optimal dividend strategy with transaction costs for an upward jump model
TRANSCRIPT
This article was downloaded by: [University of Saskatchewan Library]On: 12 September 2012, At: 14:54Publisher: RoutledgeInforma Ltd Registered in England and Wales Registered Number: 1072954 Registered office: MortimerHouse, 37-41 Mortimer Street, London W1T 3JH, UK
Quantitative FinancePublication details, including instructions for authors and subscription information:http://www.tandfonline.com/loi/rquf20
Optimal dividend strategy with transaction costs foran upward jump modelMing Zhou a & Ka Fai Cedric Yiu ba China Institute for Actuarial Science, Central University of Finance and Economics, 39South College Road, Haidian, Beijing 100081, P.R. Chinab Department of Applied Mathematics, Hong Kong Polytechnic University, Kowloon, HongKong, P.R. China
Version of record first published: 13 Jan 2012.
To cite this article: Ming Zhou & Ka Fai Cedric Yiu (2012): Optimal dividend strategy with transaction costs for an upwardjump model, Quantitative Finance, DOI:10.1080/14697688.2011.647052
To link to this article: http://dx.doi.org/10.1080/14697688.2011.647052
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions
This article may be used for research, teaching, and private study purposes. Any substantial or systematicreproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form toanyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contentswill be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug dosesshould be independently verified with primary sources. The publisher shall not be liable for any loss, actions,claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly orindirectly in connection with or arising out of the use of this material.
Quantitative Finance, 2012, 1–10, iFirst
Optimal dividend strategy with transaction costs for
an upward jump model
MING ZHOUy and KA FAI CEDRIC YIU*z
yChina Institute for Actuarial Science, Central University of Finance and Economics, 39 South College Road,Haidian, Beijing 100081, P.R. China
zDepartment of Applied Mathematics, Hong Kong Polytechnic University, Kowloon, Hong Kong, P.R. China
(Received 5 July 2010; revised 3 March 2011; in final form 2 December 2011)
In this paper, we consider the optimal dividend problem with transaction costs when theincomes of a company can be described by an upward jump model. Both fixed andproportional costs are considered in the problem. The value function is defined as the expectedtotal discounted dividends up to the time of ruin. Although the same problem has alreadybeen studied in the pure diffusion model and the spectrally negative Levy process, the optimaldividend problem in an upward jump model has two different aspects in determining theoptimal dividends barrier and in the property of the value function. First, the value function istwice continuous differentiable in the diffusion case, but it is not in the jump model. Second,under the spectrally negative Levy process, downward jumps will not cause any paymentactions; however, it might trigger dividend payments when there are upward jumps. Inderiving the optimal barriers, we show that the value function is bounded by a linear function.Using this property, we establish the verification theorem for the value function. By solvingthe quasi-variational inequalities associated with this problem, we obtain the closed-formsolution to the value function and hence the optimal dividend strategy when the income sizesfollow a common exponential distribution. In the presence of a fixed transaction cost, it isshown that the optimal strategy is a two-barrier policy, and the optimal barriers are onlydependent on the fixed cost and not the proportional cost. A numerical example is used toillustrate how the fixed cost plays a significant role in the optimal dividend strategy and alsothe value function. Moreover, an increased fixed cost results in larger but less frequentdividend payments.
Keywords: Upward jump model; Optimal dividend; Transaction costs; Stochastic jumps;Mathematical finance; Impulsive control
JEL Classification: C4, C6, C44, C61, G3, G35
1. Introduction
In recent years, upward jump models have received
increased attention in the field of finance and insurance.
Under the upward jump model, the company’s incomes
are received randomly and the expenses are paid contin-
uously at a fixed constant rate. Mathematically, the
surplus of a company at time t can be expressed as
XðtÞ ¼ xþ SðtÞ � ct,
where x is the initial surplus, c is the rate of expenses, and
S(t) is the aggregate income by time t, usually modeled by
a compound Poisson process.
There are various applications of this model in practice.
It can be used to describe the surplus of companies such
as oil exploration or firms that make natural resource
investments, in which new fields might be found from
time to time. It can also be applied to model firms
engaging in research and development and filing valuable
product patents along the way, or to model the surplus of
consultancy companies who are constantly bidding for
new projects. In these cases, the Poisson arrival incomes
are gained from a successful bid for a project or striking a
new oil field, and the costs and expenses are assumed to
be paid at a constant rate. Bayraktar and Egami (2008)
use a similar model to describe the capital of a venture
capital investment. In insurance, the upward jump model
has been proposed to describe life annuity or pension*Corresponding author. Email: [email protected]
Quantitative FinanceISSN 1469–7688 print/ISSN 1469–7696 online � 2012 Taylor & Francis
http://www.tandfonline.comhttp://dx.doi.org/10.1080/14697688.2011.647052
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
insurance, which pays benefits to annuitants and gainsprofits from the released reserve when the insured dies(see, e.g., Grandell (1991, p. 8)). The upward jump modelalso finds applications in inventory management, wherethe supplier supplies products with order arrivals follow-ing a Poisson process. For a company whose surplusprocess can be reasonably described in terms of theupward jump model, it is of interest to study the optimaldividend problem.
Within the framework of stochastic control theory, theoptimal dividend (or consumption) problem has beenstudied extensively in recent decades. In the literature, it isusually assumed that the dividend payments are distrib-uted without any cost (see, e.g., Asmussen and Taksar(1997), Kulenko and Schmidli (2008), Loeffen (2008), andSchmidli (2008) and references therein). This frictionlessassumption usually leads to a barrier or a thresholddividend strategy. With the presence of transaction costs,the optimal dividend problem has been studied by Shreveet al. (1984), Jeanblanc-Picque and Shiryave (1995, caseB), Boguslavskaya (2003, section 4), Cadenillas et al.(2007), and Paulsen (2007, 2008). In these papers, theoptimal dividend problems were established by employingstochastic impulse control theory (Harrison et al. 1983)and the asset processes were usually described by purediffusions without jumps. Bai and Guo (2010) andLoeffen (2009) studied the optimal dividends with trans-action costs for compound Poisson risk processes and thespectrally negative Levy process, respectively. Althoughthe jump-diffusion model was considered, the jump canonly be downward and therefore dividend payments willnot be triggered by jumps.
