Professional Documents
Culture Documents
To cite this article: Xiaoling Mei, Huanjun Zhu & Chongzhu Chen (2023) Mean-variance
portfolio selection with estimation risk and transaction costs, Applied Economics, 55:13,
1436-1453, DOI: 10.1080/00036846.2022.2097191
ABSTRACT KEYWORDS
There are many approaches that have been proposed to improve the empirical performance of the Portfolio optimization;
Markowitz mean-variance model. Designed to mitigate the impact of parameter uncertainty and transaction costs; market
estimation error, these approaches have delivered substantially better out-of-sample performance. impact; mean-variance
analysis
In this paper, we consider the portfolio optimization problem for a single-period investor facing
different types of transaction costs. By reformulating the rebalancing problem into a linear regres JEL CLASSIFICATION
sion framework, we show analytically that considering different transaction costs is equivalent to G10; G11
imposing additional constraints on the portfolio weights, thus providing desired properties such as
sparsity and stability in the trading strategy.
CONTACT Huanjun Zhu hzhu928@xmu.edu.cn Department of Statistics, School of Economics, Wang Yanan Institute for Studies in Economics, Xiamen,
China.
© 2022 Informa UK Limited, trading as Taylor & Francis Group
APPLIED ECONOMICS 1437
Another important way to cope with parameter solution to the model with quadratic transaction
uncertainty is to impose constraints on the portfo costs in a multiperiod setting, and Mei, DeMiguel,
lio weights. For example, Jagannathan and Ma and Nogales (2016) provide some analytical results
(2003) study the mean-variance portfolio model on the optimal portfolio policy for an investor
with no short-selling constraint and find that facing multiple risky assets in the presence of gen
imposing such constraint can help to improve the eral transaction costs. See also Constantinides
out-of-sample performance significantly. They (1986), Davis and Norman (1990), Liu (2004) and
further explain and demonstrate that constraining more, for example.
portfolio weights to be nonnegative is equivalent to In this paper, we study the impact and role of
shrinking the sample covariance matrix. In addi transaction costs from a different perspective from
tion, Lp -normal regulation is also widely adopted. the existing literature. For a single-period mean-
For instance, DeMiguel et al. (2009) propose to use variance investor who faces estimation error and
the L1 -norm constraint or the L2 -norm constraint various types of transaction costs, we show that the
on the portfolio weights. They show that the port investor’s rebalancing problem can be written into
folio performance has been largely improved even a linear regression form so that considering differ
in the presence of estimation error, and a link ent transaction costs is equivalent to imposing reg
between their approach and the existing procedure ularization constraints on the regression
has also been provided. Aside from that, Gotoh and coefficient. In this case, we can provide
Takeda (2011) suggest using L0 -norm to improve a statistical tool to establish the rebalancing strate
the out-of-sample performance, aiming to obtain gies with desired properties, such as sparsity or
a sparse portfolio, which can partially mitigate the stability.
impact of estimation error. Li (2015) also considers the Markowitz mean-
Apart from the parameter estimation error, con variance portfolio optimization and reformulates
siderable attention has been given to transaction it into a linear regression form. He suggests
costs over the past several years. For instance, imposing different constraints on the portfolio
Pogue (1970) presents an extended version of the weights so that to generate desired portfolio
Markowitz’s portfolio selection model without pro weights. Similarly, we rewrite the optimization
viding an analytical solution to this model. Since with a linear regression form, but we choose the
transaction costs can easily erode the gains from trading amount as our decision variable instead
a trading strategy, especially when there is high of portfolio weights. It becomes a different pro
turnover, Atsushi (1996) points out that portfolio blem. Moreover, our model focuses on coping
strategy without considering transaction cost is with various transaction costs, which distin
usually suboptimal. An increasing number of stu guishes this paper from Li (2015), who does
dies have emerged to seek solutions to portfolio not take into account transaction costs directly
optimization model in the presence of various in the optimization.
types of transaction costs under different model The main contributions of this paper are as
settings. Prior studies include Subrahmanyam follows. First, we connect the portfolio rebalancing
(1982), who studied the portfolio selection problem models with transaction costs to penalized least
with fixed transaction costs in a mean-variance square regression models and prove that taking
context. They provide a simple algorithm to solve different types of transaction costs into considera
the optimization problem by imposing additional tion helps to develop portfolio rebalancing strate
restrictions on the covariance matrix. Dai, Xu, and gies with desired properties. More specifically, this
Zhou (2010) study a continuous-time Markowitz’s paper shows that a rebalancing model with linear
mean-variance portfolio selection problem in transaction cost is equivalent to a classical mean-
a market with a single risky asset in the presence variance model with the L1 constraint; incorporat
of proportional transaction costs. Gârleanu and ing linear market impact costs can be rewritten into
Pedersen (2013) and DeMiguel, Martn-Utrera, a regression model with L2 penalty; the model
and Nogales (2015) study analytically the optimal considers that general market impact costs is
