure νtA (da) = c(t) (1−a)a A0 (t) da
As far as St is concerned, we have:
dSt = 1{T ≤L} 1{T ≤t−} Z dF (t − T )
since the measure A0 (dt) is, by assumption, continuous at time 0, there is a.s. no up front
payment at time T . We can also write St as a stochastic integral with respect to the real-
marked point process NA (t, ·):
Z
dSt = 1{T ≤L} 1{T ≤t−} Z F̄ (t − T )− a NA (d(t − T ), da)
[0,1]
where F̄ (t − T ) = 1 − F (t − T ). Indeed we have

Y
F (t) = 1 − (1 − ∆As )(1 − ∆At )
0≤s<t
= F (t− ) + [1 − F (t− )] ∆At
so we can write Z
dF (t) = [1 − F (t− )] a NA (dt, da)
[0,1]
The compensator of St is indeed also absolutely continuous and is given by:
Z
S
λ (t)dt = 1{T ≤L} 1{T ≤t−} Z F̄ (t − T )− a pA (d(t − T ), da)
[0,1]
= 1{T ≤L} 1{T ≤t−} Z F̄ (t − T )− A00 (t −T )dt
h i 1
since [0,1] c(t − T ) (1 − a)c(t−T )−1 da = − c(t−T
1 c(t−T ) 1
R
) (1 − a) = c(t−T
0 ).
We can also model both marked point processes NZ (t, B) and NA (t, B) through a single more
general marked point process N (t, B) as in the general framework of Section 2.2. The filtration
I will be the filtration generated by N (t, B) and the process St is indeed I-adapted.
following result:
I

I
E Q λSu |It = 1{t≤L} 1{T >t} µZ f ∗ (t, u)

(5.2)
+ 1{T ≤L} 1{T ≤t} Z F̄ (t − T ) fT (t, u)
where
0
fT (t, u) = A0 (u − T )e−(A0 (u−T )−A0 (t−T ))
Z L∧u Rv
∗
f (t, u) = fv (v, u)λ(v)e− t λ(w)dw dv
t
As far as the value process is concerned, we have:
Z T∗
∗ Z T p(t, u) ∗
Vt (ρ ) = 1{t≤L} 1{T >t} µ f (t, u) du
t Bn (t)
Z T∗
T p(t, u)
+ 1{T ≤L} 1{T ≤t} Z F̄ (t − T ) fT (t, u)du
t Bn (t)
We can now apply the general result of Proposition 3.3 to find the following martingale
representation.
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]

Z t Z
+ 1{u≤L} H1 (u, z, T ∗ ) qZ (du, dz)
0 R+
Z t Z
+ 1{T ≤u−} 1{T ≤L} Z a H2 (u, T, T ∗ ) qA (d(u − T ), da)
0 [0,1]
"Z #
Z t T∗
Z T p(u, s) ∗ ∗
+ 1{T >u−} 1{u≤L} µ σ (u, s) f (u, s) ds dWuQ
0 u Bn (u) T p
"Z #
Z t T∗
T p(u, s) ∗
+ 1{T ≤u−} 1{T ≤L} Z σ (u, s) fT (u, s) F̄ (u − T )− ds dWuQ
0 u Bn (u) T p
where
Z T∗
∗ T p(u, s)
z fu (u, s) − µZ f ∗ (u, s) ds

H1 (u, z, T ) =
u Bn (u)
and
Z T∗
I(u) T p(u, s)
H2 (u, T, T ∗ ) = − fv (u, s)F̄ (u − v)− ds
Bn (u) u Bn (u)
I
Proof. In Appendix 7, we show the martingale representation of E Q
S
λs |Iu is given by:
Z
QI
z fu (u, s) − µZ f ∗ (u, s) qZ (du, dz)
S
dE λs |Iu = 1{u≤L}
R+
Z
−1{T ≤L} 1{T <u−} Z F̄ (u − T )− a fT (u, s) qA (d(u − T ), da)
[0,1]
Moreover
Z t Z t
I(u) I(u)
dMuS = 1{T ≤u−} 1{T ≤L} Z d F (t − T ) − pF (t)dt

0 Bn (u) 0 Bn (u)
Z t Z
I(u)
= 1{T ≤u−} 1{T ≤L} Z a qA (d(u − T ), da)
0 Bn (u) [0,1]
We can now easily find the risk-minimizing strategy.


" ! #
Z T∗
Pn (t, Ti ) 0
i
= 1{t≤L} 1{T >t−} µZ T p(t, s) σT∗ p (t, s) f ∗ (t, s) ds σn∗ (t, Ti )

d J· , X· t
dt
Bn2 (t) t
" ! #
Z T∗
Pn (t, Ti ) 0
+ 1{T ≤L} 1{T ≤t−} Z F̄ (t − T ) T p(t, s) σT∗ p (t, s) fT (t, s) ds σn∗ (t, Ti ) dt
Bn2 (t) t
for all i = 1, . . . , s1 . and

" Z T∗ ! #
T p(t, T i ) 0
i
= 1{t≤L} 1{T >t−} µZ T p(t, s) σT∗ p (t, s) f ∗ (t, s) ds σT∗ p (t, Ti ) dt

d J· , X· t Bn2 (t) t
" Z T∗ ! #
T p(t, Ti ) 0
+ 1{T ≤L} 1{T ≤t−} Z F̄ (t − T ) T p(t, s) σT∗ p (t, s) fT (t, s) ds σT∗ p (t, Ti ) dt
Bn2 (t) t
for all i = s1 + 1, . . . , s.
s 1 s
Z T∗
T p(t, u) QI S X
i Pn (t, Ti )
X T p(t, Ti )
ηt∗ εit

= E λu |Gt du − εt −
t Bn (t) Bn (t) 1
Bn (t)
i=1 i=s +1
As in the previous model, we have two different financial strategies depending on whether
or not the claim has already been reported. We can again distinguish the RBNI (Reserved but
not incurred) claims and the RBNS (Reserved but not settled) claims.
Proposition 5.10. The cost process Ct (ρ∗ ) is given by:
Ct (ρ∗ ) = E Q [AT ∗ |G0 ]

