Finance Stochast.

3, 137–165 (1999)

c Springer-Verlag 1999

Optimal trading of a security when there are taxes

and transaction costs?
Abel Cadenillas1 , Stanley R. Pliska2
1 Department of Mathematical Sciences, University of Alberta, Edmonton, Alberta,
Canada T6G 2G1 (e-mail: [email protected])
2 University of Illinois at Chicago, 601 South Morgan Street, Chicago, IL 60607-7124, USA

(e-mail: [email protected])

Abstract. We study the problem of investing in securities in order to maximize

the after-tax rate of return. We consider a single stock modeled as geometric
Brownian motion along with the objective of maximizing the long-run growth
rate of after-tax wealth. We show that it is optimal not only to cut short the
losses, but also the profits, even though there is no distinction between short and
long term tax rates. This surprising result may be due to the possibility of using
the tax system to reduce after-tax volatility.

Key words: Portfolio management, stopping time, stochastic control, taxes, trans-
action costs

JEL classification: G11, H20, C63

Mathematics Subject Classification (1991): 90A09, 93E20, 60H30, 60G44,

90A16

1 Introduction

The practical management of a realistic portfolio for a tax-paying investor is

exceedingly complicated because the optimal policy depends on much of the
whole history of investment, and so computational complexity grows rapidly
with time and number of securities. There is literature on how trading strategies
? This work was initiated when we participated in the Programme on Financial Mathematics

organized by the Isaac Newton Institute for Mathematical Sciences, University of Cambridge, January
– June 1995. We are very grateful to the Newton Institute for its kind hospitality. Abel Cadenillas
is also grateful to the Instituto Tecnológico Autónomo de México, where he made some of his
contributions to this paper. His research was supported in part by the Social Sciences and Humanities
Research Council of Canada grant 410-97-0204.
Manuscript received: June 1997; final version received: February 1998
138 A. Cadenillas, S.R. Pliska

affect after-tax returns, but these studies (see, for example, Miller 1977; Miller
and Scholes 1978; Schaefer 1982; Constantinides 1983, 1984; Constantinides and
Ingersoll 1984; Dybvig and Ross 1986; Bossaerts and Dammon 1994; Dammon,
Dunn and Spatt 1989 and Dammon and Spatt 1996) tend to deal with specific
situations, to make simplifying assumptions, and to emphasize the consequences
for security market equilibrium. And the conventional wisdom emerging from
this literature seems to be that while one should cut one’s losses short, one should
never realize a gain.
The papers by Constantinides 1983, 1984; and Dammon and Spatt 1996; are
especially germane to this paper. The first of these three papers (Theorem 1, page
617) showed for a case of no transaction costs, no distinction between tax rates
for short term versus long term capital investment, and a general price model
for a single risky security that it is optimal to realize losses immediately and
to defer the realization of gains as long as possible. While the first conclusion
is consistent with our results, because it is desirable to collect tax credits, the
second is not. The difference is due, however, to restrictions on the choice of
admissible trading strategies, for Constantinides (1983, page 615) assumes for the
investor the absence of the need to speculate, rebalance his portfolio, or consume
the proceeds.
In his second paper Constantinides 1984; introduced a distinction between
tax rates for short term versus long term capital transaction, and he showed
for a binomial stock price model, where short term positions become long term
positions after the first time period, that it can be optimal to realize a long term
gain if the long term tax rate is sufficiently smaller than the short term rate. On
the other hand, if the two tax rates are the same, then the situation is the same as
in his earlier paper, and all gains should be deferred. However, Constantinides
was again implicitly looking at a restricted class of trading strategies.
Dammon and Spatt 1996; extended the Constantinides 1984; binomial model
approach by allowing N , the number of trading periods before a short term
position becomes a long term position, to be greater than one. Although their
model was designed to study a tax system where short term and long term tax
rates are different, their special case of N = 0 corresponds to symmetric taxation.
In this case their optimal strategy is described by a single non-negative number
xL : defer all gains and losses if and only if the current price divided by the basis
exceeds xL . Single critical number policies can be inconsistent with the results in
our paper, although it is difficult to reconcile the N = 0 version of the Dammon
and Spatt 1996; model with ours because the two models are rather different and
their analysis was not mathematically rigorous.
In addition to taxes, our trading model assumes a fixed transaction cost asso-
ciated with each transaction. Hence under an optimal strategy there will be only
a finite number of transactions in any finite interval of time. At the end of our
paper, however, we eliminate this transaction cost from our model, thereby lead-
ing to a form of continuous trading. We should mention that a moderate number
of recent studies have investigated optimal portfolio problems where there are
transaction costs but taxes are ignored. For example, Assaf, Taksar and Klass
Taxes and transaction costs 139

1988; Davis and Norman 1990; Dumas and Luciano 1991; Morton and Pliska
1995; and Shreve and Soner 1994 investigated markets where there is a single
risky security and a deterministic bank account (see Karatzas 1996; for other
references on this problem).
In this paper we consider a financial market in which there is a single stock
and its price is modeled by a geometric Brownian motion. There is a single tax
rate (i.e., no distinction between short and long term capital gains), and the stock
pays no dividends. There is an investor, and at each point in time he must have
all his money invested in this stock. At any point in time he can sell his stock,
but when he does so he pays a transaction cost, he pays a tax if he has a capital
gain, he receives a tax credit if he has a loss (this is a standard assumption in the
economics literature), and he must take all the remaining money and immediately
invest it back in the same stock. Can such a transaction ever be sensible, even if
the investor is holding a gain? The main point of this paper is to show that the
answer is yes.
We should point out that our model is equivalent to one where the financial
market has several or even many stocks, with each modeled by a geometric
Brownian motion, all having the same parameters. Thus the stocks are identical
from the statistical standpoint, but the investor can invest in only one at a time.
Whenever he chooses to do so, the investor can switch from one stock to another,
thereby incurring the same kinds of transaction costs and tax payments as above.
The reader may find this second perspective more useful than the first.
Our model is spelled out in detail in the following section. The investor’s
objective will be to maximize the long-run growth rate of his portfolio. Since
transaction costs will have a fixed component, transactions will occur only at
discrete times which are Markov renewal times, and so optimal trading strategies
will be described by a stopping rule which governs the time between transactions.
Sections 3 and 4 examine the no-tax case and what we call the high appre-
ciation rate case, respectively. We show that buy-and-hold is optimal in both
cases. This leaves for Sect. 5 what we call the interesting case, where the stock’s
appreciation rate is moderate, relative to volatility. We provide characterizations
of the optimal trading strategy as well as two algorithms for its computation.
We also present a numerical example where it is optimal to realize gains. The
key idea here is that the tax basis should be accounted for in the dynamic pro-
gramming formulation, in which case there may be times when gains should be
realized because the capital gains tax to be paid is compensated by the value of
increasing the basis.
Section 6 analyses various aspects of the optimal solution, including the
effects of the transaction cost and tax rate parameters. Some analysis is analytical,
some is numerical, and all is accompanied by economic interpretations.
When we consider the special case with the transaction cost rate equal to zero,
we get some explicit results. Now trading under the optimal strategy is continuous
under some conditions, the value of the corresponding portfolio grows like a
geometric Brownian motion, and we get an explicit formula for the long-run
growth rate corresponding to continuous trading. This is the subject of Sect. 7.
140 A. Cadenillas, S.R. Pliska

Finally, we conclude in Sect. 8 with some remarks about the economics.

