You are on page 1of 33

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/326861230

Optimal mean-reversion strategy in the presence of bid-ask spread and delays


in capital allocations

Article in Quantitative Finance · August 2018


DOI: 10.1080/14697688.2018.1484151

CITATION READS
1 446

1 author:

Sergey Isaenko
Concordia University Montreal
32 PUBLICATIONS 232 CITATIONS

SEE PROFILE

All content following this page was uploaded by Sergey Isaenko on 05 December 2020.

The user has requested enhancement of the downloaded file.


Optimal Mean–Reversion Strategy in the Presence of
Bid–Ask Spread and Delays in Capital Allocations
Sergey Isaenko∗†

Concordia University

Abstract
A portfolio optimization problem for an investor who trades T-bills and a mean–
reverting stock in the presence of proportional and convex transaction costs is
considered. The proportional transaction cost represents a bid–ask spread, while the
convex transaction cost is used to model delays in capital allocations. I utilize the
historical bid–ask spread in US stock market and assume that the stock reverts on
yearly basis, while an investor follows monthly changes in the stock price. It is found
that proportional transaction cost has a relatively weak effect on the expected return
and the Sharpe ratio of the investor’s portfolio. Meantime, the presence of delays in
capital allocations has a dramatic impact on the expected return and the Sharpe ratio
of investor’s portfolio. I also find the robust optimal strategy in the presence of model
uncertainty and show that the latter increases the effective risk aversion of the investor
and makes her view the stock as more risky.

Keywords: Portfolio Choice; Mean–Reverting security, Transaction Costs

JEL Classification: G11


The John Molson School of Business, Concordia University, 1455 de Maisonneuve Blvd. W. Montreal,
Quebec, H3G 1M8. Sergey Isaenko is a corresponding author: phone #1-514-848-2424, ext 2797, fax #
1-514-848-4500.

The author is thankful to anonymous referees for the helpful remarks.
1 Introduction
Many portfolio selection models assume that a stock price follows the geometrical Brownian
motion (GBM). This assumption is probably reasonable for a long horizon problem. However,
the behavior of a stock price at short horizon can be substantially different from the GBM
model. Many investors trade with short horizons and look for predictability in stock returns.
In this regard, a stock, or portfolio of stocks, that exhibits a mean–reverting behavior is
considered as one of the most attractive for investments. A high demand for mean–reverting
strategies resulted, in turn, in a reduction of profits from such strategies1 in recent years.
Hence, finding an optimal strategy and following it carefully becomes essential when trading
securities with mean–reverting returns.
Trading a mean–reverting stock requires paying an ask price and receiving a bid price
which are traditionally modeled by means of proportional transaction costs. In addition
to losing money on paying a bid–ask spread, investors face losses due to their inability to
allocate funds to an investment opportunity instantly. This inability may come from latency
in order execution, time delays required to raise capital by intermediaries, opportunity costs
from time delays caused by valuable alternative activities, difficulties in borrowing and short–
selling, transitory price impact, and sub–optimality of a trading model. The goal of my study
is to find an optimal trading strategy for an investor who trades a risk–free security and a
risky mean–reverting security in the presence of proportional transaction costs and delays in
capital allocations. The delays in capital allocation are also called capital inertia.
I make capital move slowly by using adjustment costs that an investor faces for changing
her allocations to a risky security. The adjustment costs are modeled by convex trading
costs. These costs assume that the bigger a trade the higher cost per share will be. Moreover,
because the portfolio selection problem is treated in continuous time, the convexity in the
number of shares is converted into the convexity in the rate of share trading. Consequently,
an investor endogenously chooses to trade a finite volume per time interval and is not able to
adjust her allocations instantly in response to new investment opportunities. The approach
of modeling trading delays in continuous time by means of convex trading costs was pioneered
by practitioners [see Grinold and Kahn (2000)] and is also used in academia by Almgren and
Chriss (2000), Almgren (2003), Isaenko (2010, 2015), Rogers and Singh (2010), and Vayanos
and Woolley (2013). Convex trading costs are also found in Heaton and Lucas (1996) and
Garleanu and Pedersen (2012) in the discrete time settings.
The analysis of portfolio selection in the presence of proportional transaction costs has a
long history and includes a vast number of papers that are not discussed here for the purpose
of compactness.2 It is worthy to remark that proportional transaction costs also cause delays
1
See, for example, Anrew (2008).
2
For example, see Constantinides (1986), Davis and Norman (1990), Duffie and Sun (1990), Dumas and
Luciano (1991), Shreve and Soner (1994), Vayanos (1998), Liu and Loewestein (2002), Liu (2004), and Jang,
Koo, Liu, and Loewenstein (2007), and many others. See Guasoni and Muhle-Karbe (2013), for a more
detailed literature review.

1
in capital allocations. However, these delays are relatively small under the calibration used
in the paper and, therefore, will be overlooked for the purpose of the clarity of exposition.
I consider an investor with a constant absolute risk aversion (CARA) utility function
that supports her terminal consumption. The investor trades a risk–free asset and a risky
security that follows mean–reverting Cox-Ingersoll–Ross (CIR) process. It is assumed that
changing the position in the risk–free asset is costless, while trading the mean–reverting
security requires paying proportional and convex transaction costs. The convex term in
transaction costs is used to control the variability of the stock holdings. I choose the value
for this term that allows the investor to follow long–term (monthly) changes in the stock
price but does not let her trail its daily fluctuations. Following only long–term price changes
may seem to be a reasonable strategy given that the reversion coefficient used in the stock
price implies a reversion of this price within approximately two years. This strategy results
in a relatively low portfolio turnover which is historically observed for many individual and
institutional investors.3 Furthermore, the proportional term in transaction costs is set to
correspond to the bid–ask spread of 0.5%.4
An optimal strategy for the investor in the presence and the absence of transaction costs
is found. The former case is called illiquid, while the latter case is called perfectly liquid.
The investor in the illiquid market trades in the direction to the allocations that provide the
highest utility function. The further away she is from this allocation, the faster she trades.
The investor seizes trading as soon as she gets close enough to the best allocations. The
proximity to the best allocations at which the investor stops trading defines a no–transaction
region. The latter results from the presence of a bid–ask spread. Overall, the investor
follows a seemingly common strategy by trading against the market and pursuing long–
term (monthly) changes in the stock price with some lagging, while overlooking short–term
(daily) investment opportunities. Consequently, the presence of transaction costs results in
a negative impact on the expected return and the Sharpe ratio of the optimal portfolio.
Interestingly, this impact turns out to be rather dramatic: imposing transaction costs makes
the expected return fall by 7.7 times and the portfolio’s Sharpe ratio drop by 4.8 times.
Most of the impact on portfolio characteristics in the illiquid market comes from convex
term in transaction costs. In particular, paying only the convex term in transaction costs
decreases the expected return and the Sharpe ratio of the portfolio by factors 6.5 and 4.3,
respectively. Furthermore, if it is assumed that the investor follows the strategy which is
optimal in the presence of both transaction costs but pays only the proportional costs, then
the expected return and the Sharpe ratio of the investor’s portfolio are smaller than their
perfectly liquid counterparts by 5.0 and 3.7 times, respectively. It follows that if an investor,
3
Transaction cost, either fixed, linear, or convex, has two effects on trading strategies: it causes an
investor to skip on some investment opportunities or invest with delays and it makes an investor to
underallocate if she chooses to undertake an investment opportunity. In this study I emphasize the first
effect even though the both are important for trading strategies. These effects are controlled by a trading
volume. Moreover, two types of transaction cost are used to isolate contributions from the bid–ask spread
and delays in capital allocations.
4
This bid–ask spread is close to its historical value in US stock market. See, for example, Jones (2002).