1438 X. MEI ET AL.
equivalent to a model with Lp constraint. Secondly, before rebalancing wt . More specifically, rewriting
we study a more general case incorporating both the objective function Equation (2.1) into an
linear and general market impact transaction costs, expression that is related to initial position wt
and further transform the aforementioned model yields:
into an elastic net regression form, but in a more n γ T^ o
general form. Finally, the third contribution is to w� ¼ argmax wTt μ^ w �wt
wt 2 t
analyse empirically the superior performance of the
proposed model, and further confirm the impor n γ T^ ^ γ T^ ^ γ T^ o
; argmax wTt ^μ w �wt þ γwTt �w w �wt γwTt �w t þ wt �wt
tance of incorporating transaction costs in the wt 2 t t
2 t 2
incorporate the costs into the objective function that is proportional to the amount of trading. More
Equation (2.2) to avoid overtrading. More specifi specifically, the linear transaction cost function of
cally, we consider a transaction cost function in asset i incurred at each period is:
a general form cðθ; ΔwÞ, in which θ is a set of � �
transaction cost parameters that control the type ci ¼ τ i �wi wi � (2:6)
and form of the cost function. In this case, the where wi is the weight right before rebalancing, and
optimal portfolio selection becomes: τi is the proportional transaction cost parameter of
n γ T^ o asset i that measures the amount of costs paid by the
Δw ¼ argmax ΔwTb
�
~
μ Δw �Δw cðθ; ΔwÞ : investor for each unit trading amount. For example,
Δw 2
in China’s stock market, investors would be charged
(2:4)
a commission fee from 0.02% to 0.3%. Besides,
The existence of transaction costs in the objec stamp duty is imposed on sell-side investors. All
tive function Equation (2.4) makes the trading such costs can be expressed in a proportional form,
amount in Equation (2.2) not optimal anymore. where the level of proportion is measured by the
Alternatively, the investor seeks different types of exogenous transaction cost parameter τi .
trading strategy depending on the specific type of As has been previously studied in the literature
transaction costs. To better understand the objec (see Magill and Constantinides 1976 and many
tive function, let us consider an alternative form. others), when the linear costs are imposed into
pffiffi ^ 1=2
By following Li (2015), define X ¼ γ� and a portfolio optimization model, the optimal port
1 ^ 1=2 b
~, the expression of optimal rebalancing folio weights are usually confined by a no-trade
y ¼ γ� μ
region. Consequently, the investor will not trade
strategy in (2.3) indicates that Δw� ¼ ðX T XÞ 1 XT y,
only if the portfolio weight inherited from the pre
and it coincides with the OLS estimator of the
vious period falls out of the no-trade region. In
linear regression model y ¼ XΔw þ e:
other words, in the presence of transaction costs,
� � a rational investor will weigh up the costs and
Δw�OLS ¼ argmax ky XΔwk22 : (2:5) benefits, to decide the optimal portfolio weights.
Δw
More specifically, when an investor is managing
In other words, the Markowitz mean-variance a large number of assets, trading part of the assets
model can be rewritten into a linear regression while keeping the others remained unchanged is
form, where the decision variable is the vector of a desired strategy. Equivalently, a sparse Δw is
trading weights. Consequently, the mean-variance favourable since it can reduce the transaction
model with transaction costs in Equation (2.4) can costs and further lower the portfolio variance
also be rewritten into a linear regression form with caused by the estimation error, especially when
the sample length is relatively small compared to
regularization, where the form of the regularization
the number of assets N.
depends on the specific type of transaction costs. In
To simplify the analysis, we assume an equal
general, transaction costs include direct costs and
proportional transaction cost parameter τ across
indirect costs. In subsequent sections, we will
different assets. The following proposition shows
explain the portfolio rebalance model with such that a sparse portfolio trading strategy can be
transaction costs with more details. obtained by incorporating proportional transaction
costs directly into the mean-variance portfolio
optimization problem.
Linear costs and LASSO
It is known that direct costs usually refer to bro Proposition 2 Let τbe the proportional transac
kers’ commissions and spreads, which are the dif tion cost parameter, and the total linear costs be
ferences between the bid and ask prices. In terms of cðτ; ΔwÞ ¼ τ k Δw k11 . The portfolio rebalance
such costs, it is usually expressed by a linear form model with such costs
1440 X. MEI ET AL.
n γ T^ o
ΔwLinear ¼ argmax ΔwTb
~
μ Δw �Δw τkΔwk11 costs. Incorporating proportional transaction
Δw 2 costs in the portfolio rebalance model helps to
(2:7) produce such a strategy. On the one hand, zero
trading amount reduces transaction costs and port
is equivalent to the optimization problem: folio management fees. On the other hand, L1
n γ T^ o norm regularization can partly mitigate the impact
ΔwLinear ¼ argmax ΔwTb ~
μ Δw �Δw (2:8) of estimation risk, especially when the number of
Δw 2
assets is large. Also note that as the level of pena
lization λ depends on the linear cost parameter τ,
s:t: k Δw k11 � s1 (2:9)
a higher value in τ would result in fewer transac
Thus, the existence of linear transaction costs auto tions in our model and thus decreases the overall
matically imposes a sparsity constraint on Δw, where transaction cost.
sis a proper threshold that controls the level of sparsity.
Quadratic costs and ridge regression
Remark.