Furthermore, several dividend strategies have beenproposed in recent years by including jumps in the model.In particular, Avanzi et al. (2007) and Avanzi and Gerber(2008) studied the expectation of discounted dividends upto the time of ruin using a common barrier strategy. Ng(2009) considered the same problem under a generalthreshold dividend strategy. Gerber and Smith (2008)investigated the optimal dividend barrier with incompleteinformation under the model with diffusion. Albrecheret al. (2008) studied the expected discounted taxes beforeruin in which the taxes are paid at a constant rate whenthe surplus is at a running maximum. Zhu and Yang(2008) generalized the model to a regime-switching settingand studied the ruin probabilities instead of the dividendproblem.
In this paper, inspired by the above studies, we considerthe optimal impulse dividend problem in an upward jumpmodel under the framework of stochastic impulse control.The impulse dividends are paid out, incurring both fixedand proportional costs at the same time. The valuefunction is defined to maximize the expected totaldiscounted dividends payments up to the time of ruin.Because of the fixed cost, we found that the optimalstrategy is no longer a common barrier strategy orthreshold strategy. Instead, for the upward jump modelwith exponentially distributed income sizes, we prove thatthe optimal dividend strategy is a two-barrier policy. Withthe two-barrier strategy, there are no dividends paid out
before the surplus hits or exceeds the upper barrier. Onceit does, a lump sum amount to the surplus in excess of thelower barrier is paid out as dividends together with thetransaction costs. The surplus then starts at the lowerbarrier again. Dividends will not be paid until the time thesurplus hits the upper barrier again, and the same rule isapplied. In addition, we find that the optimal barriers areonly affected by the fixed cost and are independent of theproportional cost. When the fixed cost is zero, the two-barrier strategy degenerates to the common barrierstrategy studied by Avanzi et al. (2007). When the fixedcost is high, the lower barrier in the two-barrier strategywill be close to zero, which means that the investor willpay out all the surplus when it hits the upper barrier.
Finally, we would like to point out that Øksendal andSulem (2007) attempted to seek the classical solution tothe optimal dividend strategy by extending the valuefunction from a pure diffusion case to a jump diffusioncase (see example 6.4). However, the value function given there in equation (6.2.15) is not a full solution to thequasi-integrovariational inequalities presented in theo-rem 6.2. Due to the existence of jumps, the result ofequation (x) in theorem 6.2 involves the whole domain of . Therefore, it is true that 0(x) for �15x51 is thesolution to equation (x); however, after replacement by alinear function when x4x� and by 0 when x50, thecandidate solution in (6.2.15) is no longer a validsolution to equation (x). Similarly, Zou et al. (2009)incurred the same problem in their study. Despite thisslight flaw, their studies are indeed inspiring. In thispaper, we fix the problem for the upward jump modelwhen the jump size follows an exponential distribution.
The rest of the paper is organized as follows. In section2 we present a rigorous mathematical formulation of theoptimal impulse dividend problem. In section 3 weintroduce the quasi-variational inequality with respect tothis stochastic impulse problem and establish the verifi-cation theorem. In section 4 the classical solutions to thevalue function and optimal impulse dividend strategy areobtained when the income size follows a commonexponential distribution. Numerical calculations andeconomic explanations are given in section 5.
2. Formulation of the optimal dividend problem
Let (�,F , {F t, t� 0}, P) be a complete filtered probabil-ity space on which all stochastic objects are well definedthroughout this paper. Here {F t, t� 0} is a filtration thatsatisfies the usual conditions. The surplus process of anasset is described by an upward jump risk model. Withoutdividend payments, the surplus at time t, also denoted byX(t), is
XðtÞ ¼ xþ SðtÞ � ct: ð2:1Þ
Here x is the initial surplus and c is the rate of expenses;{S(t)} is a compound Poisson process of form SðtÞ ¼PNðtÞ
k¼1 Ik, where {N(t)} is a Poisson process with jumpintensity �, and {Ik}, independent of {N(t)}, denotes aseries of incomes with a common distribution function F
2 M. Zhou and K.F.C. Yiu
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
and a finite mean �. Clearly, the process X is a Levyprocess with upward jumps, but no diffusion, and S(t) canbe written as SðtÞ ¼
R t0
R10 zNðds, dzÞ, where N(ds, dz) is a
Poisson random measure with finite Levy measure �F.The infinitesimal generator of X, denoted by A, can beexpressed as
AgðxÞ ¼ �cg0ðxÞ þ �
Z 10
½ gðxþ yÞ � gðxÞ�dFð yÞ ð2:2Þ
for any function g in the domain of A. Throughout thispaper we denote {Px, x� 0} as the family of probabilitymeasures corresponding to the surplus process {X(t)} suchthat X(0)¼x, and we denote E
x as the expectation withrespect to P
x.Assume that the dividends are paid out in an impulsive
manner. An impulse dividend strategy � is described by
� ¼ ð�i, i ¼ 1, 2, . . . ; �i, i ¼ 1, 2, . . .Þ, ð2:3Þ
where {�i, i¼ 1, 2, . . .} is a sequence of increasing stoppingtimes with respect to {F t, t� 0} and {�i, i¼ 1, 2, . . .} is asequence of positive random variables. Here, �i is F �imeasurable denoting the amount of money paid out as
dividends to shareholders at time �i. In addition, trans-action costs are required for each payment. For thedividend payment �i, we assume that the insurer will pay afixed cost � � 0 and a proportion of the transaction cost��i with �� 0. Thus, for dividends �i, the amount ofmoney deducted from the surplus is
~�i :¼ � þ ð1þ �Þ�i:
Incorporating this impulse dividend strategy �, thesurplus process of the asset can be expressed as
X�ðtÞ ¼ xþ SðtÞ � ct�X1i¼1
If�i5tg~�i
¼ XðtÞ �X1i¼1
If�i5tg½� þ ð1þ �Þ�i�, ð2:4Þ
or, equivalently, written in the form of a stochasticdifferential equation,
dX�ðtÞ ¼
Z 10
zNðdt, dzÞ � cdt, �j 5 t � �jþ1,
X�ð�jþÞ ¼ X�ð�j�Þ þ DNSð�j Þ � ~�j, j ¼ 0, 1, 2, . . . ,
with X�(0)¼ x, �0¼ 0, and DNS(�j)¼S(�j)�S(�j�)denotes the size of the jump caused by the income processat time �j.
