Vous êtes sur la page 1sur 51

Optimal control of risk exposure, reinsurance and

investments for insurance portfolios

Christian Irgens
Norwegian Hull Club
Postboks 75 Sentrum
5803 Bergen
NORWAY
email: christian.irgens@norclub.no

and
Jostein Paulsen
Department of Mathematics
University of Bergen
Johs. Brunsgt. 12
5008 Bergen
NORWAY
email: jostein@mi.uib.no

Abstract
We consider an insurance company whose insurance business follows a
diffusion perturbated classical risk process. Assets can be invested in a risk
free asset or in a risky one, the return of the latter follows an independent
diffusion perturbated classical risk process (in finance it is called a jump-
diffusion process). The company can dynamically control the proportion
of the insurance business that is ceded to reinsurers, and in addition it can
dynamically cover parts of its remaining business using excess of loss reinsur-
ance. It can also dynamically control the proportion of the assets it invests in
the risky asset. We seek to find the controls that maximize expected utility
of assets at a terminal time. We do so for three different utility functions.
Let T be a fixed time and Ty be the time of ruin, i.e. the first time assets
become nonpositive. Then the utility functions and terminal times are:
1. U (y) = (max{0, y})c , 0 < c < 1 and terminal time is T ∧ Ty .
2. U (y) = ln y when y > 0 and U (y) = −∞ when y ≤ 0. Terminal time
is T .
3. U (y) = −e−cy , c > 0, and terminal time is T .

1
It turns out that in spite of the complexity of the problem, provided there
are no complicating constraints on the controls, the problems have rather
simple solutions. Our vehicle is the Hamilton-Jacobi-Bellmann equations,
and we seek to verify that the suggested solutions are in fact optimal among
a fairly large easily verifiable admissible class of controls. We also discuss
quantitative and qualitative aspects of the solutions, and show how the re-
sults can be extended to several lines of insurance business as well as several
risky investments.

Keywords: Optimal control; diffusion perturbated risk process; Hamilton-Jacobi-


Bellmann equation: proportional reinsurance: excess of loss reinsurance; invest-
ment strategy.

2
1 Introduction
In recent years there has been a large increase in the utilization of dynamic
control theory to insurance related problems, thus continuing the classical works
by e.g. Gerber (1969), Bühlmann (1970), Dayanada (1970) and Martin-Löf (1973,
1983, 1994). Jeanblanc-Picqué and Shiryaev (1995) and Asmussen and Taksar
(1997) considered maximizing expected value of discounted dividends paid until
time of ruin for a Brownian motion with drift. Paulsen and Gjessing (1997) also
allowed for stochastic returns on investments in this model, and in Paulsen (2003)
the same problem is studied when dividends paid are subject to solvency con-
straints. Højgaard and Taksar (1999) extended the work of Asmussen and Taksar
(1997) to include control of the proportion of the insurance business that is ceded
to reinsurers. Then Højgaard and Taksar (2001) solved this problem for the more
complicated model used by Paulsen and Gjessing (1997). Finally, for this same
model, Højgaard and Taksar (2000) included control of investment strategies as
well.
A somewhat different approach was taken in Højgaard and Taksar (1997). They
seeked to choose the optimal proportional reinsurance in order to maximize ex-
pected discounted average value of assets up to time of ruin. Their model was
again a Brownian motion with drift.
All these recent papers work directly with diffusion processes, the argument is
that these processes can serve as approximations. In Asmussen, Højgaard and
Taksar (2000) this approximation aspect was part of the model formulation. As
in Højgaard and Taksar (1999) they seeked to find an optimal dividend payment
scheme for a Brownian motion with drift when reinsurance is allowed, but here it
was excess of loss (XL) reinsurance. Using weak convergence arguments their model
again boiled down to a diffusion model, but with drift and diffusion coefficients
depending in a nonlinear way on the XL limit.
Assuming the classical risk process instead of a diffusion, Højgaard (2001) let
the premium rate be dependent on the size of the business.
Another very popular optimization criterium, in particular among German speak-
ing researcher, is to control the process so that the probability of ruin is minimized.
Lots of papers have been written on this topic, and among them are Hipp and
Plum (2000, 2003), Hipp and Vogt (2003), Hipp and Schimidli (2003), Gaier and
Grandits (2002, 2003), Gaier, Grandits and Schachermayer (2003) and Schmidli
(2001, 2002). In these papers the risk process is not assumed to be a diffusion, and
some of them also allow for the possibility of investments.
A survey of control theory in insurance with main emphasis on minimization of
the ruin probability can be found in Hipp (2003).

In this paper we consider a classical risk process perturbated by a diffusion.


Furthermore the company is allowed to invest money in either a risk free asset or
a risky one. It can also buy proportional reinsurance on its insurance business,
and then cover parts of the remaining risk using XL reinsurance. The proportion
of retained business before XL as well as the excess point are constrained to be
nonnegative, but otherwise there will generally be no restrictions on the controls.

3
However, constraining the proportion held in the risky asset to be nonnegative
would not change the results noticeably.
Our aim is to choose the control parameters in order to maximize expected
utility of total assets at either a fixed terminal time T or at T ∧ Ty where Ty is
the time of ruin when initial capital is y. With our choices of utility functions it
turns out that provided no complicating restrictions on the controls are imposed,
the problems have fairly simple solutions. We will invest a fair amount of energy
to show that the candidates for optimal control obtained by solving the Hamilton-
Jacobi-Bellmann equations, really are optimal among a class of easily identifyable
admissible controls. Although this approach leaves the class of admissible controls
somewhat smaller than otherwise obtained by using conditions as in e.g. Theorem
III.8.1 in Fleming and Soner (1993), it does not suffer from the disadvantage of not
knowing which controls are admissible. A fairly large amount of energy is invested
to analyze qualitative and quantitative properties of the solutions.
Here is a brief outline of the paper. In Section 2 the model assumptions are
formulated, and a brief discussion of optimal control in our context is presented.
In Section 3 we study the problem using the utility function U(y) = (max{0, y})c,
0 < c < 1, and time horizon T ∧ Ty . It turns out that this problem has a fairly nice
solution. However, there are two competing candidates, and a substantial effort is
made to analyze and compare these. Lots can be said about them, but in the end
the optimal one have to be found by calculating both numerically and take the one
that gives highest utility.
In Section 4 a logaritmic utility function is used, and the problem is here to
maximize expected utility of total assets at a fixed terminal time T . Again the
problem has a nice solution, and it is shown that this solution is a limiting case as
c → 0 of the solution in Section 3. We also show, as in e.g. Pestien and Sudderth
(1985), that this problem is related to the problem of minimizing the expected time
to reach a given capital.
In Section 5 the utility function is exponential, and again the problem is to
maximize expected utility of total assets at a fixed terminal time T . This is a
generalization of a problem studied by Browne (1995), and it turns out that in spite
of our more complex model, the results are basically the same as those obtained
by him. Unfortunately, for this utility function we are only able to show that our
proposed control is optimal among a more restricted class of controls.
In Section 6 the models in Sections 3 to 5 are discussed when there are constraints
on the controls allowed, constraints on reinsurance as well as on investments, such
as not ceding more than 100% of the business to reinsurers or not going short
on risky investments. Another constraint discussed is to put an upper bound on
tolerated finite time ruin probabilities. It turns out that some constraints are easily
taken care of, while others turn what was originally a fairly simple problem into a
very difficult one.
Section 7 describes how results from Sections 3 to 5 can be generalized to the
situation with several lines of business as well as several risky investment opportu-
nities. Finally, Section 8 gives some concluding remarks and briefly compares the
different solutions.

4
2 Model formulation
We will assume that all processes and random variables are defined on a fil-
tered probability space (Ω, F , F, P ) satisfying the usual conditions, i.e. Ft is right-
continuous and P -complete.
It will be assumed throughout the paper that the basic insurance process P is
a diffusion perturbated classical risk process, i.e.
NP,t
X
Pt = pt + σP WP,t − SP,i
i=1

where WP is a Brownian motion independent of the compound Poisson process


PNP,·
i=1 SP,i where NP is a Poisson process with intensity λP independent of the
i.i.d. {SP,i} with distribution function FP . We assume that FP (0) = 0, and write
µP = E[SP ] which is assumed to be finite. Our interpretation of the model is that
the first term represents expected premium income, the second term small claims
and possibly variations in premium income, while the compound Poisson process
represents large claims.
The interpretation that the term σP WP in addition to variations in premium
income also represents small claims is necessary for the model to make sense. Since
in practice insurance premiums can vary hugely over a few years, σP has to be rather
large to capture such variations, but then the term pt + σP WP,t will fluctuate a lot
and may also become negative, which is hardly realisticR for premiums. A better
model would be to replace the term pt + σP WP,t with 0t ps ds where p is a mean
reverting and positive diffusion such as the CIR model in finance, but that would
complicate the calculations enormously.
We will allow the company to continuously reinsure a fraction of its business,
and then (also continuously) buy an excess of loss (XL) reinsurance to cover large
losses on its remaining risk. The insurance process then becomes
Z t Z t
Pta,D = (as p + (1 − as )qP p − qXL λP ρas (Ds ))ds + σP as dWP,s
0 0
NP,t
X
− min{aTP,i SP,i, DTP,i }.
i=1

Here TP,i is the time of the i’th jump of NP . Furthermore at is the proportion
of the retained business at time t after proportional reinsurance, and the process
a is assumed to be nonnegative and predictable. The constant qP is a commis-
sion percentage received (can be negative) for ceded reinsurance, so that the total
commission received will vary with p only and not with WP . Typically reinsurers
will pay a fee, so that qP ≥ 0. If the company at study is a reinsurer figuring out
how large a share to take, this is dealt with by setting qP = 0. Dt is the excess
point, i.e. the companys own deductible at time t. The process D is assumed
to be nonnegative and predictable. The price intensity for the XL reinsurance is
assumed to be of the form qXL λP ρat (Dt ) = qXL λP E[(at SP − Dt )+ |Ft− ], where qXL
is a loading factor, typically qXL > 1. Here α+ = max{0, α} so that ρat (Dt ) is

5
expected amount received from the XL reinsurer if a claim occurs at time t. For
constant a and D we write
def ∞ ∞ x
Z Z
ρa (D) = E[(aSP − D)+ ] = a F̄P (x)dx = F̄P ( )dx, (2.1)
D/a D a

where F̄P (x) = 1 − FP (x). The second equality was obtained using integration by
parts.

In addition to reinsurance, the company is allowed to invest parts of its surplus


in a risk free asset and the rest in a risky asset, a market fund say. Assets invested
in the risk free asset follow dVt = Vt dR̃t , while if invested in the risky asset they
follow dVt = Vt− dRt where
NR,t
X
R̃t = r̃t > 0 and Rt = rt + σR WR,t + SR,i .
i=1

The processes P and R are assumed independent and WR is a Brownian motion


PNR,·
independent of the compound Poisson process i=1 SR,i . We let λR be the intensity
of NR , and FR the distribution function of SR with µR = E[SR ] assumed finite.
If σR2 + λR = 0 there is no investment risk, and to avoid arbitrage r = r̃. When
σR2 + λR > 0 we will assume the rather natural

r̃ < r + λR µR , (2.2)

i.e. expected return from risky investments is higher than for riskless ones. In
order for (2.2) to make sense in an arbitrage free market, it is also necessary that

σR2 + λR FR (0−) > 0 or that r < r̃. (2.3)

It will be assumed throughout that σR2 + λR > 0 and that (2.3) holds.
Now let βt be the proportion of total assets held in the risky asset at time t,
where also β is required to be predictable. If initial assets of the company is y,
then using the control strategy π = (a, D, β), total assets at time t equals
Z t
Ytπ =y+ Pta,D + π
Ys− (βs dRs + (1 − βs )dR̃s ). (2.4)
0

Definition 2.1 We say that a control π is admissible if it is predictible with at and


Dt nonnegative and such that (2.4) has a unique (strong) solution. Furthermore,
for some K > 0 it must satisfy the growth condition
π π
|at | + |βt Yt− | < K(1 + |Yt− |). (2.5)

Let U be a utility function, typically increasing and concave, but see Section
3 below. Also for a control policy π = (a, D, β), let Tyπ be the time of ruin, i.e.
Tyπ = inf{t : Ytπ ≤ 0} with Tyπ = ∞ if Ytπ > 0 for all t. For a fixed time horizon T ,
π
we set Ty,T = Tyπ ∧ T . Then define

J π (t, y) = E t,y [U(YTπy,T


π )] (2.6)
6
or
J π (t, y) = E t,y [U(YTπ )]. (2.7)
It will always be clear which of (2.6) or (2.7) that is considered, therefore the same
notation J π can be used. By e.g. E t,y [U(YTπ )] we mean E[U(YTπ )|Yt = y]. We also
set
J(t, y) = sup J π (t, y),
π

where the supremum is over all admissible controls.


The Hamilton-Jacobi-Bellman (HJB) equation for this problem is then

sup(Vt (t, y) + Aπ V (t, y)) = 0 (2.8)


π

with boundary condition


V (T, y) = U(y), (2.9)
and in addition when (2.6) is considered,

V (t, y) = U(y), y ≤ 0. (2.10)


∂ ∂ ∂2
By Vt is meant ∂t V and similarly Vy = ∂y V and Vyy = ∂y 2
V . The generator Aπ in
(2.8) acts on the y variable and is given as

Aπ φ(y) = (ap + (1 − a)qP p − qXL λP ρa (D) + r̃y + β(r − r̃)y)φy (y)


1
+ (a2 σP2 + β 2 σR2 y 2)φyy (y)
2 (2.11)
+λP (E[φ(y − min{aSP , D})] − φ(y))
+λR (E[φ(y(1 + βSR ))] − φ(y))

for all twice continuously differentiable φ.


Our approach is standard. Find a C 1,2 solution V of (2.8) that satisfies (2.9),
and if applicable, (2.10). Prove that with this solution

V (t, y) ≥ J π (t, y)

for all admissible controls π, and that there exists an admissible control π ∗ so that

V (t, y) = J π (t, y).


Then this π ∗ is optimal, and the value function J(t, y) = V (t, y).

We end this section with a simple result.


Theorem 2.1 If σP2 = 0 and (1 − qP )p > qXL λP µP there is no optimal control and
a
using the control πta = (a, 0, 0), J π (t, y) → limy→∞ U(y) as a → ∞.
Proof With the control π a there is no investment risk and

Pta,D = (qP p + ((1 − qP )p − qXL λP µP )a)t.

Clearly Pta,D → ∞ as a → ∞ and the result follows.


7
3 Power utility with termination at ruin
In this section we shall consider the problem (2.6) using the utility function
U(y) = (max{0, y})c = y+
c

for some c with 0 < c < 1. This utility function is not concave, but since the
π
process is only considered up to time Ty,T , it still makes sense to use it. However,
it does not punish large deficits at ruin, so some bold strategies may be expected.
For y > 0, U(y) has constant relative risk aversion equal to 1 − c, hence an increase
in c is equivalent to a decrease in risk aversion.
It is assumed throughout this section that we have neutral proportional reinsur-
ance, i.e. that qP = 0.
π π π π
Define αt and δt by at = αt Yt− and Dt = δt Yt− . Then πt = (αt Yt− , δt Yt− , βt ) and
π
the process Y has the dynamics
dYtπ = Yt−
π
dZtπ ,
where
Z t
Ztπ = (αs p − qXL λP ραs (δs ) + r̃ + βs (r − r̃))ds
0
Z t Z t
+σP αs dWP,s + σR βs dWR,s
0 0
NP,t NR,t
X X
− min{αTP,i SP,i, δTP,i } + βTR,i SR,i .
i=1 i=1

This is just the exponential equation, and the solution is


t 1
Z
Ytπ = y exp (αs p − qXL λP ραs (δs ) − αs2 σP2
0 2
1 2 2
Z t Z t 
+r̃ + βs (r − r̃) − βs σR )ds + σP αs dWP,s + σR βs dWR,s (3.1)
2 0 0
NP,t NR,t
Y Y
× (1 − min{αTP,i SP,i , δTP,i }) (1 + βTR,i SR,i ).
i=1 i=1

In particular if (α, δ, β) are constants,


1
Ytπ = y exp{(αp − qXL λP ρα (δ) − α2 σP2
2
1 2 2
+r̃ + β(r − r̃) − β σR )t + ασP WP,t + βσR WR,t } (3.2)
2
NP,t NR,t
Y Y
× (1 − min{αSP,i, δ}) (1 + βSR,i ).
i=1 i=1

π
For constant π̃ = (α, δ, β), let TP,y = min{TP,i : SP,i ≥ α1 } if δ ≥ 1 and TP,y
π
=∞
π 1
if δ < 1. Also let TR,y = min{TR,i : SR,i ≤ − β }. Then

Tyπ = TP,y
π π
∧ TR,y . (3.3)
We are now ready for the main result of this section.
8
Theorem 3.1 Let Y π be given by (2.4) with qP = 0, i.e. risk neutral proportional
reinsurance. Let
1
hP (α, δ) = αcp − cqXL λP ρα (δ) − α2 c(1 − c)σP2
2 (3.4)
+λP (E[(1 − min{αSP , δ})c+ ] − 1)
and
1
hR (β) = cr̃ + βc(r − r̃) − β 2 c(1 − c)σR2 + λR (E[(1 + βSR )c+ ] − 1). (3.5)
2
We seek to maximize J π (t, y) = E t,y [U(YTπy,T c
π )] with U(y) = y+ , 0 < c < 1.

a) If σP2 > 0 there is a finite α∗ and a δ ∗ (may be infinite), so that hP (α, δ)


attains its maximum at (α∗ , δ ∗ ). In fact this δ ∗ can either take the value
δ ∗ = ∞ or δ ∗ = δ̂ where
1/(c−1)
δ̂ = (1 − qXL )+ .
If σR2 > 0 there is a finite β ∗ so that hR (β) attains its maximum at β ∗ .
b) If both σP2 > 0 and σR2 > 0 , the control πt∗ = (α∗ Yt−
π
, δ ∗ Yt−
π
, β ∗ ) is optimal
∗ ∗

π∗
among all admissible controls, and J (t, y) = J(t, y) = VP (t, y), where
VP (t, y) = eηP (T −t) y+
c

with
ηP∗ = hP (α∗ , δ ∗ ) + hR (β ∗ ).

c) If either σP2 = 0 or σR2 = 0 then there is no optimal control and the value
function J(t, y) = ∞ for t < T, y > 0.