� � � � Besides the linear form, transaction costs in
� �
�U Δw; b ^
~; �
μ ~; �Þ�
U ðΔw; μ a quadratic form are also widely considered in the
γ ^ portfolio selection literature. This type of transac
� kb ~ μ
μ ~k1 kΔwk1 þ k� �k1 kΔwk21 tion cost is usually modelled when large trades are
2
(2:10) made and the market price is impacted by such
pffiffi ^ 1=2 trades. More specifically, the asset prices will be
Furthermore, let us define again X ¼ γ� impacted when there is a large trade, such that
^ 1=2b
and y ¼ 1γ � ~, optimizing Equation (2.7) yields:
μ the investor will place her transaction with a less
� � favourable price. The linear costs in Section 2.1,
ΔwLinear ¼ argmin ky XΔwk22 þ λkΔwk11 ; which are usually used to measure the costs of
Δw
small trades, are not capable of capturing such
(2:11)
price impact.
where λ ¼ 2τ � 0. The optimal trading amount is However, it is difficult to fully characterize the
the solution to the foregoing penalized least impact of large trades on stocks since many factors,
squares with an L1 norm regularization (LASSO), such as the stock market value, the price, and trad
which is proposed by Tibshirani (1996) and has ing time (see Foster and Viswanathan 1993; Engle,
been widely used in statistics and economics, due Ferstenberg, and Russell 2006; Gârleanu and
to its favourable performance in variable selection Pedersen 2013) can cause the drift of stock prices.
and robustness against estimation error. Hence, In addition, institutional investors usually split lar
incorporating the proportional transaction costs ger orders into smaller ones to reduce the costs
automatically imposes a sparsity constraint on the caused by market impact. For this reason, a large
decision variable Δw. number of existing studies in the broader literature
However, unlike the traditional Lasso, the para have focused on the trade-off between the market
meter λ in the penalty term is related to the pro impact and opportunity cost to minimize the total
portional transaction cost parameter τ. More cost (for example, volume-weighted-average-price
specifically, when there are no transaction costs strategy in Berkowitz, Logue, and Noser 1988).
(τ ¼ 0), λ ¼ 0, given that the investor has no Others have also focused on deriving the explicit
need to pay for trading amount Δwi ; i ¼ 1; . . . ; N; form of the transaction costs in the presence of the
when there are linear transaction costs (τ�0), Δw� market impact. For instance, Kyle (1985) proposes
are shrunk towards zero in the sense that the assets to use a linear function of the trading amount to
receiving a small trading amount may be excluded characterize the price impact induced by large
from the trading strategy. trades. Consequently, it implies that the total
In other words, a sparse trading strategy is costs incurred by the investor are in a quadratic
desired for an investor who incurs transaction form:
APPLIED ECONOMICS 1441
� �
cit ¼ βi ðwi;t wi;t Þ2 (2:12) Δwquadratic ¼ argmin ky XΔwk22 þ λ2 kΔwk22 ;
Δw
or 0.6 other than 1.0 as considered in Kyle (1985), stable. More importantly, it characterizes more
since large orders are often split to reduce the precisely the impact of large trades on the market
market price impact. For instance, Moro et al. prices. The following proposition further shows
(2009) study the case of Madrid stock exchange that our proposed model is related to a regression
and London stock exchange, and find that p � 0:5 with an elastic net regularization (see Zou and
and p � 0:7, respectively. Toth et al. (2011) also Hastie 2005), but with a more general form:
find a value of smaller than 1 for the future market.
Consequently, if we use a function to measure such Proposition 4 Consider a portfolio rebalancing
p
impact, the resulting form would be approximately: model with total costs c ¼ τ k Δw k11 þκ k Δw kp
in Equation (2.19), which is equivalent to the opti
cit ¼ κi jwi;t wi;t jp (2:16) mization problem:
where κi is a general transaction cost parameter � p�
and p 2 ð1; 2Þ is used to measure the level of mar wgeneral ¼ argmin Py XΔwP22 þ λ1 PΔwP11 þ λp PΔwPp ;
w
ket price impact. (2:20)
For simplification, let us consider the case where κi
p
is identical for each asset: c ¼ κ k Δw kp . So the where λ1 ¼ 2τand λp ¼ 2κ. If we further define
portfolio rebalancing model in the presence of such λp
αp ¼ λ1 þλ p
and λ ¼ λ1 þ λp , then:
costs is:
n γ T^ o Δw� ¼ argmin�Py XΔwP22 þ λ� 1 αp �PΔwP11 þ αp PΔwPpp ��:
Tb p
Δwgeneral ¼ argmax Δw μ ~ Δw �Δw κkΔwkp : Δw
w 2 (2:21)
(2:17)
Our proposed model enjoys several advan
Analogously, model Equation (2.17) can be writ tages compared to the linear model with elastic
ten into the following penalized linear model: net regularization. On the one hand, our model
� � framework is more general in the sense that, the
Δwgeneral ¼ argmin ky XΔwk22 þ λp kΔwkpp ;
w market impact parameter p that controls the
(2:18) stability can provide a strategy that performs at
pffiffi ^ 1=2 least as well as the one from the elastic net, if we
where λp ¼ 2κ, X ¼ γ� ^ 1=2b
and y ¼ 1γ � ~.