In practice, the company is strictly prohibited to paydividends under deficit. Hence, in this paper, it isreasonable to constrain every dividend payment so thatit will not lead to a negative surplus, i.e.
� � ~�i � X�ð�iÞ, or, equivalently,
0 � �i �1
1þ �½X�ð�iÞ � ��: ð2:5Þ
In particular, when the current surplus is less than thefixed cost �, it is not allowed to pay dividends because thecurrent surplus is not sufficient to cover the fixed cost.
We call the impulse dividend strategy � defined in (2.3)together with (2.5) an admissible control policy and denote� as the set of all admissible control policies. For eachadmissible control �2�, the time of ruin is defined as
�� ¼ infft � 0 : X�ðtÞ ¼ 0g,
and the corresponding performance functional J isdefined as
Jðx;�Þ ¼ ExX1i¼1
e��i If�i���g�i
" #,
where 40 is the discount factor. The goal of this paper isto find the value function defined as
VðxÞ ¼ sup�2�
Jðx;�Þ, for x � 0, ð2:6Þ
and the optimal dividend strategy �� ¼ ð��i , i ¼ 1, 2, . . . ;
��i , i ¼ 1, 2, . . .Þ 2 � such that V(x)¼ J(x;��).
3. The value function and the quasi-variational
inequalities
In this section, we establish the quasi-variational inequal-ities (QVI) associated with this stochastic impulse controlproblem and prove the verification theorem for a solutionto the QVI. Before doing so, we first establish a propertyof the value function.
Proposition 3.1: The value function V(x) defined by (2.6)is bounded by a linear function. More exactly, forevery x� 0,
ðx� �Þþ
1þ �� VðxÞ �
1
1þ �xþj��� cj
� �: ð3:1Þ
Proof: Apparently, the lower bound follows, because theentire surplus can be paid out instantaneously if x4� andno dividend is paid out if x� �. In the following weestablish the upper bound. For any admissible policy�2�, by the monotone convergence theorem, we have,from (2.4),
Jðx,�Þ¼ limt!þ1
ExX1i¼1
e��i If�i�t^��g�i
" #
� limt!þ1
1
1þ�Ex
Z t^��
0
e�sdXðsÞ�
Z t^��þ
0
e�sdX�ðsÞ
� �:
For the first integral, noting that t6 �� is a boundedstopping time, we have
Ex
Z t^��
0
e�sdXðsÞ
� �¼E
x
Z t^��
0
e�sdSðtÞ�c
Z t^��
0
e�sdt
� �
¼ ð��� cÞEx
Z t^��
0
e�sds
� �
� j��� cj
Z 10
e�sds
¼j���cj
: ð3:2Þ
Optimal dividend strategy with transaction costs 3
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
For the second integral, from the Ito formula it follows
that
0 � e�ðt^��ÞX�ðt ^ ��þÞ ¼ x�
Z t^��þ
0
e�sX�ðsÞds
þ
Z t^��þ
0
e�sdX�ðsÞ:
Since X�(s)� 0 for s� t6 ��, it follows thatZ t^��þ
0
e�sdX�ðsÞ � �x: ð3:3Þ
Together with (3.2) and (3.3), we have
Jðx,�Þ �1
1þ �xþj��� cj
� �,
and (3.1) follows by taking the supremum over all
�2�. œ
For this impulse control problem, we start with the
definition of the intervention operatorM.
Definition 3.2: For each continuous function h
supported on [0, 1), the intervention operator M is
defined by
MhðxÞ
¼sup� hðx���ð1þ�Þ�Þþ �;0� �� x��
1þ�
n o, � � x51,
0, 0� x5�:
(
ð3:4Þ
The intervention operator can be intuitively interpreted
as follows. Let us view h(x) as the value function and
assume the current surplus is x. At the current time, there
are two possible choices for the decision maker: pay
dividends or not. If the decision maker decides to pay a
lump sum dividend �, it is reasonable to pay the amount
by which the supremum is attained in (3.4), otherwise the
dividend strategy will not be optimal. Therefore, if the
decision is right for the optimal strategy, we have
h(x)¼Mh(x). Otherwise, h(x)4Mh(x) follows from the
definition of the value function. The set Dh, defined by
Dh ¼ fx � 0; hðxÞ ¼ MhðxÞg,
is called the intervention region of h. Thus, for the value
function V(x) the inequality V(x)�MV(x) must hold.
This is the key step in finding the intervention region of
the value function.
Definition 3.3: A continuous function g(x) supported on
[0, 1) is said to satisfy the quasi-variational inequalities
of the impulse control problem if
gð0Þ ¼ 0, ð3:5Þ
gðxÞ �MgðxÞ � 0, x � 0, ð3:6Þ
AgðxÞ � gðxÞ � 0, x � 0, ð3:7Þ
½AgðxÞ � gðxÞ�½ gðxÞ �MgðxÞ� ¼ 0, x � 0: ð3:8Þ
We define the optimal impulse control policy associatedwith this solution in the following.
Definition 3.4: Given a solution (x) to QVI, an optimalimpulse control policy,
� ¼ ð��1 , ��2 , . . . ; ��1, �
�2, . . .Þ,
is defined inductively as follows:
��1 ¼ infft � 0 : XðtÞ 2 Dg, ��1 ¼ argMðXð��1 ÞÞ;
for j� 2,
��jþ1 ¼ infft4 ��j : X�jðtÞ 2Dg, ��jþ1 ¼ argMðX�jð��jþ1ÞÞ:
As a result, the control strategy �j is defined by
�j ¼ ð��1 , ��2 , . . . , ��j ; ��1, �
�2, . . . , ��j Þ:
The optimal impulse control policy suggests thefollowing facts. Given a candidate (x) for the valuefunction, the intervention takes place once the controlledprocess falls into the intervention region of , and theamount of dividend payment is determined by taking themaximum of the intervention operatorM(x).