Z t Z
+ 1{u≤L} H1 (u, z, T ∗ ) qZ (du, dz)
0 R+
Z t Z
+ 1{T ≤u−} 1{T ≤L} Z a H2 (u, T, T ∗ ) qA (d(u − T ), da)
0 [0,1]
"Z #
Z t T∗
Z T p(u, s) ∗ ∗ ∗
+ 1{T >u−} 1{u≤L} µ σ (u, s) f (u, s) ds − σRM (u) dWuQ
0 u Bn (u) T p
"Z #
Z t T∗
T p(u, s) ∗ ∗
+ 1{T ≤u−} 1{T ≤L} Z σ (u, s) fT (u, s) F̄ (u − T )− ds − σRM (u) dWuQ
0 u Bn (u) T p
The risk process Rt (ρ∗ ):

Z T∗ Z h i Ru
∗ ∗ 2
Rt (ρ ) = 1{T >t} 1{u≤L} E Q
H1 (u, z, T ) Ft QZ (dz)e− t λ(v)dv λ(u)du
t R+
Z L Z T∗ Z h i Rv
QI
+ 1{T >t} E [Z ]2
a2 E Q H2 (u, v, T ∗ )2 Ft pA (d(u − v), da)λ(v)e− t λ(w)dw dv

t∧L v [0,1]
Z T∗ Z h i
+ 1{T ≤t} 1{T ≤L} Z2 a2 E Q H2 (u, T, T ∗ )2 Ft ∪ F (t − T ) pA (d(u − T ), da)

t [0,1]
h R ∗ R ∗ 2 i
T T
+ EQ T p(u,s) ∗ QI ∗ (u) du G
S
t u Bn (u) σT p (u, s)E λs |Iu − ds − σRM t
where
Z T∗
∗ T p(u, s)
z fu (u, s) − µZ f ∗ (u, s) ds

H1 (u, z, T ) =
u Bn (u)
and
Z T∗
∗ I(u) T p(u, s)
H2 (u, v, T ) = − fv (u, s)F̄ (u − v)− ds
Bn (u) u Bn (u)
Again, the cost process consists in two parts. The first one is related to the policyholder’s
claim and corresponds to the two first stochastic integrals. As in the previous model, the integral
with respect to q(du, dz) represents the part that cannot be hedged due to the unpredictable
arrival and unpredictable total size of the policyholder’s claim. The integral with respect to
qA (d(u − T ), da), did not appear in the previous model. It represents the part of the claim
process that cannot be hedged due to the unpredictable settlement of the claim. As in the
previous model, the two last stochastic integrals represent the part of the claim process that
cannot be hedged due the incompleteness of the financial market.
Appendix 1.
Proposition. The martingale representation of E Q [AT ∗ |Gt ] is given by
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]

Z t
I(u)
+ dMuS
0 B n (u)
Z t Z "Z T ∗ #
T p(u, s)
+ h(s, u, z) ds q(du, dz)
0 E u Bn (u)
Z t Z T∗
T p(u, s) ∗ I
σT p (u, s)E Q λSs |Iu − ds dWuQ

−
0 u Bn (u)
262 Appendix 1.
or
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]

Z tZ " Z T∗ #
I(u) T p(u, s)
+ k(u, z) + h(s, u, z) ds q(du, dz)
0 E Bn (u) u Bn (u)
Z t Z T∗
T p(u, s) ∗ I

−
0 u Bn (u)
I
where h(s, u, z) is the predictable integrand in the martingale representation E Q [λSs |It ] =
I RtR
E Q [λSs |I0 ] + 0 E h(s, u, z)q(du, dz), k(u, z) is the predictable integrand in the martingale
RtR
representation MtS = M0S + 0 E k(u, z)q(du, dz).
Proof. We have
"Z #
T∗
I(u)
E Q [AT ∗ |Gt ] = E Q dSu |Gt
0 Bn (u)
Z t "Z ∗ #
T
I(u) I(u)
= dSu + E Q dSu |Gt
0 Bn (u) t Bn (u)
Z t Z T∗
I(u) I(t)Pr (t, u) Q S
(5.3) = dSu + E λu |Gt du
0 Bn (u) t Bn (t)
thanks to Proposition 3.2. The expression I(t)Pr (t,u)

Bn (t) corresponds to the discounted price at
time t of a TIPS with maturity u. Using the integration by parts formula, we can write for
each u:
Z t
T p(t, u) Q S T p(0, u) Q S T p(s, u) Q S
E λu |Gt = E λu |G0 + dE λu |Gs
Bn (t) Bn (0) 0 Bn (s)
Z t
Q
S T p(s, u) T p(·, u) Q S
(5.4) + E λu |Gs − d + , E λu |G·
0 Bn (s) Bn (·) t
The square bracket is null in the last formula. We saw in Section 2.1 that:
Z t
T p(t, u) T p(0, u) T p(s, u) ∗
= − σT p (u, s)dWsQ
Bn (t) Bn (0) 0 B n (s)
I
Furthermore, since λSu is I-adapted and I and F are independent, E Q λSu |Gt = E Q λSu |It .