2
who does not face delays in capital allocations, mistakenly uses the strategy that I find in
the presence of these delays, then her losses from the suboptimality of her strategy will be
comparable to the losses due to the capital inertia.
I consider comparative statics to see how transaction costs, the risk aversion of investor,
the reversion speed and volatility of the risky security affect the portfolio performance.
As expected, higher transaction costs make the investor trade slowly and miss a lot of
investment opportunities which result in a lower Sharpe ratio and a lower expected return
of the portfolio. An increase of the coefficient of risk aversion makes the investor trade more
cautiously (slowly), resulting in a lower expected return but a higher Sharpe ratio. The latter
trait is different from that in the perfectly liquid market where investors with different risk
aversions will hold portfolios with the same Sharpe ratio. Moreover, an increase in the stock
price reversion coefficient will result in a significant improvement of investment opportunities
leading to a higher Sharpe ratio and to a higher expected return on the portfolio in the
perfectly liquid market. The improvement of investment opportunities is much weaker in
the illiquid market. A higher reversion coefficient results in only marginally higher Sharpe
ratio and trading volume, while the expected return actually decreases due to an increase
in transaction costs. Finally, I document that a higher volatility of stock returns provides
more trading incentives resulting in a higher turnover, a higher expected return, and a higher
Sharpe ratio of the optimal portfolio in the illiquid market.
An investor who is seeking for a mean–reverting security has to deal with model
uncertainty in addition to a bid–ask spread and delays in capital allocations. Therefore,
besides the investor with no model uncertainty, I consider the investor who makes robust
decisions in the presence of model uncertainty along the lines of Anderson, Hansen, and
Sargent (2003; henceforth AHS).5 Following Maenhout (2004), I assume homotheticity, or
scale invariance of the decision problem, which allows to preserve tractability of the analysis.
Then it is shown that model uncertainty makes the robust investor increase her risk aversion
and assume more risk in the stock returns. A similar result was found in Maenhout (2004,
2006) but in the absence of transaction costs. Importantly, the stock returns with model
uncertainty are less risky when transaction costs are present. The effective coefficient of risk
aversion and dynamics of the stock price are defined by the parameter called “preference for
robustness” which has to be found from statistical errors of the model estimation. I find
that following the AHS approach to model uncertainty may substantially reduce the risk
premium and the portfolio turnover, but is likely to have a relatively small impact on the
portfolio Sharpe ratio. The latter is due to the fact that an increase of the investor’s risk
aversion makes the portfolio Sharpe ratio become higher, while changes in the drift of the
stock price reduces this ratio.
The literature on portfolio selection with a mean–reverting security and transaction costs
is limited. Most of it is concerned with a double stopping problem where an investor is
committed to a sequence of trades in which a buy order must be followed by a sell order
5
See also Hansen and Sargent (1995).

3
and each order must be for one share only.6 The resulted strategies are suboptimal since an
investor could be better off by changing the sizes of her orders and have sequential trades in
the same direction. This study is also related to the paper by Isaenko (2010) who analyzes
an optimal trading of an investor with CARA utility in the presence of proportional and
convex transaction costs when the stock follows the mixture of the GBM process and the
Poisson process.
Garleanu and Pederson (2013) also solve a portfolio selection problem for an investor
who faces convex (quadratic) transaction costs and trades a risky security with predictable
returns. Their paper has a few important differences from this paper. First, Garleanu and
Pederson assume that the stock price follows a different predictable process. The short–term
stock returns are uncorrelated with the short–term risk premium in their model and can
potentially offer less investment opportunities than the stock returns in my model. Second,
they assume the absence of the linear term in transaction costs. Third, the utility function
of the investor in Garleanu and Pederson (2013) has an infinite time horizon and does not
support inter–temporal consumption. Furthermore, Garleanu and Pederson (2013) analyze
the performance of the optimal portfolio based on a very specific example and are not able
to see the possibility of very dramatic effects on portfolio performance coming from delays
in capital allocations.
Glasserman and Xu (2013) take the model and the numeric example of Garleanu and
Pederson (2013) and add model uncertainty with a robust approach along the lines of Hansen
and Sargent (1995). Therefore, they inherit the differences with my model outlined for
the paper by Garleanu and Pederson (2013). Furthermore, Glasserman and Xu (2013) do
not assume homotheticity and, therefore, do not predict the equivalence of distributions
(conditional on the stock price) between the portfolio of an investor with model uncertainty
and the portfolio of an investor with a higher coefficient of risk aversion facing no model
uncertainty.
The rest of the article is organized as follows. Section 2 describes the model. Section
3 presents the optimal strategy in the absence of transaction costs. Section 4 delivers an
optimal strategy in the presence of transaction costs. Section 5 presents the comparative
analysis. Section 6 finds the portfolio rule in the presence of transaction costs and model
uncertainty. Section 7 concludes. The proofs of propositions and lemmas are given in
Appendix.

2 The Basic Model


I consider a Markovian economy with a finite horizon T and assume a filtered probability
space (Ω, F, {Ft }, Q). Uncertainty in the model is generated by a standard one–dimensional
6
See, for example, Zhang and Zhang (2008), Kong and Zhang (2010), Zervos et al. (2013), Leung and
Li (2015).

4
Brownian motion W , which is adapted.

2.1 Securities
The investor is a price–taker who has CARA–preferences that support only consumption at
time T . She continuously trades two securities: a risk–free bond and a stock. For simplicity
of exposition, it is assumed that the risk–free bond has zero interest rate. The stock price
follows the CIR process:
p
dS(t) = κ(S̄ − S(t))dt + σ S(t) dW (t), S(0) = S0 . (1)

It follows that the probability density of the stock is


 q/2
−u−v v

f (S, t) = c e Iq (2 uv), (2)
u
−κt

where c = 1−e2κ−κt σ 2 , q =
2κS̄
σ 2 − 1, u = cS0 e , v = cS, and Iq (2 uv) is modified Bessel
function of the first kind of order q. The probability density implies the following expected
price and variance of S(t) conditioned on information at time zero:

E(S(t)) = S0 e−κt + S̄(1 − e−κt ), (3)


σ2 S̄σ 2
V ar(S(t)) = S0 (e−κt − e−2κt ) + (1 − e−κt )2 . (4)
κ 2κ
It follows that the distribution becomes approximately stationary after time t > 1/κ with
2
the expected return and the variance being close to S̄ and S̄σ 2κ
, respectively.
Trading in the risky security requires paying transaction costs. In particular, trading the
risky security within a small time interval ∆t imposes trading losses equal to [α1 S(t)|u(t)| +
α2 (S(t)|u(t)|)1+ε ]∆t dollars, where αn , n = 1, 2 and ε are positive constants, and u is a
rate of trading within the time interval ∆t.7 For simplicity of exposition, we require that
coefficients αn , n = 1, 2 be the same for purchasing and selling the stock. The term α1 S|u| in
transaction costs represents proportional transaction fees which model the bid–ask spread.
Moreover, if α2 = 0 then the investor trades at infinitely fast rate. However, if α2 > 0 then
infinitely fast trade results in infinitely large losses. Therefore, the term α2 |Su|1+ε does not
allow the rate of trading to be infinite within even a very short time–interval. Naturally, the
number of shares held by an investor, N , becomes absolutely continuous and is given by

dN (t) = u(t)dt, N (0) = N0 . (5)

It follows that investors change their allocations at a finite rate which I will also term as
“slowly” to contrast with the case with no transaction costs where an investor trades with
an infinite speed. The slow change of allocations implies that the rate of trading (or trading
7
Formally, u(ti ) = (N (ti ) − N (ti−1 ))/(ti − ti−1 ), where N (t) is the number of shares held by an investor
at time t and ti − ti−1 = ∆t.

5
volume) becomes a control variable of an investor, while an allocation becomes her state
variable. The convex trading costs are appealing since they endogenize a slow movement of
capital. In this regard, they are similar to the adjustment costs that are used in the literature
on production economy to prevent firms from changing their capital stock too quickly.

2.2 Investor’s Problem


This section introduces a portfolio–choice problem faced by the investor in the illiquid
market:
 
1
sup E0 − exp(−γX(T )) (6)
u∈R γ
p
dX(t) = [N (t)κ(S̄ − S(t)) − α1 S(t)|u(t)| − α2 |S(t)u(t)|1+ε ]dt + N (t)σ S(t)dW (t),(7)

where γ > 0 is an absolute risk aversion, X stands for an investor’s total wealth, and it is
assumed that only self–financing strategies are available. Since dX depends explicitly on S
and N , the description of the problem is completed by adding equations (1) and (5).
The optimization problem (6) will be solved in the dynamic programming approach.
Furthermore, equations (1), (5), and (7) show that the Markovian set of state variables in
the given economy is (S, X, N ). Therefore, I introduce an investor’s value function as
 
1
V (t, X, S, N ) = sup Et − exp(−γX(T )) (8)
u∈R γ

and solve the corresponding Hamilton–Jacobi-Bellman (HJB) equation to find this function.

2.3 Calibration
The HJB equation for the value function can be solved only numerically. The following is
the calibration of the model used in the numerical solution: γ = 0.1, α1 = 0.0025, α2 =
0.00002, ε = 1, κ = 3, S̄ = 40, σ = 0.5, T = 1. This calibration will be used in the examples
below unless mentioned otherwise.
It should be noticed that there are many practical cases where mean–reverting security
or portfolio of securities is traded. It could be one stock traded for a relatively short time
interval or, for example, a pair of co-integrated stocks traded for a long time interval. One can
build a mean–reverting portfolio with various reversion coefficients κ and volatility coefficient
σ. The construction of such portfolios and the estimation of their underlying processes lies
outside of scope of this paper.8 Therefore, a certain calibration of the process for the stock
price is simply assumed. A somewhat arbitrary choice of the calibration is compensated
by comparative statics analysis with respect to parameters κ and σ carried in Section 5.
Furthermore, an estimation of mean–reverting processes should introduce an uncertainty
8
For empirical detection and estimation of mean–reverting portfolios see, for example, Balvers, Wu,
Gilliland (2000), D’Aspremont (2011), and Akarim and Sevim (2013).