In the following theorem, we show that a solution toQVI can be a candidate for the value function of thisstochastic impulse control problem. If the optimalimpulse control policy associated with this candidatesolution is taken, then the candidate coincides with thevalue function.
Theorem 3.5: Let (x) be a continuous function supportedon [0, 1) which is differentiable except for x¼ b� � 0. If(x) is a solution to the quasi-variational inequalitiesdefined in definition 3.3, such that
0 � ðxÞ � Kð1þ xÞ, for some constant K4 0, ð3:9Þ
then
ðxÞ � VðxÞ:
Furthermore, by taking the optimal impulse control policy� associated with , we have
ðxÞ ¼ VðxÞ ¼ Jðx;�Þ:
Proof: Choose any admissible strategy �¼ (�i, i¼ 1, 2,. . .; �i, i¼ 1, 2, . . .)2� and set �0¼ 0. By employing the Itoformula, we have, a.s. P
x (that is, excluding the null set{X�(s�)¼ b� for at most countable s� 0}), that
e�ðt^��ÞðX�ðt ^ ��þÞÞ � ðxÞ
¼
Z t^��þ
0
e�s½A � �ðX�ðs�ÞÞds
þX�i�t^��
e��i ½ðX�ð�iþÞÞ � ðX�ð�iÞÞ�:
Taking expectations on both sides of the above equality,together with (3.6) and (3.7), we obtain
Ex½e�ðt^�
�ÞðX�ðt ^ ��þÞÞ � ðxÞ � Ex
X�i�t^��
e��i�i
" #:
ð3:10Þ
4 M. Zhou and K.F.C. Yiu
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
The above inequality becomes an equality when � is
replaced by the optimal impulse policy � associated with
(x). Because X� ðs�Þ 62 D for all s� 0, then
½A � �ðX� ðs�ÞÞ ¼ 0 for all s� 0. In addition, note
that X� ð�iÞ 2 D, then it follows that
ðX� ð��i þÞÞ � ðX� ð��i ÞÞ ¼ ��
�i :
Taking t!þ1 on both sides of (3.10) and by the
monotone convergence theorem, we have
limt!þ1
Ex½e�ðt^�
�ÞðX�ðt ^ ��þÞÞ� � ðxÞ � ExX�i���
e��i�i
" #:
On the other hand, together with (0)¼ 0 we have
Ex½e�ðt^�
�ÞðX�ðt ^ ��þÞÞ� ¼ Ex½e�tðX�ðtþÞÞIf��4tg�:
Since (x) is bounded by a linear function (see
equation (3.9)), we have
0 � Ex½e�ðt^�
�ÞðX�ðt ^ ��þÞÞ� � KEx½e�tð1þ X�ðtþÞÞ�
� KEx½e�tð1þ XðtÞÞ�:
Because
limt!þ1
Ex½e�tjXðtÞj� � lim
t!þ1e�t½xþ ðcþ ��Þt� ¼ 0,
we can conclude that
limt!þ1
Ex½e�ðt^�
�ÞðX�ðt ^ ��þÞÞ� ¼ 0,
and for every admissible strategy �2�,
ðxÞ � ExX�i���
e��i�i
" #¼ Jðx;�Þ:
The above inequality becomes an equality by taking
optimal impulse policy � associated with (x). Thus it
follows that
ðxÞ ¼ VðxÞ ¼ Jðx;�Þ:
Then the proof is completed. œ
4. Closed-form solutions to the quasi-variational
inequalities
From the last section, we know that a solution to QVI will
be a candidate for the value function. In this section, we
construct the closed-form solution to the value function
and optimal impulse dividend policy when the income size
follows an exponential distribution, i.e.
dFð yÞ ¼1
�e�ð1=�Þ ydy, y � 0:
In addition, we assume that the expected increase of the
surplus per unit time is positive, i.e.
E½Sð1Þ� � c ¼ ��� c4 0, ð4:1Þ
which is the so-called positive safe loading condition.
To construct a classical solution, we guess that the
intervention region of the value function is of the form
[b�,1). Thus, we first solve A(x)� (x)¼ 0, or,
equivalently,
c0ðxÞ þ ðxÞ � �
Z 10
½ðxþ yÞ � ðxÞ�dFð yÞ ¼ 0,
for x 2 ð0, b�Þ, ð4:2Þ
with boundary condition (0)¼ 0. It is worth noting that
this integro-differential equation (4.2) involves the form
of from x to1. It is impossible to give a solution to this
equation without knowing the expression of (x) for
x4b�. Fortunately, in the case of exponential incomes,
the integro-differential equation (4.2) can be reduced to
an ordinary differential equation by applying the operator
[(d/dx)� (1/�)]. In fact, if dF(y)¼ (1/�) e�(1/�)y dy, we
can derive that
d
dx
Z 10
ðxþ yÞdFð yÞ ¼1
�
Z 10
ðxþ yÞdFð yÞ � ðxÞ
� �,
by employing integration by parts, i.e.