This term is obviously a (QI , I)-martingale for each u. Thanks to the martingale representation
property of a marked point process in its own natural filtration (see Brémaud [33]), we can
write for each u:
Z tZ
QI QI
S S
E λu |It = E λu |I0 + h(u, v, z)q(dv, dz)
0 E
for a predictable process h(u, ·, ·) which is σ(N (t, A), ∀A ∈ B(R+ ))-measurable. We can then
write Equation (5.4) as:
T p(t, u) Q S T p(0, u) Q S
E λu |Gt = E λu |G0
Bn (t) Bn (0)
Z t Z
T p(s, u)
+ h(u, s, z)q(ds, dz)
0 Bn (s) E
Z t
T p(s, u) ∗
E Q λSu |Gs − σT p (u, s)dWsQ

−
0 B n (s)
The second integral in Equation (5.3) can then be written as:
Z T∗ Z T∗
I(t)Pr (t, u) Q S T p(0, u) Q S
E λu |Gt du = E λu |G0 du
t Bn (t) t Bn (0)
Z T∗ Z t Z
T p(s, u)
+ h(u, s, z)q(ds, dz)du
t 0 Bn (s) E
Z T∗ Z t
T p(s, u) ∗
E Q λSu |Gs − σ (u, s)dWsQ du

−
t 0 Bn (s) T p
By interchanging the order of integration, we have:
Z T∗ Z T∗
I(t)Pr (t, u) Q S T p(0, u) Q S
E λu |Gt du = E λu |G0 du
t Bn (t) t Bn (0)
Z t Z T∗ Z
T p(s, u)
+ h(u, s, z)du q(ds, dz)
0 t Bn (s) E
Z t Z T∗
T p(s, u) ∗
E Q λSu |Gs − σT p (u, s)du dWsQ

−
0 t B n (s)
Z T∗
T p(0, u) Q S
= E λu |G0 du
0 Bn (0)
Z t Z T∗ Z
T p(s, u)
+ h(u, s, z)du q(ds, dz)
0 s Bn (s) E
Z t Z T∗
T p(s, u) ∗
σT p (u, s)E Q λSu |Gs − du dWsQ

−
0 s Bn (s)
Z t
T p(0, u) Q S
− E λu |G0 du
0 Bn (0)
Z tZ t Z
T p(s, u)
− h(u, s, z)du q(ds, dz)
0 s Bn (s) E
Z tZ t
T p(s, u) ∗
E Q λSu |Gs − σT p (u, s)du dWsQ

+
0 s B n (s)
264 Appendix 2.
If we denote by (I) the three last lines, and interchange the order of integration of these
integrals, we have:
Z t
T p(0, u) Q S
(I) = − E λu |G0 du
0 Bn (0)
Z tZ u Z
T p(s, u)
− h(u, s, z) q(ds, dz) du
0 0 Bn (s) E
Z tZ u
T p(s, u) ∗
σT p (u, s)E Q λSu |Gs − dWsQ du

+
0 0 Bn (s)
Z t
T p(u, u) Q S
= − E λu |Gu du
0 Bn (u)
Z t
I(u) S
= − λu du
0 Bn (u)
Rt
Since E Q [AT ∗ |G0 ] = 0 TBp(0,u) E Q λSu |G0 du, we eventually get the decomposition for E Q [AT ∗ |Gt ]:

n (0)
E Q [AT ∗ |Gt ] = E Q [AT ∗ |G0 ]

Z t
I(u)
+ dMuS
0 Bn (u)
Z t Z "Z T ∗ #
T p(u, s)
+ h(s, u, z) ds q(du, dz)
0 E u Bn (u)
Z t Z T∗
T p(u, s) ∗ I

−
0 u Bn (u)
where dMuS = dS S
u − λu du is a (Q, G)-martingale and h(s, u, z) is the integrand for the mar-
S
tingale E λu |I· .
Since MuS is also a (I, QI )-martingale, we can also find a predictable process k(u, z) such
RtR
that MtS = M0S + 0 E k(u, z)q(du, dz). We thus directly obtain the second representation of
E Q [AT ∗ |Gt ].
Appendix 2.
Proposition. For u > t, we have:
Z u Rv
R
u O
I λO (w)dw
E Q [ku |It ] = kt + λI (v)e t dv e− t λ (v)dv
t
or
Ru Rv
λO (w)dw
!
QI
Ru
λO (v)dv

E [ku |It ] = kt e− t + R tu Rv
λO (w)dw
1 − e− t
t λO (v)e t dv
Proof. We have
I I
E Q [ku |It ] = kt + E Q [ku − kt |It ]
Z u Z u
I I
= kt + λ (v)dv − λO (v)E Q [kv |It ] dv
t t
Ru
We have to find the solution f (s) of the following equation: f (u) = kt + t λI (v)dv −
Ru O 1 0 λI (u)
t λ (v)f (v)dv or of the following differential equation f (u) − λO (u) f (u) + λO (u) = 0. The
solution of this equation is of the form
Ru I Rv O
0 λ (w)dw dv
0 λ (v)e − 0u λO (v)dv
R
f (u) = Ru
O
+ Ce
e 0 λ (v)dv
where C is a constant. We know that f (t) = kt . So we have
Rt I Rv O
0 λ (w)dw dv
0 λ (v)e
Rt O
Rt
O
+ Ce− 0 λ (v)dv = kt
e 0 λ (v)dv
and thus
Rt I Rv O !
λ (v)e 0 λ (w)dw dv Rt O
C = kt − 0 Rt e 0 λ (v)dv
O
e 0 λ (v)dv
The expectation is then given by:
Ru I Rv O Rt I Rv O !
0 λ (w)dw dv λ (v)e 0 λ (w)dw dv
QI 0 λ (v)e − tu λO (v)dv
R
0
E [ku |It ] = Ru
O
+ kt − Rt e
e 0 λ (v)dv O
e 0 λ (v)dv
Z u R
− tu λO (v)dv
R Rv O u O
= kt e + I
λ (v)e 0 λ (w)dw
dv e− 0 λ (v)dv
t
Eventually we have:
Ru
Z u Rv
Ru
QI − λO (v)dv λO (w)dw λO (v)dv
E [ku |It ] = kt e t + I
λ (v)e t dv e− t
t
We can also rewrite this formula as a weighted average. Let us first notice that
Z u Rv O Z u h R
− 0t λO (v)dv t O
R i
O λ (w)dw
λ (v)e t dv = e d e 0 λ (v)dv
t t
Ru
λO (v)dv
= e t −1
So we can write:
Ru Rv
λO (w)dw
!
QI
Ru
λO (v)dv
Ru O
E [ku |It ] = kt e− t + R tu Rv
λO (w)dw
e t − 1 e− t λ (v)dv
t λO (v)e t dv
Ru O
Rv !
−
Ru
λO (v)dv λI (v)e t λ (w)dw dv Ru
λO (v)dv