6
in the underlying parameters which brings up an issue of model uncertainty. This issue is
analyzed in Section 6.
The choice of κ suggests that the security could be approximated as stationary in a few
months after beginning of trading. The stock becomes stationary with the expected return
and the standard deviation being equal to 40 and 1.67, respectively. Furthermore, it is
known that the spectrum of the Ornstein–Uhlenbeck process have most of its weight on low
frequencies with a cut-off frequency being equal to κ/2π 9 or approximately 0.5 under the
chosen calibration. Given that the cut–off frequency of the Ornstein-Uhlenbeck process does
not depend on the diffusion coefficient, I conjecture that the cut–off frequency for the CIR
process is close or equal to κ/2π implying the period of two years. Therefore, it is expected
that the stock price has a complete cycle within at least two years.
The choice of coefficient α1 is based on the assumption that the bid–ask spread is equal
to 0.5% of the traded dollar amount. Setting the value of α2 is more ambiguous and depends
significantly on the type of investor. This coefficient defines the portfolio turnover which
historically has been very low throughout most types of investors. Indeed, most investors
historically prefer to ignore (or are not able to take advantage of) short–term stock variations
and follow only a long–term (monthly or even yearly) changes in the stock price. I incorporate
this stylized fact by setting α2 to be reasonably high but low enough to make variability of
the investor’s stock position be close to this variability in the liquid market. High variability
of the stock position allows the investor to take a desired stock allocation in the future
and to follow long–term investment opportunities. Since types of investors vary a lot across
markets, I will mitigate the restrictions imposed by a chosen calibration by carrying the
sensitivity analysis with respect to parameters α1 , α2 , and γ.

3 Optimal Policies without Transaction Costs


For the purpose of comparison, I first solve problem (6) in the absence of transaction costs.

Proposition 1 Assume that equation (10) below has a unique solution,10 then the unique
value function of the investor is given by
1
V (t, X, S) = − exp(−γX + g(t, S)), (9)
γ
9
See, for example, Bibbona, Panfilo, and Tavella (2008).
10
I conjecture that parabolic partial differential equations for function g appearing in Propositions 1 to
4 have unique solutions. Proving the existence and uniqueness of these solutions lies beyond this paper. It
is also assumed that the numerical solutions of these equations converge to the actual solutions in the limit
∆t → 0.

7
20
10
0
−10
−20

37 38 39 40 41 42 43

Figure 1: Figure shows the stock holding, N (solid line) and function g(t, S) (dashed line)
versus the stock price S at time zero in the absence of transaction costs. It is assumed that
γ = 0.1, ε = 1, κ = 3, S̄ = 40, σ = 0.5, and T = 1.

where function g(t, S) solves the partial differential equation

[κ(S̄ − S)]2 1 2
0 = gt − + σ SgSS , (10)
2σ 2 S 2
g(T, S) = 0, lim g(t, S) = −∞.
S→±∞

The optimal allocation of the investor is


κ(S̄ − S) + σ 2 SgS
N ? (t, S) = . (11)
γσ 2 S
Proof. See Appendix.
I solve PDE (10) numerically and find the optimal allocations versus the stock price and
time. Figure 1 shows these allocations (solid line) at t = 0. The investor trades against
the market and takes a short position when the stock price increases above its long–term
mean value and a long position otherwise. The magnitude of the position is proportional to
the deviation of the stock from its long–term mean. It is worthwhile to notice a significant
variation in the allocations for relatively small changes in the security price due to a low risk
aversion of the investor and the long–term robustness of the stock mean–reversion. Figure 1
also shows a slight asymmetry of allocation with respect to the long–term mean of the stock.
The asymmetry is due to the dependence of the stock volatility from the stock price. Higher
stock price implies higher volatility making the investor reduce her risk exposure.
Proposition 1 implies the following

8
Lemma 1 The expected return and the standard deviation of the investor’s portfolio returns
are inversely proportional to the investor’s coefficient of risk aversion. The Sharpe ratio of
this portfolio is independent from her coefficient of risk aversion.

Proof. See Appendix.


The following example provides the reference values for the expected return and
volatilities of the investor’s portfolio:

Example 1 Let the initial value of the stock price be equal to S̄. Then the expected return
and the Sharpe ratio of the investor’s portfolio, as well as the standard deviation of her stock
allocation at time T are

E0 (rX ) = E0 (X(T )) − X0 = 25.14, (12)


E0 (rX )
Sr = = 3.50, (13)
SD0 (rX )
SD0 (N (T )) = 3.89. (14)

I note that all examples in this paper and the comparative statics are established by using
Monte–Carlo simulations.11 Example 1 confirms that the mean–reverting security brings
prosperous investment opportunities resulting in a very high Sharpe ratio and a significant
dollar return. The next section establishes how these portfolio characteristics are affected
by the bid–ask spread and delays in capital allocations.

4 Optimal Policies with Transaction Costs


This section considers the portfolio–choice problem in the presence of transaction costs. The
market without transaction costs will be labeled “perfectly liquid”, while the market with
convex transaction costs (with or without linear term) will be called “illiquid”, or the market
with delays in capital allocations (also called the market with capital inertia). Furthermore,
the market where only proportional transaction costs are present will be tagged “liquid”.
I start from the case when both terms in transaction costs are present.

Proposition 2 Assume that equation (16) below has a unique solution, then the unique
value function of the investor is given by
1
V (t, X, S, N ) = − exp(−γX + g(t, S, N )), (15)
γ
11
A path of a process dY (t) = µY (t)dt + σY (t)dW (t), √Y (0) = Y0 is simulated by using a finite–difference
approximation Y (ti+1 ) = Y (ti ) + ∆t[µY (ti )dt + σY (ti ) ∆t εi )], Y (0) = Y0 , where ∆t = ti+1 − ti , and
{εi , i = 0, 1, 2, ...} is a set of independent standard normal random variables. ∆t is set to one day. The
predictions of simulations are affected by the initial conditions only to a small extent due to a relatively long
investment horizon.

9
where function g(t, S, N ) solves the partial differential equation
1 1
gt + (σγN )2 S − γN κ(S̄ − S) + [κ(S̄ − S) − γσ 2 N S]gS + σ 2 S[(gS )2 + gSS ]
2 2
  1ε   1ε
−gN /γS − α1 gN /γS − α1
− ε 1+ε
(−gN /S − α1 γ)+ − ε 1+ε
(gN /S − α1 γ)+ = 0,
α2 (1 + ε) α2 (1 + ε)
g(T, S, N ) = 0, lim g(t, S, N ) = ∞, lim g(t, S, N ) = ∞. (16)
S→±∞ N →±∞

The optimal rate of trading is


   1ε
−g /γS−α
 1 N 1
if − gN /γS > α1


 S α2 (1+ε)


u∗ = 0 if |gN /γS| ≤ α1 (17)

   1ε
 − S1 gNα2/γS−α

1
if gN /γS > α1 .


(1+ε)

Proof. See Appendix.


Equation (17) suggests that proportional transaction costs results in the presence of
no–transaction region in the state space, where the investor does not trade. This region
disappears when α1 = 0. Meantime, coefficient α2 takes care of the trading speed: the
smaller the coefficient, the faster trading. It is worthwhile to notice that Proposition 2 holds
in the special case when α1 = 0 and α2 > 0, but cannot be applied if α1 > 0 and α2 = 0.
The latter case requires solving a singular stochastic control problem which is beyond the
scope of this paper.
Figure 2 shows the rate of trading versus the stock holding (Panel A) at fixed stock price
and the rate of trading versus the stock price (Panel B) at fixed stock holding at time zero.
Panels A and B confirm that in the presence of proportional transaction costs the trading
space (S, N ) splits into buy, sell, and no–transaction regions where the investor buys, sells,
and does not trade stock shares, respectively. If the number of shares is excessive or/and the
stock price is too high then the investor sells her shares, if the number of shares is insufficient
or/and the stock price is too low then the investor buys shares. Furthermore, the further
away from the optimal allocations the investor is, the faster she trades in the direction of
these allocations. Finally, the investor does not trade if her allocations are close to the
optimal ones.
Panel C of Figure 2 depicts the no–transaction region and the optimal allocation, N ∗ ,
in the perfectly liquid market at time zero. Panel C shows that the no–transaction region
has a smaller slope than the optimal allocations in the perfectly liquid market causing the
latter to be outside of this region when the investor has a very long position. The presence
of convexity in transaction costs makes a risk averse investor hold less securities when she is
very long in the stock and be less short when she is very short. Furthermore, it follows that
the allocations with the highest utility for an investor in the illiquid market are different
from the optimal allocations in the perfectly liquid market, N ∗ (t, S).