d
dx�
1
�
� �Z 10
ðxþ yÞdFð yÞ ¼ �1
�ðxÞ:
Therefore, applying the operator [(d/dx)� (1/�)] to (4.2),
we have
c00ðxÞ þ þ ��c
�
� �0ðxÞ �
�ðxÞ ¼ 0,
which implies that the solution is of the form erx for some
constant r2R. Thus r must satisfy the equation
cr2 þ þ ��c
�
� �r�
�¼ 0: ð4:3Þ
This equation determines two roots, r1 and r2, such that
r2505r151/�. Combining with the positive loading
condition ��4c, we know that jr2j4r1. With this choice
of r1 and r2 determined by (4.3), we have ðxÞ ¼ A1er1xþ
A2er2x. Combining with the boundary condition (0)¼ 0,
(x) can be rewritten as
ðxÞ ¼A
1þ �ðer1x � er2xÞ, for x 2 ½0, b�Þ:
Motivated by the diffusion case studied by Jeanblanc-
Picque and Shiryaev (1995) and Cadenillas et al. (2007),
we try a two-barrier dividend strategy with 0� a� � b� as
the optimal dividend strategy. If this conjecture is right,
then the value function will be of the form
ðxÞ ¼
A1þ� ðe
r1x � er2xÞ, 0 � x5 b�,
ða�Þ þ x�a���1þ� , b� � x51,
(ð4:4Þ
where the three constants A, a� and b� are yet to be
determined. In addition, the value function (x) is
continuous at b�, which means
ðb��Þ ¼b� � a� � �
1þ �þ ða�Þ: ð4:5Þ
Optimal dividend strategy with transaction costs 5
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
If D¼ [b�, 1), for any x� b�, we have
ðxÞ ¼ sup ðx� � � ð1þ �Þ�Þ þ �; 0 � � �x� �
1þ �
� �:
Thus the maximum, denoted by �ðxÞ, satisfies
x� � � ð1þ �Þ�ðxÞ ¼ a�, 0ða�Þ ¼1
1þ �and 00ða�Þ5 0:
From (4.4), we know that
0ða�Þ ¼A
1þ �ðr1e
r1a�
� r2er2a�
Þ ¼1
1þ �,
which implies
A ¼1
r1er1a�� r2er2a
� : ð4:6Þ
On the other hand, substituting (4.4) into (4.2) yields
c0ðxÞ þ ðþ �ÞðxÞ � �
Z b�
x
ð yÞdFð y� xÞ
��
1þ �
Z 10
ydFð yþ b� � xÞ � �ðb�Þ �Fðb� � xÞ ¼ 0,
05 x5 b�:
Letting x! b�� on both sides of the above equation
yields
c0ðb��Þ þ ðb�Þ ��
1þ �� ¼ 0:
Combining with (4.5), we have
A ¼��
ðcr1 þ Þer1b�� ðcr2 þ Þer2b
� : ð4:7Þ
From (4.6) and (4.7), if A is given, a� and b� can be solved
from these two equations simultaneously. Noting that
the solution to each equation might not be unique
(see figure 1), we view A as a function of a� and b�.
According to (4.6) and (4.7), we define two new functions,
Aa(x) and Ab(x), as follows:
AaðxÞ ¼1
r1er1x � r2er2x, ð4:8Þ
AbðxÞ ¼��
ðcr1 þ Þer1x � ðcr2 þ Þer2x: ð4:9Þ
Thus we need to find a�, b� and A such that
A¼Aa(a�)¼Ab(b
�). In what follows, we analyse the
property of the two function Aa(x) and Ab(x).First, at x¼ 0, we have
05Aað0Þ ¼1
r1 � r25
��=c
r1 � r2¼ Abð0Þ, ð4:10Þ
where the inequality follows from the positive loading
condition (4.1).Taking the derivative with respect to x on both sides
and after some algebra, we have
A0aðxÞ ¼ �eðr1þr2ÞxqaðxÞ
ðr1er1x � r2er2xÞ2, ð4:11Þ
A0bðxÞ ¼ ���eðr1þr2ÞxqbðxÞ
½ðcr1 þ Þer1x � ðcr2 þ Þer2x�2, ð4:12Þ
where
qaðxÞ ¼ r21e�r2x � r22e
�r1x, ð4:13Þ
qbðxÞ ¼ ðcr1 þ Þr1e�r2x � ðcr2 þ Þr2e
�r1x: ð4:14Þ
Since r2505r1 and jr2j4r1, it follows that qað0Þ ¼ ð1=cÞqbð0Þ ¼ r21 � r22 5 0 and q0aðxÞ ¼ r1r2ðr2e
�r1x� r1e�r2xÞ4 0.
As for qb(x), we have q0bðxÞ ¼ r1r2½ðcr2 þ Þe�r1x�
ðcr1 þ Þr2e�r2x�. From (4.3) and (4.1), it follows that
0 5 10 15 20 25 30 35 40 45 500
1
2
3
4
5
6
7
8Aa(x )Ab(x )A(x )
Figure 1. Shapes of the functions Aa(x), Ab(x) and A(x).
6 M. Zhou and K.F.C. Yiu
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
cr2þ ¼ (1/�) [(/r2)þ (c� ��)]50. Thus we also have
q0bðxÞ4 0. Both qa(x) and qb(x) are strictly increasing
functions. Let �xa and �xb satisfy qað �xaÞ ¼ 0 and qbð �xbÞ ¼ 0,
respectively, i.e.
�xa ¼1
r1 � r2lnr22r21, ð4:15Þ
�xb ¼1
r1 � r2lnðcr2 þ Þr2ðcr1 þ Þr1
: ð4:16Þ
With the above analysis, we can conclude that the
function Aa(x) is strictly increasing on ½0, �xa� and strictly
decreasing to zero on ½ �xa,1Þ, and the function Ab(x) is
strictly increasing on ½0, �xb� and strictly decreasing to zero
on ½ �xb,1Þ. In addition, from (4.3), it follows that
�xb ¼1
r1 � r2ln� ð��� cÞr2� ð��� cÞr1
¼1
r1 � r2lnr22ðð1=�Þ � r1Þ
r21ðð1=�Þ � r2Þ
51
r1 � r2lnr22r21¼ �xa:
In addition,
AaðxÞ�AbðxÞ ¼½ðcr1þ Þ���r1�e
r1x�½ðcr2þ Þ���r2�er2x
ðr1er1x� r2er2xÞ½ðcr1þ Þer1x�ðcr2þ Þer2x�
¼�eðr1þr2ÞxqbðxÞ
ðr1er1x� r2er2xÞ½ðcr1þ Þer1x�ðcr2þ Þer2x�,
which means Aa(x)5Ab(x) on ½0, �xbÞ, Aa(x)4Ab(x) on
½ �xb,1Þ and Aað �xbÞ ¼ Abð �xbÞ.Combining with Aa(x) and Ab(x), we define a new
function A(x)¼min{Aa(x), Ab(x)}, or, equivalently,
AðxÞ ¼AaðxÞ, 0 � x � �xb,AbðxÞ, x � �xb:
�ð4:17Þ
Thus, the function A(x) is strictly increasing on ½0, �xb� and
strictly decreasing to zero on ½ �xb,1Þ. The shapes of the
functions A(x), Aa(x) and Ab(x) are shown in figure 1.In what follows we will give the solutions of a�, b� and
A in terms of the function A(x). Define �b4 �xb such that
Að0Þ ¼ Að �bÞ. From figure 1 we know that, for any
b 2 ½ �xb, �b�, there exists a unique a 2 ½0, �xb� such that
A(a)¼A(b). Consider a as a function of b and denote it by
a(b). For any b 2 ½ �xb, �b�, we have A(a(b))¼A(b), að �xbÞ ¼�xb and að �bÞ ¼ 0. Thus, the function a(b) is strictly
decreasing from �xb to zero on ½ �xb, �b�. For b4 �b, we
define a(b)� 0.Define a new function of the form
ðb, xÞ ¼ AðbÞðer1x � er2xÞ, b � 0 and 0 � x � b,
where the function A(x) is determined by (4.17).