Proof. Only part a will be proved here. The rest is proved in the appendix. Start-
ing with hP , it is clear that it is continuous, and since σP2 > 0, limα→∞ hP (α, δ) =
−∞, hence α∗ exists and is finite.
Now by (2.1) for any δ so that F̄P is continuous at δ/α,
∂ δ
ρα (δ) = −F̄P ( ). (3.6)
∂δ α
If F̄P has a jump at δ/α, the corresponding left and right derivatives will apply.
Furthermore, using integration by parts
Z δ/α
E[(1 − min{αSP , δ})c+ ] = 1 − αc (1 − αx)c−1 F̄P (x)dx. (3.7)
0

Hence for any δ,


∂ δ
E[(1 − min{αSP , δ})c+ ] = −c(1 − δ)c−1
+ F̄P ( ) (3.8)
∂δ α
using left and right derivatives if F̄P has a jump at δ/α. From (3.6) and (3.8) we
get
∂ δ
hP (α, δ) = cλP F̄P ( )(qXL − (1 − δ)c−1
+ )
∂δ α
9
which is zero for qXL = (1 − δ)c−1
+ . Furthermore

∂− ∂ δ

hP (α, 1) = −∞ and hP (α, δ) = cqXL λP F̄P ( ) ≥ 0 for δ > 1.
∂δ ∂δ α

Also ∂δ hP (α, δ) is decreasing for δ < 1, hence for any α ≥ 0 there are two possibil-
ities for δ ∗ , namely δ ∗ = ∞ or δ ∗ = δ̂. The proof for hR is trivial.

In particular when qXL ≤ 1, i.e. cheap reinsurance, it is optimal to have either


no XL reinsurance or full XL reinsurance. When qXL > 1, δ̂ is increasing in c with
−1
δ̂ = 1 − qXL when c = 0 and δ̂ = 1 when c = 1. Furthermore, and rather obviously,
δ̂ is increasing in qXL with δ̂ = 0 when qXL ≤ 1 and δ̂ → 1 as qXL → ∞.

We will now have a closer look at hP (α, ∞) and hP (α, δ̂) together with their

maximizers α∞ and αδ̂∗ . To this end it will be assumed that σP2 > 0 in addition
to the standing assumption qP = 0. In this and in later sections FP is frequently
assumed to be of the form
Z x ∞
X
FP (x) = fP (y)dy + pP,i 1[xP,i,∞) (x), (3.9)
0 i=1

where the {xP,i } are separated, i.e. for all i there exists an εi > 0 so that xP,j ∈
/
(xP,i − εi , xP,i + εi ) for all j 6= i.

Consider first the case δ = ∞. Using integration by parts and a change of


variables
1 Z 1/α
hP (α, ∞) = αcp − α2 c(1 − c)σP2 − λP cα (1 − αx)c−1 F̄P (x)dx. (3.10)
2 0

If FP is of the form (3.9),


1
hP (α, ∞) = αcp − α2 c(1 − c)σP2

2 
Z 1/α
(1 − αx)c fP (x)dx + (1 − αxP,i )c pP,i − 1.
X
+λP 
0 αxP,i ≤1

This gives
∂− 1

hP (α, ∞) = −∞ when α = , (3.11)
∂α xP,i
hence hP (α, ∞) does not attain a maximum for α = 1/xP,i . Also
∂ 1
hP (α, ∞) = cp − αc(1 − c)σP2 − cλP E[SP (1 − αSP )c−1
+ ], α 6= , (3.12)
∂α xP,i
provided this expression is finite, and if it is infinite it will have the same sign at
both sides of that point. Furthermore,

hP (0, ∞) = c(p − λP µP )
∂α
10
∗ ∂ ∗
and if p > λP µP the optimal α∞ > 0 will satisfy h (α∞
∂α P
, ∞) = 0. Otherwise

α∞ = 0.
1
Remark 3.1 The result (3.11) has a practical explanation. If α = xP,i
, a jump
PNP,· 1
of size xP,i in i=1 SP,i will cause ruin (with YTπyπ = 0), while if α − ε for = xP,i
ε > 0, such a jump will not cause ruin, and the process can continue. On the
1
other hand, if α = xP,i + ε, this jump will again cause ruin (with YTπyπ < 0), but
since U(y) = 0 for all y ≤ 0, the consequences of this jump is no larger than when
1 1
α = xP,i . However, with α = xP,i + ε, if ruin does not occur, E[U(YTπy,T
π )] is higher

1 1
than with α = xP,i . Consequently, there is no reason to choose α = xP,i .

Now let δ = δ̂ and assume first that qXL ≤ 1, i.e. cheap reinsurance. Then δ̂ = 0
and it is easy to verify that hP (α, 0) is maximum for
(p − qXL λP µP )+
α0∗ = , (3.13)
(1 − c)σP2
and α0∗ > 0 if and only if p > qXL λP µP .

Note that if hP (α∞ , ∞) ≥ hP (α0∗ , 0) when qXL = 1, then hP (α∞

, ∞) ≥ hP (α, δ̂)
for all qXL ≥ 1 and α ≥ 0. The reason for this is that if reinsurance is not attractive
when it is cheap, then it is certainly not attractive when it is expensive.
Finally, we consider the case qXL > 1. For δ < 1 we get from (3.7),
1
hP (α, δ) = αcp − α2 c(1 − c)σP2 − qXL αcµP λP
2
Z δ/α  
(3.14)
c−1
−αcλP (1 − αx) − qXL F̄P (x)dx.
0

Now (1 − δ̂)c−1 = qXL , hence hP (α, δ̂) is differentiable and we find


∂ 1
hP (α, δ̂) = hP (α, δ̂)
∂α α
! (3.15)
1 δ̂/α
Z
−(1 − c) αcσP2 + αcλP x(1 − αx) c−2
F̄P (x)dx .
2 0

It is worth mentioning that for δ = δ0 6= δ̂, hP (α, δ0 ) is not differentiable at


α = δ0 /xP,i . Differentiability for δ = δ̂ is akin to the ”smooth fit principle” well
known in stochastic control. Some calculations show that (3.15) equals

hP (α, δ̂) = cp − αc(1 − c)σP2 − qXL cλP µP
∂α
δ̂/α
(3.16)
Z  
−cλP x (1 − αx)c−1 − qXL dFP (x)
0

= cp − αc(1 − c)σP2 − cλP E[SP (1 − min{αSP , δ̂})c−1 ].


Again this simple expression will not hold for δ = δ0 6= δ̂. It is easy to see that

h (α, δ̂) is decreasing in α, hence hP (α, δ̂) is concave in α. For α = 0,
∂α P

hP (0, δ̂) = c(p − λP µP ),
∂α
11

hence the optimal αδ̂∗ > 0 will satisfy ∂α hP (αδ̂∗ , δ̂) = 0 provided p > λP µP . Oth-
∗ ∗
erwise αδ̂ = 0. In any case the global maximum is the larger of hP (α∞ , ∞) and
hP (αδ̂∗ , δ̂).
As already noticed, δ̂ depends only on qXL and the risk aversion parameter c.
However, the optimal δ ∗ will depend on the model parameters p, σP2 , λP and FP ,
since these will determine whether δ ∗ = δ̂ or δ ∗ = ∞.

Before we continue, we will give some examples that indicate the nontrivial
behaviour of hP (α, ∞) and hP (α, δ̂).

Example 3.1. The left panel of Figure 1 gives log10 hP (αδ̂∗ , δ̂) (solid line) and

log10 hP (α∞ , ∞) (dashed line) as a function of the premium rate p for certain model
parameters (see the figure caption). For 2 ≤ p ≤ 2.21, hP (αδ̂∗ , δ̂) and hP (α∞ ∗
, ∞) are
∗ ∗
identical, while hP (αδ̂ , δ̂) > hP (α∞ , ∞) in the intervals (2.21, 3.58) and (5.08, 5.36)
(the latter is not easy to see from the figure). Thus there is no clear picture which of
hP (αδ̂∗ , δ̂) and hP (α∞

, ∞) is the largest. We also see that hP (αδ̂∗ , δ̂) and hP (α∞∗
, ∞)
∗ ∗
are both increasing in p, hence trivially the optimal hP (α , δ ) is also increasing.
The right panel gives log10 αδ̂∗ (solid line) and log10 α∞ ∗
(dashed line) again as a
function of the premium rate p, and it is seen they are equal for 2 ≤ p ≤ 2.21,
while αδ̂∗ > α∞ ∗
on the intervals (2.21, 3.13) and (4.36, 5.30). It follows from (3.16)

that log10 α∞ cannot equal log10 (1/6) = −0.778 nor log10 (1/1) = 0, and it is seen

that the jumps of log10 α∞ are across these values. Also both αδ̂∗ and α∞ ∗
are

increasing in p, and a closer look reveals that the optimal α increases as well.
However, as was seen from the left panel, the optimal δ ∗ is not monotone in p.
0

0
-1

log10 of maximizing alpha values


log10 of maximum h_P
-2

-1
-4 -3

-2
-5

-3

2 3 4 5 6 7 2 3 4 5 6 7
Premium p Premium p

Figure 1: Values of log10 hP (αδ̂∗ , δ̂), log10 hP (α∞∗


, ∞), log10 αδ̂∗ and of log10 α∞

as
2
functions of the premium rate p for c = 0.3, σP = 4, λP = 1, qXL = 1.1 and
P (SP = 1) = 0.8 and P (SP = 6) = 0.2 (giving µP = 2). Left panel: Solid line
is log10 hP (αδ̂∗ , δ̂) and dashed line is log10 hP (α∞∗
, ∞). Right panel: Solid line is
∗ ∗
log10 αδ̂ and dashed line is log10 α∞ .

12
Remark 3.2 In Example 3.1 as well as in some later examples we have chosen to
plot log10 hP (α, δ) on the y axis. The only reason to use a logaritmic scale is to
better visualize the relative behaviour of hP (αδ̂∗ , δ̂) and hP (α∞

, ∞) for the span from
p = 2 to p = 7 (and in e.g. Figure 2 below for the span from c = 0.1 to c = 0.9).
As far as we can see, log10 hP (α, δ) has no economic meaning, while exp{hP (α, δ)}
can be interpreted as the relative increase in expected utility per unit of time.
However, plotting exp{hP (α, δ)} or even hP (α, δ) will not show very clearly the
relative behaviour of hP (αδ̂∗ , δ̂) and hP (α∞

, ∞). Likewise, using a logarithmic scale
∗ ∗
for αδ̂ and α∞ is purely for visual reasons.

Example 3.2. The left panel of Figure 2 gives log10 hP (αδ̂∗ , δ̂) (solid line) and

log10 hP (α∞ , ∞) (dashed line) as a function of the risk aversion parameter c for
certain model parameters, and it can be seen that both are increasing in c. For
small c (c ≤ 0.335), i.e. high risk aversion, we see that it is optimal to have some
XL reinsurance, while for c > 0.335 it is optimal to have no XL reinsurance. From

the right panel we see that α∞ < αδ̂∗ for c < 0.319, while α∞ ∗
> αδ̂∗ for c > 0.319.
This means that for low risk aversion (high c) we get the rather surprising result
that the optimal proportion held by the company is higher when there is no XL
reinsurance than when there is. This is probably caused by the peculiar utility
function employed here, not taking into account the size of the deficit at ruin.

Furthermore both α∞ and αδ̂∗ are increasing in c, and in fact so is α∗ since α∞∗
> αδ̂∗
for c = 0.335, the changepoint. Also the optimal δ ∗ is increasing in c.
As in Example 3.1 there can be no maximum at log10 α = log10 1/6 = −0.778
nor for log10 α = log10 (1/1) = 0, and we see from Figure 2, right panel, that the

jumps are across these values. However, even though α∞ is discontinuous as a

function of c, hP (α∞ , ∞) is continuous, this is a consequence of the general fact
that hP (α, ∞) is continuous as a function of α and c.
In Figure 3 we show hP (α, δ̂) (solid line) and hP (α, ∞) (dashed line) for c =
0.25, 0.5 and 0.6. The other parameters are the same as in Figure 2. We see
that hP (α, δ̂) is smooth as shown mathematically above, and that hP (α, ∞) is not
differentiable at α = 1/6 and at α = 1, in accordance with (3.11) (and Remark

3.1). Figure 3 also exemplifies the dynamics of the jumps in α∞ as a function of
c. For c = 0.25, the maximum of hP (α, ∞) is at the top of the left ”hill”, i.e. for
α < 1/6. For c = 0.5, the maximum has shifted to the top of the middle ”hill”,
i.e. for 1/6 < α < 1, and for c = 0.6 it has again shifted to the top of the right

”hill”, i.e. for α > 1. The reason for the jumps of α∞ at c = 0.319 and c = 0.560 is
therefore that the global maximum of hP (α, ∞) goes from one ”hilltop” to another
as c changes around these points.

Example 3.3. This example is almost the same as Example 3.2, the difference is
that now SP has the continuous density
1 1
 
fP (x) = − x 1{0≤x≤6} . (3.17)
3 18
In particular this gives µP = 2 as in Example 3.2. It is seen from Figures 4 and
5 that there are no substantial differences. We have from Figure 4 that hP (αδ̂∗ , δ̂)
∗ ∗
is lower than hP (α∞ , ∞) for c < 0.437, but as c increases above 0.437, hP (α∞ , ∞)
13
1

0.5
log10 of maximizing alpha values
0
log10 of maximum h_P

0.0
-1

-0.5
-2

-1.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Risk parameter c Risk parameter c

Figure 2: Values of log10 hP (αδ̂∗ , δ̂), log10 hP (α∞∗


, ∞), log10 αδ̂∗ and of log10 α∞

as
2
functions of the risk parameter c for p = 3, σP = 4, λP = 1, qXL = 1.2 and
P (SP = 1) = 0.8 and P (SP = 6) = 0.2 (giving µP = 2). Left panel: Solid line
is log10 hP (αδ̂∗ , δ̂) and dashed line is log10 hP (α∞∗
, ∞). Right panel: Solid line is
∗ ∗
log10 αδ̂ and dashed line is log10 α∞ .

c=0.25 c=0.50 c=0.60


1.0

1.0

1.0
0.5

0.5

0.5
h_P

h_P

h_P
0.0

0.0

0.0
-0.5

-0.5

-0.5
-1.0

-1.0

-1.0

0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5
alpfa alpfa alpfa

Figure 3: Values of hP (α, δ̂) (solid line) and of hP (α, ∞) (dashed line) for c =
0.25, 0.5 and 0.6. The other parameters are as in Figure 2.