= kt e t + R tu Rv
O
1 − e− t
O t λ (w)dw dv
t λ (v)e

266 Appendix 4.
Appendix 3.
Proposition. For s > u, we have
I
Rs O
dE Q λs |Iu = λµZ e− u λ (v)dv dMuk
S
I I
Proof. Since E Q λSs |Iu = λµZ E Q [ks |Iu ], we have to find the martingale representa-

I
tion of E Q [ks |Iu ]. Using Itô’s lemma on Equation (4.3) (and the Leibniz rule), we have:
I
Rs
λO (u)du
dE Q [ks |Iu ] = λO (u)e− ku− du
u
Z s
Rs Rv O

O − u λO (u)du I λ (w)dw
+ λ (u)e λ (v)e u dv du
u
Z s
− us λO (u)du ∂
R Rv O
I λ (w)dw
+ e λ (v)e u dv du
∂u u
Rs
λO (v)dv
+ e− u dku
Rs Rs O
Z s Rv O

λO (u)du
= λO (u)e− ku− du + λO (u)e− u λ (u)du
u λI (v)e u λ (w)dw dv du
Rs O
Z s Rv O
u Ru O

− u λ (u)du O I λ (w)dw I λ (w)dw
+ e −λ (u) λ (v)e u dv − λ (u)e u du
u
Rs
− λO (v)dv
+ e u dku
− us λO (u)du
Rs Rs
λO (u)du O (v)dv
R
= λO (u)e ku− du − λI (u)e− u du + e− u λ dku
Rs Rs
− λO (v)dv − λO (v)dv
λI (u) − ku− λO (u) du

= e u dku − e u
Rs
λO (v)dv
= e− u dMuk
Appendix 4.
Proposition. For all u > t,
I
EQ λu |It = 1{t≤L} 1{T >t} µZ f ∗ (t, u)
S
+ 1{T ≤L} 1{T ≤t} Z f (u − T )
R L∧u Rv
where f ∗ (t, u) = t f (u − v)λ(v)e− t λ(w)dw
dv.
Proof. We have
I I
E Q λSu |It = E Q 1{T ≤L} 1{T ≤u−} Z f (u − T ) |It

I
= 1{T >t} E Q 1{T ≤L} 1{T ≤u−} Z f (u − T ) |{T > t}

I
+ 1{T ≤t} E Q 1{T ≤L} 1{T ≤u−} Z f (u − T ) |T ∪ Z

Z L∧u Rv
= 1{t≤L} 1{T >t} µZ
f (u − v)λ(v)e− t λ(w)dw dv
t
+ 1{T ≤L} 1{T ≤t} Z f (u − T )

Appendix 5.
I
Proposition. The martingale representation of E Q λSt |Is is given by

Z
I
dE Q λSt |Is = 1{s≤L} z f (t − s) − µZ f ∗ (s, t) q(ds, dz)

R+
I
Proof. Thanks to Appendix 4, we know E Q λSt |Is is given by:

Z L∧t Ru
QI
f (t − u)λ(u)e− s λ(v)dv du
S Z
E λt |Is = 1{s≤L} 1{T >s} µ
s
+ 1{T ≤L} 1{T ≤s} Z f (t − T )
I
Using Itô’s formula and the Leibniz rule, we can find the martingale representation of E Q λSt |Is :

Z L∧t Ru

I
dE Q λSt |Is = 1{s≤L} 1{T >s−} λ(s)µZ f (t − u)λ(u)e− s λ(v)dv du − f (t − s) ds

s
Z L∧t
− su λ(v)dv
R
Z
− 1{s≤L} µ f (t − u)λ(u)e du dHs + 1{s≤L} Z f (t − s)dHs
s
Z Z L∧t
− su λ(v)dv
R
Z
= 1{s≤L} z f (t − s) − µ f (t − u)λ(u)e du q(ds, dz)
R+ s

Appendix 6.
Proposition. We have for u ≥ t,
I
E Q λSu |It = 1{T >t} 1{t≤L} µZ f ∗ (t, u)

+ 1{T ≤L} 1{T ≤t} Z F̄ (t − T ) fT (t, u)

where
0
fT (t, u) = eA0 (t−T ) A0 (u − T )e−A0 (u−T )
Z L∧u Rv
∗
f (t, u) = fv (v, u)λ(v)e− t λ(w)dw dv
t
268 Appendix 7.
Proof. We have:
h 0
i
QI
S
E λu |It = E 1{T ≤L} 1{T ≤u−} Z F̄ (u − T )− A0 (u − T ) |It
I
h 0
i
= 1{T >t} E Q 1{T ≤L} 1{T ≤u−} Z F̄ (u − T )− A0 (u − T ) |{T > t}
I
h 0
i
+ 1{T ≤t} E Q 1{T ≤L} 1{T ≤u−} Z F̄ (u − T )− A0 (u − T ) |T ∪ Z ∪ F (t − T )
Z L∧u
Z QI
0 I
= 1{T >t} 1{t≤L} µ E F̄ (u − v)− A0 (u − v)Q (T ∈ dv |{T > t} )
t
0 I
+ 1{T ≤L} 1{T ≤t} Z A0 (u − T )E Q

F̄ (u − T )− |T ∪ F (t − T )
I
Since A0 (t) is continuous, we have E Q [1 − F (t)− ] = e−A0 (t) and since T admits a deterministic
intensity λ(t), we have:
Z L∧u Z L∧u Rv
QI
0 0
I
e−A0 (u−v) A0 (u − v)λ(v)e− t λ(w)dw dv