Combined with the fact that (x) is continuous at b�,
this implies that b� must solve the equation
ðbÞ ¼ ðaðbÞÞ þb� aðbÞ � �
1þ �,
or, equivalently,Z b
aðbÞ
1�@
@x ðb, xÞ
� �dx ¼ �:
Lemma 4.1: Define a function
gðbÞ ¼
Z b
aðbÞ
1�@
@x ðb, xÞ
� �dx, b 2 ½ �xb,1Þ,
then g(b) is increasing from zero to þ1 on ½ �xb,1Þ.
Proof: Note that A(x)¼Aa(x) on ½0, �xb� and A(x)5Aa(x) on ð �xb,1Þ. Given b 2 ½ �xb,1Þ and a(b)� x� b, we
have A(b)�Aa(x) and
@
@x ðb,xÞ ¼AðbÞðr1e
r1x� r2er2xÞ �AaðxÞðr1e
r1x� r2er2xÞ ¼ 1,
i.e. 1� (@/@x) (b, x)� 0 for x2 [a(b), b]. Also note that
a(b) is decreasing on ½ �xb, þ1Þ. Thus, for �xb � b1 5b2 51, we have [a(b1), b1]� [a(b2), b2], and
gðb1Þ ¼
Z b1
aðb1Þ
1�@
@x ðb1,xÞ
� �dx
5Z b1
aðb1Þ
1�@
@x ðb2, xÞ
� �dx
�
Z b2
aðb2Þ
1�@
@x ðb2, xÞ
� �dx ¼ gðb2Þ,
which implies that g(b) is strictly decreasing on ½ �xb,1Þ.
Since að �xbÞ ¼ �xb, we know that gð �xbÞ ¼ 0 and
gðþ1Þ ¼ limb!þ1
Z b
aðbÞ
1�@
@x ðb, xÞ
� �dx
¼ limb!1
b� AbðbÞðer1b � er2bÞ ¼ þ1:
This completes the proof. œ
According to the above lemma, we know that there
exists a unique constant b�, such that
gðb�Þ ¼
Z b�
aðb�Þ
1�@
@x ðb�, xÞ
� �dx ¼ �, b� 2 ½ �xb,1Þ:
ð4:18Þ
In addition, since A(b�)40, we know that
@
@x ðb�, xÞ ¼ Aðb�Þðr1e
r1x � r2er2xÞ4 0, ð4:19Þ
which implies that b� � a(b�)� � from (4.18). The gap
between the upper barrier and the lower barrier is always
greater than the fixed cost.Based on the above analysis, we conclude the solution
to the value function in the following theorem.
Theorem 4.2: The value function is of the form
ð1þ �ÞVðxÞ
¼Aðb�Þðer1x � er2xÞ, 0 � x5 b�,
Aðb�Þðer1a�
� er2a�
Þ þ ðx� a� � �Þ, b� � x,
�ð4:20Þ
where b� 2 ½ �xb,1Þ is determined by (4.18) and a� ¼ a(b�),
such that 0 � a� � �xb � b�51. In addition, the value
function is not differentiable at x¼ b� except for the case
of �¼ 0.
Optimal dividend strategy with transaction costs 7
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
Proof: According to the above analysis, we know thatV(x) with the form (4.20) is continuous on (0,1) anddifferentiable on (0, b�)[ (b�, 1). This function is notdifferentiable at x¼ b� except for the case where �¼ 0.This is because
ð1þ �ÞV0ðb��Þ ¼@
@x ðb�, xÞjx¼b�
¼ Aðb�Þðr1er1b�
� r2er2b�
Þ
� Aaðb�Þðr1e
r1b�
� r2er2b�
Þ
¼ 1, ð4:21Þ
where the equality follows only when �¼ 0 and thenb� ¼ �xb. So we have V0(b� � )51/(1þ �) when b�4 �xb,and V0(b��)¼ 1/(1þ �) when b� ¼ �xb. Note thatV0(b�þ)¼ 1/(1þ �). Hence V0(b��)� 1/(1þ �)¼V0(b�þ).The equality follows only when b� ¼ �xb, which is the casewhere the fixed cost �¼ 0 in (4.18).
In order to show that V(x) is a solution to QVI, weneed the following two steps.
Step 1: We verify DV¼ [b�, 1). According to thedefinition of the intervention operator, we have
MVðxÞ ¼
Vða�Þ þ x�a���1þ� , x � a� þ �,
Vðx� �Þ, �5 x5 a� þ �,
0, 0 � x � �:
8><>:
Without loss of generality, we assume that 05� �a�5a� þ �5b�.
For 0� x5a� þ �, we have V0ðxÞ ¼ Aðb�Þðr1er1x�
r2er2xÞ4 0, which implies
VðxÞ4Vð0Þ ¼ 0 ¼MVðxÞ, for 0 � x5 �,
and
VðxÞ4Vðx� �Þ ¼ MVðxÞ, for � � x5 a� þ �:
For a� þ � � x5b� we have
V0ðxÞ ¼ Aðb�Þðr1er1x � r2e
r2xÞ5Aaðb�Þðr1e
r1x � r2er2xÞ
� AaðxÞðr1er1x � r2e
r2xÞ ¼1
1þ �,
which implies
VðxÞ4Vðb�Þ �b� � x
1þ �¼ Vða�Þ þ
b� � a� � �
1þ ��b� � x
1þ �
¼ Vða�Þ þx� a� � �
1þ �¼MVðxÞ:
For x� b�, it is clear that V(x)¼MV(x).