14
becomes substantially higher than hP (αδ̂∗ , δ̂). So even though SP is continuously

distributed, we observe that α∞ still can jump from one ”hilltop” to another. This
is observed in Figure 5. Again both α∗ and δ ∗ are increasing in c.
1

0.5
log10 of maximizing alpha values
0
log10 of maximum h_P

0.0
-1

-0.5
-2

-1.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Risk parameter c Risk parameter c

Figure 4: Values of log10 hP (αδ̂∗ , δ̂), log10 hP (α∞ ∗


, ∞), log10 αδ̂∗ and of log10 α∞

as
functions of the risk aversion parameter c. The parameters are the same as in
Figure 2 except that SP now has the continuous density (3.17). Left panel: Solid
line is log10 hP (αδ̂∗ , δ̂) and dashed line is log10 hP (α∞∗
, ∞). Right panel: Solid line
is log10 αδ̂∗ and dashed line is log10 α∞ ∗
.


We will now discuss the sensitivity of α∞ , αδ̂∗ and to some extend α∗ with respect
to the model parameters as well as to the risk aversion parameter c. We will prove
that most of the qualitative structures in Figures 1 to 5 are of general validity.
Since all parameters except from c are given by the model, we feel that c is the
most interesting one, and start with that. To this end let

LP,U = sup{x : FP (x) < 1}.

The following assumption is needed.


Assumption 3.1 In addition to assuming that qP = 0 and σP2 > 0 assume that
FP satisfies (3.9) and that p > λP µP .

Although not necessary for all our results, (3.9) is of sufficient generality to
include all interesting distributions for SP , and therefore nothing is lost by assuming
that it holds.

Proposition 3.1 Under Assumption 3.1 the following hold.



a) All of α∞ , αδ̂∗ , hP (α∞

, ∞) and hP (αδ̂∗ , δ̂) are strictly increasing in c. Further-

more, with the possible exception of α∞ , they are all continuous in c.

15
c=0.25 c=0.50 c=0.60

1.0

1.0

1.0
0.5

0.5

0.5
h_P

h_P

h_P
0.0

0.0

0.0
-0.5

-0.5

-0.5
-1.0

-1.0

-1.0
0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5 0.0 0.5 1.0 1.5 2.0 2.5
alpfa alpfa alpfa

Figure 5: Values of hP (α, δ̂) (solid line) and of hP (α, ∞) (dashed line) for c =
0.25, 0.5 and 0.6. The other parameters are as in Figure 4.

∗ ∗
b) When c → 1 both α∞ and hP (α∞ , ∞) → ∞, and so does both αδ̂∗ and hP (αδ̂∗ , δ̂)
if p > qXL λP µP (i.e. full XL reinsurance is lower than premium incomes).
If p < qXL λP µP then both αδ̂∗ and hP (αδ̂∗ , δ̂) stay bounded when c → 1.

When c → 0 then α∞ → α∞ (0) ≤ 1/LP,U . In particular α∞ (0) = 0 if

LP,U = ∞. Also αδ̂ → αδ̂(0) > 0.

c) There exists a c0 > 0 so that for c ≤ c0 , hP (αδ̂∗ , δ̂) ≥ hP (α∞



, ∞) and αδ̂∗ ≥

α∞ with strict inequality if LP,U = ∞. Furthermore, there exists a c1 < 1
so that for c ≥ c1 , hP (αδ̂∗ , δ̂) < hP (α∞

, ∞) and αδ̂∗ < α∞ ∗
. Indeed, both
∗ ∗ ∗ ∗
hP (α∞ , ∞) − hP (αδ̂ , δ̂) and α∞ − αδ̂ converge towards infinity as c → 1.

For the other model parameters we need a definition.

Definition 3.1 Let F1 and F2 be distribution functions and φ a functional on the


space of distributions. We say that φ is increasing in F if φ(F1 ) ≥ φ(F2 ) whenever
F1 (x) ≤ F2 (x) for all x, and strictly increasing if φ(F1 ) > φ(F2 ) when in addition
F1 (x) < F2 (x) for at least one x.

The next result is Proposition 3.1 again, but with respect to the model param-
eters. Its proof is similar to the proof for Proposition 3.1, but typically simpler,
and is therefore omitted.

Proposition 3.2 The following holds.



a) Given Assumption 3.1. Then all of α∞ , αδ̂∗ , hP (α∞

, ∞) and hP (αδ̂∗ , δ̂) are
strictly increasing in p and strictly decreasing in σP2 , λP and FP . Also αδ̂∗
and hP (αδ̂∗ , δ̂) are decreasing in qXL , and the decrease is strict if LP,U = ∞.
Finally αδ̂∗ , hP (α∞∗
, ∞) and hP (αδ̂∗ , δ̂) are continuous in all these parameters.
16

b) Given Assumption 3.1. Then all of α∞ , αδ̂∗ and hP (α∞

, ∞) converge towards
2 ∗
infinity as σP → 0, and if p > qXL λP µP , hP (αδ̂ , δ̂) does so as well. Also all

of α∞ , αδ̂∗ , hP (α∞

, ∞) and hP (αδ̂∗ , δ̂) converge towards zero as σP2 → ∞.

c) If p ≤ λP µP then α∞ , αδ̂∗ , hP (α∞∗
, ∞) and hP (αδ̂∗ , δ̂) are all zero. Conse-
quently they all converge towards zero as p → 0 or as λP µP → ∞. Fur-

thermore they all converge towards infinity as p → ∞. Finally both α∞ and
∗ 2
αδ̂ converge towards p/((1 − c)σP ) as λP → 0 or as µP → 0, and then both

hP (α∞ , ∞) and hP (αδ̂∗ , δ̂) converge towards cp2 /(2(1 − c)σP2 ).

d) Given Assumption 3.1. Then hP (α∞ , ∞) − hP (αδ̂∗ , δ̂) → ∞ and α∞

− αδ̂ →
qXL λP µP /((1 − c)σP ) as p → ∞. Furthermore, both hP (α∞ , ∞) − hP (αδ̂∗ , δ̂)
2 ∗

and α∞ − αδ̂ converge towards infinity as σP2 → 0.

We have the following result, the proof of part b follows trivially from Proposi-
tions 3.1 and 3.2, and the proof of part a is given in the appendix.

Proposition 3.3 Given Assumption 3.1.

a) The maximizing α∗ is strictly increasing in p, strictly decreasing in σP2 , λP


and FP and decreasing in qXL .

b) hP (α∗ , δ ∗ ) is strictly increasing in p and c, strictly decreasing in σP2 , λP and


FP and decreasing in qXL . Furthermore hP (α∗ , δ ∗ ) is continuous in all these
parameters.

Based on Examples 3.2 and 3.3 it is natural to conjecture that α∗ is increasing


in c as well. Unfortunately we have not been able to prove this conjecture but it is
discussed after the proof of Proposition 3.3 in the appendix.

After this lengthy discussion of the function hP , we will now move on to the
function hR . Integration by parts gives
1
hR (β) = cr̃ + βc(r − r̃) − β 2 c(1 − c)σR2
2
Z ∞
! (3.18)
c−1
+λR βc (1 + βx) F̄R (x)dx − 1 .
−1/β

The following is needed.

Assumption 3.2 Assume that σR2 > 0 and that FR is of the form
Z x ∞
X
FR (x) = fR (y)dy + pR,i 1[xR,i ,∞)(x) (3.19)
−∞ i=1

where the {xR,i } are separated.

17
Under Assumption 3.3 we get as in (3.11),
d+ 1
hR (β) = ∞ when β=− < 0,
dβ + xR,i
(3.20)
d− 1
hR (β) = −∞ when β=− > 0.
dβ − xR,i
Consequently, β ∗ 6= −1/xR,i for all i. For β 6= −1/xR,i , it is not hard to show that
hR (β) is differentiable and
h′R (β) = c(r − r̃) − βc(1 − c)σR2 + cλR E[SR (1 + βSR )c−1
+ ] (3.21)
provided this expression is finite, and if it is infinite it will have the same sign at
both sides of that point. Therefore, β ∗ must satisfy h′R (β ∗ ) = 0. We also have
h′R (0) = c(r − r̃ + λR µR ) > 0 by (2.2). However, due to the erratic behaviour of
E[(1 + βSR )c+ ], it is still possible that h′R (β) = 0 for some β < 0. But if r ≥ r̃ it is
easy to see that h′R (β) > 0 for β < 0, so now things are nicer.
The proof of the following proposition is given in the appendix.
Proposition 3.4 Assume FR is of the form (3.19). If r ≥ r̃, then the optimal
β ∗ is positive. Furthermore, both hR (β ∗ ) and β ∗ are strictly increasing in r − r̃
and in c, and strictly decreasing in σR2 and in FR . In addition hR (β ∗ ) is (locally)
increasing in λR if and only if E[(1 + β ∗ SR )c+ ] ≥ 1, and β ∗ is (locally) increasing
in λR if and only if E[SR (1 + β ∗ SR )c−1
+ ] ≥ 0.
∗ ∗
Also both hR (β ) and β converge towards infinity as r − r̃ → ∞, and if r > r̃
this convergence also takes place when σR2 → 0 or c → 1.
Finally β ∗ → β0∗ > 0 as c → 0 where β0∗ ≤ −1/LR,L if LR,L = inf{x : FR (x) >
0} < 0.
Example 3.4 When r < r̃, Proposition 3.4 will generally not hold, but since this
is a rather uncommon investment situation we shall not go into details, just give
an example. Figure 6 shows log10 hR (β ∗ ) (left panel) and β ∗ (right panel) as a
function of the risk parameter c when r = 0 and r̃ = 0.1. The other parameters are
σR2 = 0.1, λR = 1, P (SR = −0.5) = p0 and P (SR = 0.5) = 1 − p0 . The results are
shown for p0 = 0.6, 0.8 and 1.0, giving r − r̃ + λR µR = 0.0, 0.2 and 0.4 respectively.
We see from the left panel that hR (β ∗ ) is increasing in c, while from the right panel
it is seen that β ∗ is decreasing in c, and in addition that β ∗ is always negative (no
logscale here). In this case h′R (β) = 0 for two different β values, one negative and
one positive, and hR (β) is largest at the negative root. From the right panel we
also see that β ∗ never takes the value -2, in accordance with (3.20).

4 Logarithmic utility and minimization of expected


time to reach a given capital
In this section we start with the problem (2.7) using the utility function
(
ln y; y > 0,
U(y) = (4.1)
−∞; y ≤ 0.
We then have with a = αy and D = δy.
18
1.0

0
p_0=0.6 p_0=0.6
p_0=0.8 p_0=0.8
p_0=1.0 p_0=1.0

0.5

-5
maximizing beta values
log10 of maximum h_R
0.0

-10
-1.0 -0.5

-20 -15
-1.5
-2.0

-25
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Risk parameter c Risk parameter c

Figure 6: Values of log10 hR (β ∗ ) (left panel) and of β ∗ (right panel) as a function of


the risk aversion parameter c. The parameters are r = 0, r̃ = 0.1, σR2 = 0.1, λR =
1, P (SR = −0.5) = p0 and P (SR = 0.5) = 1 − p0 . The results are shown for
p0 = 0.6, 0.8 and 1.0.

Theorem 4.1 Let YTπ be given by (2.4) with qP = 0. Let


1
gP (α, δ) = αp − qXL λP ρα (δ) − α2 σP2 + λP E[ln(1 − min{αSP , δ})] (4.2)
2
and
1
gR (β) = r̃ + β(r − r̃) − β 2 σR2 + λR E[ln(1 + βSR )]. (4.3)
2
Also let
LR,L = inf{x : FR (x) > 0} and LR,U = sup{x : FR (x) < 1}
and assume that LR,L > −∞. We seek to maximize J π (t, y) = E t,y [U(YTπ )] where
U is given by (4.1).
a) If σP2 > 0 or p < qXL λP µP there exists a finite α∗ and a δ ∗ given by δ ∗ =
−1
(1 − qXL )+ so that gP (α, δ) attains its maximum at (α∗ , δ ∗ ). Also gP (α, δ ∗ )
is concave in α.
The function gR is concave on IR = {β : min{βLR,L , βLR,U } > −1} and
minus infinity outside IR , hence there is a finite β ∗ so that hR (β) attains its
maximum at β ∗ .
b) If σP2 > 0 or p < qXL λP µP then the control πt∗ = (α∗ Yt−
π
, δ ∗ Yt−
π
, β ∗ ) is optimal
∗ ∗

π∗
among all admissible controls, and J (t, y) = J(t, y) = VL (t, y) where
VL (t, y) = ηL∗ (T − t) + ln y, (4.4)
with
ηL∗ = gP (α∗ , δ ∗ ) + gR (β ∗ ).
19
c) Assume that σP2 = 0 and p = qXL λP µP . Then there exists an optimal control
if and only if qXL ≤ 1. If qXL > 1 there is no optimal control, but the value
function J(t, y) is bounded for y > 0.

Example 4.1. Figure 7, left panel, shows gP (α, δ ∗ ) for various values of qXL .
The other model parameters are as in Figure 2. We see that α∗ decreases as a
function of qXL . In the right panel gR (β) is shown for three different choices of the
distribution of SR , and it is seen that β ∗ increases as µR increases.
0.12

p_0=1.0
p_0=0.5
p_0=0.0

0.2
0.04 0.06 0.08 0.10
g_P for optimal delta

0.1
q_XL=1.0

g_R
0.0
q_XL=1.2
0.02

q_XL=1.5
-0.1
q_XL=2.0
0.0

0.0 0.1 0.2 0.3 0.4 0.5 -3 -2 -1 0 1 2 3


alpha beta

Figure 7: Left panel: Values of gP (α, δ ∗) for various values of qXL . The other
model parameters are as in Figure 2. Right panel: Values of gR (β) when r̃ =
0, r = 0.1, σR2 = 0.1, λR = 1, P (SR = −0.1) = p0 and P (SR = 0.1) = 1 − p0 . The
figure shows gR (β) for three different values of p0 .

We will now show that the results obtained here are basically limits of those
obtained in Section 3 when we let c → 0 there with both σP2 and σR2 positive. As
1/(c−1)
in Section 3 write δ̂(c) = (1 − qXL )+ for the δ̂ given there, and similarly write

αδ̂(c) and β ∗ (c) for the αδ̂∗ and β ∗ found there respectively. Using that δ̂(c) →
−1
(1 − qXL )+ = δ ∗ and that for x ∈ (0, 1), ((1 − x)c − 1)/c → ln(1 − x) as c → 0, it
follows that for any α
1
hP (α, δ̂(c)) → gP (α, δ ∗ ) as c → 0. (4.5)
c

Since hP (α, δ̂(c))/c → −∞ uniformly as α → ∞ for c ≤ 0.5 say, αδ̂(c) → α∗ as
c → 0. Similarly
1
hR (β) → gR (β) as c → 0, (4.6)
c
and therefore β ∗ (c) → β ∗ as c → 0. The following result is proved in the appendix.

20
Proposition 4.1 Assume that Assumptions 3.1 and 3.2 hold. Let π̃ ∗ (c) =
(α∗ (c), δ ∗ (c), β ∗(c)) be the maximizers of (3.4) and (3.5), and let π̃ ∗ = (α∗ , δ ∗ , β ∗ )
be the maximizers of (4.2) and (4.3). Then

π̃ ∗ (c) → π̃ ∗ as c → 0. (4.7)
π ∗ (c) π ∗ (c)
Furthermore, let πt∗ (c) = (α∗ (c)Yt− , δ ∗ (c)Yt− , β ∗ (c)) and πt∗ = (α∗ Yt−
π
, δ ∗ Yt−
π
, β ∗)
∗ ∗

be the optimizing controls. Then


π ∗ (c)
− Ytπ | → 0 a.s and in L1 as c → 0.

sup |Yt (4.8)
t≤T

Also (using any norm on R3 ),

sup |πt∗ (c) − πt∗ | → 0 a.s. and in L1 as c → 0. (4.9)


t≤T

Furthermore, if E[|SR |q ] < ∞, the above convergence is also in Lq . Finally


1  t,y π ∗ (c)
 1
ln E [(YTy,t )c ] = ln(VP (t, y)) → VL (t, y) = E t,y [ln(YTπ )] as c → 0.

c c
As in (4.5), using (3.16) and (A.14) we get
1 ∂ ∂
hP (α, δ̂(c)) → gP (α, δ ∗ ) as c → 0, (4.10)
c ∂α ∂α
and similarly with the other derivatives. Therefore, the following result is a conse-
quence of Propositions 3.1 and 3.2 for the case when σP2 > 0 (or it can be proved
directly). If σP2 = 0 but p < qXL λP µP it can be proved directly (for the relevant
results).
Proposition 4.2 Assume that qP = 0, that FP satisfies (3.9) and that either σP2 >
0 or p < qXL λP µP . Then α∗ and gP (α∗ , δ ∗ ) have the same properties as αδ̂∗ and
hP (αδ̂∗ , δ̂) in Proposition 3.2.
The next result is a consequence of Proposition 3.4, together with arguments
such as those in (4.6) and (4.10).
Proposition 4.3 Assume FR is of the form (3.19) and that r̃ = 0. Then b∗ and
kR (b∗ ) (except from the c-dependence) have the same properties as β ∗ and hR (β ∗ )
in Proposition 3.3 provided the conditions E[(1+SR )c+ ] ≥ 1and E[SR (1+SR )c+ ] ≥ 0
there are replaced by E[ln(1 + β ∗ SR )] ≥ 0 and E[SR (1 + β ∗ SR )−1 ] ≥ 0 respectively.
Also β ∗ ≤ −1/LR,L if LR,L < 0.
Related to the above problem is the problem of finding a control that minimizes
the expected time to reach a given capital ȳ > y = Y0π before reaching zero.
This relationship was proved in a special diffusion setting by Pestien and Sudderth
(1985), and we shall see that it is valid in our model as well. To be more precise,
let τȳπ = inf{t : Ytπ ≥ ȳ} and set
(
π τȳπ if τȳπ < Tyπ ,
Ty,ȳ =
∞ if τȳπ ≥ Tyπ ,
21
where as before Tyπ = inf{t : Ytπ ≤ 0}. We shall assume that Y π is skip-free
upwards, i.e. that FR (0) = 1, implying that τȳπ = inf{t : Ytπ = ȳ}.
There is clearly no time dependence here, so we write
J π (y) = E[Ty,ȳ
π
] and π
J(y) = inf E[Ty,ȳ ].
π

The proof of the next theorem is given in the appendix.