μ allow the parameter p to be obtained through
Consequently, imposing general transaction costs cross validation. On the other hand, our model
is equivalent to considering a bridge regression allows us to incorporate simultaneously the pro
with 1 < p < 2. It results in a stable trading strategy, portional transaction costs and the general mar
similar to the case with quadratic transaction costs. ket impact costs, although the parameter p is
However, when an investor is managing a large now given exogenously.
portfolio with many assets, a sparse trading strategy Also note that when we consider only the
is also desired. general costs, the consequent model can be writ
To build such a strategy, a natural way is to ten as a penalized least-squares regression with
incorporate a variable selection regularization into Lp regularization, resulting in a stable but not
the model. Also note that, besides the market sparse trading strategy; when both general costs
impact costs, an investor also incurs linear costs. and linear costs are incorporated, the model is
Consequently, incorporating both the linear costs equivalent to a least square regression with both
and market impact costs into the portfolio rebalan Lp and Lp regularization. And due to the exis
cing model yields:
tence of a Lp norm penalty, our model also gives
n γ T^ o
�
Δw ¼ argmax Δw μ Tb
~ 1
Δw �Δw τkΔwk1 κkΔwkp : p a sparse trading strategy, which is desired for an
2
Δw
investor who manages a large number of assets,
(2:19)
thus providing better out-of-sample perfor
Model Equation (2.19) then provides an investor mance compared to the model that ignores
with a trading strategy that is both sparse and transaction costs.
APPLIED ECONOMICS 1443
III. Empirical analysis is repeated by adding the risky asset returns of the
next period to the dataset and dropping the earliest
In this section, we study empirically the out-of-
sample performance of our proposed model in return, until the end of the dataset is reached.
the presence of various transaction costs and com Using this rolling window procedure, we are able
pare it with that of the models without considering to obtain T t out-of-sample returns. We then
the transaction costs before the transaction based compare the out-of-sample performance of the
on a real dataset. More specifically, we use the strategies through the Sharpe ratio computed in
constituent stocks of CSI 300 as our asset pool, the following way:
which represents the largest stocks of China’s �p
μ
A-share market. We consider weekly returns of SR ¼ ; (3:1)
σp
these assets ranging from January 2009 to
April 2019. After removing the stocks whose IPO
1 X
T 1
dates are after 2009 and those that have been sus �p ¼
μ ðrp;tþ1 rf Þ; (3:2)
pended for more than half a year, there are 161 T s t¼s
assets remaining in our asset pool. For simplicity,
we consider the case with N ¼ 100, including the 1 X
T 1
For the setting of transaction cost parameters, leads to an extremely unstable trading strategy,
note that it is usually the case that commission fee thus resulting in overtrading in the out-of-sample
for both sellers and buyers ranges from 0.02% to period. Not surprisingly, the GMV strategy pro
0.3% in China’s stock market, depending on the vides the lowest out-of-sample portfolio variance,
bargaining power of brokers. And transfer fee (only although it leads to negative average return after
in Shanghai stock exchange) also exists and is deducting the transaction cost. Moreover, the ICF
charged bi-directionally at a level of around model we propose results in the lowest average
0.002%. Moreover, stamp duty is charged in both number of assets in the transactions, due to the
the Shanghai and Shenzhen stock exchange at existence of the cost term which is equivalent to
a rate of 0.1% (only on the seller’s side). an L1 norm penalty. Furthermore, our model
Consequently, we consider a level of 300 basis incurs the least total costs, followed by the N1 strat
points (τ ¼ 0:3%) for the model with proportional egy, which is also considered to be a simple but
transaction costs in the base case for the purpose of stable trading strategy (see DeMiguel, Garlappi,
simplification. To obtain a proper value of the and Uppal 2009).
quadratic transaction cost parameter, we look into Figure 1 shows the cumulative return net of
the price impact index, which denotes the percen linear cost for different portfolio rules. In the pre
tage of price change when the turnover is sence of proportional transaction costs, our pro
0.1 million CNY. According to The Performance posed model dominates the other strategies, with
of Shanghai Stock Exchange (2018) and The a return of 38.86% ranging from January 2016 to
Performance of Shenzhen Stock Exchange (2017), April 2019, while that of the CSI 300 index
the price impact index is about 0.1%. Note that β is 4.88%.
is related to the portfolio size M, we thus have As mentioned in Section 2.1, the benefits of
β ¼ 1=2 � ðM � 0:10%Þ=ð10 � 104 Þ ¼ 5 � 10 9 M incorporating transaction costs in the portfolio
. Equivalently, β ¼ 0:5 with 0.1 billion CNY, and rebalancing model are twofold: (i) after impos
1:0 with 0.2 billion CNY. Finally, for the case with ing transaction costs, the portfolio rebalancing
general transaction costs, the cost function is con model can be rewritten into a least square pro
p
sidered to be c ¼ κ k Δw kp , where p 2 ð1; 2Þ is the blem with an L1 penalty, thus mitigating the
market impact costs parameter, and κ is related to impact of risk in the sample estimators to
the liquidity level of the asset as well as the size of some extent; (ii) with the form of proportional
the portfolio. transaction costs, the sparsity in the trading
amount Δw reduces greatly the turnover at
each period, thus providing higher average
Results return and Sharpe ratios.
Proportional transaction costs
Table 1 reports the out-of-sample performance of Quadratic transaction costs
our proposed model with proportional transaction When we consider large transactions and linear
costs against other portfolio rules. In terms of the market impacts, quadratic transaction costs should
Sharpe ratio, the model incorporating transaction be imposed in the rebalancing model. Note that the
costs outperforms the other strategies, both before quadratic costs parameter β is related to the port
and after deducting linear costs. The main reason is folio size. Consequently, to show the advantage of
that, for the MV portfolio rebalancing model and our proposed model, we list the out-of-sample
the GMV model, the error in the sample estimators performance of different portfolio rules based on
different values of β. For example, when β ¼ 0:5, it
Table 1. Out-Of-Sample performance with linear costs. indicates that when the weight is adjusted by 1%,
Portfolio rule SR Rτ Variance Rebalance TC the return rate net of transaction cost will decrease
ICF 0.0540 0.0529 1.0000 1.5740 0.0090 by 0.005%. Table 2 shows the out-of-sample per
MV −0.0506 −0.2306 3.5205 100.0000 6.0959
GMV 0.0262 −0.0316 0.3190 100.0000 0.1752
formance of different strategies when β ¼ 0:5
1/N 0.0369 0.0340 0.4493 99.2367 0.0127 and β ¼ 1:0.