10
Panel A Panel B Panel C

40
15

10
30
10

20

5
5

10

N
u

0
0

0
−5

−10

−5
−20
−10

−10
−30

−10 −5 0 5 10 37 38 39 40 41 42 43 39.0 39.5 40.0 40.5 41.0

N S S

Figure 2: Figure shows the rate of trading u versus stock holding N at S = 40 (Panel A),
the rate of trading u versus stock price S at N = 0 (Panel B), and the no–transaction region
(Panel C) at time zero. The dashed line on Panel C shows the stock allocations in the
perfectly liquid market. It is assumed that γ = 0.1, ε = 1, κ = 3, S̄ = 40, σ = 0.5, and
T = 1.

To get a better intuition of the investor’s trading, I simulate her trading positions and
compare them with the stock path and its moving average. The results are presented on
Figure 3. The figure shows a few long cycles in the stock price combined with a lot of local
variability of various scale. The optimal allocation follows the moving average of the stock
with a lag and the investor neglects the local variation in the stock price. This strategy
seems intuitive and common.
The found strategies should be considered as approximations since investors cannot trade
continuously in time. One has to modify these strategies for discrete trading in time to make
them practical. Assume that the investor trades at equal time intervals. Then the rate of
trading at a given time translates into the number of traded shares equal to ∆N = u∆t,
where u is the average rate of trading per the given time interval for the current stock price
and current stock holdings, and ∆t is the time interval between two trades. Equation (17)
relates the number of traded shares with the coefficient of risk aversion of the investor. In
case that the investor does not know her coefficient of risk aversion, she can choose the size
of her optimal trades and then relate them to the coefficient of the risk aversion by using
equation (17). Moreover, the bid–ask spread paid for a trade in the discrete time is equal
to α1 |∆N |S, while the loss on the delays in capital allocations should be made equal to
α̂2 (|∆N |S)2 , where α̂2 = α2 ∆t. Notice that it is assumed that α̂2 remains finite in the limit
∆t → 0.

11
2
1
0
−1
−2

0.0 0.2 0.4 0.6 0.8 1.0

Figure 3: Figure shows stock holdings, N (solid line), stock price S (dashed line), and
moving average of the stock price (dotted line) over time. It is assumed that γ = 0.1, α1 =
0.0025, α2 = 0.00002, ε = 1, κ = 3, S̄ = 40, σ = 0.5, and T = 1.

The following example illustrates the optimal strategy. I will use the portfolio Sharpe
ratio to gauge the strategy performance instead of using the utility function or function g.
Despite a non–normal distribution of the portfolio return, I find that the Sharpe ratio has a
very strong consistency with function g in measuring a relative portfolio performance.

Example 2 Assume that the initial stock price is equal to S̄ and the initial stock allocation
is zero (N0 = 0). Then the expected return, the Sharpe ratio, the standard deviation of the
stock allocations at time T , and portfolio turnover are given by:

E0 (rX ) = 3.26, (18)


Sr = 0.73, (19)
SD0 (N (T )) = 3.07, (20)
Z T
T urnover = E0 |u(t)|dt = 5.07. (21)
0

Comparison of the standard deviations of the optimal holdings presented in examples 1 and
2 shows that the two are close indicating that the delays in capital allocations are not very
significant to hinder the investor ability to take a desired stock position in the future. In
addition, in agreement with Figure 3, the portfolio turnover seems to be sufficient to trace
long–term investment opportunities. Further comparison of examples 1 and 2 reveals an
intuitive relation between the expected returns and between the Sharpe ratios across the
economies with and without transaction costs: in the presence of transaction costs, the

12
15
10
5
0
−5
−10

0.0 0.2 0.4 0.6 0.8 1.0

Figure 4: Figure shows stock holdings in illiquid market (dashed line), stock holdings in
liquid market (solid line), and stock price (dotted line) over time. It is assumed that
γ = 0.1, α1 = 0.0025, α2 = 0.00002, ε = 1, κ = 10, S̄ = 40, σ = 0.5, and T = 1.

investor is always at the suboptimal allocations casing her both expected return and the
Sharpe ratio to fall. Moreover, the difference between the Sharpe ratios and the difference
between the expected returns across the markets with and without transaction costs appear
to be quite dramatic (close to five times for the Sharpe ratios and close to eight times for
the expected returns). It follows that despite a reasonable variability in her positions, the
investor loses a lot of investment opportunities in the presence of transaction costs resulting in
the crash of her Sharpe ratio and her expected return. The missed investment opportunities
are due to local variations in the stock price. Furthermore, the investor takes advantage
of long-term investment opportunities only in a limited extent since her allocations in the
perfectly liquid market are much more considerable. The latter is confirmed on Figure 4
which shows the simulation of the optimal allocations in the absence of transaction costs.
These allocations are much more sizable and change much more frequently.
It is interesting to evaluate the impact of proportional transaction costs on the portfolio
performance. I find this impact by comparing portfolios in the illiquid market in the presence
of the linear term in transaction costs and in its absence:

Example 3 Assume that the initial stock price is equal to S̄, the initial stock allocation is
zero (N0 = 0), and there are no proportional transaction costs (α1 = 0). Then the expected
return, the Sharpe ratio, the standard deviation of the stock allocation at time T , and portfolio

13
turnover are given by:

E0 (rX ) = 3.90, (22)


Sr = 0.82, (23)
SD0 (N (T )) = 3.24, (24)
T urnover = 6.21. (25)

Comparison of examples 2 and 3 shows that proportional transaction costs have a distinct
impact on the outcomes of the investor’s trading. The investor is able to trade noticeably
faster in the absence of the bid–ask spread which allows her to get a better advantage of
investment opportunities. Consequently, the terminal portfolio becomes more wealthy and
less risky with the resulting improvement of its Sharpe ratio by about 12%. Nonetheless, I
conclude that the impact of the linear term in transaction costs on the optimal strategy is
much smaller than that of the convex term since the presence of only convex term decreases
the Sharpe ratio by more than four times.
In the previous analysis it was assumed that the convex part in transaction costs have
to be paid by the investor. It is possible that the investor will follow the strategy found
above even in the absence of this part in transaction costs. For example, the investor may
not be able to determine an optimal strategy in the absence of the convex term and follow
the strategy found above based on its appealing intuition. Therefore, it is interesting to
determine the characteristics of the portfolio when the investor follows the strategy described
in Proposition 2, while assuming that she pays only proportional transaction costs:

Example 4 Assume that the initial stock price is equal to S̄, the initial stock allocation is
zero (N0 = 0), there is no convex term in transaction costs (α2 = 0), and the investor follows
the strategy which is optimal in the presence of both terms in transaction costs. Then the
expected return and the Sharpe ratio of her portfolio are given by:

E0 (rX ) = 4.99, (26)


Sr = 0.95. (27)

Comparison of examples 2 and 4 shows noticeable differences in the main characteristics of


the portfolios in the markets with and without convex term in transaction costs. Paying
an additional (convex) cost decreases both the expected return and the Sharpe ratio of
portfolio. However, these cutbacks seem not very significant in comparison to losses in the
expected return and the Sharpe ratio resulting from using a suboptimal strategy as seen from
comparison of examples 1 and 2. Indeed, following strategy in example 4 leads to the loss in
the Sharpe ratio by 3.7 times, while paying an additional convex transaction cost decreases
it further by only 30%. Based on this observation, I conclude that a portfolio in the illiquid
market is a reasonable zeroth–order approximation for a portfolio in a liquid market in which
the investor pays proportional transaction costs and follows the stock allocations found in
the presence of both terms in transaction costs. Example 4 suggests that this seemingly

14
α1 E0 (rX ) Sr SD0 (N (T )) Turnover
0.001 3.61 0.78 3.18 5.72
0.0025 3.26 0.73 3.07 5.07
0.01 1.87 0.51 2.52 2.90

Table 1: The risk premium and the Sharpe ratio of the portfolio as well as the standard
deviation of holdings at time T and the portfolio turnover for different values of coefficient
α1 . It is assumed that γ = 0.1, α2 = 0.00002, ε = 1, κ = 3, S̄ = 40, σ = 0.5, T = 1.

reasonable trading strategy can result in a dramatic under–performance of the investor’s


portfolio.