Theorem 4.2 Let π ∗ be as in Theorem 4.1b and in addition to the assumptions


there assume also that FR (0) = 1. Then for 0 < y ≤ ȳ,
1 ȳ
J(y) = J π (y) =


ln ,
ηL y
where ηL∗ is given in Theorem 4.1a.

5 Exponential utility
In this section we consider the problem (2.7) using the utility function U(y) =
−e−cy for some positive c. Here the company is allowed to continue after ruin.
π
For a control π = (a, D, β), define b by bt = βt Yt− . Then (2.4) becomes
Z t Z t
Ytπ =y+ Pta,D + bs dRs + r̃ π
(Ys− − bs )ds. (5.1)
0 0

We will find a π ∗ that is optimal among a somewhat smaller class than the class
of admissible controls defined in Section 2.
Definition 5.1 We define Π0 to be set of all admissible controls that satisfy
" #
−cYtπ
E sup e < ∞. (5.2)
t≤T

Furthermore, we let ΠB be the set of all admissible bounded controls, i.e all admis-
sible controls so that
sup max{at , |bt |}
t≤T

is bounded.
Any admissible tame portfolio, i.e. any admissible control so that inf t≤T Ytπ is
bounded from below, belongs to Π0 .

Theorem 5.1 Let YTπ be given by (2.4) with r̃ = 0 and let


kP (a, D) = acp + (1 − a)qP cp − cqXL λP ρa (D)
1  h i  (5.3)
− a2 c2 σP2 − λP E ec min{aSP ,D} − 1
2
and
1  h i 
kR (b) = bcr − b2 c2 σR2 − λR E e−cbSR − 1 . (5.4)
2
We seek to maximize E [U(YT )] where U(y) = −e−cy .
t,y π

22
a) If σP2 > 0 or (1 − qP )p < qXL λP µP , there exists a finite a∗ and a D ∗ given
by D ∗ = 1c (ln qXL )+ so that kP (a, D) attains its maximum at (a∗ , D ∗ ). Also
kP (a, D ∗ ) is concave in a.
There exists a finite b∗ so that kR (b) attains its maximum at b∗ . Also kR is a
concave function.

b) If σP2 > 0 or (1 − qP )p < qXL λP µP , then the control πt∗ = (a∗ , D ∗ , b∗ /Yt−
π ∗
)
π ∗
is optimal among all controls in Π0 ∪ ΠB , and J (t, y) = J(t, y) = VE (t, y)
where
VE (t, y) = −e−ηE (T −t)−cy ,

with
ηE∗ = kP (a∗ , D ∗) + kR (b∗ ).

c) Assume that σP2 = 0 and (1 − qP )p = qXL λP µP . Then there exists an optimal


control if and only if qXL ≤ 1.
π
The proof of part b actually shows that any π ∈ ΠB so that E[e−cYT ] < ∞ is also
in Π0 . Unfortunately we have not been able to prove the following more general
result which then would imply that π ∗ is optimal among all admissible controls.
π
Conjecture. All admissible controls π that satisfy E[e−cYT ] < ∞ belongs to Π0 .

It turns out that the dependency on the risk aversion c is fairly easy in this case.

Proposition 5.1 Assume that π ∗ exists. As functions of the risk aversion c, write

a∗ = a∗ (c), D ∗ = D ∗ (c) and b∗ = b∗ (c).

Then
a∗ (1) D ∗ (1) b∗ (1)
a∗ (c) = , D ∗ (c) = and b∗ (c) =
c c c

with D (1) = (ln qXL )+ . Furthermore,

kP (a∗ (c), D ∗ (c)) = kP (a∗ (1), D ∗(1)) + qP (c − 1)p

and
kR (b∗ (c)) = kR (b∗ (1)).
Therefore
ηE∗ = kP (a∗ (1), D ∗(1)) + kR (b∗ (1)) + qP (c − 1)p,
and in particular ηE∗ as a function of c is constant if qP = 0.

Proof. Since D ∗ (c) = 1c (ln qXL )+ , the result is trivially true for D ∗ (c). Further-
more,

kP (a, D ∗ (c)) = qP cp + ac(1 − qP )p − qXL λP E[(acSP − D ∗ (1))+ ]


1  h i 
− (ac)2 σP2 − λP E emin{acSP ,D (1)} − 1 .

2
23
From this it follows that an optimizing point must satisfy ac = a0 , a constant,
hence the result. The proof for b∗ (c) is similar.

Because of the concavity of kP (a, D ∗ ) and kR , the following result can be estab-
lished along the same lines, but typically simpler, as those for hP (αδ̂∗ , δ̂ ∗ ) and αδ̂∗ in
Section 3.

Proposition 5.2 Assume that r̃ = 0, that FP satisfies (3.9) and that either σP2 > 0
or p < qXL λP µP . Then a∗ and kP (a∗ , D ∗ ) have the same properties as αδ̂∗ and
hP (αδ̂∗ , δ̂) in Proposition 3.2 provided that statements such as p < qXL λP µP are
replaced with (1 − qP )p < qXL λP µP .

The next result is similar to Proposition 3.3.

Proposition 5.3 Assume FR is of the form (3.19) and that r̃ = 0. Then b∗ and
kR (b∗ ) (except from the c-dependence) have the same properties as β ∗ and hR (β ∗ )
in Proposition 3.3 provided the conditions E[(1+SR )c+ ] ≥ 1and E[SR (1+SR )c+ ] ≥ 0
there are replaced with E[e−cb SR ] ≤ 1 and E[SR e−cb SR ] ≥ 0 respectively.
∗ ∗

We will now allow r̃ to be nonzero, but instead we shall assume that λR = 0.


The proof of the following theorem is given in the appendix.

Theorem 5.2 Let YTπ be given by (2.4) and assume that λR = 0. Let

k̃P (a, D) = ac(1 − qP )p − cqXL λP E[(aSP − D)+ ]


1  h i  (5.5)
− a2 c2 σP2 − λP E ec min{aSP ,D} − 1
2
and
1
k̃R (b) = bc(r − r̃) − b2 c2 σR2 . (5.6)
2
We seek to maximize E [U(YT )] where U(y) = −e−cy .
t,y π

a) Part a in Theorem 5.1 applies for k̃P (a, D) and k̃R (b) giving maximizers ã∗
and D̃ ∗ equal to a∗ and D ∗ in Theorem 5.1a, while the maximizer b̃∗ is equal
to b∗ there with r replaced by r − r̃ and λR = 0.

b) Assume that σP2 > 0 or (1 − qP )p < qXL λP µP and let

ã∗t = e−r̃(T −t) ã∗ , D̃t∗ = e−r̃(T −t) D̃ ∗ and b̃∗t = e−r̃(T −t) b̃∗ .

Then π̃t∗ = (ã∗t , D̃t∗ , b̃∗t /Yt−π̃



) is optimal among all controls in Π0 ∪ ΠB . Fur-
π̃ ∗
thermore, J(t, y) = J (t, y) = ṼE (t, y), where
cqP p r̃(T −t)
  
ṼE (t, y) = − exp − η̃E∗ (T − t) + (e − 1) + cer̃(T −t) y ,

and
η̃E∗ = k̃P (ã∗ , D̃ ∗) + k̃R (b̃∗ ).

24
c) Part c of Theorem 5.1 is valid here as well.

Thus both the proportion of the insurance business that is not ceded to reinsurers
and the excess point as well as the proportion of assets invested in the risky asset
increases with time. Browne (1995) obtained similar conclusions in his model.
It follows easily that Propositions 5.1, 5.2 and 5.3 all hold in this case as well.
However, in Propositions 5.3 for all statements involving r, this r must be replaced
by r − r̃, and all statements involving λR are of course void.

6 Discussion and extension to policy constraints


In Sections 3 and 4 the optimal controls were of the form πt∗ = (α∗ Yt− π
, δ ∗ Yt−
π
∗ ∗
, β∗),
∗ −r̃(T −t) ∗ ∗ ∗ π
(a , D , b /Yt− ) with (α , δ , β ∗)
∗ ∗

while in Section 5 they were of the form πt = e
∗ ∗ ∗
and (a , D , b ) constants.
With the controls in Sections 3 and 4 the proportion of the insurance business re-
tained in the company at time t is a∗t = α∗ Yt− π∗
, and the excess point is Dt∗ = δ ∗ Yt− π∗
.
Consequently the company will have to continuously rebalance these proportions.
Furthermore, if a claim occurs at time t, a∗t will be reduced from α∗ Yt− π∗
to α∗ Ytπ ,

and if δ ∗ = δ̂, Dt∗ is reduced from δ̂Yt−π∗


to δ̂Ytπ . Unfortunately, the reinsurers may

not accept such a reduction since a (large) claim will usually be a signal to them
to decrease their exposure to that particular portfolio rather than to increase it.
Hence the optimal policy found there may sometimes be difficult to implement in
practice. But not necessarily so, and for an exception consider the deductible pro-
cess D only. If the time horizon T equals one year, it is not uncommon that after
a large claim the reinsurance contract specifies that the excess point on the next
claims will be lower. By our results, such a contract can be seen as an approxima-
tion to an optimal solution since in Dt∗ = δ ∗ Yt− π∗
, the term Yt− π∗
will be small after
a large claim. Rather common are so called AAD (annual aggregate deductible)
contracts which specify an upper limit of aggregated deductibles during a year and
then the reinsurers cover the rest. Again, the per claim deductible Dt∗ = δ ∗ Yt− π∗
will
π ∗
approximate such a contract since one or many large claims will make Yt− small.
Optimal reinsurance controls of the form (α∗ Yt− π∗ ∗ π∗
, δ Yt− ) indicate that an option
market for reinsurance capacity could be meaningful. At the beginning of the period
the insurance company buys an option that allows it to change the reinsurance
(at , Dt ) a specific number of times during the period, typically one year. Such
an option would allow the company to better approximate its optimal reinsurance
policy. Common practice today is that companies buy extra reinsurance during
the year if claims are large.
Frequently reinsurance levels stay fixed over a period of time, in which case the
optimal control (a∗ , D ∗ ) found in Section 5 when r̃ = 0 is very attractive from a
practical point of view. However, if r̃ > 0, it changes to er̃(T −t) (a∗ , D ∗ ), and this
requires continuous rebalancing. But if T is a certain number of years, a good
approximation would be to let e.g. (at , Dt ) = er̃(T +0.5−i) (a∗ , D ∗ ) for i − 1 ≤ t < i
with i = 1, . . . , T . In this way the reinsurance is only changed once a year.
When it comes to investments these are usually rebalanced frequently, so both
the optimal solution in Sections 3 and 4 with a constant fraction of wealth in the
25
risky asset as well as the solution in Section 5 with b∗t = e−r̃(T −t) b∗ are in principle
practically feasible. In the latter case total investments in the risky asset increase
exponentially in time, and remain constant if the risk free rate is zero,

It may well be that the optimal controls found in this paper are either too risky
or even not allowed legally. Then constraints may have to be imposed.
Consider the utility functions of Sections 3 and 4. If π̂ = (α, δ, β) is constrained
to lie in a closed set V ⊂ R, Theorems 3.1, 4.1 and 4.2 all hold with e.g. ηP∗ in
Theorem 3.1 replaced with

ηP∗ V = sup(hP (α, δ) + hR (β)).


π̂∈V

If V = [α0 , α1 ] × [δ0 , δ1 ] × [β0 , β1 ], a rectangle, then as before we can maximize


hP (α, δ) and hR (β) separately. It is not unreasonable to have a constraint of
the form β ∈ [β0 , β1 ], giving lower and upper bounds for the relative amount
of assets invested in the risky asset. Also, except possibly for solvency reasons,
there is usually no reason to limit δ, so δ ∈ [0, ∞] is typically unproblematic, this
is equivalent to D ∈ [0, ∞]. On the other hand, it is a and not α that represents
the fraction of the insurance income that is retained in the company. A normal
restriction would therefore be a ≤ 1, in which case a = αy for some α > 0 is not
permissible. A natural conjecture for an optimal policy under this restriction is
a = α∗ y for α∗ y ≤ 1, and then a = 1 for α∗ y > 1. A proof of this conjecture (if it
is true) and the actual calculation of the value function J π (t, y) under this policy
seems rather difficult however.
The possibility that at > 1 certainly seems a bit unnatural, meaning more or
less that the insurance business is used for speculative purposes. Since there is
no market for insurance business similar to the financial markets, this would be
difficult to implement in practice. However, a∗t > 1 can be interpreted as a sign
that the company has a wrong underwriting policy. Assume for example that it
takes a share ξt with 0 < ξt < 1 of a co-insurance contract, this is quite common
with large catastrophic-type risks. If the optimal a∗t reinsured of its own share is
larger than one, this just tells the company that it should increase its share ξt of
the total risk. If it is increased to kξt for some k > 1 (meaning that p is increased
to kp, σP2 is increased to k 2 σP2 and SP,i to kSP,i), then α∗ is decreased to α∗ /k while
hP (α∗ , δ ∗ ) remains the same. This is rather obvious, and can also easily be seen
from (3.4) by using the reparametrization α̃ = α/k.
In the case with no co-insurance, a∗t > 1 suggests that the company should
venture into more risky business by underwriting more dangerous risks. If for some
k > 1 the new risks have distribution FP′ (x) = FP (x/k), and p becomes kp and σP2
becomes k 2 σP2 , then again α∗ decreases to α∗ /k while hP (α∗ , δ ∗ ) remains the same.
On the other hand, underwriting more independent business will not change α∗ ,
but it will change hP (α∗ , δ ∗ ) and thereby J(t, y). If for example new independent
and similar business that is k − 1 times as large as the original one is added so that
p is increased to kp, σP2 is increased to kσP2 and λP is increased to kλP , then α∗ will
not change, but hP (α∗ , δ ∗ ) is increased to khP (α∗ , δ ∗ ), and the relative increase in
utility When T = 1 is exp{(k − 1)hP (α∗ , δ ∗ )}.
The same arguments basically apply for the utility function of Section 4.
26
Now consider the utility function of Section 5. Assume there are constraints of
the form π̌t = (at , Dt , bt ) ∈ V , a t-independent closed set in R3 . If r̃ = 0, Theorem
5.1 still holds, but with ηE∗ replaced by

ηEV = sup(kP (a, D) + kR (b)).
π̌∈V

If r̃ > 0 and λR = 0, let (ã∗V,t , D̃V,t∗


, b̃∗V,t ) maximize k̃P (a, D) + k̃R (b) subject to the
constraint that er̃(T −t) π̌t ∈ V . Then Theorem 5.2 holds with η̃E∗ replaced by η̃EV ∗
=
∗ ∗ ∗ ∗ ∗ ∗ ∗ π∗
k̃P (ãV,T , D̃V,T ) + k̃R (b̃V,T ). The optimal control is then πt = (ãV,t , D̃V,t , b̃V,t /Yt− ).
All this follows from an inspection of the proof of Theorem 5.2, and in particular
from (A.39).
If V = [a0 , a1 ] × [D0 , D1 ] × [b0 , b1 ], a rectangle, then as before we can maximize
kP (a, D) and kR (b) as well as k̃P (a, D) and k̃R (b) separately. Constraints of the
form a ∈ [a0 , a1 ] and D ∈ [D0 , D1 ] (often with D0 = 0 and D1 = ∞) are fairly
unproblematic and reasonable, and if r̃ = 0 we saw that they result in an easy
solution. Constraining total capital b invested in the risky asset so that b ∈ [b0 , b1 ]
also makes sense, at least if the time horizon T is not too long. An alternative as
we saw above is to constrain the proportion of total assets β invested in the risky
asset, i.e. β ∈ [β0 , β1 ]. If e.g. short sales and lending risky assets are disallowed,
the constraint becomes β ∈ [0, 1]. Similar to the analogous situation above, a
conjecture is that an optimal policy with a constraint of the form β ∈ [β0 , β1 ] is to
let b = b∗ if b∗ /y ∈ [β0 , β1 ], b = β0 y if b∗ < β0 y and b = β1 y if b∗ > β1 y. A proof
of this conjecture (if it is correct) is again a rather difficult task. However, just as
above with a∗t > 1, βt∗ > 1 may indicate that the company should look for more
risky investments.