E F̄ (u − v)− A0 (u − v)Q (T ∈ dv |{T > t} ) =
t t
On the other hand, we have

Y Y
1 − F (u − T )− = (1 − ∆As ) (1 − ∆As )
0≤s≤t−T t−T <s<u−T
Y
= [1 − F (t − T )] (1 − ∆As )
t−T <s<u−T
so we get
" #
QI
Y
E F̄ (u − T )− |T ∪ F (t − T ) = F̄ (t − T ) E (1 − ∆As ) |T ∪ F (t − T )
t−T <s<u−T
" #
Y
= F̄ (t − T )E (1 − ∆As )
t−T <s<u−T
−(A0 (u−T )−A0 (t−T ))
= F̄ (t − T )e

Appendix 7.
Proposition. We have:
Z
I
dE Q λSs |Iu = 1{t≤L} z ft (t, u) − µZ f ∗ (t, u) qZ (dt, dz)

R
Z+
− 1{T ≤L} 1{T <t−} Z F̄ (t − T )− a fT (t, u) qA (d(t − T ), da)
[0,1]
Proof. Using the Leibniz rule, we have

∂f ∗ (t, u)
Z L∧u
− tv λ(w)dw
R Rt
= fv (v, u)λ(v)λ(t)e dv − ft (t, u)λ(t)e− t λ(v)dv
∂t t
= λ(t) f ∗ (t, u) − λ(t) ft (t, u)
and
∂fT (t, u) 0
= A0 (t − T ) fT (t, u)
∂t
Using Itô’s lemma on Equation (5.2):
I
dE Q λSu |It = 1{T >t−} 1{t≤L} µZ λ(t) f ∗ (t, u) dt

− 1{t≤L} µZ f ∗ (t, u) d1{T ≤t}

− 1{T >t−} 1{t≤L} µZ λ(t) ft (t, u)dt
+ 1{T ≤L} Z F̄ (t − T )− fT (t, u) d1{T ≤t}
0
+ 1{T ≤L} 1{T ≤t−} Z F̄ (t − T )− A0 (t − T ) fT (t, u) dt
+ 1{T ≤L} 1{T ≤t−} Z fT (t, u) dF̄ (t − T )
If we notice that
Z
∗
Z
1{t≤L} µ f (t, u) d1{T ≤t} = 1{t≤L} µ Z
f ∗ (t, u) N (dt, dz)
R+
Z
1{T ≤L} Z F̄ (t − T )− fT (t, u) d1{T ≤t} = 1{t≤L} z ft (t, u) N (dt, dz)
R+
Z
dF̄ (t − T ) = −F̄ (t − T )− a NA (dt, da)
[0,1]
Z
0
F̄ (t − T )− A0 (t − T )dt = F̄ (t − T )− a pA (d(t − T ), da)
[0,1]
We eventually have:
Z
QI
z ft (t, u) − µZ f ∗ (t, u) qZ (dt, dz)
S
dE λu |It = 1{t≤L}
R
Z+
− 1{T ≤L} 1{T <t−} Z F̄ (t − T )− a fT (t, u) qA (d(t − T ), da)
[0,1]