2 The financial market model

We consider a financial market in which the price of a stock is modeled by a

geometric Brownian motion, i.e.,

dSt = St {µdt + σdWt }, (1)

where µ, the drift coefficient, and σ, the volatility parameter, are positive con-
stants. Here W is a standard Brownian motion, which equals zero at time t = 0.
We assume that the long-run growth rate of the stock is positive, i.e.,
1
λ := µ − σ 2 > 0. (2)
2
We define the stochastic process X by
 
St 1 2
Xt := = exp µt + σWt − σ t . (3)
S0 2
Thus Xt represents the price of the stock at time t relative to the tax basis, i.e.
the purchase price.
Suppose that whenever a stock is sold, the fraction α times the monetary
value of the stock is paid as the transaction cost, leaving (1 − α) times the value
of the stock as the amount of money available to invest in the next stock. In
particular, if one share of stock is purchased at time zero and this stock is sold at
time t, then (1 − α)St is available for subsequent investment. Usually, α ∈ [0, 1]
is a small positive number, such as 0.05.
We also suppose that this investor must pay taxes at a rate β ∈ [0, 1]. To
model this, suppose an investor purchases one share of stock at time 0. If τ
denotes the amount of time that the investor owns this stock, then the after-tax
profit per share is
{(1 − α)Sτ − S0 }(1 − β),
the after-tax return over τ is
 
Sτ
(1 − α) − 1 (1 − β),
S0
and the factor by which wealth is increased (i.e., the price relative) is
Sτ
β + (1 − β)(1 − α) = β + (1 − β)(1 − α)Xτ . (4)
S0
Notice that if the profit is positive, then the investor pays a tax equal to β times
the profit, whereas if the sale incurs a loss, then the investor receives a payment
equal to β times the absolute value of the loss. In other words, we assume there
is other income, perhaps from the sale of other stock or from regular income,
so the tax credit from a loss can be realized as an actual cash payment. We
Taxes and transaction costs 141

ignore complications of the tax laws such as distinctions between regular and
investment income, limitations on tax credits, wash rules concerning the length
of time between the time when a stock is sold and then repurchased, and so forth.
We also assume, for simplicity, that tax payments and receipts are made at the
time when the stock transaction occurs, not, for example, at the end of some tax
year.
Now suppose the investor wants to invest in a sequence of stocks, all of
which are identical in that they are all described by (1). The investor does this
by choosing a sequence {τi } of transactions times with 0 = τ0 < τ1 < .... The
initial funds, denoted by V0 , are all invested in the first stock at time τ0 . At time
τ1 , this position is closed out and all the after-tax proceeds are invested in the
second stock. This transaction cycle is repeated at times τ2 , τ3 , · · ·.
Let the random variable
Mi := β + (1 − β)(1 − α)Xτi /Xτi −1
represent the multiplicative factor by which the value of the investment increases
over the i-th transaction cycle, so that V0 M1 M2 ...Mi is the value of the investment
just after the i-th transaction. Note that Mi has the same probability distribution
as the random variable β + (1 − α)(1 − β)X (τi − τi −1 ), because Mi is just the
price relative associated with the i-th transaction cycle.
Problem 2.1. The investor wants to maximize the long-run growth rate of his or
her investment (this is the Kelly criterion). If Vt denotes the value of the investment
at time t, then the long-run growth rate of the investment is
1
E [log Vt ].
lim (5)
t→∞ t

This is consistent with logarithmic utility, so our investor is less risk averse than
most investors seem to be. Our investor plans to live a long time, so the infinite
planning horizon is assumed for simplicity.
It is convenient to denote
g(x ) := log{β + (1 − β)(1 − α)x }, ∀x ∈ (0, ∞). (6)
We also denote by S the class of stopping times for the process X defined in
(3) that take values in (0, ∞), and
S ˜ := {τ ∈ S : E [τ ] ∈ (0, ∞)} (7)
for the subclass of stopping times with positive but finite expectation.
Lemma 2.1. For every η ∈ (0, ∞) and τ ∈ S ˜ ,
Z τ
E [g(Xτ ) − ητ ] = g(1) + E [ Yt dt], (8)
0
where Y = Y (η) is the stochastic process defined by
   2
(1 − α)(1 − β)Xt σ2 (1 − α)(1 − β)Xt
Yt := µ− − η. (9)
β + (1 − α)(1 − β)Xt 2 β + (1 − α)(1 − β)Xt
142 A. Cadenillas, S.R. Pliska

Proof. According to Itô’s formula, for every τ ∈ S ˜ ,

Z τ Z τ  
(1 − α)(1 − β)Xt
g(Xτ ) − ητ = g(1) + Yt dt + σdWt . (10)
0 0 β + (1 − α)(1 − β)Xt

We observe now that the expected value of the stochastic integral in (10) is zero
when E [τ ] < ∞. In fact, if β = 0, then the stochastic integral is just σWτ , and we
may apply Wald’s identity for Brownian motion (see, for instance, Sect. 4.2.8 of
Lipster and Shiryayev (1977)). If β ∈ (0, 1] then the integrand of the stochastic
integral is bounded, so the claim is still valid. Hence, taking the expected value
in (10), we obtain (8). 

Lemma 2.2. If the pairs (τi , Mi ), i ∈ N, are independent and identically dis-
tributed with E [τ1 ] < ∞, then

1 E [log M1 ] E [g(Xτ1 )]
lim E [log Vt ] = = .
t→∞ t E [τ1 ] E [τ1 ]

Proof. It is easy to see that the stochastic process Y of Lemma 2.1 is bounded.
Thus if τ is a stopping time with E [τ ] < ∞, then E [g(Xτ )] < ∞. This implies
E [log M1 ] < ∞. Applying renewal theory (see Theorem 3.6.1 of Ross (1996)),
we get the conclusion of this Lemma. 

In view of (6) and Lemma 2.2, the investor’s portfolio management problem
can be restated as follows:

Problem 2.2. Select τ̂ ∈ S ˜ that maximizes

E [log{β + (1 − β)(1 − α)Xτ }]

J (τ ) := . (11)
E [τ ]

Problem 2.1 has been transformed by Markov renewal theory (see Ross
(1996)) into Problem 2.2, a problem of selecting a single stopping time. Thus,
after each transaction it is as if the process X starts fresh with a value of one, so
Xt always represents the stock price relative to the tax basis, i.e., the most recent
purchase price.
Since the long-run growth rate of our geometric Brownian motion stock
is λ = µ − 12 σ 2 , this will therefore equal the long-run growth rate of the
investment if the investor employs a buy-and-hold strategy corresponding to
ζ := inf {t ≥ 0 : Xt ∈/ (0, ∞)} = ∞. We shall prove in Sect. 6 (see Propo-
sition 6.1) that λ is also the long-run growth rate corresponding to any cut-
losses-short-and-let-profits-run strategy. These are strategies of the form ξ(a) :=
inf {t ≥ 0 : Xt ∈
/ (a, ∞)} , where 0 < a < 1. The specific form of τ̂ will vary
according to the parameters, as will be seen in the sections that follow.
Taxes and transaction costs 143

3 No taxes

In this section, we assume β = 0 and we verify an intuitive result: buy-and-hold,

and thus cut-losses-short-and-let-profits-run, strategies are optimal.
Theorem 3.1. If β = 0, then J (τ ) ≤ λ for every τ ∈ S ˜ , with strict inequality
when α ∈ (0, 1].
Proof. We see that if β = 0 and η = λ in Lemma 2.1, then
1
Yt = Yt (λ) = µ − σ 2 − λ = 0.
2
Thus, for arbitrary τ ∈ S ˜ ,

E [g(Xτ ) − λτ ] = g(1) = log{β + (1 − α)(1 − β)} ≤ 0,

that is, J (τ ) ≤ λ. Both of these inequalities are strict if α > 0. 

Since λ is the long-run growth rate of our geometric Brownian motion stock,
Theorem 3.1 says that when there are no taxes, it is optimal to buy-and-hold the
stock. In view of Proposition 6.1, it is also optimal to cut-losses-short-and-let-
profits-run.

4 The case of a stock with high appreciation rate

In this section, we assume β > 0 and consider the case in which

µ ≥ σ2 ,

and we again reach the same conclusion as in Sect. 3: buy-and-hold as well as

cut-losses-short-and-let-profits-run strategies are optimal.
Theorem 4.1. If µ ≥ σ 2 and β > 0, then J (τ ) ≤ λ for every τ ∈ S ˜ , with strict
inequality when α ∈ (0, 1].
Proof. We observe that if we take η = λ in Lemma 2.1, then
   2  
(1 − α)(1 − β)Xt σ2 (1 − α)(1 − β)Xt σ2
Yt = µ− − µ−
β + (1 − α)(1 − β)Xt 2 β +(1 − α)(1 − β)Xt 2
   
−β σ 2 β 2 + 2β(1 − α)(1 − β)Xt
= µ+
β + (1 − α)(1 − β)Xt 2 [β + (1 − α)(1 − β)Xt ]2
β 2 (−µ + 12 σ 2 ) + β(1 − α)(1 − β)Xt (−µ + σ 2 )
= .
[β + (1 − α)(1 − β)Xt ]2
The numerator in the above equation is negative because µ ≥ σ 2 > 12 σ 2 . Hence,
Yt is negative, and we obtain from Lemma 2.1

E [g(Xτ ) − λτ ] ≤ g(1) = log{β + (1 − α)(1 − β)} ≤ 0. 