5 Comparative Statics
This section studies the effects from the coefficient of risk aversion γ, mean–reversion
coefficient κ, coefficient in the volatility of the stock return σ, and coefficients α1 , α2 in
transaction costs on the portfolio performance.
First, I consider the effects from the coefficients in transaction costs. Table 5 presents
the risk premium and the Sharpe ratio of the portfolio as well as the standard deviation of
holdings at time T and the portfolio turnover for different values of coefficient α1 . A higher
bid–ask spread (coefficient α1 ) increases the size of the non–transaction region causing the
investor to trade less so that the portfolio turnover and the standard deviation of her stock
position fall. Consequently, the investor misses good investment opportunities resulting in
the smaller expected return and the smaller Sharpe ratio of her portfolio.
Table 5 presents the risk premium and the Sharpe ratio of the portfolio as well as the
standard deviation of holdings at time T and the portfolio turnover for different values of
coefficient α2 . As expected, an increase in convexity of transaction costs results in a slower
trading and in a smaller expected return on the portfolio. The investor is not able to follow
changes in the stock price and misses a lot of investment opportunities when the costs are
high. It also leads to a decrease of the Sharpe ratio of her portfolio.
The rest of the analysis concerns with the parameters affecting the investor’s portfolio
in both illiquid and perfectly liquid markets. Therefore, the comparative statics for both
types of markets are shown. Notice that the portfolio turnover is equal to infinity in the
perfectly liquid market and is not shown in Tables 5 to 5. Table 5 presents the risk premium
and the Sharpe ratio of the portfolio as well as the standard deviation of holdings at time T
and the portfolio turnover in the illiquid market for different values of the coefficient of risk
aversion. It follows that the higher the risk tolerance is, the faster investor trades making
him able to get better advantage of investment opportunities. A better advantage shows
in a higher expected return. In the meantime, faster trade in the presence of transaction

15
α2 E0 (rX ) Sr SD0 (N (T )) Turnover
0.00001 5.09 0.87 3.82 8.18
0.00002 3.26 0.73 3.07 5.07
0.0002 0.42 0.49 0.65 0.75
0.002 0.043 0.093 0.074 0.081

Table 2: The risk premium and the Sharpe ratio of the portfolio as well as the standard
deviation of holdings at time T and the portfolio turnover for different values of coefficient
α2 . It is assumed that γ = 0.1, α1 = 0.0025, ε = 1, κ = 3, S̄ = 40, σ = 0.5, T = 1.

Illiquid Perfectly liquid


γ E0 (rX ) Sr SD0 (N (T )) Turnover E0 (rX ) Sr SD0 (N (T ))
0.01 4.25 0.50 6.49 7.55 250.70 3.50 38.93
0.05 3.76 0.63 4.37 6.01 50.28 3.50 7.88
0.1 3.26 0.73 3.07 5.07 25.14 3.50 3.89
0.2 2.55 0.87 1.90 4.07 12.57 3.50 1.96
0.4 1.75 1.01 1.04 2.98 6.28 3.50 0.98

Table 3: The risk premium and the Sharpe ratio of the portfolio as well as the standard
deviation of holdings at time T and the portfolio turnover for different values of the
coefficient of risk aversion in the illiquid and perfectly liquid markets. It is assumed that
α1 = 0.0025, α2 = 0.00002, ε = 1, κ = 3, S̄ = 40, σ = 0.5, T = 1.

costs makes the stock holdings be more volatile also resulting in a higher volatility of the
portfolio. Contrary to the case of the liquid market, it appears that the volatility of the
investor’s portfolio increases faster with risk aversion than the expected return resulting in
a smaller Sharpe ratio for a more risk tolerant investor. The last trait is consistent with a
behavior of a risk averse investor who is willing to give up more of the expected return for
an extra volatility if her risk aversion decreases. Furthermore, the impact of illiquidity on
the portfolio performance is the strongest for the least risk averse investor who wants to take
extreme stock positions and significantly amplify the changes in the stock prices.
Table 5 presents the risk premium and the Sharpe ratio of the portfolio as well as the
standard deviation of holdings at time T and the portfolio turnover for different values of
the mean–reversion coefficient κ. It follows from the table that a quicker stock reversion
presents more investment opportunities, making the investor trade faster and allowing her
to earn a higher Sharpe ratio. The presence of delay in capital allocations comes at more
dramatic costs when κ is higher. For example, if κ = 10 then the Sharpe ratio in the
perfectly liquid market is about fifteen times bigger than in the illiquid market. Interestingly,
I find that the expected portfolio return decreases in illiquid market as the mean–reversion
coefficient increases from 3 to 10. A faster reversion to the mean makes the stock price less

16
Illiquid Perfectly liquid
κ E0 (rX ) Sr SD0 (N (T )) Turnover E0 (rX ) Sr SD0 (N (T ))
2. 2.40 0.58 2.51 3.88 13.73 2.34 3.17
3. 3.26 0.73 3.07 5.07 25.14 3.50 3.89
10. 2.30 0.96 2.69 5.75 234.78 14.12 7.23

Table 4: The risk premium and the Sharpe ratio of the portfolio as well as the standard
deviation of holdings at time T and the portfolio turnover for different values of coefficient
κ in the illiquid and perfectly liquid markets. It is assumed that γ = 0.1, α1 = 0.0025, α2 =
0.00002, ε = 1, S̄ = 40, σ = 0.5, T = 1.

risky and more predictable. The investor has incentives to trade more and has to pay a
higher transaction costs leading to a smaller expected return on the portfolio. The trade–off
between better investment opportunities and higher transaction costs results in a loss in the
expected return. Furthermore, better predictability in trading may also cause a decrease
of the uncertainty in the stock allocations (compare standard deviations of stock positions
when κ = 3 and when κ = 10).
Finally, Table 5 presents the risk premium and the Sharpe ratio of the portfolio as well as
the standard deviation of holdings at time T and the portfolio turnover for different values of
the volatility coefficient σ. The results lead to a few observations. First, the uncertainty of
the stock allocations in the perfectly liquid market rises with a smaller stock volatility since
the stock trends more deterministically, allowing the investor to take more sizable positions.
See solid lines on Figure 5 that show the stock allocation path when σ = 0.75 (Panel A) and
the stock allocation path when σ = 0.25 (Panel B). Larger positions also bring higher Sharpe
ratio and expected return. Second, I notice that higher volatility does not necessarily implies
a lower Sharpe ratio and expected return of the portfolio. Indeed, the investor trades more
conservatively at higher volatility which can make her better off than at lower volatility in
the perfectly liquid market. Third, the above traits change in the illiquid market. Higher
volatility of the stock price provides more trading incentives which result in a higher turnover
and amplified stock positions (see dashed lines on Panels A and B on Figure 5). The latter
produces a higher expected return and a higher Sharpe ratio of the portfolio.
The following section extends the analysis to a typical situation where the investor is
uncertain about the model that is followed by the stock price.

17
Illiquid Perfectly liquid
σ E0 (rX ) Sr SD0 (N (T )) Turnover E0 (rX ) Sr SD0 (N (T ))
0.25 0.83 0.49 2.52 2.89 28.86 4.09 7.78
0.5 3.26 0.73 3.07 5.07 25.14 3.50 3.89
0.75 5.84 0.90 2.66 6.43 26.21 3.72 2.60

Table 5: The risk premium and the Sharpe ratio of the portfolio as well as the standard
deviation of holdings at time T and the portfolio turnover for different values of coefficient
σ in the illiquid and perfectly liquid markets. It is assumed that γ = 0.1, α1 = 0.0025, α2 =
0.00002, ε = 1, κ = 3, S̄ = 40, T = 1.

Panel A Panel B
10

10
5

5
0

0
−5

−5
−10

−10

0.0 0.4 0.8 0.0 0.4 0.8

t t

Figure 5: Figure shows stock holdings in illiquid market (dashed line), stock holdings in
liquid market (solid line), and stock price (dotted line) over time when σ = 0.75 (Panel A)
and when σ = 0.25 (Panel B). It is assumed that γ = 0.1, α1 = 0.0025, α2 = 0.00002, ε =
1, κ = 3, S̄ = 40, and T = 1.

18
6 Optimal Policies with Transaction Costs and Model
Uncertainty
Suppose that the investor runs the augmented Dickey–Fuller test on a stock and determines
that it exhibits a mean–reverting behavior with some errors in the key parameters such
as κ, S̄ and σ. The latter implies that the true process for the stock may not be given by
equation (1) and may follow some alternative model. I handle the resulting model uncertainty
by following the approach developed by Anderson, Hansen, and Sargent (2002) (henceforth,
AHS). According to AHS, the decision-maker accepts the state equation (1) (also called
“reference” model) as useful, but suspects it to be misspecified. She, therefore, wants to
consider alternative state models in the Hamilton-Jacobi-Bellman equation in robust way.
A preference for robustness is then achieved by having the investor guard against the worst
alternative model that is reasonably similar to the reference model.
AHS show that an alternative model adds drifts to the stock process and the process for
the portfolio value:
p
dS(t) = [κ(S̄ − S(t)) + vx S(t)N (t)σ 2 + vs S(t)σ 2 ]dt + σ S(t)dW (t), (28)
dX(t) = [κ(S̄ − S(t))N (t) + vx S(t)N 2 (t)σ 2 + vs S(t)N (t)σ 2 − α1 S(t)|u(t)|
p
− α2 |S(t)u(t)|1+ε ]dt + N (t)σ S(t)dW (t). (29)

Functions vs and vx are found endogenously by minimizing differential value function with
2
an additional entropy penalty given by Sσ 2Ψ
(N vx + vs )2 , where Ψ is a penalty function.
The entropy penalty is used to keep an alternative model close to the reference model.
Unfortunately, the choice of penalty function Ψ is rather arbitrary and, therefore, is often
motivated by mathematical convenience. Commonly, function Ψ is chosen to be inversely
proportional to the value function which leads to homotheticity of the decision problem.
See, for example, Cagetti et al. (2002), Uppal and Wang (2003), Maenhout (2004, 2006),
and Liu et al. (2005). I will follow these footsteps and require that the decision problem be
homothetic. I call the investor who follows the AHS approach as a robust investor and the
investor who does not follow this approach as a non–robust investor.
I start by finding the optimal strategy for a robust investor in the case of model
uncertainty and no transaction costs.