APPLIED ECONOMICS 1445
For the case of β ¼ 0:5, the N1 strategy over quadratic transaction costs with β ¼ 1.
whelms the others in terms of Sharpe ratio and Figure 2(a) shows that the ICF model generates
portfolio variance, as well as total transaction the highest cumulative return of over 40%, fol
costs incurred in the out-of-sample period. Note lowed by a 21% return from the N1 strategy. With
that the portfolio rebalancing model incorporating
an equal-weight initial portfolio, each rebalancing
quadratic transaction costs is equivalent to a ridge
regression form about Δw, thus it provides a linear strategy incurs large transaction costs in the first
shrinkage on Δ based on the mean-variance model. period, while incorporating transaction costs in the
The shrinkage intensity depends on the quadratic model can partly alleviate such high costs. More
transaction cost parameter β: with a small β, the specifically, our proposed model shows that, the
shrinkage intensity is small as well, thus leading to optimal allocation for the first week is:
a trading strategy that is close to Markowitz’s � �
γ T^ 1 2
mean-variance strategy; with a higher β, our ICF Δw� ¼ argmax ΔwT b
~μ Δw �Δw βPw P (3:7)
Δw;�T w¼1 2 N 2
model with quadratic transaction costs outper
forms the others in terms of the Sharpe ratio, due when β ! 1,
to a more stable trading strategy derived from � �
a stronger shrinkage intensity. � 1
w ¼ argmax βPw P
We also show graphically the cumulative return Δw;�T w¼1 N
� �
of our proposed model against the other strategies, 1 1
¼ argmax Pw P ¼ lN�1 (3:8)
including the CSI300 index, after charging Δw;�T w¼1 N N
1446 X. MEI ET AL.
That is, as the value of β increases, the opti cost still outperforms the others, while that of
mal portfolio with quadratic cost will eventually the global minimum variance remains the
converge to the N1 portfolio, while the GMV lowest.
portfolio rebalancing policy costs much more
than the other strategies in the first period, due General transaction costs
to the absence of considering transaction costs. To characterize the market price impact more
To mitigate the impact of the initial position, we properly, the general transaction costs are also con
also show the cumulative return that excludes sidered in our model. For the cost form, c ¼ κ k
p
the first period. As it has been shown in Δw kp with p 2 ð1; 2Þ, κ is usually related to spe
Figure 2b, the model that considers quadratic cific asset liquidity and portfolio size, and p is used
APPLIED ECONOMICS 1447
Table 3. Out-Of-Sample performance with general costs. both before and after deducting transaction costs.
Portfolio SR Rτ Variance Rebalance TC
Besides, the portfolio variance of our model
p = 1.4 ICF 0.0450 0.0446 1.0000 83.3669 0.0020
MV −0.0506 −0.4531 64.9490 100.0000 132.64 declines rapidly compared to the mean-variance
GMV 0.0262 −0.1952 0.9236 100.0000 0.9016 portfolio and is close to that of the global mini
1/N 0.0369 0.0325 0.9422 99.2367 0.0191
p = 1.5 ICF 0.0594 0.0584 1.0000 94.4260 0.0051 mum variance strategy and the N1 strategy. More
MV −0.0506 −0.3992 58.2715 100.0000 123.7416
GMV 0.0262 −0.1323 0.7191 100.0000 0.5916
importantly, the total cost of the proposed model
1/N 0.0369 0.0348 0.7975 99.2367 0.0090 is the lowest and the trading amount Δw is sparse,
p = 1.6 ICF 0.0622 0.0610 1.0000 98.4320 0.0081
MV −0.0506 −0.3550 49.7963 100.0000 117.1169 due to the presence of both general transaction
GMV 0.0262 −0.0873 0.5414 100.0000 0.3965 cost that controls stability, and the linear transac
1/N 0.0369 0.0360 0.6397 99.2367 0.0042
tion costs that regulate the sparsity. Note that
these two properties are important to build an
to model the extent to which transactions move the efficient portfolio, especially when the number of
price of the assets. Table 3 shows the out-of-sample risky assets is large: when the amount of adjust
performance of our proposed model against the ment in each period is sparse, the investment
others for several values of p. manager may pay more attention to the stocks
It shows that the out-of-sample performance of with larger adjustments and ignore the assets
the model that considers general costs is much with little adjustment. This not only reduces the
better than that of other portfolios in terms of the transaction costs and management fees but also
Sharpe ratio net of costs. The variance and total increases the stability of the portfolio. Moreover,
transaction cost are far smaller than those of the the penalty term combining L1 with Lp can deal
MV model for different p, which confirms that the with problems that have multicollinearity issue,
imposing p-norm penalization can help to improve and meanwhile avoid excessive sparsity like
the stability of the model. Note that the number of LASSO2
active assets in our model is close to N, implying Analogously, Figure 4 shows the cumulative
return net of costs of our proposed model and
a non-sparse but stable trading strategy, as shown
other strategies, including CSI300 index. It
in Section 2.3.