144 A. Cadenillas, S.R. Pliska

5 The interesting case

In this section we assume that

1 2
0 < α < 1, 0 < β < 1, and σ < µ < σ2 .
2
Let θ̂ be the value of Problem 2.2, i.e.,

θ̂ := sup J (τ ),
τ ∈S ˜

and observe that θ̂ ≥ λ > 0. Since the mathematical techniques to solve Problem
2.2 are now more complicated than in the previous cases, we start by solving
Problem 2.2 for a subclass of stopping times. We denote by

T ˜ := {τ (a, b) : 0 < a < 1 < b < ∞}

the subclass of stopping times with

τ (a, b) = inf{t ∈ [0, ∞) : Xt 6∈ (a, b)}

= inf{t ∈ [0, ∞) : λt + σWt 6∈ (log a, log b)}, (12)

where λ = µ − 12 σ 2 . Note that T ˜ ⊂ S ˜ .

Theorem 5.1. The solution of Problem 2.2 in the subclass T ˜ is given by

sup J (τ ) = sup h(a, b), (13)

τ ∈T ˜ 0<a<1<b<∞

where

h(a, b) = (14)
−( 2λ2 ) −( 2λ2 )
log{β+(1−β)(1−α)a}[b σ −1] + log{β +(1 − β)(1 − α)b}[1−a σ ]
λ .
−( 2λ2 ) −( 2λ2 )
(log a)b σ −(log b)a σ +(log b) − (log a)
Proof. We observe that, for every 0 < a < 1 < b < ∞,
E [log{β + (1 − β)(1 − α)Xτ (a,b) }]
J (τ (a, b)) := .
E [τ (a, b)]
g(a)P {Xτ (a,b) = a} + g(b)P {Xτ (a,b) = b}
= , (15)
E [τ (a, b)]
where g is given by (6). The probabilities and expectation here can be computed
by analysing an ordinary Brownian motion with drift. According to Theorem
7.5.2 of Karlin and Taylor (1975),

P {Xτ (a,b) = b} = P {λτ (a, b) + σWτ (a,b) = log b}

1 − a −( σ2 )
2λ

= , (16)
b −( σ2 ) − a −( σ2 )
2λ 2λ
Taxes and transaction costs 145

and similarly
b −( σ2 ) − 1
2λ

P {Xτ (a,b) = a} = 2λ . (17)

b −( σ2 ) − a −( σ2 )
2λ

On the other hand, by a standard computation of the expected first passage time
(see, for instance, Sect. 8.4 of Ross (1996)),

(log a)b −( σ2 ) − (log b)a −( σ2 ) + (log b) − (log a)

2λ 2λ

E [τ (a, b)] = . (18)

λ[b −( σ2 ) − a −( σ2 ) ]
2λ 2λ

The theorem follows from equations (15)-(18). 

Example 5.1 Let us consider a numerical example in which

µ = 0.065, σ = 0.3, α = 0.02 and β = 0.3. (19)

We used Mathematica to solve equation (13) and obtained

a = 0.303293, b = 7.18583, θ = 0.0223115. (20)

Note that with λ = µ − 12 σ 2 = 0.02, this trading strategy beats buy-and-hold, as

well as any cut-losses-short-let-profits-run strategy.
The reader might find this kind of result counter-intuitive, thinking that it
is foolish to throw away money in the form of a capital gains tax. However, it
should be kept in mind that there will be sample paths where the investor will be
strictly better off by rebalancing after a gain than by sticking to a buy-and-hold
or a cut-losses-short-let-profits-run strategy. For example, suppose for simplicity
that α = 0.02, β = 0.3, a = 0.3, and b = 7, and consider a sample path where
the price of the stock rises from $1 per share to $7 per share and then falls back
to $2.1 per share, so a buy-and-hold strategy starting with one share will end
up worth $2.1. On the other hand, if the investor establishes a new tax basis
when the stock price (that is, the process X ) is at 7, then the capital gain is ($7
X (1-0.02)) - $1 = $5.86, the tax paid is $5.86 X 0.3 = $1.758, and the money
left after rebalancing is $7 X 0.98 - $1.758 = $5.102. Then, when the price of
the stock goes down from $7 per share to $2.1 per share, the investment goes
down to $5.102 X (2.1/7) = $1.5306, the capital loss is $5.102 - ($1.5306 X
(1-0.02))= $3.602012, the tax credit is $3.602012 X 0.3 = $1.0806036, and the
investment after the second rebalance is (see also expression (4) for the price
relative) worth $1.5306 X 0.98 + $1.0806036 = $2.5805916. Therefore, not only
will the investor have more money than with buy-and-hold, but also the updated
tax basis will be 2.1 instead of only 1 under the buy-and-hold strategy!
It is important to consider whether more general stopping rules provide su-
perior results. It will turn out, however, that one cannot do better than by taking
a stopping rule in T ˜ .
To solve Problem 2.2 for the class S ˜ of more general stopping times, we
shall transform the problem of maximizing (11) to an equivalent and traditional
optimal stopping problem. We observe that Problem 2.2 has a solution if and
only if
146 A. Cadenillas, S.R. Pliska

J (τ ) ≤ θ̂, ∀τ ∈ S ˜ ,
with equality for some τ̂ ∈ S ˜ . Equivalently, Problem 2.2 has a solution if and
only if
E [g(Xτ ) − θ̂τ ] ≤ 0, ∀τ ∈ S ˜ , (21)
with equality for some τ̂ ∈ S ˜ . We recognize that if θ̂ is a fixed constant, then
the problem of maximizing the left-hand-side of (21) is a traditional optimal
stopping problem. Thus we are interested in solving the following equivalent
problem:
Problem 5.1 First, for each θ ≥ λ, solve the optimal stopping problem with value

H (θ) := sup E [g(Xτ ) − θτ ]. (22)

τ ∈S ˜

Second, select θ̂ ≥ λ such that

H (θ̂) = 0. (23)

To apply the standard theory of dynamic programming to Problem 5.1, we

start by defining, for each θ ≥ λ, the function s : (0, ∞) 7→ < by

s(x ) = sθ (x ) := sup E x [g(Xτ ) − θτ ], (24)

τ ∈S ˜

where E x denotes expectation conditional on the event {X0 = x }. Thus sθ (x )

represents the maximum expected net payoff for stopping, with initial state X0 =
x , with reward g(Xτ ) for stopping at time τ in state Xτ , and with waiting cost rate
θ. We observe that selecting θ̂ satisfying (23) is the same as finding θ̂ satisfying

sθ̂ (1) = 0. (25)

Let us denote

C := {x ∈ (0, ∞) : s(x ) > g(x )} (26)

Σ := {x ∈ (0, ∞) : s(x ) = g(x )}. (27)

Here C is the continuation region and Σ is the stopping region. According to

Sect. 3.8 of Shiryayev 1978; the payoff s is a solution of the generalized Stefan
problem:
ds 1 2 2 d 2 s
µx + σ x = θ if x ∈ C (28)
dx 2 dx 2
and
s(x ) = g(x ) if x ∈ Σ. (29)
As pointed out in Sect. 3.8.2 of Shiryayev (1978), Stefan problems usually have
many solutions. To find the solution that coincides with the payoff s, we shall
make two conjectures. The first one is that the smooth-fit condition holds, i.e.,
that the derivatives of s and g are equal on the boundary of Σ. The second one
Taxes and transaction costs 147

is that C is an open interval. In Theorem 5.1, we shall prove rigorously that

both conjectures are valid.
The general solution of the ordinary differential equation (28) is (see, for
instance, Sect. 3.6 of Boyce and DiPrima (1992))
2µ θ
s(x ) = c + dx (1− σ2 ) + log x , (30)
(µ − 12 σ 2 )
where c and d are real numbers.
Let us conjecture that the smooth-fit condition holds and that C = (a, b),
where 0 < a < 1 < b < ∞. Then we may find θ, c, d , a, and b from the
following equations:

s(1) = 0, (31)
s(a) = g(a), (32)
s(b) = g(b), (33)
s 0 (a) = g 0 (a), (34)
0 0
s (b) = g (b). (35)