Proposition 3 Assume the AHS approach to model uncertainty. Assume that there are
no transaction costs and equation (31) below has a unique solution, then the unique value
function of the investor is given by
1
V (t, X, S) = − exp(−γX + g(t, S)), (30)
γ
where function g(t, S) solves the partial differential equation
1 κ2 (S̄ − S)2 γ
0 = gt + σ 2 SgSS − 2 , g(T, S) = 0, lim g(t, S) = −∞ (31)
2 2σ S(θ + γ) S→±∞

19
in which parameter θ is state–independent and is equal to Ψ e−γX+g(t,S) . The optimal stock
allocation and functions vx , vs in the reference model for the stock price and the portfolio
value equal to


κ(S̄ − S) + σ 2 SgS (1 + γθ )
N = , (32)
σ 2 S(γ + θ)
vx∗ = −θ, (33)
θ
vs∗ = gS . (34)
γ
Proof. See Appendix.
Parameter θ is also called preference for robustness. Proposition 3 immediately implies
Lemma 2 For a given stock price, the stock allocations of the investor with the coefficient
of risk aversion γ, who follows the AHS approach with the preference for robustness θ, are
inversely proportional to γ + θ.
Proof. See Appendix.
It follows that strategy of a non–robust investor with the coefficient of risk aversion equal
to γ + θ has the same probability distribution (conditioned on the value of the stock price)
as strategy of a robust investor with the coefficient of risk aversion γ. The latter investor
assumes that the stock price follows the model:
γκ p
dS(t) = (S̄ − S(t))dt + σ S(t)dW (t), (35)
γ+θ
which is found from equation (28).
I conclude that the model uncertainty comes down to two modifications: slower stock
γ
price reversion (reversion coefficient changes from κ to κ γ+θ ) and higher risk aversion (the
coefficient of risk aversion changes from γ to γ + θ). In the meantime, the expected stock
price and its volatility are not affected by the model uncertainty. A similar result for a
related stock price model can be found in Maenhout (2006).
The portfolio performance of a robust investor with the coefficient of risk aversion γ
and with preference for robustness θ is different from the portfolio performance of a non–
robust investor with the coefficient of risk aversion γ + θ due to the difference in the stock
distributions. I illustrate this statement assuming that θ = 0.1.12
Example 5 Let the initial values of the stock price be equal to S̄ and θ be equal to 0.1. Then
the expected return and the Sharpe ratio of the investor’s portfolio, as well as the standard
deviation of the stock allocation at time T are
E0 (rX ) = 9.19, (36)
Sr = 1.13, (37)
SD0 (N (T )) = 2.68. (38)
12
The chosen value for θ is justified by convenience. An alternative calibration for θ will not change the
intuition behind the numerical results.

20
I relate the portfolio in Example 5 with the portfolio of an investor with the coefficient of risk
aversion of θ + γ = 0.2 in the absence of model uncertainty. Comparison of the Sharpe ratio
(37) with the Sharpe ratio for the investor with the coefficient of risk aversion of 0.2 in Table
5 shows that the investor significantly outperforms in the absence of model uncertainty. The
latter is expected since the reversion coefficient in model (35) is half of its value in model
(1).
Now let me consider a robust investor who faces transaction costs and model uncertainty.

Proposition 4 Assume the AHS approach to model uncertainty and let equation (40) below
have a unique solution, then the unique value function of the investor is given by
1
V (t, X, S, N ) = − exp(−γX + g(t, S, N )), (39)
γ
where function g(t, S, N ) solves the partial differential equation
γ(γ + θ) 1 1 θ
gt + S(N σ)2 − γκ(S̄ − S)N + Sσ 2 gSS + Sσ 2 (gS )2 (1 + )
2 2 2 γ
  1ε
−gN /γS − α1
+ [κ(S̄ − S) − (γ + θ)Sσ 2 N ]gS − ε 1+ε
(−gN /S − α1 γ)+ , (40)
α2 (1 + ε)
  1ε
gN /γS − α1
− ε (gN /S − α1 γ)+ = 0,
α2 (1 + ε)1+ε
g(T, S, N ) = 0, lim g(t, S, N ) = ∞, lim g(t, S, N ) = ∞, (41)
S→±∞ N →±∞

in which the preference for robustness θ is state–independent and is equal to Ψ e−γX+g(t,S,N ) .


The optimal rate of trading is given by
 
gN
1
− γS −α1 ε
1
if − gγS


 N
> α1
 S α2 (1+ε)


u∗ = 0 if | gγS
N
| ≤ α1 (42)
   1
 g
 1 γSN −α1 ε

 − S α2 (1+ε)
 if gγSN
> α1

and functions vx , vs are equal to

vx∗ = −θ, (43)


θ
vs∗ = gS . (44)
γ
Proof. See Appendix.
Proposition 4 implies
Lemma 3 For a given time, stock price S, and stock allocation N , the trading rate of a
robust investor with the coefficient of risk aversion γ and the preference for robustness θ are
the same as the trading rate of a non–robust investor with the coefficient of risk aversion
equal to γ + θ.

21
0.30
0.25
0.20
Density

0.15
0.10
0.05
0.00

36 38 40 42 44 46

Figure 6: Figure shows the stock price density without model uncertainty (dotted line),
the stock price density with model uncertainty in the perfectly liquid market (dashed line),
and the stock price density with model uncertainty in the illiquid market (solid line). It is
assumed that θ = 0.1, γ = 0.1, α1 = 0.0025, α2 = 0.00002, ε = 1, κ = 3, S̄ = 40, and T = 1.

Proof. See Appendix.


Besides changing the trading rate, model uncertainty breaks linearity of the drift in
the stock process. This drift could be written as κ0 (S̄ 0 − S), where κ0 = κ + θσ 2 (N −
gS (t, S, N )/γ), S̄ 0 = S̄/κ0 . It follows that reversion coefficient κ0 and target price S̄ 0 become
functions of state variables and time. Numerical analysis shows that κ0 tends to be higher
than κ when N is positive and lower than κ otherwise. Since allocations are more often
negative when S is high and positive when S is low, I conclude that the stock price converges
with the rate κ0 which is smaller than κ to a higher than S̄ target price when the stock price
is high and the stock price converges with the rate faster than κ to a lower than S̄ target price
when the stock price is small. The resulted changes of the stock drift affect the distribution
of the stock price. Figure 6 depicts the distribution of the stock at time T with no model
uncertainty (dotted line), with model uncertainty in the perfectly liquid market (dashed
line) and with model uncertainty in the illiquid market (solid line). Clearly, the presence of
illiquidity makes the investor assume that the stock price is less risky than in the perfectly
liquid market but more risky than in the case of no model uncertainty.
It follows that, like in the perfectly liquid market, the model uncertainty affects the risk
aversion of the investor and the stock distribution. These changes have an impact on the
portfolio performance:

Example 6 Let the initial values of the stock price be equal to S̄ and θ be equal to 0.1. Then

22
the expected return and the Sharpe ratio of the portfolio, as well as the standard deviation of
the stock allocation at time T and portfolio turnover are

E0 (rX ) = 1.69, (45)


Sr = 0.55, (46)
SD0 (N (T )) = 2.03, (47)
T urnover = 3.96. (48)

Comparison of examples 2 and 6 shows that model uncertainty in the illiquid market has
strong effects on the expected return of the portfolio and the portfolio turnover. Interestingly,
reduction in the Sharpe ratio after model uncertainty is introduced is not very significant
(about 25%) since the effective increase of the risk aversion of the investor results in an
increase of the Sharpe ratio, while changes in the drift of the stock price reduces this ratio.
Furthermore, comparison of examples 1 and 6 suggests that imposing delays in capital
allocations and model uncertainty can bring down dramatically all characteristics of the
optimal portfolio. For example, the expected return of the portfolio drops by fifteen times
under the chosen calibration.
Finally, one has to relate to the choice of value of parameter θ. It is assumed somewhat
arbitrary in examples 5 and 6 that θ = 0.1, while in practice this parameter should be found
from statistical errors of an estimation of the process for a stock price.13 AHS suggest that
the value of parameter θ is deemed reasonable if the worst-case model associated with it is
sufficiently similar to the reference model to obscure the model selection problem based on a
time series. The parameter θ is, therefore, chosen such that the detection-error probability is
sufficiently high. See Anderson, Hansen, and Sargent (2003) and Maenhout (2006) for more
details.