shows that the model incorporating transaction
After taking into account both the linear
costs dominates the CSI300 and the other port
costs and market impact costs, the investor
folio models, by reaching 23.90% after deducting
now incurs total transaction costs:
p 1
transaction costs, while that of the CSI300 is
cðθ; ΔwÞ ¼ k k Δw kp þτ k Δw k1 . In this sec
only 4.88%.
tion, we propose to consider a portfolio reba
We also study the performance of our pro
lancing model incorporating both parts.
posed model depending on different pairs of
Table 4 shows the out-of-sample performance
cost parameters. More specifically, we would
of each rebalancing policy, where the linear
like to investigate how these parameters affect
cost parameter τ ¼ 0:003, and general cost
the sparsity and out-of-sample Sharpe ratios.
parameters κ ¼ 0:1 and p ¼ 1:5.
Intuitively, for a fixed market impact costs para
In this case, our model still outperforms the
meter κ, a higher τ means a stronger penalization
other strategies in terms of different performance
on the L1 norm, thus leading to a more sparse
evaluation criteria: it has the highest Sharpe ratio,
trading strategy Δw. Figure 5a further confirms
our discussion. In addition, for a fixed value of τ,
Table 4. Out-Of-Sample performance with general costs and an increasing κ indicates αp ! 1, thus
linear costs. a dominating weight on Lp norm penalization.
Portfolio SR Rτ Variance Rebalance TC Consequently, a first increase and then a stable
ICF 0.0418 0.0414 1.0000 20.2899 0.0016
MV −0.0506 −0.4093 70.6400 100.0000 129.8355
average number of active assets are observed.
GMV 0.0262 −0.1718 0.8807 100.0000 0.0767 To demonstrate the superior performance of
1/N 0.0369 0.0319 0.9453 99.2367 0.0216
the proposed model, Figure 5b graphs the
2
The average active asset is only 1.57 for the case with linear costs, while that of the case with both rises to 20.29..
1448 X. MEI ET AL.
Sharpe ratios of our model against N1 strategy consider other empirical settings, including differ
based on various pairs of transaction cost para ent datasets, window lengths, and adding con
meters.Not surprisingly, our model always out straints to the model.
performs the N1 strategy, after accounting for
both types of transaction costs.
Randomly selected dataset
Table 6. Out-Of-Sample performance with shortsell constraint. In addition, the N1 strategy no longer beats
Portfolio SR Rτ Variance Nonzero Rebalance TC the others, as the instability in the first three
Window Length=2 years, with shortsell constrain
ICF 0.0621 0.0610 1.0000 42.0282 19.3200 0.0178 models can be reduced by adding short-selling
MV 0.0573 −0.0974 1.1675 5.6612 6.4447 3.0777 constraint. And note that our proposed model
GMV 0.0855 −0.0050 0.5574 38.1224 40.9953 0.0855
1/N 0.0537 0.0439 0.7380 100.0000 99.6118 0.0608 now dominates in terms of out-of-sample
Sharpe ratios. This can be explained in the
sense that, since the sparse portfolio is being
produced by adding a short-selling constraint,
out-of-sample performance based on various cri
the transaction costs can thus be reduced
teria of different models. As the window length
through lower turnover. Furthermore, as has
decreases, the instability in the rebalancing strategy
been shown from the variance, adding a short-
caused by a larger error in the estimation of μ and
selling constraint prevents extreme negative
� aggravates, resulting in worse out-of-sample per
positions; therefore, it increases the stability of
formance in all models, excepting the N1 strategy
the rebalancing strategy, which is consistent
that does not require estimation. Consequently, the
1 with Jagannathan and Ma (2003).
N strategy outperforms the others when the win
dow length is 2 years. While when the estimation
window length increases to 6 years, our proposed V. Conclusion
model with transaction costs beats the others, indi
cating that a longer window helps to embody The classical Markowitz mean-variance is not
sufficient information required for parameter esti applicable in practice, usually due to the lack of
mation, and to further improve the performance of consideration of important factors, such as para
the models that rely on estimation. meter uncertainty and transaction costs. While
parameter uncertainty is mainly caused by the esti
mation error in model inputs μ and �, and transac
tion costs have different forms depending on the
Adding constraint
size of the portfolio. In this article, we investigate
In previous empirical examples, we do not consider the roles of various transaction costs in portfolio
any portfolio constraints that confine the investor’s rebalancing problems, and show the advantage of
trading strategy. In this section, we would like to imposing transaction costs in dealing with the
show how the trading strategies are affected by add impact of parameter uncertainty and producing
ing a short-selling constraint. Table 6 lists the out-of a rebalancing strategy with desired properties.