Equations (34)-(35) are the smooth-fit condition. Equations (31)-(35) are equiv-
alent to

c+d = 0, (36)
(1− 2µ2 ) θ
c + da σ + log a = log{β + (1 − α)(1 − β)a}, (37)
(µ − 12 σ 2 )
2µ θ
c + db (1− σ2 ) + log b = log{β + (1 − α)(1 − β)b}, (38)
(µ − 12 σ 2 )
2µ 2µ θ 1 (1 − α)(1 − β)
d (1 − 2 )a (− σ2 ) + = , (39)
σ (µ − 2 σ ) a
1 2 [β + (1 − α)(1 − β)a]
2µ (− 2µ2 ) θ 1 (1 − α)(1 − β)
d (1 − )b σ + = , (40)
σ 2
(µ − 2 σ ) b
1 2 [β + (1 − α)(1 − β)b]

respectively. For computational purposes, it is convenient to write equations

(36)-(40) in the form
c = −d , (41)
( )
(1 − α)(1 − β) θ 1 2µ −1 ( 2µ2 )
d= − (1 − ) a σ , (42)
[β + (1 − α)(1 − β)a] (µ − 2 σ 2 ) a
1 σ2

2µ
(1 − α)(1 − β)a ( σ2 ) θ 2µ
− a ( σ2 −1)
[β + (1 − α)(1 − β)a] (µ − 12 σ 2 )
2µ
(1 − α)(1 − β)b ( σ2 ) θ 2µ
= − b ( σ2 −1) , (43)
[β + (1 − α)(1 − β)b] (µ − 2 σ )
1 2
148 A. Cadenillas, S.R. Pliska

( )
(1 − α)(1 − β) θ 1 2µ −1  2µ

− (1 − ) a − a ( σ2 )
[β + (1 − α)(1 − β)a] (µ − 12 σ 2 ) a σ 2

θ
+ log a − log{β + (1 − α)(1 − β)a} = 0, (44)
(µ − 12 σ 2 )

( )
(1 − α)(1 − β) θ 1 2µ −1  ( 2µ2 )

− (1 − ) b − b σ
[β + (1 − α)(1 − β)b] (µ − 12 σ 2 ) b σ2
θ
+ log b − log{β + (1 − α)(1 − β)b} = 0. (45)
(µ − 12 σ 2 )

To solve this system of equations, we first find a, b, and θ from equations (43)-
(45). Then, we obtain c and d from equations (41)-(42).
It is important to realize that, up to this point, the existence of a solution
to (41)-(45) is not enough to guarantee that this is also the solution of Problem
5.1, due to nonuniqueness. However in Theorem 5.2, we shall give additional
conditions under which the solution of this system of equations is indeed a
solution of Problem 5.1.

Theorem 5.2. Let a, b, c, d , and θ, where 0 < a < 1 < b < ∞, θ ≥ λ, and
c, d ∈ <, be a solution of equations (41)-(45). We define the function V by

f (x ) if x ∈ (a, b),
V (x ) := (46)
g(x ) if x 6∈ (a, b),

where
2µ θ
f (x ) = c + dx (1− σ2 ) + log x . (47)
(µ − 12 σ 2 )
If
f (x ) ≥ g(x ), ∀x ∈ (a, b), (48)
1
β{µ − 2θ − (µ2 − 2θσ 2 ) 2 }
a≤ , (49)
2(1 − α)(1 − β)(θ − µ + 12 σ 2 )
and
1
β{µ − 2θ + (µ2 − 2θσ 2 ) 2 }
b≥ , (50)
2(1 − α)(1 − β)(θ − µ + 12 σ 2 )
then
V (x ) = sup E x [g(Xτ ) − θτ ]. (51)
τ ∈S ˜

Furthermore, the solution of Problem 5.1 (or equivalently, of Problem 2.2) is

τ̂ = τ (a, b) := inf{t ∈ [0, ∞) : Xt 6∈ (a, b)}. (52)

Taxes and transaction costs 149

Proof. Let us consider the function q : [0, ∞) 7→ < defined by

   2
(1 − β)(1 − α)x 1 (1 − β)(1 − α)x
q(x ) := µ − σ2 −θ
β + (1 − α)(1 − β)x 2 β + (1 − α)(1 − β)x
{(1 − α)2 (1 − β)2 (µ− 12 σ 2 − θ)}x 2 + {β(1 − α)(1 − β)(µ−2θ)}x −θβ 2
= .
{β +(1 − α)(1 − β)x }2
(53)

We observe that y ∈ < satisfies q(y) ≤ 0 if and only if

1
{(1 − α)2 (1 − β)2 (θ − µ + σ 2 )}y 2 + {β(1 − α)(1 − β)(2θ − µ)}y + θβ 2 ≥ 0.
2
This, in turn, is true if and only if
1 1
β{µ − 2θ − (µ2 − 2θσ 2 ) 2 } β{µ − 2θ + (µ2 − 2θσ 2 ) 2 }
y≤ or y ≥ .
2(1 − α)(1 − β)(θ − µ + 2 σ )
1 2
2(1 − α)(1 − β)(θ − µ + 12 σ 2 )
In particular, from conditions (49)-(50), we see that

q(y) ≤ 0, ∀y 6∈ (a, b). (54)

Next, define the adapted stochastic process Z by

Zt = V (Xt ) − θt. (55)

In the region D := {(t, w) ∈ [0, ∞) × Ω : Xt (ω) 6∈ (a, b)}, one has

Zt = log{β + (1 − β)(1 − α)Xt } − θt,

so applying Itô’s formula we obtain

 
(1 − β)(1 − α)Xt
dZt = q(Xt )dt + σdWt .
β + (1 − α)(1 − β)Xt
According to equation (54), the term multiplying dt is nonpositive. In addition,
the term multiplying dWt is bounded. Thus Z is a supermartingale in the region
D.
In the region E := {(t, w) ∈ [0, ∞) × Ω : Xt (ω) ∈ (a, b)}, one has
(1− 2µ2 ) θ
Zt = c + dXt σ
+ log Xt − θt,
(µ − 12 σ 2 )
so applying Itô’s formula we obtain
( )
2µ (1− σ2µ2 ) θ
dZt = d (1 − 2 )Xt + σdWt .
σ (µ − 12 σ 2 )

We observe that the term multiplying dWt is bounded. Thus Z is a martingale in

the region E .
Hence the process Z is a supermartingale in [0, ∞) × Ω. Furthermore,
150 A. Cadenillas, S.R. Pliska

Z t
Zt = Z0 + q(Xs )1{Xs 6∈(a,b)} ds
0
Z t  
(1 − β)(1 − α)Xs
+ 1{Xs 6∈(a,b)}
0 β + (1 − α)(1 − β)Xs
" # )
2µ (1− σ2µ2 ) θ
+ d (1 − 2 )Xs + 1{Xs ∈(a,b)} σdWs .
σ (µ − 12 σ 2 )
Since the first integrand is nonpositive,
Z τ 
E x
q(Xs )1{Xs 6∈(a,b)} ds ≤ 0, ∀τ ∈ S ˜ ,
0
with equality for τ = τ (a, b). In addition,
Z τ  
x (1 − β)(1 − α)Xs
E 1{Xs 6∈(a,b)}
0 β + (1 − α)(1 − β)Xs
" # ) #
2µ (1− σ2µ2 ) θ
+ d (1 − 2 )Xs + 1{Xs ∈(a,b)} σdWs = 0, ∀τ ∈ S ˜ ,
σ (µ − 12 σ 2 )
because the integrand of this stochastic integral is bounded. Thus,
E x [Zτ ] ≤ E x [Z0 ], ∀τ ∈ S ˜ ,
with equality for τ = τ (a, b). From condition (48), we then obtain
E x [g(Xτ ) − θτ ] ≤ E x [V (Xτ ) − θτ ] = E x [Zτ ]
≤ E x [Z0 ] = V (x ),
with equality for τ = τ (a, b). Therefore,
sup E x [g(Xτ ) − θτ ] = V (x ) = E x [g(Xτ (a,b) ) − θτ (a, b)]. 
τ ∈S ˜