7 Conclusion
The paper analyzes the optimal behavior of an investor who trades T–bills and a stock that
follows a mean–reverting process. The stock is a subject of a bid–ask spread and delays in
allocations modeled by linear and convex transaction costs, respectively. The low frequency
cut–off of the mean–reversion process has a period of two years. The delays in capital
allocations allow to follow long–term (monthly) changes in the stock prices but do not permit
to pursuit its daily changes. I find that the presence of delays in capital allocations has a
dramatic impact on the expected return and the Sharpe ratio of investor’s portfolio as long
as the mean–reverting security offers rich investment opportunities. Meantime, proportional
13
A comparative statics with respect to parameter θ is also carried. The found results are fully consistent
with the discussion of example 6 and are not reported for the purpose of compactness.

23
transaction costs have a relatively weak effect on the expected return and the Sharpe ratio
of the investor’s portfolio. I also find the robust optimal strategy in the presence of model
uncertainty and show that the latter increases the effective risk aversion of the investor and
changes the drift of the stock price.

Appendix
The Appendix proves the propositions and lemmas of the paper.
Proof of Proposition 1. Suppose that the investor trades a bond and a perfectly liquid
stock. Then the state space of the investor is fully described by state variables X and S and,
as follows from the dynamics of the portfolio value given by equation (7) at α1 = α2 = 0,
the HJB equation for the value function V (t, X, S) can be written as
 
1 2 2 1 2 2
sup Vt + σ N SVXX + σ SVSS + σ N SVXS + N µS VX + µS VS = 0,
N ∈R 2 2
1
V (T, X, S) = − e−γX , (A-1)
γ
2
where µS = κ(S̄ − S). The first–order condition reads N ∗ = − µS VσX2+σ SVXS
SVXX
, which after
substitution into equation (A-1) gives a PDE for V (t, X, S):

[µS VX + σ 2 SVXS ]2 1 2
Vt − + σ SVSS + µS VS = 0, (A-2)
2σ 2 SVXX 2
1
V (T, X, S) = − e−γX .
γ

Let me assume that V (t, X, S) = − γ1 e−γX+g(t,S) . Then equation (A-2) transforms into the
PDE for g given by equation (10) and the optimal allocation becomes equal to N ∗ given by
equation (11).
Q.E.D.

Proof of Lemma 1. Equation (10) suggests that function g is independent of γ. Hence,


the stock holding N ∗ and, therefore, the expected return and standard deviation of return of
the portfolio are inversely proportional to γ. Obviously, the Sharpe ratio is γ–independent.
Q.E.D.

24
Proof of Proposition 2. If the stock is illiquid then the indirect utility function depends
on state variables S, X, N and the HJB equation for V (t, X, S, N ) becomes

1 1
sup Vt + σ 2 N 2 SVXX + σ 2 SVSS + σ 2 N SVXS + µS VS + uVN
u∈R 2 2
  
1
+ N µS − α1 S|u| − α2 |Su| 1+ε
VX , V (T, X, S, N ) = − e−γX . (A-3)
γ

VN
The first–order condition implies that u is zero or a solution of α1 +α2 (1+ε)|Su|ε = SVX
.
The last equation has a solution for u only if α1 ≤ |VN /SVX |. In the latter case I find
  1ε
|VN /SVX |−α1
|Su| = α2 (1+ε)
. Therefore, the optimal rate of trading can be written as

   1ε
VN /SVX −α1
 1
if VN /SVX > α1




 S α2 (1+ε)

u= 0 if |VN /SVX | ≤ α1 (A-4)



   1ε
 − S1 −VNα2/SV X −α1

if − VN /SVX > α1 .


(1+ε)

When u is substituted into equation (A-3), I find


1 1
Vt + σ 2 N 2 SVXX + σ 2 SVSS + σ 2 N SVXS + N µS VX + µS VS (A-5)
2 2
 1  1
VN /SVX − α1 ε + −VN /SVX − α1 ε
+ ε 1+ε
(VN /S − α1 VX ) + ε 1+ε
(−VN /S − α1 VX )+
α2 (1 + ε) α2 (1 + ε)
1 −γX
= 0, V (T, X, S, N ) = − e ,
γ

where z + is a positive part of variable z.


Let me assume that V (t, X, S, N ) = − γ1 e−γX+g(t,S,N ) . Then equation (A-5) transforms
into the PDE for g given by equation (16), while the optimal rate of trading becomes equal
to u∗ given by equation (17).
Q.E.D.

Proof of Proposition 3. The stock price and the portfolio value under alternative model
follow (see, for example, Maenhout, 2006)
p
dS(t) = [κ(S̄ − S(t)) + vx S(t)N (t)σ 2 + vs S(t)σ 2 ]dt + σ S(t)dW (t), (A-6)
dX(t) = [N (t)κ(S̄ − S(t)) + vx S(t)(N (t)σ)2 + vs S(t)N (t)σ 2 ]dt
p
+ N (t)σ S(t)dW (t), (A-7)

25
where functions vx and vs are to be determined endogenously. Hence, the indirect utility
function of the investor , V (t, X, S), solves

1 1
0 = sup inf 2 Vt + (N σ)2 SVXX + σ 2 SVSS + N σ 2 SVXS + [µS + vx SN σ 2 + vs Sσ 2 ]VS
N ∈R (vx ,vs )∈R 2 2
Sσ 2

2 2 2
+ [µS N + vx S(σN ) + vs SN σ ]VX + (N vx + vs ) , (A-8)

where the last term represents the penalty for deviation from the reference model. The
first–order conditions read:

2 1 2 2 2 2 Sσ 2
0 = N σ S[VXX + 2vx VX + vx ] + σ SVXS + vx Sσ VS + (µS + vs Sσ )VX + vx vs
Ψ Ψ
1
0 = VX N + VS + (vx N + vs ).
Ψ
There are infinitely many solutions to this system of equations. I choose the following:

µS VX + σ 2 SVXS − Ψσ 2 SVX VS
N∗ = − , (A-9)
σ 2 S(VXX − ΨVX2 )
vx∗ = −ΨVX , (A-10)
vs∗ = −ΨVS . (A-11)

Assuming that
1
V (X, S, t) = − exp[−γX + g(t, S)] (A-12)
γ
results in the controls


µS + σ 2 SgS (1 + Ψγ e−γX+g(t,S) )
N = , (A-13)
σ 2 S(γ + Ψe−γX+g(t,S) )
vx∗ = −Ψe−γX+g(t,S) , (A-14)
Ψ −γX+g(t,S)
vs∗ = e gS . (A-15)
γ
θ
Furthermore, setting Ψ = − γV = θeγX−g(t,N,S) , where θ is positive and state–independent,
implies formulas (32)–(34). Finally, I find the PDE (31) for function g(t, S) by plugging
expressions (A-13)–(A-15) into equation (A-8) and replacing V with a trial function (A-12).
Q.E.D.

γ+θ
Proof of Lemma 2. Consider function f (t, S) = γ
g(t, S). Equation (31) suggests that
function f is a solution of

1 µ2
0 = ft + σ 2 SfSS − 2S , f (T, S) = 0, (A-16)
2 2σ S

26
and
µS + σ 2 SfS
N∗ = .
σ 2 S(γ + θ)
The result follows from comparison of the last equation with equation (11), since function
f (t, S) solves equation (10) which is independent from γ.

Q.E.D.