-sample performance, with a window length of 2 By rewriting the classical mean-variance model
years. The idea of considering such a short window into a least square regression problem, we find
length is to amplify the impact of adding such con that considering proportional transaction costs is
straints during rebalancing. It shows that the perfor equivalent to adding an L1 norm penalization,
mance of the mean-variance model and the global which helps to produce a sparse rebalancing strat
minimum variance model is significantly improved, egy. When market impact costs are considered,
as the short-selling constraint helps to prevent the portfolio rebalancing model can be trans
extreme weights. As it has been pointed out in formed into a least square regression with differ
DeMiguel et al. (2009), adding the short-selling con ent norm penalties, depending on how we
straint results in a similar shrinking effect when characterize the impact of the transactions on
considering the L1 constraint, since negative weights the asset prices. We further show that these penal
would be shrunk to zero, which generates sparse ties are crucial for an investor who wants to ease
portfolios. For instance, the mean-variance model the effect of estimation error. Finally, empirical
only invests 5.66 stocks on average, while that of the analysis and robustness check show the superior
global minimum variance model is 38.12, far less out-of-sample performance of our proposed
than the one before a short-selling constraint. model, and the model that ignores transaction
APPLIED ECONOMICS 1451
costs may suffer a large loss due to uncertainty in DeMiguel, V., L. Garlappi, F. J. Nogales, and R. Uppal. 2009.
model parameters. “A Generalized Approach to Portfolio Optimization:
Improving Performance by Constraining Portfolio
Norms.” Management Science 55 (5): 798–812.
Acknowledgements doi:10.1287/mnsc.1080.0986.
DeMiguel, V., L. Garlappi, and R. Uppal. 2009. “Optimal
We are grateful to the editor and the anonymous referee for versus Naive Diversification: How Inefficient is the 1/n
their constructive comments.
Portfolio Strategy?” The Review of Financial Studies
22 (5): 1915–1953. doi:10.1093/rfs/hhm075.
Disclosure statement DeMiguel, V., A. Martn-Utrera, and F. J. Nogales. 2015.
“Parameter Uncertainty in Multiperiod Portfolio
No potential conflict of interest was reported by the author(s). Optimization with Transaction Costs.” Journal of
Financial and Quantitative Analysis 50 (6): 1443–1471.
doi:10.1017/S002210901500054X.
Funding Engle, R. F., R. Ferstenberg, and J. R. Russell. 2006.
“Measuring and Modeling Execution Cost and Risk.”
The work was supported by the National Natural Science
Journal of Portfolio Management 38 (2): 14–28.
Foundation of China [71901186,71903166,71988101];
Fan, J., J. Zhang, and K. Yu. 2012. “Vast Portfolio Selection
Natural Science Foundation of Fujian Province
with Gross-Exposure Constraints.” Journal of the American
[2019J05013]; the Ministry of Education in China (MOE)
Statistical Association 107 (498): 592–606. doi:10.1080/
Project [19YJC790206]
01621459.2012.682825.
Foster, F. D., and S. Viswanathan. 1993. “Variations in
References Trading Volume, Return Volatility, and Trading Costs:
Evidence on Recent Price Formation Models.” The
Almgren, R., C. Thum, E. Hauptmann, and H. Li. 2005. Journal of Finance 48 (1): 187–211. doi:10.1111/j.1540-
“Direct Estimation of Equity Market Impact.” Risk 18 (7): 6261.1993.tb04706.x.
58–62. Garlappi, L., R. Uppal, and T. Wang. 2007. “Portfolio
Atsushi, Y. 1996. “The Mean-Variance Approach to Portfolio Selection with Parameter and Model Uncertainty: A
Optimization Subject to Transaction Costs.” Journal of the Multi-Prior Approach.” The Review of Financial Studies
Operations Research Society of Japan 39 (1): 99–117. 20 (1): 41–81. doi:10.1093/rfs/hhl003.
doi:10.15807/jorsj.39.99. Gârleanu, N., and L. H. Pedersen. 2013. “Dynamic Trading with
Ban, G.-Y., N. El Karoui, and A. E. Lim. 2018. “Machine Predictable Returns and Transaction Costs.” The Journal of
Learning and Portfolio Optimization.” Management Finance 68 (6): 2309–2340. doi:10.1111/jofi.12080.
Science 64 (3): 1136–1154. doi:10.1287/mnsc.2016.2644. Goldfarb, D., and G. Iyengar. 2003. “Robust Portfolio
Barry, C. B. 1974. “Portfolio Analysis Under Uncertain Means, Selection Problems.” Mathematics of Operations Research
Variances, and Covariances.” The Journal of Finance 29 (2): 28 (1): 1–38. doi:10.1287/moor.28.1.1.14260.
515–522. doi:10.1111/j.1540-6261.1974.tb03064.x. Gotoh, J. Y., and A. Takeda. 2011. “On the Role of Norm
Berkowitz, S. A., D. E. Logue, and E. A. Noser. 1988. “The Constraints in Portfolio Selection.” Computational
Total Cost of Transaction on the Nyse.” The Journal of Management Science 8 (4): 323–353. doi:10.1007/s10287-011-
Finance 43 (1): 97–112. doi:10.1111/j.1540-6261.1988. 0130-2.
tb02591.x. Jagannathan, R., and T. Ma. 2003. “Risk Reduction in Large
Best, M. J., and R. R. Grauer. 1991. “On the Sensitivity of Portfolios: Why Imposing the Wrong Constraints Helps.
Mean-Variance-Efficient Portfolios to Changes in Asset The Journal of Finance .” 58 (4): 1651–1684.
Means: Some Analytical and Computational Results.” The Jorion, P. 1985. “‘International Portfolio Diversification with
Review of Financial Studies 4 (2): 315–342. doi:10.1093/rfs/ Estimation Risk.” Journal of Business 58: 259–278.
4.2.315. doi:10.1086/296296.
Constantinides, G. M. 1986. “Capital Market Equilibrium Jorion, P. 1986. “Bayes-Stein Estimation for Portfolio
with Transaction Costs.” The Journal of Political Economy Analysis.” Journal of Financial and Quantitative Analysis
94 (4): 842–862. doi:10.1086/261410. 21 (3): 279–292. doi:10.2307/2331042.