One possible condition on the controls is to limit the probability of ruin up to


time T , i.e. to require that ψTπ (y) = P (Tyπ < T ) ≤ ε for some preassigned ε. Now
for the optimal control in Section 4 it is clear that ψTπ (y) = 0, so here a ruin

constraint is automatically satisfied. Looking at the situation in Section 3, for any


strategy π = (αy, δy, β) with π̃ = (α, δ, β) constant, it follows from (3.2) and (3.3)
that with LR,L = inf{x : FR (x) > 0},
   
NP,T NR,T
ψTπ (y) = 1 − P 
Y Y
(1 − min{αSP , δ})+ > 0 P  (1 + βSR )+ > 0
i=1 i=1

= 1 − (1{δ<1} + 1{δ≥1} E[FP (α−1)NP,T ])


×(1{0≤β<−1/LR,L } + 1{β<0} E[FR (−β −1 )NR,T ] + 1{β≥−1/LR,L } E[F̄R (−β −1 )NR,T ])
= 1 − e−κ(π̃)T ,

where

κ(π̃) = λP 1{δ≥1} F̄P (α−1 ) + λR (1{β<0} F̄R (−β −1 ) + 1{β≥−1/LR,L } FR (−β −1 )).

In particular if δ < 1 and 0 ≤ β < −1/LR,L , then ψTπ (y) = 0. Note that ψTπ (y)
is independent of y, making it very well suited as a control constraint since it
will not change continuously in time with Ytπ . If we want to maximize J π (0, y) =
27
E 0,y [(YTπy,T c π
π )+ ] subject to the solvency constraint ψT (y) ≤ ε, then by the above we

must find
ηP∗ U = sup(hP (α, δ) + hR (β)),
π̃∈U

where U is the set


1
{π̃ : κ(π̃) ≤ − log(1 − ε)}.
T
We just saw that when δ ∗ = δ̂ < 1 and 0 ≤ β ∗ < −1/LR,L , the unconstrained
solution given above also solves this constrained problem.
Moving on to the situation in Section 5 and considering only controls with
constant π̂ = (a, D, b), (5.1) takes the form
Nt
Ytπ = y + ξt + σWt −
X
Si , (6.1)
i=1

with

ξ = ap + (1 − a)qP p − ρa (D) + br,


σ 2 = a2 σP2 + b2 σR2 ,

and W a Brownian motion. Furthermore, N is a Poisson process with intensity


λP + λR , and the {Si } are i.i.d. with distribution function
λP x λR x
 
F (x) = FP ( )1{x<D} + 1{x≥D} + F̄R ( ).
λ a λ b
So in this case Y π is just a diffusion perturbated classical risk process, exactly
as the original processes P and R. Now ψTπ (y) depends on y and therefore it will
continuously change with the state Ytπ , making it much less clear how to apply it as
a constraint. At any rate ψTπ (y) is not easy to calculate analytically for (6.1), see e.g.
Asmussen (2000), hence the available options will either be to solve the appropriate
integro-differential equations for ψTπ (y) numerically, or to use Monte-Carlo methods.
For the latter methods, importance sampling may be useful, see e.g. Lehtonen and
Nyrhinen (1992) or Paulsen and Rasmussen (2003). When r̃ > 0 and λR = 0 in
π
Section 5, constraints limiting the ruin probability ψT (y) as for the case with r̃ = 0
can be implemented here as well. However, due to the nature of the solution in the
unconstrained case, considering controls of the form er̃(T −t) (a, D, b) with (a, D, b)
constant would be more natural. Then the controlled process Y π will not be time
homogeneous, but since the ruin problem for homogeneous processes typically has
to be solved by numerical methods anyway, this is not a serious handicap.

7 A multivariate generalization
So far there has been only one insurance process and only one risky investment.
However, as we shall see, extending to several lines of insurance as well as several
risky investment opportunities can be fairly straightforward.
Assume that the company has nP different lines of business; with the exception
of some very specialized companies this is the case in practice where nP can be fairly
28
large, depending of course on what is considered to be different lines of business.
We model each line as in Section 2, i.e.
j
NP,t
T j
X
Pj,t = pj,tt + σP,j WP,t − SP,i , j = 1, . . . , nP , (7.1)
i=1

where WP is now an mP -dimensional Brownian motion and σP,j is an mP -dimensional


vector. Furthermore, NPj is a Poisson process with intensity λjP and SP,1 j j
, SP,2 ,...
j j
are i.i.d. with distribution function FP . It is assumed that FP (0) = 0.
Although claims from different lines of business may exhibit some dependence
due to market conditions, it is not very unrealistic to assume that the compound
Poisson processes in (7.1) are all independent, and consequently we shall do so.
Premium rates on the other hand depend on the general insurance market, and
independence is less realistic. We will therefore allow the covariance matrix ΣP
T
given by (ΣP )i,j = σP,i σP,j to be nondiagonal.
Similar to what was done in Section 2 we let aj,t be the proportion of business
in line j that is retained by the company at time t, and Dj,t the excess point after
proportional reinsurance. With provision factors qPj and qXL j
, the accumulated
insurance process becomes
nP
a ,Dj
Pta,D
X
= Pj,tj ,
j=1

where
Z t Z t
a ,Dj
Pj,tj = (aj,s pj + (1 − aj,s )qPj pj − qXL
j
λjP ρaj,s (Dj,s ))ds + T
aj,s σP,j dWP,s
0 0
j
NP,t
j
X
− min{aj,T j SP,i , Dj,T j }.
P,i P,i
i=1

In addition we allow for nR risky investments, and we shall assume that they are
all independent of the insurance business. The return of risky asset j is assumed
to follow
j
NR,t
T j
X
Rj,t = rj t + σR,j WR,t + SR,i ,
i=1
where WR is an mR -dimensional Brownian motion and σR,j is an mR -dimensional
j
vector. Furthermore, NRj is a Poisson process with intensity λjR and SR,1 j
, SR,2 ,...
j j
are i.i.d. with distribution function FR , where FR (−1−) = 0. As in (2.2) and (2.3)
we also assume that
r̃ < rj + λjR µjR , j = 1, . . . , nR
and
T
σR,j σR,j + λjR FRj (0−) > 0 or that rj < r̃, j = 1, . . . , nR
Now let βj,t be the proportion held in risky asset j at time t. Let πs = (as , Ds , βs )
where as = (a1,s , . . . , anP ,s )T , Ds = (D1,s , . . . , DnP ,s )T and βs = (β1,s , . . . , βnR ,s )T .
The total asset model (2.4) becomes
 
Z t nR nR
Ytπ = y + Pta,D + π 
X X
Ys− βj,s dRj,s + (1 − βj,s )dR̃s  . (7.2)
0 j=1 j=1

29
Let α = (α1 , . . . , αnP )T , δ = (δ1 , . . . , δnP )T and β = (β1 , . . . , βnR )T . Then just
as in Theorem 1, with an identical proof.
Theorem 7.1 Let Y π be given by (7.2) with qP = 0 and let
nP 
j
 1
λjP ραj (δj ) − c(1 − c)αT ΣP α
X
hP (α, δ) = c αj pj − qXL
j=1 2
nP
λjP (E[(1 − min{αj SPj , δj })c+ ] − 1)
X
+
j=1

and
nR nR
1
βj (rj − r̃) − c(1 − c)β T ΣR β + λjR (E[(1 + βj SRj )c+ ] − 1).
X X
hR (β) = cr̃ + c
j=1 2 j=1

We seek to maximize J π (t, y) = E t,y [U(YTπy,T c


π )] with U(y) = y+ , 0 < c < 1.

a) If ΣP is nonsingular there is a finite α∗ and a δ ∗ (may be infinite) so that


hP (α, δ) attains its maximum at (α∗ , δ ∗ ). In fact each component δj∗ of δ ∗ can
either take the value δj∗ = ∞ or δj∗ = δ̂j where
j
δ̂j = (1 − (qXL )1/(c−1) )+ .
If ΣR is nonsingular there is a finite β ∗ so that hR (β) attains its maximum
at β ∗ .
b) If both ΣP and ΣR are nonsingular, the control π ∗ given by π ∗ = (α∗ Yt−
π
, δ ∗ Yt−
π
, β ∗)
∗ ∗

is optimal among all admissible controls and


J(t, y) = J π (t, y) = eηP (T −t) y+
c
∗ ∗
,
with
ηP∗ = hP (α∗ , δ ∗ ) + hR (β ∗ ).
c) If either ΣP or ΣR = 0 is singular, then there is no optimal control and the
value function J(t, y) = ∞ for t < T, y > 0.
If ΣP is diagonal so that the insurance processes Pj are independent, we get
nP
X
hP (α, δ) = hP (αj , δj ), (7.3)
j=1

where each hP (αj , δj ) is as in (3.4). In this case the hP (αj , δj ) can be maximized
separately, and we are back to the situation in Section 3. Another consequence
of (7.3) is that each line of insurance can be optimized separately, thus reducing
the need for an overall reinsurance policy. The same separation obtains if ΣR is
diagonal.
Assume that ΣP and ΣR are both nonsingular. Note that there are 2nP possibil-
ities for δ ∗ , hence the optimal (α∗ , δ ∗ ) can be found by maximizing hP (α, δ) w.r.t.
α as δ takes each of these 2nP values, and then take the pair that yields the largest
maximum. However, if nP is large this may become computationally infeasible. On
the other hand, maximizing hR (β) w.r.t. β should not be too complicated unless
nR is very large.
Generalizations of the results of Section 4 and 5 to the multivariate case can be
done in the same way, and the details are left to the reader.
30
8 Concluding remarks
In this paper we have seeked to maximize expected utility at a terminal time,
and we have demonstrated that seemingly very complex control problems can have
fairly simple solutions. The uncontrolled process was a mixed additive-geometric
model where premiums and claims constitute the additive part and return on in-
vestments constitute the geometric part. From an analytical point of view, such
models are difficult to deal with, but once control is introduced the optimal con-
trols transformed these mixed models into either pure geometric or pure additive
models, which are substantially more simple. Not surprisingly, given the concept
of risk aversion, utility functions with constant relative risk aversion turned the
optimally controlled risk process into a geometric one, while constant absolute risk
aversion utility functions turned it into an additive process. In all cases invest-
ment decisions can be made independently of the insurance decisions. This is in
accordance with what most insurance companies do in practice.
This all seems fine, but the price to pay using our general model is that natural
restrictions on the controls can make the problem untractable. So in a way the
analytical methods employed in this paper are not good enough for applications. To
have more flexibility in the model assumptions, numerical solutions will ultimately
have to be utilized.
In the paper three different utility functions have been used: power utility with
termination at ruin, logarithmic utility and exponential utility. None of these
singles out as a clear ”winner”, they all have their pros and cons. The two former
are very similar, indeed we saw that the second was a limiting case of the first.
A minus with the first two is that they do not take into account the severity
of ruin, and since the logaritmic utility gives minus infinity to negative assets, the
optimal control makes this an event of probability zero, and consequently minimizes
this probability. The power utility function does not punish ruin very hard, at
least when the relative risk aversion 1 − c is small, so in this case the optimal
control may lead to an unacceptably high ruin probability. It may be excellent for
the management of the company, but not for its customers. Exponential utility
punishes deficits at the terminal time according to their severity, and large deficits
are punished very hard. On the other hand, it does not give much reward to high
surpluses, and may therefore lead to overly prudent policies. A utility function
that balances these considerations could be


 1 − ae−c0 y y < 0,
U(y) = 
 (y + b)c1 y ≥ 0,

with 0 < c1 < 1, a, b and c0 positive and a ≥ 1 − bc1 . In this case however,
numerical methods are probably necessary. The drawback with numerical methods
is that general sensitivity results are less transparent, but as discussed above the
advantage is that additional restrictions on the controls do not necessarily change
the problem into an unmanageable one.
During the course of the paper we have seen that some of the optimal reinsurance
policies proposed are not easily implemented in practice. This should call for a

31
more flexible market where insurance and reinsurance contracts as well as options
on them can be created and traded.

9 Acknowledgments
The second author was supported in part by the Norwegian Research Council and
the University of Chicago. Parts of the paper was written while he was visiting the
Department of Statistics and the Mathematical Finance Program at the University
of Chicago. I would like to thank them, and in particular Per Mykland, for their
hospitality. We will also thank the referee for several useful suggestions.

A Appendix: Proofs
c
Proof of Theorem 3.1 Trying a solution of the form VP (t, y) = γ(t)y+ , using
(2.11) the HJB equation (2.8) becomes on y > 0,

γ ′ (t) a 1
 
c
0 = γ(t)y+ + sup cp − cqXL λP ρa (D) + cr̃ + βc(r − r̃)
γ(t) π y y
!2
1 a 1
− c(1 − c)σP2 − β 2 c(1 − c)σR2
2 y 2
" !c # !
a D

c
+λP E 1 − min{ SP , } − 1 + λR (E[(1 + βSR )+ ] − 1) .
y y +

Define α and δ by a = αy and D = δy. Separation of the above terms give

γ ′ (t)
0= + sup hP (α, δ) + sup hR (β). (A.1)
γ(t) α,δ β

Using the boundary condition γ(T ) = 1, the solution of (A.1) becomes γ(t) =
eηE (T −t) .

First we will prove that for any admissible control π,

J π (t, y) = E t,y [(YTπy,T c


π )+ ] ≤ VP (t, y). (A.2)

To this end let


1
τnπ = Ty,T
π
∧ inf{t : Ytπ ≤ } ∧ inf{t : Ytπ ≥ n},
n
and write V for VP and η ∗ for ηP∗ . For y ∈ (1/n, n), Itô’s formula gives for any
admissible control π,
Z τnπ 4
V (τnπ , Yτπnπ ) = V (t, y) + (Vt (s, Ysπ ) + Aπ V (s, Ysπ ))ds + π π
X
(Mi,τn
π − Mi,t ),
t i=1

32
where
Z t
π
M1,t = σP as Vy (s, Ysπ )dWP,s ,
0
NP,t
π
(V (TP,i , YTπP,i− − min{aTP,i SP,i, DTP,i }) − V (TP,i , YTπP,i− ))
X
M2,t =
i=1
Z tZ ∞
π π
−λP (V (s, Ys− − min{as x, Ds }) − V (s, Ys− ))dFP (x)ds,
0 0
Z t
π
M3,t = σR βs Ysπ Vy (s, Ysπ )dWR,s ,
0
NR,t
π
(V (TR,i , YTπR,i − (1 + βTR,i SR,i )) − V (TR,i , YTπR,i − ))
X
M4,t =
i=1
Z tZ ∞
π π
−λR (V (s, Ys− (1 + βs x)) − V (s, Ys− ))dFR (x)ds
0 −1−

On [0, τnπ ], Vy (s, Ysπ ) = exp{η ∗ (T − s)}c(Ys−


π c−1
) , so that by (2.5), |as Vy (s, Ysπ )| ≤
2Kcn2−c . Therefore M1π is a martingale on [0, τnπ ], hence E[M1,τ π π
π − M1,t ] = 0.
n
π π
Similarly E[M3,τnπ − M3,t ] = 0.
PNR,·
We will let νR be the Poisson random measure associated with i=1 SR,i , i.e.
for s < t and any Borel set A on (−1, ∞)
NR,t
X
νR ((s, t] × A) = 1{SR,i ∈A} . (A.3)
i=NR,s +1

Then
Z τnπ Z ∞
π π π
M4,τn
π = (V (s, Ys− (1 + βs x)) − V (s, Ys− ))(νR (ds, dx) − λR dFR (x)ds)
0 −1−
Z τπ Z ∞
n
η∗ T −η∗ s π c
= e e (Ys− ) ((1 + βs x)c − 1)(νR (ds, dx) − λR dFR (x)ds).
0 −1−

This gives for the quadratic variation process,


NR,τnπ

[M4π , M4π ]τnπ e2η


∗ (T −T
P,i )
(YTπR,i − )2c ((1 + βTR,i SR,i )c+ − 1)2 ,
X
=
i=1

so by (2.5),
 
NR,T
1
E[([M4π , M4π ]τnπ ) 2 ] c η∗ T c c
X
≤K e (1 + n) E  |SR,i | < ∞.
i=1

It follows from the Burkholder-Davis-Gundy inequality that M4π is a martingale on


[0, τnπ ], hence E[M4,τ
π
n
π
π − M4,t ] = 0. Since claims are assumed positive and as ≥ 0,
π π
it follows trivially that E[M2,τ π − M2,t ] = 0. Therefore by the HJB equation (2.8),
n

E t,y [V (τnπ , Yτπnπ )] ≤ V (t, y).