Proof of Proposition 4. The stock price and the portfolio value under an alternative
model follow
p
dS(t) = [κ(S̄ − S(t)) + vx S(t)N (t)σ 2 + vs S(t)σ 2 ]dt + σ S(t)dW (t), (A-17)
dX(t) = [N (t)κ(S̄ − S(t)) + S(t)N (t)σ 2 (vx N (t) + vs ) − α1 S(t)|u(t)|
p
− α2 |S(t)u(t)|1+ε ]dt + N (t)σ S(t)dW (t). (A-18)

Hence, the indirect utility function of the investor , V (t, X, N, S), solves

1 1
0 = sup inf 2 Vt + (N σ)2 SVXX + σ 2 SVSS + N σ 2 SVXS + uVN
u∈R (vx ,vs )∈R 2 2
+ [µS + vx SN σ 2 + vs Sσ 2 ]VS + [µS N + vx S(σN )2 + vs SN σ 2 − α1 S|u|
Sσ 2

1+ε 2
− α2 |Su| ]VX + (N vx + vs ) . (A-19)

It follows that    1ε
VN /SVX −α1
 1
if VN /SVX > α1




 S α2 (1+ε)

u∗ = 0 if |VN /SVX | ≤ α1 (A-20)



   1ε
 − S1 −VNα2/SV X −α1

if − VN /SVX > α1 ,


(1+ε)

vx∗ = −ΨVX , (A-21)


vs∗ = −ΨVS . (A-22)

I assume that V (X, N, S, t) = − γ1 exp[−γX + g(t, N, S)], then


   1ε
−gN /γS−α1
 1
if − gN /γS > α1




 S α2 (1+ε)

u∗ = 0 if |gN /γS| ≤ α1 (A-23)



   1ε
 − S1 gNα2/γS−α

1
if gN /γS > α1 ,


(1+ε)

27
vx∗ = −Ψe−γX+g(t,S) , (A-24)
Ψ −γX+g(t,S)
vs∗ = e gS , (A-25)
γ
and g solves the following PDE:

(γ)2 1
0 = gt + S(N σ)2 − γµS N + Sσ 2 [gSS + (gS )2 ] + (µS − γSσ 2 N )gS
2 2
  1ε  1
−gN /γS − α1 + gN /γS − α1 ε
− ε 1+ε
(−gN /S − α1 γ) − ε 1+ε
(gN /S − α1 γ)+
α2 (1 + ε) α2 (1 + ε)
1
+ ΨSσ 2 e−γX+g (N γ − gS )2 . (A-26)

θ
Let Ψ be equal to − γV = θeγX−g(t,N,S) , where θ is state–independent. Hence the last PDE
becomes equation (40) and functions vx and vs become equal to −θ and γθ gS , respectively.
Q.E.D.

γ+θ
Proof of Lemma 3. Consider function f (t, S, N ) = γ
g(t, S, N ). Equation (40) suggests
that function f is a solution of the following PDE
1 1 1 1
0 = ft + S[(γ + θ)N σ]2 − (γ + θ)µS N + Sσ 2 fSS + Sσ 2 (fS )2
2 2 2 γ+θ
fN  1ε  1
− (γ+θ)S − α1
+  fN
− α1 (γ + θ) ε

fN S
− ε − − α1 (γ + θ) −ε
α2 (1 + ε)1+ε S α2 (1 + ε)1+ε
 +
fN
× − α1 + [µS − (γ + θ)Sσ 2 N ]fS f (T, S, N ) = 0. (A-27)
(γ + θ)S

The last equation is the same as equation (16) for function g in which γ is replaced by γ + θ.
Furthermore, the rate of trading of the investor is given by
 
fN
1
 1 − (γ+θ)S −α1 ε

 if − gN /γS > α1
 S α2 (1+ε)


fN
u= 0 if | (γ+θ)S | ≤ α1 (A-28)
   1
 f

 1 (γ+θ)S N −α1 ε fN
 − S α (1+ε)
 if (γ+θ)S > α1 .
2

This rate is is the same as the trading rate of the investor with a coefficient of risk aversion
equal to γ + θ since function f is equal to function g of the investor with the coefficient of
risk aversion given by γ + θ.
Q.E.D.

28
References
[1] Yasemin Deniz Akarim, Serafettin Sevim, 2013,“The impact of mean reversion model
on portfolio investment strategies: Empirical evidence from emerging markets”, 31,
453–459.

[2] Almgren, R., 2003, “Optimal Execution with Nonlinear Impact Functions and Trading-
enhanced Risk,” Applied Mathematical Finance, 10, 1–18.

[3] Almgren, R., Chriss, 2000, “Optimal Execution of Portfolio Transactions,” Journal of
Risk, 3, 5–39.

[4] Andrew, S., 2008. “Where Have All the Stat Arb Profits Gone?”, Columbia University
Financial Engineering Practitioners Seminars.

[5] Balvers, R., Wu, Y., Gilliland, E. 2000,“Mean Reversion across National Stock Markets
and Parametric Contrarian Investment Strategies”, Journal of Finance, 55, 745–772.

[6] Bibbona, E., Panfilo, G., Tavella, P., 2008,“The Ornstein–Uhlenbeck process as a model
of a low pass filtered white noise,” Metrologia, 45, 117–126

[7] M. Cagetti, L. Hansen, T. Sargent, N.Williams, 2002, “Robustness and pricing with
uncertain growth”, Review of Financial Studies, 15, 363–404.

[8] Constantinides, G., 1986, “Capital Market Equilibrium with Transaction Costs,”Journal
of Political Economy, 94, 842–862.

[9] Alexandre D’Aspremont, 2011,“Identifying small mean-reverting portfolios”,


Quantitative Finance, 11, 352–364.

[10] Davis, M. H. A., Norman, A.R., 1990, “Portfolio Selection with Transaction Costs”,
Mathematics of Operations Research, 15, 676–713.

[11] Duffie, D., Sun, T., 1990, “Transaction Costs and Portfolio Choice in a Discrete–
Continuous Time Setting”, Journal of Economic Dynamics and Control, 14, 35–51.

[12] Dumas, B., Luciano, E., 1991, “An Exact Solution to a Dynamic Portfolio Choice
Problem under Transaction Costs”, Journal of Finance, 48, 577–595.

[13] Garleanu, N., B., Pedersen, L., H., 2013, “Dynamic Trading with Predictable Returns
and Transaction Costs”. Journal of Finance, 68, 2309–2340.

[14] Glasserman, P., Xu, X. 2013, “Robust Portfolio Control with Stochastic Factor
Dynamics,” Operations Research, 61, 874–893.

29
[15] Guasoni, P. , Muhle-Karbe, J., 2013, “Portfolio Choice with Transaction Costs: A User’s
Guide”. Chapter in Paris-Princeton Lectures in Mathematical Finance 2013. Springer
International Publishing.

[16] Grinold, R., Kahn, R., 2000. “Active Portfolio Management.” McGraw-Hill, second
edition.

[17] Hansen LP, Sargent TJ, 1995, “Discounted linear exponential quadratic Gaussian
control,” Automatic Control, IEEE Trans. 40(5): 968–971.

[18] Heaton, J., Lucas, D. 1996, “Evaluating the Effects of Incomplete Markets on Risk
Sharing and Asset Pricing.” Journal of Political Economy, 104, 443–487.

[19] Isaenko, S., 2010, “Portfolio Choice under Transitory Price Impact.” Journal of
Economic Dynamics and Control, 34, 2375–2389.

[20] Isaenko, S., 2015, “Equilibrium Theory of Stock Market Crashes.” Journal of Economic
Dynamics and Control, 60, 73–94.

[21] Jang, B., Koo, H., K., Liu, H., Loewenstein, M., 2007, “Liquidity Premia and
Transaction Costs”, Journal of Finance, 62, 2329–2366.

[22] Jones, C., 2002, “A Century of Stock Market Liquidity and Trading Costs.” Working
Paper, Columbia University.

[23] Leung, T., Li, X. 2015,“ Optimal Mean Reversion Trading with Transaction Costs and
Stop-Loss Exit”, International Journal of Theoretical and Applied Finance, vol 18, issue
3, 2015

[24] Leung, T., Li, X., Wang, T., 2015, “Optimal Multiple Trading Times Under the
Exponential OU Model with Transaction Costs”, with T. Leung and Z. Wang, Stochastic
Models, vol 31, issue 4, 2015

[25] Liu, H., 2004, “Optimal Consumption and Investment with Transaction Costs and
Multiple Risky Assets”, Journal of Finance, 59, 289–338.

[26] Liu, H., Loewestein, 2002, “Optimal Portfolio Selection with Transaction Costs and
Finite Horizons”, Review of Financial Studies, 15, 805–835.

[27] J. Liu, J. Pan, T.Wang, 2005, “An equilibrium model of rare-event premia and its
implication for option smirks”, Review of Financial Studies, 18, 131–164.

[28] Maenhout, P., 2004, “Robust portfolio rules and asset pricing,” Review of Financial
Studies, 17, 951983.

[29] Rogers, L. C. G., Singh, S., 2010 “The cost of illiquidity and its effects on hedging”.
Mathematical Finance 20, 597–615.

30
[30] S.E. Shreve, H.M. Soner, 1994, “Optimal investment and consumption with transaction
costs,” Annals of Applied Probabability, 4, 609–692.

[31] R. Uppal, T.Wang, 2003, “Model misspecification and underdiversification,” Journal of


Finance, 58, 2465–2486.

[32] Vayanos, D., 1998, “Transaction Costs and Asset Prices: A Dynamic Equilibrium
Model”, Review of Financial Studies, 11, 1–58.

[33] Vayanos, D., Woolley, P., 2013, “An Institutional Theory of Momentum and Reversal”,
Review of Financial Studies, 26, 1087–1145.

31

View publication stats

You might also like