Example 5.2 Let us consider the same data as in Example 5.1. Using Maple, we
obtained from equations (43)-(45) that
a = 0.3032934964, b = 7.186001883, θ = 0.02231148251. (56)
Then we obtained from equations (41)-(42)
d = 0.9348432672, c = −0.9348432672. (57)
It is easy to verify that conditions (48)-(50) are satisfied. Therefore, the solution
of Problem 5.1, or equivalently Problem 2.2, is given by the stopping time of (52),
with a, b, and θ given by (56).
Remark 5.1 We conjecture for our “interesting case” that conditions (48)-(50)
are satisfied for any solution of the system of equations (41)-(45). Indeed, these
conditions are satisfied not only for the numerical example of this section, but also
for a variety of other numerical examples that we tried, including the numerical
examples used for the figures in Sect. 6. However, we were unable to prove this
conjecture. The main difficulty is that there is not an explicit solution for the
system (41)–(45).
Taxes and transaction costs 151

6 Numerical analysis and discussion

In this section we collect some basic properties of the optimal solution for the
interesting case of the preceding section. We also present some numerical results
showing how the optimal solution varies with respect to selected parameters. Our
aim is to develop and verify some economic intuition.
Our first category of results concerns the limiting behavior of the function
h : (0, 1) × (1, ∞) 7→ < (see Theorem 5.1) near the boundary of its domain,
namely, the region where 0 < a < 1 and 1 < b < ∞. Our first such result is
Proposition 6.1. Any cut-losses-short-and-let-profits-run strategy gives the same
performance as buy-and-hold. That is,

lim h(a, b) = λ, ∀0 < a < 1. (58)

b→∞

Proof. To see this, note that

2λ
γ := − < 0,
σ2
so b γ decreases to 0 as b increases to ∞. Thus the numerator in the formula for
h (see equation (14)) behaves like log{β + (1 − α)(1 − β)b}(1 − a γ )λ + constant,
whereas the denominator behaves like (1 − a γ ) log{b} + constant. Both diverge
to −∞, but using L’Hospital’s rule we easily obtain the indicated result. 
For our next results we need to study the denominator in the formula for h;
we denote this d (a, b). Since ∂d
∂a (1, b) > 0 for all b > 1 and since d (1, b) = 0, it
follows that, for each fixed b, d (a, b) < 0 for all a < 1 in some neighborhood of
1. Of the two terms in the numerator of h, the first one converges to a positive
constant as a ↑ 1, so when divided by d (a, b), the limit as a ↑ 1 is −∞.
Meanwhile, the second term when divided by d (a, b) converges to a constant as
a ↑ 1 by L’Hospital’s rule. Hence

lim h(a, b) = −∞, ∀b > 1.

a↑1

In a similar fashion, we are able to conclude that

lim h(a, b) = −∞, ∀a < 1.

b↓1

The intuition here is simply that continuous trading is suicidal as long as (as is
being assumed) the transaction cost factor α is strictly positive.
For a final result of this type, fix b > 1 and consider the denominator d (a, b)
as a decreases to 0; it behaves like (1 − a γ ) log b, because this term dominates
(b γ −1) log a. Similarly, the numerator behaves like λ(1−a γ ) log{β +(1−α)(1−
β)b}, so we are able to conclude the following:

h(0, b) = λ log{β + (1 − α)(1 − β)b}/ log b, ∀b > 1.

152 A. Cadenillas, S.R. Pliska

Fig. 1. The function h

Note that 0 < β + (1 − α)(1 − β)b < b for all b > 1, so h(0, b) equals a positive
number less than λ if β +(1−α)(1−β)b > 1, whereas in the opposite case h(0, b)
is some negative number. In either event, h(0, b) < λ. Some intuition: if you
never cut your losses short, then you will be worse off than under buy-and-hold.
Our results on the behavior of h near the boundary of its domain are consistent
with a numerical analysis of our baseline case where µ = 0.065, σ = 0.3, α =
0.02, β = 0.3, and, thus, λ = 0.02. Figure 1 shows the graph of h for the region
where 0.1 < a < 0.99 and 1.01 < b < 35. The surface looks very flat over most
of this region, and it is difficult to see where h is maximized. But if we look
more closely at a smaller region, then we can see more curvature. For example,
Fig. 2 shows h for the region where 0.2 < a < 0.4 and 7.0 < b < 7.4, a small
region where the maximizing values of a and b are found.
In the remainder of this section we shall examine how the optimal solution
is affected by three parameters: λ, β, and α. To see how the optimal long-
run growth rate θ̂ depends on the stock’s long-run growth rate λ, let us denote
∀λ > 0, t ≥ 0, τ ∈ S ˜ :
Xtλ = exp{λt + σWt }
E [log{β + (1 − β)(1 − α)Xτλ }]
J (τ ; λ) =
E [τ ]
Taxes and transaction costs 153

Fig. 2. The function h

θ̂(λ) = sup J (τ ; λ).

τ ∈S ˜

We observe that P {Xτλ1 < Xτλ2 } = 1 for all τ ∈ S ˜ and 0 < λ1 < λ2 , so
J (τ ; λ1 ) < J (τ ; λ2 ) for all τ ∈ S ˜ and 0 < λ1 < λ2 . Hence

Proposition 6.2. The optimal long-run growth rate of the investment is a strictly
increasing function of the long-run growth rate of the stock. That is,

θ̂(λ1 ) < θ̂(λ2 ), ∀λ1 < λ2 . (59)

This is accordance with our intuition: the bigger the λ, the better the stock, and
thus the better the optimal result.
We now turn to a study of how the optimal solution is affected by the tax
rate parameter β. We are unable to obtain analytical results for this, so we shall
need to be content with numerical analyses.
Figure 3 shows the optimal long-run growth rate of the investment versus
β, with the other parameters taking the usual baseline values. As expected, this
graph is increasing to a point and then is decreasing. This is as expected because
we know the optimal long-run growth rate is the same as with buy-and-hold both
when β = 0 as well as when β = 1 (in this latter case, after any trade you can
never have more money than what you had at the time of your last transaction,
so buy-and-hold is optimal), and for some intermediate values of β we do better
than buy-and-hold. In particular,
154 A. Cadenillas, S.R. Pliska

Fig. 3. Optimal theta as a function of beta

Remark 6.1. For our baseline values, there is an intermediate value of β that
maximizes the optimal long-run growth rate of the investment.
We conclude this section by looking at the transaction cost parameter α. To
see how θ̂ depends on α, let us denote ∀α ∈ [0, 1], t > 0, τ ∈ S ˜ :

E [log{β + (1 − β)(1 − α)Xτ }]

J (τ ; α) =
E [τ ]
θ̂(α) = sup J (τ ; α).
τ ∈S ˜

We observe that ∀τ ∈ S ˜ , 0 < α1 < α2 :

E [log{β + (1 − β)(1 − α1 )Xτ }] > E [log{β + (1 − β)(1 − α2 )Xτ }],

so J (τ ; α1 ) > J (τ ; α2 ) for all τ ∈ S ˜ and 0 < α1 < α2 . Hence,

Proposition 6.3. The optimal long-run growth rate of the investment is a strictly
decreasing function of the transaction cost. That is,

θ̂(α1 ) > θ̂(α2 ), ∀0 < α1 < α2 . (60)

Taxes and transaction costs 155

Fig. 4. Continuation region vs. alpha

In other words, higher transaction cost rates are always worse.

Figure 4 shows for baseline parameters the optimal values of a and b versus
α. The optimal value of a is a decreasing function of α, because as α goes up it
becomes more expensive to cut losses short, so you should do so less often. For
a similar reason, the optimal value of b will be increasing with respect to α.

7 The interesting case with no transaction costs

In this section we examine some issues and seek some explicit results for the
case where the transaction costs are zero. From equation (60) and Fig. 4, when
α ↓ 0 for our baseline case, the optimal growth rate θ̂ increases, a increases,
and b decreases. Although it is not obvious from these figures, one possibility is
that a ↑ 1 and b ↓ 1 as α ↓ 0. In other words, some form of continuous trading
might become optimal. For example, suppose α = 0 and the investor trades at
times n1 < n2 < · · · < nk < · · · < ∞, and consider the process describing the
portfolio’s value as n ↑ ∞.
Fix n ∈ N and denote, for each k ∈ {1, 2, . . . , n} and s ≥ (k −1)
n ,

Xs
Yk (s) :=
X (k −1)
n
156 A. Cadenillas, S.R. Pliska

    
σ2 (k − 1)
= exp µ− s− + σ Ws − W (k −1) ,
2 n n

and, for each a ∈ <,

dac := greatest integer ≤ a.