33
Now τnπ → Ty,T
π
as n → ∞, so by Fatou’s lemma,

E t,y [(YTππ )c+ ] = E t,y [V (Ty,T


π
, YTππ )] ≤ lim inf E t,y [V (τnπ , Yτπnπ )] ≤ V (t, y),
y,T y,T n→∞

and (A.2) is proved.


π π
We will now show that for any control of the form π = (αYt− , δYt− , β) with
(α, δ, β) constants,
E t,y [(YTπy,T c
π )+ ] = e
ηP (T −t)
, (A.4)
where ηP = hP (α, δ) + hR (β). Part b then follows from the definition of π ∗ and
(A.2). To prove (A.4), by homogeneity it is sufficient to prove it for t = 0. Since
(YTπy,T c
π )+ = 0 if and only if (1 − min{αSP,i , δ})+ = 0 for some i ≤ NP,T or (1 +

βSR )+ = 0 for some i ≤ NR,T , we get by (3.2),


1
(YTπy,T
π )+
c
= y c exp{(αcp − cqXL λP ρα (δ) − α2 cσP2
2
1
+cr̃ + βc(r − r̃) − β 2 cσR2 )T + αcσP WP,T + βcσR WR,T }
2
NP,T NR,T
(1 − min{αSP,i, δ})c+ (1 + βSR,i )c+ .
Y Y
×
i=1 i=1

Using that
1 2 c2 σ 2 +β 2 c2 σ 2 )T
E[eαcσP WP,T +βσR WR,T ] = e 2 (α P R ,
and that e.g.  
NR,T
c
(1 + βSR,i )c+  = eλR (E[(1+βSR )+ ]−1)T ,
Y
E
i=1

the result follows easily by (3.4) and (3.5), using independence of the different
terms in (YTπy,T c
π )+ .

To prove part c it is enough to prove that J(t, y) → ∞ for a sequence of admissi-


ble controls. Assume first that σP2 = 0, and consider the control πt = (αYt−π
, ∞, 0)
with α constant. Then

J(t, y) ≥ E t,y [(YTπy,T c


π )+ ] = e
hP (α,∞)(T −t)
→ ∞ as α → ∞,

and the result follows from (A.4). If σR2 = 0, consider the control πt = (0, 0, β) with
β a constant. Then hR (β) → ∞ as β → ∞ if r ≥ r̃, and hR (β) → ∞ as β → −∞
if r ≤ r̃. The result follows.

Proof of Proposition 3.1 From (3.10),


∂ 1 1
hP (α, ∞) = hP (α, ∞) + α2 cσP2
∂c c 2
(A.5)
Z 1
c−1 x
−cλP (1 − x) ln(1 − x)F̄P ( )dx,
0 α
∗ ∗
and this is obviously positive when α = α∞ , hence hP (α∞ , ∞) is increasing in c.
34

We will now prove that α∞ is strictly increasing in c as well. We saw that
∂ ∗ ∗ ∗
h (α∞
∂α P
, ∞)= 0 and that α∞ 6= 1/xP,i. Assume first that α∞ is a unique maxi-
mum point of hP (α, ∞). Then from (3.12),
∂ ∂ ∗
hP (α∞ , ∞) = αcσP2 − cλP E[SP (1 − αSP )c−1
+ ln(1 − αSP )+ ] > 0,
∂c ∂α

hence α∞ is increasing in c at such a point. However, as we saw in Examples 3.2

and 3.3, α∞ can jump from one ”hilltop” to another, so we must show that in this
case the jumps will be from left to right. To do so it is sufficient to show that if
∂ ∂
α1 < α2 and hP (α1 , ∞) = hP (α2 , ∞), then ∂c hP (α1 , ∞) < ∂c hP (α2 , ∞). But from
(A.5)
∂ 1
(hP (α2 , ∞) − hP (α1 , ∞)) = (α22 − α12 )cσP2
∂c 2
x x
Z 1  
+cλP (1 − x)c−1 ln(1 − x) F̄P ( ) − F̄P ( ) dx.
0 α1 α2
The integrand in the last integral is always nonnegative, hence the result.
Next we prove part a for hP (αδ̂∗ , δ̂) and αδ̂∗ . If qXL ≤ 1, the result follows easily
from (3.14) and the fact that hP (α0∗ , 0) = (p − qXL λP µP )+ )/(2(1 − c)σP2 ), so assume
that qXL > 1. Since hP (α, δ̂) is strictly concave the the continuity of αδ̂∗ follows.
From (3.14), using that (1 − δ̂)c−1 = qXL ,
∂ 1 1
hP (α, δ̂) = hP (α, δ̂) + α2 cσP2
∂c c 2
(A.6)
Z δ̂/α
c−1
−αcλP (1 − αx) ln(1 − αx)F̄P (x)dx > 0
0

for α = αδ̂∗ , hence hP (αδ̂∗ , δ̂) is strictly increasing in c. Using (3.16) it is fairly easy

to see that ∂α hP (αδ̂∗ , δ̂) is increasing in c, hence αδ̂∗ is increasing in c as well. This
ends the proof of part a.
∗ ∗ ∗
For part b write α∞ (c) for α∞ . We saw in part a that α∞ (c) is increasing in c.

Setting ∂α hP (α(c), ∞) = 0 in (3.12) gives

∗ 1 ∗
α∞ (c) = (p − λP E[SP (1 − α∞ (c)SP )c−1
+ ]). (A.7)
(1 − c)σP2

But E[SP (1 − α∞ (c)SP )c−1 ∗
+ ] → µP as c → 1, hence α∞ (c) → ∞. Now consider
1 2

hP (α∞ , ∞). Replacing one of the α’s in the term 2 α c(1 − c)σP2 with (A.7) gives
for (3.10)

∗ 1 ∗  ∗

hP (α∞ , ∞) = α∞ c p + λP E[SP (1 − α∞ SP )c−1
+ ] ∗
+ λP (E[(1 − α∞ SP )c+ ] − 1).
2

The result now follows since α∞ →∞
∗ ∗
Next let c → 0 and assume that α∞ (c) → α∞ (0) > 1/LP,U . Then from (3.10),
1 x
Z

lim hP (α∞ (c), ∞) = −λP lim c (1 − x)c−1 F̄P ( ∗ (c)
)dx
c→0 c→0 0 α∞
1
≤ −λP F̄P ( ) < 0,
α0
35

a contradiction since hP (0, ∞) = 0. Therefore α∞ (0) ≤ 1/LP,U .
∗ ∗
We now write δ̂(c) and αδ̂(c) for δ̂ and αδ̂ , and assume that p > qXL λP µP . Then
by (3.16)
∗ 1 ∗
αδ̂(c) = (p − qXL λP µP − λP φ(αδ̂(c) , c)), (A.8)
(1 − c)σP2
where Z δ̂(c)/α  
φ(α, c) = x (1 − αx)c−1 − qXL dFP (x).
0
c−1
For αx < δ̂(c), (1 − αx) < qXL , hence φ(α, c) ≤ 0. Therefore it is clear that

in order to satisfy (A.8), αδ̂(c) must converge towards infinity as c → 1. Now let

p < qXL λP µP and assume that αδ̂(c) → ∞ as c → 1. Then since φ(α, c) → 0 as

α → ∞, for c large enough, p < qXL λP µP + λP φ(αδ̂(c) , c). But then by (A.8),

αδ̂(c) < 0 which is a contradiction.

Finally, let c → 0 and assume that αδ̂(c) → 0. Then by the last equation in
(3.16) and dominated convergence,

1 ∂ ∗
0= hP (αδ̂(c) , δ̂(c)) → p − λP µP > 0,
c ∂α
a contradiction. This ends the proof of part b.
For part c, assume that c → 0 and that qXL > 1, the case with qXL ≤ 1 is
similar but simpler. Also if qXL ≤ 1, reinsurance is cheaper, so if it is optimal to
use it for qXL > 1, it is certainly optimal to use it for qXL ≤ 1. We will divide the
∗ ∗ ∗
proof into several cases. Again we write α∞ (c) for α∞ and δ̂(c) and αδ̂(c) for δ̂ and
∗ ∗
αδ̂ to indicate the c-dependence. Due to monotonocity in c, the limits α∞ (0) and
∗ ∗
αδ̂(0) will exist and α∞ (0) ≤ 1/LP,U by part b. Note also that δ̂(c) is increasing in
−1
c, taking the value δ0 = 1 − qXL when c = 0.
∗ ∗ ∗ ∗ ∗
Case 1, LP,U = ∞: By part b, α∞ (c) → 0 and αδ̂(c) → αδ̂(0) > 0, hence αδ̂(c) > α∞
for c small enough. Furthermore, from (3.10) and (3.14), and using that (1−x)c−1 −
qXL is increasing in x for x ≤ 1, and that (1 − δ̂(c))c−1 = qXL ,

def
v(α, c) = hP (α, δ̂(c)) − hP (α, ∞)
!
1 x ∞ x
Z   Z
c−1
= cλP (1 − x) − qXL F̄P ( )dx − qXL F̄P ( )dx
δ̂(c) α 1 α
1 x
 Z ∞ 
> qXL λP (1 − c(1 − δ̂(c)))F̄P ( ) − c F̄P ( )dx ,
α 1 α

and this is positive for c small enough. Hence for such c, hP (α∞ (c), ∞) <
∗ ∗
hP (α∞ (c), δ̂(c)) ≤ hP (αδ̂(c) , δ̂(c)).
∗ ∗
Case 2, LP,U < ∞ and α∞ (0) < δ0 /LP,U : For all c sufficiently small, α∞ (c) <
∂ ∗ ∂ ∗
δ̂(c)/LP,U , and from (3.12) and (3.16), ∂α hP (α∞ (c), δ̂(c)) = ∂α hP (α∞ (c), ∞) = 0,
∗ ∗ ∗ ∗
giving αδ̂(c) = α∞ (c). Also from (3.10) and (3.14), hP (αδ̂(c) , δ̂(c)) = hP (α∞ (c), ∞).

36
∗ ∗
Case 3, LP,U < ∞ and δ0 /LP,U ≤ α∞ (0) ≤ 1/LP,U : Assume first that α∞ (0) <

1/LP,U . For all c sufficiently small, α∞ (c) < 1/LP,U as well, so for such c
1 ∂ ∗
0 = hP (α∞ (c), ∞)
c ∂α

= p − α∞ (c)(1 − c)σP2 − λP E[SP (1 − α∞

(c)SP )c−1 ]

< p − α∞ (c)(1 − c)σP2 − λP E[SP (1 − min{α∞∗
(c)SP , δ̂(c))c−1 ]
1 ∂ ∗
= hP (α∞ (c), δ̂(c))
c ∂α
∗ ∗ ∗
hence αδ̂(c) > α∞ (c). If α∞ (0) = 1/LP,U , it follows by Fatou’s lemma that

lim inf E[SP (1 − α∞ (c)SP )c−1 ∗ −1
+ ] ≥ E[SP (1 − α∞ (0)SP ) ]
c→0

≥ E[SP (1 − min{α∞ (0)SP , δ̂(c))−1 ] + ε

> E[SP (1 − min{α∞ (c)SP , δ̂(c))c−1]
∗ ∗
for some ε > 0 and c sufficiently small. Therefore αδ̂(c) > α∞ (c) in this case as

well. Furthermore, when δ0 /LP,U ≤ α∞ (0) < 1/LP,U and c small enough

∗ ∗ 1 2
hP (α∞ (c), ∞) = α∞ (c)cp − α∞ (c)c(1 − c)σP2 + λP (E[(1 − α∞

(c)SP )c ] − 1)
2
∗ 1 2
< α∞ (c)cp − α∞ (c)c(1 − c)σP2 + λP (E[(1 − min{α∞∗
(c)SP , δ̂(c))c ] − 1)
2

= hP (α∞ (c), δ̂(c)),
∗ ∗ ∗
hence hP (α∞ (c), ∞) < hP (αδ̂(c) , δ̂(c)). If α∞ (0) = 1/LP,U , dominated convergence
will give the same result.
Finally let c → 1. Since qXL ≤ 1 gives highest hP (αδ̂∗ , δ̂), we shall consider this
case only, but qXL > 1 is not much different. Now δ̂ = 0, hence by (3.12) and
(3.13),

∗ 1  
α∞ − α0∗ = q λ µ
XL P P − λ P E[S P (1 − α ∗
S
∞ P +) c−1
] .
(1 − c)σP2

When c → 1 it follows from part b that α∞ → ∞, hence

∗ 1 1
E[SP (1 − α∞ SP )c−1
+ ] = ∗

E[(1 − α∞ SP )c−1 ∗ c
+ ] − ∗ E[(1 − α∞ SP )+ ] → 0
α∞ α∞

as c → 1. Therefore α∞ − α0∗ → qXL λP µP /((1 − c)σP2 ) as c → 1. Furthermore,

solving ∂α hP (α, ∞) = 0 in (3.12) and inserting into (3.10) in conjunction with
(3.13) gives

∗ 1 ∗  
hP (α∞ , ∞) − hP (α0∗ , 0) = α∞ ∗
p + λP E[SP (1 − α∞ SP )c−1
+ ]
2
1
∗ c
+λP (E[(1 − α∞ SP )+ ] − 1) − α0∗ (p − qXL λP µP ) → ∞
2
as c → 1. This ends the proof of the proposition.
37
Proof of Proposition 3.3 We start with the strict increase in p. By Proposition

3.2, both α∞ and αδ̂∗ are strictly increasing in p, so what is needed to prove is
that if hP (α∗ , δ ∗ ) moves from hP (αδ̂ , δ̂) to hP (α∞∗
, ∞), then at the changepoint p0 ,
α∞ ≥ αδ̂ , or if it moves from hP (α∞ , ∞) to hP (αδ̂∗ , δ̂), then at the changepoint
∗ ∗ ∗


α∞ ≤ αδ̂∗ . So assume that hP (α∗ , δ ∗ ) moves from hP (αδ̂∗ , δ̂) to hP (α∞ ∗
, ∞) at p0 .
Then at this point
hP (αδ̂∗ , δ̂) = hP (α∞

, ∞) (A.9)
and
∂ ∂
hP (αδ̂∗ , δ̂) < ∗
hP (α∞ , ∞). (A.10)
∂p ∂p
However, by (A.10)
∂ ∗
0< (hP (α∞ , ∞) − hP (αδ̂∗ , δ̂)) = c(α∞

− αδ̂∗ ), (A.11)
∂p

and therefore α∞ > αδ̂∗ at p0 . If hP (α∗ , δ ∗ ) moves from hP (α∞∗
, ∞) to hP (αδ̂∗ , δ̂),
just reverse the inequalities in (A.10) and (A.11). The proofs for qXL and σP2 are

basically identical, just take into account that α∞ and αδ̂∗ are decreasing in these
parameters.
Now to λP , and assume that hP (α∗ , δ ∗ ) moves from hP (αδ̂∗ , δ̂) to hP (α∞∗
, ∞) at

a point λP,0. Then since α∞ and αδ̂∗ is decreasing in λP , we must prove that at the

changepoint λP,0 , α∞ ≤ αδ̂ . Assume that qXL > 1, the proof with qXL < 1 is the
same. Using (3.10) and (3.14) we find that
∂ ∗
(hP (α∞ , ∞) − hP (αδ̂∗ , δ̂))
∂λP
1 1 ∗2
 
∗ ∗
= hP (α∞ , ∞) − α∞ cp + α∞ c(1 − c)σP2
λP,0 2
1 1 2
 
− hP (αδ̂∗ , δ̂) − αδ̂∗ cp + αδ̂∗ c(1 − c)σP2 .
λP,0 2
But using (A.9) this is positive provided
1 2 1 ∗2
αδ̂∗ p − αδ̂∗ (1 − c)σP2 > α∞∗
p − α∞ (1 − c)σP2 .
2 2
Now the function f (α) = αp − 0.5α (1 − c)σP is increasing for α ≤ p/((1 − c)σP2 ),
2 2

and from (A.7) and (A.8) it is clear that both α∞ and αδ̂∗ are no larger than this
value, hence at λP,0 , ∂λ∂P (hP (α∞

, ∞) − hP (αδ̂∗ , δ̂)) > 0 if and only if αδ̂∗ ≥ α∞

and
we are done.