Step 2: We verify (3.7) of QVI. For 0� x� b�, we havethat AV(x)� V(x)¼ 0 and V(0)¼ 0. To see this,replacing (x) by V(x) (defined in (4.20)) in the equationbelow (4.6), and noting that dF(y)¼ (1/�) e�(1/�)ydy, wehave, for any 0� x� b�,
�AVðxÞ þ VðxÞ
¼ cV0ðxÞ þ ðþ �ÞVðxÞ � �
Z b�
x
Vð yÞdFð y� xÞ
��
1þ �
Z 10
ydFð yþ b� � xÞ � �Vðb�Þ �Fðb� � xÞ
¼Aðb�Þ
1þ �
�cr1 þ þ �þ
�
r1�� 1
� �er1x
� cr1 þ þ �þ�
r1�� 1
� �er2x
�
�
�Aðb�Þ
1þ �
��þ
�
r1�� 1
� �er1b
�
� �þ�
r2�� 1
� �er2b
�
�
þ��
1þ �
�e�ðb
��xÞ=�:
In addition, from (4.3), we know that
cri þ þ �þ�
ri�� 1¼ 0, i ¼ 1, 2, ð4:22Þ
and
�þ�
ri�� 1¼ �cri þ , i ¼ 1, 2: ð4:23Þ
Note that b� � �xb, thus
Aðb�Þ ¼ Abðb�Þ ¼
��
ðcr1 þ Þer1b�� ðcr2 þ Þer2b
� :
This, together with (4.22) and (4.23), implies thatAV(x)� V(x)¼ 0 for 0� x� b�.
For x4b�, noting that V0(x)¼ 1/(1þ �), we haveAV(x)¼ (��� c)/(1þ �), and therefore
AVðxÞ � VðxÞ ¼ AVðb�þÞ � VðxÞ
5AVðb�þÞ � Vðb�þÞ
� AVðb��Þ � Vðb��Þ ¼ 0,
where the last two steps follow from the continuity ofV(x) at x¼ b� and V0(b��)� 1/(1þ �)¼V0(b�þ). Thus,(3.7) of QVI follows.
With the above two steps, we know that V(x) definedin (4.20) is a solution to QVI (3.5–3.8). Then, according totheorem 3.5, this completes the proof. œ
Remark 1: According to the procedure described in thissection, the two barriers of the optimal strategy aredetermined by (4.17) and (4.18). Therefore, the barriersare only related to the fixed transaction cost and do notdepend on the proportional transaction cost.
Denote a critical value for the fixed cost by
�0 ¼ gð �bÞ,
where the function g is defined in lemma 4.1. If 0� �5�0,it follows that 05 a� � b�5 �b. Otherwise, if � � �0, thenb� � �b and a� ¼ 0. This observation implies that if thefixed cost is less than the critical value (low fixed cost),there exists a regular two-barrier dividend strategy. If thefixed cost is greater than the critical value (high fixedcost), then the optimal dividend strategy is reduced to astrategy of waiting for a chance to take the money andrun away. In particular, for the special case where thefixed cost is zero, i.e. �¼ 0,
a� ¼ b� ¼ �xb ¼1
r1 � r2lnðcr2 þ Þr2ðcr1 þ Þr1
,
8 M. Zhou and K.F.C. Yiu
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
which was obtained by Avanzi et al. (2007), where theproportional cost � is also assumed to be zero.
5. Numerical examples
In theorem 4.2, we obtain the value function and theoptimal dividend strategy in the case of exponentiallydistributed income sizes. In this section, we perform somenumerical calculations and study the impact of a fixedtransaction cost on the value function and the optimaldividend strategy.
Example 5.1: Set the parameters as follows: ¼ 0.1,�¼ 1, c¼ 1 and the jump size follows a Expð12Þ distribu-tion, i.e. f(x)¼ e�(1/2)x, x� 0. Then �¼ 2. The two rootsof (4.3) are r1¼ 0.0742 and r2¼� 0.6742. Then we have�xb ¼ 4:5436, �b ¼ 29:0024 and �0¼ 17.5191.
Figure 2 displays the optimal two-barrier strategies as afunction of the fixed transaction cost. Without the fixedcost (�¼ 0), the optimal strategy is a common barrierpolicy with level barrier �xb ¼ 4:5436, as represented bythe thin horizontal line. In the presence of a fixed cost, theoptimal strategy is a two-barrier policy. The amount ofsurplus over the lower barrier a� is paid out as dividendsplus costs once the surplus reaches the upper barrier b�.Thus, the fixed transaction has a significant impact on theoptimal dividend strategy. In addition, it should also benoted that, as the fixed cost increases, the upper barrierincreases, but the lower barrier decreases, making thedividend payments larger but less frequent. When thefixed cost exceeds �0¼ 17.5191, the lower barrier will stayat zero. When the surplus reaches the upper barrier b�,the amount of the whole surplus is paid out and ruinoccurs immediately because of the dividend payments.We call this a chance-waiting strategy to take the moneyand run away.
Set the proportional transaction cost rate �¼ 0.01.Figure 3 displays the shapes of the value function for thefixed cost �¼ 0, 10 and 20. As we can see, the valuefunction is convex when the fixed cost is zero. In thepresence of a fixed cost, the value function is no longerconvex. Thus, the fixed cost has a significant impact onthe value function. With an increase in the fixed cost, thevalue function decreases. In addition, as proved in the lastsection, the value function is not differentiable at x¼ b� inthe presence of a fixed cost, while it is differentiable whenthe fixed cost � ¼ 0.