Dai, M., Z. Q. Xu, and X. Y. Zhou. 2010. “Continuous-Time Jorion, P. 1991. “Bayesian and CAPM Estimators of the Means:
Markowitz’s Model with Transaction Costs.” SIAM Journal Implications for Portfolio Selection.” Journal of Banking and
on Financial Mathematics 1 (1): 96–125. doi:10.1137/ Finance 15 (3): 717–727. doi:10.1016/0378-4266(91)90094-3.
080742889. Kan, R., and G. Zhou. 2007. “Optimal Portfolio Choice with
Davis, M. H., and A. R. Norman. 1990. “Portfolio Selection Parameter Uncertainty.” Journal of Financial and
with Transaction Costs.” Mathematics of Operations Quantitative Analysis 42 (3): 621–656. doi:10.1017/
Research 15 (4): 676–713. doi:10.1287/moor.15.4.676. S0022109000004129.
1452 X. MEI ET AL.
Kyle, A. S. 1985. “Continuous Auctions and Insider Trading.” Pástor, L. 2000. “Portfolio Selection and Asset Pricing
Econometrica 53 (6): 1315–1335. doi:10.2307/1913210. Models.” The Journal of Finance 55 (1): 179–223.
Ledoit, O., and M. Wolf. 2003. “Improved Estimation of the doi:10.1111/0022-1082.00204.
Covariance Matrix of Stock Returns with an Application to Patel, N. R., and M. G. Subrahmanyam. 1982. “A Simple
Portfolio Selection.” Journal of Empirical Finance 10 (5): Algorithm for Optimal Portfolio Selection with Fixed
603–621. doi:10.1016/S0927-5398(03)00007-0. Transaction Costs.” Management Science 28 (3): 303–314.
Ledoit, O., and M. Wolf. 2004. “Honey, I Shrunk the Sample doi:10.1287/mnsc.28.3.303.
Covariance Matrix.” Journal of Portfolio Management Pogue, G. A. 1970. “An Extension of the Markowitz Portfolio
30 (4): 110–119. doi:10.3905/jpm.2004.110. Selection Model to Include Variable Transactions’ Costs,
Li, J. 2015. “Sparse and Stable Portfolio Selection with Short Sales, Leverage Policies and Taxes.” The Journal of
Parameter Uncertainty.” Journal of Business and Economic Finance 25 (5): 1005–1027. doi:10.1111/j.1540-6261.1970.
Statistics 33 (3): 381–392. doi:10.1080/07350015.2014. tb00865.x.
954708. Rustem, B., R. G. Becker, and W. Marty. 2000. “Robust Min–
Liu, H. 2004. “Optimal Consumption and Investment with max Portfolio Strategies for Rival Forecast and Risk
Transaction Costs and Multiple Risky Assets.” The Journal Scenarios.” Journal of Economic Dynamics & Control
of Finance 59 (1): 289–338. doi:10.1111/j.1540- 24 (11–12): 1591–1621. doi:10.1016/S0165-1889(99)
6261.2004.00634.x. 00088-3.
Magill, M. J. P., and G. M. Constantinides. 1976. “Portfolio Tibshirani, R. 1996. “Regression Shrinkage and Selection via
Selection with Transactions Costs.” Journal of Economic the Lasso.” Journal of the Royal Statistical Society 58 (1):
Theory 13 (2): 245–263. doi:10.1016/0022-0531(76)90018-1. 267–288.
Markowitz, H. M. 1952. “Portfolio Selection.” The Journal of Toth, B., Y. Lemperiere, C. Deremble, J. Lataillade,
Finance 7 (1): 77. J. Kockelkoren, and J.-P. Bouchaud. 2011. “Anomalous
Mei, X., V. DeMiguel, and F. J. Nogales. 2016. “Multiperiod Price Impact and the Critical Nature of Liquidity in
Portfolio Optimization with Multiple Risky Assets and Financial Markets.” Physical Review X 1. doi:10.1103/
General Transaction Costs.” Journal of Banking & Finance PhysRevX.1.021006.
69: 108–120. doi:10.1016/j.jbankfin.2016.04.002. Wang, S., and M. Cai. 2011. “Measurement of Risk Aversion
Michaud, R. O. 1989. “The Markowitz Optimization Enigma: Coefficient of Chinese Residents and Analysis of
Is “Optimized” Optimal?” Financial Analysts Journal Influencing Factors: An Empirical Study Based on the
45 (1): 31–42. doi:10.2469/faj.v45.n1.31. Data of Investment Behavior of Chinese Residents.”
Moro, E., J. Vicente, L. G. Moyano, A. Gerig, J. D. Farmer, Journal of Financial Research 8: 192–206.
G. Vaglica, F. Lillo, and R. N. Mantegna. 2009. “Market Zou, H., and T. Hastie. 2005. “Regularization and Variable
Impact and Trading Profile of Hidden Orders in Stock Selection via the Elastic Net.” Journal of the Royal Statistical
Markets.” Physical Review E 80 (6 Pt 2): 066102. Society 67 (2): 301–320. doi:10.1111/j.1467-9868.20
doi:10.1103/PhysRevE.80.066102. 05.00503.x.
APPLIED ECONOMICS 1453
� � 1
Appendix 1 ^ 2β ^�^ 1b
¼ �þ I � ~
μ
γ γ