Conclusion.
This conclusion is the place to discuss the possible extensions of the results described
throughout this thesis and gives some ideas of futur researches.
In Part I, we studied the pricing of a guaranteed investment. The pricing was here consid-
ered in terms of a guaranteed rate on the premium and a participation rate on the terminal
financial surplus. We gave a two-steps procedure to fix these two technical parameters, consis-
tently with the risk management policy of an insurance company. However, this procedure is
not anymore valid if we consider periodic bonuses since, in this case, the level of bonus offered
to the policyholders would influence the insolvency risk of the insurer. It would be interesting
to study how our procedure could be extended in this case.
In Part II, we studied the valuation and the (local) “risk-minimizing” strategies of insurance
contracts with a surrender option. However, for the sake of simplicity, we assumed there was no
mortality. A natural extension would be therefore to introduce this mortality. In such a case, a
general model should allow the surrender probabilities to depend on the survival probabilities
of the policyholder. If we assume the survival probabilities are deterministic, we can expect to
get results not so different of those described in Part II. However, if we assume the survival
probabilities are stochastic (as in Chapter 5), things get a little bit more complicated because
we could have to deal with a non-trivial mix of financial and mortality risk. This question is
probably worth studying.
In Part III, we studied the systematic mortality risk. More precisely, in Chapter 5, we
investigated the form of the risk-minimizing strategy of a single life insurance policy with
systematic mortality risk. We can easily extend this result to a portfolio of policies if we
assume the time of death of the different policyholders are J-independent. Indeed, in this case,
the risk-minimizing strategy of the portfolio is given by the sum of the individual strategies of
the different policies. Finding the risk-minimizing strategy of a portfolio when this conditional
independence does not hold, is another interesting further research.
In Chapter 6, we studied the application of the HJM methodology to longevity bonds. In
particular, in Section 3.6, we gave a pricing formula for these longevity bonds that suggested
that the classical pricing formula found in the intensity-based models could be extended to
much more general settings. It would be interesting to study how to bridge this gap.
In Part IV, we studied the risk-minimizing strategies of non-life insurance contracts, when
the inflation and the interest rate risks are taken into account. We first gave the form of the
risk-minimizing strategy for a general increasing claim process, assuming the financial market
and the claim process (expressed in real terms) are independent. Then, we illustrated this
result in four different claim models. We could extend these results in at least two ways. First,
271
272
we could try to remove the independence between the financial market and the claims. Second,
we could look for more realistic claim models than the four studied in Chapter 7.
Bibliography
[1] Aase, K., (1988) Contingent claim valuation when the security price is a combination of an Itô process and
a random point process, Stochastic Processes and Their Applications 28, 185-220.
[2] Aase, K., Persson, S.A., (1994) Pricing of unit-linked life insurance policies, Scandinavian Actuarial Journal
1, 26-52.
[3] Acerbi, C., Nordio, C., Sirtori, C., (2001) Expected Shortfall as a Tool for Financial Risk Management.
Working paper http://www.gloriamundi.org/
[4] Acerbi, C., Tasche, D., (2001) Expected shortfall: a natural coherent alternative to Value at Risk. Working
paper http://www.gloriamundi.org/
[5] Ahlgrim, K.C., D’Arcy, S.P., Gorvett, R.W., (2004) The Effective Duration and Convexity of Liabilities
for property-Liability Insurers Under Stochastic Interest Rates. The Geneva Papers on Risk and Insurance
Theory, 29, 75-108.
[6] Albizzati, M.-O., Geman, H., (1994) Interest rate risk management and valuation of the surrender option
in life insurance policies. The Journal of Risk and Insurance, Vol 61, No 4, 616-637.
[7] Ansel, J.P., Stricker, C., (1992) Lois de martingale, densités et décomposition de Föllmer-Schweizer. Annales
de l’Institut Henri Poincaré 28, No 3, 375-392.
[8] Ansel, J.P., Stricker, C., (1993) Unicité et existence de la loi minimale. Séminaire de Probabilités XXVII,
Lectures Notes in Mathematics 1557, Springer, 22-29.
[9] Arjas, E., (1989) The claims reserving problem in non-life insurance: some structural ideas. Astin Bulletin,
19 (2), 139-152.
[10] Artzner, P., Delbaen, F., Eber, J.-M., Heath, D., (1999) Coherent measure of risk. Mathematical Finance
9, 203-228.
[11] Artzner, P., Delbaen, F., Eber, J.-M., Heath, D., (1997) Thinking coherently. RISK 10 (11).
[12] Bacinello, A.R., (2001) Fair pricing of life insurance participating policies with a minimum interest rate
guaranteed. Astin Bulletin 31 (2),275-298.
[13] Bacinello, A.R., (2003) Pricing guaranteed life insurance participating policies with annual premiums and
surrender option. North American Actuarial Journal 7 (3), 1-17.
[14] Bacinello, A.R., (2003) Fair valuation of a Guaranteed life insurance participating contract embedding a
surrender option. The Journal of Risk and Insurance, Vol 70, No 3, 461-487.
[15] Bacinello, A.R., Pricing guaranteed life insurance participating policies with periodical premiums and
surrender options. Working Paper, Università degli Studi di Triest.
[16] Bacinello, A.R., Modelling the surrender conditions in equity-linked life insurance. Working Paper, Uni-
versità degli Studi di Triest.
[17] Bacinello, A.R., (2005) Introduction to the pricing of equity-linked life insurance policies with embedded
options. Workshop on Life Insurance Valuation. Lyon, april 27th.
[18] Bauer, D., (2006) An arbitrage-free family of longevity bonds. Working Paper, Ulm University.
[19] Barbarin, J., (2005) Risk neutral valuation of life insurance contract with surrender option. An alternative
approach with asymmetry of information. Working paper, Université Catholique de Louvain.
[20] Barbarin, J., (2007) A note on the literature on the Heath-Jarrow-Morton methodology for longevity bonds.
Working Paper, Université Catholique de Louvain.
273
274
[21] Biagini, F., Cretarola, A., (2006) Local risk-minimization for defaultable markets. Preprint, LMU University
of München and University of Bologna.
[22] Biagini, F., Cretarola, A., (2006) Local risk-minimization for defaultable claims with recovery process.
Preprint, LMU University of München and University of Bologna.
[23] Bielecki, T.R., Rutkowski, M., (2002) Credit Risk: Modeling, Valuation and Hedging. Springer.
[24] Bielecki, T.R, Jeanblanc, M., (2005) Indifference prices in Indifference Pricing, Theory and Applications.
Financial Engineering, Princeton University, R. Carmona Editor.
[25] Bielecki, T., Jeanblanc, M., Rutkowski, M., (2004) Modelling and Valuation of Credit Risk. Stochastic
methods in finance, 27-126, Lecture Notes in Math., 1856, Springer, Berlin.
[26] Bielecki, T.R., Jeanblanc, M., Rutkowski, M., (2003) Pricing and hedging of credit risk: Replication and
mean-variance approaches I. Contemporary Mathematics 351: Mathematics of Finance, AMS, Providence,
37-53.
[27] Bielecki, T.R., Jeanblanc, M., Rutkowski, M., (2003) Pricing and hedging of credit risk: Replication and
mean-variance approaches II. Contemporary Mathematics 351: Mathematics of Finance, AMS, Providence,
55-64.
[28] Bielecki, T.R., Jeanblanc, M., Rutkowski, M., (2004) Hedging of defaultable claims, In: Paris-Princeton
Lectures on Mathematical Finance 2003. R.A. Carmona et al., eds. Springer-Verlag, 1-132.