Then the value of the investment at time t is

  ( )
 X2/n Xt
Vn (t) = β + (1 − β)X1/n β + (1 − β) · · · β + (1 − β)
X1/n X dtnc
Ydtnc    n

k 
= β + (1 − β)Yk β + (1 − β)Ydtnc+1 (t) ,
k =1 n

because dtnc
n ≤t <
dtnc+1
n .
The next result says that the sequence of adapted stochastic processes {Vn }
defined by Vn := {Vn (t) : t ∈ [0, ∞)} converges in distribution as n ↑ ∞ to a
geometric Brownian motion with parameters µ̃ := (1−β)µ and σ̃ := (1−β)σ (the
reader may consult, for instance, Ethier and Kurtz 1986; for the definition and
some results on convergence in distribution of stochastic processes). In particular,
for arbitrary t ∈ [0, ∞), the sequence of random variables {Vn (t)} converges in
distribution to a lognormal random variable.
Proposition 7.1. When α = 0,
d
Vn −→ GBM(µ̃, σ̃).
d
Here −→ denotes convergence in distribution and GBM(µ̃, σ̃) denotes a geometric
Brownian motion with parameters µ̃ := (1 − β)µ and σ̃ := (1 − β)σ.
Proof. See Appendix. 
It follows from Proposition 7.1 that with this continuous trading strategy the
value of the investment has a long-run growth rate equal to

1 1
µ̃ − σ̃ 2 = (1 − β)µ − (1 − β)2 σ 2 .
2 2
It is tempting to conjecture that this form of continuous trading is optimal when
α = 0, because µ̃ − 12 σ̃ 2 is consistent with the following limiting value of h(a, b)
as a ↑ 1 and b ↓ 1.
Proposition 7.2. When α = 0,

λ(1 − β)(γ − β) 1
ρ := lim h(a, b) = = (1 − β)µ − (1 − β)2 σ 2 , (61)
a↑1,b↓1 γ 2

where γ := − 2λ
σ2 .
Taxes and transaction costs 157

Proof. For a small number  > 0 we have

h(1 − , 1 + )

log{β + (1 − β)(1 − α)(1 − )}[(1 + )γ − 1]
= λ
log(1 − )[(1 + )γ − 1)] + log(1 + )[1 − (1 − )γ ]

log{β + (1 − β)(1 − α)(1 + )}[1 − (1 − )γ ]
+ .
log(1 − )[(1 + )γ − 1)] + log(1 + )[1 − (1 − )γ ]
By doing a Taylor series expansion about  = 0, we see that the common de-
nominator in the above two terms is equal to

(−γ − (γ − 1)γ)3 + O(5 ) = −γ 2 3 + O(5 ).

Similarly, the sum of the numerators can be expressed as

1 1
λ [(1 − β)(β − 1)γ − (1 − β)(γ − 1)γ + (β − 1)(γ − 1)γ]3 + O(4 )
2 2
= λ(1 − β)(β − γ)γ3 + O(4 ).

Applying L’Hospital’s rule, we obtain the desired limits as  ↓ 0. The last equality
in the Proposition is obtained immediately by substitution. 
Note that for our baseline case when α = 0 we have
0.02(1 − 0.3)(− 49 − 0.3)
ρ= = 0.02345.
− 49

Is this the optimal long-run growth rate when α = 0? We shall answer this ques-
tion shortly, but first consider the following sufficient condition for continuous
trading to be optimal.
Theorem 7.1. Suppose that α = 0, µ = (1 − β)σ 2 , and β < 0.5, Then, for every
τ ∈ S ˜ , we have
J (τ ) ≤ ρ = J (0). (62)
Proof. We are assumming β < 0.5 only to guarantee that σ 2 /2 < µ = (1−β)σ 2 <
σ 2 , so that we are in the interesting case. As in the proof of Theorem 4.1, it
suffices to apply Lemma 2.1 with α = 0 and η = ρ, and show that the process
Y of Lemma 2.1 is nonpositive. Indeed, since g(1) = 0 when α = 0, this would
give E [g(Xτ ) − ρτ ] ≤ 0 for every τ ∈ S ˜ , with equality for τ = 0.
As in the proof of Theorem 5.2, we observe that P {Yt ≤ 0 : ∀t ≥ 0} = 1 if
and only if
   2
(1 − β)(1 − α)x 1 (1 − β)(1 − α)x
µ − σ2 − ρ ≤ 0, ∀x > 0,
β + (1 − α)(1 − β)x 2 β + (1 − α)(1 − β)x
which in turn is true if and only if

f (x ) := (1 − β)2 (λ − ρ)x 2 + {β(1 − β)(µ − 2ρ)}x − ρβ 2 ≤ 0, ∀x > 0,

158 A. Cadenillas, S.R. Pliska

where we have introduced f to denote the quadratic function on the left-hand-

side. Since λ < ρ by assumption, this function is concave. We observe that
f (0) = −ρβ 2 < 0 and f (1) = 0,
so x = 1 is always a root of the quadratic function f . In addition, the discriminant
of this quadratic function is
D = {β(1 − β)(µ − 2ρ)}2 + 4(1 − β)2 (λ − ρ)ρβ 2
= (µ − (1 − β)σ 2 )2 ≥ 0,
so
f (x ) ≤ 0, ∀x > 0 ⇐⇒ µ = (1 − β)σ 2 .
Hence, the condition µ = (1 − β)σ 2 implies that P {Yt ≤ 0 : ∀t ≥ 0} = 1, so (62)
holds. 
Thus only in a very special case do we know for sure that continuous trading
of the form examined in Proposition 7.1 is optimal when α = 0. In other cases,
it might not be optimal. For example, some simple algebra reveals that
ρ≥λ ⇐⇒ β ≤ 1 + γ,
a condition which holds for our baseline case as well as in Theorem 7.1. Hence
β > 1+γ means buy-and-hold does better than the trading strategy corresponding
to ρ.
If ρ is the optimal long-run growth rate for all β ≤ 1+γ , then it is interesting
to note there is a simple formula for the tax rate β̂ which maximizes ρ. Some
simple calculus reveals this and the corresponding long-run growth rate ρ̂ to be
1 µ −λ(1 − γ)2 µ2
β̂ = [1 + γ] = 1 − 2 and ρ̂ = = .
2 σ 4γ 2σ 2
For our baseline case we have β̂ = 5/18, which is in some accordance with
Fig. 3, and ρ̂ = 0.023472.
We now return to the question posed just before Theorem 7.1: for our baseline
case µ = 0.065, σ = 0.3, and β = 0.3, is ρ = 0.02345 corresponding to continuous
trading the optimal long-run growth rate when α = 0? Note that (1−β)σ 2 = 0.063
whereas µ = 0.065, so the sufficient condition of Theorem 7.1 is not safisfied.
It turns out that ρ is not the optimal objective value, because, for example, we
have computed
h(0.9999999, 1.3) = 0.02731136757 > ρ = 0.02345.
We thus conjecture that for many situations when α = 0 the optimal strategy
is defined by taking a = 1 and b > 1, with the interpretation that a singular
control (see Cadenillas and Haussmann 1994; and Davis and Norman 1990; for
the theory and an example of this) is now employed at the level a = 1, which acts
like a reflecting barrier to keep X from dropping below the level 1. It remains for
a future research project to provide a better understanding of the optimal trading
strategies when α = 0.
Taxes and transaction costs 159