Table 1 shows the optimal two-barrier strategies underdifferent fixed costs, and the corresponding value func-tion at x¼ b� is calculated. Figure 4 displays the valuefunction V(x) at x¼ b� as a function of the fixed cost. Aswe can see, V(b�) increases sharply when the fixed cost islow. This suggests that even a rather low fixed cost has avery significant impact on the value function.
6. Concluding remarks
The main contribution of this paper has been to studymathematically the optimal dividend strategies in upward
jump models in the presence of transaction costs.
The closed-form solution to the value function and
hence the optimal dividend strategy have been established
in the case of exponentially distributed income sizes. We
have shown that a two-barrier strategy is optimal and
both the upper barrier and the lower barrier depend only
on the fixed cost, but not the proportional cost.
Furthermore, it has been shown that an increased fixed
cost results in larger but less frequent dividend payments.
0 5 10 15 20 250
5
10
15
20
25
30
35
40
a*
b*
Figure 2. The two-barrier strategies as a function of the fixedcost.
0 5 10 15 20 25 30 350
2
4
6
8
10
12
14
16
18
20
x
V(x
)
= 20
= 10
= 0
Figure 3. The value function with different fixed costs.
Table 1. Optimal two-barrier strategies and V(b�) for differentfixed costs (�0¼ 17.5191, �xb ¼ 4:5436, �b ¼ 29:0024).
�
0 5 10 15 20 25 30
a� 4.5436 1.9470 1.0280 0.3199 0 0 0b� 4.5436 15.1120 21.1229 26.4512 31.4833 36.4833 41.4833V(b�) 9.9010 11.3690 11.3696 11.3696 11.3696 11.3696 11.3696
Optimal dividend strategy with transaction costs 9
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12
As an extension of this work, it would certainly be ofinterest to study the optimal dividend strategy with amore general distributed income size.
Acknowledgements
The first author was supported by a grant from theNatural Science Foundation of China (10701082) and bythe Scientific Research Foundation for ReturnedOverseas Chinese Scholars, State Education Ministry.The second author was supported partially by RGC grantPolyU(5001/11P) during revision of the paper, TheAMSS-PolyU Joint Research Institute and the ResearchCommittee of Hong Kong Polytechnic University.
References
Albrecher, H., Badescu, A. and Landriault, D., On the dualmodel with tax payments. Insurance: Math. Econ., 2008, 42,1086–1094.
Asmussen, S. and Taksar, M., Controlled diffusion models foroptimal dividend pay-out. Insurance: Math. Econ., 1997, 20,1–15.
Avanzi, B. and Gerber, H.U., Optimal dividends in the dualmodel with diffusion. ASTIN Bull., 2008, 38, 653–667.
Avanzi, B., Gerber, H.U. and Shiu, E.S.W., Optimal dividendsin the dual model. Insurance: Math. Econ., 2007, 41, 111–123.
Bai, L. and Guo, J., Optimal dividend payments in the classicalrisk model when payments are subject to both transactioncosts and taxes. Scand. Actuar. J., 2010, 1, 36–55.
Bayraktar, E. and Egami, M., Optimizing venture capitalinvestment in a jump diffusion model. Math. Meth. Oper.Res., 2008, 67, 21–42.
Boguslavskaya, E., On optimization of dividend flow for acompany in the presence of liquidation value. Preprint, 2003.
Cadenillas, A., Sarkar, S. and Zapatero, F., Optimal dividendpolicy with mean-reverting cash reservoir. Math. Finance,2007, 17, 81–109.
Gerber, H.U. and Smith, N., Optimal dividends with incompleteinformation in the dual model. Insurance: Math. Econ., 2008,43, 227–233.
Grandell, J., Aspects of Risk Theory, 1991 (Springer:New York).
Harrison, J., Sellke, M. and Taylor, J., Impulse control ofBrownian motion. Math. Oper. Res., 1983, 8(3), 454–466.
Jeanblanc-Picque, M. and Shiryaev, A.N., Optimization of theflow of dividends. Russ. Math. Surv., 1995, 50(2), 257–277.
Kulenko, N. and Schmidli, H., Optimal dividend strategies in aCramer–Lundberg model with capital injections. Insurance:Math. Econ., 2008, 43, 270–278.
Loeffen, R.L., On optimality of the barrier strategy in deFinetti’s dividend problem for spectrally negative Levyprocesses. Ann. Appl. Probab., 2008, 18, 1669–1680.
Loeffen, R.L., An optimal dividend problem with transactioncosts for spectrally negative Levy processes. Insurance: Math.Econ., 2009, 45, 41–48.
Ng, A.C.Y., On a dual model with a dividend threshold.Insurance: Math. Econ., 2009, 44, 315–324.
Øksendal, B. and Sulem, A., Applied Stochastic Control of JumpDiffusions, 2nd ed., 2007 (Springer: Berlin).
Paulsen, J., Optimal dividend payments until ruin of diffusionprocesses when payments are subject to both fixed andproportional costs. Adv. Appl. Probab., 2007, 39, 669–689.
Paulsen, J., Optimal dividend payments and reinvestments ofdiffusion processes with both fixed and proportional costs.SIAM J. Control Optimiz., 2008, 47(5), 2201–2226.
Schmidli, H., Stochastic Control in Insurance, 2008 (Springer:London).
Shreve, S.E., Lehoczky, J.P. and Gaver, D.P., Optimalconsumption for general diffusions with absorbing andreflecting barriers. SIAM J. Control Optimiz., 1984, 22(1),55–75.
Zhu, J. and Yang, H., Ruin probabilities of a dual Markov-modulated risk model. Commun. Statist.: Theory Meth., 2008,37, 3298–3307.
Zou, J., Zhang, Z. and Zhang, J., Optimal dividend payoutsunder jump-diffusion risk processes. Stochast. Models, 2009,25(2), 332–347.
0 2 4 6 8 109.8
10
10.2
10.4
10.6
10.8
11
11.2
11.4V
(b* )
Figure 4. The value function at b� as a function of �.
10 M. Zhou and K.F.C. Yiu
Dow
nloa
ded
by [
Uni
vers
ity o
f Sa
skat
chew
an L
ibra
ry]
at 1
4:54
12
Sept
embe
r 20
12