[29] Biffis, E., (2005) Affine processes for dynamic mortality and actuarial valuations. Insurance: Mathematics
and Economics, 37, 443-468.
[30] Blanchet-Scalliet, C., Jeanblanc, M., (2004) Hazard rate for credit risk and hedging defaultable contingent
claims. Finance and Stochastics. Vol 8, No 1. January.
[31] Boyle, P.P., Hardy, M.R., (1997) Reserving for maturity guarantees: two approaches. Insurance: Mathe-
matics and Economics, 21, 113-127.
[32] Brémaud, P., Yor, M., (1978) Changes of filtrations and of probability measures. Zeitschrift für Wahrschein-
lichkeitstheorie und Verwandte Gebiete, 45, 269-295.
[33] Brémaud, P., (1981) Point Processes and Queues. Martingale Dynamics. Springer.
[34] Brennan, M.J., Schwartz, E.S., (1976) The pricing of equity linked life insurance policies with an asset
value guarantee. Journal of Financial Economics 3, 195-213.
[35] Browne, S., (1995) Optimal investment policies for a firm with a random risk process: exponential utility
and minimizing the probability of ruin. Mathematics of Operations Research, 20 (4), 937-958.
[36] Cairns, A.J., Blake, D., Dowd, K., (2006) Pricing death: framework for the valuation and the securitization
of mortality risk. Astin Bulletin, 36 (1), 79-120.
[37] Dahl, M., (2004) Stochastic mortality in life insurance: market reserves and mortality-linked insurance
contracts. Insurance: Mathematics and Economics, 35 (1), 113-136.
[38] Dai, Q., Singleton, K. J., (2000) Specification Analysis of Affine Term Structure Models. The Journal of
Finance, 55, 1943-1978.
[39] Dahl, M., Møller, T., (2006) Valuation and hedging of life insurance liabilities with systematic mortality
risk. Insurance: Mathematics and Economics, 39 (2), 193-217.
[40] Delbaen, F., (1986) Equity linked policies. BARAB 80, 33-52.
[41] Delong, L., (2005) Optimal investment strategy for a non-life insurance company: quadratic loss. Applica-
tiones Mathematicae, 32, 263-277.
[42] Duffee, G.R., (2002) Term premia and interest rate forecasts in affine models. The Journal of Finance, 57,
405–443
[43] El Karoui, N., Martellini, L., (2001) Dynamic Asset Pricing Theory With Uncertain Time-Horizon. Working
Paper.
[44] Föllmer, H., Sondermann, D., (1986) Hedging of non-redundant contingent claims. in W. Hildenbrand and
A. Mas-Colell(eds.), “Contributions to mathematical economics”, North-Holland, 205-223.
[45] Grandell, J., (1976) Doubly stochastic Poisson processes. Lecture Notes in Mathematics, vol. 529, Springer-
Verlag.
Bibliography 275
[46] Gravereaux, J.B., Pellaumail, J., (1974) Formule de Itô pour des processus non continus à valeurs dans des
espaces de Banach. Annales de l’institut Henri Poincaré, Section B, Tome 10 (4), 339-422.
[47] Grosen, A. , Jørgensen, P.L., (2000) Fair valuation of life insurance liabilities: the impact of interest rate
guarantees, surrender options and bonus policies. Insurance: Mathematics and Economics 26 (1), 37-57.
[48] Hardy, M., (2003) Investment Guarantees. Wiley Finance.
[49] Hainaut, D., Devolder, P., (2007) Mortality modelling with Lévy processes. To appear in Insurance: Math-
ematics and Economics.
[50] Heath, D., Jarrow, R., Morton, R., (1992) Bond pricing and the term structure of interest rates: a new
methodology for contingent claim valuation. Econometrica, 60, 77-105.
[51] Hipp, C., (2004) Stochastic control with application in insurance. Stochastic Methods in Finance: Lectures
given at the C.I.M.E.-E.M.S. Summer School held in Bressanone/Brixen, Springer.
[52] Hipp, C., Plum, M., (2000) Optimal investment for insurers. Insurance: Mathematics and Economics, 27,
215-228.
[53] Hipp, C., Plum, M., (2003). Optimal investment for investors with state dependent income, and for insurers.
Finance and Stochastics, 7, 299-321.
[54] Hjort, N. L., (1990) Nonparametric bayes estimators based on beta processes in models for life history
data. The Annals of Statistics, 18 (3), 1259-1294.
[55] Hughston, L.P., (1998) Inflation Derivatives. Working Paper, Department of Mathematics King’s College
London.
[56] Jacobsen, M., (2006) Point Process Theory and Applications. Marked Point and Piecewise Deterministic
Processes. Birkhauser.
[57] Jarrow, R., Yildirim, Y., (2003) Pricing treasury inflation protected securities and related derivatives using
an HJM model. Journal of financial and quantitative analysis, 38 (2), 409-430.
[58] Jeanblanc, M., Rutkowski, M., (2000) Modelling of default risk: mathematical tools. Working Paper.
[59] Kaufmann, R., Gadmer, A., Klett, R., (2001) Introduction to Dynamic Financial Analysis, Astin Bulletin
31(1), 213-249.
[60] Korn, R., (2005) Worst-case scenario investment for insurers. Insurance: Mathematics and Economics, 36,
1-11.
[61] Last, G., Brandt, A., (1995) Marked Point Processes on the Real line. The Dynamic approach. Springer.
[62] Linnemann, P., Valuation of participating life insurance liabilities. Scandinavian Actuarial Journal 2 (2004),
p. 81-104.
[63] Lotz, C., (1998) Locally minimizing the credit risk, Working paper, University of Bonn.
[64] Luciano, E., Vigna, E., (2005) Non mean reverting affine processes for stochastic mortality. ICER working
paper.
[65] Mania, M., Schweizer, M., (2005) Dynamic exponential utility indifference valuation. The Annals of Applied
Probability, Vol 15, No. 3, 2113-2143.
[66] Musiela, M., Rutkowski, M., (1998) Martingale Methods in Financial Modelling, second edition, Springer.
[67] Miltersen, K.R., Persson, S.A., (2005) Is mortality dead? Stochastic force of mortality determined by
arbitrage? Working Paper, University of Bergen.
[68] Møller, T., (2001) Risk-minimizing hedging strategies for insurance payment processes. Finance and
Stochastics, 5 (4), 419-446.
[69] Møller, T., Risk-minimizing hedging strategies for unit-linked life insurance contracts. Astin Bulletin 28,
17-47.
[70] Nielsen, J.A., Sandmann, K., (1995) Equity-linked life insurance: a model with stochastic interest rates.
Insurance: Mathematics and economics 16, 225-253.
[71] Protter, P.E., (2005) Stochastic Integration and differential equation. Springer, Second Edition, Version 2.1
[72] Riesner, M., (2006) Hedging life insurance contracts in a Lévy process financial market. Insurance: Math-
ematics and Economics, 38, 599-608.
[73] Simon, S., Van Wouwe, M., (2005) Life insurance liabilities: analysis and valuation of embedded financial
guarantees. Working paper.
276
[74] Schrager, D.F., (2006) Affine stochastic mortality. Insurance: Mathematics and Economics, 38, 81-97.
[75] Schweizer, M., (1990) Risk-Minimality and Orthogonality of Martingales. Stochastics and Stochastics Re-
ports 30, 123-131.
[76] Schweizer, M., (1991) Option Hedging for semimartingales. Stochastic processes and their applications 37,
339-363.
[77] Schweizer, M., (1995) On the minimal martingale measure and the Föllmer-Schweizer decomposition. Sto-
chastic Analysis and Applications 13, 573-599.
[78] Schweizer, M., A guided tour through quadratic hedging approaches, in E. Jouini, J Cvitanic and M.
Musiela, option pricing, interest rates and risk management, Cambridge University Press, 538-574.
[79] Schweizer, M., (2008) Local risk-minimization for multidimensional assets and payment streams. To appear
in Banach Center Publications.
[80] Tak Kuen Siu, (2005) Fair valuation of participating policies with surrender options and regime switching.
Insurance: Mathematics and Economics, 37, 533-552.
[81] Weixi Shen, Huiping Xu, (2005) The valuation of unit linked policies with or without surrender options.
Insurance: Mathematics and Economics, 36, 79-92.
[82] Walker, S., Muliere, P., (1997) Beta-Stacy processes and a generalization of Polya-urn scheme. The Annals
of Statistics, 25 (4), 1762-1780.
[83] Wirch, J.L., Hardy, M.R., (1999) A synthesis of risk measures for capital adequacy. Insurance: Mathematics
and Economics, 25 (3), 337-347.