Based on Examples 3.2 and 3.3 it is natural to conjecture that α∗ is increasing


in c as well. However, at a changepoint c0 using (A.5), (A.6) and (A.9) we get
∂ ∗ 1 ∗2 2
(hP (α∞ , ∞) − hP (αδ̂∗ , δ̂)) = (α∞ − αδ̂∗ )
∂c 2 !
x x
Z δ̂
c0 −1
−c0 λP (1 − x) ln(1 − x) F̄P ( ∗ ) − F̄P ( ∗ ) dx
0 α∞ αδ̂
Z 1
x
−c0 λP (1 − x)c0 −1 ln(1 − x)F̄P ( ∗ )dx,
δ̂ α∞
38

and due to the last term here, this can in principle be positive even if α∞ < αδ̂∗ .

For the same reason, δ may not be increasing in c.

Proof of Proposition 3.4 The proof is much the same as that for hP (α∞ , ∞)

and α∞ given in Proposition 3.1, and we shall only give a partial proof. We start
by showing that β ∗ is increasing in c. From (3.21)

∂ ′ 1 
hR (β) = hR (β) + βσR2 + λR E[SR (1 + βSR )c−1
+ ln(1 + βSR )+ ] .
∂c c
∂ ′
For β ≥ 0, SR (1 + βSR )c−1 ∗
+ ln(1 + βSR )+ ≥ 0, hence ∂c hR (β ) > 0. Furthermore, if
0 ≤ β1 < β2 and hR (β1 ) = hR (β2 ), some calculations using (3.18) show that

∂ 1 2
(hR (β2 ) − hR (β1 )) = (β − β12 )cσR2
∂c 2 2 !
∞ x x
Z
c−1
+cλR (1 + x) ln(1 + x) F̄R ( ) − F̄R ( ) dx
−1 β2 β1

and the integrand in the last integral is always nonnegative. Therefore, using the
same arguments as in Proposition 3.1, β ∗ is increasing in c.
Now let c → 0. We will end the proof by showing the last statement of the
proposition. From (3.18) we see that if F̄R (−1/β) < 1, then because of the last
term −λR , hR (β) becomes negative as c → 0. Therefore it is necessary that
−1/β0∗ ≤ LR,L , or equivalently that β0∗ ≤ −1/LR,L if LR,L < 0. Furthermore,
from (3.21) and what we just proved, for c small enough
1
β∗ = (r − r̃ − λR E[SR (1 + β ∗ SR )c−1 ]). (A.12)
(1 − c)σR2
1
Assume that β ∗ → 0 as c → 0. Then from (2.2) and (A.12), 0 = 2 (r − r̃ −λR µR )
σR
>
0, a contradiction, hence β0∗ > 0.

Proof of Theorem 4.1 For part a we start with gP . Using (3.7) and calculations

similar to those in (3.8), we find that ∂δ gP (α, δ) = λP F̄ ( αδ )(qXL − (1 − δ)−1
+ ), hence
−1
δ ∗ = (1 − qXL )+ . This gives that

1 δ∗ /α
Z
gP (α, δ ∗) = αp − α2 σP2 − qXL αλP µP − αλP ((1 − αx)−1 − qXL )F̄P (x)dx.
2 0
(A.13)
∗ ∗
If qXL ≤ 1, gP (α, δ ) = gP (α, 0) is concave and is maximal for α given by (3.13)
with c = 0. This holds if either σP2 > 0 or p ≤ qXL λP µP , in the latter case α∗ = 0.
Assume that qXL > 1. Then as in (3.16),
" #
∂ SP
gP (α, δ ∗ ) = p − ασP2 − λP E . (A.14)
∂α 1 − min{αSP , δ ∗ }

This is decreasing in α, hence gP (α, δ ∗ ) is concave in α. Also if σP2 = 0, gP (α, δ ∗ ) →


p − qXL λP µP as α → ∞, hence α∗ exists and is finite if p < qXL λP µP .

39
It is clear that gR is concave on IR and minus infinity outside IR . On IR we have
" #
SR
gR′ (β) = r − r̃ − βσR2 + λR E .
1 + βSR
In particular gR′ (0) = r − r̃ + λR µR > 0, so that β ∗ , if it exists, is positive. It will
always exist if σR2 > 0. If LR,L < 0, β ∗ will exist also when σR2 = 0. If LR,L ≥ 0
and σR2 = 0, then β ∗ will exist if and only if r < r̃, which is necessary by the no
arbitrage condition (2.3). Therefore β ∗ will always exist and part a is proved.
Now part b. Trying a solution of the form VL (t, y) = γ(t) + ln y, the HJB
equation (2.8) becomes
γ ′ (t) + ηL∗ = 0. (A.15)
Using the condition γ(T ) = 0 and solving (A.15) yields the solution (4.4).
We will now prove that
J π (t, y) = E t,y [ln YTπ ] ≤ VL (t, y) (A.16)
for all admissible controls π. Note that Tyπ = ∞, and if for some control π,

P (Ytπ ≤ 0 for some t ≤ T ) > 0, then because of the model assumptions we must
have that P (YTπ ≤ 0) > 0 as well yielding E[U(YTπ )] = −∞. Therefore we can rule
out such controls and only consider controls where P (Tyπ > T ) = 1. In particular it
means that all the (1 −min{αTP,i SP,i, δTP,i }) and (1 +βTR,i SR,i ) of (3.1) are positive.
Writing V for VL , we therefore get
T 1
Z
V (T, YTπ ) = V (t, y) + (−ηL∗ + αs p − qXL λP ραs (δs ) − αs2 σP2
t 2
1 2 2
Z T Z T
+r̃ + βs (r − r̃) − βs σR )ds + σP αs dWP,s + σR βs dWR,s
2 t t
NP,T NR,T
X X
+ ln(1 − min{αTP,i SP,i, δTP,i }) + ln(1 + βTR,i SR,i ).
i=NP,t +1 i=NR,t +1
(A.17)
For E[V (T, YTπ )]
= E[ln YTπ ] to be well defined, we must have (Problem 3.4.11 in
Karatzas and Shreve, 1988),
"Z #
T
E (σP2 αs2 + σR2 βs2 )ds < ∞, (A.18)
0

and for E[ln YTπ ] > −∞ it is necessary that


 
NP,T
X
E ln(1 − min{αTP,i SP,i, δTP,i })
i=1 (A.19)
"Z #
T Z ∞
= λP E ln(1 − min{αs x, δs })dFP (x)ds > −∞,
0 0
 
NR,T
X
E ln(1 + βTR,i SR,i )
i=1 (A.20)
"Z #
T Z ∞
= λR E ln(1 + βs x)λP dFP (x)ds > −∞.
0 −1−

40
Now (A.17) can be written as
Z T
V (T, YTπ ) = V (t, y) + (Vt (s, Ysπ ) + Aπ V (s, Ysπ ))ds + (MTπ − Mtπ )
t (A.21)
≤ V (t, y) + (MTπ − Mtπ ),

where
Z t Z t
Mtπ = σP αs dWP,s + σR βs dWR,s
0 0
Z tZ ∞
+ ln(1 − min{αs x, δs })(νP (ds, dx) − λP dFP (x)ds), (A.22)
0 0
Z tZ ∞
+ ln(1 + βs x)(νR (ds, dx) − λR dFR (x)ds)
0 −1−

PNP,· PNR,·
and νP and νR are the random measures associated with i=1 SP,i and i=1 SR,i
respectively, see (A.3). Its quadratic variation equals
Z t
[M, M]t = (σP2 αs2 + σR2 βs2 )ds
0
NP,t
(ln(1 − min{αTP,i SP,i , δTP,i }))2
X
+
i=1
NR,t
(ln(1 + βTR,i SR,i ))2
X
+
i=1

Therefore, by the Burkholder-Davis-Gundy inequality for some postive constant


c1 ,
  !1 
h i Z T 2
E sups≤T |Msπ | ≤ c1 E  (σP2 αs2 + σR2 βs2 )ds 
0
"Z #
T Z ∞
+λP E | ln(1 − min{αs x, δs })|dFP (x)ds (A.23)
0 0
"Z #!
T Z ∞
+ λR E | ln(1 + βs x)|dFR (x)ds <∞
0 −1−

by (A.18)-(A.20). Consequently, M π is a martingale on [0, T ], so from (A.21)

E t,y [V (T, YTπ )] = E t,y [ln YTπ ] ≤ V (t, y),

proving (A.16).
π π
Now let πt = (αYt− , δYt− , β) with (α, δ, β) constant and chosen so that P (YTπ >
0) = 1. Then it follows from (3.2) that J π (t, y) = E t,y [ln YTπ ] = ηL (T −t)+ln y with
ηL = hP (α, δ) + hR (β). Setting π = π ∗ gives in particular that J π (t, y) = VL (t, y).

This ends the proof of part b.


For part c, note that with σP2 = 0 and p = qXL λP µP ,
δ∗ x
Z

gP (α, δ ) = λP (qXL − (1 − x)−1 )F̄P ( )dx.
0 α
41
If qXL ≤ 1, δ ∗ = 0 and then α∗ = 0 is optimal. If qXL > 1, gP (α, δ ∗ ) is increas-
ing in α, hence there is no finite optimal α. However by monotone convergence
gP (α, δ ∗ ) → λP (qXL − 1 − ln qXL ) as α → ∞.

Proof of Proposition 4.1 It follows from Proposition 3.1 that for c sufficiently
small, δ ∗ (c) = δ̂(c) and α∗ (c) = αδ̂(c)

. Therefore (4.7) follows from (4.5) and (4.6).
From this it is clear that (4.9) follows from (4.8). Using Hölder’s inequality, it is
easy to prove that for numbers ai and bi and q ≥ 1,
n q    
Y n n i−1 n
≤ nq−1 |bi |q  |ai − bi |q  |ai |q  .
Y X Y Y
ai − bi (A.24)



i=1 i=1 i=1 j=1 j=i+1

Setting q = 1, it follows from this inequality and (3.2) that


π ∗ (c)
− Ytπ | → 0 a.s. as c → 0.

sup |Yt
t≤T

We will now prove Lq convergence. For given ε > 0 we have for c sufficiently
small
|α∗ (c) − α∗ | < ε, |δ ∗ (c) − δ ∗ | < ε and |β ∗ (c) − β ∗ | < ε.
Hence for such c,
sup eqα
∗ (c)σ W
≤ sup eq(α
∗ +ε)σ W
P P,t P P,t
,
t≤T t≤T

sup eqβ
∗ (c)σ W
≤ sup eq(β
∗ +ε)σ W
R R,t R R,t
,
t≤T t≤T
NP,t
|1 − min{α∗ (c)SP,i, δ ∗ (c)}|q ≤ 1,
Y
sup
t≤T i=1
NR,t NR,T
|(1 + β ∗ (c)SR,i )|q ≤ |(1 + |β| + ε)SR,i )|q .
Y Y
sup
t≤T i=1 i=1

Since all the terms on the right hand side here are integrable, by independence and
(3.2) and (A.24) it is sufficient to prove
" #
α∗ σP WP,t q

E sup eα (c)σP WP,t −e → 0,


t≤T
" #
β ∗ σR WR,t q

E sup eβ (c)σR WR,t −e → 0,


t≤T
 q 
NP,t NP,t
∗ ∗ ∗ ∗ 
Y Y
E sup (1 − min{α (c)SP,i , δ (c)}) −
 (1 − min{α SP,i , δ }) → 0,
t≤T i=1 i=1
 q 
NR,t NR,t
E sup (1 + β ∗ (c)SR,i ) − (1 + β ∗ SR,i )  → 0
Y Y
t≤T i=1 i=1
(A.25)
as c → 0. To prove the first convergence in (A.25), note that for c small enough, it
follows as above
q
sup eα
∗ (c)σ W
− eα
∗σ W
≤ sup eq(α
∗ +ε)σ W
P P,t P P,t P P,t

t≤T t≤T

42
which is integrable. The result then follows by dominated convergence. The second
convergence is similar, and the third is similar to the fourth which we now prove.
Using (A.24),
q
NR,t NR,t
∗ ∗
Y Y
sup (1 + β (c)SR,i ) −
(1 + β SR,i )
t≤T i=1 i=1
q
NR,T NR,T
∗ ∗
Y Y
= (1 + β (c)SR,i ) − (1 + β SR,i )
i=1 i=1
NR,T
∗ ∗ q q−1
|SR,i |q (1 + |β ∗ | + ε)|SR,j |)q .
X Y
≤ |β (c) − β | NR,T
i=1 j6=i

Conditioning on NR,T it is not hard to see that this is integrable, and dominated
convergence gives the result.
The final convergence is easily proved.

Proof of Theorem 4.2 For this problem the HJB equation is

1 + inf Aπ V (y) = 0, (A.26)


π

with boundary conditions

V (y) = ∞ for y ≤ 0 and V (ȳ) = 0.

As before we let αy = a and δy = D. Trying a solution of the form



V (y) = κ ln = κ(ln ȳ − ln y)
y
brings (A.26) into the form

1 + κ inf (−gP (α, δ) − gR (β)) = 1 − κ sup(gP (α, δ) + gR (β)) = 1 − κηL∗ = 0.


π π

We note first that only controls π for which τȳπ < Tyπ are relevant for the problem,
so therefore 0 < Ytπ ≤ ȳ for t ∈ [0, Ty,ȳ
π
]. Consequently we can use Itô’s formula
together with (A.26) to get
π π π
V (y) ≤ V (Yt∧T π ) + t ∧ Ty,ȳ + κMt∧T π ,
y,ȳ y,ȳ
(A.27)

where M π is given by (A.22). If π = π ∗ , there is equality in (A.27). For the control


π ∗ , (3.2) gives that ln Y π is a Lévy process with positive drift ηL∗ , hence E[Ty,ȳ
π∗

]<
π
∞. Therefore we can restrict our attention to controls π so that E[Ty,ȳ ] < ∞.
By the Burkholder-Davis-Gundy inequality, for such controls E[sups≤Ty,π ȳ |Msπ |] is
π
dominated by the right hand side of (A.23) with T replaced by Ty,ȳ . For π = π ∗ ,
we get for some positive c2 ,
 
1
π∗ π 2 π ∗ ∗
E  sup |Ms | ≤ c2 (E[(Ty,ȳ ) ] + E[Ty,ȳ ]) < ∞,
π
s≤Ty, ȳ

43
π ∗ π ∗
hence Mt∧T π ∗ is a uniformly integrable martingale and therefore E[MT π ∗ ] = 0.
y,ȳ y,ȳ
So by first letting t → ∞ and then taking expectations in (A.27), it follows that
π∗
V (y) = E[Ty,ȳ ]. On the other hand, if for some control π the right side of (A.23)
π
with Ty,ȳ instead of T is infinite, it follows as in the proof of Theorem 4.1 that
E[ln YTπy,π ȳ ] = −∞, a contradiction since YTπy,π ȳ = ȳ. Therefore, for any ”good”
π π
control π that makes E[Ty,ȳ ] < ∞, Mt∧T π
y,ȳ
is a uniformly integrable martingale
π
and thus E[MTy,π ȳ ] = 0. Again letting t → ∞ and then taking expectations in
π
(A.27) yields V (y) ≤ E[Ty,ȳ ], and the theorem is proved.