8 Concluding remarks

That many readers will find some of our results counterintuitive is strong evi-
dence that (1) the results are surprising in the first place, (2) we have not been
completely successful in adding satisfactory economic explanations, and (3) there
is room for a lot more research on this problem. It is rather clear why the tax-
paying investor should cut losses short: the government is willing to subsidize
losses by paying tax credits, so the investor should take advantage of these cred-
its. It is also understandable why the strategy of cutting losses short but letting
profits run will not beat the long-run growth rate of buy-and-hold. But our result
that it can be optimal to cut your profits short is probably the most difficult to
come to terms with as far as economic intuition is concerned. Our best expla-
nation for this, as was stated several times, is that the increase in value due to
bringing the basis up to the current stock price can be greater than the decrease
due to the tax payment. This is in spite of the fact that the cumulative tax paid out
is likely to exceed the cumulative tax credit received. Simple calculations show
that from the expected value standpoint the government is not a net provider of
funds under the optimal strategy.
What seems to be happening can perhaps best be understood by considering
the continuous trading strategy of Section 7 (with α = 0): you are altering the
statistical properties of the after-tax value process in a clever way. Although the
net payments cause the appreciation rate to be reduced from µ to (1 − β)µ (see
Proposition 7.1), there is also a reduction in the volatility from σ to (1−β)σ, and
consequently the resulting long-run growth rate can be higher. This suggests that
even when transactions occur at discrete epochs, the tax system can be used to
reduce the volatility of the after-tax portfolio, thereby enhancing the portfolio’s
long-run growth rate.
It is important to remark that our results are not restricted to our studied
criterion of maximizing the long-run growth rate. Using numerical methods it is
not difficult to find examples where one should cut gains short (in order to reset
the basis) with a finite planning horizon and different utility functions. Hence
our results suggest there could be circumstances when investors would be better
off putting funds into taxable rather than tax-exempt accounts (see Remark 6.1).
Indeed, investors might even be worse off if policy makers were to eliminate
taxes on long term capital investments. Of course, whether these notions are true
for realistic, multi-security situations is a subject for future research.

Appendix.
Proof of Proposition 7.1

We observe that
dtnc
X   
k 
log Vn (t) = log β + (1 − β)Yk + log β + (1 − β)Ydtnc+1 (t) .
n
k =1
160 A. Cadenillas, S.R. Pliska

It suffices to prove that the sequence of adapted stochastic processes {log Vn }

converges in distribution to a Brownian motion with constant drift µ̃ − 12 σ̃ 2 and
constant volatility σ̃ 2 .
Let us consider the sequence of stochastic processes {Mn } defined by
Mn (t)
dtnc
X   
k 
:= log β + (1 − β)Yk + log β + (1 − β)Ydtnc+1 (t)
n
k =1
dtnc Z k
(   2 )
X n (1 − β)Yk (s) 1 2 (1 − β)Yk (s)
− µ− σ ds.
(k −1) β + (1 − β)Yk (s) 2 β + (1 − β)Yk (s)
k =1
Z t ( )
n
 
(1 − β)Ydtnc+1 (s) 1 2 (1 − β)Ydtnc+1 (s) 2
− µ− σ ds.
dtnc
n
β + (1 − β)Ydtnc+1 (s) 2 β + (1 − β)Ydtnc+1 (s)
(63)
Our first objective is to apply the Martingale Central Limit Theorem to prove
that {Mn } converges in distribution to the stochastic process (1 − β)σW .
According to Itô’s formula, for every n ∈ N, k ∈ {1, 2, . . . , n}, and r ≥
(k −1)
n ,

log {β + (1 − β)Yk (r)}

Z r (   2 )
(1 − β)Yk (s) σ2 (1 − β)Yk (s)
= µ− ds
(k −1) β + (1 − β)Yk (s) 2 β + (1 − β)Yk (s)
Z r  
n

(1 − β)Yk (s)
+ σdWs .
(k −1)
n
β + (1 − β)Yk (s)

Substracting the first integral on the right-hand-side from the expression on the
left-hand-side, and then adding the resulting expressions, we obtain that for each
n ∈ N and t ∈ [0, ∞),
Mn (t)
dtnc Z
X k   Z  
n (1 − β)Yk (s) t
(1 − β)Ydtnc+1 (s)
= σdWs + σdWs
(k −1) β + (1 − β)Yk (s) dtnc β + (1 − β)Ydtnc+1 (s)
k =1 n n
Z t
= Un (s)σdWs ,
0

where
(1 − β)Yk (s)
Un (s) = I (k −1) k .
β + (1 − β)Yk (s) { n ≤s≤ n }
Thus, for arbitrary n ∈ N, Mn = {Mn (t) : t ∈ [0, ∞)} is a martingale with
continuous sample paths and Mn (0) = 0. Let us consider now the sequence of
stochastic processes {An } defined by
An (t) := hMn (t)i, t ∈ [0, ∞).
Taxes and transaction costs 161

That is, An is the quadratic variation of the continuous martingale Mn . To apply

the Martingale Central Limit Theorem to {Mn }, we need to consider an arbitrary
t ∈ [0, ∞) and study the convergence in probability of the sequence of random
variables {An (t)}.
We observe that, for every t ∈ [0, ∞),
Z t
An (t) = Un2 (s)σ 2 ds
0
dtnc Z
X k  2
n (1 − β)Yk (s)
= σ 2 ds
(k −1) β + (1 − β)Y k (s)
k =1 n
Z t  2
(1 − β)Ydtnc+1 (s)
+ σ 2 ds
dtnc β + (1 − β)Y dtnc+1 (s)
 n

X n 
dtnc Z k
Yk (s)
2
= (1 − β)2 σ 2  ds
(k −1) β + (1 − β)Yk (s)
k =1 n

Z t  2 #
Ydtnc+1 (s)
+ ds
dtnc β + (1 − β)Ydtnc+1 (s)
 
n

Xdtnc Z t  2
Y dtnc+1 (s)
= (1 − β)2 σ 2  Znk + ds  , (64)
dtnc β + (1 − β)Ydtnc+1 (s)
k =1 n

where
1
Znk := Rnk ,
n
Z k  2
1 n Yk (s)
Rnk := ds.
1
n
(k −1)
n
β + (1 − β)Yk (s)

Thus, for each n ∈ N, the stochastic process An = {An (t) : t ∈ [0, ∞)} has
continuous sample paths and satisfies, for each t ≥ s ≥ 0, An (t) − An (s) ≥
0. We want to prove that, for arbitrary t ∈ [0, ∞), the sequence of random
variables {An (t)} converges in probability to (1 − β)2 σ 2 t. Since the integral in
Pdtnc
(64) converges in probability to 0, this is equivalent to proving that k =1 Znk
converges in probability to t.
We cannot apply the standard version of the Weak Law of Large Numbers
Pdtnc 1
Pdtnc
to prove that k =1 Znk = n k =1 Rnk converges in probability to t, because
Rnk depends not only on k but also on n. Nevertheless, we observe that for
each n ∈ N and t ∈ [0, ∞), {Znk : k ∈ {1, 2, . . . , dtnc}} are independent and
identically distributed positive random variables. Furthermore, for every n ∈ N
and k ∈ {1, 2, . . . , n},
 Z k  2 
1  1 n Yk (s) 
|Znk | =  ds 
n  n1 (k −1) β + (1 − β)Yk (s)
n

162 A. Cadenillas, S.R. Pliska

Z k  2  2
11 n 1 1 1
≤ ds = . (65)
n n1 (k −1)
n
1−β n 1−β

In addition, as n ↑ ∞,
 2
a.s. Y1 (0)
Rn1 = nZn1 −→ = 1. (66)
β + (1 − β)Y1 (0)
We also observe that, for every t ∈ [0, ∞) and  > 0,
dtnc dtnc dtnc
X 1X 1X
P {Znk ≥ } ≤ E [Znk I{Znk ≥} ] = E [Zn1 I{Zn1 ≥} ]
 
k =1 k =1 k =1
1 dtnc
= E [Rn1 I{Zn1 ≥} ]. (67)
 n
Inequality (65) implies that Rn1 is bounded by (1 − β)−2 , while the limit in (66)
a.s. a.s.
guarantees not only Rn1 −→ 1, but also I{Zn1 ≥} −→ 0. Thus, we may apply the
Lebesgue Dominated Convergence Theorem (see Corollary 4.2.3 of Chow and
Teicher (1988)) to the right-hand-side of inequality (67). Hence, for every t > 0
and  > 0,
dtnc
X
lim P {Znk ≥ } = 0.
n→∞
k =1