Valuation, Hedging and The Risk Management of Insurance Contracts (Doctoral Thesis) - by Jérôme Barbarin, Université Catholique de Louvain, 2008

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Valuation, Hedging and The Risk Management of Insurance Contracts (Doctoral Thesis) - by Jérôme Barbarin, Université Catholique de Louvain, 2008

Uploaded by

Copyright:

Available Formats

Université Catholique de Louvain

Institut des Sciences Actuarielles

Valuation, Hedging and the Risk Management of Insurance Contracts.

Membres du jury: Thèse présentée en vue de

"On se lasse de tout sauf de comprendre",

Part 1. Life Insurance Contracts and the Interest Rate Risk. 7

Chapter 1. Risk Measure and Fair Valuation of an Investment Guarantee in Life

Part 2. Life Insurance Contracts and the Surrender Risk. 41

Chapter 2. Risk-Neutral Valuation of Life Insurance Contracts with a Surrender Option. 43

Chapter 3. Risk-Minimizing Strategies of Life Insurance Contracts with Surrender

Chapter 4. Locally Risk-Minimizing Strategies of Life Insurance Contracts with Surrender

Part 3. Life Insurance Contracts and the Longevity Risk. 161

Chapter 5. Risk-Minimization with Systematic mortality risk. 163

Chapter 6. Heath-Jarrow-Morton Modelling of Longevity Bonds. 193

Part 4. Non-Life Insurance Contracts and the Inflation Risk. 235

Contribution of the dissertation.

Chapter 3: Barbarin J. “Risk-Minimizing Strategies of Life Insurance Contracts with Sur-

Chapter 4: Barbarin J. “Local Risk-Minimization for Payment Processes. Application to In-

Chapter 5: Barbarin J. “Risk-Minimizing Strategies of Life Insurance Contracts with Sys-

Chapter 6 is an extended version of: Barbarin J. “Heath-Jarrow-Morton Modelling of Longevity

Life Insurance Contracts and the Interest

Risk Measure and Fair Valuation of an Investment Guarantee in

2. Cash-Bond-Stock Model (CBS).

with initial value β(0) = 1.

where W2 (t) is a standard Brownian motion, independent of W1 (t).The correlation

Or equivalently, we can rewrite this formula as:

dV (t) dS(t) dP (t, Ts + ts ) dβ(t)

Taking the logarithm, we get:

In order to determine the distribution of V (T ), we start by determining the distribution of

dV (t) dS(t) dP (t, Ts + ts ) dβ(t)

ln V (ts+1 ) − ln V (ts ) = d ln V (u)

= xa µ − xb λσr B(Ts + ts − u)du + (xb + xc ) r(u)du

E [ln V (ts+1 ) − ln V (ts ) |F0 ] = K(Ts , ts+1 − ts )

K(Ts , ts+1 − ts ) = xa µ − xb λσr B(Ts + ts − u)du

IB1 (Ts , ts ) = B(Ts + ts − u)du

IB2 (Ts , ts ) = B(Ts + ts − u)2 du

3. Investment Guarantee and Risk Measures.

3.1. Investment Guarantee. As already said, we assume a single premium is paid at

V aR = P (0, T ) L(T ) − exp µP (T ) + σP (T )φ−1 ()

4. Fair Valuation and Participation Rate.

(4.1) F V (T ) = L(T ) + η max(V (T ) − L(T ); 0)

(4.2) F V (0) = (1 + i)T P (0, T ) + ηP (0, T )EQT ((V (T ) − (1 + i)T )+ )

where the new Brownian motions are defined by:

dW 1 (t) = dW1 (t) + λdt

Finally the risk free rate follows a modified Ornstein-Uhlenbeck process:

The T -forward martingale measure is then generated by the change:

dW 1 (t) = dW 1 (t) + σr B(T − t)dt

In this model, we have then to price a call option:

C(0, K) = P (0, T )EQT ((V (T ) − K)+ )

ln(1/K) − ln P (0, T ) + v 2 (T )/2

(4.4) C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

Table 5.1. Interest rate

Table 5.2. Stock model

Figure 5.1. Probability  = 1%

Figure 5.2. Guaranteed rate = 3.25%

Figure 5.3. Probability of Ruin = 1%

Figure 5.4. Probability of Ruin = 1%

Figure 5.5. Guaranteed rate = 3.25 %

Figure 5.6. Conditional Value at Risk = 4%

Figure 5.7. Conditional Value at Risk = 4%

6. Conclusions and Extensions.

tZs+1 tZs+1 tZs+1Zu

V aR = P (0, T ) L(T ) − exp µP (T ) + σP (T )φ−1 ()

(4.4) C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))

Figure 5.1. Probability = 1%

C(0, (1 + i )T ) = Φ(D+ (i )) − (1 + i )T P (0, T )Φ(D− (i ))