Proof of Theorem 5.1 For part a we start with kP . Using (3.7) and calculations
similar to those in (3.8) give
∂ D
kP (a, D) = cλP F̄P ( )(qXL − ecD )
∂D a
using left and right derivatives of F̄P it is has a jump at D/a. Therefore D ∗ =
1
c
(ln qXL )+ . In particular D ∗ is decreasing in c and increasing in qXL . If qXL ≤ 1,
kP (a, D ∗ ) = kP (a, 0) is concave and is maximal for
((1 − qP )p − qXL λP µP )+
a∗ = . (A.28)
cσP2
For a∗ to exist it is necessary that either σP2 > 0 or that (1 − qP )p ≤ qXL λP µP , in
the latter case a∗ = 0.
Assume now that qXL > 1 and that FP is given by (3.9). Then as in (A.13) we
get
1
kP (a, D) = qP cp + (1 − qP )acp − cqXL aλP µP − a2 c2 σP2
2
Z D/a
−acλP (ecax − qXL )F̄P (x)dx.
0

Since ecD = qXL , kP (a, D ∗ ) is differentiable in a and some calculations give


∂ h i
kP (a, D ∗ ) = (1 − qP )cp − ac2 σP2 − cλP E SP ec min{aSP ,D } .

∂a

It is easily seen that ∂a kP (a, D ∗ ) is decreasing, hence kP (a, D ∗ ) is concave in a,
and the optimal a will exist if and only if σP2 > 0 or (1 − qP )p < qXL λP µP .

It is clear that kR is concave, and



kR (b) = c(r − bcσR2 + λR E[SR e−cbSR ]).

In particular kR (0) = c(r + λR µR ) > 0, and therefore as in the proof of Theorem

4.1, b exists and is positive. This ends part a.
For part b, setting Ǔ (y) = −U(y), the problem is equivalent to minimizing

Jˇπ (t, y) = E t,y [Ǔ (YTπ )].

The HJB equation for this problem is

inf (Vt (t, y) + Aπ V (t, y)) = 0 (A.29)


π
44
with boundary condition V (T, y) = e−cy . Trying a solution of the form V (t, y) =
γ(t)e−cy , the HJB equation (A.29) becomes

γ ′ (t) γ ′ (t)
+ inf (−kP (a, D) − kR (b)) = − sup kP (a, D) − sup kR (b) = 0. (A.30)
γ(t) π γ(t) a,D b

Using the boundary condition γ(T ) = 1, the solution of (A.30) gives V (t, y) =
e−ηE (T −t) e−cy . By Itô’s formula

dV (t, Ytπ ) = V (t, Yt−


π
)dZtπ , (A.31)

where
Z t Z t Z t
Ztπ = hπs ds − cσP as dWP,s − cσR bs dWR,s
0 0 0
NP,t   NR,t  
ec min{aTP,i SP,i ,DTP,i } − 1 + cbTP,i SR,i
X X
+ e −1
i=1 i=1

with
1 1
hπs = −ηE∗ − as cp − (1 − as )qP cp + cqXL λP ρas (Ds ) + a2s c2 σP2 − bs cr + b2s c2 σR2 .
2 2
Now assume first that π ∈ Π0 . Let τnπ = inf s≥t {as > n, |bs | > n, |Ys | > n} ∧ T .
Then V (s, Ytπ ) is bounded on [t, τnπ ), and τnπ → T as n → ∞. From (A.31)
"Z #
τnπ
E[V (τnπ , Yτπnπ )] = V (t, y) + E t,y V (s, Ysπ )hπs ds
t
"Z #
τnπ Z ∞  
t,y π c min{as x,Ds }
+λP E V (s, Ys− ) e − 1 dFP (x)ds
t 0
"Z #
τnπ Z ∞  
t,y π −cbs x
+λR E V (s, Ys− ) e − 1 dFR (x)ds .
t −1−

Note that hπs is bounded on [t, τnπ ]. Therefore by (5.2),


"Z #
τnπ Z ∞  
t,y π c min{as x,Ds }
λP E V (s, Ys− ) e − 1 dFP (x)ds
t 0
"Z
τnπ Z ∞  
# (A.32)
t,y π −cbs x
+λR E V (s, Ys− ) e − 1 dFR (x)ds < ∞.
t −1−

We can thus write, using (A.31) and (A.29),


Z τnπ Z τnπ
V (τnπ , Yτπnπ ) = V (t, y) + (Vt (s, Ysπ ) +A Vπ
(s, Ysπ ))ds + π
V (s, Ys− )dMsπ (c)
t t
Z τnπ
π
≥ V (t, y) + V (s, Ys− )dMsπ (c),
t
(A.33)

45
where Vt (s, Ysπ ) + Aπ V (s, Ysπ ) = V (s, Ysπ )ξsπ (c) with
1 1
ξsπ (c) = hπs + a2s c2 σP2 − bs cr + b2s c2 σR2
2 2
Z ∞  
+λP ec min{as x,Ds } − 1 dFP (x) (A.34)
0
Z ∞  
+λR e−cbs x − 1 dFR (x)
−1−

and
Z t Z t
Mtπ (c) = −cσP as dWP,s − cσR bs dWR,s
0 0
Z tZ ∞  
+ ec min{as x,Ds} − 1 (νP (ds, dx) − λP FP (x)ds) (A.35)
0 0
Z tZ ∞  
+ e−cbs x − 1 (νR (ds, dx) − λR FR (x)ds),
0 −1−

where νP and νR are defined in the proof of Theorem 4.1. By the Burkholder-
Davis-Gundy inequality and (A.32),
E[V (τnπ , Yτπnπ )] ≥ V (t, y).
Letting n → ∞, we get by assumption and the dominated convergence theorem,
E[V (T, YTπ )] ≥ V (t, y).
π∗
If π = π ∗ , the proof that V (t, y) = E t,y [e−cYT ] can be easily proved by direct
calculations, or using that in this case there is equality in (A.33) and then use the
Burkholder-Davis-Gundy inequality to prove that the local martingale there is in
fact a uniformly integrable martingale. Hence VE (t, y) = −V (t, y) is optimal for
π ∈ Π0 .
Now assume that π ∈ ΠB . We start by proving that for any π ∈ ΠB so that
E[V (T, YTπ )] < ∞,
"Z #
T Z ∞
λP E V (s, Ysπ ) c min{as x,Ds }
e − 1 dFP (x)ds
0 0
"Z # (A.36)
T Z ∞
+λR E V (s, Ysπ ) −cbs x
e − 1 dFR (x)ds < ∞.
0 −1−

Indeed, since π ∈ ΠB , the same boundedness applies for π after a stopping time τ
as before τ . Furthermore, for any control π, the control π̂ defined as π̂t = πt for
t ≤ τ and π̂t = 0 for t > τ , τ a stopping time, is also in ΠB . Hence for all stopping
times τ ≤ T ,
inf inf P (YTπ − Yτπ ≤ 0) = inf P (YTπ ≤ 0|Y0π = 0) ≥ c0
π τ π

for some c0 > 0. Therefore, in order for E[V (T, YTπ )] to be finite it is necessary
that  

|△V (s, Ysπ )| < ∞.


X
E
s≤T

46
But this condition is just (A.36). To continue the proof when π ∈ ΠB , we may
π
assume that E[e−cYT ] < ∞, otherwise the result is trivially true. By the above
we need to prove (5.2). To this end let τnπ = inf{t : |Ytπ | ≥ n} ∧ T and set
K = maxt≤T {at , |bt |}. Itô’s formula gives

1 1
Z t∧τnπ 1 1
V (t ∧
2 τnπ , Yt∧τ
π
n
π) = V (0, y) + 2 (Vt 2 (s, Ysπ ) + Aπ V 2 (s, Ysπ ))ds
0
Z t∧τnπ 1 c
π
+ V (s, Ys−
2 )dMsπ ( )
0 2
1 1 1
where Vt 2 (s, Ysπ ) + Aπ V 2 (s, Ysπ ) = V 2 (s, Ysπ )ξsπ ( 2c ), with ξsπ ( 2c ) and Mtπ ( 2c ) given as
in (A.34) and (A.35) respectively. Taking squares we get
!2
Z t∧τnπ 1 c

c
2
V (t∧τnπ , Yt∧τ
π
π) ≤ 3V (0, y)+3 V 2 (s, Ysπ )ξsπ ( ) π
+3 M̃t∧τ π( ) , (A.37)
n
0 2 n
2

where
c c t
Z
1
M̃tπ ( )
π
V 2 (s, Ys− )dMsπ ( ). =
2 0 2
Using the boundedness of the controls together with Jensens’s inequality and the
nonnegativity of V yields for some K1 > 0,
!2
Z t∧τnπ 1 c
Z t∧τnπ
V 2 (s, Ysπ )ξsπ ( )ds ≤ K1 T V (s, Ysπ )ds
0 2 0
Z T Z ∞  c
2
+3λ2P T V (s, Ysπ ) e 2
min{as x,Ds }
−1 dFP (x)ds
0 0
Z T Z ∞  c
2
+3λ2R T V (s, Ysπ ) e− 2 bs x − 1 dFR (x)ds.
0 −1−

Now " 2 # "Z #


c t∧τnπ
π
E M̃t∧τ π( ) ≤ K2 E V (s, Ysπ )ds + K3 (A.38)
n
2 0

with
c2 2
K2 = K T (σP2 + σR2 )
4
and
"Z #
T Z ∞  c
2
K3 = λ P E V (s, Ysπ ) e 2
min{as x,Ds }
−1 dFP (x)ds
0 0
"Z #
T Z ∞ 2
− 2c bs x

λR E V (s, Ysπ ) e −1 dFR (x)ds .
0 −1−

It follows by (A.36) that K3 < ∞. Set V ∗π (t) = sups≤t V (s, Ysπ ). Then taking
supremum in (A.37) together with Doob’s inequality using (A.38) and monotone
convergence theorem yields
Z t∧τnπ
E[V ∗π (t∧τnπ )] ≤ 3V (0, y)+(9(λP +λR )T +12)K3 +(3K1 T +12K2) E[V ∗π (s)]ds.
0

47
Letting n → ∞ this becomes by monotone convergence,
Z t
E[V ∗π (t)] ≤ 3V (0, y) + (9(λP + λR )T + 12)K3 + (3K1 T + 12K2 ) E[V ∗π (s)]ds.
0

Now (5.2) follows from Gronwall’s inequality. This ends the proof of part b. Part
c is proved as part c in Theorem 4.1.

Proof of Theorem 5.2 Trying a solution of the form

Ṽ (t, y) = γ̃(t)e−cψ(t)y ,

the HJB equation (A.29) now becomes (as in (A.30))

γ̃ ′ (t)
0 = − cψ ′ (t)y − cψ(t)qP p − cψ(t)r̃y
γ̃(t) (A.39)
−sup k̃P (aψ(t), Dψ(t)) − sup k̃R (bψ(t)).
a,D b

Comparing (5.5) with (5.3) and (5.6) with (5.4) it is seen that the analysis in
Theorem 5.1 for kP and kR is equally valid for k̃P and k̃R . In fact, the values
(a∗ , D ∗ ) that maximizes (5.3) also maximizes (5.5), and the only difference between
(5.4) and (5.6) is that r in (5.4) is replaced by r − r̃.
Separating the terms depending on y in (A.39) gives the equations

ψ ′ (t) + r̃ψ(t) = 0 (A.40)

and
γ̃ ′ (t)
− cψ(t)qP p − η̃E∗ = 0. (A.41)
γ̃(t)
The boundary condition Ṽ (T, y) = e−cy implies that ψ(T ) = γ̃(T ) = 1. Solving
(A.40) with this boundary condition yields

ψ(t) = er̃(T −t) . (A.42)

Inserting this solution into (A.41) and using the boundary condition yields
cqP p r̃(T −t)
 
γ̃(t) = exp −η̃E∗ (T − t) − (e − 1) .

Also the corresponding controls become ã∗t = e−r̃(T −t) ã∗ , D̃t∗ = e−r̃(T −t) D̃ ∗ and
b̃∗t = e−r̃(T −t) b̃∗ .
The rest of the proof is similar to that of Theorem 5.1 and is omitted.

48
B References
Asmussen, S. (2000). Ruin Probabilities. World Scientific Publishing Co., River
Edge N.J.

Asmussen, S. and M. Taksar (1997). Controlled diffusion models for optimal


dividend pay-out. Insurance: Mathematics and Economics, 20, 1-15.

Asmussen, S., B. Højgaard, and M. Taksar (2000). Optimal risk control and
dividend distribution policies. Example of excess-of loss reinsurance for an
insurance corporation. Finance and Stochastics, 4, 299-324.

Browne, S. (1995). Optimal investment policies for a firm with random risk pro-
cess: Exponential utility and minimizing the probability of ruin. Mathematics
of Operations Research, 20, 937-958.

Bühlmann, H. (1970). Mathematical Methods in Risk Theory. Springer, Berlin.

Dayananda, P.W.A. (1970). Optimal reinsurance. Journal of Applied Proba-


bility, 7, 134-156.

Fleming, W.H. and H.M. Soner (1993). Controlled Markov processes and vis-
cocity solutions. Springer, New York.

J. Gaier and P. Grandits (2002). Ruin probabilities in the presence of regu-


larly varying tails and optimal investment. Insurance: Mathematics and
Economics, 30, 211-217.

J. Gaier and P. Grandits (2003). Ruin probabilities and investment under in-
terest force in the presence of regularly varying tails. Technical report, Vienna
University of Technology.

J. Gaier, P. Grandits and W. Schachermayer (2003). Asymptotic Ruin Prob-


abilities and Optimal Investment. Annals of Applied Probability, 13, 1054-
1076.

Gerber, H.U. (1969). Entscheidigungskriterien für den zusammengesetzten Pois-


son Prozess. Mitteilungen der Vereinigung Schweizer Versicherungsmathe-
matiker, 19, 185-228.

Hipp, C. (2003). Stochastic control with application in insurance. Report, Uni-


versity of Karlsruhe.

Hipp, C. and M. Plum (2000). Optimal investment for insurers. Insurance:


Mathematics and Economics, 27, 215-228.

Hipp, C. and M. Plum (2003). Optimal investment for investors with state
dependent income, and for insurers. Finance and Stochastics, 7, 299-321.

Hipp, C. and M. Vogt (2003). Optimal dynamic XL reinsurance. ASTIN Bul-


letin, 33, 193-208.
49
Hipp, C and H. Schmidli (2003) Asymptotics of the ruin probability for the
controlled risk process: the small claims case. To appear in Scandinavian
Actuarial Journal

Højgaard, B. (2001). Optimal dynamic premium control in non-life insurance.


Maximizing dividend pay-outs. Preprint, Aalborg University.

Højgaard, B. and M. Taksar (1997). Optimal proportional reinsurance poli-


cies for diffusion models. Scandinavian Actuarial Journal, 166-180.

Højgaard, B. and M. Taksar (1999). Controlling risk exposure and dividend


pay-out scemes: Insurance company example. Mathematical Finance, 9, 153-
182.

Højgaard, B. and M. Taksar (2000). Optimal dynamic portfolio selection for


a corporation with controllable risk and dividend distribution policy. Sub-
mittet for publication.

Højgaard, B. and M. Taksar (2001). Optimal risk control for a large corpo-
ration in the presence of returns on investments. Finance and Stochastics, 5,
527-547.

Jeanblanc-Picqué, M. and A.N. Shiryaev (1995). Optimization of the flow


of dividends. Russian Math. Surveys, 50, 257-277.

Karatzas, I. and S. Shreve (1988). Brownian Motion and Stochastic Calcu-


lus. Springer, New York.

Lehtonen, T. and H. Nyrhinen (1992). Simulating level-crossing probabilities


by importance sampling. Advances in Applied Probability, 24, 858-874.

Martin-Löf, A. (1973). A method for finding the optimal decision rule for a
policy holder of an insurance with a bonus system. Scandinavaian Actuarial
Journal, 23-39.

Martin-Löf, A. (1983). Premium control in an insurance system, an approach


using linear control theory. Scandinavaian Actuarial Journal, 1-27

Martin-Löf, A. (1994). Lectures on the use of control theory in insurance. Scan-


dinavaian Actuarial Journal, 1-25.

Pestien, V.C. and W.D. Sudderth (1985). Continuous time red and black:
How to control a diffusion to a goal. Mathematics of Operations Research,
11, 371-382.

Paulsen, J. (2003). Optimal dividend payouts for diffusions with solvency con-
straints. Finance and Stochastics, 4, 457-474.

Paulsen, J. and H.K. Gjessing (1997). Optimal choice of dividend barriers


for a risk process with stochastic return on investments. Insurance: Mathe-
matics and Economics, 20, 215-223.
50
Paulsen, P. and B.N. Rasmussen (2003). Simulating ruin probabilities for a
class of semimartingales by importance sampling methods. Scandinavian
Actuarial Journal, 178-216.

Schmidli, H. (2001). Optimal proportional reinsurance policies in a dynamic


setting. Scandinavaian Actuarial Journal, 55-68.

Schmidli, H. (2002). On minimising the ruin probability by investment and


reinsurance. Annals of Applied Probability, 12, 890-907.

51

Vous aimerez peut-être aussi