Another application of the Lebesgue Dominated Convergence Theorem gives

dtnc
X
E [Znk I{Znk <1} ] = dtncE [Zn1 I{Zn1 <1} ]
k =1
dtnc
= E [Rn1 I{Zn1 <1} ] −→ t,
n
a.s. a.s.
because Rn1 −→ 1, and I{Zn1 <1} −→ 1. Thus, the Weak Law of Large Numbers
for independent random variables (see Corollary 10.1.2 of Chow and Teicher
(1988)), implies that, for each t ∈ (0, ∞),
dtnc
X P
Znk −→ t, as n ↑ ∞.
k =1

According to equation (64), this means that, for each t ∈ (0, ∞),
P
An (t) −→ (1 − β)2 σ 2 t, as n ↑ ∞. (68)
P
Here, −→ denotes convergence in probability.
In summary, for each n ∈ N, Mn is a martingale with continuous sample
paths and Mn (0) = 0. Furthermore, the sequence of stochastic processes {An }
defined by An (t) = hMn (t)i has also continuous sample paths and satisfies, for
each t ≥ s ≥ 0, An (t) − An (s) ≥ 0. Thus, we may apply the Martingale Central
Taxes and transaction costs 163

Limit Theorem (see Theorem 7.1.4 of Ethier and Kurtz (1986)). From equation
(68) we then conclude that
d
Mn −→ (1 − β)σW . (69)

By the same methodology that we applied to prove (68), we can also prove
that for each t ∈ [0, ∞),
P 1
Cn (t) −→ (1 − β)µt − (1 − β)2 σ 2 t, as n ↑ ∞, (70)
2
where

Cn (t)
dtnc Z
(   2 )
X k
n (1 − β)Yk (s) 1 2 (1 − β)Yk (s)
:= µ− σ ds
(k −1) β + (1 − β)Yk (s) 2 β + (1 − β)Yk (s)
k =1
Z t ( )
n
 
(1 − β)Ydtnc+1 (s) 1 2 (1 − β)Ydtnc+1 (s) 2
+ µ− σ ds.
dtnc
n
β + (1 − β)Ydtnc+1 (s) 2 β + (1 − β)Ydtnc+1 (s)
(71)

According to the Lebesgue Dominated Convergence Theorem, we observe that

for each t ∈ [0, ∞)
1
E [Cn (t)] −→ (1 − β)µt − (1 − β)2 σ 2 t, as n ↑ ∞.
2
We want to prove that the sequence of stochastic processes {Cn } defined by (71)
converges in distribution to the deterministic process C defined by
1
C (t) := (1 − β)µt − (1 − β)2 σ 2 t. (72)
2
Let us define the sequence of stochatic processes {Dn } by

Dn (t) := Cn (t) − E [Cn (t)], n ∈ N, t ∈ [0, ∞),

and the sequence of deterministic processes {Gn } by

Gn (t) = E [Dn (t)2 ] = VAR[Cn (t)].

We observe that, for each n ∈ N, Dn is a martingale with continuous sample paths

and Dn (0) = 0. Furthermore, the sequence of deterministic processes {Gn } has
also continuous sample paths and satisfies, for each t ≥ s ≥ 0, Gn (t)−Gn (s) ≥ 0.
P
In addition, for each t ∈ [0, ∞), Dn (t) −→ 0 and {Dn2 (t)} is uniformly integrable.
An application of the Mean Convergence Criterion (see Theorem 4.2.3 of Chow
and Teicher (1988)) then gives, for each t ∈ [0, ∞), Gn (t) −→ 0 as n ↑ ∞.
Then, the Martingale Central Limit Theorem (see Theorem 7.1.4 of Ethier and
d
Kurtz (1986)) implies that Dn −→ 0. Hence,
164 A. Cadenillas, S.R. Pliska

d
Cn −→ C . (73)

Finally, equations (63), (69), and (73) imply that the sequence of stochastic
processes {log Vn } given by

log Vn (t)
dtnc
X   
k 
= log β + (1 − β)Yk + log β + (1 − β)Ydtnc+1 (t)
n
k =1
dtnc Z k
(   2 )
X n (1 − β)Yk (s) 1 2 (1 − β)Yk (s)
= Mn (t) + µ− σ ds.
(k −1) β + (1 − β)Yk (s) 2 β + (1 − β)Yk (s)
k =1
Z t ( )
n
 
(1 − β)Ydtnc+1 (s) 1 2 (1 − β)Ydtnc+1 (s) 2
+ µ− σ ds
dtnc
n
β + (1 − β)Ydtnc+1 (s) 2 β + (1 − β)Ydtnc+1 (s)
= Mn (t) + Cn (t)

converges in distribution to a Brownian motion with positive drift (1 − β)µ −

2 (1 − β) σ and volatility (1 − β)σ. Therefore,
1 2 2

d
Vn −→ GBM((1 − β)µ, (1 − β)σ). 

References

1. Assaf, D., Taksar, M., Klass, M.: A difussion model for optimal portfolio selection in the presence
of brockerage fees. Math. Oper. Res. 13, 277-294 (1988)
2. Bossaerts P., Dammon, R.: Tax-Induced Intertemporal Restrictions on Security Returns. J. Fi-
nance 49, 1347-1371 (1994)
3. Boyce, W., DiPrima, R.: Elementary Differential Equations and Boundary Value Problems. New
York: Wiley & Sons 1992
4. Cadenillas, A., Haussmann, U.: The Stochastic Maximum Principle for a Singular Control Prob-
lem. Stochast. Stochast. Reports 49, 211-237 (1994)
5. Chow, Y., Teicher, H.: Probability Theory: Independence, Interchangeability, Martingales, 2nd
edn. Berlin Heidelberg New York: Springer 1988
6. Constantinides, G.: Capital Market Equilibrium with Personal Tax. Econometrica 51, 611-636
(1983)
7. Constantinides, G.: Optimal Stock Trading with Personal Taxes: Implications for Prices and the
Abnormal January Returns. J. Financial Econ. 13, 65-89 (1984)
8. Constantinides, G., Ingersoll, J.: Optimal Bond Trading with Personal Taxes. J. Financial Econ.
13, 299-335 (1984)
9. Dammon, R., Dunn K., Spatt, C.: A Reexamination of the Value of Tax Options. Rev. Financial
Studies 2, 341-372 (1989)
10. Dammon, R., Spatt, C.: The Optimal Trading and Pricing of Securities with Asymmetric Capital
Gains Taxes and Transaction Costs. Rev. Financial Studies 9, 921-952 (1996)
11. Davis, M.H.A., Norman, A.: Portfolio selection with transaction costs. Math. Oper. Res. 15,
676-713 (1990)
12. Dumas, B., Luciano, E.: An exact solution to a dynamic portfolio choice problem under trans-
action costs. J. Finance 46, 577-595 (1991)
13. Dybvig, P., Ross, S.: Tax Clienteles and Asset Pricing. J. Finance 41, 751-762 (1986)
14. Ethier, S., Kurtz, T.: Markov Processes: Characterization and Convergence. New York: John
Wiley & Sons 1986
Taxes and transaction costs 165

15. Karatzas, I.: Lecture Notes on the Mathematics of Finance. CRM Monograph Series. Providence,
Rhode Island: AMS 1996
16. Karlin, S., Taylor, R.: A first course in stochastic processes, second edition. New York: Academic
Press, 1975
17. Lipster, R., Shiryayev, A.: Statistics of Random Processes I: General Theory. Berlin Heidelberg
New York: Springer 1977
18. Miller, M.: Debt and Taxes. J. Finance 33, 261-275 (1977)
19. Miller, M., Scholes, M.: Dividends and Taxes. J. Financial Econ. 6, 333-364 (1978)
20. Morton, A., Pliska, S.: Optimal Portfolio Management with Fixed Transaction Costs. Math.
Finance 5, 337-356 (1995)
21. Ross, S.: Stochastic Processes. New York: John Wiley & Sons 1996
22. Schaefer, S.: Taxes and Security Market Equilibrium. In: Sharpe, W. et al. (eds.) Financial
Economics: Essays in honor of Paul Cootner. Englewood Cliffs, NJ: Prentice Hall 1982, pp
159-178
23. Shiryayev, A.: Optimal Stopping Rules. Berlin Heidelberg New York: Springer 1978
24. Shreve, S., Soner, M.: Optimal Investment and Consumption with Transaction Costs. Ann. Appl.
Probab. 4, 609-692 (1994)